From Wordcels to
Shape Rotators

Language models have learned to reason through language, symbols, and software, however they still lack the ability to reason through space and acting in the physical world.

We believe we don’t have to rebuild intelligence from scratch. Today’s models already have the planning, abstraction, and long-horizon agency we need. We simply need fluency: learning to perceive, reason, and act across spatial and physical modalities.

We believe that fluency will arise from diversity rather than fidelity. Training across sufficiently diverse environments forces models to discover the universals beneath them: geometry, causality, affordances, and action. Reality becomes yet another test of generalization, not a separate domain.

We’re building the tools to transfer frontier capabilities onto new substrates. Our north star is generalized spatial intelligence: models that can form coherent representations of unfamiliar environments, understand how they work, and pursue long-horizon objectives within them.