I want to keep picking a fight about “will the AI care so little about humans that it just kills them all?” This is different from a broader sense of cosmopolitanism, and moreover I’m not objecting to the narrow claim “doesn’t come for free.” But it’s directly related to the actual emotional content of your parables and paragraphs, and it keeps coming up recently with you and Eliezer, and I think it’s an important way that this particular post looks wrong even if the literal claim is trivially true.
(Note: I believe that AI takeover has a ~50% probability of killing billions and should be strongly avoided, and would be a serious and irreversible decision by our society that’s likely to be a mistake even if it doesn’t lead to billions of deaths.)
Humans care about the preferences of other agents they interact with (not much, just a little bit!), even when those agents are weak enough to be powerless. It’s not just that we have some preferences about the aesthetics of cows, which could be better optimized by having some highly optimized cow-shaped objects. It’s that we actually care (a little bit!) about the actual cows getting what they actually want, trying our best to understand their preferences and act on them and not to do something that they would regard as crazy and perverse if they understood it.
If we kill the cows, it’s because killing them meaningfully helped us achieve some other goals. We won’t kill them for arbitrarily insignificant reasons. In fact I think it’s safe to say that we’d collectively allocate much more than 1/millionth of our resources towards protecting the preferences of whatever weak agents happen to exist in the world (obviously the cows get only a small fraction of that).
Before really getting into it, some caveats about what I want to talk about:
I don’t want to focus on whatever form of altruism you and Eliezer in particular have (which might or might not be more dependent on some potentially-idiosyncratic notion of “sentience.”) I want to talk about caring about whatever weak agents happen to actually exist, which I think is reasonably common amongst humans. Let’s call that “kindness” for the purpose of this comment. I don’t think it’s a great term but it’s the best short handle I have.
I’ll talk informally about how quantitatively kind an agent is, by which I mean something like: how much of its resources it would allocate to helping weak agents get what they want? How highly does it weigh that part of its preferences against other parts? To the extent it can be modeled as an economy of subagents, what fraction of them are kind (or were kind pre-bargain)?
I don’t want to talk about whether the aliens would be very kind. I specifically want to talk about tiny levels of kindness, sufficient to make a trivial effort to make life good for a weak species you encounter but not sufficient to make big sacrifices on its behalf.
I’m not talking about whether the AI has spite or other strong preferences that are incompatible with human survival, I’m engaging specifically with the claim that AI is likely to care so little one way or the other that it would prefer just use the humans for atoms.
You and Eliezer seem to think there’s a 90% chance that AI will be <1/trillion (perhaps even a 90% chance that they have exactly 0 kindness?). But we have one example of a smart mind, and in fact: (i) it has tons of diverse shards of preference-on-reflection, varying across and within individuals (ii) it has >1/million kindness. So it’s superficially striking to be confident AI systems will have a million times less kindness.
I have no idea under what conditions evolved or selected life would be kind. The more preferences are messy with lots of moving pieces, the more probable it is that at least 1/trillion of those preferences are kind (since the less correlated the trillion different shards of preference are with one another and so the more chances you get). And the selection pressure against small levels of kindness is ~trivial, so this is mostly a question about idiosyncrasies and inductive biases of minds rather than anything that can be settled by an appeal to selection dynamics.
I can’t tell if you think kindness is rare amongst aliens, or if you think it’s common amongst aliens but rare amongst AIs. Either way, I would like to understand why you think that. What is it that makes humans so weird in this way?
(And maybe I’m being unfair here by lumping you and Eliezer together—maybe in the previous post you were just talking about how the hypothetical AI that had 0 kindness would kill us, and in this post how kindness isn’t guaranteed. But you give really strong vibes in your writing, including this post. And in other places I think you do say things that don’t actually add up unless you think that AI is very likely to be <1/trillion kind. But at any rate, if this post is unfair to you, then you can just sympathize and consider it directed at Eliezer instead who lays out this position much more explicitly though not in a convenient place to engage with.)
