Shawn Wang: belief

Transcript context

…source it from Cbench, actually. Yeah, you can do it, you can do it super easily. And that's how we got that figure out at the other end. Um, for us being able to see it against, um, our historic models were super useful. So we could see if we were, you know, actually helping ourselves or not. And initially, one of the biggest performance gains that we saw when we were work, when we did work on the RAG a bit was giving it the ability to use the LSP to like go to definition and really try to get it to emulate how we do that, because I'm sure when you go into an editor with that, where like the LSP is not working or whatever, you suddenly feel really like disarmed and naked. You're like, Oh my god, I didn't realize how much I actually used this to get about rather than just find stuff. So we really tried to get it to do that and that gave us a big jump in performance. So we went from like 54 percent up to like the 60s, but just by adding, focusing on that. One weird trick. Yes. I'll briefly comment here. So this is the standard approach I would say most, uh, code tooling startups are pursuing. The one company that's not doing this is magic. dev. So would you do things differently if you have a 10 million token context window? If I had a 10 million context window and hundreds of millions of dollars, I wouldn't have gone and built, uh, it's an LTM, it's not a transformer, right, that they're using, right? If I'm not mistaken, I believe it's not a transformer. Yeah, Eric's going to come on at some point. Listen, they obviously know a lot more about their product than I do. I don't know a great deal about how magic works. I don't think he knows anything yet. I'm not going to speculate. Would I do it the same way as them? I like the way we've done it because fundamentally like we focus on the Active software engineering and what that looks like and showing models how to do that. Fundamentally, the underlying model that we use is kind of null to us, like, so long as it's the best one, I don't mind. And the context windows, we've already seen, like, you can get transformers to have, like, million, one and a half million token context windows. And that works perfectly well, so like, as soon as you can fine tune Gemini 1. 5, then you best be sure that Genie will run on Gemini 1. 5, and like, we'll probably get very good performance out of that. I like our approach because we can be super agile and be like, Oh, well, Anthropic have just released whatever, uh, you know, and it might have half a million tokens and it might be really smart. And I can just immediately take my JSONL file and just dump it in there and suddenly Genie works on there and it can do all the new things. Does…

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