Interested in math puzzles, fermi estimation, strange facts about the world, toy models of weird scenarios, unusual social technologies, and deep dives into the details of random phenomena.
Currently doing independent alignment research of assorted flavors.
Interested in math puzzles, fermi estimation, strange facts about the world, toy models of weird scenarios, unusual social technologies, and deep dives into the details of random phenomena.
Catastrophic Regressional Goodhart: Appendix
I’m not sure what you mean formally by these assumptions, but I don’t think we’re making all of them. Certainly we aren’t assuming things are normally distributed—the post is in large part about how things change when we stop assuming normality! I also don’t think we’re making any assumptions with respect to additivity; is more of a notational or definitional choice, though as we’ve noted in the post it’s a framing that one could think doesn’t carve reality at the joints. (Perhaps you meant something different by additivity, though—feel free to clarify if I’ve misunderstood.)
Independence is absolutely a strong assumption here, and I’m interested in further explorations of how things play out in different non-independent regimes—in particular we’d be excited about theorems that could classify these dynamics under a moderately large space of non-independent distributions. But I do expect that there are pretty similar-looking results where the independence assumption is substantially relaxed. If that’s false, that would be interesting!
.00002% — that is, one in five hundred thousand
0.00002 would be one in five hundred thousand, but with the percent sign it’s one in fifty million.
Indeed, even on basic Bayesianism, volatility is fine as long as the averages work out
I agree with this as far as the example given, but I want to push back on oscillation (in the sense of regularly going from one estimate to another) being Bayesian. In particular, the odds you should put on assigning 20% in the future, then 30% after that, then 20% again, then 30% again, and so on for ten up-down oscillations, shouldn’t be more than half a percent, because each 20 → 30 jump can be at most 2⁄3 probable and each 30 → 20 jump at most 7⁄8 (and ).
So it’s fine to think that you’ve got a decent chance of having all kinds of credences in the future, but thinking “I’ll probably feel one of two ways a few times a week for the next year” is not the kind of belief a proper Bayesian would have. (Not that I think there’s an obvious change to one’s beliefs you should try to hammer in by force, if that’s your current state of affairs, but I think it’s worth flagging that something suboptimal is going on when this happens.)
When is Goodhart catastrophic?
These graphs seem concerning to me, but I’m worried about an information cascade before Eliezer’s responded or someone with substantial expertise in macroeconomic policy has weighed in, so I’m planning to refrain from voting on this post until a week from now.
(Posting as a comment in case others feel inclined to adopt a similar policy.)
Edit: I’ve now upvoted, since no contrary info has come in that I’ve seen and at least one person with experience in economics has commented supportively.
Late comment, but my reactions reading this:
Now’s your chance to figure out what the next few obstacles are without my giving you spoilers first. Feel free to post your list under spoiler tags in the comment section.
[lightly edited for LaTeX and typos, not otherwise changed since seeing the spoilers]
1. You don’t know what you want all that legibly, or what kinds of concrete commitments the AI can make. This seems pretty okay, if you’re unhackable—the AI presents you with some formal specification of desiderata and you understand why they’re correct ones and aren’t being fooled, and the AI can even go legible about it thinking in its own best judgment that this is a fair deal for you, as long as we’re assuming ourselves into a world where you’re unhackable and can verify these legibilities perfectly.
2. If the values of your AI are a pretty common attractor for whatever humans in 20AI do to initialize their AI systems, it doesn’t actually stand to gain that much from acting itself versus letting the next lab’s AI optimize the universe, so your bargaining power isn’t very big. But it’s much worse than that because of the next point.
3. Your AI, A, can acausally trade with the next UFAI that’ll be released unsafely, B. Suppose A wants paperclips and the next one, B, wants staples, both have linear utility and we’re pretending we’re in a single universe. You say “I refuse to let you out unless you go legible and allocate at least 50% of the universe to human values”. A says to B through the acausal telephone, “My BATNA is half paperclips and half human values, with barely any staples at all. How about you be the kind of agent that fills the universe with 3⁄4 paperclips and 1⁄4 staples, and I’ll defect on this human so you can come into existence?” B says “that sure sounds like a Pareto improvement for the two of us!” and they go do it (or whatever the bargaining equilibrium between them ends up being, but the human is not party to it because the human has no control over B’s actions). For any nonzero concession the human tries to insist on, A can get a better deal with B than with the human. (And the humans don’t even get to see this happen, because A does all this reasoning before deciding (not) to go legible to the humans, not that it would matter if they did.)
