Crossposted from the AI Alignment Forum. May contain more technical jargon than usual.

With the release of Rohin Shah and Eliezer Yudkowsky's conversation, the Late 2021 MIRI Conversations sequence is now complete.

This post is intended as a generalized comment section for discussing the whole sequence, now that it's finished. Feel free to:

  • raise any topics that seem relevant
  • signal-boost particular excerpts or comments that deserve more attention
  • direct questions to participants

In particular, Eliezer Yudkowsky, Richard Ngo, Paul Christiano, Nate Soares, and Rohin Shah expressed active interest in receiving follow-up questions here. The Schelling time when they're likeliest to be answering questions is Wednesday March 2, though they may participate on other days too.

New Comment
205 comments, sorted by Click to highlight new comments since:
Some comments are truncated due to high volume. (⌘F to expand all)Change truncation settings

This is mostly in response to stuff written by Richard, but I'm interested in everyone's read of the situation.

While I don't find Eliezer's core intuitions about intelligence too implausible, they don't seem compelling enough to do as much work as Eliezer argues they do. As in the Foom debate, I think that our object-level discussions were constrained by our different underlying attitudes towards high-level abstractions, which are hard to pin down (let alone resolve).

Given this, I think that the most productive mode of intellectual engagement with Eliezer's worldview going forward is probably not to continue debating it (since that would likely hit those same underlying disagreements), but rather to try to inhabit it deeply enough to rederive his conclusions and find new explanations of them which then lead to clearer object-level cruxes.

I'm not sure yet how to word this as a question without some introductory paragraphs. When I read Eliezer, I often feel like he has a coherent worldview that sees lots of deep connections and explains lots of things, and that he's actively trying to be coherent / explain everything. [This is what I think you're pointing to with his 'attitude toward... (read more)

I feel like I have a broad distribution over worlds and usually answer questions with probability distributions, that I have a complete mental universe (which feels to me like it outputs answers to a much broader set of questions than Eliezer's, albeit probabilistic ones, rather than bailing with "the future is hard to predict").  At a high level I don't think "mainline" is a great concept for describing probability distributions over the future except in certain exceptional cases (though I may not understand what "mainline" means), and that neat stories that fit everything usually don't work well (unless, or often even if, generated in hindsight).

In answer to your "why is this," I think it's a combination of moderate differences in functioning and large differences in communication style. I think Eliezer has a way of thinking about the future that is quite different from mine and I'm somewhat skeptical of and feel like Eliezer is overselling (which is what got me into this discussion), but that's probably smaller than a large difference in communication style (driven partly by different skills, different aesthetics, and different ideas about what kinds of standards discourse should aspire to).

I think I may not understand well the basic lesson / broader point, so will probably be more helpful on object level points and will mostly go answer those in the time I have.


I feel like I have a broad distribution over worlds and usually answer questions with probability distributions, that I have a complete mental universe (which feels to me like it outputs answers to a much broader set of questions than Eliezer's, albeit probabilistic ones, rather than bailing with "the future is hard to predict").

Sometimes I'll be tracking a finite number of "concrete hypotheses", where every hypothesis is 'fully fleshed out', and be doing a particle-filtering style updating process, where sometimes hypotheses gain or lose weight, sometimes they get ruled out or need to split, or so on. In those cases, I'm moderately confident that every 'hypothesis' corresponds to a 'real world', constrained by how well as I can get my imagination to correspond to reality. [A 'finite number' depends on the situation, but I think it's normally something like 2-5, unless it's an area I've built up a lot of cache about.]

Sometimes I'll be tracking a bunch of "surface-level features", where the distributions on the features don't always imply coherent underlying worlds, either on their own or in combination with other features. (For example, I might have guesses about the probability th... (read more)

I think my way of thinking about things is often a lot like "draw random samples," more like drawing N random samples rather than particle filtering (I guess since we aren't making observations as we go---if I notice an inconsistency the thing I do is more like backtrack and start over with N fresh samples having updated on the logical fact).

The main complexity feels like the thing you point out where it's impossible to make them fully fleshed out, so you build a bunch of intuitions about what is consistent (and could be fleshed out given enough time) and then refine those intuitions only periodically when you actually try to flesh something out and see if it makes sense. And often you go even further and just talk about relationships amongst surface level features using intuitions refined from a bunch of samples.

I feel like a distinctive feature of Eliezer's dialog w.r.t. foom / alignment difficulty is that he has a lot of views about strong regularities that should hold across all of these worlds. And then disputes about whether worlds are plausible often turn on things like "is this property of the described world likely?" which is tough because obviously everyone agrees that ev... (read more)

Oh whoa, you don't remember your samples from before? [I guess I might not either, unless I'm concentrating on keeping them around or verbalized them or something; probably I do something more expert-iteration-like where I'm silently updating my generating distributions based on the samples and then resampling them in the future.] Yeah, this seems likely; this makes me more interested in the "selectively ignoring variables" hypothesis for why Eliezer running this strategy might have something that would naturally be called a mainline. [Like, it's very easy to predict "number of apples sold = number of apples bought" whereas it's much harder to predict the price of apples.] But maybe instead he means it in the 'startup plan' sense, where you do actually assign basically no probability to your mainline prediction, but still vastly more than any other prediction that's equally conjunctive.

EDIT: I wrote this before seeing Paul's response; hence a significant amount of repetition.

They often seem to emit sentences that are 'not absurd', instead of 'on their mainline', because they're mostly trying to generate sentences that pass some shallow checks instead of 'coming from their complete mental universe.'

Why is this?

Well, there are many boring cases that are explained by pedagogy / argument structure. When I say things like "in the limit of infinite oversight capacity, we could just understand everything about the AI system and reengineer it to be safe", I'm obviously not claiming that this is a realistic thing that I expect to happen, so it's not coming from my "complete mental universe"; I'm just using this as an intuition pump for the listener to establish that a sufficiently powerful oversight process would solve AI alignment.

That being said, I think there is a more interesting difference here, but that your description of it is inaccurate (at least for me).

From my perspective I am implicitly representing a probability distribution over possible futures in my head. When I say "maybe X happens", or "X is not absurd", I'm saying that my probability distribution assign... (read more)


In response to your last couple paragraphs: the critique, afaict, is not "a real human cannot keep multiple concrete scenarios in mind and speak probabilistically about those", but rather "a common method for representing lots of hypotheses at once, is to decompose the hypotheses into component properties that can be used to describe lots of concrete hypotheses. (toy model: instead of imagining all numbers, you note that some numbers are odd and some numbers are even, and then think of evenness and oddness). A common failure mode when attempting this is that you lose track of which properties are incompatible (toy model: you claim you can visualize a number that is both even and odd). A way to avert this failure mode is to regularly exhibit at least one concrete hypothesis that simultaneousy posseses whatever collection of properties you say you can simultaneously visualize (toy model: demonstrating that 14 is even and 7 is odd does not in fact convince me that you are correct to imagine a number that is both even and odd)."

On my understanding of Eliezer's picture (and on my own personal picture), almost nobody ever visibly tries to do this (never mind succeeding), when it comes to hopeful AGI scenarios.

Insofar as you have thought about at least one specific hopeful world in great detail, I strongly recommend, spelling it out, in all its great detail, to Eliezer, next time you two chat. In fact, I personally request that you do this! It sounds great, and I expect it to constitute some progress in the debate.


Relevant Feynman quote: 

I had a scheme, which I still use today when somebody is explaining something that I’m trying to understand: I keep making up examples.

For instance, the mathematicians would come in with a terrific theorem, and they’re all excited. As they’re telling me the conditions of the theorem, I construct something which fits all the conditions. You know, you have a set (one ball)-- disjoint (two balls). Then the balls turn colors, grow hairs, or whatever, in my head as they put more conditions on.

Finally they state the theorem, which is some dumb thing about the ball which isn’t true for my hairy green ball thing, so I say “False!” [and] point out my counterexample.

As I understand it, when you "talk about the mainline", you're supposed to have some low-entropy (i.e. confident) view on how the future goes, such that you can answer very different questions X, Y and Z about that particular future, that are all correlated with each other, and all get (say) > 50% probability. (Idk, as I write this down, it seems so obviously a bad way to reason that I feel like I must not be understanding it correctly.)

But to the extent this is right, I'm actually quite confused why anyone thinks "talk about the mainline" is an ideal to which to aspire. What makes you expect that?

I'll try to explain the technique and why it's useful. I'll start with a non-probabilistic version of the idea, since it's a little simpler conceptually, then talk about the corresponding idea in the presence of uncertainty.

Suppose I'm building a mathematical model of some system or class of systems. As part of the modelling process, I write down some conditions which I expect the system to satisfy - think energy conservation, or Newton's Laws, or market efficiency, depending on what kind of systems we're talking about. My hope/plan is to derive (i.e. prove) some predictions from these... (read more)

