[hansonkd]

I mean it doesn't need to compute all programs in a human length reasonable amount of time.

It just needs to be able to compute enough programs to be useful.

Even our current infrastructure of precisely defined programs and compilers isn't able to compute all programs.

It seems reasonable in the future be able to give an LLM the python language specification, a python program, and it iteratively returns the answer.

[vrighter]

If it's executing a program, then the easiest way to make it more efficient, is to ditch the LLM and just execute the program. The LLM in this case is basically only (very very very very very inefficiently) approximating the very CPU it's running on. Just use the CPU to execute the program! And you won't be running it on an approximated processor, you'll be running it on a deterministic, reliable one that will not give the wrong answer ever (given the right program and correct input, of course, and assuming no hardware failures, which would affect LLMs too)