Richard Sutton: More Compute Alone Won't Cut It. His Pitch: AI That Learns From Experience

Turing Award winner Richard Sutton argues that simply scaling up AI’s compute is unlikely to be enough to reach intelligence. The Decoder reported on November 8 (local time) on a recent post Sutton made on X. Sutton raised similar concerns in an interview published on September 26, 2025. His question comes down to one thing: what should AI be learning from?

Is making models bigger enough?

Sutton points out that today’s large language models (LLMs) learn by leaning on the text and behavior that humans have left behind. He argues that an approach in which AI acts directly in a real environment, observes the results, and keeps learning from them has far more room to grow. In other words, the answer is not just doing more computation, but changing the method of learning itself.

The argument connects back to “The Bitter Lesson,” which Sutton wrote in 2019. In that essay, he made the case that what matters are general-purpose search and learning methods that can absorb ever-increasing amounts of compute. In the earlier interview, he noted that even compute-hungry LLMs ultimately depend on human-generated data. Whether LLMs are the full realization of that principle, he suggests, remains an open question.

AI that learns from the consequences of its actions

At the center of Sutton’s alternative are agents: AI systems that handle tasks on their own. Such a system sets a goal, acts, and then revises its predictions and its next move based on the outcomes and rewards it observes. Rather than learning only from human text and human behavior, it folds the results of its own actions into its learning.

How to learn while pursuing a long-horizon goal is another problem. Sutton described “temporal difference learning” as a way to extract a learning signal at intermediate steps. Instead of waiting for the final outcome, the system corrects its predictions on the fly using information gathered along the way. The change he emphasizes is not in model size but in how experience is accumulated and used.

What about the chatbots we use today?

The interviewer pushed back, suggesting that LLMs could serve as a starting point: take a language model that has already learned from text and add experience-based learning on top. The debate, then, is not only about whether to keep scaling LLMs. It is also about how to combine capabilities acquired from human-generated data with learning that comes from acting in the world.

For users, the crux is whether AI can keep learning by feeding the results of one task into its next action. Sutton stressed that learning from experience can go further, while the interviewer offered a path that combines it with LLMs. What to watch next is how the two approaches get implemented in actual systems, and what kind of performance continual learning ends up delivering.