The inventor of the bitter lesson has a bitter lesson for LLM scaling enthusiasts

Why it matters

Turing Award winner Richard Sutton, author of the influential essay known as the bitter lesson, has published a post arguing that raw compute alone cannot produce intelligence and that AI science is far from finished. His famous essay once held that general methods leveraging compute eventually outperform human-engineered approaches, but he has long warned against relying solely on scaling large language models. He calls synthetic data a big mistake, arguing the real world is far too complex to simulate, and says human expertise becomes a bottleneck. Instead, he advocates agents with world models that keep learning continuously from their own experience. His remarks prompted wide debate. Researcher Jerry Tworek argues the bitter lesson still holds but that algorithms matter increasingly as training compute grows, and says transformer scaling's temporary success, combined with a misreading of the essay, has distorted a generation of scientists. Researcher Jayden Teoh notes the essay has been used to dismiss promising academic work prematurely. Meanwhile, OpenAI's Sam Altman and Anthropic's Dario Amodei, both with obvious stakes, insist scaling remains productive.