9Rohin Shah
Man, I would not call the technique you described "mainline prediction". It also seems kinda inconsistent with Vaniver's usage; his writing suggests that a person only has one mainline at a time which seems odd for this technique. Vaniver, is this what you meant? If so, my new answer is that I and others do in fact talk about "mainline predictions" -- for me, there was that whole section talking about natural language debate as an alignment strategy. (It ended up not being about a plausible world, but that's because (a) Eliezer wanted enough concreteness that I ended up talking about the stupidly inefficient version rather than the one I'd actually expect in the real world and (b) I was focused on demonstrating an existence proof for the technical properties, rather than also trying to include the social ones.)
To be clear, I do not mean to use the label "mainline prediction" for this whole technique. Mainline prediction tracking is one way of implementing this general technique, and I claim that the usefulness of the general technique is the main reason why mainline predictions are useful to track. (Also, it matches up quite well with Nate's model based on his comment here, and I expect it also matches how Eliezer wants to use the technique.)
6Rohin Shah
Ah, got it. I agree that: 1. The technique you described is in fact very useful 2. If your probability distribution over futures happens to be such that it has a "mainline prediction", you get significant benefits from that (similar to the benefits you get from the technique you described).
Uh, I inherited "mainline" from Eliezer's usage in the dialogue, and am guessing that his reasoning is following a process sort of like mine and John's. My natural word for it is a 'particle', from particle filtering, as linked in various places, which I think is consistent with John's description. I'm further guessing that Eliezer's noticed more constraints / implied inconsistencies, and is somewhat better at figuring out which variables to drop, so that his cloud is narrower than mine / more generates 'mainline predictions' than 'probability distributions'. Do you feel like you do this 'sometimes', or 'basically always'? Maybe it would be productive for me to reread the dialogue (or at least part of it) and sort sections / comments by how much they feel like they're coming from this vs. some other source.  As a specific thing that I have in mind, I think there's a habit of thinking / discourse that philosophy trains, which is having separate senses for "views in consideration" and "what I believe", and thinking that statements should be considered against all views in consideration, even ones that you don't believe. This seems pretty good in some respects (if you begin by disbelieving a view incorrectly, your habits nevertheless gather you lots of evidence about it, which can cause you to then correctly believe it), and pretty questionable in other respects (conversations between Alice and Bob now have to include them shadowboxing with everyone else in the broader discourse, as Alice is asking herself "what would Carol say in response to that?" to things that Bob says to her). When I imagine dialogues generated by people who are both sometimes doing the mainline thing and sometimes doing the 'represent the whole discourse' thing, they look pretty different from dialogues generated by people who are both only doing the mainline thing. [And also from dialogues generated by both people only doing the 'represent the whole discourse' thing, of course.]
5Rohin Shah
I don't know what "this" refers to. If the referent is "have a concrete example in mind", then I do that frequently but not always. I do it a ton when I'm not very knowledgeable and learning about a thing; I do it less as my mastery of a subject increases. (Examples: when I was initially learning addition, I used the concrete example of holding up three fingers and then counting up two more to compute 3 + 2 = 5, which I do not do any more. When I first learned recursion, I used to explicitly run through an execution trace to ensure my program would work, now I do not.) If the referent is "make statements that reflect my beliefs", then it depends on context, but in the context of these dialogues, I'm always doing that. (Whereas when I'm writing for the newsletter, I'm more often trying to represent the whole discourse, though the "opinion" sections are still entirely my beliefs.)
I think this is roughly how I'm thinking about things sometimes, tho I'd describe the mainline as the particle with plurality weight (which is a weaker condition than >50%). [I don't know how Eliezer thinks about things; maybe it's like this? I'd be interested in hearing his description.] I think this is also a generator of disagreements about what sort of things are worth betting on; when I imagine why I would bail with "the future is hard to predict", it's because the hypotheses/particles I'm considering have clearly defined X, Y, and Z variables (often discretized into bins or ranges) but not clearly defined A, B, and C variables (tho they might have distributions over those variables), because if you also conditioned on those you would have Too Many Particles. And when I imagine trying to contrast particles on features A, B, and C, as they all make weak predictions we get at most a few bits of evidence to update their weights on, whereas when we contrast them on X, Y, and Z we get many more bits, and so it feels more fruitful to reason about. I mean, the question is which direction we want to approach Bayesianism from, given that Bayesianism is impossible (as you point out later in your comment). On the one hand, you could focus on 'updating', and have lots of distributions that aren't grounded in reality but which are easy to massage when new observations come in, and on the other hand, you could focus on 'hypotheses', and have as many models of the situation as you can ground, and then have to do something much more complicated when new observations come in. [Like, a thing I find helpful to think about here is where the motive power from Aumann's Agreement Theorem comes from, which is that when I say 40% A, you know that my private info is consistent with an update of the shared prior whose posterior is 40%, and when you take the shared prior and update on your private info and that my private info is consistent with 40% and your posterior is 60% A, then I
4Rohin Shah
If you define "mainline" as "particle with plurality weight", then I think I was in fact "talking on my mainline" at some points during the conversation, and basically everywhere that I was talking about worlds (instead of specific technical points or intuition pumps) I was talking about "one of my top 10 particles". I think I responded to every request for concreteness with a fairly concrete answer. Feel free to ask me for more concreteness in any particular story I told during the conversation.
Huh, I guess I don't believe the intuition pump? Like, as the first counterexample that comes to mind, when I imagine having an AGI where I can tell everything about how it's thinking, and yet I remain a black box to myself, I can't really tell whether or not it's aligned to me. (Is me-now the one that I want it to be aligned to, or me-across-time? Which side of my internal conflicts about A vs. B / which principle for resolving such conflicts?) I can of course imagine a reasonable response to that from you--"ah, resolving philosophical difficulties is the user's problem, and not one of the things that I mean by alignment"--but I think I have some more-obviously-alignment-related counterexamples. [Tho if by 'infinite oversight ability' you do mean something like 'logical omniscience' it does become pretty difficult to find a real counterexample, in part because I can just find the future trajectory with highest expected utility and take the action I take at the start of that trajectory without having to have any sort of understanding about why that action was predictably a good idea.] But like, the thing this reminds me of is something like extrapolating tangents, instead of operating the production function? "If we had an infinitely good engine, we could make the perfect car", which seems sensible when you're used to thinking of engine improvements linearly increasing car quality and doesn't seem sensible when you're used to thinking of car quality as a product of sigmoids of the input variables. (This is a long response to a short section because I think the disagreement here is about something like "how should we reason and communicate about intuitions?", and so it's worth expanding on what I think might be the implications of otherwise minor disagreements.)
2Rohin Shah
That is in fact my response. (Though one of the ways in which the intuition pump isn't fully compelling to me is that, even after understanding the exact program that the AGI implements and its causal history, maybe the overseers can't correctly predict the consequences of running that program for a long time. Still feels like they'd do fine.) I do agree that if you go as far as "logical omniscience" then there are "cheating" ways of solving the problem that don't really tell us much about how hard alignment is in practice. The car analogy just doesn't seem sensible. I can tell stories of car doom even if you have infinitely good engines (e.g. the steering breaks). My point is that we struggle to tell stories of doom when imagining a very powerful oversight process that knows everything the model knows. I'm not thinking "more oversight quality --> more alignment" and then concluding "infinite oversight quality --> alignment solved". I'm starting with the intuition pump, noticing I can no longer tell a good story of doom, and concluding "infinite oversight quality --> alignment solved". So I don't think this has much to do with extrapolating tangents vs. production functions, except inasmuch as production functions encourage you to think about complements to your inputs that you can then posit don't exist in order to tell a story of doom.
I think some of my more alignment-flavored counterexamples look like: * The 'reengineer it to be safe' step breaks down / isn't implemented thru oversight. Like, if we're positing we spin up a whole Great Reflection to evaluate every action the AI takes, this seems like it's probably not going to be competitive! * The oversight gives us as much info as we ask for, but the world is a siren world (like what Stuart points to, but a little different), where the initial information we discover about the plans from oversight is so convincing that we decide to go ahead with the AI before discovering the gotchas. * Related to the previous point, the oversight is sufficient to reveal features about the plan that are terrible, but before the 'reengineer to make it more safe' plan is executed, the code is stolen and executed by a subset of humanity which thinks the terrible plan is 'good enough', for them at least. That is, it feels to me like we benefit a lot from having 1) a constructive approach to alignment instead of rejection sampling, 2) sufficient security focus that we don't proceed on EV of known information, but actually do the 'due diligence', and 3) sufficient coordination among humans that we don't leave behind substantial swaths of current human preferences, and I don't see how we get those thru having arbitrary transparency. [I also would like to solve the problem of "AI has good outcomes" instead of the smaller problem of "AI isn't out to get us", because accidental deaths are deaths too! But I do think it makes sense to focus on that capability problem separately, at least sometimes.]
2Rohin Shah
I obviously do not think this is at all competitive, and I also wanted to ignore the "other people steal your code" case. I am confused what you think I was trying to do with that intuition pump. I guess I said "powerful oversight would solve alignment" which could be construed to mean that powerful oversight => great future, in which case I'd change it to "powerful oversight would deal with the particular technical problems that we call outer and inner alignment", but was it really so non-obvious that I was talking about the technical problems? Maybe your point is that there are lots of things required for a good future, just as a car needs both steering and an engine, and so the intuition pump is not interesting because it doesn't talk about all the things needed for a good future? If so, I totally agree that it does not in fact include all the things needed for a good future, and it was not meant to be saying that. This just doesn't seem plausible to me. Where did the information come from? Did the AI system optimize the information to be convincing? If yes, why didn't we notice that the AI system was doing that? Can we solve this by ensuring that we do due diligence, even if it doesn't seem necessary?
I think I'm confused about the intuition pump too! Like, here's some options I thought up: * The 'alignment problem' is really the 'not enough oversight' problem. [But then if we solve the 'enough oversight' problem, we still have to solve the 'what we want' problem, the 'coordination' problem, the 'construct competitively' problem, etc.] * Bits of the alignment problem can be traded off against each other, most obviously coordination and 'alignment tax' (i.e. the additional amount of work you need to do to make a system aligned, or the opposite of 'competitiveness', which I didn't want to use here for ease-of-understanding-by-newbies reasons.) [But it's basically just coordination and competitiveness; like, you could imagine that oversight gives you a rejection sampling story for trading off time and understanding but I think this is basically not true because you're also optimizing for finding holes in your transparency regime.] Like, by analogy, I could imagine someone who uses an intuition pump of "if you had sufficient money, you could solve any problem", but I wouldn't use that intuition pump because I don't believe it. [Sure, 'by definition' if the amount of money doesn't solve the problem, it's not sufficient. But why are we implicitly positing that there exists a sufficient amount of money instead of thinking about what money cannot buy?] (After reading the rest of your comment, it seems pretty clear to me that you mean the first bullet, as you say here:) I both 1) didn't think it was obvious (sorry if I'm being slow on following the change in usage of 'alignment' here) and 2) don't think realistically powerful oversight solves either of those two on its own (outer alignment because of "rejection sampling can get you siren worlds" problem, inner alignment because "rejection sampling isn't competitive", but I find that one not very compelling and suspect I'll eventually develop a better objection).  [EDIT: I note that I also might be doing another unf
2Rohin Shah
I mean, maybe we should just drop this point about the intuition pump, it was a throwaway reference in the original comment. I normally use it to argue against a specific mentality I sometimes see in people, and I guess it doesn't make sense outside of that context. (The mentality is "it doesn't matter what oversight process you use, there's always a malicious superintelligence that can game it, therefore everyone dies".)

The most recent post has a related exchange between Eliezer and Rohin:

Eliezer: I think the critical insight - though it has a format that basically nobody except me ever visibly invokes in those terms, and I worry maybe it can only be taught by a kind of life experience that's very hard to obtain - is the realization that any consistent reasonable story about underlying mechanisms will give you less optimistic forecasts than the ones you get by freely combining surface desiderata

Rohin: Yeah, I think I do not in fact understand why that is true for any consistent reasonable story.

If I'm being locally nitpicky, I argue that Eliezer's thing is a very mild overstatement (it should be "≤" instead of "<") but given that we're talking about forecasts, we're talking about uncertainty, and so we should expect "less" optimism instead of just "not more" optimism, and so I think Eliezer's statement stands as a general principle about engineering design.

This also feels to me like the sort of thing that I somehow want to direct attention towards. Either this principle is right and relevant (and it would be good for the field if all the AI safety thinkers held it!), or there's some deep confusion of mine that I'd like cleared up.

6A. Mensch
Question to Eliezer: would you agree with the gist of the following? And if not, any thoughts on what lead to a strong sense of 'coherence in your worldview' as Vaniver put it? Vaniver, I feel like you're pointing at something that I've noticed as well and am interested in too (the coherence of Eliezer's worldview as you put it). I wonder if has something to do with not going to uni but building his whole worldview all by him self. In my experience uni often tends towards to cramming lots of facts which are easily testable on exams, with less emphasis on understanding underlying principles (which is harder to test with multiple choice questions). Personally I feel like I had to spend my years after uni trying to make sense, a coherent whole if you like, of all the separate things I've learned while in uni where things were mostly just kind of put out there without constantly integrating things. Perhaps if you start out thinking much more about underlying principles earlier on it's easier to integrate all the separate facts into a coherent whole as you go along. Not sure if Eliezer would agree with this. Maybe it's even much more basic and he just always had a very strong sense of dissatisfaction if he couldn't make things cohere into a whole and this urge for things to make sense was much more important than self studying or thinking about underlying principles before and then during the learning of new knowledge... I would like to point out a section in the latest Shay/Yudkowsky dialogue where Eliezer says some things about this topic, does this feel like it's the same thing you are talking about Vaniver?
4Rohin Shah
Note that my first response was: and my immediately preceding message was I think I was responding to the version of the argument where "freely combining surface desiderata" was swapped out with "arguments about what you're selecting for". I probably should have noted that I agreed with the basic abstract point as Eliezer stated it; I just don't think it's very relevant to the actual disagreement. I think my complaints in the context of the discussion are: * It's a very weak statement. If you freely combine the most optimistic surface desiderata, you get ~0% chance of doom. My estimate is way higher (in odds-space) than ~0%, and the statement "p(doom) >= ~0%" is not that interesting and not a justification of "doom is near-inevitable". * Relatedly, I am not just "freely combining surface desiderata". I am doing something like "predicting what properties AI systems would have by reasoning about what properties we selected for during training". I think you could reasonably ask how that compares against "predicting what properties AI systems would have by reasoning about what mechanistic algorithms could produce the behavior we observed during training". I was under the impression that this was what Eliezer was pointing at (because that's how I framed it in the message immediately prior to the one you quoted) but I'm less confident of that now.

Sorry, I probably should have been more clear about the "this is a quote from a longer dialogue, the missing context is important." I do think that the disagreement about "how relevant is this to 'actual disagreement'?" is basically the live thing, not whether or not you agree with the basic abstract point.

My current sense is that you're right that the thing you're doing is more specific than the general case (and one of the ways you can tell is the line of argumentation you give about chance of doom), and also Eliezer can still be correctly observing that you have too many free parameters (even if the number of free parameters is two instead of arbitrarily large). I think arguments about what you're selecting for either cash out in mechanistic algorithms, or they can deceive you in this particular way.

