I am denying that superintelligences play this game in a way that looks like “Pick an ordinal to be your level of sophistication, and whoever picks the higher ordinal gets $9.” I expect sufficiently smart agents to play this game in a way that doesn’t incentivize attempts by the opponent to be more sophisticated than you, nor will you find yourself incentivized to try to exploit an opponent by being more sophisticated than them, provided that both parties have the minimum level of sophistication to be that smart.
If faced with an opponent stupid enough to play the ordinal game, of course, you just refuse all offers less than $9, and they find that there’s no ordinal level of sophistication they can pick which makes you behave otherwise. Sucks to be them!
You have misunderstood (1) the point this post was trying to communicate and (2) the structure of the larger argument where that point appears, as follows:
First, let’s talk about (2), the larger argument that this post’s point was supposed to be relevant to.
Is the larger argument that superintelligences will misunderstand what we really meant, due to a lack of knowledge about humans?
It is incredibly unlikely that Eliezer Yudkowsky in particular would have constructed an argument like this, whether in 2007, 2017, or even 1997. At all of these points in my life, I visibly held quite a lot of respect for the epistemic prowess of superintelligences. They were always going to know everything relevant about the complexities of human preference and desire. The larger argument is about whether it’s easy to make superintelligences end up caring.
This post isn’t about the distinction between knowing and caring, to be clear; that’s something I tried to cover elsewhere. The relevant central divide falls in roughly the same conceptual place as Hume’s Guillotine between ‘is’ and ‘ought’, or the difference between the belief function and the utility function.
(I don’t see myself as having managed to reliably communicate this concept (though the central idea is old indeed within philosophy) to the field that now sometimes calls itself “AI alignment”; so if you understand this distinction yourself, you should not assume that any particulary commentary within “AI alignment” is written from a place of understanding it too.)
What this post is about is the amount of information-theoretic complexity that you need to get into the system’s preferences, in order to have that system, given unlimited or rather extremely large amounts of power, deliver to you what you want.
It doesn’t argue that superintelligences will not know this information. You’ll note that the central technology in the parable isn’t an AI; it’s an Outcome Pump.
What it says, rather, is that there might be, say, a few tens of thousands of bits—the exact number is not easy to estimate, we just need to know that it’s more than a hundred bits and less than a billion bits and anything in that range is approximately the same problem from our standpoint—that you need to get into the steering function. If you understand the Central Divide that Hume’s Razor points to, the distinction between probability and preference, etcetera, the post is trying to establish the idea that we need to get 13,333 bits or whatever into the second side of this divide.
In terms of where this point falls within the larger argument, this post is not saying that it’s particularly difficult to get those 13,333 bits into the preference function; for all this post tries to say, locally, maybe that’s as easy as having humans manually enter 13,333 yes-or-no answers into the system. It’s not talking about the difficulty of doing the work but rather the amount and nature of a kind of work that needs to be done somehow.
Definitely, the post does not say that it’s hard to get those 13,333 bits into the belief function or knowledge of a superintelligence.
Separately from understanding correctly what this post is trying to communicate, at all, in 2007, there’s the question of whether modern LLMs have anything to say about—obviously not the post’s original point—but rather, other steps of the larger question in which this post’s point appears.
Modern LLMs, if you present them with a text-based story like the one in this parable, are able to answer at least some text-based questions about whether you’d prefer your grandmother to be outside the building or be safely outside the building. Let’s admit this premised observation at face value. Have we learned thereby the conclusion that it’s easy to get all of that information into a superintelligence’s preference function?
And if we say “No”, is this Eliezer making up post-hoc excuses?
What exactly we learn from the evidence of how AI has played out in 2024 so far, is the sort of thing that deserves its own post. But I observe that if you’d asked Eliezer-2007 whether an (Earth-originating) superintelligence could correctly predict the human response pattern about what to do with the grandmother—solve the same task LLMs are solving, to at least the LLM’s performance level—Eliezer-2007 would have unhesitatingly answered “yes” and indeed “OBVIOUSLY yes”.