Here are some arguments you could make that kindness is unlikely, and my objections:
“We can’t solve alignment at all.” But evolution is making no deliberate effort to make humans kind, so this is a non-sequitur.
“This is like a Texas sharpshooter hitting the side of a barn then drawing a target around the point they hit; every evolved creature might decide that their own idiosyncrasies are common but in reality none of them are.” But all the evolved creatures wonder if a powerful AI they built would kill them or if if it would it be kind. So we’re all asking the same question, we’re not changing the question based on our own idiosyncratic properties. This would have been a bias if we’d said: humans like art, so probably our AI will like art too. In that case the fact that we were interested in “art” was downstream of the fact that humans had this property. But for kindness I think we just have n=1 sample of observing a kind mind, without any analogous selection effect undermining the inference.
“Kindness is just a consequences of misfiring [kindness for kin / attachment to babies / whatever other simple story].” AI will be selected in its own ways that could give rise to kindness (e.g. being selected to do things that humans like, or to appear kind). The a priori argument for why that selection would lead to kindness seems about as good as the a priori argument for humans. And on the other side, the incentives for humans to be not kind seem if anything stronger than the incentives for ML systems to not be kind. This mostly seems like ungrounded evolutionary psychology, though maybe there are some persuasive arguments or evidence I’ve just never seen.
“Kindness is a result of the suboptimality inherent in compressing a brain down into a genome.” ML systems are suboptimal in their own random set of ways, and I’ve never seen any persuasive argument that one kind of suboptimality would lead to kindness and the other wouldn’t (I think the reverse direction is equally plausible). Note also that humans absolutely can distinguish powerful agents from weak agents, and they can distinguish kin from unrelated weak agents, and yet we care a little bit about all of them. So the super naive arguments for suboptimality (that might have appealed to information bottlenecks in a more straightforward way) just don’t work. We are really playing a kind of complicated guessing game about what is easy for SGD vs easy for a genome shaping human development.
“Kindness seems like it should be rare a priori, we can’t update that much from n=1.” But the a priori argument is a poorly grounded guess about about the inductive biases of spaces of possible minds (and genomes), since the levels of kindness we are talking about are too small to be under meaningful direct selection pressure. So I don’t think the a priori arguments are even as strong as the n=1 observation. On top of that, the more that preferences are diverse and incoherent the more chances you have to get some kindness in the mix, so you’d have to be even more confident in your a priori reasoning.
“Kindness is a totally random thing, just like maximizing squiggles, so it should represent a vanishingly small fraction of generic preferences, much less than 1/trillion.” Setting aside my a priori objections to this argument, we have an actual observation of an evolved mind having >1/million kindness. So evidently it’s just not that rare, and the other points on this list respond to various objections you might have used to try to salvage the claim that kindness is super rare despite occurring in humans (this isn’t analogous to a Texas sharpshooter, there aren’t great debunking explanation for why humans but not ML would be kind, etc.). See this twitter thread where I think Eliezer is really off base, both on this point and on the relevance of diverse and incoherent goals to the discussion.
Note that in this comment I’m not touching on acausal trade (with successful humans) or ECL. I think those are very relevant to whether AI systems kill everyone, but are less related to this implicit claim about kindness which comes across in your parables (since acausally trading AIs are basically analogous to the ants who don’t kill us because we have power).
A final note, more explicitly lumping you with Eliezer: if we can’t get on the same page about our predictions I’m at at least aiming to get folks to stop arguing so confidently for death given takeover. It’s easy to argue that AI takeover is very scary for humans, has a significant probability of killing billions of humans from rapid industrialization and conflict, and is a really weighty decision even if we don’t all die and it’s “just” handing over control over the universe. Arguing that P(death|takeover) is 100% rather than 50% doesn’t improve your case very much, but it means that doomers are often getting into fights where I think they look unreasonable.