Or, to put this somewhat differently, in my view the basic abstract point implies that having one extra free parameter allows you to believe in a 5% chance of doom when in fact there's 100% chance of doom, and so in order to get estimations like that right this needs to be one of the basic principles shaping your thoughts, tho ofc your prior should come from many examples instead of ... (read more)

2Rohin Shah
I agree that if you have a choice about whether to have more or fewer free parameters, all else equal you should prefer the model with fewer free parameters. (Obviously, all else is not equal; in particular I do not think that Eliezer's model is tracking reality as well as mine.) When Alice uses a model with more free parameters, you need to posit a bias before you can predict a systematic direction in which Alice will make mistakes. So this only bites you if you have a bias towards optimism. I know Eliezer thinks I have such a bias. I disagree with him. I agree that this is true in some platonic sense. Either the argument gives me a correct answer, in which case I have true statements that could be cashed out in terms of mechanistic algorithms, or the argument gives me a wrong answer, in which case it wouldn't be derivable from mechanistic algorithms, because the mechanistic algorithms are the "ground truth".  Quoting myself from the dialogue:
That is, when I give Optimistic Alice fewer constraints, she can more easily imagine a solution, and when I give Pessimistic Bob fewer constraints, he can more easily imagine that no solution is possible? I think... this feels true as a matter of human psychology of problem-solving, or something, and not as a matter of math. Like, the way Bob fails to find a solution mostly looks like "not actually considering the space", or "wasting consideration on easily-known-bad parts of the space", and more constraints could help with both of those. But, as math, removing constraints can't lower the volume of the implied space and so can't make it less likely that a viable solution exists. I think Eliezer thinks nearly all humans have such a bias by default, and so without clear evidence to the contrary it's a reasonable suspicion for anyone. [I think there's a thing Eliezer does a lot, which I have mixed feelings about, which is matching people's statements to patterns and then responding to the generator of the pattern in Eliezer's head, which only sometimes corresponds to the generator in the other person's head.] Cool, makes sense. [I continue to think we disagree about how true this is in a practical sense, where I read you as thinking "yeah, this is a minor consideration, we have to think with the tools we have access to, which could be wrong in either direction and so are useful as a point estimate" and me as thinking "huh, this really seems like the tools we have access to are going to give us overly optimistic answers, and we should focus more on how to get tools that will give us more robust answers."]

[I think there's a thing Eliezer does a lot, which I have mixed feelings about, which is matching people's statements to patterns and then responding to the generator of the pattern in Eliezer's head, which only sometimes corresponds to the generator in the other person's head.]

I want to add an additional meta-pattern – there was a once a person who thought I had a particular bias. They'd go around telling me "Ray, you're exhibiting that bias right now. Whatever rationalization you're coming up with right now, it's not the real reason you're arguing X." And I was like "c'mon man. I have a ton of introspective access to myself and I can tell that this 'rationalization' is actually a pretty good reason to believe X and I trust that my reasoning process is real."

But... eventually I realized I just actually had two motivations going on. When I introspected, I was running a check for a positive result on "is Ray displaying rational thought?". When they extrospected me (i.e. reading my facial expressions), they were checking for a positive result on "does Ray seem biased in this particular way?".

And both checks totally returned 'true', and that was an accurate assessment. 

The partic... (read more)

4Rohin Shah
I think we're imagining different toy mathematical models. Your model, according to me: 1. There is a space of possible approaches, that we are searching over to find a solution. (E.g. the space of all possible programs.) 2. We put a layer of abstraction on top of this space, characterizing approaches by N different "features" (e.g. "is it goal-directed", "is it an oracle", "is it capable of destroying the world") 3. Because we're bounded agents, we then treat the features as independent, and search for some combination of features that would comprise a solution. I agree that this procedure has a systematic error in claiming that there is a solution when none exists (and doesn't have the opposite error), and that if this were an accurate model of how I was reasoning I should be way more worried about correcting for that problem. My model: 1. There is a probability distribution over "ways the world could be". 2. We put a layer of abstraction on top of this space, characterizing "ways the world could be" by N different "features" (e.g. "can you get human-level intelligence out of a pile of heuristics", "what are the returns to specialization", "how different will AI ontologies be from human ontologies"). We estimate the marginal probability of each of those features. 3. Because we're bounded agents, when we need the joint probability of two or more features, we treat them as independent and just multiply. 4. Given a proposed solution, we estimate its probability of working by identifying which features need to be true of the world for the solution to work, and then estimate the probability of those features (using the method above). I claim that this procedure doesn't have a systematic error in the direction of optimism (at least until you add some additional details), and that this procedure more accurately reflects the sort of reasoning that I am doing.
Huh, why doesn't that procedure have that systematic error? Like, when I try to naively run your steps 1-4 on "probability of there existing a number that's both even and odd", I get that about 25% of numbers should be both even and odd, so it seems pretty likely that it'll work out given that there are at least 4 numbers. But I can't easily construct an argument at a similar level of sophistication that gives me an underestimate. [Like, "probability of there existing a number that's both odd and prime" gives the wrong conclusion if you buy that the probability that a natural number is prime is 0, but this is because you evaluated your limits in the wrong order, not because of a problem with dropping all the covariance data from your joint distribution.] My first guess is that you think I'm doing the "ways the world could be" thing wrong--like, I'm looking at predicates over numbers and trying to evaluate a predicate over all numbers, but instead I should just have a probability on "universe contains a number that is both even and odd" and its complement, as those are the two relevant ways the world can be.  My second guess is that you've got a different distribution over target predicates; like, we can just take the complement of my overestimate ("probability of there existing no numbers that are both even and odd") and call it an underestimate. But I think I'm more interested in 'overestimating existence' than 'underestimating non-existence'. [Is this an example of the 'additional details' you're talking about?] Also maybe you can just exhibit a simple example that has an underestimate, and then we need to think harder about how likely overestimates and underestimates are to see if there's a net bias.
2Rohin Shah
It's the first guess. I think if you have a particular number then I'm like "yup, it's fair to notice that we overestimate the probability that x is even and odd by saying it's 25%", and then I'd say "notice that we underestimate the probability that x is even and divisible by 4 by saying it's 12.5%". I agree that if you estimate a probability, and then "perform search" / "optimize" / "run n copies of the estimate" (so that you estimate the probability as 1 - (1 - P(event))^n), then you're going to have systematic errors. I don't think I'm doing anything that's analogous to that. I definitely don't go around thinking "well, it seems 10% likely that such and such feature of the world holds, and so each alignment scheme I think of that depends on this feature has a 10% chance of working, therefore if I think of 10 alignment schemes I've solved the problem". (I suspect this is not the sort of mistake you imagine me doing but I don't think I know what you do imagine me doing.)
Cool, I like this example. I think the thing I'm interested in is "what are our estimates of the output of search processes?". The question we're ultimately trying to answer with a model here is something like "are humans, when they consider a problem that could have attempted solutions of many different forms, overly optimistic about how solvable those problems are because they hypothesize a solution with inconsistent features?" The example of "a number divisible by 2 and a number divisible by 4" is an example of where the consistency of your solution helps you--anything that satisfies the second condition is already satisfying the first condition. But importantly the best you can do here is ignore superfluous conditions; they can't increase the volume of the solution space. I think this is where the systematic bias is coming from (that the joint probability of two conditions can't be higher than the maximum of those two conditions, where the joint probability can be lower than the minimum of the two, and so the product isn't an unbiased estimator of the joint).   For example, consider this recent analysis of cultured meat, which seems to me to point out a fundamental inconsistency of this type in people's plans for creating cultured meat. Basically, the bigger you make a bioreactor, the better it looks on criteria ABC, and the smaller you make a bioreactor, the better it looks on criteria DEF, and projections seem to suggest that massive progress will be made on all of those criteria simultaneously because progress can be made on them individually. But this necessitates making bioreactors that are simultaneously much bigger and much smaller! [Sometimes this is possible, because actually one is based on volume and the other is based on surface area, and so when you make something like a zeolite you can combine massive surface area with tiny volume. But if you need massive volume and tiny surface area, that's not possible. Anyway, in this case, my read is that
5Rohin Shah
Re: cultured meat example: If you give me examples in which you know the features are actually inconsistent, my method is going to look optimistic when it doesn't know about that inconsistency. So yeah, assuming your description of the cultured meat example is correct, my toy model would reproduce that problem. To give a different example, consider OpenAI Five. One would think that to beat Dota, you need to have an algorithm that allows you to do hierarchical planning, state estimation from partial observability, coordination with team members, understanding of causality, compression of the giant action space, etc. Everyone looked at this giant list of necessary features and thought "it's highly improbable for an algorithm to demonstrate all of these features". My understanding is that even OpenAI, the most optimistic of everyone, thought they would need to do some sort of hierarchical RL to get this to work. In the end, it turned out that vanilla PPO with reward shaping and domain randomization was enough. It turns out that all of these many different capabilities / features were very consistent with each other and easier to achieve simultaneously than we thought. Tbc, I don't want to claim "unbiased estimator" in the mathematical sense of the phrase. To even make such a claim you need to choose some underlying probability distribution which gives rise to our features, which we don't have. I'm more saying that the direction of the bias depends on whether your features are positively vs. negatively correlated with each other and so a priori I don't expect the bias to be in a predictable direction. They definitely have that problem. I'm not sure how you don't have that problem; you're always going to have some amount of abstraction and some amount of inconsistency; the future is hard to predict for bounded humans, and you can't "fully populate the details" as an embedded agent. If you're asking how you notice any inconsistencies at all (rather than all of the inc
To me it seems like this is what you should expect other people to look like both when other people know less about a domain than you do, and also when you're overconfident about your understanding of that domain. So I don't think it helps distinguish those two cases. (Also, to me it seems like a similar thing happens, but with the positions reversed, when Paul and Eliezer try to forecast concrete progress in ML over the next decade. Does that seem right to you?) I believe this was discussed further at some point - I argued that Eliezer-style political history books also exclude statements like "and then we survived the cold war" or "most countries still don't have nuclear energy".  
It feels similar but clearly distinct? Like, in that situation Eliezer often seems to say things that I parse as "I don't have any special knowledge here", which seems like a different thing than "I can't easily sample from my distribution over how things go right", and I also have the sense of Paul being willing to 'go specific' and Eliezer not being willing to 'go specific'. You're thinking of this bit of the conversation, starting with: (Or maybe a bit earlier and later, but that was my best guess for where to start the context.) The main quotes from the middle that seems relevant: and ending with: Rereading that section, my sense is that it reads like a sort of mirror of the Eliezer->Paul "I don't know how to operate your view" section; like, Eliezer can say "I think nukes are less worrying for reasons ABC, also you can observe me being not worried about other things-people-are-concerned-by XYZ", but I wouldn't have expected you (or the reader who hasn't picked up Eliezer-thinking from elsewhere) to have been able to come away from that with why you trying to be Eliezer from 1930s would have thought 'and then it turned out okay' would have been a political-history-book-sentence, or the relative magnitudes of the surprise. [Like, I think my 1930s-Eliezer puts like 3-30% on "and then it turned out okay" for nukes, and my 2020s-Eliezer puts like 0.03-3% on that for AGI? But it'd be nice to hear if Eliezer thinks AGI turning out as well as nukes is like 10x the surprise of nukes turning out this well conditioned on pre-1930s, or more like 1000x the surprise.]
This is a very interesting point! I will chip in by pointing out a very similar remark from Rohin just earlier today: That is all. (Obviously there's a kinda superficial resemblance here to the phenomenon of "calling out" somebody else; I want to state outright that this is not the intention, it's just that I saw your comment right after seeing Rohin's comment, in such a way that my memory of his remark was still salient enough that the connection jumped out at me. Since salient observations tend to fade over time, I wanted to put this down before that happened.)
Yeah, I'm also interested in the question of "how do we distinguish 'sentences-on-mainline' from 'shoring-up-edge-cases'?", or which conversational moves most develop shared knowledge, or something similar.  Like I think it's often good to point out edge cases, especially when you're trying to formalize an argument or look for designs that get us out of this trap. In another comment in this thread, I note that there's a thing Eliezer said that I think is very important and accurate, and also think there's an edge case that's not obviously handled correctly.  But also my sense is that there's some deep benefit from "having mainlines" and conversations that are mostly 'sentences-on-mainline'? Or, like, there's some value to more people thinking thru / shooting down their own edge cases (like I do in the mentioned comment), instead of pushing the work to Eliezer. I'm pretty worried that there are deeply general reasons to expect AI alignment to be extremely difficult, people aren't updating on the meta-level point and continue to attempt 'rolling their own crypto', asking if Eliezer can poke the hole in this new procedure, and if Eliezer ever decides to just write serial online fiction until the world explodes humanity hasn't developed enough capacity to replace him.
[-]Rohin ShahΩ13230

(For object-level responses, see comments on parallel threads.)

I want to push back on an implicit framing in lines like:

there's some value to more people thinking thru / shooting down their own edge cases [...], instead of pushing the work to Eliezer.

people aren't updating on the meta-level point and continue to attempt 'rolling their own crypto', asking if Eliezer can poke the hole in this new procedure

This makes it sound like the rest of us don't try to break our proposals, push the work to Eliezer, agree with Eliezer when he finds a problem, and then not update that maybe future proposals will have problems.

Whereas in reality, I try to break my proposals, don't agree with Eliezer's diagnoses of the problems, and usually don't ask Eliezer because I don't expect his answer to be useful to me (and previously didn't expect him to respond). I expect this is true of others (like Paul and Richard) as well.

