text stringlengths 0 2.68k |
|---|
**Adam Jacob:** But I don't feel like I'm doing a lot of babysitting. I feel like I'm doing a lot of -- it feels like engineering. It's different, but it definitely feels like engineering to me. |
**Jerod Santo:** I would agree, and maybe it's because I don't like to plan very much, and so I feel more like I'm iterating with it doing the work, and me just... I don't want to use the word "babysitting", because then Adam will be right... |
**Adam Stacoviak:** Ha-hah! |
**Jerod Santo:** \[laughs\] I'm not babysitting so much as I'm just directing. I'm directing the work. And I'm waiting for it to be ready for the next direction, or the review... |
**Adam Stacoviak:** Traffic cop maybe is a better analogy. |
**Jerod Santo:** Yeah. Well -- |
**Adam Stacoviak:** You need more traffic cops. |
**Adam Jacob:** The volume of parallelism that you can get out of this is crazy high. It's another way that all the systems that we drive will have to change. Source code is a great example. You know what sucks in the current model? It can hallucinate syntax. And there's no way to know until late in the game, until I r... |
\[01:08:27.16\] And how much more efficient will that make that loop? ...because instead of it waiting for the compile loop and the AI wrote bad code, the AI will know it wrote bad code immediately, at the moment it wrote bad code, then get corrected by the compiler. The compiler would be like "Oh, you did it wrong", a... |
**Adam Stacoviak:** No. |
**Adam Jacob:** And that loop - that's what I mean when I'm like "Oh, it's going to change a lot about how we think about how these systems are constructed." Because once you start designing the system to make that loop delicious, it's going to be real different than, like, I'm going to crank off 10,000 lines of source... |
**Adam Stacoviak:** Yeah, how would you -- can you think about that a little bit? I didn't even think about a new language, post AI, and being AI-native, really, to use a popular term out there. |
**Adam Jacob:** Yeah. I mean, I haven't thought about it a ton, but the first example was one I'll go back to, which is when the AI agents perform better, they perform better when the feedback loops are closer at hand. So when they hallucinate, which you know they will do, you have to correct the hallucination. And so ... |
And what do you have to do in programming language land to make that loop be good? I don't know, because I'm not a programming language guy, I'm a practical technology guy. But Unison, for example, does this with a database, where when you write code in Unison, it synchronizes up to this big database that makes a hash ... |
**Adam Stacoviak:** Well, I think some are because of the AI slop, right? Like, that kind of prevents -- if you could perfect that world, even if you didn't go AI-native, which is to rebuild from the ground up with AI mind... Even if you took like Rust, or Python, and said "Look, let's bolt on that kind of world." |
**Adam Jacob:** Yeah. If you could figure it out... |
**Adam Stacoviak:** When you run Cargo, it does that job. There's a better feedback loop. Assuming the build is being run by an agent being led or directed by a human - if that assumption is true, yeah. |
**Adam Jacob:** \[01:12:16.25\] Yeah. Yeah. And that's what I mean when I say that it's all -- the technology we have now... I don't need AGI for that. I don't need the bubble to pop. I don't need any more technology than the one I already have. We could build that right now. The only thing that's in our way is that we... |
I talk to people all day who hate infrastructure as code. That's their opening gambit when they talk to me, and they're still a little hesitant to get rid of it. You know what I mean? So it would be even harder in those cases. But that doesn't mean anything... You do it anyway, because doing it anyway is the way we fig... |
And yeah, I don't know -- I think that conversation we just had about programming languages, we could have it in every piece of the stack. We could have it in every industry, and every vertical... Essentially everywhere. And we don't need any new technology at all to do it. None. We have all of it right now. |
**Adam Stacoviak:** Does that just feed into the obvious beast of OpenAI, Anthropic, those folks having -- I said tollbooth before, in a different podcast... This idea that we now have to pay the piper to play the game of software development, essentially. It's like, okay, if an LLM or some sort of agentic tool is par ... |
**Adam Jacob:** For sure. But that'll create new incentives for us to be like "Well, wouldn't it be better if we didn't have to pay Anthropic all this money? Wouldn't it be better if you could run it on your desktop?" Because the loop would be better... The reason we're building AI data centers to do inference is becau... |
So yes, I think it'll grow their moat, it'll do all those things... But then what it'll do is create a new opportunity, which is "Why am I paying all this money to Anthropic all the time? Wouldn't it be better if - blah, blah, blah, blah?" And around the cycle will go. It won't be like an end game state, where it's lik... |
