69-Year-Old Richard Sutton, Father of Reinforcement Learning, Founds Oak Lab to Build a 20-Watt Human-Level Agent

In July 2025, 2024 Turing Award winner and reinforcement learning pioneer Richard Sutton announced the co-founding of Oak Lab with his PhD student Khurram Javed. The company aims to develop AI agents that can learn and plan in real time from their own experience, with the ultimate goal of building a trillion-parameter general-purpose agent consuming only 20 watts (on par with the human brain). Sutton previously joined John Carmack's Keen Technologies in 2023 but left after less than three years due to disagreements over technical direction, though he publicly praised Carmack.

Background: From Academic Titan to Entrepreneur

Richard Sutton is a founding figure of modern reinforcement learning: he studied psychology as an undergraduate at Stanford, earned his PhD under reinforcement learning pioneer Andrew Barto, proposed the temporal difference algorithm, co-authored the globally used textbook Reinforcement Learning: An Introduction, served as a professor at the University of Alberta since 2003 and founded the RLAI lab, and worked as a Distinguished Research Scientist at DeepMind from 2017 to 2023, where he led the Edmonton team. He has trained numerous top AI talents, including AlphaGo core designer David Silver and DeepMind Montreal head Doina Precup.

Core Philosophy: Experience-Driven Real-Time Learning

Sutton believes the current deep learning paradigm is brittle and inefficient, relying on static human-curated data, with models unable to continuously update their capabilities during operation. Oak Lab's core architecture, OaK (Options and Knowledge), aims to enable agents to discover temporal abstractions from their own experience and convert them into callable skills. Unlike traditional deep reinforcement learning, Oak Lab uses real-time learning with a batch size of 1, updating immediately upon receiving each new experience without storing historical data or replay buffers. The team believes that combining this with event-driven neural networks can reduce computation and energy consumption by orders of magnitude.

Technical Approach: OaK Architecture and the Big World Hypothesis

The OaK architecture's core is to enable agents to decompose high-level skills from experience (e.g., "walk to the kitchen," "pick up a cup"), and later directly invoke and flexibly adjust these skills when encountering similar goals. Sutton and Javed jointly proposed the "Big World Hypothesis": the real world is always more complex than AI, so models must selectively remember useful content, forget outdated information in a timely manner, and engage in continuous online learning. This aligns with Sutton's 2019 essay "The Bitter Lesson"—that general learning and search methods will ultimately surpass systems relying on human knowledge. In 2025, Sutton and David Silver jointly proposed the "Age of Experience," arguing that AI will shift from relying on human data to relying on experience generated by agents interacting with their environment.

Reactions and Impact

Sutton's entrepreneurial move has drawn industry attention. When he joined Keen Technologies, he and Carmack planned to build a prototype system with "signs of AGI life" by 2030, but eventually parted ways due to route differences. Sutton emphasized in a tweet, "I can't say enough good things about John Carmack and Keen Technologies." Oak Lab's long-term goal (trillion parameters, 20 watts) remains a vision for now, but it represents a fundamental challenge from the reinforcement learning school to the current Scaling Law approach of large models. Sutton's first stop after founding the company is WAIC in Shanghai, where he will deliver a talk titled "The First Principles of Reinforcement Learning: Cultivating Superintelligence from Experience."