You know that automated proof verifiers exist, right? And also that programs can know their own source code? Well, here's a puzzle for you:
Consider a program A that knows its own source code. The algorithm of A is as follows: generate and check all possible proofs up to some huge size (3^^^^3). If A finds a proof that A returns 1, it returns 1. If the search runs to the end and fails, it returns 0. What will this program actually return?
Wait, that was the easy version. Here's a harder puzzle:
Consider programs A and B that both know their own, and each other's, source code. The algorithm of A is as follows: generate and check all proofs up to size 3^^^^3. If A finds a proof that A returns the same value as B, it returns 1. If the search fails, it returns 0. The algorithm of B is similar, but possibly with a different proof system and different length limit. What will A and B return?
This second puzzle is a formalization of a Prisoner's Dilemma strategy proposed by Eliezer: "I cooperate if and only I expect you to cooperate if and only if I cooperate". So far we only knew how to make this strategy work by "reasoning from symmetry", also known as quining. But programs A and B can be very different - a human-created AI versus an extraterrestrial crystalloid AI. Will they cooperate?
I may have a tentative proof that the answer to the first problem is 1, and that in the second problem they will cooperate. But: a) it requires you to understand some logic (the diagonal lemma and Löb's Theorem), b) I'm not sure it's correct because I've only studied the subject for the last four days, c) this margin is too small to contain it. So I put it up here. I submit this post with the hope that, even though the proof is probably wrong or incomplete, the ideas may still be useful to the community, and someone else will discover the correct answers.
Edit: by request from Vladimir Nesov, I reposted the proofs to our wiki under my user page. Many thanks to all those who took the time to parse and check them.
timtyler: I think you're confused about how proof verifiers work. A proof verifier takes a proof as input, and verifies that the given proof is correct. It doesn't have to generate the proof itself, and it gets to assume that each step is some small, easily-checkable application from a finite set of axioms.
Put another way, if the proof you give a verifier makes a large leap in reasoning, then the verifier will just tell you that the proof is wrong. Moreover, by the formalization that we demand of those proofs, the proof is wrong.
So, the problem that proof verifiers solve is computable. So, we need only to assume that the proof verifier that these programs have is correctly implemented, we don't really need to know their internal details.
(If you already understand this, though, then I'm confused about your confusion. Could happen.)
[edit:] Actually, I think I might see the point you're making; we have to assume the theory the prover understands is rich enough to model stepwise program semantics. This probably should have been specified, but we can just assume it.