Ahmed El-Kishky

Right now, science has sort of been, you know, bottlenecked a bit.

But imagine giving a tool to scientists and maybe mathematicians where it can, you know, work with them.

You basically have some sort of idea or you want to investigate an area and you have the ability to spend a lot of compute sort of thinking about it, trying things out and coming up with something novel, like something that humanity has never seen before.

And that's sort of where I see things moving, I guess, in the near future.

There's going to be a lot of focus on making sure our models can solve these very difficult problems that will take longer than a few hours.

And I think the Codex achievement is just a stepping stone towards this.

We've shown that it is possible.

Each day we're sort of like, you know, solving tasks that take longer and longer.

We are very bullish on scale still at OpenAI.

I know some people mention, oh, maybe scaling isn't the answer, but I think scaling does sort of help.

Having models that maybe are trained with more RL.

We posted in our blog when we first announced O-Series, we had this really amazing scaling law that we sort of like,

show it in our charts besides pre-training as we like put more and more trained compute into RL we see better performance so yeah in this case as we like just train longer for reinforcement learning our models get smarter but then also as we let the models think longer in this case sort of expend tokens they're also able to get smarter so we're really heavily betting on that so we want to make sure that's

our reinforcement learning algorithms are working well, that we have the compute necessary to continue scaling these to get better and better performance.

But another one is just good tool use.

Having a model that can actually engage with the world in meaningful ways is a key component here.

If we constrain the model to just thinking in text and never getting that experience maybe,

trying stuff on a computer or maybe eventually trying stuff in the physical world, we will sort of be limiting what the model can actually tackle and solve.

So my personal opinion is that the things that are necessary is to make sure that as you scale, things, you know, continue working.

Um, when you throw a lot of compute, a lot of computers at problems, uh, things always start breaking down, you know, just the, you know, how it goes as things will get more complex, everything gets way more difficult.