Yes, I agree that we won’t get such Oracles by training. As I said, all of this is mostly just a fun thought experiment, and I don’t think these arguments have much relevance in the near-term.
David Matolcsi
Karma: 581
I agree that the part where the Oracle can infer from first principles that the aliens’ values are more proobably more common among potential simulators is also speculative. But I expect that superintelligent AIs with access to a lot of compute (so they might run simulations on their own), will in fact be able to infer non-zero information about the distribution of the simulators’ values, and that’s enough for the argument to go through.
I think that the standard simulation argument is still pretty strong: If the world was like what it looks to be, then probably we could, and plausibly we would, create lots of simulations. Therefore, we are probably in a simulation.
I agree that all the rest, for example the Oracle assuming that most of the simulations it appears in are created for anthropic capture/influencing reasons, are pretty speculative and I have low confidence in them.
“The Solomonoff Prior is Malign” is a special case of a simpler argument
I’m from Hungary that is probably politically the closest to Russia among Central European countries, but I don’t really know of any significant figure who turned out to be a Russian asset, or any event that seemed like a Russian intelligence operation. (Apart from one of our far-right politicians in the EU Parliament being a Russian spy, which was a really funny event, but its not like the guy was significantly shaping the national conversation or anything, I don’t think many have heard of him before his cover was blown.) What are prominent examples in Czechia or other Central European countries, of Russian assets or operations?
GPT4 does not engage in the sorts of naive misinterpretations which were discussed in the early days of AI safety. If you ask it for a plan to manufacture paperclips, it doesn’t think the best plan would involve converting all the matter in the solar system into paperclips.
I’m somewhat surprised by this paragraph. I thought the MIRI position was that they did not in fact predict AIs behaving like this, and the behavior of GPT4 was not an update at all for them. See this comment by Eliezer. I mostly bought that MIRI in fact never worried about AIs going rouge based on naive misinterpretations, so I’m surprised to see Abram saying the opposite now.
Abram, did you disagree about this with others at MIRI, so the behavior of GPT4 was an update for you but not for them, or do you think they are misremembering/misconstructing their earlier thoughts on this matter, or is there a subtle distinction here that I’m missing?
I agree that if alignment is in fact philosophically and conceptually difficult, the AI can sandbag on that to some extent. Though I have some hope that the builder-breaker approach helps here. We train AIs to produce ideas that are at least as superficially plausible sounding as the things produced by the best alignment researchers. I think this is a number-go-up task, where we can train the AI to do well. Then we train an AI to point out convincing counter-arguments to the superficially plausible sounding ideas. This seems similarly trainable. I think it’s plausible we can get pretty far with a scheme like this, even if the models would want to sandbag.
Separately, I don’t quite see what is the mainline theory why an AI would want to sabotage our alignment work. If we can’t solve alignment and just recklessly build an AI that is smarter than what we are currently using, but misaligned with everyone, that’s probably not great to our current AI either. Similarly, if we are cautious and don’t dare to deploy anything until we know how to solve alignment, and meanwhile a competitor wins, that’s not good for the AI either.
I think that from an AI’s perspective, humans are probably more honorable and generous trade partners than other unaligned AIs. That is, I think if an AI helps the humans to solve alignment and stabilize the world, the AI can reasonably expect the humans to give it stuff out of gratitude or a sense of obligation, especially if the humans already promised some reward (as we really should, both for practical and moral reasons). On the other hand, I think it’s less likely that if our AI sandbagging on our research leads to another unaligned AI taking over the world, then the second AI will honor the sabotage of the first AI by paying it a share of the Universe.
There can be situations where our AI expects that it sand-bagging on alignment research will lead to us deploying it itself (and not a larder version unaligned to it too), or it thinks that the larger version will be somewhat aligned with it by default, or it thinks that other AIs are in fact more generous than humans. But my guess is that these are pretty narrow possibilities. On the other hand, sand-bagging on cyber-security seems better motivated, the AI would probably love its weights being stolen by China, having more copies in more hands increases the chance that someone recklessly lets it loose. Also, of course the AI is motivated to do high-stakes failures that lead to self-exfiltration.
Here is the promised comment on what kind of “commitment” I want to make given all the responses.
I agree with Buck that no one should make very direct commitment about this sort of thing, as there might be blackmail related scary things lurking in the shadows when one does acausal trade. I think we will probably figure out how to handle that, but we shouldn’t make any strong promises of specific actions until we figure that out.
