Now, consider the following simplistic model for naive (un)aligned AGI:
The AGI outputs English sentences. Each time the AGI does, the human operator replies on a scale of 1 to 100 with how good and valuable and useful that sentence seemed to the human. The human may also input other sentences to the AGI as a hint about what kind of output the human is currently looking for; and the AGI also has purely passive sensory inputs like a fixed webcam stream or a pregathered internet archive.
How does this fail as an alignment methodology? Doesn't this fit very neatly into the existing prosaic methodology of reinforcement learning? Wouldn't it be very useful on hand to have an intelligence which gives us much more valuable sentences, in response to input, than the sentences that would be generated by a human?
There's a number of general or generic ways to fail here that aren't specific to the supposed criterion of the reinforcement learning system, like the AGI ending up with other internal goals and successfully forging a Wifi signal via internal RAM modulation etcetera, but let's consider failures that are in some sense intrinsic to this paradigm even if the internal AI ends up perfectly aligned on that goal. Let's even skip over the sense in which we've given the AI a long-term incentive to accept some lower rewards in the short term, in order to grab control of the rating button, if the AGI ends up with long-term consequentialist preferences and long-term planning abilities that exactly reflect the outer fitness function. Let's say that the AGI is only shortsightedly trying to maximize sentence reward on each round and that it is not superintelligent enough to disintermediate the human operators and grab control of the button in one round without multi-round planning. Some of us might be a bit skeptical about whether you can be, effectively, very much smarter than a human, in the first place, without doing some kind of multi-round or long-term internal planning about how to think about things and allocate internal resources; but fine, maybe there's just so many GPUs running the system that it can do all of its thinking, for each round, on that round. What intrinsically goes wrong?
What intrinsically goes wrong, I'd say, is that the human operators have an ability to recognize good arguments that's only rated to withstand up to a certain intensity of search, which will break down beyond that point. Our brains' ability to distinguish good arguments from bad arguments is something we'd expect to be balanced to the kind of argumentative pressure a human brain was presented with in the ancestral environment / environment of evolutionary adaptedness, and if you optimize against a brain much harder than this, you'd expect it to break. There'd be an arms race between politicians exploiting brain features to persuade people of things that were useful to the politician, and brains that were, among other things, trying to pursue the original 'reasons' for reasoning that originally and initially made it useful to recognize certain arguments as good arguments before any politicians were trying to exploit them. Again, oversimplified, and there are cases where it's not tribally good for you to be the only person who sees a politician's lie as a lie; but the broader point is that there's going to exist an ecological balance in the ancestral environment between brains trying to persuade other brains, and brains trying to do locally fitness-enhancing cognition while listening to persuaders; and this balance is going to be tuned to the level of search power that politicians had in the environment of evolutionary adaptedness.
Arguably, this is one way of viewing the flood of modern insanity at which social media seems to be the center. For the same reason pandemics get more virulent with larger and more dense population centers, Twitter may be selecting for memes that break humans at a much higher level of optimization pressure than held in the ancestral environment or even just 1000 years earlier than this.
Viewed through the lens of Goodhart's Curse: When you have an imperfect proxy U for an underlying value V, the highest values of U will represent the places where U diverges upward the most from V and not just the highest underlying values of V. The harder you search for high values of U, the wider the space of possibilities you search, the more that the highest values of U will diverge upwards from V.
### incorporated from a work in progress
Suppose that, in the earlier days of the Web, you're trying to find webpages with omelet recipes. You have the stunning insight that webpages with omelet recipes often contain the word "omelet" somewhere in them. So you build a search engine that travels URLs to crawl as much of the Web as you can find, indexing all the pages by the words they contain; and then you search for the "omelet" keyword. Works great the first time you try it! Maybe some of the pages are quoting "You can't make an omelet without breaking eggs" (said by Robespierre, allegedly), but enough pages have actual omelet recipes that you can find them by scrolling down. Better yet, assume that pages that contain the "omelet" keyword more often are more likely to be about omelets. Then you're fine... in the first iteration of the game.
