Progress in the physical world comes at a cost. A new design has to be simulated, built, and tested. When it fails, engineers have to find out why and try again. Promising ideas for lighter structures, better cooling, and more capable robots can remain unexplored simply because they are too expensive to test.
We believe better prediction can change that. If engineers can anticipate how an unfamiliar design will behave, including where it is likely to fail, they can explore more possibilities with the time and resources they have. The value of prediction lies in what it allows us to attempt.
Our work starts with a question: how much of what a model learns from developing one product can help it understand the next?
Engineering already produces a wealth of information: simulation results, experimental measurements, manufacturing variations, and observations from products in use. We want models to learn from this experience and apply it to designs they have never encountered. Whether that learning carries over has to be demonstrated. On the next problem, the model should need less new data to make predictions engineers can trust.
Our starting point is solid mechanics and heat transfer, where measurements can inform concrete design decisions. We want to predict whether a new component will withstand a given load, or how a change in geometry will affect the flow of heat. Progress should show up in an engineer’s ability to choose which designs to pursue and which tests to run.
Working on real engineering problems should also shape the research. A failed prediction can reveal a gap in the data, a missing physical relationship, or a limitation in the model’s representation. That gives us a reason to collect a particular measurement or reconsider an architectural choice. As the model improves, engineers should be able to explore a wider range of designs. The important test is whether those improvements remain useful on the next project.
This places substantial demands on how models represent geometry, material properties, boundary conditions, and physical interactions. Architecture matters because these representations shape what a model can learn and how well it can generalize. New approaches will have to prove their value through reliable predictions on unfamiliar designs and a reduced need for additional experiments.
Over time, we want to extend this work to scientific discovery and robotics. Predicting the consequences of a design change and predicting what happens when a robot makes contact with an object involve related physical questions. We want to investigate how much of the underlying knowledge can be shared across these settings. Each extension will need its own evidence. That is how we intend to broaden the capabilities of our Physics World Model.
We want to build a company that turns experience with the physical world into a growing capacity to understand it. What we learn from one product should open up possibilities for the next. Each experiment should bring the next discovery closer.
If we can carry that understanding from one problem to the next, AI can help us build a world that is still beyond our reach.