Can a model learn directly from the world as it is?
Learning from dirty, incomplete human data without an expensive cleaning pipeline.
Introducing TasteMaxxer 1
Building models that
develop
human-like taste.
TasteMaxxer 1 is a large language model that learns from messy human data—and develops a taste of its own.
Notify meFor researchers, builders, and curious minds
The work ahead
We are solving three of the most important questions in AI and robotics.
Can it keep learning without losing what it knows?
Continual learning through weight updates, without catastrophic forgetting.
Can reinforcement learning be radically more efficient?
Learning through reinforcement without millions of environment rollouts.
Hold the string.
Join the early list for notes from the lab and news on TasteMaxxer 1.