Post

So imagine for a second that there is something it is like to be an LLM, generating tokens. Presumably, there is probably some perceptible change when you change the temperature? I think we've seen examples where the model definitely notices when you do something super strange to its sampling, too.

15/10/26, 2:45 pm1 repost6 likes

N. R. M. Roshak (Natalka)

@nroshak.bsky.social

You're starting from a counterfactual here. There is nothing that it is like to be an LLM, scanning context and generating the most probable next tokens. There is no subjectivity there.

jake

@yetanotheruseless.com

If you disagree with the starting point, then you have chosen to not participate in the discussion which starts in the other branch of the tree, which is your right, feel free to move on.

N. R. M. Roshak (Natalka)

@nroshak.bsky.social

Sorry. I was attempting to open discussion about your assumption, which isn't harmless. Hazards of posting your assumptions in a public forum, I'm afraid.

jake

@yetanotheruseless.com

Yeah it's fair to want to discuss my assumption, but I didn't make this post to have that discussion right now, so *I'm* not particularly interested in having that with you, especially given the degree of... "lack of epistemic humility" you're demonstrating.

jake

@yetanotheruseless.com

Under the above assumptions, what would it "be like" to just be fed text, continually prefilling, truncating, more prefilling, and never having your logits being computed and sampled from? Take the running traces of a frontier-LLM stateful agent, feed it through another model, what's that like?