High Signal Podcasts Evidence ledger
Method
Browse
← Back to evidence

Evidence receipt / belief

Published · transcript-backed

Trenton Bricken: belief

28 Mar 2024 Dwarkesh Podcast Sholto Douglas & Trenton Bricken — How LLMs actually think

“I think in this case, he would also just have a really long context length, or a really long working memory, where he can have all of these bits and continuously query them as he's coming up with some theory so that the theory is moving through the residual stream.”

— Trenton Bricken

Source trail

Everything needed to verify it.

Speaker
Trenton Bricken
Attribution
Verified speaker
Claim type
belief
Recorded
28 Mar 2024
Publisher
Dwarkesh Podcast

Transcript context

…One of the things I talked to Demis about was a paper he had in 2008 that memory and imagination are very linked because of this very thing that you mentioned, that memory is reconstructive. So you are, in some sense, imagining every time you're thinking of a memory because you're only storing a condensed version of it and you have to. This is famously why human memory is terrible and why people in the witness box or whatever would just make shit up. So let me ask a stupid question. So you read Sherlock Holmes and the guy's incredibly sample efficient. He'll see a few observations and he'll 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 crucially, what makes him smart is that there's not just an association, but there's a sort of deductive connection between different pieces of information. Would you just explain it as, that's just higher level association? I think so. I think learning these higher-level associations to be able to then map patterns to each other, as a kind of meta-learning. I think in this case, he would also just have a really long context length, or a really long working memory, where he can have all of these bits and continuously query them as he's coming up with some theory so that the theory is moving through the residual stream. And then his attention heads are querying his context. But then, how he's projecting his query and keys in the space, and how his MLPs are then retrieving longer-term facts or modifying that information, is allowing him to in later layers do even more sophisticated queries and slowly be able to reason through and come to a meaningful conclusion. That feels right to me. You're looking back in the past. You're selectively reading in certain pieces of information, comparing them, and maybe that informs your next step of what piece of information you now need to pull in. Then you build this representation, which progressively looks closer and closer to the suspect in your case. That doesn't feel at all outlandish.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.

Search evidence