A world model is only as good as what it encodes.
Agents that learn how the world works are bottlenecked by their encoder — how raw, high-dimensional observation gets compressed into a latent state. We build better encoders: representations learned from noise that keep only what matters for predicting the future.
On the name
In 1774, Charles Messier catalogued the faint, formless smudges in the night sky — the noise he kept mistaking for comets — into a fixed list of objects.
A good encoder does the same thing: it takes raw, high-dimensional observation — mostly noise — and compresses it into a compact, structured latent state. We take the name as a statement of method: that understanding begins by cataloguing signal out of noise.
The technical bet
Three moves toward better world-model encoders.
Encode observations, not pixels.
We compress raw, high-dimensional observation into a compact latent state, learned end-to-end from data — no handcrafted features deciding in advance what the model is allowed to see.
Shape the encoder by prediction.
Reconstruction spends capacity on everything visible; prediction spends it on what changes. We train encoders to forecast the future, so the latent state keeps only what is controllable and consequential.
Roll the world forward.
On top of that representation we learn latent dynamics that stay stable over long horizons — a model an agent can imagine inside and plan against, without drifting back into noise.
Why it matters
Where this gets used.
Sample efficiency
A strong encoder turns a handful of interactions into usable structure. Agents that learn the right latent state need far less data to reach the same competence.
Long-horizon prediction
Because dynamics run in a compact, denoised latent space, rollouts stay coherent many steps ahead instead of accumulating error into noise.
Generalization and transfer
Representations shaped by prediction capture the underlying structure of an environment, so they carry over to new tasks, layouts, and embodiments.
Robustness to sensory noise
Learning what to keep — and what to throw away — makes encoders resilient to distractors, occlusion, and the messiness of real-world observation.
Everyone is scaling the world model. We are fixing what it learns to see.