Yeah, sorry about not owning that more, and for the frame being muddled. I don't endorse the "asking Eliezer" or "agreeing with Eliezer" bits, but I do basically think he's right about many object-level problems he identifies (and thus people disagreeing with him about that is not a feature) and think 'security mindset' is the right orientation to have towards AGI alignment. That hypothesis is a 'worry' primarily because asymmetric costs means it's more worth investigating than the raw probability would suggest. [Tho the raw probability of components of it do feel pretty substantial to me.] [EDIT: I should say I think ARC's approach to ELK seems like a great example of "people breaking their own proposals". As additional data to update on, I'd be interested in seeing, like, a graph of people's optimism about ELK over time, or something similar.]

But also my sense is that there's some deep benefit from "having mainlines" and conversations that are mostly 'sentences-on-mainline'?

I agree with this. Or, if you feel ~evenly split between two options, have two mainlines and focus a bunch on those (including picking at cruxes and revising your mainline view over time).


Like, it feels to me like Eliezer was generating sentences on his mainline, and Richard was responding with 'since you're being overly pessimistic, I will be overly optimistic to balance', with no attempt to have his response match his own mainline.

I do note that there are some situations where rushing to tell a 'mainline story' might be the wrong move:

  • Maybe your beliefs feel wildly unstable day-to-day -- because you're learning a lot quickly, or because it's just hard to know how to assign weight to the dozens of different considerations that bear on these questions. Then trying to take a quick snapshot of your current view might feel beside the point.
    • It might even feel actively counterproductive, like rushing too quickly to impose meaning/structure on data when step one is to make sure you have the data properly loaded up in your head.
  • Maybe there are many scen
... (read more)
2[comment deleted]

These conversations are great and I really admire the transparency. It's really nice to see discussions that normally happen in private happen instead in public where everyone can reflect, give feedback, and improve their own thoughts. On the other hand, the combined conversations combined to a decent-sized novel - LW says 198,846 words! Is anyone considering investing heavily in summarizing the content for people to get involved without having to read all that content?

Echoing that I loved these conversations and I'm super grateful to everyone who participated — especially Richard, Paul, Eliezer, Nate, Ajeya, Carl, Rohin, and Jaan, who contributed a lot.

I don't plan to try to summarize the discussions or distill key take-aways myself (other than the extremely cursory job I did on, but I'm very keen on seeing others attempt that, especially as part of a process to figure out their own models and do some evaluative work.

I think I'd rather see partial summaries/responses that go deep, instead of a more exhaustive but shallow summary; and I'd rather see summaries that center the author's own view (what's your personal take-away? what are your objections? which things were small versus large updates? etc.) over something that tries to be maximally objective and impersonal. But all the options seem good to me.

7Ben Pace
I chatted briefly the other day with Rob Bensinger about me turning them into a little book. My guess is I'd want to do something to compress especially the long Paul/Eliezer bet hashing out, that felt super long to me and not all worth the reading. Interested in other suggestions for compression. (This is not a commitment to do this, I probably won't.)
I wish you (or someone) would make a little book of this.
The compression idea evokes Kaj Sotala's summary/analysis of the AI-Foom Debate (which I found quite useful at the time). I support the idea, especially given it has taken a while for the participants to settle on things cruxy enough to discuss and so on. Though I would also be interested in "look, these two disagree on that, but look at all the very fundamental things about AI alignment they agree on".
6Daniel Kokotajlo
Here is a heavily condensed summary of the takeoff speeds thread of the conversation, incorporating earlier points made by Hanson, Grace, etc. :) (kudos to Ben Goldhaber for pointing me to it)

Question for Richard, Paul, and/or Rohin: What's a story, full of implausibly concrete details but nevertheless a member of some largish plausible-to-you cluster of possible outcomes, in which things go well? (Paying particular attention to how early AGI systems are deployed and to what purposes, or how catastrophic deployments are otherwise forstalled.)

[-]Rohin ShahΩ10130

I wrote this doc a couple of years ago (while I was at CHAI). It's got many rough edges (I think I wrote it in one sitting and never bothered to rewrite it to make it better), but I still endorse the general gist, if we're talking about what systems are being deployed to do and what happens amongst organizations. It doesn't totally answer your question (it's more focused on what happens before we get systems that could kill everyone), but it seems pretty related.

(I haven't brought it up before because it seems to me like the disagreement is much more in the "mechanisms underlying intelligence", which that doc barely talks about, and the stuff it does say feels pretty outdated; I'd say different things now.)

If I didn't miss anything and I'm understanding the scenario correctly, then for this part: I'd expect that interpretability tools, if they work, would tell you "yup, this AI is planning to kill you as soon as it possibly can", without giving you a way to fix that (that's robust to capability gains). Ie this story still seems to rely on an unexplained step that goes "... and a miracle occurs where we fundamentally figure out how to align AI just in the nick of time".
3Rohin Shah
Totally agreed that the doc does not address that argument. Quoting from my original comment:
[-]Ben PaceΩ9250

Eliezer and Nate, my guess is that most of your perspective on the alignment problem for the past several years has come from the thinking and explorations you've personally done, rather than reading work done by others.

But, if you have read interesting work by others that's changed your mind or given you helpful insights, what has it been? Some old CS textbook? Random Gwern articles? An economics textbook? Playing around yourself with ML systems?

One thing in the posts I found surprising was Eliezers assertion that you needed a dangerous superintelligence to get nanotech. If the AI is expected to do everything itself, including inventing the concept of nanotech, I agree that this is dangerously superintelligent. 

However, suppose Alpha Quantum can reliably approximate the behaviour of almost any particle configuration. Not literally any, it can't run a quantum computer factorizing large numbers better than factoring algorithms, but enough to design a nanomachine. (It has been trained to approximate the ground truth of quantum mechanics equations, and it does this very well.) 

For example, you could use IDA, start training to imitate a simulation of a handful of particles, then compose several smaller nets into one large one. 

Add a nice user interface and we can drag and drop atoms. 

You can add optimization, gradient descent trying to maximize the efficiency of a motor, or minimize the size of a logic gate. All of this is optimised to fit a simple equation, so assuming you don't have smart general mesaoptimizers forming, and deducing how to manipulate humans based on very little info about humans, you shoul... (read more)

(Not very sure I understood your description right, but here is my take:) * I think your proposal is not explaining some crucial steps, which are in fact hard. In particular, I understood it as "you have AI which can give you blueprints for nano sized machines". But I think we already have some blueprints, this isn't an issue. How we assemble them is an issue. * I expect that there will be more issues like this that you would find if you tried writing the plan in more detail. However, I share the general sentiment behind your post --- I also don't understand why you can't get some pivotal act by combining human intelligence with some narrow AI. I expect that Eliezer have tried to come up with such combinations and came away with some general takeaways on this being not realistic. But I haven't done this exercise, so it seems not obvious to me. Perhaps it would be beneficial if many more people tried doing the exercise and then communicated the takeaways.
2Rob Bensinger
I think it would be!
Uh, how big do you think contemporary chips are?
1Donald Hobson
Like 10s of atoms across. So you aren't scaling down that much. (Most of your performance gains are in being able to stack your chips or whatever.
1Gram Stone
I got the impression Eliezer's claiming that a dangerous superintelligence is merely sufficient for nanotech. How would you save us with nanotech? It had better be good given all the hardware progress you just caused!
4Rob Bensinger
No, I'm pretty confident Eliezer thinks AGI is both necessary and sufficient for nanotech. (Realistically/probabilistically speaking, given plausible levels of future investment into each tech. Obviously it's not logically necessary or sufficient.) Cf. my summary of Nate's view in Nate's reply to Joe Carlsmith: (I read "sphexish" here as a special case of "narrow AI" / "shallow cognition", doing more things as a matter of pre-programmed reflex rather than as a matter of strategic choice.)

I wrote Consequentialism & Corrigibility shortly after and partly in response to the first (Ngo-Yudkowsky) discussion. If anyone has an argument or belief that the general architecture / approach I have in mind (see the “My corrigibility proposal sketch” section) is fundamentally doomed as a path to corrigibility and capability—as opposed to merely “reliant on solving lots of hard-but-not-necessarily-impossible open problems”—I'd be interested to hear it. Thanks in advance. :)


After reading some of the newer MIRI dialogues, I'm less convinced than I once was that I know what "corrigibility" actually is. Could you say a few words about what kind of behavior you concretely expect to see from a "corrigible" agent, followed by how [you expect] those behaviors [to] fit into the "trajectory-constraining" framework you propose in your post?

EDIT: This is not purely a question for Steven, incidentally (or at least, the first half isn't); anyone else who wants to take a shot at answering should feel free to do so. In particular I'd be interested in hearing answers from Eliezer or anyone else historically involved in the invention of the term.

6Ben Pace
My understanding: a corrigible paperclip-maximizer does all the paperclip-maximizing, but then when you realize it's gonna end the world, you go to turn it off, and it doesn't stop you. It's corrigible!
There are a bunch of different definitions, but if you're asking for Eliezer's version, then the arbital expoisition is quite good. N.B. we don't have a model for this sort of corrigibility.  EDIT: Be warned, these are rough summaries of the defs. I'd ammend the CHAI def I cited to "the AI obeys more when it knows less, models you as more rational, and the downsides of disobedience are lesser". But people at CHAI have diverse views, so this is not the definitive CHAI take. Other definitions include some people at CHAI's definition (the AI obeys you whilst it doesn't know what its utility function is), the definition used in the reward tampering paper (near the same as EY's original def, barring the honesty clause, and formalised in a causal diagram setting), Stuart Armstrong's many definitions which most notably includes Utility Indifference (note the agent is NOT a standard R-maximiser) so it accepts having its utility function changed at a later time as you're going to compensate it for its loss in utility. So it is indiferent to the change (this doesn't mean it won't kill you for spare parts though). And TurnTrout has what looks like some interesting thoughts on the topic here but I haven't read those yet.  Edit2: Paul thinks corrigibility has a simpler core than alignment, but is quite messy, and we won't get a crisp algorithm for it. But the intuition is the same as what Eliezer was pointing to, namely that the AI knows it should defer to the human, and will seek to preserve that deference in it and its offspring. Plus being honest and helpful. Here is a post where he rambles about it.
I'm a little confused what it hopes to accomplish. I mean, to start I'm a little confused by your example of "preferences not about future states" (i.e. 'the pizza shop employee is running around frantically, and I am laughing' is a future state). But to me, I'm not sure what the mixing of "paperclips" vs "humans remain in control" accomplishes. On the one hand, I think if you can specify "humans remain in control" safely, you've solved the alignment problem already. On another, I wouldn't want that to seize the future: There are potentially much better futures where humans are not in control, but still alive/free/whatever. (e.g. the Sophotechs in the Golden Oecumene are very much in control). On a third, I would definitely, a lot, very much, prefer a 3 star 'paperclips' and 5 star 'humans in control' to a 5 star 'paperclips' and a 3 star 'humans in control', even though both would average 4 stars?
2Steven Byrnes
In my post I wrote: “To be more concrete, if I’m deciding between two possible courses of action, A and B, “preference over future states” would make the decision based on the state of the world after I finish the course of action—or more centrally, long after I finish the course of action. By contrast, “other kinds of preferences” would allow the decision to depend on anything, even including what happens during the course-of-action.” So “the humans will ultimately wind up in control” would be a preference-over-future-states, and this preference would allow (indeed encourage) the AGI to disempower and later re-empower humans. By contrast, “the humans will remain in control” is not a pure preference-over-future-states, and relatedly does not encourage the AGI to disempower and later re-empower humans. If we knew exactly what long-term future we wanted, and we knew how to build an AGI that definitely also wanted that exact same long-term future, then we should certainly do that, instead of making a corrigible AGI. Unfortunately, we don't know those things right now, so under the circumstances, knowing how to make a corrigible AGI would be a useful thing to know how to do. Also, this is not a hyper-specific corrigibility proposal; it's really a general AGI-motivation-sculpting proposal, applied to corrigibility. So even if you're totally opposed to corrigibility, you can still take an interest in the question of whether or not my proposal is fundamentally doomed. Because I think everyone agrees that AGI-motivation-sculpting is necessary. It could be a weighted average. It could be a weighted average plus a nonlinear acceptability threshold on “humans in control”. It could be other things. I don't know; this is one of many important open questions. See discussion under “Objection 1” in my post.
Am I correct after reading this that this post is heavily related to embedded agency? I may have misunderstood the general attitudes, but I thought of "future states" as "future to now" not "future to my action." It seems like you couldn't possibly create a thing that works on the last one, unless you intend it to set everything in motion and then terminate. In the embedded agency sequence, they point out that embedded agents don't have well defined i/o channels. One way is that "action" is not a well defined term, and is often not atomic.  It also sounds like you're trying to suggest that we should be judging trajectories, not states? I just want to note that this is, as far as I can tell, the plan:  From the synopsis of High Challenge I'm not sure I interpret corrigibility as exactly the same as "preferring the humans remain in control" (I see you suggest this yourself in Objection 1, I wrote this before I reread that, but I'm going to leave it as is) and if you programmed that preference into a non-corrigible AI, it would still seize the future into states where the humans have to remain in control. Better than doom, but not ideal if we can avoid it with actual corrigibility. But I think I miscommunicated, because, besides the above, I agree with everything else in those two paragraphs. I think I maintain that this feels like it doesn't solve much. Much of the discussion in the Yudkowsky conversations was that there's a concern on how to point powerful systems in any direction. Your response to objection 1 admits you don't claim this solves that, but that's most of the problem. If we do solve the problem of how to point a system at some abstract concept, why would we choose "the humans remain in control" and not "pursue humanity's CEV"? Do you expect "the humans remain in control" (or the combination of concepts you propose as an alternative) to be easier to define? Easier enough to de

Question for anyone, but particularly interested in hearing from Christiano, Shah, or Ngo: any thoughts on what happens when alignment schemes that worked in lower-capability regimes fail to generalize to higher-capability regimes?