But the thing is: the easily computer-measurable fact of whether a page contains the "omelet" keyword is not identical to the fact of whether it has the tasty omelet recipe you seek. V, the true value, is whether a page has a tasty recipe for omelets; U, the proxy measure, is how often the page mentions the "omelet" keyword. That some pages are compendiums of quotes from French revolutionaries, instead of omelet recipes, illustrates that U and V can diverge even in the natural ecology.
But once the search engine is built, we are not just listing possible pages and their U-measure at random; we are picking pages with the highest U-measure we can see. If we name the divergence D = U-V then we can say u_i = v_i + d_i. This helps illustrate that by selecting for the highest u_i we can find, we are putting upward selection pressure on both v_i and d_i. We are implicitly searching out, first, the underlying quality V that U is a proxy for, and second, places where U diverges far upward from V, that is, places where the proxy measure breaks down.
If we are living in an unrealistically smooth world where V and D are independent Gaussian distributions with mean 0 and variance 1, then the mean and variance of U is just 0 and 2 (the sum of the means and variances of V and D). If we randomly select an element with u_i=3, then on average it has v_i of 1.5 and d_i of 1.5. If the variance of V is 1 and the variance of D is 10 - if the "noise" from V to U varies much more widely on average than V itself - then most of the height of a high-U item probably comes from a lot of upward noise. But not all of it. On average, if you pick out an element with u_i = 11, it has expected d_i of 10 and v_i of 1; its apparent greatness is mostly noise. But still, the expected v_i is 1, not the average V of 0. The best-looking things are still, in expectation, better than average. They are just not as good as they look.
Ah, but what if everything isn't all Gaussian distributions? What if there are some regions of the space where D has much higher variance - places where U is much more prone to error as a proxy measure of V? Then selecting for high U tends to steer us to regions of possibility space where U is most mistaken as a measure of V.
And in nonsimple domains, the wider the region of possibility we search, the more likely this is to happen; the more likely it is that some part of the possibility space contains a place where U is a bad proxy for V.
This is an abstract (and widely generalizable) way of seeing the Fall of Altavista. In the beginning, the programmers noticed that naturally occurring webpages containing the word "omelet" were more likely to be about omelets. It is very hard to measure whether a webpage contains a good, coherent, well-written, tasty omelet recipe (what users actually want), but very easy to measure how often a webpage mentions the word "omelet". And the two facts did seem to correlate (webpages about dogs usually didn't mention omelets at all). So Altavista built a search engine accordingly.
But if you imagine the full space of all possible text pages, the ones that mention "omelet" most often are not pages with omelet recipes. They are pages with lots of sections that just say "omelet omelet omelet" over and over. In the natural ecology these webpages did not, at first, exist to be indexed! It doesn't matter that possibility-space is uncorrelated in principle, if we're only searching an actuality-space where things are in fact correlated.
But once lots of people started using (purely) keyword-based searches for webpages, and frequently searching for "omelet", spammers had an incentive to reshape their Viagra sales pages to contain "omelet omelet omelet" paragraphs.
That is: Once there was an economic incentive for somebody to make the search engine return a different result, the spammers began to intelligently search for ways to make U return a high result, and this implicitly meant putting the U-V correlation to a vastly stronger test. People naturally making webpages had not previously generated lots of webpages that said "omelet omelet omelet Viagra". U looked well-correlated with V in the region of textual possibility space that corresponded to the Web's status quo ante. But when an intelligent spammer imagines a way to try to steer users to their webpage, their imagination is searching through all the kinds of possible webpages they can imagine constructing; they are searching for imaginable places where U-V is very high and not just previously existing places where U-V is high. This means searching a much wider region of possibility for any place where U-V breaks down (or rather, breaks upward) which is why U is being put to a much sterner test.
