Corey Knowles

Maybe it's like a human going to sleep.

It can be awake for so many hours.

And so if your task is, and by the way, that's a terrible analogy, but whatever.

If you're doing a really hard task that takes longer than the model's context window,

you would kind of historically get stuck.

And so this is something we've been working on, both making codecs faster, which is what we were talking about with this chart, but also making it smarter at running through long tasks.

So we have this feature here called compaction that basically we train the model how to basically prepare itself to run again in a fresh context window.

And then we built a harness to do that in a very specific way that the model is trained for.

So we're able to ship that feature super quickly.

I have a question about that.

Yeah, I mean, I don't know.

I think that's up to you.

But yeah, you could totally just like have Codex run for ages with just a single window.

There's a lot of different patterns.

Like actually one of the common patterns at OpenAI, which, you know, I think in the future you might...

you might not need this pattern as models get better and as the harness gets better.

But one thing that we commonly see is people just like, so I'm just opening another terminal here.

You see people like working with the model in one pane and this context they're managing, we're kind of skipping some steps and getting more to the hero side of what you guys are talking about.

Yeah, so this basically like if I'm working with Codex here, I might want sort of this one chat with the model where we call them threads.

where I'm very opinionated and confident about what I wanted to do.