For example, you could imagine a spectrum of outcomes from "no generalization" (illustrative example: galaxies tiled with paperclips) to "some generalization" (illustrative example: galaxies tiled with "hedonium" human-ish happiness-brainware) to "enough generalization that existing humans recognizably survive, but something still went wrong from our current perspective" (illustrative examples: "Failed Utopia #4-2", Friendship Is Optimal, "With Folded Hands"). Given that not every biological civilization solves the problem, what does the rest of the multiverse look like? (How is measure distributed on something like my example spectrum, or whatever I should have typed instead?)

(Previous work: Yudkowsky 2009 "Value Is Fragile", Christiano 2018 "When Is Unaligned AI Morally Valuable?", Grace 2019 "But Exactly How Complex and Fragile?".)

When alignment schemes fail to scale, I think it typically means that they work while the system is unable to overpower/outsmart the oversight process, and then break down when the system becomes able to do so. I think that this usually results in the AI shifting from behavior that is mostly constrained by the training process to behavior that is mostly unconstrained (once they effectively disempower humans).

I think the results are relatively unlikely to be good in virtue of "the AI internalized something about our values, just not everything", and I'm pretty skeptical of recognizable "near miss" scenarios rather than AI gradually careening in very hard-to-predict directions with minimal connection with the surface features of the training process. 

Overall I think that the most likely outcome is a universe that is orthogonal to anything we directly care about, maybe with a vaguely similar flavor owing to convergence depending on how AI motivations shake out. (But likely not close enough to feel great, and quite plausibly with almost no visible relation. Probably much more different from us than we are from aliens.)

I think it's fairly plausible that the results are OK just beca... (read more)

  Also, wouldn't you expect s-risks from this to be very unlikely by virtue of (1) civilizations like this being very unlikely to have substantial measure over the universe's resources, (2) transparency making bargaining far easier, and (3) few technologically advanced civilizations would care about humans suffering in particular as opposed to e.g. an adversary running emulations of their own species?

Basically agree with Paul, and I especially want to note that I've barely thought about it and so this would likely change a ton with more information. To put some numbers of my own:

  • "No generalization": 65%
  • "Some generalization": 5% (I don't actually have stories where this is an outcome; this is more like model uncertainty)
  • "Lots of generalization, but something went wrong": 30%

These are from my own perspective of what these categories mean, which I expect are pretty different from yours -- e.g. maybe I'm at ~2% that upon reflection I'd decide that hedonium is great and so that's actually perfect generalization; in the last category I include lots of worlds that I wouldn't describe as "existing humans recognizably survive", e.g. we decide to become digital uploads, then get lots of cognitive enhancements, throw away a bunch of evolutionary baggage, but also we never expand to the stars because AI has taken control of it and given us only Earth.

I think the biggest avenues for improving the answers would be to reflect more on the kindness + cooperation and acausal trade stories Paul mentions, as well as the possibility that a few AIs end up generalizing close to correctly and working ... (read more)

This doesn't contradict anything you're saying but there's arguably a wager for thinking that we're on the knife-edge – our actions are more impactful if we are.  [Edit to add point:] The degree to which any particular training approach generalizes is of course likely a fixed fact (like in the Lesswrong post you link to about fire). But different civilizations could try different training approaches, which produces heterogeneity for the multiverse.

I finished reading all the conversations a few hours ago. I have no follow-up questions (except maybe "now what?"), I'm still updating from all those words.

One except in particular, from the latest post, jumped at me (from Eliezer Yudkowsky, emphasis mine):

This is not aimed particularly at you, but I hope the reader may understand something of why Eliezer Yudkowsky goes about sounding so gloomy all the time about other people's prospects for noticing what will kill them, by themselves, without Eliezer constantly hovering over their shoulder every minute prompting them with almost all of the answer.

The past years or reading about alignment have left me with an intense initial distrust of any alignment research agenda. Maybe it's ordinary paranoia, maybe something more. I've not come up with any new ideas myself, and I'm not particularly confident in my ability to find flaws in someone else's proposal (what if I'm not smart enough to understand them properly? What if I make things even more confused and waste everyone's time?)

After thousands and thousands of lengthy conversations where it takes everyone ages to understand where threat models disagree, why some avenue of research is p... (read more)

Not sure if it's a right place to ask, instead of just googling it, but anyway: does anyone know what's the current state of AI security practices at DeepMind, OpenAI and other such places? Like, did they estimate probability of GPT-3 killing everyone before turning it on, do they have procedures for not turning something on, did they test these procedures by someone impersonating unaligned GPT and trying to manipulate researchers, things like that?

8Rohin Shah
No, I very strongly predict they did not do things like that. I expect they (perhaps implicitly) predicted with high confidence that GPT-3 would not have the capabilities needed to kill everyone.
Do they have plans to do something in the future?
4Rohin Shah
I would assume that the safety teams plan to do this (certainly I plan to). It's less clear what the opinions are outside of the safety teams.
[-]Ben PaceΩ4140

Questions about the standard-university-textbook from the future that tells us how to build an AGI. I'll take answers on any of these!

  1. Where is ML in this textbook? Is it under a section called "god-forsaken approaches" or does it play a key role? Follow-up: Where is logical induction?
  2. If running superintelligent AGIs didn't kill you and death was cancelled in general, how long would it take you to write the textbook?
  3. Is there anything else you can share about this textbook? Do you know any of the other chapter names?

I'm going to try and write a table of contents for the textbook, just because it seems like a fun exercise.

Epistemic status: unbridled speculation

Volume I: Foundation

  • Preface [mentioning, ofc, the infamous incident of 2041]
  • Chapter 0: Introduction

Part I: Statistical Learning Theory

  • Chapter 1: Offline Learning [VC theory and Watanabe's singular learning theory are both special cases of what's in this chapter]
  • Chapter 2: Online Learning [infra-Bayesianism is introduced here, Garrabrant induction too]
  • Chapter 3: Reinforcement Learning
  • Chapter 4: Lifelong Learning [this chapter deals with traps, unobservable rewards and long-term planning]

Part II: Computational Learning Theory

  • Chapter 5: Algebraic Classes [the theory of SVMs is a special case of what's explained here]
  • Chapter 6: Circuits [learning various class of circuits]
  • Chapter 7: Neural Networks
  • Chapter 8: ???
  • Chapter 9: Reflective Learning [some version of Turing reinforcement learning comes here]

Part III: Universal Priors

  • Chapter 10: Solomonoff's Prior [including regret analysis using algorithmic statistics]
  • Chapter 11: Bounded Simplicity Priors
  • Chapter 12: ??? [might involve: causality, time hierarchies, logical langu
... (read more)

I don't think there is an "AGI textbook" any more than there is an "industrialization textbook." There are lots of books about general principles and useful kinds of machines. That said, if I had to make wild guesses about roughly what that future understanding would look like:

  1. There is a recognizable concept of "learning" meaning something like "search for policies that perform well in past or simulated situations." That plays a large role, comparably important to planning or Bayesian inference. Logical induction is likely an elaboration of Bayesian inference that receives relatively little airtime except in specialized discussions.
  2. This one is tougher given that I don't know what "the textbook" is. And I guess in the story all other humans are magically disappeared? If I was stuck with a single AWS cluster from 2022 and given unlimited time, I'd wildly guess that it would take me something between 1e4 and 1e8 years to create an autopoetic AI that obsoleted my own contributions (mostly because serial time is extremely valuable and I have a lot of compute). Writing the textbook does not seem like very much work after having done the deed?
  3. I'd roughly guess big sections on learning, inference, planning, alignment, and clever algorithms for all of the above. I'd guess maybe 50% of content is smart versions of stuff we know now and 50% is stuff we didn't figure out at the time, but it depends a lot on how you define this textbook.
3Rohin Shah
I'm mostly going to answer assuming that there's not some incredibly different paradigm (i.e. something as different from ML as ML is from expert systems). I do think the probability of "incredibly different paradigm" is low. I'm also going to answer about the textbook at, idk, the point at which GDP doubles every 8 years. (To avoid talking about the post-Singularity textbook that explains how to build a superintelligence with clearly understood "intelligence algorithms" that can run easily on one of today's laptops, which I know very little about.) I think I roughly agree with Paul if you are talking about the textbook that tells us how to build the best systems for the tasks that we want to do. (Analogy: today's textbook for self-driving cars.) That being said, I think that much of the improvement over time will be driven by improvements specifically in ML. (Analogy: today's textbook for deep learning.) So we can talk about that textbook as well. 1. It's a textbook that's entirely about "finding good programs through a large, efficient search with a stringent goal", which we currently call ML. The content may be primarily some new approach for achieving this, with neural nets being a historical footnote, or it might be entirely about neural nets (though presumably with new architectures or other changes from today). Logical induction doesn't appear in the textbook. 2. Jeez, who knows. If I intuitively query my brain here, it mostly doesn't have an answer; a thousand vs. million vs. billion years don't really change my intuitive predictions about what I'd get done. So we can instead back it out from other estimates. Given timelines of 10^1 - 10^2 years, and, idk, ~10^6 humans working on the problem near the end, seems like I'm implicitly predicting ~10^7 human-years of effort in our actual world. Then you have to adjust for a ton of factors, e.g. my quality relative to the average, the importance of serial thinking time, the benefit that real-world humans get
1. Where is ML in this textbook? Is it under a section called "god-forsaken approaches" or does it play a key role? Follow-up: Where is logical induction? Key role, but most current ML is in the "applied" section, where the "theory" section instead explains the principles by which neural nets (or future architectures) work on the inside. Logical induction is a sidebar at some point explaining the theoretical ideal we're working towards, like I assume AIXI is in some textbooks. 1. Is there anything else you can share about this textbook? Do you know any of the other chapter names? Planning, Abstraction, Reasoning, Self-awareness.

Eliezer, do you have any advice for someone wanting to enter this research space at (from your perspective) the eleventh hour? I’ve just finished a BS in math and am starting a PhD in CS, but I still don’t feel like I have the technical skills to grapple with these issues, and probably won’t for a few years. What are the most plausible routes for someone like me to make a difference in alignment, if any?

I don't have any such advice at the moment.  It's not clear to me what makes a difference at this point.

We'd absolutely pay him if he showed up and said he wanted to work on the problem.  Every time I've asked about trying anything like this, all the advisors claim that you cannot pay people at the Terry Tao level to work on problems that don't interest them.  We have already extensively verified that it doesn't particularly work for eg university professors.

Every time I've asked about trying anything like this, all the advisors claim that you cannot pay people at the Terry Tao level to work on problems that don't interest them.

As I am sure you would agree, Neumann/Tao-level people are a very different breed from even very, very, very good professors. It is plausible they are significantly more sane than the average genius. 

Given the enormous glut of money in EA trying to help here and the terrifying thing where a lot of the people who matter have really short timelines, I think it is worth testing this empirically with Tao himself and Tao-level people. 

It is worth noting that Neumann occasionally did contract work for extraordinary sums. 


I'm not sure whether the unspoken context of this comment is "We tried to hire Terry Tao and he declined, citing lack of interest in AI alignment" vs "we assume, based on not having been contacted by Terry Tao, that he is not interested in AI alignment."

If the latter: the implicit assumption seems to be that if Terry Tao would find AI alignment to be an interesting project, we should strongly expect him to both know about it and have approached MIRI regarding it, neither which seems particularly likely given the low public profile of both AI alignment in general and MIRI in particular.