How is this coherent? Because the post’s point is about how much information needs to get into the preference function. To predict a human response pattern you need (only) epistemic knowledge. This is part of why the post is about needing to give specifications to an Outcome Pump, rather than it depicting an AI being surprised by its continually incorrect predictions about a human response pattern.
If you don’t see any important distinction between the two, then of course you’ll think that it’s incoherent to talk about that distinction. But even if you think that Hume was mistaken about there existing any sort of interesting gap between ‘is’ and ‘ought’, you might by some act of empathy be able to imagine that other people think there’s an interesting subject matter there, and they are trying to talk about it with you; otherwise you will just flatly misunderstand what they were trying to say, and mispredict their future utterances. There’s a difference between disagreeing with a point, and just flatly failing to get it, and hopefully you aspire to the first state of mind rather than the second.
Have we learned anything stunningly hopeful from modern pre-AGIs getting down part of the epistemic part of the problem at their current ability levels, to the kind of resolution that this post talked about in 2007? Or from it being possible to cajole pre-AGIs with loss functions into willingly using that knowledge to predict human text outputs? Some people think that this teaches us that alignment is hugely easy. I think they are mistaken, but that would take its own post to talk about.
But people who point to “The Hidden Complexity of Wishes” and say of it that it shows that I had a view which the current evidence already falsifies—that I predicted that no AGI would ever be able to predict human response patterns about getting grandmothers out of burning buildings—have simply: misunderstood what the post is about, not understood in particular why the post is about an Outcome Pump rather than an AI stupidly mispredicting human responses, and failed to pick up on the central point that Eliezer expects superintelligences to be smart in the sense of making excellent purely epistemic predictions.
This deserves a longer answer than I have time to allocate it, but I quickly remark that I don’t recognize the philosophy or paradigm of updatelessness as refusing to learn things or being terrified of information; a rational agent should never end up in that circumstance, unless some perverse other agent is specifically punishing them for having learned the information (and will lose of their own value thereby; it shouldn’t be possible for them to gain value by behaving “perversely” in that way, for then of course it’s not “perverse”). Updatelessness is, indeed, exactly that sort of thinking which prevents you from being harmed by information, because your updateless exposure to information doesn’t cause you to lose coordination with your counterfactual other selves or exhibit dynamic inconsistency with your past self.
From an updateless standpoint, “learning” is just the process of reacting to new information the way your past self would want you to do in that branch of possibility-space; you should never need to remain ignorant of anything. Maybe that involves not doing the thing that would then be optimal when considering only the branch of reality you turned out to be inside, but the updateless mind denies that this was ever the principle of rational choice, and so feels no need to stay ignorant in order to maintain dynamic consistency.
They can solve it however they like, once they’re past the point of expecting things to work that sometimes don’t work. I have guesses but any group that still needs my hints should wait and augment harder.
I disagree with my characterization as thinking problems can be solved on paper, and with the name “Poet”. I think the problems can’t be solved by twiddling systems weak enough to be passively safe, and hoping their behavior generalizes up to dangerous levels. I don’t think paper solutions will work either, and humanity needs to back off and augment intelligence before proceeding. I do not take the position that we need a global shutdown of this research field because I think that guessing stuff without trying it is easy, but because guessing it even with some safe weak lesser tries is still impossibly hard. My message to humanity is “back off and augment” not “back off and solve it with a clever theory”.
Not what comes up for me, when I go incognito and google AI risk lesswrong.
I rather expect that existing robotic machinery could be controlled by ASI rather than “moderately smart intelligence” into picking up the pieces of a world economy after it collapses, or that if for some weird reason it was trying to play around with static-cling spaghetti It could pick up the pieces of the economy that way too.