I think OP’s broader point seems more important and defensible: “cosmopolitanism isn’t free” is a load-bearing step in explaining why handing over the universe to AI is a weighty decision. I’d just like to decouple it from “complete lack of kindness.”
Eliezer writes:
I think this suggests a really poor understanding of what’s going on. My fairly strong guess is that OpenAI folks know that it is possible to get ChatGPT to respond to inappropriate requests. For example:
They write “While we’ve made efforts to make the model refuse inappropriate requests, it will sometimes respond to harmful instructions.” I’m not even sure what Eliezer thinks this means—that they hadn’t actually seen some examples of it responding to harmful instructions, but they inserted this language as a hedge? That they thought it randomly responded to harmful instructions with 1% chance, rather than thinking that there were ways of asking the question to which it would respond? That they found such examples but thought that Twitter wouldn’t?
These attacks aren’t hard to find and there isn’t really any evidence suggesting that they didn’t know about them. I do suspect that Twitter has found more amusing attacks and probably even more consistent attacks, but that’s extremely different from “OpenAI thought there wasn’t a way to do this but there was.” (Below I describe why I think it’s correct to release a model with ineffective precautions, rather than either not releasing or taking no precautions.)
If I’m right that this is way off base, one unfortunate effect would be to make labs (probably correctly) take Eliezer’s views less seriously about alignment failures. That is, the implicit beliefs about what labs notice, what skills they have, how decisions are made, etc. all seem quite wrong, and so it’s natural to think that worries about alignment doom are similarly ungrounded from reality. (Labs will know better whether it’s inaccurate—maybe Eliezer is right that this took OpenAI by surprise in which case it may have the opposite effect.)
(Note that I think that alignment is a big deal and labs are on track to run a large risk of catastrophic misalignment! I think it’s bad if labs feel that concern only comes from people underestimating their knowledge and ability.)
I think it makes sense from OpenAI’s perspective to release this model even if protections against harms are ineffective (rather than not releasing or having no protections):
The actual harms from increased access to information are relatively low; this information is available and easily found with Google, so at best they are adding a small amount of convenience (and if you need to do a song and dance and you get back your answer as a poem, you are not even more convenient).
It seems likely that OpenAI’s primary concern is with PR risks or nudging users in bad directions. If users need to clearly go out of their way to coax the model to say bad stuff, then that mostly addresses their concerns (especially given point #1).
OpenAI making an unsuccessful effort to solve this problem makes it a significantly more appealing target for research, both for researchers at OpenAI and externally. It makes it way more appealing for someone to outcompete OpenAI on this axis and say “see OpenAI was trying but failed, so our progress is cool” vs the world where OpenAI said “whatever, we can’t solve the problem so let’s just not even try so it does’t look like we failed.” In general I think it’s good for people to advertise their alignment failures rather than trying to somehow cover them up. (I think saying the model confidently false stuff all the time is a way bigger problem than the “jailbreaking,” but both are interesting and highlight different alignment difficulties.)
I think that OpenAI also likely has an explicit internal narrative that’s like “people will break our model in creative ways and that’s a useful source of learning, so it’s great for us to get models in front of more eyes earlier.” I think that has some truth to that (though not for alignment in particular, since these failures are well-understood internally prior to release) but I suspect it’s overstated to help rationalize shipping faster.
To the extent this release was a bad idea, I think it’s mostly because of generating hype about AI, making the space more crowded, and accelerating progress towards doom. I don’t think the jailbreaking stuff changes the calculus meaningfully and so shouldn’t be evidence about what they did or did not understand.
I think there’s also a plausible case that the hallucination problems are damaging enough to justify delaying release until there is some fix, I also think it’s quite reasonable to just display the failures prominently and to increase the focus on fixing this kind of alignment problem (e.g. by allowing other labs to clearly compete with OpenAI on alignment). But this just makes it even more wrong to say “the key talent is not the ability to imagine up precautions but the ability to break them up,” the key limit is that OpenAI doesn’t have a working strategy.