A New Direction for AI
Richard Sutton, the 2024 ACM Turing Award winner whose work laid the foundation for modern reinforcement learning, has launched a new AI research lab called Oak Lab in Toronto. The venture marks a decisive break from the dominant deep learning paradigm that Sutton himself helped enable.
Co-founded with Khurram Javed, Oak Lab emerges from the same intellectual lineage as John Carmack's Keen Technologies, where both Sutton and Javed previously worked. But Sutton is blunt about why a new lab is necessary.
'Weak and Inefficient'
In announcing the lab, Sutton called current deep learning methods 'weak and inefficient,' arguing they 'need not more tweaks, but fundamentally new ideas and a thorough reworking before they can provide a solid foundation for achieving the more ambitious goals of AI.'
The critique reflects a conviction Sutton has long held: that today's AI systems are glorified pattern matchers, not true learners. In June, he argued that generative AI excels at imitation but fundamentally cannot evaluate its own outputs, making it incapable of real discovery.
Learning From Experience
Oak Lab's core thesis is that AI should learn from experience during operation, not train once on static datasets. The lab is building agents that: - Learn continuously from their environment rather than from frozen training data - Construct internal world models that evolve as they interact with the world - Handle variation, evaluation, and selection autonomously
The long-term goal is audacious: an AI agent with 'a trillion parameters that learns and plans in real time with 20 watts of energy' — a vision that deliberately echoes the efficiency of biological brains.
A Field at a Crossroads
Sutton's move comes at a moment of ferment in AI research. While deep learning has driven remarkable progress, its limitations — massive energy consumption, inability to generalize beyond training data, and lack of true understanding — are increasingly acknowledged even by its pioneers.
Oak Lab joins a growing ecosystem of labs and startups betting that reinforcement learning, rather than ever-larger language models, holds the key to more capable and general artificial intelligence. Whether Sutton can translate his theoretical legacy into a practical breakthrough remains to be seen, but when the father of reinforcement learning starts a new lab, the field pays attention.
