AI's Third Modality Is Physics. The Bottleneck Is the Data.

AI's real frontier has been physics all along. Text and image models scale because their training data is cheap to generate. Physics-simulation models don't, because every training example demands a numerical solver grinding through a 3D scene to compute forces, collisions, and material behavior. That cost is what kept a "physics foundation model" off the map.

The MIT CSAIL and Tsinghua team's GeoPT paper attacks the bottleneck directly. Instead of leaning on solver-generated data, GeoPT pre-trains on synthesized 3D mechanical interactions, learning how particles stop against surfaces, how objects respond to wind, water, and collisions, then transfers that knowledge to downstream tasks. The reported payoff: peak performance reached twice as fast, with up to 60 percent less training data.

The repeated mechanism is a data-synthesis playbook. Wherever ground-truth labels are too expensive to generate at scale, the workaround is to pre-train on cheap, physics-consistent synthetic data and let the model generalize. Minghao Guo and his collaborators are explicit about the scope: physics is the third modality after text and pixels, and GeoPT is a step toward that foundation, not the finished one. A benchmark win is not the same as generalization across the messy physics of cars, planes, and robots. The frame, though, is the part that travels.