Hm, probably we disagree on something. I’m very confused how to mesh epistemic uncertainty with these “distribution over different Universes” types of probability. When I say “Boltzmann brains are probably very low measure”, I mean “I think Boltzmann brains are very low measure, but this is a confusing topic and there might be considerations I haven’t thought of and I might be totally mistaken”. I think this epistemic uncertainty is distinct from the type of “objective probabilities” I talk about in my post, and I don’t really know how to use language without referring to degrees of my epistemic uncertainty.
You still need a prior over worlds to calculate impacts, which is the cursed part.
Maybe we have some deeper disagreement here. It feels plausible to me that there is a measure of “realness” in the Multiverse that is an objective fact about the world, and we might be able to figure it out. When I say probabilities are cursed, I just mean that even if an objective prior over worlds and moments exist (like the Solomonoff prior), your probabilities of where you are are still hackable by simulations, so you shouldn’t rely on raw probabilities for decision-making, like the people using the Oracle do. Meanwhile, expected values are not hackable in the same way, because if they recreate you in a tiny simulation, you don’t care about that, and if they recreate you in a big simulation or promise you things in the outside world (like in my other post), then that’s not hacking your decision making, but a fair deal, and you should in fact let that influence your decisions.
Is your position that the problem is deeper than this, and there is no objective prior over worlds, it’s just a thing like ethics that we choose for ourselves, and then later can bargain and trade with other beings who have a different prior of realness?
I always assume when thinking about future dangerous models that they have access to some sort of black-box memory. Do we think there is a non-negligible chance that an AI that doesn’t have hidden memory, only English-language CoT, will be able to evade our monitoring and execute a rouge deployment? (Not a rhetorical question, there might be a way I haven’t thought of.)
So I think that assuming the AI being stateless when thinking about future risk is not a good idea, as I think the vast majority of the risk comes from AIs for which this assumption is not true. I think this makes the post’s claim significantly weaker, as the untrusted AI can just learn which of its attacks get replaced, and improve its attacks based on this data.
It still needs to avoid getting top 1% suspiciousness score and getting audited though, as it would be dangerous to the AI to ever get caught writing a text that clearly intends to jailbreak the monitor. So I agree that even an AI with black-box memory will still be severely constrained in how many and how egregious jailbreak attempts it can try.