Sutton: Large Models Lack Native Intelligence as AI Moves into the Experience Era

On July 17 at the 2026 World Artificial Intelligence Conference (WAIC), 2024 Turing Award winner and reinforcement‑learning pioneer Richard Sutton delivered a keynote speech.

Sutton observes that the AI industry is caught up in a frenzy of compute growth and large‑model parameter races, but he stresses that sheer compute scale does not equate to native intelligence; merely fitting human‑generated data cannot bridge the gap to autonomous discovery.

He warns that the era of static, labeled data has reached its limits and calls for a transition to AI that learns continuously from experience, proposing a unified “general mind science.”

Drawing on William James and Alan Turing, Sutton defines intelligence as the ability to achieve goals through adaptive behavior, emphasizing that intelligence is fundamentally goal‑directed computation.

He points out that without explicit goals, reward signals, and interaction with a dynamic environment, large language models cannot verify factual accuracy or discern truth, making them fundamentally limited.

Examples such as AlphaGo’s self‑play, AlphaProof’s success in math competitions, infants learning by manipulating toys, and athletes making split‑second decisions illustrate how learning from interaction and reward drives genuine intelligence.

Sutton argues that reinforcement learning represents the beginning of a comprehensive mind science, and that experience‑driven AI will become more powerful and reliable as agents continuously adapt.

He concludes that current AI remains weak and unreliable, yet the field is moving toward an “experience era” where agents learn from real‑world interaction, promising a future of stronger, continuously improving intelligence.