Alistair Pullen: evaluation

Transcript context

…out. Yeah, no, totally. You mentioned bias in the data and some of these things. In your announcement video, you called Genie the worst verse AI software engineering colleague. And you kind of highlighted how the data needed to train it needs to show how a human engineer works. I think maybe you're contrasting that to just putting code in it. There's kind of like a lot more than code that goes into software engineering. How do you think about the data mixture, you know, and like, uh, there's this kind of known truth that code makes models better when you put in the pre training data, but since we put so much in the pre training data, what else do you add when you turn to Genium? Yeah, I think, well, I think that sort of boils down fundamentally to the difference between a model writing code and a model doing software engineering, because the software engineering sort of discipline goes wider, because if you look at something like a PR, that is obviously a Artifact of some thought and some work that has happened and has eventually been squashed into, you know, some diffs, right? What the, very crudely, what the pre trained models are reading is they're reading those final diffs and they're emulating that and they're being able to output it, right? But of course, it's a super lossy thing, a PR. You have no idea why or how, for the most part, unless there are some comments, which, you know, anyone who's worked in a company realizes PR reviews can be a bit dodgy at times, but you see that you lose so much information at the end, and that's perfectly fine, because PRs aren't designed to be something that perfectly preserves everything that happened, but What we realized was if you want something that's a software engineer, and very crudely, we started with like something that can do PRs for you, essentially, you need to be able to figure out why those things happened. Otherwise, you're just going to rely, you essentially just have a code writing model, you have something that's good at human eval, but But, but not very good at Sweet Eng. Essentially that realization was, was part of the, the kernel of the idea of of, of the approach that we took to design the agent. That, that is genie the way that we decided we want to try to extract what happened in the past, like as forensically as possible, has been and is currently like one of the, the main things that we focus all our time on, because doing that as getting as much signal out as possible, doing that as well as possible is the biggest. thing that we've seen that determines how well we do on that benchmark at the end of the day. Once you've sorted things out, like output structure, how to get it consistently writing diffs and all the stuff that is sort of ancillary to the model actually figuring out how to solve a problem, the core bit of solving the problem is how did the human solve this problem and how can we best come up with how the human solved these problems. So all the effort went in on that. And the mix that we ended up with was, as you've probably seen in the technical report and so on, all of those different languages and different combinations of different task types, all of that has run through that pipeline, and we've extracted all that information out. How does that differ when you work with customers that have private workflows? Like, do you think, is there usually a big delta between what you get in open source and maybe public data versus like Yeah,…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.