Venture with Grace
00:32:16 · Grace Gong
[00:32:16] S0: You said that coding agents fail mostly for one dumb reason, which is obeying prompt instruction to literally. What would you say is like the most like ridiculous like agent failure you've seen and...
00:32:31 · Vitaly Gordon
[00:32:31] S1: Oh my god. Well, look, there is like, there's just, like, you know, so many so many problems from... Like, look. At the worst case scenario is, like, you know, kind of sometimes in order for, you know, like, some test, like, you know, these agents would go and delete a database. Right? Because when you... Especially, like, go in YOLO mode, that's probably, like, the... Like, you know, we have them. Obviously, those models became smarter over time, but it still it still happens. And, like, they maybe, you know, not delete a database, but would, you know, delete, you know, bunch of files that you need in your system. The... A little bit less kind of things, like, they're, you know, kind of like sometimes will fail to, you know, the test that will delete the test. Right? And so like, okay, awesome. That's not the, you know, what I, you know, I intended. So there is kinda like, you know, still, I would say, even alignment issues, you know, with coding. And and I would say probably something that just happens far more often, maybe, you know, not as annoying, is, like, in many cases, they just add a bunch of code that you didn't ask for. Right? Like, okay. You know, you asked for, like, hey. I just want to fix this, you know, one small bug and, you know, just add it. And then they, you know, added the, you know, like, a bunch of tests on documentation and some UI for you to, you know, verify it and, like, you know, they just, like, go and do a lot of things because they're trying to help. And it's just, you know, it's just hard to control these, you know, billion dollar boxes that, you know, have a mind of their own.
00:34:13 · Grace Gong
[00:34:13] S0: When, I guess, like where does human in the loop sits in this coding journey right now? Since I feel like AI is generating a bunch of the code and then you guys will track the productivity of it. I guess where do human even sit in this entire coding journey, especially with a lot of long horizon tasks that's generated by different agents?
00:34:43 · Vitaly Gordon
[00:34:43] S1: Yeah. So there's a of couple of ways people think about it. Right? So maybe, you know, the most sophisticated users, they will spend most of their time on the context and the evals and then kind of let machines kinda do their their thing. All that, there are there are people who are, you know, a little bit maybe kind of... Again, we'll go and have the local horizon task, and then they will just kinda go and validate kind of, like, the result and and do it. And there's people who are, you know, kind of a little bit more, you know, kind of guiding the agents a little bit more. So one of things that we also start now getting data for is, like, a lot of session data. And we actually, like, look and we identify that people are using AI very, very differently. And and and so this is... I was thinking of, like, sorry. I I would go on kind of a different path for for a second there. I I think we kinda underestimated of what it means to move from, like, descriptive coding languages where, you know, there are very, very concrete best practices on how you like, do things to, like, just start in English, to kind of like, machines. And the thing is, like, you can speak English, but, like, there is, like, so many ways that you can ask you, you know, your helper to achieve a task. And what we found is like, there is just very different ways of... Very different ways of kind of using using these agents in order, to achieve those goals. And so therefore, it's kind of, you know... Yeah. There was just, like, no no one way. I would add another thing that I just noticed on myself where the Terra model that just came in. So I'm a Codex user mostly, Terra just came out. The model is pretty fast. It's, like, three times faster than, you know, solo or Astra. Now what it means is actually, like, if before, I would like... My work thing would be, hey. I cannot go and define task and then I go and do something else. In this, in many cases, like, the task actually completes very quickly. And so that's gonna also change my style. And so and so it got me start thinking, we go with, like, diffusion models and there's, like, new models that appear that are now, like, let's say, 10 times or or somehow we get a lot more capacity. What does it mean if, like, now a model can finish a task, like, times faster than before? Like, I don't know.
00:37:14 · Vitaly Gordon, Grace Gong
It probably will change the way I work again. And so those are also some of the things that we're thinking about. [00:37:25] S0: So when, like, how does your current workflow look like at your own company? Like, do you guys like generate codes and then like, you know, I guess, obviously you guys generate codes, like, what I meant is like, how do you implement your own product and your own kind of like day to day? [00:37:40] S1: Yeah. So we're a little bit of a weird company because part of what the... What we offer to customers is like, is we will have answers for different ways of working. So And we actually are almost like forcing people to work in different ways just so we test the different kind of... It's could be like different harnesses, different types of model, know, more expensive, cheaper, kind of things like that. So we try we try to kind of distribute and kind of, like, understand, like, what works. So it's not necessarily the... While our product, you know, allows to, like, one optimized way, we're constantly, you know, trying to kind of increase the the exploration. But we definitely have ways that we found that some ways are... You know, could literally be 10 times more efficient than, you know, than other than other ways, and that's also something that we share both internally and with our with our customers. So while we, you know, we know the answer, actually, we kind of, like, almost, like, force ourselves to work in the not not the most efficient way. [00:38:51] S0: What do mean by like, you're like asking them to try out different things?
00:38:51 · Grace Gong, Vitaly Gordon
[00:38:51] S0: What do mean by like, you're like asking them to try out different things? [00:38:55] S1: Yeah. Like, like, I'll give an example is, you know, we have our, you know, our own harnesses that we, you know, that have a lot more context and that we do, but, like, not all of our engineers are using it because we don't want them because, like, our customers will not be using them. Right? That's kind of the things, like, things like that. And there's, you know, certain, you know, skills and, you know, and ways to, like, add context to things. So there's, like, a lot of, you know, automation. And in some cases, like you said, we have a codecs and clog, but also we have open weights models and we have other things. So and sometimes, like I said, we wanna, especially if we have customers who, let's say, are, maybe a little bit more, I would say, regulated and they actually are not using the frontier models, but they, like, literally actually running, you know, some open source models on their own kind of bare metal hardware. Like, we also wanna understand, you know, what it feels like, what... You know? So we will have some engineers who, like, literally, will buy buy machines, put them in our in our office, and have assigned some engineers to work that way just so we understand the implications, and we can go in. So it's kind of like a weird analogy of, like, how people used to, like, you know, test mobile apps on many many different mobile devices. Like, that's kind of like, we have this, like, you know, weird AI equivalent of that.