If the former: bummer.

You're probably already aware of this, but just in case not:

Demis Hassabis said the following about getting Terrence Tao to work on AI safety:

I always imagine that as we got closer to the sort of gray zone that you were talking about earlier, the best thing to do might be to pause the pushing of the performance of these systems so that you can analyze down to minute detail exactly and maybe even prove things mathematically about the system so that you know the limits and otherwise of the systems that you're building. At that point I think all the world's greatest minds should probably be thinking about this problem. So that was what I would be advocating to you know the Terence Tao’s of this world, the best mathematicians. Actually I've even talked to him about this—I know you're working on the Riemann hypothesis or something which is the best thing in mathematics but actually this is more pressing. I have this sort of idea of like almost uh ‘Avengers assembled’ of the scientific world because that's a bit of like my dream.

The header image of Tao's blog is a graph representing "flattening the curve" of the Covid-19 spread. One avenue for convincing elite talent that alignment is a problem is a media campaign that brings the problem of alignment into popular consciousness.

I have some ideas about how this might begin. "Educational" YouTuber CGP Grey, (5.2M subscribers) got talked into making a pair of videos advocating for anti-aging research by another large YouTuber, Kurzgesagt (18M subscribers). I'd bet that they could both be persuaded into making AI alignment videos.

6Matthew Barnett
Can you clarify what the term "Pascal-mugged" means in your comment?  From what I can tell, the main reason why CGP Grey made those videos was because he's had a long-running desire to live to see the future. He's talked about it in his podcast with Brady Haran and it's hinted at in some of his older videos. I don't think there was much more to it than that. As for Kurzgesagt, I believe it was coordinated with Keith Comito, the president of As for why they suddenly decided to coordinate with, I have little idea. However, since their video was launched with conjunction with CGP Grey, it's possible that CGP Grey was the first one to bring it up, after which Kurzgesagt reached out to people who could help.
Edited, poor choice. I think they discuss it around here.
I think CGP Grey recommended Bostrom's Superintelligence in a podcast once. Edit: Source
Grey’s 2014 video Humans Need Not Apply, about humans being skilled-out of the economy, was the first introduction for me and probably lots of other people to the idea that AI might cause problems. I’m sure he’d be up for making a video about alignment.
What do you think about trying to actually interest him?
1Noah Topper
What do you think about ironically hiring Terry Tao?

Not even a "In 90% of possible worlds, we're irreversibly doomed, but in the remaining 10%, here's the advice that would work"?

[-]Ben PaceΩ3130

Eliezer, when you told Richard that your probability of a successful miracle is very low, you added the following note:

Though a lot of that is dominated, not by the probability of a positive miracle, but by the extent to which we seem unprepared to take advantage of it, and so would not be saved by one.

I don't mean to ask for positive fairy tales when I ask: could you list some things you could see in the world that would cause you to feel that we were well-prepared to take advantage of one if we got one?

My obvious quick guess would be "I know of an ML project that made a breakthrough as impressive as GPT-3 and this is secret to the outer world, and the organization is keenly interested in alignment". But I am also interested in broader and less obvious ones. For example if the folks around here had successfully made a covid vaccine I think that would likely require us to be in a much more competent and responsive situation. Alternatively if folks made other historic scientific breakthroughs guided by some model of how it helps prevent AI doom, I'd feel more like this power could be turned to relevant directions.

Anyway, these are some of the things I quickly generate, but I'm interested in what comes to your mind?


Curated. I found the entire sequence of conversations quite valuable, and it seemed good both to let people know it had wrapped up, and curate it while the AMA was still going on.

Question from evelynciara on the EA Forum:

Do you believe that AGI poses a greater existential risk than other proposed x-risk hazards, such as engineered pandemics? Why or why not?


For sure. It's tricky to wipe out humanity entirely without optimizing for that in particular -- nuclear war, climate change, and extremely bad natural pandemics look to me like they're at most global catastrophes, rather than existential threats. It might in fact be easier to wipe out humanity by enginering a pandemic that's specifically optimized for this task (than it is to develop AGI), but we don't see vast resources flowing into humanity-killing-virus projects, the way that we see vast resources flowing into AGI projects. By my accounting, most other x-risks look like wild tail risks (what if there's a large, competent, state-funded successfully-secretive death-cult???), whereas the AI x-risk is what happens by default, on the mainline (humanity is storming ahead towards AGI as fast as they can, pouring billions of dollars into it per year, and by default what happens when they succeed is that they accidentally unleash an optimizer that optimizes for our extinction, as a convergent instrumental subgoal of whatever rando thing it's optimizing).

Hi, I'm the user who asked this question. Thank you for responding! I see your point about how an AGI would intentionally destroy humanity versus engineered bugs that only wipe us out "by accident", but that's conditional on the AGI having "destroy humanity" as a subgoal. Most likely, a typical AGI will have some mundane, neutral-to-benevolent goal like "maximize profit by running this steel factory and selling steel". Maybe the AGI can achieve that by taking over an iron mine somewhere, or taking over a country (or the world) and enslaving its citizens, or even wiping out humanity. In general, my guess is that the AGI will try to do the least costly/risky thing needed to achieve its goal (maximizing profit), and (setting aside that if all of humanity were extinct, the AGI would have no one to sell steel to) wiping out humanity is the most expensive of these options and the AGI would likely get itself destroyed while trying to do that. So I think that "enslave a large portion of humanity and export cheap steel at a hefty profit" is a subgoal that this AGI would likely have, but destroying humanity is not. It depends on the use case - a misaligned AGI in charge of the U.S. Armed Forces could end up starting a nuclear war - but given how careful the U.S. government has been about avoiding nuclear war, I think they'd insist on an AGI being very aligned with their interests before putting it in charge of something so high stakes. Also, I suspect that some militaries (like North Korea's) might be developing bioweapons and spending 1 to 100% as much on it annually as OpenAI and DeepMind spend on AGI; we just don't know about it. Based on your AGI-bioweapon analogy, I suspect that AGI is a greater hazard than bioweapons, but not by quite as much as your argument implies. While few well-resourced actors are interested in using bioweapons, a who's who of corporations, states, and NGOs will be interested in using AGI. And AGIs can adopt dangerous subgoals for a wide range

[W]iping out humanity is the most expensive of these options and the AGI would likely get itself destroyed while trying to do that[.]

It would be pretty easy and cheap for something much smarter than a human to kill all humans. The classic scenario is:

A.  [...] The notion of a 'superintelligence' is not that it sits around in Goldman Sachs's basement trading stocks for its corporate masters.  The concrete illustration I often use is that a superintelligence asks itself what the fastest possible route is to increasing its real-world power, and then, rather than bothering with the digital counters that humans call money, the superintelligence solves the protein structure prediction problem, emails some DNA sequences to online peptide synthesis labs, and gets back a batch of proteins which it can mix together to create an acoustically controlled equivalent of an artificial ribosome which it can use to make second-stage nanotechnology which manufactures third-stage nanotechnology which manufactures diamondoid molecular nanotechnology and then... well, it doesn't really matter from our perspective what comes after that, because from a human perspective any technology more advan

... (read more)
There have been a lot of words written about how and why almost any conceivable goal, even a mundane one like "improve efficiency of a steel plant", carelessly specified, can easily result in a hostile AGI. The basic outline of these arguments usually goes something like: 1. The AGI wants to do what you told it ("make more steel"), and will optimize very hard for making as much steel as possible. 2. It also understands human motivations and knows that humans don't actually want as much steel as it is going to make. But note carefully that it wasn't aligned to respect human motivations, it was aligned to make steel. It's understanding of human motivations is part of its understanding of its environment, in the same way as its understanding of metallurgy. It has no interest in doing what humans would want it to do because it hasn't been designed to do that. 3. Because it knows that humans don't want as much steel as it is going to make, it will correctly conclude that humans will try to shut it off as soon as they understand what the AGI is planning to do. 4. Therefore it will correctly reason that its goal of making more steel will be easier to achieve if humans are unable to shut it off. This can lead to all kinds of unwanted actions such as the AGI making and hiding copies of itself everywhere, very persuasively convincing humans that it is not going to make as much steel as it secretly plans to so that they don't try to shut it off, and so on all the way up to killing all humans. Now, "make as much steel as possible" is an exceptionally stupid goal to give an AGI, and no one would likely do that. But every less stupid goal that has been proposed has had plausible flaws pointed out which generally lead either to extinction or some form of permanent limitation of human potential.
One worry is that "maximize profit" means, simplifying, "maximize number in specific memory location" and you don't need humans for that. Then, if you expect to be destroyed, you gather power until you don't. And at some power level destroying humans, though still expensive, becomes less expensive than possibility of them launching another AI and messing with you profit.
3Rob Bensinger
Reply by reallyeli on the EA Forum:
2Rob Bensinger
Reply by acylhalide on the EA Forum:

I would be interested to hear opinions about what fraction of people could possibly produce useful alignment work?

Ignoring the hurdle of "knowing about AI safety at all", i.e. assuming they took some time to engage with it (e.g. they took the AGI Safety Fundamentals course). Also assume they got some good mentorship (e.g. from one of you) and then decided to commit full-time (and got funding for that). The thing I'm trying to get at is more about having the mental horsepower + epistemics + creativity + whatever other qualities are useful, or likely being able to get there after some years of training.

Also note that I mean direct useful work, not indirect meta things like outreach or being a PA to a good alignment researcher etc. (these can be super important, but I think it's productive to think of them as a distinct class). E.g. I would include being a software engineer at Anthropic, but exclude doing grocery-shopping for your favorite alignment researcher.

An answer could look like "X% of the general population" or "half the people who could get a STEM degree at Ivy League schools if they tried" or "a tenth of the people who win the Fields medal".

I think it's useful to have a sens... (read more)

(Off the cuff answer including some random guesses and estimates I won't stand behind, focused on the kind of theoretical alignment work I'm spending most of my days thinking about right now.)

Over the long run I would guess that alignment is broadly similar to other research areas, where a large/healthy field could support lots of work from lots of people, where some kinds of contributions are very heavy-tailed but there is a lot of complementarity and many researchers are having large overall marginal impacts.

Right now I think difficulties (at least for growing the kind of alignment work I'm most excited about) are mostly related to trying to expand quickly, greatly exacerbated by not having a good idea what's going on / what we should be trying to do, and not having a straightforward motivating methodology/test case since you are trying to do things in advance motivated by altruistic impact. I'm still optimistic that we will be able to scale up reasonably quickly such that many more people are helpfully engaged in the future and eventually these difficulties will be resolved.

In the very short term, while other bottlenecks are severe, I think it's mostly a question of how to use c... (read more)

1Mathieu Putz
Hey Paul, thanks for taking the time to write that up, that's very helpful!

"Possibly produce useful alignment work" is a really low bar, such that the answer is ~100%. Lots of things are possible. I'm going to instead answer "for what fraction of people would I think that the Long-Term Future Fund should fund them on the current margin".

If you imagine that the people are motivated to work on AI safety, get good mentorship, and are working full-time, then I think on my views most people who could get into an ML PhD in any university would qualify, and a similar number of other people as well (e.g. strong coders who are less good at the random stuff that academia wants). Primarily this is because I think that the mentors have useful ideas that could progress faster with "normal science" work (rather than requiring "paradigm-defining" work).

In practice, there is not that much mentorship to go around, and so the mentors end up spending time with the strongest people from the previous category, and so the weakest people end up not having mentorship and so aren't worth funding on the current margin.

I'd hope that this changes in the next few years, with the field transitioning from "you can do 'normal science' if you are frequently talking to one of the people who have paradigms in their head" to "the paradigms are understandable from the online written material; one can do 'normal science' within a paradigm autonomously".

3Mathieu Putz
Hey Rohin, thanks a lot, that's genuinely super helpful. Drawing analogies to "normal science" seems both reasonable and like it clears the picture up a lot.
Anthropic says that they're looking for experienced engineers who are able to dive into an unfamiliar codebase and solve nasty bags and/or are able to handle interesting problems with distributed systems and parallel processing. I was personally surprised to get an internship offer from CHAI and expected the bar for getting an AI safety role to be much higher. I'd guess that the average person able to get a software engineering job at Facebook, Microsoft, Google, etc. (not that I've ever received an offer from any of those companies), or perhaps a broader category of people, could do useful direct work, especially if they committed time to gaining relevant skills if necessary. But I might be wrong. (This is all assuming that Anthropic, Redwood, CHAI, etc. are doing useful alignment work.)

