General Plasticity
An independent AI research lab
We’re working on the breakthroughs that will let language models generalize. Our starting point: intelligence must keep learning.
[ 01 / Our thesis ]
Learning doesn’t end at deployment.
The world keeps changing. We believe models should change with it — learning from new experience as it happens, rather than relying only on what they learned during pretraining.
Our core belief is that generalization requires continual learning. For language models, that means test-time training: updating the model itself while it encounters and works through new problems.
Getting there may require a training framework beyond gradient descent. We take inspiration from the brain: a system that learns continuously, locally, and through experience.
[ 02 / Research directions ]
New ways to learn.
Continual learning
How can language models acquire new abilities during use, retain what matters, and carry that learning into unfamiliar situations?
Beyond backpropagation
We explore alternatives to backpropagation and gradient descent, looking for learning rules suited to continuous adaptation at test time.
Efficient Hebbian learning
Inspired by how the brain strengthens connections through co-activity, we investigate efficient, local learning methods for test-time training.
[ 03 / Connect ]
Working on the same questions?
We’d love to exchange ideas with people exploring continual learning, biological intelligence, and new learning algorithms.
Start a conversation (opens in a new tab)