Richard Sutton – Father of RL thinks LLMs are a dead end

Richard Sutton argues that true intelligence requires Reinforcement Learning systems to actively predict and adapt to a physical world through trial and error, contrasting this with Large Language Models that merely mimic human language patterns without establishing ground truth. He identifies the inevitability of AI succession as a new evolutionary stage, warning that future superintelligences must be engineered with high integrity and pro-social values to survive the risks of external corruption and logical divergence. Ultimately, Sutton contends that the most scalable path forward abandons human-imbued knowledge in favor of self-correcting experience-driven learning across distributed digital agents.