Reinforcement learning pioneer sets out a new course
Richard Sutton, a leading figure in reinforcement learning and winner of the 2024 Turing Award, has left John Carmack’s Keen Technologies to launch a new venture, Oak Lab, alongside his former student Khurram Javed. Sutton, who developed foundational methods such as temporal-difference learning and co-authored the textbook Reinforcement Learning: An Introduction, frames the move as a bid to rethink how artificial intelligence learns and reasons.
“At 69, it’s the perfect age to create!”
In public messages outlining the shift, Sutton argues that today’s dominant deep learning strategies are too weak and inefficient to reach higher general intelligence. Rather than more incremental tweaks, he says the field requires fresh fundamentals and a reworked architecture for learning and planning.
What Oak Lab is trying to build
The new company’s objective, as set out by Sutton, is to develop an agent of roughly trillion-parameter scale capable of real-time learning and planning while consuming only 20 watts of power — a target chosen to mirror the energy use of the human brain. That contrasts starkly with the sector’s current direction, where progress often hinges on larger data centres and more compute.
“Build an agent with a trillion-parameter scale that can learn and plan in real time, with a total power consumption of only 20 watts.”
Sutton’s stance challenges an industry trend that has favoured stacking more processors and expanding infrastructure. His previous roles — including founding the Reinforcement Learning and Artificial Intelligence Laboratory (RLAI) at the University of Alberta and serving as a Distinguished Research Scientist at DeepMind, where he helped establish the Edmonton team — lend weight to the claim that a different pathway may be necessary.
Why this matters for Sutton and London
For residents and firms in the London Borough of Sutton, the proposal matters because it speaks directly to energy use, cost and accessibility in AI. If research like Oak Lab’s can push intelligence systems to operate on far less power, the implications could include:
- More practical AI deployments on everyday hardware rather than power-hungry servers.
- Lower operating costs for small businesses and public services looking to adopt automation and decision support.
- Reduced environmental impact compared with compute-intensive models that rely on expanding data centres.
The borough’s mix of small enterprises, service providers and commuters into London’s tech clusters means changes in how AI is built can filter through quickly — in pricing, availability and suitability for on-premise or edge devices. Sutton’s argument is not about marginal gains but a wholesale reconsideration of how machines acquire and use knowledge.
Background and people behind the pivot
Over four decades, Richard Sutton has helped train many notable researchers, including David Silver (AlphaGo’s lead designer), Doina Precup (DeepMind Montreal) and Michael Bowling. He and co-founder Khurram Javed previously worked within established labs before deciding to set up Oak Lab to pursue a different approach from mainstream deep learning.
The decision to depart Keen Technologies, created by game-engine veteran John Carmack, stems from Sutton’s view that the field needs new foundations rather than scaling the same recipes. His academic grounding in psychology at Stanford, doctoral work with Andrew Barto and long-standing focus on learning from interaction have consistently emphasised systems that adapt continuously, rather than those trained in large, static batches.
How the shift contrasts with the status quo
Most leading AI efforts seek better results by training even larger models on more data with greater compute budgets. Oak Lab’s target of high capability at 20W highlights a different optimisation: intelligence per watt. For local institutions working to keep technology adoption affordable and environmentally responsible, that framing is increasingly relevant.
| Entity | Role/Note |
|---|---|
| Richard Sutton | Reinforcement learning pioneer; founder of Oak Lab; 2024 Turing Award recipient |
| Khurram Javed | Co-founder of Oak Lab; former student of Sutton |
| Keen Technologies | Company founded by John Carmack; Sutton and Javed departed to form Oak Lab |
| DeepMind | Sutton served as Distinguished Research Scientist; helped establish Edmonton team |
What to watch next
Details such as Oak Lab’s development timeline, research outputs and partnerships have not been set out in the source material. For now, Sutton’s declaration signals a push to prove that general-purpose learning and planning can be achieved within a tight energy envelope — a goal that, if realised, could reshape procurement decisions and digital strategies across London boroughs, including Sutton.
As discussion builds around the future of AI infrastructure and its power footprint, residents and local organisations can expect more debate over performance versus energy use. Oak Lab’s progress will be an early test of whether a leaner path to machine intelligence can compete with the prevailing compute-intensive model.