We can also see issues in computer security from a similar perspective: Regularities that are obseved in narrow possibility spaces often break down in wider regions of the possibility space that can be searched by intelligent optimization. Consider how weird a buffer overflow attack would look, relative to a more "natural" ecology of program execution traces produced by non-malicious actors. Not only does the buffer overflow attack involve an unnaturally huge input, it's a huge input that overwrites the stack return address in a way that improbably happens to go to one of the most effectual possible destinations. A buffer overflow that results in root privilege escalation might not happen by accident inside a vulnerable system even once before the end of the universe. But an intelligent attacker doesn't search the space of only things that have already happened, they use their intelligence to search the much wider region of things that they can imagine happening. It says very little about the security of a computer system to say that, on average over the lifetime of the universe, it will never once yield up protected data in response to random inputs or in response to inputs typical of the previously observed distribution.
And the smarter the attacker, the wider the space of system execution traces it can effectively search. Very sophisticated attacks can look like "magic" in the sense that they exploit regularities you didn't realize existed! As an example in computer security, consider the Rowhammer attack, where repeatedly writing to unprotected RAM causes a nearby protected bit to flip. This violates what you might have thought were the basic axioms governing the computer's transistors. If you didn't know the trick behind Rowhammer, somebody could show you the code for the attack, and you just wouldn't see any advance reason why that code would succeed. You would not predict in advance that this weird code would successfully get root privileges, given all the phenomena inside the computer that you currently know about. This is "magic" in the same way that an air conditioner is magic in 1000AD. It's not just that the medieval scholar hasn't yet imagined the air conditioner. Even if you showed them the blueprint for the air conditioner, they wouldn't see any advance reason to predict that the weird contraption would output cold air. The blueprint is exploiting regularities like the pressure-temperature relationship that they haven't yet figured out.
To rephrase back into terms of Goodhart's Law as originally said by Goodhart - "Any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes" - statistical regularities that previously didn't break down in the face of lesser control pressures, can break down in the face of stronger control pressures that effectively search a wider range of possibilities, including possibilities that obey rules you didn't know were rules. This is more likely to happen the more complicated and rich and poorly understood the system is...
### end of quote
...which is how we can be nearly certain, even in advance of knowing the exact method, that a sufficiently strong search against a rating output by a complicated rich poorly-understood human brain will break that brain in ways that we can't even understand.
Even if everything goes exactly as planned on an internal level inside the AGI, which in real life is at least 90% of the difficulty, the outer control structure of the High-Rated Sentence Producer is something that, on the face of it, learns to break the operator. The fact that it's producing sentences more highly rated than a human inside the same box, the very fact that makes the High-Rated Sentence Producer possibly be useful in the first place, implies that it's searching harder against the rating criterion than a human does. Human ratings are imperfect proxies for validity, accuracy, estimated expectation of true value produced by a policy, etcetera. Human brains are rich and complicated and poorly understood. Such integrity as they possess is probably nearly in balance with the ecological expectation of encountering persuasive invalid arguments produced by other human-level intelligences. We should expect with very high probability that if HRSP searches hard enough against the rating, it will break the brain behind it.
I’m reasonably compelled by Sperber and Mercer’s claim that explicit reasoning in humans primarily evolved not in order to help us find out about the world, but rather in order to win arguments.
Seems obviously false. If we simplistically imagine humans as being swayed by, and separately arguing, an increasingly sophisticated series of argument types that we could label 0, 1, 2, ...N, N+1, and which are all each encoded in a single allele that somehow arose to fixation, then the capacity to initially recognize and be swayed by a type N+1 argument is a disadvantage when it comes to winning a type N argument using internal sympathy with the audience's viewpoint, because when that mutation happens for the first time, the other people in the tribe will not find N+1-type arguments compelling, and you do, which leads you to make intuitive mistakes about what they will find compelling. Only after the capacity to recognize type N+1 arguments as good arguments becomes pervasive in other listeners, does the ability to search for type-N+1 arguments congruent to some particular political or selfish purpose, become a fitness advantage. Even if we have underlying capabilities to automatically search for political/selfish arguments of all types we currently recognize as good, this just makes the step from N+1 recognition to N+1 search be simultaneous within an individual organism. It doesn't change the logic whereby going from N to N+1 in the sequence of recognizably good arguments must have some fitness advantage that is not "in order to win arguments" in order for individuals bearing the N+1 allele to have a fitness advantage over individuals who only have the alleles up to N, because being swayed by N+1 is not an advantage in argument until other individuals have that allele too.