It’s false that currently existing robotic machinery controlled by moderately smart intelligence can pick up the pieces of a world economy after it collapses. One well-directed algae cell could, but not existing robots controlled by moderate intelligence.
What does this operationalize as? Presumably not that if we load a bone and a diamond rod under equal pressures, the diamond rod breaks first? Is it more about if we drop sudden sharp weights onto a bone rod and a diamond rod, the diamond rod breaks first? I admit I hadn’t expected that, despite a general notion that diamond is crystal and crystals are unexpectedly fragile against particular kinds of hits, and if so that modifies my sense of what’s a valid metaphor to use.
“Pandemics” aren’t a locally valid substitute step in my own larger argument, because an ASI needs its own manufacturing infrastructure before it makes sense for the ASI to kill the humans currently keeping its computers turned on. So things that kill a bunch of humans are not a valid substitute for being able to, eg, take over and repurpose the existing solar-powered micron-diameter self-replicating factory systems, aka algae, and those repurposed algae being able to build enough computing substrate to go on running the ASI after the humans die.
It’s possible this argument can and should be carried without talking about the level above biology, but I’m nervous that this causes people to start thinking in terms of Hollywood movie plots about defeating pandemics and hunting down the AI’s hidden cave of shoggoths, rather than hearing, “And this is a lower bound but actually in real life you just fall over dead.”
Why is flesh weaker than diamond? Diamond is made of carbon-carbon bonds. Proteins also have some carbon-carbon bonds! So why should a diamond blade be able to cut skin?
I reply: Because the strength of the material is determined by its weakest link, not its strongest link. A structure of steel beams held together at the vertices by Scotch tape (and lacking other clever arrangements of mechanical advantage) has the strength of Scotch tape rather than the strength of steel.
Or: Even when the load-bearing forces holding large molecular systems together are locally covalent bonds, as in lignin (what makes wood strong), if you’ve got larger molecules only held together by covalent bonds at interspersed points along their edges, that’s like having 10cm-diameter steel beams held together by 1cm welds. Again, barring other clever arrangements of mechanical advantage, that structure has the strength of 1cm of steel rather than 10cm of steel.
Bone is stronger than wood; it runs on a relatively stronger structure of ionic bonds, which are no locally weaker than carbon bonds in terms of attojoules of potential energy per bond. Bone is weaker than diamond, then, because… why?
Well, partially, IIUC, because calcium atoms are heavier than carbon atoms. So even if per-bond the ionic forces are strong, some of that is lost in the price you pay for including heavier atoms whose nuclei have more protons that are able to exert the stronger electrical forces making up that stronger bond.
But mainly, bone is so much weaker than diamond (on my understanding) because the carbon bonds in diamond have a regular crystal structure that locks the carbon atoms into relative angles, and in a solid diamond this crystal structure is tesselated globally. Hydroxyapatite (the crystal part of bone) also tesselates in an energetically favorable configuration; but (I could be wrong about this) it doesn’t have the same local resistance to local deformation; and also, the actual hydroxyapatite crystal is assembled by other tissues that layer the ionic components into place, which means that a larger structure of bone is full of fault lines. Bone cleaves along the weaker fault line, not at its strongest point.
But then, why don’t diamond bones exist already? Not just for the added strength; why make the organism look for calcium and phosphorus instead of just carbon?
The search process of evolutionary biology is not the search of engineering; natural selection can only access designs via pathways of incremental mutations that are locally advantageous, not intelligently designed simultaneous changes that compensate for each other. There were, last time I checked, only three known cases where evolutionary biology invented the freely rotating wheel. Two of those known cases are ATP synthase and the bacterial flagellum, which demonstrates that freely rotating wheels are in fact incredibly useful in biology, and are conserved when biology stumbles across them after a few hundred million years of search. But there’s no use for a freely rotating wheel without a bearing and there’s no use for a bearing without a freely rotating wheel, and a simultaneous dependency like that is a huge obstacle to biology, even though it’s a hardly noticeable obstacle to intelligent engineering.
