Richard Sutton, father of the 'bitter lesson,' says compute alone won't get AI there

The man whose essay became the scaling crowd's favorite citation is telling them they misread it. Turing Award winner Richard Sutton posted that raw compute isn't enough for intelligence, and that the field still needs clever new methods because it is "far from finished," Matthias Bastian reports for The Decoder. The post set off a big argument on X.

What the bitter lesson actually said

Sutton's 2019 essay argued that general methods that scale with compute eventually beat clever human-engineered approaches. The point was ideas that grow with compute instead of leaning on human assumptions, not "just make the transformer bigger."

He's been consistent. He has called synthetic data a big mistake, since any simulation of an infinitely complex world is tiny by comparison, and says human expertise is a bottleneck. What he wants are agents with world models that keep learning from their own experience, without wiping out what they already know. Today's language models stop learning after training, and internet text captures a sliver of the world.

Others pile on

Jerry Tworek, who left OpenAI and now runs the lab Core Automation, agrees the bitter lesson is still true, but argues that as compute grows, the choice of algorithm counts for more, not less. Some methods scale, some hit walls, and many scaling dimensions are unexplored. In his view, transformer success plus a misread essay has warped a generation of researchers.

Researcher Jayden Teoh says he's watched the bitter lesson used to dismiss promising academic work. Ideas that were later replaced by cleaner, more general methods often enabled progress at the time, he argues, and deserve credit for it.

The Decoder counts a wave of new labs betting on these gaps, including Sutton's own Oak Lab.

The other side

Sam Altman says LLM scaling is far from done and that critics are holding researchers back by being too sure of its limits. His evidence: an OpenAI model just knocked down a math conjecture that had stood for years. Dario Amodei describes scaling progress as a rainbow with no end. Both, as The Decoder notes, run companies whose whole approach rests on that bet.

For builders, the practical read is that "bigger model next year" is not the only roadmap anymore. Continual learning and agents that improve from use are where a lot of new money and talent are heading.

Scoreboard update: Draw. The bitter lesson's author says scaling isn't enough, and the scaling labs keep winning anyway.

Stay undead.