Oak Lab
Our goal is to discover and implement algorithms that allow agents to achieve goals in big worlds.
We are building the OaK architecture
OaK architecture discovers temporal abstractions grounded in experience that are both self-verifiable and useful for planning.
Watch the talk on the OaK ArchitectureOur algorithms learn in real-time without storing or replaying data
Our batch-size-one learning algorithms can learn directly from experience without storing or replaying data. These algorithms, when combined with event-driven neural networks, result in systems that learn using multiple orders of magnitude less compute and energy than existing methods; their computational efficiency makes continual real-time learning possible.
Detailed post coming soonOur focus is on learning directly from experience, not from curated datasets
By learning to assign credit to parameters that generalize well, instead of assigning credit to all parameters, our methods can learn directly from noisy data streams; existing methods, in comparison, learn best when provided with clean human-curated datasets.
Our holy grail
A trillion-parameter agent that learns and plans in real-time with 20 watts of energy.