PolymathicAl released a scientific foundation model (called Walrus) the other day.
Today they posted a follow up paper examining how the model represents the physical world.
Basically they look at the model's "thoughts" and show that it understands very abstract physical ideas (like speed, or diffusion).
To be clear Walrus itself isn’t yet a fully general purpose “physics AI” because it only works with continuum (fluid-like) data, but it feels like a big stepping stone because it is able to handle anything that is vaguely fluid like (e.g. plasma, gasses, acoustics, turbulence, astrophysics etc). It appears the model is looking at all these different systems and finding general laws/principles that underly everything.