In the source code of Y, in a place where X will apply its inference module, the programmer has included an explicit Henkin sentence (i.e. contains the ingredients for its own proof) such that when X applies its inference module, it quickly proves “if X!=a then U=-1000”
I still don’t understand what you’re saying. What’s the nature of this “Henkin sentence”, that is in what way is it “included” in Y’s source code? Is it written in the comments or something? What does it mean for X to “apply its inference module”? My only guess is that you are talking about X starting to enumerate proofs in order other than their length, checking the proofs that happen to be lying around in Y’s code just in case, so that including something in Y’s code might push X to consider a proof that Y wants it to consider earlier than some other proof. Is it what you mean?
How would “if X!=a then U=-1000” follow from a Henkin sentence (it obviously can’t make anything it likes true)?
The mental image of an inference module that I’m trying to trick is something like a chess program, performing a heuristically biased depth-first search on the space of proofs, based on the source code of the round. If the first path it takes is a garden path to an easy but spurious counterfactual proof, and if the genuine counterfactual is unreachable at the same level (by some diagonalization of X’s proof method), then X should be fooled. The question is whether, knowing X’s source code, there’s a Y which leads X’s heuristics down that garden path.
Since in the end we’ll achieve X=a, the statement “if X!=a then U=-1000” is true, and it should be provable by NDT if it finds the right path, so there should be some way of helping NDT find the proof more quickly.
I don’t have a strong intuition for how to fool the “enumerate and check all proofs up to length N” inference module.
Okay. Let’s say we write an X that will start from checking any proof given to it (and accepting its conclusion). How can we construct a proof of X=a that X can read then? It looks like the more X reads, the longer the proof must be, and so there doesn’t exist a proof that X can also read. I don’t see how to construct a counterexample to this property without corrupting X’s inference system, though I imagine some quining trick might work...
Right. That’s why I acknowledge that this is speculative.
If it turns out there’s really no need to worry about spurious counterfactuals for a reasonable inference module, then hooray! But mathematical logic is perverse enough that I expect there’s room for malice if you leave the front door unlocked.
...and Slepnev posted a proof (on the list) that in my formulation, X can be successfully deceived. It’s not so much a Henkin sentence, just a program that enumerates all proofs, looking for a particular spurious counterfactual, and doesn’t give up until it finds it. If the spurious counterfactual is provable, the program will find it, and so the agent will be tricked by it, and then the spurious counterfactual will be true. We have an implication from provability of the spurious argument to its truth, so by Loeb’s theorem it’s true, and X is misled. So you were right!
I still don’t understand what you’re saying. What’s the nature of this “Henkin sentence”, that is in what way is it “included” in Y’s source code? Is it written in the comments or something? What does it mean for X to “apply its inference module”? My only guess is that you are talking about X starting to enumerate proofs in order other than their length, checking the proofs that happen to be lying around in Y’s code just in case, so that including something in Y’s code might push X to consider a proof that Y wants it to consider earlier than some other proof. Is it what you mean?
How would “if X!=a then U=-1000” follow from a Henkin sentence (it obviously can’t make anything it likes true)?
The mental image of an inference module that I’m trying to trick is something like a chess program, performing a heuristically biased depth-first search on the space of proofs, based on the source code of the round. If the first path it takes is a garden path to an easy but spurious counterfactual proof, and if the genuine counterfactual is unreachable at the same level (by some diagonalization of X’s proof method), then X should be fooled. The question is whether, knowing X’s source code, there’s a Y which leads X’s heuristics down that garden path.
Since in the end we’ll achieve X=a, the statement “if X!=a then U=-1000” is true, and it should be provable by NDT if it finds the right path, so there should be some way of helping NDT find the proof more quickly.
I don’t have a strong intuition for how to fool the “enumerate and check all proofs up to length N” inference module.
Okay. Let’s say we write an X that will start from checking any proof given to it (and accepting its conclusion). How can we construct a proof of X=a that X can read then? It looks like the more X reads, the longer the proof must be, and so there doesn’t exist a proof that X can also read. I don’t see how to construct a counterexample to this property without corrupting X’s inference system, though I imagine some quining trick might work...
Right. That’s why I acknowledge that this is speculative.
If it turns out there’s really no need to worry about spurious counterfactuals for a reasonable inference module, then hooray! But mathematical logic is perverse enough that I expect there’s room for malice if you leave the front door unlocked.
...and Slepnev posted a proof (on the list) that in my formulation, X can be successfully deceived. It’s not so much a Henkin sentence, just a program that enumerates all proofs, looking for a particular spurious counterfactual, and doesn’t give up until it finds it. If the spurious counterfactual is provable, the program will find it, and so the agent will be tricked by it, and then the spurious counterfactual will be true. We have an implication from provability of the spurious argument to its truth, so by Loeb’s theorem it’s true, and X is misled. So you were right!
I just wrote up the proof Nesov is talking about, here.