True Intrinsics
Intelligence through experience.
True Intrinsics builds agentic foundation models that learn abstractions from interaction, keep learning after deployment, and connect what they learn to human language as an interface for task specification, goal inference, knowledge, culture and norms alignment.
Why now
Pretraining on human-generated data is one road to intelligence. There is another: agents—and the foundation models that power them—that form their own abstractions and internal models through interaction with the world.
Pretrained models make this path newly practical by providing a computational interface to human language, knowledge, culture, and goals, while an experience-learning core learns directly from what happens in the environment.
What we build
We are building that agent and the foundation models that power it. Motivated by predictive coding, our models continually update their internal predictions through interaction and use prediction error to learn from new experience. There is no fixed boundary between training and deployment: the agent remains in its environment, continues updating its parameters, and forms abstractions from the experience it accumulates there. The language interface can remain stable while the experience-learning core continues to change with interaction.