**Jerod Santo:** \[01:15:59.13\] Yeah, so the agent client protocol that Zed came out with, and is starting to get deployed out there is basically leaning into that, where it's like "We're making an awesome editor. We don't want to write the agent, we just want somebody who's putting all this effort into the agent, mak... |
**Adam Jacob:** I think everybody should. The way we should think about the agent is glue. That's the thing I've been -- that's my pithy sentence that I'm trying to get everyone to repeat, is "The agent is glue." Think of it like a Perl script that you wrote. See, I've aged myself again. |
**Jerod Santo:** What's that again? What's Perl? \[laughter\] |
**Adam Jacob:** Perl used to run the internet. We all had the -- I'm so sorry. Anyway... Think of it like the script that you write. That's more what agents are going to be. But the first generation of this, we were all like "Oh, we'll build custom embedded agents. They'll be on autopilot." There's all these startups t... |
**Jerod Santo:** So how does that show itself in System Init today? Is it just using Claude Code, or...? |
**Adam Jacob:** Yeah. It shows itself by me not embedding the agent, and not trying to build a wall. |
**Jerod Santo:** Right. |
**Adam Jacob:** So my moat is that I build -- I have the best deterministic system to be driven by an agent. If what you want to do right now is manage infrastructure in AWS, or in Azure soon, or in other places, what you're going to do is use System Initiative within an agent and you're going to have it do stuff like ... |
**Jerod Santo:** Mm-hm. So how much sauce is there in System Initiative that's not the agent doing things? |
**Adam Jacob:** A ridiculous amount of sauce. Like six years of R&D sauce. Hundreds of thousands of lines of source code... \[laughs\] |
**Adam Jacob:** What kind of stuff are you doing -- what are you bringing to the table that I couldn't get by plugging Claude into EC2, or something? |
**Adam Jacob:** Yeah, yeah. So there's a couple of things. So one is the way that the models get built. So it turns out that if you want the agent to be able to drive something, what you want to do is mimic the outside world as close as possible. So for example, we don't abstract AWS from you. We just expose AWS the wa... |
\[01:19:58.01\] The other thing we're doing is taking that one-to-one model and because we have this strict modeling language, when it hallucinates, we correct it in that moment. So when it tries to invent a property or whatever, we just tell the LLM "That property is not real here. Use this tool to read the directions... |
And then it goes and discovers it on its own. We put your own policy in that same loop. So if you want to make sure that you're compliant, that's how that happens. It doesn't happen later, it happens right when the moment the hallucination happens. |
You can create your own new models from scratch. So you can be like "Hey, I have this API document. I want to use it in System Initiative. Go build me the assets for this internal system so I can use it through System Initiative." And it will just run off and read that source code, read that description, and then build... |
And change sets - you don't want to work in the cloud, YOLOing infrastructure. You don't just want to throw the MCP server of AWS at an agent and be like "If you want to, delete the database." You need change sets. It turns out you need change control. You need the loop. And all of that stuff is what we provide. And th... |
**Adam Stacoviak:** To automate it. Yeah. |
**Adam Jacob:** And it works better. |
**Adam Stacoviak:** Yeah. |
**Adam Jacob:** It turns out it works great. But, yeah, that's where it is. And that's one of the mistakes that we made in the early era of thinking about AI systems, was we were like "Oh, all the value's in the LLM." And I would make the counter argument. It turns out the LLM is useless, pretty much. It's cool for gen... |
**Jerod Santo:** Sure. |
**Adam Jacob:** Yeah. |
**Jerod Santo:** Tell us more about this custom model. |
**Adam Jacob:** Like building your own stuff in System Initiative? |
**Jerod Santo:** You said you \[unintelligible 01:22:30.27\] a custom model, and I'm not sure what you mean by that. |
**Adam Jacob:** Oh, yeah, yeah, yeah. |
**Adam Stacoviak:** You said you \[unintelligible 01:22:33.22\] an API spec, I believe, and said "Build a model around this." |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.