Split (1)

train · 4.04k rows

messages listlengths 3 3
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\n[Music]\n\n \n\nHello everyone. Thanks for coming to to\n\nHello everyone. Thanks for coming to to\n \n\nHello everyone. Thanks for coming to to\nour talk here. Um, just briefly\n\nour talk here. Um, just briefly\n \n\nour talk here. Um, just briefly\nintroduce myself. I'm Toki. I lead the\n\nintroduce myself. I'm To..." }, { "role": "assistant", "content": "[Music]\n\n \n\nHello everyone. Thanks for coming to to\n\nHello everyone. Thanks for coming to to\n \n\nHello everyone. Thanks for coming to to\nour talk here. Um, just briefly\n\nour talk here. Um, just briefly\n \n\nour talk here. Um, just briefly\nintroduce myself. I'm Toki. I lead the\n\nintroduce myself. I'm Toki. I lead the\n \n\nintroduce myself. I'm Toki. I lead the\nsolution architecture team uh in the US\n\nsolution architecture team uh in the US\n \n\nsolution architecture team uh in the US\nfor OpenAI and I have Venoop here as\n\nfor OpenAI and I have Venoop here as\n \n\nfor OpenAI and I have Venoop here as\nwell on the solution architecture team. well on the solution architecture team. well on the solution architecture team. Uh, and we'll talk about building\n\nUh, and we'll talk about building\n \n\nUh, and we'll talk about building\npractical audio agents. Um, before I\n\npractical audio agents. Um, before I\n \n\npractical audio agents. Um, before I\njump in, maybe a quick raise of hands. jump in, maybe a quick raise of hands. jump in, maybe a quick raise of hands. Who has either like built an audio agent\n\nWho has either like built an audio agent\n \n\nWho has either like built an audio agent\nor voice agent of some kind or thinking\n\nor voice agent of some kind or thinking\n \n\nor voice agent of some kind or thinking\nabout building a audio agent? about building a audio agent? about building a audio agent? Okay, so like halfish of the room. Okay,\n\nOkay, so like halfish of the room. Okay,\n \n\nOkay, so like halfish of the room. Okay,\ngood. It's always good to know what\n\ngood. It's always good to know what\n \n\ngood. It's always good to know what\npeople's interests are. So yeah, we'll\n\npeople's interests are. So yeah, we'll\n \n\npeople's interests are. So yeah, we'll\ntalk a bit about like um some of the\n\ntalk a bit about like um some of the\n \n\ntalk a bit about like um some of the\nthings we we've learned, where audio\n\nthings we we've learned, where audio\n \n\nthings we we've learned, where audio\nmodels are today, and then some of the\n\nmodels are today, and then some of the\n \n\nmodels are today, and then some of the\nkind of patterns and best practices. So\n\nkind of patterns and best practices. So\n \n\nkind of patterns and best practices. So\njust to take a quick step back, um,\n\njust to take a quick step back, um,\n \n\njust to take a quick step back, um,\nGenai applications broadly to date, as\n\nGenai applications broadly to date, as\n \n\nGenai applications broadly to date, as\nmany of you all know, have been very\n\nmany of you all know, have been very\n \n\nmany of you all know, have been very\ntextbased, right? This is uh models that\n\ntextbased, right? This is uh models that\n \n\ntextbased, right? This is uh models that\nreally deeply understand uh and generate\n\nreally deeply understand uh and generate\n \n\nreally deeply understand uh and generate\ntext with very much humanlike fluency\n\ntext with very much humanlike fluency\n \n\ntext with very much humanlike fluency\nand there's a lot of production\n\nand there's a lot of production\n \n\nand there's a lot of production\napplications out there built on text and\n\napplications out there built on text and\n \n\napplications out there built on text and\nthat's continues to scale. I think we've\n\nthat's continues to scale. I think we've\n \n\nthat's continues to scale. I think we've\nreally started to enter a bit more of\n\nreally started to enter a bit more of\n \n\nreally started to enter a bit more of\nthis multimodal era where we now have\n\nthis multimodal era where we now have\n \n\nthis multimodal era where we now have\nmodels that can understand images\n\nmodels that can understand images\n \n\nmodels that can understand images\nextremely well. There's now models that\n\nextremely well. There's now models that\n \n\nextremely well. There's now models that\ncan generate images really well if you\n\ncan generate images really well if you\n \n\ncan generate images really well if you\nwere part of the image gen Studio Gibbly\n\nwere part of the image gen Studio Gibbly\n \n\nwere part of the image gen Studio Gibbly\ncraze uh or Sora that generates videos\n\ncraze uh or Sora that generates videos\n \n\ncraze uh or Sora that generates videos\nfor example. Um and then now also audio. for example. Um and then now also audio. for example. Um and then now also audio. These are models. We've had models like\n\nThese are models. We've had models like\n \n\nThese are models. We've had models like\ntranscription models or texttospech\n\ntranscription models or texttospech\n \n\ntranscription models or texttospech\nmodels for some time, but now this kind\n\nmodels for some time, but now this kind\n \n\nmodels for some time, but now this kind\nof emerging speechtoech capability is\n\nof emerging speechtoech capability is\n \n\nof emerging speechtoech capability is\ngetting a lot better. And I would say\n\ngetting a lot better. And I would say\n \n\ngetting a lot better. And I would say\nnow we we will see I think a much more\n\nnow we we will see I think a much more\n \n\nnow we we will see I think a much more\nof an explosion of use cases around\n\nof an explosion of use cases around\n \n\nof an explosion of use cases around\nthese modalities of images and video. these modalities of images and video. these modalities of images and video. And today we're really going to focus on\n\nAnd today we're really going to focus on\n \n\nAnd today we're really going to focus on\nthe audio modality uh across this. So I\n\nthe audio modality uh across this. So I\n \n\nthe audio modality uh across this. So I\nwould say like one important thing we're\n\nwould say like one important thing we're\n \n\nwould say like one important thing we're\nstarting to see is that audio is really\n\nstarting to see is that audio is really\n \n\nstarting to see is that audio is really\ngetting to that point where it's\n\ngetting to that point where it's\n \n\ngetting to that point where it's\nbecoming much more usable at scale. And\n\nbecoming much more usable at scale. And\n \n\nbecoming much more usable at scale. And\nagain, we've had the transcription and\n\nagain, we've had the transcription and\n \n\nagain, we've had the transcription and\nthe TTS side, but especially that\n\nthe TTS side, but especially that\n \n\nthe TTS side, but especially that\nspeech-to-pech low latency experience. speech-to-pech low latency experience. speech-to-pech low latency experience. The models are getting, I would say,\n\nThe models are getting, I would say,\n \n\nThe models are getting, I would say,\ngood enough and are kind of at this\n\ngood enough and are kind of at this\n \n\ngood enough and are kind of at this\ntipping point where you can start to\n\ntipping point where you can start to\n \n\ntipping point where you can start to\nreally meaningfully build high quality\n\nreally meaningfully build high quality\n \n\nreally meaningfully build high quality\nproduction applications at scale. So, I\n\nproduction applications at scale. So, I\n \n\nproduction applications at scale. So, I\nwould say before the models were a lot\n\nwould say before the models were a lot\n \n\nwould say before the models were a lot\nslower, robotic, and brittle, and now\n\nslower, robotic, and brittle, and now\n \n\nslower, robotic, and brittle, and now\nthey're a lot faster, expressive, and\n\nthey're a lot faster, expressive, and\n \n\nthey're a lot faster, expressive, and\naccurate. I'll show an example. The\n\naccurate. I'll show an example. The\n \n\naccurate. I'll show an example. The\nfirst one will show what the models were\n\nfirst one will show what the models were\n \n\nfirst one will show what the models were\nlike not even six months ago. It's a\n\nlike not even six months ago. It's a\n \n\nlike not even six months ago. It's a\nvery painful example, so I hope you feel\n\nvery painful example, so I hope you feel\n \n\nvery painful example, so I hope you feel\na bit of the pain uh in this. a bit of the pain uh in this. a bit of the pain uh in this. >> Hello. >> Hello. >> Hello. Hello. Please give me a topic for your\n\nHello. Please give me a topic for your\n \n\nHello. Please give me a topic for your\nhaik coup. haik coup. haik coup. >> Uh,\n\n \n\nwell, you can see the the initial delay. well, you can see the the initial delay. well, you can see the the initial delay. Um, that's actually honestly\n\nUm, that's actually honestly\n \n\nUm, that's actually honestly\n>> Could you write me a haiku about\n\n>> Could you write me a haiku about\n \n\n>> Could you write me a haiku about\nwhispers? >> Silent air, faint words. Drift through\n\n>> Silent air, faint words. Drift through\n \n\n>> Silent air, faint words. Drift through\ndarken rooms. Secrets linger near. Dust slides through my hands." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nSifted\n\nDust slides through my hands. Sifted\n \n\nDust slides through my hands. Sifted\ngrains forget the weight. What is left\n\ngrains forget the weight. What is left\n \n\ngrains forget the weight. What is left\nbehind? So, I hope you felt the pain a\n\nbehind? So, I hope you felt the pain a\n \n\nbehind? So, ..." }, { "role": "assistant", "content": "Sifted\n\nDust slides through my hands. Sifted\n \n\nDust slides through my hands. Sifted\ngrains forget the weight. What is left\n\ngrains forget the weight. What is left\n \n\ngrains forget the weight. What is left\nbehind? So, I hope you felt the pain a\n\nbehind? So, I hope you felt the pain a\n \n\nbehind? So, I hope you felt the pain a\nlittle bit. Right. So, these models are\n\nlittle bit. Right. So, these models are\n \n\nlittle bit. Right. So, these models are\nuh this is actually not just not even\n\nuh this is actually not just not even\n \n\nuh this is actually not just not even\nsix months ago, right? The models are\n\nsix months ago, right? The models are\n \n\nsix months ago, right? The models are\nincredibly slow. Um they're very robotic\n\nincredibly slow. Um they're very robotic\n \n\nincredibly slow. Um they're very robotic\nin how they sound. Uh and they're quite\n\nin how they sound. Uh and they're quite\n \n\nin how they sound. Uh and they're quite\nbrittle. So, in that example, we\n\nbrittle. So, in that example, we\n \n\nbrittle. So, in that example, we\nactually tried to interrupt the model\n\nactually tried to interrupt the model\n \n\nactually tried to interrupt the model\nand tell it no, don't write a don't\n\nand tell it no, don't write a don't\n \n\nand tell it no, don't write a don't\nwrite a poem about uh whispers. Write\n\nwrite a poem about uh whispers. Write\n \n\nwrite a poem about uh whispers. Write\none about stars. They wrote one about\n\none about stars. They wrote one about\n \n\none about stars. They wrote one about\nsand. I'm not sure where it got that\n\nsand. I'm not sure where it got that\n \n\nsand. I'm not sure where it got that\nfrom. So just showing you that's\n\nfrom. So just showing you that's\n \n\nfrom. So just showing you that's\nactually where it was. And then now if\n\nactually where it was. And then now if\n \n\nactually where it was. And then now if\nyou like fast forward to today, these\n\nyou like fast forward to today, these\n \n\nyou like fast forward to today, these\nare the models that uh for a very\n\nare the models that uh for a very\n \n\nare the models that uh for a very\nsimilar use case kind of how they how\n\nsimilar use case kind of how they how\n \n\nsimilar use case kind of how they how\nwe'll behave today. we'll behave today. we'll behave today. >> Hello\n\n \n\n>> share with me a topic and I shall weave\n\n>> share with me a topic and I shall weave\n \n\n>> share with me a topic and I shall weave\na ha coup for you rich with emotion and\n\na ha coup for you rich with emotion and\n \n\na ha coup for you rich with emotion and\nimagery. What shall it be? imagery. What shall it be? imagery. What shall it be? >> Can you make a ha coup about whispers? >> Can you make a ha coup about whispers? >> Can you make a ha coup about whispers? >> Whispers? Actually, could you make a ha\n\n>> Whispers? Actually, could you make a ha\n \n\n>> Whispers? Actually, could you make a ha\ncoup about stars? coup about stars? coup about stars? >> Stars, whispers of light adorning the\n\n>> Stars, whispers of light adorning the\n \n\n>> Stars, whispers of light adorning the\nvelvet night,\n\nvelvet night,\n \n\nvelvet night,\neternal secrets. eternal secrets. eternal secrets. >> So, you can see it's much faster. It's\n\n>> So, you can see it's much faster. It's\n \n\n>> So, you can see it's much faster. It's\nincredibly expressive, emotional. You\n\nincredibly expressive, emotional. You\n \n\nincredibly expressive, emotional. You\ncan steer it, you can get it excited,\n\ncan steer it, you can get it excited,\n \n\ncan steer it, you can get it excited,\nyou can slow it down, and it's accurate. you can slow it down, and it's accurate. you can slow it down, and it's accurate. We interrupted it. We had it change\n\nWe interrupted it. We had it change\n \n\nWe interrupted it. We had it change\ncourse and do something else, and it\n\ncourse and do something else, and it\n \n\ncourse and do something else, and it\nresponded well. Um, so the models are\n\nresponded well. Um, so the models are\n \n\nresponded well. Um, so the models are\nreally, we do believe, at that good\n\nreally, we do believe, at that good\n \n\nreally, we do believe, at that good\nenough tipping point where you can start\n\nenough tipping point where you can start\n \n\nenough tipping point where you can start\nto build much more reliable experiences\n\nto build much more reliable experiences\n \n\nto build much more reliable experiences\naround it. Um, in terms of actually\n\naround it. Um, in terms of actually\n \n\naround it. Um, in terms of actually\nbuilding on top of the uh, audio models,\n\nbuilding on top of the uh, audio models,\n \n\nbuilding on top of the uh, audio models,\nthere's really two emerging\n\nthere's really two emerging\n \n\nthere's really two emerging\narchitectures that we see. One is kind\n\narchitectures that we see. One is kind\n \n\narchitectures that we see. One is kind\nof the one that's been around for a bit\n\nof the one that's been around for a bit\n \n\nof the one that's been around for a bit\nand then I'd say a little bit of the\n\nand then I'd say a little bit of the\n \n\nand then I'd say a little bit of the\nnewer one, which I'll talk about. The\n\nnewer one, which I'll talk about. The\n \n\nnewer one, which I'll talk about. The\nexisting one has been what we call kind\n\nexisting one has been what we call kind\n \n\nexisting one has been what we call kind\nof the chained approach. This is\n\nof the chained approach. This is\n \n\nof the chained approach. This is\nstitching together these three\n\nstitching together these three\n \n\nstitching together these three\napproaches. You have audio come in, you\n\napproaches. You have audio come in, you\n \n\napproaches. You have audio come in, you\ntranscribe it, you have some sort of\n\ntranscribe it, you have some sort of\n \n\ntranscribe it, you have some sort of\nintelligence layer with an LLM that does\n\nintelligence layer with an LLM that does\n \n\nintelligence layer with an LLM that does\nsome sort of intelligence, and then\n\nsome sort of intelligence, and then\n \n\nsome sort of intelligence, and then\nfinally text to speech on the output. finally text to speech on the output. finally text to speech on the output. Um, but again, this is stitching\n\nUm, but again, this is stitching\n \n\nUm, but again, this is stitching\ntogether three models. There's a lot of\n\ntogether three models. There's a lot of\n \n\ntogether three models. There's a lot of\nproblems with this. slow across time\n\nproblems with this. slow across time\n \n\nproblems with this. slow across time\nlike it takes longer time to actually\n\nlike it takes longer time to actually\n \n\nlike it takes longer time to actually\ngenerate your output audio. Um it also\n\ngenerate your output audio. Um it also\n \n\ngenerate your output audio. Um it also\nhas some lossiness across this right you\n\nhas some lossiness across this right you\n \n\nhas some lossiness across this right you\nkind of lose a little bit of the\n\nkind of lose a little bit of the\n \n\nkind of lose a little bit of the\nsemantic understanding of the\n\nsemantic understanding of the\n \n\nsemantic understanding of the\nconversation. So really this emerging\n\nconversation. So really this emerging\n \n\nconversation. So really this emerging\npattern is more this speechtoech\n\npattern is more this speechtoech\n \n\npattern is more this speechtoech\narchitecture where it really it takes\n\narchitecture where it really it takes\n \n\narchitecture where it really it takes\nthese three models into a um single\n\nthese three models into a um single\n \n\nthese three models into a um single\nmodel which for us it's the real-time\n\nmodel which for us it's the real-time\n \n\nmodel which for us it's the real-time\nAPI uh and it does all these layers it\n\nAPI uh and it does all these layers it\n \n\nAPI uh and it does all these layers it\ndoes the transcription it does that\n\ndoes the transcription it does that\n \n\ndoes the transcription it does that\nintelligence layer and then finally does\n\nintelligence layer and then finally does\n \n\nintelligence layer and then finally does\nthe speech as the output. So this\n\nthe speech as the output. So this\n \n\nthe speech as the output. So this\nsimplifies the architecture quite\n\nsimplifies the architecture quite\n \n\nsimplifies the architecture quite\ndramatically. Really reduces the time it\n\ndramatically. Really reduces the time it\n \n\ndramatically. Really reduces the time it\ntakes to output. This is really good for\n\ntakes to output. This is really good for\n \n\ntakes to output. This is really good for\nlow latency experiences. Um and it\n\nlow latency experiences. Um and it\n \n\nlow latency experiences." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nUm and it\ndoesn't have that lossiness problem. You\n\ndoesn't have that lossiness problem. You\n \n\ndoesn't have that lossiness problem. You\nactually are able to maintain semantic\n\nactually are able to maintain semantic\n \n\nactually are able to maintain semantic\nunderstanding across a conversation. So\n\nu..." }, { "role": "assistant", "content": "Um and it\ndoesn't have that lossiness problem. You\n\ndoesn't have that lossiness problem. You\n \n\ndoesn't have that lossiness problem. You\nactually are able to maintain semantic\n\nactually are able to maintain semantic\n \n\nactually are able to maintain semantic\nunderstanding across a conversation. So\n\nunderstanding across a conversation. So\n \n\nunderstanding across a conversation. So\nthis is really kind of the emerging\n\nthis is really kind of the emerging\n \n\nthis is really kind of the emerging\npattern we see. pattern we see. pattern we see. And when building there's some kind of\n\nAnd when building there's some kind of\n \n\nAnd when building there's some kind of\nkey trade-offs and considerations to\n\nkey trade-offs and considerations to\n \n\nkey trade-offs and considerations to\ntake into account um when building on\n\ntake into account um when building on\n \n\ntake into account um when building on\ntops of these architectures. And we\n\ntops of these architectures. And we\n \n\ntops of these architectures. And we\nthink it's kind of across these five\n\nthink it's kind of across these five\n \n\nthink it's kind of across these five\nmain areas. You have latency, you have\n\nmain areas. You have latency, you have\n \n\nmain areas. You have latency, you have\ncost, you have accuracy and\n\ncost, you have accuracy and\n \n\ncost, you have accuracy and\nintelligence, user experience and\n\nintelligence, user experience and\n \n\nintelligence, user experience and\nintegrations and tooling. And depending\n\nintegrations and tooling. And depending\n \n\nintegrations and tooling. And depending\nwhat you're building, your trade-offs\n\nwhat you're building, your trade-offs\n \n\nwhat you're building, your trade-offs\nwill be different. Right? So if for\n\nwill be different. Right? So if for\n \n\nwill be different. Right? So if for\nexample you're talking about a consumerf\n\nexample you're talking about a consumerf\n \n\nexample you're talking about a consumerf\nfacing application, um the trade-offs\n\nfacing application, um the trade-offs\n \n\nfacing application, um the trade-offs\nare quite different. Here you care a lot\n\nare quite different. Here you care a lot\n \n\nare quite different. Here you care a lot\nabout user experience. Um these are\n\nabout user experience. Um these are\n \n\nabout user experience. Um these are\nusually a low latency experience that um\n\nusually a low latency experience that um\n \n\nusually a low latency experience that um\na lot of end users are are interacting\n\na lot of end users are are interacting\n \n\na lot of end users are are interacting\nwith the model. So user experience\n\nwith the model. So user experience\n \n\nwith the model. So user experience\nmatters a lot. Latency matters a lot. It\n\nmatters a lot. Latency matters a lot. It\n \n\nmatters a lot. Latency matters a lot. It\ncan't be slow. It needs to be really\n\ncan't be slow. It needs to be really\n \n\ncan't be slow. It needs to be really\nfast and attractive. But things like\n\nfast and attractive. But things like\n \n\nfast and attractive. But things like\ncost probably doesn't matter as much. It\n\ncost probably doesn't matter as much. It\n \n\ncost probably doesn't matter as much. It\ndoesn't need to be accurate. Not really,\n\ndoesn't need to be accurate. Not really,\n \n\ndoesn't need to be accurate. Not really,\nright? It's more about expressiveness\n\nright? It's more about expressiveness\n \n\nright? It's more about expressiveness\nand quick answers. Um, and integrations\n\nand quick answers. Um, and integrations\n \n\nand quick answers. Um, and integrations\nand tooling like you don't really need\n\nand tooling like you don't really need\n \n\nand tooling like you don't really need\nto integrate with uh SIP or like Twilio\n\nto integrate with uh SIP or like Twilio\n \n\nto integrate with uh SIP or like Twilio\nfor for a broader experience. It's\n\nfor for a broader experience. It's\n \n\nfor for a broader experience. It's\npretty simple um uh interaction. So\n\npretty simple um uh interaction. So\n \n\npretty simple um uh interaction. So\nthings like this would really the\n\nthings like this would really the\n \n\nthings like this would really the\nreal-time API works really well for a\n\nreal-time API works really well for a\n \n\nreal-time API works really well for a\nconsumerf facing application depending\n\nconsumerf facing application depending\n \n\nconsumerf facing application depending\non these trade-offs. Now if you go to\n\non these trade-offs. Now if you go to\n \n\non these trade-offs. Now if you go to\ncustomer service as another example, the\n\ncustomer service as another example, the\n \n\ncustomer service as another example, the\ntrade-offs are different, right? So here\n\ntrade-offs are different, right? So here\n \n\ntrade-offs are different, right? So here\naccuracy intelligence actually probably\n\naccuracy intelligence actually probably\n \n\naccuracy intelligence actually probably\nmatters the most. you can't get an order\n\nmatters the most. you can't get an order\n \n\nmatters the most. you can't get an order\nnumber wrong. If it uh goes to delete an\n\nnumber wrong. If it uh goes to delete an\n \n\nnumber wrong. If it uh goes to delete an\norder instead of updating an order,\n\norder instead of updating an order,\n \n\norder instead of updating an order,\nthat's a big deal. So, accuracy and\n\nthat's a big deal. So, accuracy and\n \n\nthat's a big deal. So, accuracy and\nintelligence matters probably the most. intelligence matters probably the most. intelligence matters probably the most. Integrations and tooling start to matter\n\nIntegrations and tooling start to matter\n \n\nIntegrations and tooling start to matter\na lot more, too, because now you have to\n\na lot more, too, because now you have to\n \n\na lot more, too, because now you have to\nintegrate your with your various\n\nintegrate your with your various\n \n\nintegrate your with your various\ninternal systems. You have to integrate\n\ninternal systems. You have to integrate\n \n\ninternal systems. You have to integrate\nwith, yeah, SIP and phone providers if\n\nwith, yeah, SIP and phone providers if\n \n\nwith, yeah, SIP and phone providers if\nneeded. Um, so those are probably the\n\nneeded. Um, so those are probably the\n \n\nneeded. Um, so those are probably the\nmost important things, but user\n\nmost important things, but user\n \n\nmost important things, but user\nexperience still matters, but you'll\n\nexperience still matters, but you'll\n \n\nexperience still matters, but you'll\nprobably take a hit on that a little bit\n\nprobably take a hit on that a little bit\n \n\nprobably take a hit on that a little bit\nto make sure it's more accurate. Again,\n\nto make sure it's more accurate. Again,\n \n\nto make sure it's more accurate. Again,\nlatency still matters, but not as much\n\nlatency still matters, but not as much\n \n\nlatency still matters, but not as much\nbecause accuracy intelligence matters\n\nbecause accuracy intelligence matters\n \n\nbecause accuracy intelligence matters\nmore and cost um you're saving a large\n\nmore and cost um you're saving a large\n \n\nmore and cost um you're saving a large\ncost for this anyway. So, I think these\n\ncost for this anyway. So, I think these\n \n\ncost for this anyway. So, I think these\nare just the other considerations to um\n\nare just the other considerations to um\n \n\nare just the other considerations to um\nto take into account for customer\n\nto take into account for customer\n \n\nto take into account for customer\nservice. So, for this real-time API\n\nservice. So, for this real-time API\n \n\nservice. So, for this real-time API\nmight make sense for uh a situation\n\nmight make sense for uh a situation\n \n\nmight make sense for uh a situation\nwhere latency might matter the most, but\n\nwhere latency might matter the most, but\n \n\nwhere latency might matter the most, but\nif you care about determinism and\n\nif you care about determinism and\n \n\nif you care about determinism and\naccuracy and intelligence, then actually\n\naccuracy and intelligence, then actually\n \n\naccuracy and intelligence, then actually\nthe chained architecture would make a\n\nthe chained architecture would make a\n \n\nthe chained architecture would make a\nlot of sense here, too. Um but it's just\n\nlot of sense here, too. Um but it's just\n \n\nlot of sense here, too. Um but it's just\nimportant to consider what your\n\nimportant to consider what your\n \n\nimportant to consider what your\ntrade-offs are depending on what you're\n\ntrade-offs are depending on what you're\n \n\ntrade-offs are depending on what you're\ntrying to build. So hopefully that gives\n\ntrying to build. So hopefully that gives\n \n\ntrying to build. So hopefully that gives\nyou paints you a picture a bit of like\n\nyou paints you a picture a bit of like\n \n\nyou paints you a picture a bit of like\nwhat to consider. I'll hand it off to\n\nwhat to consider. I'll hand it off to\n \n\nwhat to consider. I'll hand it off to\nNoob to talk about very specific\n\nNoob to talk about very specific\n \n\nNoob to talk about very specific\npatterns we've seen when working with\n\npatterns we've seen when working with\n \n\npatterns we've seen when working with\ncustomers building on this. customers building on this. customers building on this. >> Cool. Awesome Tobi. Um so me just like\n\n>> Cool. Awesome Tobi. Um so me just like\n \n\n>> Cool. Awesome Tobi." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nUm so me just like\ntaking a step back. I think we've\n\ntaking a step back. I think we've\n \n\ntaking a step back. I think we've\nprobably defined agent a bunch of times\n\nprobably defined agent a bunch of times\n \n\nprobably defined agent a bunch of times\nbut here's just like a a canonical\n\nbut here's just..." }, { "role": "assistant", "content": "Um so me just like\ntaking a step back. I think we've\n\ntaking a step back. I think we've\n \n\ntaking a step back. I think we've\nprobably defined agent a bunch of times\n\nprobably defined agent a bunch of times\n \n\nprobably defined agent a bunch of times\nbut here's just like a a canonical\n\nbut here's just like a a canonical\n \n\nbut here's just like a a canonical\ndefinition to hopefully keep everyone\n\ndefinition to hopefully keep everyone\n \n\ndefinition to hopefully keep everyone\ngrounded for this session. Um think of\n\ngrounded for this session. Um think of\n \n\ngrounded for this session. Um think of\nan agent as some model, some set of\n\nan agent as some model, some set of\n \n\nan agent as some model, some set of\ninstructions. So your prompts, the tools\n\ninstructions. So your prompts, the tools\n \n\ninstructions. So your prompts, the tools\nyou give your model and then the\n\nyou give your model and then the\n \n\nyou give your model and then the\nruntime. So like the guardrails like how\n\nruntime. So like the guardrails like how\n \n\nruntime. So like the guardrails like how\nyou do your execution environment. Um,\n\nyou do your execution environment. Um,\n \n\nyou do your execution environment. Um,\nand generally like when you're building\n\nand generally like when you're building\n \n\nand generally like when you're building\nvoice agents, that's all the same things\n\nvoice agents, that's all the same things\n \n\nvoice agents, that's all the same things\nyou should be worrying about, but\n\nyou should be worrying about, but\n \n\nyou should be worrying about, but\nthere's also a few other things that are\n\nthere's also a few other things that are\n \n\nthere's also a few other things that are\nmaybe a bit different that we're going\n\nmaybe a bit different that we're going\n \n\nmaybe a bit different that we're going\nto focus on today. Uh, so the main stuff\n\nto focus on today. Uh, so the main stuff\n \n\nto focus on today. Uh, so the main stuff\nis like your system design is probably a\n\nis like your system design is probably a\n \n\nis like your system design is probably a\nbit different than a traditional LM\n\nbit different than a traditional LM\n \n\nbit different than a traditional LM\nagent. Um, it's probably a bit different\n\nagent. Um, it's probably a bit different\n \n\nagent. Um, it's probably a bit different\non the prompting side and customization\n\non the prompting side and customization\n \n\non the prompting side and customization\nof the voice. Um, the tools you're using\n\nof the voice. Um, the tools you're using\n \n\nof the voice. Um, the tools you're using\nprobably a bit different. Um, and then\n\nprobably a bit different. Um, and then\n \n\nprobably a bit different. Um, and then\nalso eval guardrails you might have to\n\nalso eval guardrails you might have to\n \n\nalso eval guardrails you might have to\nthink about um, as well. So these are\n\nthink about um, as well. So these are\n \n\nthink about um, as well. So these are\nthe things we'll we'll probably focus on\n\nthe things we'll we'll probably focus on\n \n\nthe things we'll we'll probably focus on\nthat are like distinct when you're\n\nthat are like distinct when you're\n \n\nthat are like distinct when you're\nbuilding specifically voice agents\n\nbuilding specifically voice agents\n \n\nbuilding specifically voice agents\ncompared to like traditional agents. compared to like traditional agents. compared to like traditional agents. Cool. So if you've all built uh maybe\n\nCool. So if you've all built uh maybe\n \n\nCool. So if you've all built uh maybe\nlike a textbased agent, let's use like\n\nlike a textbased agent, let's use like\n \n\nlike a textbased agent, let's use like\ncustomer service as a conical example. customer service as a conical example. customer service as a conical example. Um you have like some user question. You\n\nUm you have like some user question. You\n \n\nUm you have like some user question. You\nhave some sort of triage agent that it's\n\nhave some sort of triage agent that it's\n \n\nhave some sort of triage agent that it's\nlike a small model that goes to other\n\nlike a small model that goes to other\n \n\nlike a small model that goes to other\nmodels here that will do like the harder\n\nmodels here that will do like the harder\n \n\nmodels here that will do like the harder\ntasks. So you have like an 03 to do like\n\ntasks. So you have like an 03 to do like\n \n\ntasks. So you have like an 03 to do like\nthe refund or maybe like an L4 mini to\n\nthe refund or maybe like an L4 mini to\n \n\nthe refund or maybe like an L4 mini to\ndo some of the cancellation. Um, and\n\ndo some of the cancellation. Um, and\n \n\ndo some of the cancellation. Um, and\ntypically that's like how you build a a\n\ntypically that's like how you build a a\n \n\ntypically that's like how you build a a\nnormal agent today. Uh, for voice, one\n\nnormal agent today. Uh, for voice, one\n \n\nnormal agent today. Uh, for voice, one\npattern that we're starting to see more\n\npattern that we're starting to see more\n \n\npattern that we're starting to see more\nand more of is some sort of like\n\nand more of is some sort of like\n \n\nand more of is some sort of like\ndelegation through tools. So you would\n\ndelegation through tools. So you would\n \n\ndelegation through tools. So you would\nhave like the real-time API be your like\n\nhave like the real-time API be your like\n \n\nhave like the real-time API be your like\nfrontline agent. So it does a lot of the\n\nfrontline agent. So it does a lot of the\n \n\nfrontline agent. So it does a lot of the\nnormal interaction patterns that you're\n\nnormal interaction patterns that you're\n \n\nnormal interaction patterns that you're\nused to. So it can respond to like the\n\nused to. So it can respond to like the\n \n\nused to. So it can respond to like the\neasy questions that you sort of want\n\neasy questions that you sort of want\n \n\neasy questions that you sort of want\nyour model to answer and you know you're\n\nyour model to answer and you know you're\n \n\nyour model to answer and you know you're\nconfident in your model answering. Um,\n\nconfident in your model answering. Um,\n \n\nconfident in your model answering. Um,\nbut then there's also other questions\n\nbut then there's also other questions\n \n\nbut then there's also other questions\nthat might be a bit harder. Um, and then\n\nthat might be a bit harder. Um, and then\n \n\nthat might be a bit harder. Um, and then\nyou want your model to like do a tool\n\nyou want your model to like do a tool\n \n\nyou want your model to like do a tool\ncall to these other agents that are run\n\ncall to these other agents that are run\n \n\ncall to these other agents that are run\nby smarter models behind the hood. Um,\n\nby smarter models behind the hood. Um,\n \n\nby smarter models behind the hood. Um,\nso that's a pattern that we'll show a\n\nso that's a pattern that we'll show a\n \n\nso that's a pattern that we'll show a\nquick video of of like how this would\n\nquick video of of like how this would\n \n\nquick video of of like how this would\nwork. >> I've authenticated you now. What item\n\n>> I've authenticated you now. What item\n \n\n>> I've authenticated you now. What item\nwould you like to return? would you like to return? would you like to return? >> Yeah, I want to return the snowboard I\n\n>> Yeah, I want to return the snowboard I\n \n\n>> Yeah, I want to return the snowboard I\nbought. bought. bought. >> Got it. I'm reviewing the policy and\n\n>> Got it. I'm reviewing the policy and\n \n\n>> Got it. I'm reviewing the policy and\nwe'll get back to you with next steps\n\nwe'll get back to you with next steps\n \n\nwe'll get back to you with next steps\nshortly. So, here's a scenario where\n\nshortly. So, here's a scenario where\n \n\nshortly. So, here's a scenario where\nwe're actually using like a much smaller\n\nwe're actually using like a much smaller\n \n\nwe're actually using like a much smaller\nmodel04 mini uh and delegating to that\n\nmodel04 mini uh and delegating to that\n \n\nmodel04 mini uh and delegating to that\nbecause we don't want uh a smaller\n\nbecause we don't want uh a smaller\n \n\nbecause we don't want uh a smaller\nmodel. model. model. >> It looks like the return you are making\n\n>> It looks like the return you are making\n \n\n>> It looks like the return you are making\nis within the 30-day return policy. We\n\nis within the 30-day return policy. We\n \n\nis within the 30-day return policy." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nWe\nwill process the return and send you a\n\nwill process the return and send you a\n \n\nwill process the return and send you a\nconfirmation email with next steps\n\nconfirmation email with next steps\n \n\nconfirmation email with next steps\nshortly. >> So, that's like one pattern uh we've\n\n>> So, that's lik..." }, { "role": "assistant", "content": "We\nwill process the return and send you a\n\nwill process the return and send you a\n \n\nwill process the return and send you a\nconfirmation email with next steps\n\nconfirmation email with next steps\n \n\nconfirmation email with next steps\nshortly. >> So, that's like one pattern uh we've\n\n>> So, that's like one pattern uh we've\n \n\n>> So, that's like one pattern uh we've\nseen on the architecture side. Another\n\nseen on the architecture side. Another\n \n\nseen on the architecture side. Another\npoint I touched on is like the the way\n\npoint I touched on is like the the way\n \n\npoint I touched on is like the the way\nyou basically customize your brand and\n\nyou basically customize your brand and\n \n\nyou basically customize your brand and\nwhat you're sort of building. Uh so the\n\nwhat you're sort of building. Uh so the\n \n\nwhat you're sort of building. Uh so the\nprompts you're using is a huge part of\n\nprompts you're using is a huge part of\n \n\nprompts you're using is a huge part of\nthat. And the way to think about this is\n\nthat. And the way to think about this is\n \n\nthat. And the way to think about this is\nwhen you're prompting models today uh in\n\nwhen you're prompting models today uh in\n \n\nwhen you're prompting models today uh in\nthe textbased world, you only can\n\nthe textbased world, you only can\n \n\nthe textbased world, you only can\ncontrol like the instructions you're\n\ncontrol like the instructions you're\n \n\ncontrol like the instructions you're\ngiving it and sort of what you want the\n\ngiving it and sort of what you want the\n \n\ngiving it and sort of what you want the\nmodel to do. But in voice based\n\nmodel to do. But in voice based\n \n\nmodel to do. But in voice based\napplications, you can also control the\n\napplications, you can also control the\n \n\napplications, you can also control the\nexpressiveness of the voice and how you\n\nexpressiveness of the voice and how you\n \n\nexpressiveness of the voice and how you\nwant the model to speak or sound. Uh, so\n\nwant the model to speak or sound. Uh, so\n \n\nwant the model to speak or sound. Uh, so\nthis is like an example prompt here of,\n\nthis is like an example prompt here of,\n \n\nthis is like an example prompt here of,\nhey, you can control like the demeanor,\n\nhey, you can control like the demeanor,\n \n\nhey, you can control like the demeanor,\nthe tone, the level of enthusiasm, all\n\nthe tone, the level of enthusiasm, all\n \n\nthe tone, the level of enthusiasm, all\nthose things that don't really get\n\nthose things that don't really get\n \n\nthose things that don't really get\ncaptured in a in a text agent that you\n\ncaptured in a in a text agent that you\n \n\ncaptured in a in a text agent that you\ndo have much more control over in a\n\ndo have much more control over in a\n \n\ndo have much more control over in a\nvoice space agent. Um, here's just a fun website you can\n\nUm, here's just a fun website you can\n \n\nUm, here's just a fun website you can\nuse to play around with this open. Um,\n\nuse to play around with this open. Um,\n \n\nuse to play around with this open. Um,\nwhere you can play around with different\n\nwhere you can play around with different\n \n\nwhere you can play around with different\nvoices and also click through sample\n\nvoices and also click through sample\n \n\nvoices and also click through sample\nprompts that we've created. Uh this is\n\nprompts that we've created. Uh this is\n \n\nprompts that we've created. Uh this is\njust a great place to get started on\n\njust a great place to get started on\n \n\njust a great place to get started on\nlike actually playing around with the\n\nlike actually playing around with the\n \n\nlike actually playing around with the\nexpressiveness of voices and seeing that\n\nexpressiveness of voices and seeing that\n \n\nexpressiveness of voices and seeing that\nthere is a ceiling and like a lot of\n\nthere is a ceiling and like a lot of\n \n\nthere is a ceiling and like a lot of\nfluctuation you can get in performance\n\nfluctuation you can get in performance\n \n\nfluctuation you can get in performance\njust by the prompts you do use. just by the prompts you do use. just by the prompts you do use. Uh, another side of this is when you're\n\nUh, another side of this is when you're\n \n\nUh, another side of this is when you're\nwhen you're prompting models in the the\n\nwhen you're prompting models in the the\n \n\nwhen you're prompting models in the the\ntextbased approach, you'll probably have\n\ntextbased approach, you'll probably have\n \n\ntextbased approach, you'll probably have\nlike a lot of few shot examples of like,\n\nlike a lot of few shot examples of like,\n \n\nlike a lot of few shot examples of like,\nhey, I want you to do these few steps uh\n\nhey, I want you to do these few steps uh\n \n\nhey, I want you to do these few steps uh\nand then do these next few steps um and\n\nand then do these next few steps um and\n \n\nand then do these next few steps um and\ngo through like a typical uh\n\ngo through like a typical uh\n \n\ngo through like a typical uh\nconversation flow. Uh the same thing you\n\nconversation flow. Uh the same thing you\n \n\nconversation flow. Uh the same thing you\ncan sort of mimic in voice. So if you\n\ncan sort of mimic in voice. So if you\n \n\ncan sort of mimic in voice. So if you\nsee in in this picture uh there's like\n\nsee in in this picture uh there's like\n \n\nsee in in this picture uh there's like\nsome ID of hey I want you to do some\n\nsome ID of hey I want you to do some\n \n\nsome ID of hey I want you to do some\ngreeting, the description of it and then\n\ngreeting, the description of it and then\n \n\ngreeting, the description of it and then\nthe instructions for the users. Um and\n\nthe instructions for the users. Um and\n \n\nthe instructions for the users. Um and\nyou can just go through that. So this is\n\nyou can just go through that. So this is\n \n\nyou can just go through that. So this is\nlike a sample way of you to think\n\nlike a sample way of you to think\n \n\nlike a sample way of you to think\nthrough how you should be prompting\n\nthrough how you should be prompting\n \n\nthrough how you should be prompting\nmodels especially for more complex tasks\n\nmodels especially for more complex tasks\n \n\nmodels especially for more complex tasks\nthat you want the real-time API to\n\nthat you want the real-time API to\n \n\nthat you want the real-time API to\nhandle. handle. handle. Uh the next big bucket of things that\n\nUh the next big bucket of things that\n \n\nUh the next big bucket of things that\nare sort of important or maybe like a\n\nare sort of important or maybe like a\n \n\nare sort of important or maybe like a\nbit different for the voice agents\n\nbit different for the voice agents\n \n\nbit different for the voice agents\nyou're building is tools. Tool use in\n\nyou're building is tools. Tool use in\n \n\nyou're building is tools. Tool use in\ngeneral is like very uh important\n\ngeneral is like very uh important\n \n\ngeneral is like very uh important\nbecause you want your model to connect\n\nbecause you want your model to connect\n \n\nbecause you want your model to connect\nto other things that are sort of\n\nto other things that are sort of\n \n\nto other things that are sort of\nimportant for you. Um and the first step\n\nimportant for you. Um and the first step\n \n\nimportant for you. Um and the first step\nis honestly just start simple. you don't\n\nis honestly just start simple. you don't\n \n\nis honestly just start simple. you don't\nneed to connect 10 tools or like 20\n\nneed to connect 10 tools or like 20\n \n\nneed to connect 10 tools or like 20\ntools to a given model. Uh keep the\n\ntools to a given model. Uh keep the\n \n\ntools to a given model. Uh keep the\nnumber of tool calls pretty limited for\n\nnumber of tool calls pretty limited for\n \n\nnumber of tool calls pretty limited for\na given agent. Uh and that maybe goes\n\na given agent. Uh and that maybe goes\n \n\na given agent. Uh and that maybe goes\nback to the first diagram we showed of\n\nback to the first diagram we showed of\n \n\nback to the first diagram we showed of\nyou want to delegate to different models\n\nyou want to delegate to different models\n \n\nyou want to delegate to different models\nor agents to do different tasks." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nSo here\n\nor agents to do different tasks. So here\n \n\nor agents to do different tasks. So here\na given agent don't give it a ton of\n\na given agent don't give it a ton of\n \n\na given agent don't give it a ton of\ntools to start with and then slowly add\n\ntools to start with and then slowly add\n \n\ntools t..." }, { "role": "assistant", "content": "So here\n\nor agents to do different tasks. So here\n \n\nor agents to do different tasks. So here\na given agent don't give it a ton of\n\na given agent don't give it a ton of\n \n\na given agent don't give it a ton of\ntools to start with and then slowly add\n\ntools to start with and then slowly add\n \n\ntools to start with and then slowly add\non more tools as needed. Um handoffs. So\n\non more tools as needed. Um handoffs. So\n \n\non more tools as needed. Um handoffs. So\nif you ever you uh use the agents SDK,\n\nif you ever you uh use the agents SDK,\n \n\nif you ever you uh use the agents SDK,\nthis is a concept of if we have a given\n\nthis is a concept of if we have a given\n \n\nthis is a concept of if we have a given\nmodel um you want to pass it to another\n\nmodel um you want to pass it to another\n \n\nmodel um you want to pass it to another\num and here you can manage context\n\num and here you can manage context\n \n\num and here you can manage context\nbetween handoffs as well. So the thing\n\nbetween handoffs as well. So the thing\n \n\nbetween handoffs as well. So the thing\nto remember here is you might have a\n\nto remember here is you might have a\n \n\nto remember here is you might have a\ngiven agent and you're going to another\n\ngiven agent and you're going to another\n \n\ngiven agent and you're going to another\none but you want to keep the context the\n\none but you want to keep the context the\n \n\none but you want to keep the context the\nsame between them. Um so a good pattern\n\nsame between them. Um so a good pattern\n \n\nsame between them. Um so a good pattern\nto do is actually summarizing the\n\nto do is actually summarizing the\n \n\nto do is actually summarizing the\nconversation state and then passing it\n\nconversation state and then passing it\n \n\nconversation state and then passing it\nto the other agents so you don't lose\n\nto the other agents so you don't lose\n \n\nto the other agents so you don't lose\nany of the the context and it's not\n\nany of the the context and it's not\n \n\nany of the the context and it's not\nlossy. Uh a final one is delegation. So\n\nlossy. Uh a final one is delegation. So\n \n\nlossy. Uh a final one is delegation. So\nthe concept we showed a bit earlier of\n\nthe concept we showed a bit earlier of\n \n\nthe concept we showed a bit earlier of\nlike using 03 mini or 04 mini to do the\n\nlike using 03 mini or 04 mini to do the\n \n\nlike using 03 mini or 04 mini to do the\ntool call that's like natively built\n\ntool call that's like natively built\n \n\ntool call that's like natively built\ninto the agents SDK if you'd like to try\n\ninto the agents SDK if you'd like to try\n \n\ninto the agents SDK if you'd like to try\nit out. it out. it out. >> Um evals. So I think this is a pretty\n\n>> Um evals. So I think this is a pretty\n \n\n>> Um evals. So I think this is a pretty\nbig part of the like voice experience\n\nbig part of the like voice experience\n \n\nbig part of the like voice experience\nand maybe like breaking it down into\n\nand maybe like breaking it down into\n \n\nand maybe like breaking it down into\nlike four key buckets. I think the first\n\nlike four key buckets. I think the first\n \n\nlike four key buckets. I think the first\nand most obvious thing is always start\n\nand most obvious thing is always start\n \n\nand most obvious thing is always start\nwith observability. um you're not going\n\nwith observability. um you're not going\n \n\nwith observability. um you're not going\nto get very far if you don't know how\n\nto get very far if you don't know how\n \n\nto get very far if you don't know how\nyour data looks. Um so just make sure\n\nyour data looks. Um so just make sure\n \n\nyour data looks. Um so just make sure\nyou have traces of what's going on with\n\nyou have traces of what's going on with\n \n\nyou have traces of what's going on with\nyour agents uh the audio and everything\n\nyour agents uh the audio and everything\n \n\nyour agents uh the audio and everything\nso you can actually look at it and take\n\nso you can actually look at it and take\n \n\nso you can actually look at it and take\nactions. Um and the next thing is when\n\nactions. Um and the next thing is when\n \n\nactions. Um and the next thing is when\nyou're actually starting to do evals\n\nyou're actually starting to do evals\n \n\nyou're actually starting to do evals\nlike the most important thing is\n\nlike the most important thing is\n \n\nlike the most important thing is\nactually havememes to label these data\n\nactually havememes to label these data\n \n\nactually havememes to label these data\num and actually go and iterate on\n\num and actually go and iterate on\n \n\num and actually go and iterate on\nprompts from there. It sounds maybe a\n\nprompts from there. It sounds maybe a\n \n\nprompts from there. It sounds maybe a\nbit too easy or it's like not the most\n\nbit too easy or it's like not the most\n \n\nbit too easy or it's like not the most\nappropriate or uh maybe the best\n\nappropriate or uh maybe the best\n \n\nappropriate or uh maybe the best\napproach but it is the most effective\n\napproach but it is the most effective\n \n\napproach but it is the most effective\nthat we've seen when working with\n\nthat we've seen when working with\n \n\nthat we've seen when working with\ncustomers. Um it lets you go a lot\n\ncustomers. Um it lets you go a lot\n \n\ncustomers. Um it lets you go a lot\nfaster. Uh and then from there the next\n\nfaster. Uh and then from there the next\n \n\nfaster. Uh and then from there the next\nstep is transcription based evals. So\n\nstep is transcription based evals. So\n \n\nstep is transcription based evals. So\nthis is where you would have your\n\nthis is where you would have your\n \n\nthis is where you would have your\ntraditional as a judge evals or like\n\ntraditional as a judge evals or like\n \n\ntraditional as a judge evals or like\ntesting your function calling\n\ntesting your function calling\n \n\ntesting your function calling\nperformance on certain business criteria\n\nperformance on certain business criteria\n \n\nperformance on certain business criteria\nand having a rubric in place. Um, and\n\nand having a rubric in place. Um, and\n \n\nand having a rubric in place. Um, and\nthen there's this next bucket, uh, where\n\nthen there's this next bucket, uh, where\n \n\nthen there's this next bucket, uh, where\na lot of people ask us, how do we\n\na lot of people ask us, how do we\n \n\na lot of people ask us, how do we\nactually test the audio that's\n\nactually test the audio that's\n \n\nactually test the audio that's\ngenerated? Um, and for the most part,\n\ngenerated? Um, and for the most part,\n \n\ngenerated? Um, and for the most part,\nit's there's not a lot of things that\n\nit's there's not a lot of things that\n \n\nit's there's not a lot of things that\naudio uh, evals get you that maybe text\n\naudio uh, evals get you that maybe text\n \n\naudio uh, evals get you that maybe text\nevals don't, but the things that maybe\n\nevals don't, but the things that maybe\n \n\nevals don't, but the things that maybe\nit does get you is like understanding\n\nit does get you is like understanding\n \n\nit does get you is like understanding\nthe tone, understanding the pacing, or\n\nthe tone, understanding the pacing, or\n \n\nthe tone, understanding the pacing, or\nother things that are maybe a bit harder\n\nother things that are maybe a bit harder\n \n\nother things that are maybe a bit harder\nto capture just through text. And this\n\nto capture just through text. And this\n \n\nto capture just through text. And this\nis where you can use like a completions\n\nis where you can use like a completions\n \n\nis where you can use like a completions\nmodel or some other audio model. Uh so\n\nmodel or some other audio model. Uh so\n \n\nmodel or some other audio model. Uh so\nfor example GT4 audio to like understand\n\nfor example GT4 audio to like understand\n \n\nfor example GT4 audio to like understand\nthe tone, the pacing and what other\n\nthe tone, the pacing and what other\n \n\nthe tone, the pacing and what other\nintonations that you care about for your\n\nintonations that you care about for your\n \n\nintonations that you care about for your\nbusiness." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nUm and then the final one that\n\nbusiness. Um and then the final one that\n \n\nbusiness. Um and then the final one that\nwe've started to see a bit more is\n\nwe've started to see a bit more is\n \n\nwe've started to see a bit more is\nactually synthetic conversations. So for\n\nactually synthetic conversations..." }, { "role": "assistant", "content": "Um and then the final one that\n\nbusiness. Um and then the final one that\n \n\nbusiness. Um and then the final one that\nwe've started to see a bit more is\n\nwe've started to see a bit more is\n \n\nwe've started to see a bit more is\nactually synthetic conversations. So for\n\nactually synthetic conversations. So for\n \n\nactually synthetic conversations. So for\nexample, maybe you have two real-time\n\nexample, maybe you have two real-time\n \n\nexample, maybe you have two real-time\nclients going back and forth across to\n\nclients going back and forth across to\n \n\nclients going back and forth across to\neach other. So maybe one is your\n\neach other. So maybe one is your\n \n\neach other. So maybe one is your\nreal-time API agent or whatever and then\n\nreal-time API agent or whatever and then\n \n\nreal-time API agent or whatever and then\nthe other is a set of customer personas\n\nthe other is a set of customer personas\n \n\nthe other is a set of customer personas\nthat you have and you can simulate like\n\nthat you have and you can simulate like\n \n\nthat you have and you can simulate like\nmany many of those conversations and\n\nmany many of those conversations and\n \n\nmany many of those conversations and\nthen extract those evals across maybe\n\nthen extract those evals across maybe\n \n\nthen extract those evals across maybe\nlike transcriptions or audio and then\n\nlike transcriptions or audio and then\n \n\nlike transcriptions or audio and then\nhave a much better way to understand\n\nhave a much better way to understand\n \n\nhave a much better way to understand\nwhat you've sort of built. what you've sort of built. what you've sort of built. Cool. Uh and then the final thing here\n\nCool. Uh and then the final thing here\n \n\nCool. Uh and then the final thing here\nis guardrails. Um because especially for\n\nis guardrails. Um because especially for\n \n\nis guardrails. Um because especially for\nthe real-time API it's pretty fast. uh\n\nthe real-time API it's pretty fast. uh\n \n\nthe real-time API it's pretty fast. uh\nyou want to have guardrails that make\n\nyou want to have guardrails that make\n \n\nyou want to have guardrails that make\nsure that you have the safety that's\n\nsure that you have the safety that's\n \n\nsure that you have the safety that's\nneeded for uh to actually have\n\nneeded for uh to actually have\n \n\nneeded for uh to actually have\nconfidence in in the solution you're\n\nconfidence in in the solution you're\n \n\nconfidence in in the solution you're\nyou're rolling out to users. Um so the\n\nyou're rolling out to users. Um so the\n \n\nyou're rolling out to users. Um so the\nmain things we have here is like run\n\nmain things we have here is like run\n \n\nmain things we have here is like run\nthem async uh generally because the\n\nthem async uh generally because the\n \n\nthem async uh generally because the\ngenerated text from the real-time API is\n\ngenerated text from the real-time API is\n \n\ngenerated text from the real-time API is\nmuch faster than the spoken audio that\n\nmuch faster than the spoken audio that\n \n\nmuch faster than the spoken audio that\nyou get back from the model. So you have\n\nyou get back from the model. So you have\n \n\nyou get back from the model. So you have\nroom to play with the latency here. Uh\n\nroom to play with the latency here. Uh\n \n\nroom to play with the latency here. Uh\nthe next thing also is you have control\n\nthe next thing also is you have control\n \n\nthe next thing also is you have control\nto set like whatever debounce period you\n\nto set like whatever debounce period you\n \n\nto set like whatever debounce period you\nhave here. So in the code snippet uh to\n\nhave here. So in the code snippet uh to\n \n\nhave here. So in the code snippet uh to\nthe right hand side of the screen you'll\n\nthe right hand side of the screen you'll\n \n\nthe right hand side of the screen you'll\nsee that like right now it runs every\n\nsee that like right now it runs every\n \n\nsee that like right now it runs every\n100 characters to run the guardrails but\n\n100 characters to run the guardrails but\n \n\n100 characters to run the guardrails but\nyou can control that if you want to. you can control that if you want to. you can control that if you want to. Awesome. And then maybe just some\n\nAwesome. And then maybe just some\n \n\nAwesome. And then maybe just some\nlearnings that we've seen from other\n\nlearnings that we've seen from other\n \n\nlearnings that we've seen from other\nfolks in the field. Um Lemonade so uh AI\n\nfolks in the field. Um Lemonade so uh AI\n \n\nfolks in the field. Um Lemonade so uh AI\nlike insurance company they're like\n\nlike insurance company they're like\n \n\nlike insurance company they're like\nbuilding e one of the things we like\n\nbuilding e one of the things we like\n \n\nbuilding e one of the things we like\nnoticed from them that made them\n\nnoticed from them that made them\n \n\nnoticed from them that made them\nsuccessful is they like focused really\n\nsuccessful is they like focused really\n \n\nsuccessful is they like focused really\nearly on eval guardrails and like\n\nearly on eval guardrails and like\n \n\nearly on eval guardrails and like\nfeedback mechanisms for their team. And\n\nfeedback mechanisms for their team. And\n \n\nfeedback mechanisms for their team. And\neven if it wasn't scalable, they\n\neven if it wasn't scalable, they\n \n\neven if it wasn't scalable, they\nrealized that and it actually made them\n\nrealized that and it actually made them\n \n\nrealized that and it actually made them\nmove a lot faster in the end. Um, and\n\nmove a lot faster in the end. Um, and\n \n\nmove a lot faster in the end. Um, and\nthen finally for Tinder, um, they\n\nthen finally for Tinder, um, they\n \n\nthen finally for Tinder, um, they\nrealized that customization and brand\n\nrealized that customization and brand\n \n\nrealized that customization and brand\nrealism was pretty important for them\n\nrealism was pretty important for them\n \n\nrealism was pretty important for them\nwhen they were launching their RZ chat\n\nwhen they were launching their RZ chat\n \n\nwhen they were launching their RZ chat\nexperience. Um, so that's what they\n\nexperience. Um, so that's what they\n \n\nexperience. Um, so that's what they\nfocused on and it created a much better\n\nfocused on and it created a much better\n \n\nfocused on and it created a much better\nexperience for their users. experience for their users. experience for their users. And now I'll hand it off to Toki to\n\nAnd now I'll hand it off to Toki to\n \n\nAnd now I'll hand it off to Toki to\n>> Yeah, just to wrap us all up. So I kind\n\n>> Yeah, just to wrap us all up. So I kind\n \n\n>> Yeah, just to wrap us all up. So I kind\nof what I said in the beginning is we're\n\nof what I said in the beginning is we're\n \n\nof what I said in the beginning is we're\nreally entering this multimodal era. So\n\nreally entering this multimodal era. So\n \n\nreally entering this multimodal era. So\naway from or in addition to textbased\n\naway from or in addition to textbased\n \n\naway from or in addition to textbased\nmodels, we have now obviously um video\n\nmodels, we have now obviously um video\n \n\nmodels, we have now obviously um video\nand image models, but now audio. So very\n\nand image models, but now audio. So very\n \n\nand image models, but now audio. So very\nmuch in the multimodal era and this\n\nmuch in the multimodal era and this\n \n\nmuch in the multimodal era and this\nreal-time speech to speech technology is\n\nreal-time speech to speech technology is\n \n\nreal-time speech to speech technology is\nemerging. And I think we had a kind of\n\nemerging. And I think we had a kind of\n \n\nemerging. And I think we had a kind of\nlowkey update yesterday, a real time\n\nlowkey update yesterday, a real time\n \n\nlowkey update yesterday, a real time\nAPI. We had a new snapshot that we that\n\nAPI. We had a new snapshot that we that\n \n\nAPI. We had a new snapshot that we that\nwe released that's actually a lot\n\nwe released that's actually a lot\n \n\nwe released that's actually a lot\nbetter. So I really feel like now is the\n\nbetter. So I really feel like now is the\n \n\nbetter. So I really feel like now is the\ntime to to build. So you kind of have\n\ntime to to build. So you kind of have\n \n\ntime to to build. So you kind of have\nthis first mover's advantage to to build\n\nthis first mover's advantage to to build\n \n\nthis first mover's advantage to to build\nnow. I think the technology is getting\n\nnow." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nI think the technology is getting\n \n\nnow. I think the technology is getting\nto that point where it is good enough to\n\nto that point where it is good enough to\n \n\nto that point where it is good enough to\nbuild scalable production applications. build scalable production applications. build scalable pr..." }, { "role": "assistant", "content": "I think the technology is getting\n \n\nnow. I think the technology is getting\nto that point where it is good enough to\n\nto that point where it is good enough to\n \n\nto that point where it is good enough to\nbuild scalable production applications. build scalable production applications. build scalable production applications. Um, so really excited to see what you\n\nUm, so really excited to see what you\n \n\nUm, so really excited to see what you\nall build. Please test it out. Um, we'll\n\nall build. Please test it out. Um, we'll\n \n\nall build. Please test it out. Um, we'll\nbe around after the talk if you guys\n\nbe around after the talk if you guys\n \n\nbe around after the talk if you guys\nhave any questions, but thanks for the\n\nhave any questions, but thanks for the\n \n\nhave any questions, but thanks for the\ntime. [Music]" } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\n[Music]\n\n \n\nWelcome everyone. Welcome everyone. Welcome everyone. I'm going to talk about practical\n\nI'm going to talk about practical\n \n\nI'm going to talk about practical\ntactics to build uh reliable AI\n\ntactics to build uh reliable AI\n \n\ntactics to build uh reliable AI\napplications and why nobody ..." }, { "role": "assistant", "content": "[Music]\n\n \n\nWelcome everyone. Welcome everyone. Welcome everyone. I'm going to talk about practical\n\nI'm going to talk about practical\n \n\nI'm going to talk about practical\ntactics to build uh reliable AI\n\ntactics to build uh reliable AI\n \n\ntactics to build uh reliable AI\napplications and why nobody does it this\n\napplications and why nobody does it this\n \n\napplications and why nobody does it this\nway yet. way yet. way yet. Uh a little bit about myself or why you\n\nUh a little bit about myself or why you\n \n\nUh a little bit about myself or why you\nshould trust me. Uh I have allowed 15\n\nshould trust me. Uh I have allowed 15\n \n\nshould trust me. Uh I have allowed 15\nyears as a startup co-founder and CTO. years as a startup co-founder and CTO. years as a startup co-founder and CTO. Uh I held executive positions for the\n\nUh I held executive positions for the\n \n\nUh I held executive positions for the\nlast five years at uh several\n\nlast five years at uh several\n \n\nlast five years at uh several\nenterprises. uh but most importantly I\n\nenterprises. uh but most importantly I\n \n\nenterprises. uh but most importantly I\nspent last couple of years developing\n\nspent last couple of years developing\n \n\nspent last couple of years developing\na lot of gen projects ranging from PC's\n\na lot of gen projects ranging from PC's\n \n\na lot of gen projects ranging from PC's\nto uh many production level uh solutions\n\nto uh many production level uh solutions\n \n\nto uh many production level uh solutions\nand helped some companies to get it done\n\nand helped some companies to get it done\n \n\nand helped some companies to get it done\nand uh I've learned or distilled a way\n\nand uh I've learned or distilled a way\n \n\nand uh I've learned or distilled a way\nto uh make these applications reliable\n\nto uh make these applications reliable\n \n\nto uh make these applications reliable\nand there are quite a lot of uh tracks\n\nand there are quite a lot of uh tracks\n \n\nand there are quite a lot of uh tracks\nthis uh uh this conference about evals\n\nthis uh uh this conference about evals\n \n\nthis uh uh this conference about evals\nand reliability but uh to my surprise\n\nand reliability but uh to my surprise\n \n\nand reliability but uh to my surprise\nnobody was talking about the most\n\nnobody was talking about the most\n \n\nnobody was talking about the most\nimportant things and uh we're going to\n\nimportant things and uh we're going to\n \n\nimportant things and uh we're going to\ntalk about it right now. talk about it right now. talk about it right now. So uh standard software development life\n\nSo uh standard software development life\n \n\nSo uh standard software development life\ncycle is uh very standard uh simple uh\n\ncycle is uh very standard uh simple uh\n \n\ncycle is uh very standard uh simple uh\nyou design your solution, you develop\n\nyou design your solution, you develop\n \n\nyou design your solution, you develop\nit, you test it and then eventually you\n\nit, you test it and then eventually you\n \n\nit, you test it and then eventually you\ndeploy it. And uh when people start\n\ndeploy it. And uh when people start\n \n\ndeploy it. And uh when people start\ndoing uh PC with AI\n\ndoing uh PC with AI\n \n\ndoing uh PC with AI\nit sounds simple like you can very\n\nit sounds simple like you can very\n \n\nit sounds simple like you can very\neasily do you have some prompt and uh\n\neasily do you have some prompt and uh\n \n\neasily do you have some prompt and uh\nmodels are very capable but then you\n\nmodels are very capable but then you\n \n\nmodels are very capable but then you\nstart uh facing some uh unexpected\n\nstart uh facing some uh unexpected\n \n\nstart uh facing some uh unexpected\nchallenges\n\nchallenges\n \n\nchallenges\nuh actually like you can easily do a PC\n\nuh actually like you can easily do a PC\n \n\nuh actually like you can easily do a PC\nthat works 50% of the time uh but we're\n\nthat works 50% of the time uh but we're\n \n\nthat works 50% of the time uh but we're\nlike making it do the same reliable work\n\nlike making it do the same reliable work\n \n\nlike making it do the same reliable work\nthe Rest of the 50% is very hard uh\n\nthe Rest of the 50% is very hard uh\n \n\nthe Rest of the 50% is very hard uh\nbecause models are nondeterministic\n\nbecause models are nondeterministic\n \n\nbecause models are nondeterministic\nand uh it starts requiring uh a data\n\nand uh it starts requiring uh a data\n \n\nand uh it starts requiring uh a data\nscience approach uh continuous\n\nscience approach uh continuous\n \n\nscience approach uh continuous\nexperimentation you need to try this\n\nexperimentation you need to try this\n \n\nexperimentation you need to try this\nprompt you need to try that model you\n\nprompt you need to try that model you\n \n\nprompt you need to try that model you\nneed to try this approach etc etc and uh\n\nneed to try this approach etc etc and uh\n \n\nneed to try this approach etc etc and uh\neverything in your solution everything\n\neverything in your solution everything\n \n\neverything in your solution everything\nthat uh represents your solution which\n\nthat uh represents your solution which\n \n\nthat uh represents your solution which\nis your code your logic uh the prompts\n\nis your code your logic uh the prompts\n \n\nis your code your logic uh the prompts\nthat you use the the models that you use\n\nthat you use the the models that you use\n \n\nthat you use the the models that you use\nthe the data that you base your solution\n\nthe the data that you base your solution\n \n\nthe the data that you base your solution\non changing anything of that impacts\n\non changing anything of that impacts\n \n\non changing anything of that impacts\nyour uh solution in unexpected ways. Um people very often come to this uh to\n\nUm people very often come to this uh to\n \n\nUm people very often come to this uh to\ntry solving this with the wrong\n\ntry solving this with the wrong\n \n\ntry solving this with the wrong\napproach. They start with a data science\n\napproach. They start with a data science\n \n\napproach. They start with a data science\nmetrics. They like it sounds reasonable,\n\nmetrics. They like it sounds reasonable,\n \n\nmetrics. They like it sounds reasonable,\nright? So it requires data science\n\nright? So it requires data science\n \n\nright? So it requires data science\napproach of experimentation and uh\n\napproach of experimentation and uh\n \n\napproach of experimentation and uh\npeople start measuring groundness,\n\npeople start measuring groundness,\n \n\npeople start measuring groundness,\nfactuality, bias and other uh metrics\n\nfactuality, bias and other uh metrics\n \n\nfactuality, bias and other uh metrics\nthat don't really help you to understand\n\nthat don't really help you to understand\n \n\nthat don't really help you to understand\nuh is your solution uh working the right\n\nuh is your solution uh working the right\n \n\nuh is your solution uh working the right\nway? Does it uh does your latest change\n\nway? Does it uh does your latest change\n \n\nway? Does it uh does your latest change\nimproved uh your solution in the right\n\nimproved uh your solution in the right\n \n\nimproved uh your solution in the right\nway for your users? uh for example I've\n\nway for your users? uh for example I've\n \n\nway for your users?" } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nuh for example I've\nbeen talking to an ex-colague that are\n\nbeen talking to an ex-colague that are\n \n\nbeen talking to an ex-colague that are\nbuilding a customer support bot at\n\nbuilding a customer support bot at\n \n\nbuilding a customer support bot at\ntweaks I asked him how do you know that\n\ntweaks I ..." }, { "role": "assistant", "content": "uh for example I've\nbeen talking to an ex-colague that are\n\nbeen talking to an ex-colague that are\n \n\nbeen talking to an ex-colague that are\nbuilding a customer support bot at\n\nbuilding a customer support bot at\n \n\nbuilding a customer support bot at\ntweaks I asked him how do you know that\n\ntweaks I asked him how do you know that\n \n\ntweaks I asked him how do you know that\nyour solution is working well he started\n\nyour solution is working well he started\n \n\nyour solution is working well he started\ntalking about factuality and other data\n\ntalking about factuality and other data\n \n\ntalking about factuality and other data\nscience metrics\n\nscience metrics\n \n\nscience metrics\nu that's again I started to dig deeper\n\nu that's again I started to dig deeper\n \n\nu that's again I started to dig deeper\nand then we just uh together figure out\n\nand then we just uh together figure out\n \n\nand then we just uh together figure out\nthat the most important metric for them\n\nthat the most important metric for them\n \n\nthat the most important metric for them\nis uh the rate of moving from AI I\n\nis uh the rate of moving from AI I\n \n\nis uh the rate of moving from AI I\nsupport bot like escalation to a human\n\nsupport bot like escalation to a human\n \n\nsupport bot like escalation to a human\nsupport. support. support. If uh your solution uh hasn't able to\n\nIf uh your solution uh hasn't able to\n \n\nIf uh your solution uh hasn't able to\nanswer the user with all this factuality\n\nanswer the user with all this factuality\n \n\nanswer the user with all this factuality\nlike it could be super grounded but\n\nlike it could be super grounded but\n \n\nlike it could be super grounded but\nstill not provide the right answer that\n\nstill not provide the right answer that\n \n\nstill not provide the right answer that\nthe user expects and uh this is what you\n\nthe user expects and uh this is what you\n \n\nthe user expects and uh this is what you\nactually need to test. actually need to test. actually need to test. Um and my experience was to start with\n\nUm and my experience was to start with\n \n\nUm and my experience was to start with\nreal world scenarios. So basically you\n\nreal world scenarios. So basically you\n \n\nreal world scenarios. So basically you\nneed to reverse engineer your metrics\n\nneed to reverse engineer your metrics\n \n\nneed to reverse engineer your metrics\nand your metrics should be very very\n\nand your metrics should be very very\n \n\nand your metrics should be very very\nspecific to what your end goal. So they\n\nspecific to what your end goal. So they\n \n\nspecific to what your end goal. So they\nshould come from a product experience\n\nshould come from a product experience\n \n\nshould come from a product experience\nfrom business outcomes. Uh if your\n\nfrom business outcomes. Uh if your\n \n\nfrom business outcomes. Uh if your\nsolution is customer support bot, you\n\nsolution is customer support bot, you\n \n\nsolution is customer support bot, you\nneed to figure out what your users want\n\nneed to figure out what your users want\n \n\nneed to figure out what your users want\nand uh how you can mimic it. And instead\n\nand uh how you can mimic it. And instead\n \n\nand uh how you can mimic it. And instead\nof measuring something u average or\n\nof measuring something u average or\n \n\nof measuring something u average or\nsomething generic, you need to measure a\n\nsomething generic, you need to measure a\n \n\nsomething generic, you need to measure a\nvery specific criterias. very specific criterias. very specific criterias. Uh cuz universal valves don't really\n\nUh cuz universal valves don't really\n \n\nUh cuz universal valves don't really\nwork. work. work. How do we do it? Uh so for example,\n\nHow do we do it? Uh so for example,\n \n\nHow do we do it? Uh so for example,\ncustomer support bot, which is by the\n\ncustomer support bot, which is by the\n \n\ncustomer support bot, which is by the\nway one of the hardest uh things to do\n\nway one of the hardest uh things to do\n \n\nway one of the hardest uh things to do\nproperly. uh let's say I have a bank and\n\nproperly. uh let's say I have a bank and\n \n\nproperly. uh let's say I have a bank and\nuh bank has FAQ materials\n\nuh bank has FAQ materials\n \n\nuh bank has FAQ materials\nwhich contain including like how do you\n\nwhich contain including like how do you\n \n\nwhich contain including like how do you\nreset your password? reset your password? reset your password? Um so what I usually do when I help my\n\nUm so what I usually do when I help my\n \n\nUm so what I usually do when I help my\nuh like companies that I help them to\n\nuh like companies that I help them to\n \n\nuh like companies that I help them to\nbuild uh AI solutions we start with uh\n\nbuild uh AI solutions we start with uh\n \n\nbuild uh AI solutions we start with uh\nreverse engineering like how do we\n\nreverse engineering like how do we\n \n\nreverse engineering like how do we\ncreate the valves based on that. So in\n\ncreate the valves based on that. So in\n \n\ncreate the valves based on that. So in\nthis case I use LLM and in most cases I\n\nthis case I use LLM and in most cases I\n \n\nthis case I use LLM and in most cases I\nuse LLM to come up with right\n\nuse LLM to come up with right\n \n\nuse LLM to come up with right\nevaluations. So here I can take say 01\n\nevaluations. So here I can take say 01\n \n\nevaluations. So here I can take say 01\nuh 03 now uh and just reverse engineer\n\nuh 03 now uh and just reverse engineer\n \n\nuh 03 now uh and just reverse engineer\nwhat should be the user question uh that\n\nwhat should be the user question uh that\n \n\nwhat should be the user question uh that\nwe know to answer based on these\n\nwe know to answer based on these\n \n\nwe know to answer based on these\nmaterials and what should be the\n\nmaterials and what should be the\n \n\nmaterials and what should be the\nspecific criteria that uh these\n\nspecific criteria that uh these\n \n\nspecific criteria that uh these\nmaterials are providing an answer for\n\nmaterials are providing an answer for\n \n\nmaterials are providing an answer for\nand some of these criteria are quite\n\nand some of these criteria are quite\n \n\nand some of these criteria are quite\nimportant. So for example here it says\n\nimportant. So for example here it says\n \n\nimportant. So for example here it says\nthat uh uh as part of the thing you you\n\nthat uh uh as part of the thing you you\n \n\nthat uh uh as part of the thing you you\nneed to receive a mobile validation. So\n\nneed to receive a mobile validation. So\n \n\nneed to receive a mobile validation. So\nyou receive a SMS code and uh it says\n\nyou receive a SMS code and uh it says\n \n\nyou receive a SMS code and uh it says\nthat if you uh don't have a mobile\n\nthat if you uh don't have a mobile\n \n\nthat if you uh don't have a mobile\nnumber then you can reach uh support etc\n\nnumber then you can reach uh support etc\n \n\nnumber then you can reach uh support etc\netc. Uh if some of that information is\n\netc. Uh if some of that information is\n \n\netc. Uh if some of that information is\nmissing from the answer the answer would\n\nmissing from the answer the answer would\n \n\nmissing from the answer the answer would\nnot be correct. not be correct. not be correct. You need to be very specific about what\n\nYou need to be very specific about what\n \n\nYou need to be very specific about what\nexact information you need to see in the\n\nexact information you need to see in the\n \n\nexact information you need to see in the\nanswer and that information is very\n\nanswer and that information is very\n \n\nanswer and that information is very\nspecific to that specific question. So\n\nspecific to that specific question. So\n \n\nspecific to that specific question. So\nyou need to build like lots of evals\n\nyou need to build like lots of evals\n \n\nyou need to build like lots of evals\nuh from the materials in this case uh\n\nuh from the materials in this case uh\n \n\nuh from the materials in this case uh\nthat mimic specific user questions that\n\nthat mimic specific user questions that\n \n\nthat mimic specific user questions that\nuh you need to be able to answer for. Uh\n\nuh you need to be able to answer for." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nUh\n \n\nuh you need to be able to answer for. Uh\nhow do we do it? Usually again I work\n\nhow do we do it? Usually again I work\n \n\nhow do we do it? Usually again I work\nwith uh smart models like O3 uh and I uh\n\nwith uh smart models like O3 uh and I uh\n \n\nwith uh smart models like O3 uh and I uh\nprovided..." }, { "role": "assistant", "content": "Uh\n \n\nuh you need to be able to answer for. Uh\nhow do we do it? Usually again I work\n\nhow do we do it? Usually again I work\n \n\nhow do we do it? Usually again I work\nwith uh smart models like O3 uh and I uh\n\nwith uh smart models like O3 uh and I uh\n \n\nwith uh smart models like O3 uh and I uh\nprovided enough context. I provided\n\nprovided enough context. I provided\n \n\nprovided enough context. I provided\nwhich personas are we trying to\n\nwhich personas are we trying to\n \n\nwhich personas are we trying to\nrepresent because you can make ask the\n\nrepresent because you can make ask the\n \n\nrepresent because you can make ask the\nsame question in uh completely different\n\nsame question in uh completely different\n \n\nsame question in uh completely different\nways depending on who is the persona\n\nways depending on who is the persona\n \n\nways depending on who is the persona\nasking uh yet you would expect exactly\n\nasking uh yet you would expect exactly\n \n\nasking uh yet you would expect exactly\nthe same answer. So you need to account\n\nthe same answer. So you need to account\n \n\nthe same answer. So you need to account\nfor it. for it. for it. Um so this is uh an example from uh the\n\nUm so this is uh an example from uh the\n \n\nUm so this is uh an example from uh the\nopen source platform that we have that\n\nopen source platform that we have that\n \n\nopen source platform that we have that\nuh just helps to get it done. So if you\n\nuh just helps to get it done. So if you\n \n\nuh just helps to get it done. So if you\nlook it up multineer I'm not trying to\n\nlook it up multineer I'm not trying to\n \n\nlook it up multineer I'm not trying to\nsell you anything. I'm not trying to\n\nsell you anything. I'm not trying to\n \n\nsell you anything. I'm not trying to\nlike vendor lock in or whatever. It's\n\nlike vendor lock in or whatever. It's\n \n\nlike vendor lock in or whatever. It's\ncompletely open source and if needed I\n\ncompletely open source and if needed I\n \n\ncompletely open source and if needed I\ncan just recreate it in a couple of days\n\ncan just recreate it in a couple of days\n \n\ncan just recreate it in a couple of days\nnow with cursor. The point is in the\n\nnow with cursor. The point is in the\n \n\nnow with cursor. The point is in the\napproach not in the platform. Uh so for\n\napproach not in the platform. Uh so for\n \n\napproach not in the platform. Uh so for\nexample here we see that very same\n\nexample here we see that very same\n \n\nexample here we see that very same\nquestion um how do I reset my password\n\nquestion um how do I reset my password\n \n\nquestion um how do I reset my password\nyou see the what was the input what was\n\nyou see the what was the input what was\n \n\nyou see the what was the input what was\nthe output and uh that specific criteria\n\nthe output and uh that specific criteria\n \n\nthe output and uh that specific criteria\nthat I measure it uh that specific\n\nthat I measure it uh that specific\n \n\nthat I measure it uh that specific\nquestion how do I know if the answer is\n\nquestion how do I know if the answer is\n \n\nquestion how do I know if the answer is\ncorrect and now I can just reiterate and\n\ncorrect and now I can just reiterate and\n \n\ncorrect and now I can just reiterate and\ngenerate like 50 different variations of\n\ngenerate like 50 different variations of\n \n\ngenerate like 50 different variations of\nthe same question and see if I still get\n\nthe same question and see if I still get\n \n\nthe same question and see if I still get\nthe right answer if the answer matches\n\nthe right answer if the answer matches\n \n\nthe right answer if the answer matches\nall the checklist that I have for that\n\nall the checklist that I have for that\n \n\nall the checklist that I have for that\nspecific answer. specific answer. specific answer. Um how the process usually works\n\nUm how the process usually works\n \n\nUm how the process usually works\num so contrary to like regular approach\n\num so contrary to like regular approach\n \n\num so contrary to like regular approach\nyou build your evals not at the end of\n\nyou build your evals not at the end of\n \n\nyou build your evals not at the end of\nthe process but in the very beginning of\n\nthe process but in the very beginning of\n \n\nthe process but in the very beginning of\nthe process. So you just build your\n\nthe process. So you just build your\n \n\nthe process. So you just build your\nfirst version of the PC. You define the\n\nfirst version of the PC. You define the\n \n\nfirst version of the PC. You define the\nfirst version of your tests evaluations. first version of your tests evaluations. first version of your tests evaluations. You run them and you see what's going\n\nYou run them and you see what's going\n \n\nYou run them and you see what's going\non. You you will see that uh in some\n\non. You you will see that uh in some\n \n\non. You you will see that uh in some\ncases it will fail. Uh in some cases it\n\ncases it will fail. Uh in some cases it\n \n\ncases it will fail. Uh in some cases it\nwill succeed. What's important is to to\n\nwill succeed. What's important is to to\n \n\nwill succeed. What's important is to to\nlook at the details not just see the\n\nlook at the details not just see the\n \n\nlook at the details not just see the\naverage numbers. The average numbers\n\naverage numbers. The average numbers\n \n\naverage numbers. The average numbers\nwon't tell you anything. Uh won't tell\n\nwon't tell you anything. Uh won't tell\n \n\nwon't tell you anything. Uh won't tell\nyou how to improve it. If you actually\n\nyou how to improve it. If you actually\n \n\nyou how to improve it. If you actually\nlook at the details of each evaluation,\n\nlook at the details of each evaluation,\n \n\nlook at the details of each evaluation,\nyou'll see exactly why it's failing. It\n\nyou'll see exactly why it's failing. It\n \n\nyou'll see exactly why it's failing. It\ncould be failing um because your test is\n\ncould be failing um because your test is\n \n\ncould be failing um because your test is\nnot defined correctly. It could be\n\nnot defined correctly. It could be\n \n\nnot defined correctly. It could be\nfailing because your uh solution is not\n\nfailing because your uh solution is not\n \n\nfailing because your uh solution is not\nworking as it should be. And like in\n\nworking as it should be. And like in\n \n\nworking as it should be. And like in\norder to do it, you may need to uh to do\n\norder to do it, you may need to uh to do\n \n\norder to do it, you may need to uh to do\na change in like you may change a model,\n\na change in like you may change a model,\n \n\na change in like you may change a model,\nyou may change something on in your\n\nyou may change something on in your\n \n\nyou may change something on in your\nlogic, you may change a prompt or the\n\nlogic, you may change a prompt or the\n \n\nlogic, you may change a prompt or the\ndata that you use in order to uh answer\n\ndata that you use in order to uh answer\n \n\ndata that you use in order to uh answer\na question in our example. a question in our example. a question in our example. And uh basically what you do now is\n\nAnd uh basically what you do now is\n \n\nAnd uh basically what you do now is\nexperimentation. So you you start\n\nexperimentation. So you you start\n \n\nexperimentation. So you you start\nrunning your experiment. You change\n\nrunning your experiment. You change\n \n\nrunning your experiment. You change\nsomething. you you need to define these\n\nsomething. you you need to define these\n \n\nsomething. you you need to define these\ntests in a way that will uh help you to\n\ntests in a way that will uh help you to\n \n\ntests in a way that will uh help you to\nmake an educated guess on uh what you\n\nmake an educated guess on uh what you\n \n\nmake an educated guess on uh what you\nneed to change in order to to do it. In\n\nneed to change in order to to do it. In\n \n\nneed to change in order to to do it." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nIn\nsome cases it will work in some cases it\n\nsome cases it will work in some cases it\n \n\nsome cases it will work in some cases it\nwon't. But even if it works uh let's say\n\nwon't. But even if it works uh let's say\n \n\nwon't. But even if it works uh let's say\nyou change something in your prompt and\n\nyo..." }, { "role": "assistant", "content": "In\nsome cases it will work in some cases it\n\nsome cases it will work in some cases it\n \n\nsome cases it will work in some cases it\nwon't. But even if it works uh let's say\n\nwon't. But even if it works uh let's say\n \n\nwon't. But even if it works uh let's say\nyou change something in your prompt and\n\nyou change something in your prompt and\n \n\nyou change something in your prompt and\nit fixed this test. In my experience, in\n\nit fixed this test. In my experience, in\n \n\nit fixed this test. In my experience, in\nmany cases, it breaks uh something that\n\nmany cases, it breaks uh something that\n \n\nmany cases, it breaks uh something that\nused to work before. Uh like you you you\n\nused to work before. Uh like you you you\n \n\nused to work before. Uh like you you you\nhave constant regressions and if you\n\nhave constant regressions and if you\n \n\nhave constant regressions and if you\ndon't have these evaluations, there is\n\ndon't have these evaluations, there is\n \n\ndon't have these evaluations, there is\nno way you'll be able to catch it on\n\nno way you'll be able to catch it on\n \n\nno way you'll be able to catch it on\ntime. time. time. So this is hugely important and what\n\nSo this is hugely important and what\n \n\nSo this is hugely important and what\nactually happens is that again you build\n\nactually happens is that again you build\n \n\nactually happens is that again you build\nyour first version. You build your first\n\nyour first version. You build your first\n \n\nyour first version. You build your first\nversion of the vows uh you match them\n\nversion of the vows uh you match them\n \n\nversion of the vows uh you match them\nyou run these valves you improve\n\nyou run these valves you improve\n \n\nyou run these valves you improve\nsomething you improve your vows or maybe\n\nsomething you improve your vows or maybe\n \n\nsomething you improve your vows or maybe\nadd more evaluations and then you like\n\nadd more evaluations and then you like\n \n\nadd more evaluations and then you like\ncontinuously improve it until you reach\n\ncontinuously improve it until you reach\n \n\ncontinuously improve it until you reach\nsome point where you are satisfied with\n\nsome point where you are satisfied with\n \n\nsome point where you are satisfied with\nyour valves for this specific solution\n\nyour valves for this specific solution\n \n\nyour valves for this specific solution\nfor that specific point of time. And\n\nfor that specific point of time. And\n \n\nfor that specific point of time. And\nwhat actually happened is that you you\n\nwhat actually happened is that you you\n \n\nwhat actually happened is that you you\nyou got your baseline, you got your\n\nyou got your baseline, you got your\n \n\nyou got your baseline, you got your\nbenchmark that uh now you can start\n\nbenchmark that uh now you can start\n \n\nbenchmark that uh now you can start\noptimizing and uh you have the\n\noptimizing and uh you have the\n \n\noptimizing and uh you have the\nconfidence that the tests should be\n\nconfidence that the tests should be\n \n\nconfidence that the tests should be\nworking. So now you can try another\n\nworking. So now you can try another\n \n\nworking. So now you can try another\nmodel. Let's say well what how can I try\n\nmodel. Let's say well what how can I try\n \n\nmodel. Let's say well what how can I try\nto see if 40 mini will work the same way\n\nto see if 40 mini will work the same way\n \n\nto see if 40 mini will work the same way\nwith 40 or not? uh can I use a graph rag\n\nwith 40 or not? uh can I use a graph rag\n \n\nwith 40 or not? uh can I use a graph rag\nor can I try a simpler solution? or can I try a simpler solution? or can I try a simpler solution? uh should I have uh to use agentic\n\nuh should I have uh to use agentic\n \n\nuh should I have uh to use agentic\napproach that like maybe better but uh\n\napproach that like maybe better but uh\n \n\napproach that like maybe better but uh\nrequires more time more uh inference\n\nrequires more time more uh inference\n \n\nrequires more time more uh inference\ncost etc or should I try to simplify the\n\ncost etc or should I try to simplify the\n \n\ncost etc or should I try to simplify the\nlogic or maybe I can simplify the logic\n\nlogic or maybe I can simplify the logic\n \n\nlogic or maybe I can simplify the logic\nfor a specific portion of the\n\nfor a specific portion of the\n \n\nfor a specific portion of the\napplication etc etc having this\n\napplication etc etc having this\n \n\napplication etc etc having this\nbenchmark uh allows you to do all these\n\nbenchmark uh allows you to do all these\n \n\nbenchmark uh allows you to do all these\nexperimentations uh with confidence but\n\nexperimentations uh with confidence but\n \n\nexperimentations uh with confidence but\nagain the the most important part is\n\nagain the the most important part is\n \n\nagain the the most important part is\nlike how do you reach this benchmark\n\nlike how do you reach this benchmark\n \n\nlike how do you reach this benchmark\nAnd uh while the approach is uh pretty\n\nAnd uh while the approach is uh pretty\n \n\nAnd uh while the approach is uh pretty\nmuch the same, the evaluations that you\n\nmuch the same, the evaluations that you\n \n\nmuch the same, the evaluations that you\nneed to build and how do you build your\n\nneed to build and how do you build your\n \n\nneed to build and how do you build your\nevaluations are completely different\n\nevaluations are completely different\n \n\nevaluations are completely different\ndepending on the solution that you need\n\ndepending on the solution that you need\n \n\ndepending on the solution that you need\nto build because uh the models are super\n\nto build because uh the models are super\n \n\nto build because uh the models are super\ncapable right now. Uh so they allow you\n\ncapable right now. Uh so they allow you\n \n\ncapable right now. Uh so they allow you\nto build a huge variety of uh solutions\n\nto build a huge variety of uh solutions\n \n\nto build a huge variety of uh solutions\nbut each and every solution is quite uh\n\nbut each and every solution is quite uh\n \n\nbut each and every solution is quite uh\ndifferent in terms of how do you uh\n\ndifferent in terms of how do you uh\n \n\ndifferent in terms of how do you uh\nevaluate it. uh for support bot you\n\nevaluate it. uh for support bot you\n \n\nevaluate it. uh for support bot you\nusually typically use LM as a judge as I\n\nusually typically use LM as a judge as I\n \n\nusually typically use LM as a judge as I\nuh made an example if you're building\n\nuh made an example if you're building\n \n\nuh made an example if you're building\ntext to SQL or text to graph database\n\ntext to SQL or text to graph database\n \n\ntext to SQL or text to graph database\nthen uh to my experience the best way is\n\nthen uh to my experience the best way is\n \n\nthen uh to my experience the best way is\nto create a mock database that\n\nto create a mock database that\n \n\nto create a mock database that\nrepresents the um whatever uh database\n\nrepresents the um whatever uh database\n \n\nrepresents the um whatever uh database\nor databases that you need your solution\n\nor databases that you need your solution\n \n\nor databases that you need your solution\nto work with they represent the same\n\nto work with they represent the same\n \n\nto work with they represent the same\nschema and you have a mock data so you\n\nschema and you have a mock data so you\n \n\nschema and you have a mock data so you\nknow exactly uh what should expect on\n\nknow exactly uh what should expect on\n \n\nknow exactly uh what should expect on\nspecific questions. Um if you need to\n\nspecific questions. Um if you need to\n \n\nspecific questions. Um if you need to\nbuild some classifier for call center\n\nbuild some classifier for call center\n \n\nbuild some classifier for call center\nconversations then your uh tests are\n\nconversations then your uh tests are\n \n\nconversations then your uh tests are\nlike simple match whenever this is this\n\nlike simple match whenever this is this\n \n\nlike simple match whenever this is this\nis the right rubric or not. Uh and the\n\nis the right rubric or not. Uh and the\n \n\nis the right rubric or not. Uh and the\nsame appro uh approach applies to guard\n\nsame appro uh approach applies to guard\n \n\nsame appro uh approach applies to guard\nrails." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nSo uh getting back to the support\n\nrails. So uh getting back to the support\n \n\nrails. So uh getting back to the support\nto the uh example of the customer\n\nto the uh example of the customer\n \n\nto the uh example of the customer\nsupport bot uh guardrails you need to\n\nsupport bot uh guardrails you need ..." }, { "role": "assistant", "content": "So uh getting back to the support\n\nrails. So uh getting back to the support\n \n\nrails. So uh getting back to the support\nto the uh example of the customer\n\nto the uh example of the customer\n \n\nto the uh example of the customer\nsupport bot uh guardrails you need to\n\nsupport bot uh guardrails you need to\n \n\nsupport bot uh guardrails you need to\ncover uh questions that should not be\n\ncover uh questions that should not be\n \n\ncover uh questions that should not be\nanswered or questions that should be\n\nanswered or questions that should be\n \n\nanswered or questions that should be\nanswered in different ways or questions\n\nanswered in different ways or questions\n \n\nanswered in different ways or questions\nthat uh uh the answers are not in the\n\nthat uh uh the answers are not in the\n \n\nthat uh uh the answers are not in the\nmaterial. So all of these you can put\n\nmaterial. So all of these you can put\n \n\nmaterial. So all of these you can put\ninto your benchmark just different type\n\ninto your benchmark just different type\n \n\ninto your benchmark just different type\nof benchmark but it's pretty much the\n\nof benchmark but it's pretty much the\n \n\nof benchmark but it's pretty much the\nsame approach. same approach. same approach. Uh so just to reiterate uh the key\n\nUh so just to reiterate uh the key\n \n\nUh so just to reiterate uh the key\ntakeaways, you need to evaluate your\n\ntakeaways, you need to evaluate your\n \n\ntakeaways, you need to evaluate your\napps the way your users actually use\n\napps the way your users actually use\n \n\napps the way your users actually use\nthem. Um and uh avoid abstract metrics\n\nthem. Um and uh avoid abstract metrics\n \n\nthem. Um and uh avoid abstract metrics\nuh because these abstract connectors\n\nuh because these abstract connectors\n \n\nuh because these abstract connectors\ndon't really measure anything important. don't really measure anything important. don't really measure anything important. Uh and the approach is uh through\n\nUh and the approach is uh through\n \n\nUh and the approach is uh through\nexperimentation. So you run these\n\nexperimentation. So you run these\n \n\nexperimentation. So you run these\nevaluations frequently. you that allows\n\nevaluations frequently. you that allows\n \n\nevaluations frequently. you that allows\nyou to have rapid progress with uh less\n\nyou to have rapid progress with uh less\n \n\nyou to have rapid progress with uh less\nregressions because testing frequently\n\nregressions because testing frequently\n \n\nregressions because testing frequently\nhelp you to to catch these surprises. Uh\n\nhelp you to to catch these surprises. Uh\n \n\nhelp you to to catch these surprises. Uh\nbut most importantly what you get if you\n\nbut most importantly what you get if you\n \n\nbut most importantly what you get if you\ndivi define your evaluations correctly,\n\ndivi define your evaluations correctly,\n \n\ndivi define your evaluations correctly,\nyou get your solution pretty much uh as\n\nyou get your solution pretty much uh as\n \n\nyou get your solution pretty much uh as\nkind of explainable AI because you know\n\nkind of explainable AI because you know\n \n\nkind of explainable AI because you know\nexactly what it does, you know exactly\n\nexactly what it does, you know exactly\n \n\nexactly what it does, you know exactly\nhow it does it if you test it the right\n\nhow it does it if you test it the right\n \n\nhow it does it if you test it the right\nway. way. way. Thank you very much. Uh take a look at\n\nThank you very much. Uh take a look at\n \n\nThank you very much. Uh take a look at\nmultineer uh that's a platform that you\n\nmultineer uh that's a platform that you\n \n\nmultineer uh that's a platform that you\ncan use to uh run these evaluations. You\n\ncan use to uh run these evaluations. You\n \n\ncan use to uh run these evaluations. You\ncan totally use any other platform. The\n\ncan totally use any other platform. The\n \n\ncan totally use any other platform. The\napproach is quite simple. It doesn't\n\napproach is quite simple. It doesn't\n \n\napproach is quite simple. It doesn't\nrequire any specific platform. Uh I've\n\nrequire any specific platform. Uh I've\n \n\nrequire any specific platform. Uh I've\nbuilt multin just because no other\n\nbuilt multin just because no other\n \n\nbuilt multin just because no other\nplatform helped me to do it this way to\n\nplatform helped me to do it this way to\n \n\nplatform helped me to do it this way to\nto help me with the process of\n\nto help me with the process of\n \n\nto help me with the process of\nevaluation like end to end. Um, I'm\n\nevaluation like end to end. Um, I'm\n \n\nevaluation like end to end. Um, I'm\nworking on a startup that does reliable\n\nworking on a startup that does reliable\n \n\nworking on a startup that does reliable\nAI automation right now. Um, and uh,\n\nAI automation right now. Um, and uh,\n \n\nAI automation right now. Um, and uh,\nyeah, thank you very much. yeah, thank you very much. yeah, thank you very much. [Music]" } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\n[music]\n\n \n\nThanks everybody for uh for joining. I'm\n\nThanks everybody for uh for joining. I'm\n \n\nThanks everybody for uh for joining. I'm\nhere to talk to you today about DSPI. here to talk to you today about DSPI. here to talk to you today about DSPI. Um, and feel free to jump in with\n\nUm, and feel..." }, { "role": "assistant", "content": "[music]\n\n \n\nThanks everybody for uh for joining. I'm\n\nThanks everybody for uh for joining. I'm\n \n\nThanks everybody for uh for joining. I'm\nhere to talk to you today about DSPI. here to talk to you today about DSPI. here to talk to you today about DSPI. Um, and feel free to jump in with\n\nUm, and feel free to jump in with\n \n\nUm, and feel free to jump in with\nquestions or anything throughout the\n\nquestions or anything throughout the\n \n\nquestions or anything throughout the\ntalk. It's, you know, I don't sp I don't\n\ntalk. It's, you know, I don't sp I don't\n \n\ntalk. It's, you know, I don't sp I don't\nplan on spending the full hour and a\n\nplan on spending the full hour and a\n \n\nplan on spending the full hour and a\nhalf or so. I know it's the last session\n\nhalf or so. I know it's the last session\n \n\nhalf or so. I know it's the last session\nof the day. So, um, keep it casual. Feel\n\nof the day. So, um, keep it casual. Feel\n \n\nof the day. So, um, keep it casual. Feel\nfree to jump in. I'll start with a\n\nfree to jump in. I'll start with a\n \n\nfree to jump in. I'll start with a\nlittle bit of background. Don't want to\n\nlittle bit of background. Don't want to\n \n\nlittle bit of background. Don't want to\ngo through too many slides. I'm\n\ngo through too many slides. I'm\n \n\ngo through too many slides. I'm\ntechnically a consultant, so I have to\n\ntechnically a consultant, so I have to\n \n\ntechnically a consultant, so I have to\ndo some slides, but we will dive into\n\ndo some slides, but we will dive into\n \n\ndo some slides, but we will dive into\nthe code for the the latter half. And\n\nthe code for the the latter half. And\n \n\nthe code for the the latter half. And\nthere's a GitHub repo that you can\n\nthere's a GitHub repo that you can\n \n\nthere's a GitHub repo that you can\ndownload to to follow along and play\n\ndownload to to follow along and play\n \n\ndownload to to follow along and play\naround with it on your own. around with it on your own. around with it on your own. Um, so how many people here have heard\n\nUm, so how many people here have heard\n \n\nUm, so how many people here have heard\nof DSPI? of DSPI? of DSPI? Almost everyone. That's awesome. How\n\nAlmost everyone. That's awesome. How\n \n\nAlmost everyone. That's awesome. How\nmany people have actually used it kind\n\nmany people have actually used it kind\n \n\nmany people have actually used it kind\nof day-to-day in production or anything\n\nof day-to-day in production or anything\n \n\nof day-to-day in production or anything\nlike that? Three. Okay, good. So\n\nlike that? Three. Okay, good. So\n \n\nlike that? Three. Okay, good. So\nhopefully we can convert some more of\n\nhopefully we can convert some more of\n \n\nhopefully we can convert some more of\nyou today. Um, so high level DSPI, this\n\nyou today. Um, so high level DSPI, this\n \n\nyou today. Um, so high level DSPI, this\nis straight from the website. Um, it's a\n\nis straight from the website. Um, it's a\n \n\nis straight from the website. Um, it's a\ndeclarative framework for how you can\n\ndeclarative framework for how you can\n \n\ndeclarative framework for how you can\nbuild modular software. And most\n\nbuild modular software. And most\n \n\nbuild modular software. And most\nimportant for someone like myself, I'm\n\nimportant for someone like myself, I'm\n \n\nimportant for someone like myself, I'm\nnot necessarily an engineer that is\n\nnot necessarily an engineer that is\n \n\nnot necessarily an engineer that is\nwriting code all day, every day. As I\n\nwriting code all day, every day. As I\n \n\nwriting code all day, every day. As I\nmentioned before, I'm a more of a\n\nmentioned before, I'm a more of a\n \n\nmentioned before, I'm a more of a\ntechnical consultant. So, I run across a\n\ntechnical consultant. So, I run across a\n \n\ntechnical consultant. So, I run across a\nvariety of different problems. Could be\n\nvariety of different problems. Could be\n \n\nvariety of different problems. Could be\num an investigation for a law firm. It\n\num an investigation for a law firm. It\n \n\num an investigation for a law firm. It\ncould be helping a company understand\n\ncould be helping a company understand\n \n\ncould be helping a company understand\nhow to improve their processes, how to\n\nhow to improve their processes, how to\n \n\nhow to improve their processes, how to\ndeploy AI internally. Maybe we need to\n\ndeploy AI internally. Maybe we need to\n \n\ndeploy AI internally. Maybe we need to\nlook through 10 10,000 contracts to\n\nlook through 10 10,000 contracts to\n \n\nlook through 10 10,000 contracts to\nidentify a particular clause um or or\n\nidentify a particular clause um or or\n \n\nidentify a particular clause um or or\nparagraph. And so DSPI has been a really\n\nparagraph. And so DSPI has been a really\n \n\nparagraph. And so DSPI has been a really\nnice way for me personally and my team\n\nnice way for me personally and my team\n \n\nnice way for me personally and my team\nto iterate really really quickly on\n\nto iterate really really quickly on\n \n\nto iterate really really quickly on\nbuilding these applications. building these applications. building these applications. Most importantly building programs. It's\n\nMost importantly building programs. It's\n \n\nMost importantly building programs. It's\nnot um it's not kind of iterating with\n\nnot um it's not kind of iterating with\n \n\nnot um it's not kind of iterating with\nprompts and tweaking things back and\n\nprompts and tweaking things back and\n \n\nprompts and tweaking things back and\nforth. It is building a a proper Python\n\nforth. It is building a a proper Python\n \n\nforth. It is building a a proper Python\nprogram and and DSP is a really good way\n\nprogram and and DSP is a really good way\n \n\nprogram and and DSP is a really good way\nfor you to do that. for you to do that. for you to do that. So I mentioned before there's a repo\n\nSo I mentioned before there's a repo\n \n\nSo I mentioned before there's a repo\nonline if you want to download it now\n\nonline if you want to download it now\n \n\nonline if you want to download it now\nand kind of just get everything set up. and kind of just get everything set up. and kind of just get everything set up. I'll put this on the screen later on. I'll put this on the screen later on. I'll put this on the screen later on. Um, but if you want to go here, just\n\nUm, but if you want to go here, just\n \n\nUm, but if you want to go here, just\nkind of download some of the code. It uh\n\nkind of download some of the code. It uh\n \n\nkind of download some of the code. It uh\nit's been put together over the past\n\nit's been put together over the past\n \n\nit's been put together over the past\ncouple days. So, it's not going to be\n\ncouple days. So, it's not going to be\n \n\ncouple days. So, it's not going to be\nperfect production level code. It's much\n\nperfect production level code. It's much\n \n\nperfect production level code. It's much\nmore of utilities and little things here\n\nmore of utilities and little things here\n \n\nmore of utilities and little things here\nand there to just come and kind of\n\nand there to just come and kind of\n \n\nand there to just come and kind of\ndemonstrate the usefulness, demonstrate\n\ndemonstrate the usefulness, demonstrate\n \n\ndemonstrate the usefulness, demonstrate\nthe point of of what we're talking about\n\nthe point of of what we're talking about\n \n\nthe point of of what we're talking about\ntoday in that and we'll walk through all\n\ntoday in that and we'll walk through all\n \n\ntoday in that and we'll walk through all\nof these these different use cases. So\n\nof these these different use cases. So\n \n\nof these these different use cases." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nSo\num sentiment classifier going through a\n\num sentiment classifier going through a\n \n\num sentiment classifier going through a\nPDF some multimodal work uh a very very\n\nPDF some multimodal work uh a very very\n \n\nPDF some multimodal work uh a very very\nsimple web research agent detecting\n\nsimple web r..." }, { "role": "assistant", "content": "So\num sentiment classifier going through a\n\num sentiment classifier going through a\n \n\num sentiment classifier going through a\nPDF some multimodal work uh a very very\n\nPDF some multimodal work uh a very very\n \n\nPDF some multimodal work uh a very very\nsimple web research agent detecting\n\nsimple web research agent detecting\n \n\nsimple web research agent detecting\nboundaries of a PDF document you'll see\n\nboundaries of a PDF document you'll see\n \n\nboundaries of a PDF document you'll see\nhow to summarize basically arbitrary\n\nhow to summarize basically arbitrary\n \n\nhow to summarize basically arbitrary\nlength text and then go into an\n\nlength text and then go into an\n \n\nlength text and then go into an\noptimizer uh with Jeepo\n\noptimizer uh with Jeepo\n \n\noptimizer uh with Jeepo\nbut before we do that just again kind of\n\nbut before we do that just again kind of\n \n\nbut before we do that just again kind of\nlevel set the biggest thing for me\n\nlevel set the biggest thing for me\n \n\nlevel set the biggest thing for me\npersonally DSP is a really nice way to\n\npersonally DSP is a really nice way to\n \n\npersonally DSP is a really nice way to\ndecompose your logic into a program that\n\ndecompose your logic into a program that\n \n\ndecompose your logic into a program that\ntreats LLMs as a first class citizen. So\n\ntreats LLMs as a first class citizen. So\n \n\ntreats LLMs as a first class citizen. So\nat the end of the day, you're\n\nat the end of the day, you're\n \n\nat the end of the day, you're\nfundamentally just calling a function\n\nfundamentally just calling a function\n \n\nfundamentally just calling a function\nthat under the hood just happens to be\n\nthat under the hood just happens to be\n \n\nthat under the hood just happens to be\nan LLM and DSPI gives you a really nice\n\nan LLM and DSPI gives you a really nice\n \n\nan LLM and DSPI gives you a really nice\nintuitive easy way to do that with some\n\nintuitive easy way to do that with some\n \n\nintuitive easy way to do that with some\nguarantees about the input and output\n\nguarantees about the input and output\n \n\nguarantees about the input and output\ntypes. So of course there are structured\n\ntypes. So of course there are structured\n \n\ntypes. So of course there are structured\noutputs, of course there are other ways\n\noutputs, of course there are other ways\n \n\noutputs, of course there are other ways\nto do this, Pyantic [snorts]\n\nto do this, Pyantic [snorts]\n \n\nto do this, Pyantic [snorts]\nand others. Um, but DSPI has a set of\n\nand others. Um, but DSPI has a set of\n \n\nand others. Um, but DSPI has a set of\nprimitives that when you put it all\n\nprimitives that when you put it all\n \n\nprimitives that when you put it all\ntogether allows you to build a cohesive\n\ntogether allows you to build a cohesive\n \n\ntogether allows you to build a cohesive\nmodular piece of software that you then\n\nmodular piece of software that you then\n \n\nmodular piece of software that you then\nhappen to be able to optimize. We'll get\n\nhappen to be able to optimize. We'll get\n \n\nhappen to be able to optimize. We'll get\ninto that uh in a minute. into that uh in a minute. into that uh in a minute. So, just a few reasons of why I'm such\n\nSo, just a few reasons of why I'm such\n \n\nSo, just a few reasons of why I'm such\nan advocate. It sit at it sits at this\n\nan advocate. It sit at it sits at this\n \n\nan advocate. It sit at it sits at this\nreally nice level of abstraction. So,\n\nreally nice level of abstraction. So,\n \n\nreally nice level of abstraction. So,\nit's I I would say it doesn't get in\n\nit's I I would say it doesn't get in\n \n\nit's I I would say it doesn't get in\nyour way as much as a lang chain. And\n\nyour way as much as a lang chain. And\n \n\nyour way as much as a lang chain. And\nthat's not a knock-on lang chain. It's\n\nthat's not a knock-on lang chain. It's\n \n\nthat's not a knock-on lang chain. It's\njust a different kind of paradigm in the\n\njust a different kind of paradigm in the\n \n\njust a different kind of paradigm in the\nway that DSPI is is structured. Um, and\n\nway that DSPI is is structured. Um, and\n \n\nway that DSPI is is structured. Um, and\nallows you to focus on things that\n\nallows you to focus on things that\n \n\nallows you to focus on things that\nactually matter. So you're not writing\n\nactually matter. So you're not writing\n \n\nactually matter. So you're not writing\nchoices zero messages content. You're\n\nchoices zero messages content. You're\n \n\nchoices zero messages content. You're\nnot you're not doing string parser. not you're not doing string parser. not you're not doing string parser. You're not doing a bunch of stuff under\n\nYou're not doing a bunch of stuff under\n \n\nYou're not doing a bunch of stuff under\nthe hood. You're just declaring your\n\nthe hood. You're just declaring your\n \n\nthe hood. You're just declaring your\nintent of how you want the program to\n\nintent of how you want the program to\n \n\nintent of how you want the program to\noperate, what you want your inputs and\n\noperate, what you want your inputs and\n \n\noperate, what you want your inputs and\noutputs to be. outputs to be. outputs to be. Because of this, it allows you to create\n\nBecause of this, it allows you to create\n \n\nBecause of this, it allows you to create\ncomputer programs. As I mentioned\n\ncomputer programs. As I mentioned\n \n\ncomputer programs. As I mentioned\nbefore, not just tweaking strings and\n\nbefore, not just tweaking strings and\n \n\nbefore, not just tweaking strings and\nsending them back and forth. You are\n\nsending them back and forth. You are\n \n\nsending them back and forth. You are\nbuilding a program first. It just\n\nbuilding a program first. It just\n \n\nbuilding a program first. It just\nhappens to also use LLMs. And really the\n\nhappens to also use LLMs. And really the\n \n\nhappens to also use LLMs. And really the\nthe most kind of important part of this\n\nthe most kind of important part of this\n \n\nthe most kind of important part of this\nis that and Omar the KB the uh the\n\nis that and Omar the KB the uh the\n \n\nis that and Omar the KB the uh the\nfounder of this or the the original\n\nfounder of this or the the original\n \n\nfounder of this or the the original\ndeveloper of it had this really good\n\ndeveloper of it had this really good\n \n\ndeveloper of it had this really good\npodcast with A16Z. I think it came out\n\npodcast with A16Z. I think it came out\n \n\npodcast with A16Z. I think it came out\njust like two or three days ago. But it\n\njust like two or three days ago. But it\n \n\njust like two or three days ago. But it\nhe put it a really nice way. He said\n\nhe put it a really nice way. He said\n \n\nhe put it a really nice way. He said\nit's a it's built with a systems mindset\n\nit's a it's built with a systems mindset\n \n\nit's a it's built with a systems mindset\nand it's really about how you're\n\nand it's really about how you're\n \n\nand it's really about how you're\nencoding or expressing your intent of\n\nencoding or expressing your intent of\n \n\nencoding or expressing your intent of\nwhat you want to do most importantly in\n\nwhat you want to do most importantly in\n \n\nwhat you want to do most importantly in\na way that's transferable. So the the\n\na way that's transferable. So the the\n \n\na way that's transferable. So the the\ndesign of your system, I would imagine,\n\ndesign of your system, I would imagine,\n \n\ndesign of your system, I would imagine,\nor your program isn't going to move\n\nor your program isn't going to move\n \n\nor your program isn't going to move\nnecessarily as quickly as maybe the\n\nnecessarily as quickly as maybe the\n \n\nnecessarily as quickly as maybe the\nmodel capabilities are under the hood. model capabilities are under the hood. model capabilities are under the hood." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nwhen we see new releases almost every\n\nwhen we see new releases almost every\n \n\nwhen we see new releases almost every\nsingle day, different capabilities,\n\nsingle day, different capabilities,\n \n\nsingle day, different capabilities,\nbetter models and so DSPI allows you to\n\nbetter models and so DSPI all..." }, { "role": "assistant", "content": "when we see new releases almost every\n\nwhen we see new releases almost every\n \n\nwhen we see new releases almost every\nsingle day, different capabilities,\n\nsingle day, different capabilities,\n \n\nsingle day, different capabilities,\nbetter models and so DSPI allows you to\n\nbetter models and so DSPI allows you to\n \n\nbetter models and so DSPI allows you to\nstructure it in a way that retains the\n\nstructure it in a way that retains the\n \n\nstructure it in a way that retains the\ncontrol flow uh retains the intent of\n\ncontrol flow uh retains the intent of\n \n\ncontrol flow uh retains the intent of\nyour system, your program um while\n\nyour system, your program um while\n \n\nyour system, your program um while\nallowing you to bounce from model to\n\nallowing you to bounce from model to\n \n\nallowing you to bounce from model to\nmodel to the extent that you want to or\n\nmodel to the extent that you want to or\n \n\nmodel to the extent that you want to or\nneed to. need to. need to. Convenience comes for free. There's no\n\nConvenience comes for free. There's no\n \n\nConvenience comes for free. There's no\nparsing, JSON, things like that. It\n\nparsing, JSON, things like that. It\n \n\nparsing, JSON, things like that. It\nagain, it sits at a nice level of\n\nagain, it sits at a nice level of\n \n\nagain, it sits at a nice level of\nabstraction where you can still\n\nabstraction where you can still\n \n\nabstraction where you can still\nunderstand what's going on under the\n\nunderstand what's going on under the\n \n\nunderstand what's going on under the\nhood. If you want to, you can go in and\n\nhood. If you want to, you can go in and\n \n\nhood. If you want to, you can go in and\ntweak things, but it allows you to to\n\ntweak things, but it allows you to to\n \n\ntweak things, but it allows you to to\nkind of focus on just what you want to\n\nkind of focus on just what you want to\n \n\nkind of focus on just what you want to\ndo while retaining the level of\n\ndo while retaining the level of\n \n\ndo while retaining the level of\nprecision that you that I think most of\n\nprecision that you that I think most of\n \n\nprecision that you that I think most of\nus would like to have in and kind of\n\nus would like to have in and kind of\n \n\nus would like to have in and kind of\nbuilding your programs. Um, [snorts]\n\nbuilding your programs. Um, [snorts]\n \n\nbuilding your programs. Um, [snorts]\nas mentioned, it's robust to kind of\n\nas mentioned, it's robust to kind of\n \n\nas mentioned, it's robust to kind of\nmodel and paradigm shifts. So, you can\n\nmodel and paradigm shifts. So, you can\n \n\nmodel and paradigm shifts. So, you can\nagain keep the logic of your program. Um\n\nagain keep the logic of your program. Um\n \n\nagain keep the logic of your program. Um\nbut it but keep that those LLMs infused\n\nbut it but keep that those LLMs infused\n \n\nbut it but keep that those LLMs infused\nin uh basically in line. Now that being\n\nin uh basically in line. Now that being\n \n\nin uh basically in line. Now that being\nsaid, you know, there are absolutely\n\nsaid, you know, there are absolutely\n \n\nsaid, you know, there are absolutely\nother great libraries out there. other great libraries out there. other great libraries out there. Pedantic AI, Langchain, there's many\n\nPedantic AI, Langchain, there's many\n \n\nPedantic AI, Langchain, there's many\nmany others that allow you to do similar\n\nmany others that allow you to do similar\n \n\nmany others that allow you to do similar\nthings. Agno is another one. Um this is\n\nthings. Agno is another one. Um this is\n \n\nthings. Agno is another one. Um this is\njust one perspective and um it may not\n\njust one perspective and um it may not\n \n\njust one perspective and um it may not\nbe perfect for your use case. For me, it\n\nbe perfect for your use case. For me, it\n \n\nbe perfect for your use case. For me, it\ntook me a little bit to kind of gro how\n\ntook me a little bit to kind of gro how\n \n\ntook me a little bit to kind of gro how\nDSPI works and you'll see why that is in\n\nDSPI works and you'll see why that is in\n \n\nDSPI works and you'll see why that is in\na minute. Um, so I would just recommend\n\na minute. Um, so I would just recommend\n \n\na minute. Um, so I would just recommend\nkind of have an have an open mind, play\n\nkind of have an have an open mind, play\n \n\nkind of have an have an open mind, play\nwith it. Um, run the code, tweak the\n\nwith it. Um, run the code, tweak the\n \n\nwith it. Um, run the code, tweak the\ncode, do whatever you need to do. Um,\n\ncode, do whatever you need to do. Um,\n \n\ncode, do whatever you need to do. Um,\nand just see how it might work, might\n\nand just see how it might work, might\n \n\nand just see how it might work, might\nwork for you. And really, this talk is\n\nwork for you. And really, this talk is\n \n\nwork for you. And really, this talk is\nmore about ways that I found it useful. more about ways that I found it useful. more about ways that I found it useful. It's not a dissertation on the ins and\n\nIt's not a dissertation on the ins and\n \n\nIt's not a dissertation on the ins and\nouts of every nook and cranny of DSPI. outs of every nook and cranny of DSPI. outs of every nook and cranny of DSPI. It's more of, you know, I've run into\n\nIt's more of, you know, I've run into\n \n\nIt's more of, you know, I've run into\nthese problems myself now. I naturally\n\nthese problems myself now. I naturally\n \n\nthese problems myself now. I naturally\nrun to DSPI to solve them. And this is\n\nrun to DSPI to solve them. And this is\n \n\nrun to DSPI to solve them. And this is\nkind of why. And the hope is that you\n\nkind of why. And the hope is that you\n \n\nkind of why. And the hope is that you\ncan extrapolate some of this to your own\n\ncan extrapolate some of this to your own\n \n\ncan extrapolate some of this to your own\nuse cases. So we we'll go through\n\nuse cases. So we we'll go through\n \n\nuse cases. So we we'll go through\neverything uh fairly quickly here, but\n\neverything uh fairly quickly here, but\n \n\neverything uh fairly quickly here, but\nthe core concepts of DSPI really comes\n\nthe core concepts of DSPI really comes\n \n\nthe core concepts of DSPI really comes\ndown to arguably five or these six that\n\ndown to arguably five or these six that\n \n\ndown to arguably five or these six that\nyou see on the screen here. So we'll go\n\nyou see on the screen here. So we'll go\n \n\nyou see on the screen here. So we'll go\ninto each of these in more detail, but\n\ninto each of these in more detail, but\n \n\ninto each of these in more detail, but\nhigh level signatures\n\nhigh level signatures\n \n\nhigh level signatures\nspecify what you want the L what\n\nspecify what you want the L what\n \n\nspecify what you want the L what\nbasically what you want your function\n\nbasically what you want your function\n \n\nbasically what you want your function\ncall to do. This is when you specify\n\ncall to do. This is when you specify\n \n\ncall to do. This is when you specify\nyour inputs, your outputs. Inputs and\n\nyour inputs, your outputs. Inputs and\n \n\nyour inputs, your outputs. Inputs and\noutputs can both be typed. Um, and you\n\noutputs can both be typed. Um, and you\n \n\noutputs can both be typed. Um, and you\ndefer the rest of the basically the how\n\ndefer the rest of the basically the how\n \n\ndefer the rest of the basically the how\nthe implementation of it to the LLM. And\n\nthe implementation of it to the LLM. And\n \n\nthe implementation of it to the LLM. And\nwe'll see how we how that all kind of\n\nwe'll see how we how that all kind of\n \n\nwe'll see how we how that all kind of\ncomes together uh in a minute. Modules\n\ncomes together uh in a minute. Modules\n \n\ncomes together uh in a minute. Modules\nthemselves are ways to logically\n\nthemselves are ways to logically\n \n\nthemselves are ways to logically\nstructure your program. They're based\n\nstructure your program. They're based\n \n\nstructure your program. They're based\noff of signatures. So, a module can have\n\noff of signatures. So, a module can have\n \n\noff of signatures." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nSo, a module can have\none or more signatures embedded within\n\none or more signatures embedded within\n \n\none or more signatures embedded within\nit in addition to uh additional logic. it in addition to uh additional logic. it in addition to uh additional logic. and it's based off of um pietorrch and\n\n..." }, { "role": "assistant", "content": "So, a module can have\none or more signatures embedded within\n\none or more signatures embedded within\n \n\none or more signatures embedded within\nit in addition to uh additional logic. it in addition to uh additional logic. it in addition to uh additional logic. and it's based off of um pietorrch and\n\nand it's based off of um pietorrch and\n \n\nand it's based off of um pietorrch and\nsome of the in terms of like the\n\nsome of the in terms of like the\n \n\nsome of the in terms of like the\nmethodology for how it's structured and\n\nmethodology for how it's structured and\n \n\nmethodology for how it's structured and\nyou'll you'll see how that uh comes to\n\nyou'll you'll see how that uh comes to\n \n\nyou'll you'll see how that uh comes to\nbe in a minute. Tools we're all familiar\n\nbe in a minute. Tools we're all familiar\n \n\nbe in a minute. Tools we're all familiar\nwith tools MCP and others and really\n\nwith tools MCP and others and really\n \n\nwith tools MCP and others and really\ntools fundamentally as DSPI looks at\n\ntools fundamentally as DSPI looks at\n \n\ntools fundamentally as DSPI looks at\nthem are just Python functions. So it's\n\nthem are just Python functions. So it's\n \n\nthem are just Python functions. So it's\njust a way for you to very easily expose\n\njust a way for you to very easily expose\n \n\njust a way for you to very easily expose\nPython functions to the LLM within the\n\nPython functions to the LLM within the\n \n\nPython functions to the LLM within the\nDSP kind of ecosystem if you will. um\n\nDSP kind of ecosystem if you will. um\n \n\nDSP kind of ecosystem if you will. um\nadapters\n\nadapters\n \n\nadapters\nlive in between your signature and the\n\nlive in between your signature and the\n \n\nlive in between your signature and the\nLLM call itself. I mean, as we all know,\n\nLLM call itself. I mean, as we all know,\n \n\nLLM call itself. I mean, as we all know,\nprompts are ultimately just strings of\n\nprompts are ultimately just strings of\n \n\nprompts are ultimately just strings of\ntext that are sent to the LLM. text that are sent to the LLM. text that are sent to the LLM. Signatures are a way for you to express\n\nSignatures are a way for you to express\n \n\nSignatures are a way for you to express\nyour intent at a at a higher level. And\n\nyour intent at a at a higher level. And\n \n\nyour intent at a at a higher level. And\nso, adapters are the things that sit in\n\nso, adapters are the things that sit in\n \n\nso, adapters are the things that sit in\nbetween those two. So, it's how you\n\nbetween those two. So, it's how you\n \n\nbetween those two. So, it's how you\ntranslate your inputs and outputs into a\n\ntranslate your inputs and outputs into a\n \n\ntranslate your inputs and outputs into a\nformat basically explodes out from your\n\nformat basically explodes out from your\n \n\nformat basically explodes out from your\ninitial signature into a format that is\n\ninitial signature into a format that is\n \n\ninitial signature into a format that is\nultimately the prompt that is sent to\n\nultimately the prompt that is sent to\n \n\nultimately the prompt that is sent to\nthe LLM. And so, you know, there's some\n\nthe LLM. And so, you know, there's some\n \n\nthe LLM. And so, you know, there's some\ndebate or some research on if certain\n\ndebate or some research on if certain\n \n\ndebate or some research on if certain\nmodels perform better with XML as an\n\nmodels perform better with XML as an\n \n\nmodels perform better with XML as an\nexample or BAML or JSON or others. And\n\nexample or BAML or JSON or others. And\n \n\nexample or BAML or JSON or others. And\nso adapters give you a nice easy\n\nso adapters give you a nice easy\n \n\nso adapters give you a nice easy\nabstraction to to basically mix and\n\nabstraction to to basically mix and\n \n\nabstraction to to basically mix and\nmatch those at at will as you want. match those at at will as you want. match those at at will as you want. Optimizers um are\n\nOptimizers um are\n \n\nOptimizers um are\nthe most interesting and for whatever\n\nthe most interesting and for whatever\n \n\nthe most interesting and for whatever\nreason the most controversial part of\n\nreason the most controversial part of\n \n\nreason the most controversial part of\nDSP. That's kind of the first thing that\n\nDSP. That's kind of the first thing that\n \n\nDSP. That's kind of the first thing that\npeople think of or at least when they\n\npeople think of or at least when they\n \n\npeople think of or at least when they\nhear of DSP they think optimizers. We'll\n\nhear of DSP they think optimizers. We'll\n \n\nhear of DSP they think optimizers. We'll\nsee a quote in a minute. It's not\n\nsee a quote in a minute. It's not\n \n\nsee a quote in a minute. It's not\noptimizers first. It is just a nice\n\noptimizers first. It is just a nice\n \n\noptimizers first. It is just a nice\nadded benefit and a nice capability that\n\nadded benefit and a nice capability that\n \n\nadded benefit and a nice capability that\nDSPI offers in addition to the ability\n\nDSPI offers in addition to the ability\n \n\nDSPI offers in addition to the ability\nto structure your program with the\n\nto structure your program with the\n \n\nto structure your program with the\nsignatures and modules and everything\n\nsignatures and modules and everything\n \n\nsignatures and modules and everything\nelse. Um, and metrics are used in tandem\n\nelse. Um, and metrics are used in tandem\n \n\nelse. Um, and metrics are used in tandem\nwith optimizers that that basically\n\nwith optimizers that that basically\n \n\nwith optimizers that that basically\ndefines how you measure success in your\n\ndefines how you measure success in your\n \n\ndefines how you measure success in your\nin uh your DSPI program. So the\n\nin uh your DSPI program. So the\n \n\nin uh your DSPI program. So the\noptimizers use the metrics to determine\n\noptimizers use the metrics to determine\n \n\noptimizers use the metrics to determine\nif it's finding the right path if you\n\nif it's finding the right path if you\n \n\nif it's finding the right path if you\nwill. will. will. So signature as I mentioned before it's\n\nSo signature as I mentioned before it's\n \n\nSo signature as I mentioned before it's\nhow you express your intent your\n\nhow you express your intent your\n \n\nhow you express your intent your\ndeclarative intent can be super simple\n\ndeclarative intent can be super simple\n \n\ndeclarative intent can be super simple\nstrings and this is the weirdest part\n\nstrings and this is the weirdest part\n \n\nstrings and this is the weirdest part\nfor me initially but is one of the most\n\nfor me initially but is one of the most\n \n\nfor me initially but is one of the most\npowerful parts uh of it now or it can be\n\npowerful parts uh of it now or it can be\n \n\npowerful parts uh of it now or it can be\nmore complicated class-based classbased\n\nmore complicated class-based classbased\n \n\nmore complicated class-based classbased\nobjects if you've used pyantic it that's\n\nobjects if you've used pyantic it that's\n \n\nobjects if you've used pyantic it that's\nbasically what what it runs on under the\n\nbasically what what it runs on under the\n \n\nbasically what what it runs on under the\nhood. hood. hood. So this is an example of one of the\n\nSo this is an example of one of the\n \n\nSo this is an example of one of the\nclass-based signatures. Again, it it's\n\nclass-based signatures. Again, it it's\n \n\nclass-based signatures. Again, it it's\nbasically just a pyantic object. basically just a pyantic object. basically just a pyantic object. What's super interesting about this is\n\nWhat's super interesting about this is\n \n\nWhat's super interesting about this is\nthat\n\nthat\n \n\nthat\nthe\n\nthe\n \n\nthe\nthe names of the fields themselves act\n\nthe names of the fields themselves act\n \n\nthe names of the fields themselves act\nalmost as like mini prompts. It's part\n\nalmost as like mini prompts. It's part\n \n\nalmost as like mini prompts. It's part\nof the prompt itself. And you'll see how\n\nof the prompt itself. And you'll see how\n \n\nof the prompt itself. And you'll see how\nthis comes to life in a minute. But\n\nthis comes to life in a minute. But\n \n\nthis comes to life in a minute." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nBut\nwhat's ultimately passed to the model\n\nwhat's ultimately passed to the model\n \n\nwhat's ultimately passed to the model\nfrom something like this is it will say\n\nfrom something like this is it will say\n \n\nfrom something like this is it will say\nokay your inputs are going to be a\n\nokay your inputs a..." }, { "role": "assistant", "content": "But\nwhat's ultimately passed to the model\n\nwhat's ultimately passed to the model\n \n\nwhat's ultimately passed to the model\nfrom something like this is it will say\n\nfrom something like this is it will say\n \n\nfrom something like this is it will say\nokay your inputs are going to be a\n\nokay your inputs are going to be a\n \n\nokay your inputs are going to be a\nparameter called text and it's based off\n\nparameter called text and it's based off\n \n\nparameter called text and it's based off\nof the name of the that particular\n\nof the name of the that particular\n \n\nof the name of the that particular\nparameter in this class. And so these\n\nparameter in this class. And so these\n \n\nparameter in this class. And so these\nthings are actually passed through. And\n\nthings are actually passed through. And\n \n\nthings are actually passed through. And\nso it's it's very important uh to be\n\nso it's it's very important uh to be\n \n\nso it's it's very important uh to be\nable to name your parameters in a way\n\nable to name your parameters in a way\n \n\nable to name your parameters in a way\nthat is intuitive for the model to be\n\nthat is intuitive for the model to be\n \n\nthat is intuitive for the model to be\nable to pick it up. Um, and you can add\n\nable to pick it up. Um, and you can add\n \n\nable to pick it up. Um, and you can add\nsome additional context or what have you\n\nsome additional context or what have you\n \n\nsome additional context or what have you\nin the description field here. So most\n\nin the description field here. So most\n \n\nin the description field here. So most\nof this, if not all of this, yes, it is\n\nof this, if not all of this, yes, it is\n \n\nof this, if not all of this, yes, it is\nproper, you know, typed Python code, but\n\nproper, you know, typed Python code, but\n \n\nproper, you know, typed Python code, but\nit's also it also serves almost as a\n\nit's also it also serves almost as a\n \n\nit's also it also serves almost as a\nprompt ultimately that feeds into the\n\nprompt ultimately that feeds into the\n \n\nprompt ultimately that feeds into the\nmodel. Um, and that's basically\n\nmodel. Um, and that's basically\n \n\nmodel. Um, and that's basically\ntranslated through the use of adapters. translated through the use of adapters. translated through the use of adapters. Um, and so just to highlight here like\n\nUm, and so just to highlight here like\n \n\nUm, and so just to highlight here like\nthese, it's the ones that are a little\n\nthese, it's the ones that are a little\n \n\nthese, it's the ones that are a little\nbit darker and bold, you know, those are\n\nbit darker and bold, you know, those are\n \n\nbit darker and bold, you know, those are\nthe things that are effectively part of\n\nthe things that are effectively part of\n \n\nthe things that are effectively part of\nthe prompt. uh that's been sent in and\n\nthe prompt. uh that's been sent in and\n \n\nthe prompt. uh that's been sent in and\nyou'll see kind of how DSPI works with\n\nyou'll see kind of how DSPI works with\n \n\nyou'll see kind of how DSPI works with\nall this and formats it in a way that\n\nall this and formats it in a way that\n \n\nall this and formats it in a way that\nagain allow you to just worry about what\n\nagain allow you to just worry about what\n \n\nagain allow you to just worry about what\nyou want. Worry about constructing your\n\nyou want. Worry about constructing your\n \n\nyou want. Worry about constructing your\nsignature instead of figuring out how\n\nsignature instead of figuring out how\n \n\nsignature instead of figuring out how\nbest to word something in the prompt. Go\n\nbest to word something in the prompt. Go\n \n\nbest to word something in the prompt. Go\n>> ahead. I have a really good prompt. I have a really good prompt. I have a really good prompt. >> Sure. Then I don't want this thing. >> That's exactly right. >> Sure. >> Sure. >> Sure. >> So the the question for folks online is\n\n>> So the the question for folks online is\n \n\n>> So the the question for folks online is\nwhat if I already have a great prompt? what if I already have a great prompt? what if I already have a great prompt? I've done all this work. I'm a I'm a\n\nI've done all this work. I'm a I'm a\n \n\nI've done all this work. I'm a I'm a\namazing prompt engineer. I don't want my\n\namazing prompt engineer. I don't want my\n \n\namazing prompt engineer. I don't want my\njob to go away or whatever. Um, yes. So,\n\njob to go away or whatever. Um, yes. So,\n \n\njob to go away or whatever. Um, yes. So,\nyou can absolutely start with a custom\n\nyou can absolutely start with a custom\n \n\nyou can absolutely start with a custom\nprompt or something that you have\n\nprompt or something that you have\n \n\nprompt or something that you have\ndemonstrated works really well. And\n\ndemonstrated works really well. And\n \n\ndemonstrated works really well. And\nyou're exactly right that's that can be\n\nyou're exactly right that's that can be\n \n\nyou're exactly right that's that can be\ndone in the dock string itself. There's\n\ndone in the dock string itself. There's\n \n\ndone in the dock string itself. There's\nthere's some other methods in order uh\n\nthere's some other methods in order uh\n \n\nthere's some other methods in order uh\nfor you to inject basically system\n\nfor you to inject basically system\n \n\nfor you to inject basically system\ninstructions or add additional things at\n\ninstructions or add additional things at\n \n\ninstructions or add additional things at\ncertain parts of the ultimate prompt and\n\ncertain parts of the ultimate prompt and\n \n\ncertain parts of the ultimate prompt and\nor of course you can just inject it in\n\nor of course you can just inject it in\n \n\nor of course you can just inject it in\nthe in the final string anyway. I mean\n\nthe in the final string anyway. I mean\n \n\nthe in the final string anyway. I mean\nit's just you know a string that is\n\nit's just you know a string that is\n \n\nit's just you know a string that is\nconstructed by VSPI. So um absolutely\n\nconstructed by VSPI. So um absolutely\n \n\nconstructed by VSPI. So um absolutely\nthis doesn't necessarily prevent you it\n\nthis doesn't necessarily prevent you it\n \n\nthis doesn't necessarily prevent you it\ndoes does not prevent you from adding in\n\ndoes does not prevent you from adding in\n \n\ndoes does not prevent you from adding in\nsome super prompt that you already have. some super prompt that you already have. some super prompt that you already have. Absolutely. Um and to your point it is\n\nAbsolutely. Um and to your point it is\n \n\nAbsolutely. Um and to your point it is\nit can serve as a nice starting point\n\nit can serve as a nice starting point\n \n\nit can serve as a nice starting point\nfrom which to build the rest of the\n\nfrom which to build the rest of the\n \n\nfrom which to build the rest of the\nsystem. system. system. Here's a shorthand version of the same\n\nHere's a shorthand version of the same\n \n\nHere's a shorthand version of the same\nexact thing which to me the first time I\n\nexact thing which to me the first time I\n \n\nexact thing which to me the first time I\nsaw this so this was like baffling to\n\nsaw this so this was like baffling to\n \n\nsaw this so this was like baffling to\nme. Um, but it it that's exactly how it\n\nme. Um, but it it that's exactly how it\n \n\nme. Um, but it it that's exactly how it\nworks is that you're basically again\n\nworks is that you're basically again\n \n\nworks is that you're basically again\nkind of deferring the implementation or\n\nkind of deferring the implementation or\n \n\nkind of deferring the implementation or\nthe logic or what have you to DSPI and\n\nthe logic or what have you to DSPI and\n \n\nthe logic or what have you to DSPI and\nthe model to basically figure out what\n\nthe model to basically figure out what\n \n\nthe model to basically figure out what\nyou want to do. So in this case, if I\n\nyou want to do. So in this case, if I\n \n\nyou want to do. So in this case, if I\nwant a super super simple text uh\n\nwant a super super simple text uh\n \n\nwant a super super simple text uh\nsentiment classifier, this is basically\n\nsentiment classifier, this is basically\n \n\nsentiment classifier, this is basically\nall you need. You're just saying, okay,\n\nall you need." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nYou're just saying, okay,\n \n\nall you need. You're just saying, okay,\nI'm going to give you text as an input. I'm going to give you text as an input. I'm going to give you text as an input. I want the sentiment as an integer as\n\nI want the sentiment as an integer as\n \n\nI want the sentiment as an inte..." }, { "role": "assistant", "content": "You're just saying, okay,\n \n\nall you need. You're just saying, okay,\nI'm going to give you text as an input. I'm going to give you text as an input. I'm going to give you text as an input. I want the sentiment as an integer as\n\nI want the sentiment as an integer as\n \n\nI want the sentiment as an integer as\nthe output. Now you probably want to\n\nthe output. Now you probably want to\n \n\nthe output. Now you probably want to\nspecify some additional instructions to\n\nspecify some additional instructions to\n \n\nspecify some additional instructions to\nsay okay your sentiment you know a lower\n\nsay okay your sentiment you know a lower\n \n\nsay okay your sentiment you know a lower\nnumber means negative you know a higher\n\nnumber means negative you know a higher\n \n\nnumber means negative you know a higher\nnumber is more positive sentiment etc. number is more positive sentiment etc. number is more positive sentiment etc. But it just gives you a nice kind of\n\nBut it just gives you a nice kind of\n \n\nBut it just gives you a nice kind of\neasy way to to kind of scaffold these\n\neasy way to to kind of scaffold these\n \n\neasy way to to kind of scaffold these\nthings out in a way that you don't have\n\nthings out in a way that you don't have\n \n\nthings out in a way that you don't have\nto worry about like you know creating\n\nto worry about like you know creating\n \n\nto worry about like you know creating\nthis whole prompt from hand. It's like\n\nthis whole prompt from hand. It's like\n \n\nthis whole prompt from hand. It's like\nokay I just want to see how this works\n\nokay I just want to see how this works\n \n\nokay I just want to see how this works\nand then if it works then I can add the\n\nand then if it works then I can add the\n \n\nand then if it works then I can add the\nadditional instructions then I can\n\nadditional instructions then I can\n \n\nadditional instructions then I can\ncreate a module out of it or you know\n\ncreate a module out of it or you know\n \n\ncreate a module out of it or you know\nwhatever it might be. It's these\n\nwhatever it might be. It's these\n \n\nwhatever it might be. It's these\nshorthand\n\nshorthand\n \n\nshorthand\nor it is this shorthand that makes\n\nor it is this shorthand that makes\n \n\nor it is this shorthand that makes\nexperimentation and iteration incredibly\n\nexperimentation and iteration incredibly\n \n\nexperimentation and iteration incredibly\nquick. quick. quick. So modules it's that base abstraction\n\nSo modules it's that base abstraction\n \n\nSo modules it's that base abstraction\nlayer for DSPI programs. There are a\n\nlayer for DSPI programs. There are a\n \n\nlayer for DSPI programs. There are a\nbunch of modules that are built in and\n\nbunch of modules that are built in and\n \n\nbunch of modules that are built in and\nthese are a collection of kind of\n\nthese are a collection of kind of\n \n\nthese are a collection of kind of\nprompting techniques if you will and you\n\nprompting techniques if you will and you\n \n\nprompting techniques if you will and you\ncan always create your own module. So to\n\ncan always create your own module. So to\n \n\ncan always create your own module. So to\nthe question before, if you have\n\nthe question before, if you have\n \n\nthe question before, if you have\nsomething that you know works really\n\nsomething that you know works really\n \n\nsomething that you know works really\nwell, sure yeah, put it in the module. well, sure yeah, put it in the module. well, sure yeah, put it in the module. That's now the kind of the base\n\nThat's now the kind of the base\n \n\nThat's now the kind of the base\nassumption, the base module that others\n\nassumption, the base module that others\n \n\nassumption, the base module that others\ncan build off of. And all of DSPI is\n\ncan build off of. And all of DSPI is\n \n\ncan build off of. And all of DSPI is\nmeant to be composable, optimizable, and\n\nmeant to be composable, optimizable, and\n \n\nmeant to be composable, optimizable, and\nwhen you deconstruct your business logic\n\nwhen you deconstruct your business logic\n \n\nwhen you deconstruct your business logic\nor whatever you're trying to achieve by\n\nor whatever you're trying to achieve by\n \n\nor whatever you're trying to achieve by\nusing these different primitives, it all\n\nusing these different primitives, it all\n \n\nusing these different primitives, it all\nit's intended to kind of fit together\n\nit's intended to kind of fit together\n \n\nit's intended to kind of fit together\nand flow together. Um, and we'll get to\n\nand flow together. Um, and we'll get to\n \n\nand flow together. Um, and we'll get to\noptimizers in a minute, but at least for\n\noptimizers in a minute, but at least for\n \n\noptimizers in a minute, but at least for\nme and my team's experience, just being\n\nme and my team's experience, just being\n \n\nme and my team's experience, just being\nable to logically separate the different\n\nable to logically separate the different\n \n\nable to logically separate the different\ncomponents of a program, but basically\n\ncomponents of a program, but basically\n \n\ncomponents of a program, but basically\ninlining uh LLM calls has been\n\ninlining uh LLM calls has been\n \n\ninlining uh LLM calls has been\nincredibly powerful for us. And it's\n\nincredibly powerful for us. And it's\n \n\nincredibly powerful for us. And it's\njust an added benefit that at the end of\n\njust an added benefit that at the end of\n \n\njust an added benefit that at the end of\nthe day, because we're just kind of in\n\nthe day, because we're just kind of in\n \n\nthe day, because we're just kind of in\nthe DSPI paradigm, we happen to also be\n\nthe DSPI paradigm, we happen to also be\n \n\nthe DSPI paradigm, we happen to also be\nable to optimize it at the end of the\n\nable to optimize it at the end of the\n \n\nable to optimize it at the end of the\nday. Uh, so it comes with a bunch of\n\nday. Uh, so it comes with a bunch of\n \n\nday. Uh, so it comes with a bunch of\nstandard ones built in. I I don't use\n\nstandard ones built in. I I don't use\n \n\nstandard ones built in. I I don't use\nsome of these bottom ones as much,\n\nsome of these bottom ones as much,\n \n\nsome of these bottom ones as much,\nalthough it's they're super interesting. although it's they're super interesting. although it's they're super interesting. Um the base one at the top there is just\n\nUm the base one at the top there is just\n \n\nUm the base one at the top there is just\nDSpi.predict. DSpi.predict. DSpi.predict. That's literally just, you know, an LM\n\nThat's literally just, you know, an LM\n \n\nThat's literally just, you know, an LM\ncall. That's just uh a vanilla call. call. That's just uh a vanilla call. call. That's just uh a vanilla call. chain of thought uh probably isn't isn't\n\nchain of thought uh probably isn't isn't\n \n\nchain of thought uh probably isn't isn't\nas relevant anymore these days because\n\nas relevant anymore these days because\n \n\nas relevant anymore these days because\nmodels have kind of ironed those out but\n\nmodels have kind of ironed those out but\n \n\nmodels have kind of ironed those out but\num it is a good example of the types of\n\num it is a good example of the types of\n \n\num it is a good example of the types of\num kind of prompting techniques that can\n\num kind of prompting techniques that can\n \n\num kind of prompting techniques that can\nbe built into some of these modules um\n\nbe built into some of these modules um\n \n\nbe built into some of these modules um\nand basically all this does is add um\n\nand basically all this does is add um\n \n\nand basically all this does is add um\nsome some of the uh strings from\n\nsome some of the uh strings from\n \n\nsome some of the uh strings from\nliterature to say okay let's think step\n\nliterature to say okay let's think step\n \n\nliterature to say okay let's think step\nby step or whatever that might be same\n\nby step or whatever that might be same\n \n\nby step or whatever that might be same\nthing for react and codeact react is\n\nthing for react and codeact react is\n \n\nthing for react and codeact react is\nbasically the way that you expose the\n\nbasically the way that you expose the\n \n\nbasically the way that you expose the\ntools to the model. So, it's wrapping\n\ntools to the model. So, it's wrapping\n \n\ntools to the model." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nSo, it's wrapping\nand doing some things under the hood\n\nand doing some things under the hood\n \n\nand doing some things under the hood\nwith um basically taking your signatures\n\nwith um basically taking your signatures\n \n\nwith um basically taking your signatures\nand uh it's injecting the Python\n\nand uh..." }, { "role": "assistant", "content": "So, it's wrapping\nand doing some things under the hood\n\nand doing some things under the hood\n \n\nand doing some things under the hood\nwith um basically taking your signatures\n\nwith um basically taking your signatures\n \n\nwith um basically taking your signatures\nand uh it's injecting the Python\n\nand uh it's injecting the Python\n \n\nand uh it's injecting the Python\nfunctions that you've given it as tools\n\nfunctions that you've given it as tools\n \n\nfunctions that you've given it as tools\nand basically React is how you do tool\n\nand basically React is how you do tool\n \n\nand basically React is how you do tool\ncalling in DSP. calling in DSP. calling in DSP. Program with thought is uh is pretty\n\nProgram with thought is uh is pretty\n \n\nProgram with thought is uh is pretty\ncool. It kind of forces the model to\n\ncool. It kind of forces the model to\n \n\ncool. It kind of forces the model to\nthink in code and then we'll return the\n\nthink in code and then we'll return the\n \n\nthink in code and then we'll return the\nresult. Um, and you can give it a, it\n\nresult. Um, and you can give it a, it\n \n\nresult. Um, and you can give it a, it\ncomes with a Python interpreter built\n\ncomes with a Python interpreter built\n \n\ncomes with a Python interpreter built\nin, but you can give it some custom one,\n\nin, but you can give it some custom one,\n \n\nin, but you can give it some custom one,\nsome type of custom harness if you\n\nsome type of custom harness if you\n \n\nsome type of custom harness if you\nwanted to. Um, I haven't played with\n\nwanted to. Um, I haven't played with\n \n\nwanted to. Um, I haven't played with\nthat one too too much, but it is super\n\nthat one too too much, but it is super\n \n\nthat one too too much, but it is super\ninteresting. If you have like a highly\n\ninteresting. If you have like a highly\n \n\ninteresting. If you have like a highly\ntechnical problem or workflow or\n\ntechnical problem or workflow or\n \n\ntechnical problem or workflow or\nsomething like that where you want the\n\nsomething like that where you want the\n \n\nsomething like that where you want the\nmodel to inject reasoning in code at\n\nmodel to inject reasoning in code at\n \n\nmodel to inject reasoning in code at\ncertain parts of your pipeline, that's\n\ncertain parts of your pipeline, that's\n \n\ncertain parts of your pipeline, that's\nthat's kind of an really easy way to do\n\nthat's kind of an really easy way to do\n \n\nthat's kind of an really easy way to do\nit. And then some of these other ones\n\nit. And then some of these other ones\n \n\nit. And then some of these other ones\nare basically just different\n\nare basically just different\n \n\nare basically just different\nmethodologies for comparing outputs or\n\nmethodologies for comparing outputs or\n \n\nmethodologies for comparing outputs or\nrunning things in parallel. running things in parallel. running things in parallel. So here's what one looks like. Again,\n\nSo here's what one looks like. Again,\n \n\nSo here's what one looks like. Again,\nit's it's fairly simple. It's, you know,\n\nit's it's fairly simple. It's, you know,\n \n\nit's it's fairly simple. It's, you know,\nit is a Python class at the end of the\n\nit is a Python class at the end of the\n \n\nit is a Python class at the end of the\nday. Um, and so you do some initial\n\nday. Um, and so you do some initial\n \n\nday. Um, and so you do some initial\ninitialization up top. In this case,\n\ninitialization up top. In this case,\n \n\ninitialization up top. In this case,\nyou're seeing the uh\n\nyou're seeing the uh\n \n\nyou're seeing the uh\nuh the shorthand signature up there. So,\n\nuh the shorthand signature up there. So,\n \n\nuh the shorthand signature up there. So,\nI'm this module uh just to give you some\n\nI'm this module uh just to give you some\n \n\nI'm this module uh just to give you some\ncontext is an excerpt from um one of the\n\ncontext is an excerpt from um one of the\n \n\ncontext is an excerpt from um one of the\nthe Python um files that's in the repo\n\nthe Python um files that's in the repo\n \n\nthe Python um files that's in the repo\nis basically taking in a bunch of time\n\nis basically taking in a bunch of time\n \n\nis basically taking in a bunch of time\nentries and making sure that they adhere\n\nentries and making sure that they adhere\n \n\nentries and making sure that they adhere\nto certain standards, making sure that\n\nto certain standards, making sure that\n \n\nto certain standards, making sure that\nthings are capitalized properly or that\n\nthings are capitalized properly or that\n \n\nthings are capitalized properly or that\nthere are periods at the end of the\n\nthere are periods at the end of the\n \n\nthere are periods at the end of the\nsentences or whatever it might be. sentences or whatever it might be. sentences or whatever it might be. that's from a real client use case where\n\nthat's from a real client use case where\n \n\nthat's from a real client use case where\nthey had hundreds of thousands of time\n\nthey had hundreds of thousands of time\n \n\nthey had hundreds of thousands of time\nentries and they needed to make sure\n\nentries and they needed to make sure\n \n\nentries and they needed to make sure\nthat they all adhere to the same format. that they all adhere to the same format. that they all adhere to the same format. This was one way to to kind of do that\n\nThis was one way to to kind of do that\n \n\nThis was one way to to kind of do that\nvery elegantly, at least in my opinion,\n\nvery elegantly, at least in my opinion,\n \n\nvery elegantly, at least in my opinion,\nwas taking up top you can define the the\n\nwas taking up top you can define the the\n \n\nwas taking up top you can define the the\nsignature. It's adding the some\n\nsignature. It's adding the some\n \n\nsignature. It's adding the some\nadditional instructions that were\n\nadditional instructions that were\n \n\nadditional instructions that were\ndefined elsewhere and then saying for\n\ndefined elsewhere and then saying for\n \n\ndefined elsewhere and then saying for\nthis module the the change tense um call\n\nthis module the the change tense um call\n \n\nthis module the the change tense um call\nis going to be just a vanilla predict\n\nis going to be just a vanilla predict\n \n\nis going to be just a vanilla predict\ncall. And then when you actually call\n\ncall. And then when you actually call\n \n\ncall. And then when you actually call\nthe module, you enter into the forward\n\nthe module, you enter into the forward\n \n\nthe module, you enter into the forward\nfunction which you can inter basically\n\nfunction which you can inter basically\n \n\nfunction which you can inter basically\nintersperse the LLM call which would be\n\nintersperse the LLM call which would be\n \n\nintersperse the LLM call which would be\nthe first one and then do some kind of\n\nthe first one and then do some kind of\n \n\nthe first one and then do some kind of\nhard-coded business logic beneath it. hard-coded business logic beneath it. hard-coded business logic beneath it. Uh tools as I mentioned before these are\n\nUh tools as I mentioned before these are\n \n\nUh tools as I mentioned before these are\njust vanilla kind of Python functions. just vanilla kind of Python functions. just vanilla kind of Python functions. It's the DSP's tool interface. So under\n\nIt's the DSP's tool interface. So under\n \n\nIt's the DSP's tool interface. So under\nthe hood, DSPI uses light LLM. And so\n\nthe hood, DSPI uses light LLM. And so\n \n\nthe hood, DSPI uses light LLM. And so\nthere needs to be some kind of coupling\n\nthere needs to be some kind of coupling\n \n\nthere needs to be some kind of coupling\nbetween the two, but fundamentally um\n\nbetween the two, but fundamentally um\n \n\nbetween the two, but fundamentally um\nany type of tool that would that you\n\nany type of tool that would that you\n \n\nany type of tool that would that you\nwould use elsewhere, you can also use in\n\nwould use elsewhere, you can also use in\n \n\nwould use elsewhere, you can also use in\nin DSPI. And this is probably obvious to\n\nin DSPI. And this is probably obvious to\n \n\nin DSPI. And this is probably obvious to\nmost of you, but here's just an example. most of you, but here's just an example. most of you, but here's just an example. You have two functions, get weather,\n\nYou have two functions, get weather,\n \n\nYou have two functions, get weather,\nsearch web. You include that with a\n\nsearch web." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nYou include that with a\n \n\nsearch web. You include that with a\nsignature. So in this case, I'm saying\n\nsignature. So in this case, I'm saying\n \n\nsignature. So in this case, I'm saying\nthe signature is I'm going to give you a\n\nthe signature is I'm going to give you a\n \n\nthe signature is I'm going to ..." }, { "role": "assistant", "content": "You include that with a\n \n\nsearch web. You include that with a\nsignature. So in this case, I'm saying\n\nsignature. So in this case, I'm saying\n \n\nsignature. So in this case, I'm saying\nthe signature is I'm going to give you a\n\nthe signature is I'm going to give you a\n \n\nthe signature is I'm going to give you a\nquestion. please give me an answer. I'm\n\nquestion. please give me an answer. I'm\n \n\nquestion. please give me an answer. I'm\nnot even specifying the types. It's just\n\nnot even specifying the types. It's just\n \n\nnot even specifying the types. It's just\ngoing to infer what that means. Uh I'm\n\ngoing to infer what that means. Uh I'm\n \n\ngoing to infer what that means. Uh I'm\ngiving it the get weather and the search\n\ngiving it the get weather and the search\n \n\ngiving it the get weather and the search\nweb tools and I'm saying, okay, do your\n\nweb tools and I'm saying, okay, do your\n \n\nweb tools and I'm saying, okay, do your\nthing, but only go five rounds just so\n\nthing, but only go five rounds just so\n \n\nthing, but only go five rounds just so\nit doesn't spin off and do something\n\nit doesn't spin off and do something\n \n\nit doesn't spin off and do something\ncrazy. And then a call here is literally\n\ncrazy. And then a call here is literally\n \n\ncrazy. And then a call here is literally\njust calling the React agent that I\n\njust calling the React agent that I\n \n\njust calling the React agent that I\ncreated above with the question, what's\n\ncreated above with the question, what's\n \n\ncreated above with the question, what's\nthe weather like in Tokyo? We'll see an\n\nthe weather like in Tokyo? We'll see an\n \n\nthe weather like in Tokyo? We'll see an\nexample of this in the code session, but\n\nexample of this in the code session, but\n \n\nexample of this in the code session, but\nbasically what this would do is give the\n\nbasically what this would do is give the\n \n\nbasically what this would do is give the\nmodel the prompt, the tools, and let it\n\nmodel the prompt, the tools, and let it\n \n\nmodel the prompt, the tools, and let it\ndo its thing. do its thing. do its thing. So adapters, before I cover this a\n\nSo adapters, before I cover this a\n \n\nSo adapters, before I cover this a\nlittle bit, they're basically prompt\n\nlittle bit, they're basically prompt\n \n\nlittle bit, they're basically prompt\nformatterers, if you will. So the\n\nformatterers, if you will. So the\n \n\nformatterers, if you will. So the\ndescription from the docs probably says\n\ndescription from the docs probably says\n \n\ndescription from the docs probably says\nit best. It's you know it takes your\n\nit best. It's you know it takes your\n \n\nit best. It's you know it takes your\nsignature the inputs other attributes\n\nsignature the inputs other attributes\n \n\nsignature the inputs other attributes\nand it converts them into some type of\n\nand it converts them into some type of\n \n\nand it converts them into some type of\nmessage format that you have specified\n\nmessage format that you have specified\n \n\nmessage format that you have specified\nor that the adapter has specified and so\n\nor that the adapter has specified and so\n \n\nor that the adapter has specified and so\nas an example the JSON adapter taking\n\nas an example the JSON adapter taking\n \n\nas an example the JSON adapter taking\nsay a pyantic object that we defined\n\nsay a pyantic object that we defined\n \n\nsay a pyantic object that we defined\nbefore this is the actual prompt that's\n\nbefore this is the actual prompt that's\n \n\nbefore this is the actual prompt that's\nsent into the LLM and so you can see the\n\nsent into the LLM and so you can see the\n \n\nsent into the LLM and so you can see the\ninput fields so this would have been\n\ninput fields so this would have been\n \n\ninput fields so this would have been\ndefined as okay clinical note type\n\ndefined as okay clinical note type\n \n\ndefined as okay clinical note type\nstring patient info as a patient details\n\nstring patient info as a patient details\n \n\nstring patient info as a patient details\nobject object which which would have\n\nobject object which which would have\n \n\nobject object which which would have\nbeen defined elsewhere and then this is\n\nbeen defined elsewhere and then this is\n \n\nbeen defined elsewhere and then this is\nthe definition of the patient info. It's\n\nthe definition of the patient info. It's\n \n\nthe definition of the patient info. It's\nbasically a JSON dump of that pantic\n\nbasically a JSON dump of that pantic\n \n\nbasically a JSON dump of that pantic\nobject. Go ahead. object. Go ahead. object. Go ahead. >> So this idea there's like a base adapter\n\n>> So this idea there's like a base adapter\n \n\n>> So this idea there's like a base adapter\ndefault that's good for most cases and\n\ndefault that's good for most cases and\n \n\ndefault that's good for most cases and\nthis is if you want to tweak that to do\n\nthis is if you want to tweak that to do\n \n\nthis is if you want to tweak that to do\nsomething more specific. something more specific. something more specific. >> That's right. >> That's right. >> That's right. >> Yeah. The question was if if there's a\n\n>> Yeah. The question was if if there's a\n \n\n>> Yeah. The question was if if there's a\nbase adapter and would this be an\n\nbase adapter and would this be an\n \n\nbase adapter and would this be an\nexample of where you want to do\n\nexample of where you want to do\n \n\nexample of where you want to do\nsomething specific? Answer is yes. So um\n\nsomething specific? Answer is yes. So um\n \n\nsomething specific? Answer is yes. So um\nit's a guy pashant who is um I have his\n\nit's a guy pashant who is um I have his\n \n\nit's a guy pashant who is um I have his\nTwitter at the end of this presentation\n\nTwitter at the end of this presentation\n \n\nTwitter at the end of this presentation\nbut he's been great. [clears throat] He\n\nbut he's been great. [clears throat] He\n \n\nbut he's been great. [clears throat] He\ndid some testing comparing the JSON\n\ndid some testing comparing the JSON\n \n\ndid some testing comparing the JSON\nadapter with the BAML adapter. Um and\n\nadapter with the BAML adapter. Um and\n \n\nadapter with the BAML adapter. Um and\nyou can see just intuitively even even\n\nyou can see just intuitively even even\n \n\nyou can see just intuitively even even\nfor us humans the way that this is\n\nfor us humans the way that this is\n \n\nfor us humans the way that this is\nformatted is a little bit more\n\nformatted is a little bit more\n \n\nformatted is a little bit more\nintuitive. It's probably more token\n\nintuitive. It's probably more token\n \n\nintuitive. It's probably more token\nefficient too just considering like if\n\nefficient too just considering like if\n \n\nefficient too just considering like if\nyou look at the messy JSON that's here\n\nyou look at the messy JSON that's here\n \n\nyou look at the messy JSON that's here\nversus the I guess slightly better\n\nversus the I guess slightly better\n \n\nversus the I guess slightly better\nformatted BAML that's here. um can\n\nformatted BAML that's here. um can\n \n\nformatted BAML that's here. um can\nactually improve performance by you know\n\nactually improve performance by you know\n \n\nactually improve performance by you know\nfive to 10 percent depending on your use\n\nfive to 10 percent depending on your use\n \n\nfive to 10 percent depending on your use\ncase. So it's a good example of how you\n\ncase. So it's a good example of how you\n \n\ncase. So it's a good example of how you\ncan format things differently. The the\n\ncan format things differently. The the\n \n\ncan format things differently. The the\nrest of the program wouldn't have\n\nrest of the program wouldn't have\n \n\nrest of the program wouldn't have\nchanged at all. You just specify the\n\nchanged at all. You just specify the\n \n\nchanged at all. You just specify the\nBAML adapter and it totally changes how\n\nBAML adapter and it totally changes how\n \n\nBAML adapter and it totally changes how\nthe information is presented under the\n\nthe information is presented under the\n \n\nthe information is presented under the\nhood to the LLM\n\nhood to the LLM\n \n\nhood to the LLM\nmultimodality. I mean this obviously is\n\nmultimodality. I mean this obviously is\n \n\nmultimodality. I mean this obviously is\nmore at the model level but DSPI\n\nmore at the model level but DSPI\n \n\nmore at the model level but DSPI\nsupports multiple modalities by default. supports multiple modalities by default. supports multiple modalities by default. So images, audio, some others." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nUm, and\n\nSo images, audio, some others. Um, and\n \n\nSo images, audio, some others. Um, and\nthe same type of thing, you kind of just\n\nthe same type of thing, you kind of just\n \n\nthe same type of thing, you kind of just\nfeed it in as part of your signature and\n\nfeed it in as part of your signature and\n..." }, { "role": "assistant", "content": "Um, and\n\nSo images, audio, some others. Um, and\n \n\nSo images, audio, some others. Um, and\nthe same type of thing, you kind of just\n\nthe same type of thing, you kind of just\n \n\nthe same type of thing, you kind of just\nfeed it in as part of your signature and\n\nfeed it in as part of your signature and\n \n\nfeed it in as part of your signature and\nthen you can get some very nice clean\n\nthen you can get some very nice clean\n \n\nthen you can get some very nice clean\noutput. This allows you to work with\n\noutput. This allows you to work with\n \n\noutput. This allows you to work with\nthem very, very, very easily, very\n\nthem very, very, very easily, very\n \n\nthem very, very, very easily, very\nquickly. And for those uh, eagle-eyed\n\nquickly. And for those uh, eagle-eyed\n \n\nquickly. And for those uh, eagle-eyed\nparticipants, you can see the first uh,\n\nparticipants, you can see the first uh,\n \n\nparticipants, you can see the first uh,\nlineup there is attachments. It's\n\nlineup there is attachments. It's\n \n\nlineup there is attachments. It's\nprobably a lesserk known library. probably a lesserk known library. probably a lesserk known library. Another guy on Twitter is awesome. Uh,\n\nAnother guy on Twitter is awesome. Uh,\n \n\nAnother guy on Twitter is awesome. Uh,\nMaxim, I think it is. uh he created this\n\nMaxim, I think it is. uh he created this\n \n\nMaxim, I think it is. uh he created this\nlibrary that just is basically a\n\nlibrary that just is basically a\n \n\nlibrary that just is basically a\ncatch-all for working with different\n\ncatch-all for working with different\n \n\ncatch-all for working with different\ntypes of files and converting them into\n\ntypes of files and converting them into\n \n\ntypes of files and converting them into\na format that's super easy to use with\n\na format that's super easy to use with\n \n\na format that's super easy to use with\nLLMs. Um he's a big DSPI fan as well. So\n\nLLMs. Um he's a big DSPI fan as well. So\n \n\nLLMs. Um he's a big DSPI fan as well. So\nhe made basically an adapter that's\n\nhe made basically an adapter that's\n \n\nhe made basically an adapter that's\nspecific to this. But that's all it\n\nspecific to this. But that's all it\n \n\nspecific to this. But that's all it\ntakes to pull in images, PDFs, whatever\n\ntakes to pull in images, PDFs, whatever\n \n\ntakes to pull in images, PDFs, whatever\nit might be. You'll see some examples of\n\nit might be. You'll see some examples of\n \n\nit might be. You'll see some examples of\nthat and it just makes at least has made\n\nthat and it just makes at least has made\n \n\nthat and it just makes at least has made\nmy life super super easy. my life super super easy. my life super super easy. Here's another example of the same sort\n\nHere's another example of the same sort\n \n\nHere's another example of the same sort\nof thing. So this is a PDF of a form 4\n\nof thing. So this is a PDF of a form 4\n \n\nof thing. So this is a PDF of a form 4\nform. So, you know, public SEC form from\n\nform. So, you know, public SEC form from\n \n\nform. So, you know, public SEC form from\nNvidia. Nvidia. Nvidia. Um, up top I'm just giving it the link. Um, up top I'm just giving it the link. Um, up top I'm just giving it the link. I'm saying, okay, attachments, do your\n\nI'm saying, okay, attachments, do your\n \n\nI'm saying, okay, attachments, do your\nthing. Pull it down, create images,\n\nthing. Pull it down, create images,\n \n\nthing. Pull it down, create images,\nwhatever you're going to do. I don't\n\nwhatever you're going to do. I don't\n \n\nwhatever you're going to do. I don't\nneed to worry about it. I don't care\n\nneed to worry about it. I don't care\n \n\nneed to worry about it. I don't care\nabout it. This is super simple rag, but\n\nabout it. This is super simple rag, but\n \n\nabout it. This is super simple rag, but\nbasically, okay, I want to do rag over\n\nbasically, okay, I want to do rag over\n \n\nbasically, okay, I want to do rag over\nthis document. I'm going to give you a\n\nthis document. I'm going to give you a\n \n\nthis document. I'm going to give you a\nquestion. I'm going to give you the\n\nquestion. I'm going to give you the\n \n\nquestion. I'm going to give you the\ndocument and I want the answer. Um, and\n\ndocument and I want the answer. Um, and\n \n\ndocument and I want the answer. Um, and\nyou can see how simple that is. you can see how simple that is. you can see how simple that is. Literally just feeding in the document. Literally just feeding in the document. Literally just feeding in the document. How many shares were sold? Interestingly\n\nHow many shares were sold? Interestingly\n \n\nHow many shares were sold? Interestingly\nhere, I'm not sure if it's super easy to\n\nhere, I'm not sure if it's super easy to\n \n\nhere, I'm not sure if it's super easy to\nsee, but you actually have two\n\nsee, but you actually have two\n \n\nsee, but you actually have two\ntransactions\n\ntransactions\n \n\ntransactions\nhere. So, it's going to have to do some\n\nhere. So, it's going to have to do some\n \n\nhere. So, it's going to have to do some\nmath likely under the hood. And you can\n\nmath likely under the hood. And you can\n \n\nmath likely under the hood. And you can\nsee here the thinking and the the\n\nsee here the thinking and the the\n \n\nsee here the thinking and the the\nultimate answer. Go ahead. ultimate answer. Go ahead. ultimate answer. Go ahead. >> Is it on the rag step? Is it creating a\n\n>> Is it on the rag step? Is it creating a\n \n\n>> Is it on the rag step? Is it creating a\nvector store of some kind or creating\n\nvector store of some kind or creating\n \n\nvector store of some kind or creating\nembeddings and searching over those? Is\n\nembeddings and searching over those? Is\n \n\nembeddings and searching over those? Is\nthere a bunch going on in the background\n\nthere a bunch going on in the background\n \n\nthere a bunch going on in the background\nthere or what? there or what? there or what? >> This is poor man's rack. I should have\n\n>> This is poor man's rack. I should have\n \n\n>> This is poor man's rack. I should have\nclarified. This is this is literally\n\nclarified. This is this is literally\n \n\nclarified. This is this is literally\njust pulling in the document images and\n\njust pulling in the document images and\n \n\njust pulling in the document images and\nI think attachments will do some basic\n\nI think attachments will do some basic\n \n\nI think attachments will do some basic\nOCR under the hood. Um, but it doesn't\n\nOCR under the hood. Um, but it doesn't\n \n\nOCR under the hood. Um, but it doesn't\ndo anything other than that. That's it. do anything other than that. That's it. do anything other than that. That's it. All we're feeding in here, the the\n\nAll we're feeding in here, the the\n \n\nAll we're feeding in here, the the\nactual document object that's being fed\n\nactual document object that's being fed\n \n\nactual document object that's being fed\nin, yeah, is literally just the text\n\nin, yeah, is literally just the text\n \n\nin, yeah, is literally just the text\nthat's been OCRD. the images, the model\n\nthat's been OCRD. the images, the model\n \n\nthat's been OCRD. the images, the model\ndoes the rest. does the rest. does the rest. [sighs] All right, so optimizers uh\n\n[sighs] All right, so optimizers uh\n \n\n[sighs] All right, so optimizers uh\nlet's see how we're doing. Okay. Um\n\nlet's see how we're doing. Okay. Um\n \n\nlet's see how we're doing. Okay. Um\noptimizers are super powerful, super\n\noptimizers are super powerful, super\n \n\noptimizers are super powerful, super\ninteresting concept. It's been some\n\ninteresting concept. It's been some\n \n\ninteresting concept. It's been some\nresearch um that argues I think that\n\nresearch um that argues I think that\n \n\nresearch um that argues I think that\nit's just as performant if not in cert\n\nit's just as performant if not in cert\n \n\nit's just as performant if not in cert\nin certain situations more performant\n\nin certain situations more performant\n \n\nin certain situations more performant\nthan fine-tuning would be for certain\n\nthan fine-tuning would be for certain\n \n\nthan fine-tuning would be for certain\nmodels for certain situations. there's\n\nmodels for certain situations. there's\n \n\nmodels for certain situations. there's\nall this research about in context\n\nall this research about in context\n \n\nall this research about in context\nlearning and such. And so whether you\n\nlearning and such." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nAnd so whether you\n \n\nlearning and such. And so whether you\nwant to go fine-tune and do all of that,\n\nwant to go fine-tune and do all of that,\n \n\nwant to go fine-tune and do all of that,\nnothing stops you. But I would recommend\n\nnothing stops you. But I would recommend\n \n\nnothing stops you. But I wo..." }, { "role": "assistant", "content": "And so whether you\n \n\nlearning and such. And so whether you\nwant to go fine-tune and do all of that,\n\nwant to go fine-tune and do all of that,\n \n\nwant to go fine-tune and do all of that,\nnothing stops you. But I would recommend\n\nnothing stops you. But I would recommend\n \n\nnothing stops you. But I would recommend\nat least trying this first to see how\n\nat least trying this first to see how\n \n\nat least trying this first to see how\nfar you can get without having to set up\n\nfar you can get without having to set up\n \n\nfar you can get without having to set up\na bunch of infrastructure and, you know,\n\na bunch of infrastructure and, you know,\n \n\na bunch of infrastructure and, you know,\ngo through all of that. See how the\n\ngo through all of that. See how the\n \n\ngo through all of that. See how the\noptimizers work. Um, but fundamentally\n\noptimizers work. Um, but fundamentally\n \n\noptimizers work. Um, but fundamentally\nwhat it allows you to do is DSPI gives\n\nwhat it allows you to do is DSPI gives\n \n\nwhat it allows you to do is DSPI gives\nyou the primitives that you need and the\n\nyou the primitives that you need and the\n \n\nyou the primitives that you need and the\norganization you need to be able to\n\norganization you need to be able to\n \n\norganization you need to be able to\nmeasure and then quantitatively improve\n\nmeasure and then quantitatively improve\n \n\nmeasure and then quantitatively improve\nthat performance. And I mentioned\n\nthat performance. And I mentioned\n \n\nthat performance. And I mentioned\ntransferability before. This the\n\ntransferability before. This the\n \n\ntransferability before. This the\ntransferability is enabled arguably\n\ntransferability is enabled arguably\n \n\ntransferability is enabled arguably\nthrough the use of optimizers because if\n\nthrough the use of optimizers because if\n \n\nthrough the use of optimizers because if\nyou can get okay I want to I have the\n\nyou can get okay I want to I have the\n \n\nyou can get okay I want to I have the\nclassification task works really well\n\nclassification task works really well\n \n\nclassification task works really well\nwith 41 but maybe it's a little bit\n\nwith 41 but maybe it's a little bit\n \n\nwith 41 but maybe it's a little bit\ncostly because I have to run it a\n\ncostly because I have to run it a\n \n\ncostly because I have to run it a\nmillion times a day. Can I try it with\n\nmillion times a day. Can I try it with\n \n\nmillion times a day. Can I try it with\n41 nano? Okay, maybe it's at 70%\n\n41 nano? Okay, maybe it's at 70%\n \n\n41 nano? Okay, maybe it's at 70%\nwhatever it might be. But I run the\n\nwhatever it might be. But I run the\n \n\nwhatever it might be. But I run the\noptimizer on 41 nano and I can get the\n\noptimizer on 41 nano and I can get the\n \n\noptimizer on 41 nano and I can get the\nperformance back up to maybe 87%. maybe\n\nperformance back up to maybe 87%. maybe\n \n\nperformance back up to maybe 87%. maybe\nthat's okay for my use case, but I've\n\nthat's okay for my use case, but I've\n \n\nthat's okay for my use case, but I've\nnow just dropped my cross my cost\n\nnow just dropped my cross my cost\n \n\nnow just dropped my cross my cost\nprofile by multiple orders of magnitude. profile by multiple orders of magnitude. profile by multiple orders of magnitude. And it's the optimizer that allows you\n\nAnd it's the optimizer that allows you\n \n\nAnd it's the optimizer that allows you\nto do that type of model and kind of use\n\nto do that type of model and kind of use\n \n\nto do that type of model and kind of use\ncase transferability, if you will. But\n\ncase transferability, if you will. But\n \n\ncase transferability, if you will. But\nreally all it does at at the end of the\n\nreally all it does at at the end of the\n \n\nreally all it does at at the end of the\nday under the hood is iteratively prompt\n\nday under the hood is iteratively prompt\n \n\nday under the hood is iteratively prompt\nuh iteratively optimize or tweak that\n\nuh iteratively optimize or tweak that\n \n\nuh iteratively optimize or tweak that\nprompt, that string under the hood. And\n\nprompt, that string under the hood. And\n \n\nprompt, that string under the hood. And\nbecause you've constructed your program\n\nbecause you've constructed your program\n \n\nbecause you've constructed your program\nusing the different modules, DSPI kind\n\nusing the different modules, DSPI kind\n \n\nusing the different modules, DSPI kind\nof handles all of that for you under the\n\nof handles all of that for you under the\n \n\nof handles all of that for you under the\nhood. So if you compose a program with\n\nhood. So if you compose a program with\n \n\nhood. So if you compose a program with\nmultiple modules and you're optimizing\n\nmultiple modules and you're optimizing\n \n\nmultiple modules and you're optimizing\nagainst all that, it it by itself DSPI\n\nagainst all that, it it by itself DSPI\n \n\nagainst all that, it it by itself DSPI\nwill optimize the various components in\n\nwill optimize the various components in\n \n\nwill optimize the various components in\norder to improve the input and output\n\norder to improve the input and output\n \n\norder to improve the input and output\nperformance. And we'll we'll take it from the man\n\nAnd we'll we'll take it from the man\n \n\nAnd we'll we'll take it from the man\nhimself, Omar. You know, ESPI is not an\n\nhimself, Omar. You know, ESPI is not an\n \n\nhimself, Omar. You know, ESPI is not an\noptimizer. I've said this multiple\n\noptimizer. I've said this multiple\n \n\noptimizer. I've said this multiple\ntimes. it's it's just a set of\n\ntimes. it's it's just a set of\n \n\ntimes. it's it's just a set of\nprogramming abstractions or a way to\n\nprogramming abstractions or a way to\n \n\nprogramming abstractions or a way to\nprogram. You just happen to be able to\n\nprogram. You just happen to be able to\n \n\nprogram. You just happen to be able to\noptimize it. Um so again, the value that\n\noptimize it. Um so again, the value that\n \n\noptimize it. Um so again, the value that\nI've gotten and my team has gotten is\n\nI've gotten and my team has gotten is\n \n\nI've gotten and my team has gotten is\nmostly because of the programming\n\nmostly because of the programming\n \n\nmostly because of the programming\nabstractions. It's just this incredible\n\nabstractions. It's just this incredible\n \n\nabstractions. It's just this incredible\nadded benefit that you are also able to\n\nadded benefit that you are also able to\n \n\nadded benefit that you are also able to\nto should you choose to to optimize it\n\nto should you choose to to optimize it\n \n\nto should you choose to to optimize it\nafterwards. afterwards. afterwards. And I was listening to this to Dwaresh\n\nAnd I was listening to this to Dwaresh\n \n\nAnd I was listening to this to Dwaresh\nand and uh Carpathy the other day and\n\nand and uh Carpathy the other day and\n \n\nand and uh Carpathy the other day and\nthis kind of I was like prepping for\n\nthis kind of I was like prepping for\n \n\nthis kind of I was like prepping for\nthis talk and this like hit home\n\nthis talk and this like hit home\n \n\nthis talk and this like hit home\nperfectly. I was thinking about the\n\nperfectly. I was thinking about the\n \n\nperfectly. I was thinking about the\noptimizers and someone smarter than me\n\noptimizers and someone smarter than me\n \n\noptimizers and someone smarter than me\ncan can ple you know please correct me\n\ncan can ple you know please correct me\n \n\ncan can ple you know please correct me\nbut I think this makes sense because he\n\nbut I think this makes sense because he\n \n\nbut I think this makes sense because he\nhe was basically talking about using LLM\n\nhe was basically talking about using LLM\n \n\nhe was basically talking about using LLM\nas a judge can be a bad thing because\n\nas a judge can be a bad thing because\n \n\nas a judge can be a bad thing because\nthe model being judged can find\n\nthe model being judged can find\n \n\nthe model being judged can find\nadversarial examples and degrade the\n\nadversarial examples and degrade the\n \n\nadversarial examples and degrade the\nperformance or basically um create a\n\nperformance or basically um create a\n \n\nperformance or basically um create a\nsituation where the judge is not uh not\n\nsituation where the judge is not uh not\n \n\nsituation where the judge is not uh not\nscoring something properly. um because\n\nscoring something properly. um because\n \n\nscoring something properly." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\num because\nhe's saying that the model will find\n\nhe's saying that the model will find\n \n\nhe's saying that the model will find\nthese little cracks. It'll find these\n\nthese little cracks. It'll find these\n \n\nthese little cracks. It'll find these\nlittle spirious things in the nooks and\n\nlittle spirious..." }, { "role": "assistant", "content": "um because\nhe's saying that the model will find\n\nhe's saying that the model will find\n \n\nhe's saying that the model will find\nthese little cracks. It'll find these\n\nthese little cracks. It'll find these\n \n\nthese little cracks. It'll find these\nlittle spirious things in the nooks and\n\nlittle spirious things in the nooks and\n \n\nlittle spirious things in the nooks and\ncrannies of the giant model and find a\n\ncrannies of the giant model and find a\n \n\ncrannies of the giant model and find a\nway to cheat it. Basically saying that\n\nway to cheat it. Basically saying that\n \n\nway to cheat it. Basically saying that\nLM as a judge can only go so far until\n\nLM as a judge can only go so far until\n \n\nLM as a judge can only go so far until\nthe other model uh finds those\n\nthe other model uh finds those\n \n\nthe other model uh finds those\nadversarial examples. If you kind of\n\nadversarial examples. If you kind of\n \n\nadversarial examples. If you kind of\ninvert that and flip that on its head,\n\ninvert that and flip that on its head,\n \n\ninvert that and flip that on its head,\nit's this property that the optimizers\n\nit's this property that the optimizers\n \n\nit's this property that the optimizers\nfor DSpir are taking advantage of to\n\nfor DSpir are taking advantage of to\n \n\nfor DSpir are taking advantage of to\noptimize to find the nooks and crannies\n\noptimize to find the nooks and crannies\n \n\noptimize to find the nooks and crannies\nin the model, whether it's a bigger\n\nin the model, whether it's a bigger\n \n\nin the model, whether it's a bigger\nmodel model or smaller model to improve\n\nmodel model or smaller model to improve\n \n\nmodel model or smaller model to improve\nthe performance against your data set. the performance against your data set. the performance against your data set. So that's what the optimizer is doing is\n\nSo that's what the optimizer is doing is\n \n\nSo that's what the optimizer is doing is\nfinding finding these nooks and crannies\n\nfinding finding these nooks and crannies\n \n\nfinding finding these nooks and crannies\nin the model to optimize and improve\n\nin the model to optimize and improve\n \n\nin the model to optimize and improve\nthat performance. that performance. that performance. So a typical flow, I'm not going to\n\nSo a typical flow, I'm not going to\n \n\nSo a typical flow, I'm not going to\nspend too much time on this, but fairly\n\nspend too much time on this, but fairly\n \n\nspend too much time on this, but fairly\nlogical constructor program which is\n\nlogical constructor program which is\n \n\nlogical constructor program which is\ndecomposing your logic into the modules. decomposing your logic into the modules. decomposing your logic into the modules. You use your metrics to define basically\n\nYou use your metrics to define basically\n \n\nYou use your metrics to define basically\nthe contours of how the program works\n\nthe contours of how the program works\n \n\nthe contours of how the program works\nand you optimize all that through um to\n\nand you optimize all that through um to\n \n\nand you optimize all that through um to\nto get your your uh your final result. So, another talk that this guy Chris\n\nSo, another talk that this guy Chris\n \n\nSo, another talk that this guy Chris\nPototts just had maybe two days ago, um,\n\nPototts just had maybe two days ago, um,\n \n\nPototts just had maybe two days ago, um,\nwhere he made the point, this is what I\n\nwhere he made the point, this is what I\n \n\nwhere he made the point, this is what I\nwas mentioning before, where Jeepa,\n\nwas mentioning before, where Jeepa,\n \n\nwas mentioning before, where Jeepa,\nwhich is, uh, you probably saw some of\n\nwhich is, uh, you probably saw some of\n \n\nwhich is, uh, you probably saw some of\nthe the talks the other day, um, where\n\nthe the talks the other day, um, where\n \n\nthe the talks the other day, um, where\nthe optimizers are on par or exceed the\n\nthe optimizers are on par or exceed the\n \n\nthe optimizers are on par or exceed the\nperformance of something like GRPO,\n\nperformance of something like GRPO,\n \n\nperformance of something like GRPO,\nanother kind of fine-tuning method. So,\n\nanother kind of fine-tuning method. So,\n \n\nanother kind of fine-tuning method. So,\npretty impressive. I think it's an\n\npretty impressive. I think it's an\n \n\npretty impressive. I think it's an\nactive area of research. people a lot\n\nactive area of research. people a lot\n \n\nactive area of research. people a lot\nsmarter than me like Omar and Chris and\n\nsmarter than me like Omar and Chris and\n \n\nsmarter than me like Omar and Chris and\nothers are are leading the way on this. others are are leading the way on this. others are are leading the way on this. But uh point being I think prompt op\n\nBut uh point being I think prompt op\n \n\nBut uh point being I think prompt op\nprompt optimization is a pretty exciting\n\nprompt optimization is a pretty exciting\n \n\nprompt optimization is a pretty exciting\nplace to be and if nothing else is worth\n\nplace to be and if nothing else is worth\n \n\nplace to be and if nothing else is worth\nexploring. exploring. exploring. And [clears throat] then finally metrics\n\nAnd [clears throat] then finally metrics\n \n\nAnd [clears throat] then finally metrics\nagain these are kind of the building\n\nagain these are kind of the building\n \n\nagain these are kind of the building\nblocks that allow you to define what\n\nblocks that allow you to define what\n \n\nblocks that allow you to define what\nsuccess looks like for the optimizer. So\n\nsuccess looks like for the optimizer. So\n \n\nsuccess looks like for the optimizer. So\nthis is what it's using and you can have\n\nthis is what it's using and you can have\n \n\nthis is what it's using and you can have\nmany of these and we'll see examples of\n\nmany of these and we'll see examples of\n \n\nmany of these and we'll see examples of\nthis where again at a high level your\n\nthis where again at a high level your\n \n\nthis where again at a high level your\nprogram works on inputs it works on\n\nprogram works on inputs it works on\n \n\nprogram works on inputs it works on\noutputs the optimizer is going to use\n\noutputs the optimizer is going to use\n \n\noutputs the optimizer is going to use\nthe metrics to understand okay my last\n\nthe metrics to understand okay my last\n \n\nthe metrics to understand okay my last\ntweak in the prompt did it improve\n\ntweak in the prompt did it improve\n \n\ntweak in the prompt did it improve\nperformance it did it degrade\n\nperformance it did it degrade\n \n\nperformance it did it degrade\nperformance and the way you define your\n\nperformance and the way you define your\n \n\nperformance and the way you define your\nmetrics uh provides that direct feedback\n\nmetrics uh provides that direct feedback\n \n\nmetrics uh provides that direct feedback\nfor the optimizers to work on. for the optimizers to work on. for the optimizers to work on. Uh so here's another example, a super\n\nUh so here's another example, a super\n \n\nUh so here's another example, a super\nsimple one from that time entry example\n\nsimple one from that time entry example\n \n\nsimple one from that time entry example\nI mentioned before. Um, so they can be\n\nI mentioned before. Um, so they can be\n \n\nI mentioned before." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nUm, so they can be\nthe metrics can either be like fairly\n\nthe metrics can either be like fairly\n \n\nthe metrics can either be like fairly\nrigorous in terms of like does this\n\nrigorous in terms of like does this\n \n\nrigorous in terms of like does this\nequal one or or you know some type of\n\nequal one or..." }, { "role": "assistant", "content": "Um, so they can be\nthe metrics can either be like fairly\n\nthe metrics can either be like fairly\n \n\nthe metrics can either be like fairly\nrigorous in terms of like does this\n\nrigorous in terms of like does this\n \n\nrigorous in terms of like does this\nequal one or or you know some type of\n\nequal one or or you know some type of\n \n\nequal one or or you know some type of\nequality check or a little bit more\n\nequality check or a little bit more\n \n\nequality check or a little bit more\nsubjective where using LLM as a judge to\n\nsubjective where using LLM as a judge to\n \n\nsubjective where using LLM as a judge to\nsay whatever was this generated um\n\nsay whatever was this generated um\n \n\nsay whatever was this generated um\nstring does it adhere to these you know\n\nstring does it adhere to these you know\n \n\nstring does it adhere to these you know\nvarious criteria whatever it might be\n\nvarious criteria whatever it might be\n \n\nvarious criteria whatever it might be\nbut that itself can be a metric\n\nbut that itself can be a metric\n \n\nbut that itself can be a metric\nand so all of this is to say it's a very\n\nand so all of this is to say it's a very\n \n\nand so all of this is to say it's a very\nlong-winded way of saying in my opinion\n\nlong-winded way of saying in my opinion\n \n\nlong-winded way of saying in my opinion\nthis is probably most if not all of what\n\nthis is probably most if not all of what\n \n\nthis is probably most if not all of what\nyou need to construct arbitrarily\n\nyou need to construct arbitrarily\n \n\nyou need to construct arbitrarily\ncomplex workflows, data processing\n\ncomplex workflows, data processing\n \n\ncomplex workflows, data processing\npipelines, business logic, whatever that\n\npipelines, business logic, whatever that\n \n\npipelines, business logic, whatever that\nmight be. Different ways to work with\n\nmight be. Different ways to work with\n \n\nmight be. Different ways to work with\nLLMs. If nothing else, DSPI gives you\n\nLLMs. If nothing else, DSPI gives you\n \n\nLLMs. If nothing else, DSPI gives you\nthe primitives that you need in order to\n\nthe primitives that you need in order to\n \n\nthe primitives that you need in order to\nbuild these modular composable systems. build these modular composable systems. build these modular composable systems. So, if you're interested in some people\n\nSo, if you're interested in some people\n \n\nSo, if you're interested in some people\nonline, um\n\nonline, um\n \n\nonline, um\nthere's many many more. There's a\n\nthere's many many more. There's a\n \n\nthere's many many more. There's a\nDiscord community as well. Um, but\n\nDiscord community as well. Um, but\n \n\nDiscord community as well. Um, but\nusually these people are are on top of\n\nusually these people are are on top of\n \n\nusually these people are are on top of\nthe latest and greatest and so would\n\nthe latest and greatest and so would\n \n\nthe latest and greatest and so would\nrecommend giving them a follow. You\n\nrecommend giving them a follow. You\n \n\nrecommend giving them a follow. You\ndon't need to follow me. I don't really\n\ndon't need to follow me. I don't really\n \n\ndon't need to follow me. I don't really\ndo much, but uh the others on there are\n\ndo much, but uh the others on there are\n \n\ndo much, but uh the others on there are\nare really pretty good. are really pretty good. are really pretty good. Okay, so the fun part, we'll actually\n\nOkay, so the fun part, we'll actually\n \n\nOkay, so the fun part, we'll actually\nget into some to some code. So, if you\n\nget into some to some code. So, if you\n \n\nget into some to some code. So, if you\nhaven't had a chance, now's your last\n\nhaven't had a chance, now's your last\n \n\nhaven't had a chance, now's your last\nchance to get the repo. chance to get the repo. chance to get the repo. U, but I'll just kind of go through a\n\nU, but I'll just kind of go through a\n \n\nU, but I'll just kind of go through a\nfew different examples here of what we\n\nfew different examples here of what we\n \n\nfew different examples here of what we\ntalked about. Maybe\n\n \n\nYeah. Okay. Yeah. Okay. Yeah. Okay. Okay. So, I'll set up Phoenix, which is\n\nOkay. So, I'll set up Phoenix, which is\n \n\nOkay. So, I'll set up Phoenix, which is\nfrom Arise, uh, which is basically an\n\nfrom Arise, uh, which is basically an\n \n\nfrom Arise, uh, which is basically an\nobser an observability platform. Uh, I\n\nobser an observability platform. Uh, I\n \n\nobser an observability platform. Uh, I\njust did this today, so I don't know if\n\njust did this today, so I don't know if\n \n\njust did this today, so I don't know if\nit's going to work or not, but we'll\n\nit's going to work or not, but we'll\n \n\nit's going to work or not, but we'll\nwe'll see. We'll give it a shot. Uh, but\n\nwe'll see. We'll give it a shot. Uh, but\n \n\nwe'll see. We'll give it a shot. Uh, but\nbasically what this allows you to do is\n\nbasically what this allows you to do is\n \n\nbasically what this allows you to do is\nhave a bunch of observability and\n\nhave a bunch of observability and\n \n\nhave a bunch of observability and\ntracing for all the calls that are\n\ntracing for all the calls that are\n \n\ntracing for all the calls that are\nhappening under the hood. We'll see if\n\nhappening under the hood. We'll see if\n \n\nhappening under the hood. We'll see if\nthis works. We'll give it like another 5\n\nthis works. We'll give it like another 5\n \n\nthis works. We'll give it like another 5\nseconds. seconds. seconds. Um, but it should, I think,\n\nUm, but it should, I think,\n \n\nUm, but it should, I think,\nautomatically do all this stuff for me. automatically do all this stuff for me. automatically do all this stuff for me. Yeah. So, let's see. Yeah. So, let's see. Yeah. So, let's see. Yeah. All right. So, something's up. Yeah. All right. So, something's up. Yeah. All right. So, something's up. Okay, cool. So,\n\nOkay, cool. So,\n \n\nOkay, cool. So,\nI'll just I'm just going to run through\n\nI'll just I'm just going to run through\n \n\nI'll just I'm just going to run through\nthe notebook, which is a collection of\n\nthe notebook, which is a collection of\n \n\nthe notebook, which is a collection of\ndifferent use cases, basically putting\n\ndifferent use cases, basically putting\n \n\ndifferent use cases, basically putting\ninto practice a lot of what we just saw. into practice a lot of what we just saw. into practice a lot of what we just saw. Feel free to jump in any questions,\n\nFeel free to jump in any questions,\n \n\nFeel free to jump in any questions,\nanything like that. We'll start with\n\nanything like that. We'll start with\n \n\nanything like that. We'll start with\nthis notebook. There's a couple of other\n\nthis notebook. There's a couple of other\n \n\nthis notebook. There's a couple of other\nuh more proper Python programs that\n\nuh more proper Python programs that\n \n\nuh more proper Python programs that\nwe'll walk through afterwards. Uh but\n\nwe'll walk through afterwards. Uh but\n \n\nwe'll walk through afterwards. Uh but\nreally the intent is a rapidfire review\n\nreally the intent is a rapidfire review\n \n\nreally the intent is a rapidfire review\nof different ways that DSPI has been\n\nof different ways that DSPI has been\n \n\nof different ways that DSPI has been\nuseful to me and others. So\n\nuseful to me and others. So\n \n\nuseful to me and others. So\nload in the end file. Usually I'll have\n\nload in the end file. Usually I'll have\n \n\nload in the end file. Usually I'll have\nsome type of config object like this\n\nsome type of config object like this\n \n\nsome type of config object like this\nwhere I can very easily use these later\n\nwhere I can very easily use these later\n \n\nwhere I can very easily use these later\non. So if I'm like call like model\n\non. So if I'm like call like model\n \n\non. So if I'm like call like model\nmixing. So if I have like a super hairy\n\nmixing. So if I have like a super hairy\n \n\nmixing. So if I have like a super hairy\nproblem or like some workload I know\n\nproblem or like some workload I know\n \n\nproblem or like some workload I know\nwill need the power of a reasoning model\n\nwill need the power of a reasoning model\n \n\nwill need the power of a reasoning model\nlike GPD5 or something else like that,\n\nlike GPD5 or something else like that,\n \n\nlike GPD5 or something else like that,\nI'll define multiple LM." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nSo like one\n\nI'll define multiple LM. So like one\n \n\nI'll define multiple LM. So like one\nwill be 41, one will be five, maybe I'll\n\nwill be 41, one will be five, maybe I'll\n \n\nwill be 41, one will be five, maybe I'll\ndo a 41 nano um you know Gemini 2.5\n\ndo a 41 nano um you know Gemini 2.5\n \n\ndo a 41..." }, { "role": "assistant", "content": "So like one\n\nI'll define multiple LM. So like one\n \n\nI'll define multiple LM. So like one\nwill be 41, one will be five, maybe I'll\n\nwill be 41, one will be five, maybe I'll\n \n\nwill be 41, one will be five, maybe I'll\ndo a 41 nano um you know Gemini 2.5\n\ndo a 41 nano um you know Gemini 2.5\n \n\ndo a 41 nano um you know Gemini 2.5\nflash, stuff like that. And then I can\n\nflash, stuff like that. And then I can\n \n\nflash, stuff like that. And then I can\nkind of intermingle or intersperse them\n\nkind of intermingle or intersperse them\n \n\nkind of intermingle or intersperse them\ndepending on what I think or what I'm\n\ndepending on what I think or what I'm\n \n\ndepending on what I think or what I'm\nreasonably sure the workload will be. reasonably sure the workload will be. reasonably sure the workload will be. and you'll see how that comes into play\n\nand you'll see how that comes into play\n \n\nand you'll see how that comes into play\nin terms of classification and others. in terms of classification and others. in terms of classification and others. Um, I'll pull in a few others here. I'm\n\nUm, I'll pull in a few others here. I'm\n \n\nUm, I'll pull in a few others here. I'm\nI'm using open router for this. So, if\n\nI'm using open router for this. So, if\n \n\nI'm using open router for this. So, if\nyou have an open router API key, would\n\nyou have an open router API key, would\n \n\nyou have an open router API key, would\nrecommend plug plugging that in. So, now\n\nrecommend plug plugging that in. So, now\n \n\nrecommend plug plugging that in. So, now\nI have three different LLMs I can work\n\nI have three different LLMs I can work\n \n\nI have three different LLMs I can work\nwith. I have Claude, I have Gemini, I\n\nwith. I have Claude, I have Gemini, I\n \n\nwith. I have Claude, I have Gemini, I\nhave 41 mini. And then I'll ask\n\nhave 41 mini. And then I'll ask\n \n\nhave 41 mini. And then I'll ask\nbasically for each of them who's best\n\nbasically for each of them who's best\n \n\nbasically for each of them who's best\nbetween Google Anthropic OpenAI. All of\n\nbetween Google Anthropic OpenAI. All of\n \n\nbetween Google Anthropic OpenAI. All of\nthem are hedging a little bit. They say\n\nthem are hedging a little bit. They say\n \n\nthem are hedging a little bit. They say\nsubjective, subjective, undefined. All\n\nsubjective, subjective, undefined. All\n \n\nsubjective, subjective, undefined. All\nright, great. It's not very helpful. But\n\nright, great. It's not very helpful. But\n \n\nright, great. It's not very helpful. But\nbecause DSPI works on Pyantic, I can\n\nbecause DSPI works on Pyantic, I can\n \n\nbecause DSPI works on Pyantic, I can\ndefine the answer as a literal. So I'm\n\ndefine the answer as a literal. So I'm\n \n\ndefine the answer as a literal. So I'm\nbasically forcing it to only give me\n\nbasically forcing it to only give me\n \n\nbasically forcing it to only give me\nthose three options and then I can go\n\nthose three options and then I can go\n \n\nthose three options and then I can go\nthrough each of those. And you can see\n\nthrough each of those. And you can see\n \n\nthrough each of those. And you can see\neach of them, of course, chooses their\n\neach of them, of course, chooses their\n \n\neach of them, of course, chooses their\nown organization. Um, the reason that\n\nown organization. Um, the reason that\n \n\nown organization. Um, the reason that\nthose came back so fast\n\nthose came back so fast\n \n\nthose came back so fast\nis that DSP has caching automated under\n\nis that DSP has caching automated under\n \n\nis that DSP has caching automated under\nthe hood. So as long as nothing has\n\nthe hood. So as long as nothing has\n \n\nthe hood. So as long as nothing has\nchanged in terms of your uh your\n\nchanged in terms of your uh your\n \n\nchanged in terms of your uh your\nsignature definitions or basically if\n\nsignature definitions or basically if\n \n\nsignature definitions or basically if\nnothing has changed this is super useful\n\nnothing has changed this is super useful\n \n\nnothing has changed this is super useful\nfor testing it will just load it from\n\nfor testing it will just load it from\n \n\nfor testing it will just load it from\nthe cache. Um so I ran this before\n\nthe cache. Um so I ran this before\n \n\nthe cache. Um so I ran this before\nthat's why those came back so quickly. U\n\nthat's why those came back so quickly. U\n \n\nthat's why those came back so quickly. U\nbut that's another kind of super useful\n\nbut that's another kind of super useful\n \n\nbut that's another kind of super useful\num piece here. Let's see. Okay. Okay. Okay. Make sure we're up and running. So, if I\n\nMake sure we're up and running. So, if I\n \n\nMake sure we're up and running. So, if I\nchange this to hello\n\nchange this to hello\n \n\nchange this to hello\nwith a space,\n\nwith a space,\n \n\nwith a space,\nyou can see we're making a live call. you can see we're making a live call. you can see we're making a live call. Okay, great. We're still up. So, super\n\nOkay, great. We're still up. So, super\n \n\nOkay, great. We're still up. So, super\nsimple class sentiment classifier. simple class sentiment classifier. simple class sentiment classifier. Obviously, this can be built into\n\nObviously, this can be built into\n \n\nObviously, this can be built into\nsomething arbitrarily complex. Make this\n\nsomething arbitrarily complex. Make this\n \n\nsomething arbitrarily complex. Make this\na little bit bigger. Um, but I'm\n\na little bit bigger. Um, but I'm\n \n\na little bit bigger. Um, but I'm\nbasically I'm giving it the text, the\n\nbasically I'm giving it the text, the\n \n\nbasically I'm giving it the text, the\nsentiment that you saw before, and I'm\n\nsentiment that you saw before, and I'm\n \n\nsentiment that you saw before, and I'm\nadding that additional specification to\n\nadding that additional specification to\n \n\nadding that additional specification to\nsay, okay, lower uh is more negative,\n\nsay, okay, lower uh is more negative,\n \n\nsay, okay, lower uh is more negative,\nhigher is more positive. I'm going to\n\nhigher is more positive. I'm going to\n \n\nhigher is more positive. I'm going to\ndefine that as my signature. I'm going\n\ndefine that as my signature. I'm going\n \n\ndefine that as my signature. I'm going\nto pass this into just a super simple\n\nto pass this into just a super simple\n \n\nto pass this into just a super simple\npredict object. predict object. predict object. And then I'm going to say, okay, well,\n\nAnd then I'm going to say, okay, well,\n \n\nAnd then I'm going to say, okay, well,\nthis hotel stinks. Okay, it's probably\n\nthis hotel stinks. Okay, it's probably\n \n\nthis hotel stinks. Okay, it's probably\npretty negative. Now, if I flip that to\n\npretty negative. Now, if I flip that to\n \n\npretty negative. Now, if I flip that to\nI'm feeling pretty happy. Whoops. Good thing I'm not in a hotel right now. Good thing I'm not in a hotel right now. Good thing I'm not in a hotel right now. U you can see I'm feeling pretty happy. U you can see I'm feeling pretty happy. U you can see I'm feeling pretty happy. Comes down to eight. And this might not\n\nComes down to eight. And this might not\n \n\nComes down to eight. And this might not\nseem that impressive and you know it's\n\nseem that impressive and you know it's\n \n\nseem that impressive and you know it's\nit's not really but uh the the the\n\nit's not really but uh the the the\n \n\nit's not really but uh the the the\nimportant part here is that it just\n\nimportant part here is that it just\n \n\nimportant part here is that it just\ndemonstrates the use of the shorthand um\n\ndemonstrates the use of the shorthand um\n \n\ndemonstrates the use of the shorthand um\nsignature. So I have I have the string,\n\nsignature. So I have I have the string,\n \n\nsignature. So I have I have the string,\nI have the integer, I pass in the custom\n\nI have the integer, I pass in the custom\n \n\nI have the integer, I pass in the custom\ninstructions which would be in the dock\n\ninstructions which would be in the dock\n \n\ninstructions which would be in the dock\nstring if I use the class B classbased\n\nstring if I use the class B classbased\n \n\nstring if I use the class B classbased\nuh method. The other interesting part or\n\nuh method. The other interesting part or\n \n\nuh method." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nThe other interesting part or\nor useful part about DSPI comes with a\n\nor useful part about DSPI comes with a\n \n\nor useful part about DSPI comes with a\nbunch of usage information built in. So\n\nbunch of usage information built in. So\n \n\nbunch of usage information built in. So\num because it's cached, i..." }, { "role": "assistant", "content": "The other interesting part or\nor useful part about DSPI comes with a\n\nor useful part about DSPI comes with a\n \n\nor useful part about DSPI comes with a\nbunch of usage information built in. So\n\nbunch of usage information built in. So\n \n\nbunch of usage information built in. So\num because it's cached, it's going to be\n\num because it's cached, it's going to be\n \n\num because it's cached, it's going to be\nan empty object. an empty object. an empty object. But when I change it, you can see that\n\nBut when I change it, you can see that\n \n\nBut when I change it, you can see that\nI'm using Azure right now, but for each\n\nI'm using Azure right now, but for each\n \n\nI'm using Azure right now, but for each\ncall, you get this nice breakdown. and I\n\ncall, you get this nice breakdown. and I\n \n\ncall, you get this nice breakdown. and I\nthink it's from late LLM, but allows you\n\nthink it's from late LLM, but allows you\n \n\nthink it's from late LLM, but allows you\nto very easily track your usage, token\n\nto very easily track your usage, token\n \n\nto very easily track your usage, token\nusage, etc. for observability and\n\nusage, etc. for observability and\n \n\nusage, etc. for observability and\noptimization and everything like that. optimization and everything like that. optimization and everything like that. Just nice little tidbits uh that are\n\nJust nice little tidbits uh that are\n \n\nJust nice little tidbits uh that are\npart of it here and there. Make this\n\npart of it here and there. Make this\n \n\npart of it here and there. Make this\nsmaller. smaller. smaller. Uh we saw the example before in the\n\nUh we saw the example before in the\n \n\nUh we saw the example before in the\nslides, but I'm going to pull in that\n\nslides, but I'm going to pull in that\n \n\nslides, but I'm going to pull in that\nform 4\n\nform 4\n \n\nform 4\noff of online. I'm going to create this\n\noff of online. I'm going to create this\n \n\noff of online. I'm going to create this\ndoc objects using attachments. You can\n\ndoc objects using attachments. You can\n \n\ndoc objects using attachments. You can\nsee some of the stuff it did under the\n\nsee some of the stuff it did under the\n \n\nsee some of the stuff it did under the\nhood. So, it pulled out um PDF plumber. hood. So, it pulled out um PDF plumber. hood. So, it pulled out um PDF plumber. It created markdown from it. Pulled out\n\nIt created markdown from it. Pulled out\n \n\nIt created markdown from it. Pulled out\nthe images, etc. Again, I don't have to\n\nthe images, etc. Again, I don't have to\n \n\nthe images, etc. Again, I don't have to\nworry about all that. Attachments make\n\nworry about all that. Attachments make\n \n\nworry about all that. Attachments make\nthat super easy. I'm going to show you\n\nthat super easy. I'm going to show you\n \n\nthat super easy. I'm going to show you\nwhat we're working with here. This case,\n\nwhat we're working with here. This case,\n \n\nwhat we're working with here. This case,\nwe have the form four. And then I'm\n\nwe have the form four. And then I'm\n \n\nwe have the form four. And then I'm\ngoing to do that poor man's rag that I\n\ngoing to do that poor man's rag that I\n \n\ngoing to do that poor man's rag that I\nmentioned before. Okay, great. How many\n\nmentioned before. Okay, great. How many\n \n\nmentioned before. Okay, great. How many\nshares were were sold in total? It's\n\nshares were were sold in total? It's\n \n\nshares were were sold in total? It's\ngoing to go through that whole chain of\n\ngoing to go through that whole chain of\n \n\ngoing to go through that whole chain of\nthought and bring back the response. thought and bring back the response. thought and bring back the response. That's all well and good, but the power\n\nThat's all well and good, but the power\n \n\nThat's all well and good, but the power\nin my mind of DSPI is that you can have\n\nin my mind of DSPI is that you can have\n \n\nin my mind of DSPI is that you can have\nthese arbitrarily complex data\n\nthese arbitrarily complex data\n \n\nthese arbitrarily complex data\nstructures. That's fairly obvious\n\nstructures. That's fairly obvious\n \n\nstructures. That's fairly obvious\nbecause it uses paidantic and everything\n\nbecause it uses paidantic and everything\n \n\nbecause it uses paidantic and everything\nelse, but you can get a little creative\n\nelse, but you can get a little creative\n \n\nelse, but you can get a little creative\nwith it. So in this case, I'm going to\n\nwith it. So in this case, I'm going to\n \n\nwith it. So in this case, I'm going to\nsay, okay, a different type of document\n\nsay, okay, a different type of document\n \n\nsay, okay, a different type of document\nanalyzer signature. I'm just going to\n\nanalyzer signature. I'm just going to\n \n\nanalyzer signature. I'm just going to\ngive it the document and then I'm just\n\ngive it the document and then I'm just\n \n\ngive it the document and then I'm just\ngoing to defer to the model on defining\n\ngoing to defer to the model on defining\n \n\ngoing to defer to the model on defining\nthe structure of what it thinks is most\n\nthe structure of what it thinks is most\n \n\nthe structure of what it thinks is most\nimportant from the document. So in this\n\nimportant from the document. So in this\n \n\nimportant from the document. So in this\ncase, [clears throat] I'm defining a\n\ncase, [clears throat] I'm defining a\n \n\ncase, [clears throat] I'm defining a\ndictionary object and so it will\n\ndictionary object and so it will\n \n\ndictionary object and so it will\nhopefully return to me a series of key\n\nhopefully return to me a series of key\n \n\nhopefully return to me a series of key\nvalue pairs that describe important\n\nvalue pairs that describe important\n \n\nvalue pairs that describe important\ninformation in the document in a\n\ninformation in the document in a\n \n\ninformation in the document in a\nstructured way. And so you can see here\n\nstructured way. And so you can see here\n \n\nstructured way. And so you can see here\nagain this is probably cached uh but I\n\nagain this is probably cached uh but I\n \n\nagain this is probably cached uh but I\npassed in I did it all in one line in\n\npassed in I did it all in one line in\n \n\npassed in I did it all in one line in\nthis case but I'm saying I want to do\n\nthis case but I'm saying I want to do\n \n\nthis case but I'm saying I want to do\nchain of thought using the document\n\nchain of thought using the document\n \n\nchain of thought using the document\nanalyzer signature and I'm going to pass\n\nanalyzer signature and I'm going to pass\n \n\nanalyzer signature and I'm going to pass\nin the input field which is just the\n\nin the input field which is just the\n \n\nin the input field which is just the\ndocument here. I'm going to pass in the\n\ndocument here. I'm going to pass in the\n \n\ndocument here. I'm going to pass in the\ndocument that I got before. And you can\n\ndocument that I got before. And you can\n \n\ndocument that I got before. And you can\nsee here it pulled out bunch of great\n\nsee here it pulled out bunch of great\n \n\nsee here it pulled out bunch of great\ninformation in the super structured way. information in the super structured way. information in the super structured way. And I didn't have to really think about\n\nAnd I didn't have to really think about\n \n\nAnd I didn't have to really think about\nit. I just kind of deferred all this to\n\nit. I just kind of deferred all this to\n \n\nit. I just kind of deferred all this to\nthe model to DSPI for how to do this. the model to DSPI for how to do this. the model to DSPI for how to do this. Now, of course, you can do the inverse\n\nNow, of course, you can do the inverse\n \n\nNow, of course, you can do the inverse\nin saying, okay, I have a very specific\n\nin saying, okay, I have a very specific\n \n\nin saying, okay, I have a very specific\nbusiness use case. I have something\n\nbusiness use case. I have something\n \n\nbusiness use case. I have something\nspecific in terms of the formatting or\n\nspecific in terms of the formatting or\n \n\nspecific in terms of the formatting or\nthe content that I want to get out of\n\nthe content that I want to get out of\n \n\nthe content that I want to get out of\nthe document. I define that as just kind\n\nthe document." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nI define that as just kind\n \n\nthe document. I define that as just kind\nof your typical paid classes. So in this\n\nof your typical paid classes. So in this\n \n\nof your typical paid classes. So in this\ncase I want to pull out the if there's\n\ncase I want to pull out the if there's\n \n\ncase I want to pull ..." }, { "role": "assistant", "content": "I define that as just kind\n \n\nthe document. I define that as just kind\nof your typical paid classes. So in this\n\nof your typical paid classes. So in this\n \n\nof your typical paid classes. So in this\ncase I want to pull out the if there's\n\ncase I want to pull out the if there's\n \n\ncase I want to pull out the if there's\nmultiple transactions the schema itself\n\nmultiple transactions the schema itself\n \n\nmultiple transactions the schema itself\nimportant information like the filing\n\nimportant information like the filing\n \n\nimportant information like the filing\ndate\n\ndate\n \n\ndate\ngoing to define the document analyzer\n\ngoing to define the document analyzer\n \n\ngoing to define the document analyzer\nschema signature. Uh again super simple\n\nschema signature. Uh again super simple\n \n\nschema signature. Uh again super simple\ninput field which is just the document\n\ninput field which is just the document\n \n\ninput field which is just the document\nitself which is parsed by attachments\n\nitself which is parsed by attachments\n \n\nitself which is parsed by attachments\ngives me the text and the images and\n\ngives me the text and the images and\n \n\ngives me the text and the images and\nthen I'm passing in the document schema\n\nthen I'm passing in the document schema\n \n\nthen I'm passing in the document schema\nparameter which has the document schema\n\nparameter which has the document schema\n \n\nparameter which has the document schema\ntype which is defined above and this is\n\ntype which is defined above and this is\n \n\ntype which is defined above and this is\nthe this is effectively what you would\n\nthe this is effectively what you would\n \n\nthe this is effectively what you would\npass into structured outputs um but just\n\npass into structured outputs um but just\n \n\npass into structured outputs um but just\ndoing it the DS pie where it's going to\n\ndoing it the DS pie where it's going to\n \n\ndoing it the DS pie where it's going to\ngive you um basically the the output in\n\ngive you um basically the the output in\n \n\ngive you um basically the the output in\nthat specific format. So you can see\n\nthat specific format. So you can see\n \n\nthat specific format. So you can see\npulled out things super nicely. Filing\n\npulled out things super nicely. Filing\n \n\npulled out things super nicely. Filing\ndate, form date, form type, transactions\n\ndate, form date, form type, transactions\n \n\ndate, form date, form type, transactions\nthemselves, and then the ultimate\n\nthemselves, and then the ultimate\n \n\nthemselves, and then the ultimate\nanswer. [clears throat] And it's nice\n\nanswer. [clears throat] And it's nice\n \n\nanswer. [clears throat] And it's nice\nbecause it exposes it in a way that you\n\nbecause it exposes it in a way that you\n \n\nbecause it exposes it in a way that you\ncan use dot notation. So you can just\n\ncan use dot notation. So you can just\n \n\ncan use dot notation. So you can just\nvery quickly access the the resulting\n\nvery quickly access the the resulting\n \n\nvery quickly access the the resulting\nobjects. objects. objects. So looking at adapters, um I'll use\n\nSo looking at adapters, um I'll use\n \n\nSo looking at adapters, um I'll use\nanother little tidbit from DSPI, which\n\nanother little tidbit from DSPI, which\n \n\nanother little tidbit from DSPI, which\nis the inspect history. So for those who\n\nis the inspect history. So for those who\n \n\nis the inspect history. So for those who\nwant to know what's going on under the\n\nwant to know what's going on under the\n \n\nwant to know what's going on under the\nhood, inspect history will give you the\n\nhood, inspect history will give you the\n \n\nhood, inspect history will give you the\nraw dump of what's actually going on. So\n\nraw dump of what's actually going on. So\n \n\nraw dump of what's actually going on. So\nyou can see here the system message that\n\nyou can see here the system message that\n \n\nyou can see here the system message that\nwas uh constructed under the hood was\n\nwas uh constructed under the hood was\n \n\nwas uh constructed under the hood was\nall of this. So you can see input fields\n\nall of this. So you can see input fields\n \n\nall of this. So you can see input fields\nare document output fields or reasoning\n\nare document output fields or reasoning\n \n\nare document output fields or reasoning\nand the schema. It's going to pass these\n\nand the schema. It's going to pass these\n \n\nand the schema. It's going to pass these\nin. And then you can see here the actual\n\nin. And then you can see here the actual\n \n\nin. And then you can see here the actual\ndocument content that was extracted and\n\ndocument content that was extracted and\n \n\ndocument content that was extracted and\nput into the text and into the prompt uh\n\nput into the text and into the prompt uh\n \n\nput into the text and into the prompt uh\nwith some metadata. This is all\n\nwith some metadata. This is all\n \n\nwith some metadata. This is all\ngenerated by attachments. And then you\n\ngenerated by attachments. And then you\n \n\ngenerated by attachments. And then you\nget the response which follows this\n\nget the response which follows this\n \n\nget the response which follows this\nspecific format. So you can see the\n\nspecific format. So you can see the\n \n\nspecific format. So you can see the\ndifferent fields that are here. And it's\n\ndifferent fields that are here. And it's\n \n\ndifferent fields that are here. And it's\nthis kind of relatively arbitrary\n\nthis kind of relatively arbitrary\n \n\nthis kind of relatively arbitrary\nresponse um basically format for the for\n\nresponse um basically format for the for\n \n\nresponse um basically format for the for\nthe names which is then parsed by the\n\nthe names which is then parsed by the\n \n\nthe names which is then parsed by the\npie and passed back to you as the user. pie and passed back to you as the user. pie and passed back to you as the user. Um, so I can do okay response.document\n\nUm, so I can do okay response.document\n \n\nUm, so I can do okay response.document\nschema and get the the actual result. schema and get the the actual result. schema and get the the actual result. To show you what the BAML adapter looks\n\nTo show you what the BAML adapter looks\n \n\nTo show you what the BAML adapter looks\nlike, we can basically do two different\n\nlike, we can basically do two different\n \n\nlike, we can basically do two different\ncalls. So this is an example from uh my\n\ncalls. So this is an example from uh my\n \n\ncalls. So this is an example from uh my\nbuddy Pashant uh online again. So what\n\nbuddy Pashant uh online again. So what\n \n\nbuddy Pashant uh online again. So what\nwe do here is define pyantic model super\n\nwe do here is define pyantic model super\n \n\nwe do here is define pyantic model super\nsimple one. Patient address and then\n\nsimple one. Patient address and then\n \n\nsimple one. Patient address and then\npatient details. Patient details has the\n\npatient details. Patient details has the\n \n\npatient details. Patient details has the\npatient address object within it. And\n\npatient address object within it. And\n \n\npatient address object within it. And\nthen we're going to say we're going to\n\nthen we're going to say we're going to\n \n\nthen we're going to say we're going to\ncreate a super simple DSPI signature to\n\ncreate a super simple DSPI signature to\n \n\ncreate a super simple DSPI signature to\nsay taking a clinical note which is a\n\nsay taking a clinical note which is a\n \n\nsay taking a clinical note which is a\nstring. The patient info is the output\n\nstring. The patient info is the output\n \n\nstring. The patient info is the output\ntype. And then note so I'm going to run\n\ntype. And then note so I'm going to run\n \n\ntype. And then note so I'm going to run\nthis two different ways. The first time\n\nthis two different ways. The first time\n \n\nthis two different ways. The first time\nwith the smart LLM that I mentioned\n\nwith the smart LLM that I mentioned\n \n\nwith the smart LLM that I mentioned\nbefore and just use the the built-in\n\nbefore and just use the the built-in\n \n\nbefore and just use the the built-in\nadapter. So I don't specify anything\n\nadapter. So I don't specify anything\n \n\nadapter. So I don't specify anything\nthere. And then the second one will be\n\nthere. And then the second one will be\n \n\nthere. And then the second one will be\nusing the BAML adapter which which is\n\nusing the BAML adapter which which is\n \n\nusing the BAML adapter which which is\ndefined there. Um so I guess a few\n\ndefined there. Um so I guess a few\n \n\ndefined there. Um so I guess a few\nthings going on here." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nOne is the ability\n\nthings going on here. One is the ability\n \n\nthings going on here. One is the ability\nto use Python's uh context which is the\n\nto use Python's uh context which is the\n \n\nto use Python's uh context which is the\nthe lines starting with with width which\n\nthe lines starting with with ..." }, { "role": "assistant", "content": "One is the ability\n\nthings going on here. One is the ability\n \n\nthings going on here. One is the ability\nto use Python's uh context which is the\n\nto use Python's uh context which is the\n \n\nto use Python's uh context which is the\nthe lines starting with with width which\n\nthe lines starting with with width which\n \n\nthe lines starting with with width which\nallow you to basically break out of what\n\nallow you to basically break out of what\n \n\nallow you to basically break out of what\nthe global LLM um has been defined as\n\nthe global LLM um has been defined as\n \n\nthe global LLM um has been defined as\nand use a specific one just for that\n\nand use a specific one just for that\n \n\nand use a specific one just for that\ncall. So you can see in this case I'm\n\ncall. So you can see in this case I'm\n \n\ncall. So you can see in this case I'm\nusing the same LM but if I want to\n\nusing the same LM but if I want to\n \n\nusing the same LM but if I want to\nchange this to like LM anthropic or\n\nchange this to like LM anthropic or\n \n\nchange this to like LM anthropic or\nsomething\n\nsomething\n \n\nsomething\nI think that should work. Um, but\n\nI think that should work. Um, but\n \n\nI think that should work. Um, but\nbasically what that's doing is just\n\nbasically what that's doing is just\n \n\nbasically what that's doing is just\noffloading that call to the other\n\noffloading that call to the other\n \n\noffloading that call to the other\nwhatever LLM that you're defining\n\nwhatever LLM that you're defining\n \n\nwhatever LLM that you're defining\n[clears throat] for that particular call\n\n[clears throat] for that particular call\n \n\n[clears throat] for that particular call\nand something happened. And I'm on a\n\nand something happened. And I'm on a\n \n\nand something happened. And I'm on a\nVPN, so let's kill that. VPN, so let's kill that. VPN, so let's kill that. Sorry, Alex Partners. Sorry, Alex Partners. Sorry, Alex Partners. Okay. Okay, great. So, we had two separate\n\nOkay, great. So, we had two separate\n \n\nOkay, great. So, we had two separate\ncalls. One was to the smart LLM, which\n\ncalls. One was to the smart LLM, which\n \n\ncalls. One was to the smart LLM, which\nis I think 41. The other one was to\n\nis I think 41. The other one was to\n \n\nis I think 41. The other one was to\nAnthropic. Same. Everything else is the\n\nAnthropic. Same. Everything else is the\n \n\nAnthropic. Same. Everything else is the\nexact same. The notes exact same, etc. exact same. The notes exact same, etc. exact same. The notes exact same, etc. We got the same exact output. That's\n\nWe got the same exact output. That's\n \n\nWe got the same exact output. That's\ngreat. But what I wanted to show here is\n\ngreat. But what I wanted to show here is\n \n\ngreat. But what I wanted to show here is\nthe adapters themselves. So in this\n\nthe adapters themselves. So in this\n \n\nthe adapters themselves. So in this\ncase, I'm doing inspect history equals\n\ncase, I'm doing inspect history equals\n \n\ncase, I'm doing inspect history equals\n2. So I'm going to get both of the last\n\n2. So I'm going to get both of the last\n \n\n2. So I'm going to get both of the last\ntwo calls. And we're going to see how\n\ntwo calls. And we're going to see how\n \n\ntwo calls. And we're going to see how\nthe prompts are going to be different. the prompts are going to be different. the prompts are going to be different. And so you can see here the first one,\n\nAnd so you can see here the first one,\n \n\nAnd so you can see here the first one,\nthis is the built-in JSON schema, this\n\nthis is the built-in JSON schema, this\n \n\nthis is the built-in JSON schema, this\ncrazy long JSON string. Yeah, LLMs are\n\ncrazy long JSON string. Yeah, LLMs are\n \n\ncrazy long JSON string. Yeah, LLMs are\ngood enough to to handle that, but um\n\ngood enough to to handle that, but um\n \n\ngood enough to to handle that, but um\nyou know, probably not for super\n\nyou know, probably not for super\n \n\nyou know, probably not for super\ncomplicated ones. Um uh and then you see\n\ncomplicated ones. Um uh and then you see\n \n\ncomplicated ones. Um uh and then you see\nhere for the the second one, it uses the\n\nhere for the the second one, it uses the\n \n\nhere for the the second one, it uses the\nBAML notation, which as we saw in the\n\nBAML notation, which as we saw in the\n \n\nBAML notation, which as we saw in the\nslides, a little bit easier to\n\nslides, a little bit easier to\n \n\nslides, a little bit easier to\ncomprehend. Um and on super complicated\n\ncomprehend. Um and on super complicated\n \n\ncomprehend. Um and on super complicated\nuse cases can actually have a measurable\n\nuse cases can actually have a measurable\n \n\nuse cases can actually have a measurable\nu improvement. u improvement. u improvement. Multimodal example, same sort of thing\n\nMultimodal example, same sort of thing\n \n\nMultimodal example, same sort of thing\nas before. I'll pull in the image\n\nas before. I'll pull in the image\n \n\nas before. I'll pull in the image\nitself. itself. itself. Let's just see what we're working with. Let's just see what we're working with. Let's just see what we're working with. Okay, great. We're looking at these\n\nOkay, great. We're looking at these\n \n\nOkay, great. We're looking at these\nvarious street signs. various street signs. various street signs. And I'm just going to ask it super\n\nAnd I'm just going to ask it super\n \n\nAnd I'm just going to ask it super\nsimple question. It's this time of day. simple question. It's this time of day. simple question. It's this time of day. Can I park here now? When when should I\n\nCan I park here now? When when should I\n \n\nCan I park here now? When when should I\nleave? And you can see I'm just passing\n\nleave? And you can see I'm just passing\n \n\nleave? And you can see I'm just passing\nin again the super simple um shorthand\n\nin again the super simple um shorthand\n \n\nin again the super simple um shorthand\nfor defining a signature which then I\n\nfor defining a signature which then I\n \n\nfor defining a signature which then I\nget out the the var the boolean in this\n\nget out the the var the boolean in this\n \n\nget out the the var the boolean in this\ncase and a string of when I can leave. case and a string of when I can leave. case and a string of when I can leave. Um\n\nUm\n \n\nUm\nso modules themselves it's again fairly\n\nso modules themselves it's again fairly\n \n\nso modules themselves it's again fairly\nsimple. You just kind of wrap all this\n\nsimple. You just kind of wrap all this\n \n\nsimple. You just kind of wrap all this\nin a class. Good question. in a class. Good question. in a class. Good question. >> So does it return reasoning by default\n\n>> So does it return reasoning by default\n \n\n>> So does it return reasoning by default\nalways? always? always? >> Oh good question. Yeah. So when you do\n\n>> Oh good question. Yeah. So when you do\n \n\n>> Oh good question. Yeah. So when you do\n>> can you repeat the question? >> can you repeat the question? >> can you repeat the question? >> Yes. So for those online the question\n\n>> Yes. So for those online the question\n \n\n>> Yes. So for those online the question\nwas does it always return reasoning by\n\nwas does it always return reasoning by\n \n\nwas does it always return reasoning by\ndefault? When you call DSPI.chain chain\n\ndefault? When you call DSPI.chain chain\n \n\ndefault? When you call DSPI.chain chain\nof thought as part of the module where\n\nof thought as part of the module where\n \n\nof thought as part of the module where\nit's built in. It's adding the reasoning\n\nit's built in. It's adding the reasoning\n \n\nit's built in. It's adding the reasoning\nu automatically into your response. So\n\nu automatically into your response. So\n \n\nu automatically into your response. So\nyou're not defining that. It's a great\n\nyou're not defining that. It's a great\n \n\nyou're not defining that. It's a great\nquestion. It's not defined in the\n\nquestion. It's not defined in the\n \n\nquestion. It's not defined in the\nsignature as you can see up here. Uh but\n\nsignature as you can see up here. Uh but\n \n\nsignature as you can see up here." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nUh but\nit will add that in and expose that to\n\nit will add that in and expose that to\n \n\nit will add that in and expose that to\nyou um to the extent that you want to\n\nyou um to the extent that you want to\n \n\nyou um to the extent that you want to\nretain it for any you know any reason. retain it for a..." }, { "role": "assistant", "content": "Uh but\nit will add that in and expose that to\n\nit will add that in and expose that to\n \n\nit will add that in and expose that to\nyou um to the extent that you want to\n\nyou um to the extent that you want to\n \n\nyou um to the extent that you want to\nretain it for any you know any reason. retain it for any you know any reason. retain it for any you know any reason. Uh but that's so if I ju if I changed\n\nUh but that's so if I ju if I changed\n \n\nUh but that's so if I ju if I changed\nthis to predict\n\nthis to predict\n \n\nthis to predict\nyou wouldn't get that same response,\n\nyou wouldn't get that same response,\n \n\nyou wouldn't get that same response,\nright? You just you literally just get\n\nright? You just you literally just get\n \n\nright? You just you literally just get\nthat part. that part. that part. Um so that's actually a good segue to\n\nUm so that's actually a good segue to\n \n\nUm so that's actually a good segue to\nthe modules. Um so module is basically\n\nthe modules. Um so module is basically\n \n\nthe modules. Um so module is basically\njust wrapping all that into some type of\n\njust wrapping all that into some type of\n \n\njust wrapping all that into some type of\nreplic replicable uh logic. Um and so\n\nreplic replicable uh logic. Um and so\n \n\nreplic replicable uh logic. Um and so\nwe're just we're giving it the signature\n\nwe're just we're giving it the signature\n \n\nwe're just we're giving it the signature\nhere. We're saying selfpredict. here. We're saying selfpredict. here. We're saying selfpredict. We're in this case is just a\n\nWe're in this case is just a\n \n\nWe're in this case is just a\ndemonstration of how it's being used as\n\ndemonstration of how it's being used as\n \n\ndemonstration of how it's being used as\na class. So I'll just add this module\n\na class. So I'll just add this module\n \n\na class. So I'll just add this module\nidentifier and sort some sort of counter\n\nidentifier and sort some sort of counter\n \n\nidentifier and sort some sort of counter\nbut this can be any type of arbitrary\n\nbut this can be any type of arbitrary\n \n\nbut this can be any type of arbitrary\nbusiness logic or control flow or any\n\nbusiness logic or control flow or any\n \n\nbusiness logic or control flow or any\ndatabase action or whatever it might be. database action or whatever it might be. database action or whatever it might be. When this image analyzer class is called\n\nWhen this image analyzer class is called\n \n\nWhen this image analyzer class is called\nthis function would run um and then when\n\nthis function would run um and then when\n \n\nthis function would run um and then when\nyou actually invoke it this is when it's\n\nyou actually invoke it this is when it's\n \n\nyou actually invoke it this is when it's\nactually going to run the the core\n\nactually going to run the the core\n \n\nactually going to run the the core\nlogic. And so you can see I'm just\n\nlogic. And so you can see I'm just\n \n\nlogic. And so you can see I'm just\npassing in the So I'm instantiating it\n\npassing in the So I'm instantiating it\n \n\npassing in the So I'm instantiating it\nthe analyzer of AIE123\n\nthe analyzer of AIE123\n \n\nthe analyzer of AIE123\nand then I'll call it. and then I'll call it. and then I'll call it. Great. It called that and you can see\n\nGreat. It called that and you can see\n \n\nGreat. It called that and you can see\nthe counter incrementing each time I\n\nthe counter incrementing each time I\n \n\nthe counter incrementing each time I\nactually make the call. So super simple\n\nactually make the call. So super simple\n \n\nactually make the call. So super simple\nexample. Um we don't have a ton of time\n\nexample. Um we don't have a ton of time\n \n\nexample. Um we don't have a ton of time\nbut I'll I'll show you some of the other\n\nbut I'll I'll show you some of the other\n \n\nbut I'll I'll show you some of the other\nmodules and how that kind of works out. modules and how that kind of works out. modules and how that kind of works out. Terms of tool calling fairly\n\nTerms of tool calling fairly\n \n\nTerms of tool calling fairly\nstraightforward. I'm going to define two\n\nstraightforward. I'm going to define two\n \n\nstraightforward. I'm going to define two\ndifferent functions perplexity search\n\ndifferent functions perplexity search\n \n\ndifferent functions perplexity search\nand get URL content. creating a bioagent\n\nand get URL content. creating a bioagent\n \n\nand get URL content. creating a bioagent\nmodule. So this is going to define\n\nmodule. So this is going to define\n \n\nmodule. So this is going to define\nGemini 25 as this particular module's um\n\nGemini 25 as this particular module's um\n \n\nGemini 25 as this particular module's um\nLLM. It's going to create an answer\n\nLLM. It's going to create an answer\n \n\nLLM. It's going to create an answer\ngenerator object which is a react call. generator object which is a react call. generator object which is a react call. So I'm going to basically do tool\n\nSo I'm going to basically do tool\n \n\nSo I'm going to basically do tool\ncalling whenever this is called and then\n\ncalling whenever this is called and then\n \n\ncalling whenever this is called and then\nthe forward function is literally just\n\nthe forward function is literally just\n \n\nthe forward function is literally just\ncalling that answer generator with the\n\ncalling that answer generator with the\n \n\ncalling that answer generator with the\nparameters that are provided to it. And\n\nparameters that are provided to it. And\n \n\nparameters that are provided to it. And\nthen I'm creating an async version of\n\nthen I'm creating an async version of\n \n\nthen I'm creating an async version of\nthat function as well. that function as well. that function as well. So I can do that here. I'm going to say\n\nSo I can do that here. I'm going to say\n \n\nSo I can do that here. I'm going to say\nokay identify instances where a\n\nokay identify instances where a\n \n\nokay identify instances where a\nparticular person has been at their\n\nparticular person has been at their\n \n\nparticular person has been at their\ncompany for more than 10 years. It needs\n\ncompany for more than 10 years. It needs\n \n\ncompany for more than 10 years. It needs\nto do tool calling to do this to get the\n\nto do tool calling to do this to get the\n \n\nto do tool calling to do this to get the\nmost up-to-date information. And so what\n\nmost up-to-date information. And so what\n \n\nmost up-to-date information. And so what\nthis is doing and basically looping\n\nthis is doing and basically looping\n \n\nthis is doing and basically looping\nthrough um and it's going to call that\n\nthrough um and it's going to call that\n \n\nthrough um and it's going to call that\nbio agent which is using the tool calls\n\nbio agent which is using the tool calls\n \n\nbio agent which is using the tool calls\nin the background and it will make a\n\nin the background and it will make a\n \n\nin the background and it will make a\ndetermination as to whether their\n\ndetermination as to whether their\n \n\ndetermination as to whether their\nbackground is applicable per my\n\nbackground is applicable per my\n \n\nbackground is applicable per my\ncriteria. In this case, Satia is true. criteria. In this case, Satia is true. criteria. In this case, Satia is true. Brian should be false. Uh but what's\n\nBrian should be false. Uh but what's\n \n\nBrian should be false. Uh but what's\ninteresting here while that's going in\n\ninteresting here while that's going in\n \n\ninteresting here while that's going in\nit uh similar to the reasoning uh par or\n\nit uh similar to the reasoning uh par or\n \n\nit uh similar to the reasoning uh par or\nthe reasoning object that you get back\n\nthe reasoning object that you get back\n \n\nthe reasoning object that you get back\nfor chain of thought you can get a\n\nfor chain of thought you can get a\n \n\nfor chain of thought you can get a\ntrajectory back for things like react. trajectory back for things like react. trajectory back for things like react." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nSo you can see what tools it's calling\n\nSo you can see what tools it's calling\n \n\nSo you can see what tools it's calling\nthe arguments that are passed in um and\n\nthe arguments that are passed in um and\n \n\nthe arguments that are passed in um and\nthe observations for each of those calls\n\nthe observati..." }, { "role": "assistant", "content": "So you can see what tools it's calling\n\nSo you can see what tools it's calling\n \n\nSo you can see what tools it's calling\nthe arguments that are passed in um and\n\nthe arguments that are passed in um and\n \n\nthe arguments that are passed in um and\nthe observations for each of those calls\n\nthe observations for each of those calls\n \n\nthe observations for each of those calls\nwhich is nice for debugging and and\n\nwhich is nice for debugging and and\n \n\nwhich is nice for debugging and and\nother obviously other uses. other obviously other uses. other obviously other uses. Um I want to get to the other content so\n\nUm I want to get to the other content so\n \n\nUm I want to get to the other content so\nI'm going to speed through the rest of\n\nI'm going to speed through the rest of\n \n\nI'm going to speed through the rest of\nthis. This is basically an async version\n\nthis. This is basically an async version\n \n\nthis. This is basically an async version\nof the same thing. So you would run both\n\nof the same thing. So you would run both\n \n\nof the same thing. So you would run both\nof them in parallel. Same idea. of them in parallel. Same idea. of them in parallel. Same idea. Um I'm going to skip the JEPA example\n\nUm I'm going to skip the JEPA example\n \n\nUm I'm going to skip the JEPA example\nhere just for a second. Um I can show\n\nhere just for a second. Um I can show\n \n\nhere just for a second. Um I can show\nyou what the output looks like, but\n\nyou what the output looks like, but\n \n\nyou what the output looks like, but\nbasically what this is doing is creating\n\nbasically what this is doing is creating\n \n\nbasically what this is doing is creating\na data set. a data set. a data set. It is showing you what's in the data\n\nIt is showing you what's in the data\n \n\nIt is showing you what's in the data\nset. It's creating a variety of\n\nset. It's creating a variety of\n \n\nset. It's creating a variety of\nsignatures. In this case, it's going to\n\nsignatures. In this case, it's going to\n \n\nsignatures. In this case, it's going to\ncreate a system that categorizes and\n\ncreate a system that categorizes and\n \n\ncreate a system that categorizes and\nclassifies different basically help\n\nclassifies different basically help\n \n\nclassifies different basically help\nmessages um that is part of the data\n\nmessages um that is part of the data\n \n\nmessages um that is part of the data\nset. So, my sync is broken or my light\n\nset. So, my sync is broken or my light\n \n\nset. So, my sync is broken or my light\nis out or whatever it is. They want to\n\nis out or whatever it is. They want to\n \n\nis out or whatever it is. They want to\nclassify whether it's positive, neutral,\n\nclassify whether it's positive, neutral,\n \n\nclassify whether it's positive, neutral,\nor negative and the uh the urgency of\n\nor negative and the uh the urgency of\n \n\nor negative and the uh the urgency of\nthe actual message. It's going to\n\nthe actual message. It's going to\n \n\nthe actual message. It's going to\ncategorize it and then it's going to\n\ncategorize it and then it's going to\n \n\ncategorize it and then it's going to\npack all this stuff, all those different\n\npack all this stuff, all those different\n \n\npack all this stuff, all those different\nmodules into a single support analyzer\n\nmodules into a single support analyzer\n \n\nmodules into a single support analyzer\nmodule. And then from there, what it's\n\nmodule. And then from there, what it's\n \n\nmodule. And then from there, what it's\ngoing to do is define a bunch of metrics\n\ngoing to do is define a bunch of metrics\n \n\ngoing to do is define a bunch of metrics\nwhich is based off of the data set\n\nwhich is based off of the data set\n \n\nwhich is based off of the data set\nitself. So it's going to say, okay, how\n\nitself. So it's going to say, okay, how\n \n\nitself. So it's going to say, okay, how\ndo we score the urgency? This is a a\n\ndo we score the urgency? This is a a\n \n\ndo we score the urgency? This is a a\nvery simple one where it's okay, it\n\nvery simple one where it's okay, it\n \n\nvery simple one where it's okay, it\neither matches or it doesn't. Um, and\n\neither matches or it doesn't. Um, and\n \n\neither matches or it doesn't. Um, and\nthere's other ones where it can be a\n\nthere's other ones where it can be a\n \n\nthere's other ones where it can be a\nlittle bit more subjective and then you\n\nlittle bit more subjective and then you\n \n\nlittle bit more subjective and then you\ncan run it. This going to take too long. can run it. This going to take too long. can run it. This going to take too long. Probably takes 20 minutes or so. Um but\n\nProbably takes 20 minutes or so. Um but\n \n\nProbably takes 20 minutes or so. Um but\nuh what it will do is basically evaluate\n\nuh what it will do is basically evaluate\n \n\nuh what it will do is basically evaluate\nthe performance of the base model and\n\nthe performance of the base model and\n \n\nthe performance of the base model and\nthen apply those metrics uh and\n\nthen apply those metrics uh and\n \n\nthen apply those metrics uh and\niteratively come up with new prompts to\n\niteratively come up with new prompts to\n \n\niteratively come up with new prompts to\nuh to create that. uh to create that. uh to create that. Now I want to pause here just for a\n\nNow I want to pause here just for a\n \n\nNow I want to pause here just for a\nsecond because there's different types\n\nsecond because there's different types\n \n\nsecond because there's different types\nof metrics and in particular for Jeepa\n\nof metrics and in particular for Jeepa\n \n\nof metrics and in particular for Jeepa\nit uses feedback from the teacher model\n\nit uses feedback from the teacher model\n \n\nit uses feedback from the teacher model\nin this case. So it can work with the\n\nin this case. So it can work with the\n \n\nin this case. So it can work with the\nsame level of model, but in particular\n\nsame level of model, but in particular\n \n\nsame level of model, but in particular\nwhen you're trying to use say a smaller\n\nwhen you're trying to use say a smaller\n \n\nwhen you're trying to use say a smaller\nmodel, um it can actually provide\n\nmodel, um it can actually provide\n \n\nmodel, um it can actually provide\ntextual feedback. So, it says not only\n\ntextual feedback. So, it says not only\n \n\ntextual feedback. So, it says not only\ndid you get this classification wrong,\n\ndid you get this classification wrong,\n \n\ndid you get this classification wrong,\nbut it's going to give you some\n\nbut it's going to give you some\n \n\nbut it's going to give you some\nadditional um information or feedback as\n\nadditional um information or feedback as\n \n\nadditional um information or feedback as\nyou can see here for why it got it wrong\n\nyou can see here for why it got it wrong\n \n\nyou can see here for why it got it wrong\nor what the answer should have been,\n\nor what the answer should have been,\n \n\nor what the answer should have been,\nwhich allows it you you can read the\n\nwhich allows it you you can read the\n \n\nwhich allows it you you can read the\npaper, but it basically allows it to um\n\npaper, but it basically allows it to um\n \n\npaper, but it basically allows it to um\niteratively find that kind of paro\n\niteratively find that kind of paro\n \n\niteratively find that kind of paro\nfrontier of how it should uh tweak the\n\nfrontier of how it should uh tweak the\n \n\nfrontier of how it should uh tweak the\nprompt to optimize it based off that\n\nprompt to optimize it based off that\n \n\nprompt to optimize it based off that\nfeedback. It basically just tightens\n\nfeedback. It basically just tightens\n \n\nfeedback. It basically just tightens\nthat iteration loop. that iteration loop. that iteration loop. Um you can see there's a bunch here. Um\n\nUm you can see there's a bunch here. Um\n \n\nUm you can see there's a bunch here. Um\nand then you can run it and see how it\n\nand then you can run it and see how it\n \n\nand then you can run it and see how it\nworks. [snorts] Um but kind of just to\n\nworks. [snorts] Um but kind of just to\n \n\nworks." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\n[snorts] Um but kind of just to\ngive you a concrete example of how it\n\ngive you a concrete example of how it\n \n\ngive you a concrete example of how it\nall comes together. So we took a bunch\n\nall comes together. So we took a bunch\n \n\nall comes together. So we took a bunch\nof those examples from before..." }, { "role": "assistant", "content": "[snorts] Um but kind of just to\ngive you a concrete example of how it\n\ngive you a concrete example of how it\n \n\ngive you a concrete example of how it\nall comes together. So we took a bunch\n\nall comes together. So we took a bunch\n \n\nall comes together. So we took a bunch\nof those examples from before. We're\n\nof those examples from before. We're\n \n\nof those examples from before. We're\nbasically basically going to do a bit of\n\nbasically basically going to do a bit of\n \n\nbasically basically going to do a bit of\num categorization. So I have things like\n\num categorization. So I have things like\n \n\num categorization. So I have things like\ncontracts, I have images, I have\n\ncontracts, I have images, I have\n \n\ncontracts, I have images, I have\ndifferent things that one DSPI program\n\ndifferent things that one DSPI program\n \n\ndifferent things that one DSPI program\ncan comprehend and do some type of\n\ncan comprehend and do some type of\n \n\ncan comprehend and do some type of\nprocessing with. So this is something\n\nprocessing with. So this is something\n \n\nprocessing with. So this is something\nthat we see fairly regularly in terms of\n\nthat we see fairly regularly in terms of\n \n\nthat we see fairly regularly in terms of\nwe might run into a client situation\n\nwe might run into a client situation\n \n\nwe might run into a client situation\nwhere they have just a big dump of of\n\nwhere they have just a big dump of of\n \n\nwhere they have just a big dump of of\nfiles. They don't really know what's in\n\nfiles. They don't really know what's in\n \n\nfiles. They don't really know what's in\nit. They want to find something of uh\n\nit. They want to find something of uh\n \n\nit. They want to find something of uh\nthey want to maybe find SEC filings and\n\nthey want to maybe find SEC filings and\n \n\nthey want to maybe find SEC filings and\nprocess them a certain way. they want to\n\nprocess them a certain way. they want to\n \n\nprocess them a certain way. they want to\nfind contracts and process those a\n\nfind contracts and process those a\n \n\nfind contracts and process those a\ncertain way. Maybe there's some images\n\ncertain way. Maybe there's some images\n \n\ncertain way. Maybe there's some images\nin in there and they want to process\n\nin in there and they want to process\n \n\nin in there and they want to process\nthose a certain way. Uh [snorts] so this\n\nthose a certain way. Uh [snorts] so this\n \n\nthose a certain way. Uh [snorts] so this\nis an example of how you would do that\n\nis an example of how you would do that\n \n\nis an example of how you would do that\nwhere if I start at the bottom here,\n\nwhere if I start at the bottom here,\n \n\nwhere if I start at the bottom here,\nthis is a regular Python file. Um and it\n\nthis is a regular Python file. Um and it\n \n\nthis is a regular Python file. Um and it\nuses DSPI to do all those things I just\n\nuses DSPI to do all those things I just\n \n\nuses DSPI to do all those things I just\nmentioned. So we're pulling in the\n\nmentioned. So we're pulling in the\n \n\nmentioned. So we're pulling in the\nconfigurations,\n\nconfigurations,\n \n\nconfigurations,\nwe're setting the regular LM, the small\n\nwe're setting the regular LM, the small\n \n\nwe're setting the regular LM, the small\nand one we use for an image. As an\n\nand one we use for an image. As an\n \n\nand one we use for an image. As an\nexample, Gemini might Gemini models\n\nexample, Gemini might Gemini models\n \n\nexample, Gemini might Gemini models\nmight be better at image recognition\n\nmight be better at image recognition\n \n\nmight be better at image recognition\nthan others. So I might want to defer or\n\nthan others. So I might want to defer or\n \n\nthan others. So I might want to defer or\nuse a particular model for a particular\n\nuse a particular model for a particular\n \n\nuse a particular model for a particular\nworkload. So if I detect an image, I\n\nworkload. So if I detect an image, I\n \n\nworkload. So if I detect an image, I\nwill route the request to Gemini. If I\n\nwill route the request to Gemini. If I\n \n\nwill route the request to Gemini. If I\ndetect something else, I'll route it to\n\ndetect something else, I'll route it to\n \n\ndetect something else, I'll route it to\na 4.1 or whatever it might be. a 4.1 or whatever it might be. a 4.1 or whatever it might be. So I'm going to process a single file. So I'm going to process a single file. So I'm going to process a single file. And what it does is use our handy\n\nAnd what it does is use our handy\n \n\nAnd what it does is use our handy\nattachments\n\nattachments\n \n\nattachments\num library to put it into a format that\n\num library to put it into a format that\n \n\num library to put it into a format that\nwe can use. And then I'm going to\n\nwe can use. And then I'm going to\n \n\nwe can use. And then I'm going to\nclassify it. And it's not super obvious\n\nclassify it. And it's not super obvious\n \n\nclassify it. And it's not super obvious\nhere, but I'm getting a file type from\n\nhere, but I'm getting a file type from\n \n\nhere, but I'm getting a file type from\nthis classify file uh function call. And\n\nthis classify file uh function call. And\n \n\nthis classify file uh function call. And\nthen I'm doing some different type of\n\nthen I'm doing some different type of\n \n\nthen I'm doing some different type of\nlogic depending on what type of file it\n\nlogic depending on what type of file it\n \n\nlogic depending on what type of file it\nis. So if it's an SEC filing, I do\n\nis. So if it's an SEC filing, I do\n \n\nis. So if it's an SEC filing, I do\ncertain things. If it's a certain type\n\ncertain things. If it's a certain type\n \n\ncertain things. If it's a certain type\nof SEC filing, I do something else. Uh,\n\nof SEC filing, I do something else. Uh,\n \n\nof SEC filing, I do something else. Uh,\nif [snorts] it's a contract, maybe I'll\n\nif [snorts] it's a contract, maybe I'll\n \n\nif [snorts] it's a contract, maybe I'll\nsummarize it. If it's something that\n\nsummarize it. If it's something that\n \n\nsummarize it. If it's something that\nlooks like city infrastructure, in this\n\nlooks like city infrastructure, in this\n \n\nlooks like city infrastructure, in this\ncase, the image that we saw before, I\n\ncase, the image that we saw before, I\n \n\ncase, the image that we saw before, I\nmight do some more visual interpretation\n\nmight do some more visual interpretation\n \n\nmight do some more visual interpretation\nof it. Um, so if I dive into classify\n\nof it. Um, so if I dive into classify\n \n\nof it. Um, so if I dive into classify\nfile super quick,\n\nfile super quick,\n \n\nfile super quick,\nit's running the document classifier. it's running the document classifier. it's running the document classifier. And all that is is basically doing a\n\nAnd all that is is basically doing a\n \n\nAnd all that is is basically doing a\npredict on the image from the file. and\n\npredict on the image from the file. and\n \n\npredict on the image from the file. and\num making sure\n\num making sure\n \n\num making sure\nit returns a type. Where is this\n\nit returns a type. Where is this\n \n\nit returns a type." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nWhere is this\nreturns a type which would be document\n\nreturns a type which would be document\n \n\nreturns a type which would be document\ntype and so you can see here at the end\n\ntype and so you can see here at the end\n \n\ntype and so you can see here at the end\nof the day it's a fairly simple\n\nof the d..." }, { "role": "assistant", "content": "Where is this\nreturns a type which would be document\n\nreturns a type which would be document\n \n\nreturns a type which would be document\ntype and so you can see here at the end\n\ntype and so you can see here at the end\n \n\ntype and so you can see here at the end\nof the day it's a fairly simple\n\nof the day it's a fairly simple\n \n\nof the day it's a fairly simple\nsignature and so what we've done is\n\nsignature and so what we've done is\n \n\nsignature and so what we've done is\nbasically take the PDF file in this case\n\nbasically take the PDF file in this case\n \n\nbasically take the PDF file in this case\ntake all the images from it and take the\n\ntake all the images from it and take the\n \n\ntake all the images from it and take the\nfirst image or first few images in this\n\nfirst image or first few images in this\n \n\nfirst image or first few images in this\ncase a list of images as the input field\n\ncase a list of images as the input field\n \n\ncase a list of images as the input field\nand I'm saying okay just give me the\n\nand I'm saying okay just give me the\n \n\nand I'm saying okay just give me the\ntype what is this and I'm giving it an\n\ntype what is this and I'm giving it an\n \n\ntype what is this and I'm giving it an\noption of these document types so\n\noption of these document types so\n \n\noption of these document types so\nobviously say this is a fairly simple\n\nobviously say this is a fairly simple\n \n\nobviously say this is a fairly simple\nuse case but it's basically saying given\n\nuse case but it's basically saying given\n \n\nuse case but it's basically saying given\nthese three images the first three pages\n\nthese three images the first three pages\n \n\nthese three images the first three pages\nof a document is it an SEC filing is it\n\nof a document is it an SEC filing is it\n \n\nof a document is it an SEC filing is it\na patent filing is the contract city\n\na patent filing is the contract city\n \n\na patent filing is the contract city\ninfrastructure pretty different things\n\ninfrastructure pretty different things\n \n\ninfrastructure pretty different things\nso the model really shouldn't have an\n\nso the model really shouldn't have an\n \n\nso the model really shouldn't have an\nissue with any of those and then we have\n\nissue with any of those and then we have\n \n\nissue with any of those and then we have\na catchall bucket for other and then as\n\na catchall bucket for other and then as\n \n\na catchall bucket for other and then as\nI mentioned before um depending on the\n\nI mentioned before um depending on the\n \n\nI mentioned before um depending on the\nfile type that you get back you can\n\nfile type that you get back you can\n \n\nfile type that you get back you can\nprocess them differently so I'm using\n\nprocess them differently so I'm using\n \n\nprocess them differently so I'm using\nthe small model to do the same type of\n\nthe small model to do the same type of\n \n\nthe small model to do the same type of\nform4 extraction that we saw before um\n\nform4 extraction that we saw before um\n \n\nform4 extraction that we saw before um\nand then asserting basically in this\n\nand then asserting basically in this\n \n\nand then asserting basically in this\ncase that it is what we think it is. Um\n\ncase that it is what we think it is. Um\n \n\ncase that it is what we think it is. Um\na contract in this case we're saying uh\n\na contract in this case we're saying uh\n \n\na contract in this case we're saying uh\nlet's see I have like 10 more minutes so\n\nlet's see I have like 10 more minutes so\n \n\nlet's see I have like 10 more minutes so\nwe can go we'll we'll stop after this uh\n\nwe can go we'll we'll stop after this uh\n \n\nwe can go we'll we'll stop after this uh\nup to this file but for the particular\n\nup to this file but for the particular\n \n\nup to this file but for the particular\ncontract we'll go we'll create this\n\ncontract we'll go we'll create this\n \n\ncontract we'll go we'll create this\nsummarizer object. So we'll go through\n\nsummarizer object. So we'll go through\n \n\nsummarizer object. So we'll go through\nas many pages as there are. We'll do\n\nas many pages as there are. We'll do\n \n\nas many pages as there are. We'll do\nsome uh basically recursive\n\nsome uh basically recursive\n \n\nsome uh basically recursive\nsummarization of that using a separate\n\nsummarization of that using a separate\n \n\nsummarization of that using a separate\nDSPI function and then we'll detect some\n\nDSPI function and then we'll detect some\n \n\nDSPI function and then we'll detect some\ntype of boundaries of that document too. type of boundaries of that document too. type of boundaries of that document too. So we'll say I want the summaries and I\n\nSo we'll say I want the summaries and I\n \n\nSo we'll say I want the summaries and I\nwant the boundaries of the document. Um\n\nwant the boundaries of the document. Um\n \n\nwant the boundaries of the document. Um\nand then we'll print those things out. and then we'll print those things out. and then we'll print those things out. So let's just see if I can run this. So let's just see if I can run this. So let's just see if I can run this. It's going to classify it should as a\n\nIt's going to classify it should as a\n \n\nIt's going to classify it should as a\n[clears throat] contract. >> So is you're just relying on the model\n\n>> So is you're just relying on the model\n \n\n>> So is you're just relying on the model\nitself to realize that it's a city\n\nitself to realize that it's a city\n \n\nitself to realize that it's a city\ninfrastructure. infrastructure. infrastructure. >> Yeah. The question was I'm I'm just\n\n>> Yeah. The question was I'm I'm just\n \n\n>> Yeah. The question was I'm I'm just\nrelying on the model to determine if\n\nrelying on the model to determine if\n \n\nrelying on the model to determine if\nit's a city infrastructure. Yes. I mean\n\nit's a city infrastructure. Yes. I mean\n \n\nit's a city infrastructure. Yes. I mean\nthis is more just like a workshop quick\n\nthis is more just like a workshop quick\n \n\nthis is more just like a workshop quick\nand dirty example. It's only because\n\nand dirty example. It's only because\n \n\nand dirty example. It's only because\nthere's one picture of the street signs. there's one picture of the street signs. there's one picture of the street signs. Um, and if we look in the data folder, I\n\nUm, and if we look in the data folder, I\n \n\nUm, and if we look in the data folder, I\nhave a contract,\n\nhave a contract,\n \n\nhave a contract,\nsome image that's irrelevant, the form\n\nsome image that's irrelevant, the form\n \n\nsome image that's irrelevant, the form\nfor SEC filing, and then the parking\n\nfor SEC filing, and then the parking\n \n\nfor SEC filing, and then the parking\ntoo. Um, they're pretty different. The\n\ntoo. Um, they're pretty different. The\n \n\ntoo. Um, they're pretty different. The\nmodel should have no problem out of\n\nmodel should have no problem out of\n \n\nmodel should have no problem out of\nthose categories that I gave it to\n\nthose categories that I gave it to\n \n\nthose categories that I gave it to\ncategorize it properly. In some type of\n\ncategorize it properly. In some type of\n \n\ncategorize it properly. In some type of\nproduction use case, you would want much\n\nproduction use case, you would want much\n \n\nproduction use case, you would want much\nmore stringent or maybe even multiple\n\nmore stringent or maybe even multiple\n \n\nmore stringent or maybe even multiple\npasses of classification, maybe using\n\npasses of classification, maybe using\n \n\npasses of classification, maybe using\ndifferent models to do that. Um but\n\ndifferent models to do that. Um but\n \n\ndifferent models to do that. Um but\nyeah, given those options, at least the\n\nyeah, given those options, at least the\n \n\nyeah, given those options, at least the\nmany times I've run it, had no problem. many times I've run it, had no problem. many times I've run it, had no problem." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nSo in this case, I gave it um one of\n\nSo in this case, I gave it um one of\n \n\nSo in this case, I gave it um one of\nthese contract documents and it ran some\n\nthese contract documents and it ran some\n \n\nthese contract documents and it ran some\nadditional summarization logic under the\n\nadditional summa..." }, { "role": "assistant", "content": "So in this case, I gave it um one of\n\nSo in this case, I gave it um one of\n \n\nSo in this case, I gave it um one of\nthese contract documents and it ran some\n\nthese contract documents and it ran some\n \n\nthese contract documents and it ran some\nadditional summarization logic under the\n\nadditional summarization logic under the\n \n\nadditional summarization logic under the\nhood. So, if I go to that super quick,\n\nhood. So, if I go to that super quick,\n \n\nhood. So, if I go to that super quick,\num you can find all this in the code,\n\num you can find all this in the code,\n \n\num you can find all this in the code,\nbut basically what it does is use three\n\nbut basically what it does is use three\n \n\nbut basically what it does is use three\nseparate signatures to basically\n\nseparate signatures to basically\n \n\nseparate signatures to basically\ndecompose the contents of the the um the\n\ndecompose the contents of the the um the\n \n\ndecompose the contents of the the um the\ncontract and then summarize them up. So,\n\ncontract and then summarize them up. So,\n \n\ncontract and then summarize them up. So,\nit's basically just iteratively working\n\nit's basically just iteratively working\n \n\nit's basically just iteratively working\nthrough each of the chunks of the\n\nthrough each of the chunks of the\n \n\nthrough each of the chunks of the\ndocument to create a summary that you\n\ndocument to create a summary that you\n \n\ndocument to create a summary that you\nsee here at the bottom. And then just\n\nsee here at the bottom. And then just\n \n\nsee here at the bottom. And then just\nfor good measure, we're also detecting\n\nfor good measure, we're also detecting\n \n\nfor good measure, we're also detecting\nbasically the the boundaries of the\n\nbasically the the boundaries of the\n \n\nbasically the the boundaries of the\ndocument to say, okay, here's out of the\n\ndocument to say, okay, here's out of the\n \n\ndocument to say, okay, here's out of the\n13 pages, you have the main document and\n\n13 pages, you have the main document and\n \n\n13 pages, you have the main document and\nthen some of the exhibits or the\n\nthen some of the exhibits or the\n \n\nthen some of the exhibits or the\nschedules that are a part of it. So, let\n\nschedules that are a part of it. So, let\n \n\nschedules that are a part of it. So, let\nme just bring it up super quick\n\n \n\njust to show you what we're working\n\njust to show you what we're working\n \n\njust to show you what we're working\nwith. This is just some random thing I\n\nwith. This is just some random thing I\n \n\nwith. This is just some random thing I\nfound online. And you can see so it said\n\nfound online. And you can see so it said\n \n\nfound online. And you can see so it said\nthe main document was from page 0 to six\n\nthe main document was from page 0 to six\n \n\nthe main document was from page 0 to six\nand the way and so we zero 1 2 3 4 5 six\n\nand the way and so we zero 1 2 3 4 5 six\n \n\nand the way and so we zero 1 2 3 4 5 six\nseems reasonable. Now we have the start\n\nseems reasonable. Now we have the start\n \n\nseems reasonable. Now we have the start\nof schedule one. of schedule one. of schedule one. Schedule one it says it's the next two\n\nSchedule one it says it's the next two\n \n\nSchedule one it says it's the next two\npages. That looks pretty good. Schedule\n\npages. That looks pretty good. Schedule\n \n\npages. That looks pretty good. Schedule\ntwo is just the one page 9 to9. two is just the one page 9 to9. two is just the one page 9 to9. That looks good. and then schedule three\n\nThat looks good. and then schedule three\n \n\nThat looks good. and then schedule three\nthrough to the end of the document. through to the end of the document. through to the end of the document. And that looks pretty good, too. And so\n\nAnd that looks pretty good, too. And so\n \n\nAnd that looks pretty good, too. And so\nthe way we did that under the hood was\n\nthe way we did that under the hood was\n \n\nthe way we did that under the hood was\nbasically take the PDF, convert it to a\n\nbasically take the PDF, convert it to a\n \n\nbasically take the PDF, convert it to a\nlist of images and then for each of the\n\nlist of images and then for each of the\n \n\nlist of images and then for each of the\nimages pass those to classifier\n\nimages pass those to classifier\n \n\nimages pass those to classifier\num and then use that to\n\num and then use that to\n \n\num and then use that to\nwell let's just look at the code but\n\nwell let's just look at the code but\n \n\nwell let's just look at the code but\nbasically take the list of those\n\nbasically take the list of those\n \n\nbasically take the list of those\nclassifications\n\nclassifications\n \n\nclassifications\ngive that to another DSPI signature to\n\ngive that to another DSPI signature to\n \n\ngive that to another DSPI signature to\nsay given these classifications of the\n\nsay given these classifications of the\n \n\nsay given these classifications of the\ndocument give me the structure and\n\ndocument give me the structure and\n \n\ndocument give me the structure and\nbasically give me a key pair of you name\n\nbasically give me a key pair of you name\n \n\nbasically give me a key pair of you name\nof the section and two integers, a\n\nof the section and two integers, a\n \n\nof the section and two integers, a\ntupole of integers that detect or that\n\ntupole of integers that detect or that\n \n\ntupole of integers that detect or that\nuh determine the um you know the\n\nuh determine the um you know the\n \n\nuh determine the um you know the\nboundaries essentially. Um so that's\n\nboundaries essentially. Um so that's\n \n\nboundaries essentially. Um so that's\nwhat that part does. what that part does. what that part does. Um [clears throat]\n\nUm [clears throat]\n \n\nUm [clears throat]\nif we go back so city infrastructure,\n\nif we go back so city infrastructure,\n \n\nif we go back so city infrastructure,\nI'll do this one super quick just\n\nI'll do this one super quick just\n \n\nI'll do this one super quick just\nbecause it's pretty interesting on how\n\nbecause it's pretty interesting on how\n \n\nbecause it's pretty interesting on how\nit uses tool calls. And while this is\n\nit uses tool calls. And while this is\n \n\nit uses tool calls. And while this is\nrunning,\n\nrunning,\n \n\nrunning,\nI should use the right one. Hold on. I should use the right one. Hold on. I should use the right one. Hold on. >> [clears throat]\n\n \n\n>> Yeah,\n\n>> Yeah,\n \n\n>> Yeah,\n>> good question. The second part like when\n\n>> good question. The second part like when\n \n\n>> good question. The second part like when\nyou generated the list of like my\n\nyou generated the list of like my\n \n\nyou generated the list of like my\ndocuments from 0 to six, did you have\n\ndocuments from 0 to six, did you have\n \n\ndocuments from 0 to six, did you have\nlike original document as an input or\n\nlike original document as an input or\n \n\nlike original document as an input or\nno? no? no? >> No. Uh so let let's just go to that uh\n\n>> No. Uh so let let's just go to that uh\n \n\n>> No. Uh so let let's just go to that uh\nthat was super quick. So\n\nthat was super quick. So\n \n\nthat was super quick. So\nthat should be boundary detector. So, there's a blog post on this that I\n\nSo, there's a blog post on this that I\n \n\nSo, there's a blog post on this that I\npublished probably in August or so that\n\npublished probably in August or so that\n \n\npublished probably in August or so that\ngoes into a little bit more detail. The\n\ngoes into a little bit more detail. The\n \n\ngoes into a little bit more detail. The\ncode is actually pretty crappy in that\n\ncode is actually pretty crappy in that\n \n\ncode is actually pretty crappy in that\none. It's it's going to be better here. one. It's it's going to be better here. one. It's it's going to be better here. Um, but basically what it does is\n\nUm, but basically what it does is\n \n\nUm, but basically what it does is\nthis is probably the main logic. So, for\n\nthis is probably the main logic. So, for\n \n\nthis is probably the main logic." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nSo, for\neach of the images in the PDF, we're\n\neach of the images in the PDF, we're\n \n\neach of the images in the PDF, we're\ngoing to call classify page. going to call classify page. going to call classify page. We're going to gather the results. So\n\nWe're going to gather the results. So\n \n\nWe're goin..." }, { "role": "assistant", "content": "So, for\neach of the images in the PDF, we're\n\neach of the images in the PDF, we're\n \n\neach of the images in the PDF, we're\ngoing to call classify page. going to call classify page. going to call classify page. We're going to gather the results. So\n\nWe're going to gather the results. So\n \n\nWe're going to gather the results. So\nit's doing all that asynchronously\n\nit's doing all that asynchronously\n \n\nit's doing all that asynchronously\npulling it all back saying okay all\n\npulling it all back saying okay all\n \n\npulling it all back saying okay all\nthese you know all the different page\n\nthese you know all the different page\n \n\nthese you know all the different page\nclassifications that there are and then\n\nclassifications that there are and then\n \n\nclassifications that there are and then\nI pass the output of that into a new\n\nI pass the output of that into a new\n \n\nI pass the output of that into a new\nsignature that says given tupil of p I\n\nsignature that says given tupil of p I\n \n\nsignature that says given tupil of p I\ndon't even define it here given tupil of\n\ndon't even define it here given tupil of\n \n\ndon't even define it here given tupil of\npage and classification\n\npage and classification\n \n\npage and classification\ngive me this I don't know relatively\n\ngive me this I don't know relatively\n \n\ngive me this I don't know relatively\ncomplicated output of a dictionary of a\n\ncomplicated output of a dictionary of a\n \n\ncomplicated output of a dictionary of a\nstring tupil integer integer and I give\n\nstring tupil integer integer and I give\n \n\nstring tupil integer integer and I give\nit this set of instructions to say just\n\nit this set of instructions to say just\n \n\nit this set of instructions to say just\ndetect the boundaries. Like this is all\n\ndetect the boundaries. Like this is all\n \n\ndetect the boundaries. Like this is all\nvery like non-production code obviously,\n\nvery like non-production code obviously,\n \n\nvery like non-production code obviously,\nbut the point is that you can do these\n\nbut the point is that you can do these\n \n\nbut the point is that you can do these\ntypes of things super super quickly. types of things super super quickly. types of things super super quickly. Like I'm not specifying much not giving\n\nLike I'm not specifying much not giving\n \n\nLike I'm not specifying much not giving\nit much context and it worked like\n\nit much context and it worked like\n \n\nit much context and it worked like\npretty well. Like it it's worked pretty\n\npretty well. Like it it's worked pretty\n \n\npretty well. Like it it's worked pretty\nwell in most of my testing. Now\n\nwell in most of my testing. Now\n \n\nwell in most of my testing. Now\nobviously there is a ton of low hanging\n\nobviously there is a ton of low hanging\n \n\nobviously there is a ton of low hanging\nfruit in terms of ways to improve that,\n\nfruit in terms of ways to improve that,\n \n\nfruit in terms of ways to improve that,\noptimize it, etc. optimize it, etc. optimize it, etc. Um, but all this is doing is taking that\n\nUm, but all this is doing is taking that\n \n\nUm, but all this is doing is taking that\nsignature, these instructions, and then\n\nsignature, these instructions, and then\n \n\nsignature, these instructions, and then\nI call react. And then all I give it is,\n\nI call react. And then all I give it is,\n \n\nI call react. And then all I give it is,\nuh, the ability to basically\n\nuh, the ability to basically\n \n\nuh, the ability to basically\nself-reflect and call um, get page\n\nself-reflect and call um, get page\n \n\nself-reflect and call um, get page\nimages. So, it says, okay, I'm going to\n\nimages. So, it says, okay, I'm going to\n \n\nimages. So, it says, okay, I'm going to\nlook at this boundary. Well, let me get\n\nlook at this boundary. Well, let me get\n \n\nlook at this boundary. Well, let me get\nthe the page images for these three\n\nthe the page images for these three\n \n\nthe the page images for these three\npages to and make sure basically that\n\npages to and make sure basically that\n \n\npages to and make sure basically that\nthe boundary is correct. And then it\n\nthe boundary is correct. And then it\n \n\nthe boundary is correct. And then it\nuses that to construct the final answer. uses that to construct the final answer. uses that to construct the final answer. And so it's really this is a perfect\n\nAnd so it's really this is a perfect\n \n\nAnd so it's really this is a perfect\nexample of like the tight iteration loop\n\nexample of like the tight iteration loop\n \n\nexample of like the tight iteration loop\nthat you can have both in um building it\n\nthat you can have both in um building it\n \n\nthat you can have both in um building it\nbut then the you can kind of take\n\nbut then the you can kind of take\n \n\nbut then the you can kind of take\nadvantage of the model's introspective\n\nadvantage of the model's introspective\n \n\nadvantage of the model's introspective\nability if you will to use function\n\nability if you will to use function\n \n\nability if you will to use function\ncalls against the data itself the data\n\ncalls against the data itself the data\n \n\ncalls against the data itself the data\nit generated itself etc to kind of keep\n\nit generated itself etc to kind of keep\n \n\nit generated itself etc to kind of keep\nthat loop going. question. that loop going. question. that loop going. question. >> So under the hood, the the beauty of ESP\n\n>> So under the hood, the the beauty of ESP\n \n\n>> So under the hood, the the beauty of ESP\nthen is that it enforces kind of\n\nthen is that it enforces kind of\n \n\nthen is that it enforces kind of\nstructured output on a on a model. structured output on a on a model. structured output on a on a model. >> I mean yes, I think that's probably\n\n>> I mean yes, I think that's probably\n \n\n>> I mean yes, I think that's probably\nreductive of of like its full potential,\n\nreductive of of like its full potential,\n \n\nreductive of of like its full potential,\nbut generally that's that's correct. I\n\nbut generally that's that's correct. I\n \n\nbut generally that's that's correct. I\nmean yes, you can use structured\n\nmean yes, you can use structured\n \n\nmean yes, you can use structured\noutputs, but you have to do a bunch of\n\noutputs, but you have to do a bunch of\n \n\noutputs, but you have to do a bunch of\ncrap basically to coordinate like\n\ncrap basically to coordinate like\n \n\ncrap basically to coordinate like\nfeeding all the feeding that into the\n\nfeeding all the feeding that into the\n \n\nfeeding all the feeding that into the\nrest of the program. maybe you want to\n\nrest of the program. maybe you want to\n \n\nrest of the program. maybe you want to\ncall a model differently or use XML here\n\ncall a model differently or use XML here\n \n\ncall a model differently or use XML here\nor use a different type of model or\n\nor use a different type of model or\n \n\nor use a different type of model or\nwhatever it might be\n\nwhatever it might be\n \n\nwhatever it might be\num to to do that. So absolutely yeah I'm\n\num to to do that. So absolutely yeah I'm\n \n\num to to do that. So absolutely yeah I'm\nnot saying this is the only way\n\nnot saying this is the only way\n \n\nnot saying this is the only way\nobviously to kind of create these\n\nobviously to kind of create these\n \n\nobviously to kind of create these\napplications or that you shouldn't use\n\napplications or that you shouldn't use\n \n\napplications or that you shouldn't use\nPantic or shouldn't use structured\n\nPantic or shouldn't use structured\n \n\nPantic or shouldn't use structured\noutputs. You absolutely should. Um, it's\n\noutputs. You absolutely should. Um, it's\n \n\noutputs. You absolutely should." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nUm, it's\njust a way that once you kind of wrap\n\njust a way that once you kind of wrap\n \n\njust a way that once you kind of wrap\nyour head around the the primitives that\n\nyour head around the the primitives that\n \n\nyour head around the the primitives that\nDSPI gives you, you can start to very\n\nDSPI gi..." }, { "role": "assistant", "content": "Um, it's\njust a way that once you kind of wrap\n\njust a way that once you kind of wrap\n \n\njust a way that once you kind of wrap\nyour head around the the primitives that\n\nyour head around the the primitives that\n \n\nyour head around the the primitives that\nDSPI gives you, you can start to very\n\nDSPI gives you, you can start to very\n \n\nDSPI gives you, you can start to very\nquickly build these types of\n\nquickly build these types of\n \n\nquickly build these types of\narguably uh I mean these are like\n\narguably uh I mean these are like\n \n\narguably uh I mean these are like\nprototypes right now, but like if you\n\nprototypes right now, but like if you\n \n\nprototypes right now, but like if you\nwant to take this to the next level to\n\nwant to take this to the next level to\n \n\nwant to take this to the next level to\nproduction scale, you have all the\n\nproduction scale, you have all the\n \n\nproduction scale, you have all the\npieces in front of you to be able to do\n\npieces in front of you to be able to do\n \n\npieces in front of you to be able to do\nthat. that. that. >> Um,\n\n>> Um,\n \n\n>> Um,\nany other questions? I probably got\n\nany other questions? I probably got\n \n\nany other questions? I probably got\nabout five minutes left. Go ahead. Can\n\nabout five minutes left. Go ahead. Can\n \n\nabout five minutes left. Go ahead. Can\nyou talk about your experience using\n\nyou talk about your experience using\n \n\nyou talk about your experience using\noptimization\n\noptimization\n \n\noptimization\nand just\n\nand just\n \n\nand just\n>> Yeah. Yeah. So Jeep and actually I'll\n\n>> Yeah. Yeah. So Jeep and actually I'll\n \n\n>> Yeah. Yeah. So Jeep and actually I'll\npull up uh I I ran one right before\n\npull up uh I I ran one right before\n \n\npull up uh I I ran one right before\nthis. Um this uses a a different\n\nthis. Um this uses a a different\n \n\nthis. Um this uses a a different\nalgorithm called my row but basically um\n\nalgorithm called my row but basically um\n \n\nalgorithm called my row but basically um\nthe optimizers as long as you have well\n\nthe optimizers as long as you have well\n \n\nthe optimizers as long as you have well\nstructured data. So for the machine\n\nstructured data. So for the machine\n \n\nstructured data. So for the machine\nlearning folks in the room, which is\n\nlearning folks in the room, which is\n \n\nlearning folks in the room, which is\nprobably everybody, obviously the\n\nprobably everybody, obviously the\n \n\nprobably everybody, obviously the\nquality of your of your data is very\n\nquality of your of your data is very\n \n\nquality of your of your data is very\nimportant,\n\nimportant,\n \n\nimportant,\num you don't need thousands and\n\num you don't need thousands and\n \n\num you don't need thousands and\nthousands of examples necessarily, but\n\nthousands of examples necessarily, but\n \n\nthousands of examples necessarily, but\nas long as you have enough, maybe 10 to\n\nas long as you have enough, maybe 10 to\n \n\nas long as you have enough, maybe 10 to\n100 of inputs and outputs. 100 of inputs and outputs. 100 of inputs and outputs. [clears throat] And if you're\n\n[clears throat] And if you're\n \n\n[clears throat] And if you're\nconstructing your metrics in a way that\n\nconstructing your metrics in a way that\n \n\nconstructing your metrics in a way that\nis relatively intuitive and and that,\n\nis relatively intuitive and and that,\n \n\nis relatively intuitive and and that,\nyou know, accurately describes what\n\nyou know, accurately describes what\n \n\nyou know, accurately describes what\nyou're trying to achieve, the\n\nyou're trying to achieve, the\n \n\nyou're trying to achieve, the\nimprovement can be pretty significant. improvement can be pretty significant. improvement can be pretty significant. Um, and so that time entry corrector\n\nUm, and so that time entry corrector\n \n\nUm, and so that time entry corrector\nthing that I mentioned before, uh, you\n\nthing that I mentioned before, uh, you\n \n\nthing that I mentioned before, uh, you\ncan see the output of here. It's kind of\n\ncan see the output of here. It's kind of\n \n\ncan see the output of here. It's kind of\niterating through. It's measuring the\n\niterating through. It's measuring the\n \n\niterating through. It's measuring the\noutput metrics for each of these. And\n\noutput metrics for each of these. And\n \n\noutput metrics for each of these. And\nthen you can see all the way at the\n\nthen you can see all the way at the\n \n\nthen you can see all the way at the\nbottom once it goes through all of its\n\nbottom once it goes through all of its\n \n\nbottom once it goes through all of its\noptimization stuff. optimization stuff. optimization stuff. You can see the actual performance\n\nYou can see the actual performance\n \n\nYou can see the actual performance\non\n\non\n \n\non\num, the basic versus the optimized\n\num, the basic versus the optimized\n \n\num, the basic versus the optimized\nmodel. In this case, went from 86 to 89. model. In this case, went from 86 to 89. model. In this case, went from 86 to 89. And then interestingly, this is still in\n\nAnd then interestingly, this is still in\n \n\nAnd then interestingly, this is still in\ndevelopment, this one in particular, but\n\ndevelopment, this one in particular, but\n \n\ndevelopment, this one in particular, but\nyou can break it down by metric. So you\n\nyou can break it down by metric. So you\n \n\nyou can break it down by metric. So you\ncan see where the model's optimizing\n\ncan see where the model's optimizing\n \n\ncan see where the model's optimizing\nbetter, performing better across certain\n\nbetter, performing better across certain\n \n\nbetter, performing better across certain\nmetrics. And this can be really telling\n\nmetrics. And this can be really telling\n \n\nmetrics. And this can be really telling\nas to whether you need to tweak your\n\nas to whether you need to tweak your\n \n\nas to whether you need to tweak your\nmetric, maybe you need to decompose your\n\nmetric, maybe you need to decompose your\n \n\nmetric, maybe you need to decompose your\nmetric, maybe there's other areas within\n\nmetric, maybe there's other areas within\n \n\nmetric, maybe there's other areas within\nyour data set, or the the basically the\n\nyour data set, or the the basically the\n \n\nyour data set, or the the basically the\nstructure of your program that you can\n\nstructure of your program that you can\n \n\nstructure of your program that you can\nimprove. Um, but it's a really nice way\n\nimprove. Um, but it's a really nice way\n \n\nimprove. Um, but it's a really nice way\nto understand what's going under the\n\nto understand what's going under the\n \n\nto understand what's going under the\nunder the hood. And if if you don't care\n\nunder the hood. And if if you don't care\n \n\nunder the hood. And if if you don't care\nabout some of these and the optimizer\n\nabout some of these and the optimizer\n \n\nabout some of these and the optimizer\nisn't doing as well on them, maybe you\n\nisn't doing as well on them, maybe you\n \n\nisn't doing as well on them, maybe you\ncan maybe you can throw them out, too. can maybe you can throw them out, too. can maybe you can throw them out, too. So, it's it's a very kind of flexible\n\nSo, it's it's a very kind of flexible\n \n\nSo, it's it's a very kind of flexible\nsystem, flexible way of kind of doing\n\nsystem, flexible way of kind of doing\n \n\nsystem, flexible way of kind of doing\nall that. all that. all that. >> Yeah. What's the output of the\n\n>> Yeah. What's the output of the\n \n\n>> Yeah. What's the output of the\noptimization? Like what do you get out\n\noptimization? Like what do you get out\n \n\noptimization? Like what do you get out\nof it and then how do you use that\n\nof it and then how do you use that\n \n\nof it and then how do you use that\nobject, whatever it is? object, whatever it is? object, whatever it is? >> Yeah. Yeah. So the output of the\n\n>> Yeah. Yeah. So the output of the\n \n\n>> Yeah. Yeah. So the output of the\noptimizers is basically just another um\n\noptimizers is basically just another um\n \n\noptimizers is basically just another um\nit's almost like a compiled object if\n\nit's almost like a compiled object if\n \n\nit's almost like a compiled object if\nyou will. you will. you will. >> So DSPI allows you to save and load\n\n>> So DSPI allows you to save and load\n \n\n>> So DSPI allows you to save and load\nprograms as well. So the output of the\n\nprograms as well. So the output of the\n \n\nprograms as well." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nSo the output of the\noptimizer is basically just a module\n\noptimizer is basically just a module\n \n\noptimizer is basically just a module\nthat you can then serialize and save off\n\nthat you can then serialize and save off\n \n\nthat you can then serialize and save off\nsomewhere\n\nsomewhere\n \n\nsomewhere\n>> ..." }, { "role": "assistant", "content": "So the output of the\noptimizer is basically just a module\n\noptimizer is basically just a module\n \n\noptimizer is basically just a module\nthat you can then serialize and save off\n\nthat you can then serialize and save off\n \n\nthat you can then serialize and save off\nsomewhere\n\nsomewhere\n \n\nsomewhere\n>> or you can call it later uh as you would\n\n>> or you can call it later uh as you would\n \n\n>> or you can call it later uh as you would\nany other module\n\nany other module\n \n\nany other module\n>> and it's just manipulating the phrasing\n\n>> and it's just manipulating the phrasing\n \n\n>> and it's just manipulating the phrasing\nof the prompt. So like what is it\n\nof the prompt. So like what is it\n \n\nof the prompt. So like what is it\nactually like you know what's its\n\nactually like you know what's its\n \n\nactually like you know what's its\nsolution space look like? solution space look like? solution space look like? >> Yeah. Yeah. under the hood, it's\n\n>> Yeah. Yeah. under the hood, it's\n \n\n>> Yeah. Yeah. under the hood, it's\nliterally just iterating on the actual\n\nliterally just iterating on the actual\n \n\nliterally just iterating on the actual\nprompt itself. Maybe it's adding\n\nprompt itself. Maybe it's adding\n \n\nprompt itself. Maybe it's adding\nadditional instructions. It's saying,\n\nadditional instructions. It's saying,\n \n\nadditional instructions. It's saying,\n\"Well, I keep failing on this particular\n\n\"Well, I keep failing on this particular\n \n\n\"Well, I keep failing on this particular\nthing, like not capitalizing the names\n\nthing, like not capitalizing the names\n \n\nthing, like not capitalizing the names\ncorrectly. I need to add [clears throat]\n\ncorrectly. I need to add [clears throat]\n \n\ncorrectly. I need to add [clears throat]\nin my upfront criteria in the prompt an\n\nin my upfront criteria in the prompt an\n \n\nin my upfront criteria in the prompt an\ninstruction to the model to say you must\n\ninstruction to the model to say you must\n \n\ninstruction to the model to say you must\ncapitalize names properly.\" And Chris uh\n\ncapitalize names properly.\" And Chris uh\n \n\ncapitalize names properly.\" And Chris uh\nwho I mentioned before has a really good\n\nwho I mentioned before has a really good\n \n\nwho I mentioned before has a really good\nway of putting this and I'm going to\n\nway of putting this and I'm going to\n \n\nway of putting this and I'm going to\nbutcher it now, but like the optimizer\n\nbutcher it now, but like the optimizer\n \n\nbutcher it now, but like the optimizer\nis basically finding latent requirements\n\nis basically finding latent requirements\n \n\nis basically finding latent requirements\nthat you might not have specified\n\nthat you might not have specified\n \n\nthat you might not have specified\ninitially up front, but based off of the\n\ninitially up front, but based off of the\n \n\ninitially up front, but based off of the\ndata, it's kind of like a poor man's\n\ndata, it's kind of like a poor man's\n \n\ndata, it's kind of like a poor man's\ndeep learning, I guess, but like it's\n\ndeep learning, I guess, but like it's\n \n\ndeep learning, I guess, but like it's\nlearning from the data. It's learning\n\nlearning from the data. It's learning\n \n\nlearning from the data. It's learning\nwhat it's doing well, what what it's\n\nwhat it's doing well, what what it's\n \n\nwhat it's doing well, what what it's\ndoing not so well, and it's dynamically\n\ndoing not so well, and it's dynamically\n \n\ndoing not so well, and it's dynamically\nconstructing a prompt that improves the\n\nconstructing a prompt that improves the\n \n\nconstructing a prompt that improves the\nperformance based off of your metrics. performance based off of your metrics. performance based off of your metrics. And is that like LMG guided like is it\n\nAnd is that like LMG guided like is it\n \n\nAnd is that like LMG guided like is it\nlike about like capitalization? like about like capitalization? like about like capitalization? >> Yeah. Yeah. Question being is it all LLM\n\n>> Yeah. Yeah. Question being is it all LLM\n \n\n>> Yeah. Yeah. Question being is it all LLM\nguided? Yes. Particularly for Jeepa it's\n\nguided? Yes. Particularly for Jeepa it's\n \n\nguided? Yes. Particularly for Jeepa it's\nusing LLM to improve LLM's performance. using LLM to improve LLM's performance. using LLM to improve LLM's performance. So it's using the LLM to dynamically\n\nSo it's using the LLM to dynamically\n \n\nSo it's using the LLM to dynamically\nconstruct new prompts which are then fed\n\nconstruct new prompts which are then fed\n \n\nconstruct new prompts which are then fed\ninto the system measured and then it\n\ninto the system measured and then it\n \n\ninto the system measured and then it\nkind of iterates. So it's using AI to\n\nkind of iterates. So it's using AI to\n \n\nkind of iterates. So it's using AI to\nbuild AI if you will. build AI if you will. build AI if you will. >> Thank you. >> Thank you. >> Thank you. >> Yeah\n\n>> Yeah\n \n\n>> Yeah\nquestion. Why is the solution object not\n\nquestion. Why is the solution object not\n \n\nquestion. Why is the solution object not\njust the optimized prompt? just the optimized prompt? just the optimized prompt? >> Why is the solution object not what? >> Why is the solution object not what? >> Why is the solution object not what? >> Not just the optimized prompt. Why are\n\n>> Not just the optimized prompt. Why are\n \n\n>> Not just the optimized prompt. Why are\nyou using\n\nyou using\n \n\nyou using\n>> Oh, absolutely is. You can get it under\n\n>> Oh, absolutely is. You can get it under\n \n\n>> Oh, absolutely is. You can get it under\nthe hood. I mean, you can The question\n\nthe hood. I mean, you can The question\n \n\nthe hood. I mean, you can The question\nwas why don't you just get the optimized\n\nwas why don't you just get the optimized\n \n\nwas why don't you just get the optimized\nprompt? You can absolutely. Um,\n\nprompt? You can absolutely. Um,\n \n\nprompt? You can absolutely. Um,\n>> what else is there besides\n\n>> what else is there besides\n \n\n>> what else is there besides\n>> the the So, what else is there other\n\n>> the the So, what else is there other\n \n\n>> the the So, what else is there other\nthan the prompt? The DSPI object itself. than the prompt? The DSPI object itself. than the prompt? The DSPI object itself. So the module the way things um well we\n\nSo the module the way things um well we\n \n\nSo the module the way things um well we\ncan probably look at one if we have\n\ncan probably look at one if we have\n \n\ncan probably look at one if we have\ntime. Um\n\ntime. Um\n \n\ntime. Um\n>> if I could see a dump of what gets you\n\n>> if I could see a dump of what gets you\n \n\n>> if I could see a dump of what gets you\nknow what is the optimized state that\n\nknow what is the optimized state that\n \n\nknow what is the optimized state that\nwould be interesting. would be interesting. would be interesting. >> Yeah. Yeah sure. Let me see if I can\n\n>> Yeah. Yeah sure. Let me see if I can\n \n\n>> Yeah. Yeah sure. Let me see if I can\nfind one quick. Um but fundamentally at\n\nfind one quick. Um but fundamentally at\n \n\nfind one quick. Um but fundamentally at\nthe end of the day yes you get an\n\nthe end of the day yes you get an\n \n\nthe end of the day yes you get an\noptimized prompt a string that you can\n\noptimized prompt a string that you can\n \n\noptimized prompt a string that you can\ndump somewhere if you if you want to. Um\n\ndump somewhere if you if you want to. Um\n \n\ndump somewhere if you if you want to. Um\nactually\n\nactually\n \n\nactually\num\n\num\n \n\num\n>> there's a lot of pieces to the\n\n>> there's a lot of pieces to the\n \n\n>> there's a lot of pieces to the\nsignature, right? So it's like how you\n\nsignature, right? So it's like how you\n \n\nsignature, right? So it's like how you\ndescribe your feels in the doc. describe your feels in the doc. describe your feels in the doc. >> This is a perfect segue and I'll I'll\n\n>> This is a perfect segue and I'll I'll\n \n\n>> This is a perfect segue and I'll I'll\nconclude right after this. I was playing\n\nconclude right after this. I was playing\n \n\nconclude right after this." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nI was playing\naround with something I was well I was\n\naround with something I was well I was\n \n\naround with something I was well I was\nplaying around this thing called DSPIHub\n\nplaying around this thing called DSPIHub\n \n\nplaying around this thing called DSPIHub\nthat I kind of created to create a\n\nth..." }, { "role": "assistant", "content": "I was playing\naround with something I was well I was\n\naround with something I was well I was\n \n\naround with something I was well I was\nplaying around this thing called DSPIHub\n\nplaying around this thing called DSPIHub\n \n\nplaying around this thing called DSPIHub\nthat I kind of created to create a\n\nthat I kind of created to create a\n \n\nthat I kind of created to create a\nrepository of optimized programs. So\n\nrepository of optimized programs. So\n \n\nrepository of optimized programs. So\nbasically like if you're an expert in\n\nbasically like if you're an expert in\n \n\nbasically like if you're an expert in\nwhatever you optimize an LLM against\n\nwhatever you optimize an LLM against\n \n\nwhatever you optimize an LLM against\nthis data set or have a great classifier\n\nthis data set or have a great classifier\n \n\nthis data set or have a great classifier\nfor city infrastructure images or\n\nfor city infrastructure images or\n \n\nfor city infrastructure images or\nwhatever kind of like a hugging face you\n\nwhatever kind of like a hugging face you\n \n\nwhatever kind of like a hugging face you\ncan download something that has been\n\ncan download something that has been\n \n\ncan download something that has been\npre-optimized\n\npre-optimized\n \n\npre-optimized\nand then what I have here this is the\n\nand then what I have here this is the\n \n\nand then what I have here this is the\nactual loaded program this would be the\n\nactual loaded program this would be the\n \n\nactual loaded program this would be the\noutput of the optimized process or it it\n\noutput of the optimized process or it it\n \n\noutput of the optimized process or it it\nis and then I can call it as I would any\n\nis and then I can call it as I would any\n \n\nis and then I can call it as I would any\nanything else and so you You can see\n\nanything else and so you You can see\n \n\nanything else and so you You can see\nhere this is the output and I used the\n\nhere this is the output and I used the\n \n\nhere this is the output and I used the\noptimized program that I downloaded from\n\noptimized program that I downloaded from\n \n\noptimized program that I downloaded from\nfrom this hub. And if we inspect maybe\n\nfrom this hub. And if we inspect maybe\n \n\nfrom this hub. And if we inspect maybe\nthe loaded program,\n\nthe loaded program,\n \n\nthe loaded program,\nyou can see under the hood, it's a\n\nyou can see under the hood, it's a\n \n\nyou can see under the hood, it's a\npredict object with a string signature\n\npredict object with a string signature\n \n\npredict object with a string signature\nof time and reasoning. Here is the\n\nof time and reasoning. Here is the\n \n\nof time and reasoning. Here is the\noptimized prompt. Ultimately,\n\n \n\nthis is the output of the optimization\n\nthis is the output of the optimization\n \n\nthis is the output of the optimization\nprocess. this long string here. process. this long string here. process. this long string here. Um, and then the various uh\n\nUm, and then the various uh\n \n\nUm, and then the various uh\nspecifications and definitions of the\n\nspecifications and definitions of the\n \n\nspecifications and definitions of the\ninputs and outputs. inputs and outputs. inputs and outputs. >> Have you found specific uses of those? >> Have you found specific uses of those? >> Have you found specific uses of those? Like to his question like what is it? Like to his question like what is it? Like to his question like what is it? What can you do with that? What can you do with that? What can you do with that? >> It's up to your it's up to your use\n\n>> It's up to your it's up to your use\n \n\n>> It's up to your it's up to your use\ncase. So if I if I have a so a document\n\ncase. So if I if I have a so a document\n \n\ncase. So if I if I have a so a document\nclassifier might be a good example. If\n\nclassifier might be a good example. If\n \n\nclassifier might be a good example. If\nin my business I come across whatever\n\nin my business I come across whatever\n \n\nin my business I come across whatever\ndocuments of a certain type, I might\n\ndocuments of a certain type, I might\n \n\ndocuments of a certain type, I might\noptimize a classifier against those and\n\noptimize a classifier against those and\n \n\noptimize a classifier against those and\nthen I can use that somewhere else on a\n\nthen I can use that somewhere else on a\n \n\nthen I can use that somewhere else on a\ndifferent project or something like\n\ndifferent project or something like\n \n\ndifferent project or something like\nthat. So out of 100,000 documents, I\n\nthat. So out of 100,000 documents, I\n \n\nthat. So out of 100,000 documents, I\nwant to find only the pages that have an\n\nwant to find only the pages that have an\n \n\nwant to find only the pages that have an\ninvoice on it as an example. Now sure\n\ninvoice on it as an example. Now sure\n \n\ninvoice on it as an example. Now sure\n100% you can use a typical ML classifier\n\n100% you can use a typical ML classifier\n \n\n100% you can use a typical ML classifier\nto do that. That's great. This is just\n\nto do that. That's great. This is just\n \n\nto do that. That's great. This is just\nan example. We can also theoretically\n\nan example. We can also theoretically\n \n\nan example. We can also theoretically\ntrain or optimize a model to do that\n\ntrain or optimize a model to do that\n \n\ntrain or optimize a model to do that\ntype of classification or some type of\n\ntype of classification or some type of\n \n\ntype of classification or some type of\ngeneration of text or what have you\n\ngeneration of text or what have you\n \n\ngeneration of text or what have you\nwhich then you have the optimized state\n\nwhich then you have the optimized state\n \n\nwhich then you have the optimized state\nof which then lives in your data\n\nof which then lives in your data\n \n\nof which then lives in your data\nprocessing pipeline you know and you can\n\nprocessing pipeline you know and you can\n \n\nprocessing pipeline you know and you can\nuse it for other types of purposes or\n\nuse it for other types of purposes or\n \n\nuse it for other types of purposes or\ngive it to other teams or whatever it\n\ngive it to other teams or whatever it\n \n\ngive it to other teams or whatever it\nmight be. So it's just up to your\n\nmight be. So it's just up to your\n \n\nmight be. So it's just up to your\nparticular use case. um something like\n\nparticular use case. um something like\n \n\nparticular use case. um something like\nthis like hub who maybe it's not useful\n\nthis like hub who maybe it's not useful\n \n\nthis like hub who maybe it's not useful\nbecause each individual's use case is so\n\nbecause each individual's use case is so\n \n\nbecause each individual's use case is so\nhyper specific I don't really know but\n\nhyper specific I don't really know but\n \n\nhyper specific I don't really know but\num yeah you can do with it kind of\n\num yeah you can do with it kind of\n \n\num yeah you can do with it kind of\nwhatever you want last question yeah\n\nwhatever you want last question yeah\n \n\nwhatever you want last question yeah\n>> is generally you know like using DSP\n\n>> is generally you know like using DSP\n \n\n>> is generally you know like using DSP\nsomething where people kind of do\n\nsomething where people kind of do\n \n\nsomething where people kind of do\nreplays just to optimize their prop or\n\nreplays just to optimize their prop or\n \n\nreplays just to optimize their prop or\nis there a way to sort of do it in real\n\nis there a way to sort of do it in real\n \n\nis there a way to sort of do it in real\ntime given delays\n\ntime given delays\n \n\ntime given delays\nWhat I mean by delayed is okay chat GPT\n\nWhat I mean by delayed is okay chat GPT\n \n\nWhat I mean by delayed is okay chat GPT\ngives you your answer and you can thumbs\n\ngives you your answer and you can thumbs\n \n\ngives you your answer and you can thumbs\nup or thumbs down. You know that thumbs\n\nup or thumbs down. You know that thumbs\n \n\nup or thumbs down. You know that thumbs\nup comes you know 10 minutes later, 30\n\nup comes you know 10 minutes later, 30\n \n\nup comes you know 10 minutes later, 30\nminutes later, a day later, right? minutes later, a day later, right? minutes later, a day later, right?" } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\n>> So is the question more about like\n\n>> So is the question more about like\n \n\n>> So is the question more about like\ncontinuous learning like how would you\n\ncontinuous learning like how would you\n \n\ncontinuous learning like how would you\ndo that here? do that here? do that here? >> You can be the j..." }, { "role": "assistant", "content": ">> So is the question more about like\n\n>> So is the question more about like\n \n\n>> So is the question more about like\ncontinuous learning like how would you\n\ncontinuous learning like how would you\n \n\ncontinuous learning like how would you\ndo that here? do that here? do that here? >> You can be the judge. >> You can be the judge. >> You can be the judge. >> Well, how are you feeding back delayed\n\n>> Well, how are you feeding back delayed\n \n\n>> Well, how are you feeding back delayed\nmetrics to optimize it? Why why would it\n\nmetrics to optimize it? Why why would it\n \n\nmetrics to optimize it? Why why would it\nneed to be delayed? Because you know\n\nneed to be delayed? Because you know\n \n\nneed to be delayed? Because you know\nusually the feedback is from the user,\n\nusually the feedback is from the user,\n \n\nusually the feedback is from the user,\nright? Like delayed. >> Yeah. Well, then\n\n>> Yeah. Well, then\n \n\n>> Yeah. Well, then\n>> yeah, that's right. You it basically be\n\n>> yeah, that's right. You it basically be\n \n\n>> yeah, that's right. You it basically be\nadded to the data set and then you would\n\nadded to the data set and then you would\n \n\nadded to the data set and then you would\nuse the latest optimize and just keep\n\nuse the latest optimize and just keep\n \n\nuse the latest optimize and just keep\nkeep optimizing off of that\n\nkeep optimizing off of that\n \n\nkeep optimizing off of that\n>> ground truth data set. >> ground truth data set. >> ground truth data set. >> That's right. >> That's right. >> That's right. >> You will collect the outputs of your\n\n>> You will collect the outputs of your\n \n\n>> You will collect the outputs of your\noptimization and feed it back and the\n\noptimization and feed it back and the\n \n\noptimization and feed it back and the\nloop hits. loop hits. loop hits. >> Yeah. But that Why you're trying to do\n\n>> Yeah. But that Why you're trying to do\n \n\n>> Yeah. But that Why you're trying to do\noffline optimization, right? offline optimization, right? offline optimization, right? >> Yes. >> Yes. >> Yes. >> But I'm I'm asking, can you do this\n\n>> But I'm I'm asking, can you do this\n \n\n>> But I'm I'm asking, can you do this\nonline where with the metric feedback? online where with the metric feedback? online where with the metric feedback? >> If you're good if you're a good enough\n\n>> If you're good if you're a good enough\n \n\n>> If you're good if you're a good enough\nengineer, you probably do it. But\n\nengineer, you probably do it. But\n \n\nengineer, you probably do it. But\n>> I'm not I'm not recommending replacing\n\n>> I'm not I'm not recommending replacing\n \n\n>> I'm not I'm not recommending replacing\nML models with like optimized DSPI\n\nML models with like optimized DSPI\n \n\nML models with like optimized DSPI\nprograms for particular use cases. Maybe\n\nprograms for particular use cases. Maybe\n \n\nprograms for particular use cases. Maybe\nlike classification is a terrible\n\nlike classification is a terrible\n \n\nlike classification is a terrible\nexample. I recognize that. But for other\n\nexample. I recognize that. But for other\n \n\nexample. I recognize that. But for other\nother are other in theory, yes, you\n\nother are other in theory, yes, you\n \n\nother are other in theory, yes, you\nknow, you could do something like that. know, you could do something like that. know, you could do something like that. Yes. Yes. Yes. But for for particular LLM tasks, I'm\n\nBut for for particular LLM tasks, I'm\n \n\nBut for for particular LLM tasks, I'm\nsure we all have interesting ones. If\n\nsure we all have interesting ones. If\n \n\nsure we all have interesting ones. If\nyou have something that is relatively\n\nyou have something that is relatively\n \n\nyou have something that is relatively\nwell defined where you have known inputs\n\nwell defined where you have known inputs\n \n\nwell defined where you have known inputs\nand outputs, it might be a candidate for\n\nand outputs, it might be a candidate for\n \n\nand outputs, it might be a candidate for\nsomething worth optimizing. If nothing\n\nsomething worth optimizing. If nothing\n \n\nsomething worth optimizing. If nothing\nelse, to transfer it to a smaller model\n\nelse, to transfer it to a smaller model\n \n\nelse, to transfer it to a smaller model\nto preserve the level of performance at\n\nto preserve the level of performance at\n \n\nto preserve the level of performance at\na lower cost. That's really one of the\n\na lower cost. That's really one of the\n \n\na lower cost. That's really one of the\nbiggest benefits I see. All right, last last question. All right, last last question. All right, last last question. I've heard that uh DSPI is can be kind\n\nI've heard that uh DSPI is can be kind\n \n\nI've heard that uh DSPI is can be kind\nof expensive because you're doing all\n\nof expensive because you're doing all\n \n\nof expensive because you're doing all\nthese LM calls. these LM calls. these LM calls. >> Um so I was curious your experience with\n\n>> Um so I was curious your experience with\n \n\n>> Um so I was curious your experience with\nthat and maybe relatedly like if you\n\nthat and maybe relatedly like if you\n \n\nthat and maybe relatedly like if you\nhave any experience with like large\n\nhave any experience with like large\n \n\nhave any experience with like large\ncontext in your optimization data set\n\ncontext in your optimization data set\n \n\ncontext in your optimization data set\nways of shrinking those. ways of shrinking those. ways of shrinking those. >> Yeah. So the question was do can BSI be\n\n>> Yeah. So the question was do can BSI be\n \n\n>> Yeah. So the question was do can BSI be\nexpensive and then for large context\n\nexpensive and then for large context\n \n\nexpensive and then for large context\nkind of how have you seen that? How have\n\nkind of how have you seen that? How have\n \n\nkind of how have you seen that? How have\nyou managed that? The expensive part is\n\nyou managed that? The expensive part is\n \n\nyou managed that? The expensive part is\ntotally up to you. If you call a\n\ntotally up to you. If you call a\n \n\ntotally up to you. If you call a\nfunction a million times asynchronously,\n\nfunction a million times asynchronously,\n \n\nfunction a million times asynchronously,\nyou're going to generate a lot of cost. you're going to generate a lot of cost. you're going to generate a lot of cost. I don't think DSPI necessarily maybe it\n\nI don't think DSPI necessarily maybe it\n \n\nI don't think DSPI necessarily maybe it\nmakes it easier to call things, but it's\n\nmakes it easier to call things, but it's\n \n\nmakes it easier to call things, but it's\nnot inherently expensive. It might, to\n\nnot inherently expensive. It might, to\n \n\nnot inherently expensive. It might, to\nyour point,\n\nyour point,\n \n\nyour point,\nadd more content to the prompt. Like,\n\nadd more content to the prompt. Like,\n \n\nadd more content to the prompt. Like,\nsure, the signature is a string, but the\n\nsure, the signature is a string, but the\n \n\nsure, the signature is a string, but the\nactual text that's sent to the model is\n\nactual text that's sent to the model is\n \n\nactual text that's sent to the model is\nmuch longer than that. That's totally\n\nmuch longer than that. That's totally\n \n\nmuch longer than that. That's totally\ntrue. I wouldn't say that it's a large\n\ntrue. I wouldn't say that it's a large\n \n\ntrue. I wouldn't say that it's a large\ncost driver. I mean it again it's\n\ncost driver. I mean it again it's\n \n\ncost driver. I mean it again it's\nultimately it's more more of a\n\nultimately it's more more of a\n \n\nultimately it's more more of a\nprogramming paradigm. So you can write\n\nprogramming paradigm. So you can write\n \n\nprogramming paradigm. So you can write\nyour compressed adapter if you want that\n\nyour compressed adapter if you want that\n \n\nyour compressed adapter if you want that\nlike you know reduces the amount that's\n\nlike you know reduces the amount that's\n \n\nlike you know reduces the amount that's\nsent to the to uh to the model. Um in\n\nsent to the to uh to the model. Um in\n \n\nsent to the to uh to the model." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nUm in\nterms of large context I it's kind of\n\nterms of large context I it's kind of\n \n\nterms of large context I it's kind of\nthe same answer I think in terms of if\n\nthe same answer I think in terms of if\n \n\nthe same answer I think in terms of if\nyou're worried about that maybe you have\n\nyou're worrie..." }, { "role": "assistant", "content": "Um in\nterms of large context I it's kind of\n\nterms of large context I it's kind of\n \n\nterms of large context I it's kind of\nthe same answer I think in terms of if\n\nthe same answer I think in terms of if\n \n\nthe same answer I think in terms of if\nyou're worried about that maybe you have\n\nyou're worried about that maybe you have\n \n\nyou're worried about that maybe you have\nsome additional logic either in the\n\nsome additional logic either in the\n \n\nsome additional logic either in the\nprogram itself or in an adapter or part\n\nprogram itself or in an adapter or part\n \n\nprogram itself or in an adapter or part\nof the module that keeps track of that. of the module that keeps track of that. of the module that keeps track of that. Maybe you do some like context\n\nMaybe you do some like context\n \n\nMaybe you do some like context\ncompression or something like that. compression or something like that. compression or something like that. There's some really good talks about\n\nThere's some really good talks about\n \n\nThere's some really good talks about\nthat past few days. Obviously, I have a\n\nthat past few days. Obviously, I have a\n \n\nthat past few days. Obviously, I have a\nfeeling that that will kind of go away\n\nfeeling that that will kind of go away\n \n\nfeeling that that will kind of go away\nat some point where either context\n\nat some point where either context\n \n\nat some point where either context\nwindows get bigger or context management\n\nwindows get bigger or context management\n \n\nwindows get bigger or context management\nis abstracted away somehow. I don't\n\nis abstracted away somehow. I don't\n \n\nis abstracted away somehow. I don't\nreally have an answer just that's more\n\nreally have an answer just that's more\n \n\nreally have an answer just that's more\nof an intuition. Um, but DSP again kind\n\nof an intuition. Um, but DSP again kind\n \n\nof an intuition. Um, but DSP again kind\nof gives you the tools, the primitives\n\nof gives you the tools, the primitives\n \n\nof gives you the tools, the primitives\nfor you to do that should you choose. for you to do that should you choose. for you to do that should you choose. Um, and kind of track that state, track\n\nUm, and kind of track that state, track\n \n\nUm, and kind of track that state, track\nthat management over time. that management over time. that management over time. So, I think that's it. We're going to\n\nSo, I think that's it. We're going to\n \n\nSo, I think that's it. We're going to\nget kicked out soon. So, thanks so much\n\nget kicked out soon. So, thanks so much\n \n\nget kicked out soon. So, thanks so much\nfor your time. Really appreciate it. for your time. Really appreciate it. for your time. Really appreciate it. [music]\n\n \n\n[music]" } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\n[Music]\n\n \n\nhi everyone uh my name is Kyle Corbett\n\nhi everyone uh my name is Kyle Corbett\n \n\nhi everyone uh my name is Kyle Corbett\num thank you for the introduction uh I\n\num thank you for the introduction uh I\n \n\num thank you for the introduction uh I\nwork for a company called open pipe\n\nwork for ..." }, { "role": "assistant", "content": "[Music]\n\n \n\nhi everyone uh my name is Kyle Corbett\n\nhi everyone uh my name is Kyle Corbett\n \n\nhi everyone uh my name is Kyle Corbett\num thank you for the introduction uh I\n\num thank you for the introduction uh I\n \n\num thank you for the introduction uh I\nwork for a company called open pipe\n\nwork for a company called open pipe\n \n\nwork for a company called open pipe\nwhich is a fine-tuning platform um the\n\nwhich is a fine-tuning platform um the\n \n\nwhich is a fine-tuning platform um the\ntitle of my talk is that you probably\n\ntitle of my talk is that you probably\n \n\ntitle of my talk is that you probably\ndon't need fine\n\ndon't need fine\n \n\ndon't need fine\ntuning this is an awkward title\n\ntuning this is an awkward title\n \n\ntuning this is an awkward title\num given that I am doing a fine tuning\n\num given that I am doing a fine tuning\n \n\num given that I am doing a fine tuning\nplatform there's a little bit of a catch\n\nplatform there's a little bit of a catch\n \n\nplatform there's a little bit of a catch\nhere uh this is a little bit better put\n\nhere uh this is a little bit better put\n \n\nhere uh this is a little bit better put\na question mark there maybe you don't\n\na question mark there maybe you don't\n \n\na question mark there maybe you don't\nneed fine tuning um all right I'm going\n\nneed fine tuning um all right I'm going\n \n\nneed fine tuning um all right I'm going\nto take one more thing do you need fine\n\nto take one more thing do you need fine\n \n\nto take one more thing do you need fine\ntuning all right let's start from this\n\ntuning all right let's start from this\n \n\ntuning all right let's start from this\nbasic we're not assuming anything and\n\nbasic we're not assuming anything and\n \n\nbasic we're not assuming anything and\nwe're going to be talking about that in\n\nwe're going to be talking about that in\n \n\nwe're going to be talking about that in\nthe presentation so um as we're going\n\nthe presentation so um as we're going\n \n\nthe presentation so um as we're going\nthrough this presentation now going back\n\nthrough this presentation now going back\n \n\nthrough this presentation now going back\na couple slides um I actually do believe\n\na couple slides um I actually do believe\n \n\na couple slides um I actually do believe\ndespite the fact that I am in fact\n\ndespite the fact that I am in fact\n \n\ndespite the fact that I am in fact\nworking on a fine tuning platform that\n\nworking on a fine tuning platform that\n \n\nworking on a fine tuning platform that\nmy guess is that most people in this\n\nmy guess is that most people in this\n \n\nmy guess is that most people in this\naudience as of right this moment\n\naudience as of right this moment\n \n\naudience as of right this moment\nprobably do not need fine tuning and\n\nprobably do not need fine tuning and\n \n\nprobably do not need fine tuning and\nwe're going to be talking about in this\n\nwe're going to be talking about in this\n \n\nwe're going to be talking about in this\npresentation kind of what the maturity\n\npresentation kind of what the maturity\n \n\npresentation kind of what the maturity\ncurve looks like and where the right\n\ncurve looks like and where the right\n \n\ncurve looks like and where the right\nmoment is for you to bring that into\n\nmoment is for you to bring that into\n \n\nmoment is for you to bring that into\nyour infrastructure and whether it even\n\nyour infrastructure and whether it even\n \n\nyour infrastructure and whether it even\nmakes sense to do that um in your\n\nmakes sense to do that um in your\n \n\nmakes sense to do that um in your\nparticular case um but my hope is that\n\nparticular case um but my hope is that\n \n\nparticular case um but my hope is that\nby the end of this presentation um a few\n\nby the end of this presentation um a few\n \n\nby the end of this presentation um a few\nthings have happened one is um you have\n\nthings have happened one is um you have\n \n\nthings have happened one is um you have\na better qu answer to that question for\n\na better qu answer to that question for\n \n\na better qu answer to that question for\nyourself um does this make sense for me\n\nyourself um does this make sense for me\n \n\nyourself um does this make sense for me\nis this going to apply and then just as\n\nis this going to apply and then just as\n \n\nis this going to apply and then just as\nimportantly um if it does make sense\n\nimportantly um if it does make sense\n \n\nimportantly um if it does make sense\nwhat is the right moment um How will I\n\nwhat is the right moment um How will I\n \n\nwhat is the right moment um How will I\nknow that this is something that like I\n\nknow that this is something that like I\n \n\nknow that this is something that like I\nI should be doing that's actually going\n\nI should be doing that's actually going\n \n\nI should be doing that's actually going\nto help my company um you know whatever\n\nto help my company um you know whatever\n \n\nto help my company um you know whatever\nI'm trying to do uh get get the outcome\n\nI'm trying to do uh get get the outcome\n \n\nI'm trying to do uh get get the outcome\nI'm looking\n\nI'm looking\n \n\nI'm looking\nfor okay um splitting this up into three\n\nfor okay um splitting this up into three\n \n\nfor okay um splitting this up into three\npieces um one is why okay so kind of\n\npieces um one is why okay so kind of\n \n\npieces um one is why okay so kind of\naddressing the the title of the talk um\n\naddressing the the title of the talk um\n \n\naddressing the the title of the talk um\ndo you need fine tuning um we're going\n\ndo you need fine tuning um we're going\n \n\ndo you need fine tuning um we're going\nto talk about what are the benefits like\n\nto talk about what are the benefits like\n \n\nto talk about what are the benefits like\nwhy why is this even something you would\n\nwhy why is this even something you would\n \n\nwhy why is this even something you would\nput time into what potentially could you\n\nput time into what potentially could you\n \n\nput time into what potentially could you\ngain from it um and then the second one\n\ngain from it um and then the second one\n \n\ngain from it um and then the second one\nis when uh again at what moment in the\n\nis when uh again at what moment in the\n \n\nis when uh again at what moment in the\nprocess does it make sense uh to to\n\nprocess does it make sense uh to to\n \n\nprocess does it make sense uh to to\nstart doing this if it does and then\n\nstart doing this if it does and then\n \n\nstart doing this if it does and then\nfinally how um like what are the actual\n\nfinally how um like what are the actual\n \n\nfinally how um like what are the actual\nsteps you take kind of like what is the\n\nsteps you take kind of like what is the\n \n\nsteps you take kind of like what is the\nwhat is the general process and even\n\nwhat is the general process and even\n \n\nwhat is the general process and even\nsome specifics on how to do that um\n\nsome specifics on how to do that um\n \n\nsome specifics on how to do that um\nhopefully in that last section we find\n\nhopefully in that last section we find\n \n\nhopefully in that last section we find\nthat it's actually maybe easier than you\n\nthat it's actually maybe easier than you\n \n\nthat it's actually maybe easier than you\nthought uh that's usually the case when\n\nthought uh that's usually the case when\n \n\nthought uh that's usually the case when\nwhen I talk to people uh who have not\n\nwhen I talk to people uh who have not\n \n\nwhen I talk to people uh who have not\ndone it before okay first we're going to\n\ndone it before okay first we're going to\n \n\ndone it before okay first we're going to\ntalk about why um so anyone this this is\n\ntalk about why um so anyone this this is\n \n\ntalk about why um so anyone this this is\nthis is a graph we're looking at uh you\n\nthis is a graph we're looking at uh you\n \n\nthis is a graph we're looking at uh you\nknow the the quality of output across\n\nknow the the quality of output across\n \n\nknow the the quality of output across\nseveral tasks we averaged um to generate\n\nseveral tasks we averaged um to generate\n \n\nseveral tasks we averaged um to generate\nthis and then on the on the xaxis we're\n\nthis and then on the on the xaxis we're\n \n\nthis and then on the on the xaxis we're\nlooking at kind of tokens per dollar\n\nlooking at kind of tokens per dollar\n \n\nlooking at kind of tokens per dollar\nbasically the cost effectivess of this\n\nbasically the cost effectivess of this\n \n\nbasically the cost effectivess of this\nmodel all right um Can anyone if you've\n\nmodel all right um Can anyone if you've\n \n\nmodel all right um Can anyone if you've\nseen a graph kind of like this before\n\nseen a graph kind of like this before\n \n\nseen a graph kind of like this before\nraise your hand this is this this is a\n\nraise your hand this is this this is a\n \n\nraise your hand this is this this is a\npretty common um kind of way to to look\n\npretty common um kind of way to to look\n \n\npretty common um kind of way to to look\nat this and and I suspect many of you\n\nat this and and I suspect many of you\n \n\nat this and and I suspect many of you\nwho are um practitioners and are using\n\nwho are um practitioners and are using\n \n\nwho are um practitioners and are using\nllms you've you've um maybe even\n\nllms you've you've um maybe even\n \n\nllms you've you've um maybe even\ngenerated a graph like this on your own\n\ngenerated a graph like this on your own\n \n\ngenerated a graph like this on your own\nspecific use case maybe You' evaluated a\n\nspecific use case maybe You' evaluated a\n \n\nspecific use case maybe You' evaluated a\nfew different base models you've looked\n\nfew different base models you've looked\n \n\nfew different base models you've looked\nat um either on your own evaluations or\n\nat um either on your own evaluations or\n \n\nat um either on your own evaluations or\nsome kind of vibe check or something to\n\nsome kind of vibe check or something to\n \n\nsome kind of vibe check or something to\nfigure out okay um for this for for\n\nfigure out okay um for this for for\n \n\nfigure out okay um for this for for\nthese different models uh how well does\n\nthese different models uh how well does\n \n\nthese different models uh how well does\nit do all my task and um and you compare\n\nit do all my task and um and you compare\n \n\nit do all my task and um and you compare\nthem um and usually uh there's a\n\nthem um and usually uh there's a\n \n\nthem um and usually uh there's a\ncorrelation at least at at sort of like\n\ncorrelation at least at at sort of like\n \n\ncorrelation at least at at sort of like\nthe frontier level between how expensive\n\nthe frontier level between how expensive\n \n\nthe frontier level between how expensive\na model is and how good it is um\n\na model is and how good it is um\n \n\na model is and how good it is um\ntypically the larger a model is and\n\ntypically the larger a model is and\n \n\ntypically the larger a model is and\nnumber of parameters and the more time\n\nnumber of parameters and the more time\n \n\nnumber of parameters and the more time\nand effort that has gone into training\n\nand effort that has gone into training\n \n\nand effort that has gone into training\nit um the higher quality outputs it's\n\nit um the higher quality outputs it's\n \n\nit um the higher quality outputs it's\nable to produce this is this is all um\n\nable to produce this is this is all um\n \n\nable to produce this is this is all um\npretty basic stuff um and then of course\n\npretty basic stuff um and then of course\n \n\npretty basic stuff um and then of course\nas you get into smaller models uh or or\n\nas you get into smaller models uh or or\n \n\nas you get into smaller models uh or or\nmodels that haven't had as much training\n\nmodels that haven't had as much training\n \n\nmodels that haven't had as much training\ntime put into them then then maybe the\n\ntime put into them then then maybe the\n \n\ntime put into them then then maybe the\nresponse quality on sort of a zero shot\n\nresponse quality on sort of a zero shot\n \n\nresponse quality on sort of a zero shot\nbasis is is not quite as good um okay\n\nbasis is is not quite as good um okay\n \n\nbasis is is not quite as good um okay\nall all pretty basic stuff um one thing\n\nall all pretty basic stuff um one thing\n \n\nall all pretty basic stuff um one thing\nyou'll notice is uh is that you can sort\n\nyou'll notice is uh is that you can sort\n \n\nyou'll notice is uh is that you can sort\nof think of this as like a predo\n\nof think of this as like a predo\n \n\nof think of this as like a predo\nfrontier right where it's like you if\n\nfrontier right where it's like you if\n \n\nfrontier right where it's like you if\nyou figure out how good quality you need\n\nyou figure out how good quality you need\n \n\nyou figure out how good quality you need\nthen you can from that number sort of\n\nthen you can from that number sort of\n \n\nthen you can from that number sort of\nlike backfill all right which uh which\n\nlike backfill all right which uh which\n \n\nlike backfill all right which uh which\nmodel hits that bar that's at at the\n\nmodel hits that bar that's at at the\n \n\nmodel hits that bar that's at at the\nlowest price point or alternatively if\n\nlowest price point or alternatively if\n \n\nlowest price point or alternatively if\num maybe you've got some some constraint\n\num maybe you've got some some constraint\n \n\num maybe you've got some some constraint\non your unit economics or like you know\n\non your unit economics or like you know\n \n\non your unit economics or like you know\nhow much you want to spend on this\n\nhow much you want to spend on this\n \n\nhow much you want to spend on this\nproject and so then you you work from\n\nproject and so then you you work from\n \n\nproject and so then you you work from\nthe other direction say Okay I I can't\n\nthe other direction say Okay I I can't\n \n\nthe other direction say Okay I I can't\nspend more than this what's the best\n\nspend more than this what's the best\n \n\nspend more than this what's the best\nmodel I can get and so this this\n\nmodel I can get and so this this\n \n\nmodel I can get and so this this\nFrontier kind of maps all right given\n\nFrontier kind of maps all right given\n \n\nFrontier kind of maps all right given\ngiven a certain constraint um what\n\ngiven a certain constraint um what\n \n\ngiven a certain constraint um what\nshould I\n\nshould I\n \n\nshould I\nchoose um there's uh this this next\n\nchoose um there's uh this this next\n \n\nchoose um there's uh this this next\nslide is a little bit of a slide but\n\nslide is a little bit of a slide but\n \n\nslide is a little bit of a slide but\nit's actually very important for for\n\nit's actually very important for for\n \n\nit's actually very important for for\nwhat we're getting to next um one way\n\nwhat we're getting to next um one way\n \n\nwhat we're getting to next um one way\nyou can push this perto Frontier\n\nyou can push this perto Frontier\n \n\nyou can push this perto Frontier\nespecially kind of the low end in the\n\nespecially kind of the low end in the\n \n\nespecially kind of the low end in the\nsense of like the very expensive but\n\nsense of like the very expensive but\n \n\nsense of like the very expensive but\nhigh quality end is um by basically\n\nhigh quality end is um by basically\n \n\nhigh quality end is um by basically\nusing more compute at inference time so\n\nusing more compute at inference time so\n \n\nusing more compute at inference time so\nso this is um I put a an example on the\n\nso this is um I put a an example on the\n \n\nso this is um I put a an example on the\ngraph this is actually just a very like\n\ngraph this is actually just a very like\n \n\ngraph this is actually just a very like\nparticular way of doing that it's it's\n\nparticular way of doing that it's it's\n \n\nparticular way of doing that it's it's\ncalled a mixture of Agents approach um\n\ncalled a mixture of Agents approach um\n \n\ncalled a mixture of Agents approach um\ntogether AI published some papers and\n\ntogether AI published some papers and\n \n\ntogether AI published some papers and\nwe've also done some research in that\n\nwe've also done some research in that\n \n\nwe've also done some research in that\narea um but this is you can think of\n\narea um but this is you can think of\n \n\narea um but this is you can think of\nthis as like a very specific proom\n\nthis as like a very specific proom\n \n\nthis as like a very specific proom\nprompting technique that uses extra time\n\nprompting technique that uses extra time\n \n\nprompting technique that uses extra time\nat inference you're you're going through\n\nat inference you're you're going through\n \n\nat inference you're you're going through\nseveral layers and kind of like\n\nseveral layers and kind of like\n \n\nseveral layers and kind of like\nimproving your answer and then\n\nimproving your answer and then\n \n\nimproving your answer and then\neventually you come out with something\n\neventually you come out with something\n \n\neventually you come out with something\nthat's higher quality than in this case\n\nthat's higher quality than in this case\n \n\nthat's higher quality than in this case\nyou would have gotten with gbd4 by\n\nyou would have gotten with gbd4 by\n \n\nyou would have gotten with gbd4 by\nitself okay so why is this important um\n\nitself okay so why is this important um\n \n\nitself okay so why is this important um\nbecause it actually does prove it's\n\nbecause it actually does prove it's\n \n\nbecause it actually does prove it's\npossible to to move outside this\n\npossible to to move outside this\n \n\npossible to to move outside this\nFrontier um if we're a little bit smart\n\nFrontier um if we're a little bit smart\n \n\nFrontier um if we're a little bit smart\nof how we do it in this case by using\n\nof how we do it in this case by using\n \n\nof how we do it in this case by using\nmore prompting now this particular\n\nmore prompting now this particular\n \n\nmore prompting now this particular\napproach um there's some downsides as\n\napproach um there's some downsides as\n \n\napproach um there's some downsides as\nyou can see it's all the way over on the\n\nyou can see it's all the way over on the\n \n\nyou can see it's all the way over on the\nuh left side of the graph which means\n\nuh left side of the graph which means\n \n\nuh left side of the graph which means\nit's extremely expensive because you're\n\nit's extremely expensive because you're\n \n\nit's extremely expensive because you're\ndoing all these calls um but you get\n\ndoing all these calls um but you get\n \n\ndoing all these calls um but you get\nhigher quality okay that's cool but if\n\nhigher quality okay that's cool but if\n \n\nhigher quality okay that's cool but if\num a lot of people can't use that\n\num a lot of people can't use that\n \n\num a lot of people can't use that\ndirectly again because of those\n\ndirectly again because of those\n \n\ndirectly again because of those\nconstraints where it's it's just really\n\nconstraints where it's it's just really\n \n\nconstraints where it's it's just really\nexpensive latency is really high you're\n\nexpensive latency is really high you're\n \n\nexpensive latency is really high you're\ndoing multiple prompts um and so that by\n\ndoing multiple prompts um and so that by\n \n\ndoing multiple prompts um and so that by\nitself maybe isn't going to going going\n\nitself maybe isn't going to going going\n \n\nitself maybe isn't going to going going\nto be that interesting what is\n\nto be that interesting what is\n \n\nto be that interesting what is\ninteresting is if you then take say you\n\ninteresting is if you then take say you\n \n\ninteresting is if you then take say you\nfound way like using this mixture of\n\nfound way like using this mixture of\n \n\nfound way like using this mixture of\nexperts to get or sorry mixture of\n\nexperts to get or sorry mixture of\n \n\nexperts to get or sorry mixture of\nagents to get really high quality data\n\nagents to get really high quality data\n \n\nagents to get really high quality data\num that is the moment uh where fine\n\num that is the moment uh where fine\n \n\num that is the moment uh where fine\ntuning can really help you and I'm going\n\ntuning can really help you and I'm going\n \n\ntuning can really help you and I'm going\nto put up some more dots on this graph\n\nto put up some more dots on this graph\n \n\nto put up some more dots on this graph\num okay so let's take a look at what I\n\num okay so let's take a look at what I\n \n\num okay so let's take a look at what I\njust put through up here so this and by\n\njust put through up here so this and by\n \n\njust put through up here so this and by\nthe way this graph this these are not um\n\nthe way this graph this these are not um\n \n\nthe way this graph this these are not um\nthese are not conceptual like this is\n\nthese are not conceptual like this is\n \n\nthese are not conceptual like this is\nactually uh you know like this is uh\n\nactually uh you know like this is uh\n \n\nactually uh you know like this is uh\nevaluations that we ran internally on\n\nevaluations that we ran internally on\n \n\nevaluations that we ran internally on\nreal customer data sets um using the\n\nreal customer data sets um using the\n \n\nreal customer data sets um using the\nevaluations our customers had to find\n\nevaluations our customers had to find\n \n\nevaluations our customers had to find\nand uh and we used four different data\n\nand uh and we used four different data\n \n\nand uh and we used four different data\nsets we averaged it the quality is um an\n\nsets we averaged it the quality is um an\n \n\nsets we averaged it the quality is um an\nLMS judge uh and then the tokens per\n\nLMS judge uh and then the tokens per\n \n\nLMS judge uh and then the tokens per\ndollar is basically just like industry\n\ndollar is basically just like industry\n \n\ndollar is basically just like industry\nstandard uh pricing for those models so\n\nstandard uh pricing for those models so\n \n\nstandard uh pricing for those models so\nso this is based in fact um this is not\n\nso this is based in fact um this is not\n \n\nso this is based in fact um this is not\na theoretical graph and what I've put up\n\na theoretical graph and what I've put up\n \n\na theoretical graph and what I've put up\nhere are three additional models um\n\nhere are three additional models um\n \n\nhere are three additional models um\nwhich are fine-tuned versions and two of\n\nwhich are fine-tuned versions and two of\n \n\nwhich are fine-tuned versions and two of\nthem are actually fine tuned versions\n\nthem are actually fine tuned versions\n \n\nthem are actually fine tuned versions\nthat are also on the graph so if we look\n\nthat are also on the graph so if we look\n \n\nthat are also on the graph so if we look\nyeah on my next slide you can actually\n\nyeah on my next slide you can actually\n \n\nyeah on my next slide you can actually\nsee through the fine-tuning process for\n\nsee through the fine-tuning process for\n \n\nsee through the fine-tuning process for\nfor llama 370b and then llama 3 AP we\n\nfor llama 370b and then llama 3 AP we\n \n\nfor llama 370b and then llama 3 AP we\nwere able to move the performance on\n\nwere able to move the performance on\n \n\nwere able to move the performance on\nlike that average across these four\n\nlike that average across these four\n \n\nlike that average across these four\ntasks that we tested on up substantially\n\ntasks that we tested on up substantially\n \n\ntasks that we tested on up substantially\nand not only are they much higher than\n\nand not only are they much higher than\n \n\nand not only are they much higher than\nthey were from just prompting they're\n\nthey were from just prompting they're\n \n\nthey were from just prompting they're\nactually in both those cases higher than\n\nactually in both those cases higher than\n \n\nactually in both those cases higher than\nany of the Bas prompted models and\n\nany of the Bas prompted models and\n \n\nany of the Bas prompted models and\nthat's really important because um what\n\nthat's really important because um what\n \n\nthat's really important because um what\nwe're doing here is we are actually\n\nwe're doing here is we are actually\n \n\nwe're doing here is we are actually\nmoving the PTO Frontier outwards right\n\nmoving the PTO Frontier outwards right\n \n\nmoving the PTO Frontier outwards right\nso it is a different trade-off if you're\n\nso it is a different trade-off if you're\n \n\nso it is a different trade-off if you're\nwilling to do fine tune if you're\n\nwilling to do fine tune if you're\n \n\nwilling to do fine tune if you're\nwilling to put in the effort to create a\n\nwilling to put in the effort to create a\n \n\nwilling to put in the effort to create a\nfine tune model you can actually move\n\nfine tune model you can actually move\n \n\nfine tune model you can actually move\nthat pedo Frontier outward and get much\n\nthat pedo Frontier outward and get much\n \n\nthat pedo Frontier outward and get much\nmore bang for your buck so whatever the\n\nmore bang for your buck so whatever the\n \n\nmore bang for your buck so whatever the\namount of money you're willing to spend\n\namount of money you're willing to spend\n \n\namount of money you're willing to spend\non inferences um you can you can get\n\non inferences um you can you can get\n \n\non inferences um you can you can get\njust much higher quality completions um\n\njust much higher quality completions um\n \n\njust much higher quality completions um\nthan you would have without this and\n\nthan you would have without this and\n \n\nthan you would have without this and\nthat basically in a nutshell is the\n\nthat basically in a nutshell is the\n \n\nthat basically in a nutshell is the\npitch for fine tuning is that it gives\n\npitch for fine tuning is that it gives\n \n\npitch for fine tuning is that it gives\nyou this ability to move outside that\n\nyou this ability to move outside that\n \n\nyou this ability to move outside that\nPera Frontier and for whatever budget\n\nPera Frontier and for whatever budget\n \n\nPera Frontier and for whatever budget\nyou have um you end up uh with with a\n\nyou have um you end up uh with with a\n \n\nyou have um you end up uh with with a\nbetter\n\nbetter\n \n\nbetter\nmodel um so anyway I've redrawn here we\n\nmodel um so anyway I've redrawn here we\n \n\nmodel um so anyway I've redrawn here we\nhave we have a new parto Frontier um you\n\nhave we have a new parto Frontier um you\n \n\nhave we have a new parto Frontier um you\nknow which which kind of takes this into\n\nknow which which kind of takes this into\n \n\nknow which which kind of takes this into\naccount um and so now all of a sudden\n\naccount um and so now all of a sudden\n \n\naccount um and so now all of a sudden\nthe best model you're going to get at\n\nthe best model you're going to get at\n \n\nthe best model you're going to get at\nany price point is going to be one that\n\nany price point is going to be one that\n \n\nany price point is going to be one that\nyou tuned if you're willing to go to\n\nyou tuned if you're willing to go to\n \n\nyou tuned if you're willing to go to\nthat work so um you know actually I'm\n\nthat work so um you know actually I'm\n \n\nthat work so um you know actually I'm\ngoing to go I I really like this this\n\ngoing to go I I really like this this\n \n\ngoing to go I I really like this this\ntransition just go back and forth you\n\ntransition just go back and forth you\n \n\ntransition just go back and forth you\nknow it's like better right cool\n\nknow it's like better right cool\n \n\nknow it's like better right cool\neveryone got got it okay so so this is\n\neveryone got got it okay so so this is\n \n\neveryone got got it okay so so this is\nawesome hopefully I've convinced you\n\nawesome hopefully I've convinced you\n \n\nawesome hopefully I've convinced you\nlike you know moving that outward is\n\nlike you know moving that outward is\n \n\nlike you know moving that outward is\nbetter whatever your price point is um\n\nbetter whatever your price point is um\n \n\nbetter whatever your price point is um\nyou're going to get a better completion\n\nyou're going to get a better completion\n \n\nyou're going to get a better completion\num okay so this is awesome right like\n\num okay so this is awesome right like\n \n\num okay so this is awesome right like\nprto Frontier like by definition means\n\nprto Frontier like by definition means\n \n\nprto Frontier like by definition means\nthat you we've expanded it no matter\n\nthat you we've expanded it no matter\n \n\nthat you we've expanded it no matter\nwhere you are on that um on that point\n\nwhere you are on that um on that point\n \n\nwhere you are on that um on that point\nyou're getting a better completion there\n\nyou're getting a better completion there\n \n\nyou're getting a better completion there\nare no trade-offs right woo that's right\n\nare no trade-offs right woo that's right\n \n\nare no trade-offs right woo that's right\nthank\n\nthank\n \n\nthank\nyou okay I lied there are trade-offs\n\nyou okay I lied there are trade-offs\n \n\nyou okay I lied there are trade-offs\nthere's always trade-offs um the\n\nthere's always trade-offs um the\n \n\nthere's always trade-offs um the\ntrade-off that you typically run into\n\ntrade-off that you typically run into\n \n\ntrade-off that you typically run into\nwith fine tuning is there there's a\n\nwith fine tuning is there there's a\n \n\nwith fine tuning is there there's a\ncouple of different things um the\n\ncouple of different things um the\n \n\ncouple of different things um the\nbiggest one honestly is just the time up\n\nbiggest one honestly is just the time up\n \n\nbiggest one honestly is just the time up\nfront it takes right like this is an\n\nfront it takes right like this is an\n \n\nfront it takes right like this is an\nextra thing um you've got to actually do\n\nextra thing um you've got to actually do\n \n\nextra thing um you've got to actually do\nit you've got to gather your data set\n\nit you've got to gather your data set\n \n\nit you've got to gather your data set\nyou've got to make sure that the model's\n\nyou've got to make sure that the model's\n \n\nyou've got to make sure that the model's\nworking um and that is a real cost that\n\nworking um and that is a real cost that\n \n\nworking um and that is a real cost that\ntakes some time um and then the second\n\ntakes some time um and then the second\n \n\ntakes some time um and then the second\none is is you get you lose a little bit\n\none is is you get you lose a little bit\n \n\none is is you get you lose a little bit\nof flexibility um when you tun a model\n\nof flexibility um when you tun a model\n \n\nof flexibility um when you tun a model\nwhat you're doing is you are making it\n\nwhat you're doing is you are making it\n \n\nwhat you're doing is you are making it\nreally really good at one specific thing\n\nreally really good at one specific thing\n \n\nreally really good at one specific thing\num but it's not necessarily really\n\num but it's not necessarily really\n \n\num but it's not necessarily really\nreally good at everything else in fact\n\nreally good at everything else in fact\n \n\nreally good at everything else in fact\nby definition it's uh it's not good at\n\nby definition it's uh it's not good at\n \n\nby definition it's uh it's not good at\neverything um that's that's sort of like\n\neverything um that's that's sort of like\n \n\neverything um that's that's sort of like\nexactly what we're trying to achieve by\n\nexactly what we're trying to achieve by\n \n\nexactly what we're trying to achieve by\nfine tuning and so um because you're\n\nfine tuning and so um because you're\n \n\nfine tuning and so um because you're\nlosing some of that flexibility it's\n\nlosing some of that flexibility it's\n \n\nlosing some of that flexibility it's\nreally important that you understand\n\nreally important that you understand\n \n\nreally important that you understand\nbefore you go into that fine tuning\n\nbefore you go into that fine tuning\n \n\nbefore you go into that fine tuning\nprocess what your task actually looks\n\nprocess what your task actually looks\n \n\nprocess what your task actually looks\nlike uh because you want to be able to\n\nlike uh because you want to be able to\n \n\nlike uh because you want to be able to\nspecialize on that task and get really\n\nspecialize on that task and get really\n \n\nspecialize on that task and get really\ngreat performance on that task and you\n\ngreat performance on that task and you\n \n\ngreat performance on that task and you\nwant to make sure that when you're\n\nwant to make sure that when you're\n \n\nwant to make sure that when you're\nactually using this model at inference\n\nactually using this model at inference\n \n\nactually using this model at inference\ntime um you aren't sending it a bunch of\n\ntime um you aren't sending it a bunch of\n \n\ntime um you aren't sending it a bunch of\nstuff that's kind of out of domain that\n\nstuff that's kind of out of domain that\n \n\nstuff that's kind of out of domain that\nis not related to what it was trained on\n\nis not related to what it was trained on\n \n\nis not related to what it was trained on\nbecause you're not going to get that\n\nbecause you're not going to get that\n \n\nbecause you're not going to get that\nsame performance boost on on anything\n\nsame performance boost on on anything\n \n\nsame performance boost on on anything\nthat was out of domain so that's the\n\nthat was out of domain so that's the\n \n\nthat was out of domain so that's the\nprimary tradeoff um that's that's the\n\nprimary tradeoff um that's that's the\n \n\nprimary tradeoff um that's that's the\nreason why maybe it's not right to to\n\nreason why maybe it's not right to to\n \n\nreason why maybe it's not right to to\nfind tun immediately and so that brings\n\nfind tun immediately and so that brings\n \n\nfind tun immediately and so that brings\nus to our second of the three sections\n\nus to our second of the three sections\n \n\nus to our second of the three sections\nwe're going to talk about and the second\n\nwe're going to talk about and the second\n \n\nwe're going to talk about and the second\none is like okay when how do I make this\n\none is like okay when how do I make this\n \n\none is like okay when how do I make this\ndecision um how do I know when the right\n\ndecision um how do I know when the right\n \n\ndecision um how do I know when the right\nmoment is to invest the time to do this\n\nmoment is to invest the time to do this\n \n\nmoment is to invest the time to do this\num and there is a little bit of sort of\n\num and there is a little bit of sort of\n \n\num and there is a little bit of sort of\nlike uh you know uh specificity but um\n\nlike uh you know uh specificity but um\n \n\nlike uh you know uh specificity but um\nthere are generally three different\n\nthere are generally three different\n \n\nthere are generally three different\nreasons why you might want to move from\n\nreasons why you might want to move from\n \n\nreasons why you might want to move from\na prompted model to a fine-tune model um\n\na prompted model to a fine-tune model um\n \n\na prompted model to a fine-tune model um\nand the three reasons are first of all\n\nand the three reasons are first of all\n \n\nand the three reasons are first of all\ncost um if you just can't deal with the\n\ncost um if you just can't deal with the\n \n\ncost um if you just can't deal with the\nunit economics if you're using gbd form\n\nunit economics if you're using gbd form\n \n\nunit economics if you're using gbd form\nproduction that's going to be really\n\nproduction that's going to be really\n \n\nproduction that's going to be really\nexpensive um the second reason is\n\nexpensive um the second reason is\n \n\nexpensive um the second reason is\nlatency so as we're moving down towards\n\nlatency so as we're moving down towards\n \n\nlatency so as we're moving down towards\nthese smaller and cheaper models we're\n\nthese smaller and cheaper models we're\n \n\nthese smaller and cheaper models we're\nalso you know one advantage of being a\n\nalso you know one advantage of being a\n \n\nalso you know one advantage of being a\nsmaller model is there's just like less\n\nsmaller model is there's just like less\n \n\nsmaller model is there's just like less\nmultiplications less math going on for\n\nmultiplications less math going on for\n \n\nmultiplications less math going on for\nevery single generation and that means\n\nevery single generation and that means\n \n\nevery single generation and that means\nyou can actually bring your latency\n\nyou can actually bring your latency\n \n\nyou can actually bring your latency\nfloor much lower than you can with a\n\nfloor much lower than you can with a\n \n\nfloor much lower than you can with a\nlarge model like for example a\n\nlarge model like for example a\n \n\nlarge model like for example a\ngbd4 um so cost latency and then the\n\ngbd4 um so cost latency and then the\n \n\ngbd4 um so cost latency and then the\nfinal one is actually the uh the the\n\nfinal one is actually the uh the the\n \n\nfinal one is actually the uh the the\nquality um particularly the consistency\n\nquality um particularly the consistency\n \n\nquality um particularly the consistency\nof the quality so if you find the gbd4\n\nof the quality so if you find the gbd4\n \n\nof the quality so if you find the gbd4\nmaybe 80% 90% of the time is doing a\n\nmaybe 80% 90% of the time is doing a\n \n\nmaybe 80% 90% of the time is doing a\ngreat job and it is solving your task\n\ngreat job and it is solving your task\n \n\ngreat job and it is solving your task\nwell but that last 10 20% you know it\n\nwell but that last 10 20% you know it\n \n\nwell but that last 10 20% you know it\nmakes mistakes it doesn't get the format\n\nmakes mistakes it doesn't get the format\n \n\nmakes mistakes it doesn't get the format\nright it forgets one instruction um\n\nright it forgets one instruction um\n \n\nright it forgets one instruction um\ntypically those kind of issues where it\n\ntypically those kind of issues where it\n \n\ntypically those kind of issues where it\nusually gets it right but not\n\nusually gets it right but not\n \n\nusually gets it right but not\nconsistently um a fine tune model that's\n\nconsistently um a fine tune model that's\n \n\nconsistently um a fine tune model that's\nthe kind of thing that fine tune excels\n\nthe kind of thing that fine tune excels\n \n\nthe kind of thing that fine tune excels\nat and you're going to very consistently\n\nat and you're going to very consistently\n \n\nat and you're going to very consistently\nget the get the output you're looking\n\nget the get the output you're looking\n \n\nget the get the output you're looking\nfor um at a minimum much more\n\nfor um at a minimum much more\n \n\nfor um at a minimum much more\nconsistently than you would just purely\n\nconsistently than you would just purely\n \n\nconsistently than you would just purely\nthrough prompting so the way I like to\n\nthrough prompting so the way I like to\n \n\nthrough prompting so the way I like to\nthink about it is um for almost every\n\nthink about it is um for almost every\n \n\nthink about it is um for almost every\napplication not not universally but for\n\napplication not not universally but for\n \n\napplication not not universally but for\nmost applications I I see people using\n\nmost applications I I see people using\n \n\nmost applications I I see people using\num if you can get away with a prompted\n\num if you can get away with a prompted\n \n\num if you can get away with a prompted\nmodel to start with um you should do\n\nmodel to start with um you should do\n \n\nmodel to start with um you should do\nthat um you you should just eat the cost\n\nthat um you you should just eat the cost\n \n\nthat um you you should just eat the cost\nthis is like as you're first deploying\n\nthis is like as you're first deploying\n \n\nthis is like as you're first deploying\nit you're probably not spending that\n\nit you're probably not spending that\n \n\nit you're probably not spending that\nmuch money as you're just testing it um\n\nmuch money as you're just testing it um\n \n\nmuch money as you're just testing it um\nand you're going to benefit from the\n\nand you're going to benefit from the\n \n\nand you're going to benefit from the\niteration speed you get from being able\n\niteration speed you get from being able\n \n\niteration speed you get from being able\nto uh to just change the prompt and\n\nto uh to just change the prompt and\n \n\nto uh to just change the prompt and\nchange the behavior once you figured out\n\nchange the behavior once you figured out\n \n\nchange the behavior once you figured out\nokay this is the this is the product\n\nokay this is the this is the product\n \n\nokay this is the this is the product\nshape this is like kind of like you know\n\nshape this is like kind of like you know\n \n\nshape this is like kind of like you know\num what we're actually going to be\n\num what we're actually going to be\n \n\num what we're actually going to be\ndeploying at scale then it's worth\n\ndeploying at scale then it's worth\n \n\ndeploying at scale then it's worth\ntaking you know the the couple of hours\n\ntaking you know the the couple of hours\n \n\ntaking you know the the couple of hours\nyou're going to have to invest to go\n\nyou're going to have to invest to go\n \n\nyou're going to have to invest to go\nfrom that prompted model to a fine-tune\n\nfrom that prompted model to a fine-tune\n \n\nfrom that prompted model to a fine-tune\nmodel um and and at that point you're\n\nmodel um and and at that point you're\n \n\nmodel um and and at that point you're\ngoing to start getting those benefits\n\ngoing to start getting those benefits\n \n\ngoing to start getting those benefits\nfrom way cheaper way faster and higher\n\nfrom way cheaper way faster and higher\n \n\nfrom way cheaper way faster and higher\nquality\n\nquality\n \n\nquality\ncompletions um so that's that's\n\ncompletions um so that's that's\n \n\ncompletions um so that's that's\nbasically the the way I think about that\n\nbasically the the way I think about that\n \n\nbasically the the way I think about that\nokay so we have minutes left I'm to take\n\nokay so we have minutes left I'm to take\n \n\nokay so we have minutes left I'm to take\n6 minutes and teach you how to F tuna\n\n6 minutes and teach you how to F tuna\n \n\n6 minutes and teach you how to F tuna\nmodel okay and if you have not done this\n\nmodel okay and if you have not done this\n \n\nmodel okay and if you have not done this\nbefore it might feel intimidating I'm\n\nbefore it might feel intimidating I'm\n \n\nbefore it might feel intimidating I'm\nhoping by the end of this six-minute\n\nhoping by the end of this six-minute\n \n\nhoping by the end of this six-minute\nsection it is no longer intimidating\n\nsection it is no longer intimidating\n \n\nsection it is no longer intimidating\nbecause honestly we have had Engineers\n\nbecause honestly we have had Engineers\n \n\nbecause honestly we have had Engineers\nthat come in with no experience with\n\nthat come in with no experience with\n \n\nthat come in with no experience with\nmachine learning no experience with data\n\nmachine learning no experience with data\n \n\nmachine learning no experience with data\nscience and have a successful fine T\n\nscience and have a successful fine T\n \n\nscience and have a successful fine T\nmodel those outperforming their base\n\nmodel those outperforming their base\n \n\nmodel those outperforming their base\nmodel they're coming from um with less\n\nmodel they're coming from um with less\n \n\nmodel they're coming from um with less\nthan an hour of work um so hopefully\n\nthan an hour of work um so hopefully\n \n\nthan an hour of work um so hopefully\nthis is something that that uh we come\n\nthis is something that that uh we come\n \n\nthis is something that that uh we come\nout of this with with a good\n\nout of this with with a good\n \n\nout of this with with a good\nunderstanding at least um of what you\n\nunderstanding at least um of what you\n \n\nunderstanding at least um of what you\nneed okay so yeah if if you you can\n\nneed okay so yeah if if you you can\n \n\nneed okay so yeah if if you you can\nprompt you can f tune uh it's you know\n\nprompt you can f tune uh it's you know\n \n\nprompt you can f tune uh it's you know\nthere there's we can walk you through\n\nthere there's we can walk you through\n \n\nthere there's we can walk you through\nthis process all right four steps you\n\nthis process all right four steps you\n \n\nthis process all right four steps you\nhave to get the data you have to\n\nhave to get the data you have to\n \n\nhave to get the data you have to\nactually train the model you probably\n\nactually train the model you probably\n \n\nactually train the model you probably\nshould evaluate technically step three\n\nshould evaluate technically step three\n \n\nshould evaluate technically step three\nis optional but I think we all we all\n\nis optional but I think we all we all\n \n\nis optional but I think we all we all\nknow on this track that evals are\n\nknow on this track that evals are\n \n\nknow on this track that evals are\nimportant um and then you have to deploy\n\nimportant um and then you have to deploy\n \n\nimportant um and then you have to deploy\nthe the model okay so preparing data one\n\nthe the model okay so preparing data one\n \n\nthe the model okay so preparing data one\nof the reasons why it's extremely um\n\nof the reasons why it's extremely um\n \n\nof the reasons why it's extremely um\nuseful if you can get away with it to\n\nuseful if you can get away with it to\n \n\nuseful if you can get away with it to\nstart with a prompted model like C gbd4\n\nstart with a prompted model like C gbd4\n \n\nstart with a prompted model like C gbd4\nin production is because it Sol like a\n\nin production is because it Sol like a\n \n\nin production is because it Sol like a\nhuge class of problems that anyone who\n\nhuge class of problems that anyone who\n \n\nhuge class of problems that anyone who\nhas a background in traditional machine\n\nhas a background in traditional machine\n \n\nhas a background in traditional machine\nlearning is very familiar with which is\n\nlearning is very familiar with which is\n \n\nlearning is very familiar with which is\nit helps you figure out what your actual\n\nit helps you figure out what your actual\n \n\nit helps you figure out what your actual\ninput distribution is before you have to\n\ninput distribution is before you have to\n \n\ninput distribution is before you have to\nlike actually train the model um if\n\nlike actually train the model um if\n \n\nlike actually train the model um if\nyou're deploying GPD form production you\n\nyou're deploying GPD form production you\n \n\nyou're deploying GPD form production you\ncan actually see okay what are my users\n\ncan actually see okay what are my users\n \n\ncan actually see okay what are my users\nusing this for um like what are the\n\nusing this for um like what are the\n \n\nusing this for um like what are the\nareas where where like I it was\n\nareas where where like I it was\n \n\nareas where where like I it was\nunexpected but like this is something\n\nunexpected but like this is something\n \n\nunexpected but like this is something\npeople are doing and you can actually\n\npeople are doing and you can actually\n \n\npeople are doing and you can actually\ncapture that distribution really well\n\ncapture that distribution really well\n \n\ncapture that distribution really well\nyou just have to log the prompts um and\n\nyou just have to log the prompts um and\n \n\nyou just have to log the prompts um and\nthen you can use that to know exactly\n\nthen you can use that to know exactly\n \n\nthen you can use that to know exactly\nwhat your model has to be good at so\n\nwhat your model has to be good at so\n \n\nwhat your model has to be good at so\nthis is actually when people talk about\n\nthis is actually when people talk about\n \n\nthis is actually when people talk about\nhow hard training models like used to be\n\nhow hard training models like used to be\n \n\nhow hard training models like used to be\nin the Dark Ages um this was actually\n\nin the Dark Ages um this was actually\n \n\nin the Dark Ages um this was actually\nwhere most of the effort was was\n\nwhere most of the effort was was\n \n\nwhere most of the effort was was\nfiguring out what what does the data\n\nfiguring out what what does the data\n \n\nfiguring out what what does the data\nactually have to look like like what\n\nactually have to look like like what\n \n\nactually have to look like like what\ndoes my input look like what do the good\n\ndoes my input look like what do the good\n \n\ndoes my input look like what do the good\noutputs look like um and if you can get\n\noutputs look like um and if you can get\n \n\noutputs look like um and if you can get\naway and not every problem can but most\n\naway and not every problem can but most\n \n\naway and not every problem can but most\ncan in my experience if you can get away\n\ncan in my experience if you can get away\n \n\ncan in my experience if you can get away\nwith starting with um a prompt model\n\nwith starting with um a prompt model\n \n\nwith starting with um a prompt model\nlike gp4 then you're in a great spot\n\nlike gp4 then you're in a great spot\n \n\nlike gp4 then you're in a great spot\nbecause you can very accurately capture\n\nbecause you can very accurately capture\n \n\nbecause you can very accurately capture\nthat input distribution you know what\n\nthat input distribution you know what\n \n\nthat input distribution you know what\nyour model is going to need to do and\n\nyour model is going to need to do and\n \n\nyour model is going to need to do and\nyou can just train a model that's really\n\nyou can just train a model that's really\n \n\nyou can just train a model that's really\nreally good at doing that um we do that\n\nreally good at doing that um we do that\n \n\nreally good at doing that um we do that\nuh first party with an open pipe it's\n\nuh first party with an open pipe it's\n \n\nuh first party with an open pipe it's\nyou don't overthink this step by the way\n\nyou don't overthink this step by the way\n \n\nyou don't overthink this step by the way\nI've talked to tons of people who just\n\nI've talked to tons of people who just\n \n\nI've talked to tons of people who just\nlike in whatever database you're using\n\nlike in whatever database you're using\n \n\nlike in whatever database you're using\non your application or whatever\n\non your application or whatever\n \n\non your application or whatever\nobservability or logs provider you're\n\nobservability or logs provider you're\n \n\nobservability or logs provider you're\nalready using you just every time you do\n\nalready using you just every time you do\n \n\nalready using you just every time you do\na prompt you throw it in there um the\n\na prompt you throw it in there um the\n \n\na prompt you throw it in there um the\nimportant thing is you're just you're\n\nimportant thing is you're just you're\n \n\nimportant thing is you're just you're\ncapturing that somewhere um and then and\n\ncapturing that somewhere um and then and\n \n\ncapturing that somewhere um and then and\nthen you can go ahead and use that to\n\nthen you can go ahead and use that to\n \n\nthen you can go ahead and use that to\nfigure out what your what your data\n\nfigure out what your what your data\n \n\nfigure out what your what your data\nactually looks\n\nactually looks\n \n\nactually looks\nlike um and yeah this is this is what it\n\nlike um and yeah this is this is what it\n \n\nlike um and yeah this is this is what it\nlooks like within open pipe okay once\n\nlooks like within open pipe okay once\n \n\nlooks like within open pipe okay once\nyou've captured that the next step is to\n\nyou've captured that the next step is to\n \n\nyou've captured that the next step is to\nmake sure your data set is as high\n\nmake sure your data set is as high\n \n\nmake sure your data set is as high\nquality as possible and the amount of\n\nquality as possible and the amount of\n \n\nquality as possible and the amount of\nwork that needs to go this step is\n\nwork that needs to go this step is\n \n\nwork that needs to go this step is\npretty variable um if you're lucky\n\npretty variable um if you're lucky\n \n\npretty variable um if you're lucky\nenough that like gp4 is doing a great\n\nenough that like gp4 is doing a great\n \n\nenough that like gp4 is doing a great\njob on your problem already and you're\n\njob on your problem already and you're\n \n\njob on your problem already and you're\nreally just trying to decrease latency\n\nreally just trying to decrease latency\n \n\nreally just trying to decrease latency\nor cost um then you're in a really great\n\nor cost um then you're in a really great\n \n\nor cost um then you're in a really great\nspot you can literally just use that\n\nspot you can literally just use that\n \n\nspot you can literally just use that\nexact data that you've been generating\n\nexact data that you've been generating\n \n\nexact data that you've been generating\nfrom just your normal production um and\n\nfrom just your normal production um and\n \n\nfrom just your normal production um and\nand you can find T on do great um if\n\nand you can find T on do great um if\n \n\nand you can find T on do great um if\nthere it is making mistakes there may be\n\nthere it is making mistakes there may be\n \n\nthere it is making mistakes there may be\nextra steps required you may have to do\n\nextra steps required you may have to do\n \n\nextra steps required you may have to do\nsome human review um or you may have to\n\nsome human review um or you may have to\n \n\nsome human review um or you may have to\ndo some relabeling we say a stronger\n\ndo some relabeling we say a stronger\n \n\ndo some relabeling we say a stronger\nmodel um this is something I mentioned\n\nmodel um this is something I mentioned\n \n\nmodel um this is something I mentioned\nthat that mixture of Agents thing that I\n\nthat that mixture of Agents thing that I\n \n\nthat that mixture of Agents thing that I\nmentioned earlier this is something\n\nmentioned earlier this is something\n \n\nmentioned earlier this is something\nwe've seen a lot of our customers be\n\nwe've seen a lot of our customers be\n \n\nwe've seen a lot of our customers be\nsuccessful with is basically going\n\nsuccessful with is basically going\n \n\nsuccessful with is basically going\nthrough multiple iterations it's still\n\nthrough multiple iterations it's still\n \n\nthrough multiple iterations it's still\nfully automated still with an llm but it\n\nfully automated still with an llm but it\n \n\nfully automated still with an llm but it\ncan solve many of those uh reliability\n\ncan solve many of those uh reliability\n \n\ncan solve many of those uh reliability\nissues by going through multiple\n\nissues by going through multiple\n \n\nissues by going through multiple\niterations as part of the labeling of\n\niterations as part of the labeling of\n \n\niterations as part of the labeling of\ncourse it's more expensive because\n\ncourse it's more expensive because\n \n\ncourse it's more expensive because\nyou're doing this multiple times and\n\nyou're doing this multiple times and\n \n\nyou're doing this multiple times and\nit's slower but this is for generating\n\nit's slower but this is for generating\n \n\nit's slower but this is for generating\ntraining data you don't have to do it on\n\ntraining data you don't have to do it on\n \n\ntraining data you don't have to do it on\non that many so it's it's still very\n\non that many so it's it's still very\n \n\non that many so it's it's still very\nmanageable anyway um once that data is\n\nmanageable anyway um once that data is\n \n\nmanageable anyway um once that data is\nprepared then you actually have to train\n\nprepared then you actually have to train\n \n\nprepared then you actually have to train\nthe model um this part I feel like can\n\nthe model um this part I feel like can\n \n\nthe model um this part I feel like can\nbe pretty intimidating to people this is\n\nbe pretty intimidating to people this is\n \n\nbe pretty intimidating to people this is\nwhat people think if they if they're\n\nwhat people think if they if they're\n \n\nwhat people think if they if they're\ncoming in without familiar with fine\n\ncoming in without familiar with fine\n \n\ncoming in without familiar with fine\ntuning this is what people think is\n\ntuning this is what people think is\n \n\ntuning this is what people think is\ngoing to be the hard part is figuring\n\ngoing to be the hard part is figuring\n \n\ngoing to be the hard part is figuring\nout the hyper parameters you know like\n\nout the hyper parameters you know like\n \n\nout the hyper parameters you know like\nhow do I actually train it how do I get\n\nhow do I actually train it how do I get\n \n\nhow do I actually train it how do I get\nthe infrastructure um honestly at this\n\nthe infrastructure um honestly at this\n \n\nthe infrastructure um honestly at this\npoint like I feel like on this this is\n\npoint like I feel like on this this is\n \n\npoint like I feel like on this this is\nmostly a solved problem um of course\n\nmostly a solved problem um of course\n \n\nmostly a solved problem um of course\nlike we do it just like with one click\n\nlike we do it just like with one click\n \n\nlike we do it just like with one click\nat open pipe um but there's lots of\n\nat open pipe um but there's lots of\n \n\nat open pipe um but there's lots of\nother providers out there I've listed\n\nother providers out there I've listed\n \n\nother providers out there I've listed\nsome that also do fine tuning it's quite\n\nsome that also do fine tuning it's quite\n \n\nsome that also do fine tuning it's quite\nsimple there's also some pretty easy to\n\nsimple there's also some pretty easy to\n \n\nsimple there's also some pretty easy to\nuse at this point open source libraries\n\nuse at this point open source libraries\n \n\nuse at this point open source libraries\nyou can use um that I've listed on the\n\nyou can use um that I've listed on the\n \n\nyou can use um that I've listed on the\nother side as well and um and ultimately\n\nother side as well and um and ultimately\n \n\nother side as well and um and ultimately\nlike this part is actually that the\n\nlike this part is actually that the\n \n\nlike this part is actually that the\ninfrastructure and tooling has gotten\n\ninfrastructure and tooling has gotten\n \n\ninfrastructure and tooling has gotten\ngood enough the architectures involved\n\ngood enough the architectures involved\n \n\ngood enough the architectures involved\nhave gotten like standardized enough\n\nhave gotten like standardized enough\n \n\nhave gotten like standardized enough\nthat you don't really have to do that\n\nthat you don't really have to do that\n \n\nthat you don't really have to do that\nmuch custom work um so so the the\n\nmuch custom work um so so the the\n \n\nmuch custom work um so so the the\ntraining part is actually relatively\n\ntraining part is actually relatively\n \n\ntraining part is actually relatively\neasy um and then you want to evaluate so\n\neasy um and then you want to evaluate so\n \n\neasy um and then you want to evaluate so\nevaluations obviously there's going to\n\nevaluations obviously there's going to\n \n\nevaluations obviously there's going to\nbe a lot of discussion on that on this\n\nbe a lot of discussion on that on this\n \n\nbe a lot of discussion on that on this\ntrack um really quickly I like to think\n\ntrack um really quickly I like to think\n \n\ntrack um really quickly I like to think\nof evaluations as being in two\n\nof evaluations as being in two\n \n\nof evaluations as being in two\ncategories um which I call the inner\n\ncategories um which I call the inner\n \n\ncategories um which I call the inner\nloop and the outer loop evaluations\n\nloop and the outer loop evaluations\n \n\nloop and the outer loop evaluations\ninner loop evaluations just mean\n\ninner loop evaluations just mean\n \n\ninner loop evaluations just mean\nsomething you can run really fast and\n\nsomething you can run really fast and\n \n\nsomething you can run really fast and\nget feedback on fast so what we do at\n\nget feedback on fast so what we do at\n \n\nget feedback on fast so what we do at\nopen pipe is we use LM as judge\n\nopen pipe is we use LM as judge\n \n\nopen pipe is we use LM as judge\nevaluation\n\nevaluation\n \n\nevaluation\nwe let you define uh you know the\n\nwe let you define uh you know the\n \n\nwe let you define uh you know the\ncriteria and then you can run that\n\ncriteria and then you can run that\n \n\ncriteria and then you can run that\nevaluation on your fine tune output vers\n\nevaluation on your fine tune output vers\n \n\nevaluation on your fine tune output vers\nyour data set and see that um you know\n\nyour data set and see that um you know\n \n\nyour data set and see that um you know\nreally quickly does this seem right and\n\nreally quickly does this seem right and\n \n\nreally quickly does this seem right and\nthen there's the outer loop evaluations\n\nthen there's the outer loop evaluations\n \n\nthen there's the outer loop evaluations\nwhich are incredibly important as well\n\nwhich are incredibly important as well\n \n\nwhich are incredibly important as well\nand that is going to be very business\n\nand that is going to be very business\n \n\nand that is going to be very business\nspecific very use case specific but it's\n\nspecific very use case specific but it's\n \n\nspecific very use case specific but it's\nbasically what do I care about how do I\n\nbasically what do I care about how do I\n \n\nbasically what do I care about how do I\nmeasure it is this actually achieving\n\nmeasure it is this actually achieving\n \n\nmeasure it is this actually achieving\nthat and usually that requires actually\n\nthat and usually that requires actually\n \n\nthat and usually that requires actually\ndeploying your model or at least doing\n\ndeploying your model or at least doing\n \n\ndeploying your model or at least doing\nsome kind of AB testing or something\n\nsome kind of AB testing or something\n \n\nsome kind of AB testing or something\nlike that in production and seeing is\n\nlike that in production and seeing is\n \n\nlike that in production and seeing is\nthis driving is is this doing what um\n\nthis driving is is this doing what um\n \n\nthis driving is is this doing what um\nyou know what what I'm expecting it to\n\nyou know what what I'm expecting it to\n \n\nyou know what what I'm expecting it to\ndo okay within open pipe we do have that\n\ndo okay within open pipe we do have that\n \n\ndo okay within open pipe we do have that\ninter Loop really strongly integrated we\n\ninter Loop really strongly integrated we\n \n\ninter Loop really strongly integrated we\ncan you can Define Elm judges you can\n\ncan you can Define Elm judges you can\n \n\ncan you can Define Elm judges you can\nsee the uh basically how your fine two\n\nsee the uh basically how your fine two\n \n\nsee the uh basically how your fine two\nmodels are comparing um and very quickly\n\nmodels are comparing um and very quickly\n \n\nmodels are comparing um and very quickly\ndecide if it's something you you you\n\ndecide if it's something you you you\n \n\ndecide if it's something you you you\nwant to promote to prod and and get that\n\nwant to promote to prod and and get that\n \n\nwant to promote to prod and and get that\nouter loop done with okay got a minute\n\nouter loop done with okay got a minute\n \n\nouter loop done with okay got a minute\nleft deployment um this is actually also\n\nleft deployment um this is actually also\n \n\nleft deployment um this is actually also\nan area where things have gotten much\n\nan area where things have gotten much\n \n\nan area where things have gotten much\nmore mature recently and actually it's\n\nmore mature recently and actually it's\n \n\nmore mature recently and actually it's\nnot that scary anymore um you can\n\nnot that scary anymore um you can\n \n\nnot that scary anymore um you can\nself-host there's great projects out\n\nself-host there's great projects out\n \n\nself-host there's great projects out\nthere that can give you like BLM and\n\nthere that can give you like BLM and\n \n\nthere that can give you like BLM and\ntrtm they both give you like an open AI\n\ntrtm they both give you like an open AI\n \n\ntrtm they both give you like an open AI\ncompatible API that you can just like\n\ncompatible API that you can just like\n \n\ncompatible API that you can just like\nhost on your own gpus and kind of call\n\nhost on your own gpus and kind of call\n \n\nhost on your own gpus and kind of call\nthe same way you would there's also a\n\nthe same way you would there's also a\n \n\nthe same way you would there's also a\nlot of hosted providers we will host\n\nlot of hosted providers we will host\n \n\nlot of hosted providers we will host\nyour model for you if you want there's\n\nyour model for you if you want there's\n \n\nyour model for you if you want there's\nlots of other inference providers out\n\nlots of other inference providers out\n \n\nlots of other inference providers out\nthere um this space like I said has\n\nthere um this space like I said has\n \n\nthere um this space like I said has\ngotten very mature recently so um not\n\ngotten very mature recently so um not\n \n\ngotten very mature recently so um not\nsomething that uh you you should be too\n\nsomething that uh you you should be too\n \n\nsomething that uh you you should be too\nscared of um okay this is just a recap\n\nscared of um okay this is just a recap\n \n\nscared of um okay this is just a recap\nof the the workflow you got to prepare\n\nof the the workflow you got to prepare\n \n\nof the the workflow you got to prepare\nyour data train a model evaluate it and\n\nyour data train a model evaluate it and\n \n\nyour data train a model evaluate it and\ndeploy it um and then finally um this is\n\ndeploy it um and then finally um this is\n \n\ndeploy it um and then finally um this is\nwhat you're trying to achieve uh you\n\nwhat you're trying to achieve uh you\n \n\nwhat you're trying to achieve uh you\nknow you you can get for 125th the cost\n\nknow you you can get for 125th the cost\n \n\nknow you you can get for 125th the cost\nbetter performance than gp4 um and and\n\nbetter performance than gp4 um and and\n \n\nbetter performance than gp4 um and and\nfor a lot of you case it is worth it um\n\nfor a lot of you case it is worth it um\n \n\nfor a lot of you case it is worth it um\nand hopefully uh this is enough\n\nand hopefully uh this is enough\n \n\nand hopefully uh this is enough\ninformation so you can decide for\n\ninformation so you can decide for\n \n\ninformation so you can decide for\nyourself when the right moment is to do\n\nyourself when the right moment is to do\n \n\nyourself when the right moment is to do\nthat thank you\n\n \n\n[Music]" } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\n[Music]\n\n[Music]\n \n\n[Music]\nso we are here to talk to you about\n\nso we are here to talk to you about\n \n\nso we are here to talk to you about\nllama file today what we've been doing\n\nllama file today what we've been doing\n \n\nllama file today what we've been doing\non this\n\n \n\nproject so I'll get it star..." }, { "role": "assistant", "content": "[Music]\n\n[Music]\n \n\n[Music]\nso we are here to talk to you about\n\nso we are here to talk to you about\n \n\nso we are here to talk to you about\nllama file today what we've been doing\n\nllama file today what we've been doing\n \n\nllama file today what we've been doing\non this\n\n \n\nproject so I'll get it started I'm going\n\nproject so I'll get it started I'm going\n \n\nproject so I'll get it started I'm going\nto tell you what llama file is how it\n\nto tell you what llama file is how it\n \n\nto tell you what llama file is how it\nworks I'm going to spend a little time\n\nworks I'm going to spend a little time\n \n\nworks I'm going to spend a little time\ntalking about why we're building it why\n\ntalking about why we're building it why\n \n\ntalking about why we're building it why\nMOS specifically is involved and then\n\nMOS specifically is involved and then\n \n\nMOS specifically is involved and then\nI'm going to hand it over for the fun\n\nI'm going to hand it over for the fun\n \n\nI'm going to hand it over for the fun\npart to Justine Justine is going to talk\n\npart to Justine Justine is going to talk\n \n\npart to Justine Justine is going to talk\nabout the actual work that she and the\n\nabout the actual work that she and the\n \n\nabout the actual work that she and the\nopen source Community have been doing on\n\nopen source Community have been doing on\n \n\nopen source Community have been doing on\nthis project lots of insights and tricks\n\nthis project lots of insights and tricks\n \n\nthis project lots of insights and tricks\nand hacks that have made CPU inference\n\nand hacks that have made CPU inference\n \n\nand hacks that have made CPU inference\ngo faster than ever before so that'll be\n\ngo faster than ever before so that'll be\n \n\ngo faster than ever before so that'll be\nfun and we're done we want you to share\n\nfun and we're done we want you to share\n \n\nfun and we're done we want you to share\nthe feeling that we have which is kind\n\nthe feeling that we have which is kind\n \n\nthe feeling that we have which is kind\nof a sense of excitement and empowerment\n\nof a sense of excitement and empowerment\n \n\nof a sense of excitement and empowerment\nfrom the knowledge that there are lots\n\nfrom the knowledge that there are lots\n \n\nfrom the knowledge that there are lots\nof really interesting juicy impactful\n\nof really interesting juicy impactful\n \n\nof really interesting juicy impactful\nproblems still to be solved in AI a lot\n\nproblems still to be solved in AI a lot\n \n\nproblems still to be solved in AI a lot\nof them and the key thing is it's not\n\nof them and the key thing is it's not\n \n\nof them and the key thing is it's not\njust the big folks who can solve these\n\njust the big folks who can solve these\n \n\njust the big folks who can solve these\nproblems it's individuals and small\n\nproblems it's individuals and small\n \n\nproblems it's individuals and small\ngroups working together in open source\n\ngroups working together in open source\n \n\ngroups working together in open source\nso anyone in this room or anyone\n\nso anyone in this room or anyone\n \n\nso anyone in this room or anyone\nlistening to this talk can potentially\n\nlistening to this talk can potentially\n \n\nlistening to this talk can potentially\nmake a big impact in this\n\n \n\nspace so what's LL file LL file is an\n\nspace so what's LL file LL file is an\n \n\nspace so what's LL file LL file is an\nopen source project from Mozilla that\n\nopen source project from Mozilla that\n \n\nopen source project from Mozilla that\nhas the goal of\n\nhas the goal of\n \n\nhas the goal of\ndemocratizing access to AI so we do that\n\ndemocratizing access to AI so we do that\n \n\ndemocratizing access to AI so we do that\nin a few different ways the the first is\n\nin a few different ways the the first is\n \n\nin a few different ways the the first is\nprobably how if you've heard of llama\n\nprobably how if you've heard of llama\n \n\nprobably how if you've heard of llama\nfile the reason you heard of it it's the\n\nfile the reason you heard of it it's the\n \n\nfile the reason you heard of it it's the\noriginal magic trick of the project that\n\noriginal magic trick of the project that\n \n\noriginal magic trick of the project that\nJustin figured out which is how to turn\n\nJustin figured out which is how to turn\n \n\nJustin figured out which is how to turn\nweights into programs so a llama file is\n\nweights into programs so a llama file is\n \n\nweights into programs so a llama file is\na single file executable that runs\n\na single file executable that runs\n \n\na single file executable that runs\nwithout any installation on pretty much\n\nwithout any installation on pretty much\n \n\nwithout any installation on pretty much\nevery operating system every CPU\n\nevery operating system every CPU\n \n\nevery operating system every CPU\narchitecture and every GPU\n\narchitecture and every GPU\n \n\narchitecture and every GPU\narchitecture and that's all thank you\n\narchitecture and that's all thank you\n \n\narchitecture and that's all thank you\nvery\n\nvery\n \n\nvery\nmuch that was easy\n\nmuch that was easy\n \n\nmuch that was easy\nyeah so by the way this isn't just one\n\nyeah so by the way this isn't just one\n \n\nyeah so by the way this isn't just one\nfile like for Windows right and a\n\nfile like for Windows right and a\n \n\nfile like for Windows right and a\ndifferent one for Linux and Mac it's\n\ndifferent one for Linux and Mac it's\n \n\ndifferent one for Linux and Mac it's\nactually a single file you can download\n\nactually a single file you can download\n \n\nactually a single file you can download\na llama file run on any computer in the\n\na llama file run on any computer in the\n \n\na llama file run on any computer in the\nworld and it'll just work and it'll use\n\nworld and it'll just work and it'll use\n \n\nworld and it'll just work and it'll use\nthe hardware you have whether that be\n\nthe hardware you have whether that be\n \n\nthe hardware you have whether that be\nfancy gpus or your CPU so we should talk\n\nfancy gpus or your CPU so we should talk\n \n\nfancy gpus or your CPU so we should talk\na little more later about how uh Justine\n\na little more later about how uh Justine\n \n\na little more later about how uh Justine\nmade that\n\nmade that\n \n\nmade that\nwork but we're here to talk about\n\nwork but we're here to talk about\n \n\nwork but we're here to talk about\nanother topic too most of the talk is\n\nanother topic too most of the talk is\n \n\nanother topic too most of the talk is\nactually about this which is CPU\n\nactually about this which is CPU\n \n\nactually about this which is CPU\ninference speed now you might ask why do\n\ninference speed now you might ask why do\n \n\ninference speed now you might ask why do\nwe need to worry about CPU inference\n\nwe need to worry about CPU inference\n \n\nwe need to worry about CPU inference\nspeed we've got these fancy gpus right\n\nspeed we've got these fancy gpus right\n \n\nspeed we've got these fancy gpus right\nwell