Richard Sutton warns that compute alone will not advance AI

What Sutton sees as the limitation

In “The Bitter Lesson,” Sutton argued that, over time, general methods able to use more computing resources outperform solutions based on rules carefully designed by people. What matters are approaches that improve as compute increases, rather than ones that depend on human assumptions.

In a recent post, Sutton reiterated that compute alone is not enough to advance intelligence, and that AI science is far from finished. His remarks were widely discussed in the AI community on X.

Sutton has previously criticized the idea of simply scaling large language models. He called synthetic data a major mistake: the real world is infinitely complex, and any simulation of it covers only a microscopic fraction. Dependence on human expertise also limits scaling.

Instead, Sutton proposes building AI agents with world models that continuously learn from their own experience. The world is far more complex than its description in internet text, and today’s language models do not keep learning after training ends. In Sutton’s view, continuous learning methods are needed to preserve knowledge acquired earlier.

The debate over scaling

AI researcher Jerry Tworek, like many others, left OpenAI to study new approaches to AI. He believes the “Bitter Lesson” still applies, but the more compute is devoted to training models, the more important the algorithm becomes.

Some algorithms scale; others hit a dead end, and many avenues for scaling have barely been explored. In Tworek’s view, the temporary success of scaling transformers and a misreading of the “Bitter Lesson” have distorted the thinking of an entire generation of scientists.

AI researcher Jaden Theo adds that he has repeatedly seen the “Bitter Lesson” hold back academic research: promising ideas were rejected on the grounds that they supposedly ran counter to its principles. Theo believes the problem is not Sutton himself, but the misinterpretation of his original argument.

He also notes that ideas that later seemed incompatible with the “Bitter Lesson” often gave way to simpler, more general approaches. That does not mean the earlier ideas were wrong: at the time, they may have opened a path to progress. Theo urges researchers not to dismiss such work outright. An idea does not have to maintain a scaling advantage forever to contribute to the development of AI.

Disagreements over new learning methods and scaling transformers are also visible among startups. In the latest edition of Frontier Radar #5, the editorial team tracked 45 new AI labs betting on these areas. Tworek himself leads Core Automation, one of those labs, which is looking for learning algorithms beyond pretraining and reinforcement learning. The founders of these labs, including Sutton at Oak Lab, are aiming for much more capable AI, but doubt that the current recipe alone will get them there.

Scaling advocates disagree

Leaders of major companies take a different position. OpenAI CEO Sam Altman believes large language models still have room to scale. He accuses critics of the approach of being so confident in its limitations that they are holding back an entire generation of researchers.

As evidence, Altman points to an OpenAI model that recently refuted a mathematical conjecture that had long eluded mathematicians. World models are useful, for example, in robotics, he says, but the available evidence supports further scaling of language models.

Anthropic CEO Dario Amodei also believes scaling has no visible limit, comparing its progress to a rainbow. Since OpenAI’s and Anthropic’s technical strategies rely entirely on this bet, Amodei and Altman have an interest in seeing it succeed.

Dataist’s take: the debate is not only about how much compute to devote to training, but also about which algorithms can make use of that growth. The evidence cited by the two sides supports different conclusions so far: the success of OpenAI’s model shows what scaling can do, but it does not answer whether language models can learn from experience after training ends.