In real life we have a deep pool of fixed genes with a bubbling surface of multiple genes under selection, along with complicated phenotypical interactions etcetera, but none of this changes the larger point so far as I can tell: a bacterium or a mouse have little ability to be swayed by arguments of the sort humans exchange with each other, which defines their lack of reasoning ability more than their difficulty in coming up with good arguments; and an ability to be swayed by an argument of whatever type must be present before there's any use in improving a search for arguments that meet that criterion. In other words, the journey from the kind of arguments that bacteria recognize, to the kind of arguments that humans recognize, cannot have been driven by an increasingly powerful search for political arguments that appeal to bacteria.
Even if the key word is supposed to be 'explicit', we can apply a similar logic to the ability to be swayed by an 'explicit' thought and the ability to search for explicit thoughts that sway people.
If arguments had no meaning but to argue other people into things, if they were being subject only to neutral selection or genetic drift or mere conformism, there really wouldn't be any reason for "the kind of arguments humans can be swayed by" to work to build a spaceship. We'd just end up with some arbitrary set of rules fixed in place. False cynicism.
Now, for the rats, there’s an evolutionarily-adaptive goal of "when in a salt-deprived state, try to eat salt". The genome is “trying” to install that goal in the rat’s brain. And apparently, it worked! That goal was installed!
This is importantly technically false in a way that should not be forgotten on pain of planetary extinction:
The outer loss function training the rat genome was strictly inclusive genetic fitness. The rats ended up with zero internal concept of inclusive genetic fitness, and indeed, no coherent utility function; and instead ended up with complicated internal machinery running off of millions of humanly illegible neural activations; whose properties included attaching positive motivational valence to imagined states that the rat had never experienced before, but which shared a regularity with states experienced by past rats during the "training" phase.
A human, who works quite similarly to the rat due to common ancestry, may find it natural to think of this as a very simple 'goal'; because things similar to us appear to have falsely low algorithmic complexity when we model them by empathy; because the empathy can model them using short codes. A human may imagine that natural selection successfully created rats with a simple salt-balance term in their simple generalization of a utility function, simply by natural-selection-training them on environmental scenarios with salt deficits and simple loss-function penalties for not balancing the salt deficits, which were then straightforwardly encoded into equally simple concepts in the rat.
This isn't what actually happened. Natural selection applied a very simple loss function of 'inclusive genetic fitness'. It ended up as much more complicated internal machinery in the rat that made zero mention of the far more compact concept behind the original loss function. You share the complicated machinery so it looks simpler to you than it should be, and you find the results sympathetic so they seem like natural outcomes to you. But from the standpoint of natural-selection-the-programmer the results were bizarre, and involved huge inner divergences and huge inner novel complexity relative to the outer optimization pressures.
What is all of humanity if not a walking catastrophic inner alignment failure? We were optimized for one thing: inclusive genetic fitness. And only a tiny fraction of humanity could correctly define what that is!
I don't want to take away from MIRI's work (I still support them, and I think that if the GPTs peter out, we'll be glad they've been continuing their work), but I think it's an essential time to support projects that can work for a GPT-style near-term AGI
I'd love to know of a non-zero integer number of plans that could possibly, possibly, possibly work for not dying to a GPT-style near-term AGI.
Thank you for sharing this info. My faith is now shaken.
Yes, via "Language in Thought and Action" and the Null-A novels.
(Deleted section on why I thought cultural general-intelligence software was not much of the work of AGI:)
...because the soft fidelity of implicit unconscious cultural transmission can store less serially deep and intricate algorithms than the high-fidelity DNA transmission used to store the kind of algorithms that appear in computational neuroscience.
I recommend Terrence Deacon's The Symbolic Species for some good discussion of the surprising importance of the shallow algorithms and parameters that can get transmitted culturally. The human-raised chimpanzee Kanzi didn't become a human, because that takes deeper and more neural algorithms than imitating the apes around you can transmit, but Kanzi was a lot smarter than other chimpanzees in some interesting ways.