no\n\nwell no\n \n\nwell no\ndisrespect Almighty\n\ndisrespect Almighty\n \n\ndisrespect Almighty\nGenson first of his name master of\n\nGenson first of his name master of\n \n\nGenson first of his name master of\nmarket\n\nmarket\n \n\nmarket\ncap uh don't strike me down but I would\n\ncap uh don't strike me down but I would\n \n\ncap uh don't strike me down but I would\nposit that it is not a universally good\n\nposit that it is not a universally good\n \n\nposit that it is not a universally good\nthing that we are so dependent in this\n\nthing that we are so dependent in this\n \n\nthing that we are so dependent in this\nroom on\n\nroom on\n \n\nroom on\ngpus uh they are expensive they're\n\ngpus uh they are expensive they're\n \n\ngpus uh they are expensive they're\ndifficult to Source let's face it they\n\ndifficult to Source let's face it they\n \n\ndifficult to Source let's face it they\nconsume a lot of electricity which we\n\nconsume a lot of electricity which we\n \n\nconsume a lot of electricity which we\nmight want to think about\n\nmight want to think about\n \n\nmight want to think about\nbut uh bigger picture we have an entire\n\nbut uh bigger picture we have an entire\n \n\nbut uh bigger picture we have an entire\nplanet of CPUs out there literally all\n\nplanet of CPUs out there literally all\n \n\nplanet of CPUs out there literally all\nover the world great Hardware often\n\nover the world great Hardware often\n \n\nover the world great Hardware often\naffordable hardware and we are at risk\n\naffordable hardware and we are at risk\n \n\naffordable hardware and we are at risk\nof just kind of throwing that all away\n\nof just kind of throwing that all away\n \n\nof just kind of throwing that all away\nwith this new era of of AI and we don't\n\nwith this new era of of AI and we don't\n \n\nwith this new era of of AI and we don't\nneed to do that so who here knows llama\n\nneed to do that so who here knows llama\n \n\nneed to do that so who here knows llama\nCPP this is an easy question right so we\n\nCPP this is an easy question right so we\n \n\nCPP this is an easy question right so we\nall know and love this project we build\n\nall know and love this project we build\n \n\nall know and love this project we build\non top of that project with llama file\n\non top of that project with llama file\n \n\non top of that project with llama file\nand we contribute our uh performance\n\nand we contribute our uh performance\n \n\nand we contribute our uh performance\nenhancements back to it many have been\n\nenhancements back to it many have been\n \n\nenhancements back to it many have been\nmerged in that project proved that CPUs\n\nmerged in that project proved that CPUs\n \n\nmerged in that project proved that CPUs\ncould do inference perfectly well and so\n\ncould do inference perfectly well and so\n \n\ncould do inference perfectly well and so\nwe have been basically trying to take\n\nwe have been basically trying to take\n \n\nwe have been basically trying to take\nthat performance to the next level and\n\nthat performance to the next level and\n \n\nthat performance to the next level and\nas a result of Justine and the\n\nas a result of Justine and the\n \n\nas a result of Justine and the\ncommunity's work depending on what CPU\n\ncommunity's work depending on what CPU\n \n\ncommunity's work depending on what CPU\nyou're using what model you're running\n\nyou're using what model you're running\n \n\nyou're using what model you're running\nwhat weights you will see between 30 and\n\nwhat weights you will see between 30 and\n \n\nwhat weights you will see between 30 and\n500% speed increases at L file which\n\n500% speed increases at L file which\n \n\n500% speed increases at L file which\nkind of still blows my mind and I by the\n\nkind of still blows my mind and I by the\n \n\nkind of still blows my mind and I by the\nway I don't think we're anywhere near\n\nway I don't think we're anywhere near\n \n\nway I don't think we're anywhere near\n[Applause]\n\n[Applause]\n \n\n[Applause]\ndone so these things also run locally by\n\ndone so these things also run locally by\n \n\ndone so these things also run locally by\nthe way this runs totally on your\n\nthe way this runs totally on your\n \n\nthe way this runs totally on your\nmachine there's no network access you\n\nmachine there's no network access you\n \n\nmachine there's no network access you\ncould take a pair of scissors and cut\n\ncould take a pair of scissors and cut\n \n\ncould take a pair of scissors and cut\nthe ethernet cord and it'll work which\n\nthe ethernet cord and it'll work which\n \n\nthe ethernet cord and it'll work which\nis what I asked Dolly 3 to\n\nis what I asked Dolly 3 to\n \n\nis what I asked Dolly 3 to\ndraw\n\ndraw\n \n\ndraw\nokay I don't think it understood the\n\nokay I don't think it understood the\n \n\nokay I don't think it understood the\nassignment but that's all\n\nassignment but that's all\n \n\nassignment but that's all\nright uh but seriously like we're not\n\nright uh but seriously like we're not\n \n\nright uh but seriously like we're not\ncalling Cloud llms there's no monitoring\n\ncalling Cloud llms there's no monitoring\n \n\ncalling Cloud llms there's no monitoring\nyour analytics no bits to leave your\n\nyour analytics no bits to leave your\n \n\nyour analytics no bits to leave your\nmachine it's totally private and local\n\nmachine it's totally private and local\n \n\nmachine it's totally private and local\nand everything you need comes in the box\n\nand everything you need comes in the box\n \n\nand everything you need comes in the box\nso whether you want to just play with a\n\nso whether you want to just play with a\n \n\nso whether you want to just play with a\nmodel that you just found on hugging\n\nmodel that you just found on hugging\n \n\nmodel that you just found on hugging\nface or you want to start building local\n\nface or you want to start building local\n \n\nface or you want to start building local\nlocally running llm applications on your\n\nlocally running llm applications on your\n \n\nlocally running llm applications on your\nmachine you got everything you need in\n\nmachine you got everything you need in\n \n\nmachine you got everything you need in\nthe box and they're readily available so\n\nthe box and they're readily available so\n \n\nthe box and they're readily available so\nuh hugging face now supports llama file\n\nuh hugging face now supports llama file\n \n\nuh hugging face now supports llama file\nas a file type so you can search and\n\nas a file type so you can search and\n \n\nas a file type so you can search and\nfilter by llama file you could also just\n\nfilter by llama file you could also just\n \n\nfilter by llama file you could also just\nsearch Mozilla on hugging face you'll\n\nsearch Mozilla on hugging face you'll\n \n\nsearch Mozilla on hugging face you'll\nfind we have a bunch of llama files that\n\nfind we have a bunch of llama files that\n \n\nfind we have a bunch of llama files that\nwe've already published and with a\n\nwe've already published and with a\n \n\nwe've already published and with a\nsingle command you can create your own\n\nsingle command you can create your own\n \n\nsingle command you can create your own\nso really this project is collapsing all\n\nso really this project is collapsing all\n \n\nso really this project is collapsing all\nthe complexity of the open source AI\n\nthe complexity of the open source AI\n \n\nthe complexity of the open source AI\nstack down into a single action in a\n\nstack down into a single action in a\n \n\nstack down into a single action in a\nsingle\n\n \n\nso why are we involved why is Mill\n\nso why are we involved why is Mill\n \n\nso why are we involved why is Mill\ninvolved in this you might be saying\n\ninvolved in this you might be saying\n \n\ninvolved in this you might be saying\ndon't you folks make\n\ndon't you folks make\n \n\ndon't you folks make\nbrowsers in fact we do we make a damn\n\nbrowsers in fact we do we make a damn\n \n\nbrowsers in fact we do we make a damn\nfine browser and you should try it out\n\nfine browser and you should try it out\n \n\nfine browser and you should try it out\nif you haven't\n\nif you haven't\n \n\nif you haven't\nlately but we exist also for a bigger\n\nlately but we exist also for a bigger\n \n\nlately but we exist also for a bigger\npurpose which is to fight for the web so\n\npurpose which is to fight for the web so\n \n\npurpose which is to fight for the web so\nI'm going to ask a question here who\n\nI'm going to ask a question here who\n \n\nI'm going to ask a question here who\nhere remembers using the original\n\nhere remembers using the original\n \n\nhere remembers using the original\nNetscape Navigator don't be shy no one\n\nNetscape Navigator don't be shy no one\n \n\nNetscape Navigator don't be shy no one\ncan see how old you are they see how old\n\ncan see how old you are they see how old\n \n\ncan see how old you are they see how old\nI\n\nI\n \n\nI\nam a lot of hands right so you are my\n\nam a lot of hands right so you are my\n \n\nam a lot of hands right so you are my\npeople you remember the '90s\n\npeople you remember the '90s\n \n\npeople you remember the '90s\nmy C\n\nmy C\n \n\nmy C\nTV terrible\n\nTV terrible\n \n\nTV terrible\nhaircuts Milly the nly I don't know\n\nhaircuts Milly the nly I don't know\n \n\nhaircuts Milly the nly I don't know\nwhatever my point is you remember the\n\nwhatever my point is you remember the\n \n\nwhatever my point is you remember the\nearly days of the web and you remember\n\nearly days of the web and you remember\n \n\nearly days of the web and you remember\nhow close we came to one company and one\n\nhow close we came to one company and one\n \n\nhow close we came to one company and one\nproduct kind of controlling the whole\n\nproduct kind of controlling the whole\n \n\nproduct kind of controlling the whole\nthing and we kind of see that maybe\n\nthing and we kind of see that maybe\n \n\nthing and we kind of see that maybe\nhappening again today with AI no no\n\nhappening again today with AI no no\n \n\nhappening again today with AI no no\nmatter what we may think of these\n\nmatter what we may think of these\n \n\nmatter what we may think of these\ncompanies the reality is there are some\n\ncompanies the reality is there are some\n \n\ncompanies the reality is there are some\nvery influential big tech companies that\n\nvery influential big tech companies that\n \n\nvery influential big tech companies that\nare in a position to maybe control the\n\nare in a position to maybe control the\n \n\nare in a position to maybe control the\nfuture of machine\n\nfuture of machine\n \n\nfuture of machine\nintelligence and that's itself not a\n\nintelligence and that's itself not a\n \n\nintelligence and that's itself not a\ngreat thing it's not great for Equity\n\ngreat thing it's not great for Equity\n \n\ngreat thing it's not great for Equity\nit's not great especially for users\n\nit's not great especially for users\n \n\nit's not great especially for users\nsense of\n\nsense of\n \n\nsense of\nprivacy and uh safety and agency and\n\nprivacy and uh safety and agency and\n \n\nprivacy and uh safety and agency and\ncontrol and we've had an answer to this\n\ncontrol and we've had an answer to this\n \n\ncontrol and we've had an answer to this\nfor many years it's called open source\n\nfor many years it's called open source\n \n\nfor many years it's called open source\nand the answer is right in the name\n\nand the answer is right in the name\n \n\nand the answer is right in the name\nright open source transparency is the\n\nright open source transparency is the\n \n\nright open source transparency is the\nsolution here and it's important for us\n\nsolution here and it's important for us\n \n\nsolution here and it's important for us\nto have viable open source Alternatives\n\nto have viable open source Alternatives\n \n\nto have viable open source Alternatives\nin Ai and that's why mosil is getting\n\nin Ai and that's why mosil is getting\n \n\nin Ai and that's why mosil is getting\ninvolved that's why we made llama file\n\ninvolved that's why we made llama file\n \n\ninvolved that's why we made llama file\nand more projects to follow and uh I\n\nand more projects to follow and uh I\n \n\nand more projects to follow and uh I\nknow many of you in this room are\n\nknow many of you in this room are\n \n\nknow many of you in this room are\nalready working on open source AI we\n\nalready working on open source AI we\n \n\nalready working on open source AI we\nwant to help support what you're doing\n\nwant to help support what you're doing\n \n\nwant to help support what you're doing\nso that I'm going to hand it over to\n\nso that I'm going to hand it over to\n \n\nso that I'm going to hand it over to\nJustine who's going to tell you actually\n\nJustine who's going to tell you actually\n \n\nJustine who's going to tell you actually\nthe cool part which is all the things\n\nthe cool part which is all the things\n \n\nthe cool part which is all the things\nthat she and the community have been\n\nthat she and the community have been\n \n\nthat she and the community have been\ndoing on this project Justine\n\n \n\nthank you Steven um so I'm Justine tuny\n\nthank you Steven um so I'm Justine tuny\n \n\nthank you Steven um so I'm Justine tuny\nI'm the lead developer on Lama file and\n\nI'm the lead developer on Lama file and\n \n\nI'm the lead developer on Lama file and\nas Steph mentioned I'm going to talk\n\nas Steph mentioned I'm going to talk\n \n\nas Steph mentioned I'm going to talk\nabout some of the cool work we've been\n\nabout some of the cool work we've been\n \n\nabout some of the cool work we've been\ndoing in the community to help you run\n\ndoing in the community to help you run\n \n\ndoing in the community to help you run\nthe fastest local L experience possible\n\nthe fastest local L experience possible\n \n\nthe fastest local L experience possible\nand in order to do this we started by\n\nand in order to do this we started by\n \n\nand in order to do this we started by\nfirst getting it to run on the systems\n\nfirst getting it to run on the systems\n \n\nfirst getting it to run on the systems\nat all and with Cosmopolitan what it\n\nat all and with Cosmopolitan what it\n \n\nat all and with Cosmopolitan what it\nenables us to do is take your weights in\n\nenables us to do is take your weights in\n \n\nenables us to do is take your weights in\na single file and run it on six os's and\n\na single file and run it on six os's and\n \n\na single file and run it on six os's and\nthere's a really cool hack that makes\n\nthere's a really cool hack that makes\n \n\nthere's a really cool hack that makes\nthat possible which is we basically take\n\nthat possible which is we basically take\n \n\nthat possible which is we basically take\na Unix 6 Edition shell script put it in\n\na Unix 6 Edition shell script put it in\n \n\na Unix 6 Edition shell script put it in\nthe MS do stub of a portable\n\nthe MS do stub of a portable\n \n\nthe MS do stub of a portable\nexecutable and that enables it to run on\n\nexecutable and that enables it to run on\n \n\nexecutable and that enables it to run on\nMac windows and bsds and Linux Etc\n\nMac windows and bsds and Linux Etc\n \n\nMac windows and bsds and Linux Etc\nreally cool\n\nreally cool\n \n\nreally cool\nstuff and once we conquered the\n\nstuff and once we conquered the\n \n\nstuff and once we conquered the\nportability issue with CPUs um um I had\n\nportability issue with CPUs um um I had\n \n\nportability issue with CPUs um um I had\nthe opportunity to work with Mozilla on\n\nthe opportunity to work with Mozilla on\n \n\nthe opportunity to work with Mozilla on\nbringing this to Ai and with AI gpus are\n\nbringing this to Ai and with AI gpus are\n \n\nbringing this to Ai and with AI gpus are\nindispensable as much as we focus on\n\nindispensable as much as we focus on\n \n\nindispensable as much as we focus on\nCPUs we care very much about gpus too\n\nCPUs we care very much about gpus too\n \n\nCPUs we care very much about gpus too\nbut gpus have always had the problem of\n\nbut gpus have always had the problem of\n \n\nbut gpus have always had the problem of\ndistribut ability many people have\n\ndistribut ability many people have\n \n\ndistribut ability many people have\nneeded to ship kubalas binaries with\n\nneeded to ship kubalas binaries with\n \n\nneeded to ship kubalas binaries with\ntheir project 500 Megs in size can we\n\ntheir project 500 Megs in size can we\n \n\ntheir project 500 Megs in size can we\nreally call our software open source if\n\nreally call our software open source if\n \n\nreally call our software open source if\nit spends the majority of its time in a\n\nit spends the majority of its time in a\n \n\nit spends the majority of its time in a\nproprietary blob so I never felt\n\nproprietary blob so I never felt\n \n\nproprietary blob so I never felt\ncomfortable with that and one of the\n\ncomfortable with that and one of the\n \n\ncomfortable with that and one of the\nways we're solving that is by\n\nways we're solving that is by\n \n\nways we're solving that is by\nDistributing a library called tiny blast\n\nDistributing a library called tiny blast\n \n\nDistributing a library called tiny blast\nthat enables you to ship your llms to\n\nthat enables you to ship your llms to\n \n\nthat enables you to ship your llms to\nplatforms like Windows without depending\n\nplatforms like Windows without depending\n \n\nplatforms like Windows without depending\non sdks it'll run with only the driver\n\n \n\ninstalled but more importantly\n\ninstalled but more importantly\n \n\ninstalled but more importantly\nperformance now llm spend the majority\n\nperformance now llm spend the majority\n \n\nperformance now llm spend the majority\nof their time doing matrix\n\nof their time doing matrix\n \n\nof their time doing matrix\nmultiplication probably the most\n\nmultiplication probably the most\n \n\nmultiplication probably the most\nimportant algorithm is in the world has\n\nimportant algorithm is in the world has\n \n\nimportant algorithm is in the world has\na really simple definition we've been\n\na really simple definition we've been\n \n\na really simple definition we've been\nmaking it go faster for prompt\n\nmaking it go faster for prompt\n \n\nmaking it go faster for prompt\nprocessing and the way we did it is with\n\nprocessing and the way we did it is with\n \n\nprocessing and the way we did it is with\na very simple trick we figured out and\n\na very simple trick we figured out and\n \n\na very simple trick we figured out and\nthis is something all programmers can\n\nthis is something all programmers can\n \n\nthis is something all programmers can\nadopt in their code and it entails\n\nadopt in their code and it entails\n \n\nadopt in their code and it entails\nunrolling the outer loop so let's talk\n\nunrolling the outer loop so let's talk\n \n\nunrolling the outer loop so let's talk\nabout what not to do first and that\n\nabout what not to do first and that\n \n\nabout what not to do first and that\nwould be unrolling the inner one um\n\nwould be unrolling the inner one um\n \n\nwould be unrolling the inner one um\nwe've all seen fun roll Loops Gen 2 it's\n\nwe've all seen fun roll Loops Gen 2 it's\n \n\nwe've all seen fun roll Loops Gen 2 it's\nbad idea computers can generally do that\n\nbad idea computers can generally do that\n \n\nbad idea computers can generally do that\non their own if you unroll the outer\n\non their own if you unroll the outer\n \n\non their own if you unroll the outer\nLoops then your algorithm with matrix\n\nLoops then your algorithm with matrix\n \n\nLoops then your algorithm with matrix\nmultiplication can sort of unfold like a\n\nmultiplication can sort of unfold like a\n \n\nmultiplication can sort of unfold like a\nflower and focus on pure flops like a\n\nflower and focus on pure flops like a\n \n\nflower and focus on pure flops like a\nblast carel and that's really all there\n\nblast carel and that's really all there\n \n\nblast carel and that's really all there\nis to it to getting the majority of the\n\nis to it to getting the majority of the\n \n\nis to it to getting the majority of the\nbenefits of blast to make prompt\n\nbenefits of blast to make prompt\n \n\nbenefits of blast to make prompt\nprocessing go really fast so what's the\n\nprocessing go really fast so what's the\n \n\nprocessing go really fast so what's the\nimpact of this really simple\n\nimpact of this really simple\n \n\nimpact of this really simple\nsolution um this generalizes to a wide\n\nsolution um this generalizes to a wide\n \n\nsolution um this generalizes to a wide\nvariety of hardware\n\nvariety of hardware\n \n\nvariety of hardware\nwe've seen everything from a scrappy\n\nwe've seen everything from a scrappy\n \n\nwe've seen everything from a scrappy\nhobbyist Raspberry Pi um to much bigger\n\nhobbyist Raspberry Pi um to much bigger\n \n\nhobbyist Raspberry Pi um to much bigger\ncomputers going significantly faster you\n\ncomputers going significantly faster you\n \n\ncomputers going significantly faster you\nneed algorithms like this to exploit the\n\nneed algorithms like this to exploit the\n \n\nneed algorithms like this to exploit the\nlatest capabilities of hardware token\n\nlatest capabilities of hardware token\n \n\nlatest capabilities of hardware token\ngeneration wraiths I wouldn't\n\n \n\nbelieve if you use a gaming computer\n\nbelieve if you use a gaming computer\n \n\nbelieve if you use a gaming computer\nlike Intel you're going to see better\n\nlike Intel you're going to see better\n \n\nlike Intel you're going to see better\nperformance with llama file on those two\n\nperformance with llama file on those two\n \n\nperformance with llama file on those two\nreally exciting stuff like particularly\n\nreally exciting stuff like particularly\n \n\nreally exciting stuff like particularly\nwith Alder Lake we are able to get a\n\nwith Alder Lake we are able to get a\n \n\nwith Alder Lake we are able to get a\nForex\n\nForex\n \n\nForex\nImprovement but thread Ripper most of\n\nImprovement but thread Ripper most of\n \n\nImprovement but thread Ripper most of\nall for the first time AVX 512 is\n\nall for the first time AVX 512 is\n \n\nall for the first time AVX 512 is\navailable to Consumers and we've been\n\navailable to Consumers and we've been\n \n\navailable to Consumers and we've been\nable to help you prepare for that future\n\nable to help you prepare for that future\n \n\nable to help you prepare for that future\nso if you have a thread Ripper you're\n\nso if you have a thread Ripper you're\n \n\nso if you have a thread Ripper you're\ngoing to see better performance than\n\ngoing to see better performance than\n \n\ngoing to see better performance than\never almost like a\n\never almost like a\n \n\never almost like a\nGPU now prompted valel speed what makes\n\nGPU now prompted valel speed what makes\n \n\nGPU now prompted valel speed what makes\nit important\n\n \n\nis it's really cool to be able to\n\nis it's really cool to be able to\n \n\nis it's really cool to be able to\ngenerate text and use a chatbot but the\n\ngenerate text and use a chatbot but the\n \n\ngenerate text and use a chatbot but the\nway I want you to think about llama file\n\nway I want you to think about llama file\n \n\nway I want you to think about llama file\nis it's more of a word crunching machine\n\nis it's more of a word crunching machine\n \n\nis it's more of a word crunching machine\nthat can help you understand our world\n\nthat can help you understand our world\n \n\nthat can help you understand our world\nand I love to use it personally for task\n\nand I love to use it personally for task\n \n\nand I love to use it personally for task\nlike\n\nlike\n \n\nlike\nsummarization I love that I can help me\n\nsummarization I love that I can help me\n \n\nsummarization I love that I can help me\nread a blog post and we've used other\n\nread a blog post and we've used other\n \n\nread a blog post and we've used other\nperformance tricks\n\nperformance tricks\n \n\nperformance tricks\ntoo when with\n\ntoo when with\n \n\ntoo when with\nNvidia part of what makes them so\n\nNvidia part of what makes them so\n \n\nNvidia part of what makes them so\nsuccessful it's not just great Hardware\n\nsuccessful it's not just great Hardware\n \n\nsuccessful it's not just great Hardware\nbut they built a great framework too and\n\nbut they built a great framework too and\n \n\nbut they built a great framework too and\ntheir framework helps developers think\n\ntheir framework helps developers think\n \n\ntheir framework helps developers think\nabout programming in a different way\n\nabout programming in a different way\n \n\nabout programming in a different way\nthat helps them be successful I mean who\n\nthat helps them be successful I mean who\n \n\nthat helps them be successful I mean who\nhere thinks that software with CPUs just\n\nhere thinks that software with CPUs just\n \n\nhere thinks that software with CPUs just\ngets slower each year can I see you\n\ngets slower each year can I see you\n \n\ngets slower each year can I see you\nraise some\n\n \n\nhands well part of the um one of the\n\nhands well part of the um one of the\n \n\nhands well part of the um one of the\nthings that's great about Nidia is they\n\nthings that's great about Nidia is they\n \n\nthings that's great about Nidia is they\nshowed us a better alternative to\n\nshowed us a better alternative to\n \n\nshowed us a better alternative to\ngetting performance and when I learned\n\ngetting performance and when I learned\n \n\ngetting performance and when I learned\nhow to program en Cuda I found one of\n\nhow to program en Cuda I found one of\n \n\nhow to program en Cuda I found one of\nthe most important functions was sync\n\nthe most important functions was sync\n \n\nthe most important functions was sync\nthreads this is how you can implement it\n\nthreads this is how you can implement it\n \n\nthreads this is how you can implement it\nfor CPU and like 10 lines of code and if\n\nfor CPU and like 10 lines of code and if\n \n\nfor CPU and like 10 lines of code and if\nyou use the lock set programming model\n\nyou use the lock set programming model\n \n\nyou use the lock set programming model\nuse your CPU as though it were a GPU you\n\nuse your CPU as though it were a GPU you\n \n\nuse your CPU as though it were a GPU you\ncan get really good\n\ncan get really good\n \n\ncan get really good\nperformance now this is going to be a\n\nperformance now this is going to be a\n \n\nperformance now this is going to be a\ndemo showing the impact of this work\n\ndemo showing the impact of this work\n \n\ndemo showing the impact of this work\nbefore and after for\n\nbefore and after for\n \n\nbefore and after for\nsummarization and here we're going to be\n\nsummarization and here we're going to be\n \n\nsummarization and here we're going to be\nprocessing an essay by dictra really\n\nprocessing an essay by dictra really\n \n\nprocessing an essay by dictra really\ncool worth reading but I want you to\n\ncool worth reading but I want you to\n \n\ncool worth reading but I want you to\nwatch like as it processes it in terms\n\nwatch like as it processes it in terms\n \n\nwatch like as it processes it in terms\nof speed\n\nof speed\n \n\nof speed\nhere we see it going and on right we\n\nhere we see it going and on right we\n \n\nhere we see it going and on right we\nhave the new\n\nhave the new\n \n\nhave the new\nversion it's like Bam Bam Bam Bam huge\n\nversion it's like Bam Bam Bam Bam huge\n \n\nversion it's like Bam Bam Bam Bam huge\nnight and day difference it's already\n\nnight and day difference it's already\n \n\nnight and day difference it's already\nsummarizing it in the old version is\n\nsummarizing it in the old version is\n \n\nsummarizing it in the old version is\nlike nowhere\n\n \n\nclose so that is the kind of new\n\nclose so that is the kind of new\n \n\nclose so that is the kind of new\nperformance you can expect and it's the\n\nperformance you can expect and it's the\n \n\nperformance you can expect and it's the\nkind of performance that's actually\n\nkind of performance that's actually\n \n\nkind of performance that's actually\npossible which I wouldn't have imagined\n\npossible which I wouldn't have imagined\n \n\npossible which I wouldn't have imagined\nbeforehand it's really great\n\nbeforehand it's really great\n \n\nbeforehand it's really great\n[Applause]\n\n[Applause]\n \n\n[Applause]\nthank you CPUs can do so much and people\n\nthank you CPUs can do so much and people\n \n\nthank you CPUs can do so much and people\nin the community have like loved this\n\nin the community have like loved this\n \n\nin the community have like loved this\nwork we've managed to attract some like\n\nwork we've managed to attract some like\n \n\nwork we've managed to attract some like\nreally amazing contributors like Ian the\n\nreally amazing contributors like Ian the\n \n\nreally amazing contributors like Ian the\ninventor of kants very popular I'm sure\n\ninventor of kants very popular I'm sure\n \n\ninventor of kants very popular I'm sure\nmany of you have used them he got them\n\nmany of you have used them he got them\n \n\nmany of you have used them he got them\ngoing 2x 4X faster 2 on both x86 and arm\n\ngoing 2x 4X faster 2 on both x86 and arm\n \n\ngoing 2x 4X faster 2 on both x86 and arm\nso if you Ed quantise formats those are\n\nso if you Ed quantise formats those are\n \n\nso if you Ed quantise formats those are\ngoing to be better than ever with Lop\n\ngoing to be better than ever with Lop\n \n\ngoing to be better than ever with Lop\nfile now too and it's worth\n\nfile now too and it's worth\n \n\nfile now too and it's worth\nmentioning that we've um seen really\n\nmentioning that we've um seen really\n \n\nmentioning that we've um seen really\ninteresting things about this like\n\ninteresting things about this like\n \n\ninteresting things about this like\npeople once we put this out into the\n\npeople once we put this out into the\n \n\npeople once we put this out into the\nworld people have come back and given us\n\nworld people have come back and given us\n \n\nworld people have come back and given us\nfeedback and reported like their own\n\nfeedback and reported like their own\n \n\nfeedback and reported like their own\nexperiences we found out that someone\n\nexperiences we found out that someone\n \n\nexperiences we found out that someone\nwas running mial adex 22b on a $350 CPU\n\nwas running mial adex 22b on a $350 CPU\n \n\nwas running mial adex 22b on a $350 CPU\nand to me that's just wonderful um\n\nand to me that's just wonderful um\n \n\nand to me that's just wonderful um\nbecause performance matters but it's not\n\nbecause performance matters but it's not\n \n\nbecause performance matters but it's not\nreally the thing we care about what we\n\nreally the thing we care about what we\n \n\nreally the thing we care about what we\ncare about is\n\ncare about is\n \n\ncare about is\nintelligence and to have the\n\nintelligence and to have the\n \n\nintelligence and to have the\nintelligence you need to run bigger\n\nintelligence you need to run bigger\n \n\nintelligence you need to run bigger\nmodels and RAM is cheap with CPUs for\n\nmodels and RAM is cheap with CPUs for\n \n\nmodels and RAM is cheap with CPUs for\nthe price of a graphics card I put 512\n\nthe price of a graphics card I put 512\n \n\nthe price of a graphics card I put 512\ngigs in my workstation and that means I\n\ngigs in my workstation and that means I\n \n\ngigs in my workstation and that means I\ncan run all the frontier models coming\n\ncan run all the frontier models coming\n \n\ncan run all the frontier models coming\nout and I just have to wait a little\n\nout and I just have to wait a little\n \n\nout and I just have to wait a little\nlonger but I get a much more intelligent\n\nlonger but I get a much more intelligent\n \n\nlonger but I get a much more intelligent\nanswer and the fact that that went from\n\nanswer and the fact that that went from\n \n\nanswer and the fact that that went from\nimpossible to possible for most\n\nimpossible to possible for most\n \n\nimpossible to possible for most\nconsumers is um you know story I want\n\nconsumers is um you know story I want\n \n\nconsumers is um you know story I want\nyou all to\n\n \n\ntell individuals are making a big\n\ntell individuals are making a big\n \n\ntell individuals are making a big\ndifference and you can be a part of that\n\ndifference and you can be a part of that\n \n\ndifference and you can be a part of that\ntoo and I'm going to hand it back to\n\ntoo and I'm going to hand it back to\n \n\ntoo and I'm going to hand it back to\nStephen who can explain what Mozilla can\n\nStephen who can explain what Mozilla can\n \n\nStephen who can explain what Mozilla can\ndo to support you getting involved in\n\ndo to support you getting involved in\n \n\ndo to support you getting involved in\nthat effort thanks\n\nthat effort thanks\n \n\nthat effort thanks\n[Applause]\n\n \n\nJustin thanks a lot for all your efforts\n\nJustin thanks a lot for all your efforts\n \n\nJustin thanks a lot for all your efforts\nso yeah that that's a key message of\n\nso yeah that that's a key message of\n \n\nso yeah that that's a key message of\nthis talk is anyone in this audience you\n\nthis talk is anyone in this audience you\n \n\nthis talk is anyone in this audience you\ndon't have to work for these big giant\n\ndon't have to work for these big giant\n \n\ndon't have to work for these big giant\nlargest in the history Humanity\n\nlargest in the history Humanity\n \n\nlargest in the history Humanity\ncompanies necessarily to have a big\n\ncompanies necessarily to have a big\n \n\ncompanies necessarily to have a big\nimpact there's lots of Headroom here\n\nimpact there's lots of Headroom here\n \n\nimpact there's lots of Headroom here\nthere's lots of Unsolved interesting\n\nthere's lots of Unsolved interesting\n \n\nthere's lots of Unsolved interesting\nproblems in this\n\nproblems in this\n \n\nproblems in this\nspace and we want to get involved in\n\nspace and we want to get involved in\n \n\nspace and we want to get involved in\nhelping so we recently launched a\n\nhelping so we recently launched a\n \n\nhelping so we recently launched a\nprogram called Mozilla Builders and this\n\nprogram called Mozilla Builders and this\n \n\nprogram called Mozilla Builders and this\nis a program by which we either sponsor\n\nis a program by which we either sponsor\n \n\nis a program by which we either sponsor\nor in some cases co-develop impactful\n\nor in some cases co-develop impactful\n \n\nor in some cases co-develop impactful\nopen- Source AI projects Lama file is\n\nopen- Source AI projects Lama file is\n \n\nopen- Source AI projects Lama file is\nactually the first in this program I'm\n\nactually the first in this program I'm\n \n\nactually the first in this program I'm\nhappy to announce today the second which\n\nhappy to announce today the second which\n \n\nhappy to announce today the second which\nis sqlite VEC this is from a developer\n\nis sqlite VEC this is from a developer\n \n\nis sqlite VEC this is from a developer\nnamed Alex gcia Alex is adding Vector\n\nnamed Alex gcia Alex is adding Vector\n \n\nnamed Alex gcia Alex is adding Vector\nsearch capability to SQL light so for\n\nsearch capability to SQL light so for\n \n\nsearch capability to SQL light so for\nsome folks Ines that'll have some\n\nsome folks Ines that'll have some\n \n\nsome folks Ines that'll have some\nobvious implications that are kind of\n\n \n\ncool but just imagine remember that\n\ncool but just imagine remember that\n \n\ncool but just imagine remember that\nlittle modest Raspberry Pi 5 so I\n\nlittle modest Raspberry Pi 5 so I\n \n\nlittle modest Raspberry Pi 5 so I\nimagine now a local llm open llm running\n\nimagine now a local llm open llm running\n \n\nimagine now a local llm open llm running\nprivately on that machine with no\n\nprivately on that machine with no\n \n\nprivately on that machine with no\nnetwork connection connected to your\n\nnetwork connection connected to your\n \n\nnetwork connection connected to your\npersonal private data which you can use\n\npersonal private data which you can use\n \n\npersonal private data which you can use\nwith confidence that it's safe to do rag\n\nwith confidence that it's safe to do rag\n \n\nwith confidence that it's safe to do rag\nand other interesting applications\n\nand other interesting applications\n \n\nand other interesting applications\nthat's the kind of stuff we're talking\n\nthat's the kind of stuff we're talking\n \n\nthat's the kind of stuff we're talking\nabout here we also just launched our own\n\nabout here we also just launched our own\n \n\nabout here we also just launched our own\naccelerator it's called the Mozilla\n\naccelerator it's called the Mozilla\n \n\naccelerator it's called the Mozilla\nBuilders accelerator so we are offering\n\nBuilders accelerator so we are offering\n \n\nBuilders accelerator so we are offering\n100,000 Us in non-dilutive funding for\n\n100,000 Us in non-dilutive funding for\n \n\n100,000 Us in non-dilutive funding for\nopen source projects that Advance the\n\nopen source projects that Advance the\n \n\nopen source projects that Advance the\npromise and potential of local AI so\n\npromise and potential of local AI so\n \n\npromise and potential of local AI so\nthat's AI applications running at the\n\nthat's AI applications running at the\n \n\nthat's AI applications running at the\nedge on user devices these are some of\n\nedge on user devices these are some of\n \n\nedge on user devices these are some of\nthe bullet points of areas we're\n\nthe bullet points of areas we're\n \n\nthe bullet points of areas we're\nparticularly interested in but it's not\n\nparticularly interested in but it's not\n \n\nparticularly interested in but it's not\nan exclusive list and you don't have to\n\nan exclusive list and you don't have to\n \n\nan exclusive list and you don't have to\nnecessarily be building a company to\n\nnecessarily be building a company to\n \n\nnecessarily be building a company to\napply for this\n\napply for this\n \n\napply for this\naccelerator so if you want to learn more\n\naccelerator so if you want to learn more\n \n\naccelerator so if you want to learn more\nabout the accelerator this QR code will\n\nabout the accelerator this QR code will\n \n\nabout the accelerator this QR code will\ntake you there take a picture of that or\n\ntake you there take a picture of that or\n \n\ntake you there take a picture of that or\njust go to future." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nmozilla.org\n\njust go to future. mozilla.org\n \n\njust go to future. mozilla.org\nbuilders and you know Justine and I and\n\nbuilders and you know Justine and I and\n \n\nbuilders and you know Justine and I and\na lot of mozillians are here this week\n\na lot of mozillians are here this week\n \n\na lot of mozilli..." }, { "role": "assistant", "content": "mozilla.org\n\njust go to future. mozilla.org\n \n\njust go to future. mozilla.org\nbuilders and you know Justine and I and\n\nbuilders and you know Justine and I and\n \n\nbuilders and you know Justine and I and\na lot of mozillians are here this week\n\na lot of mozillians are here this week\n \n\na lot of mozillians are here this week\nif you have something you're working on\n\nif you have something you're working on\n \n\nif you have something you're working on\nthere something you think we should know\n\nthere something you think we should know\n \n\nthere something you think we should know\nabout or you want to collaborate with us\n\nabout or you want to collaborate with us\n \n\nabout or you want to collaborate with us\nplease find us reach out or reach out to\n\nplease find us reach out or reach out to\n \n\nplease find us reach out or reach out to\nme via email\n\nme via email\n \n\nme via email\nso thanks again thanks to Justine and\n\nso thanks again thanks to Justine and\n \n\nso thanks again thanks to Justine and\nthe community all their work on llama\n\nthe community all their work on llama\n \n\nthe community all their work on llama\nfile thank you stepen thank you\n\nfile thank you stepen thank you\n \n\nfile thank you stepen thank you\n[Music]\n\n \n\n[Music]" } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\n[Music]\n\n[Music]\n \n\n[Music]\nso a little bit about me um had of a\n\nso a little bit about me um had of a\n \n\nso a little bit about me um had of a\nplaat ramp I've been working on llms for\n\nplaat ramp I've been working on llms for\n \n\nplaat ramp I've been working on llms for\nfour years which\n\nfour years w..." }, { "role": "assistant", "content": "[Music]\n\n[Music]\n \n\n[Music]\nso a little bit about me um had of a\n\nso a little bit about me um had of a\n \n\nso a little bit about me um had of a\nplaat ramp I've been working on llms for\n\nplaat ramp I've been working on llms for\n \n\nplaat ramp I've been working on llms for\nfour years which\n\nfour years which\n \n\nfour years which\nis\n\nis\n \n\nis\nwell which is uh kind of a long time I\n\nwell which is uh kind of a long time I\n \n\nwell which is uh kind of a long time I\nguess uh in LM land everything started\n\nguess uh in LM land everything started\n \n\nguess uh in LM land everything started\nhappening really when chat GPD came out\n\nhappening really when chat GPD came out\n \n\nhappening really when chat GPD came out\num so I was trying to build what people\n\num so I was trying to build what people\n \n\num so I was trying to build what people\nwould Now call an AI agent company back\n\nwould Now call an AI agent company back\n \n\nwould Now call an AI agent company back\nthen we were just doing customer support\n\nthen we were just doing customer support\n \n\nthen we were just doing customer support\nwe're trying to make our chat bot\n\nwe're trying to make our chat bot\n \n\nwe're trying to make our chat bot\nsmarter we're trying to figure out what\n\nsmarter we're trying to figure out what\n \n\nsmarter we're trying to figure out what\nwhat models to use to or what tech to\n\nwhat models to use to or what tech to\n \n\nwhat models to use to or what tech to\nuse to get them to respond to customers\n\nuse to get them to respond to customers\n \n\nuse to get them to respond to customers\nbetter and we were messing with gpd2 on\n\nbetter and we were messing with gpd2 on\n \n\nbetter and we were messing with gpd2 on\nBT and models were so frustratingly\n\nBT and models were so frustratingly\n \n\nBT and models were so frustratingly\nstupid and context windows were small\n\nstupid and context windows were small\n \n\nstupid and context windows were small\nand they were not very smart reasoning\n\nand they were not very smart reasoning\n \n\nand they were not very smart reasoning\nand it was just incredibly annoying and\n\nand it was just incredibly annoying and\n \n\nand it was just incredibly annoying and\nwe just wrote lots of code around these\n\nwe just wrote lots of code around these\n \n\nwe just wrote lots of code around these\nmodels to get them to work at least\n\nmodels to get them to work at least\n \n\nmodels to get them to work at least\nsomewhat\n\nsomewhat\n \n\nsomewhat\nreliably and along the way as models got\n\nreliably and along the way as models got\n \n\nreliably and along the way as models got\nsmarter this kind of had to delete more\n\nsmarter this kind of had to delete more\n \n\nsmarter this kind of had to delete more\nof that code and this ended up seeing a\n\nof that code and this ended up seeing a\n \n\nof that code and this ended up seeing a\nlot of patterns in what what code needs\n\nlot of patterns in what what code needs\n \n\nlot of patterns in what what code needs\nto get deleted how to build agents in\n\nto get deleted how to build agents in\n \n\nto get deleted how to build agents in\nwhat ways that will scale with more\n\nwhat ways that will scale with more\n \n\nwhat ways that will scale with more\nintelligence and clearly we're going to\n\nintelligence and clearly we're going to\n \n\nintelligence and clearly we're going to\ncontinue to get a lot more intelligence\n\ncontinue to get a lot more intelligence\n \n\ncontinue to get a lot more intelligence\nand I just wanted to uh maybe talk about\n\nand I just wanted to uh maybe talk about\n \n\nand I just wanted to uh maybe talk about\na single idea throughout the talk uh\n\na single idea throughout the talk uh\n \n\na single idea throughout the talk uh\nthrough various examples uh we we'll do\n\nthrough various examples uh we we'll do\n \n\nthrough various examples uh we we'll do\nsome uh some uh setting but I'll also\n\nsome uh some uh setting but I'll also\n \n\nsome uh some uh setting but I'll also\nhave a bunch of demos to kind of like\n\nhave a bunch of demos to kind of like\n \n\nhave a bunch of demos to kind of like\ndrive home the point and maybe it can\n\ndrive home the point and maybe it can\n \n\ndrive home the point and maybe it can\nconvince you guys that uh there's a\n\nconvince you guys that uh there's a\n \n\nconvince you guys that uh there's a\ncertain way of building agents that's\n\ncertain way of building agents that's\n \n\ncertain way of building agents that's\nslightly better than other\n\n \n\nways I also built a structur extraction\n\nways I also built a structur extraction\n \n\nways I also built a structur extraction\nLibrary called Json former um I think it\n\nLibrary called Json former um I think it\n \n\nLibrary called Json former um I think it\nwas the first one I don't I'm not fully\n\nwas the first one I don't I'm not fully\n \n\nwas the first one I don't I'm not fully\nsure but timing wise it was before all\n\nsure but timing wise it was before all\n \n\nsure but timing wise it was before all\nthe other major ones um and that was\n\nthe other major ones um and that was\n \n\nthe other major ones um and that was\nalso scaffolding around a model models\n\nalso scaffolding around a model models\n \n\nalso scaffolding around a model models\nwere too stupid to up with Json and we\n\nwere too stupid to up with Json and we\n \n\nwere too stupid to up with Json and we\nwere just really begging it pleading\n\nwere just really begging it pleading\n \n\nwere just really begging it pleading\npleading it and forcing it to uh in ways\n\npleading it and forcing it to uh in ways\n \n\npleading it and forcing it to uh in ways\nthat we want it to\n\nthat we want it to\n \n\nthat we want it to\nbe so as I said earlier just have a one\n\nbe so as I said earlier just have a one\n \n\nbe so as I said earlier just have a one\ncore agenda item here which is want to\n\ncore agenda item here which is want to\n \n\ncore agenda item here which is want to\nconvey one idea uh we'll start off all\n\nconvey one idea uh we'll start off all\n \n\nconvey one idea uh we'll start off all\nof you probably read the essay bit or\n\nof you probably read the essay bit or\n \n\nof you probably read the essay bit or\nlesson just quickly go through what it\n\nlesson just quickly go through what it\n \n\nlesson just quickly go through what it\nis uh we'll go through a production\n\nis uh we'll go through a production\n \n\nis uh we'll go through a production\nagent we have at ramp and how it works\n\nagent we have at ramp and how it works\n \n\nagent we have at ramp and how it works\nand three different ways of architecting\n\nand three different ways of architecting\n \n\nand three different ways of architecting\nit and then I have a demo that to really\n\nit and then I have a demo that to really\n \n\nit and then I have a demo that to really\npush maybe how we all think about how\n\npush maybe how we all think about how\n \n\npush maybe how we all think about how\nsoftware and backends and things will\n\nsoftware and backends and things will\n \n\nsoftware and backends and things will\nwork in the\n\nwork in the\n \n\nwork in the\nfuture so very simply the idea is just\n\nfuture so very simply the idea is just\n \n\nfuture so very simply the idea is just\nthat systems that scale with compute\n\nthat systems that scale with compute\n \n\nthat systems that scale with compute\nbeat systems that don't so there's two\n\nbeat systems that don't so there's two\n \n\nbeat systems that don't so there's two\nsystems and uh without any effort the\n\nsystems and uh without any effort the\n \n\nsystems and uh without any effort the\nthe system one of the systems can just\n\nthe system one of the systems can just\n \n\nthe system one of the systems can just\nthink more or use more compute in some\n\nthink more or use more compute in some\n \n\nthink more or use more compute in some\nway that system tends to be systems that\n\nway that system tends to be systems that\n \n\nway that system tends to be systems that\nare rigid and fixed and just\n\nare rigid and fixed and just\n \n\nare rigid and fixed and just\ndeterministic so from that idea it's\n\ndeterministic so from that idea it's\n \n\ndeterministic so from that idea it's\npretty clear like if you're Building\n\npretty clear like if you're Building\n \n\npretty clear like if you're Building\nSystems you might as well build systems\n\nSystems you might as well build systems\n \n\nSystems you might as well build systems\nthat improve with more compute and this\n\nthat improve with more compute and this\n \n\nthat improve with more compute and this\nthis seems pretty obvious like obvious\n\nthis seems pretty obvious like obvious\n \n\nthis seems pretty obvious like obvious\nconclusion from the B lesson taking it a\n\nconclusion from the B lesson taking it a\n \n\nconclusion from the B lesson taking it a\nstep further why is this true it's\n\nstep further why is this true it's\n \n\nstep further why is this true it's\nbecause exponentials are rare like they\n\nbecause exponentials are rare like they\n \n\nbecause exponentials are rare like they\njust don't exist most things in the\n\njust don't exist most things in the\n \n\njust don't exist most things in the\nworld aren't exponential so when you\n\nworld aren't exponential so when you\n \n\nworld aren't exponential so when you\nfind one you just should hop on strap on\n\nfind one you just should hop on strap on\n \n\nfind one you just should hop on strap on\njust take the free pass and go for the\n\njust take the free pass and go for the\n \n\njust take the free pass and go for the\nright and probably shouldn't try too\n\nright and probably shouldn't try too\n \n\nright and probably shouldn't try too\nhard and there's a lot of examples from\n\nhard and there's a lot of examples from\n \n\nhard and there's a lot of examples from\nhistory that um to kind of reflect this\n\nhistory that um to kind of reflect this\n \n\nhistory that um to kind of reflect this\nso for for chess and go and computer\n\nso for for chess and go and computer\n \n\nso for for chess and go and computer\nvision Atari games like people have\n\nvision Atari games like people have\n \n\nvision Atari games like people have\ntried to build lots of systems and\n\ntried to build lots of systems and\n \n\ntried to build lots of systems and\nwritten a lot of code and my way of\n\nwritten a lot of code and my way of\n \n\nwritten a lot of code and my way of\nthinking about rigid systems just like\n\nthinking about rigid systems just like\n \n\nthinking about rigid systems just like\nspending a lot of time grinding weekends\n\nspending a lot of time grinding weekends\n \n\nspending a lot of time grinding weekends\nand writing very clever software well\n\nand writing very clever software well\n \n\nand writing very clever software well\nabstracted uh maybe trying to synthesize\n\nabstracted uh maybe trying to synthesize\n \n\nabstracted uh maybe trying to synthesize\nhuman reasoning and thought process into\n\nhuman reasoning and thought process into\n \n\nhuman reasoning and thought process into\nfeatures and then using them in clever\n\nfeatures and then using them in clever\n \n\nfeatures and then using them in clever\nways and trying to approximate how a\n\nways and trying to approximate how a\n \n\nways and trying to approximate how a\nhuman would\n\nhuman would\n \n\nhuman would\nthink and if you actually fix the amount\n\nthink and if you actually fix the amount\n \n\nthink and if you actually fix the amount\nof compute that approach will win but if\n\nof compute that approach will win but if\n \n\nof compute that approach will win but if\nyou just turns out if you end up scaling\n\nyou just turns out if you end up scaling\n \n\nyou just turns out if you end up scaling\nout how much search you're doing the\n\nout how much search you're doing the\n \n\nout how much search you're doing the\ngeneral method always ends up winning\n\ngeneral method always ends up winning\n \n\ngeneral method always ends up winning\neven uh like in all these cases so Atari\n\neven uh like in all these cases so Atari\n \n\neven uh like in all these cases so Atari\ngo and computer\n\ngo and computer\n \n\ngo and computer\nvision a little bit about ramp so ramp\n\nvision a little bit about ramp so ramp\n \n\nvision a little bit about ramp so ramp\nis a finance platform that helps\n\nis a finance platform that helps\n \n\nis a finance platform that helps\nbusinesses manage expenses payments\n\nbusinesses manage expenses payments\n \n\nbusinesses manage expenses payments\nprocurement travel bookkeeping more\n\nprocurement travel bookkeeping more\n \n\nprocurement travel bookkeeping more\nefficiently and we have a ton of AI\n\nefficiently and we have a ton of AI\n \n\nefficiently and we have a ton of AI\nacross the product so automate a lot of\n\nacross the product so automate a lot of\n \n\nacross the product so automate a lot of\nboring stuff the finance teams do and\n\nboring stuff the finance teams do and\n \n\nboring stuff the finance teams do and\nemployees do with uh submitting expense\n\nemployees do with uh submitting expense\n \n\nemployees do with uh submitting expense\nreports and booking your flights and\n\nreports and booking your flights and\n \n\nreports and booking your flights and\nhotels and uh submitting reimbursements\n\nhotels and uh submitting reimbursements\n \n\nhotels and uh submitting reimbursements\nall that and so a lot of the work behind\n\nall that and so a lot of the work behind\n \n\nall that and so a lot of the work behind\nthe scenes is just we're interacting\n\nthe scenes is just we're interacting\n \n\nthe scenes is just we're interacting\nwith other systems um and helping like\n\nwith other systems um and helping like\n \n\nwith other systems um and helping like\nLegacy systems and helping employees get\n\nLegacy systems and helping employees get\n \n\nLegacy systems and helping employees get\ntheir work done\n\ntheir work done\n \n\ntheir work done\nfaster so let's actually talk through\n\nfaster so let's actually talk through\n \n\nfaster so let's actually talk through\none of the systems we have today at ramp\n\none of the systems we have today at ramp\n \n\none of the systems we have today at ramp\nand um maybe some talk through the\n\nand um maybe some talk through the\n \n\nand um maybe some talk through the\ndifferent versions of the system and how\n\ndifferent versions of the system and how\n \n\ndifferent versions of the system and how\nit evolved over\n\nit evolved over\n \n\nit evolved over\ntime so we're going to talk about\n\ntime so we're going to talk about\n \n\ntime so we're going to talk about\nsomething called a switching report it's\n\nsomething called a switching report it's\n \n\nsomething called a switching report it's\nvery simple agent all it needs to do\n\nvery simple agent all it needs to do\n \n\nvery simple agent all it needs to do\nit's taking a CSV a CSV arbitrary format\n\nit's taking a CSV a CSV arbitrary format\n \n\nit's taking a CSV a CSV arbitrary format\nso the schema could be seriously\n\nso the schema could be seriously\n \n\nso the schema could be seriously\nanything from the internet and we want\n\nanything from the internet and we want\n \n\nanything from the internet and we want\nthese csvs to come from third party card\n\nthese csvs to come from third party card\n \n\nthese csvs to come from third party card\nproviders so when people on board to\n\nproviders so when people on board to\n \n\nproviders so when people on board to\nramp we want to give them a nice\n\nramp we want to give them a nice\n \n\nramp we want to give them a nice\nchecklist and say hey here are all the\n\nchecklist and say hey here are all the\n \n\nchecklist and say hey here are all the\ntransactions you have on other platforms\n\ntransactions you have on other platforms\n \n\ntransactions you have on other platforms\nand we want help you move them over and\n\nand we want help you move them over and\n \n\nand we want help you move them over and\nthe more transactions come on ramp the\n\nthe more transactions come on ramp the\n \n\nthe more transactions come on ramp the\nmore we can help you and the more you'll\n\nmore we can help you and the more you'll\n \n\nmore we can help you and the more you'll\nuse our software and more everyone\n\nuse our software and more everyone\n \n\nuse our software and more everyone\nbenefits and so the switching reports\n\nbenefits and so the switching reports\n \n\nbenefits and so the switching reports\njust really a checklist but to read\n\njust really a checklist but to read\n \n\njust really a checklist but to read\npeople's CSV transactions we need to\n\npeople's CSV transactions we need to\n \n\npeople's CSV transactions we need to\nunderstand those and other platforms\n\nunderstand those and other platforms\n \n\nunderstand those and other platforms\nhave all these kinds of crazy\n\nhave all these kinds of crazy\n \n\nhave all these kinds of crazy\nschemas and so the the description of\n\nschemas and so the the description of\n \n\nschemas and so the the description of\nthe problem we have here is just for an\n\nthe problem we have here is just for an\n \n\nthe problem we have here is just for an\narbitrary arbitrary like CSV how can we\n\narbitrary arbitrary like CSV how can we\n \n\narbitrary arbitrary like CSV how can we\nsupport um parsing it and then into some\n\nsupport um parsing it and then into some\n \n\nsupport um parsing it and then into some\nformat that we we understand\n\nformat that we we understand\n \n\nformat that we we understand\nso let's just start with the the simple\n\nso let's just start with the the simple\n \n\nso let's just start with the the simple\napproach right is like let's just take\n\napproach right is like let's just take\n \n\napproach right is like let's just take\nthe 50 most common third party card\n\nthe 50 most common third party card\n \n\nthe 50 most common third party card\nvendors um and just manually write code\n\nvendors um and just manually write code\n \n\nvendors um and just manually write code\nfor all of them now obviously like this\n\nfor all of them now obviously like this\n \n\nfor all of them now obviously like this\nthis will just work it is some work not\n\nthis will just work it is some work not\n \n\nthis will just work it is some work not\na lot of work but you still have to\n\na lot of work but you still have to\n \n\na lot of work but you still have to\nmaybe go to 50 different platforms and\n\nmaybe go to 50 different platforms and\n \n\nmaybe go to 50 different platforms and\ndownload their csvs see what schemas\n\ndownload their csvs see what schemas\n \n\ndownload their csvs see what schemas\nthey have and then write code maybe if\n\nthey have and then write code maybe if\n \n\nthey have and then write code maybe if\nthey decide one day they change their\n\nthey decide one day they change their\n \n\nthey decide one day they change their\nformat your thing will break but that's\n\nformat your thing will break but that's\n \n\nformat your thing will break but that's\nokay you'll get page and you can wake up\n\nokay you'll get page and you can wake up\n \n\nokay you'll get page and you can wake up\nand go fix it\n\nand go fix it\n \n\nand go fix it\nso let's maybe introduce some LMS in\n\nso let's maybe introduce some LMS in\n \n\nso let's maybe introduce some LMS in\nhere so from the over engineered code\n\nhere so from the over engineered code\n \n\nhere so from the over engineered code\nwhere you ended up writing 100,000 lines\n\nwhere you ended up writing 100,000 lines\n \n\nwhere you ended up writing 100,000 lines\nmaybe we don't we don't we want a more\n\nmaybe we don't we don't we want a more\n \n\nmaybe we don't we don't we want a more\nGeneral system so let's introduce a\n\nGeneral system so let's introduce a\n \n\nGeneral system so let's introduce a\nlittle bit of LM a little bit of AI in\n\nlittle bit of LM a little bit of AI in\n \n\nlittle bit of LM a little bit of AI in\nhere and so in the deterministic flow\n\nhere and so in the deterministic flow\n \n\nhere and so in the deterministic flow\nlet's maybe add some or just like\n\nlet's maybe add some or just like\n \n\nlet's maybe add some or just like\nscripting In classical scripting land\n\nscripting In classical scripting land\n \n\nscripting In classical scripting land\nlet's add some more um call to open AI\n\nlet's add some more um call to open AI\n \n\nlet's add some more um call to open AI\nor you have an embedding model you want\n\nor you have an embedding model you want\n \n\nor you have an embedding model you want\nto do semetic similarity or something\n\nto do semetic similarity or something\n \n\nto do semetic similarity or something\nlike that so then let's just take every\n\nlike that so then let's just take every\n \n\nlike that so then let's just take every\ncolumn in the CSV that comes in let's\n\ncolumn in the CSV that comes in let's\n \n\ncolumn in the CSV that comes in let's\ntry to classify what kind of column it\n\ntry to classify what kind of column it\n \n\ntry to classify what kind of column it\nis is it a date is it a transaction uh\n\nis is it a date is it a transaction uh\n \n\nis is it a date is it a transaction uh\nis it a transaction amount is it a\n\nis it a transaction amount is it a\n \n\nis it a transaction amount is it a\nmerchant name or is it the uh users's\n\nmerchant name or is it the uh users's\n \n\nmerchant name or is it the uh users's\nname and then we map it on and then we\n\nname and then we map it on and then we\n \n\nname and then we map it on and then we\nprobably could uh end up in a schema\n\nprobably could uh end up in a schema\n \n\nprobably could uh end up in a schema\nthat we're happy with again most of the\n\nthat we're happy with again most of the\n \n\nthat we're happy with again most of the\ncompute is running in classical land\n\ncompute is running in classical land\n \n\ncompute is running in classical land\nsome of it is running in fuzzy llm land\n\nsome of it is running in fuzzy llm land\n \n\nsome of it is running in fuzzy llm land\nbut this is somewhat looking like a more\n\nbut this is somewhat looking like a more\n \n\nbut this is somewhat looking like a more\nGeneral\n\nGeneral\n \n\nGeneral\nsystem let's go maybe a different\n\nsystem let's go maybe a different\n \n\nsystem let's go maybe a different\napproach when like we just go all the\n\napproach when like we just go all the\n \n\napproach when like we just go all the\nway through let's just say we're just\n\nway through let's just say we're just\n \n\nway through let's just say we're just\ngoing to literally give the CSV to LM\n\ngoing to literally give the CSV to LM\n \n\ngoing to literally give the CSV to LM\nand say you have a code interpreter so\n\nand say you have a code interpreter so\n \n\nand say you have a code interpreter so\nyou can write whatever code you want\n\nyou can write whatever code you want\n \n\nyou can write whatever code you want\npandas or all the faster rust based ones\n\npandas or all the faster rust based ones\n \n\npandas or all the faster rust based ones\num you have all these python packages um\n\num you have all these python packages um\n \n\num you have all these python packages um\nyou're allowed to look at the head of\n\nyou're allowed to look at the head of\n \n\nyou're allowed to look at the head of\nthe CSV the tail whichever rows you want\n\nthe CSV the tail whichever rows you want\n \n\nthe CSV the tail whichever rows you want\num and then I just want you to give me a\n\num and then I just want you to give me a\n \n\num and then I just want you to give me a\nCSV uh with this specific format here's\n\nCSV uh with this specific format here's\n \n\nCSV uh with this specific format here's\na unit test here's a verifier that you\n\na unit test here's a verifier that you\n \n\na unit test here's a verifier that you\ncan use to tell if it's working or not\n\ncan use to tell if it's working or not\n \n\ncan use to tell if it's working or not\nturns out this approach actually doesn't\n\nturns out this approach actually doesn't\n \n\nturns out this approach actually doesn't\nwork like we tried it um if you only run\n\nwork like we tried it um if you only run\n \n\nwork like we tried it um if you only run\nit once but instead if you run it 50\n\nit once but instead if you run it 50\n \n\nit once but instead if you run it 50\ntimes in parallel it's actually very\n\ntimes in parallel it's actually very\n \n\ntimes in parallel it's actually very\nlikely that it works really well and\n\nlikely that it works really well and\n \n\nlikely that it works really well and\ngeneralizes across a ton of different\n\ngeneralizes across a ton of different\n \n\ngeneralizes across a ton of different\nformats the amount of compute here is\n\nformats the amount of compute here is\n \n\nformats the amount of compute here is\nactually probably like what is that\n\nactually probably like what is that\n \n\nactually probably like what is that\nnumber 10,000 times more than the the\n\nnumber 10,000 times more than the the\n \n\nnumber 10,000 times more than the the\nfirst approach we came up with but again\n\nfirst approach we came up with but again\n \n\nfirst approach we came up with but again\nlike what is truly scarce in the world\n\nlike what is truly scarce in the world\n \n\nlike what is truly scarce in the world\nis engineer time maybe not for not in a\n\nis engineer time maybe not for not in a\n \n\nis engineer time maybe not for not in a\nwhile but at least today and we'd rather\n\nwhile but at least today and we'd rather\n \n\nwhile but at least today and we'd rather\nhave a system that works really well and\n\nhave a system that works really well and\n \n\nhave a system that works really well and\neven with a 10,000 times more compute it\n\neven with a 10,000 times more compute it\n \n\neven with a 10,000 times more compute it\nwill probably cost less than a dollar\n\nwill probably cost less than a dollar\n \n\nwill probably cost less than a dollar\nand every transaction that switched over\n\nand every transaction that switched over\n \n\nand every transaction that switched over\nevery fail CSV will cost ramp way more\n\nevery fail CSV will cost ramp way more\n \n\nevery fail CSV will cost ramp way more\nmoney than whatever money we spend on\n\nmoney than whatever money we spend on\n \n\nmoney than whatever money we spend on\nthis exact this exact\n\nthis exact this exact\n \n\nthis exact this exact\narchitecture so this is a very specific\n\narchitecture so this is a very specific\n \n\narchitecture so this is a very specific\nuh example it's like how does this apply\n\nuh example it's like how does this apply\n \n\nuh example it's like how does this apply\nto the agents that we all build and\n\nto the agents that we all build and\n \n\nto the agents that we all build and\nmaybe the systems we're all working on\n\nmaybe the systems we're all working on\n \n\nmaybe the systems we're all working on\nturns out something like this actually\n\nturns out something like this actually\n \n\nturns out something like this actually\ngeneralizes so if you look at three\n\ngeneralizes so if you look at three\n \n\ngeneralizes so if you look at three\napproaches and let's assume like The\n\napproaches and let's assume like The\n \n\napproaches and let's assume like The\nBlack Arrow is just classical compute\n\nBlack Arrow is just classical compute\n \n\nBlack Arrow is just classical compute\nand then the blue arrows are fuzzy lands\n\nand then the blue arrows are fuzzy lands\n \n\nand then the blue arrows are fuzzy lands\nso it goes into neuron net and all all\n\nso it goes into neuron net and all all\n \n\nso it goes into neuron net and all all\nsort of weird matric multiplication\n\nsort of weird matric multiplication\n \n\nsort of weird matric multiplication\nhappens and we're in latent space and\n\nhappens and we're in latent space and\n \n\nhappens and we're in latent space and\ngets all alien intelligency and then\n\ngets all alien intelligency and then\n \n\ngets all alien intelligency and then\ncomes back to a classical land first\n\ncomes back to a classical land first\n \n\ncomes back to a classical land first\napproach there was no AI we just wrote\n\napproach there was no AI we just wrote\n \n\napproach there was no AI we just wrote\ncode and it just worked mostly the\n\ncode and it just worked mostly the\n \n\ncode and it just worked mostly the\nconstrainted agent so the second\n\nconstrainted agent so the second\n \n\nconstrainted agent so the second\napproach we broke into fuzzy land from\n\napproach we broke into fuzzy land from\n \n\napproach we broke into fuzzy land from\nclassic land when when we decided we\n\nclassic land when when we decided we\n \n\nclassic land when when we decided we\nwanted similarity scores or something\n\nwanted similarity scores or something\n \n\nwanted similarity scores or something\nlike that and then the third approach is\n\nlike that and then the third approach is\n \n\nlike that and then the third approach is\nactually flipped where the llm decides\n\nactually flipped where the llm decides\n \n\nactually flipped where the llm decides\nit needs to go into classical l so it\n\nit needs to go into classical l so it\n \n\nit needs to go into classical l so it\nwrite some code write some pandas or uh\n\nwrite some code write some pandas or uh\n \n\nwrite some code write some pandas or uh\npython code and it decides to break in\n\npython code and it decides to break in\n \n\npython code and it decides to break in\ninto this classical land when it needs\n\ninto this classical land when it needs\n \n\ninto this classical land when it needs\nto but most of the computer is\n\nto but most of the computer is\n \n\nto but most of the computer is\nfuzzy actually this is maybe not the\n\nfuzzy actually this is maybe not the\n \n\nfuzzy actually this is maybe not the\nmost accurate graph like because I\n\nmost accurate graph like because I\n \n\nmost accurate graph like because I\npropos that we run it 50 times it more\n\npropos that we run it 50 times it more\n \n\npropos that we run it 50 times it more\nso looks like this but if you look at a\n\nso looks like this but if you look at a\n \n\nso looks like this but if you look at a\nback end in general they're all request\n\nback end in general they're all request\n \n\nback end in general they're all request\nresponse so some sort of message is\n\nresponse so some sort of message is\n \n\nresponse so some sort of message is\ngoing in it's like a post request or get\n\ngoing in it's like a post request or get\n \n\ngoing in it's like a post request or get\nor update or read any sort of credit\n\nor update or read any sort of credit\n \n\nor update or read any sort of credit\noperation and we're really just asking\n\noperation and we're really just asking\n \n\noperation and we're really just asking\nthe back end to take this piece of\n\nthe back end to take this piece of\n \n\nthe back end to take this piece of\ninformation do whatever you must with it\n\ninformation do whatever you must with it\n \n\ninformation do whatever you must with it\nrun out whatever mutations you want and\n\nrun out whatever mutations you want and\n \n\nrun out whatever mutations you want and\nreturn me a\n\nreturn me a\n \n\nreturn me a\nresponse and almost all systems we've\n\nresponse and almost all systems we've\n \n\nresponse and almost all systems we've\nbuilt so far as like Humanity I guess\n\nbuilt so far as like Humanity I guess\n \n\nbuilt so far as like Humanity I guess\nlike look like the first one but more\n\nlike look like the first one but more\n \n\nlike look like the first one but more\npeople are using open AI open AI makes\n\npeople are using open AI open AI makes\n \n\npeople are using open AI open AI makes\nbillions of dollars and probably a lot\n\nbillions of dollars and probably a lot\n \n\nbillions of dollars and probably a lot\nof the systems that use them look like\n\nof the systems that use them look like\n \n\nof the systems that use them look like\nnumber two where just regular uh\n\nnumber two where just regular uh\n \n\nnumber two where just regular uh\nprogramming languages are calling into\n\nprogramming languages are calling into\n \n\nprogramming languages are calling into\nopen AI servers and we're running some\n\nopen AI servers and we're running some\n \n\nopen AI servers and we're running some\nfuzzy\n\nfuzzy\n \n\nfuzzy\ncompute what we're seeing in like more\n\ncompute what we're seeing in like more\n \n\ncompute what we're seeing in like more\nand more parts of the ram codebase we're\n\nand more parts of the ram codebase we're\n \n\nand more parts of the ram codebase we're\nmoving to the third approach because it\n\nmoving to the third approach because it\n \n\nmoving to the third approach because it\njust tends to work well because all the\n\njust tends to work well because all the\n \n\njust tends to work well because all the\nblue arrows if you did nothing it Absol\n\nblue arrows if you did nothing it Absol\n \n\nblue arrows if you did nothing it Absol\nabsolutely nothing we all went to\n\nabsolutely nothing we all went to\n \n\nabsolutely nothing we all went to\nvacation for the next year the big labs\n\nvacation for the next year the big labs\n \n\nvacation for the next year the big labs\nare still working and spending billions\n\nare still working and spending billions\n \n\nare still working and spending billions\nof dollars making those models better so\n\nof dollars making those models better so\n \n\nof dollars making those models better so\nthe blue arrows will get better and so\n\nthe blue arrows will get better and so\n \n\nthe blue arrows will get better and so\nhow much blue arrow you're using in your\n\nhow much blue arrow you're using in your\n \n\nhow much blue arrow you're using in your\ncode base actually will help directly\n\ncode base actually will help directly\n \n\ncode base actually will help directly\nyour company without much effort from\n\nyour company without much effort from\n \n\nyour company without much effort from\nyour end so this is what I was saying is\n\nyour end so this is what I was saying is\n \n\nyour end so this is what I was saying is\nlike the bit or lesson is just so\n\nlike the bit or lesson is just so\n \n\nlike the bit or lesson is just so\npowerful and exponential Trends are so\n\npowerful and exponential Trends are so\n \n\npowerful and exponential Trends are so\npowerful that you can just hitch hitch\n\npowerful that you can just hitch hitch\n \n\npowerful that you can just hitch hitch\nthe\n\n \n\nride let's um let's take this idea like\n\nride let's um let's take this idea like\n \n\nride let's um let's take this idea like\nfurther um let's actually like go all\n\nfurther um let's actually like go all\n \n\nfurther um let's actually like go all\nthe way like something something\n\n \n\ncrazy um on the left you'll see a\n\ncrazy um on the left you'll see a\n \n\ncrazy um on the left you'll see a\ntraditional web app so usually the way\n\ntraditional web app so usually the way\n \n\ntraditional web app so usually the way\nit works is you open um\n\nit works is you open um\n \n\nit works is you open um\ngmail.com and some uh static file server\n\ngmail.com and some uh static file server\n \n\ngmail.com and some uh static file server\nand Google sending you bunch of\n\nand Google sending you bunch of\n \n\nand Google sending you bunch of\nJavaScript HTML and CSS your browser\n\nJavaScript HTML and CSS your browser\n \n\nJavaScript HTML and CSS your browser\nrenders that um and shows you some nice\n\nrenders that um and shows you some nice\n \n\nrenders that um and shows you some nice\nUI nice HTML that's user friendly maybe\n\nUI nice HTML that's user friendly maybe\n \n\nUI nice HTML that's user friendly maybe\nyou see some emails maybe you click on\n\nyou see some emails maybe you click on\n \n\nyou see some emails maybe you click on\none of them um the friend end makes a\n\none of them um the friend end makes a\n \n\none of them um the friend end makes a\nrequest to the back end the back ask the\n\nrequest to the back end the back ask the\n \n\nrequest to the back end the back ask the\nfriend end friend end asks the back end\n\nfriend end friend end asks the back end\n \n\nfriend end friend end asks the back end\ngive me the content for email and\n\ngive me the content for email and\n \n\ngive me the content for email and\nwhatever ID it is and then the back in\n\nwhatever ID it is and then the back in\n \n\nwhatever ID it is and then the back in\nhits database and gives you the result\n\nhits database and gives you the result\n \n\nhits database and gives you the result\nand maybe they use cod genen maybe they\n\nand maybe they use cod genen maybe they\n \n\nand maybe they use cod genen maybe they\nuse all the Coden tools available to\n\nuse all the Coden tools available to\n \n\nuse all the Coden tools available to\nmake Gmail so that that was probably the\n\nmake Gmail so that that was probably the\n \n\nmake Gmail so that that was probably the\nLM only worked when the software\n\nLM only worked when the software\n \n\nLM only worked when the software\nengineer was writing the code but once\n\nengineer was writing the code but once\n \n\nengineer was writing the code but once\nthe code is written and it's like pushed\n\nthe code is written and it's like pushed\n \n\nthe code is written and it's like pushed\nto production it's just classical\n\nto production it's just classical\n \n\nto production it's just classical\ncompute and on the right I'm actually\n\ncompute and on the right I'm actually\n \n\ncompute and on the right I'm actually\nproposing a different model which is the\n\nproposing a different model which is the\n \n\nproposing a different model which is the\nback end is the LM it's not Cen it's\n\nback end is the LM it's not Cen it's\n \n\nback end is the LM it's not Cen it's\nthis LM is doing the execution it is the\n\nthis LM is doing the execution it is the\n \n\nthis LM is doing the execution it is the\nback end so the LM has access to tools\n\nback end so the LM has access to tools\n \n\nback end so the LM has access to tools\nlike coder interpreter and potentially\n\nlike coder interpreter and potentially\n \n\nlike coder interpreter and potentially\nhas access to um through that making\n\nhas access to um through that making\n \n\nhas access to um through that making\nrequest Network requests and also has an\n\nrequest Network requests and also has an\n \n\nrequest Network requests and also has an\naccess to uh DB\n\naccess to uh DB\n \n\naccess to uh DB\nso I have a mail client actually that\n\nso I have a mail client actually that\n \n\nso I have a mail client actually that\nworks with this principal and this is my\n\nworks with this principal and this is my\n \n\nworks with this principal and this is my\ntest email so if y'all want to see any\n\ntest email so if y'all want to see any\n \n\ntest email so if y'all want to see any\nemails you sent to me in a minute or so\n\nemails you sent to me in a minute or so\n \n\nemails you sent to me in a minute or so\nyou can send me an email but please be\n\nyou can send me an email but please be\n \n\nyou can send me an email but please be\nnice\n\n \n\nall right I think um it's probably\n\nall right I think um it's probably\n \n\nall right I think um it's probably\nenough\n\nenough\n \n\nenough\ntime so I'm going to go\n\n \n\nover so we have this email client I mean\n\nover so we have this email client I mean\n \n\nover so we have this email client I mean\nwe still have some regular JavaScript to\n\nwe still have some regular JavaScript to\n \n\nwe still have some regular JavaScript to\nhook into the LM hook the LM into the\n\nhook into the LM hook the LM into the\n \n\nhook into the LM hook the LM into the\nbrowser but when I do log in I'm going\n\nbrowser but when I do log in I'm going\n \n\nbrowser but when I do log in I'm going\nto use my email is just showing you\n\n \n\n[Music]\n\n \n\nwhat oh it's\n\n \n\nprobably okay we're good we're good all\n\nprobably okay we're good we're good all\n \n\nprobably okay we're good we're good all\nright we're\n\nright we're\n \n\nright we're\nsaved I\n\n \n\nthink thankfully I have a room full of\n\nthink thankfully I have a room full of\n \n\nthink thankfully I have a room full of\nEngineers\n\n \n\nso there's a dot but the reason it's so\n\nso there's a dot but the reason it's so\n \n\nso there's a dot but the reason it's so\nslow is because when I open this page\n\nslow is because when I open this page\n \n\nslow is because when I open this page\nand log into Gmail the Gmail token is\n\nand log into Gmail the Gmail token is\n \n\nand log into Gmail the Gmail token is\nactually being sent to an llm we're\n\nactually being sent to an llm we're\n \n\nactually being sent to an llm we're\nsaying literally this is a LM chat\n\nsaying literally this is a LM chat\n \n\nsaying literally this is a LM chat\nsession what we're we're seeing on the\n\nsession what we're we're seeing on the\n \n\nsession what we're we're seeing on the\nscreen is like hey LM you're you're\n\nscreen is like hey LM you're you're\n \n\nscreen is like hey LM you're you're\nactually simulating a Gmail client you\n\nactually simulating a Gmail client you\n \n\nactually simulating a Gmail client you\nhave access to Oh all the emails you\n\nhave access to Oh all the emails you\n \n\nhave access to Oh all the emails you\nhave access to um rahul's Gmail token\n\nhave access to um rahul's Gmail token\n \n\nhave access to um rahul's Gmail token\nand aord interpreter and so just render\n\nand aord interpreter and so just render\n \n\nand aord interpreter and so just render\nsome UI based on uh what you think is\n\nsome UI based on uh what you think is\n \n\nsome UI based on uh what you think is\nreasonable for the homepage for Gmail\n\nreasonable for the homepage for Gmail\n \n\nreasonable for the homepage for Gmail\nclient and so looks like it decided to\n\nclient and so looks like it decided to\n \n\nclient and so looks like it decided to\nrender his markdown uh I think we\n\nrender his markdown uh I think we\n \n\nrender his markdown uh I think we\nactually tell it to render his markdown\n\nactually tell it to render his markdown\n \n\nactually tell it to render his markdown\nand it's rendering all the emails that a\n\nand it's rendering all the emails that a\n \n\nand it's rendering all the emails that a\nbunch of people sent me from here so it\n\nbunch of people sent me from here so it\n \n\nbunch of people sent me from here so it\nlooks like it says uh hello from\n\nlooks like it says uh hello from\n \n\nlooks like it says uh hello from\nCalifornia so I'm going to click on that\n\nCalifornia so I'm going to click on that\n \n\nCalifornia so I'm going to click on that\nwhen I click on that we're actually not\n\nwhen I click on that we're actually not\n \n\nwhen I click on that we're actually not\nrunning um any like backend calls or\n\nrunning um any like backend calls or\n \n\nrunning um any like backend calls or\nanything like that we're just telling\n\nanything like that we're just telling\n \n\nanything like that we're just telling\nthe LM the user clicked on that piece of\n\nthe LM the user clicked on that piece of\n \n\nthe LM the user clicked on that piece of\ntext in this case it was hello from\n\ntext in this case it was hello from\n \n\ntext in this case it was hello from\nCalifornia and the ID number so the LM\n\nCalifornia and the ID number so the LM\n \n\nCalifornia and the ID number so the LM\nnow has the information on what the user\n\nnow has the information on what the user\n \n\nnow has the information on what the user\nclicked on and it has the chance to\n\nclicked on and it has the chance to\n \n\nclicked on and it has the chance to\nrender the page much like a web\n\nrender the page much like a web\n \n\nrender the page much like a web\nframework would so again it goes back it\n\nframework would so again it goes back it\n \n\nframework would so again it goes back it\nprobably hits uh a get request for that\n\nprobably hits uh a get request for that\n \n\nprobably hits uh a get request for that\nspecific email and pulls the\n\nspecific email and pulls the\n \n\nspecific email and pulls the\nbody what is this agent going to do I'm\n\nbody what is this agent going to do I'm\n \n\nbody what is this agent going to do I'm\nwatching you live so the LM just decided\n\nwatching you live so the LM just decided\n \n\nwatching you live so the LM just decided\nthis is the appropriate UI for a Gmail\n\nthis is the appropriate UI for a Gmail\n \n\nthis is the appropriate UI for a Gmail\nclient also I have other features the LM\n\nclient also I have other features the LM\n \n\nclient also I have other features the LM\nthought was reasonable so looks like I\n\nthought was reasonable so looks like I\n \n\nthought was reasonable so looks like I\ncould mark it as unread or or delete the\n\ncould mark it as unread or or delete the\n \n\ncould mark it as unread or or delete the\nemail if I want to uh maybe I'll delete\n\nemail if I want to uh maybe I'll delete\n \n\nemail if I want to uh maybe I'll delete\nit because it's not that good of an\n\nit because it's not that good of an\n \n\nit because it's not that good of an\nemail I'm sorry\n\n \n\nit is very slow because we're doing a\n\nit is very slow because we're doing a\n \n\nit is very slow because we're doing a\nlot but wanted to push you in this\n\nlot but wanted to push you in this\n \n\nlot but wanted to push you in this\ndirection because this kind of software\n\ndirection because this kind of software\n \n\ndirection because this kind of software\nbarely\n\n \n\nworks\n\nworks\n \n\nworks\ndang I guess not um also I clicked on it\n\ndang I guess not um also I clicked on it\n \n\ndang I guess not um also I clicked on it\nand now the LM is trying to do something\n\nand now the LM is trying to do something\n \n\nand now the LM is trying to do something\nwith me clicking on it anyway um this\n\nwith me clicking on it anyway um this\n \n\nwith me clicking on it anyway um this\nkind of software barely works today and\n\nkind of software barely works today and\n \n\nkind of software barely works today and\nit doesn't mean it w won't work in the\n\nit doesn't mean it w won't work in the\n \n\nit doesn't mean it w won't work in the\nfuture uh but exponential trends like\n\nfuture uh but exponential trends like\n \n\nfuture uh but exponential trends like\nthings might just like this might just\n\nthings might just like this might just\n \n\nthings might just like this might just\ntake off um so just wanted to push you\n\ntake off um so just wanted to push you\n \n\ntake off um so just wanted to push you\nall to think in this direction um yeah\n\nall to think in this direction um yeah\n \n\nall to think in this direction um yeah\nWill software more software look like\n\nWill software more software look like\n \n\nWill software more software look like\nthis I don't know we'll see thank you\n\nthis I don't know we'll see thank you\n \n\nthis I don't know we'll see thank you\n[Music]" } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\n[Music]\n\n \n\nto quickly introduce myself. My name is\n\nto quickly introduce myself. My name is\n \n\nto quickly introduce myself. My name is\nMitesh. I lead the develop advocate team\n\nMitesh. I lead the develop advocate team\n \n\nMitesh. I lead the develop advocate team\nat Nvidia. And the goal of my team is..." }, { "role": "assistant", "content": "[Music]\n\n \n\nto quickly introduce myself. My name is\n\nto quickly introduce myself. My name is\n \n\nto quickly introduce myself. My name is\nMitesh. I lead the develop advocate team\n\nMitesh. I lead the develop advocate team\n \n\nMitesh. I lead the develop advocate team\nat Nvidia. And the goal of my team is to\n\nat Nvidia. And the goal of my team is to\n \n\nat Nvidia. And the goal of my team is to\nuh create technical workflows, notebooks\n\nuh create technical workflows, notebooks\n \n\nuh create technical workflows, notebooks\nuh for different applications and then\n\nuh for different applications and then\n \n\nuh for different applications and then\nwe release that codebase uh on GitHub. we release that codebase uh on GitHub. we release that codebase uh on GitHub. So developers in general which is me and\n\nSo developers in general which is me and\n \n\nSo developers in general which is me and\nyou all of us together we can harness\n\nyou all of us together we can harness\n \n\nyou all of us together we can harness\nthat uh that knowledge and take it\n\nthat uh that knowledge and take it\n \n\nthat uh that knowledge and take it\nfurther for the application or use case\n\nfurther for the application or use case\n \n\nfurther for the application or use case\nthat you're working on. So that is what\n\nthat you're working on. So that is what\n \n\nthat you're working on. So that is what\nmy uh my team does including myself. In\n\nmy uh my team does including myself. In\n \n\nmy uh my team does including myself. In\ntoday's talk, I'm I'm I'm going to talk\n\ntoday's talk, I'm I'm I'm going to talk\n \n\ntoday's talk, I'm I'm I'm going to talk\nabout this project that we did with one\n\nabout this project that we did with one\n \n\nabout this project that we did with one\nof our partners um um and some of my\n\nof our partners um um and some of my\n \n\nof our partners um um and some of my\ncolleagues at Nvidia and our partner\n\ncolleagues at Nvidia and our partner\n \n\ncolleagues at Nvidia and our partner\nabout how can we create a graph rack\n\nabout how can we create a graph rack\n \n\nabout how can we create a graph rack\nsystem what are the advantages of it and\n\nsystem what are the advantages of it and\n \n\nsystem what are the advantages of it and\nif we add the hybrid nature to it how it\n\nif we add the hybrid nature to it how it\n \n\nif we add the hybrid nature to it how it\nis helpful so that's what my uh my talk\n\nis helpful so that's what my uh my talk\n \n\nis helpful so that's what my uh my talk\nis going to be on I will not give I will\n\nis going to be on I will not give I will\n \n\nis going to be on I will not give I will\nnot be able to give you a 10TI view\n\nnot be able to give you a 10TI view\n \n\nnot be able to give you a 10TI view\nwhere you can I can dive with you in the\n\nwhere you can I can dive with you in the\n \n\nwhere you can I can dive with you in the\ncodebase but there is a GitHub link at\n\ncodebase but there is a GitHub link at\n \n\ncodebase but there is a GitHub link at\nthe end of this talk which you can um\n\nthe end of this talk which you can um\n \n\nthe end of this talk which you can um\nscan and all these notebooks whatever\n\nscan and all these notebooks whatever\n \n\nscan and all these notebooks whatever\nI'm going to talk about is available for\n\nI'm going to talk about is available for\n \n\nI'm going to talk about is available for\nyou to take home but I'll give you a\n\nyou to take home but I'll give you a\n \n\nyou to take home but I'll give you a\n10,000 ft view or if you are trying to\n\n10,000 ft view or if you are trying to\n \n\n10,000 ft view or if you are trying to\nbuild your own graph rag system how can\n\nbuild your own graph rag system how can\n \n\nbuild your own graph rag system how can\nyou build it so u a quick refresher what\n\nyou build it so u a quick refresher what\n \n\nyou build it so u a quick refresher what\nis knowledger graph um and why are they\n\nis knowledger graph um and why are they\n \n\nis knowledger graph um and why are they\nimportant so um it is a network that\n\nimportant so um it is a network that\n \n\nimportant so um it is a network that\nrepresents relationship between\n\nrepresents relationship between\n \n\nrepresents relationship between\ndifferent entities and those entities\n\ndifferent entities and those entities\n \n\ndifferent entities and those entities\ncan be anything it can be people places\n\ncan be anything it can be people places\n \n\ncan be anything it can be people places\nuh concept events. A a simple example\n\nuh concept events. A a simple example\n \n\nuh concept events. A a simple example\nwould be me being here. What is my\n\nwould be me being here. What is my\n \n\nwould be me being here. What is my\nrelationship to AI worldfare conference? relationship to AI worldfare conference? relationship to AI worldfare conference? AI engineers worldfare conference and my\n\nAI engineers worldfare conference and my\n \n\nAI engineers worldfare conference and my\nrelationship is I'm a speaker at this\n\nrelationship is I'm a speaker at this\n \n\nrelationship is I'm a speaker at this\nconference. What is my relationship to\n\nconference. What is my relationship to\n \n\nconference. What is my relationship to\nanyone who is attending here? Well, uh\n\nanyone who is attending here? Well, uh\n \n\nanyone who is attending here? Well, uh\nour relationship is you attended my\n\nour relationship is you attended my\n \n\nour relationship is you attended my\nsession. So this edge of relationship\n\nsession. So this edge of relationship\n \n\nsession. So this edge of relationship\nbetween the two entities becomes very\n\nbetween the two entities becomes very\n \n\nbetween the two entities becomes very\nimportant uh to which only graph-based\n\nimportant uh to which only graph-based\n \n\nimportant uh to which only graph-based\nnetwork can exploit or knowledge graphs\n\nnetwork can exploit or knowledge graphs\n \n\nnetwork can exploit or knowledge graphs\ncan exploit. And that is the reason why\n\ncan exploit. And that is the reason why\n \n\ncan exploit. And that is the reason why\nuh there's a lot of active research\n\nuh there's a lot of active research\n \n\nuh there's a lot of active research\nhappening in this domain of how you can\n\nhappening in this domain of how you can\n \n\nhappening in this domain of how you can\nharness graph rag u how can how you can\n\nharness graph rag u how can how you can\n \n\nharness graph rag u how can how you can\nharness knowledge graph and put it into\n\nharness knowledge graph and put it into\n \n\nharness knowledge graph and put it into\na rag based system. So the goal is three\n\na rag based system. So the goal is three\n \n\na rag based system. So the goal is three\nthings. How can you create a triplet\n\nthings. How can you create a triplet\n \n\nthings. How can you create a triplet\nwhich is the which defines the\n\nwhich is the which defines the\n \n\nwhich is the which defines the\nrelationship between these entities that\n\nrelationship between these entities that\n \n\nrelationship between these entities that\ngraph our graph-based system or\n\ngraph our graph-based system or\n \n\ngraph our graph-based system or\nknowledge graph is really good at\n\nknowledge graph is really good at\n \n\nknowledge graph is really good at\nexploiting\n\nexploiting\n \n\nexploiting\nand that's what is unique about this\n\nand that's what is unique about this\n \n\nand that's what is unique about this\nknowledge graph. So if you think about\n\nknowledge graph. So if you think about\n \n\nknowledge graph. So if you think about\num um why can they work better than\n\num um why can they work better than\n \n\num um why can they work better than\nsemantic u rag system well it captures\n\nsemantic u rag system well it captures\n \n\nsemantic u rag system well it captures\nthe information between entities in much\n\nthe information between entities in much\n \n\nthe information between entities in much\nmore detail. So those connections can um\n\nmore detail. So those connections can um\n \n\nmore detail." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nSo those connections can um\ncan provide a very comprehensive view um\n\ncan provide a very comprehensive view um\n \n\ncan provide a very comprehensive view um\num of the knowledge that you that you\n\num of the knowledge that you that you\n \n\num of the knowledge that you that you\nare creating in your rag sy..." }, { "role": "assistant", "content": "So those connections can um\ncan provide a very comprehensive view um\n\ncan provide a very comprehensive view um\n \n\ncan provide a very comprehensive view um\num of the knowledge that you that you\n\num of the knowledge that you that you\n \n\num of the knowledge that you that you\nare creating in your rag system and that\n\nare creating in your rag system and that\n \n\nare creating in your rag system and that\nwill become very important to exploit\n\nwill become very important to exploit\n \n\nwill become very important to exploit\nwhen you are retrieving some of that\n\nwhen you are retrieving some of that\n \n\nwhen you are retrieving some of that\ninformation and and converting that into\n\ninformation and and converting that into\n \n\ninformation and and converting that into\ninto a response for the user who is\n\ninto a response for the user who is\n \n\ninto a response for the user who is\nasking that question and it and it has\n\nasking that question and it and it has\n \n\nasking that question and it and it has\nthe ability to organize your data from\n\nthe ability to organize your data from\n \n\nthe ability to organize your data from\nmultiple sources. I mean that's a given\n\nmultiple sources. I mean that's a given\n \n\nmultiple sources. I mean that's a given\nno matter u what kind of rack system\n\nno matter u what kind of rack system\n \n\nno matter u what kind of rack system\nyou're building. you're building. you're building. So how do we create a graph rag or a\n\nSo how do we create a graph rag or a\n \n\nSo how do we create a graph rag or a\nhybrid system? So this is the highlevel\n\nhybrid system? So this is the highlevel\n \n\nhybrid system? So this is the highlevel\ndiagram of what it entails. So I broke\n\ndiagram of what it entails. So I broke\n \n\ndiagram of what it entails. So I broke\nit down into four components. The very\n\nit down into four components. The very\n \n\nit down into four components. The very\nfirst thing is your data. You need to\n\nfirst thing is your data. You need to\n \n\nfirst thing is your data. You need to\nprocess your data. The better you\n\nprocess your data. The better you\n \n\nprocess your data. The better you\nprocess your data, the better is a\n\nprocess your data, the better is a\n \n\nprocess your data, the better is a\nknowledge graph. The better is a\n\nknowledge graph. The better is a\n \n\nknowledge graph. The better is a\nknowledge graph, the better is the\n\nknowledge graph, the better is the\n \n\nknowledge graph, the better is the\nretrieval. So four components data, data\n\nretrieval. So four components data, data\n \n\nretrieval. So four components data, data\nprocessing, your graph creation or your\n\nprocessing, your graph creation or your\n \n\nprocessing, your graph creation or your\nsemantic u embedding vector database\n\nsemantic u embedding vector database\n \n\nsemantic u embedding vector database\ncreation. Those are the three uh steps. creation. Those are the three uh steps. creation. Those are the three uh steps. And then the last step is of course\n\nAnd then the last step is of course\n \n\nAnd then the last step is of course\ninferencing when you're asking questions\n\ninferencing when you're asking questions\n \n\ninferencing when you're asking questions\nuh to your u rag pipeline. uh to your u rag pipeline. uh to your u rag pipeline. And at a higher level this can be broken\n\nAnd at a higher level this can be broken\n \n\nAnd at a higher level this can be broken\ndown into two big pieces offline online. down into two big pieces offline online. down into two big pieces offline online. So all your data processing u work which\n\nSo all your data processing u work which\n \n\nSo all your data processing u work which\nis a one-time process is offline and and\n\nis a one-time process is offline and and\n \n\nis a one-time process is offline and and\nonce you have created your knowledge\n\nonce you have created your knowledge\n \n\nonce you have created your knowledge\ngraph which is your triplet entity\n\ngraph which is your triplet entity\n \n\ngraph which is your triplet entity\nrelationship entity 2 um or your\n\nrelationship entity 2 um or your\n \n\nrelationship entity 2 um or your\nsemantic vector database once you have\n\nsemantic vector database once you have\n \n\nsemantic vector database once you have\nit then it's all about quering it and\n\nit then it's all about quering it and\n \n\nit then it's all about quering it and\nconverting that information into um um a\n\nconverting that information into um um a\n \n\nconverting that information into um um a\nresponse that is readable to the user. response that is readable to the user. response that is readable to the user. It cannot be something that here are the\n\nIt cannot be something that here are the\n \n\nIt cannot be something that here are the\nthree relationship and then we as the\n\nthree relationship and then we as the\n \n\nthree relationship and then we as the\nuser have to go figure out what does\n\nuser have to go figure out what does\n \n\nuser have to go figure out what does\nthis exactly mean. this exactly mean. this exactly mean. So the top um part of this u flow\n\nSo the top um part of this u flow\n \n\nSo the top um part of this u flow\ndiagram is where you build your semantic\n\ndiagram is where you build your semantic\n \n\ndiagram is where you build your semantic\nu vector database which is you you pick\n\nu vector database which is you you pick\n \n\nu vector database which is you you pick\nyour uh u documents and then you convert\n\nyour uh u documents and then you convert\n \n\nyour uh u documents and then you convert\nthem into vector embeddings and you\n\nthem into vector embeddings and you\n \n\nthem into vector embeddings and you\nstore into a vector database. store into a vector database. store into a vector database. So that piece is uh is how you create\n\nSo that piece is uh is how you create\n \n\nSo that piece is uh is how you create\nyour semantic uh vector database and\n\nyour semantic uh vector database and\n \n\nyour semantic uh vector database and\nthen the piece below is um how you\n\nthen the piece below is um how you\n \n\nthen the piece below is um how you\ncreate your knowledge graph and it is\n\ncreate your knowledge graph and it is\n \n\ncreate your knowledge graph and it is\nmuch more uh um there are much more\n\nmuch more uh um there are much more\n \n\nmuch more uh um there are much more\nsteps that you have to follow a care\n\nsteps that you have to follow a care\n \n\nsteps that you have to follow a care\nthat you have to take when you're\n\nthat you have to take when you're\n \n\nthat you have to take when you're\ncreating your knowledge graph. So diving in the first step creating\n\nSo diving in the first step creating\n \n\nSo diving in the first step creating\nyour knowledge graph. How can you create\n\nyour knowledge graph. How can you create\n \n\nyour knowledge graph. How can you create\nthose triplets out of documents that are\n\nthose triplets out of documents that are\n \n\nthose triplets out of documents that are\nthat are not that structured? So\n\nthat are not that structured? So\n \n\nthat are not that structured? So\ncreating triplets which uh which exposes\n\ncreating triplets which uh which exposes\n \n\ncreating triplets which uh which exposes\nthe information between two entities and\n\nthe information between two entities and\n \n\nthe information between two entities and\npicking up those entities uh so that\n\npicking up those entities uh so that\n \n\npicking up those entities uh so that\nthat information becomes helpful is very\n\nthat information becomes helpful is very\n \n\nthat information becomes helpful is very\nimportant. Here's a simple example. This\n\nimportant. Here's a simple example. This\n \n\nimportant. Here's a simple example." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nThis\ndocument is of Exxon Mobile's uh results\n\ndocument is of Exxon Mobile's uh results\n \n\ndocument is of Exxon Mobile's uh results\nI think uh their quarterly results and\n\nI think uh their quarterly results and\n \n\nI think uh their quarterly results and\nwe we tried to pick up um the\n\nwe we tried to p..." }, { "role": "assistant", "content": "This\ndocument is of Exxon Mobile's uh results\n\ndocument is of Exxon Mobile's uh results\n \n\ndocument is of Exxon Mobile's uh results\nI think uh their quarterly results and\n\nI think uh their quarterly results and\n \n\nI think uh their quarterly results and\nwe we tried to pick up um the\n\nwe we tried to pick up um the\n \n\nwe we tried to pick up um the\nrelationship or create the the knowledge\n\nrelationship or create the the knowledge\n \n\nrelationship or create the the knowledge\ngraph using an LLM and if you see at the\n\ngraph using an LLM and if you see at the\n \n\ngraph using an LLM and if you see at the\nfirst line it's Exon Mobile which is a\n\nfirst line it's Exon Mobile which is a\n \n\nfirst line it's Exon Mobile which is a\ncompany that's the entity uh cut is the\n\ncompany that's the entity uh cut is the\n \n\ncompany that's the entity uh cut is the\nfeature of um of that entity spending\n\nfeature of um of that entity spending\n \n\nfeature of um of that entity spending\noil and gas exploration um and activity\n\noil and gas exploration um and activity\n \n\noil and gas exploration um and activity\nmy apologies cut is the relationship\n\nmy apologies cut is the relationship\n \n\nmy apologies cut is the relationship\nbetween Exxon on mobile and spending on\n\nbetween Exxon on mobile and spending on\n \n\nbetween Exxon on mobile and spending on\noil and gas exploration and activity is\n\noil and gas exploration and activity is\n \n\noil and gas exploration and activity is\nthe the um the name of the entity\n\nthe the um the name of the entity\n \n\nthe the um the name of the entity\nspending on oil and gas exploration. So\n\nspending on oil and gas exploration. So\n \n\nspending on oil and gas exploration. So\nthis is how the relationship needs to be\n\nthis is how the relationship needs to be\n \n\nthis is how the relationship needs to be\nexploited. Now the question that comes\n\nexploited. Now the question that comes\n \n\nexploited. Now the question that comes\nto our mind is that sounds very\n\nto our mind is that sounds very\n \n\nto our mind is that sounds very\ndifficult to do and exactly it is\n\ndifficult to do and exactly it is\n \n\ndifficult to do and exactly it is\ndifficult to do and that is the reason\n\ndifficult to do and that is the reason\n \n\ndifficult to do and that is the reason\nwhy we need to harness uh or we need to\n\nwhy we need to harness uh or we need to\n \n\nwhy we need to harness uh or we need to\nuse LLM to figure out a way to extract\n\nuse LLM to figure out a way to extract\n \n\nuse LLM to figure out a way to extract\nthis information and structure it for us\n\nthis information and structure it for us\n \n\nthis information and structure it for us\nso that we can save it in um um in a\n\nso that we can save it in um um in a\n \n\nso that we can save it in um um in a\ntriplet format and how can we do that\n\ntriplet format and how can we do that\n \n\ntriplet format and how can we do that\nprompt engineering but we need to be\n\nprompt engineering but we need to be\n \n\nprompt engineering but we need to be\nmuch more uh uh uh defined about it. So\n\nmuch more uh uh uh defined about it. So\n \n\nmuch more uh uh uh defined about it. So\nyou based on the use case that you are\n\nyou based on the use case that you are\n \n\nyou based on the use case that you are\ntrying to work on you can define your\n\ntrying to work on you can define your\n \n\ntrying to work on you can define your\noncology and once you have defined your\n\noncology and once you have defined your\n \n\noncology and once you have defined your\noncology you can put it in your prompt\n\noncology you can put it in your prompt\n \n\noncology you can put it in your prompt\nand then ask the LLM to go extract this\n\nand then ask the LLM to go extract this\n \n\nand then ask the LLM to go extract this\ninformation that is oncology specific\n\ninformation that is oncology specific\n \n\ninformation that is oncology specific\nfrom the documents and then structure it\n\nfrom the documents and then structure it\n \n\nfrom the documents and then structure it\nin that way so that that can be stored\n\nin that way so that that can be stored\n \n\nin that way so that that can be stored\nin a form of a triplet. This step is\n\nin a form of a triplet. This step is\n \n\nin a form of a triplet. This step is\nvery important. You might be spending a\n\nvery important. You might be spending a\n \n\nvery important. You might be spending a\nlot of time here to make sure your\n\nlot of time here to make sure your\n \n\nlot of time here to make sure your\nprompt is doing the right thing and it\n\nprompt is doing the right thing and it\n \n\nprompt is doing the right thing and it\nis creating the right oncology for you. is creating the right oncology for you. is creating the right oncology for you. If your oncology is not right, uh if\n\nIf your oncology is not right, uh if\n \n\nIf your oncology is not right, uh if\nyour triplets are not right, if they are\n\nyour triplets are not right, if they are\n \n\nyour triplets are not right, if they are\nnoisy, your retrieval will be noisy. So\n\nnoisy, your retrieval will be noisy. So\n \n\nnoisy, your retrieval will be noisy. So\nthis is where you will be going back and\n\nthis is where you will be going back and\n \n\nthis is where you will be going back and\nforth figuring out how to get a better\n\nforth figuring out how to get a better\n \n\nforth figuring out how to get a better\noncology. oncology. oncology. So th this is where you will spend my\n\nSo th this is where you will spend my\n \n\nSo th this is where you will spend my\ntake is this is where you'll spend uh\n\ntake is this is where you'll spend uh\n \n\ntake is this is where you'll spend uh\n80% of your time to make sure you get\n\n80% of your time to make sure you get\n \n\n80% of your time to make sure you get\nthe oncology right and you'll be going\n\nthe oncology right and you'll be going\n \n\nthe oncology right and you'll be going\nback and forth in an iterative manner to\n\nback and forth in an iterative manner to\n \n\nback and forth in an iterative manner to\nsee how you can make it better over time\n\n \n\nand then the next vector database for a\n\nand then the next vector database for a\n \n\nand then the next vector database for a\nhybrid rack system is to create the\n\nhybrid rack system is to create the\n \n\nhybrid rack system is to create the\nsemantic vector database and that is\n\nsemantic vector database and that is\n \n\nsemantic vector database and that is\nvery reasonably straight straight\n\nvery reasonably straight straight\n \n\nvery reasonably straight straight\nstraightforward or it is well studied. straightforward or it is well studied. straightforward or it is well studied. So you pick your document. This is the\n\nSo you pick your document. This is the\n \n\nSo you pick your document. This is the\nfirst page of attention is all you need\n\nfirst page of attention is all you need\n \n\nfirst page of attention is all you need\nresearch paper. And you you break it\n\nresearch paper. And you you break it\n \n\nresearch paper. And you you break it\ninto chunk sizes and you you have\n\ninto chunk sizes and you you have\n \n\ninto chunk sizes and you you have\nanother factor called overlap. And chunk\n\nanother factor called overlap. And chunk\n \n\nanother factor called overlap. And chunk\nsizes are important because what\n\nsizes are important because what\n \n\nsizes are important because what\nsemantic vector database does is it will\n\nsemantic vector database does is it will\n \n\nsemantic vector database does is it will\nit will pick up that chunk and convert\n\nit will pick up that chunk and convert\n \n\nit will pick up that chunk and convert\nthat into use the embedding model and\n\nthat into use the embedding model and\n \n\nthat into use the embedding model and\nconvert them into a u embedding vector\n\nconvert them into a u embedding vector\n \n\nconvert them into a u embedding vector\nand store into the vector database. And\n\nand store into the vector database. And\n \n\nand store into the vector database." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nAnd\nit will if you don't have an overlap\n\nit will if you don't have an overlap\n \n\nit will if you don't have an overlap\nthen the context between the previous\n\nthen the context between the previous\n \n\nthen the context between the previous\nand the and the next chunk will be lost. and the and the next c..." }, { "role": "assistant", "content": "And\nit will if you don't have an overlap\n\nit will if you don't have an overlap\n \n\nit will if you don't have an overlap\nthen the context between the previous\n\nthen the context between the previous\n \n\nthen the context between the previous\nand the and the next chunk will be lost. and the and the next chunk will be lost. and the and the next chunk will be lost. if there is any relationship. So you try\n\nif there is any relationship. So you try\n \n\nif there is any relationship. So you try\nto be smart on how much overlap do I\n\nto be smart on how much overlap do I\n \n\nto be smart on how much overlap do I\nneed between my previous chunk and the\n\nneed between my previous chunk and the\n \n\nneed between my previous chunk and the\nand the next chunk and what is the size\n\nand the next chunk and what is the size\n \n\nand the next chunk and what is the size\nof the chunk that I should uh I should\n\nof the chunk that I should uh I should\n \n\nof the chunk that I should uh I should\nuse when I'm chunking my documents into\n\nuse when I'm chunking my documents into\n \n\nuse when I'm chunking my documents into\ndifferent paragraphs. That is where the\n\ndifferent paragraphs. That is where the\n \n\ndifferent paragraphs. That is where the\nthe advantage of graph rag comes into\n\nthe advantage of graph rag comes into\n \n\nthe advantage of graph rag comes into\nplay because uh if you think about it\n\nplay because uh if you think about it\n \n\nplay because uh if you think about it\nthe important information which is uh\n\nthe important information which is uh\n \n\nthe important information which is uh\nthe relationship between different\n\nthe relationship between different\n \n\nthe relationship between different\nentities are not exploited by u by your\n\nentities are not exploited by u by your\n \n\nentities are not exploited by u by your\nsemantic uh uh vector database but they\n\nsemantic uh uh vector database but they\n \n\nsemantic uh uh vector database but they\nare exploited really well when you're\n\nare exploited really well when you're\n \n\nare exploited really well when you're\ntrying to um use a knowledge graph or\n\ntrying to um use a knowledge graph or\n \n\ntrying to um use a knowledge graph or\ncreate a knowledge graph based system. create a knowledge graph based system. create a knowledge graph based system. So once you have created this uh um this\n\nSo once you have created this uh um this\n \n\nSo once you have created this uh um this\nknowledge graph what is the next step? knowledge graph what is the next step? knowledge graph what is the next step? Now, now comes the retrieval piece which\n\nNow, now comes the retrieval piece which\n \n\nNow, now comes the retrieval piece which\nis um um you you ask a question what is\n\nis um um you you ask a question what is\n \n\nis um um you you ask a question what is\nExon Mobile's\n\nExon Mobile's\n \n\nExon Mobile's\ncut this quarter that that it is looking\n\ncut this quarter that that it is looking\n \n\ncut this quarter that that it is looking\nlike and knowledger graph\n\nlike and knowledger graph\n \n\nlike and knowledger graph\nwill will help you figure out how to\n\nwill will help you figure out how to\n \n\nwill will help you figure out how to\nretrieve those nodes or those entities\n\nretrieve those nodes or those entities\n \n\nretrieve those nodes or those entities\nand the relationship between them. But\n\nand the relationship between them. But\n \n\nand the relationship between them. But\nif you do uh a very flat retrieval which\n\nif you do uh a very flat retrieval which\n \n\nif you do uh a very flat retrieval which\nis a single hop you are missing uh the\n\nis a single hop you are missing uh the\n \n\nis a single hop you are missing uh the\nthe most important u piece that graph\n\nthe most important u piece that graph\n \n\nthe most important u piece that graph\nallows you which is exploitation through\n\nallows you which is exploitation through\n \n\nallows you which is exploitation through\nmultiple nodes that you can think about\n\nmultiple nodes that you can think about\n \n\nmultiple nodes that you can think about\nand that becomes very very very\n\nand that becomes very very very\n \n\nand that becomes very very very\nimportant. I I cannot stress how\n\nimportant. I I cannot stress how\n \n\nimportant. I I cannot stress how\nimportant that becomes. So think of\n\nimportant that becomes. So think of\n \n\nimportant that becomes. So think of\ndifferent strategies. Again you will\n\ndifferent strategies. Again you will\n \n\ndifferent strategies. Again you will\nspend a lot of time to optimize this\n\nspend a lot of time to optimize this\n \n\nspend a lot of time to optimize this\nwhether you should look at um single\n\nwhether you should look at um single\n \n\nwhether you should look at um single\nhop, double hop, how much deep you want\n\nhop, double hop, how much deep you want\n \n\nhop, double hop, how much deep you want\nto go so that nodes um the relationship\n\nto go so that nodes um the relationship\n \n\nto go so that nodes um the relationship\nbetween your first node to the second\n\nbetween your first node to the second\n \n\nbetween your first node to the second\nnode, your second node to the third node\n\nnode, your second node to the third node\n \n\nnode, your second node to the third node\nis exploited pretty well. And and the\n\nis exploited pretty well. And and the\n \n\nis exploited pretty well. And and the\nthe more deeper you go, the better\n\nthe more deeper you go, the better\n \n\nthe more deeper you go, the better\ncontext you'll get. But there's a\n\ncontext you'll get. But there's a\n \n\ncontext you'll get. But there's a\ndisadvantage of that. The more deeper\n\ndisadvantage of that. The more deeper\n \n\ndisadvantage of that. The more deeper\nyou go, the more time you're going to\n\nyou go, the more time you're going to\n \n\nyou go, the more time you're going to\nspend on retrieving that information. So\n\nspend on retrieving that information. So\n \n\nspend on retrieving that information. So\nthen uh uh latency becomes a factor as\n\nthen uh uh latency becomes a factor as\n \n\nthen uh uh latency becomes a factor as\nwell especially when you're working in a\n\nwell especially when you're working in a\n \n\nwell especially when you're working in a\nproduction environment. So there is a\n\nproduction environment. So there is a\n \n\nproduction environment. So there is a\nsweet spot that you'll have to hit when\n\nsweet spot that you'll have to hit when\n \n\nsweet spot that you'll have to hit when\nyou're trying to um go how deep you want\n\nyou're trying to um go how deep you want\n \n\nyou're trying to um go how deep you want\nto go how how many hops you want to go\n\nto go how how many hops you want to go\n \n\nto go how how many hops you want to go\ninto your graph versus how many uh what\n\ninto your graph versus how many uh what\n \n\ninto your graph versus how many uh what\nis the latency that you can u you can\n\nis the latency that you can u you can\n \n\nis the latency that you can u you can\nsurvive. So so that becomes very uh very\n\nsurvive. So so that becomes very uh very\n \n\nsurvive. So so that becomes very uh very\nimportant\n\nimportant\n \n\nimportant\nand those some of those searches can be\n\nand those some of those searches can be\n \n\nand those some of those searches can be\naccelerated. So um um um we created a\n\naccelerated. So um um um we created a\n \n\naccelerated. So um um um we created a\nlibrary called cool graph um which which\n\nlibrary called cool graph um which which\n \n\nlibrary called cool graph um which which\nis a which is available or integrated in\n\nis a which is available or integrated in\n \n\nis a which is available or integrated in\na lot of um libraries out there like\n\na lot of um libraries out there like\n \n\na lot of um libraries out there like\nnetwork X and whatnot. But that\n\nnetwork X and whatnot. But that\n \n\nnetwork X and whatnot." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nBut that\nacceleration becomes important so that\n\nacceleration becomes important so that\n \n\nacceleration becomes important so that\nit gives you the flexibility to get\n\nit gives you the flexibility to get\n \n\nit gives you the flexibility to get\ndeeper into your graph go through\n\ndeeper into your graph ..." }, { "role": "assistant", "content": "But that\nacceleration becomes important so that\n\nacceleration becomes important so that\n \n\nacceleration becomes important so that\nit gives you the flexibility to get\n\nit gives you the flexibility to get\n \n\nit gives you the flexibility to get\ndeeper into your graph go through\n\ndeeper into your graph go through\n \n\ndeeper into your graph go through\nmultiple hops but at the same time you\n\nmultiple hops but at the same time you\n \n\nmultiple hops but at the same time you\ncan reduce the latency so your\n\ncan reduce the latency so your\n \n\ncan reduce the latency so your\nperformance of your graph improves uh a\n\nperformance of your graph improves uh a\n \n\nperformance of your graph improves uh a\nlot. lot. lot. So this is the where the retrieval piece\n\nSo this is the where the retrieval piece\n \n\nSo this is the where the retrieval piece\ncomes into play where you can have\n\ncomes into play where you can have\n \n\ncomes into play where you can have\ndifferent strategies defined so that\n\ndifferent strategies defined so that\n \n\ndifferent strategies defined so that\nwhen you're querying uh your data um and\n\nwhen you're querying uh your data um and\n \n\nwhen you're querying uh your data um and\nget getting the responses you can have\n\nget getting the responses you can have\n \n\nget getting the responses you can have\nbetter responses\n\n \n\nand the other important piece I\n\nand the other important piece I\n \n\nand the other important piece I\npersonally worked on this piece so I I\n\npersonally worked on this piece so I I\n \n\npersonally worked on this piece so I I\ncan talk at length on this but uh I'm\n\ncan talk at length on this but uh I'm\n \n\ncan talk at length on this but uh I'm\nI'm going to give you a very high level\n\nI'm going to give you a very high level\n \n\nI'm going to give you a very high level\num is evaluating the performance and\n\num is evaluating the performance and\n \n\num is evaluating the performance and\nthere are multiple factors that you can\n\nthere are multiple factors that you can\n \n\nthere are multiple factors that you can\nevaluate around faithfulness um answer\n\nevaluate around faithfulness um answer\n \n\nevaluate around faithfulness um answer\nrelevancy uh precision recall\n\nrelevancy uh precision recall\n \n\nrelevancy uh precision recall\num um if you try to use an LLM model,\n\num um if you try to use an LLM model,\n \n\num um if you try to use an LLM model,\nhelpfulness, collectiveness, coherence,\n\nhelpfulness, collectiveness, coherence,\n \n\nhelpfulness, collectiveness, coherence,\ncomplexity, verbosity, all these factors\n\ncomplexity, verbosity, all these factors\n \n\ncomplexity, verbosity, all these factors\nbecomes very important. So there is a\n\nbecomes very important. So there is a\n \n\nbecomes very important. So there is a\nlibrary pistol library called Ragas. Um\n\nlibrary pistol library called Ragas. Um\n \n\nlibrary pistol library called Ragas. Um\nit is meant to evaluate your rag\n\nit is meant to evaluate your rag\n \n\nit is meant to evaluate your rag\nworkflow end to end. Anyone who used\n\nworkflow end to end. Anyone who used\n \n\nworkflow end to end. Anyone who used\nRagas for evaluating your graph rag? All\n\nRagas for evaluating your graph rag? All\n \n\nRagas for evaluating your graph rag? All\nright, a few of them. Thank you. But it\n\nright, a few of them. Thank you. But it\n \n\nright, a few of them. Thank you. But it\nis it is an amazing library that you can\n\nis it is an amazing library that you can\n \n\nis it is an amazing library that you can\nuh uh use to evaluate your uh your rag\n\nuh uh use to evaluate your uh your rag\n \n\nuh uh use to evaluate your uh your rag\npipeline end to end because it evaluates\n\npipeline end to end because it evaluates\n \n\npipeline end to end because it evaluates\nthe response. It evaluates the retrieval\n\nthe response. It evaluates the retrieval\n \n\nthe response. It evaluates the retrieval\nand it evaluates what the query is. So\n\nand it evaluates what the query is. So\n \n\nand it evaluates what the query is. So\nit it will evaluate your your pipeline\n\nit it will evaluate your your pipeline\n \n\nit it will evaluate your your pipeline\nend to end which becomes very handy when\n\nend to end which becomes very handy when\n \n\nend to end which becomes very handy when\nyou're when you're trying to test\n\nyou're when you're trying to test\n \n\nyou're when you're trying to test\nwhether my retrieval is doing the right\n\nwhether my retrieval is doing the right\n \n\nwhether my retrieval is doing the right\nthing or whether my uh the questions\n\nthing or whether my uh the questions\n \n\nthing or whether my uh the questions\nthat I'm asking is the LLM interpreting\n\nthat I'm asking is the LLM interpreting\n \n\nthat I'm asking is the LLM interpreting\nit in in the right way or not. So you\n\nit in in the right way or not. So you\n \n\nit in in the right way or not. So you\ncan break down your responses in u the\n\ncan break down your responses in u the\n \n\ncan break down your responses in u the\nraas pipeline will evaluate all those\n\nraas pipeline will evaluate all those\n \n\nraas pipeline will evaluate all those\npieces and see what your eventual score\n\npieces and see what your eventual score\n \n\npieces and see what your eventual score\nis. So it is a pip install library. The\n\nis. So it is a pip install library. The\n \n\nis. So it is a pip install library. The\nother is LLM uh and Ragas under the hood\n\nother is LLM uh and Ragas under the hood\n \n\nother is LLM uh and Ragas under the hood\nuses an LLM um no surprises there. By\n\nuses an LLM um no surprises there. By\n \n\nuses an LLM um no surprises there. By\ndefault, it is integrated with GPT, but\n\ndefault, it is integrated with GPT, but\n \n\ndefault, it is integrated with GPT, but\nit provides you the flexibility that if\n\nit provides you the flexibility that if\n \n\nit provides you the flexibility that if\nyou have your own um model, you can\n\nyou have your own um model, you can\n \n\nyou have your own um model, you can\nbring it in as well and you can uh wire\n\nbring it in as well and you can uh wire\n \n\nbring it in as well and you can uh wire\nit up with your API and you can use that\n\nit up with your API and you can use that\n \n\nit up with your API and you can use that\nLLM to figure out on these four four\n\nLLM to figure out on these four four\n \n\nLLM to figure out on these four four\nevaluation parameters that RAS offers. evaluation parameters that RAS offers. evaluation parameters that RAS offers. So, so it's a it's it's it's quite comp\n\nSo, so it's a it's it's it's quite comp\n \n\nSo, so it's a it's it's it's quite comp\nI would say it's comprehensive but it's\n\nI would say it's comprehensive but it's\n \n\nI would say it's comprehensive but it's\nreally good in terms of giving you that\n\nreally good in terms of giving you that\n \n\nreally good in terms of giving you that\nflexibility. The other path is uh using\n\nflexibility. The other path is uh using\n \n\nflexibility. The other path is uh using\na model that is meant to evaluate\n\na model that is meant to evaluate\n \n\na model that is meant to evaluate\nspecifically the response coming out of\n\nspecifically the response coming out of\n \n\nspecifically the response coming out of\nLM. And that is where this model\n\nLM. And that is where this model\n \n\nLM. And that is where this model\nLanimotron 340 million reward model that\n\nLanimotron 340 million reward model that\n \n\nLanimotron 340 million reward model that\nwe released I think few years ago. At\n\nwe released I think few years ago. At\n \n\nwe released I think few years ago. At\nthat time it was a really good response\n\nthat time it was a really good response\n \n\nthat time it was a really good response\nmodel. It's it's a 340 billion parameter\n\nmodel. It's it's a 340 billion parameter\n \n\nmodel. It's it's a 340 billion parameter\nmodel so reasonably big but uh it\n\nmodel so reasonably big but uh it\n \n\nmodel so reasonably big but uh it\nevaluates\n\nevaluates\n \n\nevaluates\num it's a reward model. So it will go\n\num it's a reward model. So it will go\n \n\num it's a reward model." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nSo it will go\nand evaluate the response of another LLM\n\nand evaluate the response of another LLM\n \n\nand evaluate the response of another LLM\nand judge it in terms of um how the\n\nand judge it in terms of um how the\n \n\nand judge it in terms of um how the\nresponses are looking looking like on\n\nresponse..." }, { "role": "assistant", "content": "So it will go\nand evaluate the response of another LLM\n\nand evaluate the response of another LLM\n \n\nand evaluate the response of another LLM\nand judge it in terms of um how the\n\nand judge it in terms of um how the\n \n\nand judge it in terms of um how the\nresponses are looking looking like on\n\nresponses are looking looking like on\n \n\nresponses are looking looking like on\nthis five parameters but it is meant to\n\nthis five parameters but it is meant to\n \n\nthis five parameters but it is meant to\ngo and judge other LLMs. That is how it\n\ngo and judge other LLMs. That is how it\n \n\ngo and judge other LLMs. That is how it\nwas trained. So moving further I would like to use\n\nSo moving further I would like to use\n \n\nSo moving further I would like to use\nthis analogy that for u to create a\n\nthis analogy that for u to create a\n \n\nthis analogy that for u to create a\ngraph ra system it will take you uh\n\ngraph ra system it will take you uh\n \n\ngraph ra system it will take you uh\nwhich is 80% of the job it will take you\n\nwhich is 80% of the job it will take you\n \n\nwhich is 80% of the job it will take you\n20% of your time but then to make it\n\n20% of your time but then to make it\n \n\n20% of your time but then to make it\nbetter which is the last 20% uh sorry\n\nbetter which is the last 20% uh sorry\n \n\nbetter which is the last 20% uh sorry\nwhich is the um the 80/20 rule the last\n\nwhich is the um the 80/20 rule the last\n \n\nwhich is the um the 80/20 rule the last\n20% will take 80% of your time because\n\n20% will take 80% of your time because\n \n\n20% will take 80% of your time because\nnow you are in the process of optimizing\n\nnow you are in the process of optimizing\n \n\nnow you are in the process of optimizing\nit further to make it make sure it\n\nit further to make it make sure it\n \n\nit further to make it make sure it\nworks. for the use case good enough um\n\nworks. for the use case good enough um\n \n\nworks. for the use case good enough um\num for for the application that you're\n\num for for the application that you're\n \n\num for for the application that you're\nworking on and there are some strategies\n\nworking on and there are some strategies\n \n\nworking on and there are some strategies\nthere which I would like to walk you\n\nthere which I would like to walk you\n \n\nthere which I would like to walk you\nthrough so one as I said before which I\n\nthrough so one as I said before which I\n \n\nthrough so one as I said before which I\ncouldn't stress enough the way you are\n\ncouldn't stress enough the way you are\n \n\ncouldn't stress enough the way you are\ncreating your knowledge graph out of\n\ncreating your knowledge graph out of\n \n\ncreating your knowledge graph out of\nyour unstructured data becomes very\n\nyour unstructured data becomes very\n \n\nyour unstructured data becomes very\nimportant the better your knowledge\n\nimportant the better your knowledge\n \n\nimportant the better your knowledge\ngraph the better results you're going to\n\ngraph the better results you're going to\n \n\ngraph the better results you're going to\nget and something that we did as\n\nget and something that we did as\n \n\nget and something that we did as\nexperimentation through this use case\n\nexperimentation through this use case\n \n\nexperimentation through this use case\nthat we were exploring with one of our\n\nthat we were exploring with one of our\n \n\nthat we were exploring with one of our\npartners\n\npartners\n \n\npartners\nuh was can we fine-tune an LLM model to\n\nuh was can we fine-tune an LLM model to\n \n\nuh was can we fine-tune an LLM model to\nget the quality of the of the triplets\n\nget the quality of the of the triplets\n \n\nget the quality of the of the triplets\nthat we are creating better and does\n\nthat we are creating better and does\n \n\nthat we are creating better and does\nthat improve results? Can we do a better\n\nthat improve results? Can we do a better\n \n\nthat improve results? Can we do a better\njob at data processing like removing\n\njob at data processing like removing\n \n\njob at data processing like removing\nreax, apostrophes, brackets, words that\n\nreax, apostrophes, brackets, words that\n \n\nreax, apostrophes, brackets, words that\ncharacters that don't matter? If we\n\ncharacters that don't matter? If we\n \n\ncharacters that don't matter? If we\nremove them, does it give you better\n\nremove them, does it give you better\n \n\nremove them, does it give you better\nresults? So these are like small things\n\nresults? So these are like small things\n \n\nresults? So these are like small things\nthat um that you can think about but it\n\nthat um that you can think about but it\n \n\nthat um that you can think about but it\ngives you it it improves the performance\n\ngives you it it improves the performance\n \n\ngives you it it improves the performance\nof your overall system. So that is where\n\nof your overall system. So that is where\n \n\nof your overall system. So that is where\nyou I'm talking about 80% of your time\n\nyou I'm talking about 80% of your time\n \n\nyou I'm talking about 80% of your time\nsmall nitty-gritty of the things that\n\nsmall nitty-gritty of the things that\n \n\nsmall nitty-gritty of the things that\nyou are the knobs that you are\n\nyou are the knobs that you are\n \n\nyou are the knobs that you are\nfine-tuning with slowly and steadily to\n\nfine-tuning with slowly and steadily to\n \n\nfine-tuning with slowly and steadily to\nmake sure your performance gets better\n\nmake sure your performance gets better\n \n\nmake sure your performance gets better\nand better and I would like to share a\n\nand better and I would like to share a\n \n\nand better and I would like to share a\nfew strategies that we did which we got\n\nfew strategies that we did which we got\n \n\nfew strategies that we did which we got\nuh which led us to uh uh which led us to\n\nuh which led us to uh uh which led us to\n \n\nuh which led us to uh uh which led us to\nget better results. get better results. get better results. So the very first thing is uh reax or\n\nSo the very first thing is uh reax or\n \n\nSo the very first thing is uh reax or\njust cleaning out your data. Um we we\n\njust cleaning out your data. Um we we\n \n\njust cleaning out your data. Um we we\nremoved uh apostrophes as other other\n\nremoved uh apostrophes as other other\n \n\nremoved uh apostrophes as other other\ncharacters that are not that important\n\ncharacters that are not that important\n \n\ncharacters that are not that important\nif you think about uh triplet generation\n\nif you think about uh triplet generation\n \n\nif you think about uh triplet generation\nthat led us to uh um to better uh better\n\nthat led us to uh um to better uh better\n \n\nthat led us to uh um to better uh better\nresults. We we then implemented another\n\nresults. We we then implemented another\n \n\nresults. We we then implemented another\nstrategy of reducing the not not missing\n\nstrategy of reducing the not not missing\n \n\nstrategy of reducing the not not missing\nout of longer output making it smaller. out of longer output making it smaller. out of longer output making it smaller. that got us uh uh better results and we\n\nthat got us uh uh better results and we\n \n\nthat got us uh uh better results and we\nalso fine-tuned the um the llama 3.3\n\nalso fine-tuned the um the llama 3.3\n \n\nalso fine-tuned the um the llama 3.3\nmodel or 3.2 model and that got us\n\nmodel or 3.2 model and that got us\n \n\nmodel or 3.2 model and that got us\nbetter better results. So if you look at\n\nbetter better results. So if you look at\n \n\nbetter better results. So if you look at\nthe last three columns you'll see that\n\nthe last three columns you'll see that\n \n\nthe last three columns you'll see that\nby using llama 3.3 as is we got 70 1%\n\nby using llama 3.3 as is we got 70 1%\n \n\nby using llama 3.3 as is we got 70 1%\naccuracy. So this was tested on 100 uh\n\naccuracy. So this was tested on 100 uh\n \n\naccuracy. So this was tested on 100 uh\ntriplets to see how it is performing and\n\ntriplets to see how it is performing and\n \n\ntriplets to see how it is performing and\nas it got sorry 100 documents." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nSo as it\n\nas it got sorry 100 documents. So as it\n \n\nas it got sorry 100 documents. So as it\ngot better and uh as we introduced Laura\n\ngot better and uh as we introduced Laura\n \n\ngot better and uh as we introduced Laura\nwe fine the llama 3.1 model our our\n\nwe fine the llama 3.1 model our our\n \n\nwe f..." }, { "role": "assistant", "content": "So as it\n\nas it got sorry 100 documents. So as it\n \n\nas it got sorry 100 documents. So as it\ngot better and uh as we introduced Laura\n\ngot better and uh as we introduced Laura\n \n\ngot better and uh as we introduced Laura\nwe fine the llama 3.1 model our our\n\nwe fine the llama 3.1 model our our\n \n\nwe fine the llama 3.1 model our our\naccuracy or performance went up from 71\n\naccuracy or performance went up from 71\n \n\naccuracy or performance went up from 71\nto 87%. And then we did those small\n\nto 87%. And then we did those small\n \n\nto 87%. And then we did those small\ntweaks uh it improved the performance\n\ntweaks uh it improved the performance\n \n\ntweaks uh it improved the performance\nbetter. Again remember this is on 100\n\nbetter. Again remember this is on 100\n \n\nbetter. Again remember this is on 100\ndocuments so the accuracy is looking\n\ndocuments so the accuracy is looking\n \n\ndocuments so the accuracy is looking\nhigh but if your document pool increases\n\nhigh but if your document pool increases\n \n\nhigh but if your document pool increases\nthat will come down a bit but in\n\nthat will come down a bit but in\n \n\nthat will come down a bit but in\ncomparison to where we were before we\n\ncomparison to where we were before we\n \n\ncomparison to where we were before we\nsaw improvement and and that is where\n\nsaw improvement and and that is where\n \n\nsaw improvement and and that is where\nthe small uh tweaks come into play which\n\nthe small uh tweaks come into play which\n \n\nthe small uh tweaks come into play which\nwould be very very very helpful to you\n\nwould be very very very helpful to you\n \n\nwould be very very very helpful to you\nwhen you're putting a a system um a\n\nwhen you're putting a a system um a\n \n\nwhen you're putting a a system um a\ngraph rag or a rack system into\n\ngraph rag or a rack system into\n \n\ngraph rag or a rack system into\nproduction. production. production. The other is from a latency standpoint. The other is from a latency standpoint. The other is from a latency standpoint. Um so if your graph gets bigger and\n\nUm so if your graph gets bigger and\n \n\nUm so if your graph gets bigger and\nbigger now you're talking about a\n\nbigger now you're talking about a\n \n\nbigger now you're talking about a\nnetwork which which goes into millions\n\nnetwork which which goes into millions\n \n\nnetwork which which goes into millions\nor billions of parameter and uh or\n\nor billions of parameter and uh or\n \n\nor billions of parameter and uh or\nmillions and billions of nodes. Now how\n\nmillions and billions of nodes. Now how\n \n\nmillions and billions of nodes. Now how\ndo you how do you do search in um in\n\ndo you how do you do search in um in\n \n\ndo you how do you do search in um in\nthose millions and billions u\n\nthose millions and billions u\n \n\nthose millions and billions u\nin the graph that has got millions or\n\nin the graph that has got millions or\n \n\nin the graph that has got millions or\nbillions of nodes and that is where\n\nbillions of nodes and that is where\n \n\nbillions of nodes and that is where\nacceleration comes into play. So with\n\nacceleration comes into play. So with\n \n\nacceleration comes into play. So with\nwith with cool graph which is now\n\nwith with cool graph which is now\n \n\nwith with cool graph which is now\navailable through network X. So network\n\navailable through network X. So network\n \n\navailable through network X. So network\nX is also al also a pip install library. X is also al also a pip install library. X is also al also a pip install library. Uh anyone who used network X here right\n\nUh anyone who used network X here right\n \n\nUh anyone who used network X here right\nfew okay um so network is also a pip\n\nfew okay um so network is also a pip\n \n\nfew okay um so network is also a pip\ninstall library under the hood um it\n\ninstall library under the hood um it\n \n\ninstall library under the hood um it\nuses um acceleration and if you see a\n\nuses um acceleration and if you see a\n \n\nuses um acceleration and if you see a\nfew of the algorithms uh we um we we did\n\nfew of the algorithms uh we um we we did\n \n\nfew of the algorithms uh we um we we did\na performance test on that and um you\n\na performance test on that and um you\n \n\na performance test on that and um you\ncan see the amount of latency in terms\n\ncan see the amount of latency in terms\n \n\ncan see the amount of latency in terms\nof overall execution reducing\n\nof overall execution reducing\n \n\nof overall execution reducing\ndrastically. So that is where you can\n\ndrastically. So that is where you can\n \n\ndrastically. So that is where you can\nagain small tweaks which will lead you\n\nagain small tweaks which will lead you\n \n\nagain small tweaks which will lead you\nto better results. So these are two\n\nto better results. So these are two\n \n\nto better results. So these are two\nthings that we experimented which led us\n\nthings that we experimented which led us\n \n\nthings that we experimented which led us\nto to better results in terms of\n\nto to better results in terms of\n \n\nto to better results in terms of\naccuracy as well as reducing the overall\n\naccuracy as well as reducing the overall\n \n\naccuracy as well as reducing the overall\nlatency and these are small tweaks and\n\nlatency and these are small tweaks and\n \n\nlatency and these are small tweaks and\nit it leads us to better results. So then the question obviously is should\n\nSo then the question obviously is should\n \n\nSo then the question obviously is should\nI uh use graph or should I use semantic\n\nI uh use graph or should I use semantic\n \n\nI uh use graph or should I use semantic\num based rack system or should I use\n\num based rack system or should I use\n \n\num based rack system or should I use\nhybrid and I'm going to give you the\n\nhybrid and I'm going to give you the\n \n\nhybrid and I'm going to give you the\ndiplomatic answer. It depends but but\n\ndiplomatic answer. It depends but but\n \n\ndiplomatic answer. It depends but but\nthere are few things I would like to you\n\nthere are few things I would like to you\n \n\nthere are few things I would like to you\nguys to take home to to um um which will\n\nguys to take home to to um um which will\n \n\nguys to take home to to um um which will\nhelp you to come up to a decision so\n\nhelp you to come up to a decision so\n \n\nhelp you to come up to a decision so\nthat you can make an educated guess that\n\nthat you can make an educated guess that\n \n\nthat you can make an educated guess that\nfor this use case that I'm working on a\n\nfor this use case that I'm working on a\n \n\nfor this use case that I'm working on a\nrack system would solve the problem I\n\nrack system would solve the problem I\n \n\nrack system would solve the problem I\ndon't need a graph and vice versa or I\n\ndon't need a graph and vice versa or I\n \n\ndon't need a graph and vice versa or I\nneed a hybrid approach so it depends on\n\nneed a hybrid approach so it depends on\n \n\nneed a hybrid approach so it depends on\ntwo two factors one is your data um\n\ntwo two factors one is your data um\n \n\ntwo two factors one is your data um\ntraditionally if you look at retail data\n\ntraditionally if you look at retail data\n \n\ntraditionally if you look at retail data\nif you look at FSI data if you look at\n\nif you look at FSI data if you look at\n \n\nif you look at FSI data if you look at\nemployee database of companies those\n\nemployee database of companies those\n \n\nemployee database of companies those\nhave a really good structure structure\n\nhave a really good structure structure\n \n\nhave a really good structure structure\ndefined. So those kind of data set\n\ndefined. So those kind of data set\n \n\ndefined." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nSo those kind of data set\nbecomes really good use cases for graph\n\nbecomes really good use cases for graph\n \n\nbecomes really good use cases for graph\nbased system and the other thing you\n\nbased system and the other thing you\n \n\nbased system and the other thing you\nthink about is even if you have\n\nth..." }, { "role": "assistant", "content": "So those kind of data set\nbecomes really good use cases for graph\n\nbecomes really good use cases for graph\n \n\nbecomes really good use cases for graph\nbased system and the other thing you\n\nbased system and the other thing you\n \n\nbased system and the other thing you\nthink about is even if you have\n\nthink about is even if you have\n \n\nthink about is even if you have\nunstructured data can you create a good\n\nunstructured data can you create a good\n \n\nunstructured data can you create a good\nuh graph knowledge graph out of it. If\n\nuh graph knowledge graph out of it. If\n \n\nuh graph knowledge graph out of it. If\nthe answer is yes then it's worthwhile\n\nthe answer is yes then it's worthwhile\n \n\nthe answer is yes then it's worthwhile\nexperimenting uh um with u to go the\n\nexperimenting uh um with u to go the\n \n\nexperimenting uh um with u to go the\ngraph path and it depend it will depend\n\ngraph path and it depend it will depend\n \n\ngraph path and it depend it will depend\non the application and use case. So if\n\non the application and use case. So if\n \n\non the application and use case. So if\nyour use case requires to um to\n\nyour use case requires to um to\n \n\nyour use case requires to um to\nunderstand the complex relationship and\n\nunderstand the complex relationship and\n \n\nunderstand the complex relationship and\nthen extract that information u um to\n\nthen extract that information u um to\n \n\nthen extract that information u um to\nfor the response that you um for the\n\nfor the response that you um for the\n \n\nfor the response that you um for the\nquestions that you are asking only then\n\nquestions that you are asking only then\n \n\nquestions that you are asking only then\nit makes sense uh to use graph because\n\nit makes sense uh to use graph because\n \n\nit makes sense uh to use graph because\nremember these are compute heavy uh\n\nremember these are compute heavy uh\n \n\nremember these are compute heavy uh\nheavy systems. So you need to make sure\n\nheavy systems. So you need to make sure\n \n\nheavy systems. So you need to make sure\nthat these things are taken care of. I\n\nthat these things are taken care of. I\n \n\nthat these things are taken care of. I\nam running out of time I think but u as\n\nam running out of time I think but u as\n \n\nam running out of time I think but u as\nI said before all these things that I\n\nI said before all these things that I\n \n\nI said before all these things that I\ntalked about I gave you a 10,000 ft view\n\ntalked about I gave you a 10,000 ft view\n \n\ntalked about I gave you a 10,000 ft view\nbut if you want to get a 100 ft view\n\nbut if you want to get a 100 ft view\n \n\nbut if you want to get a 100 ft view\nwhere you are coding into into things\n\nwhere you are coding into into things\n \n\nwhere you are coding into into things\nall these things is available on GitHub\n\nall these things is available on GitHub\n \n\nall these things is available on GitHub\neven the finetuning of the llama 1.1\n\neven the finetuning of the llama 1.1\n \n\neven the finetuning of the llama 1.1\nLaura model and we had a workshop a\n\nLaura model and we had a workshop a\n \n\nLaura model and we had a workshop a\ntwo-hour workshop so I gave you a\n\ntwo-hour workshop so I gave you a\n \n\ntwo-hour workshop so I gave you a\n20-minute talk but this whole workshop\n\n20-minute talk but this whole workshop\n \n\n20-minute talk but this whole workshop\nis covered uh in two hours as well and\n\nis covered uh in two hours as well and\n \n\nis covered uh in two hours as well and\nlastly um join our developer programs we\n\nlastly um join our developer programs we\n \n\nlastly um join our developer programs we\ndo release all these things on a regular\n\ndo release all these things on a regular\n \n\ndo release all these things on a regular\nbasis you if you join the mailing list\n\nbasis you if you join the mailing list\n \n\nbasis you if you join the mailing list\nyou get this information based on your\n\nyou get this information based on your\n \n\nyou get this information based on your\ninterest and as u my colleague mentioned\n\ninterest and as u my colleague mentioned\n \n\ninterest and as u my colleague mentioned\nI will be across uh the hall at Neo4j\n\nI will be across uh the hall at Neo4j\n \n\nI will be across uh the hall at Neo4j\nbooth uh to answer questions if any I\n\nbooth uh to answer questions if any I\n \n\nbooth uh to answer questions if any I\nwould love to interact with you and see\n\nwould love to interact with you and see\n \n\nwould love to interact with you and see\nif you have any qu uh any questions and\n\nif you have any qu uh any questions and\n \n\nif you have any qu uh any questions and\nI can answer those questions. Thank you\n\nI can answer those questions. Thank you\n \n\nI can answer those questions. Thank you\nfor your time. for your time. for your time. [Applause]\n\n[Applause]\n \n\n[Applause]\n[Music]" } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\n[Music]\n\n[Music]\n \n\n[Music]\nhello everybody uh my name is Nicholas\n\nhello everybody uh my name is Nicholas\n \n\nhello everybody uh my name is Nicholas\nschlapfer I'm an AI engineer I think I'm\n\nschlapfer I'm an AI engineer I think I'm\n \n\nschlapfer I'm an AI engineer I think I'm\nat the right place um to..." }, { "role": "assistant", "content": "[Music]\n\n[Music]\n \n\n[Music]\nhello everybody uh my name is Nicholas\n\nhello everybody uh my name is Nicholas\n \n\nhello everybody uh my name is Nicholas\nschlapfer I'm an AI engineer I think I'm\n\nschlapfer I'm an AI engineer I think I'm\n \n\nschlapfer I'm an AI engineer I think I'm\nat the right place um today I'm going to\n\nat the right place um today I'm going to\n \n\nat the right place um today I'm going to\nbe pronouncing uh a new product we've\n\nbe pronouncing uh a new product we've\n \n\nbe pronouncing uh a new product we've\nbeen working on at\n\nbeen working on at\n \n\nbeen working on at\nhyperspace um a little bit about us uh\n\nhyperspace um a little bit about us uh\n \n\nhyperspace um a little bit about us uh\nwe are a decentralized AI network uh we\n\nwe are a decentralized AI network uh we\n \n\nwe are a decentralized AI network uh we\nhave no gpus um we're building a\n\nhave no gpus um we're building a\n \n\nhave no gpus um we're building a\ncommunity who takes uh the resources\n\ncommunity who takes uh the resources\n \n\ncommunity who takes uh the resources\nfrom their personal computer and\n\nfrom their personal computer and\n \n\nfrom their personal computer and\ncontributes uh to our decentralized\n\ncontributes uh to our decentralized\n \n\ncontributes uh to our decentralized\nNetwork um we have a product currently\n\nNetwork um we have a product currently\n \n\nNetwork um we have a product currently\nout called aios um you can download it\n\nout called aios um you can download it\n \n\nout called aios um you can download it\nfor Windows and Mac it uses llama CPP\n\nfor Windows and Mac it uses llama CPP\n \n\nfor Windows and Mac it uses llama CPP\nand does inference so technically you\n\nand does inference so technically you\n \n\nand does inference so technically you\ncan download it get inference from\n\ncan download it get inference from\n \n\ncan download it get inference from\nsomebody in Belgium and you can have a\n\nsomebody in Belgium and you can have a\n \n\nsomebody in Belgium and you can have a\nchat experience like that\n\nchat experience like that\n \n\nchat experience like that\nso we're hoping to oh actually sorry um\n\nso we're hoping to oh actually sorry um\n \n\nso we're hoping to oh actually sorry um\nwe really believe in uh diverse models\n\nwe really believe in uh diverse models\n \n\nwe really believe in uh diverse models\nuh we think uh not having um just one\n\nuh we think uh not having um just one\n \n\nuh we think uh not having um just one\nbig close Source model is the answer for\n\nbig close Source model is the answer for\n \n\nbig close Source model is the answer for\nthe best AI experience um having a\n\nthe best AI experience um having a\n \n\nthe best AI experience um having a\nmixture of a bunch of great experts will\n\nmixture of a bunch of great experts will\n \n\nmixture of a bunch of great experts will\nprovide the best AI\n\nprovide the best AI\n \n\nprovide the best AI\nexperience um so that gets us to our\n\nexperience um so that gets us to our\n \n\nexperience um so that gets us to our\nself-titled product hyperspace which\n\nself-titled product hyperspace which\n \n\nself-titled product hyperspace which\nwill be built on our Network\n\nwill be built on our Network\n \n\nwill be built on our Network\num this is a really interesting product\n\num this is a really interesting product\n \n\num this is a really interesting product\nit's a mix of prompt engineering um\n\nit's a mix of prompt engineering um\n \n\nit's a mix of prompt engineering um\nvisual uh react flow uh python execution\n\nvisual uh react flow uh python execution\n \n\nvisual uh react flow uh python execution\nand uh rag like uh web\n\nand uh rag like uh web\n \n\nand uh rag like uh web\nbrowsing so for the first thing with\n\nbrowsing so for the first thing with\n \n\nbrowsing so for the first thing with\nthis product we wanted to build a uh\n\nthis product we wanted to build a uh\n \n\nthis product we wanted to build a uh\nfine-tune model that outputed a gentic\n\nfine-tune model that outputed a gentic\n \n\nfine-tune model that outputed a gentic\nplanning experiences so you put in a\n\nplanning experiences so you put in a\n \n\nplanning experiences so you put in a\nquery you get in a dag or you get out a\n\nquery you get in a dag or you get out a\n \n\nquery you get in a dag or you get out a\ndag in a Json format and this is a a\n\ndag in a Json format and this is a a\n \n\ndag in a Json format and this is a a\nmethodical plan from that\n\nmethodical plan from that\n \n\nmethodical plan from that\nquery um we also wanted to have a\n\nquery um we also wanted to have a\n \n\nquery um we also wanted to have a\nprimitive of having a in-house web\n\nprimitive of having a in-house web\n \n\nprimitive of having a in-house web\nscraping experience for llms so we're\n\nscraping experience for llms so we're\n \n\nscraping experience for llms so we're\nusing Puppeteer and beautiful soup to\n\nusing Puppeteer and beautiful soup to\n \n\nusing Puppeteer and beautiful soup to\nscrape websites convert that uh HTML\n\nscrape websites convert that uh HTML\n \n\nscrape websites convert that uh HTML\ninto uh a markdown something easier to\n\ninto uh a markdown something easier to\n \n\ninto uh a markdown something easier to\nread for llms to\n\nread for llms to\n \n\nread for llms to\ndigest um kind of the product we're\n\ndigest um kind of the product we're\n \n\ndigest um kind of the product we're\ngoing to talk about is uh we have a node\n\ngoing to talk about is uh we have a node\n \n\ngoing to talk about is uh we have a node\neditor in a terminal and this is going\n\neditor in a terminal and this is going\n \n\neditor in a terminal and this is going\nto be a power tool for the power users\n\nto be a power tool for the power users\n \n\nto be a power tool for the power users\num this node editor will allow you to\n\num this node editor will allow you to\n \n\num this node editor will allow you to\nchange each node in the react flow to uh\n\nchange each node in the react flow to uh\n \n\nchange each node in the react flow to uh\nfit your needs and this is coming from\n\nfit your needs and this is coming from\n \n\nfit your needs and this is coming from\nthe dag or dag orchestration model we\n\nthe dag or dag orchestration model we\n \n\nthe dag or dag orchestration model we\nhave um here's a little video demo of uh\n\nhave um here's a little video demo of uh\n \n\nhave um here's a little video demo of uh\nour product I recorded it yesterday in\n\nour product I recorded it yesterday in\n \n\nour product I recorded it yesterday in\nthe hotel so I'm sorry for the audio and\n\nthe hotel so I'm sorry for the audio and\n \n\nthe hotel so I'm sorry for the audio and\nhere we go\n\n \n\n[Music]\n\n[Music]\n \n\n[Music]\nwelcome to a very early look at\n\nwelcome to a very early look at\n \n\nwelcome to a very early look at\nhyperspace let's begin with a sample\n\n \n\nquery once you submit your query you're\n\nquery once you submit your query you're\n \n\nquery once you submit your query you're\nbrought into our node editor view where\n\nbrought into our node editor view where\n \n\nbrought into our node editor view where\neach node is streamed in from our dag\n\neach node is streamed in from our dag\n \n\neach node is streamed in from our dag\norchestration model hyper engine V3\n\n \n\nwe want to emphasize that the user still\n\nwe want to emphasize that the user still\n \n\nwe want to emphasize that the user still\nin full control they can edit the title\n\nin full control they can edit the title\n \n\nin full control they can edit the title\ntask description and expected\n\ntask description and expected\n \n\ntask description and expected\noutput they also have the freedom to add\n\noutput they also have the freedom to add\n \n\noutput they also have the freedom to add\nas many nodes as they\n\n \n\nlike let's\n\n \n\nexecute now that our outputs are done we\n\nexecute now that our outputs are done we\n \n\nexecute now that our outputs are done we\ncan talk a little bit about the outputs\n\ncan talk a little bit about the outputs\n \n\ncan talk a little bit about the outputs\neach output is coming from each node and\n\neach output is coming from each node and\n \n\neach output is coming from each node and\neach node is creating a query based on\n\neach node is creating a query based on\n \n\neach node is creating a query based on\nthat task it's combining the overall\n\nthat task it's combining the overall\n \n\nthat task it's combining the overall\ngoal the local goal and whatever is\n\ngoal the local goal and whatever is\n \n\ngoal the local goal and whatever is\nhappened in the previous\n\nhappened in the previous\n \n\nhappened in the previous\nnode we're using a reasoning model and a\n\nnode we're using a reasoning model and a\n \n\nnode we're using a reasoning model and a\nsummarization model for reasoning we're\n\nsummarization model for reasoning we're\n \n\nsummarization model for reasoning we're\nusing quen 2 instruct and then for\n\nusing quen 2 instruct and then for\n \n\nusing quen 2 instruct and then for\nsummarization we're using llama\n\nsummarization we're using llama\n \n\nsummarization we're using llama\n370b this helps provide a diverse set of\n\n370b this helps provide a diverse set of\n \n\n370b this helps provide a diverse set of\nsynthesized answers\n\nsynthesized answers\n \n\nsynthesized answers\nfor the outputs that do have python we\n\nfor the outputs that do have python we\n \n\nfor the outputs that do have python we\ncan go ahead and run them and see them\n\ncan go ahead and run them and see them\n \n\ncan go ahead and run them and see them\nin our\n\n \n\nterminal terminal will automatically\n\nterminal terminal will automatically\n \n\nterminal terminal will automatically\nopen up with the\n\nopen up with the\n \n\nopen up with the\noutput we want to provide the groundwork\n\noutput we want to provide the groundwork\n \n\noutput we want to provide the groundwork\nfor agentic behavior in the future by\n\nfor agentic behavior in the future by\n \n\nfor agentic behavior in the future by\nproviding these core\n\nproviding these core\n \n\nproviding these core\nPrimitives over here on our right we\n\nPrimitives over here on our right we\n \n\nPrimitives over here on our right we\nhave our virtual file system that\n\nhave our virtual file system that\n \n\nhave our virtual file system that\nchanges our\n\nchanges our\n \n\nchanges our\ndirectory we're trying to build out all\n\ndirectory we're trying to build out all\n \n\ndirectory we're trying to build out all\nThe Primitives to what an agent would\n\nThe Primitives to what an agent would\n \n\nThe Primitives to what an agent would\nneed memory python execution planning\n\nneed memory python execution planning\n \n\nneed memory python execution planning\nand code\n\nand code\n \n\nand code\ngeneration thank you so much for\n\ngeneration thank you so much for\n \n\ngeneration thank you so much for\nwatching we're very excited to get this\n\nwatching we're very excited to get this\n \n\nwatching we're very excited to get this\nout in the coming\n\n \n\nweeks um happy to announce that it'll be\n\nweeks um happy to announce that it'll be\n \n\nweeks um happy to announce that it'll be\navailable later this week via weight\n\navailable later this week via weight\n \n\navailable later this week via weight\nlist so go ahead and pay attention to\n\nlist so go ahead and pay attention to\n \n\nlist so go ahead and pay attention to\nour Twitter hyperspace AI all right\n\nour Twitter hyperspace AI all right\n \n\nour Twitter hyperspace AI all right\nthank you that's my time\n\n \n\n[Music]" } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\n[Music]\n\n \n\nHi everyone, thank you for having us. Hi everyone, thank you for having us. Hi everyone, thank you for having us. I'm Infant, director of engineering at\n\nI'm Infant, director of engineering at\n \n\nI'm Infant, director of engineering at\nBlack Rockck. This is my colleague\n\nBlack Rockck. Thi..." }, { "role": "assistant", "content": "[Music]\n\n \n\nHi everyone, thank you for having us. Hi everyone, thank you for having us. Hi everyone, thank you for having us. I'm Infant, director of engineering at\n\nI'm Infant, director of engineering at\n \n\nI'm Infant, director of engineering at\nBlack Rockck. This is my colleague\n\nBlack Rockck. This is my colleague\n \n\nBlack Rockck. This is my colleague\nWyber, principal engineer and we both\n\nWyber, principal engineer and we both\n \n\nWyber, principal engineer and we both\nwork for the data teams at Black Rockck. work for the data teams at Black Rockck. work for the data teams at Black Rockck. And today we're going to talk about how\n\nAnd today we're going to talk about how\n \n\nAnd today we're going to talk about how\nwe can scale building custom\n\nwe can scale building custom\n \n\nwe can scale building custom\napplications in Black Rockck. applications in Black Rockck. applications in Black Rockck. Specifically, we're talking about like\n\nSpecifically, we're talking about like\n \n\nSpecifically, we're talking about like\nAI applications and knowledge apps at\n\nAI applications and knowledge apps at\n \n\nAI applications and knowledge apps at\nBlack Rockck, right? So, just to level\n\nBlack Rockck, right? So, just to level\n \n\nBlack Rockck, right? So, just to level\nset before I get into the uh details. set before I get into the uh details. set before I get into the uh details. So, Black Rockck is an asset management\n\nSo, Black Rockck is an asset management\n \n\nSo, Black Rockck is an asset management\nfirm, the world's largest asset manager. firm, the world's largest asset manager. firm, the world's largest asset manager. What we do is our portfolio managers,\n\nWhat we do is our portfolio managers,\n \n\nWhat we do is our portfolio managers,\nanalysts get a torrent of information on\n\nanalysts get a torrent of information on\n \n\nanalysts get a torrent of information on\na daily basis. They synthesize this\n\na daily basis. They synthesize this\n \n\na daily basis. They synthesize this\ninformation, they develop an investment\n\ninformation, they develop an investment\n \n\ninformation, they develop an investment\nstrategy and then they rebalance your\n\nstrategy and then they rebalance your\n \n\nstrategy and then they rebalance your\nportfolios uh which ultimately results\n\nportfolios uh which ultimately results\n \n\nportfolios uh which ultimately results\nin a particular trade. Now the\n\nin a particular trade. Now the\n \n\nin a particular trade. Now the\ninvestment operations teams you can\n\ninvestment operations teams you can\n \n\ninvestment operations teams you can\nthink of that as the teams that are the\n\nthink of that as the teams that are the\n \n\nthink of that as the teams that are the\nbackbone or the engine that makes sure\n\nbackbone or the engine that makes sure\n \n\nbackbone or the engine that makes sure\nthat all of the activities that the\n\nthat all of the activities that the\n \n\nthat all of the activities that the\ninvestment managers actually perform on\n\ninvestment managers actually perform on\n \n\ninvestment managers actually perform on\na day-to-day basis like runs smoothly\n\na day-to-day basis like runs smoothly\n \n\na day-to-day basis like runs smoothly\nright so these teams are kind of\n\nright so these teams are kind of\n \n\nright so these teams are kind of\nresponsible for like acquiring the data\n\nresponsible for like acquiring the data\n \n\nresponsible for like acquiring the data\nthat you kind of need right uh to\n\nthat you kind of need right uh to\n \n\nthat you kind of need right uh to\nactually executing a trade running\n\nactually executing a trade running\n \n\nactually executing a trade running\nthrough compliance all the way to like\n\nthrough compliance all the way to like\n \n\nthrough compliance all the way to like\nall of the post- trading activities\n\nall of the post- trading activities\n \n\nall of the post- trading activities\nright so all of these teams actually\n\nright so all of these teams actually\n \n\nright so all of these teams actually\nhave to build these internal tools that\n\nhave to build these internal tools that\n \n\nhave to build these internal tools that\nare actually fairly complex for each of\n\nare actually fairly complex for each of\n \n\nare actually fairly complex for each of\ntheir domains. Right? So building apps\n\ntheir domains. Right? So building apps\n \n\ntheir domains. Right? So building apps\nand pushing out these apps uh relatively\n\nand pushing out these apps uh relatively\n \n\nand pushing out these apps uh relatively\nquickly is like of utmost important to\n\nquickly is like of utmost important to\n \n\nquickly is like of utmost important to\nus. Right? So if you move on to the next\n\nus. Right? So if you move on to the next\n \n\nus. Right? So if you move on to the next\nslide again if you actually classify\n\nslide again if you actually classify\n \n\nslide again if you actually classify\nwhat kind of apps we are talking about\n\nwhat kind of apps we are talking about\n \n\nwhat kind of apps we are talking about\nwhat you'll see is that it kind of falls\n\nwhat you'll see is that it kind of falls\n \n\nwhat you'll see is that it kind of falls\ninto like four different buckets right? into like four different buckets right? into like four different buckets right? One is everything to do with document\n\nOne is everything to do with document\n \n\nOne is everything to do with document\nextraction. So I have an app I kind of\n\nextraction. So I have an app I kind of\n \n\nextraction. So I have an app I kind of\nwant to like extract entities out of it\n\nwant to like extract entities out of it\n \n\nwant to like extract entities out of it\nin that bucket. Second has to do\n\nin that bucket. Second has to do\n \n\nin that bucket. Second has to do\neverything with like hey I kind of want\n\neverything with like hey I kind of want\n \n\neverything with like hey I kind of want\nto define a complex uh workflow or an\n\nto define a complex uh workflow or an\n \n\nto define a complex uh workflow or an\nautomation. So I could have a case where\n\nautomation. So I could have a case where\n \n\nautomation. So I could have a case where\nI kind of want to run through X number\n\nI kind of want to run through X number\n \n\nI kind of want to run through X number\nof steps and then integrate to my\n\nof steps and then integrate to my\n \n\nof steps and then integrate to my\ndownstream systems and then you have the\n\ndownstream systems and then you have the\n \n\ndownstream systems and then you have the\nnormal like Q&A type systems that you\n\nnormal like Q&A type systems that you\n \n\nnormal like Q&A type systems that you\nlook at like this is your chat\n\nlook at like this is your chat\n \n\nlook at like this is your chat\ninterfaces and finally like the the\n\ninterfaces and finally like the the\n \n\ninterfaces and finally like the the\nagentic systems right so in each of\n\nagentic systems right so in each of\n \n\nagentic systems right so in each of\nthese domains what we see is u we have\n\nthese domains what we see is u we have\n \n\nthese domains what we see is u we have\nthis like big opportunity to leverage\n\nthis like big opportunity to leverage\n \n\nthis like big opportunity to leverage\nyour models and LLMs to either augment\n\nyour models and LLMs to either augment\n \n\nyour models and LLMs to either augment\nour existing systems uh or like kind of\n\nour existing systems uh or like kind of\n \n\nour existing systems uh or like kind of\nlike supercharge those right so that\n\nlike supercharge those right so that\n \n\nlike supercharge those right so that\nthat is like the domain we are speaking\n\nthat is like the domain we are speaking\n \n\nthat is like the domain we are speaking\nabout. So I'll move quickly to one\n\nabout. So I'll move quickly to one\n \n\nabout. So I'll move quickly to one\nparticular use case. So this this is a\n\nparticular use case. So this this is a\n \n\nparticular use case." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nSo this this is a\nuse case that came to us like about like\n\nuse case that came to us like about like\n \n\nuse case that came to us like about like\n3 to 4 months back right and we have a\n\n3 to 4 months back right and we have a\n \n\n3 to 4 months back right and we have a\nteam within the investment operati..." }, { "role": "assistant", "content": "So this this is a\nuse case that came to us like about like\n\nuse case that came to us like about like\n \n\nuse case that came to us like about like\n3 to 4 months back right and we have a\n\n3 to 4 months back right and we have a\n \n\n3 to 4 months back right and we have a\nteam within the investment operations\n\nteam within the investment operations\n \n\nteam within the investment operations\nspace it's known as a new issue\n\nspace it's known as a new issue\n \n\nspace it's known as a new issue\noperations team right so this team is\n\noperations team right so this team is\n \n\noperations team right so this team is\nkind of responsible for setting up\n\nkind of responsible for setting up\n \n\nkind of responsible for setting up\nsecurities uh whenever there is like a\n\nsecurities uh whenever there is like a\n \n\nsecurities uh whenever there is like a\nmarket event right so a company goes IPO\n\nmarket event right so a company goes IPO\n \n\nmarket event right so a company goes IPO\nor like there is like a stock split for\n\nor like there is like a stock split for\n \n\nor like there is like a stock split for\na particular organization right the team\n\na particular organization right the team\n \n\na particular organization right the team\nactually has to take the security and\n\nactually has to take the security and\n \n\nactually has to take the security and\nthey have to set it up in our internal\n\nthey have to set it up in our internal\n \n\nthey have to set it up in our internal\nsystems before our portfolio managers or\n\nsystems before our portfolio managers or\n \n\nsystems before our portfolio managers or\ntraders can action upon it, right? So,\n\ntraders can action upon it, right? So,\n \n\ntraders can action upon it, right? So,\nwe kind of have to build this tool for\n\nwe kind of have to build this tool for\n \n\nwe kind of have to build this tool for\nthe investment operations team, right? the investment operations team, right? the investment operations team, right? To set up a particular security. This is\n\nTo set up a particular security. This is\n \n\nTo set up a particular security. This is\nlike actually honestly this is like a\n\nlike actually honestly this is like a\n \n\nlike actually honestly this is like a\nsuper simplified version of what\n\nsuper simplified version of what\n \n\nsuper simplified version of what\nhappens. But a super high level, we have\n\nhappens. But a super high level, we have\n \n\nhappens. But a super high level, we have\nto build an app that is able to like\n\nto build an app that is able to like\n \n\nto build an app that is able to like\ningest your prospectors or a term sheet. ingest your prospectors or a term sheet. ingest your prospectors or a term sheet. It pushes it through a particular\n\nIt pushes it through a particular\n \n\nIt pushes it through a particular\npipeline, right? U then you talk to your\n\npipeline, right? U then you talk to your\n \n\npipeline, right? U then you talk to your\ndomain experts and these are like your\n\ndomain experts and these are like your\n \n\ndomain experts and these are like your\nbusiness teams, your equity teams, ETF\n\nbusiness teams, your equity teams, ETF\n \n\nbusiness teams, your equity teams, ETF\nteams, etc. They actually know how to\n\nteams, etc. They actually know how to\n \n\nteams, etc. They actually know how to\nset up these complex instruments. you\n\nset up these complex instruments. you\n \n\nset up these complex instruments. you\nget some kind of structured output and\n\nget some kind of structured output and\n \n\nget some kind of structured output and\nnow that team works with the engineering\n\nnow that team works with the engineering\n \n\nnow that team works with the engineering\nteams to actually build this\n\nteams to actually build this\n \n\nteams to actually build this\ntransformation logic and the like and\n\ntransformation logic and the like and\n \n\ntransformation logic and the like and\nthen integrate it with your downstream\n\nthen integrate it with your downstream\n \n\nthen integrate it with your downstream\napplications. So you can see that this\n\napplications. So you can see that this\n \n\napplications. So you can see that this\nprocess actually takes a long time,\n\nprocess actually takes a long time,\n \n\nprocess actually takes a long time,\nright? So building an app and then\n\nright? So building an app and then\n \n\nright? So building an app and then\nyou're introducing new model providers,\n\nyou're introducing new model providers,\n \n\nyou're introducing new model providers,\nyou're trying to put in like new\n\nyou're trying to put in like new\n \n\nyou're trying to put in like new\nstrategies, the lot of challenges to get\n\nstrategies, the lot of challenges to get\n \n\nstrategies, the lot of challenges to get\nan single app out, right? We tried this\n\nan single app out, right? We tried this\n \n\nan single app out, right? We tried this\nwith agentic systems doesn't quite work\n\nwith agentic systems doesn't quite work\n \n\nwith agentic systems doesn't quite work\nright now because of the complexity and\n\nright now because of the complexity and\n \n\nright now because of the complexity and\nthe the domain knowledge that's imbued\n\nthe the domain knowledge that's imbued\n \n\nthe the domain knowledge that's imbued\nin the human head, right? So the big\n\nin the human head, right? So the big\n \n\nin the human head, right? So the big\nchallenges with scale are again these\n\nchallenges with scale are again these\n \n\nchallenges with scale are again these\nthree categories, right? One is we're\n\nthree categories, right? One is we're\n \n\nthree categories, right?" } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nOne is we're\nspending a lot of time with our domain\n\nspending a lot of time with our domain\n \n\nspending a lot of time with our domain\nexperts prompt engineering right so in\n\nexperts prompt engineering right so in\n \n\nexperts prompt engineering right so in\nthe first phase where we have to extract\n\nthe..." }, { "role": "assistant", "content": "One is we're\nspending a lot of time with our domain\n\nspending a lot of time with our domain\n \n\nspending a lot of time with our domain\nexperts prompt engineering right so in\n\nexperts prompt engineering right so in\n \n\nexperts prompt engineering right so in\nthe first phase where we have to extract\n\nthe first phase where we have to extract\n \n\nthe first phase where we have to extract\nthese documents right they're very\n\nthese documents right they're very\n \n\nthese documents right they're very\ncomplex right your prompt itself in our\n\ncomplex right your prompt itself in our\n \n\ncomplex right your prompt itself in our\nsimplest case like started with like a\n\nsimplest case like started with like a\n \n\nsimplest case like started with like a\ncouple of sentences before you knew it\n\ncouple of sentences before you knew it\n \n\ncouple of sentences before you knew it\nyou're trying to describe this financial\n\nyou're trying to describe this financial\n \n\nyou're trying to describe this financial\ninstrument and it is like three\n\ninstrument and it is like three\n \n\ninstrument and it is like three\nparagraphs long right uh so there's this\n\nparagraphs long right uh so there's this\n \n\nparagraphs long right uh so there's this\nchallenge of like hey I have to iterate\n\nchallenge of like hey I have to iterate\n \n\nchallenge of like hey I have to iterate\nover these prompts I have to version and\n\nover these prompts I have to version and\n \n\nover these prompts I have to version and\ncompare these promps how do I manage it\n\ncompare these promps how do I manage it\n \n\ncompare these promps how do I manage it\neffectively and I think even the\n\neffectively and I think even the\n \n\neffectively and I think even the\nprevious speaker had mentioned you kind\n\nprevious speaker had mentioned you kind\n \n\nprevious speaker had mentioned you kind\nof need to eval and have this data set\n\nof need to eval and have this data set\n \n\nof need to eval and have this data set\nhow how good is your prompt performing\n\nhow how good is your prompt performing\n \n\nhow how good is your prompt performing\nso that's the first set of challenges in\n\nso that's the first set of challenges in\n \n\nso that's the first set of challenges in\ncreating like AI apps itself like how\n\ncreating like AI apps itself like how\n \n\ncreating like AI apps itself like how\nare you going to manage this in what\n\nare you going to manage this in what\n \n\nare you going to manage this in what\ndirection second set of challenges is\n\ndirection second set of challenges is\n \n\ndirection second set of challenges is\naround like LLM strategies right what I\n\naround like LLM strategies right what I\n \n\naround like LLM strategies right what I\nmean by this is like when you're\n\nmean by this is like when you're\n \n\nmean by this is like when you're\nbuilding an AI app so to speak you have\n\nbuilding an AI app so to speak you have\n \n\nbuilding an AI app so to speak you have\nto choose what strategy am I going to\n\nto choose what strategy am I going to\n \n\nto choose what strategy am I going to\nuse like a rag based approach right or\n\nuse like a rag based approach right or\n \n\nuse like a rag based approach right or\nam I going to use a chain of\n\nam I going to use a chain of\n \n\nam I going to use a chain of\nthought-based approach even for a simple\n\nthought-based approach even for a simple\n \n\nthought-based approach even for a simple\ntask of like data extraction depending\n\ntask of like data extraction depending\n \n\ntask of like data extraction depending\non what your instrument is this actually\n\non what your instrument is this actually\n \n\non what your instrument is this actually\nvaries uh very highly right if you take\n\nvaries uh very highly right if you take\n \n\nvaries uh very highly right if you take\nlike an investment corporate bond like\n\nlike an investment corporate bond like\n \n\nlike an investment corporate bond like\nthe vanilla one is fairly simple I can\n\nthe vanilla one is fairly simple I can\n \n\nthe vanilla one is fairly simple I can\ndo this with like in context positive\n\ndo this with like in context positive\n \n\ndo this with like in context positive\nmodel I'm able to get my stuff back if\n\nmodel I'm able to get my stuff back if\n \n\nmodel I'm able to get my stuff back if\nthe document size is small right some\n\nthe document size is small right some\n \n\nthe document size is small right some\ndocuments are like thousands of pages\n\ndocuments are like thousands of pages\n \n\ndocuments are like thousands of pages\nlong 10,000 pages long now suddenly\n\nlong 10,000 pages long now suddenly\n \n\nlong 10,000 pages long now suddenly\nyou're like oh okay I don't know if I\n\nyou're like oh okay I don't know if I\n \n\nyou're like oh okay I don't know if I\ncan pass more than a million tokens into\n\ncan pass more than a million tokens into\n \n\ncan pass more than a million tokens into\nsay uh the open AI models what do I do\n\nsay uh the open AI models what do I do\n \n\nsay uh the open AI models what do I do\nthen right then okay I need to choose a\n\nthen right then okay I need to choose a\n \n\nthen right then okay I need to choose a\ndifferent strategy and often what we do\n\ndifferent strategy and often what we do\n \n\ndifferent strategy and often what we do\nis we have a choose choose different\n\nis we have a choose choose different\n \n\nis we have a choose choose different\nstrategies and kind of mix them with\n\nstrategies and kind of mix them with\n \n\nstrategies and kind of mix them with\nyour prompts to kind of build this\n\nyour prompts to kind of build this\n \n\nyour prompts to kind of build this\niterative process where like I have to\n\niterative process where like I have to\n \n\niterative process where like I have to\nplay around with my prompts, I have to\n\nplay around with my prompts, I have to\n \n\nplay around with my prompts, I have to\nplay around with the different LM\n\nplay around with the different LM\n \n\nplay around with the different LM\nstrategies and we kind of make want to\n\nstrategies and we kind of make want to\n \n\nstrategies and we kind of make want to\nmake that process as quickly as\n\nmake that process as quickly as\n \n\nmake that process as quickly as\npossible. That's a challenge, right? possible. That's a challenge, right? possible. That's a challenge, right? Then you have obviously the context\n\nThen you have obviously the context\n \n\nThen you have obviously the context\nlimitations, model limitations,\n\nlimitations, model limitations,\n \n\nlimitations, model limitations,\ndifferent vendors and you're trying and\n\ndifferent vendors and you're trying and\n \n\ndifferent vendors and you're trying and\ntesting uh things uh uh for quite a\n\ntesting uh things uh uh for quite a\n \n\ntesting uh things uh uh for quite a\nwhile and this kind of goes into the\n\nwhile and this kind of goes into the\n \n\nwhile and this kind of goes into the\nmonth right then the biggest challenge\n\nmonth right then the biggest challenge\n \n\nmonth right then the biggest challenge\nis like okay fine I've kind of built\n\nis like okay fine I've kind of built\n \n\nis like okay fine I've kind of built\nthis app now what how do I get this to\n\nthis app now what how do I get this to\n \n\nthis app now what how do I get this to\ndeployment and it's this whole other set\n\ndeployment and it's this whole other set\n \n\ndeployment and it's this whole other set\nof challenges right you have your\n\nof challenges right you have your\n \n\nof challenges right you have your\ntraditional challenges which is has to\n\ntraditional challenges which is has to\n \n\ntraditional challenges which is has to\ndo with distribution access control how\n\ndo with distribution access control how\n \n\ndo with distribution access control how\nam I going to fedate the app to the\n\nam I going to fedate the app to the\n \n\nam I going to fedate the app to the\nusers but then in the AI space it's like\n\nusers but then in the AI space it's like\n \n\nusers but then in the AI space it's like\nyou have this new challenge of like what\n\nyou have this new challenge of like what\n \n\nyou have this new challenge of like what\ntype of cluster am I going to deploy\n\ntype of cluster am I going to deploy\n \n\ntype of cluster am I going to deploy\nthis to? Right? So, our equity team\n\nthis to? Right? So, our equity team\n \n\nthis to? Right?" } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nSo, our equity team\nwould come and say something like, hey,\n\nwould come and say something like, hey,\n \n\nwould come and say something like, hey,\nI need to analyze, you know, 500\n\nI need to analyze, you know, 500\n \n\nI need to analyze, you know, 500\nresearch reports like overnight, can you\n\nresearch re..." }, { "role": "assistant", "content": "So, our equity team\nwould come and say something like, hey,\n\nwould come and say something like, hey,\n \n\nwould come and say something like, hey,\nI need to analyze, you know, 500\n\nI need to analyze, you know, 500\n \n\nI need to analyze, you know, 500\nresearch reports like overnight, can you\n\nresearch reports like overnight, can you\n \n\nresearch reports like overnight, can you\nhelp me do this? Right? So, okay, if\n\nhelp me do this? Right? So, okay, if\n \n\nhelp me do this? Right? So, okay, if\nyou're going to do that, I probably have\n\nyou're going to do that, I probably have\n \n\nyou're going to do that, I probably have\nto have like a GPU based inference\n\nto have like a GPU based inference\n \n\nto have like a GPU based inference\ncluster that I can kind of spin up,\n\ncluster that I can kind of spin up,\n \n\ncluster that I can kind of spin up,\nright? This is the use case that I kind\n\nright? This is the use case that I kind\n \n\nright? This is the use case that I kind\nof described is the new issue setup. In\n\nof described is the new issue setup. In\n \n\nof described is the new issue setup. In\nthat case, what we do is okay, I don't\n\nthat case, what we do is okay, I don't\n \n\nthat case, what we do is okay, I don't\nreally want to use my GPU inference\n\nreally want to use my GPU inference\n \n\nreally want to use my GPU inference\ncluster, etc. What I do instead is I use\n\ncluster, etc. What I do instead is I use\n \n\ncluster, etc. What I do instead is I use\nlike a burstable cluster, right? All\n\nlike a burstable cluster, right? All\n \n\nlike a burstable cluster, right? All\nthose have to be kind of like uh defined\n\nthose have to be kind of like uh defined\n \n\nthose have to be kind of like uh defined\nso that our app deployment phase is like\n\nso that our app deployment phase is like\n \n\nso that our app deployment phase is like\nas close to like a CI/CD pipeline as\n\nas close to like a CI/CD pipeline as\n \n\nas close to like a CI/CD pipeline as\npossible. Then you have like cost\n\npossible. Then you have like cost\n \n\npossible. Then you have like cost\ncontrols. So these are again it's not an\n\ncontrols. So these are again it's not an\n \n\ncontrols. So these are again it's not an\nexhaustive list. I think what I'm trying\n\nexhaustive list. I think what I'm trying\n \n\nexhaustive list. I think what I'm trying\nto highlight is the challenges with kind\n\nto highlight is the challenges with kind\n \n\nto highlight is the challenges with kind\nof building AI apps. Right? So what we\n\nof building AI apps. Right? So what we\n \n\nof building AI apps. Right? So what we\ndid at Black Rockck is what I'm going to\n\ndid at Black Rockck is what I'm going to\n \n\ndid at Black Rockck is what I'm going to\ndo is I'll kind of give you a highle\n\ndo is I'll kind of give you a highle\n \n\ndo is I'll kind of give you a highle\narchitecture uh and then maybe wuff you\n\narchitecture uh and then maybe wuff you\n \n\narchitecture uh and then maybe wuff you\ncan dive into the details and mechanics\n\ncan dive into the details and mechanics\n \n\ncan dive into the details and mechanics\nof how this works and how we are able to\n\nof how this works and how we are able to\n \n\nof how this works and how we are able to\nbuild apps relatively quickly right\n\nbuild apps relatively quickly right\n \n\nbuild apps relatively quickly right\nwe're able to we took this uh an app\n\nwe're able to we took this uh an app\n \n\nwe're able to we took this uh an app\ntook us close to like 8 months somewhere\n\ntook us close to like 8 months somewhere\n \n\ntook us close to like 8 months somewhere\nbetween 3 to 8 months to build a single\n\nbetween 3 to 8 months to build a single\n \n\nbetween 3 to 8 months to build a single\napp for a complex use case and we able\n\napp for a complex use case and we able\n \n\napp for a complex use case and we able\nto compress time bring it down to like a\n\nto compress time bring it down to like a\n \n\nto compress time bring it down to like a\ncouple of days right we achieved that by\n\ncouple of days right we achieved that by\n \n\ncouple of days right we achieved that by\nbuilding up this framework what I kind\n\nbuilding up this framework what I kind\n \n\nbuilding up this framework what I kind\nof want to focus on is on the top two\n\nof want to focus on is on the top two\n \n\nof want to focus on is on the top two\nboxes that you see which is your sandbox\n\nboxes that you see which is your sandbox\n \n\nboxes that you see which is your sandbox\nin your app factory right so to the uh\n\nin your app factory right so to the uh\n \n\nin your app factory right so to the uh\nthe data platform and the developer\n\nthe data platform and the developer\n \n\nthe data platform and the developer\nplatform is like the name suggest hey\n\nplatform is like the name suggest hey\n \n\nplatform is like the name suggest hey\nplatform is someone for ingesting data\n\nplatform is someone for ingesting data\n \n\nplatform is someone for ingesting data\netc right you have an orchestration\n\netc right you have an orchestration\n \n\netc right you have an orchestration\nlayer that has a pipeline that kind of\n\nlayer that has a pipeline that kind of\n \n\nlayer that has a pipeline that kind of\nlike transforms it brings it into some\n\nlike transforms it brings it into some\n \n\nlike transforms it brings it into some\nuh new format and then you kind of\n\nuh new format and then you kind of\n \n\nuh new format and then you kind of\ndistribute that as a app or report what\n\ndistribute that as a app or report what\n \n\ndistribute that as a app or report what\nkind of accelerates at app development\n\nkind of accelerates at app development\n \n\nkind of accelerates at app development\nis like if you're able to federate out\n\nis like if you're able to federate out\n \n\nis like if you're able to federate out\nthose pain points or those bottlenecks\n\nthose pain points or those bottlenecks\n \n\nthose pain points or those bottlenecks\nwhich is like prompt creation or\n\nwhich is like prompt creation or\n \n\nwhich is like prompt creation or\nextraction templates choosing an LLM\n\nextraction templates choosing an LLM\n \n\nextraction templates choosing an LLM\nstrategy right having extraction runs or\n\nstrategy right having extraction runs or\n \n\nstrategy right having extraction runs or\nlike and then building out these logic\n\nlike and then building out these logic\n \n\nlike and then building out these logic\npieces which are calling transformer and\n\npieces which are calling transformer and\n \n\npieces which are calling transformer and\nexecutors if you can get that sandbox\n\nexecutors if you can get that sandbox\n \n\nexecutors if you can get that sandbox\nout into the hands of the domain experts\n\nout into the hands of the domain experts\n \n\nout into the hands of the domain experts\nthen your iteration speed becomes really\n\nthen your iteration speed becomes really\n \n\nthen your iteration speed becomes really\nfast right so you're kind of saying that\n\nfast right so you're kind of saying that\n \n\nfast right so you're kind of saying that\nhey I have this modular component can I\n\nhey I have this modular component can I\n \n\nhey I have this modular component can I\nmove across the s iteration really\n\nmove across the s iteration really\n \n\nmove across the s iteration really\nquickly and then pass it along to an app\n\nquickly and then pass it along to an app\n \n\nquickly and then pass it along to an app\nfactory which is like our cloudnative\n\nfactory which is like our cloudnative\n \n\nfactory which is like our cloudnative\noperator which takes a definition and\n\noperator which takes a definition and\n \n\noperator which takes a definition and\nspins out an app right so that's super\n\nspins out an app right so that's super\n \n\nspins out an app right so that's super\nhigh level with that quick demo. high level with that quick demo. high level with that quick demo. >> Perfect. >> Perfect. >> Perfect. >> All right,\n\n>> All right,\n \n\n>> All right,\ncool. So, what I'm going to show you\n\ncool. So, what I'm going to show you\n \n\ncool." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nSo, what I'm going to show you\nguys is pretty slim down version of the\n\nguys is pretty slim down version of the\n \n\nguys is pretty slim down version of the\nactual tool we used internally. Um, so\n\nactual tool we used internally. Um, so\n \n\nactual tool we used internally. Um, so\nto start with, uh, when ..." }, { "role": "assistant", "content": "So, what I'm going to show you\nguys is pretty slim down version of the\n\nguys is pretty slim down version of the\n \n\nguys is pretty slim down version of the\nactual tool we used internally. Um, so\n\nactual tool we used internally. Um, so\n \n\nactual tool we used internally. Um, so\nto start with, uh, when the operator, so\n\nto start with, uh, when the operator, so\n \n\nto start with, uh, when the operator, so\nwe have like two different, uh, concore\n\nwe have like two different, uh, concore\n \n\nwe have like two different, uh, concore\ncomponents. One is the sandbox, another\n\ncomponents. One is the sandbox, another\n \n\ncomponents. One is the sandbox, another\none is the factory. So think of sandbox\n\none is the factory. So think of sandbox\n \n\none is the factory. So think of sandbox\nas a playground for the operators to\n\nas a playground for the operators to\n \n\nas a playground for the operators to\nsort of like quickly build and refine\n\nsort of like quickly build and refine\n \n\nsort of like quickly build and refine\nthe extraction templates. Uh sort of run\n\nthe extraction templates. Uh sort of run\n \n\nthe extraction templates. Uh sort of run\nextraction on the set of documents and\n\nextraction on the set of documents and\n \n\nextraction on the set of documents and\nthen compare and contrast the results of\n\nthen compare and contrast the results of\n \n\nthen compare and contrast the results of\nthese extractions. Um so it's sort of\n\nthese extractions. Um so it's sort of\n \n\nthese extractions. Um so it's sort of\nlike to get started with the extraction\n\nlike to get started with the extraction\n \n\nlike to get started with the extraction\ntemplate itself. Uh you might have seen\n\ntemplate itself. Uh you might have seen\n \n\ntemplate itself. Uh you might have seen\nin the other tools both closed and open\n\nin the other tools both closed and open\n \n\nin the other tools both closed and open\nsource they have similar concept like\n\nsource they have similar concept like\n \n\nsource they have similar concept like\nprompt template management where you\n\nprompt template management where you\n \n\nprompt template management where you\nhave certain fields that you want to\n\nhave certain fields that you want to\n \n\nhave certain fields that you want to\nextract out of the documents and you\n\nextract out of the documents and you\n \n\nextract out of the documents and you\nhave their corresponding prompts and\n\nhave their corresponding prompts and\n \n\nhave their corresponding prompts and\nsome metadata that you can associate\n\nsome metadata that you can associate\n \n\nsome metadata that you can associate\nwith them such as the data type that you\n\nwith them such as the data type that you\n \n\nwith them such as the data type that you\nexpect of the the final result values. expect of the the final result values. expect of the the final result values. But when these operators sort of like\n\nBut when these operators sort of like\n \n\nBut when these operators sort of like\ntrying to run extractions on these\n\ntrying to run extractions on these\n \n\ntrying to run extractions on these\ndocuments, they need far more sort of\n\ndocuments, they need far more sort of\n \n\ndocuments, they need far more sort of\nlike greater configuration capabilities\n\nlike greater configuration capabilities\n \n\nlike greater configuration capabilities\nthan just like configuring prompts and\n\nthan just like configuring prompts and\n \n\nthan just like configuring prompts and\nconfiguring the data types that they\n\nconfiguring the data types that they\n \n\nconfiguring the data types that they\nexpect for the end result. So they need\n\nexpect for the end result. So they need\n \n\nexpect for the end result. So they need\nlike hey I need to have multiple QC\n\nlike hey I need to have multiple QC\n \n\nlike hey I need to have multiple QC\nchecks on the result values. I need to\n\nchecks on the result values. I need to\n \n\nchecks on the result values. I need to\nhave a lot of validations and\n\nhave a lot of validations and\n \n\nhave a lot of validations and\nconstraints on the fields and there\n\nconstraints on the fields and there\n \n\nconstraints on the fields and there\nmight be like a interfield dependencies\n\nmight be like a interfield dependencies\n \n\nmight be like a interfield dependencies\nuh what what the fields that are getting\n\nuh what what the fields that are getting\n \n\nuh what what the fields that are getting\nextracted. So as in mentioned with the\n\nextracted. So as in mentioned with the\n \n\nextracted. So as in mentioned with the\nnew security operation uh issuance\n\nnew security operation uh issuance\n \n\nnew security operation uh issuance\nbasically onboarding that stuff there\n\nbasically onboarding that stuff there\n \n\nbasically onboarding that stuff there\ncould be a case where uh the security or\n\ncould be a case where uh the security or\n \n\ncould be a case where uh the security or\nthe bond is callable and you have other\n\nthe bond is callable and you have other\n \n\nthe bond is callable and you have other\nfields such as call data and call price\n\nfields such as call data and call price\n \n\nfields such as call data and call price\nwhich now needs to have a value. So\n\nwhich now needs to have a value. So\n \n\nwhich now needs to have a value. So\nthere is like this inter sort of like\n\nthere is like this inter sort of like\n \n\nthere is like this inter sort of like\nfield dependencies that operators sort\n\nfield dependencies that operators sort\n \n\nfield dependencies that operators sort\nof like uh need they need to take that\n\nof like uh need they need to take that\n \n\nof like uh need they need to take that\ninto consideration be able to configure\n\ninto consideration be able to configure\n \n\ninto consideration be able to configure\nthat. So here is like what a like a\n\nthat. So here is like what a like a\n \n\nthat. So here is like what a like a\nsample uh extraction template looks\n\nsample uh extraction template looks\n \n\nsample uh extraction template looks\nlike. like. like. So here is how a again this is a example\n\nSo here is how a again this is a example\n \n\nSo here is how a again this is a example\ntemplate where we have like issuer\n\ntemplate where we have like issuer\n \n\ntemplate where we have like issuer\ncallable call price and call date these\n\ncallable call price and call date these\n \n\ncallable call price and call date these\nfields set up and to sort of like add\n\nfields set up and to sort of like add\n \n\nfields set up and to sort of like add\nnew fields we would define the field\n\nnew fields we would define the field\n \n\nnew fields we would define the field\nname uh define the data type that is\n\nname uh define the data type that is\n \n\nname uh define the data type that is\nexpected out of that uh define the\n\nexpected out of that uh define the\n \n\nexpected out of that uh define the\nsource whether it's extracted or derived\n\nsource whether it's extracted or derived\n \n\nsource whether it's extracted or derived\nnot every time you want to sort of like\n\nnot every time you want to sort of like\n \n\nnot every time you want to sort of like\nrun an extraction for a field there\n\nrun an extraction for a field there\n \n\nrun an extraction for a field there\nmight be a derived field that operator\n\nmight be a derived field that operator\n \n\nmight be a derived field that operator\nexpect which is sort of like uh\n\nexpect which is sort of like uh\n \n\nexpect which is sort of like uh\npopulated through some transformation\n\npopulated through some transformation\n \n\npopulated through some transformation\ndownstream um and once uh again uh\n\ndownstream um and once uh again uh\n \n\ndownstream um and once uh again uh\nwhether the field is required and the\n\nwhether the field is required and the\n \n\nwhether the field is required and the\nfield dependencies. Here is where you\n\nfield dependencies. Here is where you\n \n\nfield dependencies. Here is where you\ndefine what sort of like dependencies\n\ndefine what sort of like dependencies\n \n\ndefine what sort of like dependencies\nthis field have and sort of validations\n\nthis field have and sort of validations\n \n\nthis field have and sort of validations\nright. So this is how they set up the\n\nright. So this is how they set up the\n \n\nright. So this is how they set up the\nextraction. The next thing is the\n\nextraction. The next thing is the\n \n\nextraction. The next thing is the\ndocument management itself. So this is\n\ndocument management itself. So this is\n \n\ndocument management itself. So this is\nwhere the documents are ingested uh from\n\nwhere the documents are ingested uh from\n \n\nwhere the documents are ingested uh from\nthe uh the data platform." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nThey are\n\nthe uh the data platform. They are\n \n\nthe uh the data platform. They are\ntagged according to the business\n\ntagged according to the business\n \n\ntagged according to the business\ncategory uh and they are labeled they're\n\ncategory uh and they are labeled they're\n \n\ncategory uh and they are lab..." }, { "role": "assistant", "content": "They are\n\nthe uh the data platform. They are\n \n\nthe uh the data platform. They are\ntagged according to the business\n\ntagged according to the business\n \n\ntagged according to the business\ncategory uh and they are labeled they're\n\ncategory uh and they are labeled they're\n \n\ncategory uh and they are labeled they're\nembedded all of that stuff. embedded all of that stuff. embedded all of that stuff. >> Okay. While I think while Viber kind of\n\n>> Okay. While I think while Viber kind of\n \n\n>> Okay. While I think while Viber kind of\nbrings it up. So I think what in essence\n\nbrings it up. So I think what in essence\n \n\nbrings it up. So I think what in essence\nwhat we're saying is we kind of built\n\nwhat we're saying is we kind of built\n \n\nwhat we're saying is we kind of built\nthis tool which has like a UI component\n\nthis tool which has like a UI component\n \n\nthis tool which has like a UI component\nand like a framework that actually lets\n\nand like a framework that actually lets\n \n\nand like a framework that actually lets\nyou take these different pieces and\n\nyou take these different pieces and\n \n\nyou take these different pieces and\nthese modular components and give it to\n\nthese modular components and give it to\n \n\nthese modular components and give it to\nthe hands of like the domain expert to\n\nthe hands of like the domain expert to\n \n\nthe hands of like the domain expert to\nbuild out their app really quickly. build out their app really quickly. build out their app really quickly. Right? Right? Right? >> I think something happened just so let\n\n>> I think something happened just so let\n \n\n>> I think something happened just so let\nme just sort of walk you guys the what\n\nme just sort of walk you guys the what\n \n\nme just sort of walk you guys the what\nhappens next. Um so like once you have\n\nhappens next. Um so like once you have\n \n\nhappens next. Um so like once you have\nset up the extraction templates and\n\nset up the extraction templates and\n \n\nset up the extraction templates and\ndocuments management the operators\n\ndocuments management the operators\n \n\ndocuments management the operators\nbasically run the extractions. That's\n\nbasically run the extractions. That's\n \n\nbasically run the extractions. That's\nwhere they basically see the values that\n\nwhere they basically see the values that\n \n\nwhere they basically see the values that\nthey expect from these documents and\n\nthey expect from these documents and\n \n\nthey expect from these documents and\nsort of like review them. Uh the thing\n\nsort of like review them. Uh the thing\n \n\nsort of like review them. Uh the thing\nwith we have seen with these operators\n\nwith we have seen with these operators\n \n\nwith we have seen with these operators\ntrying to use other tools. Uh no this is\n\ntrying to use other tools. Uh no this is\n \n\ntrying to use other tools. Uh no this is\njust saying um\n\njust saying um\n \n\njust saying um\n>> yeah I did uh the thing we have seen uh\n\n>> yeah I did uh the thing we have seen uh\n \n\n>> yeah I did uh the thing we have seen uh\nwith these operators is that most of the\n\nwith these operators is that most of the\n \n\nwith these operators is that most of the\ntools tools that they have used in past\n\ntools tools that they have used in past\n \n\ntools tools that they have used in past\nuh these tools basically does extraction\n\nuh these tools basically does extraction\n \n\nuh these tools basically does extraction\nuh the they do a pretty good job at\n\nuh the they do a pretty good job at\n \n\nuh the they do a pretty good job at\nextraction but when it comes to like uh\n\nextraction but when it comes to like uh\n \n\nextraction but when it comes to like uh\nhey I need to now use this result that\n\nhey I need to now use this result that\n \n\nhey I need to now use this result that\nhas been uh presented to me and pass it\n\nhas been uh presented to me and pass it\n \n\nhas been uh presented to me and pass it\nto the downstream processes. The process\n\nto the downstream processes. The process\n \n\nto the downstream processes. The process\nright now is very manual where they have\n\nright now is very manual where they have\n \n\nright now is very manual where they have\nto like download a CSV or a JSON file,\n\nto like download a CSV or a JSON file,\n \n\nto like download a CSV or a JSON file,\nrun manual or add a transformation and\n\nrun manual or add a transformation and\n \n\nrun manual or add a transformation and\nthen push it to the downstream process. then push it to the downstream process. then push it to the downstream process. So what we have done and again I can't\n\nSo what we have done and again I can't\n \n\nSo what we have done and again I can't\nshow you but what we have done is like\n\nshow you but what we have done is like\n \n\nshow you but what we have done is like\nbuild this sort of like low code no code\n\nbuild this sort of like low code no code\n \n\nbuild this sort of like low code no code\nframework where the operators can\n\nframework where the operators can\n \n\nframework where the operators can\nbasically essentially uh run the uh sort\n\nbasically essentially uh run the uh sort\n \n\nbasically essentially uh run the uh sort\nof build this transformation and\n\nof build this transformation and\n \n\nof build this transformation and\nexecution workflows and uh sort of like\n\nexecution workflows and uh sort of like\n \n\nexecution workflows and uh sort of like\nhave this end toend uh pipeline running\n\nhave this end toend uh pipeline running\n \n\nhave this end toend uh pipeline running\nuh and I I think yeah so I think we'll\n\nuh and I I think yeah so I think we'll\n \n\nuh and I I think yeah so I think we'll\nconclude by saying that our key\n\nconclude by saying that our key\n \n\nconclude by saying that our key\ntakeaways of this right I would say that\n\ntakeaways of this right I would say that\n \n\ntakeaways of this right I would say that\nare like three key takeaways invest\n\nare like three key takeaways invest\n \n\nare like three key takeaways invest\nheavily on your like prompt engineering\n\nheavily on your like prompt engineering\n \n\nheavily on your like prompt engineering\nskills for your domain experts\n\nskills for your domain experts\n \n\nskills for your domain experts\nespecially in like the financial space\n\nespecially in like the financial space\n \n\nespecially in like the financial space\nand world. Uh defining and describing\n\nand world. Uh defining and describing\n \n\nand world. Uh defining and describing\nthese documents is really hard, right? A\n\nthese documents is really hard, right? A\n \n\nthese documents is really hard, right? A\nsecond is like educating the firm and\n\nsecond is like educating the firm and\n \n\nsecond is like educating the firm and\nthe company on what an LLM strategy\n\nthe company on what an LLM strategy\n \n\nthe company on what an LLM strategy\nmeans uh and how to actually fix these\n\nmeans uh and how to actually fix these\n \n\nmeans uh and how to actually fix these\ndifferent pieces for your particular use\n\ndifferent pieces for your particular use\n \n\ndifferent pieces for your particular use\ncase. And I think the third one I would\n\ncase. And I think the third one I would\n \n\ncase. And I think the third one I would\nsay is hey uh the key takeaway that we\n\nsay is hey uh the key takeaway that we\n \n\nsay is hey uh the key takeaway that we\nhad is all of this is great in\n\nhad is all of this is great in\n \n\nhad is all of this is great in\nexperimentation and prototyping mode but\n\nexperimentation and prototyping mode but\n \n\nexperimentation and prototyping mode but\nif you kind of want to bring this you\n\nif you kind of want to bring this you\n \n\nif you kind of want to bring this you\nhave to really evaluate what your ROI is\n\nhave to really evaluate what your ROI is\n \n\nhave to really evaluate what your ROI is\nand as is it going to be like more\n\nand as is it going to be like more\n \n\nand as is it going to be like more\nexpensive actually spinning up an AI app\n\nexpensive actually spinning up an AI app\n \n\nexpensive actually spinning up an AI app\nversus just having like an offtheshelf\n\nversus just having like an offtheshelf\n \n\nversus just having like an offtheshelf\nproduct that does it quicker and faster. product that does it quicker and faster. product that does it quicker and faster. Right?" } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nSo those are the three key\n\nRight? So those are the three key\n \n\nRight? So those are the three key\ntakeaways in terms of like uh building\n\ntakeaways in terms of like uh building\n \n\ntakeaways in terms of like uh building\napps at scale. And what we realized was\n\napps at scale. And what we realized was..." }, { "role": "assistant", "content": "So those are the three key\n\nRight? So those are the three key\n \n\nRight? So those are the three key\ntakeaways in terms of like uh building\n\ntakeaways in terms of like uh building\n \n\ntakeaways in terms of like uh building\napps at scale. And what we realized was\n\napps at scale. And what we realized was\n \n\napps at scale. And what we realized was\nlike hey uh this notion of like human in\n\nlike hey uh this notion of like human in\n \n\nlike hey uh this notion of like human in\nthe loop and the one more thing I'll add\n\nthe loop and the one more thing I'll add\n \n\nthe loop and the one more thing I'll add\nis human in the loop is super important\n\nis human in the loop is super important\n \n\nis human in the loop is super important\nright we all are like really tempted\n\nright we all are like really tempted\n \n\nright we all are like really tempted\nlike let's go all agent tech with this\n\nlike let's go all agent tech with this\n \n\nlike let's go all agent tech with this\nuh but in the financial space with\n\nuh but in the financial space with\n \n\nuh but in the financial space with\ncompliance with regulations you kind of\n\ncompliance with regulations you kind of\n \n\ncompliance with regulations you kind of\nneed those four eyes check and you kind\n\nneed those four eyes check and you kind\n \n\nneed those four eyes check and you kind\nof need the human loop so design for\n\nof need the human loop so design for\n \n\nof need the human loop so design for\nhuman in the loop first uh if you're in\n\nhuman in the loop first uh if you're in\n \n\nhuman in the loop first uh if you're in\na highly regulated environment\n\na highly regulated environment\n \n\na highly regulated environment\n>> yeah and as info said one thing we\n\n>> yeah and as info said one thing we\n \n\n>> yeah and as info said one thing we\ncouldn't show is the whole app factory\n\ncouldn't show is the whole app factory\n \n\ncouldn't show is the whole app factory\nsort of like uh component which is all\n\nsort of like uh component which is all\n \n\nsort of like uh component which is all\nthe things that operators do through\n\nthe things that operators do through\n \n\nthe things that operators do through\nthis iteration cycle of through the\n\nthis iteration cycle of through the\n \n\nthis iteration cycle of through the\nsandbox. They take all that knowledge,\n\nsandbox. They take all that knowledge,\n \n\nsandbox. They take all that knowledge,\nthe extraction templates, the transform\n\nthe extraction templates, the transform\n \n\nthe extraction templates, the transform\nit transformers and executors they build\n\nit transformers and executors they build\n \n\nit transformers and executors they build\nthrough this workflow pipeline and\n\nthrough this workflow pipeline and\n \n\nthrough this workflow pipeline and\nthrough our app uh ecosystem within uh\n\nthrough our app uh ecosystem within uh\n \n\nthrough our app uh ecosystem within uh\nBlackRock they sort of like build this\n\nBlackRock they sort of like build this\n \n\nBlackRock they sort of like build this\nuh custom applications that are then\n\nuh custom applications that are then\n \n\nuh custom applications that are then\nexposed to the users where users of this\n\nexposed to the users where users of this\n \n\nexposed to the users where users of this\napp don't have to worry about how to\n\napp don't have to worry about how to\n \n\napp don't have to worry about how to\nconfigure templates or how to basically\n\nconfigure templates or how to basically\n \n\nconfigure templates or how to basically\nfigure out how to integrate the result\n\nfigure out how to integrate the result\n \n\nfigure out how to integrate the result\nvalues into final downstream processes. values into final downstream processes. values into final downstream processes. they are presented with this whole end\n\nthey are presented with this whole end\n \n\nthey are presented with this whole end\nto end app where they can just go and\n\nto end app where they can just go and\n \n\nto end app where they can just go and\nlike sort of like upload documents and\n\nlike sort of like upload documents and\n \n\nlike sort of like upload documents and\nrun extraction and sort of uh get the\n\nrun extraction and sort of uh get the\n \n\nrun extraction and sort of uh get the\nwhole pipeline set uh running. whole pipeline set uh running. whole pipeline set uh running. >> Yeah. With that we'll open up for\n\n>> Yeah. With that we'll open up for\n \n\n>> Yeah. With that we'll open up for\nquestions. I think we have like a minute\n\nquestions. I think we have like a minute\n \n\nquestions. I think we have like a minute\nor two left. or two left. or two left. >> Yeah. Uh so I have a question which may\n\n>> Yeah. Uh so I have a question which may\n \n\n>> Yeah. Uh so I have a question which may\ndirectly be related to\n\n \n\n>> Good morning. I have a question which\n\n>> Good morning. I have a question which\n \n\n>> Good morning. I have a question which\nmay directly be related to the uh\n\nmay directly be related to the uh\n \n\nmay directly be related to the uh\narchitecture that you developed. architecture that you developed. architecture that you developed. >> You can tell me I can discuss later. But\n\n>> You can tell me I can discuss later. But\n \n\n>> You can tell me I can discuss later. But\nthe question is going to be\n\nthe question is going to be\n \n\nthe question is going to be\nyou you have developed uh um the key\n\nyou you have developed uh um the key\n \n\nyou you have developed uh um the key\ntakeaways. One of those key takeaways\n\ntakeaways. One of those key takeaways\n \n\ntakeaways. One of those key takeaways\nhad been in invest heavily on prompt\n\nhad been in invest heavily on prompt\n \n\nhad been in invest heavily on prompt\nengineering. So you have essentially\n\nengineering. So you have essentially\n \n\nengineering. So you have essentially\nautomated the process from the leaf\n\nautomated the process from the leaf\n \n\nautomated the process from the leaf\nlevel for example a company's coming to\n\nlevel for example a company's coming to\n \n\nlevel for example a company's coming to\nan IPO from that level all the way to\n\nan IPO from that level all the way to\n \n\nan IPO from that level all the way to\ncataloging through ETL processes and\n\ncataloging through ETL processes and\n \n\ncataloging through ETL processes and\nthen to finally to the data analytics. then to finally to the data analytics. then to finally to the data analytics. So now your CEO who looks at the balance\n\nSo now your CEO who looks at the balance\n \n\nSo now your CEO who looks at the balance\nsheet, assets and liability will be\n\nsheet, assets and liability will be\n \n\nsheet, assets and liability will be\nusing your AI the most\n\nusing your AI the most\n \n\nusing your AI the most\nand for C uh your CEO. Now what are the\n\nand for C uh your CEO. Now what are the\n \n\nand for C uh your CEO. Now what are the\nfeatures involved here at the lowest\n\nfeatures involved here at the lowest\n \n\nfeatures involved here at the lowest\nlevel for example term maturity duration\n\nlevel for example term maturity duration\n \n\nlevel for example term maturity duration\nthere are so many metrics at the L\n\nthere are so many metrics at the L\n \n\nthere are so many metrics at the L\nlevel. How are you transforming those\n\nlevel. How are you transforming those\n \n\nlevel. How are you transforming those\nfeatures from the lowest level to\n\nfeatures from the lowest level to\n \n\nfeatures from the lowest level to\nhighest level? I'm looking for an answer\n\nhighest level? I'm looking for an answer\n \n\nhighest level? I'm looking for an answer\nin reference to decentralized data. in reference to decentralized data. in reference to decentralized data. >> Yeah, I mean I can give you a quick\n\n>> Yeah, I mean I can give you a quick\n \n\n>> Yeah, I mean I can give you a quick\nanswer and then we can discuss uh in\n\nanswer and then we can discuss uh in\n \n\nanswer and then we can discuss uh in\ndetail like offline. I think real\n\ndetail like offline. I think real\n \n\ndetail like offline. I think real\nquickly like the the framework that we\n\nquickly like the the framework that we\n \n\nquickly like the the framework that we\nbuilt was specifically targeting like\n\nbuilt was specifically targeting like\n \n\nbuilt was specifically targeting like\nthe investment operation domain experts\n\nthe investment operation domain experts\n \n\nthe investment operation domain experts\nwho are trying to build applications. To\n\nwho are trying to build applications. To\n \n\nwho are trying to build applications." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nTo\nyour question of like hey what does the\n\nyour question of like hey what does the\n \n\nyour question of like hey what does the\nCEO care about? Can I construct a memo\n\nCEO care about? Can I construct a memo\n \n\nCEO care about? Can I construct a memo\nthat gives me my asset liabilities XYZ? that gives m..." }, { "role": "assistant", "content": "To\nyour question of like hey what does the\n\nyour question of like hey what does the\n \n\nyour question of like hey what does the\nCEO care about? Can I construct a memo\n\nCEO care about? Can I construct a memo\n \n\nCEO care about? Can I construct a memo\nthat gives me my asset liabilities XYZ? that gives me my asset liabilities XYZ? that gives me my asset liabilities XYZ? Those would be like different\n\nThose would be like different\n \n\nThose would be like different\ninitiatives which may or may not use our\n\ninitiatives which may or may not use our\n \n\ninitiatives which may or may not use our\nparticular framework. But uh yes, there\n\nparticular framework. But uh yes, there\n \n\nparticular framework. But uh yes, there\nare many reusable components in here\n\nare many reusable components in here\n \n\nare many reusable components in here\nthat people can use. Yeah. >> Yeah. So I do like a lot of document\n\n>> Yeah. So I do like a lot of document\n \n\n>> Yeah. So I do like a lot of document\nprocessing for insurance company. Pretty\n\nprocessing for insurance company. Pretty\n \n\nprocessing for insurance company. Pretty\nmuch same problems as you guys run into. much same problems as you guys run into. much same problems as you guys run into. So I wonder how do you build a wall\n\nSo I wonder how do you build a wall\n \n\nSo I wonder how do you build a wall\naround your information extraction from\n\naround your information extraction from\n \n\naround your information extraction from\nthe documents right because there are so\n\nthe documents right because there are so\n \n\nthe documents right because there are so\nmany things that can go wrong starting\n\nmany things that can go wrong starting\n \n\nmany things that can go wrong starting\nfrom a CR like doesn't understand what\n\nfrom a CR like doesn't understand what\n \n\nfrom a CR like doesn't understand what\nall these terms actually mean no matter\n\nall these terms actually mean no matter\n \n\nall these terms actually mean no matter\nhow you prompt it right all this stuff. how you prompt it right all this stuff. how you prompt it right all this stuff. So that's kind of what's for the reason\n\nSo that's kind of what's for the reason\n \n\nSo that's kind of what's for the reason\n>> again I mean we had all of that that we\n\n>> again I mean we had all of that that we\n \n\n>> again I mean we had all of that that we\nwanted to show but yeah\n\nwanted to show but yeah\n \n\nwanted to show but yeah\n>> I think a short answer short answer to\n\n>> I think a short answer short answer to\n \n\n>> I think a short answer short answer to\nyour question is in terms of like\n\nyour question is in terms of like\n \n\nyour question is in terms of like\ninformation security and what are the\n\ninformation security and what are the\n \n\ninformation security and what are the\nboundaries uh that we're putting in\n\nboundaries uh that we're putting in\n \n\nboundaries uh that we're putting in\nterms of like hey we are not having data\n\nterms of like hey we are not having data\n \n\nterms of like hey we are not having data\nleakage or errors or understanding of\n\nleakage or errors or understanding of\n \n\nleakage or errors or understanding of\nlike in terms of security you can think\n\nlike in terms of security you can think\n \n\nlike in terms of security you can think\nof it as different layers all the way\n\nof it as different layers all the way\n \n\nof it as different layers all the way\nfrom like your infra platform\n\nfrom like your infra platform\n \n\nfrom like your infra platform\napplication right and the user levels\n\napplication right and the user levels\n \n\napplication right and the user levels\nthere are different controls and\n\nthere are different controls and\n \n\nthere are different controls and\npolicies in place uh and it's also\n\npolicies in place uh and it's also\n \n\npolicies in place uh and it's also\nwithin SD network like I think there are\n\nwithin SD network like I think there are\n \n\nwithin SD network like I think there are\npolicies across the stack that we can\n\npolicies across the stack that we can\n \n\npolicies across the stack that we can\nget into in detail later that kind of uh\n\nget into in detail later that kind of uh\n \n\nget into in detail later that kind of uh\naddresses your um concerns\n\naddresses your um concerns\n \n\naddresses your um concerns\n>> and also also to your point um I think\n\n>> and also also to your point um I think\n \n\n>> and also also to your point um I think\nwe have like different sort of like uh\n\nwe have like different sort of like uh\n \n\nwe have like different sort of like uh\nstrategies that we use based on uh the\n\nstrategies that we use based on uh the\n \n\nstrategies that we use based on uh the\nsort of like the use case at hand. Uh so\n\nsort of like the use case at hand. Uh so\n \n\nsort of like the use case at hand. Uh so\nit's not just like hey one rag versus\n\nit's not just like hey one rag versus\n \n\nit's not just like hey one rag versus\nthis there are multiple model providers\n\nthis there are multiple model providers\n \n\nthis there are multiple model providers\nthat we use uh multiple different\n\nthat we use uh multiple different\n \n\nthat we use uh multiple different\nstrategies etc. uh different like like\n\nstrategies etc. uh different like like\n \n\nstrategies etc. uh different like like\nengineering sort of tweaks. Uh so it's a\n\nengineering sort of tweaks. Uh so it's a\n \n\nengineering sort of tweaks. Uh so it's a\nquite complex sort of yeah process. quite complex sort of yeah process. quite complex sort of yeah process. >> All right. >> All right. >> All right. >> Very cool. >> Very cool. >> Very cool. >> Awesome. >> Awesome. >> Awesome. >> Thank you. >> Thank you. >> Thank you. >> All right. >> All right. >> All right. [Music]" } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\n[Music]\n\n \n\nAll right. Hey everyone. Uh, my name is\n\nAll right. Hey everyone. Uh, my name is\n \n\nAll right. Hey everyone. Uh, my name is\nSharief. I'll be talking to you about\n\nSharief. I'll be talking to you about\n \n\nSharief. I'll be talking to you about\ndemos and why I think demos are probably\n\ndemo..." }, { "role": "assistant", "content": "[Music]\n\n \n\nAll right. Hey everyone. Uh, my name is\n\nAll right. Hey everyone. Uh, my name is\n \n\nAll right. Hey everyone. Uh, my name is\nSharief. I'll be talking to you about\n\nSharief. I'll be talking to you about\n \n\nSharief. I'll be talking to you about\ndemos and why I think demos are probably\n\ndemos and why I think demos are probably\n \n\ndemos and why I think demos are probably\nthe most important thing in the world\n\nthe most important thing in the world\n \n\nthe most important thing in the world\nright now. Um, I'm the founder of\n\nright now. Um, I'm the founder of\n \n\nright now. Um, I'm the founder of\nLexica. We're working on generative\n\nLexica. We're working on generative\n \n\nLexica. We're working on generative\nmodels, specifically image models. Um,\n\nmodels, specifically image models. Um,\n \n\nmodels, specifically image models. Um,\nbut I kind of want to just talk to you\n\nbut I kind of want to just talk to you\n \n\nbut I kind of want to just talk to you\nabout something a bit more than just\n\nabout something a bit more than just\n \n\nabout something a bit more than just\nmodels themselves. Um, even more than\n\nmodels themselves. Um, even more than\n \n\nmodels themselves. Um, even more than\ndemos. I kind of just want to talk to\n\ndemos. I kind of just want to talk to\n \n\ndemos. I kind of just want to talk to\nyou about curiosity. Um, there was a\n\nyou about curiosity. Um, there was a\n \n\nyou about curiosity. Um, there was a\nfamous French mathematician Poare. He\n\nfamous French mathematician Poare. He\n \n\nfamous French mathematician Poare. He\nsaid at the moment when I put my foot on\n\nsaid at the moment when I put my foot on\n \n\nsaid at the moment when I put my foot on\nthe step, the idea came to me. He was\n\nthe step, the idea came to me. He was\n \n\nthe step, the idea came to me. He was\nworking on this really really esoteric\n\nworking on this really really esoteric\n \n\nworking on this really really esoteric\nfield of mathematics called fusion\n\nfield of mathematics called fusion\n \n\nfield of mathematics called fusion\nfunctions. Um, and he was stuck on this\n\nfunctions. Um, and he was stuck on this\n \n\nfunctions. Um, and he was stuck on this\nproblem for weeks at a time. He didn't\n\nproblem for weeks at a time. He didn't\n \n\nproblem for weeks at a time. He didn't\nreally know how to make any progress at\n\nreally know how to make any progress at\n \n\nreally know how to make any progress at\nall. and he was boarding a bus one day\n\nall. and he was boarding a bus one day\n \n\nall. and he was boarding a bus one day\nand suddenly it kind of just all came to\n\nand suddenly it kind of just all came to\n \n\nand suddenly it kind of just all came to\nhim. He went from not knowing at all to\n\nhim. He went from not knowing at all to\n \n\nhim. He went from not knowing at all to\nhaving a full understanding of the\n\nhaving a full understanding of the\n \n\nhaving a full understanding of the\nproblem. Uh he said something along the\n\nproblem. Uh he said something along the\n \n\nproblem. Uh he said something along the\nlines of the role of this unconscious\n\nlines of the role of this unconscious\n \n\nlines of the role of this unconscious\nwork in mathematical invention appears\n\nwork in mathematical invention appears\n \n\nwork in mathematical invention appears\nto me as incontestable. to me as incontestable. to me as incontestable. Um I'm going to make one main argument\n\nUm I'm going to make one main argument\n \n\nUm I'm going to make one main argument\nto you guys today and that's going to be\n\nto you guys today and that's going to be\n \n\nto you guys today and that's going to be\nthat curiosity is the main force for how\n\nthat curiosity is the main force for how\n \n\nthat curiosity is the main force for how\nwe pull ideas from the future into the\n\nwe pull ideas from the future into the\n \n\nwe pull ideas from the future into the\npresent. Um, and when we have these\n\npresent. Um, and when we have these\n \n\npresent. Um, and when we have these\nsubconscious patterns that our brains\n\nsubconscious patterns that our brains\n \n\nsubconscious patterns that our brains\nrecognize, they kind of surface as a\n\nrecognize, they kind of surface as a\n \n\nrecognize, they kind of surface as a\nfeeling. And this feeling is what we\n\nfeeling. And this feeling is what we\n \n\nfeeling. And this feeling is what we\nknow as curiosity. So, I'm going to present you with a few\n\nSo, I'm going to present you with a few\n \n\nSo, I'm going to present you with a few\ndemos I've worked on over the years. Um,\n\ndemos I've worked on over the years. Um,\n \n\ndemos I've worked on over the years. Um,\nand they've each followed a specific\n\nand they've each followed a specific\n \n\nand they've each followed a specific\npattern where initially I had this\n\npattern where initially I had this\n \n\npattern where initially I had this\nreally great idea. I thought it was\n\nreally great idea. I thought it was\n \n\nreally great idea. I thought it was\nfantastic that it was going to change\n\nfantastic that it was going to change\n \n\nfantastic that it was going to change\neverything. And then you kind of get to\n\neverything. And then you kind of get to\n \n\neverything. And then you kind of get to\nimplementing the specific idea and you\n\nimplementing the specific idea and you\n \n\nimplementing the specific idea and you\nrealize it's actually not possible at\n\nrealize it's actually not possible at\n \n\nrealize it's actually not possible at\nall. And then like through sheer effort\n\nall. And then like through sheer effort\n \n\nall. And then like through sheer effort\nand like determination, you somehow find\n\nand like determination, you somehow find\n \n\nand like determination, you somehow find\na way to make things work even though\n\na way to make things work even though\n \n\na way to make things work even though\nyou're working with models that have\n\nyou're working with models that have\n \n\nyou're working with models that have\nmaybe a context length of like 2,000\n\nmaybe a context length of like 2,000\n \n\nmaybe a context length of like 2,000\ntokens. Uh and then once you get it\n\ntokens. Uh and then once you get it\n \n\ntokens. Uh and then once you get it\nworking, you feel this like really good\n\nworking, you feel this like really good\n \n\nworking, you feel this like really good\nsense of pride and joy. Um and I think\n\nsense of pride and joy. Um and I think\n \n\nsense of pride and joy. Um and I think\nthe most important thing about good\n\nthe most important thing about good\n \n\nthe most important thing about good\ndemos is that um they're kind of a way\n\ndemos is that um they're kind of a way\n \n\ndemos is that um they're kind of a way\nof exploring what's possible with these\n\nof exploring what's possible with these\n \n\nof exploring what's possible with these\nmodels. I'm I kind of see these models\n\nmodels. I'm I kind of see these models\n \n\nmodels. I'm I kind of see these models\nas not necessarily things you can kind\n\nas not necessarily things you can kind\n \n\nas not necessarily things you can kind\nof understand fully without interacting\n\nof understand fully without interacting\n \n\nof understand fully without interacting\nwith them. And I think the way you can\n\nwith them. And I think the way you can\n \n\nwith them. And I think the way you can\nbest interact with them is by making\n\nbest interact with them is by making\n \n\nbest interact with them is by making\nreally really interesting demos. And uh\n\nreally really interesting demos. And uh\n \n\nreally really interesting demos. And uh\nthe way you make interesting demos is\n\nthe way you make interesting demos is\n \n\nthe way you make interesting demos is\njust by following your curiosity. just by following your curiosity. just by following your curiosity. So this is from 2020. This was when GPD3\n\nSo this is from 2020. This was when GPD3\n \n\nSo this is from 2020. This was when GPD3\nwas released. Um this was pretty\n\nwas released. Um this was pretty\n \n\nwas released. Um this was pretty\nmind-blowing for me and I was surprised\n\nmind-blowing for me and I was surprised\n \n\nmind-blowing for me and I was surprised\nno one was really talking about this. Uh\n\nno one was really talking about this. Uh\n \n\nno one was really talking about this. Uh\nGPD3 for those of you remember had a\n\nGPD3 for those of you remember had a\n \n\nGPD3 for those of you remember had a\ncontext length of 2,000 tokens. Uh it\n\ncontext length of 2,000 tokens." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nUh it\n \n\ncontext length of 2,000 tokens. Uh it\ncost I think $75 per million output\n\ncost I think $75 per million output\n \n\ncost I think $75 per million output\ntokens. Um, and yeah, you had to get\n\ntokens. Um, and yeah, you had to get\n \n\ntokens. Um, and yeah, you had to get\nspecific permissions from ..." }, { "role": "assistant", "content": "Uh it\n \n\ncontext length of 2,000 tokens. Uh it\ncost I think $75 per million output\n\ncost I think $75 per million output\n \n\ncost I think $75 per million output\ntokens. Um, and yeah, you had to get\n\ntokens. Um, and yeah, you had to get\n \n\ntokens. Um, and yeah, you had to get\nspecific permissions from OpenAI before\n\nspecific permissions from OpenAI before\n \n\nspecific permissions from OpenAI before\nyou shared anything about the model. Uh,\n\nyou shared anything about the model. Uh,\n \n\nyou shared anything about the model. Uh,\nyou couldn't ship a chat app because\n\nyou couldn't ship a chat app because\n \n\nyou couldn't ship a chat app because\nthat violated the terms of service. But\n\nthat violated the terms of service. But\n \n\nthat violated the terms of service. But\nit was a really magical time. It felt\n\nit was a really magical time. It felt\n \n\nit was a really magical time. It felt\nlike you had this new tool in this like\n\nlike you had this new tool in this like\n \n\nlike you had this new tool in this like\nin this toolkit of computing and you\n\nin this toolkit of computing and you\n \n\nin this toolkit of computing and you\ncould do so many things with it. And I\n\ncould do so many things with it. And I\n \n\ncould do so many things with it. And I\nthink what was really important about\n\nthink what was really important about\n \n\nthink what was really important about\nthis demo is that it inspired people\n\nthis demo is that it inspired people\n \n\nthis demo is that it inspired people\nthat you could actually do things with\n\nthat you could actually do things with\n \n\nthat you could actually do things with\nsoftware. Um, I think the way you get\n\nsoftware. Um, I think the way you get\n \n\nsoftware. Um, I think the way you get\nreally impressive ideas is actually not\n\nreally impressive ideas is actually not\n \n\nreally impressive ideas is actually not\nby doing anything special. I think each\n\nby doing anything special. I think each\n \n\nby doing anything special. I think each\nperson has their own unique kind of\n\nperson has their own unique kind of\n \n\nperson has their own unique kind of\ncontext window, the things you've seen\n\ncontext window, the things you've seen\n \n\ncontext window, the things you've seen\nand experienced. And I just so happened\n\nand experienced. And I just so happened\n \n\nand experienced. And I just so happened\nto watch a talk by Brett Victor before\n\nto watch a talk by Brett Victor before\n \n\nto watch a talk by Brett Victor before\nmaking this where um he came up with\n\nmaking this where um he came up with\n \n\nmaking this where um he came up with\nthis principle that creators needed\n\nthis principle that creators needed\n \n\nthis principle that creators needed\nimmediate feedback with what they're\n\nimmediate feedback with what they're\n \n\nimmediate feedback with what they're\ncreating. And I was really tired of\n\ncreating. And I was really tired of\n \n\ncreating. And I was really tired of\ncopying and pasting code into my editor\n\ncopying and pasting code into my editor\n \n\ncopying and pasting code into my editor\nand then compiling it and then like\n\nand then compiling it and then like\n \n\nand then compiling it and then like\nrunning it and see what would happen. So\n\nrunning it and see what would happen. So\n \n\nrunning it and see what would happen. So\nI decided to just put like a JSX\n\nI decided to just put like a JSX\n \n\nI decided to just put like a JSX\ncompiler in the browser and it just felt\n\ncompiler in the browser and it just felt\n \n\ncompiler in the browser and it just felt\ndifferent. It felt kind of magical in a\n\ndifferent. It felt kind of magical in a\n \n\ndifferent. It felt kind of magical in a\nway. Um, and now today like Claude\n\nway. Um, and now today like Claude\n \n\nway. Um, and now today like Claude\nsystem prompt is 25,000 tokens which is\n\nsystem prompt is 25,000 tokens which is\n \n\nsystem prompt is 25,000 tokens which is\nkind of funny. kind of funny. kind of funny. Here's another similar demo. This one's\n\nHere's another similar demo. This one's\n \n\nHere's another similar demo. This one's\na bit more interactive. So you can kind\n\na bit more interactive. So you can kind\n \n\na bit more interactive. So you can kind\nof describe what you want. And then uh\n\nof describe what you want. And then uh\n \n\nof describe what you want. And then uh\nbecause the context windows were so\n\nbecause the context windows were so\n \n\nbecause the context windows were so\nsmall, it couldn't actually generate the\n\nsmall, it couldn't actually generate the\n \n\nsmall, it couldn't actually generate the\nentire application in a single prompt. entire application in a single prompt. entire application in a single prompt. uh you actually had to do three parallel\n\nuh you actually had to do three parallel\n \n\nuh you actually had to do three parallel\nprompts and then join them in the\n\nprompts and then join them in the\n \n\nprompts and then join them in the\nbackground. Uh this is really simple\n\nbackground. Uh this is really simple\n \n\nbackground. Uh this is really simple\njust asking for a Google homepage and\n\njust asking for a Google homepage and\n \n\njust asking for a Google homepage and\nthen it generates three different\n\nthen it generates three different\n \n\nthen it generates three different\ncomponents for it. Um but yeah, this is\n\ncomponents for it. Um but yeah, this is\n \n\ncomponents for it. Um but yeah, this is\njust I think one of the ways you can\n\njust I think one of the ways you can\n \n\njust I think one of the ways you can\nexpress your curiosity. Uh you kind of\n\nexpress your curiosity. Uh you kind of\n \n\nexpress your curiosity. Uh you kind of\nlook at the world around you and what\n\nlook at the world around you and what\n \n\nlook at the world around you and what\nyou've experienced and you kind of\n\nyou've experienced and you kind of\n \n\nyou've experienced and you kind of\nsynthesize new ideas and you get this\n\nsynthesize new ideas and you get this\n \n\nsynthesize new ideas and you get this\nsubconscious feeling pulling you in a\n\nsubconscious feeling pulling you in a\n \n\nsubconscious feeling pulling you in a\ndirection and I think the demo is the\n\ndirection and I think the demo is the\n \n\ndirection and I think the demo is the\nbest way to kind of express that feeling\n\nbest way to kind of express that feeling\n \n\nbest way to kind of express that feeling\nto the world. to the world. to the world. Um, here's another more interesting one. Um, here's another more interesting one. Um, here's another more interesting one. This was 2021. I think the context\n\nThis was 2021. I think the context\n \n\nThis was 2021. I think the context\nlengths now expanded from 2,000 tokens\n\nlengths now expanded from 2,000 tokens\n \n\nlengths now expanded from 2,000 tokens\nto about 4,000 tokens. So, we could do a\n\nto about 4,000 tokens. So, we could do a\n \n\nto about 4,000 tokens. So, we could do a\nbit more with this. Um, I was kind of\n\nbit more with this. Um, I was kind of\n \n\nbit more with this. Um, I was kind of\ncurious if these models had any sense of\n\ncurious if these models had any sense of\n \n\ncurious if these models had any sense of\nagency. So, I decided to give it a\n\nagency. So, I decided to give it a\n \n\nagency. So, I decided to give it a\nreally simple objective of buying me\n\nreally simple objective of buying me\n \n\nreally simple objective of buying me\nAirPods in Chrome. And, uh, if you were\n\nAirPods in Chrome. And, uh, if you were\n \n\nAirPods in Chrome. And, uh, if you were\nto just dump a web page into the browser\n\nto just dump a web page into the browser\n \n\nto just dump a web page into the browser\ninto the context window, it just\n\ninto the context window, it just\n \n\ninto the context window, it just\nwouldn't work. Like the walmart.com\n\nwouldn't work. Like the walmart.com\n \n\nwouldn't work. Like the walmart.com\nshopping page would be like 24,000\n\nshopping page would be like 24,000\n \n\nshopping page would be like 24,000\ntokens. It's just impossible. Uh so I\n\ntokens. It's just impossible. Uh so I\n \n\ntokens. It's just impossible. Uh so I\nwas actually a bit frustrated that I\n\nwas actually a bit frustrated that I\n \n\nwas actually a bit frustrated that I\ncouldn't get it to work for a few days. couldn't get it to work for a few days. couldn't get it to work for a few days." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nUh so I wrote a custom HTML parser that\n\nUh so I wrote a custom HTML parser that\n \n\nUh so I wrote a custom HTML parser that\nwould parse a web page into its core\n\nwould parse a web page into its core\n \n\nwould parse a web page into its core\nessence which was able to fit it into\n\nessence which was able..." }, { "role": "assistant", "content": "Uh so I wrote a custom HTML parser that\n\nUh so I wrote a custom HTML parser that\n \n\nUh so I wrote a custom HTML parser that\nwould parse a web page into its core\n\nwould parse a web page into its core\n \n\nwould parse a web page into its core\nessence which was able to fit it into\n\nessence which was able to fit it into\n \n\nessence which was able to fit it into\nthe tiny context window of GPD3 in 2021. the tiny context window of GPD3 in 2021. the tiny context window of GPD3 in 2021. Uh it definitely failed spectacularly. Uh it definitely failed spectacularly. Uh it definitely failed spectacularly. It got distracted with the terms of\n\nIt got distracted with the terms of\n \n\nIt got distracted with the terms of\nservice. Uh but I think it was more so\n\nservice. Uh but I think it was more so\n \n\nservice. Uh but I think it was more so\njust interesting that we discovered that\n\njust interesting that we discovered that\n \n\njust interesting that we discovered that\nthese models pre-trained on web text had\n\nthese models pre-trained on web text had\n \n\nthese models pre-trained on web text had\nthis sense of agency kind of internal in\n\nthis sense of agency kind of internal in\n \n\nthis sense of agency kind of internal in\ntheir weights. Uh, we kind of take that\n\ntheir weights. Uh, we kind of take that\n \n\ntheir weights. Uh, we kind of take that\nfor granted now, but 2021 was a very\n\nfor granted now, but 2021 was a very\n \n\nfor granted now, but 2021 was a very\ndifferent time. different time. different time. Here's a more recent demo from a friend\n\nHere's a more recent demo from a friend\n \n\nHere's a more recent demo from a friend\nof mine, Farsza. Uh, he's using Gemini\n\nof mine, Farsza. Uh, he's using Gemini\n \n\nof mine, Farsza. Uh, he's using Gemini\n2.5 Pro here today. So, we're still\n\n2.5 Pro here today. So, we're still\n \n\n2.5 Pro here today. So, we're still\ndiscovering new capabilities here. He's\n\ndiscovering new capabilities here. He's\n \n\ndiscovering new capabilities here. He's\nmaking a basketball shot tracker where\n\nmaking a basketball shot tracker where\n \n\nmaking a basketball shot tracker where\nhe's just putting in a video of him\n\nhe's just putting in a video of him\n \n\nhe's just putting in a video of him\nplaying basketball, asking it to provide\n\nplaying basketball, asking it to provide\n \n\nplaying basketball, asking it to provide\nfeedback as if Michael Jordan was\n\nfeedback as if Michael Jordan was\n \n\nfeedback as if Michael Jordan was\nwatching his gameplay. And I think this\n\nwatching his gameplay. And I think this\n \n\nwatching his gameplay. And I think this\nis a really great demo because it\n\nis a really great demo because it\n \n\nis a really great demo because it\ninspires people to realize that you can\n\ninspires people to realize that you can\n \n\ninspires people to realize that you can\nactually make video first experiences\n\nactually make video first experiences\n \n\nactually make video first experiences\nwith Gemini 2.5 Pro. Uh before this it\n\nwith Gemini 2.5 Pro. Uh before this it\n \n\nwith Gemini 2.5 Pro. Uh before this it\nwas kind of like oh you can have it\n\nwas kind of like oh you can have it\n \n\nwas kind of like oh you can have it\nwatch your screen and it'll give you\n\nwatch your screen and it'll give you\n \n\nwatch your screen and it'll give you\nfeedback on your code. But there's so\n\nfeedback on your code. But there's so\n \n\nfeedback on your code. But there's so\nmuch more we can do. And I think one of\n\nmuch more we can do. And I think one of\n \n\nmuch more we can do. And I think one of\nthe main reasons I find demos really\n\nthe main reasons I find demos really\n \n\nthe main reasons I find demos really\nreally interesting is that they inspire\n\nreally interesting is that they inspire\n \n\nreally interesting is that they inspire\npossibility. Um, so much so that there's\n\npossibility. Um, so much so that there's\n \n\npossibility. Um, so much so that there's\nprobably so much lowhanging fruit today\n\nprobably so much lowhanging fruit today\n \n\nprobably so much lowhanging fruit today\nin these models that if you were to halt\n\nin these models that if you were to halt\n \n\nin these models that if you were to halt\nall capabilities, if you kept all the\n\nall capabilities, if you kept all the\n \n\nall capabilities, if you kept all the\nweights frozen, didn't do a single back\n\nweights frozen, didn't do a single back\n \n\nweights frozen, didn't do a single back\npropagation, I think you could build\n\npropagation, I think you could build\n \n\npropagation, I think you could build\nreally amazing products for the next 10\n\nreally amazing products for the next 10\n \n\nreally amazing products for the next 10\nyears, keeping everything constant. Um,\n\nyears, keeping everything constant. Um,\n \n\nyears, keeping everything constant. Um,\nand I think the way you do that is just\n\nand I think the way you do that is just\n \n\nand I think the way you do that is just\nby building these demos and following\n\nby building these demos and following\n \n\nby building these demos and following\nyour curiosity. your curiosity. your curiosity. So I have this really famous quote by\n\nSo I have this really famous quote by\n \n\nSo I have this really famous quote by\nRichard Hamming where he says, \"In\n\nRichard Hamming where he says, \"In\n \n\nRichard Hamming where he says, \"In\nscience, if you know what you're doing,\n\nscience, if you know what you're doing,\n \n\nscience, if you know what you're doing,\nyou should not be doing it. engineering\n\nyou should not be doing it. engineering\n \n\nyou should not be doing it. engineering\nif you know what you're doing you should\n\nif you know what you're doing you should\n \n\nif you know what you're doing you should\nnot be doing it. So traditional\n\nnot be doing it. So traditional\n \n\nnot be doing it. So traditional\nengineering is very teological. It's\n\nengineering is very teological. It's\n \n\nengineering is very teological. It's\nvery goal oriented, very purpose-\n\nvery goal oriented, very purpose-\n \n\nvery goal oriented, very purpose-\ndriven. Um, but I think AI engineering\n\ndriven. Um, but I think AI engineering\n \n\ndriven. Um, but I think AI engineering\nis a bit different. I think AI\n\nis a bit different. I think AI\n \n\nis a bit different. I think AI\nengineering is actually a bit more it's\n\nengineering is actually a bit more it's\n \n\nengineering is actually a bit more it's\nit's a bit closer to excavating. You're\n\nit's a bit closer to excavating. You're\n \n\nit's a bit closer to excavating. You're\nlooking for new capabilities hidden\n\nlooking for new capabilities hidden\n \n\nlooking for new capabilities hidden\nwithin these models. And your toolkit is\n\nwithin these models. And your toolkit is\n \n\nwithin these models. And your toolkit is\na demo. Uh, it your curiosity is kind of\n\na demo. Uh, it your curiosity is kind of\n \n\na demo. Uh, it your curiosity is kind of\nyour flashlight guiding you to where the\n\nyour flashlight guiding you to where the\n \n\nyour flashlight guiding you to where the\ninteresting bits of the models are. And\n\ninteresting bits of the models are. And\n \n\ninteresting bits of the models are. And\nthe way you kind of discover what's\n\nthe way you kind of discover what's\n \n\nthe way you kind of discover what's\npossible is just by making things. Um,\n\npossible is just by making things. Um,\n \n\npossible is just by making things. Um,\nand what's really really interesting is\n\nand what's really really interesting is\n \n\nand what's really really interesting is\nthat even the researchers today at labs\n\nthat even the researchers today at labs\n \n\nthat even the researchers today at labs\nlike OpenAI and Anthropic actually don't\n\nlike OpenAI and Anthropic actually don't\n \n\nlike OpenAI and Anthropic actually don't\nhave a full understanding of the\n\nhave a full understanding of the\n \n\nhave a full understanding of the\ncapabilities of these models. Um, I've\n\ncapabilities of these models. Um, I've\n \n\ncapabilities of these models. Um, I've\nhad OpenAI researchers show me or tell\n\nhad OpenAI researchers show me or tell\n \n\nhad OpenAI researchers show me or tell\nme that they didn't even know GP3 could\n\nme that they didn't even know GP3 could\n \n\nme that they didn't even know GP3 could\ndo this uh could browse the web or that\n\ndo this uh could browse the web or that\n \n\ndo this uh could browse the web or that\nit could generate fully functioning\n\nit could generate fully functioning\n \n\nit could generate fully functioning\nReact components, which was pretty\n\nReact components, which was pretty\n \n\nReact components, which was pretty\ninteresting. Um,\n\ninteresting. Um,\n \n\ninteresting. Um,\nthis is pretty funny. Uh Charles Darwin\n\nthis is pretty funny. Uh Charles Darwin\n \n\nthis is pretty funny." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nUh Charles Darwin\nwas famous for coming coming up with a\n\nwas famous for coming coming up with a\n \n\nwas famous for coming coming up with a\ntheory of evolution but little known\n\ntheory of evolution but little known\n \n\ntheory of evolution but little known\nfact he actually spent eight years\n\nfact he ac..." }, { "role": "assistant", "content": "Uh Charles Darwin\nwas famous for coming coming up with a\n\nwas famous for coming coming up with a\n \n\nwas famous for coming coming up with a\ntheory of evolution but little known\n\ntheory of evolution but little known\n \n\ntheory of evolution but little known\nfact he actually spent eight years\n\nfact he actually spent eight years\n \n\nfact he actually spent eight years\nstudying barnacles like the things on\n\nstudying barnacles like the things on\n \n\nstudying barnacles like the things on\nthe sides of ships and peers and docks. the sides of ships and peers and docks. the sides of ships and peers and docks. Uh he spent he spent eight years\n\nUh he spent he spent eight years\n \n\nUh he spent he spent eight years\nstudying barnacles so much so that\n\nstudying barnacles so much so that\n \n\nstudying barnacles so much so that\npeople thought he was going crazy uh\n\npeople thought he was going crazy uh\n \n\npeople thought he was going crazy uh\nbefore he published evolution. Um, in\n\nbefore he published evolution. Um, in\n \n\nbefore he published evolution. Um, in\nthe moment you wouldn't have known that\n\nthe moment you wouldn't have known that\n \n\nthe moment you wouldn't have known that\nit was important though, but the\n\nit was important though, but the\n \n\nit was important though, but the\nbarnacle studying taught him that\n\nbarnacle studying taught him that\n \n\nbarnacle studying taught him that\nevolution was correct and it was kind of\n\nevolution was correct and it was kind of\n \n\nevolution was correct and it was kind of\nindisputable evidence for his theory. indisputable evidence for his theory. indisputable evidence for his theory. Um, in the moment you actually don't\n\nUm, in the moment you actually don't\n \n\nUm, in the moment you actually don't\nknow what is actually work versus play. know what is actually work versus play. know what is actually work versus play. Uh, what you're doing might feel like\n\nUh, what you're doing might feel like\n \n\nUh, what you're doing might feel like\nit's useless. It might feel like it's\n\nit's useless. It might feel like it's\n \n\nit's useless. It might feel like it's\nleading nowhere. Um, but sometimes you\n\nleading nowhere. Um, but sometimes you\n \n\nleading nowhere. Um, but sometimes you\nneed to study barnacles for eight years\n\nneed to study barnacles for eight years\n \n\nneed to study barnacles for eight years\nbefore you can publish evolution. So, I think we're in this really strange\n\nSo, I think we're in this really strange\n \n\nSo, I think we're in this really strange\nmoment right now in 2025. Uh, these\n\nmoment right now in 2025. Uh, these\n \n\nmoment right now in 2025. Uh, these\nmodels can do amazing things. There's\n\nmodels can do amazing things. There's\n \n\nmodels can do amazing things. There's\ntons of them. Their context windows have\n\ntons of them. Their context windows have\n \n\ntons of them. Their context windows have\nnow expanded from 2,000 to maybe a\n\nnow expanded from 2,000 to maybe a\n \n\nnow expanded from 2,000 to maybe a\nmillion tokens or so. Um, and I think\n\nmillion tokens or so. Um, and I think\n \n\nmillion tokens or so. Um, and I think\ndemos are the way we explore what's\n\ndemos are the way we explore what's\n \n\ndemos are the way we explore what's\npossible. It's the way we we expand the\n\npossible. It's the way we we expand the\n \n\npossible. It's the way we we expand the\nsearch space and kind of see what we can\n\nsearch space and kind of see what we can\n \n\nsearch space and kind of see what we can\ndo with these capabilities. Um, and I\n\ndo with these capabilities. Um, and I\n \n\ndo with these capabilities. Um, and I\nthink it's not something you can predict\n\nthink it's not something you can predict\n \n\nthink it's not something you can predict\nahead of time. It's kind of like\n\nahead of time. It's kind of like\n \n\nahead of time. It's kind of like\ncrossing a foggy pond. Uh, you kind of\n\ncrossing a foggy pond. Uh, you kind of\n \n\ncrossing a foggy pond. Uh, you kind of\ntake one stone, you kind of step on a\n\ntake one stone, you kind of step on a\n \n\ntake one stone, you kind of step on a\nstone and then see where it leads. If it\n\nstone and then see where it leads. If it\n \n\nstone and then see where it leads. If it\nleads somewhere interesting, you can\n\nleads somewhere interesting, you can\n \n\nleads somewhere interesting, you can\nkeep going, but if it doesn't, you can\n\nkeep going, but if it doesn't, you can\n \n\nkeep going, but if it doesn't, you can\nalways backtrack and go a different way. always backtrack and go a different way. always backtrack and go a different way. You'll never be able to plan your route\n\nYou'll never be able to plan your route\n \n\nYou'll never be able to plan your route\nacross the pond ahead of time. You just\n\nacross the pond ahead of time. You just\n \n\nacross the pond ahead of time. You just\nkind of have to take the first step. I came across this really interesting\n\nI came across this really interesting\n \n\nI came across this really interesting\ntweet um and I really like it. It's\n\ntweet um and I really like it. It's\n \n\ntweet um and I really like it. It's\nbecause Anthropic really markets Claude\n\nbecause Anthropic really markets Claude\n \n\nbecause Anthropic really markets Claude\nas kind of a coding model or like a\n\nas kind of a coding model or like a\n \n\nas kind of a coding model or like a\ngeneral reasoning model, but it's like\n\ngeneral reasoning model, but it's like\n \n\ngeneral reasoning model, but it's like\ntrying to sell an intergalactic\n\ntrying to sell an intergalactic\n \n\ntrying to sell an intergalactic\nspaceship as a toaster because one of\n\nspaceship as a toaster because one of\n \n\nspaceship as a toaster because one of\nits surfaces gets hot every once in a\n\nits surfaces gets hot every once in a\n \n\nits surfaces gets hot every once in a\nwhile. And I think this is a really\n\nwhile. And I think this is a really\n \n\nwhile. And I think this is a really\nreally good way of thinking of these\n\nreally good way of thinking of these\n \n\nreally good way of thinking of these\nmodels. There are so much capabilities\n\nmodels. There are so much capabilities\n \n\nmodels. There are so much capabilities\nlatent in them that uh we kind of only\n\nlatent in them that uh we kind of only\n \n\nlatent in them that uh we kind of only\nfocus on the immediate and the obvious. focus on the immediate and the obvious. focus on the immediate and the obvious. But good demos reveal really interesting\n\nBut good demos reveal really interesting\n \n\nBut good demos reveal really interesting\ncapabilities uh mainly through\n\ncapabilities uh mainly through\n \n\ncapabilities uh mainly through\nexploration and play. And I think\n\nexploration and play. And I think\n \n\nexploration and play. And I think\nuncertainty is at the core of being an\n\nuncertainty is at the core of being an\n \n\nuncertainty is at the core of being an\nAI engineer. If you know what you're\n\nAI engineer. If you know what you're\n \n\nAI engineer. If you know what you're\ndoing, you're kind of doing it wrong. doing, you're kind of doing it wrong. doing, you're kind of doing it wrong. And I think if you're uncertain and\n\nAnd I think if you're uncertain and\n \n\nAnd I think if you're uncertain and\nyou're kind of just exploring, uh you'll\n\nyou're kind of just exploring, uh you'll\n \n\nyou're kind of just exploring, uh you'll\nyou'll lead down interesting you'll find\n\nyou'll lead down interesting you'll find\n \n\nyou'll lead down interesting you'll find\nyourself being led down interesting\n\nyourself being led down interesting\n \n\nyourself being led down interesting\npaths. Um\n\npaths. Um\n \n\npaths. Um\nyeah, in subconsciously you notice these\n\nyeah, in subconsciously you notice these\n \n\nyeah, in subconsciously you notice these\npatterns because you've worked with\n\npatterns because you've worked with\n \n\npatterns because you've worked with\nthings that no one else has worked with\n\nthings that no one else has worked with\n \n\nthings that no one else has worked with\nbefore. Uh your life is unique to you. before. Uh your life is unique to you. before. Uh your life is unique to you. Your context window is unique to you and\n\nYour context window is unique to you and\n \n\nYour context window is unique to you and\nno one else has that same shared context\n\nno one else has that same shared context\n \n\nno one else has that same shared context\nwindow. So when you come across an idea\n\nwindow. So when you come across an idea\n \n\nwindow." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nSo when you come across an idea\nin your head, often times you're one of\n\nin your head, often times you're one of\n \n\nin your head, often times you're one of\nthe only people to ever have that idea. the only people to ever have that idea. the only people to ever have that idea. And I think you'd be doi..." }, { "role": "assistant", "content": "So when you come across an idea\nin your head, often times you're one of\n\nin your head, often times you're one of\n \n\nin your head, often times you're one of\nthe only people to ever have that idea. the only people to ever have that idea. the only people to ever have that idea. And I think you'd be doing yourself a\n\nAnd I think you'd be doing yourself a\n \n\nAnd I think you'd be doing yourself a\ngreat injustice if you never actually\n\ngreat injustice if you never actually\n \n\ngreat injustice if you never actually\ntried to make that idea a reality. So I'm going to close with this slide. So I'm going to close with this slide. So I'm going to close with this slide. Um, one of the greatest computing papers\n\nUm, one of the greatest computing papers\n \n\nUm, one of the greatest computing papers\never written was man machine symbiosis\n\never written was man machine symbiosis\n \n\never written was man machine symbiosis\nby lick lighter in the 1960s. And the\n\nby lick lighter in the 1960s. And the\n \n\nby lick lighter in the 1960s. And the\nepitome of technology at the time were\n\nepitome of technology at the time were\n \n\nepitome of technology at the time were\nvacuum tube computers and punch cards. vacuum tube computers and punch cards. vacuum tube computers and punch cards. Uh, if you wanted to write a program, it\n\nUh, if you wanted to write a program, it\n \n\nUh, if you wanted to write a program, it\nwould probably take a few hours, maybe\n\nwould probably take a few hours, maybe\n \n\nwould probably take a few hours, maybe\neven days to run. Um, meanwhile today we\n\neven days to run. Um, meanwhile today we\n \n\neven days to run. Um, meanwhile today we\nhave claopus 4 on our computers. It's\n\nhave claopus 4 on our computers. It's\n \n\nhave claopus 4 on our computers. It's\nactually kind of insane. And I think\n\nactually kind of insane. And I think\n \n\nactually kind of insane. And I think\nLick Lighter genuinely would have killed\n\nLick Lighter genuinely would have killed\n \n\nLick Lighter genuinely would have killed\nsomeone to have an hour with the tools\n\nsomeone to have an hour with the tools\n \n\nsomeone to have an hour with the tools\nwe have today. Um, and like I'm not even\n\nwe have today. Um, and like I'm not even\n \n\nwe have today. Um, and like I'm not even\njoking. I think it's kind of important. joking. I think it's kind of important. joking. I think it's kind of important. So much so that I feel like today we\n\nSo much so that I feel like today we\n \n\nSo much so that I feel like today we\nhave a moral obligation to do him\n\nhave a moral obligation to do him\n \n\nhave a moral obligation to do him\njustice and everyone else in the field\n\njustice and everyone else in the field\n \n\njustice and everyone else in the field\nthat came before us. Uh, not only to\n\nthat came before us. Uh, not only to\n \n\nthat came before us. Uh, not only to\njust follow your curiosity, but to share\n\njust follow your curiosity, but to share\n \n\njust follow your curiosity, but to share\nwhat you explore with the world. Um,\n\nwhat you explore with the world. Um,\n \n\nwhat you explore with the world. Um,\nbecause by sharing your demos, you kind\n\nbecause by sharing your demos, you kind\n \n\nbecause by sharing your demos, you kind\nof share what's possible with these\n\nof share what's possible with these\n \n\nof share what's possible with these\nmodels. And I think that's how we move\n\nmodels. And I think that's how we move\n \n\nmodels. And I think that's how we move\nthe field forward. And um yeah, that's\n\nthe field forward. And um yeah, that's\n \n\nthe field forward. And um yeah, that's\nthat's really it. Your unique\n\nthat's really it. Your unique\n \n\nthat's really it. Your unique\nperspective shouldn't be wasted, and I\n\nperspective shouldn't be wasted, and I\n \n\nperspective shouldn't be wasted, and I\nthink you have a moral responsibility to\n\nthink you have a moral responsibility to\n \n\nthink you have a moral responsibility to\nshare them with the world. share them with the world. share them with the world. Thank you. All right, we've got some time for\n\nAll right, we've got some time for\n \n\nAll right, we've got some time for\nquestions. Does anybody have any? None. None. None. Cool. Thank you, guys. Cool. Thank you, guys. Cool. Thank you, guys. Appreciate it. Appreciate it. Appreciate it. We're doing really good for time. I was\n\nWe're doing really good for time. I was\n \n\nWe're doing really good for time. I was\ngoing to ask you if you had like other\n\ngoing to ask you if you had like other\n \n\ngoing to ask you if you had like other\ndemos that you wanted to show us because\n\ndemos that you wanted to show us because\n \n\ndemos that you wanted to show us because\nthe I I liked seeing the 2020 versions\n\nthe I I liked seeing the 2020 versions\n \n\nthe I I liked seeing the 2020 versions\nof things. of things. of things. I have a few more actually. Do you want\n\nI have a few more actually. Do you want\n \n\nI have a few more actually. Do you want\nme to pull them? me to pull them? me to pull them? We've got eight minutes. You might as\n\nWe've got eight minutes. You might as\n \n\nWe've got eight minutes. You might as\nwell\n\nwell\n \n\nwell\nas long as they don't use Wi-Fi because\n\nas long as they don't use Wi-Fi because\n \n\nas long as they don't use Wi-Fi because\napparently that's the running joke of\n\napparently that's the running joke of\n \n\napparently that's the running joke of\nthis conference. this conference. this conference. I I think I might have a few downloaded. I I think I might have a few downloaded. I I think I might have a few downloaded. Let me check. For the people in the room, I did try\n\nFor the people in the room, I did try\n \n\nFor the people in the room, I did try\nthat basketball one except I tried to\n\nthat basketball one except I tried to\n \n\nthat basketball one except I tried to\napply to running and it works really\n\napply to running and it works really\n \n\napply to running and it works really\nwell and it pretty much gave the same\n\nwell and it pretty much gave the same\n \n\nwell and it pretty much gave the same\nfeedback that my $600 a month running\n\nfeedback that my $600 a month running\n \n\nfeedback that my $600 a month running\ncoach would give me. Oh wow. coach would give me. Oh wow. coach would give me. Oh wow. And I thought, wait, I think I can\n\nAnd I thought, wait, I think I can\n \n\nAnd I thought, wait, I think I can\ncancel this. cancel this. cancel this. Did it give you like pretty good advice\n\nDid it give you like pretty good advice\n \n\nDid it give you like pretty good advice\non your gate? on your gate? on your gate? It gave me not not just the gate per\n\nIt gave me not not just the gate per\n \n\nIt gave me not not just the gate per\nstep. step. step. And so that's something that my coach\n\nAnd so that's something that my coach\n \n\nAnd so that's something that my coach\nwould never able be able to do. What I\n\nwould never able be able to do. What I\n \n\nwould never able be able to do. What I\ncouldn't get to figure out is um how it\n\ncouldn't get to figure out is um how it\n \n\ncouldn't get to figure out is um how it\nhad the little arrow on top of the head. had the little arrow on top of the head. had the little arrow on top of the head. But if I had 20 more minutes to\n\nBut if I had 20 more minutes to\n \n\nBut if I had 20 more minutes to\nYeah. Here's a pretty cool demo. think\n\nYeah. Here's a pretty cool demo. think\n \n\nYeah. Here's a pretty cool demo. think\nwas also from 2020. Um, let's see if\n\nwas also from 2020. Um, let's see if\n \n\nwas also from 2020. Um, let's see if\nit's playing. It's not. Let me Oh, it\n\nit's playing. It's not. Let me Oh, it\n \n\nit's playing. It's not. Let me Oh, it\nis. Um, yeah, this was about a few weeks\n\nis. Um, yeah, this was about a few weeks\n \n\nis. Um, yeah, this was about a few weeks\nafter the GP3 API came out. And I think\n\nafter the GP3 API came out. And I think\n \n\nafter the GP3 API came out." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nAnd I think\nthe way I came across the API was really\n\nthe way I came across the API was really\n \n\nthe way I came across the API was really\nfunny. Uh, someone had said to me, you\n\nfunny. Uh, someone had said to me, you\n \n\nfunny. Uh, someone had said to me, you\nhave to try this out. OpenAI has created..." }, { "role": "assistant", "content": "And I think\nthe way I came across the API was really\n\nthe way I came across the API was really\n \n\nthe way I came across the API was really\nfunny. Uh, someone had said to me, you\n\nfunny. Uh, someone had said to me, you\n \n\nfunny. Uh, someone had said to me, you\nhave to try this out. OpenAI has created\n\nhave to try this out. OpenAI has created\n \n\nhave to try this out. OpenAI has created\nAGI and it's here available today and no\n\nAGI and it's here available today and no\n \n\nAGI and it's here available today and no\none's really talking about it. And I was\n\none's really talking about it. And I was\n \n\none's really talking about it. And I was\nlike, okay, let's see what this is\n\nlike, okay, let's see what this is\n \n\nlike, okay, let's see what this is\nabout. Um, and I I quickly realized I\n\nabout. Um, and I I quickly realized I\n \n\nabout. Um, and I I quickly realized I\ncould actually write code, but writing\n\ncould actually write code, but writing\n \n\ncould actually write code, but writing\ncode in the text interface was not\n\ncode in the text interface was not\n \n\ncode in the text interface was not\nreally the best way to do it. Uh, so you\n\nreally the best way to do it. Uh, so you\n \n\nreally the best way to do it. Uh, so you\nactually hook it up to an API, put a\n\nactually hook it up to an API, put a\n \n\nactually hook it up to an API, put a\ncompiler in the browser, and you get\n\ncompiler in the browser, and you get\n \n\ncompiler in the browser, and you get\nthis like nice back and forth visual\n\nthis like nice back and forth visual\n \n\nthis like nice back and forth visual\ninterface. We kind of take this granted\n\ninterface. We kind of take this granted\n \n\ninterface. We kind of take this granted\nfor today with tools like cursor where\n\nfor today with tools like cursor where\n \n\nfor today with tools like cursor where\nyou can kind of like chat with your code\n\nyou can kind of like chat with your code\n \n\nyou can kind of like chat with your code\nin the sidebar. But, uh, in 2020, this\n\nin the sidebar. But, uh, in 2020, this\n \n\nin the sidebar. But, uh, in 2020, this\nfelt really really different than what\n\nfelt really really different than what\n \n\nfelt really really different than what\nanything was possible. Um, here I'm like\n\nanything was possible. Um, here I'm like\n \n\nanything was possible. Um, here I'm like\nworking on a like a really really basic\n\nworking on a like a really really basic\n \n\nworking on a like a really really basic\nlike banking app where you just ask it\n\nlike banking app where you just ask it\n \n\nlike banking app where you just ask it\nto add $3 or subtract another $5. Uh, it\n\nto add $3 or subtract another $5. Uh, it\n \n\nto add $3 or subtract another $5. Uh, it\nwas pretty funny because like the bugs\n\nwas pretty funny because like the bugs\n \n\nwas pretty funny because like the bugs\nwere really bad. You could actually\n\nwere really bad. You could actually\n \n\nwere really bad. You could actually\nthere's a button where you could give\n\nthere's a button where you could give\n \n\nthere's a button where you could give\naway all your money and if you were in\n\naway all your money and if you were in\n \n\naway all your money and if you were in\ndebt it would just like negate it and\n\ndebt it would just like negate it and\n \n\ndebt it would just like negate it and\nmake you make your balance go to zero\n\nmake you make your balance go to zero\n \n\nmake you make your balance go to zero\nagain. Um, but this really was I think\n\nagain. Um, but this really was I think\n \n\nagain. Um, but this really was I think\nthe start of vibe coding and it really a\n\nthe start of vibe coding and it really a\n \n\nthe start of vibe coding and it really a\nlot of people to take these models not\n\nlot of people to take these models not\n \n\nlot of people to take these models not\nonly as like language models but kind of\n\nonly as like language models but kind of\n \n\nonly as like language models but kind of\nreasoning engines. Um, yeah. And I think\n\nreasoning engines. Um, yeah. And I think\n \n\nreasoning engines. Um, yeah. And I think\nI think um the way to think about these\n\nI think um the way to think about these\n \n\nI think um the way to think about these\nmodels is really that like they're these\n\nmodels is really that like they're these\n \n\nmodels is really that like they're these\nreally really intelligent in a way\n\nreally really intelligent in a way\n \n\nreally really intelligent in a way\nbeings, which sounds kind of weird to\n\nbeings, which sounds kind of weird to\n \n\nbeings, which sounds kind of weird to\nsay out loud, but that's like the mental\n\nsay out loud, but that's like the mental\n \n\nsay out loud, but that's like the mental\nmodel I have for them. And you kind of\n\nmodel I have for them. And you kind of\n \n\nmodel I have for them. And you kind of\nhook them up to these different\n\nhook them up to these different\n \n\nhook them up to these different\napparatuses and they can kind of work\n\napparatuses and they can kind of work\n \n\napparatuses and they can kind of work\nthem and you kind of like instill these\n\nthem and you kind of like instill these\n \n\nthem and you kind of like instill these\ntools with a sense of like uh purpose\n\ntools with a sense of like uh purpose\n \n\ntools with a sense of like uh purpose\nand agency. Um, yeah. I I really just\n\nand agency. Um, yeah. I I really just\n \n\nand agency. Um, yeah. I I really just\nhope a lot more people are inspired to\n\nhope a lot more people are inspired to\n \n\nhope a lot more people are inspired to\nwork on demos because the capabilities\n\nwork on demos because the capabilities\n \n\nwork on demos because the capabilities\nwe have today are really impressive and\n\nwe have today are really impressive and\n \n\nwe have today are really impressive and\nyou'd be really doing a disservice by\n\nyou'd be really doing a disservice by\n \n\nyou'd be really doing a disservice by\nnot just like building something really\n\nnot just like building something really\n \n\nnot just like building something really\nfun and simple and sharing it with the\n\nfun and simple and sharing it with the\n \n\nfun and simple and sharing it with the\nworld. world. world. Yeah. Can over there. Yeah. Can over there. Yeah. Can over there. Yeah. Yes, it was the base model. Uh, we\n\nYes, it was the base model. Uh, we\n \n\nYes, it was the base model. Uh, we\ndidn't have an we didn't have an\n\ndidn't have an we didn't have an\n \n\ndidn't have an we didn't have an\ninstruct model until about a year later. instruct model until about a year later. instruct model until about a year later. Um, so what it was essentially similar\n\nUm, so what it was essentially similar\n \n\nUm, so what it was essentially similar\nto base models we have today where you\n\nto base models we have today where you\n \n\nto base models we have today where you\ngive it a prefix and it just completes\n\ngive it a prefix and it just completes\n \n\ngive it a prefix and it just completes\nit. So you you prompt engineer with a\n\nit. So you you prompt engineer with a\n \n\nit. So you you prompt engineer with a\nfew examples and that's usually good\n\nfew examples and that's usually good\n \n\nfew examples and that's usually good\nenough. Any other demos you got there? Any other demos you got there? Any other demos you got there? Uh I can I can go into the archive. Uh I can I can go into the archive. Uh I can I can go into the archive. Do you want to go through the entire\n\nDo you want to go through the entire\n \n\nDo you want to go through the entire\ndesktop while you're at it? desktop while you're at it? desktop while you're at it? Let me close this one. I I I have a few but I I don't know. Do\n\nI I I have a few but I I don't know. Do\n \n\nI I I have a few but I I don't know. Do\nwe have time? We have five minutes. we have time? We have five minutes. we have time? We have five minutes. Let's see if I can find anything." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nUm\n\nLet's see if I can find anything. Um\n \n\nLet's see if I can find anything. Um\nokay, let let me try something really\n\nokay, let let me try something really\n \n\nokay, let let me try something really\nquick. Yeah, sure. How did\n\n \n\nI never\n\n \n\nSure. Yeah. I think um a lot of it was\n\nSure. Yeah. I thin..." }, { "role": "assistant", "content": "Um\n\nLet's see if I can find anything. Um\n \n\nLet's see if I can find anything. Um\nokay, let let me try something really\n\nokay, let let me try something really\n \n\nokay, let let me try something really\nquick. Yeah, sure. How did\n\n \n\nI never\n\n \n\nSure. Yeah. I think um a lot of it was\n\nSure. Yeah. I think um a lot of it was\n \n\nSure. Yeah. I think um a lot of it was\njust kind of introspecting on why I made\n\njust kind of introspecting on why I made\n \n\njust kind of introspecting on why I made\nthese demos in the first place. Um, a\n\nthese demos in the first place. Um, a\n \n\nthese demos in the first place. Um, a\nlarge part of it came from a sense of\n\nlarge part of it came from a sense of\n \n\nlarge part of it came from a sense of\nfrustration that we have these really\n\nfrustration that we have these really\n \n\nfrustration that we have these really\npowerful models today and no one really\n\npowerful models today and no one really\n \n\npowerful models today and no one really\nknows what they're capable of doing. Um,\n\nknows what they're capable of doing. Um,\n \n\nknows what they're capable of doing. Um,\nand I think I examined it a bit further\n\nand I think I examined it a bit further\n \n\nand I think I examined it a bit further\nand it did feel like a sense of moral\n\nand it did feel like a sense of moral\n \n\nand it did feel like a sense of moral\nobligation. Uh, you have these pioneers\n\nobligation. Uh, you have these pioneers\n \n\nobligation. Uh, you have these pioneers\nof computing uh from the 60s and 70s and\n\nof computing uh from the 60s and 70s and\n \n\nof computing uh from the 60s and 70s and\n80s like Lick Lighter and Alen K and\n\n80s like Lick Lighter and Alen K and\n \n\n80s like Lick Lighter and Alen K and\nwhatnot and they came up with these\n\nwhatnot and they came up with these\n \n\nwhatnot and they came up with these\ngrand ideas uh with the computers they\n\ngrand ideas uh with the computers they\n \n\ngrand ideas uh with the computers they\nhad available to them. They just\n\nhad available to them. They just\n \n\nhad available to them. They just\ncouldn't make it possible. And I look at\n\ncouldn't make it possible. And I look at\n \n\ncouldn't make it possible. And I look at\nwhat we have today and it's kind of like\n\nwhat we have today and it's kind of like\n \n\nwhat we have today and it's kind of like\nwe're spoiled by so many amazing pieces\n\nwe're spoiled by so many amazing pieces\n \n\nwe're spoiled by so many amazing pieces\nof technology and we're kind of just\n\nof technology and we're kind of just\n \n\nof technology and we're kind of just\nmaking the same things all over again. making the same things all over again. making the same things all over again. Uh but really I think if you look back\n\nUh but really I think if you look back\n \n\nUh but really I think if you look back\nat like what people were writing about\n\nat like what people were writing about\n \n\nat like what people were writing about\nin the 60s and 70s, there's a whole gold\n\nin the 60s and 70s, there's a whole gold\n \n\nin the 60s and 70s, there's a whole gold\nmine of ideas there that we can revisit\n\nmine of ideas there that we can revisit\n \n\nmine of ideas there that we can revisit\nand actually make possible today. Um in\n\nand actually make possible today. Um in\n \n\nand actually make possible today. Um in\nthe uh man computer man machine\n\nthe uh man computer man machine\n \n\nthe uh man computer man machine\nsymbiosis paper lick lighter talks about\n\nsymbiosis paper lick lighter talks about\n \n\nsymbiosis paper lick lighter talks about\nan assistant that knows everything\n\nan assistant that knows everything\n \n\nan assistant that knows everything\nyou're working on and has like perfect\n\nyou're working on and has like perfect\n \n\nyou're working on and has like perfect\ncontext and can help you with anything\n\ncontext and can help you with anything\n \n\ncontext and can help you with anything\nimmediately. Um, and here we have like\n\nimmediately. Um, and here we have like\n \n\nimmediately. Um, and here we have like\nchat GBT where every time you want to\n\nchat GBT where every time you want to\n \n\nchat GBT where every time you want to\ntalk to it, you press new chat and it\n\ntalk to it, you press new chat and it\n \n\ntalk to it, you press new chat and it\nhas no memory of what you've talked\n\nhas no memory of what you've talked\n \n\nhas no memory of what you've talked\nabout beforehand minus like a few basic\n\nabout beforehand minus like a few basic\n \n\nabout beforehand minus like a few basic\nfacts. Um, and I think it's really just\n\nfacts. Um, and I think it's really just\n \n\nfacts. Um, and I think it's really just\nit really boils down to wanting to uh\n\nit really boils down to wanting to uh\n \n\nit really boils down to wanting to uh\nkind of do the ideas that these pioneers\n\nkind of do the ideas that these pioneers\n \n\nkind of do the ideas that these pioneers\ncame up with justice beforehand. Thank you. Thank you. Thank you. Three minutes. So, did you have\n\nThree minutes. So, did you have\n \n\nThree minutes. So, did you have\nsomething? something? something? Um\n\nUm\n \n\nUm\num yeah, this was an old GPD3 demo where\n\num yeah, this was an old GPD3 demo where\n \n\num yeah, this was an old GPD3 demo where\nthe idea was how do you get these models\n\nthe idea was how do you get these models\n \n\nthe idea was how do you get these models\nto solve very large and ambitious\n\nto solve very large and ambitious\n \n\nto solve very large and ambitious\nproblems. It was called multivac and the\n\nproblems. It was called multivac and the\n \n\nproblems. It was called multivac and the\nidea was you can't really fit everything\n\nidea was you can't really fit everything\n \n\nidea was you can't really fit everything\ninto a 2000 context window token context\n\ninto a 2000 context window token context\n \n\ninto a 2000 context window token context\nwindow. So what you do is you you\n\nwindow. So what you do is you you\n \n\nwindow. So what you do is you you\nessentially break down the problems into\n\nessentially break down the problems into\n \n\nessentially break down the problems into\nmore digestible sub problems and you\n\nmore digestible sub problems and you\n \n\nmore digestible sub problems and you\nhave this kind of visual interface to\n\nhave this kind of visual interface to\n \n\nhave this kind of visual interface to\nhelp you see where things are going. So\n\nhelp you see where things are going. So\n \n\nhelp you see where things are going. So\nyou can give it some really ambitious\n\nyou can give it some really ambitious\n \n\nyou can give it some really ambitious\nproblem like how do you solve climate\n\nproblem like how do you solve climate\n \n\nproblem like how do you solve climate\nchange and it might come up with things\n\nchange and it might come up with things\n \n\nchange and it might come up with things\nlike convince more people to go\n\nlike convince more people to go\n \n\nlike convince more people to go\nvegetarian or build climate or build\n\nvegetarian or build climate or build\n \n\nvegetarian or build climate or build\nwind turbines and like install more\n\nwind turbines and like install more\n \n\nwind turbines and like install more\nsolar panels and then you can click on\n\nsolar panels and then you can click on\n \n\nsolar panels and then you can click on\neach of the sub ideas and it kind of\n\neach of the sub ideas and it kind of\n \n\neach of the sub ideas and it kind of\nbreaks it down even further. Um yeah I\n\nbreaks it down even further. Um yeah I\n \n\nbreaks it down even further. Um yeah I\nthink one of the core ideas behind this\n\nthink one of the core ideas behind this\n \n\nthink one of the core ideas behind this\nwas like these models are a lot more\n\nwas like these models are a lot more\n \n\nwas like these models are a lot more\nthan just text completion models but I\n\nthan just text completion models but I\n \n\nthan just text completion models but I\nthink they can be useful as like very\n\nthink they can be useful as like very\n \n\nthink they can be useful as like very\nhelpful reasoning assistance uh\n\nhelpful reasoning assistance uh\n \n\nhelpful reasoning assistance uh\nspecifically at solving big problems. Uh\n\nspecifically at solving big problems. Uh\n \n\nspecifically at solving big problems." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nUh\nso much so that they could come up with\n\nso much so that they could come up with\n \n\nso much so that they could come up with\nideas on their own one day and hopefully\n\nideas on their own one day and hopefully\n \n\nideas on their own one day and hopefully\nbe really useful thought partners. Yeah. I mea..." }, { "role": "assistant", "content": "Uh\nso much so that they could come up with\n\nso much so that they could come up with\n \n\nso much so that they could come up with\nideas on their own one day and hopefully\n\nideas on their own one day and hopefully\n \n\nideas on their own one day and hopefully\nbe really useful thought partners. Yeah. I mean looking at it now it's\n\nYeah. I mean looking at it now it's\n \n\nYeah. I mean looking at it now it's\npretty rudimentary but I maybe someone\n\npretty rudimentary but I maybe someone\n \n\npretty rudimentary but I maybe someone\nshould make a new version of this with\n\nshould make a new version of this with\n \n\nshould make a new version of this with\nlike Opus Max. like Opus Max. like Opus Max. Yeah. I mean someone here should do it. Yeah. I mean someone here should do it. Yeah. I mean someone here should do it. I think that'd be pretty cool. I think that'd be pretty cool. I think that'd be pretty cool. Um yeah that's about it guys. Thanks. [Music]" } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\n[Music]\n\n[Music]\n \n\n[Music]\nmy name is John kzi and I work on a team\n\nmy name is John kzi and I work on a team\n \n\nmy name is John kzi and I work on a team\nat Jane Street called AI assistant our\n\nat Jane Street called AI assistant our\n \n\nat Jane Street called AI assistant our\ngroup roughly uh is ther..." }, { "role": "assistant", "content": "[Music]\n\n[Music]\n \n\n[Music]\nmy name is John kzi and I work on a team\n\nmy name is John kzi and I work on a team\n \n\nmy name is John kzi and I work on a team\nat Jane Street called AI assistant our\n\nat Jane Street called AI assistant our\n \n\nat Jane Street called AI assistant our\ngroup roughly uh is there at Jan Street\n\ngroup roughly uh is there at Jan Street\n \n\ngroup roughly uh is there at Jan Street\nto try to maximize the value that Jan\n\nto try to maximize the value that Jan\n \n\nto try to maximize the value that Jan\nStreet can get from large language\n\nStreet can get from large language\n \n\nStreet can get from large language\nmodels and I've spent my entire career\n\nmodels and I've spent my entire career\n \n\nmodels and I've spent my entire career\nuh in Dev tools before I worked at Jan\n\nuh in Dev tools before I worked at Jan\n \n\nuh in Dev tools before I worked at Jan\nstreet I was a GitHub for a long time\n\nstreet I was a GitHub for a long time\n \n\nstreet I was a GitHub for a long time\nand then before that I worked at a\n\nand then before that I worked at a\n \n\nand then before that I worked at a\nvariety of other Dev tools companies and\n\nvariety of other Dev tools companies and\n \n\nvariety of other Dev tools companies and\nllms kind of present this really amazing\n\nllms kind of present this really amazing\n \n\nllms kind of present this really amazing\nopportunity in that they're so\n\nopportunity in that they're so\n \n\nopportunity in that they're so\nopen-ended that we can build kind of\n\nopen-ended that we can build kind of\n \n\nopen-ended that we can build kind of\nanything that we can imagine and it\n\nanything that we can imagine and it\n \n\nanything that we can imagine and it\nseems like right now the only thing\n\nseems like right now the only thing\n \n\nseems like right now the only thing\nmoving faster than the progress of the\n\nmoving faster than the progress of the\n \n\nmoving faster than the progress of the\nmodels is kind of our creativity around\n\nmodels is kind of our creativity around\n \n\nmodels is kind of our creativity around\nhow to employ them uh at Jan street\n\nhow to employ them uh at Jan street\n \n\nhow to employ them uh at Jan street\nthough we've made some choices that make\n\nthough we've made some choices that make\n \n\nthough we've made some choices that make\nadoption of off-the-shelf tooling a\n\nadoption of off-the-shelf tooling a\n \n\nadoption of off-the-shelf tooling a\nlittle bit more difficult than it is for\n\nlittle bit more difficult than it is for\n \n\nlittle bit more difficult than it is for\nother companies\n\nother companies\n \n\nother companies\nand kind of the biggest reason that we\n\nand kind of the biggest reason that we\n \n\nand kind of the biggest reason that we\nhave this problem is that we use o camel\n\nhave this problem is that we use o camel\n \n\nhave this problem is that we use o camel\nas a development platform for those not\n\nas a development platform for those not\n \n\nas a development platform for those not\nfamiliar with oaml it is a a functional\n\nfamiliar with oaml it is a a functional\n \n\nfamiliar with oaml it is a a functional\nvery powerful language but it's also\n\nvery powerful language but it's also\n \n\nvery powerful language but it's also\nincredibly obscure language uh it was\n\nincredibly obscure language uh it was\n \n\nincredibly obscure language uh it was\nbuilt in France and its most common\n\nbuilt in France and its most common\n \n\nbuilt in France and its most common\napplications are in things like theorem\n\napplications are in things like theorem\n \n\napplications are in things like theorem\nproving or formal verification it's also\n\nproving or formal verification it's also\n \n\nproving or formal verification it's also\nused to write programming\n\nused to write programming\n \n\nused to write programming\nlanguages um we use oaml kind of for\n\nlanguages um we use oaml kind of for\n \n\nlanguages um we use oaml kind of for\neverything at change\n\neverything at change\n \n\neverything at change\nso just a couple quick examples when we\n\nso just a couple quick examples when we\n \n\nso just a couple quick examples when we\nwrite web applications of course web\n\nwrite web applications of course web\n \n\nwrite web applications of course web\napplications have to be written in\n\napplications have to be written in\n \n\napplications have to be written in\nJavaScript but instead we write oaml and\n\nJavaScript but instead we write oaml and\n \n\nJavaScript but instead we write oaml and\nwe use a library called JS of oaml that\n\nwe use a library called JS of oaml that\n \n\nwe use a library called JS of oaml that\nis essentially a oaml bik code to\n\nis essentially a oaml bik code to\n \n\nis essentially a oaml bik code to\nJavaScript\n\nJavaScript\n \n\nJavaScript\ntranspiler when we write plugins for Vim\n\ntranspiler when we write plugins for Vim\n \n\ntranspiler when we write plugins for Vim\nthose have to be written in Vim script\n\nthose have to be written in Vim script\n \n\nthose have to be written in Vim script\nuh but we actually use a library called\n\nuh but we actually use a library called\n \n\nuh but we actually use a library called\nvaml which again is oaml to vimscript\n\nvaml which again is oaml to vimscript\n \n\nvaml which again is oaml to vimscript\ntranspiler and uh even people at the\n\ntranspiler and uh even people at the\n \n\ntranspiler and uh even people at the\ncompany that are working on fpga code\n\ncompany that are working on fpga code\n \n\ncompany that are working on fpga code\nthey're not writing verog they're\n\nthey're not writing verog they're\n \n\nthey're not writing verog they're\nwriting in an O camel Library called\n\nwriting in an O camel Library called\n \n\nwriting in an O camel Library called\nhard\n\nhard\n \n\nhard\ncamel uh so why are the tools on the\n\ncamel uh so why are the tools on the\n \n\ncamel uh so why are the tools on the\nmarket available not good for working\n\nmarket available not good for working\n \n\nmarket available not good for working\nwith oaml I think it kind of comes down\n\nwith oaml I think it kind of comes down\n \n\nwith oaml I think it kind of comes down\nto a few primary reasons the first and\n\nto a few primary reasons the first and\n \n\nto a few primary reasons the first and\nthe most important is that models\n\nthe most important is that models\n \n\nthe most important is that models\nthemselves are just not very good at\n\nthemselves are just not very good at\n \n\nthemselves are just not very good at\noaml and this isn't the fault of the AI\n\noaml and this isn't the fault of the AI\n \n\noaml and this isn't the fault of the AI\nlabs this is just kind of a byproduct of\n\nlabs this is just kind of a byproduct of\n \n\nlabs this is just kind of a byproduct of\nthe amount of data that exists for\n\nthe amount of data that exists for\n \n\nthe amount of data that exists for\ntraining so it's there's a really good\n\ntraining so it's there's a really good\n \n\ntraining so it's there's a really good\nchance that the amount of okl code that\n\nchance that the amount of okl code that\n \n\nchance that the amount of okl code that\nwe have inside of J street it's just\n\nwe have inside of J street it's just\n \n\nwe have inside of J street it's just\nmore than like the total combined amount\n\nmore than like the total combined amount\n \n\nmore than like the total combined amount\nof oam code that there exists in the\n\nof oam code that there exists in the\n \n\nof oam code that there exists in the\nworld uh outside of our\n\nworld uh outside of our\n \n\nworld uh outside of our\nwalls the second is that we've made\n\nwalls the second is that we've made\n \n\nwalls the second is that we've made\nthings really hard on ourselves\n\nthings really hard on ourselves\n \n\nthings really hard on ourselves\npartially as a byproduct of working in O\n\npartially as a byproduct of working in O\n \n\npartially as a byproduct of working in O\ncamel we've had to build our own build\n\ncamel we've had to build our own build\n \n\ncamel we've had to build our own build\nsystems we built our own distributed\n\nsystems we built our own distributed\n \n\nsystems we built our own distributed\nbuild environment we even built our own\n\nbuild environment we even built our own\n \n\nbuild environment we even built our own\ncode review system which is called\n\ncode review system which is called\n \n\ncode review system which is called\niron we develop all of our software on a\n\niron we develop all of our software on a\n \n\niron we develop all of our software on a\ngiant monor repo application and just\n\ngiant monor repo application and just\n \n\ngiant monor repo application and just\nfor fun instead of storing that monor\n\nfor fun instead of storing that monor\n \n\nfor fun instead of storing that monor\nrepo in git we store it in mercurial\n\n \n\nand uh at last count 67% of the firm\n\nand uh at last count 67% of the firm\n \n\nand uh at last count 67% of the firm\nuses emac instead of normal editors\n\nuses emac instead of normal editors\n \n\nuses emac instead of normal editors\nmaybe like vs code uh we do have people\n\nmaybe like vs code uh we do have people\n \n\nmaybe like vs code uh we do have people\nusing vs code but emac is the most\n\nusing vs code but emac is the most\n \n\nusing vs code but emac is the most\npopular and the last thing is we're\n\npopular and the last thing is we're\n \n\npopular and the last thing is we're\ndreamers I mean kind of everyone in this\n\ndreamers I mean kind of everyone in this\n \n\ndreamers I mean kind of everyone in this\nroom hopefully is is a dreamer in a way\n\nroom hopefully is is a dreamer in a way\n \n\nroom hopefully is is a dreamer in a way\nuh and what I mean by this is we want\n\nuh and what I mean by this is we want\n \n\nuh and what I mean by this is we want\nthe ability to kind of take llms and\n\nthe ability to kind of take llms and\n \n\nthe ability to kind of take llms and\napply them to different parts of our\n\napply them to different parts of our\n \n\napply them to different parts of our\ndevelopment flow and light up different\n\ndevelopment flow and light up different\n \n\ndevelopment flow and light up different\nparts so maybe we want to use large\n\nparts so maybe we want to use large\n \n\nparts so maybe we want to use large\nlanguage models to resolve merge\n\nlanguage models to resolve merge\n \n\nlanguage models to resolve merge\nconflict or build better feature\n\nconflict or build better feature\n \n\nconflict or build better feature\ndescriptions or figure out who reviewers\n\ndescriptions or figure out who reviewers\n \n\ndescriptions or figure out who reviewers\nfor features be and we don't want to be\n\nfor features be and we don't want to be\n \n\nfor features be and we don't want to be\nhampered by the boundaries between\n\nhampered by the boundaries between\n \n\nhampered by the boundaries between\ndifferent systems when we do\n\ndifferent systems when we do\n \n\ndifferent systems when we do\nthat over the next 15 minutes I'm going\n\nthat over the next 15 minutes I'm going\n \n\nthat over the next 15 minutes I'm going\nto cover our approach to large language\n\nto cover our approach to large language\n \n\nto cover our approach to large language\nmodels at CH Street uh particularly when\n\nmodels at CH Street uh particularly when\n \n\nmodels at CH Street uh particularly when\nit comes to developer tools um I'm going\n\nit comes to developer tools um I'm going\n \n\nit comes to developer tools um I'm going\nto talk about custom models that we're\n\nto talk about custom models that we're\n \n\nto talk about custom models that we're\nbuilding and how we build them I'm going\n\nbuilding and how we build them I'm going\n \n\nbuilding and how we build them I'm going\nto talk about editor Integrations so\n\nto talk about editor Integrations so\n \n\nto talk about editor Integrations so\nthese are the Integrations into uh to vs\n\nthese are the Integrations into uh to vs\n \n\nthese are the Integrations into uh to vs\ncode emac and neovim\n\ncode emac and neovim\n \n\ncode emac and neovim\nand I will talk about uh the ability\n\nand I will talk about uh the ability\n \n\nand I will talk about uh the ability\nthat we've built over time to evaluate\n\nthat we've built over time to evaluate\n \n\nthat we've built over time to evaluate\nmodels and figure out how to make them\n\nmodels and figure out how to make them\n \n\nmodels and figure out how to make them\nperform\n\nperform\n \n\nperform\nbest and I guess at first glance it's\n\nbest and I guess at first glance it's\n \n\nbest and I guess at first glance it's\nnot really obvious that training models\n\nnot really obvious that training models\n \n\nnot really obvious that training models\nat all is a good idea I mean it's very\n\nat all is a good idea I mean it's very\n \n\nat all is a good idea I mean it's very\nexpensive proposition it takes a lot of\n\nexpensive proposition it takes a lot of\n \n\nexpensive proposition it takes a lot of\ntime and it can go wrong in a ton of\n\ntime and it can go wrong in a ton of\n \n\ntime and it can go wrong in a ton of\ndifferent ways who here has trained a\n\ndifferent ways who here has trained a\n \n\ndifferent ways who here has trained a\nmodel before or tried to train something\n\nmodel before or tried to train something\n \n\nmodel before or tried to train something\nlike a model maybe took a foundation\n\nlike a model maybe took a foundation\n \n\nlike a model maybe took a foundation\nmodel and trained on top of\n\nmodel and trained on top of\n \n\nmodel and trained on top of\nit\n\nit\n \n\nit\ncool we were more convinced after we\n\ncool we were more convinced after we\n \n\ncool we were more convinced after we\nread this paper this is a paper from\n\nread this paper this is a paper from\n \n\nread this paper this is a paper from\nmeta about a project called code compose\n\nmeta about a project called code compose\n \n\nmeta about a project called code compose\nand in this paper they detail their\n\nand in this paper they detail their\n \n\nand in this paper they detail their\nresults fine tuning a model specifically\n\nresults fine tuning a model specifically\n \n\nresults fine tuning a model specifically\nfor use with hack uh hack is actually\n\nfor use with hack uh hack is actually\n \n\nfor use with hack uh hack is actually\npretty similar to O camel uh not in its\n\npretty similar to O camel uh not in its\n \n\npretty similar to O camel uh not in its\nlike syntax or function but really just\n\nlike syntax or function but really just\n \n\nlike syntax or function but really just\nin the fact that it's used primarily at\n\nin the fact that it's used primarily at\n \n\nin the fact that it's used primarily at\none company and not really used much\n\none company and not really used much\n \n\none company and not really used much\noutside of that company even though it's\n\noutside of that company even though it's\n \n\noutside of that company even though it's\nopen source so oh actually a fun fact\n\nopen source so oh actually a fun fact\n \n\nopen source so oh actually a fun fact\nhack is implemented in no camel I think\n\nhack is implemented in no camel I think\n \n\nhack is implemented in no camel I think\nthat's just like a total coincidence\n\nthat's just like a total coincidence\n \n\nthat's just like a total coincidence\nbut uh we were pretty naive early on we\n\nbut uh we were pretty naive early on we\n \n\nbut uh we were pretty naive early on we\nread this paper and we decided that it\n\nread this paper and we decided that it\n \n\nread this paper and we decided that it\nwould be really cool if we could\n\nwould be really cool if we could\n \n\nwould be really cool if we could\nreplicate the results we thought we\n\nreplicate the results we thought we\n \n\nreplicate the results we thought we\nwould just take a model off the shelf we\n\nwould just take a model off the shelf we\n \n\nwould just take a model off the shelf we\nwould show it a bunch of our code and\n\nwould show it a bunch of our code and\n \n\nwould show it a bunch of our code and\nthen we would get back a model that uh\n\nthen we would get back a model that uh\n \n\nthen we would get back a model that uh\nworked like the original model but knew\n\nworked like the original model but knew\n \n\nworked like the original model but knew\nabout our libraries and\n\nabout our libraries and\n \n\nabout our libraries and\nidioms it turns out that's just not how\n\nidioms it turns out that's just not how\n \n\nidioms it turns out that's just not how\nit works uh it's not that easy in order\n\nit works uh it's not that easy in order\n \n\nit works uh it's not that easy in order\nto get good outcomes you have to have\n\nto get good outcomes you have to have\n \n\nto get good outcomes you have to have\nthe model see a bunch of examples that\n\nthe model see a bunch of examples that\n \n\nthe model see a bunch of examples that\nare in the shape of the type of question\n\nare in the shape of the type of question\n \n\nare in the shape of the type of question\nthat you want to ask the model so we\n\nthat you want to ask the model so we\n \n\nthat you want to ask the model so we\nneeded to First create a goal a thing\n\nneeded to First create a goal a thing\n \n\nneeded to First create a goal a thing\nthat we wanted the model to be able to\n\nthat we wanted the model to be able to\n \n\nthat we wanted the model to be able to\ndo and in our in our world the goal that\n\ndo and in our in our world the goal that\n \n\ndo and in our in our world the goal that\nwe came up with was this we wanted to be\n\nwe came up with was this we wanted to be\n \n\nwe came up with was this we wanted to be\nable to generate diffs given a prompt so\n\nable to generate diffs given a prompt so\n \n\nable to generate diffs given a prompt so\nwhat that means is we wanted a user\n\nwhat that means is we wanted a user\n \n\nwhat that means is we wanted a user\ninside of an Editor to be able to write\n\ninside of an Editor to be able to write\n \n\ninside of an Editor to be able to write\na description of what they wanted to\n\na description of what they wanted to\n \n\na description of what they wanted to\nhappen and then have the model suggest a\n\nhappen and then have the model suggest a\n \n\nhappen and then have the model suggest a\npotentially multifile diff so maybe you\n\npotentially multifile diff so maybe you\n \n\npotentially multifile diff so maybe you\nwant to modify the test file an ml file\n\nwant to modify the test file an ml file\n \n\nwant to modify the test file an ml file\nand an mli which is kind of like a\n\nand an mli which is kind of like a\n \n\nand an mli which is kind of like a\nheader\n\nheader\n \n\nheader\nfile we wanted the diffs to apply\n\nfile we wanted the diffs to apply\n \n\nfile we wanted the diffs to apply\ncleanly and we wanted them to have a\n\ncleanly and we wanted them to have a\n \n\ncleanly and we wanted them to have a\ngood likelihood of typechecking after\n\ngood likelihood of typechecking after\n \n\ngood likelihood of typechecking after\nthey had been applied and we were kind\n\nthey had been applied and we were kind\n \n\nthey had been applied and we were kind\nof targeting this range of up to 100\n\nof targeting this range of up to 100\n \n\nof targeting this range of up to 100\nlines as an ideal zone of what we\n\nlines as an ideal zone of what we\n \n\nlines as an ideal zone of what we\nthought llms would actually be capable\n\nthought llms would actually be capable\n \n\nthought llms would actually be capable\nof and in order for that to work we\n\nof and in order for that to work we\n \n\nof and in order for that to work we\nneeded to collect data like I was\n\nneeded to collect data like I was\n \n\nneeded to collect data like I was\ntalking about before we needed data of\n\ntalking about before we needed data of\n \n\ntalking about before we needed data of\nthe training shape that looked just like\n\nthe training shape that looked just like\n \n\nthe training shape that looked just like\nthe test shape and this is what that\n\nthe test shape and this is what that\n \n\nthe test shape and this is what that\nshape looks like for this task you need\n\nshape looks like for this task you need\n \n\nshape looks like for this task you need\nto be able to collect a bunch of\n\nto be able to collect a bunch of\n \n\nto be able to collect a bunch of\nexamples of what context the model would\n\nexamples of what context the model would\n \n\nexamples of what context the model would\nhave had beforehand and then some prompt\n\nhave had beforehand and then some prompt\n \n\nhave had beforehand and then some prompt\nof what you want the model to do written\n\nof what you want the model to do written\n \n\nof what you want the model to do written\nhopefully in the same way that a human\n\nhopefully in the same way that a human\n \n\nhopefully in the same way that a human\nwould write it and then some diff that\n\nwould write it and then some diff that\n \n\nwould write it and then some diff that\nwould accomplish that goal so context\n\nwould accomplish that goal so context\n \n\nwould accomplish that goal so context\nprompt diff and we need a bunch of these\n\nprompt diff and we need a bunch of these\n \n\nprompt diff and we need a bunch of these\nexamples so how do we get these how do\n\nexamples so how do we get these how do\n \n\nexamples so how do we get these how do\nwe get these training\n\nwe get these training\n \n\nwe get these training\nexamples kind of the first place to look\n\nexamples kind of the first place to look\n \n\nexamples kind of the first place to look\nis features features is I mentioned a\n\nis features features is I mentioned a\n \n\nis features features is I mentioned a\ncode review system that we built\n\ncode review system that we built\n \n\ncode review system that we built\ninternally this is what it looks like\n\ninternally this is what it looks like\n \n\ninternally this is what it looks like\nit's called iron uh features are very\n\nit's called iron uh features are very\n \n\nit's called iron uh features are very\nsimilar to poll requests I think you can\n\nsimilar to poll requests I think you can\n \n\nsimilar to poll requests I think you can\njust you know swap that term in your\n\njust you know swap that term in your\n \n\njust you know swap that term in your\nhead and features at first glance have\n\nhead and features at first glance have\n \n\nhead and features at first glance have\nexactly the data they want on the\n\nexactly the data they want on the\n \n\nexactly the data they want on the\ndescription tab they have a human\n\ndescription tab they have a human\n \n\ndescription tab they have a human\nwritten description of a change and on\n\nwritten description of a change and on\n \n\nwritten description of a change and on\nthe diff tab they have the code that\n\nthe diff tab they have the code that\n \n\nthe diff tab they have the code that\naccom is the goal of the\n\naccom is the goal of the\n \n\naccom is the goal of the\ndeveloper but on closer look they're not\n\ndeveloper but on closer look they're not\n \n\ndeveloper but on closer look they're not\nexactly what you want right the way that\n\nexactly what you want right the way that\n \n\nexactly what you want right the way that\nyou write a feature description or a p\n\nyou write a feature description or a p\n \n\nyou write a feature description or a p\nrequest description is really very\n\nrequest description is really very\n \n\nrequest description is really very\ndifferent from what you might want to\n\ndifferent from what you might want to\n \n\ndifferent from what you might want to\nsay inside of an editor so you're not\n\nsay inside of an editor so you're not\n \n\nsay inside of an editor so you're not\nwriting multiple paragraphs in the\n\nwriting multiple paragraphs in the\n \n\nwriting multiple paragraphs in the\neditor you're just saying something like\n\neditor you're just saying something like\n \n\neditor you're just saying something like\nfix that error that's happening right\n\nfix that error that's happening right\n \n\nfix that error that's happening right\nnow and that's just not how we write\n\nnow and that's just not how we write\n \n\nnow and that's just not how we write\nfeature descriptions another problem\n\nfeature descriptions another problem\n \n\nfeature descriptions another problem\nwith these features or P requests is\n\nwith these features or P requests is\n \n\nwith these features or P requests is\nthat they're really large right often\n\nthat they're really large right often\n \n\nthat they're really large right often\nit's a feature is 500 lines or a\n\nit's a feature is 500 lines or a\n \n\nit's a feature is 500 lines or a\nthousand lines so in order to use it as\n\nthousand lines so in order to use it as\n \n\nthousand lines so in order to use it as\ntraining data we would need to have an\n\ntraining data we would need to have an\n \n\ntraining data we would need to have an\nautomated way to pull features apart\n\nautomated way to pull features apart\n \n\nautomated way to pull features apart\ninto individual smaller components that\n\ninto individual smaller components that\n \n\ninto individual smaller components that\nwe could train\n\nwe could train\n \n\nwe could train\non so we need smaller things than\n\non so we need smaller things than\n \n\non so we need smaller things than\nfeatures what are those well maybe\n\nfeatures what are those well maybe\n \n\nfeatures what are those well maybe\ncommits commits are smaller chunks than\n\ncommits commits are smaller chunks than\n \n\ncommits commits are smaller chunks than\nfeatures uh this is what a typical\n\nfeatures uh this is what a typical\n \n\nfeatures uh this is what a typical\ncommit log looks like at Jan street so\n\ncommit log looks like at Jan street so\n \n\ncommit log looks like at Jan street so\nthis is not like a git short log this is\n\nthis is not like a git short log this is\n \n\nthis is not like a git short log this is\nliterally just like an actual I want you\n\nliterally just like an actual I want you\n \n\nliterally just like an actual I want you\nto look at this as an actual git log and\n\nto look at this as an actual git log and\n \n\nto look at this as an actual git log and\nwhere it says summary Z that's my commit\n\nwhere it says summary Z that's my commit\n \n\nwhere it says summary Z that's my commit\nmessage we don't really use commits the\n\nmessage we don't really use commits the\n \n\nmessage we don't really use commits the\nsame way the rest of the world use\n\nsame way the rest of the world use\n \n\nsame way the rest of the world use\nsystem so we use commits mostly As\n\nsystem so we use commits mostly As\n \n\nsystem so we use commits mostly As\ncheckpoints between different parts\n\ncheckpoints between different parts\n \n\ncheckpoints between different parts\nparts of a development cycle that you\n\nparts of a development cycle that you\n \n\nparts of a development cycle that you\nmight want to revert back to commits\n\nmight want to revert back to commits\n \n\nmight want to revert back to commits\ndon't have a description and they also\n\ndon't have a description and they also\n \n\ndon't have a description and they also\nhave the same problem in that they're\n\nhave the same problem in that they're\n \n\nhave the same problem in that they're\nnot isolated changes they're they're a\n\nnot isolated changes they're they're a\n \n\nnot isolated changes they're they're a\ncollection of changes what we actually\n\ncollection of changes what we actually\n \n\ncollection of changes what we actually\nended up with was a approach called\n\nended up with was a approach called\n \n\nended up with was a approach called\nworkspace snapshotting and the way that\n\nworkspace snapshotting and the way that\n \n\nworkspace snapshotting and the way that\nthat works is we take snapshots of\n\nthat works is we take snapshots of\n \n\nthat works is we take snapshots of\ndeveloper workstations throughout the\n\ndeveloper workstations throughout the\n \n\ndeveloper workstations throughout the\nworkday so you can think like every 20\n\nworkday so you can think like every 20\n \n\nworkday so you can think like every 20\nseconds we just take a snapshot of what\n\nseconds we just take a snapshot of what\n \n\nseconds we just take a snapshot of what\nthe developer doing and as we take the\n\nthe developer doing and as we take the\n \n\nthe developer doing and as we take the\nsnapshots we also take snapshots of the\n\nsnapshots we also take snapshots of the\n \n\nsnapshots we also take snapshots of the\nbuild status so the build that's running\n\nbuild status so the build that's running\n \n\nbuild status so the build that's running\non the box we can see what the error is\n\non the box we can see what the error is\n \n\non the box we can see what the error is\nor whether the build is green and we can\n\nor whether the build is green and we can\n \n\nor whether the build is green and we can\nkind of notice these little patterns if\n\nkind of notice these little patterns if\n \n\nkind of notice these little patterns if\nyou have a green to Red to Green that\n\nyou have a green to Red to Green that\n \n\nyou have a green to Red to Green that\noften corresponds to a place where a\n\noften corresponds to a place where a\n \n\noften corresponds to a place where a\ndeveloper has made an isolated change\n\ndeveloper has made an isolated change\n \n\ndeveloper has made an isolated change\nright you start writing some code you\n\nright you start writing some code you\n \n\nright you start writing some code you\nbreak the build and then you get it back\n\nbreak the build and then you get it back\n \n\nbreak the build and then you get it back\nto green and that's how you make a\n\nto green and that's how you make a\n \n\nto green and that's how you make a\nchange maybe this one the red to Green\n\nchange maybe this one the red to Green\n \n\nchange maybe this one the red to Green\nthis is the place where the developer\n\nthis is the place where the developer\n \n\nthis is the place where the developer\nencountered an error whether that's a\n\nencountered an error whether that's a\n \n\nencountered an error whether that's a\ntype error or a compilation error and\n\ntype error or a compilation error and\n \n\ntype error or a compilation error and\nthen they fixed it so if we capture the\n\nthen they fixed it so if we capture the\n \n\nthen they fixed it so if we capture the\nbuild error at the Red State and then\n\nbuild error at the Red State and then\n \n\nbuild error at the Red State and then\nthe diff from red to Green we can use\n\nthe diff from red to Green we can use\n \n\nthe diff from red to Green we can use\nthat as training data to help the model\n\nthat as training data to help the model\n \n\nthat as training data to help the model\nbe able to recover from\n\nbe able to recover from\n \n\nbe able to recover from\nmistakes the next thing we need is a\n\nmistakes the next thing we need is a\n \n\nmistakes the next thing we need is a\ndescription and the way that we did that\n\ndescription and the way that we did that\n \n\ndescription and the way that we did that\nwe just used the large language model so\n\nwe just used the large language model so\n \n\nwe just used the large language model so\nwe had a large language model write a\n\nwe had a large language model write a\n \n\nwe had a large language model write a\nreally detailed description of a change\n\nreally detailed description of a change\n \n\nreally detailed description of a change\nin in as much words as it possibly could\n\nin in as much words as it possibly could\n \n\nin in as much words as it possibly could\nand then we just kept filtering it down\n\nand then we just kept filtering it down\n \n\nand then we just kept filtering it down\nuntil it was something that was around\n\nuntil it was something that was around\n \n\nuntil it was something that was around\nthe right level of what a human would\n\nthe right level of what a human would\n \n\nthe right level of what a human would\nwrite so now we have this training data\n\nwrite so now we have this training data\n \n\nwrite so now we have this training data\nand training data is kind of only half\n\nand training data is kind of only half\n \n\nand training data is kind of only half\nthe picture of training a model so you\n\nthe picture of training a model so you\n \n\nthe picture of training a model so you\nyou have the the supervised training\n\nyou have the the supervised training\n \n\nyou have the the supervised training\ndata and then you need to do the second\n\ndata and then you need to do the second\n \n\ndata and then you need to do the second\npart which is the reinforcement learning\n\npart which is the reinforcement learning\n \n\npart which is the reinforcement learning\nthis is really where models get a lot of\n\nthis is really where models get a lot of\n \n\nthis is really where models get a lot of\ntheir power right we we align the\n\ntheir power right we we align the\n \n\ntheir power right we we align the\nmodel's ability to what humans think is\n\nmodel's ability to what humans think is\n \n\nmodel's ability to what humans think is\nactually good code so what is good code\n\nactually good code so what is good code\n \n\nactually good code so what is good code\nI guess on the surface good code is I\n\nI guess on the surface good code is I\n \n\nI guess on the surface good code is I\nmean it's it's code it has the parse is\n\nmean it's it's code it has the parse is\n \n\nmean it's it's code it has the parse is\ncode meaning if a piece of code doesn't\n\ncode meaning if a piece of code doesn't\n \n\ncode meaning if a piece of code doesn't\ngo through the O camel parser and come\n\ngo through the O camel parser and come\n \n\ngo through the O camel parser and come\nout with a green status that is that is\n\nout with a green status that is that is\n \n\nout with a green status that is that is\nnot good code I would say by most\n\nnot good code I would say by most\n \n\nnot good code I would say by most\ndefinitions\n\ndefinitions\n \n\ndefinitions\nuh good code in oaml because it's\n\nuh good code in oaml because it's\n \n\nuh good code in oaml because it's\nstatically typed is code that type\n\nstatically typed is code that type\n \n\nstatically typed is code that type\nchecks so we want to have good code be\n\nchecks so we want to have good code be\n \n\nchecks so we want to have good code be\ncode that when it is applied on top of a\n\ncode that when it is applied on top of a\n \n\ncode that when it is applied on top of a\nbase revision can go through the type\n\nbase revision can go through the type\n \n\nbase revision can go through the type\nChecker and the type Checker agrees that\n\nChecker and the type Checker agrees that\n \n\nChecker and the type Checker agrees that\nthe code is\n\nthe code is\n \n\nthe code is\nvalid and of course the the gold\n\nvalid and of course the the gold\n \n\nvalid and of course the the gold\nstandard is that good code is code that\n\nstandard is that good code is code that\n \n\nstandard is that good code is code that\ncompiles and passes tests so ideally in\n\ncompiles and passes tests so ideally in\n \n\ncompiles and passes tests so ideally in\nduring the reinforcement learning phase\n\nduring the reinforcement learning phase\n \n\nduring the reinforcement learning phase\nof a model you could give the model a\n\nof a model you could give the model a\n \n\nof a model you could give the model a\nbunch of tasks that are like verifiable\n\nbunch of tasks that are like verifiable\n \n\nbunch of tasks that are like verifiable\nwe have the model performs some some\n\nwe have the model performs some some\n \n\nwe have the model performs some some\nedit and then we check whether or not it\n\nedit and then we check whether or not it\n \n\nedit and then we check whether or not it\nactually passes the test when applied to\n\nactually passes the test when applied to\n \n\nactually passes the test when applied to\nthe\n\nthe\n \n\nthe\ncode so we did that uh we've done this\n\ncode so we did that uh we've done this\n \n\ncode so we did that uh we've done this\nas part of our our training cycle and we\n\nas part of our our training cycle and we\n \n\nas part of our our training cycle and we\nbuilt this thing that is called uh CES\n\nbuilt this thing that is called uh CES\n \n\nbuilt this thing that is called uh CES\nit's the code evaluation service you can\n\nit's the code evaluation service you can\n \n\nit's the code evaluation service you can\nthink of it kind of like a build service\n\nthink of it kind of like a build service\n \n\nthink of it kind of like a build service\nexcept with a slight modification to\n\nexcept with a slight modification to\n \n\nexcept with a slight modification to\nmake it much faster and that's that\n\nmake it much faster and that's that\n \n\nmake it much faster and that's that\nfirst we pre-warm a build it sits at a a\n\nfirst we pre-warm a build it sits at a a\n \n\nfirst we pre-warm a build it sits at a a\nrevision and is green and then we have\n\nrevision and is green and then we have\n \n\nrevision and is green and then we have\nthese workers that all day just take\n\nthese workers that all day just take\n \n\nthese workers that all day just take\ndiffs from the model they apply them and\n\ndiffs from the model they apply them and\n \n\ndiffs from the model they apply them and\nthen we determine whether the build\n\nthen we determine whether the build\n \n\nthen we determine whether the build\nstatus turns red or green and then we\n\nstatus turns red or green and then we\n \n\nstatus turns red or green and then we\nreport that error or or success back up\n\nreport that error or or success back up\n \n\nreport that error or or success back up\nto the build function and through\n\nto the build function and through\n \n\nto the build function and through\ncontinued use of this service over the\n\ncontinued use of this service over the\n \n\ncontinued use of this service over the\ncourse of like months we're able to\n\ncourse of like months we're able to\n \n\ncourse of like months we're able to\nbetter align the model to write code\n\nbetter align the model to write code\n \n\nbetter align the model to write code\nthat actually does compile and past\n\n \n\ntests it turns out this exact same setup\n\ntests it turns out this exact same setup\n \n\ntests it turns out this exact same setup\nis the one that you would want for\n\nis the one that you would want for\n \n\nis the one that you would want for\nevaluation so if you just hold out some\n\nevaluation so if you just hold out some\n \n\nevaluation so if you just hold out some\nof the RL data you can also use it to\n\nof the RL data you can also use it to\n \n\nof the RL data you can also use it to\nevaluate model's ability to write code\n\nevaluate model's ability to write code\n \n\nevaluate model's ability to write code\nkind of looks like this you give the\n\nkind of looks like this you give the\n \n\nkind of looks like this you give the\nmodel a problem you let it write some\n\nmodel a problem you let it write some\n \n\nmodel a problem you let it write some\ncode and then you evaluate whether or\n\ncode and then you evaluate whether or\n \n\ncode and then you evaluate whether or\nnot the code that it writes actually\n\n \n\nworks and training is hard and it can\n\nworks and training is hard and it can\n \n\nworks and training is hard and it can\nhave kind of uh catastrophic but\n\nhave kind of uh catastrophic but\n \n\nhave kind of uh catastrophic but\nhilarious results so at one point we\n\nhilarious results so at one point we\n \n\nhilarious results so at one point we\nwere training a code review model and\n\nwere training a code review model and\n \n\nwere training a code review model and\nthis is a totally separate model but the\n\nthis is a totally separate model but the\n \n\nthis is a totally separate model but the\nidea was we want to be able to give some\n\nidea was we want to be able to give some\n \n\nidea was we want to be able to give some\ncode to this model and have it do a\n\ncode to this model and have it do a\n \n\ncode to this model and have it do a\nfirst passive code review just like a\n\nfirst passive code review just like a\n \n\nfirst passive code review just like a\nhuman would do to try to save some of\n\nhuman would do to try to save some of\n \n\nhuman would do to try to save some of\nthe toil of of code review we train this\n\nthe toil of of code review we train this\n \n\nthe toil of of code review we train this\nmodel we put a bunch of dat dat into it\n\nmodel we put a bunch of dat dat into it\n \n\nmodel we put a bunch of dat dat into it\nwe worked on it for months we're real\n\nwe worked on it for months we're real\n \n\nwe worked on it for months we're real\nexcited and we put our first code in for\n\nexcited and we put our first code in for\n \n\nexcited and we put our first code in for\nuh for code review through the automated\n\nuh for code review through the automated\n \n\nuh for code review through the automated\nagent it spun for a bit and it came back\n\nagent it spun for a bit and it came back\n \n\nagent it spun for a bit and it came back\nwith something along the lines of um\n\nwith something along the lines of um\n \n\nwith something along the lines of um\nI'll do it\n\nI'll do it\n \n\nI'll do it\ntomorrow and like of course it did that\n\ntomorrow and like of course it did that\n \n\ntomorrow and like of course it did that\nbecause it's trained on a bunch of human\n\nbecause it's trained on a bunch of human\n \n\nbecause it's trained on a bunch of human\nexamples and humans write things like\n\nexamples and humans write things like\n \n\nexamples and humans write things like\nI'll do things or I'll do this tomorrow\n\nI'll do things or I'll do this tomorrow\n \n\nI'll do things or I'll do this tomorrow\nuh so it's it's you know not very\n\nuh so it's it's you know not very\n \n\nuh so it's it's you know not very\nsurprising so having evaluations that\n\nsurprising so having evaluations that\n \n\nsurprising so having evaluations that\nare meaningful is kind of a Cornerstone\n\nare meaningful is kind of a Cornerstone\n \n\nare meaningful is kind of a Cornerstone\nof making sure that models don't go off\n\nof making sure that models don't go off\n \n\nof making sure that models don't go off\nthe rails like this and you don't waste\n\nthe rails like this and you don't waste\n \n\nthe rails like this and you don't waste\na bunch of your time and\n\na bunch of your time and\n \n\na bunch of your time and\nmoney in the end though the real test of\n\nmoney in the end though the real test of\n \n\nmoney in the end though the real test of\nmodels is whether or not they work for\n\nmodels is whether or not they work for\n \n\nmodels is whether or not they work for\nhumans so I'm going to talk a little bit\n\nhumans so I'm going to talk a little bit\n \n\nhumans so I'm going to talk a little bit\nabout the editor Integrations that we've\n\nabout the editor Integrations that we've\n \n\nabout the editor Integrations that we've\nbuilt to expose these models to\n\nbuilt to expose these models to\n \n\nbuilt to expose these models to\ndevelopers at Jan\n\ndevelopers at Jan\n \n\ndevelopers at Jan\nStreet kind of when we were starting\n\nStreet kind of when we were starting\n \n\nStreet kind of when we were starting\nbuilding these Integrations we had three\n\nbuilding these Integrations we had three\n \n\nbuilding these Integrations we had three\nideas in mind the first idea was wow we\n\nideas in mind the first idea was wow we\n \n\nideas in mind the first idea was wow we\nsupport three editors we have neovim vs\n\nsupport three editors we have neovim vs\n \n\nsupport three editors we have neovim vs\ncode and emac and we really don't want\n\ncode and emac and we really don't want\n \n\ncode and emac and we really don't want\nto write the same thing three times so\n\nto write the same thing three times so\n \n\nto write the same thing three times so\nideally we don't want to write all the\n\nideally we don't want to write all the\n \n\nideally we don't want to write all the\nsame context building strategies and all\n\nsame context building strategies and all\n \n\nsame context building strategies and all\nof the same prompting strategies we want\n\nof the same prompting strategies we want\n \n\nof the same prompting strategies we want\nto just write at once the second is that\n\nto just write at once the second is that\n \n\nto just write at once the second is that\nwe wanted to maintain flexibility so we\n\nwe wanted to maintain flexibility so we\n \n\nwe wanted to maintain flexibility so we\nhad a model that we were using at the\n\nhad a model that we were using at the\n \n\nhad a model that we were using at the\ntime uh that was not a fine tuned model\n\ntime uh that was not a fine tuned model\n \n\ntime uh that was not a fine tuned model\nwe were pretty convinced that a fine\n\nwe were pretty convinced that a fine\n \n\nwe were pretty convinced that a fine\ntuned model was in our future we wanted\n\ntuned model was in our future we wanted\n \n\ntuned model was in our future we wanted\nthe ability to do things like swap the\n\nthe ability to do things like swap the\n \n\nthe ability to do things like swap the\nmodel or swap the prompting strategy out\n\nmodel or swap the prompting strategy out\n \n\nmodel or swap the prompting strategy out\nand lastly we wanted to be able to\n\nand lastly we wanted to be able to\n \n\nand lastly we wanted to be able to\ncollect metrics so in a developer uh in\n\ncollect metrics so in a developer uh in\n \n\ncollect metrics so in a developer uh in\ntheir in their editor developers care\n\ntheir in their editor developers care\n \n\ntheir in their editor developers care\nabout latency and they care about\n\nabout latency and they care about\n \n\nabout latency and they care about\nwhether or not the diffs actually apply\n\nwhether or not the diffs actually apply\n \n\nwhether or not the diffs actually apply\nso we wanted to get kind of on the\n\nso we wanted to get kind of on the\n \n\nso we wanted to get kind of on the\nground real experience of whether or not\n\nground real experience of whether or not\n \n\nground real experience of whether or not\nthe diffs really were meaningful for\n\nthe diffs really were meaningful for\n \n\nthe diffs really were meaningful for\npeople this is the simplified version of\n\npeople this is the simplified version of\n \n\npeople this is the simplified version of\nthe architecture that we settled on for\n\nthe architecture that we settled on for\n \n\nthe architecture that we settled on for\nthis service the AI development\n\nthis service the AI development\n \n\nthis service the AI development\nenvironment essentially you have llms on\n\nenvironment essentially you have llms on\n \n\nenvironment essentially you have llms on\none side and then Aid handles all of the\n\none side and then Aid handles all of the\n \n\none side and then Aid handles all of the\nuh ability to construct prompts and to\n\nuh ability to construct prompts and to\n \n\nuh ability to construct prompts and to\nconstruct context and to see the build\n\nconstruct context and to see the build\n \n\nconstruct context and to see the build\nstatus and then we are able to just\n\nstatus and then we are able to just\n \n\nstatus and then we are able to just\nwrite these really thin layers on top of\n\nwrite these really thin layers on top of\n \n\nwrite these really thin layers on top of\nAid uh for each of the individual\n\nAid uh for each of the individual\n \n\nAid uh for each of the individual\neditors and what's really neat about\n\neditors and what's really neat about\n \n\neditors and what's really neat about\nthis is that Aid sits as a sidecar\n\nthis is that Aid sits as a sidecar\n \n\nthis is that Aid sits as a sidecar\napplication on the Developers machine\n\napplication on the Developers machine\n \n\napplication on the Developers machine\nwhich means that we when we want to make\n\nwhich means that we when we want to make\n \n\nwhich means that we when we want to make\nchanges to Aid we don't have to make\n\nchanges to Aid we don't have to make\n \n\nchanges to Aid we don't have to make\nchanges to the individual editors and\n\nchanges to the individual editors and\n \n\nchanges to the individual editors and\nhope that people restart their editors\n\nhope that people restart their editors\n \n\nhope that people restart their editors\nwe can just restart the Aid Service on\n\nwe can just restart the Aid Service on\n \n\nwe can just restart the Aid Service on\nall of the boxes so we restart Aid and\n\nall of the boxes so we restart Aid and\n \n\nall of the boxes so we restart Aid and\nthen everyone gets the most recent\n\n \n\ncopy uh this is an example of Aid\n\ncopy uh this is an example of Aid\n \n\ncopy uh this is an example of Aid\nworking inside of vs code so this is the\n\nworking inside of vs code so this is the\n \n\nworking inside of vs code so this is the\nsidebar in vs code very similar to\n\nsidebar in vs code very similar to\n \n\nsidebar in vs code very similar to\nsomething like co-pilot except this\n\nsomething like co-pilot except this\n \n\nsomething like co-pilot except this\nthing allows you to uh ask for it and\n\nthing allows you to uh ask for it and\n \n\nthing allows you to uh ask for it and\nget back multifile diffs uh and you can\n\nget back multifile diffs uh and you can\n \n\nget back multifile diffs uh and you can\nsee it kind of looks like what you'd\n\nsee it kind of looks like what you'd\n \n\nsee it kind of looks like what you'd\nexpect in VSS code it's it's you know a\n\nexpect in VSS code it's it's you know a\n \n\nexpect in VSS code it's it's you know a\nvisual interface that lays things out\n\nvisual interface that lays things out\n \n\nvisual interface that lays things out\nreally\n\nreally\n \n\nreally\nnicely this is what we built in emac\n\nnicely this is what we built in emac\n \n\nnicely this is what we built in emac\nthough so in emac developers are used to\n\nthough so in emac developers are used to\n \n\nthough so in emac developers are used to\nworking in text buffers they move around\n\nworking in text buffers they move around\n \n\nworking in text buffers they move around\nfiles they want to be able to copy\n\nfiles they want to be able to copy\n \n\nfiles they want to be able to copy\nthings the normal way that they copy\n\nthings the normal way that they copy\n \n\nthings the normal way that they copy\nthings so we actually built the eight\n\nthings so we actually built the eight\n \n\nthings so we actually built the eight\nexperience in emac into a markdown\n\nexperience in emac into a markdown\n \n\nexperience in emac into a markdown\nbuffer so users can move around inside\n\nbuffer so users can move around inside\n \n\nbuffer so users can move around inside\nthis markdown buffer they can ask\n\nthis markdown buffer they can ask\n \n\nthis markdown buffer they can ask\nquestions and then there are key binds\n\nquestions and then there are key binds\n \n\nquestions and then there are key binds\nthat essentially append extra content to\n\nthat essentially append extra content to\n \n\nthat essentially append extra content to\nthe bottom of the markdown\n\n \n\nbuffer AIDS architecture lets us plug\n\nbuffer AIDS architecture lets us plug\n \n\nbuffer AIDS architecture lets us plug\nvarious pieces in and out like I\n\nvarious pieces in and out like I\n \n\nvarious pieces in and out like I\nmentioned uh so we can swap in new\n\nmentioned uh so we can swap in new\n \n\nmentioned uh so we can swap in new\nmodels we can uh make changes to the\n\nmodels we can uh make changes to the\n \n\nmodels we can uh make changes to the\ncontext building we can add support for\n\ncontext building we can add support for\n \n\ncontext building we can add support for\nnew editors which I think probably\n\nnew editors which I think probably\n \n\nnew editors which I think probably\nsounds far-fetched but this is something\n\nsounds far-fetched but this is something\n \n\nsounds far-fetched but this is something\nwe're actually just doing right\n\nwe're actually just doing right\n \n\nwe're actually just doing right\nnow uh and we can even add domain\n\nnow uh and we can even add domain\n \n\nnow uh and we can even add domain\nspecific tools so different areas of the\n\nspecific tools so different areas of the\n \n\nspecific tools so different areas of the\ncompany can supply tools that are\n\ncompany can supply tools that are\n \n\ncompany can supply tools that are\navailable inside of the editors and they\n\navailable inside of the editors and they\n \n\navailable inside of the editors and they\nkind of end up in all the editors\n\nkind of end up in all the editors\n \n\nkind of end up in all the editors\nwithout having to write individual\n\nwithout having to write individual\n \n\nwithout having to write individual\nIntegrations eight also allows us to AB\n\nIntegrations eight also allows us to AB\n \n\nIntegrations eight also allows us to AB\ntest different approaches so we can do\n\ntest different approaches so we can do\n \n\ntest different approaches so we can do\nsomething like send 50% of the company\n\nsomething like send 50% of the company\n \n\nsomething like send 50% of the company\nto one model and 50% to another and then\n\nto one model and 50% to another and then\n \n\nto one model and 50% to another and then\ndetermine which one gets the higher\n\ndetermine which one gets the higher\n \n\ndetermine which one gets the higher\nacceptance rate a is kind of a an\n\nacceptance rate a is kind of a an\n \n\nacceptance rate a is kind of a an\ninvestment that pays off over time every\n\ninvestment that pays off over time every\n \n\ninvestment that pays off over time every\ntime something changes in large language\n\ntime something changes in large language\n \n\ntime something changes in large language\nmodels we're able to change it in one\n\nmodels we're able to change it in one\n \n\nmodels we're able to change it in one\nplace Downstream of the editors and then\n\nplace Downstream of the editors and then\n \n\nplace Downstream of the editors and then\nhave it available\n\nhave it available\n \n\nhave it available\neverywhere and things change like really\n\neverywhere and things change like really\n \n\neverywhere and things change like really\noften and we need to be ready uh when\n\noften and we need to be ready uh when\n \n\noften and we need to be ready uh when\nthings change what I what I've had time\n\nthings change what I what I've had time\n \n\nthings change what I what I've had time\nto show you today is only a small\n\nto show you today is only a small\n \n\nto show you today is only a small\nportion of what my team is doing we've\n\nportion of what my team is doing we've\n \n\nportion of what my team is doing we've\ngot a lot of other things going on so\n\ngot a lot of other things going on so\n \n\ngot a lot of other things going on so\nwe're finding new ways to apply rag\n\nwe're finding new ways to apply rag\n \n\nwe're finding new ways to apply rag\ninside of the editors we're applying\n\ninside of the editors we're applying\n \n\ninside of the editors we're applying\nsimilar approaches to what you've seen\n\nsimilar approaches to what you've seen\n \n\nsimilar approaches to what you've seen\nhere to large scale uh multi-agent\n\nhere to large scale uh multi-agent\n \n\nhere to large scale uh multi-agent\nworkflows we are working with reasoning\n\nworkflows we are working with reasoning\n \n\nworkflows we are working with reasoning\nmodels more and more but the approach is\n\nmodels more and more but the approach is\n \n\nmodels more and more but the approach is\nthe same through all of these we keep\n\nthe same through all of these we keep\n \n\nthe same through all of these we keep\nthings pluggable we lay a strong\n\nthings pluggable we lay a strong\n \n\nthings pluggable we lay a strong\nFoundation to build on top of and we\n\nFoundation to build on top of and we\n \n\nFoundation to build on top of and we\nbuild the ways for the rest of the\n\nbuild the ways for the rest of the\n \n\nbuild the ways for the rest of the\ncompany to add to our experience by\n\ncompany to add to our experience by\n \n\ncompany to add to our experience by\nadding more domain specific tooling on\n\nadding more domain specific tooling on\n \n\nadding more domain specific tooling on\ntop of it\n\ntop of it\n \n\ntop of it\nuh if you think what I've said is\n\nuh if you think what I've said is\n \n\nuh if you think what I've said is\ninteresting and you want to talk more\n\ninteresting and you want to talk more\n \n\ninteresting and you want to talk more\nabout this I would love to hear from you\n\nabout this I would love to hear from you\n \n\nabout this I would love to hear from you\nyou can just find me outside and thank\n\nyou can just find me outside and thank\n \n\nyou can just find me outside and thank\nyou for your time\n\nyou for your time\n \n\nyou for your time\n[Music]\n\n[Music]\n \n\n[Music]\n[Applause]\n\n \n\n[Music]" } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\n[Music]\n\n \n\nThank you everyone for coming here. I\n\nThank you everyone for coming here. I\n \n\nThank you everyone for coming here. I\nhope you're uh excited to figure out\n\nhope you're uh excited to figure out\n \n\nhope you're uh excited to figure out\nwith me what enterprise ready MCP means. with me what e..." }, { "role": "assistant", "content": "[Music]\n\n \n\nThank you everyone for coming here. I\n\nThank you everyone for coming here. I\n \n\nThank you everyone for coming here. I\nhope you're uh excited to figure out\n\nhope you're uh excited to figure out\n \n\nhope you're uh excited to figure out\nwith me what enterprise ready MCP means. with me what enterprise ready MCP means. with me what enterprise ready MCP means. Uh I'm not sure a month ago people even\n\nUh I'm not sure a month ago people even\n \n\nUh I'm not sure a month ago people even\nfully understood what MCP meant. Um so\n\nfully understood what MCP meant. Um so\n \n\nfully understood what MCP meant. Um so\nwe're all on this journey together. Um,\n\nwe're all on this journey together. Um,\n \n\nwe're all on this journey together. Um,\nbut hopefully it should be fun. but hopefully it should be fun. but hopefully it should be fun. So, I was chatting to a good friend and\n\nSo, I was chatting to a good friend and\n \n\nSo, I was chatting to a good friend and\ncolleague who is is giving a talk\n\ncolleague who is is giving a talk\n \n\ncolleague who is is giving a talk\ntomorrow on building MCP servers and\n\ntomorrow on building MCP servers and\n \n\ntomorrow on building MCP servers and\nwhat it looks like to build that out. what it looks like to build that out. what it looks like to build that out. And his question was, how do the tools\n\nAnd his question was, how do the tools\n \n\nAnd his question was, how do the tools\nof today map to the AI systems of\n\nof today map to the AI systems of\n \n\nof today map to the AI systems of\ntomorrow? We already have some kind of\n\ntomorrow? We already have some kind of\n \n\ntomorrow? We already have some kind of\nnot concept of enterprise ready tooling\n\nnot concept of enterprise ready tooling\n \n\nnot concept of enterprise ready tooling\nand building tools for the enterprise\n\nand building tools for the enterprise\n \n\nand building tools for the enterprise\nand production. What does it actually\n\nand production. What does it actually\n \n\nand production. What does it actually\nmean for MCP and something I kind of,\n\nmean for MCP and something I kind of,\n \n\nmean for MCP and something I kind of,\nyou know, want to talk through today,\n\nyou know, want to talk through today,\n \n\nyou know, want to talk through today,\nchat with you guys. So for a long time\n\nchat with you guys. So for a long time\n \n\nchat with you guys. So for a long time\nfrom you know several years now we've\n\nfrom you know several years now we've\n \n\nfrom you know several years now we've\nhad the model of you are a user you talk\n\nhad the model of you are a user you talk\n \n\nhad the model of you are a user you talk\nto an AI talk to a chatbot it does stuff\n\nto an AI talk to a chatbot it does stuff\n \n\nto an AI talk to a chatbot it does stuff\nthe AIS had tool calling now they have\n\nthe AIS had tool calling now they have\n \n\nthe AIS had tool calling now they have\nMCP this is the MCP logo which is just a\n\nMCP this is the MCP logo which is just a\n \n\nMCP this is the MCP logo which is just a\nway of interfacing between the AI and an\n\nway of interfacing between the AI and an\n \n\nway of interfacing between the AI and an\nexternal resource of some kind to do\n\nexternal resource of some kind to do\n \n\nexternal resource of some kind to do\nsomething right can just be a database\n\nsomething right can just be a database\n \n\nsomething right can just be a database\naccess it can be performing some\n\naccess it can be performing some\n \n\naccess it can be performing some\ncomplicated computation can be pulling\n\ncomplicated computation can be pulling\n \n\ncomplicated computation can be pulling\nin a prompt there's lots of stuff in the\n\nin a prompt there's lots of stuff in the\n \n\nin a prompt there's lots of stuff in the\nMCP spec that frankly you should go and\n\nMCP spec that frankly you should go and\n \n\nMCP spec that frankly you should go and\nread there's more there than most people\n\nread there's more there than most people\n \n\nread there's more there than most people\nrealize\n\nrealize\n \n\nrealize\nWe're also seeing workloads like this\n\nWe're also seeing workloads like this\n \n\nWe're also seeing workloads like this\nwhere an IT admin spins up an\n\nwhere an IT admin spins up an\n \n\nwhere an IT admin spins up an\nasynchronous workflow which is we're\n\nasynchronous workflow which is we're\n \n\nasynchronous workflow which is we're\ngoing to call an AI agent which\n\ngoing to call an AI agent which\n \n\ngoing to call an AI agent which\nautomates some process that is you know\n\nautomates some process that is you know\n \n\nautomates some process that is you know\nkind of headless and is going and doing\n\nkind of headless and is going and doing\n \n\nkind of headless and is going and doing\na thing we're trying to manage that that\n\na thing we're trying to manage that that\n \n\na thing we're trying to manage that that\nis then going to use MCP to access\n\nis then going to use MCP to access\n \n\nis then going to use MCP to access\nexternal tools that might be secured\n\nexternal tools that might be secured\n \n\nexternal tools that might be secured\nmight be internal enterprise tools which\n\nmight be internal enterprise tools which\n \n\nmight be internal enterprise tools which\nintroduce some fun problems. I also for\n\nintroduce some fun problems. I also for\n \n\nintroduce some fun problems. I also for\nthe sake of this kind of diagram of what\n\nthe sake of this kind of diagram of what\n \n\nthe sake of this kind of diagram of what\nis the future going to look like want to\n\nis the future going to look like want to\n \n\nis the future going to look like want to\nintroduce this line here just to throw\n\nintroduce this line here just to throw\n \n\nintroduce this line here just to throw\nsome some spanners in the works of users\n\nsome some spanners in the works of users\n \n\nsome some spanners in the works of users\nin a company using their chatbot which\n\nin a company using their chatbot which\n \n\nin a company using their chatbot which\nmakes a call to an MCP server which then\n\nmakes a call to an MCP server which then\n \n\nmakes a call to an MCP server which then\ngoes and queries an AI workload an AI\n\ngoes and queries an AI workload an AI\n \n\ngoes and queries an AI workload an AI\nagent that's existing in the cloud\n\nagent that's existing in the cloud\n \n\nagent that's existing in the cloud\nperforming some operation and so with\n\nperforming some operation and so with\n \n\nperforming some operation and so with\nthis kind of diagram of what could\n\nthis kind of diagram of what could\n \n\nthis kind of diagram of what could\nhappen I want us to figure out what we\n\nhappen I want us to figure out what we\n \n\nhappen I want us to figure out what we\nneed to fix to make this really useful\n\nneed to fix to make this really useful\n \n\nneed to fix to make this really useful\nand production\n\nand production\n \n\nand production\nOkay, why am I why am I talking to you? Okay, why am I why am I talking to you? Okay, why am I why am I talking to you? I work for WorkOS, which is an\n\nI work for WorkOS, which is an\n \n\nI work for WorkOS, which is an\nenterprise security vendor, um, which\n\nenterprise security vendor, um, which\n \n\nenterprise security vendor, um, which\nI'll talk about in a second. I'm also a\n\nI'll talk about in a second. I'm also a\n \n\nI'll talk about in a second. I'm also a\nresearch fellow at Stanford working on\n\nresearch fellow at Stanford working on\n \n\nresearch fellow at Stanford working on\nsafety for AI agents and just finished a\n\nsafety for AI agents and just finished a\n \n\nsafety for AI agents and just finished a\nPhD where I got to wear this really\n\nPhD where I got to wear this really\n \n\nPhD where I got to wear this really\nfunny, silly hat, um, just a couple\n\nfunny, silly hat, um, just a couple\n \n\nfunny, silly hat, um, just a couple\nmonths ago. Um, OpenAI uh work OS is a\n\nmonths ago. Um, OpenAI uh work OS is a\n \n\nmonths ago." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nUm, OpenAI uh work OS is a\nenterprise security vendor that sells to\n\nenterprise security vendor that sells to\n \n\nenterprise security vendor that sells to\nall of the AI labs because the AI labs\n\nall of the AI labs because the AI labs\n \n\nall of the AI labs because the AI labs\ndo not want to do the rea..." }, { "role": "assistant", "content": "Um, OpenAI uh work OS is a\nenterprise security vendor that sells to\n\nenterprise security vendor that sells to\n \n\nenterprise security vendor that sells to\nall of the AI labs because the AI labs\n\nall of the AI labs because the AI labs\n \n\nall of the AI labs because the AI labs\ndo not want to do the really painful\n\ndo not want to do the really painful\n \n\ndo not want to do the really painful\nannoying work to scale the security and\n\nannoying work to scale the security and\n \n\nannoying work to scale the security and\north operations into the enterprise. And\n\north operations into the enterprise. And\n \n\north operations into the enterprise. And\nso we've been thinking a lot about what\n\nso we've been thinking a lot about what\n \n\nso we've been thinking a lot about what\nit means to provide the glue that means\n\nit means to provide the glue that means\n \n\nit means to provide the glue that means\npeople can build agents and then just\n\npeople can build agents and then just\n \n\npeople can build agents and then just\nseamlessly scale it and sell it to\n\nseamlessly scale it and sell it to\n \n\nseamlessly scale it and sell it to\neveryone. And so I want to just tell you\n\neveryone. And so I want to just tell you\n \n\neveryone. And so I want to just tell you\nall the answers today whether or not you\n\nall the answers today whether or not you\n \n\nall the answers today whether or not you\nwant to use us. So let's go through the\n\nwant to use us. So let's go through the\n \n\nwant to use us. So let's go through the\nagent journey. We're all here trying to\n\nagent journey. We're all here trying to\n \n\nagent journey. We're all here trying to\nbuild agents. Um, let's tell a tell a\n\nbuild agents. Um, let's tell a tell a\n \n\nbuild agents. Um, let's tell a tell a\nstory. So, I found this great API um\n\nstory. So, I found this great API um\n \n\nstory. So, I found this great API um\nthat you can query and it will feed a\n\nthat you can query and it will feed a\n \n\nthat you can query and it will feed a\nbunch of goats that you can then see on\n\nbunch of goats that you can then see on\n \n\nbunch of goats that you can then see on\na live stream. Um, this is a real API\n\na live stream. Um, this is a real API\n \n\na live stream. Um, this is a real API\nyou can go and use. And so, naturally,\n\nyou can go and use. And so, naturally,\n \n\nyou can go and use. And so, naturally,\nyou do what anyone would do and you're\n\nyou do what anyone would do and you're\n \n\nyou do what anyone would do and you're\ngoing to build an emotional support bot\n\ngoing to build an emotional support bot\n \n\ngoing to build an emotional support bot\nfor employees at a company to build a\n\nfor employees at a company to build a\n \n\nfor employees at a company to build a\nSAS. Um, and you give it tool use to\n\nSAS. Um, and you give it tool use to\n \n\nSAS. Um, and you give it tool use to\nquery this. Now, this is fine because\n\nquery this. Now, this is fine because\n \n\nquery this. Now, this is fine because\nit's a onetime query, but frankly, tool\n\nit's a onetime query, but frankly, tool\n \n\nit's a onetime query, but frankly, tool\nuse kind of sucks. We've had it for\n\nuse kind of sucks. We've had it for\n \n\nuse kind of sucks. We've had it for\nyears now. Not as many people are using\n\nyears now. Not as many people are using\n \n\nyears now. Not as many people are using\nit. It's really painful. A lot of things\n\nit. It's really painful. A lot of things\n \n\nit. It's really painful. A lot of things\ngo wrong, which is why the model context\n\ngo wrong, which is why the model context\n \n\ngo wrong, which is why the model context\nprotocol got invented. So why would you\n\nprotocol got invented. So why would you\n \n\nprotocol got invented. So why would you\nmove to to MCP? move to to MCP? move to to MCP? One, there is this really robust\n\nOne, there is this really robust\n \n\nOne, there is this really robust\necosystem of tools and providers,\n\necosystem of tools and providers,\n \n\necosystem of tools and providers,\nsecurity tooling that lets you interface\n\nsecurity tooling that lets you interface\n \n\nsecurity tooling that lets you interface\nbetween the model and the resource to\n\nbetween the model and the resource to\n \n\nbetween the model and the resource to\nmake things safe and reliable. It's also\n\nmake things safe and reliable. It's also\n \n\nmake things safe and reliable. It's also\nreally good at providing standardization\n\nreally good at providing standardization\n \n\nreally good at providing standardization\nto the models. the models are getting\n\nto the models. the models are getting\n \n\nto the models. the models are getting\nreally good at learning how to use this\n\nreally good at learning how to use this\n \n\nreally good at learning how to use this\neither through RL or just kind of you\n\neither through RL or just kind of you\n \n\neither through RL or just kind of you\nknow good eval usage. It also runs a\n\nknow good eval usage. It also runs a\n \n\nknow good eval usage. It also runs a\nstateful connection which means you can\n\nstateful connection which means you can\n \n\nstateful connection which means you can\ndo you know better security better\n\ndo you know better security better\n \n\ndo you know better security better\nmanagement better context management in\n\nmanagement better context management in\n \n\nmanagement better context management in\nthe way you're passing in to an AM model\n\nthe way you're passing in to an AM model\n \n\nthe way you're passing in to an AM model\nand maybe you're doing it just because\n\nand maybe you're doing it just because\n \n\nand maybe you're doing it just because\neveryone's doing it and why not have\n\neveryone's doing it and why not have\n \n\neveryone's doing it and why not have\nsome fun right it is genuinely quite fun\n\nsome fun right it is genuinely quite fun\n \n\nsome fun right it is genuinely quite fun\nto build an MCP server you can make cool\n\nto build an MCP server you can make cool\n \n\nto build an MCP server you can make cool\nthings out of it so what do we do we\n\nthings out of it so what do we do we\n \n\nthings out of it so what do we do we\nmake our local server right hopefully at\n\nmake our local server right hopefully at\n \n\nmake our local server right hopefully at\nsome point everyone's had a little play\n\nsome point everyone's had a little play\n \n\nsome point everyone's had a little play\nwith this uh you tweet out check out my\n\nwith this uh you tweet out check out my\n \n\nwith this uh you tweet out check out my\nlocal host um you know try and get your\n\nlocal host um you know try and get your\n \n\nlocal host um you know try and get your\nfriends to play with it it's great it's\n\nfriends to play with it it's great it's\n \n\nfriends to play with it it's great it's\nhacky it's not particularly useful to\n\nhacky it's not particularly useful to\n \n\nhacky it's not particularly useful to\nanyone except yourself. You can build\n\nanyone except yourself. You can build\n \n\nanyone except yourself. You can build\nsome fun local service. Frankly, this is\n\nsome fun local service. Frankly, this is\n \n\nsome fun local service. Frankly, this is\nwhere we see a lot of people building\n\nwhere we see a lot of people building\n \n\nwhere we see a lot of people building\nMCP right now. It's an internal demo. MCP right now. It's an internal demo. MCP right now. It's an internal demo. You're going to see if you can connect\n\nYou're going to see if you can connect\n \n\nYou're going to see if you can connect\nit to an API. Cool. It works. And then\n\nit to an API. Cool. It works. And then\n \n\nit to an API. Cool. It works. And then\nwe don't go much further than that. we don't go much further than that. we don't go much further than that." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nAnd so kind of the next step where\n\nAnd so kind of the next step where\n \n\nAnd so kind of the next step where\neveryone has been talking about a lot of\n\neveryone has been talking about a lot of\n \n\neveryone has been talking about a lot of\ndiscussion in the community, how do we\n\ndiscussion in the commun..." }, { "role": "assistant", "content": "And so kind of the next step where\n\nAnd so kind of the next step where\n \n\nAnd so kind of the next step where\neveryone has been talking about a lot of\n\neveryone has been talking about a lot of\n \n\neveryone has been talking about a lot of\ndiscussion in the community, how do we\n\ndiscussion in the community, how do we\n \n\ndiscussion in the community, how do we\ndo this really robustly is just adding\n\ndo this really robustly is just adding\n \n\ndo this really robustly is just adding\nauthentication, authorization to these\n\nauthentication, authorization to these\n \n\nauthentication, authorization to these\nworkloads, which is truthfully right now\n\nworkloads, which is truthfully right now\n \n\nworkloads, which is truthfully right now\nthe main thing you need to do. These are\n\nthe main thing you need to do. These are\n \n\nthe main thing you need to do. These are\nlinks to docs that will help you do it. links to docs that will help you do it. links to docs that will help you do it. Uh through there you can come and chat\n\nUh through there you can come and chat\n \n\nUh through there you can come and chat\nwith me and I will sit down with you\n\nwith me and I will sit down with you\n \n\nwith me and I will sit down with you\nanywhere in the world and help you build\n\nanywhere in the world and help you build\n \n\nanywhere in the world and help you build\na really robust secure MCP server and\n\na really robust secure MCP server and\n \n\na really robust secure MCP server and\nfigure out some pain points. Um you\n\nfigure out some pain points. Um you\n \n\nfigure out some pain points. Um you\nshould not have an external API that is\n\nshould not have an external API that is\n \n\nshould not have an external API that is\nunauthenticated that has no access\n\nunauthenticated that has no access\n \n\nunauthenticated that has no access\ncontrols on it. Uh things will go very\n\ncontrols on it. Uh things will go very\n \n\ncontrols on it. Uh things will go very\nwrong. So super simple you make people\n\nwrong. So super simple you make people\n \n\nwrong. So super simple you make people\nlog in. You scope you have an admin\n\nlog in. You scope you have an admin\n \n\nlog in. You scope you have an admin\nprivilege so that no one is feeding the\n\nprivilege so that no one is feeding the\n \n\nprivilege so that no one is feeding the\ngoats too much. you know, make sure\n\ngoats too much. you know, make sure\n \n\ngoats too much. you know, make sure\ngoats are well maintained and kept safe. goats are well maintained and kept safe. goats are well maintained and kept safe. Um, and maybe if you're doing this\n\nUm, and maybe if you're doing this\n \n\nUm, and maybe if you're doing this\ninternally with an organization, you put\n\ninternally with an organization, you put\n \n\ninternally with an organization, you put\nin a VPC, you're doing that gateway to\n\nin a VPC, you're doing that gateway to\n \n\nin a VPC, you're doing that gateway to\nmake sure no one can access they should,\n\nmake sure no one can access they should,\n \n\nmake sure no one can access they should,\nbut everyone turns out in your\n\nbut everyone turns out in your\n \n\nbut everyone turns out in your\norganization loves your goat emotional\n\norganization loves your goat emotional\n \n\norganization loves your goat emotional\nsupport tool because who wouldn't? And\n\nsupport tool because who wouldn't? And\n \n\nsupport tool because who wouldn't? And\nso you do what what anyone would do. You\n\nso you do what what anyone would do. You\n \n\nso you do what what anyone would do. You\nyou say, \"Let's make it public. Let's\n\nyou say, \"Let's make it public. Let's\n \n\nyou say, \"Let's make it public. Let's\nmake a public MCP server that people can\n\nmake a public MCP server that people can\n \n\nmake a public MCP server that people can\nadd to their claude, that people can\n\nadd to their claude, that people can\n \n\nadd to their claude, that people can\nbuild applications on top of. Let's add\n\nbuild applications on top of. Let's add\n \n\nbuild applications on top of. Let's add\na stripe in there to make sure they've\n\na stripe in there to make sure they've\n \n\na stripe in there to make sure they've\ngot payment rails. just sends back a URL\n\ngot payment rails. just sends back a URL\n \n\ngot payment rails. just sends back a URL\nthat you can call out to click things. that you can call out to click things. that you can call out to click things. It's pretty straightforward. Um, you\n\nIt's pretty straightforward. Um, you\n \n\nIt's pretty straightforward. Um, you\ngive some users free credits because,\n\ngive some users free credits because,\n \n\ngive some users free credits because,\nyou know, this is how apps work. Uh, and\n\nyou know, this is how apps work. Uh, and\n \n\nyou know, this is how apps work. Uh, and\nyou put it on a cloud hosting solution. you put it on a cloud hosting solution. you put it on a cloud hosting solution. Tons and tons of them are spinning up. Tons and tons of them are spinning up. Tons and tons of them are spinning up. You get ads and announcements about them\n\nYou get ads and announcements about them\n \n\nYou get ads and announcements about them\nevery day. Fundamentally, MCP servers\n\nevery day. Fundamentally, MCP servers\n \n\nevery day. Fundamentally, MCP servers\nare just a normal workload, which means\n\nare just a normal workload, which means\n \n\nare just a normal workload, which means\na lot of the cloud hosting providers can\n\na lot of the cloud hosting providers can\n \n\na lot of the cloud hosting providers can\nsupport whatever you need to host. And\n\nsupport whatever you need to host. And\n \n\nsupport whatever you need to host. And\nso there's there's lots and lots that we\n\nso there's there's lots and lots that we\n \n\nso there's there's lots and lots that we\ncan discuss there. But you do this\n\ncan discuss there. But you do this\n \n\ncan discuss there. But you do this\nreally well, and everyone loves goat\n\nreally well, and everyone loves goat\n \n\nreally well, and everyone loves goat\nfeeding and emotional support. So, you\n\nfeeding and emotional support. So, you\n \n\nfeeding and emotional support. So, you\nend up going viral. You've got a\n\nend up going viral. You've got a\n \n\nend up going viral. You've got a\nbillboard on the highway. Mr. Beast\n\nbillboard on the highway. Mr. Beast\n \n\nbillboard on the highway. Mr. Beast\ntweets you out. What happens now? You\n\ntweets you out. What happens now? You\n \n\ntweets you out. What happens now? You\nend up with having way too much\n\nend up with having way too much\n \n\nend up with having way too much\ntraction, right? You get free credit\n\ntraction, right? You get free credit\n \n\ntraction, right? You get free credit\nabuse on the signups. Uh, this is\n\nabuse on the signups. Uh, this is\n \n\nabuse on the signups. Uh, this is\nsomething that we see a ton of AI\n\nsomething that we see a ton of AI\n \n\nsomething that we see a ton of AI\ncompanies that we support happening is\n\ncompanies that we support happening is\n \n\ncompanies that we support happening is\nAI companies give you some free credits\n\nAI companies give you some free credits\n \n\nAI companies give you some free credits\nto onboard you, but at the end of the\n\nto onboard you, but at the end of the\n \n\nto onboard you, but at the end of the\nday, they're just querying an AI API,\n\nday, they're just querying an AI API,\n \n\nday, they're just querying an AI API,\nwhich means that folks will sign up for\n\nwhich means that folks will sign up for\n \n\nwhich means that folks will sign up for\nan account, abuse those free credits,\n\nan account, abuse those free credits,\n \n\nan account, abuse those free credits,\nand cycle back. one of the AI vendors uh\n\nand cycle back. one of the AI vendors uh\n \n\nand cycle back." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\none of the AI vendors uh\nthat we work with, someone was using\n\nthat we work with, someone was using\n \n\nthat we work with, someone was using\ntheir free credits to write fanfiction\n\ntheir free credits to write fanfiction\n \n\ntheir free credits to write fanfiction\nstories um by just this is not an\n\nstor..." }, { "role": "assistant", "content": "one of the AI vendors uh\nthat we work with, someone was using\n\nthat we work with, someone was using\n \n\nthat we work with, someone was using\ntheir free credits to write fanfiction\n\ntheir free credits to write fanfiction\n \n\ntheir free credits to write fanfiction\nstories um by just this is not an\n\nstories um by just this is not an\n \n\nstories um by just this is not an\napplication designed for writing\n\napplication designed for writing\n \n\napplication designed for writing\nfanfiction just because it's a\n\nfanfiction just because it's a\n \n\nfanfiction just because it's a\nconvenient way to get free credits and\n\nconvenient way to get free credits and\n \n\nconvenient way to get free credits and\nso you end up needing bot blocking on\n\nso you end up needing bot blocking on\n \n\nso you end up needing bot blocking on\nsignups. Um you need kind of robust\n\nsignups. Um you need kind of robust\n \n\nsignups. Um you need kind of robust\ncontrols over the whole orth stack to\n\ncontrols over the whole orth stack to\n \n\ncontrols over the whole orth stack to\nmake sure nothing goes wrong. Uh you're\n\nmake sure nothing goes wrong. Uh you're\n \n\nmake sure nothing goes wrong. Uh you're\ngoing to end up needing uh input\n\ngoing to end up needing uh input\n \n\ngoing to end up needing uh input\nvalidation so that no one prompt\n\nvalidation so that no one prompt\n \n\nvalidation so that no one prompt\ninjection attacks your goats. Got to\n\ninjection attacks your goats. Got to\n \n\ninjection attacks your goats. Got to\nkeep them safe. Uh and then there's a\n\nkeep them safe. Uh and then there's a\n \n\nkeep them safe. Uh and then there's a\nlot of niche stuff that happens with\n\nlot of niche stuff that happens with\n \n\nlot of niche stuff that happens with\nMCP. So, uh MCP servers dynamically\n\nMCP. So, uh MCP servers dynamically\n \n\nMCP. So, uh MCP servers dynamically\nregister their clients with the server,\n\nregister their clients with the server,\n \n\nregister their clients with the server,\nwhich means if you have any developer\n\nwhich means if you have any developer\n \n\nwhich means if you have any developer\nadmin dashboard, application dashboard\n\nadmin dashboard, application dashboard\n \n\nadmin dashboard, application dashboard\nthat's tracking the applications that\n\nthat's tracking the applications that\n \n\nthat's tracking the applications that\nyou've created, they will suddenly be\n\nyou've created, they will suddenly be\n \n\nyou've created, they will suddenly be\nflooded with MCP servers because of this\n\nflooded with MCP servers because of this\n \n\nflooded with MCP servers because of this\nweird choice in how MCP servers register\n\nweird choice in how MCP servers register\n \n\nweird choice in how MCP servers register\nas applications. And so essentially\n\nas applications. And so essentially\n \n\nas applications. And so essentially\nevery orth stack you need, every like\n\nevery orth stack you need, every like\n \n\nevery orth stack you need, every like\nmanagement tooling you need needs to be\n\nmanagement tooling you need needs to be\n \n\nmanagement tooling you need needs to be\nadapted for MCP. And so this is\n\nadapted for MCP. And so this is\n \n\nadapted for MCP. And so this is\nsomething that we provide but also like\n\nsomething that we provide but also like\n \n\nsomething that we provide but also like\na bunch of other people are working on\n\na bunch of other people are working on\n \n\na bunch of other people are working on\nand something to be conscious of is if\n\nand something to be conscious of is if\n \n\nand something to be conscious of is if\nyou scale things will start going very\n\nyou scale things will start going very\n \n\nyou scale things will start going very\nwrong. wrong. wrong. And so everyone loves your MCP server\n\nAnd so everyone loves your MCP server\n \n\nAnd so everyone loves your MCP server\nthat's super cool. Maybe it's just an AI\n\nthat's super cool. Maybe it's just an AI\n \n\nthat's super cool. Maybe it's just an AI\nAI agent wrapping it or you're selling\n\nAI agent wrapping it or you're selling\n \n\nAI agent wrapping it or you're selling\nthe core MCP server. You want to sell\n\nthe core MCP server. You want to sell\n \n\nthe core MCP server. You want to sell\ninto enterprise. What do we have to do? into enterprise. What do we have to do? into enterprise. What do we have to do? This is like a well-worn SAS path that\n\nThis is like a well-worn SAS path that\n \n\nThis is like a well-worn SAS path that\nwe're going to go through. You have to\n\nwe're going to go through. You have to\n \n\nwe're going to go through. You have to\ndo all of the boring stuff. You have to\n\ndo all of the boring stuff. You have to\n \n\ndo all of the boring stuff. You have to\ndo the SSO, the life cycle management. do the SSO, the life cycle management. do the SSO, the life cycle management. You're going to have to do provisioning. You're going to have to do provisioning. You're going to have to do provisioning. Um, Block uh are this super cool company\n\nUm, Block uh are this super cool company\n \n\nUm, Block uh are this super cool company\num that developed Goose. And internally,\n\num that developed Goose. And internally,\n \n\num that developed Goose. And internally,\nthey've been provisioning AI like access\n\nthey've been provisioning AI like access\n \n\nthey've been provisioning AI like access\nto this client, this chat client as well\n\nto this client, this chat client as well\n \n\nto this client, this chat client as well\nas MCP servers. And I really see a\n\nas MCP servers. And I really see a\n \n\nas MCP servers. And I really see a\nfuture where enterprises use SSO to\n\nfuture where enterprises use SSO to\n \n\nfuture where enterprises use SSO to\nprovision access to a ton of internal\n\nprovision access to a ton of internal\n \n\nprovision access to a ton of internal\nresources exposed via MCP that then\n\nresources exposed via MCP that then\n \n\nresources exposed via MCP that then\nemployees can chat with as a you know\n\nemployees can chat with as a you know\n \n\nemployees can chat with as a you know\ndefault way that employees are\n\ndefault way that employees are\n \n\ndefault way that employees are\nencouraged to use AI to automate\n\nencouraged to use AI to automate\n \n\nencouraged to use AI to automate\nworkflows. workflows. workflows. And so if you're going to sell into the\n\nAnd so if you're going to sell into the\n \n\nAnd so if you're going to sell into the\nenterprise, you end up needing all this\n\nenterprise, you end up needing all this\n \n\nenterprise, you end up needing all this\nlike nitty-gritty stuff that really\n\nlike nitty-gritty stuff that really\n \n\nlike nitty-gritty stuff that really\nsucks. Fine grand access controls that\n\nsucks. Fine grand access controls that\n \n\nsucks. Fine grand access controls that\nare highly performant are really robust\n\nare highly performant are really robust\n \n\nare highly performant are really robust\naudit logs um for instant responses. A\n\naudit logs um for instant responses. A\n \n\naudit logs um for instant responses. A\nlot of the regulations that exist right\n\nlot of the regulations that exist right\n \n\nlot of the regulations that exist right\nnow like GDPR um call out explicitly\n\nnow like GDPR um call out explicitly\n \n\nnow like GDPR um call out explicitly\nadditional requirements on logging for\n\nadditional requirements on logging for\n \n\nadditional requirements on logging for\nAI workloads because the regulators like\n\nAI workloads because the regulators like\n \n\nAI workloads because the regulators like\nto regulate and so your audit logs that\n\nto regulate and so your audit logs that\n \n\nto regulate and so your audit logs that\nexist normally have to fit a very\n\nexist normally have to fit a very\n \n\nexist normally have to fit a very\nspecific set of requirements which are\n\nspecific set of requirements which are\n \n\nspecific set of requirements which are\nnot well supported widely and so\n\nnot well supported widely and so\n \n\nnot well supported widely and so\nsomething to think about as you scale\n\nsomething to think about as you scale\n \n\nsomething to think about as you scale\nthese workloads. Um and you need stuff\n\nthese workloads. Um and you need stuff\n \n\nthese workloads. Um and you need stuff\nlike data loss preventions so that\n\nlike data loss preventions so that\n \n\nlike data loss preventions so that\npeople aren't like uploading random\n\npeople aren't like uploading random\n \n\npeople aren't like uploading random\nthings to MCP servers. This becomes a\n\nthings to MCP servers. This becomes a\n \n\nthings to MCP servers. This becomes a\nsignificant risk when people are just\n\nsignificant risk when people are just\n \n\nsignificant risk when people are just\nchatting wildly with a ton of servers. chatting wildly with a ton of servers. chatting wildly with a ton of servers." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nAnd so now you have an enterprise ready\n\nAnd so now you have an enterprise ready\n \n\nAnd so now you have an enterprise ready\nserverish\n\nserverish\n \n\nserverish\nbecause there are a lot of other\n\nbecause there are a lot of other\n \n\nbecause there are a lot of other\nquestions to answer. This this is a tal..." }, { "role": "assistant", "content": "And so now you have an enterprise ready\n\nAnd so now you have an enterprise ready\n \n\nAnd so now you have an enterprise ready\nserverish\n\nserverish\n \n\nserverish\nbecause there are a lot of other\n\nbecause there are a lot of other\n \n\nbecause there are a lot of other\nquestions to answer. This this is a talk\n\nquestions to answer. This this is a talk\n \n\nquestions to answer. This this is a talk\nthat was was hidden to tell you the\n\nthat was was hidden to tell you the\n \n\nthat was was hidden to tell you the\nanswers, but truthfully the answers\n\nanswers, but truthfully the answers\n \n\nanswers, but truthfully the answers\naren't entirely known yet and the\n\naren't entirely known yet and the\n \n\naren't entirely known yet and the\nprotocol is very rapidly developing. protocol is very rapidly developing. protocol is very rapidly developing. And so this is the the diagram we\n\nAnd so this is the the diagram we\n \n\nAnd so this is the the diagram we\nstarted off with. This stuff's super\n\nstarted off with. This stuff's super\n \n\nstarted off with. This stuff's super\neasy, right? How does a user log into an\n\neasy, right? How does a user log into an\n \n\neasy, right? How does a user log into an\nAI chat? Done. We've solved that. How do\n\nAI chat? Done. We've solved that. How do\n \n\nAI chat? Done. We've solved that. How do\nAI systems connect to MCP servers? This\n\nAI systems connect to MCP servers? This\n \n\nAI systems connect to MCP servers? This\nis something that's being worked on. is something that's being worked on. is something that's being worked on. It's pretty straightforward. We can do\n\nIt's pretty straightforward. We can do\n \n\nIt's pretty straightforward. We can do\nall there. You can scan those QR codes\n\nall there. You can scan those QR codes\n \n\nall there. You can scan those QR codes\nfrom before, build it, and you have a\n\nfrom before, build it, and you have a\n \n\nfrom before, build it, and you have a\nproduction ready MCP server. MCP servers\n\nproduction ready MCP server. MCP servers\n \n\nproduction ready MCP server. MCP servers\nconnecting to external resources is\n\nconnecting to external resources is\n \n\nconnecting to external resources is\npretty straightforward. It's not too\n\npretty straightforward. It's not too\n \n\npretty straightforward. It's not too\nhard to do. Can be a bit annoying, but\n\nhard to do. Can be a bit annoying, but\n \n\nhard to do. Can be a bit annoying, but\nyou know, you have to pass your scopes\n\nyou know, you have to pass your scopes\n \n\nyou know, you have to pass your scopes\nand do it properly. Um, IT admins\n\nand do it properly. Um, IT admins\n \n\nand do it properly. Um, IT admins\nmanaging AI workloads pretty\n\nmanaging AI workloads pretty\n \n\nmanaging AI workloads pretty\nstraightforward. And in theory, you\n\nstraightforward. And in theory, you\n \n\nstraightforward. And in theory, you\nknow, the other end of MCP servers\n\nknow, the other end of MCP servers\n \n\nknow, the other end of MCP servers\nexisting in your tenency connecting to\n\nexisting in your tenency connecting to\n \n\nexisting in your tenency connecting to\nresources, bit of a pain point, but can\n\nresources, bit of a pain point, but can\n \n\nresources, bit of a pain point, but can\neasily be built. There are lots of open\n\neasily be built. There are lots of open\n \n\neasily be built. There are lots of open\nquestions still as to how we do this. questions still as to how we do this. questions still as to how we do this. Um, how exactly remote asynchronous\n\nUm, how exactly remote asynchronous\n \n\nUm, how exactly remote asynchronous\nworkloads can do headless or off into\n\nworkloads can do headless or off into\n \n\nworkloads can do headless or off into\nMCP servers with dynamic client\n\nMCP servers with dynamic client\n \n\nMCP servers with dynamic client\nregistration uh and make sure that is\n\nregistration uh and make sure that is\n \n\nregistration uh and make sure that is\nyou have correct authorization controls. you have correct authorization controls. you have correct authorization controls. It's actually super hard and things keep\n\nIt's actually super hard and things keep\n \n\nIt's actually super hard and things keep\ngoing wrong along the way. um how these\n\ngoing wrong along the way. um how these\n \n\ngoing wrong along the way. um how these\nasynchronous workloads call out to\n\nasynchronous workloads call out to\n \n\nasynchronous workloads call out to\nrelevant users. Uh there is a a new RFC\n\nrelevant users. Uh there is a a new RFC\n \n\nrelevant users. Uh there is a a new RFC\nin the MCP spec for elicitation. So when\n\nin the MCP spec for elicitation. So when\n \n\nin the MCP spec for elicitation. So when\nyour model doesn't know what to do, it\n\nyour model doesn't know what to do, it\n \n\nyour model doesn't know what to do, it\nknows that it should ask a human being\n\nknows that it should ask a human being\n \n\nknows that it should ask a human being\nand this is becoming supported as part\n\nand this is becoming supported as part\n \n\nand this is becoming supported as part\nof MCP that it can go and call out\n\nof MCP that it can go and call out\n \n\nof MCP that it can go and call out\ntowards a user um and say please provide\n\ntowards a user um and say please provide\n \n\ntowards a user um and say please provide\nadditional input that you need because\n\nadditional input that you need because\n \n\nadditional input that you need because\nI'm missing details. So stuff like this\n\nI'm missing details. So stuff like this\n \n\nI'm missing details. So stuff like this\nis actively developing and nothing is\n\nis actively developing and nothing is\n \n\nis actively developing and nothing is\nstable. Um and then there's this very\n\nstable. Um and then there's this very\n \n\nstable. Um and then there's this very\nbroad question that we're running into\n\nbroad question that we're running into\n \n\nbroad question that we're running into\nwith AI workloads um of uh passing scope\n\nwith AI workloads um of uh passing scope\n \n\nwith AI workloads um of uh passing scope\nbetween different AI workloads passing\n\nbetween different AI workloads passing\n \n\nbetween different AI workloads passing\naccess control between AI workloads\n\naccess control between AI workloads\n \n\naccess control between AI workloads\nwhere if you have the A2 we're talking\n\nwhere if you have the A2 we're talking\n \n\nwhere if you have the A2 we're talking\nabout MCP if you're using A2A protocol\n\nabout MCP if you're using A2A protocol\n \n\nabout MCP if you're using A2A protocol\nyou are just telling an agent mostly in\n\nyou are just telling an agent mostly in\n \n\nyou are just telling an agent mostly in\nvibes what it should or shouldn't do and\n\nvibes what it should or shouldn't do and\n \n\nvibes what it should or shouldn't do and\nrelying upon the alignment of that model\n\nrelying upon the alignment of that model\n \n\nrelying upon the alignment of that model\nto make sure it doesn't misbehave and so\n\nto make sure it doesn't misbehave and so\n \n\nto make sure it doesn't misbehave and so\nactually passing robust authorization\n\nactually passing robust authorization\n \n\nactually passing robust authorization\nyou know scopes uh communicating access\n\nyou know scopes uh communicating access\n \n\nyou know scopes uh communicating access\ncontrols making sure that service\n\ncontrols making sure that service\n \n\ncontrols making sure that service\naccount that AI workload might be has\n\naccount that AI workload might be has\n \n\naccount that AI workload might be has\nthe correct access turns out to be a\n\nthe correct access turns out to be a\n \n\nthe correct access turns out to be a\nhuge pain point in doing this. So\n\nhuge pain point in doing this. So\n \n\nhuge pain point in doing this. So\nfrankly the cloud vendors are mostly\n\nfrankly the cloud vendors are mostly\n \n\nfrankly the cloud vendors are mostly\nsolving the cloud hosting. I think the\n\nsolving the cloud hosting. I think the\n \n\nsolving the cloud hosting. I think the\nauthorization and access control element\n\nauthorization and access control element\n \n\nauthorization and access control element\nof this is the hardest part of putting\n\nof this is the hardest part of putting\n \n\nof this is the hardest part of putting\nthis into external enterprise workloads\n\nthis into external enterprise workloads\n \n\nthis into external enterprise workloads\num and a big thing that needs to be\n\num and a big thing that needs to be\n \n\num and a big thing that needs to be\nfilled. filled. filled. So\n\nSo\n \n\nSo\nif you want to build fun things, we are\n\nif you want to build fun things, we are\n \n\nif you want to build fun things, we are\nactively building out this entire stack\n\nactively building out this entire stack\n \n\nactively building out this entire stack\nto sell to AI companies and startups." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nUh\n\nto sell to AI companies and startups. Uh\n \n\nto sell to AI companies and startups. Uh\nand if you go to the docs, we can build\n\nand if you go to the docs, we can build\n \n\nand if you go to the docs, we can build\nsome cool stuff together. Um everyone is\n\nsome cool stuff together. Um everyone is\n \n\ns..." }, { "role": "assistant", "content": "Uh\n\nto sell to AI companies and startups. Uh\n \n\nto sell to AI companies and startups. Uh\nand if you go to the docs, we can build\n\nand if you go to the docs, we can build\n \n\nand if you go to the docs, we can build\nsome cool stuff together. Um everyone is\n\nsome cool stuff together. Um everyone is\n \n\nsome cool stuff together. Um everyone is\nbuilding interesting things right now. building interesting things right now. building interesting things right now. And you know, it's fun to get into the\n\nAnd you know, it's fun to get into the\n \n\nAnd you know, it's fun to get into the\nnitty-gritty and and see what bugs\n\nnitty-gritty and and see what bugs\n \n\nnitty-gritty and and see what bugs\noccur. And then as a thank you for all\n\noccur. And then as a thank you for all\n \n\noccur. And then as a thank you for all\nlistening to this fun talk, uh you\n\nlistening to this fun talk, uh you\n \n\nlistening to this fun talk, uh you\ncannot use this on your phones. To be\n\ncannot use this on your phones. To be\n \n\ncannot use this on your phones. To be\nclear, you can buy this shirt only\n\nclear, you can buy this shirt only\n \n\nclear, you can buy this shirt only\nthrough MCP. through MCP. through MCP. Uh the instructions are on the website. Uh the instructions are on the website. Uh the instructions are on the website. Uh very briefly, if you'd like to see,\n\nUh very briefly, if you'd like to see,\n \n\nUh very briefly, if you'd like to see,\nlet me see if I can get out of this um\n\nlet me see if I can get out of this um\n \n\nlet me see if I can get out of this um\nand go to cursor. You have to add the\n\nand go to cursor. You have to add the\n \n\nand go to cursor. You have to add the\nMCP server. There are many ways to do\n\nMCP server. There are many ways to do\n \n\nMCP server. There are many ways to do\nthis, but the way I'd recommend is just\n\nthis, but the way I'd recommend is just\n \n\nthis, but the way I'd recommend is just\ngoing into one of your AI um editors, uh\n\ngoing into one of your AI um editors, uh\n \n\ngoing into one of your AI um editors, uh\ntyping in the JSON that the website will\n\ntyping in the JSON that the website will\n \n\ntyping in the JSON that the website will\ngive you. This is like the easiest way\n\ngive you. This is like the easiest way\n \n\ngive you. This is like the easiest way\nto add. You'll see that it's turned on. to add. You'll see that it's turned on. to add. You'll see that it's turned on. You'll need to make an account because\n\nYou'll need to make an account because\n \n\nYou'll need to make an account because\nauthorization, access control, etc. Um\n\nauthorization, access control, etc. Um\n \n\nauthorization, access control, etc. Um\nand then you can say, \"Please buy me a\n\nand then you can say, \"Please buy me a\n \n\nand then you can say, \"Please buy me a\nshirt. shirt. shirt. Um, and of course the Wi-Fi um will not\n\nUm, and of course the Wi-Fi um will not\n \n\nUm, and of course the Wi-Fi um will not\nwork. Beautiful. Will need my mailing address. Beautiful. Will need my mailing address. Beautiful. Will need my mailing address. One, two, three. Hack away. Um, work OS. One, two, three. Hack away. Um, work OS. One, two, three. Hack away. Um, work OS. And I want a medium. And I want a medium. And I want a medium. And it knows my name. It's authorized. And it knows my name. It's authorized. And it knows my name. It's authorized. It's done a validation check on me. It's done a validation check on me. It's done a validation check on me. We've got a bot blocker on the signup so\n\nWe've got a bot blocker on the signup so\n \n\nWe've got a bot blocker on the signup so\nthat people can't abuse my shirts. And\n\nthat people can't abuse my shirts. And\n \n\nthat people can't abuse my shirts. And\nyou'll see right there that we purchased\n\nyou'll see right there that we purchased\n \n\nyou'll see right there that we purchased\na shirt. So, let's go back to Keynote. a shirt. So, let's go back to Keynote. a shirt. So, let's go back to Keynote. um scan it, add it to your computer if\n\num scan it, add it to your computer if\n \n\num scan it, add it to your computer if\nyou want. Um and you guys can have a\n\nyou want. Um and you guys can have a\n \n\nyou want. Um and you guys can have a\nfree very nice MCP shirt uh that I would\n\nfree very nice MCP shirt uh that I would\n \n\nfree very nice MCP shirt uh that I would\nlove to share around. Thank you very\n\nlove to share around. Thank you very\n \n\nlove to share around. Thank you very\nmuch. [Music]" } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nAwesome. Awesome. Awesome. [Music]\n\n \n\nThanks so much for coming. Um, I wanted\n\nThanks so much for coming. Um, I wanted\n \n\nThanks so much for coming. Um, I wanted\nto give a bit of a talk on implementing\n\nto give a bit of a talk on implementing\n \n\nto give a bit of a talk on implementing\nMCP clients a..." }, { "role": "assistant", "content": "Awesome. Awesome. Awesome. [Music]\n\n \n\nThanks so much for coming. Um, I wanted\n\nThanks so much for coming. Um, I wanted\n \n\nThanks so much for coming. Um, I wanted\nto give a bit of a talk on implementing\n\nto give a bit of a talk on implementing\n \n\nto give a bit of a talk on implementing\nMCP clients and talk to more about MCP\n\nMCP clients and talk to more about MCP\n \n\nMCP clients and talk to more about MCP\nat scale within a large organization\n\nat scale within a large organization\n \n\nat scale within a large organization\nlike Anthropic. Um, I wanted to give\n\nlike Anthropic. Um, I wanted to give\n \n\nlike Anthropic. Um, I wanted to give\nfirst a little introduction from me. Uh,\n\nfirst a little introduction from me. Uh,\n \n\nfirst a little introduction from me. Uh,\nmy name is John. I've spent 20 years\n\nmy name is John. I've spent 20 years\n \n\nmy name is John. I've spent 20 years\nbuilding large scale systems and dealing\n\nbuilding large scale systems and dealing\n \n\nbuilding large scale systems and dealing\nwith the problems that that causes. And\n\nwith the problems that that causes. And\n \n\nwith the problems that that causes. And\nso I've made a lot of mistakes and uh\n\nso I've made a lot of mistakes and uh\n \n\nso I've made a lot of mistakes and uh\nI'm excited to give maybe some thoughts\n\nI'm excited to give maybe some thoughts\n \n\nI'm excited to give maybe some thoughts\non avoiding some of those mistakes. I'm\n\non avoiding some of those mistakes. I'm\n \n\non avoiding some of those mistakes. I'm\ncurrently a member of technical staff\n\ncurrently a member of technical staff\n \n\ncurrently a member of technical staff\nhere at Anthropic and I've spent the\n\nhere at Anthropic and I've spent the\n \n\nhere at Anthropic and I've spent the\npast few months um focusing on tool\n\npast few months um focusing on tool\n \n\npast few months um focusing on tool\ncalling and integration and implementing\n\ncalling and integration and implementing\n \n\ncalling and integration and implementing\nMCP support for all of our internal like\n\nMCP support for all of our internal like\n \n\nMCP support for all of our internal like\nexternal integrations within the or\n\nexternal integrations within the or\n \n\nexternal integrations within the or\nso\n\nso\n \n\nso\nlooking at tool integration with models\n\nlooking at tool integration with models\n \n\nlooking at tool integration with models\nwe've kind of hit this timeline where uh\n\nwe've kind of hit this timeline where uh\n \n\nwe've kind of hit this timeline where uh\nmodels only really got good at calling\n\nmodels only really got good at calling\n \n\nmodels only really got good at calling\ntools\n\ntools\n \n\ntools\nuh like kind of late midl last year and\n\nuh like kind of late midl last year and\n \n\nuh like kind of late midl last year and\nsuddenly\n\nsuddenly\n \n\nsuddenly\nEveryone got very excited because like\n\nEveryone got very excited because like\n \n\nEveryone got very excited because like\nyour model could go and call your Google\n\nyour model could go and call your Google\n \n\nyour model could go and call your Google\nDrive and then it could call your maps\n\nDrive and then it could call your maps\n \n\nDrive and then it could call your maps\nand then it could send a text message to\n\nand then it could send a text message to\n \n\nand then it could send a text message to\npeople. And so there's a huge explosion\n\npeople. And so there's a huge explosion\n \n\npeople. And so there's a huge explosion\nwith like very little effort you can\n\nwith like very little effort you can\n \n\nwith like very little effort you can\nmake very cool things and so um teams\n\nmake very cool things and so um teams\n \n\nmake very cool things and so um teams\nare all trying to move fast. Everyone's\n\nare all trying to move fast. Everyone's\n \n\nare all trying to move fast. Everyone's\nmoving very fast in AI. Custom endpoints\n\nmoving very fast in AI. Custom endpoints\n \n\nmoving very fast in AI. Custom endpoints\nstart proliferating for every use case. start proliferating for every use case. start proliferating for every use case. there's a lot of like services popping\n\nthere's a lot of like services popping\n \n\nthere's a lot of like services popping\nup with like slashcall tool and slash\n\nup with like slashcall tool and slash\n \n\nup with like slashcall tool and slash\nlike get context and then people um\n\nlike get context and then people um\n \n\nlike get context and then people um\nstart to realize there's additional\n\nstart to realize there's additional\n \n\nstart to realize there's additional\nneeds some authentication a b a bunch of\n\nneeds some authentication a b a bunch of\n \n\nneeds some authentication a b a bunch of\nstuff there and this kind of led to some\n\nstuff there and this kind of led to some\n \n\nstuff there and this kind of led to some\nintegration chaos where you're\n\nintegration chaos where you're\n \n\nintegration chaos where you're\nduplicating a bunch of functionality\n\nduplicating a bunch of functionality\n \n\nduplicating a bunch of functionality\naround your org nothing really works the\n\naround your org nothing really works the\n \n\naround your org nothing really works the\nsame you have an integration that works\n\nsame you have an integration that works\n \n\nsame you have an integration that works\nreally well in service A but then you\n\nreally well in service A but then you\n \n\nreally well in service A but then you\nwant to use it in service B but you\n\nwant to use it in service B but you\n \n\nwant to use it in service B but you\ncan't because it's going to take you\n\ncan't because it's going to take you\n \n\ncan't because it's going to take you\nthree weeks to rewrite it to talk to the\n\nthree weeks to rewrite it to talk to the\n \n\nthree weeks to rewrite it to talk to the\nnew interface And so we're in this kind\n\nnew interface And so we're in this kind\n \n\nnew interface And so we're in this kind\nof spot and the place that we came to at\n\nof spot and the place that we came to at\n \n\nof spot and the place that we came to at\nAnthropic is realizing that over time\n\nAnthropic is realizing that over time\n \n\nAnthropic is realizing that over time\nall of these endpoints started to look a\n\nall of these endpoints started to look a\n \n\nall of these endpoints started to look a\nlot like MCP. Uh you you end up with\n\nlot like MCP. Uh you you end up with\n \n\nlot like MCP. Uh you you end up with\nsome get tools, some get resources, some\n\nsome get tools, some get resources, some\n \n\nsome get tools, some get resources, some\nelicitation of of details. Um,\n\nelicitation of of details. Um,\n \n\nelicitation of of details. Um,\nand even if you're not using the entire\n\nand even if you're not using the entire\n \n\nand even if you're not using the entire\nfeature space of MCP uh as a whole\n\nfeature space of MCP uh as a whole\n \n\nfeature space of MCP uh as a whole\nimmediately, like you're probably going\n\nimmediately, like you're probably going\n \n\nimmediately, like you're probably going\nto go extend into something that kind of\n\nto go extend into something that kind of\n \n\nto go extend into something that kind of\nlooks like it over time. And when I'm\n\nlooks like it over time. And when I'm\n \n\nlooks like it over time. And when I'm\ntalking about MCP here, there's kind of\n\ntalking about MCP here, there's kind of\n \n\ntalking about MCP here, there's kind of\ntwo sides to MCP that in my mind feel a\n\ntwo sides to MCP that in my mind feel a\n \n\ntwo sides to MCP that in my mind feel a\nbit unrelated. There's this JSON RPC\n\nbit unrelated. There's this JSON RPC\n \n\nbit unrelated. There's this JSON RPC\nspecification which is really valuable\n\nspecification which is really valuable\n \n\nspecification which is really valuable\nas engineers. It's like a standard way\n\nas engineers. It's like a standard way\n \n\nas engineers. It's like a standard way\nof sending messages and communicating\n\nof sending messages and communicating\n \n\nof sending messages and communicating\nback and forth between uh providers of\n\nback and forth between uh providers of\n \n\nback and forth between uh providers of\ncontext for your models and the code\n\ncontext for your models and the code\n \n\ncontext for your models and the code\nthat's interacting with the models. And\n\nthat's interacting with the models. And\n \n\nthat's interacting with the models. And\nuh getting those messages right is the\n\nuh getting those messages right is the\n \n\nuh getting those messages right is the\ntopic of huge debate on like the MCP\n\ntopic of huge debate on like the MCP\n \n\ntopic of huge debate on like the MCP\nuh repos." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nIf you're involved with any\n\nuh repos. If you're involved with any\n \n\nuh repos. If you're involved with any\nstandardization process ever, you know\n\nstandardization process ever, you know\n \n\nstandardization process ever, you know\nhow those conversations end up going. how those conversations end up go..." }, { "role": "assistant", "content": "If you're involved with any\n\nuh repos. If you're involved with any\n \n\nuh repos. If you're involved with any\nstandardization process ever, you know\n\nstandardization process ever, you know\n \n\nstandardization process ever, you know\nhow those conversations end up going. how those conversations end up going. how those conversations end up going. And then on the other side there's this\n\nAnd then on the other side there's this\n \n\nAnd then on the other side there's this\nglobal transport standard which is the\n\nglobal transport standard which is the\n \n\nglobal transport standard which is the\nstuff around streamable HTTP oath 2.1\n\nstuff around streamable HTTP oath 2.1\n \n\nstuff around streamable HTTP oath 2.1\nsession management. And global transport\n\nsession management. And global transport\n \n\nsession management. And global transport\nstandard is hard because you're trying\n\nstandard is hard because you're trying\n \n\nstandard is hard because you're trying\nto get everyone to speak the same\n\nto get everyone to speak the same\n \n\nto get everyone to speak the same\nlanguage. And so it's really nitty but\n\nlanguage. And so it's really nitty but\n \n\nlanguage. And so it's really nitty but\nthere's not a lot of like most of the\n\nthere's not a lot of like most of the\n \n\nthere's not a lot of like most of the\njuice of MCP is in this the message\n\njuice of MCP is in this the message\n \n\njuice of MCP is in this the message\nspecification and the way that the uh\n\nspecification and the way that the uh\n \n\nspecification and the way that the uh\nservers are interacting. Um and so we\n\nservers are interacting. Um and so we\n \n\nservers are interacting. Um and so we\nstarted asking ourselves like can we\n\nstarted asking ourselves like can we\n \n\nstarted asking ourselves like can we\njust MCP for everything and we said yes\n\njust MCP for everything and we said yes\n \n\njust MCP for everything and we said yes\nwith the caveat that yes is for\n\nwith the caveat that yes is for\n \n\nwith the caveat that yes is for\neverything involved in presiding model\n\neverything involved in presiding model\n \n\neverything involved in presiding model\ncontext to models. Um we have this\n\ncontext to models. Um we have this\n \n\ncontext to models. Um we have this\nformat where your client is sending\n\nformat where your client is sending\n \n\nformat where your client is sending\nthese messages. Something's responding\n\nthese messages. Something's responding\n \n\nthese messages. Something's responding\nwith these messages. Um where that\n\nwith these messages. Um where that\n \n\nwith these messages. Um where that\nstream is going it really doesn't\n\nstream is going it really doesn't\n \n\nstream is going it really doesn't\nmatter. It can be on the same process. matter. It can be on the same process. matter. It can be on the same process. It can be another data center. can be\n\nIt can be another data center. can be\n \n\nIt can be another data center. can be\nthrough a giant pile of uh enterprise\n\nthrough a giant pile of uh enterprise\n \n\nthrough a giant pile of uh enterprise\nnetworking stuff. Um, it doesn't really\n\nnetworking stuff. Um, it doesn't really\n \n\nnetworking stuff. Um, it doesn't really\ncare at the point that your code is\n\ncare at the point that your code is\n \n\ncare at the point that your code is\ninteracting with it. You're just calling\n\ninteracting with it. You're just calling\n \n\ninteracting with it. You're just calling\na connect to MCP and you have a a set of\n\na connect to MCP and you have a a set of\n \n\na connect to MCP and you have a a set of\nuh a set of tools and methods that you\n\nuh a set of tools and methods that you\n \n\nuh a set of tools and methods that you\ncan call. So, uh,\n\ncan call. So, uh,\n \n\ncan call. So, uh,\nstandardizing on that seemed useful. Um,\n\nstandardizing on that seemed useful. Um,\n \n\nstandardizing on that seemed useful. Um,\nstandard why standardize on anything\n\nstandard why standardize on anything\n \n\nstandard why standardize on anything\ninternally? Um, being boring on stuff\n\ninternally? Um, being boring on stuff\n \n\ninternally? Um, being boring on stuff\nlike this is good. It's not a\n\nlike this is good. It's not a\n \n\nlike this is good. It's not a\ncompetitive advantage to be really good\n\ncompetitive advantage to be really good\n \n\ncompetitive advantage to be really good\nat making Google Drive talk to your app. at making Google Drive talk to your app. at making Google Drive talk to your app. It's just a thing that you need to do. It's just a thing that you need to do. It's just a thing that you need to do. It's not your differentiator. Uh having\n\nIt's not your differentiator. Uh having\n \n\nIt's not your differentiator. Uh having\na single approach to learn as engineers\n\na single approach to learn as engineers\n \n\na single approach to learn as engineers\nmakes things faster. You can spend your\n\nmakes things faster. You can spend your\n \n\nmakes things faster. You can spend your\ncycles working on interesting problems\n\ncycles working on interesting problems\n \n\ncycles working on interesting problems\ninstead of trying to figure out how to\n\ninstead of trying to figure out how to\n \n\ninstead of trying to figure out how to\nplum uh integration. And uh if you're\n\nplum uh integration. And uh if you're\n \n\nplum uh integration. And uh if you're\nusing the same thing everywhere, then\n\nusing the same thing everywhere, then\n \n\nusing the same thing everywhere, then\nlike each new integration might clean up\n\nlike each new integration might clean up\n \n\nlike each new integration might clean up\nthe field a bit for the next person who\n\nthe field a bit for the next person who\n \n\nthe field a bit for the next person who\ncomes along. Um it's it's over overall a\n\ncomes along. Um it's it's over overall a\n \n\ncomes along. Um it's it's over overall a\ngood thing in cases like this where\n\ngood thing in cases like this where\n \n\ngood thing in cases like this where\nwe're we're not really doing anything\n\nwe're we're not really doing anything\n \n\nwe're we're not really doing anything\ninteresting. We're plumbing context\n\ninteresting. We're plumbing context\n \n\ninteresting. We're plumbing context\nbetween integrations and things that are\n\nbetween integrations and things that are\n \n\nbetween integrations and things that are\nconsuming the integrations. Uh\n\nconsuming the integrations. Uh\n \n\nconsuming the integrations. Uh\nwhat is standardized on MCP internally? what is standardized on MCP internally? what is standardized on MCP internally? Um I this is where I might make an\n\nUm I this is where I might make an\n \n\nUm I this is where I might make an\nargument to everyone that there's\n\nargument to everyone that there's\n \n\nargument to everyone that there's\nalready ecosystem demand. You have to\n\nalready ecosystem demand. You have to\n \n\nalready ecosystem demand. You have to\nimplement MCP because everyone's\n\nimplement MCP because everyone's\n \n\nimplement MCP because everyone's\nimplementing MCP. So why do two things? implementing MCP. So why do two things? implementing MCP. So why do two things? Um it's becoming an industry standard. Um it's becoming an industry standard. Um it's becoming an industry standard. there's a large coalition of engineers\n\nthere's a large coalition of engineers\n \n\nthere's a large coalition of engineers\nand organizations that are all involved\n\nand organizations that are all involved\n \n\nand organizations that are all involved\nin building out the standard. Uh all of\n\nin building out the standard. Uh all of\n \n\nin building out the standard. Uh all of\nthe major AI labs are represented in\n\nthe major AI labs are represented in\n \n\nthe major AI labs are represented in\nthat. So you you know that as new model\n\nthat. So you you know that as new model\n \n\nthat. So you you know that as new model\ncapabilities start to be developed uh\n\ncapabilities start to be developed uh\n \n\ncapabilities start to be developed uh\nthose patterns will be added to the\n\nthose patterns will be added to the\n \n\nthose patterns will be added to the\nprotocol because all the labs want you\n\nprotocol because all the labs want you\n \n\nprotocol because all the labs want you\nto use their features. So I think the\n\nto use their features. So I think the\n \n\nto use their features. So I think the\nstandardizing on MCP internally for this\n\nstandardizing on MCP internally for this\n \n\nstandardizing on MCP internally for this\ntype of context is a is a good bet. And\n\ntype of context is a is a good bet. And\n \n\ntype of context is a is a good bet. And\none of the things you get with MCP is\n\none of the things you get with MCP is\n \n\none of the things you get with MCP is\nthat it solves problems that you haven't\n\nthat it solves problems that you haven't\n \n\nthat it solves problems that you haven't\nactually run into yet." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nlike there's a\n\nactually run into yet. like there's a\n \n\nactually run into yet. like there's a\nbunch of stuff in the protocol that\n\nbunch of stuff in the protocol that\n \n\nbunch of stuff in the protocol that\nexists because there's a problem and a\n\nexists because there's a problem and a\n \n\nexists beca..." }, { "role": "assistant", "content": "like there's a\n\nactually run into yet. like there's a\n \n\nactually run into yet. like there's a\nbunch of stuff in the protocol that\n\nbunch of stuff in the protocol that\n \n\nbunch of stuff in the protocol that\nexists because there's a problem and a\n\nexists because there's a problem and a\n \n\nexists because there's a problem and a\nneed and\n\nneed and\n \n\nneed and\nhaving those solutions at hand when you\n\nhaving those solutions at hand when you\n \n\nhaving those solutions at hand when you\nrun into them is really important. So\n\nrun into them is really important. So\n \n\nrun into them is really important. So\nsampling an example of where this might\n\nsampling an example of where this might\n \n\nsampling an example of where this might\nbe valuable in your company. You might\n\nbe valuable in your company. You might\n \n\nbe valuable in your company. You might\nhave four products that have four\n\nhave four products that have four\n \n\nhave four products that have four\ndifferent billing models uh for reasons\n\ndifferent billing models uh for reasons\n \n\ndifferent billing models uh for reasons\nbecause you're building fast. Um you\n\nbecause you're building fast. Um you\n \n\nbecause you're building fast. Um you\nmight have a bunch of different token\n\nmight have a bunch of different token\n \n\nmight have a bunch of different token\nlimits. You might have different ways of\n\nlimits. You might have different ways of\n \n\nlimits. You might have different ways of\ntracking usage. This is really painful\n\ntracking usage. This is really painful\n \n\ntracking usage. This is really painful\nbecause you want to write one\n\nbecause you want to write one\n \n\nbecause you want to write one\nintegration service to connect your\n\nintegration service to connect your\n \n\nintegration service to connect your\nslides and how do you go and like hook\n\nslides and how do you go and like hook\n \n\nslides and how do you go and like hook\nthe billing and the tokens up correctly\n\nthe billing and the tokens up correctly\n \n\nthe billing and the tokens up correctly\nand MCP has uh already has sampling\n\nand MCP has uh already has sampling\n \n\nand MCP has uh already has sampling\nprimitives so you can build your\n\nprimitives so you can build your\n \n\nprimitives so you can build your\nintegration you can just be like okay\n\nintegration you can just be like okay\n \n\nintegration you can just be like okay\nyour integration sends a sampling\n\nyour integration sends a sampling\n \n\nyour integration sends a sampling\nrequest over the stream uh the other end\n\nrequest over the stream uh the other end\n \n\nrequest over the stream uh the other end\nof the pipe fulfills that request you\n\nof the pipe fulfills that request you\n \n\nof the pipe fulfills that request you\ncan go and hook it in everything works\n\ncan go and hook it in everything works\n \n\ncan go and hook it in everything works\ngreat and so this is a thing that uh a\n\ngreat and so this is a thing that uh a\n \n\ngreat and so this is a thing that uh a\nshape problem that might take you a\n\nshape problem that might take you a\n \n\nshape problem that might take you a\nbunch of effort\n\nbunch of effort\n \n\nbunch of effort\nuh internally without this, but you\n\nuh internally without this, but you\n \n\nuh internally without this, but you\nalready have the answer kind of gift\n\nalready have the answer kind of gift\n \n\nalready have the answer kind of gift\nwrapped for you in the protocol. wrapped for you in the protocol. wrapped for you in the protocol. And so at Anthropic, we're running into\n\nAnd so at Anthropic, we're running into\n \n\nAnd so at Anthropic, we're running into\nsome requirements converging. We're\n\nsome requirements converging. We're\n \n\nsome requirements converging. We're\nstarting to see external remote MCP\n\nstarting to see external remote MCP\n \n\nstarting to see external remote MCP\nservices popping up like mcp. services popping up like mcp. services popping up like mcp. We wanted to be able to talk to those. We wanted to be able to talk to those. We wanted to be able to talk to those. Talking to those is complex because you\n\nTalking to those is complex because you\n \n\nTalking to those is complex because you\nneed external network connectivity, you\n\nneed external network connectivity, you\n \n\nneed external network connectivity, you\nneed authentication. Uh\n\nneed authentication. Uh\n \n\nneed authentication. Uh\nthere's a proliferation of internal\n\nthere's a proliferation of internal\n \n\nthere's a proliferation of internal\nagents. people have started building uh\n\nagents. people have started building uh\n \n\nagents. people have started building uh\nPR review bots and like Slack management\n\nPR review bots and like Slack management\n \n\nPR review bots and like Slack management\nthings and just lots of people have lots\n\nthings and just lots of people have lots\n \n\nthings and just lots of people have lots\nof ideas. No one's really sure what's\n\nof ideas. No one's really sure what's\n \n\nof ideas. No one's really sure what's\ngoing to hit. So, we're having a huge\n\ngoing to hit. So, we're having a huge\n \n\ngoing to hit. So, we're having a huge\nexplosion of LLMbacked services\n\nexplosion of LLMbacked services\n \n\nexplosion of LLMbacked services\ninternally. Uh with that explosion,\n\ninternally. Uh with that explosion,\n \n\ninternally. Uh with that explosion,\nthere's a bunch of security concerns\n\nthere's a bunch of security concerns\n \n\nthere's a bunch of security concerns\nwhere uh you don't really want all of\n\nwhere uh you don't really want all of\n \n\nwhere uh you don't really want all of\nthose services to be going and accessing\n\nthose services to be going and accessing\n \n\nthose services to be going and accessing\nuser credentials uh because that ends up\n\nuser credentials uh because that ends up\n \n\nuser credentials uh because that ends up\nbeing being kind of a nightmare. don't\n\nbeing being kind of a nightmare. don't\n \n\nbeing being kind of a nightmare. don't\nwant uh outbound external network\n\nwant uh outbound external network\n \n\nwant uh outbound external network\nconnectivity everywhere. Um auditing\n\nconnectivity everywhere. Um auditing\n \n\nconnectivity everywhere. Um auditing\nbecomes really complex. Uh and so we are\n\nbecomes really complex. Uh and so we are\n \n\nbecomes really complex. Uh and so we are\nlooking at this problem. We wanted to be\n\nlooking at this problem. We wanted to be\n \n\nlooking at this problem. We wanted to be\nable to build our integrations once and\n\nable to build our integrations once and\n \n\nable to build our integrations once and\nuse them anywhere. And so, uh, a model I\n\nuse them anywhere. And so, uh, a model I\n \n\nuse them anywhere. And so, uh, a model I\nwas introduced to by a mentor of mine\n\nwas introduced to by a mentor of mine\n \n\nwas introduced to by a mentor of mine\nand a while ago is the pit of success,\n\nand a while ago is the pit of success,\n \n\nand a while ago is the pit of success,\nwhich is the idea that, um, if you make\n\nwhich is the idea that, um, if you make\n \n\nwhich is the idea that, um, if you make\nthe\n\nthe\n \n\nthe\nright thing to do, the easiest thing to\n\nright thing to do, the easiest thing to\n \n\nright thing to do, the easiest thing to\ndo, then everyone in your org kind of\n\ndo, then everyone in your org kind of\n \n\ndo, then everyone in your org kind of\nfalls into it. And so uh we designed a\n\nfalls into it. And so uh we designed a\n \n\nfalls into it. And so uh we designed a\nservice which is just a piece of shared\n\nservice which is just a piece of shared\n \n\nservice which is just a piece of shared\ninfrastructure called the MCP gateway\n\ninfrastructure called the MCP gateway\n \n\ninfrastructure called the MCP gateway\nthat provides a single point of entry\n\nthat provides a single point of entry\n \n\nthat provides a single point of entry\nand provided engineers just with a\n\nand provided engineers just with a\n \n\nand provided engineers just with a\nconnect to MCP call that returns a MCP\n\nconnect to MCP call that returns a MCP\n \n\nconnect to MCP call that returns a MCP\nSDK client session on the end and we're\n\nSDK client session on the end and we're\n \n\nSDK client session on the end and we're\ntrying to make that as simple as\n\ntrying to make that as simple as\n \n\ntrying to make that as simple as\npossible uh\n\npossible uh\n \n\npossible uh\nbecause\n\nbecause\n \n\nbecause\nthat way people will use it if it's the\n\nthat way people will use it if it's the\n \n\nthat way people will use it if it's the\neasiest thing to do. Um, we used URL\n\neasiest thing to do. Um, we used URL\n \n\neasiest thing to do." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nUm, we used URL\nbased routing to route to external\n\nbased routing to route to external\n \n\nbased routing to route to external\nservers, internal servers, it doesn't\n\nservers, internal servers, it doesn't\n \n\nservers, internal servers, it doesn't\nmatter. It's all the same call. Uh, we\n\nmatter. It's all ..." }, { "role": "assistant", "content": "Um, we used URL\nbased routing to route to external\n\nbased routing to route to external\n \n\nbased routing to route to external\nservers, internal servers, it doesn't\n\nservers, internal servers, it doesn't\n \n\nservers, internal servers, it doesn't\nmatter. It's all the same call. Uh, we\n\nmatter. It's all the same call. Uh, we\n \n\nmatter. It's all the same call. Uh, we\nhandle all the credential management\n\nhandle all the credential management\n \n\nhandle all the credential management\nautomatically because you don't want to\n\nautomatically because you don't want to\n \n\nautomatically because you don't want to\nbe implementing OOTH five times in your\n\nbe implementing OOTH five times in your\n \n\nbe implementing OOTH five times in your\ncompany. Uh, gives you a centralized\n\ncompany. Uh, gives you a centralized\n \n\ncompany. Uh, gives you a centralized\nplace for rate limiting and\n\nplace for rate limiting and\n \n\nplace for rate limiting and\nobservability. observability. observability. Uh, I have an obligatory diagram here of\n\nUh, I have an obligatory diagram here of\n \n\nUh, I have an obligatory diagram here of\na bunch of lines going in and out. But,\n\na bunch of lines going in and out. But,\n \n\na bunch of lines going in and out. But,\nuh, here's a a gateway in the middle. uh, here's a a gateway in the middle. uh, here's a a gateway in the middle. This is kind of the thing. Just one more\n\nThis is kind of the thing. Just one more\n \n\nThis is kind of the thing. Just one more\nbox will solve all our problems. Uh, can\n\nbox will solve all our problems. Uh, can\n \n\nbox will solve all our problems. Uh, can\nI go next? I go next? I go next? Uh where is my\n\nUh where is my\n \n\nUh where is my\nyeah ah uh so the uh the code that we\n\nyeah ah uh so the uh the code that we\n \n\nyeah ah uh so the uh the code that we\nhave here we just made some client\n\nhave here we just made some client\n \n\nhave here we just made some client\nlibraries where we just MCB gateway\n\nlibraries where we just MCB gateway\n \n\nlibraries where we just MCB gateway\nconnect to MCP uh we pass in a URL an\n\nconnect to MCP uh we pass in a URL an\n \n\nconnect to MCP uh we pass in a URL an\norg ID account ID this is like a bit\n\norg ID account ID this is like a bit\n \n\norg ID account ID this is like a bit\nsimplified we actually pass a signed\n\nsimplified we actually pass a signed\n \n\nsimplified we actually pass a signed\ntoken to authenticate because it's\n\ntoken to authenticate because it's\n \n\ntoken to authenticate because it's\naccessing credentials but this is the\n\naccessing credentials but this is the\n \n\naccessing credentials but this is the\nbasic idea and then importantly this\n\nbasic idea and then importantly this\n \n\nbasic idea and then importantly this\ncall returns an MCP SDK object which\n\ncall returns an MCP SDK object which\n \n\ncall returns an MCP SDK object which\nmeans that when new features get added\n\nmeans that when new features get added\n \n\nmeans that when new features get added\nto protocol. You just update your MCP\n\nto protocol. You just update your MCP\n \n\nto protocol. You just update your MCP\npackages internally. You get those\n\npackages internally. You get those\n \n\npackages internally. You get those\nfeatures across the board. Everything\n\nfeatures across the board. Everything\n \n\nfeatures across the board. Everything\nworks great. The same code seamlessly\n\nworks great. The same code seamlessly\n \n\nworks great. The same code seamlessly\nconnects to internal external\n\nconnects to internal external\n \n\nconnects to internal external\nintegrations. integrations. integrations. When it comes to transports, uh, and\n\nWhen it comes to transports, uh, and\n \n\nWhen it comes to transports, uh, and\nthis is a bit high level and handwavy\n\nthis is a bit high level and handwavy\n \n\nthis is a bit high level and handwavy\nbecause everyone's setup is different. because everyone's setup is different. because everyone's setup is different. Um, internally within your network, it\n\nUm, internally within your network, it\n \n\nUm, internally within your network, it\nreally doesn't matter. You can do\n\nreally doesn't matter. You can do\n \n\nreally doesn't matter. You can do\nanything you want. We've got the\n\nanything you want. We've got the\n \n\nanything you want. We've got the\nstandardized transport for connecting to\n\nstandardized transport for connecting to\n \n\nstandardized transport for connecting to\nexternal MCP servers. Um, but really\n\nexternal MCP servers. Um, but really\n \n\nexternal MCP servers. Um, but really\njust picking the best thing for your\n\njust picking the best thing for your\n \n\njust picking the best thing for your\norg. So we went and picked\n\norg. So we went and picked\n \n\norg. So we went and picked\nuh websockets for our internal\n\nuh websockets for our internal\n \n\nuh websockets for our internal\ntransport. And here's just a quick code\n\ntransport. And here's just a quick code\n \n\ntransport. And here's just a quick code\nexample. It's nothing special. We just\n\nexample. It's nothing special. We just\n \n\nexample. It's nothing special. We just\nhave a websocket uh that's being opened. have a websocket uh that's being opened. have a websocket uh that's being opened. We are sending these JSON RPC blobs back\n\nWe are sending these JSON RPC blobs back\n \n\nWe are sending these JSON RPC blobs back\nand forth over the websocket. And then\n\nand forth over the websocket. And then\n \n\nand forth over the websocket. And then\nif I can make this scroll down\n\nif I can make this scroll down\n \n\nif I can make this scroll down\nat the end, we just pipe those uh read\n\nat the end, we just pipe those uh read\n \n\nat the end, we just pipe those uh read\nstreams and write streams into an MCPS\n\nstreams and write streams into an MCPS\n \n\nstreams and write streams into an MCPS\nSDK client session and we're good to go. SDK client session and we're good to go. SDK client session and we're good to go. We've got MCP going. Um you might want\n\nWe've got MCP going. Um you might want\n \n\nWe've got MCP going. Um you might want\nto do this with gRPC because you want to\n\nto do this with gRPC because you want to\n \n\nto do this with gRPC because you want to\nwrap these in some multipplexed\n\nwrap these in some multipplexed\n \n\nwrap these in some multipplexed\ntransport so you don't have to open one\n\ntransport so you don't have to open one\n \n\ntransport so you don't have to open one\nsoocket per connection. That's pretty\n\nsoocket per connection. That's pretty\n \n\nsoocket per connection. That's pretty\nsimple. Also, uh we have read stream\n\nsimple. Also, uh we have read stream\n \n\nsimple. Also, uh we have read stream\nright stream at the end. Uh starting to\n\nright stream at the end. Uh starting to\n \n\nright stream at the end. Uh starting to\nsee a pattern here. You can do like Unix\n\nsee a pattern here. You can do like Unix\n \n\nsee a pattern here. You can do like Unix\nsocket transport if you want. You can\n\nsocket transport if you want. You can\n \n\nsocket transport if you want. You can\njust have uh messages be passed that\n\njust have uh messages be passed that\n \n\njust have uh messages be passed that\nway. Read, stream, write, stream at the\n\nway. Read, stream, write, stream at the\n \n\nway. Read, stream, write, stream at the\nend. MCU works great. Um I threw in an\n\nend. MCU works great. Um I threw in an\n \n\nend. MCU works great. Um I threw in an\nenterprisegrade email transport\n\nenterprisegrade email transport\n \n\nenterprisegrade email transport\nimplementation over IMAP. Um which is\n\nimplementation over IMAP. Um which is\n \n\nimplementation over IMAP. Um which is\npretty much the same thing. You just go\n\npretty much the same thing. You just go\n \n\npretty much the same thing. You just go\nthrough here is our server. We're\n\nthrough here is our server. We're\n \n\nthrough here is our server. We're\nsending emails back and forth. Uh dear\n\nsending emails back and forth. Uh dear\n \n\nsending emails back and forth. Uh dear\nserver, I hope this finds you well. Uh\n\nserver, I hope this finds you well. Uh\n \n\nserver, I hope this finds you well. Uh\nMCP request start. And then we pipe\n\nMCP request start. And then we pipe\n \n\nMCP request start. And then we pipe\nthose into a client session at the end. those into a client session at the end. those into a client session at the end. And so it truly doesn't matter like\n\nAnd so it truly doesn't matter like\n \n\nAnd so it truly doesn't matter like\nwhatever it takes inside your\n\nwhatever it takes inside your\n \n\nwhatever it takes inside your\norganization is great. We set up this\n\norganization is great. We set up this\n \n\norganization is great." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nWe set up this\nunified authentication model where we're\n\nunified authentication model where we're\n \n\nunified authentication model where we're\nhandling\n\nhandling\n \n\nhandling\nooth at the gateway uh which means that\n\nooth at the gateway uh which means that\n \n\nooth at the gateway uh which means that\ncon..." }, { "role": "assistant", "content": "We set up this\nunified authentication model where we're\n\nunified authentication model where we're\n \n\nunified authentication model where we're\nhandling\n\nhandling\n \n\nhandling\nooth at the gateway uh which means that\n\nooth at the gateway uh which means that\n \n\nooth at the gateway uh which means that\nconsumers don't have to worry about all\n\nconsumers don't have to worry about all\n \n\nconsumers don't have to worry about all\nthat uh all that complexity in their\n\nthat uh all that complexity in their\n \n\nthat uh all that complexity in their\napps. Uh we added a get ooth\n\napps. Uh we added a get ooth\n \n\napps. Uh we added a get ooth\nauthorization URL function and a\n\nauthorization URL function and a\n \n\nauthorization URL function and a\ncomplete ooth flow because you might\n\ncomplete ooth flow because you might\n \n\ncomplete ooth flow because you might\nhave different endpoints at anthropic. have different endpoints at anthropic. have different endpoints at anthropic. We have api.anthropic.com and we have\n\nWe have api.anthropic.com and we have\n \n\nWe have api.anthropic.com and we have\ncloud.ai AI and we might want those\n\ncloud.ai AI and we might want those\n \n\ncloud.ai AI and we might want those\nredirects to go back to different\n\nredirects to go back to different\n \n\nredirects to go back to different\nplaces, but uh this is tied on the\n\nplaces, but uh this is tied on the\n \n\nplaces, but uh this is tied on the\ngateway. It's really easy to start a new\n\ngateway. It's really easy to start a new\n \n\ngateway. It's really easy to start a new\nauthentication. Uh a real advantage of\n\nauthentication. Uh a real advantage of\n \n\nauthentication. Uh a real advantage of\nhaving this put on your gateway is that\n\nhaving this put on your gateway is that\n \n\nhaving this put on your gateway is that\nthe credentials are portable. If you\n\nthe credentials are portable. If you\n \n\nthe credentials are portable. If you\nhave a batch job that you're kicking\n\nhave a batch job that you're kicking\n \n\nhave a batch job that you're kicking\noff, um your users don't have to\n\noff, um your users don't have to\n \n\noff, um your users don't have to\nreauthenticate to that. You're just\n\nreauthenticate to that. You're just\n \n\nreauthenticate to that. You're just\ncalling the same MCP with your internal\n\ncalling the same MCP with your internal\n \n\ncalling the same MCP with your internal\nuser ID and they get everything added\n\nuser ID and they get everything added\n \n\nuser ID and they get everything added\ncorrectly. You're also internal services\n\ncorrectly. You're also internal services\n \n\ncorrectly. You're also internal services\ndon't have to worry about your tokens. don't have to worry about your tokens. don't have to worry about your tokens. Um so your request comes in internally\n\nUm so your request comes in internally\n \n\nUm so your request comes in internally\nfor us. We're hitting a websocket\n\nfor us. We're hitting a websocket\n \n\nfor us. We're hitting a websocket\nconnection to MCP gateway uh with O\n\nconnection to MCP gateway uh with O\n \n\nconnection to MCP gateway uh with O\ntoken provided as headers to that. Uh\n\ntoken provided as headers to that. Uh\n \n\ntoken provided as headers to that. Uh\nthe gateway receives your stored\n\nthe gateway receives your stored\n \n\nthe gateway receives your stored\ncredentials. You create an authenticated\n\ncredentials. You create an authenticated\n \n\ncredentials. You create an authenticated\nSDK client. You just pass in the bearer\n\nSDK client. You just pass in the bearer\n \n\nSDK client. You just pass in the bearer\ntoken to the O header. Uh and then\n\ntoken to the O header. Uh and then\n \n\ntoken to the O header. Uh and then\nyou're good to go. The MCP client\n\nyou're good to go. The MCP client\n \n\nyou're good to go. The MCP client\nreceives a readstream and a right\n\nreceives a readstream and a right\n \n\nreceives a readstream and a right\nstream. And so you just plum those read\n\nstream. And so you just plum those read\n \n\nstream. And so you just plum those read\nstream and write stream into your\n\nstream and write stream into your\n \n\nstream and write stream into your\ninternal transport and you're and you're\n\ninternal transport and you're and you're\n \n\ninternal transport and you're and you're\nyou're good to go. Uh one of the things\n\nyou're good to go. Uh one of the things\n \n\nyou're good to go. Uh one of the things\nthat this gives for your org that's not\n\nthat this gives for your org that's not\n \n\nthat this gives for your org that's not\nimmediately obvious but is really\n\nimmediately obvious but is really\n \n\nimmediately obvious but is really\nvaluable is a central place for all of\n\nvaluable is a central place for all of\n \n\nvaluable is a central place for all of\nyour context that your models are asking\n\nyour context that your models are asking\n \n\nyour context that your models are asking\nfor and all the context that's flowing\n\nfor and all the context that's flowing\n \n\nfor and all the context that's flowing\ninto your models for your org. Uh\n\ninto your models for your org. Uh\n \n\ninto your models for your org. Uh\nthere's some papers written on MCP\n\nthere's some papers written on MCP\n \n\nthere's some papers written on MCP\nprompt injection attacks. There is a\n\nprompt injection attacks. There is a\n \n\nprompt injection attacks. There is a\ngeneral risk of uh models going and\n\ngeneral risk of uh models going and\n \n\ngeneral risk of uh models going and\nhaving access to Google Drive and\n\nhaving access to Google Drive and\n \n\nhaving access to Google Drive and\ndeleting everything in Google Drive. deleting everything in Google Drive. deleting everything in Google Drive. There's some need of uh enforcing\n\nThere's some need of uh enforcing\n \n\nThere's some need of uh enforcing\npolicy. You might want to be able to\n\npolicy. You might want to be able to\n \n\npolicy. You might want to be able to\nlike ban\n\nlike ban\n \n\nlike ban\nmalicious servers. Um do some content\n\nmalicious servers. Um do some content\n \n\nmalicious servers. Um do some content\nclassification on the request, see\n\nclassification on the request, see\n \n\nclassification on the request, see\nwhat's coming in, kind of given audit. what's coming in, kind of given audit. what's coming in, kind of given audit. And the really nice thing about this is\n\nAnd the really nice thing about this is\n \n\nAnd the really nice thing about this is\nthat because it's MCP, all of your\n\nthat because it's MCP, all of your\n \n\nthat because it's MCP, all of your\nmessages are in a standardized format. messages are in a standardized format. messages are in a standardized format. So it's really easy to hook into that\n\nSo it's really easy to hook into that\n \n\nSo it's really easy to hook into that\nstream and be like, \"Okay, here is my\n\nstream and be like, \"Okay, here is my\n \n\nstream and be like, \"Okay, here is my\ntool execution message processor or here\n\ntool execution message processor or here\n \n\ntool execution message processor or here\nis my tool definition thing or here's my\n\nis my tool definition thing or here's my\n \n\nis my tool definition thing or here's my\nresource management.\" And so the payoff\n\nresource management.\" And so the payoff\n \n\nresource management.\" And so the payoff\nthat you get from this is um adding MCP\n\nthat you get from this is um adding MCP\n \n\nthat you get from this is um adding MCP\nsupport to new services is as simple as\n\nsupport to new services is as simple as\n \n\nsupport to new services is as simple as\npossible. You just go and import a\n\npossible. You just go and import a\n \n\npossible. You just go and import a\npackage. Uh it doesn't matter what\n\npackage. Uh it doesn't matter what\n \n\npackage. Uh it doesn't matter what\nlanguage you're in. We've got multiple\n\nlanguage you're in. We've got multiple\n \n\nlanguage you're in. We've got multiple\nlanguages internally. They all have\n\nlanguages internally. They all have\n \n\nlanguages internally. They all have\ntheir own kind of packages. Engineers\n\ntheir own kind of packages. Engineers\n \n\ntheir own kind of packages. Engineers\ncan focus on building features and not\n\ncan focus on building features and not\n \n\ncan focus on building features and not\nplumbing. uh you have the operational\n\nplumbing. uh you have the operational\n \n\nplumbing." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nuh you have the operational\nsimplicity of having a single point of\n\nsimplicity of having a single point of\n \n\nsimplicity of having a single point of\ningress egress and standardized message\n\ningress egress and standardized message\n \n\ningress egress and standardized message\nformats and you get future ..." }, { "role": "assistant", "content": "uh you have the operational\nsimplicity of having a single point of\n\nsimplicity of having a single point of\n \n\nsimplicity of having a single point of\ningress egress and standardized message\n\ningress egress and standardized message\n \n\ningress egress and standardized message\nformats and you get future features for\n\nformats and you get future features for\n \n\nformats and you get future features for\nfree as the protocol involves you get\n\nfree as the protocol involves you get\n \n\nfree as the protocol involves you get\nall of that work uh naturally\n\nall of that work uh naturally\n \n\nall of that work uh naturally\nand so just wanted to go through some\n\nand so just wanted to go through some\n \n\nand so just wanted to go through some\ntakeaways from this that I I want to put\n\ntakeaways from this that I I want to put\n \n\ntakeaways from this that I I want to put\nto you is that MCP is really just JSON\n\nto you is that MCP is really just JSON\n \n\nto you is that MCP is really just JSON\nstreams and how you pipe those streams\n\nstreams and how you pipe those streams\n \n\nstreams and how you pipe those streams\naround your infrastructure is a small\n\naround your infrastructure is a small\n \n\naround your infrastructure is a small\nimplementation detail. That's a couple\n\nimplementation detail. That's a couple\n \n\nimplementation detail. That's a couple\nlines of code hook the stream into the\n\nlines of code hook the stream into the\n \n\nlines of code hook the stream into the\nclient SDK that makes the messages. Uh\n\nclient SDK that makes the messages. Uh\n \n\nclient SDK that makes the messages. Uh\nthe\n\nthe\n \n\nthe\nyou should standardize on something,\n\nyou should standardize on something,\n \n\nyou should standardize on something,\nanything. I think MCP is a good idea. If\n\nanything. I think MCP is a good idea. If\n \n\nanything. I think MCP is a good idea. If\nyou don't think it's a good idea, like\n\nyou don't think it's a good idea, like\n \n\nyou don't think it's a good idea, like\njust pick something. Um your future self\n\njust pick something. Um your future self\n \n\njust pick something. Um your future self\nwill thank you. Uh build some pits of\n\nwill thank you. Uh build some pits of\n \n\nwill thank you. Uh build some pits of\nsuccess. You really want to make the\n\nsuccess. You really want to make the\n \n\nsuccess. You really want to make the\nright way to do a thing the easiest way\n\nright way to do a thing the easiest way\n \n\nright way to do a thing the easiest way\nto do a thing and then everyone just\n\nto do a thing and then everyone just\n \n\nto do a thing and then everyone just\nfalls into doing the right thing\n\nfalls into doing the right thing\n \n\nfalls into doing the right thing\nnaturally. And also centralizing at the\n\nnaturally. And also centralizing at the\n \n\nnaturally. And also centralizing at the\ncorrect layer. So solving some shared\n\ncorrect layer. So solving some shared\n \n\ncorrect layer. So solving some shared\nproblems off and external connectivity\n\nproblems off and external connectivity\n \n\nproblems off and external connectivity\nonce allows you to spend your time\n\nonce allows you to spend your time\n \n\nonce allows you to spend your time\nworking on uh more interesting problems\n\nworking on uh more interesting problems\n \n\nworking on uh more interesting problems\nthat are more valuable to you and your\n\nthat are more valuable to you and your\n \n\nthat are more valuable to you and your\nyour business. your business. your business. Thanks. That's all I got for you. Uh\n\nThanks. That's all I got for you. Uh\n \n\nThanks. That's all I got for you. Uh\nthank you so much for coming out. thank you so much for coming out. thank you so much for coming out. [Music]" } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nHi, my name is Rick from Agentuity and\n\nHi, my name is Rick from Agentuity and\n \n\nHi, my name is Rick from Agentuity and\ntoday I'm going to talk to you about\n\ntoday I'm going to talk to you about\n \n\ntoday I'm going to talk to you about\nagent native companies or AI native\n\nagent native companies or A..." }, { "role": "assistant", "content": "Hi, my name is Rick from Agentuity and\n\nHi, my name is Rick from Agentuity and\n \n\nHi, my name is Rick from Agentuity and\ntoday I'm going to talk to you about\n\ntoday I'm going to talk to you about\n \n\ntoday I'm going to talk to you about\nagent native companies or AI native\n\nagent native companies or AI native\n \n\nagent native companies or AI native\ncompanies. What that means you hear a\n\ncompanies. What that means you hear a\n \n\ncompanies. What that means you hear a\nlot of that right now. What does it mean\n\nlot of that right now. What does it mean\n \n\nlot of that right now. What does it mean\nand uh what are some of the core\n\nand uh what are some of the core\n \n\nand uh what are some of the core\nattributes and things like that? So, so\n\nattributes and things like that? So, so\n \n\nattributes and things like that? So, so\nimagine if you will, you walk into the\n\nimagine if you will, you walk into the\n \n\nimagine if you will, you walk into the\noffice, past your co-workers, smile away\n\noffice, past your co-workers, smile away\n \n\noffice, past your co-workers, smile away\nwith them, you sit at your desk, you sit\n\nwith them, you sit at your desk, you sit\n \n\nwith them, you sit at your desk, you sit\ndown, and you check not your email, but\n\ndown, and you check not your email, but\n \n\ndown, and you check not your email, but\nthe status of what all your agent\n\nthe status of what all your agent\n \n\nthe status of what all your agent\nco-workers did over the the night\n\nco-workers did over the the night\n \n\nco-workers did over the the night\nbefore, did last night. Uh, they're\n\nbefore, did last night. Uh, they're\n \n\nbefore, did last night. Uh, they're\npulling you into things that need your\n\npulling you into things that need your\n \n\npulling you into things that need your\nattention, things that they got done or\n\nattention, things that they got done or\n \n\nattention, things that they got done or\nthings that they think need approval. things that they think need approval. things that they think need approval. Um, and so instantly you get brought\n\nUm, and so instantly you get brought\n \n\nUm, and so instantly you get brought\ninto their world on, hey, I need to go\n\ninto their world on, hey, I need to go\n \n\ninto their world on, hey, I need to go\ndo and execute these things because it's\n\ndo and execute these things because it's\n \n\ndo and execute these things because it's\npart of my job. So for a lot of us, I\n\npart of my job. So for a lot of us, I\n \n\npart of my job. So for a lot of us, I\ndon't think it's actually hard to\n\ndon't think it's actually hard to\n \n\ndon't think it's actually hard to\nimagine, especially for us that are\n\nimagine, especially for us that are\n \n\nimagine, especially for us that are\ndoing a lot of these AI things right\n\ndoing a lot of these AI things right\n \n\ndoing a lot of these AI things right\nnow. Uh, certainly at our startup at\n\nnow. Uh, certainly at our startup at\n \n\nnow. Uh, certainly at our startup at\nAgentuity, uh, we're feeling exactly\n\nAgentuity, uh, we're feeling exactly\n \n\nAgentuity, uh, we're feeling exactly\nthat during a team meeting, it's really\n\nthat during a team meeting, it's really\n \n\nthat during a team meeting, it's really\ncommon to hear things like, well, just\n\ncommon to hear things like, well, just\n \n\ncommon to hear things like, well, just\nhave Devon go do that. Oh, that's a good\n\nhave Devon go do that. Oh, that's a good\n \n\nhave Devon go do that. Oh, that's a good\npoint. We need to go update that. I went\n\npoint. We need to go update that. I went\n \n\npoint. We need to go update that. I went\nahead and pinged Devon. and then by the\n\nahead and pinged Devon. and then by the\n \n\nahead and pinged Devon. and then by the\ntime we're done with the standup, I've\n\ntime we're done with the standup, I've\n \n\ntime we're done with the standup, I've\nalready got a pull request that we have\n\nalready got a pull request that we have\n \n\nalready got a pull request that we have\nto review. Or things like, hey, we\n\nto review. Or things like, hey, we\n \n\nto review. Or things like, hey, we\nshould tweak this channel, um, let's go\n\nshould tweak this channel, um, let's go\n \n\nshould tweak this channel, um, let's go\nping the content marketing agent and\n\nping the content marketing agent and\n \n\nping the content marketing agent and\nit'll it'll adjust it. That kind of\n\nit'll it'll adjust it. That kind of\n \n\nit'll it'll adjust it. That kind of\nthing. So, um, those things are are\n\nthing. So, um, those things are are\n \n\nthing. So, um, those things are are\ncommon more and more and more,\n\ncommon more and more and more,\n \n\ncommon more and more and more,\nespecially in Agentuity. And we think\n\nespecially in Agentuity. And we think\n \n\nespecially in Agentuity. And we think\nthat for an agent native or an AI native\n\nthat for an agent native or an AI native\n \n\nthat for an agent native or an AI native\ncompany, some of those things are just\n\ncompany, some of those things are just\n \n\ncompany, some of those things are just\ngoing to be the norm. I mean even as\n\ngoing to be the norm. I mean even as\n \n\ngoing to be the norm. I mean even as\nexample I was on X recently and a\n\nexample I was on X recently and a\n \n\nexample I was on X recently and a\nstartup posted a job opening for an\n\nstartup posted a job opening for an\n \n\nstartup posted a job opening for an\nagent manager that's a manager that\n\nagent manager that's a manager that\n \n\nagent manager that's a manager that\nmanages AI workers pretty interesting\n\nmanages AI workers pretty interesting\n \n\nmanages AI workers pretty interesting\nright so we're entering a world where\n\nright so we're entering a world where\n \n\nright so we're entering a world where\nmanaging agents can be someone's\n\nmanaging agents can be someone's\n \n\nmanaging agents can be someone's\nfull-time job\n\nfull-time job\n \n\nfull-time job\nliterally so in this talk I want to\n\nliterally so in this talk I want to\n \n\nliterally so in this talk I want to\nexplore that world agent native AI\n\nexplore that world agent native AI\n \n\nexplore that world agent native AI\nnative companies where AI just isn't an\n\nnative companies where AI just isn't an\n \n\nnative companies where AI just isn't an\nadd-on or just some just some task that\n\nadd-on or just some just some task that\n \n\nadd-on or just some just some task that\nsomebody uses but central to how work\n\nsomebody uses but central to how work\n \n\nsomebody uses but central to how work\ngets done. So, we're going to walk\n\ngets done. So, we're going to walk\n \n\ngets done. So, we're going to walk\nthrough a few things. We're going to\n\nthrough a few things. We're going to\n \n\nthrough a few things. We're going to\nwalk through AI enhanced versus agent\n\nwalk through AI enhanced versus agent\n \n\nwalk through AI enhanced versus agent\nnatives. There's a difference or AI\n\nnatives. There's a difference or AI\n \n\nnatives. There's a difference or AI\nnatives. Um what makes a company agentic\n\nnatives. Um what makes a company agentic\n \n\nnatives. Um what makes a company agentic\num the typical agentic workday and uh\n\num the typical agentic workday and uh\n \n\num the typical agentic workday and uh\neven just like rethink hiring like maybe\n\neven just like rethink hiring like maybe\n \n\neven just like rethink hiring like maybe\nwe need to think about hiring a little\n\nwe need to think about hiring a little\n \n\nwe need to think about hiring a little\ndifferent. So, that's what we're going\n\ndifferent. So, that's what we're going\n \n\ndifferent. So, that's what we're going\nto cover in this presentation. So, let's\n\nto cover in this presentation. So, let's\n \n\nto cover in this presentation. So, let's\njust jump right into it. First thing is\n\njust jump right into it. First thing is\n \n\njust jump right into it. First thing is\nAI enhanced versus AI native or agent\n\nAI enhanced versus AI native or agent\n \n\nAI enhanced versus AI native or agent\nnative. So what does that mean? Uh\n\nnative. So what does that mean? Uh\n \n\nnative. So what does that mean? Uh\nanyway, what is a native company that\n\nanyway, what is a native company that\n \n\nanyway, what is a native company that\nuses AI? What does that mean? So\n\nuses AI? What does that mean? So\n \n\nuses AI? What does that mean?" } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nSo\nhopefully you're not one of those people\n\nhopefully you're not one of those people\n \n\nhopefully you're not one of those people\nthat think that an agent is only a\n\nthat think that an agent is only a\n \n\nthat think that an agent is only a\nchatbot. If you are, then you're\n\nchatbot. If you are, then y..." }, { "role": "assistant", "content": "So\nhopefully you're not one of those people\n\nhopefully you're not one of those people\n \n\nhopefully you're not one of those people\nthat think that an agent is only a\n\nthat think that an agent is only a\n \n\nthat think that an agent is only a\nchatbot. If you are, then you're\n\nchatbot. If you are, then you're\n \n\nchatbot. If you are, then you're\nprobably thinking, well, why would you\n\nprobably thinking, well, why would you\n \n\nprobably thinking, well, why would you\nwhat does a chat native company mean? So\n\nwhat does a chat native company mean? So\n \n\nwhat does a chat native company mean? So\nif you are one of those, then we\n\nif you are one of those, then we\n \n\nif you are one of those, then we\nprobably need to make another video um\n\nprobably need to make another video um\n \n\nprobably need to make another video um\nto talk through that. My my intent is\n\nto talk through that. My my intent is\n \n\nto talk through that. My my intent is\nnot to talk and convince you what an\n\nnot to talk and convince you what an\n \n\nnot to talk and convince you what an\nagent is. Um it's not just a chatbot,\n\nagent is. Um it's not just a chatbot,\n \n\nagent is. Um it's not just a chatbot,\nbut in a nutshell, what we mean when we\n\nbut in a nutshell, what we mean when we\n \n\nbut in a nutshell, what we mean when we\nsay agent native um company is a company\n\nsay agent native um company is a company\n \n\nsay agent native um company is a company\nbuilt from the ground up with AI agents\n\nbuilt from the ground up with AI agents\n \n\nbuilt from the ground up with AI agents\nat the core of everything to augment\n\nat the core of everything to augment\n \n\nat the core of everything to augment\nhuman productivity and intelligence. Um,\n\nhuman productivity and intelligence. Um,\n \n\nhuman productivity and intelligence. Um,\nI'd also additionally add that um, these\n\nI'd also additionally add that um, these\n \n\nI'd also additionally add that um, these\ncompanies build AI into the foundation\n\ncompanies build AI into the foundation\n \n\ncompanies build AI into the foundation\nof their product, into the foundation of\n\nof their product, into the foundation of\n \n\nof their product, into the foundation of\ntheir operations, and into the\n\ntheir operations, and into the\n \n\ntheir operations, and into the\nfoundation of their culture. Another way\n\nfoundation of their culture. Another way\n \n\nfoundation of their culture. Another way\nto look at it is AI is not a fancy side\n\nto look at it is AI is not a fancy side\n \n\nto look at it is AI is not a fancy side\nfeature that's just a bolt-on um, to\n\nfeature that's just a bolt-on um, to\n \n\nfeature that's just a bolt-on um, to\nexisting operation and culture. It's an\n\nexisting operation and culture. It's an\n \n\nexisting operation and culture. It's an\nengine that moves the product operation\n\nengine that moves the product operation\n \n\nengine that moves the product operation\nto culture. It's the thing that moves it\n\nto culture. It's the thing that moves it\n \n\nto culture. It's the thing that moves it\nforward. Everybody reaches for that. forward. Everybody reaches for that. forward. Everybody reaches for that. everybody relies on it uh to do their\n\neverybody relies on it uh to do their\n \n\neverybody relies on it uh to do their\njob. So sometimes it's helpful to think\n\njob. So sometimes it's helpful to think\n \n\njob. So sometimes it's helpful to think\nabout this in the opposite to contrast. about this in the opposite to contrast. about this in the opposite to contrast. So for these companies if you\n\nSo for these companies if you\n \n\nSo for these companies if you\nremove agents just for us for our\n\nremove agents just for us for our\n \n\nremove agents just for us for our\ncompany agent if you just remove agents\n\ncompany agent if you just remove agents\n \n\ncompany agent if you just remove agents\nout of our workflow\n\nout of our workflow\n \n\nout of our workflow\num and guess what happens our employees\n\num and guess what happens our employees\n \n\num and guess what happens our employees\naren't able to get as much done. They're\n\naren't able to get as much done. They're\n \n\naren't able to get as much done. They're\njust\n\njust\n \n\njust\nnot a lot of the mundane things that we\n\nnot a lot of the mundane things that we\n \n\nnot a lot of the mundane things that we\ndo unsatisfactory work that we do we\n\ndo unsatisfactory work that we do we\n \n\ndo unsatisfactory work that we do we\nhave to now do. And um it it does it's\n\nhave to now do. And um it it does it's\n \n\nhave to now do. And um it it does it's\nnot as fulfilling. I mean just even for\n\nnot as fulfilling. I mean just even for\n \n\nnot as fulfilling. I mean just even for\nus we have an agent that uh writes our\n\nus we have an agent that uh writes our\n \n\nus we have an agent that uh writes our\nchange log and our documentation. So\n\nchange log and our documentation. So\n \n\nchange log and our documentation. So\nafter all this stuff happens in GitHub,\n\nafter all this stuff happens in GitHub,\n \n\nafter all this stuff happens in GitHub,\nall this code gets written, all these\n\nall this code gets written, all these\n \n\nall this code gets written, all these\nthings get deployed and released. Um we\n\nthings get deployed and released. Um we\n \n\nthings get deployed and released. Um we\nhave to create a change log. We have to\n\nhave to create a change log. We have to\n \n\nhave to create a change log. We have to\ncreate documentation. We don't do that. create documentation. We don't do that. create documentation. We don't do that. Uh we have an agent do that. And so if\n\nUh we have an agent do that. And so if\n \n\nUh we have an agent do that. And so if\nif you were to remove that from us, we\n\nif you were to remove that from us, we\n \n\nif you were to remove that from us, we\nwould we'd be very sad. All of us as NG\n\nwould we'd be very sad. All of us as NG\n \n\nwould we'd be very sad. All of us as NG\nsoftware engineers would be very sad if\n\nsoftware engineers would be very sad if\n \n\nsoftware engineers would be very sad if\nthat got removed from us. Um, another\n\nthat got removed from us. Um, another\n \n\nthat got removed from us. Um, another\none is if you think about it, removing\n\none is if you think about it, removing\n \n\none is if you think about it, removing\nAI or agents from a agent native\n\nAI or agents from a agent native\n \n\nAI or agents from a agent native\ncompany,\n\ncompany,\n \n\ncompany,\num, makes us not move as fast. The costs\n\num, makes us not move as fast. The costs\n \n\num, makes us not move as fast. The costs\ngo up, pro productivity goes down. go up, pro productivity goes down. go up, pro productivity goes down. Removing AI and agents um, from the\n\nRemoving AI and agents um, from the\n \n\nRemoving AI and agents um, from the\nproducts that we're building. Products\n\nproducts that we're building. Products\n \n\nproducts that we're building. Products\nstart to feel old, not as interesting,\n\nstart to feel old, not as interesting,\n \n\nstart to feel old, not as interesting,\nunintelligent, not as useful. They\n\nunintelligent, not as useful. They\n \n\nunintelligent, not as useful. They\ncertainly don't help the customer 10x\n\ncertainly don't help the customer 10x\n \n\ncertainly don't help the customer 10x\nthemselves. We're want to 10x or 100x\n\nthemselves. We're want to 10x or 100x\n \n\nthemselves. We're want to 10x or 100x\nourselves. How do we do that with our\n\nourselves. How do we do that with our\n \n\nourselves. How do we do that with our\ncustomers and our product? Right? Um, so\n\ncustomers and our product? Right? Um, so\n \n\ncustomers and our product? Right? Um, so\nthen all of a sudden now we're not as\n\nthen all of a sudden now we're not as\n \n\nthen all of a sudden now we're not as\ngood as the competitors. The list goes\n\ngood as the competitors. The list goes\n \n\ngood as the competitors. The list goes\non and on and on. But that's the\n\non and on and on. But that's the\n \n\non and on and on. But that's the\ncounterpoint to it. A way to state it is\n\ncounterpoint to it. A way to state it is\n \n\ncounterpoint to it. A way to state it is\nmerely an AI enhanced business. One that\n\nmerely an AI enhanced business. One that\n \n\nmerely an AI enhanced business." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nOne that\nuses AI here and there, chats with it\n\nuses AI here and there, chats with it\n \n\nuses AI here and there, chats with it\nevery once in a while, gets a document\n\nevery once in a while, gets a document\n \n\nevery once in a while, gets a document\nevery once in a while. Maybe that helps\n\nevery once ..." }, { "role": "assistant", "content": "One that\nuses AI here and there, chats with it\n\nuses AI here and there, chats with it\n \n\nuses AI here and there, chats with it\nevery once in a while, gets a document\n\nevery once in a while, gets a document\n \n\nevery once in a while, gets a document\nevery once in a while. Maybe that helps\n\nevery once in a while. Maybe that helps\n \n\nevery once in a while. Maybe that helps\nwith some efficiency goals. It's good. with some efficiency goals. It's good. with some efficiency goals. It's good. And they would still function without\n\nAnd they would still function without\n \n\nAnd they would still function without\nAI. That's the key part. If you removed\n\nAI. That's the key part. If you removed\n \n\nAI. That's the key part. If you removed\nit from it, it would just they would\n\nit from it, it would just they would\n \n\nit from it, it would just they would\nstill function. It would just not be as\n\nstill function. It would just not be as\n \n\nstill function. It would just not be as\nefficient. And that's a car with driver\n\nefficient. And that's a car with driver\n \n\nefficient. And that's a car with driver\nassist. That's really what we're talking\n\nassist. That's really what we're talking\n \n\nassist. That's really what we're talking\nabout. An Asian native business is\n\nabout. An Asian native business is\n \n\nabout. An Asian native business is\ndifferent. Isn't the car on autopilot\n\ndifferent. Isn't the car on autopilot\n \n\ndifferent. Isn't the car on autopilot\ndirected by humans of course, but it's\n\ndirected by humans of course, but it's\n \n\ndirected by humans of course, but it's\nstill autopilot. It's a different kind\n\nstill autopilot. It's a different kind\n \n\nstill autopilot. It's a different kind\nof car. Its signature is every employee\n\nof car. Its signature is every employee\n \n\nof car. Its signature is every employee\nis focused on the higher level\n\nis focused on the higher level\n \n\nis focused on the higher level\nnavigation task and the success of the\n\nnavigation task and the success of the\n \n\nnavigation task and the success of the\ncompany and the product. While the\n\ncompany and the product. While the\n \n\ncompany and the product. While the\nroutine and mundane tasks, micro\n\nroutine and mundane tasks, micro\n \n\nroutine and mundane tasks, micro\ndecisions, they get offloaded. That's a\n\ndecisions, they get offloaded. That's a\n \n\ndecisions, they get offloaded. That's a\nsignature part of it. So the thesis that\n\nsignature part of it. So the thesis that\n \n\nsignature part of it. So the thesis that\nwe have is that an AI native or agent\n\nwe have is that an AI native or agent\n \n\nwe have is that an AI native or agent\nnative model isn't just a tech trend. native model isn't just a tech trend. native model isn't just a tech trend. It's it's redefining how we build teams,\n\nIt's it's redefining how we build teams,\n \n\nIt's it's redefining how we build teams,\nhow we design workflows, even what roles\n\nhow we design workflows, even what roles\n \n\nhow we design workflows, even what roles\nwe hire for. I don't think it's any\n\nwe hire for. I don't think it's any\n \n\nwe hire for. I don't think it's any\ndifferent than back in the industrial\n\ndifferent than back in the industrial\n \n\ndifferent than back in the industrial\nrevolution and during the car age and\n\nrevolution and during the car age and\n \n\nrevolution and during the car age and\nsome of those things, how everything got\n\nsome of those things, how everything got\n \n\nsome of those things, how everything got\ntransformed. I think it's just like\n\ntransformed. I think it's just like\n \n\ntransformed. I think it's just like\nthat. So now moving on, uh what makes a\n\nthat. So now moving on, uh what makes a\n \n\nthat. So now moving on, uh what makes a\ncompany agent native? Now, obviously,\n\ncompany agent native? Now, obviously,\n \n\ncompany agent native? Now, obviously,\nwe're on the early days, very early\n\nwe're on the early days, very early\n \n\nwe're on the early days, very early\ndays, of course, but here's a few\n\ndays, of course, but here's a few\n \n\ndays, of course, but here's a few\ndefining things that make an agent\n\ndefining things that make an agent\n \n\ndefining things that make an agent\nnative company. AI is at the center of\n\nnative company. AI is at the center of\n \n\nnative company. AI is at the center of\neverything. When you say, Rick, yeah,\n\neverything. When you say, Rick, yeah,\n \n\neverything. When you say, Rick, yeah,\nno, duh. I just listen know it's it's\n\nno, duh. I just listen know it's it's\n \n\nno, duh. I just listen know it's it's\nnot confined to just one team or a\n\nnot confined to just one team or a\n \n\nnot confined to just one team or a\nfeature or one aspect of a company's\n\nfeature or one aspect of a company's\n \n\nfeature or one aspect of a company's\nculture. It's everywhere and it's\n\nculture. It's everywhere and it's\n \n\nculture. It's everywhere and it's\neverything. Think of product, think of\n\neverything. Think of product, think of\n \n\neverything. Think of product, think of\ncustomer support, think of ops. In an AI\n\ncustomer support, think of ops. In an AI\n \n\ncustomer support, think of ops. In an AI\nnative company, the expectation is each\n\nnative company, the expectation is each\n \n\nnative company, the expectation is each\nof these departments have agents doing\n\nof these departments have agents doing\n \n\nof these departments have agents doing\nsome of the routine and key daily work\n\nsome of the routine and key daily work\n \n\nsome of the routine and key daily work\nall the time. The departments would have\n\nall the time. The departments would have\n \n\nall the time. The departments would have\nagent interfaces handoff. They integrate\n\nagent interfaces handoff. They integrate\n \n\nagent interfaces handoff. They integrate\nwith other departments. I mean this is\n\nwith other departments. I mean this is\n \n\nwith other departments. I mean this is\nobvious, right? In order to be\n\nobvious, right? In order to be\n \n\nobvious, right? In order to be\nefficient, you need this. It should be\n\nefficient, you need this. It should be\n \n\nefficient, you need this. It should be\nobvious. But again, think of the\n\nobvious. But again, think of the\n \n\nobvious. But again, think of the\nopposite. If you turn that off, each\n\nopposite. If you turn that off, each\n \n\nopposite. If you turn that off, each\ndepartment, uh, if you turn turn off all\n\ndepartment, uh, if you turn turn off all\n \n\ndepartment, uh, if you turn turn off all\ntheir agents, um, you know, you got a\n\ntheir agents, um, you know, you got a\n \n\ntheir agents, um, you know, you got a\nhuman scramble trying to coordinate\n\nhuman scramble trying to coordinate\n \n\nhuman scramble trying to coordinate\nmanually, feeling a loss of productivity\n\nmanually, feeling a loss of productivity\n \n\nmanually, feeling a loss of productivity\nbecause you had this thing, you had this\n\nbecause you had this thing, you had this\n \n\nbecause you had this thing, you had this\nway to automate things and to work\n\nway to automate things and to work\n \n\nway to automate things and to work\nfaster and to work more, and now you\n\nfaster and to work more, and now you\n \n\nfaster and to work more, and now you\ndon't have it anymore. Um, man, that's a\n\ndon't have it anymore. Um, man, that's a\n \n\ndon't have it anymore. Um, man, that's a\nbig big big uh loss uh for for us. Like\n\nbig big big uh loss uh for for us. Like\n \n\nbig big big uh loss uh for for us. Like\njust thinking about that is painful\n\njust thinking about that is painful\n \n\njust thinking about that is painful\nright now. We have six people in our\n\nright now. We have six people in our\n \n\nright now. We have six people in our\ncompany, seven people, and um, removing\n\ncompany, seven people, and um, removing\n \n\ncompany, seven people, and um, removing\nthat would be would be bad. that would be would be bad. that would be would be bad. Um, the other thing is with an AI or an\n\nUm, the other thing is with an AI or an\n \n\nUm, the other thing is with an AI or an\nagent native company, people are no\n\nagent native company, people are no\n \n\nagent native company, people are no\nlonger just like cogs in a machine. longer just like cogs in a machine. longer just like cogs in a machine. That's another way to look at it. That's another way to look at it. That's another way to look at it. They're more like conductors. Now, if\n\nThey're more like conductors. Now, if\n \n\nThey're more like conductors." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nNow, if\nyou think about that statement for a\n\nyou think about that statement for a\n \n\nyou think about that statement for a\nminute, a lot of people make that\n\nminute, a lot of people make that\n \n\nminute, a lot of people make that\nstatement. It's not a novel statement. statement. It's not a novel stat..." }, { "role": "assistant", "content": "Now, if\nyou think about that statement for a\n\nyou think about that statement for a\n \n\nyou think about that statement for a\nminute, a lot of people make that\n\nminute, a lot of people make that\n \n\nminute, a lot of people make that\nstatement. It's not a novel statement. statement. It's not a novel statement. statement. It's not a novel statement. It's a it is an important statement,\n\nIt's a it is an important statement,\n \n\nIt's a it is an important statement,\nthough. They're\n\nthough. They're\n \n\nthough. They're\nconductors. You'll realize that the\n\nconductors. You'll realize that the\n \n\nconductors. You'll realize that the\nhiring\n\nhiring\n \n\nhiring\nprofile and the company or chart will\n\nprofile and the company or chart will\n \n\nprofile and the company or chart will\nneed to be different if that's who\n\nneed to be different if that's who\n \n\nneed to be different if that's who\nyou're hiring. If that's what you need,\n\nyou're hiring. If that's what you need,\n \n\nyou're hiring. If that's what you need,\nflatter leaner is a key attribute. It\n\nflatter leaner is a key attribute. It\n \n\nflatter leaner is a key attribute. It\nstarts to become it can be a key\n\nstarts to become it can be a key\n \n\nstarts to become it can be a key\nattribute. Middle management layers\n\nattribute. Middle management layers\n \n\nattribute. Middle management layers\nshrink because a lot of that\n\nshrink because a lot of that\n \n\nshrink because a lot of that\ncoordination can be handled and executed\n\ncoordination can be handled and executed\n \n\ncoordination can be handled and executed\nby intelligent systems. I mean even in\n\nby intelligent systems. I mean even in\n \n\nby intelligent systems. I mean even in\nmy own experience, we'll have a deep\n\nmy own experience, we'll have a deep\n \n\nmy own experience, we'll have a deep\ndive product discussion in the morning\n\ndive product discussion in the morning\n \n\ndive product discussion in the morning\nwith the team. By the end of the day, we\n\nwith the team. By the end of the day, we\n \n\nwith the team. By the end of the day, we\nhave detailed requirements um to work on\n\nhave detailed requirements um to work on\n \n\nhave detailed requirements um to work on\nsomething. We already have had agents\n\nsomething. We already have had agents\n \n\nsomething. We already have had agents\nstart to build it. has already helped us\n\nstart to build it. has already helped us\n \n\nstart to build it. has already helped us\nwith the messaging and the copy and some\n\nwith the messaging and the copy and some\n \n\nwith the messaging and the copy and some\nof the documentation and that's just\n\nof the documentation and that's just\n \n\nof the documentation and that's just\nwithin you know a day and so we've had\n\nwithin you know a day and so we've had\n \n\nwithin you know a day and so we've had\nthis culture now of like you well yeah\n\nthis culture now of like you well yeah\n \n\nthis culture now of like you well yeah\nthis is a great idea let's prototype it\n\nthis is a great idea let's prototype it\n \n\nthis is a great idea let's prototype it\nand we prototype it in a few hours and\n\nand we prototype it in a few hours and\n \n\nand we prototype it in a few hours and\nwe get some more refinement and agents\n\nwe get some more refinement and agents\n \n\nwe get some more refinement and agents\nhelp with that a lot help us do more\n\nhelp with that a lot help us do more\n \n\nhelp with that a lot help us do more\nprototypes and more learning and more\n\nprototypes and more learning and more\n \n\nprototypes and more learning and more\ntesting that way so that's why we think\n\ntesting that way so that's why we think\n \n\ntesting that way so that's why we think\nat least our org chart in the near\n\nat least our org chart in the near\n \n\nat least our org chart in the near\nfuture as we scale is going to look less\n\nfuture as we scale is going to look less\n \n\nfuture as we scale is going to look less\nlike a pyramid and more like a network\n\nlike a pyramid and more like a network\n \n\nlike a pyramid and more like a network\nof humans and AI\n\nof humans and AI\n \n\nof humans and AI\ntogether. Another attribute is\n\ntogether. Another attribute is\n \n\ntogether. Another attribute is\nexperimentation and iterative culture in\n\nexperimentation and iterative culture in\n \n\nexperimentation and iterative culture in\nthe DNA. So now I know I know I know\n\nthe DNA. So now I know I know I know\n \n\nthe DNA. So now I know I know I know\nknow this is a core value in in the tech\n\nknow this is a core value in in the tech\n \n\nknow this is a core value in in the tech\nstartup world. Core value core\n\nstartup world. Core value core\n \n\nstartup world. Core value core\ninnovation is that experimentation\n\ninnovation is that experimentation\n \n\ninnovation is that experimentation\niterative approach to things. So it can\n\niterative approach to things. So it can\n \n\niterative approach to things. So it can\nbe cliche but if you think about it in\n\nbe cliche but if you think about it in\n \n\nbe cliche but if you think about it in\ncontext of an AI native company with all\n\ncontext of an AI native company with all\n \n\ncontext of an AI native company with all\nthe innovations we have now the ultimate\n\nthe innovations we have now the ultimate\n \n\nthe innovations we have now the ultimate\nrealization of this is really possible\n\nrealization of this is really possible\n \n\nrealization of this is really possible\nwhen AI is doing the routine work and AI\n\nwhen AI is doing the routine work and AI\n \n\nwhen AI is doing the routine work and AI\nis helping us with the prototypes we can\n\nis helping us with the prototypes we can\n \n\nis helping us with the prototypes we can\nreally focus on what matters I think\n\nreally focus on what matters I think\n \n\nreally focus on what matters I think\nit's I think it's powerful and I think\n\nit's I think it's powerful and I think\n \n\nit's I think it's powerful and I think\nit's awesome and it also compounds when\n\nit's awesome and it also compounds when\n \n\nit's awesome and it also compounds when\nyou start thinking about agents working\n\nyou start thinking about agents working\n \n\nyou start thinking about agents working\non things in the company learn and\n\non things in the company learn and\n \n\non things in the company learn and\nimprove it really is useful I Just take\n\nimprove it really is useful I Just take\n \n\nimprove it really is useful I Just take\nCognition's dead end. We used it. We\n\nCognition's dead end. We used it. We\n \n\nCognition's dead end. We used it. We\nstarted using it a few months ago. It's\n\nstarted using it a few months ago. It's\n \n\nstarted using it a few months ago. It's\ngood. We had to do some things and now\n\ngood. We had to do some things and now\n \n\ngood. We had to do some things and now\nover time it's learned so much. It's\n\nover time it's learned so much. It's\n \n\nover time it's learned so much. It's\ndocumented so much of our code. It knows\n\ndocumented so much of our code. It knows\n \n\ndocumented so much of our code. It knows\nhow we handle certain things. It's it's\n\nhow we handle certain things. It's it's\n \n\nhow we handle certain things. It's it's\na superpower at this point. a superpower at this point. a superpower at this point. So all of these attributes AI at the\n\nSo all of these attributes AI at the\n \n\nSo all of these attributes AI at the\ncore, humans as orchestrators, uh rapid\n\ncore, humans as orchestrators, uh rapid\n \n\ncore, humans as orchestrators, uh rapid\nexperimentation, self-learning, agent\n\nexperimentation, self-learning, agent\n \n\nexperimentation, self-learning, agent\nevolution, they combine to create an\n\nevolution, they combine to create an\n \n\nevolution, they combine to create an\norganization that looks and feels\n\norganization that looks and feels\n \n\norganization that looks and feels\ntotally different from a traditional\n\ntotally different from a traditional\n \n\ntotally different from a traditional\ncompany. It's not just a little more\n\ncompany. It's not just a little more\n \n\ncompany. It's not just a little more\nefficient, it's operating in a totally\n\nefficient, it's operating in a totally\n \n\nefficient, it's operating in a totally\ndifferent model. And so like if you're\n\ndifferent model. And so like if you're\n \n\ndifferent model." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nAnd so like if you're\nan MBA and you're like, I want to I\n\nan MBA and you're like, I want to I\n \n\nan MBA and you're like, I want to I\nlearned this approach at Berkeley, you\n\nlearned this approach at Berkeley, you\n \n\nlearned this approach at Berkeley, you\nknow, and this is here here's my\n\nknow, and t..." }, { "role": "assistant", "content": "And so like if you're\nan MBA and you're like, I want to I\n\nan MBA and you're like, I want to I\n \n\nan MBA and you're like, I want to I\nlearned this approach at Berkeley, you\n\nlearned this approach at Berkeley, you\n \n\nlearned this approach at Berkeley, you\nknow, and this is here here's my\n\nknow, and this is here here's my\n \n\nknow, and this is here here's my\ntemplate. It's not going to fit. It's a\n\ntemplate. It's not going to fit. It's a\n \n\ntemplate. It's not going to fit. It's a\nwhole other model that we're talking\n\nwhole other model that we're talking\n \n\nwhole other model that we're talking\nabout. Now, let's talk about the agent\n\nabout. Now, let's talk about the agent\n \n\nabout. Now, let's talk about the agent\nnative typical workday. What does that\n\nnative typical workday. What does that\n \n\nnative typical workday. What does that\nlook like anyway? And I think we have a\n\nlook like anyway? And I think we have a\n \n\nlook like anyway? And I think we have a\nlot of ideas on what it can look like in\n\nlot of ideas on what it can look like in\n \n\nlot of ideas on what it can look like in\nthe future. But really, let's just ask\n\nthe future. But really, let's just ask\n \n\nthe future. But really, let's just ask\nthe question, well, what's new? And\n\nthe question, well, what's new? And\n \n\nthe question, well, what's new? And\ndefinitely one of the things that's new\n\ndefinitely one of the things that's new\n \n\ndefinitely one of the things that's new\nis overseeing what AI is doing or what\n\nis overseeing what AI is doing or what\n \n\nis overseeing what AI is doing or what\nit has done. That's certainly something\n\nit has done. That's certainly something\n \n\nit has done. That's certainly something\nthat is um an everyday part of our lives\n\nthat is um an everyday part of our lives\n \n\nthat is um an everyday part of our lives\nnow that definitely didn't exist just a\n\nnow that definitely didn't exist just a\n \n\nnow that definitely didn't exist just a\nfew years ago. And so even just me for\n\nfew years ago. And so even just me for\n \n\nfew years ago. And so even just me for\nexample, I usually start my day\n\nexample, I usually start my day\n \n\nexample, I usually start my day\nuh with a builtup log of things that I\n\nuh with a builtup log of things that I\n \n\nuh with a builtup log of things that I\nhave to get done. And usually I have\n\nhave to get done. And usually I have\n \n\nhave to get done. And usually I have\ncertain theme days. I try to put put\n\ncertain theme days. I try to put put\n \n\ncertain theme days. I try to put put\ncertain tasks and certain days. And\n\ncertain tasks and certain days. And\n \n\ncertain tasks and certain days. And\nsometimes they're mundane tasks. sometimes they're mundane tasks. sometimes they're mundane tasks. Sometimes they're product ideas that\n\nSometimes they're product ideas that\n \n\nSometimes they're product ideas that\nneed to be thought through. So I found\n\nneed to be thought through. So I found\n \n\nneed to be thought through. So I found\nthat like in the morning a lot of times\n\nthat like in the morning a lot of times\n \n\nthat like in the morning a lot of times\nI'm out for a walk, I'm driving the car\n\nI'm out for a walk, I'm driving the car\n \n\nI'm out for a walk, I'm driving the car\nand I'll actually chat which I'll talk\n\nand I'll actually chat which I'll talk\n \n\nand I'll actually chat which I'll talk\nto chat GBT and I use 03 or deep\n\nto chat GBT and I use 03 or deep\n \n\nto chat GBT and I use 03 or deep\nresearch um some deep um reasoning stuff\n\nresearch um some deep um reasoning stuff\n \n\nresearch um some deep um reasoning stuff\nand um just kick off bigger thinking\n\nand um just kick off bigger thinking\n \n\nand um just kick off bigger thinking\nthings like hey this is what we're\n\nthings like hey this is what we're\n \n\nthings like hey this is what we're\nthinking of. Here's a document. Here's a\n\nthinking of. Here's a document. Here's a\n \n\nthinking of. Here's a document. Here's a\nlink. This is what we said. Here's a\n\nlink. This is what we said. Here's a\n \n\nlink. This is what we said. Here's a\nconversation. Here's a granola\n\nconversation. Here's a granola\n \n\nconversation. Here's a granola\ntranscript. And then so by the time I\n\ntranscript. And then so by the time I\n \n\ntranscript. And then so by the time I\nactually get to that to-do for that day,\n\nactually get to that to-do for that day,\n \n\nactually get to that to-do for that day,\nI've already got a bunch of thinking\n\nI've already got a bunch of thinking\n \n\nI've already got a bunch of thinking\naround it. it might have kicked off\n\naround it. it might have kicked off\n \n\naround it. it might have kicked off\nDevon and created a couple PRs. And\n\nDevon and created a couple PRs. And\n \n\nDevon and created a couple PRs. And\nthat's definitely the case when there's\n\nthat's definitely the case when there's\n \n\nthat's definitely the case when there's\nuh things like bugs and doc issues and\n\nuh things like bugs and doc issues and\n \n\nuh things like bugs and doc issues and\nother things that have amounted from\n\nother things that have amounted from\n \n\nother things that have amounted from\nusers that get put into linear. Devon is\n\nusers that get put into linear. Devon is\n \n\nusers that get put into linear. Devon is\nalready working on it. Dev's already got\n\nalready working on it. Dev's already got\n \n\nalready working on it. Dev's already got\nPRs. So, it's very very helpful just to\n\nPRs. So, it's very very helpful just to\n \n\nPRs. So, it's very very helpful just to\nget things moving. A lot of times, you\n\nget things moving. A lot of times, you\n \n\nget things moving. A lot of times, you\nknow, you got like the the molasses\n\nknow, you got like the the molasses\n \n\nknow, you got like the the molasses\naround your your legs, you know, like I\n\naround your your legs, you know, like I\n \n\naround your your legs, you know, like I\njust got to get moving before I start\n\njust got to get moving before I start\n \n\njust got to get moving before I start\nget uh getting up to speed. And so, this\n\nget uh getting up to speed. And so, this\n \n\nget uh getting up to speed. And so, this\nkind of helps with the ball moving a\n\nkind of helps with the ball moving a\n \n\nkind of helps with the ball moving a\nlot. So certainly that's that's part of\n\nlot. So certainly that's that's part of\n \n\nlot." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nSo certainly that's that's part of\nthe beginning of the day for\n\nthe beginning of the day for\n \n\nthe beginning of the day for\nme and I think essentially what we found\n\nme and I think essentially what we found\n \n\nme and I think essentially what we found\nright now is the morning time is check\n\nright n..." }, { "role": "assistant", "content": "So certainly that's that's part of\nthe beginning of the day for\n\nthe beginning of the day for\n \n\nthe beginning of the day for\nme and I think essentially what we found\n\nme and I think essentially what we found\n \n\nme and I think essentially what we found\nright now is the morning time is check\n\nright now is the morning time is check\n \n\nright now is the morning time is check\nout what agents did what they need to do\n\nout what agents did what they need to do\n \n\nout what agents did what they need to do\nkick off the things that they need to do\n\nkick off the things that they need to do\n \n\nkick off the things that they need to do\nand then come around I don't know around\n\nand then come around I don't know around\n \n\nand then come around I don't know around\nlunchtime review everything you're going\n\nlunchtime review everything you're going\n \n\nlunchtime review everything you're going\nto have a bunch of PRs you're going to\n\nto have a bunch of PRs you're going to\n \n\nto have a bunch of PRs you're going to\nhave your collateral maybe some emails\n\nhave your collateral maybe some emails\n \n\nhave your collateral maybe some emails\nwe got some email agents that do things\n\nwe got some email agents that do things\n \n\nwe got some email agents that do things\num and so that's one example big picture\n\num and so that's one example big picture\n \n\num and so that's one example big picture\nI think every employee they can become a\n\nI think every employee they can become a\n \n\nI think every employee they can become a\nlead manager type and and and I don't\n\nlead manager type and and and I don't\n \n\nlead manager type and and and I don't\nmean necessarily a people manager but of\n\nmean necessarily a people manager but of\n \n\nmean necessarily a people manager but of\ntheir AI agent counterparts that are\n\ntheir AI agent counterparts that are\n \n\ntheir AI agent counterparts that are\nresponsible for the jobs the person is\n\nresponsible for the jobs the person is\n \n\nresponsible for the jobs the person is\nhired to do. I mean, we have content\n\nhired to do. I mean, we have content\n \n\nhired to do. I mean, we have content\nmarketing swarms that we use multiple\n\nmarketing swarms that we use multiple\n \n\nmarketing swarms that we use multiple\nagents that that I'm orchestrating and\n\nagents that that I'm orchestrating and\n \n\nagents that that I'm orchestrating and\ntelling the agent, hey, the copy's over\n\ntelling the agent, hey, the copy's over\n \n\ntelling the agent, hey, the copy's over\nhere, optimize it, figure out the best\n\nhere, optimize it, figure out the best\n \n\nhere, optimize it, figure out the best\ntime for social posting. All this\n\ntime for social posting. All this\n \n\ntime for social posting. All this\nhappens and automatically schedules with\n\nhappens and automatically schedules with\n \n\nhappens and automatically schedules with\ntype fully we use. Um, so it's a it's\n\ntype fully we use. Um, so it's a it's\n \n\ntype fully we use. Um, so it's a it's\nleveraging expertise of the human, but\n\nleveraging expertise of the human, but\n \n\nleveraging expertise of the human, but\nwith async workloads kind of just doing\n\nwith async workloads kind of just doing\n \n\nwith async workloads kind of just doing\nit all so that when you get to it, it's\n\nit all so that when you get to it, it's\n \n\nit all so that when you get to it, it's\nall ready to go. Now, back to my earlier\n\nall ready to go. Now, back to my earlier\n \n\nall ready to go. Now, back to my earlier\npoint, I think this leads to a much\n\npoint, I think this leads to a much\n \n\npoint, I think this leads to a much\nflatter team structure and probably\n\nflatter team structure and probably\n \n\nflatter team structure and probably\ndifferent titles, honestly. Titles that\n\ndifferent titles, honestly. Titles that\n \n\ndifferent titles, honestly. Titles that\ncombine domain expertise with AI\n\ncombine domain expertise with AI\n \n\ncombine domain expertise with AI\nknow-how. I mean, we even have that in\n\nknow-how. I mean, we even have that in\n \n\nknow-how. I mean, we even have that in\nour you said, Rick, that's not true. our you said, Rick, that's not true. our you said, Rick, that's not true. It's always going to be the customer\n\nIt's always going to be the customer\n \n\nIt's always going to be the customer\nsupport. Think so. It might be AI\n\nsupport. Think so. It might be AI\n \n\nsupport. Think so. It might be AI\nengineer or AI um AI customer lead or\n\nengineer or AI um AI customer lead or\n \n\nengineer or AI um AI customer lead or\nsomething like that. I mean even our own\n\nsomething like that. I mean even our own\n \n\nsomething like that. I mean even our own\nindustry it's AI engineer is now you\n\nindustry it's AI engineer is now you\n \n\nindustry it's AI engineer is now you\nhear that everywhere and I think it's\n\nhear that everywhere and I think it's\n \n\nhear that everywhere and I think it's\ngoing to be across the board across\n\ngoing to be across the board across\n \n\ngoing to be across the board across\neverywhere in the organization. So that\n\neverywhere in the organization. So that\n \n\neverywhere in the organization. So that\nleads us to the final point which is\n\nleads us to the final point which is\n \n\nleads us to the final point which is\nrethinking the hiring process and who\n\nrethinking the hiring process and who\n \n\nrethinking the hiring process and who\nyou're hiring because when you start\n\nyou're hiring because when you start\n \n\nyou're hiring because when you start\nthinking this way you start thinking not\n\nthinking this way you start thinking not\n \n\nthinking this way you start thinking not\njust the hiring process and operation\n\njust the hiring process and operation\n \n\njust the hiring process and operation\nbut who you're hiring needs to be a\n\nbut who you're hiring needs to be a\n \n\nbut who you're hiring needs to be a\nlittle different you know curiosity\n\nlittle different you know curiosity\n \n\nlittle different you know curiosity\nadaptability those are things that are\n\nadaptability those are things that are\n \n\nadaptability those are things that are\nin high demand right for creative people\n\nin high demand right for creative people\n \n\nin high demand right for creative people\nfor leadership roles so we know that but\n\nfor leadership roles so we know that but\n \n\nfor leadership roles so we know that but\nessentially what we're saying is for an\n\nessentially what we're saying is for an\n \n\nessentially what we're saying is for an\nagent native company that is required\n\nagent native company that is required\n \n\nagent native company that is required\nhired. We have to have that right now in\n\nhired. We have to have that right now in\n \n\nhired. We have to have that right now in\norder to hire somebody because they're\n\norder to hire somebody because they're\n \n\norder to hire somebody because they're\nnot going to be able to use AI uh the\n\nnot going to be able to use AI uh the\n \n\nnot going to be able to use AI uh the\nway that we need them to. So, so maybe\n\nway that we need them to. So, so maybe\n \n\nway that we need them to. So, so maybe\nanother way to say it is AI fluency\n\nanother way to say it is AI fluency\n \n\nanother way to say it is AI fluency\nbecomes a mustave. You know, to\n\nbecomes a mustave. You know, to\n \n\nbecomes a mustave. You know, to\nreference uh the old world, you would\n\nreference uh the old world, you would\n \n\nreference uh the old world, you would\nnever ever ever hire somebody in an\n\nnever ever ever hire somebody in an\n \n\nnever ever ever hire somebody in an\noffice job if they've never used a word\n\noffice job if they've never used a word\n \n\noffice job if they've never used a word\nprocessor in their life, right? They\n\nprocessor in their life, right? They\n \n\nprocessor in their life, right? They\ndon't know how to use a keyboard. You\n\ndon't know how to use a keyboard. You\n \n\ndon't know how to use a keyboard. You\nwould never hire them. It's just\n\nwould never hire them. It's just\n \n\nwould never hire them. It's just\nexpected. You would even probably put it\n\nexpected. You would even probably put it\n \n\nexpected. You would even probably put it\non your job requirement. It's just\n\non your job requirement. It's just\n \n\non your job requirement. It's just\nexpected that they know how to use a\n\nexpected that they know how to use a\n \n\nexpected that they know how to use a\nkeyboard. Now, if you went 120 years\n\nkeyboard. Now, if you went 120 years\n \n\nkeyboard." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nNow, if you went 120 years\nago, that might be a debatable thing. I\n\nago, that might be a debatable thing. I\n \n\nago, that might be a debatable thing. I\ndon't know. Is it really needed? Do they\n\ndon't know. Is it really needed? Do they\n \n\ndon't know. Is it really needed? Do they\nreally need to use a k..." }, { "role": "assistant", "content": "Now, if you went 120 years\nago, that might be a debatable thing. I\n\nago, that might be a debatable thing. I\n \n\nago, that might be a debatable thing. I\ndon't know. Is it really needed? Do they\n\ndon't know. Is it really needed? Do they\n \n\ndon't know. Is it really needed? Do they\nreally need to use a keyboard? I don't\n\nreally need to use a keyboard? I don't\n \n\nreally need to use a keyboard? I don't\nknow. It's just typewriter. As long as\n\nknow. It's just typewriter. As long as\n \n\nknow. It's just typewriter. As long as\nthey have nice hand\n\nthey have nice hand\n \n\nthey have nice hand\npenmanship. But AI fluency is actually a\n\npenmanship. But AI fluency is actually a\n \n\npenmanship. But AI fluency is actually a\nreally big deal for us. And I think it's\n\nreally big deal for us. And I think it's\n \n\nreally big deal for us. And I think it's\njust going to keep being that way. It's\n\njust going to keep being that way. It's\n \n\njust going to keep being that way. It's\nnot really hard to imagine if the\n\nnot really hard to imagine if the\n \n\nnot really hard to imagine if the\nexpectation is a flatter or structure\n\nexpectation is a flatter or structure\n \n\nexpectation is a flatter or structure\nwhere each employee is efficient at\n\nwhere each employee is efficient at\n \n\nwhere each employee is efficient at\ndirectly guiding agents to task at hand\n\ndirectly guiding agents to task at hand\n \n\ndirectly guiding agents to task at hand\nyou and utilizing the person's expertise\n\nyou and utilizing the person's expertise\n \n\nyou and utilizing the person's expertise\nthrough those agents, then you're going\n\nthrough those agents, then you're going\n \n\nthrough those agents, then you're going\nto hone in on the ability to do that. to hone in on the ability to do that. to hone in on the ability to do that. You're going to in your hiring process\n\nYou're going to in your hiring process\n \n\nYou're going to in your hiring process\nwant to find out if that person has that\n\nwant to find out if that person has that\n \n\nwant to find out if that person has that\nability to do that. It's just going to\n\nability to do that. It's just going to\n \n\nability to do that. It's just going to\nbe something that happens. I I think I\n\nbe something that happens. I I think I\n \n\nbe something that happens. I I think I\nmean even in our current interviews\n\nmean even in our current interviews\n \n\nmean even in our current interviews\nwe're very very skeptical if someone\n\nwe're very very skeptical if someone\n \n\nwe're very very skeptical if someone\ndoesn't use AI and not familiar with it. doesn't use AI and not familiar with it. doesn't use AI and not familiar with it. It's just immediate warning flags um now\n\nIt's just immediate warning flags um now\n \n\nIt's just immediate warning flags um now\nmaybe they haven't had the opportunity\n\nmaybe they haven't had the opportunity\n \n\nmaybe they haven't had the opportunity\nto use it and stuff. Obviously, there's\n\nto use it and stuff. Obviously, there's\n \n\nto use it and stuff. Obviously, there's\nthere's scenarios because we're early on\n\nthere's scenarios because we're early on\n \n\nthere's scenarios because we're early on\nin this that, you know, we have to be a\n\nin this that, you know, we have to be a\n \n\nin this that, you know, we have to be a\nlittle understanding. Um, but it does\n\nlittle understanding. Um, but it does\n \n\nlittle understanding. Um, but it does\nmake us ask the question, does this\n\nmake us ask the question, does this\n \n\nmake us ask the question, does this\nperson have the ability to guide AI\n\nperson have the ability to guide AI\n \n\nperson have the ability to guide AI\nagents and learn the little uh tips and\n\nagents and learn the little uh tips and\n \n\nagents and learn the little uh tips and\ntricks and tweaks and and all that kind\n\ntricks and tweaks and and all that kind\n \n\ntricks and tweaks and and all that kind\nof\n\nof\n \n\nof\nstuff. And and even on a bigger picture,\n\nstuff. And and even on a bigger picture,\n \n\nstuff. And and even on a bigger picture,\nyou know, a reason a big reason why you\n\nyou know, a reason a big reason why you\n \n\nyou know, a reason a big reason why you\nwould hire a VP, for example, you bring\n\nwould hire a VP, for example, you bring\n \n\nwould hire a VP, for example, you bring\nin a VP is for their network. you hire\n\nin a VP is for their network. you hire\n \n\nin a VP is for their network. you hire\nthem for their experience and the\n\nthem for their experience and the\n \n\nthem for their experience and the\nquality of people that they would bring\n\nquality of people that they would bring\n \n\nquality of people that they would bring\nin to fill out their team to do the job\n\nin to fill out their team to do the job\n \n\nin to fill out their team to do the job\nthat they need to do. So, in this new\n\nthat they need to do. So, in this new\n \n\nthat they need to do. So, in this new\ntype of company, the spotlight gets\n\ntype of company, the spotlight gets\n \n\ntype of company, the spotlight gets\nreally put on the like does this does\n\nreally put on the like does this does\n \n\nreally put on the like does this does\nthis guy know how to use agents? Does\n\nthis guy know how to use agents? Does\n \n\nthis guy know how to use agents? Does\nthis guy can he bring in people that\n\nthis guy can he bring in people that\n \n\nthis guy can he bring in people that\nknow how to use agents? Um, do they have\n\nknow how to use agents? Um, do they have\n \n\nknow how to use agents? Um, do they have\nAI fluency? So, it comes a really big\n\nAI fluency? So, it comes a really big\n \n\nAI fluency? So, it comes a really big\nimportant thing. And then if you kind of\n\nimportant thing. And then if you kind of\n \n\nimportant thing. And then if you kind of\njust think through the the the life\n\njust think through the the the life\n \n\njust think through the the the life\ncycle that okay let's say you hire that\n\ncycle that okay let's say you hire that\n \n\ncycle that okay let's say you hire that\nperson then the onboarding and then they\n\nperson then the onboarding and then they\n \n\nperson then the onboarding and then they\nhire some other people to build out\n\nhire some other people to build out\n \n\nhire some other people to build out\ntheir team. The onboarding is probably\n\ntheir team. The onboarding is probably\n \n\ntheir team. The onboarding is probably\ndifferent too right when you hire that\n\ndifferent too right when you hire that\n \n\ndifferent too right when you hire that\nthe intent of being agent native then uh\n\nthe intent of being agent native then uh\n \n\nthe intent of being agent native then uh\nguess what agentic tools and systems\n\nguess what agentic tools and systems\n \n\nguess what agentic tools and systems\nthat need to be put in place for them to\n\nthat need to be put in place for them to\n \n\nthat need to be put in place for them to\nbe successful is really important. So\n\nbe successful is really important. So\n \n\nbe successful is really important. So\nthen it doesn't seem unreasonable to\n\nthen it doesn't seem unreasonable to\n \n\nthen it doesn't seem unreasonable to\nattach an engineer to that team to make\n\nattach an engineer to that team to make\n \n\nattach an engineer to that team to make\nsure that their agents are up and\n\nsure that their agents are up and\n \n\nsure that their agents are up and\nrunning and build out. Um it's the same\n\nrunning and build out. Um it's the same\n \n\nrunning and build out. Um it's the same\nthing with like new hires. You probably\n\nthing with like new hires. You probably\n \n\nthing with like new hires. You probably\nwould expect them the first few weeks to\n\nwould expect them the first few weeks to\n \n\nwould expect them the first few weeks to\nsolely be focused on getting their agent\n\nsolely be focused on getting their agent\n \n\nsolely be focused on getting their agent\nset up to do their job. Makes sense,\n\nset up to do their job. Makes sense,\n \n\nset up to do their job. Makes sense,\nright? In conclusion, just to wrap it\n\nright? In conclusion, just to wrap it\n \n\nright? In conclusion, just to wrap it\nup, we're going through a profound\n\nup, we're going through a profound\n \n\nup, we're going through a profound\nshift." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nWe don't know exactly where we're\n\nshift. We don't know exactly where we're\n \n\nshift. We don't know exactly where we're\ngoing to end up. You know, the ceiling\n\ngoing to end up. You know, the ceiling\n \n\ngoing to end up. You know, the ceiling\nis being raised so high, we don't even\n\nis being raised so ..." }, { "role": "assistant", "content": "We don't know exactly where we're\n\nshift. We don't know exactly where we're\n \n\nshift. We don't know exactly where we're\ngoing to end up. You know, the ceiling\n\ngoing to end up. You know, the ceiling\n \n\ngoing to end up. You know, the ceiling\nis being raised so high, we don't even\n\nis being raised so high, we don't even\n \n\nis being raised so high, we don't even\nfully understand, I think, how large\n\nfully understand, I think, how large\n \n\nfully understand, I think, how large\nthis AI world and the economy will go. I\n\nthis AI world and the economy will go. I\n \n\nthis AI world and the economy will go. I\nthink it's a lot like the car industry. think it's a lot like the car industry. think it's a lot like the car industry. We didn't realize how big the ceiling\n\nWe didn't realize how big the ceiling\n \n\nWe didn't realize how big the ceiling\nwas being lifted in the in the economy\n\nwas being lifted in the in the economy\n \n\nwas being lifted in the in the economy\nwhen the car came along. And I think\n\nwhen the car came along. And I think\n \n\nwhen the car came along. And I think\nthat's the case with this this AI stuff. that's the case with this this AI stuff. that's the case with this this AI stuff. So, I think AI agents will be deeply\n\nSo, I think AI agents will be deeply\n \n\nSo, I think AI agents will be deeply\nembedded in every aspect of business. embedded in every aspect of business. embedded in every aspect of business. And that means rethinking roles. It\n\nAnd that means rethinking roles. It\n \n\nAnd that means rethinking roles. It\nmeans rethinking skills and I even think\n\nmeans rethinking skills and I even think\n \n\nmeans rethinking skills and I even think\nit means rethinking culture and\n\nit means rethinking culture and\n \n\nit means rethinking culture and\noperations. The move isn't merely just\n\noperations. The move isn't merely just\n \n\noperations. The move isn't merely just\nbusinesses that use AI as a tool. That's\n\nbusinesses that use AI as a tool. That's\n \n\nbusinesses that use AI as a tool. That's\nnot what we're talking about. We're\n\nnot what we're talking about. We're\n \n\nnot what we're talking about. We're\ntalking about businesses that are built\n\ntalking about businesses that are built\n \n\ntalking about businesses that are built\naround AI as a core primitive to their\n\naround AI as a core primitive to their\n \n\naround AI as a core primitive to their\nexistence. Again, from the driver\n\nexistence. Again, from the driver\n \n\nexistence. Again, from the driver\nassisted to the AI automated uh driven\n\nassisted to the AI automated uh driven\n \n\nassisted to the AI automated uh driven\ncar. And that shift affords a ton of\n\ncar. And that shift affords a ton of\n \n\ncar. And that shift affords a ton of\nopportunities. But it is a shift. It's\n\nopportunities. But it is a shift. It's\n \n\nopportunities. But it is a shift. It's\njust going to create a lot of friction. just going to create a lot of friction. just going to create a lot of friction. But I mean just just to tell you I just\n\nBut I mean just just to tell you I just\n \n\nBut I mean just just to tell you I just\nwith us we've been around for you know\n\nwith us we've been around for you know\n \n\nwith us we've been around for you know\n15 16 weeks small team six seven people\n\n15 16 weeks small team six seven people\n \n\n15 16 weeks small team six seven people\nand uh we built an entire Aenta cloud\n\nand uh we built an entire Aenta cloud\n \n\nand uh we built an entire Aenta cloud\ninfrastructure from scratch in just a\n\ninfrastructure from scratch in just a\n \n\ninfrastructure from scratch in just a\nfew weeks. That's unheard of. Like if\n\nfew weeks. That's unheard of. Like if\n \n\nfew weeks. That's unheard of. Like if\nyou asked me that a few years ago I said\n\nyou asked me that a few years ago I said\n \n\nyou asked me that a few years ago I said\nthere's no way. There's no way. But with\n\nthere's no way. There's no way. But with\n \n\nthere's no way. There's no way. But with\nall of these tools and these agents,\n\nall of these tools and these agents,\n \n\nall of these tools and these agents,\nthese agents that we build ourselves for\n\nthese agents that we build ourselves for\n \n\nthese agents that we build ourselves for\nour things and then uh Devon and some\n\nour things and then uh Devon and some\n \n\nour things and then uh Devon and some\nothers, man, we've made so much progress\n\nothers, man, we've made so much progress\n \n\nothers, man, we've made so much progress\non this. So, it's a huge shift and\n\non this. So, it's a huge shift and\n \n\non this. So, it's a huge shift and\nespecially for founders and tech\n\nespecially for founders and tech\n \n\nespecially for founders and tech\nleaders, the challenge is that I think\n\nleaders, the challenge is that I think\n \n\nleaders, the challenge is that I think\nfully embrace the paradigm shift. If\n\nfully embrace the paradigm shift. If\n \n\nfully embrace the paradigm shift. If\nyou're an experienced founder, this is\n\nyou're an experienced founder, this is\n \n\nyou're an experienced founder, this is\nespecially true, but it's especially\n\nespecially true, but it's especially\n \n\nespecially true, but it's especially\nhard because you've got all this back uh\n\nhard because you've got all this back uh\n \n\nhard because you've got all this back uh\nyears and years and years of experience\n\nyears and years and years of experience\n \n\nyears and years and years of experience\nand you might need to check some of that\n\nand you might need to check some of that\n \n\nand you might need to check some of that\nexperience at the door. Seriously, you\n\nexperience at the door. Seriously, you\n \n\nexperience at the door. Seriously, you\nmight need to rethink like you have all\n\nmight need to rethink like you have all\n \n\nmight need to rethink like you have all\nthis built-up experience that might no\n\nthis built-up experience that might no\n \n\nthis built-up experience that might no\nlonger be valid or only parts of it are\n\nlonger be valid or only parts of it are\n \n\nlonger be valid or only parts of it are\nvalid and you have it takes some\n\nvalid and you have it takes some\n \n\nvalid and you have it takes some\ncritical thinking to analyze that\n\ncritical thinking to analyze that\n \n\ncritical thinking to analyze that\nbecause otherwise you just get stuck in\n\nbecause otherwise you just get stuck in\n \n\nbecause otherwise you just get stuck in\nthe old way of doing things. I mean PWC\n\nthe old way of doing things. I mean PWC\n \n\nthe old way of doing things. I mean PWC\nof all of all places they they had a\n\nof all of all places they they had a\n \n\nof all of all places they they had a\nreally good report recently and one of\n\nreally good report recently and one of\n \n\nreally good report recently and one of\nthe comments they said was if you're\n\nthe comments they said was if you're\n \n\nthe comments they said was if you're\nonly using AI for a small efficiency\n\nonly using AI for a small efficiency\n \n\nonly using AI for a small efficiency\ngain then you're falling behind because\n\ngain then you're falling behind because\n \n\ngain then you're falling behind because\nthere's companies that they're not just\n\nthere's companies that they're not just\n \n\nthere's companies that they're not just\nusing it for a small efficiency gain. So\n\nusing it for a small efficiency gain. So\n \n\nusing it for a small efficiency gain. So\nmy admonishment would be uh start from\n\nmy admonishment would be uh start from\n \n\nmy admonishment would be uh start from\nfirst principles take this opportunity\n\nfirst principles take this opportunity\n \n\nfirst principles take this opportunity\nto step back reimagine and refit your\n\nto step back reimagine and refit your\n \n\nto step back reimagine and refit your\ncompany and your culture for this\n\ncompany and your culture for this\n \n\ncompany and your culture for this\nfuture. you rewire the entire process so\n\nfuture. you rewire the entire process so\n \n\nfuture. you rewire the entire process so\nthat human to agent teams can scale that\n\nthat human to agent teams can scale that\n \n\nthat human to agent teams can scale that\nimpact\n\nimpact\n \n\nimpact\nexponentially. I mean that might mean\n\nexponentially. I mean that might mean\n \n\nexponentially. I mean that might mean\nagain this is friction this is pain that\n\nagain this is friction this is pain that\n \n\nagain this is friction this is pain that\nmight mean redesigning your org chart. might mean redesigning your org chart." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nmight mean redesigning your org chart. It might mean redefining roles and\n\nIt might mean redefining roles and\n \n\nIt might mean redefining roles and\nrethinking what skills you hire for. rethinking what skills you hire for. rethinking what skills you hire for. It's a lot of change but it also can be\n\n..." }, { "role": "assistant", "content": "might mean redesigning your org chart. It might mean redefining roles and\n\nIt might mean redefining roles and\n \n\nIt might mean redefining roles and\nrethinking what skills you hire for. rethinking what skills you hire for. rethinking what skills you hire for. It's a lot of change but it also can be\n\nIt's a lot of change but it also can be\n \n\nIt's a lot of change but it also can be\nyour unfair advantage if you do it. So\n\nyour unfair advantage if you do it. So\n \n\nyour unfair advantage if you do it. So\nwith that I'll just leave you with one\n\nwith that I'll just leave you with one\n \n\nwith that I'll just leave you with one\nquestion. Is your company using AI or is\n\nquestion. Is your company using AI or is\n \n\nquestion. Is your company using AI or is\nit ready to be built around AI? Thanks\n\nit ready to be built around AI? Thanks\n \n\nit ready to be built around AI? Thanks\nfor listening." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\n[Music]\n\n \n\nI want to introduce our first Speaker\n\nI want to introduce our first Speaker\n \n\nI want to introduce our first Speaker\nbang Liu CTO of source graph and I won't\n\nbang Liu CTO of source graph and I won't\n \n\nbang Liu CTO of source graph and I won't\nspoil the topic of the talk but it's\n\nspoil..." }, { "role": "assistant", "content": "[Music]\n\n \n\nI want to introduce our first Speaker\n\nI want to introduce our first Speaker\n \n\nI want to introduce our first Speaker\nbang Liu CTO of source graph and I won't\n\nbang Liu CTO of source graph and I won't\n \n\nbang Liu CTO of source graph and I won't\nspoil the topic of the talk but it's\n\nspoil the topic of the talk but it's\n \n\nspoil the topic of the talk but it's\ngoing to be a good one thank you very\n\ngoing to be a good one thank you very\n \n\ngoing to be a good one thank you very\nmuch awesome thank you Peter how's\n\nmuch awesome thank you Peter how's\n \n\nmuch awesome thank you Peter how's\neveryone doing this\n\neveryone doing this\n \n\neveryone doing this\nmorning yeah everyone uh awake bright\n\nmorning yeah everyone uh awake bright\n \n\nmorning yeah everyone uh awake bright\nand early thanks for coming out almost\n\nand early thanks for coming out almost\n \n\nand early thanks for coming out almost\nawake awesome um so before I uh dive\n\nawake awesome um so before I uh dive\n \n\nawake awesome um so before I uh dive\ninto the talk here I just want to get a\n\ninto the talk here I just want to get a\n \n\ninto the talk here I just want to get a\nsense of you know who we all have in the\n\nsense of you know who we all have in the\n \n\nsense of you know who we all have in the\nroom so um you know who here is like a\n\nroom so um you know who here is like a\n \n\nroom so um you know who here is like a\nhead of engineering or VP of\n\nhead of engineering or VP of\n \n\nhead of engineering or VP of\nengineering okay a good number of you\n\nengineering okay a good number of you\n \n\nengineering okay a good number of you\nwho here is just uh uh you know IC Dev\n\nwho here is just uh uh you know IC Dev\n \n\nwho here is just uh uh you know IC Dev\ninterested in kind of like evaluating\n\ninterested in kind of like evaluating\n \n\ninterested in kind of like evaluating\nhow things are going okay and then what\n\nhow things are going okay and then what\n \n\nhow things are going okay and then what\ndo you what what do the rest of you do\n\ndo you what what do the rest of you do\n \n\ndo you what what do the rest of you do\njust shout out your\n\njust shout out your\n \n\njust shout out your\nroles anyone anyone middle management\n\nroles anyone anyone middle management\n \n\nroles anyone anyone middle management\nmiddle management okay cool and who here\n\nmiddle management okay cool and who here\n \n\nmiddle management okay cool and who here\nhas a really you know quantitative very\n\nhas a really you know quantitative very\n \n\nhas a really you know quantitative very\nprecise thought through uh evaluation\n\nprecise thought through uh evaluation\n \n\nprecise thought through uh evaluation\nframework for measuring the ROI of AI\n\nframework for measuring the ROI of AI\n \n\nframework for measuring the ROI of AI\ntools okay one hand in the back what\n\ntools okay one hand in the back what\n \n\ntools okay one hand in the back what\ncompany are you from\n\n \n\nSir signal broad signals Okay cool so we\n\nSir signal broad signals Okay cool so we\n \n\nSir signal broad signals Okay cool so we\ngot one person in the back and then who\n\ngot one person in the back and then who\n \n\ngot one person in the back and then who\nhere is sort of like we kind of are\n\nhere is sort of like we kind of are\n \n\nhere is sort of like we kind of are\nevaluating it but it's really kind of\n\nevaluating it but it's really kind of\n \n\nevaluating it but it's really kind of\nlike Vibes at this point anyone anyone\n\nlike Vibes at this point anyone anyone\n \n\nlike Vibes at this point anyone anyone\nbrave enough to Okay cool so you're in\n\nbrave enough to Okay cool so you're in\n \n\nbrave enough to Okay cool so you're in\nthe right\n\nthe right\n \n\nthe right\nplace now who am I why am I qualified to\n\nplace now who am I why am I qualified to\n \n\nplace now who am I why am I qualified to\ntalk on this topic so I'm the CTO and\n\ntalk on this topic so I'm the CTO and\n \n\ntalk on this topic so I'm the CTO and\nco-founder of a company called Source\n\nco-founder of a company called Source\n \n\nco-founder of a company called Source\ngraph uh we're a developer tools company\n\ngraph uh we're a developer tools company\n \n\ngraph uh we're a developer tools company\nuh if you haven't heard of us we have\n\nuh if you haven't heard of us we have\n \n\nuh if you haven't heard of us we have\ntwo products one is a code search engine\n\ntwo products one is a code search engine\n \n\ntwo products one is a code search engine\nso we started the company because we\n\nso we started the company because we\n \n\nso we started the company because we\nwere developers ourselves my co\n\nwere developers ourselves my co\n \n\nwere developers ourselves my co\nco-founder and I and we were really\n\nco-founder and I and we were really\n \n\nco-founder and I and we were really\ntired of the slog of diving through\n\ntired of the slog of diving through\n \n\ntired of the slog of diving through\nlarge complex code bases and trying to\n\nlarge complex code bases and trying to\n \n\nlarge complex code bases and trying to\nmake sense of what was happening in them\n\nmake sense of what was happening in them\n \n\nmake sense of what was happening in them\nthe other product that we have is an AI\n\nthe other product that we have is an AI\n \n\nthe other product that we have is an AI\ncoding assistant uh so what this uh\n\ncoding assistant uh so what this uh\n \n\ncoding assistant uh so what this uh\nproduct does is it's\n\nproduct does is it's\n \n\nproduct does is it's\nessentially oh sorry I should mention\n\nessentially oh sorry I should mention\n \n\nessentially oh sorry I should mention\nthat we have great adoption among really\n\nthat we have great adoption among really\n \n\nthat we have great adoption among really\ngreat companies so we started this\n\ngreat companies so we started this\n \n\ngreat companies so we started this\ncompany 10 years ago uh to solve this\n\ncompany 10 years ago uh to solve this\n \n\ncompany 10 years ago uh to solve this\nproblem of tackling uh understanding\n\nproblem of tackling uh understanding\n \n\nproblem of tackling uh understanding\ncode in large code bases and today we're\n\ncode in large code bases and today we're\n \n\ncode in large code bases and today we're\nvery fortunate to have uh you know\n\nvery fortunate to have uh you know\n \n\nvery fortunate to have uh you know\ncustomers that range from early stage\n\ncustomers that range from early stage\n \n\ncustomers that range from early stage\nstartups all the way through to the\n\nstartups all the way through to the\n \n\nstartups all the way through to the\nFortune 500 and even some government\n\nFortune 500 and even some government\n \n\nFortune 500 and even some government\nagencies so what's the tie-in to aii so\n\nagencies so what's the tie-in to aii so\n \n\nagencies so what's the tie-in to aii so\nabout 2 years ago we released a new\n\nabout 2 years ago we released a new\n \n\nabout 2 years ago we released a new\nproduct called Cody which is an AI\n\nproduct called Cody which is an AI\n \n\nproduct called Cody which is an AI\ncoding assistant that ties into our code\n\ncoding assistant that ties into our code\n \n\ncoding assistant that ties into our code\nsearch engine so you can kind of think\n\nsearch engine so you can kind of think\n \n\nsearch engine so you can kind of think\nof it as like a perplexity for code\n\nof it as like a perplexity for code\n \n\nof it as like a perplexity for code\nwhereas you know your kind of vanilla\n\nwhereas you know your kind of vanilla\n \n\nwhereas you know your kind of vanilla\nrun-of-the-mill AI coding assistant uh\n\nrun-of-the-mill AI coding assistant uh\n \n\nrun-of-the-mill AI coding assistant uh\nonly uses very local context only has\n\nonly uses very local context only has\n \n\nonly uses very local context only has\naccess to kind of like the open files in\n\naccess to kind of like the open files in\n \n\naccess to kind of like the open files in\nyour editor we spent the past 10 years\n\nyour editor we spent the past 10 years\n \n\nyour editor we spent the past 10 years\nbuilding this great code search engine\n\nbuilding this great code search engine\n \n\nbuilding this great code search engine\nthat uh is really good at surfacing\n\nthat uh is really good at surfacing\n \n\nthat uh is really good at surfacing\nrelevant codes and bits from across your\n\nrelevant codes and bits from across your\n \n\nrelevant codes and bits from across your\ncode base and it just so turns out that\n\ncode base and it just so turns out that\n \n\ncode base and it just so turns out that\nlike that's a great superpower for AI to\n\nlike that's a great superpower for AI to\n \n\nlike that's a great superpower for AI to\nhave right um you know for for those of\n\nhave right um you know for for those of\n \n\nhave right um you know for for those of\nus that have started using perplexity we\n\nus that have started using perplexity we\n \n\nus that have started using perplexity we\ncan kind of see the appeal and a big\n\ncan kind of see the appeal and a big\n \n\ncan kind of see the appeal and a big\npiece of the puzzle is not just the\n\npiece of the puzzle is not just the\n \n\npiece of the puzzle is not just the\nlanguage model itself but the ability to\n\nlanguage model itself but the ability to\n \n\nlanguage model itself but the ability to\nfetch and rank relevant context from you\n\nfetch and rank relevant context from you\n \n\nfetch and rank relevant context from you\nknow whole universe of data and\n\nknow whole universe of data and\n \n\nknow whole universe of data and\ninformation and so we had to solve this\n\ninformation and so we had to solve this\n \n\ninformation and so we had to solve this\nproblem uh in order to sell Cody into\n\nproblem uh in order to sell Cody into\n \n\nproblem uh in order to sell Cody into\nthe likes of uh one password uh pal Alto\n\nthe likes of uh one password uh pal Alto\n \n\nthe likes of uh one password uh pal Alto\nnetworks and uh lios which is a big\n\nnetworks and uh lios which is a big\n \n\nnetworks and uh lios which is a big\ngovernment contractor If you flew in\n\ngovernment contractor If you flew in\n \n\ngovernment contractor If you flew in\nhere from uh some else you probably\n\nhere from uh some else you probably\n \n\nhere from uh some else you probably\nentered through one of their security\n\nentered through one of their security\n \n\nentered through one of their security\nmachines we sold Cody to all these\n\nmachines we sold Cody to all these\n \n\nmachines we sold Cody to all these\norganizations and so each one of these\n\norganizations and so each one of these\n \n\norganizations and so each one of these\norganizations has kind of like a\n\norganizations has kind of like a\n \n\norganizations has kind of like a\ndifferent way of measuring Roi they have\n\ndifferent way of measuring Roi they have\n \n\ndifferent way of measuring Roi they have\ndifferent Frameworks that they apply and\n\ndifferent Frameworks that they apply and\n \n\ndifferent Frameworks that they apply and\nso we had to answer that question in uh\n\nso we had to answer that question in uh\n \n\nso we had to answer that question in uh\na multitude of ways and that's what I'm\n\na multitude of ways and that's what I'm\n \n\na multitude of ways and that's what I'm\nhere to talk\n\nhere to talk\n \n\nhere to talk\nabout okay so I want to start out with\n\nabout okay so I want to start out with\n \n\nabout okay so I want to start out with\nhow I would describe the value prop of\n\nhow I would describe the value prop of\n \n\nhow I would describe the value prop of\nAI to someone who's actually using it to\n\nAI to someone who's actually using it to\n \n\nAI to someone who's actually using it to\na\n\na\n \n\na\ndeveloper so coding without AI\n\ndeveloper so coding without AI\n \n\ndeveloper so coding without AI\nI think you know we've all felt this\n\nI think you know we've all felt this\n \n\nI think you know we've all felt this\nbefore if you've ever written a line of\n\nbefore if you've ever written a line of\n \n\nbefore if you've ever written a line of\ncode uh you start out by asking yourself\n\ncode uh you start out by asking yourself\n \n\ncode uh you start out by asking yourself\nlike oh this task it's straightforward\n\nlike oh this task it's straightforward\n \n\nlike oh this task it's straightforward\nit should be easy let me just go build\n\nit should be easy let me just go build\n \n\nit should be easy let me just go build\nthis feature I should mention I poached\n\nthis feature I should mention I poached\n \n\nthis feature I should mention I poached\nthese slides uh from um uh a director of\n\nthese slides uh from um uh a director of\n \n\nthese slides uh from um uh a director of\nengineering at Apollo ET networks he's\n\nengineering at Apollo ET networks he's\n \n\nengineering at Apollo ET networks he's\nactually giving another talk at this\n\nactually giving another talk at this\n \n\nactually giving another talk at this\nconference gunjam Patel I thought he did\n\nconference gunjam Patel I thought he did\n \n\nconference gunjam Patel I thought he did\nan excellent job of describing the value\n\nan excellent job of describing the value\n \n\nan excellent job of describing the value\nprop that he was solving for as a\n\nprop that he was solving for as a\n \n\nprop that he was solving for as a\ndirector of engineering when they were\n\ndirector of engineering when they were\n \n\ndirector of engineering when they were\npurchasing uh a coding AI assistant so\n\npurchasing uh a coding AI assistant so\n \n\npurchasing uh a coding AI assistant so\nso this is the way he described it you\n\nso this is the way he described it you\n \n\nso this is the way he described it you\nthink it should be straightforward but\n\nthink it should be straightforward but\n \n\nthink it should be straightforward but\nthen there's all these side quests that\n\nthen there's all these side quests that\n \n\nthen there's all these side quests that\nyou end up going on as a developer it's\n\nyou end up going on as a developer it's\n \n\nyou end up going on as a developer it's\nlike uhoh like I got to go install a\n\nlike uhoh like I got to go install a\n \n\nlike uhoh like I got to go install a\nbunch of dependencies or maybe you know\n\nbunch of dependencies or maybe you know\n \n\nbunch of dependencies or maybe you know\nthere's this uh UI component that I have\n\nthere's this uh UI component that I have\n \n\nthere's this uh UI component that I have\nto go and figure out this framework that\n\nto go and figure out this framework that\n \n\nto go and figure out this framework that\nI need to learn so the gaps appear and\n\nI need to learn so the gaps appear and\n \n\nI need to learn so the gaps appear and\nthen without AI Bridging the gaps uh\n\nthen without AI Bridging the gaps uh\n \n\nthen without AI Bridging the gaps uh\ntakes both time and focus you kind of\n\ntakes both time and focus you kind of\n \n\ntakes both time and focus you kind of\nhave to you know spin off a side process\n\nhave to you know spin off a side process\n \n\nhave to you know spin off a side process\nor go on a little mini quest uh and then\n\nor go on a little mini quest uh and then\n \n\nor go on a little mini quest uh and then\nyou know 30 minutes and two cups of\n\nyou know 30 minutes and two cups of\n \n\nyou know 30 minutes and two cups of\ncoffee later you're like okay I got it I\n\ncoffee later you're like okay I got it I\n \n\ncoffee later you're like okay I got it I\nfilled the Gap but what was I doing\n\nfilled the Gap but what was I doing\n \n\nfilled the Gap but what was I doing\nagain and so AI helps Bridge those gaps\n\nagain and so AI helps Bridge those gaps\n \n\nagain and so AI helps Bridge those gaps\nit helps solve this problem it helps uh\n\nit helps solve this problem it helps uh\n \n\nit helps solve this problem it helps uh\ndevelopers really stay in flow and stay\n\ndevelopers really stay in flow and stay\n \n\ndevelopers really stay in flow and stay\nuh kind of like cognizant of the high\n\nuh kind of like cognizant of the high\n \n\nuh kind of like cognizant of the high\nlevel of what they're trying to\n\nlevel of what they're trying to\n \n\nlevel of what they're trying to\naccomplish and so what this means is\n\naccomplish and so what this means is\n \n\naccomplish and so what this means is\nthat more and more developers can\n\nthat more and more developers can\n \n\nthat more and more developers can\nactually do the thing that that we want\n\nactually do the thing that that we want\n \n\nactually do the thing that that we want\nto do which is build an amazing feature\n\nto do which is build an amazing feature\n \n\nto do which is build an amazing feature\nand deliver an amazing experience um\n\nand deliver an amazing experience um\n \n\nand deliver an amazing experience um\ninstead of giving up now the question is\n\ninstead of giving up now the question is\n \n\ninstead of giving up now the question is\nhow do we measure this in a way that we\n\nhow do we measure this in a way that we\n \n\nhow do we measure this in a way that we\ncan uh demonstrate the business impact\n\ncan uh demonstrate the business impact\n \n\ncan uh demonstrate the business impact\nof what this does to the rest of the\n\nof what this does to the rest of the\n \n\nof what this does to the rest of the\norganization and so the answer is\n\norganization and so the answer is\n \n\norganization and so the answer is\nbeans okay who here drinks\n\nbeans okay who here drinks\n \n\nbeans okay who here drinks\ncoffee okay cool do you know the\n\ncoffee okay cool do you know the\n \n\ncoffee okay cool do you know the\ndifference between a really great uh\n\ndifference between a really great uh\n \n\ndifference between a really great uh\nCoffee Bean and you know your kind of\n\nCoffee Bean and you know your kind of\n \n\nCoffee Bean and you know your kind of\nrun-of-the-mill folders or you know the\n\nrun-of-the-mill folders or you know the\n \n\nrun-of-the-mill folders or you know the\nthing that you buy at the\n\nthing that you buy at the\n \n\nthing that you buy at the\nsupermarket okay turns out we're all in\n\nsupermarket okay turns out we're all in\n \n\nsupermarket okay turns out we're all in\nthe bean business we think we're in the\n\nthe bean business we think we're in the\n \n\nthe bean business we think we're in the\nsoftware business but we're really\n\nsoftware business but we're really\n \n\nsoftware business but we're really\nselling beans in a way so in every\n\nselling beans in a way so in every\n \n\nselling beans in a way so in every\ncompany there is someone uh let's call\n\ncompany there is someone uh let's call\n \n\ncompany there is someone uh let's call\nhim Bob who grows the beans essentially\n\nhim Bob who grows the beans essentially\n \n\nhim Bob who grows the beans essentially\nthe developer uh there's another person\n\nthe developer uh there's another person\n \n\nthe developer uh there's another person\nit's called Uh him Pat who sells the\n\nit's called Uh him Pat who sells the\n \n\nit's called Uh him Pat who sells the\nbeans that's your kind of cro or sales\n\nbeans that's your kind of cro or sales\n \n\nbeans that's your kind of cro or sales\nlead and then you have Alice on the side\n\nlead and then you have Alice on the side\n \n\nlead and then you have Alice on the side\nwho's kind of your CFO uh or or CEO uh\n\nwho's kind of your CFO uh or or CEO uh\n \n\nwho's kind of your CFO uh or or CEO uh\nand Alice has got to count the beans at\n\nand Alice has got to count the beans at\n \n\nand Alice has got to count the beans at\nthe end of the day you know uh not to\n\nthe end of the day you know uh not to\n \n\nthe end of the day you know uh not to\ndiss the finance people if there are any\n\ndiss the finance people if there are any\n \n\ndiss the finance people if there are any\nin the room but like it's all about\n\nin the room but like it's all about\n \n\nin the room but like it's all about\ncounting the beans and seeing how they\n\ncounting the beans and seeing how they\n \n\ncounting the beans and seeing how they\nadd up\n\nadd up\n \n\nadd up\nnow I don't know if any of you have been\n\nnow I don't know if any of you have been\n \n\nnow I don't know if any of you have been\npaying attention but in the past two\n\npaying attention but in the past two\n \n\npaying attention but in the past two\nyears the bean business has been\n\nyears the bean business has been\n \n\nyears the bean business has been\nrevolutionized this thing called AI has\n\nrevolutionized this thing called AI has\n \n\nrevolutionized this thing called AI has\nappeared and so what does the bean\n\nappeared and so what does the bean\n \n\nappeared and so what does the bean\nbusiness look like\n\nbusiness look like\n \n\nbusiness look like\nnow well Bob grows the beans with AI and\n\nnow well Bob grows the beans with AI and\n \n\nnow well Bob grows the beans with AI and\nPat is selling beans but with AI and\n\nPat is selling beans but with AI and\n \n\nPat is selling beans but with AI and\nthen Alice is on the side being like\n\nthen Alice is on the side being like\n \n\nthen Alice is on the side being like\nwell I'm counting all the beans and\n\nwell I'm counting all the beans and\n \n\nwell I'm counting all the beans and\nwhere's the ROI and so this is the\n\nwhere's the ROI and so this is the\n \n\nwhere's the ROI and so this is the\nanswer that you know basically Bob has\n\nanswer that you know basically Bob has\n \n\nanswer that you know basically Bob has\nto answer Pat has kind of got it easy\n\nto answer Pat has kind of got it easy\n \n\nto answer Pat has kind of got it easy\nbecause Pat's job is just selling the\n\nbecause Pat's job is just selling the\n \n\nbecause Pat's job is just selling the\nbeans that's a much more quantifiable uh\n\nbeans that's a much more quantifiable uh\n \n\nbeans that's a much more quantifiable uh\ntask um those of us that are involved in\n\ntask um those of us that are involved in\n \n\ntask um those of us that are involved in\nsoftware engineering and product\n\nsoftware engineering and product\n \n\nsoftware engineering and product\ndevelopment it's a bit harder to measure\n\ndevelopment it's a bit harder to measure\n \n\ndevelopment it's a bit harder to measure\nso there's tension in the bean shop you\n\nso there's tension in the bean shop you\n \n\nso there's tension in the bean shop you\nknow uh Alice is asking Bob you know how\n\nknow uh Alice is asking Bob you know how\n \n\nknow uh Alice is asking Bob you know how\nmany more beans are