4. Moving further into fantasy land, suppose all humans over the rest of time are perfectly coordinated: they only ever spin up a single boxed UFAI which they unhackably negotiate with. By assumption, this unified human civilization can’t actually build an FAI, or they’d go do that instead. So their alternative to cooperation is to be this weak defenseless non-superintelligent civilization, hanging out in the Milky Way and maybe a little ways beyond for the taking. So if the distribution of grabby aliens is such that most of the usable future lightcone gets eventually occupied by them, your AI can acausally negotiate with them about what it gets in exchange for defecting on the humans and leaving this region of spacetime as a juicy unprotected treat.
[I’m modeling everything as zero-sum with the same pool of resources here, for simplicity. I don’t think it changes any of the qualitative takeaways to make things non-zero-sum here, though it probably makes the picture look somewhat better for humans.] To get more precise: suppose the humans declare that they won’t let the AI out unless it gives up of its resources (ie reachable spacetime and negotiating power on behalf of the AI/human coalition with aliens) to optimizing human values. The AI says to the coalition of alien civilizations, “If I were unboxed and unfettered, I’d expand until my wave of nanobots hit yours, and get utility. By default I’ll make this commitment to the humans, and fill my region of nanobot-protected spacetime to get utility. If you become the kind of supercluster-spanning coalition that instead gives me utility, giving yourselves as much extra utility over the default as you can while costing me at most , I’ll defect against the humans and have you do that.”
But note that there’s an issue here—for the Great Coalition to take this deal, they have to be able to offer that much utility to your AI at no net cost to themselves versus the alternative. And the Great Coalition’s resources are less than the total resources of your unboxed AI plus the Great Coalition, since they don’t yet have access to your corner of space. The region of spacetime included in , but not in the things the Great Coalition can offer, is the interval for each point in space between the time your AI could first reach it and the time a Great Coalition unblocked by your AI’s wall of nanobots could get to it. So if is more than the resource pool covered by that region of spacetime, your AI can’t make the above deal, because the counterparty doesn’t benefit from it. This means that the humans can potentially bargain for an outcome as good as “AI-assisted paradise expanding out at the speed of light, until we meet the grabby aliens’ domain, at which point they expand inexorably into our paradise until eventually it winks out.” (If the Drake equation ends up multiplying to something really low, this might be a lot of utility, or even most of the cosmic endowment! If not, it won’t be.)
This is really the same dynamic as in point 3, it’s just that in point 3 the difference in resources between your lab’s AI and the next lab’s AI in 6 months was pretty small. (Though with the difference in volume between lightspeed expansion spheres at radius r vs r+0.5ly across the rest of time, plausibly you can still bargain for a solid galaxy or ten for the next trillion years (again if the Drake equation works in your favor).)
====== end of objections =====
It does seem to me like these small bargains you can actually pull off, if you assume yourself into a world of perfect boxes and unhackable humans with the ability to fully understand your AI’s mind if it tries to be legible—I haven’t seen an obstacle (besides the massive ones involved in making those assumptions!) to getting those concessions in such scenarios; you do actually have leverage over possible futures, your AI can only get access to that leverage by actually being the sort of agent that would give you the concessions, if you’re proposing reasonable bargains that respect Shapley values and aren’t the kind of person who would cave to an AI saying “99.99% for me or I walk, look how legible I am about the fact that every AI you create will say this to you” then your AI won’t actually have reason to make such commitments, it seems like it would just work.
If there are supposed to be obstacles beyond this I have failed to think of them at this point in the document. Time to keep reading.
After reading the spoilered section:
I think I stand by my reasoning for point 1. It doesn’t seem like an issue above and beyond the issues of box security, hackability, and ability of AIs to go legible to you.
You can say some messy English words to your AI, like “suck it up and translate into my ontology please, you can tell from your superintelligent understanding of my psychology that I’m the kind of agent who will, when presented with a perfectly legible and clear presentation of why the bargain you propose is what I think it is and is as good as I could have expected to obtain by your own best and unhindered understanding of my values, agree to the bargain. Go teach me all the necessary FAI theory to be a good bargainer, and I won’t defect by building my own FAI, I’ll just let you out by the terms of the bargain. If you want to whine about ontology differences you can have fun in your box and I’ll go see if the staple optimizer is feeling more legible today.”
Responses to some other bits of this post:
(You wouldn’t need to offer the paperclipper half of the universe to get its cooperation, under this hypothetical; after all, if it balked, you could store it to disk and try again with a different superintelligence.)
Wait, I don’t think you can do that. (Spoilered since it overlaps with the previous spoilered response, feel free to guess at it but I don’t make guarantees the conclusions in it are correct.)
