Speaker unverified: preference

Transcript context

…Okay. We'll come back to that later, but let's talk about... Yeah. So, so just like shortly, and then we can double click on Alpha Codium. But Alpha Codium is a open source tool. You can go and try it and lets you compete on CodeForce. This is a website and a competition and actually reach a master level level, like 95% with a click of a button. You don't need to do anything. And part of what we did there is taking a problem and breaking it to different, like smaller blocks. And then the models are doing a much better job. Like we all know it by now that taking small tasks and solving them, by the way, even O1, which is supposed to be able to do system two thinking like Greg from OpenAI like hinted, is doing better on these kinds of problems. But still, it's very useful to break it down for O1, despite O1 being able to think by itself. And that's what we presented like just a month ago, OpenAI released that now they are doing 93 percentile with O1 IOI left and International Olympiad of Formation. Sorry, I forgot. Exactly. I told you I forgot. And we took their O1 preview with Alpha Codium and did better. Like it just shows like, and there is a big difference between the preview and the IOI. It shows like that these models are not still system two thinkers, and there is a big difference. So maybe they're not complete system two. Yeah, they need some guidance. I call them system 1.5. We can, we can have it. I thought about it. Like, you know, I care about this philosophy stuff. And I think like we didn't see it even close to a system two thinking. I can elaborate later. But closing the brackets, like we take Alpha Codium and as our principle of thinking, we take tasks and break them down to smaller tasks. And then we want to exploit the best model to solve them. So I want to enable anyone to enjoy O1 and SONET and Gemini 1.5, etc. But at the same time, I need to develop my own models as well, because some of the Fortune 500 want to have all air gapped or whatever. So that's a challenge. Now you need to support so many models. And to some extent, I would say that the flow engineering, the breaking down to two different blocks is a necessity for us. Why? Because when you take a big block, a big problem, you need a very different prompt for each one of the models to actually work. But when you take a big problem and break it into small tasks, we can talk how we do that, then the prompt matters less. What I want to say, like all this, like as a startup trying to do different deployment, getting all the juice that you can get from models, etc. is a big problem. And one need to think about it. And one of our mitigation is that process of taking tasks and breaking them down. That's why I'm really interested to know how you guys are doing it. And part of what we do is also open source. So you can see. There's a lot in there. But yeah, flow over prompt. I do believe that that does make sense. I feel like there's a lot that both of you can sort of exchange notes on breaking down problems. And I just want you guys to just go for it. This is fun to watch.…

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