2024 Turing Award winner and reinforcement learning pioneer Richard Sutton has left John Carmack’s AGI startup Keen Technologies, together with longtime collaborator Khurram Javed, to found a new lab called Oak Lab. Javed, a former student of Sutton’s, earned his computer science PhD at the University of Alberta and has spent years working on real-time reinforcement learning and continual learning.
https://twitter.com/RichardSSutton/status/2076663628301058329
Keen was already focused on reinforcement learning and long-running agents, but Oak Lab is spinning out the continuous learning approach that Sutton and Javed have long championed. Most current AI models are frozen after deployment — they don’t update based on new experiences. Oak wants its agents to keep learning from their own experiences while they run.
Sutton has called current deep learning methods “weak and inefficient.” He believes it’s not enough to patch things up — the field needs to redo the fundamentals. Oak Lab will push forward the OaK (Options and Knowledge) architecture, where agents form knowledge from interaction, use a world model to predict the consequences of their actions, and plan what to do next.
The company’s long-term goal: build a trillion-parameter agent that consumes just 20 watts for real-time learning and planning.