The entire human body, faced with a strong impact like being gored by a rhinocerous horn, will fail at its weakest point, not its strongest point. How much evolutionary advantage is there to stronger bone, if what fails first is torn muscle? How much advantage is there to an impact-resistant kidney, if most fights that destroy a kidney will kill you anyways? Evolution is not the sort of optimizer that says, “Okay, let’s design an entire stronger body.” (Analogously, the collection of faults that add up to “old age” is large enough that a little more age resistance in one place is not much of an advantage if other aging systems or outward accidents will soon kill you anyways.)
I don’t even think we have much of a reason to believe that it’d be physically (rather than informationally) difficult to have a set of enzymes that synthesize diamond. It could just require 3 things to go right simultaneously, and so be much much harder to stumble across than tossing more hydroxyapatite to lock into place in a bone crystal. And then even if somehow evolution hit on the right set of 3 simultaneous mutations, sometime over the history of Earth, the resulting little isolated chunk of diamond probably would not be somewhere in the phenotype that had previously constituted the weakest point in a mechanical system that frequently failed. If evolution has huge difficulty inventing wheels, why expect that it could build diamond chainmail, even assuming that diamond chainmail is physically possible and could be useful to an organism that had it?
Talking to the general public is hard. The first concept I’m trying to convey to them is that there’s an underlying physical, mechanical reason that flesh is weaker than diamond; and that this reason isn’t that things animated by vitalic spirit, elan vital, can self-heal and self-reproduce at the cost of being weaker than the cold steel making up lifeless machines, as is the price of magic imposed by the universe to maintain game balance. This is a very natural way for humans to think; and the thing I am trying to come in and do is say, “Actually, no, it’s not a mystical balance, it’s that diamond is held together by bonds that are hundreds of kJ/mol; and the mechanical strength of proteins is determined by forces a hundred times as weak as that, the part where proteins fold up like spaghetti held together by static cling.”
There is then a deeper story that’s even harder to explain, about why evolution doesn’t build freely rotating wheels or diamond chainmail; why evolutionary design doesn’t find the physically possible stronger systems. But first you need to give people a mechanical intuition for why, in a very rough intuitive sense, it is physically possible to have stuff that moves and lives and self-repairs but is strong like diamond instead of flesh, without this violating a mystical balance where the price of vitalic animation is lower material strength.
And that mechanical intuition is: Deep down is a bunch of stuff that, if you could see videos of it, would look more like tiny machines than like magic, though they would not look like familiar machines (very few freely rotating wheels). Then why aren’t these machines strong like human machines of steel are strong? Because iron atoms are stronger than carbon atoms? Actually no, diamond is made of carbon and that’s still quite strong. The reason is that these tiny systems of machinery are held together (at the weakest joints, not the strongest joints!) by static cling.
And then the deeper question: Why does evolution build that way? And the deeper answer: Because everything evolution builds is arrived at as an error, a mutation, from something else that it builds. Very tight bonds fold up along very deterministic pathways. So (in the average case, not every case) the neighborhood of functionally similar designs is densely connected along shallow energy gradients and sparsely connected along deep energy gradients. Intelligence can leap long distances through that design space using coordinated changes, but evolutionary exploration usually cannot.
And I do try to explain that too. But it is legitimately more abstract and harder to understand. So I lead with the idea that proteins are held together by static cling. This is, I think, validly the first fact you lead with if the audience does not already know it, and just has no clue why anyone could possibly possibly think that there might even be machinery that does what bacterial machinery does but better. The typical audience is not starting out with the intuition that one would naively think that of course you could put together stronger molecular machinery, given the physics of stronger bonds, and then we debate whether (as I believe) the naive intuition is actually just valid and correct; they don’t understand what the naive intuition is about, and that’s the first thing to convey.
