Hacker News Reader

When the answers can be checked by another program and best-of-n results get better with increasing n, yes.

I think TDD code development is this, at least in principle.

If you can write a program to check results you've already solved the issue.

Sometimes.

Some things are easier to verify than to generate.

LLMs can often make a distribution of generated outputs that is much closer to a good answer than other methods.