2026-09-16 08:49 UTC
And was a Bayesian calculating agent that enumerates every possible explanation for the information it takes in and ranks them according to evidence, along with every trajectory through the world that could fulfill its utility function, rather than approximating the high dimensional shape of text
Replies (1)
-
@scruiser@awful.systems 2026-09-18 02:40
It annoys me how much credit for predicting AI Eliezer gets away with claiming among his cult (and even outside his cult, like in interviews with reporters that could know better if they bothered to do their job) when all the actual details he predicted are nothing like LLM-based agents. Likewise for claiming success on predicting instrumental convergence. LLM agents often can’t stay focused on the goal they are post trained and directly prompted to pursue, they are the furthest thing from trying to optimally maximize a bunch of secondary values in pursuit of master goal.