If somebody then says, “How can you be so ignorant of chemistry? Some atoms in protein are held together by covalent bonds, not by static cling! There’s even eg sulfur bonds whereby some parts of the folded-spaghetti systems end up glued together with real glue!” then this does not validly address the original point because: the underlying point about why flesh is more easily cleaved than diamond, is about the weakest points of flesh rather than the strongest points in flesh, because that’s what determines the mechanical strength of the larger system.
I think there is an important way of looking at questions like these where, at the final end, you ask yourself, “Okay, but does my argument prove that flesh is in fact as strong as diamond? Why isn’t flesh as strong as diamond, then, if I’ve refuted the original argument for why it isn’t?” and this is the question that leads you to realize that some local strong covalent bonds don’t matter to the argument if those bonds aren’t the parts that break under load.
My main moral qualm about using the Argument From Folded Spaghetti Held Together By Static Cling as an intuition pump is that the local ionic bonds in bone are legitimately as strong per-bond as the C-C bonds in diamond, and the reason that bone is weaker than diamond is (iiuc) actually more about irregularity, fault lines, and resistance to local deformation than about kJ/mol of the underlying bonds. If somebody says “Okay, fine, you’ve validly explained why flesh is weaker than diamond, but why is bone weaker than diamond?” I have to reply “Valid, iiuc that’s legit more about irregularity and fault lines and interlaced weaker superstructure and local deformation resistance of the bonds, rather than the raw potential energy deltas of the load-bearing welds.”
Depends on how much of a superintelligence, how implemented. I wouldn’t be surprised if somebody got far superhuman theorem-proving from a mind that didn’t generalize beyond theorems. Presuming you were asking it to prove old-school fancy-math theorems, and not to, eg, arbitrarily speed up a bunch of real-world computations like asking it what GPT-4 would say about things, etc.
Solution (in retrospect this should’ve been posted a few years earlier):
let’Na’ = box N contains angry frog’Ng’ = N gold’Nf’ = N’s inscription false’Nt’ = N’s inscription trueconsistent states must have 1f 2t or 1t 2f, and 1a 2g or 1g 2athen:1a 1t, 2g 2f ⇒ 1t, 2f1a 1f, 2g 2t ⇒ 1f, 2t1g 1t, 2a 2f ⇒ 1t, 2t1g 1f, 2a 2t ⇒ 1f, 2f
I currently guess that a research community of non-upgraded alignment researchers with a hundred years to work, picks out a plausible-sounding non-solution and kills everyone at the end of the hundred years.
I don’t think that faster alignment researchers get you to victory, but uploading should also allow for upgrading and while that part is not trivial I expect it to work.
AI happening through deep learning at all is a huge update against alignment success, because deep learning is incredibly opaque. LLMs possibly ending up at the center is a small update in favor of alignment success, because it means we might (through some clever sleight, this part is not trivial) be able to have humanese sentences play an inextricable role at the center of thought (hence MIRI’s early interest in the Visible Thoughts Project).
The part where LLMs are to predict English answers to some English questions about values, and show common-sense relative to their linguistic shadow of the environment as it was presented to them by humans within an Internet corpus, is not actually very much hope because a sane approach doesn’t involve trying to promote an LLM’s predictive model of human discourse about morality to be in charge of a superintelligence’s dominion of the galaxy. What you would like to promote to values are concepts like “corrigibility”, eg “low impact” or “soft optimization”, which aren’t part of everyday human life and aren’t in the training set because humans do not have those values.
I have never since 1996 thought that it would be hard to get superintelligences to accurately model reality with respect to problems as simple as “predict what a human will thumbs-up or thumbs-down”. The theoretical distinction between producing epistemic rationality (theoretically straightforward) and shaping preference (theoretically hard) is present in my mind at every moment that I am talking about these issues; it is to me a central divide of my ontology.
If you think you’ve demonstrated by clever textual close reading that Eliezer-2018 or Eliezer-2008 thought that it would be hard to get a superintelligence to understand humans, you have arrived at a contradiction and need to back up and start over.
