Andy Halliday

Because you're not activating every single, you know, perceptron in the network every single time.

You're saying, okay, we've identified that this area over here is expert in this area and this query relates to that.

So we're only going to activate that area.

And we have eight such areas in our model and we only activate, you know, up to two of them at any one time.

in response to any one query.

But that's what they were talking about.

But now add conversation across those in reasoning models, which has just been observed.

And this team at the University of Chicago is focused on trying to understand what's happening to get to reasoning.

And now they're saying, okay, it's not just that there are different areas of representation inside the model, but they're talking to each other.

They're talking to each other without it being directed to do so.

There's a term that's used for this issue, which is called capability overhang.

Like AI's capabilities are way out ahead of where people are using it.

So that's the capability overhang.

And how do we take advantage of that?

Well, you really have to learn what the capabilities of AI systems are.

And, you know, I would argue that, Carl, you're really in the vanguard of that.

You're finding application of AI using a very up-to-the-moment understanding of what the range of capabilities are of various tools, not just one.

And you're finding real applications of that at the edge of where it's actually been demonstrated before.

Yeah.

Shall we wrap it?