Ahmed El-Kishky

And it's honestly very data-driven, very scientific.

And sometimes it takes a bit of coding.

Like, we're not all just, like, running experiments.

We code a bunch.

We sort of see how well these models then, you know, solve something.

And we just reiterate.

Once we're sure that...

what we propose is actually correct, makes our model smarter, makes your model better in some aspects.

Eventually it gets incorporated into what becomes Chai GPT.

But a lot of it is just, you know, going about your day, just banging your head sometimes being like, why won't this work?

It's honestly, in that way, it's very similar to a lot of other work, but it's really satisfying when you get to those situations where you see the fruits of your labor.

That could be like, you know, weeks, months, sometimes years away from when you started projects.

So it's often worth it in the end.

So shout out first to Andrei Mischenko, Hansen, Katie Shi, and the rest of the Codex team.

They've done an amazing job with the new Codex model, and I see it all online, people just raving about it.

Yeah.

So one of the things I think they're really excited about is how much it thinks.

So if you ask it an easy problem, it responds really quickly.

If you ask it a very difficult problem, it can take hours and hours to sort of think about it, work it, try things out in a very agentic manner.

And I think that's...