AI researcher Richard Sutton, winner of the Turing Award in 2024 and one of the “founders of reinforcement learning,” has founded his own startup. The name alludes to Sutton’s “OaK” architecture, a blueprint for continuously learning agents.
Read more after the ad
Canada’s Oak Lab wants to build AI agents that continue to learn from experience on the fly, rather than being frozen after training. Sutton himself announced the move to X, the co-founder is his former student Khurram Javed. Both come from Keen Technologies, the AI company owned by game developer John Carmack, which they say they left for this.
Learning from the data stream without saving
Technically, Oak Lab consciously differentiates itself from today’s language models. Instead of learning from a curated data set and then remaining static – or, like systems now praised as “self-learning AI”, constantly receiving new knowledge databases or being retrained every now and then – the agents should learn in real time, according to the company. And all without saving or replaying data. With this idea, Sutton separates himself from the mainstream; Providers such as OpenAI, Google and Meta are currently training their models with more and more data and in ever larger data centers. Oaks AI is designed to require less computing power and energy than current methods, making continuous learning possible during ongoing operations.
An idea, not a product
Oak Lab is currently basic research: there is no product, no funding mentioned and no timeline. Whether reinforcement learning leads to the goal will only become clear in the coming years.
Oak Lab cites a long-term goal as an agent with a trillion parameters that learns and plans in real time while using only 20 watts – about as much as the human brain. However, this should be understood as a direction, as its “holy grail”, not as a product promise, the startup said.
Read more after the ad
(rie)
