Time Travel, AI and Transparent Newcomb

by johnswentworth 1 min read22nd Aug 20197 comments

12


Epistemic status: has "time travel" in the title.

Let's suppose, for the duration of this post, that the local physics of our universe allows for time travel. The obvious question is: how are paradoxes prevented?

We may not have any idea how paradoxes are prevented, but presumably there must be some prevention mechanism. So, in a purely Bayesian sense, we can condition on paradoxes somehow not happening, and then ask what becomes more or less likely. In general, anything which would make a time machine more likely to be built should become less likely, and anything which would prevent a time machine being built should become more likely.

In other words: if we're trying to do something which would make time machines more likely to be built, this argument says that we should expect things to mysteriously go wrong.

For instance, let's say we're trying to build some kind of powerful optimization process which might find time machines instrumentally useful for some reason. To the extent that such a process is likely to build time machines and induce paradoxes, we would expect things to mysteriously go wrong when trying to build the optimizer in the first place.

On the flip side: we could commit to designing our powerful optimization process so that it not only avoids building time machines, but also actively prevents time machines from being built. Then the mysterious force should work in our favor: we would expect things to mysteriously go well. We don't need time-travel-prevention to be the optimization process' sole objective here, it just needs to make time machines sufficiently less likely to get an overall drop in the probability of paradox.