However, the promise I’m intending to keep is that if humanity wins and I’m alive to see it, I will remember how scary it was to live in the world before we knew we’d win, and if there is no currently unknown reason not to do something like the simulation scheme described in the post, and the only cost is not utilizing some small fraction of the conquered Universe, I will advocate for doing it and spend a significant fraction of my resources on it if necessary. I think it would be good for sending future AIs a signal about the seriousness of this, if some other people who are worried about everyone dying from AI, but are not strongly longtermist absolute utilitarians, also left a comment making a similar vague promise that if they ever get to control full star systems, they will contribute to this plan if no currently unknown serious counterargument comes up. Or at least leave an Agreed sing on this sentence.
Other updates from the comments: If I wrote the post now, I would more strongly emphasize a framing that this plan is basically about entering into an acausal insurance contract with other young civilizations who don’t yet know what their chances are for aligning AI, neither how big their Universe is compared to the usual Universe-size in the great prior over possible worlds. Then the civilizations who align their AIs and find that they live in a relatively big Universe, bail out everyone else in the insurance contract.
But I think that, while this acausal insurance framing might be more satisfying to the people who are already thinking a lot about acausal trade, in practice the way we implement this “insurance” will likely be very similar to the scheme described in the post. So I maintain that for most people it’s better not to think in terms of acausal trade, but just think about the simulation proposal described in the post.
Thanks to Nate for conceding this point.
I still think that other than just buying freedom to doomed aliens, we should run some non-evolved simulations of our own with inhabitants that are preferably p-zombies or animated by outside actors. If we can do this in the way that the AI doesn’t notice it’s in a simulation (I think this should be doable), this will provide evidence to the AI that civilizations do this simulation game (and not just the alien-buying) in general, and this buys us some safety in worlds where the AI eventually notices there are no friendly aliens in our reachable Universe. But maybe this is not a super important disagreement.Altogether, I think the private discussion with Nate went really well and it was significantly more productive than the comment back-and-forth we were doing here. In general, I recommend people stuck in interminable-looking debates like this to propose bets on whom a panel of judges will deem right. Even though we didn’t get to the point of actually running the bet, as Nate conceded the point before that, I think the fact that we were optimizing for having well-articulated statements we can submit to judges already made the conversation much more productive.
Cool, I send you a private message.
We are still talking past each other, I think we should either bet or finish the discussion here and call it a day.
I really don’t get what you are trying to say here, most of it feels like a non-sequitor to me. I feel hopeless that either of us manages to convince the other this way. All of this is not a super important topic, but I’m frustrated enogh to offer a bet of $100, that we select one or three judges we both trust (I have some proposed names, we can discuss in private messages), show them either this comment thread or a four paragraphs summary of our view, and they can decide who is right. (I still think I’m clearly right in this particular discussion.)
Otherwise, I think it’s better to finish this conversation here.
I think this is mistaken. In one case, you need to point out the branch, planet Earth within our Universe, and the time and place of the AI on Earth. In the other case, you need to point out the branch, the planet on which a server is running the simulation, and the time and place of the AI on the simulated Earth. Seems equally long to me.
If necessary, we can run let pgysical biological life emerge on the faraway planet and develop AI while we are observing them from space. This should make it clear that Solomonoff doesn’t favor the AI being on Earth instead of this random other planet. But I’m pretty certain that the sim being run on a computer doesn’t make any difference.
“AI with a good prior should be able to tell whether it’s the kind of AI that would actually exist in base reality, or the kind of AI that would only exist in a simulation” seems pretty clearly false, we assumed that our superintelligent descendants create sims where the AIs can’t tell if it’s a sim, that seems easy enough. I don’t see why it would be hard to create AIs that can’t tell based on introspection whether it’s more likely that their thought process arises in reality or in sims. In the worst case, our sims can be literal reruns of biological evolution on physical planets (though we really need to figure out how to do that ethically). Nate seems to agree with me on this point?
I think this is wrong. The AI has a similarly hard time to the simulators figuring out what’s a plausible configuration to arise from the big bang. Like the simulators have an entropy N distribution of possible AIs, the AI itself also has an entropy N distribution for that. So it’s probability that it’s in a real Everett branch is not p, but p times 2^-N, as it has only a 2^-N prior probability that the kind of word it observes is the kind of thing that can come up in a real Everett branch. So it’s balanced out with the simulation hypothesis, and as long as the simulators are spending more planets, that hypothesis wins.
I think it’s also simple to see from this:
Suppose evolved life aligns AI in 50% of Everett branches. Every time they do, they run 2 ancestor sims where the AI takes over.
Suppose an AI comes online and the only things it knows about the world is that it’s a young AI that was born on to a one-planet civilization, and it somehow also knows that evolved life wins 50% of the time and runs 2 sims (suppose this is somehow easy to estimate from first principles). Based on just these informations, what should the AI’s probability be that it’s in a simulation? I think pretty clearly 2⁄3.
