In our world, there are processes that cannot be compressed into functions. For example, turbulent air is very chaotic. Lets say you take a room and divide it into a 3d grid, where each cell is 0.1x0.1x0.1mm in dimentions, and has some properties of air associated around it (velocity vector, density, temperature, pressure, e.t.c).

At some time T, that grid is going to have a new values in every cell through the physics of air and turbulence.

The main question is, can an AI learn a compressed equation to where the inputs are the grid initial values and time T, and the output is the accurate final grid values for every cell?And by compressed, I mean the entire network takes up less space than a computer program that simulates the interactions at every boundary.

Or the more general way to put it, could we simulate reality faster than reality is happening, and have AI learn some compressed version of it to where you can ask it 2 "states" of reality (like the initial conditions of today, and then some time in the future where you have more money in your bank account and no criminal record and are still alive), and it can output a set of actions for you to take.

The answer, while not 100% definite, is very highly likely to be no.