Richard Sutton, recognized for founding modern reinforcement learning and co-recipient of the 2024 Turing Award, is departing Keen Technologies to establish Oak Lab, a startup focused on fundamentally rethinking how AI systems learn and reason efficiently from experience.
- Sutton leaves Keen Technologies to launch Oak Lab focused on new AI learning methods
- Oak Lab targets AI that learns from real-time experience, not fixed datasets
- Goal: a trillion-parameter agent running in real time on just 20 watts
What happened
Richard Sutton, a leading figure in artificial intelligence and winner of the 2024 Turing Award for his role in founding modern reinforcement learning, announced his departure from John Carmack's AI startup, Keen Technologies. Sutton revealed that he and collaborator Khurram Javed have launched their own new startup, Oak Lab, aiming to pursue a different approach to understanding and building intelligence.
Sutton publicly praised Carmack and Keen, but stated that their new venture intends to follow a path distinct from mainstream deep learning trends. Oak Lab’s initial research highlights fundamental inefficiencies in current AI techniques and proposes new algorithms designed to enable AI systems to learn from ongoing experience with greater precision and efficiency.
Why it matters
Sutton’s critique of current AI methods centers on their dependence on large, curated datasets and heavy computation, which limits adaptability and energy efficiency. Existing optimization algorithms like SGD treat all errors equally, leading to noisy learning. Oak Lab’s innovation, NetworkIDBD, selectively credits only informative signals, enabling better learning from unpredictable, real-world data streams.
The lab’s ambition is to create an AI agent capable of continuously learning and planning in real time while consuming as little as 20 watts of power—mirroring human brain efficiency far better than today’s megawatt-scale training setups. This focus on real-time continual learning and compute efficiency challenges prevailing beliefs in AI that more data and larger models alone will achieve true intelligence.
What to watch next
The AI research community will be closely monitoring Oak Lab’s progress as it tackles one of the field’s most pressing challenges: how to build AI systems that learn persistently and adaptively without massive energy demands. Success could redefine AI research priorities toward sustainable, real-time learning agents rather than ever-larger pretrained models.
Key developments include Oak Lab’s forthcoming research outputs and demonstrations that validate their claim of effective learning with drastically reduced compute. Sutton’s move also aligns with broader industry shifts, seen in similar bets by high-profile researchers like Yann LeCun and David Silver, who advocate for experience-driven AI over scale-driven approaches.