A possible limit of AI

In order to fully replace humans in the real world, it would seem that  AI would have to be trained on the real world. Training AI on the actual real world would be problematic, because it is time consuming and because the mistakes it would make during training would affect people in the real world. So the ideal way to train an AI would be to simulate the real world.

Simulating the real world does not at first seem like an insurmountable task. I mean we already have video games that look incredibly realistic and operate according to realistic physics. So it would seem that a full simulation of the real world is not that far off.

But reflect on the effects of physics. Physics not only governs how quickly a dropped ball hits floor, or how momentum is transferred in a collision, but it also governs molecular interactions, giving rise to chemistry. And then chemistry gives rise to biology, and biology to neuroscience, and neuroscience to psychology, and psychology to economics. All ultimately from physics. So in a truly realistic simulation of the real world, it would be told nothing more than the laws of physics, and using only this it would simulate ultimately even the economy of the world. This seems close to impossible. Even simulating something as simple as an orbiting satellite directly from the law of gravity is difficult to do because computer rounding errors can drastically affect the outcome of the simulation as it progresses, causing it to spiral inward or outward. I know there are ways to compensate for this specific problem, but the fact remains that small mistakes in simulation can cause big problems. Just think how accurate one would have to be if his aim were to simulate all of economics from the laws of physics, and all the things that could go wrong.

After thinking about this some more, I think it may be possible to do a simulation that is good enough to eliminate many jobs without doing the kind of deep simulation that I described. But still not engineering jobs, because how could it train in engineering if the materials are not the same?

But then again, AI has surprised me before, and done things I thought AI could never be capable of, so maybe they will figure out a way to make simulations good enough to train AIs to replace all jobs, without doing the deep simulation like I described.

According to Al Bielik who travelled 700 years into the future says that despite majority of the population live in floating cities with society taken care of by AI machines, there are selective minority of people who reject the convenience of AI and live in more primitive societies on the ground