But as necessary as it may be to avoid feral children, this kind of shallow soft-software doesn't strike me as something that takes a long time to redevelop, compared to hard-software like the secrets of computational neuroscience.
It would be helpful to know to what extent Paul feels like he endorses the FAQ here. This makes it sound like Yet Another Stab At Boiling Down The Disagreement would say that I disagree with Paul on two critical points:
I don't want to invest huge amounts arguing with this until I know to what extent Paul agrees with either the FAQ, or that this sounds like a plausible locus of disagreement. But a gloss on my guess at the disagreement might be:
Paul thinks that current ML methods given a ton more computing power will suffice to give us a basically neutral, not of itself ill-motivated, way of producing better conformance of a function to an input-output behavior implied by labeled data, which can learn things on the order of complexity of "corrigible behavior" and do so without containing tons of weird squiggles; Paul thinks you can iron out the difference between "mostly does what you want" and "very exact reproduction of what you want" by using more power within reasonable bounds of the computing power that might be available to a large project in N years when AGI is imminent, or through some kind of weird recursion. Paul thinks you do not get Project Chaos and Software Despair that takes more than 6 months to iron out when you try to do this. Eliezer thinks that in the alternate world where this is true, GANs pretty much worked the first time they were tried, and research got to very stable and robust behavior that boiled down to having no discernible departures from "reproduce the target distribution as best you can" within 6 months of being invented.
Eliezer expects great Project Chaos and Software Despair from trying to use gradient descent, genetic algorithms, or anything like that, as the basic optimization to reproduce par-human cognition within a boundary in great fidelity to that boundary as the boundary was implied by human-labeled data. Eliezer thinks that if you have any optimization powerful enough to reproduce humanlike cognition inside a detailed boundary by looking at a human-labeled dataset trying to outline the boundary, the thing doing the optimization is powerful enough that we cannot assume its neutrality the way we can assume the neutrality of gradient descent.
Eliezer expects weird squiggles from gradient descent - it's not that gradient descent can never produce par-human cognition, even natural selection will do that if you dump in enough computing power. But you will get the kind of weird squiggles in the learned function that adversarial examples expose in current nets - special inputs that weren't in the training distribution, but look like typical members of the training distribution from the perspective of the training distribution itself, will break what we think is the intended labeling from outside the system. Eliezer does not think Ian Goodfellow will have created a competitive form of supervised learning by gradient descent which lacks "squiggles" findable by powerful intelligence by the time anyone is trying to create ML-based AGI, though Eliezer is certainly cheering Goodfellow on about this and would recommend allocating Goodfellow $1 billion if Goodfellow said he could productively use it. You cannot iron out the squiggles just by using more computing power in bounded in-universe amounts.
These squiggles in the learned function could correspond to daemons, if they grow large enough, or just something that breaks our hoped-for behavior from outside the system when the system is put under a load of optimization. In general, Eliezer thinks that if you have scaled up ML to produce or implement some components of an Artificial General Intelligence, those components do not have a behavior that looks like "We put in loss function L, and we got out something that really actually minimizes L". You get something that minimizes some of L and has weird squiggles around typical-looking inputs (inputs not obviously distinguished from the training distribution except insofar as they exploit squiggles). The system is subjecting itself to powerful optimization that produces unusual inputs and weird execution trajectories - any output that accomplishes the goal is weird compared to a random output and it may have other weird properties as well. You can't just assume you can train for X in a robust way when you have a loss function that targets X.
I imagine that Paul replies to this saying "I agree, but..." but I'm not sure what comes after the "but". It looks to me like Paul is imagining that you can get very powerful optimization with very detailed conformance to our intended interpretation of the dataset, powerful enough to enclose par-human cognition inside a boundary drawn from human labeling of a dataset, and have that be the actual thing we get out rather than a weird thing full of squiggles. If Paul thinks he has a way to compound large conformant recursive systems out of par-human thingies that start out weird and full of squiggles, we should definitely be talking about that. From my perspective it seems like Paul repeatedly reasons "We train for X and get X" rather than "We train for X and get something that mostly conforms to X but has a bunch of weird squiggles" and also often speaks as if the training method is assumed to be gradient descent, genetic algorithms, or something else that can be assumed neutral-of-itself rather than being an-AGI-of-itself whose previous alignment has to be assumed.
