Sutton launches Oak Lab to build low-energy AI agents
Turing Award winner Richard Sutton has left Keen Technologies to found Oak Lab, betting that the future of artificial intelligence lies in continuous, low-power learning rather than scaling massive data centres.
Richard Sutton, who shared the 2024 Turing Award for founding modern reinforcement learning, announced on Monday that he is leaving John Carmack’s startup Keen Technologies. He and collaborator Khurram Javed have founded Oak Lab to pursue what Sutton calls a “slightly different path toward understanding intelligence.”
Sutton’s departure is rooted in a blunt diagnosis of the current AI industry. He argues that modern deep-learning methods are “weak and inefficient, and need not more tweaks, but fundamentally new ideas and a thorough reworking.” For an industry spending billions on power-hungry infrastructure, this represents a fundamental challenge to the dominant scaling paradigm.
Oak Lab’s research focuses on how intelligence is actually generated. Sutton contends that true intelligence comes from run-time experience, not human-curated datasets. In their first research post, Sutton and Javed demonstrated that standard optimisers like SGD and Adam fail to separate predictable signals from random noise, blindly spreading blame across all parameters.
To fix this, Oak Lab is reviving an older algorithm of Sutton’s called IDBD, introducing a neural version named NetworkIDBD. This method learns to assign credit selectively, rewarding only signals that make accurate predictions. In testing, it successfully identified real patterns where standard optimisers drowned in irrelevant data.
The ultimate goal is radical efficiency. By learning from a continuous stream of experience without storing or replaying data, Oak Lab claims its approach requires orders of magnitude less compute and energy. The startup’s stated objective is a trillion-parameter agent that learns and plans in real time on 20 watts—roughly the power consumption of a human brain.
This target contrasts sharply with today’s frontier models, which train once in facilities drawing megawatts of power before remaining frozen. If successful, such technology would drastically reduce the capital and energy barriers to competing in AI. That is a prospect with profound implications for European companies constrained by high power costs and strict environmental regulations.
Sutton is not alone in this contrarian view. Yann LeCun recently left Meta to bet $1bn on world models, and AlphaGo’s David Silver is pursuing a similar route. As the AI race matures, the focus is shifting from simply building the largest models to improving reasoning and cutting costs. Sutton is betting his career on the idea that the future of AI “looks less like a bigger brain in a bigger building, and more like a small one that never stops learning.”
