Debunking the impossibility of utility maximization, exploring tree search and AIXI implementation

jd_pressman · x · 2026-08-23

Rebutting the claim that 'no physical object will ever maximize utility over an indefinite time horizon,' the author argues it is a category error, noting that physical laws do not seek goals. The discussion suggests that autonomous agents can be implemented using tree search algorithms, citing examples like MC-AIXI, combining LLMs with AlphaZero, or extracting reward programs from LLMs to realize VNM utility theory. The referenced paper, A Monte Carlo AIXI Approximation, introduces a principled approach for designing scalable general reinforcement learning agents based on a Monte Carlo Tree Search algorithm.

Original post →

More from AGI Musings

AGI Musings channel →