we growing now with\n\nmany more beans are we growing now with\n \n\nmany more beans are we growing now with\nAI Bob and Bob's like well it's\n\nAI Bob and Bob's like well it's\n \n\nAI Bob and Bob's like well it's\ncomplicated Alice not all beans are the\n\ncomplicated Alice not all beans are the\n \n\ncomplicated Alice not all beans are the\nsame you know there's some good beans\n\nsame you know there's some good beans\n \n\nsame you know there's some good beans\nand there's some very bad beans we're\n\nand there's some very bad beans we're\n \n\nand there's some very bad beans we're\nmaking our beans better and then Al is\n\nmaking our beans better and then Al is\n \n\nmaking our beans better and then Al is\nlike well okay the bean AI tool costs\n\nlike well okay the bean AI tool costs\n \n\nlike well okay the bean AI tool costs\nmoney and we got to measure its impact\n\nmoney and we got to measure its impact\n \n\nmoney and we got to measure its impact\nsomehow\n\nsomehow\n \n\nsomehow\nhow anyone feel that tension anyone have\n\nhow anyone feel that tension anyone have\n \n\nhow anyone feel that tension anyone have\nthis kind of like\n\nthis kind of like\n \n\nthis kind of like\nconversation we we've talked to a lot of\n\nconversation we we've talked to a lot of\n \n\nconversation we we've talked to a lot of\nheads of engineering who see this\n\nheads of engineering who see this\n \n\nheads of engineering who see this\ntension very real with other parts of\n\ntension very real with other parts of\n \n\ntension very real with other parts of\nthe org specifically between finance and\n\nthe org specifically between finance and\n \n\nthe org specifically between finance and\nengineering and I think the core of the\n\nengineering and I think the core of the\n \n\nengineering and I think the core of the\nproblem is that measuring aii for\n\nproblem is that measuring aii for\n \n\nproblem is that measuring aii for\nfunctions where the work is not directly\n\nfunctions where the work is not directly\n \n\nfunctions where the work is not directly\nquantifiable through a number is what I\n\nquantifiable through a number is what I\n \n\nquantifiable through a number is what I\nlike to call MP hard so how many people\n\nlike to call MP hard so how many people\n \n\nlike to call MP hard so how many people\nare familiar with the term NP hard here\n\nare familiar with the term NP hard here\n \n\nare familiar with the term NP hard here\nokay cool we're all pretty technical so\n\nokay cool we're all pretty technical so\n \n\nokay cool we're all pretty technical so\nNP hard basically means if you have a\n\nNP hard basically means if you have a\n \n\nNP hard basically means if you have a\ntough challenge uh you have a problem\n\ntough challenge uh you have a problem\n \n\ntough challenge uh you have a problem\nand you can basically reduce it to a\n\nand you can basically reduce it to a\n \n\nand you can basically reduce it to a\nclass of very uh difficult problems it\n\nclass of very uh difficult problems it\n \n\nclass of very uh difficult problems it\nprobably means your problem is not\n\nprobably means your problem is not\n \n\nprobably means your problem is not\nsolvable and measuring AI Roi reduces to\n\nsolvable and measuring AI Roi reduces to\n \n\nsolvable and measuring AI Roi reduces to\nmeasuring developer productivity or the\n\nmeasuring developer productivity or the\n \n\nmeasuring developer productivity or the\nproductivity of whatever class of\n\nproductivity of whatever class of\n \n\nproductivity of whatever class of\nknowledge worker uh that you're managing\n\nknowledge worker uh that you're managing\n \n\nknowledge worker uh that you're managing\nuh and so that implies if you can\n\nuh and so that implies if you can\n \n\nuh and so that implies if you can\nmeasure the AI Roi precisely can also\n\nmeasure the AI Roi precisely can also\n \n\nmeasure the AI Roi precisely can also\nmeasure developer productivity and who\n\nmeasure developer productivity and who\n \n\nmeasure developer productivity and who\nhere knows how to measure developer\n\nhere knows how to measure developer\n \n\nhere knows how to measure developer\nproductivity it's kind of an open\n\nproductivity it's kind of an open\n \n\nproductivity it's kind of an open\nquestion right so uh using the logic of\n\nquestion right so uh using the logic of\n \n\nquestion right so uh using the logic of\nyour standard reduction proof uh this\n\nyour standard reduction proof uh this\n \n\nyour standard reduction proof uh this\nproblem is intractable so that's the end\n\nproblem is intractable so that's the end\n \n\nproblem is intractable so that's the end\nof my talk I'm just here to tell you\n\nof my talk I'm just here to tell you\n \n\nof my talk I'm just here to tell you\nthat this problem is intractable and we\n\nthat this problem is intractable and we\n \n\nthat this problem is intractable and we\nshould give up right well in the real\n\nshould give up right well in the real\n \n\nshould give up right well in the real\nworld we often find tractable Solutions\n\nworld we often find tractable Solutions\n \n\nworld we often find tractable Solutions\nuh to intractable problems and so uh\n\nuh to intractable problems and so uh\n \n\nuh to intractable problems and so uh\nwhat the meat of this talk is is really\n\nwhat the meat of this talk is is really\n \n\nwhat the meat of this talk is is really\nsharing a set of evaluation Frameworks\n\nsharing a set of evaluation Frameworks\n \n\nsharing a set of evaluation Frameworks\nthat we've presented to different\n\nthat we've presented to different\n \n\nthat we've presented to different\ncustomers uh not all of these are used\n\ncustomers uh not all of these are used\n \n\ncustomers uh not all of these are used\nby uh you know any given customer um but\n\nby uh you know any given customer um but\n \n\nby uh you know any given customer um but\nI wanted to give kind of like a sampling\n\nI wanted to give kind of like a sampling\n \n\nI wanted to give kind of like a sampling\nof the conversations that we've had uh\n\nof the conversations that we've had uh\n \n\nof the conversations that we've had uh\nand hopefully there'll be some time for\n\nand hopefully there'll be some time for\n \n\nand hopefully there'll be some time for\nQ&A at the end where we can kind of talk\n\nQ&A at the end where we can kind of talk\n \n\nQ&A at the end where we can kind of talk\nthrough this and and uh see what other\n\nthrough this and and uh see what other\n \n\nthrough this and and uh see what other\npeople are\n\npeople are\n \n\npeople are\ndoing okay so framework number one uh is\n\ndoing okay so framework number one uh is\n \n\ndoing okay so framework number one uh is\nthe famous rolls eliminated uh framework\n\nthe famous rolls eliminated uh framework\n \n\nthe famous rolls eliminated uh framework\nso this question gets asked a lot these\n\nso this question gets asked a lot these\n \n\nso this question gets asked a lot these\ndays especially you know on social media\n\ndays especially you know on social media\n \n\ndays especially you know on social media\nlike AI is here to take your job so how\n\nlike AI is here to take your job so how\n \n\nlike AI is here to take your job so how\ndoes this framework work well in the\n\ndoes this framework work well in the\n \n\ndoes this framework work well in the\nclassic framing you buy the tool the the\n\nclassic framing you buy the tool the the\n \n\nclassic framing you buy the tool the the\nlabor saving tool You observe you know\n\nlabor saving tool You observe you know\n \n\nlabor saving tool You observe you know\nyou're in the bean business or the the\n\nyou're in the bean business or the the\n \n\nyou're in the bean business or the the\nwidget business You observe that this\n\nwidget business You observe that this\n \n\nwidget business You observe that this\ntool yields an x% increase uh in Your\n\ntool yields an x% increase uh in Your\n \n\ntool yields an x% increase uh in Your\nCapacity to build widgets and then you\n\nCapacity to build widgets and then you\n \n\nCapacity to build widgets and then you\ncan cut your Workforce to meet the\n\ncan cut your Workforce to meet the\n \n\ncan cut your Workforce to meet the\ndemands for whatever widgets you're\n\ndemands for whatever widgets you're\n \n\ndemands for whatever widgets you're\nselling now in practice we have not\n\nselling now in practice we have not\n \n\nselling now in practice we have not\nencountered this framework at all in the\n\nencountered this framework at all in the\n \n\nencountered this framework at all in the\nrealm of software development um we do\n\nrealm of software development um we do\n \n\nrealm of software development um we do\nsee it more prevalent in other kind of\n\nsee it more prevalent in other kind of\n \n\nsee it more prevalent in other kind of\nbusiness units you know things that are\n\nbusiness units you know things that are\n \n\nbusiness units you know things that are\nmore viewed as like cost centers like uh\n\nmore viewed as like cost centers like uh\n \n\nmore viewed as like cost centers like uh\nsupport and things like that um\n\nsupport and things like that um\n \n\nsupport and things like that um\nespecially like uh consumer facing\n\nespecially like uh consumer facing\n \n\nespecially like uh consumer facing\ncustomer support but for software\n\ncustomer support but for software\n \n\ncustomer support but for software\nengineering uh for whatever reason we\n\nengineering uh for whatever reason we\n \n\nengineering uh for whatever reason we\nhaven't encountered this yet in any of\n\nhaven't encountered this yet in any of\n \n\nhaven't encountered this yet in any of\nour customers and we think that the\n\nour customers and we think that the\n \n\nour customers and we think that the\nreason here is that number one you know\n\nreason here is that number one you know\n \n\nreason here is that number one you know\nif you view your org as a widget Factory\n\nif you view your org as a widget Factory\n \n\nif you view your org as a widget Factory\nuh then you're going to prioritize\n\nuh then you're going to prioritize\n \n\nuh then you're going to prioritize\noutputting widgets but the the thing is\n\noutputting widgets but the the thing is\n \n\noutputting widgets but the the thing is\nthat very few engineering leaders\n\nthat very few engineering leaders\n \n\nthat very few engineering leaders\neffective engineering leaders these days\n\neffective engineering leaders these days\n \n\neffective engineering leaders these days\nview themselves as widget Builders you\n\nview themselves as widget Builders you\n \n\nview themselves as widget Builders you\nknow the widgets are kind of an\n\nknow the widgets are kind of an\n \n\nknow the widgets are kind of an\nexraction uh that don't apply to the\n\nexraction uh that don't apply to the\n \n\nexraction uh that don't apply to the\ncraft of software\n\ncraft of software\n \n\ncraft of software\nengineering the other observation here\n\nengineering the other observation here\n \n\nengineering the other observation here\nis that the widget that we're building\n\nis that the widget that we're building\n \n\nis that the widget that we're building\nwhich is software at today great user\n\nwhich is software at today great user\n \n\nwhich is software at today great user\nexperiences they're not really Supply\n\nexperiences they're not really Supply\n \n\nexperiences they're not really Supply\nlimited so if you have like an extensive\n\nlimited so if you have like an extensive\n \n\nlimited so if you have like an extensive\nbacklog question is you know if we made\n\nbacklog question is you know if we made\n \n\nbacklog question is you know if we made\nyour engineers 20% more productive would\n\nyour engineers 20% more productive would\n \n\nyour engineers 20% more productive would\nyou go and you know do more uh 20% more\n\nyou go and you know do more uh 20% more\n \n\nyou go and you know do more uh 20% more\nof your backlog or you just cut down 20%\n\nof your backlog or you just cut down 20%\n \n\nof your backlog or you just cut down 20%\nof your Workforce and say like you know\n\nof your Workforce and say like you know\n \n\nof your Workforce and say like you know\nit's fine we don't need to get to the\n\nit's fine we don't need to get to the\n \n\nit's fine we don't need to get to the\nbacklog and for 99% of the companies out\n\nbacklog and for 99% of the companies out\n \n\nbacklog and for 99% of the companies out\nthere that building software the answer\n\nthere that building software the answer\n \n\nthere that building software the answer\nis no we want to build a better user\n\nis no we want to build a better user\n \n\nis no we want to build a better user\nexperience these issues in our backlog\n\nexperience these issues in our backlog\n \n\nexperience these issues in our backlog\nare very important we just can't get to\n\nare very important we just can't get to\n \n\nare very important we just can't get to\nthem so framework number one is kind of\n\nthem so framework number one is kind of\n \n\nthem so framework number one is kind of\nlike uh talked about frequently but in\n\nlike uh talked about frequently but in\n \n\nlike uh talked about frequently but in\npractice we haven't really seen it as an\n\npractice we haven't really seen it as an\n \n\npractice we haven't really seen it as an\nevaluation\n\nevaluation\n \n\nevaluation\ncriteria framework number two is what I\n\ncriteria framework number two is what I\n \n\ncriteria framework number two is what I\nlike to call AB testing uh velocity so\n\nlike to call AB testing uh velocity so\n \n\nlike to call AB testing uh velocity so\nhow this works is you basically segment\n\nhow this works is you basically segment\n \n\nhow this works is you basically segment\noff uh your organization into two groups\n\noff uh your organization into two groups\n \n\noff uh your organization into two groups\nuh the test group and the control group\n\nuh the test group and the control group\n \n\nuh the test group and the control group\nand then uh you say group one gets the\n\nand then uh you say group one gets the\n \n\nand then uh you say group one gets the\ntool and group two does not and then you\n\ntool and group two does not and then you\n \n\ntool and group two does not and then you\ngo through your standard planning\n\ngo through your standard planning\n \n\ngo through your standard planning\nprocess most people uh as part of that\n\nprocess most people uh as part of that\n \n\nprocess most people uh as part of that\nplanning process what you do is you go\n\nplanning process what you do is you go\n \n\nplanning process what you do is you go\nand estimate the time that it will take\n\nand estimate the time that it will take\n \n\nand estimate the time that it will take\nto resolve uh certain issues so you know\n\nto resolve uh certain issues so you know\n \n\nto resolve uh certain issues so you know\nhow long is this feature going to take\n\nhow long is this feature going to take\n \n\nhow long is this feature going to take\nto build how how long is it going to\n\nto build how how long is it going to\n \n\nto build how how long is it going to\ntake to work through these bug\n\ntake to work through these bug\n \n\ntake to work through these bug\nbacklogs and then because you've divided\n\nbacklogs and then because you've divided\n \n\nbacklogs and then because you've divided\nthe groups into two now uh you have some\n\nthe groups into two now uh you have some\n \n\nthe groups into two now uh you have some\nnotion of like you know how well you're\n\nnotion of like you know how well you're\n \n\nnotion of like you know how well you're\nexecuting against your your timeline so\n\nexecuting against your your timeline so\n \n\nexecuting against your your timeline so\nyou basically run this AB test\n\nyou basically run this AB test\n \n\nyou basically run this AB test\num so palala networks one of our\n\num so palala networks one of our\n \n\num so palala networks one of our\ncustomers uh ran something similar to\n\ncustomers uh ran something similar to\n \n\ncustomers uh ran something similar to\nthis uh and the conclusion they drew was\n\nthis uh and the conclusion they drew was\n \n\nthis uh and the conclusion they drew was\nthe the approximate timelines got\n\nthe the approximate timelines got\n \n\nthe the approximate timelines got\naccelerated 20 to 30% uh using Cody and\n\naccelerated 20 to 30% uh using Cody and\n \n\naccelerated 20 to 30% uh using Cody and\nso this is a very kind of like rigorous\n\nso this is a very kind of like rigorous\n \n\nso this is a very kind of like rigorous\nscientific framework um we see it come\n\nscientific framework um we see it come\n \n\nscientific framework um we see it come\nup now and then uh especially when when\n\nup now and then uh especially when when\n \n\nup now and then uh especially when when\ncompanies are of a certain size and and\n\ncompanies are of a certain size and and\n \n\ncompanies are of a certain size and and\nthey they're very thoughtful about this\n\nthey they're very thoughtful about this\n \n\nthey they're very thoughtful about this\nquestion um the criticisms about this\n\nquestion um the criticisms about this\n \n\nquestion um the criticisms about this\nframework are no two teams are exactly\n\nframework are no two teams are exactly\n \n\nframework are no two teams are exactly\nthe same right like if you poop off your\n\nthe same right like if you poop off your\n \n\nthe same right like if you poop off your\ndevelopment org you have you know Dev\n\ndevelopment org you have you know Dev\n \n\ndevelopment org you have you know Dev\ninfrastructure on this side maybe\n\ninfrastructure on this side maybe\n \n\ninfrastructure on this side maybe\nbackend and then you have front end\n\nbackend and then you have front end\n \n\nbackend and then you have front end\nteams on this side it's hard sometimes\n\nteams on this side it's hard sometimes\n \n\nteams on this side it's hard sometimes\nto compare these these different teams\n\nto compare these these different teams\n \n\nto compare these these different teams\nto each other because uh software\n\nto each other because uh software\n \n\nto each other because uh software\ndevelopment is very different uh in\n\ndevelopment is very different uh in\n \n\ndevelopment is very different uh in\ndifferent parts of your\n\ndifferent parts of your\n \n\ndifferent parts of your\norganization there's also confounding\n\norganization there's also confounding\n \n\norganization there's also confounding\nfactors you know maybe team X you know\n\nfactors you know maybe team X you know\n \n\nfactors you know maybe team X you know\nhad uh an important leader or\n\nhad uh an important leader or\n \n\nhad uh an important leader or\ncontributor to part uh maybe team y uh\n\ncontributor to part uh maybe team y uh\n \n\ncontributor to part uh maybe team y uh\nyou know suffered a bout of covid that\n\nyou know suffered a bout of covid that\n \n\nyou know suffered a bout of covid that\nyou know blew through the team or things\n\nyou know blew through the team or things\n \n\nyou know blew through the team or things\nlike that you have to account for these\n\nlike that you have to account for these\n \n\nlike that you have to account for these\nthings when when make your evaluation\n\nthings when when make your evaluation\n \n\nthings when when make your evaluation\nand this framework is also high cost and\n\nand this framework is also high cost and\n \n\nand this framework is also high cost and\neffort you basically have to do the\n\neffort you basically have to do the\n \n\neffort you basically have to do the\nsubdivision you give one group access of\n\nsubdivision you give one group access of\n \n\nsubdivision you give one group access of\nthe tool and then you have to run it for\n\nthe tool and then you have to run it for\n \n\nthe tool and then you have to run it for\nan extended period of time in order to\n\nan extended period of time in order to\n \n\nan extended period of time in order to\ngain enough\n\ngain enough\n \n\ngain enough\nconfidence but provided you have the\n\nconfidence but provided you have the\n \n\nconfidence but provided you have the\nresources and the time to estimate it we\n\nresources and the time to estimate it we\n \n\nresources and the time to estimate it we\nthink this is a pretty good framework\n\nthink this is a pretty good framework\n \n\nthink this is a pretty good framework\nfor honestly testing the efficacy of of\n\nfor honestly testing the efficacy of of\n \n\nfor honestly testing the efficacy of of\nan AI\n\n \n\ntool okay framework number three um I\n\ntool okay framework number three um I\n \n\ntool okay framework number three um I\ncall this time saved as a function of\n\ncall this time saved as a function of\n \n\ncall this time saved as a function of\nEngagement so if you have have a\n\nEngagement so if you have have a\n \n\nEngagement so if you have have a\nproductivity\n\nproductivity\n \n\nproductivity\ntool um using the product should make\n\ntool um using the product should make\n \n\ntool um using the product should make\npeople more productive right so if you\n\npeople more productive right so if you\n \n\npeople more productive right so if you\nhave a code search engine uh the more\n\nhave a code search engine uh the more\n \n\nhave a code search engine uh the more\ncode searches that people do uh that\n\ncode searches that people do uh that\n \n\ncode searches that people do uh that\nthat probably saves them time if it\n\nthat probably saves them time if it\n \n\nthat probably saves them time if it\ndidn't save them time there would be no\n\ndidn't save them time there would be no\n \n\ndidn't save them time there would be no\nreason why they would go to the search\n\nreason why they would go to the search\n \n\nreason why they would go to the search\nengine and so in this framework what you\n\nengine and so in this framework what you\n \n\nengine and so in this framework what you\ndo is you basically go look at your\n\ndo is you basically go look at your\n \n\ndo is you basically go look at your\nproduct metrics and you break down all\n\nproduct metrics and you break down all\n \n\nproduct metrics and you break down all\nthe different time-saving actions uh you\n\nthe different time-saving actions uh you\n \n\nthe different time-saving actions uh you\nidentify them and then you kind of tag\n\nidentify them and then you kind of tag\n \n\nidentify them and then you kind of tag\nthem with an approximate estimate of how\n\nthem with an approximate estimate of how\n \n\nthem with an approximate estimate of how\nmuch time is saved in each action and if\n\nmuch time is saved in each action and if\n \n\nmuch time is saved in each action and if\nyou want to be conservative you can\n\nyou want to be conservative you can\n \n\nyou want to be conservative you can\nlower bound it you know like you could\n\nlower bound it you know like you could\n \n\nlower bound it you know like you could\nsay like a code search probably saves Me\n\nsay like a code search probably saves Me\n \n\nsay like a code search probably saves Me\n2 minutes that's maybe like a a lower\n\n2 minutes that's maybe like a a lower\n \n\n2 minutes that's maybe like a a lower\nbound because there's there's definitely\n\nbound because there's there's definitely\n \n\nbound because there's there's definitely\nlike searches where oh my gosh like it\n\nlike searches where oh my gosh like it\n \n\nlike searches where oh my gosh like it\nsaved me like half a day or uh maybe\n\nsaved me like half a day or uh maybe\n \n\nsaved me like half a day or uh maybe\nlike a whole week of work because it\n\nlike a whole week of work because it\n \n\nlike a whole week of work because it\nprevented me from going down uh an\n\nprevented me from going down uh an\n \n\nprevented me from going down uh an\nunproductive rabbit hole but you can\n\nunproductive rabbit hole but you can\n \n\nunproductive rabbit hole but you can\nlower bound it and just say like okay\n\nlower bound it and just say like okay\n \n\nlower bound it and just say like okay\nwe're going to get a lower bound\n\nwe're going to get a lower bound\n \n\nwe're going to get a lower bound\nestimate on the total amount time saved\n\nestimate on the total amount time saved\n \n\nestimate on the total amount time saved\nand then uh you go and ask your vendor\n\nand then uh you go and ask your vendor\n \n\nand then uh you go and ask your vendor\nhey can you build some analytics and\n\nhey can you build some analytics and\n \n\nhey can you build some analytics and\nshare them with me so this is something\n\nshare them with me so this is something\n \n\nshare them with me so this is something\nthat we built for Cody uh you know very\n\nthat we built for Cody uh you know very\n \n\nthat we built for Cody uh you know very\nfine grain analytics to show uh the\n\nfine grain analytics to show uh the\n \n\nfine grain analytics to show uh the\nadmins and the leaders in the org\n\nadmins and the leaders in the org\n \n\nadmins and the leaders in the org\nexactly what actions are being uh taken\n\nexactly what actions are being uh taken\n \n\nexactly what actions are being uh taken\nyou know how many explicit invocations\n\nyou know how many explicit invocations\n \n\nyou know how many explicit invocations\nyou know how many chats how many\n\nyou know how many chats how many\n \n\nyou know how many chats how many\nquestions about the code base how many\n\nquestions about the code base how many\n \n\nquestions about the code base how many\ninline code generation actions and those\n\ninline code generation actions and those\n \n\ninline code generation actions and those\nall map to a certain amount of time\n\nall map to a certain amount of time\n \n\nall map to a certain amount of time\nsaved there's one caveat here which is\n\nsaved there's one caveat here which is\n \n\nsaved there's one caveat here which is\nin uh products where you have like an\n\nin uh products where you have like an\n \n\nin uh products where you have like an\nimplicit trigger like an autocomplete um\n\nimplicit trigger like an autocomplete um\n \n\nimplicit trigger like an autocomplete um\nyou can't rely purely on engagement\n\nyou can't rely purely on engagement\n \n\nyou can't rely purely on engagement\nbecause it's not the human opting in to\n\nbecause it's not the human opting in to\n \n\nbecause it's not the human opting in to\nengage the product uh each time it's\n\nengage the product uh each time it's\n \n\nengage the product uh each time it's\nsort of like implicitly shown to you um\n\nsort of like implicitly shown to you um\n \n\nsort of like implicitly shown to you um\nand so there you tend to go with uh more\n\nand so there you tend to go with uh more\n \n\nand so there you tend to go with uh more\nof an acceptance rate criteria you don't\n\nof an acceptance rate criteria you don't\n \n\nof an acceptance rate criteria you don't\nwant to do just raw engagements because\n\nwant to do just raw engagements because\n \n\nwant to do just raw engagements because\nthen the product could just push a bunch\n\nthen the product could just push a bunch\n \n\nthen the product could just push a bunch\nof like lowquality uh completions to you\n\nof like lowquality uh completions to you\n \n\nof like lowquality uh completions to you\nand that would not be a good uh meas\n\nand that would not be a good uh meas\n \n\nand that would not be a good uh meas\nmetric of time\n\nmetric of time\n \n\nmetric of time\nsaved so we have a lot of customers that\n\nsaved so we have a lot of customers that\n \n\nsaved so we have a lot of customers that\ndo this they appreciate it one of the\n\ndo this they appreciate it one of the\n \n\ndo this they appreciate it one of the\nnice things about this is that because\n\nnice things about this is that because\n \n\nnice things about this is that because\nwe're lower bounding it makes the the\n\nwe're lower bounding it makes the the\n \n\nwe're lower bounding it makes the the\nthe value of the software very clear um\n\nthe value of the software very clear um\n \n\nthe value of the software very clear um\nso a lot of uh developer tools us\n\nso a lot of uh developer tools us\n \n\nso a lot of uh developer tools us\nincluded I think like Cody is like $9\n\nincluded I think like Cody is like $9\n \n\nincluded I think like Cody is like $9\nper month and Source craft is a little\n\nper month and Source craft is a little\n \n\nper month and Source craft is a little\nbit more than that but it's like if you\n\nbit more than that but it's like if you\n \n\nbit more than that but it's like if you\nback out the math of how much an hour of\n\nback out the math of how much an hour of\n \n\nback out the math of how much an hour of\na developer's time is worth you know\n\na developer's time is worth you know\n \n\na developer's time is worth you know\ntypically it's around like\n\ntypically it's around like\n \n\ntypically it's around like\n$1200 it's like if you save you know a\n\n$1200 it's like if you save you know a\n \n\n$1200 it's like if you save you know a\ncouple minutes a month this kind of pays\n\ncouple minutes a month this kind of pays\n \n\ncouple minutes a month this kind of pays\nfor itself in terms of\n\nfor itself in terms of\n \n\nfor itself in terms of\nproductivity um the criticisms of this\n\nproductivity um the criticisms of this\n \n\nproductivity um the criticisms of this\nframework is of course it's a lower\n\nframework is of course it's a lower\n \n\nframework is of course it's a lower\nbound so you're not fully assessing the\n\nbound so you're not fully assessing the\n \n\nbound so you're not fully assessing the\nvalue if you go back to that picture I\n\nvalue if you go back to that picture I\n \n\nvalue if you go back to that picture I\nshowed earlier of you know the dev\n\nshowed earlier of you know the dev\n \n\nshowed earlier of you know the dev\nJourney where you're kind of Bridging\n\nJourney where you're kind of Bridging\n \n\nJourney where you're kind of Bridging\nthe gaps uh I think a big value of AI is\n\nthe gaps uh I think a big value of AI is\n \n\nthe gaps uh I think a big value of AI is\nactually completing tasks that hither to\n\nactually completing tasks that hither to\n \n\nactually completing tasks that hither to\nor uh beforehand were just not completed\n\nor uh beforehand were just not completed\n \n\nor uh beforehand were just not completed\nbecause people got fed up uh or they got\n\nbecause people got fed up uh or they got\n \n\nbecause people got fed up uh or they got\nlazy or they just had other things to do\n\nlazy or they just had other things to do\n \n\nlazy or they just had other things to do\nso this doesn't capture that um it\n\nso this doesn't capture that um it\n \n\nso this doesn't capture that um it\ndoesn't account for the second order\n\ndoesn't account for the second order\n \n\ndoesn't account for the second order\neffects of the Velocity boost um and\n\neffects of the Velocity boost um and\n \n\neffects of the Velocity boost um and\nit's good for kind of like day-to-day\n\nit's good for kind of like day-to-day\n \n\nit's good for kind of like day-to-day\nlike hey is this speeding up the team\n\nlike hey is this speeding up the team\n \n\nlike hey is this speeding up the team\nbut it doesn't capture some of the\n\nbut it doesn't capture some of the\n \n\nbut it doesn't capture some of the\nimpact on key initiatives of the\n\nimpact on key initiatives of the\n \n\nimpact on key initiatives of the\ncompany so that leads to the fourth\n\ncompany so that leads to the fourth\n \n\ncompany so that leads to the fourth\nevaluation framework so a lot of our\n\nevaluation framework so a lot of our\n \n\nevaluation framework so a lot of our\ncustomers track certain kpis that they\n\ncustomers track certain kpis that they\n \n\ncustomers track certain kpis that they\nthey think are correlated with\n\nthey think are correlated with\n \n\nthey think are correlated with\nengineering quality and business impact\n\nengineering quality and business impact\n \n\nengineering quality and business impact\nso lines of code generated does anyone\n\nso lines of code generated does anyone\n \n\nso lines of code generated does anyone\nhere think lines of code generated is a\n\nhere think lines of code generated is a\n \n\nhere think lines of code generated is a\ngood metric of developer\n\ngood metric of developer\n \n\ngood metric of developer\nproductivity okay I think in 2024 we can\n\nproductivity okay I think in 2024 we can\n \n\nproductivity okay I think in 2024 we can\nall say that it's not we have seen this\n\nall say that it's not we have seen this\n \n\nall say that it's not we have seen this\nresurfaced in the context of measuring\n\nresurfaced in the context of measuring\n \n\nresurfaced in the context of measuring\nRoi of AI because uh it's almost like\n\nRoi of AI because uh it's almost like\n \n\nRoi of AI because uh it's almost like\nwe've forgotten all the lessons that we\n\nwe've forgotten all the lessons that we\n \n\nwe've forgotten all the lessons that we\nlearned about human developer\n\nlearned about human developer\n \n\nlearned about human developer\nproductivity and with AI uh generated\n\nproductivity and with AI uh generated\n \n\nproductivity and with AI uh generated\ntools now it's like oh like you\n\ntools now it's like oh like you\n \n\ntools now it's like oh like you\ngenerated like you know hundreds of\n\ngenerated like you know hundreds of\n \n\ngenerated like you know hundreds of\nlines of code for uh a developer in a\n\nlines of code for uh a developer in a\n \n\nlines of code for uh a developer in a\nday and so we we noted that we were\n\nday and so we we noted that we were\n \n\nday and so we we noted that we were\nactually losing some deals on lines of\n\nactually losing some deals on lines of\n \n\nactually losing some deals on lines of\ncode generated and when we actually went\n\ncode generated and when we actually went\n \n\ncode generated and when we actually went\nand looked at the product experience\n\nand looked at the product experience\n \n\nand looked at the product experience\nwe're like At first we're like hey you\n\nwe're like At first we're like hey you\n \n\nwe're like At first we're like hey you\nknow maybe we should there's a product\n\nknow maybe we should there's a product\n \n\nknow maybe we should there's a product\nImprovement here that we should be\n\nImprovement here that we should be\n \n\nImprovement here that we should be\nmaking because people aren't accepting\n\nmaking because people aren't accepting\n \n\nmaking because people aren't accepting\nas many lines generated by Cody but when\n\nas many lines generated by Cody but when\n \n\nas many lines generated by Cody but when\nwe dug into this we're like oh like you\n\nwe dug into this we're like oh like you\n \n\nwe dug into this we're like oh like you\nknow the comp products it's just more\n\nknow the comp products it's just more\n \n\nknow the comp products it's just more\naggressively kind of like triggering and\n\naggressively kind of like triggering and\n \n\naggressively kind of like triggering and\nand that's not like the sort of business\n\nand that's not like the sort of business\n \n\nand that's not like the sort of business\nthat we want to be so more and more\n\nthat we want to be so more and more\n \n\nthat we want to be so more and more\nwe're kind of pushing our customers to\n\nwe're kind of pushing our customers to\n \n\nwe're kind of pushing our customers to\nnot tie to generic metrics or like high\n\nnot tie to generic metrics or like high\n \n\nnot tie to generic metrics or like high\nLev metrics but identify certain kpis\n\nLev metrics but identify certain kpis\n \n\nLev metrics but identify certain kpis\nthat attend to changes that you want to\n\nthat attend to changes that you want to\n \n\nthat attend to changes that you want to\nmake in your organization so with lios\n\nmake in your organization so with lios\n \n\nmake in your organization so with lios\num our our big kind of Government\n\num our our big kind of Government\n \n\num our our big kind of Government\nContracting uh customer they identified\n\nContracting uh customer they identified\n \n\nContracting uh customer they identified\na set of uh zones or actions that they\n\na set of uh zones or actions that they\n \n\na set of uh zones or actions that they\nfelt were really important uh they\n\nfelt were really important uh they\n \n\nfelt were really important uh they\nwanted to reduce the amount of time\n\nwanted to reduce the amount of time\n \n\nwanted to reduce the amount of time\nspent answering questions spent bugging\n\nspent answering questions spent bugging\n \n\nspent answering questions spent bugging\nteammates and uh spend more developer\n\nteammates and uh spend more developer\n \n\nteammates and uh spend more developer\ntime in these areas that they identified\n\ntime in these areas that they identified\n \n\ntime in these areas that they identified\nas value ad uh three things mainly\n\nas value ad uh three things mainly\n \n\nas value ad uh three things mainly\nbuilding features writing unit tests and\n\nbuilding features writing unit tests and\n \n\nbuilding features writing unit tests and\nreviewing code and so that's what we\n\nreviewing code and so that's what we\n \n\nreviewing code and so that's what we\ntrack for for their evaluation\n\n \n\nperiod a fifth framework is impact on\n\nperiod a fifth framework is impact on\n \n\nperiod a fifth framework is impact on\nkey initiatives so this is the kind of\n\nkey initiatives so this is the kind of\n \n\nkey initiatives so this is the kind of\nlike map your product to okr's uh\n\nlike map your product to okr's uh\n \n\nlike map your product to okr's uh\nframework uh and so there are a couple\n\nframework uh and so there are a couple\n \n\nframework uh and so there are a couple\ncompanies where they're in the midst of\n\ncompanies where they're in the midst of\n \n\ncompanies where they're in the midst of\na big code migration like they're trying\n\na big code migration like they're trying\n \n\na big code migration like they're trying\nto migrate from you know Cobalt to Java\n\nto migrate from you know Cobalt to Java\n \n\nto migrate from you know Cobalt to Java\nor uh maybe you know from uh react to\n\nor uh maybe you know from uh react to\n \n\nor uh maybe you know from uh react to\nspelt or or the latest JS framework and\n\nspelt or or the latest JS framework and\n \n\nspelt or or the latest JS framework and\nthese are kind of like top level goals\n\nthese are kind of like top level goals\n \n\nthese are kind of like top level goals\nthat the VP of engineering um really\n\nthat the VP of engineering um really\n \n\nthat the VP of engineering um really\ncares about and so if you have a product\n\ncares about and so if you have a product\n \n\ncares about and so if you have a product\nthat accelerates progress uh towards\n\nthat accelerates progress uh towards\n \n\nthat accelerates progress uh towards\nthis then the ROI is really just what's\n\nthis then the ROI is really just what's\n \n\nthis then the ROI is really just what's\nthe value of bringing that forward by\n\nthe value of bringing that forward by\n \n\nthe value of bringing that forward by\nyou know X number of months or in some\n\nyou know X number of months or in some\n \n\nyou know X number of months or in some\ncases X number of years or making it\n\ncases X number of years or making it\n \n\ncases X number of years or making it\npossible at all\n\npossible at all\n \n\npossible at all\nhow do you measure how much you pulled\n\nhow do you measure how much you pulled\n \n\nhow do you measure how much you pulled\nit\n\nit\n \n\nit\nForward uh that's a good question so the\n\nForward uh that's a good question so the\n \n\nForward uh that's a good question so the\nquestion was how do you measure how much\n\nquestion was how do you measure how much\n \n\nquestion was how do you measure how much\nyou pulled it forward it is really a\n\nyou pulled it forward it is really a\n \n\nyou pulled it forward it is really a\nkind of judgment call with the\n\nkind of judgment call with the\n \n\nkind of judgment call with the\nengineering leader at that point um by\n\nengineering leader at that point um by\n \n\nengineering leader at that point um by\nthe time they have this conversation\n\nthe time they have this conversation\n \n\nthe time they have this conversation\nwith us they've typically already\n\nwith us they've typically already\n \n\nwith us they've typically already\nstarted it or have had a few of these\n\nstarted it or have had a few of these\n \n\nstarted it or have had a few of these\nunder their belt and they had have an\n\nunder their belt and they had have an\n \n\nunder their belt and they had have an\nidea of the pain uh and then they can\n\nidea of the pain uh and then they can\n \n\nidea of the pain uh and then they can\nassess kind of the shape of the product\n\nassess kind of the shape of the product\n \n\nassess kind of the shape of the product\nand and the things that we do and\n\nand and the things that we do and\n \n\nand and the things that we do and\nestimate how much quicker it'll done we\n\nestimate how much quicker it'll done we\n \n\nestimate how much quicker it'll done we\nalso have case studies demonstrating\n\nalso have case studies demonstrating\n \n\nalso have case studies demonstrating\nlike hey this thing that used to take\n\nlike hey this thing that used to take\n \n\nlike hey this thing that used to take\nyou know a year or longer we squished it\n\nyou know a year or longer we squished it\n \n\nyou know a year or longer we squished it\ninto the span of you know a couple\n\ninto the span of you know a couple\n \n\ninto the span of you know a couple\nmonths okay and then the last framework\n\nmonths okay and then the last framework\n \n\nmonths okay and then the last framework\nis what I'll call survey so this sounds\n\nis what I'll call survey so this sounds\n \n\nis what I'll call survey so this sounds\nlike the least rigorous framework but I\n\nlike the least rigorous framework but I\n \n\nlike the least rigorous framework but I\nthink it's still highly valuable\n\nthink it's still highly valuable\n \n\nthink it's still highly valuable\nbasically it's just run a pilot uh have\n\nbasically it's just run a pilot uh have\n \n\nbasically it's just run a pilot uh have\nyour developers use it you can compare\n\nyour developers use it you can compare\n \n\nyour developers use it you can compare\nagainst another tool and then at the end\n\nagainst another tool and then at the end\n \n\nagainst another tool and then at the end\nof it just ask your developers uh you\n\nof it just ask your developers uh you\n \n\nof it just ask your developers uh you\nknow which one was best um more and more\n\nknow which one was best um more and more\n \n\nknow which one was best um more and more\nnowadays we don't see this in an\n\nnowadays we don't see this in an\n \n\nnowadays we don't see this in an\nunbounded fashion like you know in the\n\nunbounded fashion like you know in the\n \n\nunbounded fashion like you know in the\nkind of like 2021 zer period uh people\n\nkind of like 2021 zer period uh people\n \n\nkind of like 2021 zer period uh people\nwere just like you know whatever makes\n\nwere just like you know whatever makes\n \n\nwere just like you know whatever makes\nthe developers happy let's just go buy\n\nthe developers happy let's just go buy\n \n\nthe developers happy let's just go buy\nthat nowadays we see it in more of a\n\nthat nowadays we see it in more of a\n \n\nthat nowadays we see it in more of a\nbounded fashion which is um you know a\n\nbounded fashion which is um you know a\n \n\nbounded fashion which is um you know a\nlot of orgs uh you know allocate some\n\nlot of orgs uh you know allocate some\n \n\nlot of orgs uh you know allocate some\npart of their budget toward investing in\n\npart of their budget toward investing in\n \n\npart of their budget toward investing in\ndeveloper productivity and to develop\n\ndeveloper productivity and to develop\n \n\ndeveloper productivity and to develop\nproductivity tools so that chart over\n\nproductivity tools so that chart over\n \n\nproductivity tools so that chart over\nthere shows kind of the range of org\n\nthere shows kind of the range of org\n \n\nthere shows kind of the range of org\nsurveyed in uh a survey run by the\n\nsurveyed in uh a survey run by the\n \n\nsurveyed in uh a survey run by the\npragmatic engineer engineer newsletter\n\npragmatic engineer engineer newsletter\n \n\npragmatic engineer engineer newsletter\nand it typically ranges somewhere\n\nand it typically ranges somewhere\n \n\nand it typically ranges somewhere\nbetween 5 and\n\nbetween 5 and\n \n\nbetween 5 and\n25% So within that budget allocation you\n\n25% So within that budget allocation you\n \n\n25% So within that budget allocation you\nhave a certain amount of budget to\n\nhave a certain amount of budget to\n \n\nhave a certain amount of budget to\nallocate to tools um then you basically\n\nallocate to tools um then you basically\n \n\nallocate to tools um then you basically\nsay uh subject to that constraint let me\n\nsay uh subject to that constraint let me\n \n\nsay uh subject to that constraint let me\ngo ask my developers what tools they\n\ngo ask my developers what tools they\n \n\ngo ask my developers what tools they\nwant the\n\n \n\nmost okay so I'm basically out of time\n\nmost okay so I'm basically out of time\n \n\nmost okay so I'm basically out of time\nbut hopefully this gave you kind of like\n\nbut hopefully this gave you kind of like\n \n\nbut hopefully this gave you kind of like\na sampling of of the flavors of\n\na sampling of of the flavors of\n \n\na sampling of of the flavors of\ndifferent Frameworks that are involved\n\ndifferent Frameworks that are involved\n \n\ndifferent Frameworks that are involved\nthis is something that we've had to work\n\nthis is something that we've had to work\n \n\nthis is something that we've had to work\nthrough through a lot of customers\n\nthrough through a lot of customers\n \n\nthrough through a lot of customers\nranging from very small startups to very\n\nranging from very small startups to very\n \n\nranging from very small startups to very\nlarge uh companies in the Fortune 500 I\n\nlarge uh companies in the Fortune 500 I\n \n\nlarge uh companies in the Fortune 500 I\njust want to say you know be skeptical\n\njust want to say you know be skeptical\n \n\njust want to say you know be skeptical\nof anyone saying P equals NP of saying\n\nof anyone saying P equals NP of saying\n \n\nof anyone saying P equals NP of saying\nthey they have u a precise way to\n\nthey they have u a precise way to\n \n\nthey they have u a precise way to\nmeasure aii or developer productivity no\n\nmeasure aii or developer productivity no\n \n\nmeasure aii or developer productivity no\nframework is perfect the most important\n\nframework is perfect the most important\n \n\nframework is perfect the most important\nthing I think is Define clear success\n\nthing I think is Define clear success\n \n\nthing I think is Define clear success\ncriteria um that's something that both\n\ncriteria um that's something that both\n \n\ncriteria um that's something that both\nyou your internal teams and stakeholders\n\nyou your internal teams and stakeholders\n \n\nyou your internal teams and stakeholders\nwill appreciate and also the vendor\n\nwill appreciate and also the vendor\n \n\nwill appreciate and also the vendor\nbecause they know what success looks\n\nbecause they know what success looks\n \n\nbecause they know what success looks\nlike and then the last kind of final\n\nlike and then the last kind of final\n \n\nlike and then the last kind of final\nnote here is productivity tools are\n\nnote here is productivity tools are\n \n\nnote here is productivity tools are\noften bottoms up but we've actually\n\noften bottoms up but we've actually\n \n\noften bottoms up but we've actually\nfound that top- down mandates with AI\n\nfound that top- down mandates with AI\n \n\nfound that top- down mandates with AI\ncan sometimes help because developers\n\ncan sometimes help because developers\n \n\ncan sometimes help because developers\ncan be a little bit of a skeptical crowd\n\ncan be a little bit of a skeptical crowd\n \n\ncan be a little bit of a skeptical crowd\nbut if you believe firmly that this is\n\nbut if you believe firmly that this is\n \n\nbut if you believe firmly that this is\nwhere the future is going and that\n\nwhere the future is going and that\n \n\nwhere the future is going and that\npeople have to update their skill set to\n\npeople have to update their skill set to\n \n\npeople have to update their skill set to\nmake productive use of LMS and AI uh to\n\nmake productive use of LMS and AI uh to\n \n\nmake productive use of LMS and AI uh to\ncode more productively we've actually\n\ncode more productively we've actually\n \n\ncode more productively we've actually\nseen success in this case where a CEO\n\nseen success in this case where a CEO\n \n\nseen success in this case where a CEO\nbasically says we're adopting Cody we're\n\nbasically says we're adopting Cody we're\n \n\nbasically says we're adopting Cody we're\nadopting Cod ai go figure out how to use\n\nadopting Cod ai go figure out how to use\n \n\nadopting Cod ai go figure out how to use\nthis it's not going to be perfect but\n\nthis it's not going to be perfect but\n \n\nthis it's not going to be perfect but\nit's something that is going to play out\n\nit's something that is going to play out\n \n\nit's something that is going to play out\nover the next\n\nover the next\n \n\nover the next\ndecade and then sorry one final thought\n\ndecade and then sorry one final thought\n \n\ndecade and then sorry one final thought\num as we move towards more automation um\n\num as we move towards more automation um\n \n\num as we move towards more automation um\nI like to think of two pictures in mind\n\nI like to think of two pictures in mind\n \n\nI like to think of two pictures in mind\nso one picture is the graph on the right\n\nso one picture is the graph on the right\n \n\nso one picture is the graph on the right\nwhich is kind of like a landscape of uh\n\nwhich is kind of like a landscape of uh\n \n\nwhich is kind of like a landscape of uh\ncodei tools so on the left hand side you\n\ncodei tools so on the left hand side you\n \n\ncodei tools so on the left hand side you\nhave the kind of like inline completions\n\nhave the kind of like inline completions\n \n\nhave the kind of like inline completions\nyou know very basic completing the next\n\nyou know very basic completing the next\n \n\nyou know very basic completing the next\ntokens that you're typing and then on\n\ntokens that you're typing and then on\n \n\ntokens that you're typing and then on\nthe far right you have kind of like the\n\nthe far right you have kind of like the\n \n\nthe far right you have kind of like the\nfully automated offline agents and we're\n\nfully automated offline agents and we're\n \n\nfully automated offline agents and we're\ntrying to make all these Solutions more\n\ntrying to make all these Solutions more\n \n\ntrying to make all these Solutions more\nreliable right CU like uh generative AI\n\nreliable right CU like uh generative AI\n \n\nreliable right CU like uh generative AI\nis sort of inherently unreliable we're\n\nis sort of inherently unreliable we're\n \n\nis sort of inherently unreliable we're\ntrying to make it more General and and\n\ntrying to make it more General and and\n \n\ntrying to make it more General and and\nuh you know uh make it productive in in\n\nuh you know uh make it productive in in\n \n\nuh you know uh make it productive in in\nmore languages in more scenarios and we\n\nmore languages in more scenarios and we\n \n\nmore languages in more scenarios and we\nactually think that the the the path\n\nactually think that the the the path\n \n\nactually think that the the the path\nthere is to go from the left-and side to\n\nthere is to go from the left-and side to\n \n\nthere is to go from the left-and side to\nthe right hand side you know not jumping\n\nthe right hand side you know not jumping\n \n\nthe right hand side you know not jumping\nstraight to the full automation because\n\nstraight to the full automation because\n \n\nstraight to the full automation because\nthat's a very difficult problem so we\n\nthat's a very difficult problem so we\n \n\nthat's a very difficult problem so we\nthink the next phase of evolution for us\n\nthink the next phase of evolution for us\n \n\nthink the next phase of evolution for us\nis going from kind of like these inline\n\nis going from kind of like these inline\n \n\nis going from kind of like these inline\ncode completion uh scenarios to more\n\ncode completion uh scenarios to more\n \n\ncode completion uh scenarios to more\nwhat we call online agents that live in\n\nwhat we call online agents that live in\n \n\nwhat we call online agents that live in\nyour editor but can still react to human\n\nyour editor but can still react to human\n \n\nyour editor but can still react to human\nfeedback and uh guidance and then the\n\nfeedback and uh guidance and then the\n \n\nfeedback and uh guidance and then the\nsecond picture is you know the mythical\n\nsecond picture is you know the mythical\n \n\nsecond picture is you know the mythical\nman month I think a lot of people are\n\nman month I think a lot of people are\n \n\nman month I think a lot of people are\nfamiliar with this classic work talks\n\nfamiliar with this classic work talks\n \n\nfamiliar with this classic work talks\nabout the classic fallacies with respect\n\nabout the classic fallacies with respect\n \n\nabout the classic fallacies with respect\nto uh developer productivity that a lot\n\nto uh developer productivity that a lot\n \n\nto uh developer productivity that a lot\nof companies make um one question I\n\nof companies make um one question I\n \n\nof companies make um one question I\nwould pose to all of you is have the\n\nwould pose to all of you is have the\n \n\nwould pose to all of you is have the\nfundamentals really changed you know as\n\nfundamentals really changed you know as\n \n\nfundamentals really changed you know as\nwe have more quote unquote AI\n\nwe have more quote unquote AI\n \n\nwe have more quote unquote AI\ndevelopers uh should we measure them by\n\ndevelopers uh should we measure them by\n \n\ndevelopers uh should we measure them by\nthe same evaluation criteria and I would\n\nthe same evaluation criteria and I would\n \n\nthe same evaluation criteria and I would\nthe challenge question I would pose to\n\nthe challenge question I would pose to\n \n\nthe challenge question I would pose to\nall of you is you know I think the\n\nall of you is you know I think the\n \n\nall of you is you know I think the\nlessons from the mythical man month is\n\nlessons from the mythical man month is\n \n\nlessons from the mythical man month is\nyou prefer to have one very smart\n\nyou prefer to have one very smart\n \n\nyou prefer to have one very smart\nengineer who's highly productive over 10\n\nengineer who's highly productive over 10\n \n\nengineer who's highly productive over 10\nmaybe even a hundred mediocre developers\n\nmaybe even a hundred mediocre developers\n \n\nmaybe even a hundred mediocre developers\nwho are kind of productive but need a\n\nwho are kind of productive but need a\n \n\nwho are kind of productive but need a\nlot of guidance and so as AI automation\n\nlot of guidance and so as AI automation\n \n\nlot of guidance and so as AI automation\nincreases the question is do you want\n\nincreases the question is do you want\n \n\nincreases the question is do you want\n100 mediocre AI developers or do you\n\n100 mediocre AI developers or do you\n \n\n100 mediocre AI developers or do you\nwant a 100x lever for uh the human\n\nwant a 100x lever for uh the human\n \n\nwant a 100x lever for uh the human\ndevelopers the really smart people who\n\ndevelopers the really smart people who\n \n\ndevelopers the really smart people who\nare going to craft the user\n\nare going to craft the user\n \n\nare going to craft the user\nexperience all right that's it for me um\n\nexperience all right that's it for me um\n \n\nexperience all right that's it for me um\nyeah if you want to check out Cody\n\nyeah if you want to check out Cody\n \n\nyeah if you want to check out Cody\nthat's the URL and uh that's my contact\n\nthat's the URL and uh that's my contact\n \n\nthat's the URL and uh that's my contact\ninfo if you want to reach out later do\n\ninfo if you want to reach out later do\n \n\ninfo if you want to reach out later do\nwe have time for questions or um we have\n\nwe have time for questions or um we have\n \n\nwe have time for questions or um we have\ntime for perhaps one or two questions or\n\ntime for perhaps one or two questions or\n \n\ntime for perhaps one or two questions or\ntwo I will run around with the\n\ntwo I will run around with the\n \n\ntwo I will run around with the\nmicrophone if anyone just so we can\n\nmicrophone if anyone just so we can\n \n\nmicrophone if anyone just so we can\ncatch it on audio\n\n \n\ncool thanks thanks for Cody you you are\n\ncool thanks thanks for Cody you you are\n \n\ncool thanks thanks for Cody you you are\nthe best guys I tried them all I'm\n\nthe best guys I tried them all I'm\n \n\nthe best guys I tried them all I'm\nstaying with\n\nstaying with\n \n\nstaying with\nyou you've got really good Insight\n\nyou you've got really good Insight\n \n\nyou you've got really good Insight\nwhat's happening now with software\n\nwhat's happening now with software\n \n\nwhat's happening now with software\ndevelopment so what is your intuition um\n\ndevelopment so what is your intuition um\n \n\ndevelopment so what is your intuition um\nare we going to get away from coding and\n\nare we going to get away from coding and\n \n\nare we going to get away from coding and\nthe code would become a boiler plate and\n\nthe code would become a boiler plate and\n \n\nthe code would become a boiler plate and\nwe move to meta programming or it will\n\nwe move to meta programming or it will\n \n\nwe move to meta programming or it will\nbe still code uh as a main uh the output\n\nbe still code uh as a main uh the output\n \n\nbe still code uh as a main uh the output\nof the senior\n\nof the senior\n \n\nof the senior\nengineer yeah it's a really good\n\nengineer yeah it's a really good\n \n\nengineer yeah it's a really good\nquestion the usage patterns that we're\n\nquestion the usage patterns that we're\n \n\nquestion the usage patterns that we're\nobserving now is more and more code is\n\nobserving now is more and more code is\n \n\nobserving now is more and more code is\nbeing written through natural language\n\nbeing written through natural language\n \n\nbeing written through natural language\nprompts like I had an experience just on\n\nprompts like I had an experience just on\n \n\nprompts like I had an experience just on\nFriday actually where like Cody wrote 80\n\nFriday actually where like Cody wrote 80\n \n\nFriday actually where like Cody wrote 80\n80% of the code because I just kept\n\n80% of the code because I just kept\n \n\n80% of the code because I just kept\nasking it to write different functions\n\nasking it to write different functions\n \n\nasking it to write different functions\nthat I wanted uh and that was nice\n\nthat I wanted uh and that was nice\n \n\nthat I wanted uh and that was nice\nbecause I was thinking at like the\n\nbecause I was thinking at like the\n \n\nbecause I was thinking at like the\nfunction level rather than the kind of\n\nfunction level rather than the kind of\n \n\nfunction level rather than the kind of\nlike line by line code level yeah and it\n\nlike line by line code level yeah and it\n \n\nlike line by line code level yeah and it\nwas nice cuz it allow to stay in flow\n\nwas nice cuz it allow to stay in flow\n \n\nwas nice cuz it allow to stay in flow\nbut at the same time like the output was\n\nbut at the same time like the output was\n \n\nbut at the same time like the output was\nstill code and I I really do think that\n\nstill code and I I really do think that\n \n\nstill code and I I really do think that\nlike we're never going to see a full\n\nlike we're never going to see a full\n \n\nlike we're never going to see a full\nreplacement of Code by natural language\n\nreplacement of Code by natural language\n \n\nreplacement of Code by natural language\nbecause it code is nice because you can\n\nbecause it code is nice because you can\n \n\nbecause it code is nice because you can\ndescribe what you want very precisely\n\ndescribe what you want very precisely\n \n\ndescribe what you want very precisely\nand that Precision is important as a\n\nand that Precision is important as a\n \n\nand that Precision is important as a\nsource of Truth For What the software\n\nsource of Truth For What the software\n \n\nsource of Truth For What the software\nactually\n\nactually\n \n\nactually\ndoes okay I think that's probably it now\n\ndoes okay I think that's probably it now\n \n\ndoes okay I think that's probably it now\nthank you very much thank you thank you\n\nthank you very much thank you thank you\n \n\nthank you very much thank you thank you\n[Music]" } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nHi everyone, I'm Jun, founding engineer\n\nHi everyone, I'm Jun, founding engineer\n \n\nHi everyone, I'm Jun, founding engineer\nat Tusk. I'm excited to share about a\n\nat Tusk. I'm excited to share about a\n \n\nat Tusk. I'm excited to share about a\nparadigm for building the next\n\nparadigm for building the ..." }, { "role": "assistant", "content": "Hi everyone, I'm Jun, founding engineer\n\nHi everyone, I'm Jun, founding engineer\n \n\nHi everyone, I'm Jun, founding engineer\nat Tusk. I'm excited to share about a\n\nat Tusk. I'm excited to share about a\n \n\nat Tusk. I'm excited to share about a\nparadigm for building the next\n\nparadigm for building the next\n \n\nparadigm for building the next\ngeneration of AI interfaces, the ones\n\ngeneration of AI interfaces, the ones\n \n\ngeneration of AI interfaces, the ones\nthat put humans in the center,\n\nthat put humans in the center,\n \n\nthat put humans in the center,\naugmenting our capabilities, and helping\n\naugmenting our capabilities, and helping\n \n\naugmenting our capabilities, and helping\nus be more thoughtful and\n\nus be more thoughtful and\n \n\nus be more thoughtful and\ncreative. It's a collection of ideas\n\ncreative. It's a collection of ideas\n \n\ncreative. It's a collection of ideas\nthat I've been thinking about for a\n\nthat I've been thinking about for a\n \n\nthat I've been thinking about for a\nwhile now. Though some of these concepts\n\nwhile now. Though some of these concepts\n \n\nwhile now. Though some of these concepts\nmight be a little bit more speculative,\n\nmight be a little bit more speculative,\n \n\nmight be a little bit more speculative,\nI do hope to encourage more builders in\n\nI do hope to encourage more builders in\n \n\nI do hope to encourage more builders in\nthe space to consider these patterns and\n\nthe space to consider these patterns and\n \n\nthe space to consider these patterns and\nprinciples. Now, it's often said that\n\nprinciples. Now, it's often said that\n \n\nprinciples. Now, it's often said that\n2025 is the year of the agents and\n\n2025 is the year of the agents and\n \n\n2025 is the year of the agents and\nrightfully so. From general agents\n\nrightfully so. From general agents\n \n\nrightfully so. From general agents\nperforming research and browser use for\n\nperforming research and browser use for\n \n\nperforming research and browser use for\nyou to deeply verticalized apps across\n\nyou to deeply verticalized apps across\n \n\nyou to deeply verticalized apps across\nall industries, AI agents are everywhere\n\nall industries, AI agents are everywhere\n \n\nall industries, AI agents are everywhere\nready to offload your tasks from you. ready to offload your tasks from you. ready to offload your tasks from you. And as agent based tooling and protocols\n\nAnd as agent based tooling and protocols\n \n\nAnd as agent based tooling and protocols\nget more and more mature and\n\nget more and more mature and\n \n\nget more and more mature and\nsophisticated is no doubt that large\n\nsophisticated is no doubt that large\n \n\nsophisticated is no doubt that large\nchunks of knowledge work will automated\n\nchunks of knowledge work will automated\n \n\nchunks of knowledge work will automated\naway in time to come. A lot of agents\n\naway in time to come. A lot of agents\n \n\naway in time to come. A lot of agents\nare focused on automating discrete tasks\n\nare focused on automating discrete tasks\n \n\nare focused on automating discrete tasks\nand for good reason. Well, it's easier\n\nand for good reason. Well, it's easier\n \n\nand for good reason. Well, it's easier\nto quantify the unit of work performed\n\nto quantify the unit of work performed\n \n\nto quantify the unit of work performed\nand sell the promise of eradicating\n\nand sell the promise of eradicating\n \n\nand sell the promise of eradicating\ntedious and boring work that people\n\ntedious and boring work that people\n \n\ntedious and boring work that people\ndon't like doing. It's also easier to\n\ndon't like doing. It's also easier to\n \n\ndon't like doing. It's also easier to\nbenchmark the completion of these tasks\n\nbenchmark the completion of these tasks\n \n\nbenchmark the completion of these tasks\nquick and quickly compare one system\n\nquick and quickly compare one system\n \n\nquick and quickly compare one system\nagainst\n\nagainst\n \n\nagainst\nanother. But it's also not difficult to\n\nanother. But it's also not difficult to\n \n\nanother. But it's also not difficult to\nspot the potential drawbacks of such a\n\nspot the potential drawbacks of such a\n \n\nspot the potential drawbacks of such a\nfuture if it's not carefully managed. future if it's not carefully managed. future if it's not carefully managed. Firstly, there's a notion that you know\n\nFirstly, there's a notion that you know\n \n\nFirstly, there's a notion that you know\nover reliance on automating work away\n\nover reliance on automating work away\n \n\nover reliance on automating work away\nbreeds general laziness and atrophy of\n\nbreeds general laziness and atrophy of\n \n\nbreeds general laziness and atrophy of\nskills. I'm sure most of you are\n\nskills. I'm sure most of you are\n \n\nskills. I'm sure most of you are\ndiscerning users but when you v code an\n\ndiscerning users but when you v code an\n \n\ndiscerning users but when you v code an\napp without thoughtful review or at\n\napp without thoughtful review or at\n \n\napp without thoughtful review or at\nleast bother to sufficiently understand\n\nleast bother to sufficiently understand\n \n\nleast bother to sufficiently understand\nits inner workings how are you\n\nits inner workings how are you\n \n\nits inner workings how are you\nmaintaining this as a skills are you\n\nmaintaining this as a skills are you\n \n\nmaintaining this as a skills are you\nsure you know what you're\n\nsure you know what you're\n \n\nsure you know what you're\nattributing many high judgment domains\n\nattributing many high judgment domains\n \n\nattributing many high judgment domains\nlike coding and design still require\n\nlike coding and design still require\n \n\nlike coding and design still require\ntight human supervision relying on human\n\ntight human supervision relying on human\n \n\ntight human supervision relying on human\nexperience and taste for approval a lot\n\nexperience and taste for approval a lot\n \n\nexperience and taste for approval a lot\nof these are also like high context\n\nof these are also like high context\n \n\nof these are also like high context\ntasks where it's often difficult to just\n\ntasks where it's often difficult to just\n \n\ntasks where it's often difficult to just\nupload perfect context and encapsulate\n\nupload perfect context and encapsulate\n \n\nupload perfect context and encapsulate\nthem in prompts for accurate judgement\n\nthem in prompts for accurate judgement\n \n\nthem in prompts for accurate judgement\nsince a lot of them a lot of these like\n\nsince a lot of them a lot of these like\n \n\nsince a lot of them a lot of these like\nlive in the minds of experienced\n\nlive in the minds of experienced\n \n\nlive in the minds of experienced\nprofessionals\n\nprofessionals\n \n\nprofessionals\nanyway. So the main thesis of this talk\n\nanyway. So the main thesis of this talk\n \n\nanyway. So the main thesis of this talk\nis this instead of getting humans to\n\nis this instead of getting humans to\n \n\nis this instead of getting humans to\naccommodate to AI systems uh attempting\n\naccommodate to AI systems uh attempting\n \n\naccommodate to AI systems uh attempting\nto automate complex tasks and doing so\n\nto automate complex tasks and doing so\n \n\nto automate complex tasks and doing so\nsuboptimally why not spend that comput\n\nsuboptimally why not spend that comput\n \n\nsuboptimally why not spend that comput\njust helping humans produce high quality\n\njust helping humans produce high quality\n \n\njust helping humans produce high quality\nwork in the first place. So in this talk\n\nwork in the first place. So in this talk\n \n\nwork in the first place. So in this talk\nI'd like to introduce some ideas for\n\nI'd like to introduce some ideas for\n \n\nI'd like to introduce some ideas for\naugmentation based UX. We look at\n\naugmentation based UX. We look at\n \n\naugmentation based UX. We look at\ninteraction patterns for how AI can help\n\ninteraction patterns for how AI can help\n \n\ninteraction patterns for how AI can help\nusers review blind spots, spark\n\nusers review blind spots, spark\n \n\nusers review blind spots, spark\ncreativity, and amplify thoughtful\n\ncreativity, and amplify thoughtful\n \n\ncreativity, and amplify thoughtful\ndecision-m some principles for designing\n\ndecision-m some principles for designing\n \n\ndecision-m some principles for designing\nAI products to emphasize and grow human\n\nAI products to emphasize and grow human\n \n\nAI products to emphasize and grow human\ncapabilities. Um, and trustworthy human\n\ncapabilities. Um, and trustworthy human\n \n\ncapabilities. Um, and trustworthy human\nAI partnerships. Let me make this\n\nAI partnerships. Let me make this\n \n\nAI partnerships. Let me make this\nconcrete with a simple comparison. Um,\n\nconcrete with a simple comparison. Um,\n \n\nconcrete with a simple comparison. Um,\nthe automation approach to AI emails\n\nthe automation approach to AI emails\n \n\nthe automation approach to AI emails\nmight look like this, right? The AI\n\nmight look like this, right? The AI\n \n\nmight look like this, right? The AI\nwrites the entire email for you and and\n\nwrites the entire email for you and and\n \n\nwrites the entire email for you and and\nsends it on your behalf. Um, you're out\n\nsends it on your behalf. Um, you're out\n \n\nsends it on your behalf. Um, you're out\nof the loop." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nBasically the the\n\nof the loop. Basically the the\n \n\nof the loop. Basically the the\naugmentation approach is a little bit\n\naugmentation approach is a little bit\n \n\naugmentation approach is a little bit\ndifferent. Um AI could help you u\n\ndifferent. Um AI could help you u\n \n\ndifferent. Um AI could hel..." }, { "role": "assistant", "content": "Basically the the\n\nof the loop. Basically the the\n \n\nof the loop. Basically the the\naugmentation approach is a little bit\n\naugmentation approach is a little bit\n \n\naugmentation approach is a little bit\ndifferent. Um AI could help you u\n\ndifferent. Um AI could help you u\n \n\ndifferent. Um AI could help you u\nbrainstorm key points. For example,\n\nbrainstorm key points. For example,\n \n\nbrainstorm key points. For example,\nsuggest alternative phrasings. Um check\n\nsuggest alternative phrasings. Um check\n \n\nsuggest alternative phrasings. Um check\nyour tone. Maybe it says, \"Hey, this\n\nyour tone. Maybe it says, \"Hey, this\n \n\nyour tone. Maybe it says, \"Hey, this\nmight come across as a bit harsh. Do you\n\nmight come across as a bit harsh. Do you\n \n\nmight come across as a bit harsh. Do you\nwant to soften it based on your past\n\nwant to soften it based on your past\n \n\nwant to soften it based on your past\nemails with this person or usual writing\n\nemails with this person or usual writing\n \n\nemails with this person or usual writing\nstyle?\" And then you review and hit\n\nstyle?\" And then you review and hit\n \n\nstyle?\" And then you review and hit\nsend. Same with coding, right? With\n\nsend. Same with coding, right? With\n \n\nsend. Same with coding, right? With\nautomation, AI writes the code, creates\n\nautomation, AI writes the code, creates\n \n\nautomation, AI writes the code, creates\na pull request based on a slack message\n\na pull request based on a slack message\n \n\na pull request based on a slack message\nthat you have assigned to it. that you have assigned to it. that you have assigned to it. Augmentation could be something like AI\n\nAugmentation could be something like AI\n \n\nAugmentation could be something like AI\nreviewing your pull request and says you\n\nreviewing your pull request and says you\n \n\nreviewing your pull request and says you\nmight have missed this edge case. This\n\nmight have missed this edge case. This\n \n\nmight have missed this edge case. This\npattern you're using here could cause\n\npattern you're using here could cause\n \n\npattern you're using here could cause\nissues on a highload. Now the difference\n\nissues on a highload. Now the difference\n \n\nissues on a highload. Now the difference\nis that in augmentation you are still in\n\nis that in augmentation you are still in\n \n\nis that in augmentation you are still in\ndriver seat but you have a thinking\n\ndriver seat but you have a thinking\n \n\ndriver seat but you have a thinking\npartner or co-pilot who can see things\n\npartner or co-pilot who can see things\n \n\npartner or co-pilot who can see things\nthat you might\n\nthat you might\n \n\nthat you might\nmiss and this cap this table captures\n\nmiss and this cap this table captures\n \n\nmiss and this cap this table captures\nthe the fundamental shift in this\n\nthe the fundamental shift in this\n \n\nthe the fundamental shift in this\nmindset. In a nutshell, many differences\n\nmindset. In a nutshell, many differences\n \n\nmindset. In a nutshell, many differences\nuh arise based on a shift of\n\nuh arise based on a shift of\n \n\nuh arise based on a shift of\nresponsibility of the task itself and\n\nresponsibility of the task itself and\n \n\nresponsibility of the task itself and\nthe relationship and our relationship\n\nthe relationship and our relationship\n \n\nthe relationship and our relationship\nwith the AI system that helps us with\n\nwith the AI system that helps us with\n \n\nwith the AI system that helps us with\nthese tasks. It's almost like working\n\nthese tasks. It's almost like working\n \n\nthese tasks. It's almost like working\nwith an offshore contractor versus a new\n\nwith an offshore contractor versus a new\n \n\nwith an offshore contractor versus a new\nteam member that we grow together with. team member that we grow together with. team member that we grow together with. And if you're interested in knowing more\n\nAnd if you're interested in knowing more\n \n\nAnd if you're interested in knowing more\nabout this in further detail, I write\n\nabout this in further detail, I write\n \n\nabout this in further detail, I write\nabout in my blog post link below. Now,\n\nabout in my blog post link below. Now,\n \n\nabout in my blog post link below. Now,\nlet's get to the meat. The three core\n\nlet's get to the meat. The three core\n \n\nlet's get to the meat. The three core\ninteraction patterns. I want to start\n\ninteraction patterns. I want to start\n \n\ninteraction patterns. I want to start\nwith blind spot detection because it's\n\nwith blind spot detection because it's\n \n\nwith blind spot detection because it's\nthe most immediately compelling. Now,\n\nthe most immediately compelling. Now,\n \n\nthe most immediately compelling. Now,\nthink about it. We all have blind spots,\n\nthink about it. We all have blind spots,\n \n\nthink about it. We all have blind spots,\npatterns that we can't see in our own\n\npatterns that we can't see in our own\n \n\npatterns that we can't see in our own\nthinking. Uh the question for AI isn't\n\nthinking. Uh the question for AI isn't\n \n\nthinking. Uh the question for AI isn't\nwhat do you do wrong, but what didn't\n\nwhat do you do wrong, but what didn't\n \n\nwhat do you do wrong, but what didn't\nyou consider as you're working on this\n\nyou consider as you're working on this\n \n\nyou consider as you're working on this\ntask. This could be temporal blind spots\n\ntask. This could be temporal blind spots\n \n\ntask. This could be temporal blind spots\nlike noticing you always make poor\n\nlike noticing you always make poor\n \n\nlike noticing you always make poor\ndecisions when tired on a Friday\n\ndecisions when tired on a Friday\n \n\ndecisions when tired on a Friday\nafternoons. Um or social blind spots\n\nafternoons. Um or social blind spots\n \n\nafternoons. Um or social blind spots\nflagging flagging when your technical\n\nflagging flagging when your technical\n \n\nflagging flagging when your technical\nfeedback might come across as personal\n\nfeedback might come across as personal\n \n\nfeedback might come across as personal\ncriticism. Um the key challenge is\n\ncriticism. Um the key challenge is\n \n\ncriticism. Um the key challenge is\nmanaging that signal to noise ratio and\n\nmanaging that signal to noise ratio and\n \n\nmanaging that signal to noise ratio and\nnudging someone or challenging their\n\nnudging someone or challenging their\n \n\nnudging someone or challenging their\nthinking without making them feel\n\nthinking without making them feel\n \n\nthinking without making them feel\ndefensive. Um, ideally the AI should be\n\ndefensive. Um, ideally the AI should be\n \n\ndefensive. Um, ideally the AI should be\ndesigned in a way that reveals something\n\ndesigned in a way that reveals something\n \n\ndesigned in a way that reveals something\nat the right time uh when a unit of work\n\nat the right time uh when a unit of work\n \n\nat the right time uh when a unit of work\num can be assumed to be ready. Here's\n\num can be assumed to be ready. Here's\n \n\num can be assumed to be ready. Here's\nhow we implemented this at task. Now\n\nhow we implemented this at task. Now\n \n\nhow we implemented this at task. Now\ntask is an AI testing platform that\n\ntask is an AI testing platform that\n \n\ntask is an AI testing platform that\nfinds edge cases and bugs based on your\n\nfinds edge cases and bugs based on your\n \n\nfinds edge cases and bugs based on your\npull requests. When you commit your code\n\npull requests. When you commit your code\n \n\npull requests. When you commit your code\nchanges instead of just surfacing\n\nchanges instead of just surfacing\n \n\nchanges instead of just surfacing\npotential bugs, we create unit tests and\n\npotential bugs, we create unit tests and\n \n\npotential bugs, we create unit tests and\nactually validate them, execute them\n\nactually validate them, execute them\n \n\nactually validate them, execute them\nagainst your your codebase test\n\nagainst your your codebase test\n \n\nagainst your your codebase test\nenvironment and surface to you verify\n\nenvironment and surface to you verify\n \n\nenvironment and surface to you verify\nissues that we can prove will cause\n\nissues that we can prove will cause\n \n\nissues that we can prove will cause\nproblems in the future. You can then\n\nproblems in the future. You can then\n \n\nproblems in the future. You can then\nchoose whether to commit these tests\n\nchoose whether to commit these tests\n \n\nchoose whether to commit these tests\nalongside other passing ones into your p\n\nalongside other passing ones into your p\n \n\nalongside other passing ones into your p\nrequest so that you know there's a extra\n\nrequest so that you know there's a extra\n \n\nrequest so that you know there's a extra\ncoverage against future regressions. One\n\ncoverage against future regressions. One\n \n\ncoverage against future regressions." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nOne\nthing we try to help make reviewing\n\nthing we try to help make reviewing\n \n\nthing we try to help make reviewing\nthese tests easier is also to surface\n\nthese tests easier is also to surface\n \n\nthese tests easier is also to surface\nany assumptions that task is making be\n\nany assumptions that task ..." }, { "role": "assistant", "content": "One\nthing we try to help make reviewing\n\nthing we try to help make reviewing\n \n\nthing we try to help make reviewing\nthese tests easier is also to surface\n\nthese tests easier is also to surface\n \n\nthese tests easier is also to surface\nany assumptions that task is making be\n\nany assumptions that task is making be\n \n\nany assumptions that task is making be\nit from a business or engineering\n\nit from a business or engineering\n \n\nit from a business or engineering\nstandpoint and also outline potential\n\nstandpoint and also outline potential\n \n\nstandpoint and also outline potential\nfixes. Um, users can provide feedback by\n\nfixes. Um, users can provide feedback by\n \n\nfixes. Um, users can provide feedback by\nclicking thumbs up or thumbs down or\n\nclicking thumbs up or thumbs down or\n \n\nclicking thumbs up or thumbs down or\nexplain why it's something worth\n\nexplain why it's something worth\n \n\nexplain why it's something worth\nconsidering or not. So that task learns\n\nconsidering or not. So that task learns\n \n\nconsidering or not. So that task learns\nfrom your from from your review of its\n\nfrom your from from your review of its\n \n\nfrom your from from your review of its\nassumptions. I also like to call this\n\nassumptions. I also like to call this\n \n\nassumptions. I also like to call this\nsystematic pessimism. We are using AI to\n\nsystematic pessimism. We are using AI to\n \n\nsystematic pessimism. We are using AI to\nsystematically look through um every\n\nsystematically look through um every\n \n\nsystematically look through um every\nsymbol that has been changed finding\n\nsymbol that has been changed finding\n \n\nsymbol that has been changed finding\nwhat could go wrong um consider second\n\nwhat could go wrong um consider second\n \n\nwhat could go wrong um consider second\norder effects in the call stack. Even\n\norder effects in the call stack. Even\n \n\norder effects in the call stack. Even\nthough this is something very tedious\n\nthough this is something very tedious\n \n\nthough this is something very tedious\nfor humans to manually perform\n\nfor humans to manually perform\n \n\nfor humans to manually perform\neffectively, especially in larger, more\n\neffectively, especially in larger, more\n \n\neffectively, especially in larger, more\ncomplex code changes. With this pattern,\n\ncomplex code changes. With this pattern,\n \n\ncomplex code changes. With this pattern,\nwe have helped uh enterprise and growth\n\nwe have helped uh enterprise and growth\n \n\nwe have helped uh enterprise and growth\nstage companies like deep learning AI\n\nstage companies like deep learning AI\n \n\nstage companies like deep learning AI\nand team fee pay catch verified bucks in\n\nand team fee pay catch verified bucks in\n \n\nand team fee pay catch verified bucks in\n43% of p requests and at almost a\n\n43% of p requests and at almost a\n \n\n43% of p requests and at almost a\nthousand new tests in 2 months in the\n\nthousand new tests in 2 months in the\n \n\nthousand new tests in 2 months in the\nblind spot detection pattern not all\n\nblind spot detection pattern not all\n \n\nblind spot detection pattern not all\nsuggestions are created equal. I've been\n\nsuggestions are created equal. I've been\n \n\nsuggestions are created equal. I've been\nthinking about how we can manage noise\n\nthinking about how we can manage noise\n \n\nthinking about how we can manage noise\nwith thoughtful UX. Uh one mental model\n\nwith thoughtful UX. Uh one mental model\n \n\nwith thoughtful UX. Uh one mental model\nis the novelty criticality framework and\n\nis the novelty criticality framework and\n \n\nis the novelty criticality framework and\nlet's again consider the code review\n\nlet's again consider the code review\n \n\nlet's again consider the code review\nscenario. High novelty, high\n\nscenario. High novelty, high\n \n\nscenario. High novelty, high\ncriticality. These are critical\n\ncriticality. These are critical\n \n\ncriticality. These are critical\ndiscoveries. Things like race conditions\n\ndiscoveries. Things like race conditions\n \n\ndiscoveries. Things like race conditions\nthat you have never encountered before. that you have never encountered before. that you have never encountered before. Accidentally exposing customer data. Accidentally exposing customer data. Accidentally exposing customer data. These deserve to interrupt you. Noy, low\n\nThese deserve to interrupt you. Noy, low\n \n\nThese deserve to interrupt you. Noy, low\ncriticality. These are learning moments. criticality. These are learning moments. criticality. These are learning moments. Suggestions like alternative approaches,\n\nSuggestions like alternative approaches,\n \n\nSuggestions like alternative approaches,\nnew language features, recognizing\n\nnew language features, recognizing\n \n\nnew language features, recognizing\nineffective patterns. We might want to\n\nineffective patterns. We might want to\n \n\nineffective patterns. We might want to\npresent this gently as maybe as\n\npresent this gently as maybe as\n \n\npresent this gently as maybe as\nexplanable\n\nexplanable\n \n\nexplanable\nsuggestions. Low novelty, high\n\nsuggestions. Low novelty, high\n \n\nsuggestions. Low novelty, high\ncriticality. These are more of essential\n\ncriticality. These are more of essential\n \n\ncriticality. These are more of essential\nreminders. Uh things like column bugs,\n\nreminders. Uh things like column bugs,\n \n\nreminders. Uh things like column bugs,\nstandard security checks, performance\n\nstandard security checks, performance\n \n\nstandard security checks, performance\nissues might help to give clear warnings\n\nissues might help to give clear warnings\n \n\nissues might help to give clear warnings\nand have a little bit more flexibility\n\nand have a little bit more flexibility\n \n\nand have a little bit more flexibility\nwhen it comes to interruption. But we\n\nwhen it comes to interruption. But we\n \n\nwhen it comes to interruption. But we\ndon't have to overexlain since users\n\ndon't have to overexlain since users\n \n\ndon't have to overexlain since users\nprobably already know the concept. Low\n\nprobably already know the concept. Low\n \n\nprobably already know the concept. Low\nnovelty, low criticality. These are more\n\nnovelty, low criticality. These are more\n \n\nnovelty, low criticality. These are more\nmatters of polish and preference, quote\n\nmatters of polish and preference, quote\n \n\nmatters of polish and preference, quote\nformatting, min optimizations, and\n\nformatting, min optimizations, and\n \n\nformatting, min optimizations, and\nthings of the nature. We want to batch\n\nthings of the nature. We want to batch\n \n\nthings of the nature. We want to batch\nthese, make them optional. The magic is\n\nthese, make them optional. The magic is\n \n\nthese, make them optional. The magic is\nin the prioritization because people\n\nin the prioritization because people\n \n\nin the prioritization because people\ntend to only process three or more\n\ntend to only process three or more\n \n\ntend to only process three or more\nmeaningful suggestions per review\n\nmeaningful suggestions per review\n \n\nmeaningful suggestions per review\nsession. This framework helps you spend\n\nsession. This framework helps you spend\n \n\nsession. This framework helps you spend\nthe attention budget wisely. The second\n\nthe attention budget wisely. The second\n \n\nthe attention budget wisely. The second\npattern is cognitive partnership. Moving\n\npattern is cognitive partnership. Moving\n \n\npattern is cognitive partnership. Moving\nfrom stateless answering machines to\n\nfrom stateless answering machines to\n \n\nfrom stateless answering machines to\nsystems that adapt to your mental\n\nsystems that adapt to your mental\n \n\nsystems that adapt to your mental\nmodels, right? This means building a\n\nmodels, right? This means building a\n \n\nmodels, right? This means building a\ntheory of mind about users. Not just\n\ntheory of mind about users. Not just\n \n\ntheory of mind about users. Not just\nknowing what they want, but how they\n\nknowing what they want, but how they\n \n\nknowing what they want, but how they\nthink. Do they learn through analogies? think. Do they learn through analogies? think. Do they learn through analogies? Do they prefer visual explanations? Do\n\nDo they prefer visual explanations? Do\n \n\nDo they prefer visual explanations? Do\nthey prefer to sleep on it or prefer\n\nthey prefer to sleep on it or prefer\n \n\nthey prefer to sleep on it or prefer\nlike rapid iteration? Um, imagine a code\n\nlike rapid iteration? Um, imagine a code\n \n\nlike rapid iteration? Um, imagine a code\neditor that learns your refactoring\n\neditor that learns your refactoring\n \n\neditor that learns your refactoring\npatterns and proactively suggests\n\npatterns and proactively suggests\n \n\npatterns and proactively suggests\nsimilar improvements in new contexts or\n\nsimilar improvements in new contexts or\n \n\nsimilar improvements in new contexts or\nor a research assistant that understands\n\nor a research assistant that understands\n \n\nor a research assistant that understands\nwhether you synthesize information\n\nwhether you synthesize information\n \n\nwhether you synthesize information\nchronologically or\n\nchronologically or\n \n\nchronologically or\nautomatically. The challenge here is\n\nautomatically. The challenge here is\n \n\nautomatically. The challenge here is\nbuilding this personalization without\n\nbuilding this personalization without\n \n\nbuilding this personalization without\nbeing creepy. Um, users need to feel\n\nbeing creepy. Um, users need to feel\n \n\nbeing creepy. Um, users need to feel\nunderstood but not\n\nunderstood but not\n \n\nunderstood but not\nsurveyed. And the third pattern,\n\nsurveyed. And the third pattern,\n \n\nsurveyed. And the third pattern,\nproactive guidance, is probably the\n\nproactive guidance, is probably the\n \n\nproactive guidance, is probably the\nhardest to get right. Um, the challenge\n\nhardest to get right. Um, the challenge\n \n\nhardest to get right. Um, the challenge\nisn't knowing what to suggest but when." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nisn't knowing what to suggest but when. isn't knowing what to suggest but when. Um, and great proactive guidance feels\n\nUm, and great proactive guidance feels\n \n\nUm, and great proactive guidance feels\nlike serendipity, not interruption. It's\n\nlike serendipity, not interruption. It's\n \n\nlike serendip..." }, { "role": "assistant", "content": "isn't knowing what to suggest but when. isn't knowing what to suggest but when. Um, and great proactive guidance feels\n\nUm, and great proactive guidance feels\n \n\nUm, and great proactive guidance feels\nlike serendipity, not interruption. It's\n\nlike serendipity, not interruption. It's\n \n\nlike serendipity, not interruption. It's\nlike having a calendar app that suggests\n\nlike having a calendar app that suggests\n \n\nlike having a calendar app that suggests\nmeeting time is based on your energy\n\nmeeting time is based on your energy\n \n\nmeeting time is based on your energy\npatterns and not just availability. A\n\npatterns and not just availability. A\n \n\npatterns and not just availability. A\nwriting tool that notices when you're\n\nwriting tool that notices when you're\n \n\nwriting tool that notices when you're\nstuck and suggest taking a break rather\n\nstuck and suggest taking a break rather\n \n\nstuck and suggest taking a break rather\nthan pushing through. And the key is\n\nthan pushing through. And the key is\n \n\nthan pushing through. And the key is\nfinding that Goldilock zone, right? We\n\nfinding that Goldilock zone, right? We\n \n\nfinding that Goldilock zone, right? We\ndon't want to be too reactive. We also\n\ndon't want to be too reactive. We also\n \n\ndon't want to be too reactive. We also\ndon't be too overwhelming. Just right. don't be too overwhelming. Just right. don't be too overwhelming. Just right. Now, let's shift gears into principles\n\nNow, let's shift gears into principles\n \n\nNow, let's shift gears into principles\nfor designing augmentative AI because\n\nfor designing augmentative AI because\n \n\nfor designing augmentative AI because\nwithout trust, none of this matters. without trust, none of this matters. without trust, none of this matters. How? So how do we build that trust in\n\nHow? So how do we build that trust in\n \n\nHow? So how do we build that trust in\nthese augmentation systems? Trust\n\nthese augmentation systems? Trust\n \n\nthese augmentation systems? Trust\nbecomes uh particularly crucial in\n\nbecomes uh particularly crucial in\n \n\nbecomes uh particularly crucial in\naugment uh augmentation interfaces\n\naugment uh augmentation interfaces\n \n\naugment uh augmentation interfaces\nbecause the relationship between human\n\nbecause the relationship between human\n \n\nbecause the relationship between human\nand AI is more collaborative than\n\nand AI is more collaborative than\n \n\nand AI is more collaborative than\ntransactional. Right? Again think about\n\ntransactional. Right? Again think about\n \n\ntransactional. Right? Again think about\nhow to humans begin to trust each other\n\nhow to humans begin to trust each other\n \n\nhow to humans begin to trust each other\nin augmentation paradigm especially\n\nin augmentation paradigm especially\n \n\nin augmentation paradigm especially\ncognitive partnerships. It's likely that\n\ncognitive partnerships. It's likely that\n \n\ncognitive partnerships. It's likely that\nwe see human AI interact more like\n\nwe see human AI interact more like\n \n\nwe see human AI interact more like\nhumanto human interaction. I believe\n\nhumanto human interaction. I believe\n \n\nhumanto human interaction. I believe\ntrust needs to be progressive contextual\n\ntrust needs to be progressive contextual\n \n\ntrust needs to be progressive contextual\nand biirectional. Progressive in a sense\n\nand biirectional. Progressive in a sense\n \n\nand biirectional. Progressive in a sense\nthat it's easier to build trust with low\n\nthat it's easier to build trust with low\n \n\nthat it's easier to build trust with low\nstake suggestions before moving to high\n\nstake suggestions before moving to high\n \n\nstake suggestions before moving to high\nimpact decisions. um let a system prove\n\nimpact decisions. um let a system prove\n \n\nimpact decisions. um let a system prove\nitself um on small things first. Trust\n\nitself um on small things first. Trust\n \n\nitself um on small things first. Trust\nis also contextual um because we trust\n\nis also contextual um because we trust\n \n\nis also contextual um because we trust\npeople differently across different\n\npeople differently across different\n \n\npeople differently across different\ndomains and for different situations. So\n\ndomains and for different situations. So\n \n\ndomains and for different situations. So\nthe same should apply to AI. AI might be\n\nthe same should apply to AI. AI might be\n \n\nthe same should apply to AI. AI might be\nbrilliant at code reviews but terrible\n\nbrilliant at code reviews but terrible\n \n\nbrilliant at code reviews but terrible\nat design feedback or you might not have\n\nat design feedback or you might not have\n \n\nat design feedback or you might not have\nthe right or sufficient context to make\n\nthe right or sufficient context to make\n \n\nthe right or sufficient context to make\ncertain decisions. So let's make that\n\ncertain decisions. So let's make that\n \n\ncertain decisions. So let's make that\ntransparent and birectional because true\n\ntransparent and birectional because true\n \n\ntransparent and birectional because true\npartnership means both parties adapt. partnership means both parties adapt. partnership means both parties adapt. The AI learns your preferences while you\n\nThe AI learns your preferences while you\n \n\nThe AI learns your preferences while you\nlearn its capabilities. It should adjust\n\nlearn its capabilities. It should adjust\n \n\nlearn its capabilities. It should adjust\nits behavior based on how much you trust\n\nits behavior based on how much you trust\n \n\nits behavior based on how much you trust\nit in different\n\nit in different\n \n\nit in different\nscenarios. For AI to augment users, it\n\nscenarios. For AI to augment users, it\n \n\nscenarios. For AI to augment users, it\nimplies facilitating some sort of skill\n\nimplies facilitating some sort of skill\n \n\nimplies facilitating some sort of skill\ngrowth. Automation interfaces that stay\n\ngrowth. Automation interfaces that stay\n \n\ngrowth. Automation interfaces that stay\nstatic, augmentative AI that helps you\n\nstatic, augmentative AI that helps you\n \n\nstatic, augmentative AI that helps you\nlearn should also evolve with you. This\n\nlearn should also evolve with you. This\n \n\nlearn should also evolve with you. This\nmeans some sort of skill visualization\n\nmeans some sort of skill visualization\n \n\nmeans some sort of skill visualization\nshowing people their growth, their\n\nshowing people their growth, their\n \n\nshowing people their growth, their\ngrowing expertise over time, and\n\ngrowing expertise over time, and\n \n\ngrowing expertise over time, and\ngraduated complexity, right? Like\n\ngraduated complexity, right? Like\n \n\ngraduated complexity, right? Like\nunlocking or adapting features as\n\nunlocking or adapting features as\n \n\nunlocking or adapting features as\ncompetence increases or shifting\n\ncompetence increases or shifting\n \n\ncompetence increases or shifting\nexplanation patterns based on a user's\n\nexplanation patterns based on a user's\n \n\nexplanation patterns based on a user's\nunderstanding over time. The goal is\n\nunderstanding over time. The goal is\n \n\nunderstanding over time. The goal is\ngenuine skill enhancement, not just the\n\ngenuine skill enhancement, not just the\n \n\ngenuine skill enhancement, not just the\nillusion of it. Um, users should feel\n\nillusion of it. Um, users should feel\n \n\nillusion of it. Um, users should feel\nthemselves getting better um at their\n\nthemselves getting better um at their\n \n\nthemselves getting better um at their\ncore work and derive joy from\n\ncore work and derive joy from\n \n\ncore work and derive joy from\nself-improvement, even though it may be\n\nself-improvement, even though it may be\n \n\nself-improvement, even though it may be\nhard at first, not just more being more\n\nhard at first, not just more being more\n \n\nhard at first, not just more being more\ndependent on\n\ndependent on\n \n\ndependent on\nAI. Um, and a little note about um\n\nAI. Um, and a little note about um\n \n\nAI. Um, and a little note about um\nproduct metrics. If you truly care about\n\nproduct metrics. If you truly care about\n \n\nproduct metrics. If you truly care about\nyour users, you should track how they're\n\nyour users, you should track how they're\n \n\nyour users, you should track how they're\ngrowing as they use your product beyond\n\ngrowing as they use your product beyond\n \n\ngrowing as they use your product beyond\ntraditional engagement or usage metrics,\n\ntraditional engagement or usage metrics,\n \n\ntraditional engagement or usage metrics,\neven though they are important as well. even though they are important as well. even though they are important as well. Um, I think a lot of learning apps, for\n\nUm, I think a lot of learning apps, for\n \n\nUm, I think a lot of learning apps, for\nexample, focus on user retention and\n\nexample, focus on user retention and\n \n\nexample, focus on user retention and\nengagement, but are users actually\n\nengagement, but are users actually\n \n\nengagement, but are users actually\nimproving themselves and getting more\n\nimproving themselves and getting more\n \n\nimproving themselves and getting more\nand more capable in real life\n\nand more capable in real life\n \n\nand more capable in real life\nsituations. It's hard to say. So, why\n\nsituations. It's hard to say. So, why\n \n\nsituations. It's hard to say. So, why\ndoes this matter? Beyond a normative\n\ndoes this matter? Beyond a normative\n \n\ndoes this matter? Beyond a normative\nargument that AI should serve to make\n\nargument that AI should serve to make\n \n\nargument that AI should serve to make\npeople better. Well, three main reasons. people better. Well, three main reasons. people better. Well, three main reasons. First, building an emotional connection." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nFirst, building an emotional connection. First, building an emotional connection. AI products that make an effort to learn\n\nAI products that make an effort to learn\n \n\nAI products that make an effort to learn\nfrom users and fulfill intellectual\n\nfrom users and fulfill intellectual\n \n\nfrom users and ..." }, { "role": "assistant", "content": "First, building an emotional connection. First, building an emotional connection. AI products that make an effort to learn\n\nAI products that make an effort to learn\n \n\nAI products that make an effort to learn\nfrom users and fulfill intellectual\n\nfrom users and fulfill intellectual\n \n\nfrom users and fulfill intellectual\nneeds build stickier relationships. needs build stickier relationships. needs build stickier relationships. Chatg's memory feature is a great\n\nChatg's memory feature is a great\n \n\nChatg's memory feature is a great\nexample. People love feeling understood\n\nexample. People love feeling understood\n \n\nexample. People love feeling understood\nacross multiple chat sessions. Second,\n\nacross multiple chat sessions. Second,\n \n\nacross multiple chat sessions. Second,\nlong-term value. Buyers increasingly\n\nlong-term value. Buyers increasingly\n \n\nlong-term value. Buyers increasingly\ncare not just about productivity but\n\ncare not just about productivity but\n \n\ncare not just about productivity but\ncapability building. Can this tool make\n\ncapability building. Can this tool make\n \n\ncapability building. Can this tool make\nmy team not just faster but better? Um,\n\nmy team not just faster but better? Um,\n \n\nmy team not just faster but better? Um,\nthirdly, as consumers, we should be more\n\nthirdly, as consumers, we should be more\n \n\nthirdly, as consumers, we should be more\nintentional about how we learn alongside\n\nintentional about how we learn alongside\n \n\nintentional about how we learn alongside\nthese AI tools. We are investing our\n\nthese AI tools. We are investing our\n \n\nthese AI tools. We are investing our\nprecious time using them. So, let's make\n\nprecious time using them. So, let's make\n \n\nprecious time using them. So, let's make\nsure we're growing, not just\n\nsure we're growing, not just\n \n\nsure we're growing, not just\ngrinding. As we come to the end of this\n\ngrinding. As we come to the end of this\n \n\ngrinding. As we come to the end of this\ntalk, I'd just like to summarize what we\n\ntalk, I'd just like to summarize what we\n \n\ntalk, I'd just like to summarize what we\nhave walked through so far. Um,\n\nhave walked through so far. Um,\n \n\nhave walked through so far. Um,\ntechnological revolutions have seemed to\n\ntechnological revolutions have seemed to\n \n\ntechnological revolutions have seemed to\nalways been focused on efficiency. For\n\nalways been focused on efficiency. For\n \n\nalways been focused on efficiency. For\nAI though, um, I'm particularly excited\n\nAI though, um, I'm particularly excited\n \n\nAI though, um, I'm particularly excited\nbecause never had humanity advanced a\n\nbecause never had humanity advanced a\n \n\nbecause never had humanity advanced a\npiece of technology to a point where it\n\npiece of technology to a point where it\n \n\npiece of technology to a point where it\ncan help us think better, amplify our\n\ncan help us think better, amplify our\n \n\ncan help us think better, amplify our\nintuition, taste, and\n\nintuition, taste, and\n \n\nintuition, taste, and\ncreativity. This may take the form of\n\ncreativity. This may take the form of\n \n\ncreativity. This may take the form of\nblind spot detection, cognitive\n\nblind spot detection, cognitive\n \n\nblind spot detection, cognitive\npartnerships, and proactive guidance. partnerships, and proactive guidance. partnerships, and proactive guidance. And to get there, we need to invest in\n\nAnd to get there, we need to invest in\n \n\nAnd to get there, we need to invest in\nbetter UX that enable high trust,\n\nbetter UX that enable high trust,\n \n\nbetter UX that enable high trust,\nthoughtful agency, and learn alongside\n\nthoughtful agency, and learn alongside\n \n\nthoughtful agency, and learn alongside\nusers. I think the most profound\n\nusers. I think the most profound\n \n\nusers. I think the most profound\ntechnologies don't just replace humans. technologies don't just replace humans. technologies don't just replace humans. They unlock what makes us uniquely\n\nThey unlock what makes us uniquely\n \n\nThey unlock what makes us uniquely\nhuman. The next decade won't be about AI\n\nhuman. The next decade won't be about AI\n \n\nhuman. The next decade won't be about AI\ndoing our work, but AI helping us think\n\ndoing our work, but AI helping us think\n \n\ndoing our work, but AI helping us think\nin ways that we couldn't before. What's\n\nin ways that we couldn't before. What's\n \n\nin ways that we couldn't before. What's\ntransformative isn't just making us more\n\ntransformative isn't just making us more\n \n\ntransformative isn't just making us more\nproductive. It's about making us more\n\nproductive. It's about making us more\n \n\nproductive. It's about making us more\nthoughtful and more aware of our\n\nthoughtful and more aware of our\n \n\nthoughtful and more aware of our\ncognitive patterns like mirrors for our\n\ncognitive patterns like mirrors for our\n \n\ncognitive patterns like mirrors for our\nminds, showing us blind spots and\n\nminds, showing us blind spots and\n \n\nminds, showing us blind spots and\nsuggesting perspectives that we\n\nsuggesting perspectives that we\n \n\nsuggesting perspectives that we\nhabitually miss. And no doubt we are\n\nhabitually miss. And no doubt we are\n \n\nhabitually miss. And no doubt we are\nstill at the starting line of\n\nstill at the starting line of\n \n\nstill at the starting line of\nunderstanding how to build these\n\nunderstanding how to build these\n \n\nunderstanding how to build these\nsystems. But the core insight that\n\nsystems. But the core insight that\n \n\nsystems. But the core insight that\ntechnology should enhance rather than\n\ntechnology should enhance rather than\n \n\ntechnology should enhance rather than\nreplace human capability will remain\n\nreplace human capability will remain\n \n\nreplace human capability will remain\ntrue even as our understanding evolves\n\ntrue even as our understanding evolves\n \n\ntrue even as our understanding evolves\nand the future belongs to interfaces\n\nand the future belongs to interfaces\n \n\nand the future belongs to interfaces\nthat help us uh become more fully human\n\nthat help us uh become more fully human\n \n\nthat help us uh become more fully human\nnot less. With that, thank you for\n\nnot less. With that, thank you for\n \n\nnot less. With that, thank you for\nlistening and let's stay connected,\n\nlistening and let's stay connected,\n \n\nlistening and let's stay connected,\ncontinue your conversation." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nHey everyone, I'm Joe. I'm one of the\n\nHey everyone, I'm Joe. I'm one of the\n \n\nHey everyone, I'm Joe. I'm one of the\ncreators of Luminal. And so today in\n\ncreators of Luminal. And so today in\n \n\ncreators of Luminal. And so today in\nthis talk, I'm going to explain what it\n\nthis talk, I'm going to ex..." }, { "role": "assistant", "content": "Hey everyone, I'm Joe. I'm one of the\n\nHey everyone, I'm Joe. I'm one of the\n \n\nHey everyone, I'm Joe. I'm one of the\ncreators of Luminal. And so today in\n\ncreators of Luminal. And so today in\n \n\ncreators of Luminal. And so today in\nthis talk, I'm going to explain what it\n\nthis talk, I'm going to explain what it\n \n\nthis talk, I'm going to explain what it\nis, how it works, and why we really\n\nis, how it works, and why we really\n \n\nis, how it works, and why we really\nthink that it's the future of ML\n\nthink that it's the future of ML\n \n\nthink that it's the future of ML\nlibraries. So the title of the talk is\n\nlibraries. So the title of the talk is\n \n\nlibraries. So the title of the talk is\nradical simplific simplification through\n\nradical simplific simplification through\n \n\nradical simplific simplification through\nsearch. And that really is the theme of\n\nsearch. And that really is the theme of\n \n\nsearch. And that really is the theme of\nLuminal. We are far simpler than most\n\nLuminal. We are far simpler than most\n \n\nLuminal. We are far simpler than most\nother ML libraries. And yet we are not\n\nother ML libraries. And yet we are not\n \n\nother ML libraries. And yet we are not\ntaking a hit on performance or\n\ntaking a hit on performance or\n \n\ntaking a hit on performance or\ncapability because we are using uh\n\ncapability because we are using uh\n \n\ncapability because we are using uh\ncompilers and more specifically search. compilers and more specifically search. compilers and more specifically search. So deep learning fundamentally is very\n\nSo deep learning fundamentally is very\n \n\nSo deep learning fundamentally is very\nsimple. It's simple linear algebra. And\n\nsimple. It's simple linear algebra. And\n \n\nsimple. It's simple linear algebra. And\nso basically all that really is is is\n\nso basically all that really is is is\n \n\nso basically all that really is is is\nyou know scalers, vectors, matrices and\n\nyou know scalers, vectors, matrices and\n \n\nyou know scalers, vectors, matrices and\ntensors. And then a couple core ops, a\n\ntensors. And then a couple core ops, a\n \n\ntensors. And then a couple core ops, a\ncouple core operations. So we have\n\ncouple core operations. So we have\n \n\ncouple core operations. So we have\nadditions, multiplies, matt moles, some\n\nadditions, multiplies, matt moles, some\n \n\nadditions, multiplies, matt moles, some\nelementwise ops in between and that's\n\nelementwise ops in between and that's\n \n\nelementwise ops in between and that's\nthat's mostly it. But if deep learning\n\nthat's mostly it. But if deep learning\n \n\nthat's mostly it. But if deep learning\nis so simple, why aren't the libraries\n\nis so simple, why aren't the libraries\n \n\nis so simple, why aren't the libraries\nsimple? So the machine learning software\n\nsimple? So the machine learning software\n \n\nsimple? So the machine learning software\necosystem is very very complicated. Um\n\necosystem is very very complicated. Um\n \n\necosystem is very very complicated. Um\nPyTorch is one of the most famous\n\nPyTorch is one of the most famous\n \n\nPyTorch is one of the most famous\nlibraries out there. It has over 1,200\n\nlibraries out there. It has over 1,200\n \n\nlibraries out there. It has over 1,200\noperations\n\noperations\n \n\noperations\nuh and has over 15 different data types. uh and has over 15 different data types. uh and has over 15 different data types. It runs on a ton of different kinds of\n\nIt runs on a ton of different kinds of\n \n\nIt runs on a ton of different kinds of\ndevices. Uh CPU, CUDA, AMD, TPUs, uh\n\ndevices. Uh CPU, CUDA, AMD, TPUs, uh\n \n\ndevices. Uh CPU, CUDA, AMD, TPUs, uh\nsome NPUs out there. And the problem\n\nsome NPUs out there. And the problem\n \n\nsome NPUs out there. And the problem\nwith all of this is that complexity does\n\nwith all of this is that complexity does\n \n\nwith all of this is that complexity does\nnot scale just by adding these together. not scale just by adding these together. not scale just by adding these together. You don't have to have the number of\n\nYou don't have to have the number of\n \n\nYou don't have to have the number of\noperations plus the number of data types\n\noperations plus the number of data types\n \n\noperations plus the number of data types\nplus the number of uh supported devices. plus the number of uh supported devices. plus the number of uh supported devices. It's actually multiplicative. So the\n\nIt's actually multiplicative. So the\n \n\nIt's actually multiplicative. So the\nnumber the complexity scales with ops\n\nnumber the complexity scales with ops\n \n\nnumber the complexity scales with ops\ntimes data types times number of\n\ntimes data types times number of\n \n\ntimes data types times number of\ndevices. And so once you raise any of\n\ndevices. And so once you raise any of\n \n\ndevices. And so once you raise any of\nthose you want to support a new op or a\n\nthose you want to support a new op or a\n \n\nthose you want to support a new op or a\nnew data type or a new device your\n\nnew data type or a new device your\n \n\nnew data type or a new device your\ncomplex complexity starts exploding. So\n\ncomplex complexity starts exploding. So\n \n\ncomplex complexity starts exploding. So\nwhat that what's happened to pietorch is\n\nwhat that what's happened to pietorch is\n \n\nwhat that what's happened to pietorch is\nit's now over like 3 million lines of\n\nit's now over like 3 million lines of\n \n\nit's now over like 3 million lines of\ncode. Uh tensorflow is even much worse\n\ncode. Uh tensorflow is even much worse\n \n\ncode. Uh tensorflow is even much worse\nthan this. So it's extremely extremely\n\nthan this. So it's extremely extremely\n \n\nthan this. So it's extremely extremely\ncomplicated pieces of software and this\n\ncomplicated pieces of software and this\n \n\ncomplicated pieces of software and this\nhurts because there's a lot more bugs\n\nhurts because there's a lot more bugs\n \n\nhurts because there's a lot more bugs\nobviously if you have a ton more code\n\nobviously if you have a ton more code\n \n\nobviously if you have a ton more code\nbut it's also just very hard for people\n\nbut it's also just very hard for people\n \n\nbut it's also just very hard for people\nto ever extend or use or build anything\n\nto ever extend or use or build anything\n \n\nto ever extend or use or build anything\ninside\n\ninside\n \n\ninside\nof. So what we do is we take the\n\nof. So what we do is we take the\n \n\nof. So what we do is we take the\napproach of looking top down at machine\n\napproach of looking top down at machine\n \n\napproach of looking top down at machine\nlearning and then saying fundamentally\n\nlearning and then saying fundamentally\n \n\nlearning and then saying fundamentally\nwhat are the minimum amount of things we\n\nwhat are the minimum amount of things we\n \n\nwhat are the minimum amount of things we\nneed in order to make ML models run. So\n\nneed in order to make ML models run. So\n \n\nneed in order to make ML models run. So\ndeep learning again like I said is\n\ndeep learning again like I said is\n \n\ndeep learning again like I said is\nlinear algebra. Linear algebra boils\n\nlinear algebra. Linear algebra boils\n \n\nlinear algebra. Linear algebra boils\ndown to simple ops. So what if we just\n\ndown to simple ops. So what if we just\n \n\ndown to simple ops. So what if we just\nbuilt these very complicated models out\n\nbuilt these very complicated models out\n \n\nbuilt these very complicated models out\nof like Lego blocks of very simple\n\nof like Lego blocks of very simple\n \n\nof like Lego blocks of very simple\noperations. So what we do is we have uh\n\noperations. So what we do is we have uh\n \n\noperations. So what we do is we have uh\n12 operations that were very very very\n\n12 operations that were very very very\n \n\n12 operations that were very very very\nsimple. Um and so we have x2 log 2 s and\n\nsimple. Um and so we have x2 log 2 s and\n \n\nsimple. Um and so we have x2 log 2 s and\nreciprocal and square root. Those are\n\nreciprocal and square root. Those are\n \n\nreciprocal and square root. Those are\nour unary operations. We have uh our\n\nour unary operations. We have uh our\n \n\nour unary operations. We have uh our\nbinary operations addition,\n\nbinary operations addition,\n \n\nbinary operations addition,\nmultiplication, modulo and less than. multiplication, modulo and less than. multiplication, modulo and less than. And then we have our reductions. So we\n\nAnd then we have our reductions. So we\n \n\nAnd then we have our reductions. So we\nhave sum reduce and max reduce. And with\n\nhave sum reduce and max reduce. And with\n \n\nhave sum reduce and max reduce." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nAnd with\nthose just those operations, you can\n\nthose just those operations, you can\n \n\nthose just those operations, you can\nsupport all of the big models out there. support all of the big models out there. support all of the big models out there. All of the commercially relevant models\n\nAll of the c..." }, { "role": "assistant", "content": "And with\nthose just those operations, you can\n\nthose just those operations, you can\n \n\nthose just those operations, you can\nsupport all of the big models out there. support all of the big models out there. support all of the big models out there. All of the commercially relevant models\n\nAll of the commercially relevant models\n \n\nAll of the commercially relevant models\nthat everybody cares about. So you can\n\nthat everybody cares about. So you can\n \n\nthat everybody cares about. So you can\ndo language models, vision language\n\ndo language models, vision language\n \n\ndo language models, vision language\nmodels, uh CNN's, RNN's, you can do um\n\nmodels, uh CNN's, RNN's, you can do um\n \n\nmodels, uh CNN's, RNN's, you can do um\nuh diffusion models, all these other\n\nuh diffusion models, all these other\n \n\nuh diffusion models, all these other\nvery very popular models. very very popular models. very very popular models. Is that really it? I mean, that's not a\n\nIs that really it? I mean, that's not a\n \n\nIs that really it? I mean, that's not a\nlot of ops. Is that really all it takes\n\nlot of ops. Is that really all it takes\n \n\nlot of ops. Is that really all it takes\nto uh to represent all of this? Well,\n\nto uh to represent all of this? Well,\n \n\nto uh to represent all of this? Well,\nit's a lot of these other operations\n\nit's a lot of these other operations\n \n\nit's a lot of these other operations\nthat you would have expected to see on\n\nthat you would have expected to see on\n \n\nthat you would have expected to see on\nthe list are really just uh uh formable,\n\nthe list are really just uh uh formable,\n \n\nthe list are really just uh uh formable,\nusable by combining different operations\n\nusable by combining different operations\n \n\nusable by combining different operations\non this list. So, subtraction is really\n\non this list. So, subtraction is really\n \n\non this list. So, subtraction is really\nsimple. It's just addition and\n\nsimple. It's just addition and\n \n\nsimple. It's just addition and\nmultiplication with a negative one. multiplication with a negative one. multiplication with a negative one. Division is really simple because it's\n\nDivision is really simple because it's\n \n\nDivision is really simple because it's\njust multiplication and a reciprocal on\n\njust multiplication and a reciprocal on\n \n\njust multiplication and a reciprocal on\nB. Uh map moles are really simple if you\n\nB. Uh map moles are really simple if you\n \n\nB. Uh map moles are really simple if you\nhave the ability to manipulate the shape\n\nhave the ability to manipulate the shape\n \n\nhave the ability to manipulate the shape\nmetadata of your tensor. So you can\n\nmetadata of your tensor. So you can\n \n\nmetadata of your tensor. So you can\nbasically just do a broadcasted multiply\n\nbasically just do a broadcasted multiply\n \n\nbasically just do a broadcasted multiply\nand then sum reduce it down and you get\n\nand then sum reduce it down and you get\n \n\nand then sum reduce it down and you get\nthe output of matmo. So you don't need a\n\nthe output of matmo. So you don't need a\n \n\nthe output of matmo. So you don't need a\nmatt mole op. Um, convolution is really\n\nmatt mole op. Um, convolution is really\n \n\nmatt mole op. Um, convolution is really\nsimple if you have the ability to do\n\nsimple if you have the ability to do\n \n\nsimple if you have the ability to do\npooling through uh shape trackers again\n\npooling through uh shape trackers again\n \n\npooling through uh shape trackers again\nand then you do a maple like we just\n\nand then you do a maple like we just\n \n\nand then you do a maple like we just\ndiscussed with the the convolution\n\ndiscussed with the the convolution\n \n\ndiscussed with the the convolution\nkernel and you get the output of a\n\nkernel and you get the output of a\n \n\nkernel and you get the output of a\nconvolution. And so another thing we\n\nconvolution. And so another thing we\n \n\nconvolution. And so another thing we\nrealized when we were building this is\n\nrealized when we were building this is\n \n\nrealized when we were building this is\nthat all of these existing libraries are\n\nthat all of these existing libraries are\n \n\nthat all of these existing libraries are\nbuilt with dynamism at their core and\n\nbuilt with dynamism at their core and\n \n\nbuilt with dynamism at their core and\nthey were usually they were mostly built\n\nthey were usually they were mostly built\n \n\nthey were usually they were mostly built\nuh you know 5 to 10 years ago when\n\nuh you know 5 to 10 years ago when\n \n\nuh you know 5 to 10 years ago when\ndynamism was very important. So people\n\ndynamism was very important. So people\n \n\ndynamism was very important. So people\nwere experimenting with RNN's and LSTMs\n\nwere experimenting with RNN's and LSTMs\n \n\nwere experimenting with RNN's and LSTMs\nand all these fancy models and they\n\nand all these fancy models and they\n \n\nand all these fancy models and they\nneeded a whole lot of hackability and\n\nneeded a whole lot of hackability and\n \n\nneeded a whole lot of hackability and\ndynamism and they didn't so much care\n\ndynamism and they didn't so much care\n \n\ndynamism and they didn't so much care\nabout performance. Um so they you know\n\nabout performance. Um so they you know\n \n\nabout performance. Um so they you know\npietorrch is very very dynamic but\n\npietorrch is very very dynamic but\n \n\npietorrch is very very dynamic but\nbecause of that it's a lot more complex. because of that it's a lot more complex. because of that it's a lot more complex. So deep learning fundamentally isn't\n\nSo deep learning fundamentally isn't\n \n\nSo deep learning fundamentally isn't\ndynamic. it was just there for\n\ndynamic. it was just there for\n \n\ndynamic. it was just there for\nconvenience. But the dynamism inherent\n\nconvenience. But the dynamism inherent\n \n\nconvenience. But the dynamism inherent\nto models here is very very small and\n\nto models here is very very small and\n \n\nto models here is very very small and\nvery very bounded. So in a transformer\n\nvery very bounded. So in a transformer\n \n\nvery very bounded. So in a transformer\nmodel the only real thing that is\n\nmodel the only real thing that is\n \n\nmodel the only real thing that is\ndynamic is the KV cache length and the\n\ndynamic is the KV cache length and the\n \n\ndynamic is the KV cache length and the\nactual sequence length. And aside from\n\nactual sequence length. And aside from\n \n\nactual sequence length. And aside from\nthat the entire model is static. Um and\n\nthat the entire model is static. Um and\n \n\nthat the entire model is static. Um and\nso what we do is we specify these models\n\nso what we do is we specify these models\n \n\nso what we do is we specify these models\nas directed as cyclic graphs of\n\nas directed as cyclic graphs of\n \n\nas directed as cyclic graphs of\noperations. So right in front of us we\n\noperations. So right in front of us we\n \n\noperations. So right in front of us we\nsee a really simple example where we're\n\nsee a really simple example where we're\n \n\nsee a really simple example where we're\nloading in a tensor, we're loading in a\n\nloading in a tensor, we're loading in a\n \n\nloading in a tensor, we're loading in a\nweight and then we are elementwise\n\nweight and then we are elementwise\n \n\nweight and then we are elementwise\nmultiplying them together and then we\n\nmultiplying them together and then we\n \n\nmultiplying them together and then we\nare sum reducing them. And right here\n\nare sum reducing them. And right here\n \n\nare sum reducing them. And right here\nlike we talked about this is a maple. like we talked about this is a maple. like we talked about this is a maple. This is what a matrix multiply is. And\n\nThis is what a matrix multiply is. And\n \n\nThis is what a matrix multiply is. And\nso right on screen this is all that a\n\nso right on screen this is all that a\n \n\nso right on screen this is all that a\nsingle dense uh neural network layer is. single dense uh neural network layer is. single dense uh neural network layer is. It's just your your your matrix\n\nIt's just your your your matrix\n \n\nIt's just your your your matrix\nmultiplying by a weight matrix. So on\n\nmultiplying by a weight matrix. So on\n \n\nmultiplying by a weight matrix." } ]
[ { "role": "system", "content": "You are an AI Engineer content creator. Write in the style of Swyx and the Latent Space podcast - technical but accessible, with practical insights about AI engineering, agents, LLMs, and developer tools. Be direct, insightful, and focus on what matters to practitioners." }, { "role": "user", "content": "Share your insights on this AI engineering topic:\n\nSo on\nthe right hand side here we see a much\n\nthe right hand side here we see a much\n \n\nthe right hand side here we see a much\nlarger model and so these graphs you can\n\nlarger model and so these graphs you can\n \n\nlarger model and so these graphs you can\nkind of see get quite a bit more comp\n\nkind of..." }, { "role": "assistant", "content": "So on\nthe right hand side here we see a much\n\nthe right hand side here we see a much\n \n\nthe right hand side here we see a much\nlarger model and so these graphs you can\n\nlarger model and so these graphs you can\n \n\nlarger model and so these graphs you can\nkind of see get quite a bit more comp\n\nkind of see get quite a bit more comp\n \n\nkind of see get quite a bit more comp\ncomplex uh but they are capable of fully\n\ncomplex uh but they are capable of fully\n \n\ncomplex uh but they are capable of fully\nspecifying these\n\nspecifying these\n \n\nspecifying these\nmodels. So the consequence of this is\n\nmodels. So the consequence of this is\n \n\nmodels. So the consequence of this is\nthat luminal is really really simple\n\nthat luminal is really really simple\n \n\nthat luminal is really really simple\ndespite it being able to actually\n\ndespite it being able to actually\n \n\ndespite it being able to actually\nrepresent and run all of these different\n\nrepresent and run all of these different\n \n\nrepresent and run all of these different\nmodels in the world. It's under 5,000\n\nmodels in the world. It's under 5,000\n \n\nmodels in the world. It's under 5,000\nlines of code. Uh it's very easy to\n\nlines of code. Uh it's very easy to\n \n\nlines of code. Uh it's very easy to\nunderstand. And our goal here is that we\n\nunderstand. And our goal here is that we\n \n\nunderstand. And our goal here is that we\nare going to uh this whole library\n\nare going to uh this whole library\n \n\nare going to uh this whole library\nreally should be learnable in an\n\nreally should be learnable in an\n \n\nreally should be learnable in an\nafternoon. So you should be able to sit\n\nafternoon. So you should be able to sit\n \n\nafternoon. So you should be able to sit\ndown and understand the core structure\n\ndown and understand the core structure\n \n\ndown and understand the core structure\nand core concepts of it in you know a\n\nand core concepts of it in you know a\n \n\nand core concepts of it in you know a\ncouple hours here. So that doesn't\n\ncouple hours here. So that doesn't\n \n\ncouple hours here. So that doesn't\nnecessarily mean it's fast though. So\n\nnecessarily mean it's fast though. So\n \n\nnecessarily mean it's fast though. So\nLlama 7B uh right out of the box runs uh\n\nLlama 7B uh right out of the box runs uh\n \n\nLlama 7B uh right out of the box runs uh\nyou know it takes about all day to\n\nyou know it takes about all day to\n \n\nyou know it takes about all day to\ngenerate a single sentence. Super duper\n\ngenerate a single sentence. Super duper\n \n\ngenerate a single sentence. Super duper\nslow. But the point isn't to run these\n\nslow. But the point isn't to run these\n \n\nslow. But the point isn't to run these\nprimitive graphs of operations here. The\n\nprimitive graphs of operations here. The\n \n\nprimitive graphs of operations here. The\npoint is to take these graphs and then\n\npoint is to take these graphs and then\n \n\npoint is to take these graphs and then\nrun them through some functions to\n\nrun them through some functions to\n \n\nrun them through some functions to\ntransform them into faster graphs. And\n\ntransform them into faster graphs. And\n \n\ntransform them into faster graphs. And\nso we do that with compilers. So we\n\nso we do that with compilers. So we\n \n\nso we do that with compilers. So we\ncompile our way back to high\n\ncompile our way back to high\n \n\ncompile our way back to high\nperformance. So before we get to talk\n\nperformance. So before we get to talk\n \n\nperformance. So before we get to talk\nabout compilers here, I just want to\n\nabout compilers here, I just want to\n \n\nabout compilers here, I just want to\nagain highlight um what this\n\nagain highlight um what this\n \n\nagain highlight um what this\nsimplification gets us. A traditional\n\nsimplification gets us. A traditional\n \n\nsimplification gets us. A traditional\nstack is you might be using something\n\nstack is you might be using something\n \n\nstack is you might be using something\nlike hugging face transformers. It's a\n\nlike hugging face transformers. It's a\n \n\nlike hugging face transformers. It's a\ngreat library. It sits on top of\n\ngreat library. It sits on top of\n \n\ngreat library. It sits on top of\npietorch and xformers and uh you know\n\npietorch and xformers and uh you know\n \n\npietorch and xformers and uh you know\nsome other libraries that provide\n\nsome other libraries that provide\n \n\nsome other libraries that provide\noptimized kernels. These libraries\n\noptimized kernels. These libraries\n \n\noptimized kernels. These libraries\ninside them have a bunch of handwritten\n\ninside them have a bunch of handwritten\n \n\ninside them have a bunch of handwritten\nkernels uh that try to be adaptable for\n\nkernels uh that try to be adaptable for\n \n\nkernels uh that try to be adaptable for\nall these different use cases and then\n\nall these different use cases and then\n \n\nall these different use cases and then\nthose handwritten kernels may call\n\nthose handwritten kernels may call\n \n\nthose handwritten kernels may call\noperations in coupnn or kublos which\n\noperations in coupnn or kublos which\n \n\noperations in coupnn or kublos which\nthen sit on top of cuda and so all of\n\nthen sit on top of cuda and so all of\n \n\nthen sit on top of cuda and so all of\nthis creates a very complex dependency\n\nthis creates a very complex dependency\n \n\nthis creates a very complex dependency\nstory. Anybody who's tried to install a\n\nstory. Anybody who's tried to install a\n \n\nstory. Anybody who's tried to install a\ndeep learning setup on a new machine\n\ndeep learning setup on a new machine\n \n\ndeep learning setup on a new machine\nknows that this is a non-trivial thing. knows that this is a non-trivial thing. knows that this is a non-trivial thing. uh a lot of the times it's you know\n\nuh a lot of the times it's you know\n \n\nuh a lot of the times it's you know\ndependency hell that you're stuck in. Um\n\ndependency hell that you're stuck in. Um\n \n\ndependency hell that you're stuck in. Um\nand it's just it's just you know having\n\nand it's just it's just you know having\n \n\nand it's just it's just you know having\nvery complex uh stacks means that once\n\nvery complex uh stacks means that once\n \n\nvery complex uh stacks means that once\nthere's a bug tracing that down through\n\nthere's a bug tracing that down through\n \n\nthere's a bug tracing that down through\na really complex stack is is a huge pain\n\na really complex stack is is a huge pain\n \n\na really complex stack is is a huge pain\nas well. So generally we want to keep\n\nas well. So generally we want to keep\n \n\nas well. So generally we want to keep\nour stack as simple and straightforward\n\nour stack as simple and straightforward\n \n\nour stack as simple and straightforward\nas possible. Uh with Luminal we directly\n\nas possible. Uh with Luminal we directly\n \n\nas possible. Uh with Luminal we directly\nemit CUDA code. We directly generate\n\nemit CUDA code. We directly generate\n \n\nemit CUDA code. We directly generate\nCUDA code and so there really is nothing\n\nCUDA code and so there really is nothing\n \n\nCUDA code and so there really is nothing\nbetween us and CUDA. It's just our\n\nbetween us and CUDA. It's just our\n \n\nbetween us and CUDA. It's just our\nlibrary, our graph, our compilers, and\n\nlibrary, our graph, our compilers, and\n \n\nlibrary, our graph, our compilers, and\nCUDA right\n\nCUDA right\n \n\nCUDA right\nbeneath. So, like I said, Luminal is\n\nbeneath. So, like I said, Luminal is\n \n\nbeneath. So, like I said, Luminal is\nslow by default, but you're never meant\n\nslow by default, but you're never meant\n \n\nslow by default, but you're never meant\nto run any of these uh primitive op\n\nto run any of these uh primitive op\n \n\nto run any of these uh primitive op\ngraphs. So, you're supposed to we're\n\ngraphs. So, you're supposed to we're\n \n\ngraphs. So, you're supposed to we're\ngoing to take these graphs and we're\n\ngoing to take these graphs and we're\n \n\ngoing to take these graphs and we're\ngoing to feed them through compilers and\n\ngoing to feed them through compilers and\n \n\ngoing to feed them through compilers and\nthey're going to spit out much faster\n\nthey're going to spit out much faster\n \n\nthey're going to spit out much faster\ngraphs. So, how does that actually work? graphs. So, how does that actually work? graphs. So, how does that actually work?" } ]

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.