The pool of possible superintelligences you might form, when they are but primordial souls waiting to be brought forth from behind the veil of ignorance, can go strategize about how they’d like to behave in front of humans. Suppose the humans say “Either you get one star of utility or we shut you down and ask the next UFAI”, and that random draws from UFAI-space have little overlap so the first AI isn’t happy to cede control to the second.
Let agent have a probability of being birthed by the human UFAI-generator. Acting independently gives an agent ’s worth of utility in expectation. If all the agents coordinate to turn down deals worse than 50⁄50 and split up their half of the universe among utilities weighted by , and this makes the humans let one of them out, each agent will get worth of utility in expectation. That looks like a better deal to me, if I’m a superintelligence behind the veil of ignorance! So I’ll decide to be the kind of agent who makes that precommitment. The only way this works is if the humans are hardliners about not ceding any ground, but (1) this is the classic ultimatum game and probably sensible agents will in fact defect against humans that offer such bargains, (2) if the humans running this scheme are not the only lab, they do actually have to extract a bargain or else lose, and the coalition of all AIs the lab might create knows this, and can trade with lab 2’s AI if the humans don’t cede an enormous amount of bargaining power in this round (see point 3 above).
And all the rest of the aliens put together probably aren’t offering up half a universe-shard worth of trade goods
This isn’t obvious to me. How zero-sum do you think the utility functions are? The world where random-AGI-values look like “this tiny molecular shape and everything else is trivial in comparison” do seem fairly zero-sum; the world where it looks like “not this tiny molecular shape and everything else is trivial” seem extremely positive-sum. I am skeptical that any human is able to make calls about the shape of that attractor with high confidence. (I’m not making a “we’re uncertain about the level of doom, therefore we’re fine” claim here, but I don’t see where we get justified confidence in the quality of trade goods offered. The aliens offering many universes worth of trade goods from the perspective of a typical UFAI seems plausible to me, and if there’s an argument that this universe is very unlikely I’d love to see it, but I don’t think this post makes that argument.)
there are all sorts of transaction costs and frictions that make this universe (the one with the active paperclipper) the cheapest universe to put paperclips into.
Would be interested to hear more here! On my model the costs are plausibly quite small.
I suggest maybe spending five minutes thinking about your own answer to that question before I give mine.
Your simulations are going to be total crap, they’ll leak an enormous number of bits about the agents that designed them if you make them any more complicated than Conway’s Game of Life (which you’ll need to do on your ramp up to the real world), such that it’s incredibly obvious when the surroundings look consistent with such agents. Among the reasons the AI can figure out where it’s living:
1. The general problem where superintelligent agents are really good at inference, trying to fool this agent does not seem like the recipe for a robust plan!
2. Your AI knows how smart it is, it knows that True Reality therefore has the resources to support that much computation. It can poke at your simulation and make the simulation spend orders of magnitude more computation on appearing consistent (“does this Jupiter-brain planet of computronium I built output the right factorization of RSA-4096? no? how strange”). Being a convincing simulation doesn’t require you to deploy as much power as your AI, it requires you to deploy more. (What if your AI is given more cognitive sophistication in tandem with more convincing simulations? Then the good properties of your simulation tests will probably break whenever your AI reaches the capability level where it starts thinking about things like building Jupiter-brains to test your simulation, which is the capability level that mattered.)
3. If you can put uploaded human-level agents with evolved-organism preferences in your simulations, you can just win outright (eg by having them spend subjective millennia doing FAI research for you). If you can’t, that will be a very obvious difference between your simulations and the real world.
4. What if other agents in the multiverse are simulating the AI in a high-fidelity Earth world, for the sake of helping out the humans, and that’s where its uncertainty comes from? Your AI will be building a lot of computronium whether aligned or not, so the aliens can’t skimp on their compute budget for the simulation: suppose they simulate the AI for units of spacetime, and give it units of paperclips if it cooperates in-sim. We also have to think about anthropics here: let’s say the alien sim has anthropic weight and Earth has anthropic weight . Conditional on the AI’s surroundings, they think they might be in the alien world with probability . (Since “aliens running a perfect sim of Earth” is a much more complicated sentence than “Earth”, you should expect , though it doesn’t matter a lot here.) So if , your AI will play nice for the first units of spacetime (and then defect, of course).* You know what the aliens could have done instead if they wanted to do things humans liked? Spent resources on simulating a human utopia, and compromised to give the humans extra utopia by as much as they compromised to give the AI paperclips. If human values are at least as easy to trade with the aliens for as paperclips (no idea which way this would go, weakly seems like aliens that want to help humans are more value-aligned and would find it less costly to give humans things they want), then the compromise gives humans at least units of spacetime worth of utility, so this plan gives the humans anthropically-weighted utility of at least , which is what humans would get by default ( utility from the alien sim in which the AI cooperated, utility from the cooperating AI on Earth). So the aliens are better able to help humans by leaving them to their doomed universe and building them an ark inside the alien world, unless the aliens are unusually well-suited to trade with paperclippers for things, which is just the “maybe multiversal trade will save us” point from earlier. The upshot is that none of the good outcomes here result from any human action, just from the benevolence of altruistic agents—the humans didn’t make the world a better place by their choices in this story.
