30,000 particles. Same three equations. All start within 0.00001 of one point.

For 10 seconds they look like a single dot. By 12 seconds, they're everywhere.

This is the problem physical AI doesn't talk about enough.

Genie 4 just dropped. DeepMind calls it their most realistic and interactive world model yet. World models are the bet behind robots, self-driving, and AI that "understands physics."

But realistic and predictive are not the same thing.

A world model can make a scene look physically right. Predicting what that scene will actually do is a different problem. Much of the real world is nonlinear: fluids, weather, crowds, friction and contact in a robot's hand. In these systems a tiny error in the starting state doesn't stay tiny. It grows exponentially.

That's what you're watching. The model here is perfect. The equations are exact. The only error is in the starting point, and it still takes over.

So no world model, however large, predicts a chaotic system forever. More compute buys seconds, not certainty.

What physical AI actually needs:

1. Know its own horizon. A model should know how far ahead it can be trusted.

2. Look again. Predict short, correct constantly with fresh observations. Weather forecasting has worked this way for decades. Robots should too.

3. Learn the shape, not the path. You can't say where one particle will be. You can predict the butterfly. Structure and statistics stay learnable even when single trajectories don't.

The next leap in physical AI won't be a model that looks more real.

It'll be one that knows when it's about to be wrong.