I put myself through my PhD by programming video games. The hardest part was always simulating reality—getting the light, textures, and all those details just right. In those days, “Computer Graphics” by Foley, van Dam, Feiner, Hughes was my bible.

It was done mainly through Brute Force Ray Tracing—calculating every photon, every bounce, and every shadow. It was deterministic, expensive, and slow. A couple of days ago, OpenAI released Sora, and my brain is spinning. We aren’t just looking at a video generator; we are looking at a learned simulator of physical worlds.

The Diffusion of Reality

What strikes me most about the Sora samples—specifically the now-famous Tokyo Walk—isn’t just the aesthetic. It’s the space-temporal consistency.

In previous generative video models, objects would hallucinate out of existence the moment the camera panned. Sora doesn’t do that. From a data analysis perspective, it seems the model has developed an internal, latent understanding of 3D geometry. When the woman in the video walks past a shop window, the reflections aren’t just random noise; they obey a rudimentary version of the laws of optics.

A Physics Engine without Equations

We spend days at the University teaching the Navier-Stokes equations to describe fluid dynamics or Newton’s laws for motion. Sora doesn’t know these equations—so they say. Yet, when it renders a coffee cup splashing, it understands fluid density and gravity through probabilistic inference.

It has ingested so much visual data that it has effectively reverse-engineered the physical world. It treats video patches like the tokens we are now accustomed to talk when discussing LLMs. By predicting the next patch in a 3D space over time, it simulates a world that feels physically grounded.

The Computational Cost

Make no mistake: the compute required for this is gargantuan. Running these diffusion transformers (DiT) at this scale requires the kind of H100 clusters we only dreamed of a few years ago. We are moving from the era of Artificial Intelligence to the era of Artificial Simulation.

The Skeptic’s Corner

Is it perfect? No. You’ll see physics glitches—people merging into chairs or gravity-defying liquid. But as a detector builder, I know that once you have a signal, you can refine the resolution. Sora is the first high-fidelity signal that we can simulate a complex, four-dimensional world without a single line of traditional C++ physics code.

We are no longer just teaching machines to think; we are teaching them to visualize the rules of our Universe.