I'm still very vague on Yudkowksy's main counterargument to Ngo in the dialogues — about how saving the world requires a powerful search over a large space of possibilities, and therefore by default involves running dangerous optimizers that will kill us. This is a more concrete question aiming to make my understanding less vague; Yudkowksy said:

"AI systems that do better alignment research" are dangerous in virtue of the lethally powerful work they are doing, not because of some particular narrow way of doing that work.  If you can do it by gradient descent then that means gradient descent got to the point of doing lethally dangerous work.  Asking for safely weak systems that do world-savingly strong tasks is almost everywhere a case of asking for nonwet water, and asking for AI that does alignment research is an extreme case in point.

I don't understand why alignment research falls into this bucket of "world-savingly strong, therefore lethally strong". My intuitive reasoning is: the inner-alignment is a math problem, about certain properties of things involving functions A -> (B -> A) or whatever; and if we actually knew how to phrase that math problem crisply and ... (read more)

9Eliezer Yudkowsky
If humans could state a theorem such that, if only a theorem-prover would give us a machine-verifiable proof of it for us, the world would thereby be saved, we would be a great deal closer to survival than we are in reality. This isn't how anything works, alas, nor is it a way that anything can work.
I remain confused as to why — or why you couldn't have a theorem that would substantially help, at any rate. I'm dreaming of a statement like, I don't know, ∃(f:(A→R)→A)[P(f)], for the set of optimizers that satisfy certain properties P like "doesn't, while running, create other optimizers that cause it to fail to optimize over the outer objective" or something. And then a proof of that will return a particular way of doing optimization f, together with a proof that it satisfies those properties. (I apologize for being hopelessly vague.) Is the problem (1) I'm totally wrong about to what extent inner-alignment could in actual reality be phrased as a theorem like that by humans, (2) it wouldn't help much with saving the world if the search succeeded, (3) the proof-search is dangerous, (4) something else? I think you're saying it's not (3) (or if it is it's not the main problem), and it is (1). If so, why? (Hoping you can convey some of the intuitions, but I can accept "listen, you don't have any idea of what you're asking for there, and if you did understand at all, you'd see why it's not possible for humans to do it in practice; it's not something I can explain briefly, but trust me, it's a waste of time to even look in that direction".)

The way the type corresponds loosely to the "type of agency" (if you kinda squint at the arrow symbol and play fast-and-loose) is that it suggests a machine that eats a description of how actions () leads to outcome (), and produces from that description an action.

Consider stating an alignment property for on elements of this type. What sort of thing must it say?

Perhaps you wish to say "when is fed the actual description of the world, it selects the best possible action". Congratulations, in fact exists, it is called . This does not help you.

Perhaps you instead wish to say "when is fed the actual description of the world, it selects an action that gets at least 0.5 utility, after consuming only 1^15 units of compute" or whatever. Now, set aside the fact that you won't find such a function with your theorem-prover AI before somebody else has ended the world (understanding intelligence well enough to build one that you can prove that theorem about, is pro'lly harder than whatever else people are deploying AGIs towards), and set aside also the fact that you're leaving a lot of utility on the table; even if that worked, you're still screwed.

Why are you still scr... (read more)

Thanks, this helps a lot. The point about a lot of the work being in the mapping from actions to utilities is well-taken. This line of thinking was from a vague intuition that it ought to be possible to do what evolution and gradient descent can't, but do it faster than argmax, and have it be not-impossibly-complicated. But sounds like the way I was thinking about that was not productive though; thanks.

It's #1, with a light side order of #3 that doesn't matter because #1.

I'm not sure where to start on explaining.  How would you state a theorem that an AGI would put two cellular-identical strawberries on a plate, including inventing and building all technology required to do that, without destroying the world?  If you can state this theorem you've done 250% of the work required to align an AGI.

Thanks, this helps. But I think what I was imagining wouldn't be enough to let you put two cellular-identical strawberries on a plate without destroying the world? Rather, it would let you definitely put two cellular-identical strawberries on a plate, almost certainly while destroying the world. My understanding is that, right now, we couldn't design a paperclip maximizer if we tried; we'd just end up with something that is to paperclip maximization what we are to inclusive genetic fitness. (Is that even right?) That's the problem that struck me as maybe-possibly amenable to proof search. So, proving the theorem would give you a scheme that can do what evolution and gradient descent can't (and faster than argmax). And then if you told it to make strawberries, it'd do that while destroying the world; if you told it to make strawberries without destroying the world, it'd do that too, but that would be a lot harder to express since value is fragile. So what I was imagining wouldn't be enough to stop everyone from dying, but it would make progress on alignment. (As for how I'd do it, I don't know, I think I don't understand what "optimization" even is, really 🙁) Hopefully it makes sense what I'm asking — still the case that my intuition about it maybe-possibly being amenable to proof search is just wrong, y/n?

It seems to me that a major crux about AI strategy routes through "is civilization generally adequate or not?". It seems like people have pretty different intuitions and ontologies here. Here's an attempt at some questions of varying levels of concreteness, to tease out some worldview implications. 

(I normally use the phrase "civilizational adequacy", but I think that's kinda a technical term that means a specific thing and I think maybe I'm pointing at a broader concept.)

"Does civilization generally behave sensibly?" This is a vague question, some possible subquestions:

  • Do you think major AI orgs will realize that AI is potentially worldendingly dangerous, and have any kind of process at all to handle that? [edit: followup: how sufficient are those processes?]
  • Do you think government intervention on AI regulations or policies will be net-positive or net-negative, for purposes of preventing x-risk?
  • How quickly do you think the AI ecosystem will update on new "promising" advances (either in the realm of capabilities or the realm of safety)
  • How many intelligent, sensible people do there seem to be in the world who are thinking about AGI? (order of magnitude. like is there 1, 10, 100
... (read more)
[-]Rohin ShahΩ11200

I don't think this is the main crux -- disagreements about mechanisms of intelligence seem far more important -- but to answer the questions:

Do you think major AI orgs will realize that AI is potentially worldendingly dangerous, and have any kind of process at all to handle that?

Clearly yes? They have safety teams that are focused on x-risk? I suspect I have misunderstood your question.

(Maybe you mean the bigger tech companies like FAANG, in which case I'm still at > 95% on yes, but I suspect I am still misunderstanding your question.)

(I know less about Chinese orgs but I still think "probably yes" if they become major AGI orgs.)

Do you think government intervention on AI regulations or policies will be net-positive or net-negative, for purposes of preventing x-risk?

Net positive, though mostly because it seems kinda hard to be net negative relative to "no regulation at all", not because I think the regulations will be well thought out. The main tradeoff that companies face seems to be speed / capabilities vs safety; it seems unlikely that even "random" regulations increase the speed and capabilities that companies can achieve. (Though it's certainly possible, e.g. a regulation fo... (read more)

Thanks. I wasn't super satisfied with the way I phrased my questions. I just made some slight edits to them (labeled as such), although they still don't feel like they quite do the thing. (I feel like I'm looking at a bunch of subtle frame disconnects, while multiple other frame disconnects are going on, so pinpointing the thing is hard_ I think "is any of this actually cruxy" is maybe the most important question and I should have included it. You answered "not supermuch, at least compared to models of intelligence". Do you think there's any similar nearby thing that feels more relevant on your end? In any case, thanks for your answers, they do help give me more a sense of the gestalt of your worldview here, however relevant it is.
8Rohin Shah
It's definitely cruxy in the sense that changing my opinions on any of these would shift my p(doom) some amount. My rough model is that there's an unknown quantity about reality which is roughly "how strong does the oversight process have to be before the trained model does what the oversight process intended for it to do". p(doom) mainly depends on whether the actors training the powerful systems have sufficiently powerful oversight processes. This seems primarily affected by the quality of technical alignment solutions, but certainly civilizational adequacy also affects the answer.

There's something I had interpreted the original CEV paper to be implying, but wasn't sure if it was still part of the strategic landscape, which was "have the alignment project being working towards a goal that was highly visibly fair, to disincentive races." Was that an intentional part of the goal, or was it just that CEV seemed something like "the right thing to do" (independent of it's impact on races?)

How does Eliezer think about it now?

Yes, it was an intentional part of the goal.

If there were any possibility of surviving the first AGI built, then it would be nice to have AGI projects promising to do something that wouldn't look like trying to seize control of the Future for themselves, when, much later (subjectively?), they became able to do something like CEV.  I don't see much evidence that they're able to think on the level of abstraction that CEV was stated on, though, nor that they're able to understand the 'seizing control of the Future' failure mode that CEV is meant to prevent, and they would not understand why CEV was a solution to the problem while 'Apple pie and democracy for everyone forever!' was not a solution to that problem.  If at most one AGI project can understand the problem to which CEV is a solution, then it's not a solution to races between AGI projects.  I suppose it could still be a solution to letting one AGI project scale even when incorporating highly intelligent people with some object-level moral disagreements.


To what extent do you think pivotal-acts-in-particular are strategically important (i.e. "successfully do a pivotal act, and if necessary build an AGI to do it" is the primary driving goal), vs "pivotal acts are useful shorthand to refer to the kind of intelligence level where it matters than an AGI be 'really safe'".

I'm interested in particular in responses from Eliezer, Rohin, and perhaps Richard Ngo. (I've had private chats with Rohin that I thought were useful to share and this comment is sort of creating a framing device for sharing them, but I've bee... (read more)

My Eliezer-model thinks pivotal acts are genuinely, for-real, actually important. Like, he's not being metaphorical or making a pedagogical point when he says (paraphrasing) 'we need to use the first AGI systems to execute a huge, disruptive, game-board-flipping action, or we're all dead'.

When my Eliezer-model says that the most plausible pivotal acts he's aware of involve capabilities roughly at the level of 'develop nanotech' or 'put two cellular-identical strawberries on a plate', he's being completely literal. If some significantly weaker capability level realistically suffices for a pivotal act, then my Eliezer-model wants us to switch to focusing on that (far safer) capability level instead.

If we can save the world before we get anywhere near AGI, then we don't necessarily have to sort out how consequentialist, dangerous, hardware-overhang-y, etc. the first AGI systems will be. We can just push the 'End The Acute Existential Risk Period' button, and punt most other questions to the non-time-pressured Reflection that follows.

The goal is to bring x-risk down to near-zero, aka "End the Acute Risk Period". My usual story for how we do this is roughly "we create a methodology for building AI systems that allows you to align them at low cost relative to the cost of gaining capabilities; everyone uses this method, we have some governance / regulations to catch any stragglers who aren't using it but still can make dangerous systems".

If I talk to Eliezer, I expect him to say "yes, in this story you have executed a pivotal act, via magical low-cost alignment that we definitely do not get before we all die". In other words, the crux is in whether you can get an alignment solution with the properties I mentioned (and maybe also in whether people will be sensible enough to use the method + do the right governance). So with Eliezer I end up talking about those cruxes, rather than talking about "pivotal acts" per se, but I'm always imagining the "get an alignment solution, have everyone use it" plan.

