Ahmed El-Kishky

So maybe it'll happen in a later version of GPT-5.

Maybe it'll happen in GPT-6.

But ultimately, our goal is whatever sort of we're experimenting on should eventually make it into not just a experimental model just available to a small group of people, but to everybody.

It's honestly like being a graduate student sometimes.

So we all have projects that we think are incredibly useful and projects that are going to make our models a lot smarter.

Usually we talk about them.

We discuss them with our colleagues.

We get some good constructive feedback.

Sometimes people are like, actually, no, that's not going to work.

And then maybe you have to actually push through.

Nobody bets a thousand.

Exactly.

And then we basically, we're just scientists.

It's a bunch of nerds just really liking the work we do, really believing in AI.

We are working on it.

We run these models.

We train them.

We sort of do an experiment, see if our hypotheses are actually correct or not.

So we're measuring, like, oh, did this intervention make the model smarter this way?

Or did it, like, cause the model to do this?