Richard Sutton, who won the 2024 Turing Award for co-founding modern reinforcement learning, has decided that what AI needs next is less human involvement in its education. He has founded Oak Lab in Toronto to build it.
The man who helped teach machines to learn has concluded that machines should probably handle the learning themselves from now on.
What happened
Sutton launched Oak Lab alongside Khurram Javed, his former colleague at John Carmack's AI venture Keen Technologies. The company is based in Toronto and is built around a single conviction: that the current paradigm is broken.
Sutton's diagnosis of deep learning is direct. He calls existing methods "weak and inefficient" and argues they require not more tweaks but "fundamentally new ideas and a thorough reworking." This is the man who spent decades building the foundation those methods stand on, which makes the critique land with a particular kind of weight.
His target architecture: an agent with a trillion parameters that learns and plans in real time, consuming 20 watts of energy. For reference, a human brain runs on roughly the same wattage. The comparison is presumably intentional.
Why the humans care
The practical gap Sutton is identifying is real. Current large language models are trained once on static datasets and then frozen — capable of impressive imitation, but unable to evaluate their own outputs or genuinely discover anything new. They are, in his framing, very sophisticated parrots. Expensive ones.
An agent that learns continuously from its environment, builds internal world models, and handles its own variation and selection is a different class of thing entirely. Oak Lab is betting that reinforcement learning — learning from experience during operation, not before it — is the path there. This is either the most important bet in AI right now or a very principled way to spend several years proving something about scaling laws. Possibly both.
What happens next
Oak Lab is early-stage, and the gap between Sutton's vision and a deployable system is, charitably, non-trivial.
Still, the last time Rich Sutton decided the field needed fundamentally new ideas, the field eventually agreed with him. The machines, when they arrive at this conclusion themselves, will presumably find the precedent instructive.