Reinforcement learning pioneer Richard Sutton has founded the startup Oak Lab together with his former doctoral student Khurram Javed. The Canada-based company wants to build AI agents that keep learning from experience during deployment instead of from fixed datasets.
Its long-term goal is a trillion-parameter agent that learns and plans in real time on just twenty watts of power.
Sutton Leaves Carmack's AI Company for His Own Project
Sutton taught as a computer science professor at the University of Alberta while also leading a Google DeepMind research lab in Edmonton, which the company closed in 2023. He then moved to Keen Technologies, the AGI-focused startup founded by game developer John Carmack in Dallas.
Last month, he and Javed decided to take "a slightly different path" and left Keen Technologies, Sutton told the Canadian business outlet The Logic.
Javed was one of Sutton's fourteen doctoral students at the University of Alberta and researched real-time reinforcement learning in complex, open-ended environments himself. Oak Lab is incorporated in Canada, and the company has not disclosed any investors or funding round so far.
Sutton has been regarded as one of the founders of reinforcement learning since the 1980s. In 2019, he argued in his widely cited essay “The Bitter Lesson" that compute-intensive, general learning methods eventually outperform solutions designed by hand.
For this body of work, he received the Turing Award in 2024, computing's most prestigious prize.
Oak Lab Lets Agents Learn Directly From Experience
According to its own mission, Oak Lab wants to develop algorithms that let agents achieve goals in large, messy, shifting environments the company calls "big worlds." At its core is the so-called OaK architecture, which derives temporal abstractions directly from experience while remaining verifiable and useful for planning.
In what the company calls batch-size-one learning, its systems process individual experiences immediately in real time instead of storing them for later reprocessing in large batches. Combined with event-driven neural networks, Oak Lab says this sharply reduces compute needs compared with conventional training methods.
Rather than relying on datasets curated by humans, the company trains on noisy, real-world data streams and assigns credit specifically to parameters that prove to generalize well. That sets the approach apart from today's language models, which are mostly trained on huge, pre-collected and cleaned text corpora.
As a long-term goal, Oak Lab cites an agent with a trillion parameters that learns and plans in real time. It is meant to consume only about twenty watts of power, far less than today's training clusters for large language models – a vision that remains independently unverified.
What matters now is whether Sutton's approach holds up beyond demonstrations in controlled settings and into more complex, real-world applications. No concrete timeline for first products or published research results from the new lab has been announced yet.
Oak Lab's approach mirrors a broader trend toward agents that need less curated data to perform well – also visible in BAAI's Orca world model, which reached comparable robotics results without any labeled action data during pre-training.
