Post

Not gonna QT, because I don't want to add to the pile on this person is currently experiencing (even if deserved to some extent). But that argument drives me crazy. Just because *you* don't understand how LLMs work doesn't mean nobody understands how they work. This is a fiction.

I see people saying "we don't understand how it works" and that it's making its own decisions when really they mean it's non-deterministic. Idk about this person in particular. But there's a lot of magical thinking about non-determinism.

Fwiw LLMs are deterministic algorithms (like all software), they just usually use a (deterministic) pseudorandom generator algorithm to not use just the single highest statistical probability of token continuation but pick one among the set of highly ranked ones.

0 reposts3 likes

I'm talking about the mental model of how the software works, not the details of its technical underpinnings. Yes, these randomizers are technically deterministic, but the point is that when I listen to people who say these things explain their thinking, it's about non-determinism.