Overview
We design scalable priors for AI. Foundation models learn mostly from data, but data alone misses much of what we already know about the world. We build that knowledge into every step of an intelligent system: how it perceives the world, how it reasons about what it perceives, and how it acts.
Perception
Vision & Multimodal
Perception should see the world behind the pixels: its objects, its structure, and the regularities that data only hints at.
Visual Structure
Structural priors for seeing how the world is put together.
Learning in the Wild
Statistical priors for learning from an imperfect, open-ended world.
Reliable AI
Social priors for AI that people can trust, in what it sees and what it makes.
Reasoning
Language & Reasoning
Reasoning should rest on more than text: on evidence, abstraction, and relations.
Reasoning & Agents
Cognitive priors for minds that know what they know.
Efficient LLMs
Algorithmic priors for doing more with less.
Structured Data
Symbolic priors for knowledge that lives in rules and relations.
Action
Robotics & Embodied AI
Action should start from people and the physical world: what each person wants, and what a body can sense.
VLA & World Models
Behavioral priors for robots that understand the people they serve.
Multisensory Learning
Sensory priors for robots that touch, hear, and feel the world.
In 2026, we apply these priors to agents, robots, and factories. Full publication list →