Breaking

Turing Award Winner Richard Sutton Launches Oak Lab to Build Self-Learning AI Agents

July 14, 2026 at 04:24 EDT

AI Agents
Research & Papers
Foundation Models

Richard Sutton, the reinforcement-learning pioneer who shared the 2024 Turing Award, has co-founded a Toronto-based startup called Oak Lab that aims to build AI agents which learn continuously from experience rather than from static, curated datasets. The lab, founded with researcher Khurram Javed after both departed John Carmack's Keen Technologies, has set an ambitious long-term goal: a trillion-parameter agent that learns and plans in real time while consuming only about 20 watts of energy — roughly the power budget of a human brain.

Oak Lab · Toronto

A Turing Laureate Bets Against the Frozen Model

Reinforcement-learning pioneer Richard Sutton has co-founded Oak Lab to build AI agents that learn continuously from experience — not from static, curated datasets. The moonshot: a trillion-parameter agent that plans in real time on the power budget of a human brain.

parameter agent, the long-term target

~20W

energy budget — roughly a human brain

batch size — learns from single samples, no replay

The efficiency claim

"Multiple orders of magnitude" less compute & energy

Batch-size-one learning + event-driven neural networks vs. today's methods. Even a conservative 100× gap looks like this:

1×

OaK target

100×+

current deep learning

The OaK loop — Options & Knowledge

Learn during operation, not once and freeze

Act in a big world

pursue goals through ongoing interaction

→

Discover skills

self-verifiable temporal abstractions

→

Reuse & improve

update from experience, keep learning

Enthusiasm

A credible push toward continual, self-directed learning — an alternative to ever-larger frozen models, backed by one of RL's most influential figures.

Skepticism

No product, model, benchmark or price yet — and no large end-to-end demonstration. Doubts remain over whether the approach holds up when scaled.

Near-term goal: a prototype of the complete OaK architecture within a few years — likened to a baby learning in its first year of life.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.