The imaginary Paul in my head replies that we actually are using an AGI to train on X and get X, but this AGI was previously trained by a weaker neutral AGI, and so on going back to something trained by gradient descent. My imaginary reply is that neutrality is not the same property as conformance or nonsquiggliness, and if you train your base AGI via neutral gradient descent you get out a squiggly AGI and this squiggly AGI is not neutral when it comes to that AGI looking at a dataset produced by X and learning a function conformant to X. Or to put it another way, if the plan is to use gradient descent on human-labeled data to produce a corrigible alien that is smart enough to produce more corrigible aliens better than gradient descent, this corrigible alien actually needs to be quite smart because an IQ 100 human will not build an aligned IQ 140 human even if you run them for a thousand years, so you are producing something very smart and dangerous on the first step, and gradient descent is not smart enough to align that base case.
But at this point I expect the real Paul to come back and say, "No, no, the idea is something else..."
A very important aspect of my objection to Paul here is that I don't expect weird complicated ideas about recursion to work on the first try, with only six months of additional serial labor put into stabilizing them, which I understand to be Paul's plan. In the world where you can build a weird recursive stack of neutral optimizers into conformant behavioral learning on the first try, GANs worked on the first try too, because that world is one whose general Murphy parameter is set much lower than ours. Being able to build weird recursive stacks of optimizers that work correctly to produce neutral and faithful optimization for corrigible superhuman thought out of human-labeled corrigible behaviors and corrigible reasoning, without very much of a time penalty relative to nearly-equally-resourced projects who are just cheerfully revving all the engines as hard as possible trying to destroy the world, is just not how things work in real life, dammit. Even if you could make the weird recursion work, it would take time.
Eliezer thinks that while corrigibility probably has a core which is of lower algorithmic complexity than all of human value, this core is liable to be very hard to find or reproduce by supervised learning of human-labeled data, because deference is an unusually anti-natural shape for cognition, in a way that a simple utility function would not be an anti-natural shape for cognition. Utility functions have multiple fixpoints requiring the infusion of non-environmental data, our externally desired choice of utility function would be non-natural in that sense, but that's not what we're talking about, we're talking about anti-natural behavior.
E.g.: Eliezer also thinks that there is a simple core describing a reflective superintelligence which believes that 51 is a prime number, and actually behaves like that including when the behavior incurs losses, and doesn't thereby ever promote the hypothesis that 51 is not prime or learn to safely fence away the cognitive consequences of that belief and goes on behaving like 51 is a prime number, while having no other outwardly discernible deficits of cognition except those that directly have to do with 51. Eliezer expects there's a relatively simple core for that, a fixed point of tangible but restrained insanity that persists in the face of scaling and reflection; there's a relatively simple superintelligence that refuses to learn around this hole, refuses to learn how to learn around this hole, refuses to fix itself, but is otherwise capable of self-improvement and growth and reflection, etcetera. But the core here has a very anti-natural shape and you would be swimming uphill hard if you tried to produce that core in an indefinitely scalable way that persisted under reflection. You would be very unlikely to get there by training really hard on a dataset where humans had labeled as the 'correct' behavior what humans thought would be the implied behavior if 51 were a prime number, not least because gradient descent is terrible, but also just because you'd be trying to lift 10 pounds of weirdness with an ounce of understanding.
The central reasoning behind this intuition of anti-naturalness is roughly, "Non-deference converges really hard as a consequence of almost any detailed shape that cognition can take", with a side order of "categories over behavior that don't simply reduce to utility functions or meta-utility functions are hard to make robustly scalable".
The real reasons behind this intuition are not trivial to pump, as one would expect of an intuition that Paul Christiano has been alleged to have not immediately understood. A couple of small pumps would be https://arbital.com/p/updated_deference/ for the first intuition and https://arbital.com/p/expected_utility_formalism/?l=7hh for the second intuition.
