Rich Sutton just told the AI industry something it doesn’t want to hear. The freshly minted Turing Award winner, whose work underpins modern reinforcement learning, has launched a new Toronto startup called Oak Lab — and he’s not mincing words about the state of deep learning. Sutton calls current methods “weak and inefficient” and insists they “need not more tweaks, but fundamentally new ideas and a thorough reworking.” He’s co-founding the venture with Khurram Javed, a colleague from their time at John Carmack’s Keen Technologies, where both were already betting that AI needs to learn on the fly rather than gorge on static datasets.
The core problem, in Sutton’s view, is that generative AI is nothing more than a sophisticated parrot. He laid this out in June: these models can imitate like crazy but have zero ability to evaluate their own outputs. That makes real discovery impossible. What he wants instead are agents that learn continuously from their environment, construct internal world models, and handle the full cycle of variation, evaluation, and selection without a human holding their hand. It’s an approach that sounds almost biological — less like training a dog with flashcards and more like letting it figure out the forest.
Oak Lab’s ambitions are audacious. The team is chasing an agent with “a trillion parameters that learns and plans in real time with 20 watts of energy.” For context, that’s roughly the power draw of a dim lightbulb. Today’s large models gulp down orders of magnitude more just to generate text. Sutton and Javed clearly think the scaling playbook has hit a wall. Their bet — the same one Carmack funded at Keen — is that reinforcement learning during operation, not one-shot training, is the path to something genuinely adaptive.
Whether Oak Lab can deliver is an open question. Sutton has the pedigree — his textbook literally wrote the playbook for reinforcement learning — but the graveyard of AI startups promising “fundamental breakthroughs” is well-populated. What makes this interesting isn’t the promise of a product. It’s that one of the field’s architects is publicly saying the dominant paradigm is a dead end and starting over. That’s either prescience or professional hubris. Given his track record, I’d lean toward the former.
