Daniel Jeffries (Unknown)

And what else are you excited about in this space?

Yeah, how might that work, for example?

So I mean, if you were to apply this to something like reinforcement learning, I mean, what would that look like?

And this brings us on nicely to TRACK, which is your kind of evolution of this data modeling work.

Can you introduce it?

So you said that this model that you've built now, TRAC, can work for parameterized differentiable models.

And I think you said of a size roughly commensurate to about 300 million parameters.

Is that roughly right?

Yeah, I mean, because you were talking about this trade off between speed and efficiency.

I mean, can you tell us about that?

Yes, and the first one you did in your paper was this LDS score, this linear data modeling score.

How does that work?

So can you explain the derivation?

Very cool.

Very cool.

And you also did some experiments with language models.

And what does that look like?

I mean, how interpretable is it in practice?

Yeah, that would be really useful for concept scrubbing.

Yeah, interesting.