The argument we are trying to explain has an additional step that you’re missing. You think that we are pointing to the hidden complexity of wishes in order to establish in one step that it would therefore be hard to get an AI to output a correct wish shape, because the wishes are complex, so it would be difficult to get an AI to predict them. This is not what we are trying to say. We are trying to say that because wishes have a lot of hidden complexity, the thing you are trying to get into the AI’s preferences has a lot of hidden complexity. This makes the nonstraightforward and shaky problem of getting a thing into the AI’s preferences, be harder and more dangerous than if we were just trying to get a single information-theoretic bit in there. Getting a shape into the AI’s preferences is different from getting it into the AI’s predictive model. MIRI is always in every instance talking about the first thing and not the second.
You obviously need to get a thing into the AI at all, in order to get it into the preferences, but getting it into the AI’s predictive model is not sufficient. It helps, but only in the same sense that having low-friction smooth ball-bearings would help in building a perpetual motion machine; the low-friction ball-bearings are not the main problem, they are a kind of thing it is much easier to make progress on compared to the main problem. Even if, in fact, the ball-bearings would legitimately be part of the mechanism if you could build one! Making lots of progress on smoother, lower-friction ball-bearings is even so not the sort of thing that should cause you to become much more hopeful about the perpetual motion machine. It is on the wrong side of a theoretical divide between what is straightforward and what is not.
You will probably protest that we phrased our argument badly relative to the sort of thing that you could only possibly be expected to hear, from your perspective. If so this is not surprising, because explaining things is very hard. Especially when everyone in the audience comes in with a different set of preconceptions and a different internal language about this nonstandardized topic. But mostly, explaining this thing is hard and I tried taking lots of different angles on trying to get the idea across.
In modern times, and earlier, it is of course very hard for ML folk to get their AI to make completely accurate predictions about human behavior. They have to work very hard and put a lot of sweat into getting more accurate predictions out! When we try to say that this is on the shallow end of a shallow-deep theoretical divide (corresponding to Hume’s Razor) it often sounds to them like their hard work is being devalued and we could not possibly understand how hard it is to get an AI to make good predictions.
Now that GPT-4 is making surprisingly good predictions, they feel they have learned something very surprising and shocking! They cannot possibly hear our words when we say that this is still on the shallow end of a shallow-deep theoretical divide! They think we are refusing to come to grips with this surprising shocking thing and that it surely ought to overturn all of our old theories; which were, yes, phrased and taught in a time before GPT-4 was around, and therefore do not in fact carefully emphasize at every point of teaching how in principle a superintelligence would of course have no trouble predicting human text outputs. We did not expect GPT-4 to happen, in fact, intermediate trajectories are harder to predict than endpoints, so we did not carefully phrase all our explanations in a way that would make them hard to misinterpret after GPT-4 came around.
But if you had asked us back then if a superintelligence would automatically be very good at predicting human text outputs, I guarantee we would have said yes. You could then have asked us in a shocked tone how this could possibly square up with the notion of “the hidden complexity of wishes” and we could have explained that part in advance. Alas, nobody actually predicted GPT-4 so we do not have that advance disclaimer down in that format. But it is not a case where we are just failing to process the collision between two parts of our belief system; it actually remains quite straightforward theoretically. I wish that all of these past conversations were archived to a common place, so that I could search and show you many pieces of text which would talk about this critical divide between prediction and preference (as I would now term it) and how I did in fact expect superintelligences to be able to predict things!
There’s perhaps more detail in Project Lawful and in some nearby stories (“for no laid course prepare”, “aviation is the most dangerous routine activity”).
Have you ever seen or even heard of a person who is obese who doesn’t eat hyperpalatable foods? (That is, they only eat naturally tasting, unprocessed, “healthy” foods).
Tried this for many years. Paleo diet; eating mainly broccoli and turkey; trying to get most of my calories from giant salads. Nothing.