text stringlengths 0 2.68k |
|---|
**Jerod Santo:** The other reason you don't want to disrespect it is because in the robot uprising - and Adam has made this point many times - they're gonna be like "This is the guy that called us a golden retriever", you know? |
**Nathan Sobo:** Yeah, exactly. |
**Adam Stacoviak:** On acid. |
**Jerod Santo:** On acid... \[laughs\] |
**Nathan Sobo:** They're gonna have the record. |
**Jerod Santo:** "He did not respect us, and now he will respect us." |
**Adam Stacoviak:** You did say genius level, though... So you complimented -- |
**Jerod Santo:** That's true. |
**Adam Stacoviak:** Genius-level golden retriever on acid. |
**Jerod Santo:** It is such a mixed bag though, isn't it? |
**Nathan Sobo:** Yeah. And that's what makes it really fun, and mysterious, and frustrating, and all of the above to like build product around, so far. And I think we're definitely a team that's taken a certain kind of like algorithmic product development discipline, if that makes sense, sort of... We're the most extre... |
**Jerod Santo:** Yes. |
**Nathan Sobo:** And that's just been like an interesting edge, I think, for us to explore as a team. We're traversing the learning curve... I'm proud of us. And yeah, I think the cool thing is we'll have something unique to offer because of our deep understanding of the underlying tech. |
**Jerod Santo:** So how do you test a genius-level golden retriever on acid? |
**Nathan Sobo:** \[laughs\] Or yeah, how do you test a system that integrates as a fundamental component this inherently stochastic process? And we're like in the middle of traversing that learning curve. But I think for me, it was this realization that at the end of the day it's still kind of just a test, right? And s... |
**Jerod Santo:** Right. |
**Nathan Sobo:** And so we've got this -- I think we'll ship something interesting, standalone. Or we should. We want to move this into a procedural macro in Rust, but right now it's just a function that's just like eval... And run this test 100 freaking times, right? We run it kind of once up to a certain spot and sna... |
**Jerod Santo:** Okay. |
**Nathan Sobo:** And then we get a score. Like, how many of them passed, how many of them failed... It's sampling. |
**Jerod Santo:** And there's a threshold? Like, if 80% pass, then we're good? Or...? |
**Nathan Sobo:** \[12:00\] Yeah, yeah. We're playing with setting thresholds. You can commit a threshold that would potentially fail the build. We're being pretty conservative with those right now, because failing builds are annoying, and I think we're still kind of traversing the learning curve on all of this, quite h... |
And so getting that UI in place, getting it nice, the checkpointing for which the CRDT stuff has been dope to have, the review experience for which the multi buffers that we've invested in are really great... Like, seeing everywhere that the agent did the editing. And kind of the overall chrome around the experience - ... |
**Jerod Santo:** Right. \[laughter\] That reminds me in a small way of Gary Bernhardt's functional core imperative shell. And maybe it's the other way around. Where it's like, "Whatever it is, shell." And then, like, golden retriever core. \[laughter\] You're wrapping all the -- you're like blackboxing this thing that ... |
**Nathan Sobo:** Yeah. Yeah. And so this thing we were just working on, it'll be shipped by the time this airs, right? |
**Jerod Santo:** Yeah. |
**Nathan Sobo:** But what we were working on earlier is streaming edits out of the LLM. So we're using tool calls. But the problem with tool calls, at least with the \[unintelligible 00:14:26.07\] API as it stands is like the key-value pairs will stream, but the values will not stream. We had started with a little thin... |
**Jerod Santo:** \[laughs\] Yeah... |
**Nathan Sobo:** And we ran the eval again. |
**Jerod Santo:** Tell me that worked... |
**Nathan Sobo:** \[15:46\] It doesn't work. It didn't even work. And then Antonio was like "No, no, no, Nathan. Make it simple." And I was like "Okay, how do we do that?" He said -- I think it was sort of... After saying open tag, you must say close tag before you open another tag. So after saying old text, you must sa... |
**Jerod Santo:** Nice. |
**Nathan Sobo:** And then we're like "Okay, I guess we've got to go into the parser and just make it tolerant of this last 5% for a second." Anyway, that's just like a new little world for me. I'm used to just like having so much more control... But I'm enjoying it. It's like a fun challenge. It's a new type of product... |
**Adam Stacoviak:** It's almost like network latency. It's like, you can't control the network. You have no control over whether the call goes through or not. I mean, you can with the response code that comes back, but it's almost kind of like that. You can't predict whether the network will be there or not, or how fas... |
**Jerod Santo:** It's like a good comparison, but the prompting is the weird part. |
**Adam Stacoviak:** It is. |
**Jerod Santo:** Like, you can't just tell the network "Hey, now, when you send the packet..." You're literally like putting this little -- we're back to magic incantations, right? Like, deep down inside of Zed's agentic coding there's something that Nathan and your partner, your teammate came up with... |
**Nathan Sobo:** Antonio, yes. |
**Jerod Santo:** ...to like tell it how to spit out the right thing. It's so weird that \[unintelligible 00:17:12.28\] |
**Nathan Sobo:** And the cool thing is those rules and stuff, because we're open source, it's like, they're all online. |
**Adam Stacoviak:** They're there. |
**Jerod Santo:** Right, right. That is cool. |
**Nathan Sobo:** We have Zed.dev/leaked-prompts is the URL I want to put it on... |
**Jerod Santo:** Yeah. \[laughs\] |
**Nathan Sobo:** I'm like "Yeah, there's our leaked system prompts." We leak them ourselves, because we're open source. |
**Jerod Santo:** Yeah, get some journalists who don't know what they're talking about to write about you, you know... Like, "Look at this..." \[laughter\] Get some free press out of it. |
**Nathan Sobo:** "These people leak their own system prompts. What are they doing...?" \[laughs\] |
**Jerod Santo:** Well, it turns out either you leak them yourself, or someone else leaks them for you. Those things seem like they are unable to be held. |
**Nathan Sobo:** Yeah, that does seem true. What I will say about the networking thing is - yes, but it's even harder... Because we have simulated random network latencies, right? As I was describing, you can actually in Rust build your own custom scheduler that you drive with a random number generator, and every singl... |
**Jerod Santo:** I've always felt like especially web development, but software development in general is a house of cards, that is held together with silly putty, and duct tape. But I feel like at this point we've completely lost the plot and we've just gone insane, and just like amped up the level of complexity, obsc... |
**Nathan Sobo:** Yeah. I mean, I think it's always been true, but maybe more true than ever, that I really don't think you can make it be a house of cards. Like, that was kind of a lot of the premise of Zed in a lot of ways... It's like "How about we don't build this one like a house of cards?" I mean, no software is p... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.