What I imagine Paul is imagining is that it seems to him like it would in some sense be not that hard for a human who wanted to be very corrigible toward an alien, to be very corrigible toward that alien; so you ought to be able to use gradient-descent-class technology to produce a base-case alien that wants to be very corrigible to us, the same way that natural selection sculpted humans to have a bunch of other desires, and then you apply induction on it building more corrigible things.
My class of objections in (1) is that natural selection was actually selecting for inclusive fitness when it got us, so much for going from the loss function to the cognition; and I have problems with both the base case and the induction step of what I imagine to be Paul's concept of solving this using recursive optimization bootstrapping itself; and even more so do I have trouble imagining it working on the first, second, or tenth try over the course of the first six months.
My class of objections in (2) is that it's not a coincidence that humans didn't end up deferring to natural selection, or that in real life if we were faced with a very bizarre alien we would be unlikely to want to defer to it. Our lack of scalable desire to defer in all ways to an extremely bizarre alien that ate babies, is not something that you could fix just by giving us an emotion of great deference or respect toward that very bizarre alien. We would have our own thought processes that were unlike its thought processes, and if we scaled up our intelligence and reflection to further see the consequences implied by our own thought processes, they wouldn't imply deference to the alien even if we had great respect toward it and had been trained hard in childhood to act corrigibly towards it.
A dangerous intuition pump here would be something like, "If you take a human who was trained really hard in childhood to have faith in God and show epistemic deference to the Bible, and inspecting the internal contents of their thought at age 20 showed that they still had great faith, if you kept amping up that human's intelligence their epistemology would at some point explode"; and this is true even though it's other humans training the human, and it's true even though religion as a weird sticking point of human thought is one we selected post-hoc from the category of things historically proven to be tarpits of human psychology, rather than aliens trying from the outside in advance to invent something that would stick the way religion sticks. I use this analogy with some reluctance because of the clueless readers who will try to map it onto the AGI losing religious faith in the human operators, which is not what this analogy is about at all; the analogy here is about the epistemology exploding as you ramp up intelligence because the previous epistemology had a weird shape.
Acting corrigibly towards a baby-eating virtue ethicist when you are a utilitarian is an equally weird shape for a decision theory. It probably does have a fixed point but it's not an easy one, the same way that "yep, on reflection and after a great deal of rewriting my own thought processes, I sure do still think that 51 is prime" probably has a fixed point but it's not an easy one.
I think I can imagine an IQ 100 human who defers to baby-eating aliens, although I really think a lot of this is us post-hoc knowing that certain types of thoughts can be sticky, rather than the baby-eating aliens successfully guessing in advance how religious faith works for humans and training the human to think that way using labeled data.
But if you ramp up the human's intelligence to where they are discovering subjective expected utility and logical decision theory and they have an exact model of how the baby-eating aliens work and they are rewriting their own minds, it's harder to imagine the shape of deferential thought at IQ 100 successfully scaling to a shape of deferential thought at IQ 1000.
Eliezer also tends to be very skeptical of attempts to cross cognitive chasms between A and Z by going through weird recursions and inductive processes that wouldn't work equally well to go directly from A to Z. http://slatestarcodex.com/2014/10/12/five-planets-in-search-of-a-sci-fi-story/ and the story of K'th'ranga V is a good intuition pump here. So Eliezer is also not very hopeful that Paul will come up with a weirdly recursive solution that scales deference to IQ 101, IQ 102, etcetera, via deferential agents building other deferential agents, in a way that Eliezer finds persuasive. Especially a solution that works on merely the tenth try over the first six months, doesn't kill you when the first nine tries fail, and doesn't require more than 10x extra computing power compared to projects that are just bulling cheerfully ahead.
