Corey Knowles

It deleted the cape.

That's just, you know...

me preferring no capes all right there we go no capes so you know under the hood basically what happens what we have is we have a bunch of code where was it here

in this repository that everyone can go read.

And so when I send this query, the model sees this and then it decides, OK, like what can I do?

And so there's some prompts in here and some tools that it can use in the prompt.

Say like hey, you're able to you to execute commands in the users terminal.

You can run these commands in this way and to make sure that it's safe.

We have this thing called a sandbox which limits what you can do to only things that the user would want you to do.

because we're both training the model and building the harness, we're able to do these things together.

Um, so to, to build them, to work really well together, um, that means we're able to like move really fast and like try experiments.

Uh, and you know, it's kind of like, instead of building like a general purpose thing, we're able to build something very specific.

So like when you're, when you're using the codex models, if you wanna like just try using them directly.

They're a bit more like a sports car that's very good at being used a specific way and a bit less good at being used.

If you want to use them for something outside of Codex, they're not as useful.

But then you get a sports car.

And we get to build cool features.

I think we actually talked about it here.

One of the cool things you can do with Codex is you can run these really long running tasks.

Models have something called a context window, which is basically how long they can think for before you need to reset them.