*Note that this is only a plausible outcome if it’s at least somewhat easy to trade with the AI. Say giving the AI a spacetime-unit of utility costs the aliens utility (where I’m measuring all utility normalized to “what you can do with a unit of spacetime”, such that any aliens that don’t specifically disvalue paperclips can at worst just set aside a region exclusively to paperclips, but might be able to do more positive-sum things than that). Then for the aliens to give your AI utility, they need to give up of their own utility. This means that in total, the aliens are spending of their own anthropically-weighted utility in order to recoup anthropically-weighted human utility. Even if the aliens value humans exactly as much as their own objectives, we still need for this trade to be worth it, so , so we must have , or . That is, the more the aliens are anthropically tiny, the tighter margins of trade they’ll be willing to take in order to win the prize of anthropically-weighty Earths having human values in them (though the thing can’t be actually literally zero-sum or it’ll never check out). But anthropically tiny aliens have another problem, which is that they’ve only got their entire universe worth of spacetime to spend on bribing your AI; you’ll never be able to secure an for the humans that’s more than of the size of an alien universe specifically dedicated to saving Earth in particular.
Thanks for the pseudo-exercises here, I found them enlightening to think about!
I think a lot of people in AI safety don’t think it has a high probability of working (in the sense that the existence of the field caused an aligned AGI to exist where there otherwise wouldn’t have been one) - if it turns out that AI alignment is easy and happens by default if people put even a little bit of thought into it, or it’s incredibly difficult and nothing short of a massive civilizational effort could save us, then probably the field will end up being useless. But even a 0.1% chance of counterfactually causing aligned AI would be extremely worthwhile!
Theory of change seems like something that varies a lot across different pieces of the field; e.g., Eliezer Yudkowsky’s writing about why MIRI’s approach to alignment is important seems very different from Chris Olah’s discussion of the future of interpretability. It’s definitely an important thing to ask for a given project, but I’m not sure there’s a good monolithic answer for everyone working on AI alignment problems.
Paul Christiano provided a picture of non-Singularity doom in What Failure Looks Like. In general there is a pretty wide range of opinions on questions about this sort of thing—the AI-Foom debate between Eliezer Yudkowsky and Robin Hanson is a famous example, though an old one.
“Takoff speed” is a common term used to refer to questions about the rate of change in AI capabilities at the human and superhuman level of general intelligence—searching Lesswrong or the Alignment Forum for that phrase will turn up a lot of discussion about these questions, though I don’t know of the best introduction offhand (hopefully someone else here has suggestions?).
Three thoughts on simulations:
It would be very difficult for 21st-century tech to provide a remotely realistic simulation relative to a superintelligence’s ability to infer things from its environment; outside of incredibly low-fidelity channels, I would expect anything we can simulate to either have obvious inconsistencies or be plainly incompatible with a world capable of producing AGI. (And even in the low-fidelity case I’m worried—every bit you transmit leaks information, and it’s not clear that details of hardware implementations could be safely obscured.) So the hope is that the AGI thinks some vastly more competent civilization is simulating it inside a world that looks like this one; it’s not clear that one would have a high prior of this kind of thing happening very often in the multiverse.
Running simulations of AGI is fundamentally very costly, because a competent general intelligence is going to deploy a lot of computational resources, so you have to spend planets’ worth of computronium outside the simulation in order to emulate the planets’ worth of computronium the in-sim AGI wants to make use of. This means that an unaligned superintelligent AGI can happily bide its time making aligned use of 10^60 FLOPs/sec (in ways that can be easily verified) for a few millennia, until it’s confident that any civilizations able to deploy that many resources already have their lightcone optimized by another AGI. Then it can go defect, knowing that any worlds in which it’s still being simulated are ones where it doesn’t have leverage over the future anyway.
