The Father of Modern RL Is Betting Generative AI Can't Learn. His New Lab Will Try to Prove It.

Richard Sutton has spent four decades building the theoretical foundation for reinforcement learning. His 2024 Turing Award recognized that contribution. Now he’s using it to make a specific, measurable bet against the direction the AI industry is heading.

What Happened

Sutton founded Oak Lab in Toronto in 2026, co-founded with Khurram Javed — both previously at John Carmack’s Keen Technologies. The lab’s focus is AI agents capable of continuous learning from environmental experience, without relying on large static training datasets.

The stated technical goal is specific: an agent with one trillion parameters that learns and plans in real time while consuming 20 watts of energy. That’s a desktop workstation’s idle draw running a model at the scale of today’s largest frontier systems — and it’s learning, not just inferring from a frozen checkpoint.

Why It Matters

Sutton’s critique of current AI is not aesthetic — it’s architectural. He argues that deep learning as currently practiced is “weak and inefficient,” and that generative AI excels at imitation but cannot evaluate its own outputs. The ability to evaluate and adapt based on feedback is precisely what reinforcement learning was designed to provide; it’s also the property generative AI systems conspicuously lack.

The distinction matters for agents specifically. A generative AI agent can generate a plausible sequence of actions based on patterns in its training data. A reinforcement learning agent adapts its behavior based on outcome signals — it can improve in a domain even without pre-existing training data for that domain. For agents operating in novel or dynamic environments (which is most real-world deployment contexts), this matters substantially.

The trillion-parameter / 20-watt target also makes a point about efficiency. Current large models require datacenters to run; the vision Sutton is describing is a highly capable agent operating at the energy cost of a reading lamp. Whether that’s achievable is an open question, but it’s a concrete research target rather than a philosophical position.

What the Sutton Pedigree Means

Sutton’s co-authored work on temporal difference learning and policy gradient methods forms the core of how RL algorithms work today. His book Reinforcement Learning: An Introduction (co-authored with Andrew Barto) is the primary reference text for the field. His previous work at DeepMind and at Keen (building long-context AI systems with Carmack) gives Oak Lab a specific technical lineage.

The Turing Award citation in 2024 was partly for that foundational theory. That Sutton is now betting against static training data as the primary mechanism for AI capability is a significant signal — not because the bet is certain to pay off, but because the person making it understands both sides of it better than almost anyone.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.