Trenton Bricken: belief

Transcript context

…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. I think that the people who aren't doing this research can overlook how after your first layer of the model, every query key and value that you're using for attention comes from the combination of all the previous tokens. So my first layer, I'll query my previous tokens and just extract information from them. But all of a sudden, let's say that I attended to tokens 1, 2, and 4 in equal amounts. Then the vector in my residual stream–assuming that they wrote out the same thing to the value vectors, but, but ignore that for a second–is a third of each of those. So when I'm querying in the future, my query is actually a third of each of those things. But they might be written to different subspaces.…

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