Corey Knowles

I might say, okay, how many, this is a really simple question, right?

Folders are in here.

We're going to see that Codex is going to like think a little bit about what my question is and then what it should do.

And then it's going to run a command and it's going to like think about the answer and then it's going to give me an answer, right?

So yeah, this like it had to think for a little bit, even though this is simple question, because I didn't tell it like run this command.

I told it I asked it this like natural language question, right?

So it thought for a very short amount of time.

And so the more stuff is here, the more thinking.

The further along this axis we go in terms of thinking tokens.

And then if you take this exam, this eval, so we've been verified, there's like, okay, well, how good is it?

What score does it get?

And basically, my analogy would be like, if I give you an exam, it's quite a hard exam.

And I say, you have 30 minutes total to finish the exam versus like 15 minutes versus you have two hours.

We can kind of tell the model like, hey, like think kind of medium hard about it.

Think very hard.

Think extra hard.

And so what's really cool here is that not only is the model like able, if it has a lot of time to like way outclass the prior models in terms of how well it can do in this exam, but also it can achieve similar scores with many fewer thinking tokens, which means just faster for you as a user.

And also it means it's cheaper for you.

Like you can use more.

Yeah.