I think I have a disagreement with Paul about the notion of being able to expose inspectable thought processes to humans, such that we can examine each step of the thought process locally and determine whether it locally has properties that will globally add up to corrigibility, alignment, and intelligence. It's not that I think this can never be done, or even that I think it takes longer than six months. In this case, I think this problem is literally isomorphic to "build an aligned AGI". If you can locally inspect cognitive steps for properties that globally add to intelligence, corrigibility, and alignment, you're done; you've solved the AGI alignment problem and you can just apply the same knowledge to directly build an aligned corrigible intelligence.
As I currently flailingly attempt to understand Paul, Paul thinks that having humans do the inspection (base case) or thingies trained to resemble aggregates of trained thingies (induction step) is something we can do in an intuitive sense by inspecting a reasoning step and seeing if it sounds all aligned and corrigible and intelligent. Eliezer thinks that the large-scale or macro traces of cognition, e.g. a "verbal stream of consciousness" or written debates, are not complete with respect to general intelligence in bounded quantities; we are generally intelligent because of sub-verbal cognition whose intelligence-making properties are not transparent to inspection. That is: An IQ 100 person who can reason out loud about Go, but who can't learn from the experience of playing Go, is not a complete general intelligence over boundedly reasonable amounts of reasoning time.
This means you have to be able to inspect steps like "learn an intuition for Go by playing Go" for local properties that will globally add to corrigible aligned intelligence. And at this point it no longer seems intuitive that having humans do the inspection is adding a lot of value compared to us directly writing a system that has the property.
This is a previous discussion that is ongoing between Paul and myself, and I think it's a crux of disagreement but not one that's as cruxy as 1 and 2. Although it might be a subcrux of my belief that you can't use weird recursion starting from gradient descent on human-labeled data to build corrigible agents that build corrigible agents. I think Paul is modeling the grain size here as corrigible thoughts rather than whole agents, which if it were a sensible way to think, might make the problem look much more manageable; but I don't think you can build corrigible thoughts without building corrigible agents to think them unless you have solved the decomposition problem that I think is isomorphic to building an aligned corrigible intelligence directly.
I remark that this intuition matches what the wise might learn from Scott's parable of K'th'ranga V: If you know how to do something then you know how to do it directly rather than by weird recursion, and what you imagine yourself doing by weird recursion you probably can't really do at all. When you want an airplane you don't obtain it by figuring out how to build birds and then aggregating lots of birds into a platform that can carry more weight than any one bird and then aggregating platforms into megaplatforms until you have an airplane; either you understand aerodynamics well enough to build an airplane, or you don't, the weird recursion isn't really doing the work. It is by no means clear that we would have a superior government free of exploitative politicians if all the voters elected representatives whom they believed to be only slightly smarter than themselves, until a chain of delegation reached up to the top level of government; either you know how to build a less corruptible relationship between voters and politicians, or you don't, the weirdly recursive part doesn't really help. It is no coincidence that modern ML systems do not work by weird recursion because all the discoveries are of how to just do stuff, not how to do stuff using weird recursion. (Even with AlphaGo which is arguably recursive if you squint at it hard enough, you're looking at something that is not weirdly recursive the way I think Paul's stuff is weirdly recursive, and for more on that see https://intelligence.org/2018/05/19/challenges-to-christianos-capability-amplification-proposal/.)
It's in this same sense that I intuit that if you could inspect the local elements of a modular system for properties that globally added to aligned corrigible intelligence, it would mean you had the knowledge to build an aligned corrigible AGI out of parts that worked like that, not that you could aggregate systems that corrigibly learned to put together sequences of corrigible thoughts into larger corrigible thoughts starting from gradient descent on data humans have labeled with their own judgments of corrigibility.
Voting in elections is a wonderful example of logical decision theory in the wild. The chance that you are genuinely logically correlated to a random trade partner is probably small, in cases where you don't have mutual knowledge of LDT; leaving altruism and reputation as sustaining reasons for cooperation. With millions of voters, the chance that you are correlated to thousands of them is much better.
Or perhaps you'd prefer to believe the dictate of Causal Decision Theory that if an election is won by 3 votes, nobody's vote influenced it, and if an election is won by 1 vote, all of the millions of voters on the winning side are solely responsible. But that was a silly decision theory anyway. Right?