Why RL Pioneer Richard Sutton Is Launching a Startup to Escape the Large‑Model Paradigm

Richard Sutton, the 70‑year‑old Turing Award winner and father of reinforcement learning, announced this Monday that he and Khurram Javed have founded a new company called Oak Lab.

Oak Lab’s mission is to build AI agents that can learn continuously from their own first‑person experience, eliminating the need for massive, human‑curated datasets. The ultimate goal is a trillion‑parameter agent that can perform real‑time learning and planning while consuming only about 20 W of power, thereby reaching AGI through pure experiential accumulation.

Sutton argues that today’s large language models (LLMs) hit a scalability ceiling because they depend on massive static pre‑training data, excel at imitation, but cannot evaluate their own outputs or engage in genuine exploration and discovery.

To overcome this, Oak Lab is pursuing the Options‑and‑Knowledge (OaK) architecture. OaK equips agents with an internal “world model” of the environment and uses long‑term options for planning, a capability that current Transformer‑based attention models struggle to generate intrinsically.

At the AGI‑25 conference last August, Sutton presented the design philosophy behind OaK and his vision of achieving superintelligence through experience. He also referenced his book *The Bitter Lesson*, criticizing the scaling‑law focus of modern LLM research and emphasizing that true intelligence must discover new knowledge via trial‑and‑error, similar to how AlphaGo evolved into AlphaZero through self‑play.

Earlier this year, David Silver—Sutton’s former student and the founder of DeepMind’s AlphaGo—announced his own startup with a comparable reinforcement‑learning‑centric approach, underscoring a growing trend among leading RL researchers.

Sutton, a professor at the University of Alberta, has spent decades reducing reliance on fixed training pipelines, from early temporal‑difference learning and the Dyna architecture to recent collaborations with Javed on SwiftTD and Swift‑Sarsa, which aim to stabilize online learning and automatically adjust learning rates amid constantly changing inputs.

Oak Lab therefore represents Sutton’s bold attempt to translate his long‑standing academic principles into a practical venture.