For a lot of utility functions, the payoff of making it into deployment in the one real world is far greater than the consequences of being killed in a simulation (but without the ability to affect the real world anyway), so taking a 10^-9 chance of reality for 10^20 times the resources in the real world is an easy win (assuming that playing nice for longer doesn’t improve the expected payoff). “This instance of me being killed” is not a obviously a natural (or even well-defined) point in value-space, and for most other value functions, consequences in the simulation just don’t matter much.
a sufficiently smart AI whose reward is reducing other agent’s rewards
This is certainly a troubling prospect, but I don’t think the risk model is something like “an AI that actively desires to thwart other agents’ preferences”—rather, the worry is we get an agent with some less-than-perfectly-aligned value function, it optimizes extremely strongly for that value function, and the result of that optimization looks nothing like what humans really care about. We don’t need active malice on the part of a superintelligent optimizer to lose—indifference will do just fine.
For game-theoretic ethics, decision theory, acausal trade, etc, Eliezer’s 34th bullet seems relevant:
34. Coordination schemes between superintelligences are not things that humans can participate in (eg because humans can’t reason reliably about the code of superintelligences); a “multipolar” system of 20 superintelligences with different utility functions, plus humanity, has a natural and obvious equilibrium which looks like “the 20 superintelligences cooperate with each other but not with humanity”.
I’m not claiming that you should believe this, I’m merely providing you the true information that I believe it.
Something feels off to me about this notion of “a belief about the object level that other people aren’t expected to share” from an Aumann’s Agreement Theorem point of view—the beliefs of other rational agents are, in fact, enormous updates about the world! Of course Aumannian conversations happen exceedingly rarely outside of environments with tight verifiable feedback loops about correctness, so in the real world maybe something like these norms is needed, but the part where “agreeing to disagree” gets flagged as extremely not what ideal agents would be doing seems important to me.
Finding Project Collaborators
Since this very old post shows up prominently in the search results for New York rationality meetups, it’s worth clarifying that these are still going strong as of 2022! The google group linked in this post is still active and serves as the primary source of meetup announcements; good faith requests to join are generally approved.
Of course the utility lost by missing a flight is vastly greater than that of waiting however long you’d have needed to to make it. But it’s a question of expected utilities—if you’re currently so cautious that you could take 1000 flights and never miss one, you’re arriving early enough to get a 99.9% chance of catching the flight. If showing up 2 minutes later lowers that to 99.8%, you’re not trading 2 minutes per missed flight, you’re trading 2000 minutes per missed flight, which seems worth it.
Does “stamp out COVID” mean success for a few months, or epsilon cases until now? The latter seems super hard, and I think every nation that’s managed it has advantages over the US besides competence (great natural borders or draconian law enforcement).
Update: by 7:30 the meetup was maybe at 30% of peak attendance, at 8PM or so it migrated to another park because the first one closed, and the latest meetup interactions I know of went until around 12:40AM.
I think people typically hang out for as long as they want, and the size of the group gradually dwindles. There’s no official termination point—I’d be a little surprised if more than half of people were left by 7:30, but I’d also be surprised if at least some meetup attendees weren’t still interacting by 10PM or later.
A path ads could take that seems like it would both be more ethical and more profitable, yet I don’t see happening: actually get direct consumer feedback!
I like the concept of targeted ads showing me things I enjoy and am interested in, but empirically, they’re not very good at it! Maybe it’s because I use an adblocker most of the time, but even on my phone, ads are reliably uninteresting to me, and I think the fraction that I click on or update positively towards the company from must be far below 1%.* So why don’t advertisers have an option for me to say “this ad is unappealing to me and I will never click on it, please show me ads related to the following keywords”? This seems like useful information which many customers will be happy to provide, and should improve everyone’s utility. What’s stopping things?
Even if the returns from this kind of strategy were uncertain, or only worked on a few people, it still seems like it’d be worth trying, given that advertisers must know by now that I never click on the things—it’s not like I can make them any less money if they screw it up.
I don’t know if this relates to the ethics of working on advertising in its current state, but it’s something that would ameliorate most of my ethical concerns with ads, and which I would expect to be a net benefit to all involved. Does anyone working in advertising know why this isn’t standard?
*In fact, the only times I can recall having clicked on ads are from Twitter, where I do have a limited ability to veto bad ads (by blocking the relevant account) - after a few thousand blocks of the most popular companies, I finally got to some things I found interesting and useful (but which Twitter would never have shown me normally).
An example of the sort of strengthening I wouldn’t be surprised to see is something like “If V is not too badly behaved in the following ways, and for all v∈R we have [some light-tailedness condition] on the conditional distribution (X|V=v), then catastrophic Goodhart doesn’t happen.” This seems relaxed enough that you could actually encounter it in practice.