Richard Sutton, 2024 Turing Award winner, founded Oak Lab in Toronto with Khurram Javed. The startup aims to build AI agents that learn continuously from their environment, rejecting static dataset training.
Key facts
- 2024 Turing Award winner co-founded Oak Lab.
- Sutton calls deep learning 'weak and inefficient'.
- Goal: trillion-parameter agent at 20 watts.
- Both founders previously at Keen Technologies.
- Oak Lab located in Toronto.
Richard Sutton, co-founder of modern reinforcement learning and 2024 Turing Award winner, has launched Oak Lab in Toronto with Khurram Javed. Both previously worked at John Carmack's AI company Keen Technologies. Sutton calls current deep learning methods "weak and inefficient" and says 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" According to The Decoder.
What Oak Lab is building
Oak Lab bets on reinforcement learning and the conviction that AI should learn from experience during operation rather than train once on static datasets. In June, Sutton argued that generative AI is good at imitation but can't evaluate its own outputs, making it incapable of real discovery. He wants to build AI agents that construct internal world models and handle variation, evaluation, and selection on their own. The long-term goal is an agent with "a trillion parameters that learns and plans in real time with 20 watts of energy."
This contrasts sharply with the prevailing scaling paradigm, where companies like Anthropic and Google train ever-larger models on fixed corpora. ByteDance recently found AI agents double learning speed every 3 months [per our coverage], but those agents still rely on pre-trained LLMs. Sutton's approach would eliminate the pre-training phase entirely.
Why this matters
Sutton's critique carries weight: he co-founded the field that powers AlphaGo and modern game-playing AI. If Oak Lab succeeds in building agents that learn in real-time with 20W power, it would undercut the economics of data-center-scale training. The AISI recently estimated that fixed compute budgets underestimate AI agents by 60% [per our coverage], suggesting the industry may already be undervaluing continuous learning approaches.
Key Takeaways
- Rich Sutton founded Oak Lab to build self-learning AI agents.
- He rejects static datasets for real-time reinforcement learning with a trillion-parameter goal at 20W.
What to watch
Watch for Oak Lab's first technical publication or model release, likely within 12 months. If Sutton demonstrates a real-time learning agent that matches or exceeds GPT-4 class performance on a continuous learning benchmark, it would force a re-evaluation of the static dataset paradigm.
Source: the-decoder.com
