Sholto Douglas

Okay, so let me ask a stupid question.

So you like read Sherlock Holmes, right?

And like the guy is incredibly sample efficient.

He'll like see a few observations and he'll like basically figure out who committed the crime because there's a series of deductive steps that leads from somebody's tattoo and what's on the wall to the implications of that.

How does that fit into this picture?

Because like crucially, what makes them smart is that there's not like an association, but there's a sort of deductive connection between different pieces of information.

Would you just explain it as that's just like higher level association?

Like, yeah.

Yeah.

I wonder if you'd get anything at all.

Um, Sherlock Holmes is probably already in the trading data.

You're going to get like a mystery novel that was written in the,

Oh, we can?

How do you?

There's two different threads I want to follow up on.

Let's go to the long context one and then we'll come back to this.

So in the Gemini 1.5 paper, the eval that was used was something with Paul Graham essays.

Yeah, the needle in a haystack.

Which, yeah, I mean, we don't necessarily just care about its ability to recall one specific fact from the context.

I'll step back and ask the question,