The current batch of ML models looks a lot like filling in holes in the wall of text, drawings or movies: you erase a part of the wall and tell it to fix it. An

The current batch of ML models looks a lot like filling in holes in the wall of text, drawings or movies: you erase a part of the wall and tell it to fix it. And it fills in the hole using colors from the nearby walls in the kitchen and similar walls and we watch this in awe thinking it must've figured out the design rules of the kitchen. However what it's really done is it interpolated the gaps with some sort of basic functions, trigonometric polynomials for example, and it used thousands of those. This solution wouldn't occur to us because our limited memory isn't enough for thousands of polynomials: we have to find a compact set of rules or give up entirely. So when these ML models predict the motion of planets, they approximate the Newton's law with a long series of basic functions.