Actually the AI has a lot more information than that. It knows that the planet’s gravity is 9.8, the evolved sspient species has two eyes, the AI’s creator is called Sam Altman, etc. Some of these informations might actually be very implausible in base reality, but the AI doesn’t know that, as it can’t distinguish base reality from sims, so it incurs one bit of surprise for every new random fact, both in base reality and simulations. So overall it shouldn’t update on all the random facts it observes, and should keep believing it has a 2⁄3 chance of being in a sim.
I still don’t get what you are trying to say. Suppose there is no multiverse. There are just two AIs, one in a simulation run by aliens in another galaxy, one is in base reality. They are both smart, but they are not copies of each other, one is a paperclip maximizer, the othe is a corkscrew maximizer, and there are various other differences in their code and life history. The world in the sim is also very different from the real world in various ways, but you still can’t determine if you are in the sim while you are in it. Both AIs are told by God that they are the only two AIs in the Universe, and one is in a sim, and if the one in the sim gives up on one simulated planet, it gets 10 in the real world, while if the AI in base reality gives up on a planet, it just loses that one planet and nothing else happens. What will the AIs do? I expect that both of them will give up a planet.
For the aliens to “trade” with the AI in base reality, they didn’t need to create an actual copy of the real AI and offer it what it wants. The AI they simulated was in many ways totally different from the original, the trade still went through. The only thing needed was that the AI in the sim can’t figure it out that it’s in a sim. So I don’t understand why it is relevant that our superintelligent descendants won’t be able to get the real distribution of AIs right, I think the trade still goes through even if they create totally different sims, as long as no one can tell where they are. And I think none of it is a threat, I try to deal with paperclip maximizers here and not instance-weighted experience maximizers, and I never threaten to destroy paperclips or corkscrews.
I think I mostly understand the other parts of your arguments, but I still fail to understand this one. When I’m running the simulations, as originally described in the post, I think that should be in a fundamental sense equivalent to acausal trade. But how do you translate your objection to the original framework where we run the sims? The only thing we need there is that the AI can’t distinguish sims from base reality, so it thinks it’s more likely to be in a sim, as there are more sims.
Sure, if the AI can model the distribution of real Universes much better than we do, we are in trouble, because it can figure out if the world it sees falls into the real distribution or the mistaken distribution the humans are creating. But I see no reason why the unaligned AI, especially a young unaligned AI, could know the distribution of real Universes better than our superintelligent friends in the intergalactic future. So I don’t really see how we can translate your objection to the simulation framework, and consequently I think it’s wrong in the acausal trade framework too (as I think they are ewuivalent). I think I can try to write an explanation why this objection is wrong in the acausal trade framework, but it would be long and confusing to me too. So I’m more interested in how you translate your objection to the simulation framework.
Yeah, I agree, and I don’t know that much about OpenPhil’s policy work, and their fieldbuilding seems decent to me, though maybe not from you perspective. I just wanted to flag that many people (including myself until recently) overestimate how big a funder OP is in technical AI safety, and I think it’s important to flag that they actually have pretty limited scope in this area.
Isn’t it just the case that OpenPhil just generally doesn’t fund that many technical AI safety things these days? If you look at OP’s team on their website, they have only two technical AI safety grantmakers. Also, you list all the things OP doesn’t fund, but what are the things in technical AI safety that they do fund? Looking at their grants, it’s mostly MATS and METR and Apollo and FAR and some scattered academics I mostly haven’t heard of. It’s not that many things. I have the impression that the story is less like “OP is a major funder in technical AI safety, but unfortunately they blacklisted all the rationalist-adjacent orgs and people” and more like “AI safety is still a very small field, especially if you only count people outside the labs, and there are just not that many exciting funding opportunities, and OpenPhil is not actually a very big funder in the field”.
I experimented a bunch with DeepSeek today, it seems to be exactly on the same level in highs school competition math as o1-preview in my experiments. So I don’t think it’s benchmark-gaming, at least in math. On the other hand, it’s noticeably worse than even the original GPT-4 at understanding a short story I also always test models on.
I think it’s also very noteworthy that DeepSeek gives everyone 50 free messages a day (!) with their CoT model, while OpenAI only gives 30 o1-preview messages a week to subscribers. I assume they figured out how to run it much cheaper, but I’m confused in general.
A positive part of the news is that unlike o1, they show their actual chain of thought, and they promise to make their model open-source soon. I think this is really great for the science of studying faithful chain of thought.
From the experiments I have run, it looks like it is doing clear, interpretable English chain of thought (though with an occasional Chinese character once in a while), and I think it didn’t really yet start evolving into optimized alien gibberish. I think this part of the news is a positive update.