TensorWorld
We create frontier models that faithfully understand, simulate, and interact with the physical world.
Explore our vision
Physical AI promises intelligence that can understand, simulate, and act in the physical world. Delivering on that promise requires as much attention to the world AI encounters as to AI itself.
Faithfulness to physical reality cannot be taken for granted. It must be built into how models understand, predict, and act. This is why we pursue physical reality: to make it foundational to intelligence itself.
The physical world follows its own laws, regardless of how convincing a generated scene, an explanation, or a plan may be. When AI interacts with this world, those laws govern what happens. What makes sense within a model must also hold true beyond it.
01
Video models generate scenes. Code constructs virtual worlds. We seek models that learn the physical structure and dynamics behind observations—inferring hidden properties, anticipating interactions, and generalizing beyond the assumptions built into hand-crafted simulations.
ProjectionSimplificationWe are building foundation models that internalize the structure and dynamics of the physical world. Our goal is to make faithfulness to physical reality intrinsic to these models—shaping how they understand the world, predict its behavior, and act within it.
We want these models to carry physical knowledge across tasks and environments: reasoning about unfamiliar situations, anticipating the consequences of actions, and learning from new evidence.
With these capabilities, we aim to empower robots to interact reliably with their surroundings, help engineers create better designs, and enable scientists to make new discoveries.
We have helped shape how machines understand the world—its physical laws, its temporal patterns, and its three-dimensional structure.
We advanced physics foundation models with Transolver and GeoPT, bringing the dynamics of physical systems into deep learning. We pushed the frontiers of learning from limited experience with EfficientZero, enabling agents to plan through learned models of the world. We developed Autoformer and TimesNet to uncover structure across time, and scaled 3D representation learning with Uni3D.
In previous roles, we led products serving over 100 million users and oversaw compute infrastructure spanning 100,000 accelerators. We bring that experience to TensorWorld: building frontier models at scale and turning research into products.
Today, we unite these strengths around a shared ambition: intelligence that understands how the world actually works—and acts reliably on that understanding.
