Reinforcement Learning Authority Sutton Launches New Startup Oak Lab

Richard Sutton, a researcher who laid the theoretical foundations for reinforcement learning—a technique where AI learns through trial and error—has established a new startup called "Oak Lab" in Toronto, Canada. Sutton won the 2024 Turing Award, considered the highest honor in computer science, and is known as one of the most influential figures in AI research.

Oak Lab will focus on developing AI agents that learn autonomously and continuously from their environment. Sutton characterizes current mainstream deep learning methods as "weak and inefficient," positioning himself against the view that current AI technology is fundamentally limited. Such critical remarks from someone at the forefront of AI research merit attention.

The current mainstream in AI development relies on large language models (LLMs) trained on vast amounts of data beforehand. Tools like ChatGPT are built using this approach and have a "static" nature—once training is complete, they typically do not undergo additional learning. The "AI that continuously learns from its environment" that Sutton envisions represents an approach contrasting with this static learning model.

Reinforcement learning originally gained attention as a technique through which AI demonstrated superhuman capabilities in games like Go and Shogi. In this approach, an agent takes actions, receives rewards based on the results, and learns optimal behaviors through self-improvement. Sutton has long constructed the theoretical foundation for this methodology. Oak Lab is positioned as an attempt to apply reinforcement learning's conceptual framework to broader AI agent development.

Details regarding the startup's scale, funding situation, and specific products or services remain undisclosed at this time. Nevertheless, the fact that a prominent AI researcher has declared that "current deep learning is insufficient" and subsequently established an independent organization can be regarded as a challenge to the direction of current research.

The AI agent field is currently an area where many companies and research institutions are investing effort. However, realizing "AI that continuously learns autonomously" is technically challenging, and careful discussion is required from the perspectives of safety and control. What specific technologies and methodologies Oak Lab will ultimately adopt remains to be seen in future announcements.

A Turing Prize laureate explicitly stating that "current AI is weak" and launching a new company to explore an alternative path represents a significant contribution to industry-wide technical discourse in AI. How the research community and industry will evaluate this alternative approach centered on "self-learning AI agents" in contrast to the current large-model-focused trend will likely depend critically on Oak Lab's forthcoming technical announcements.