Richard Sutton: raw compute alone isn't enough for intelligence

Sutton's original 2019 essay argued that general methods which scale with compute eventually outperform approaches built on human assumptions. His latest post extends that logic to today's large language models, which he says do not keep learning after training. He has previously called synthetic data a 'big mistake' because the real world is infinitely complex and any simulation is only 'microscopic.' Instead, he calls for agents with world models that continuously learn from their own experience without destroying old knowledge.

Jerry Tworek, who left OpenAI to explore new approaches, puts the debate in context: the bitter lesson still holds, but the more compute goes into training, the more the algorithm matters. He warns that some algorithms scale while others hit dead ends, and that the temporary success of transformer scaling, combined with a misreading of the bitter lesson, 'has reward-hacked the brains of a whole generation of scientists.' Jayden Teoh adds that the bitter lesson has sometimes held back academic research by dismissing promising ideas, and argues that an idea doesn't have to survive scaling forever to be an important contribution.

The disagreement is visible in industry leadership. Sam Altman considers LLM scaling far from exhausted and accuses critics of slowing down researchers, pointing to an OpenAI model that recently disproved a long-standing mathematical conjecture. Dario Amodei similarly calls scaling progress a rainbow with no end. As CEOs of companies whose technical approach rests on scaling, both are obviously biased in its favor.