[darkmighty]

I don't think you quite grasp how hard a problem can be for a fixed size Turing machine. The state size is 2^(Memory), naively solving the halting problem for all cases that can be solved in a reasonably-sized Turing machine still makes the age of the universe look like nothing. 2^10000000000 is nothing to scoff at.

For the Collatz problem you can simply run a (fixed version) of that program with arbitrary-precision arithmetic and it would pretty much run until you run out of memory, which is on the order of 2^RAM. For me that would be about 2^8000000000.

I think an important argument that hasn't been made is the simulation speed. If we want to make sure a program running at clock C never ill-behaves, it suffices to simulate it at a clock C'>C and halt it if and when it misbehaves. Any constant factor speedup is sufficient to manage that even in the same hardware. A problem starts to occur if C is susceptible to random mutations, with a branching factor B every T seconds. Then to simulate t seconds into the future requires B^T/t more computing power. If there's one 1-bit mutation per second, it gets unpredictible less than 1 hour in the future.