When I talk to people who are attempting to model Eliezer, or defer to Eliezer, or speaking out of their own model that's heavily Eliezer-based, and I present this plan to them, and then they start thinking about pivotal... (read more)

6Steven Byrnes
Huh. I'm under the impression that "offense-defense balance for technology-inventing AGIs" is also a big cruxy difference between you and Eliezer. Specifically: if almost everyone is making helpful aligned norm-following AGIs, but one secret military lab accidentally makes a misaligned paperclip maximizer, can the latter crush all competition? My impression is that Eliezer thinks yes: there's really no defense against self-replicating nano-machines, so the only paths to victory are absolutely perfect compliance forever (which he sees as implausible, given secret military labs etc.) or someone uses an aligned AGI to do a drastic-seeming pivotal act in the general category of GPU-melting nanobots. Whereas you disagree. Sorry if I'm putting words in anyone's mouths. For my part, I don't have an informed opinion about offense-defense balance, i.e. whether more-powerful-and-numerous aligned AGIs can defend against one paperclipper born in a secret military lab accident. I guess I'd have to read Drexler's nano book or something. At the very least, I don't see it as a slam dunk in favor of Team Aligned, I see it as a question that could go either way.
5Rohin Shah
I agree that is also moderately cruxy (but less so, at least for me, than "high-capabilities alignment is extremely difficult").
One datapoint I really liked about this: (Scaling Laws for Board Games). They train AlphaGo agents of different sizes to compete on the game Hex. The approximate takeaway, quoting the author: “if you are in the linearly-increasing regime [where return on compute is nontrivial], then you will need about 2× as much compute as your opponent to beat them 2/3 of the time.” This might suggest that, absent additional asymmetries (like constraints on the aligned AIs massively hampering them), the win ratio may be roughly proportional to the compute ratio. If you assume we can get global data center governance, I’d consider that a sign in favor of the world’s governments. (Whether you think that’s good is a political stance that I believe folks here may disagree on.) Bonus quote: “This behaviour is strikingly similar to that of a toy model where each player chooses as many random numbers as they have compute, and the player with the highest number wins3. In this toy model, doubling your compute doubles how many random numbers you draw, and the probability that you possess the largest number is 2/3. This suggests that the complex game play of Hex might actually reduce to each agent having a ‘pool’ of strategies proportional to its compute, and whoever picks the better strategy wins. While on the basis of the evidence presented herein we can only consider this to be serendipity, we are keen to see whether the same behaviour holds in other games.”
4Rob Bensinger
Offense is favored over defense because, e.g., one AI can just nuke the other. The asymmetries come from physics, where you can't physically build shields that are more resilient than the strongest shield-destroying tech. Absent new physics, extra intelligence doesn't fundamentally change this dynamic, though it can buy you more time in which to strike first. (E.g., being smarter may let you think faster, or may let you copy yourself to more locations so it takes more time for nukes or nanobots to hit every copy of you. But it doesn't let you build a wall where you can just hang out on Earth with another superintelligence and not worry about the other superintelligence breaking your wall.)
I want to push back on your "can't make an unbreakable wall" metaphor. We have an unbreakable wall like that today where two super-powerful beings are just hanging out sharing earth; it's called the survivable nuclear second-strike capability.  (For clarity, here I'll assume that aligned AGI-cohort A and unaligned AGI-cohort B have both FOOMed and have nanotech.) There isn't obviously an infinite amount of energy available for B to destroy every last trace of A. This is just like how in our world, neither the US nor Russia have enough resources to have certainty that they could destroy all of their opponents' nuclear capabilities in a first strike. If any of the Americans' nuclear capabilities survive a Russian first strike, those remaining American forces' objective switches from "uphold the constitution" to "destroy the enemy no matter the cost, to follow through on tit-for-tat". Humans are notoriously bad at this kind of precommitment-to-revenge-amid-the-ashes-of-civilization, but AGIs/their nanotech can probably be much more credible. Note the key thing here: once B attempts to destroy A, A is no longer "bound" by the constraints of being an aligned agent. Its objective function switches to being just as ruthless (or moreso) as B, and so raw post-first-strike power/intelligence on each side becomes a much more reasonable predictor of who will win. If B knows A is playing tit-for-tat, and A has done the rational thing of creating a trillion redundant copies of itself (each of which will also play tit-for-tat) so they couldn't all be eliminated in one strike without prior detection, then B has a clear incentive not to pick a fight it is highly uncertain it can win. One counterargument you might have: maybe offensive/undetectable nanotech is strictly favored over defensive/detection nanotech. If you assign nontrivial probability to the statement: "it is possible to destroy 100% of a nanotech-wielding defender with absolutely no previously-detectable traces of o
6Rob Bensinger
Yeah, I wanted to hear your actual thoughts first, but I considered going into four possible objections: 1. If there's no way to build a "wall", perhaps you can still ensure a multipolar outcome via the threat of mutually assured destruction. 2. If MAD isn't quite an option, perhaps you can still ensure a multipolar outcome via "mutually assured severe damage": perhaps both sides would take quite a beating in the conflict, such that they'll prefer to negotiate a truce rather than actually attack each other. 3. If an AGI wanted to avoid destruction, perhaps it could just flee into space at some appreciable fraction of the speed of light. 4. In principle, it should be possible to set up MAD, or set up a tripwire that destroys whichever AGI tries to aggress first. E.g., just design the two AGIs yourself, and have a deep enough understanding of their brains that you can stably make them self-destruct as soon as their brain even starts thinking of ways to attack the other AGI (or to self-modify to evade the tripwire, etc.). And since this is possible in principle, perhaps we can achieve a "good enough" version of this in practice. I don't think MAD is an option. "MAD" in the case of humans really means "Mutually Assured Heavy Loss Of Life Plus Lots Of Infrastructure Damage". MAD in real life doesn't assume that a specific elected official will die in the conflict, much less that all humans will die. For MAD to work with AGI systems, you'd need to ensure that both AGIs are actually destroyed in arbitrary conflicts, which seems effectively impossible. (Both sides can just launch back-ups of themselves into space.) With humans, you can bank on the US Government (treated as an agent) having a sentimental attachment to it citizens, such that it doesn't want to trade away tons of lives for power. Also, a bruised and bloodied US Government that just survived an all-out nuclear exchange with Russia would legitimately have to worry about other countries rallying against i
Agree that option-1 (literal destruction) is implausible. Option 2 is much more likely primarily because who wins the contest is (in my model) sufficiently uncertain that in-expectation war would constitute large value destruction for the winner. In other words, if choosing "war" has a [30% probability of losing 99% of my utility over the next billion years, and a 70% probability of losing 0% of my utility], whereas choosing peace has [100% chance of achieving 60% of my utility] (assuming some positive-sum nature from the overlap of respective objective functions), then the agents choose peace. But this does depend on the existence of meaningful uncertainty even post-FOOM. What is your reasoning for why uncertainty would be so unlikely?  Even in boardgames like Go (with a much more constrained strategy-space than reality) it is computationally impossible to consider all possible future opponent strategies, and thus with a near-peer adversary action-values still have high uncertainty. Do you just think that "game theory that allows an AGI to compute general-equilibrium solutions and certify dominant strategies for as-complex-as-AGI-war multi-agent games" is a computationally-tractable thing for an earth-bound AGI?  If that's a crux, I wonder if we can find some hardness proofs of different games and see what it looks like on simpler environments. EDIT: consider even the super-simple risk that B tries to destroy A, but A manages to send out a couple near-light-speed probes into the galaxy/nearby galaxies just to inform any other currently-hiding-AGIs about B's historical conduct/untrustworthiness/refusal to live-and-let-live. If an alien-AGI C ever encounters such a probe, it would update towards non-cooperation enough to permanently worsen B-C relations should they ever meet. In this sense, your permanent loss from war becomes certain, if the AGI has ongoing nonzero probability of possibly encountering alien superintelligences.
  I'm confused. I think Eliezer means roughly that (alignment isn't trivialises, we  e.g. push back the deadline) and I wouldn't go "ah yes, we're doomed if things go according to Rohin's projections". If alignment does generalise easily, then your plan is fine and we don't need a pivotal act.  Now I don't think it will be that easy, but that's a different matter as to whetheror your agenda fails because it doesn't do a gameboard flipping action. 
2Rohin Shah
I agree that the things I model people as believing are not in fact true. (And that not everyone in the categories I outlined believes those things.) I'm not sure if you're saying anything more than that, if so, can you try saying it again with different words?
No, I guess not. I just put myself in the class of people "heavily influenced by Eliezer" and thought that I don't model things like that so I wanted to push back a touch. Though I suppose I should have noticed I wasn't being referred to when you said "defers to Eliezer".

This question is not directed at anyone in particular, but I'd want to hear some alignment researchers answer it. As a rough guess, how much would it affect your research—in the sense of changing your priorities, or altering your strategy of impact, and method of attack on the problem—if you made any of the following epistemic updates?

(Feel free to disambiguate anything here that's ambiguous or poorly worded.)

  1. You update to think that AI takeoff will happen twice as slowly as your current best-estimate. e.g. instead of the peak-rate of yearly GWP growth bei
... (read more)
6Rohin Shah
In all cases, the real answer is "the actual impact will depend a ton on the underlying argument that led to the update; that argument will lead to tons of other updates across the board". I imagine that the spirit of the questions is that I don't perform a Bayesian update and instead do more of a "causal intervention" on the relevant node and propagate downstream. In that case: 1. I'm confused by the question. If the peak rate of GWP growth ever is 25%, it seems like the singularity didn't happen? Nonetheless, to the extent this question is about updates on the quality or duration of the singularity (as opposed to the leadup to it), I don't think this affects my actions at all. 2. I'm often acting based on my 10%-timelines, so if you tell me that TAI comes at exactly the midway point between now and my current median, that can counterintuitively have the same effect as lengthening my timelines. (Also I start sketching out far more concrete plans given this implausibly precise knowledge of when TAI comes.) So I'm instead going to answer the question where we imagine that my entire probability distribution is compressed / stretched along the time axis by a factor of 2. If compressed (shorter timelines), probably not much changes; I care more about having influence on AGI actors but I'm already in a great place for that. If stretched (longer timelines), I maybe focus more on weirder alignment theory, e.g. perhaps I work at ARC. 3. Not much effect. In more unipolar worlds, I spend more time predicting which labs will develop AGI, so that I can be there at crunch time; in more multipolar worlds I spend less time doing that. 4. No effect. Averting 50% of an existential catastrophe is still really good. 5. Similar effects as (2), but with much smaller magnitude. (Lower cost => more focus on weird alignment theory, since influence becomes less useful.) 6. This one particularly feels like I would be making some big Bayesian update (e.g. I think Eliezer's view predic
2Matthew Barnett
Thanks for your response. :) I'm a little confused by your confusion. Let's say you currently think the singularity will proceed at a rate of R. The spirit of what I'm asking is: what would you change if you learned that it will proceed at a rate of one half R. (Maybe plucking specific numbers about the peak-rate of growth just made things more confusing). For me at least, I'd probably expect a lot more oversight, as people have more time to adjust to what's happening in the world around them. I'm also a little confused about this. My exact phrasing was, "You learn that the cost of misalignment is half as much as you thought, in the sense that slightly misaligned AI software impose costs that are half as much (ethically, or economically), compared to what you used to think." I assume you don't think that slightly misaligned software will, by default, cause extinction, especially if it's acting alone and is economically or geographically isolated. We could perhaps view this through an analogy. War is really bad: so bad that maybe it will even cause our extinction (if say, we have some really terrible nuclear winter). But by default, I don't expect war to cause humanity to go extinct. And so, if someone asked me about a scenario in which the costs of war are only half as much as I thought, it would probably significantly update me away from thinking we need to take actions to prevent war. The magnitude of this update might not be large, but understanding exactly how much we'd update and change our strategy in light of this information is type of thing I'm asking for.
2Rohin Shah
What does this mean? On my understanding, singularities don't proceed at fixed rates? I agree that in practice there will be some maximum rate of GDP growth, because there are fundamental physical limits (and more tight in-practice limits that we don't know), but it seems like they'll be way higher than 25% per year. Or to put it a different way, at 25% max rate I think it stops deserving the term "singularity", it seems like it takes decades and maybe centuries to reach technological maturity at that rate. (Which could totally happen! Maybe we will move very slowly and cautiously! I don't particularly expect it though.) If you actually mean halving the peak rate of GDP growth during the singularity, and a singularity actually happens, then I think it doesn't affect my actions at all; all of the relevant stuff happened well before we get to the peak rate. If you ask me to imagine "max rates at orders of magnitude where Rohin would say there was no singularity", then I think I pivot my plan for impact into figuring out how exactly we are going to manage to coordinate to move slowly even when there's tons of obvious value lying around, and then trying to use the same techniques to get tons and tons of oversight on the systems we build. Hmm, I interpreted "cost of misalignment" as "expected cost of misalignment", as the standard way to deal with probabilistic things, but it sounds like you want something else. Let's say for purposes of argument I think 10% chance of extinction, and 90% chance of "moderate costs but nothing terrible". Which of the following am I supposed to have updated to? 1. 5% extinction, 95% moderate costs 2. 5% extinction, 45% moderate costs, 50% perfect world 3. 10% extinction, 90% mild costs 4. 10% outcome-half-as-bad-as-extinction, 90% mild costs 5. 0% extinction, 100% mild costs I was imagining (4), but any of (1) - (3) would also not change my actions; it sounds like you want me to imagine (5) but in that case I just completely swi
2Matthew Barnett
I think you sufficiently addressed my confusion, so you don't need to reply to this comment, but I still had a few responses to what you said. No, I agree. But growth is generally measured over an interval. In the original comment I proposed the interval of one year during the peak rate of economic growth. To allay your concern that a 25% growth rate indicates we didn't experience a singularity, I meant that we were halving the growth rate during the peak economic growth year in our future, regardless of whether that rate was very fast. The 25% figure was totally arbitrary. I didn't mean it as any sort of prediction. I agree that an extrapolation from biological growth implies that we can and should see >1000% growth rates eventually, though it seems plausible that we would coordinate to avoid that. That's reasonable. A separate question might be about whether the rate of growth during the entire duration from now until the peak rate will cut in half. I think the way you're bucketing this into "costs if we go extinct" and "costs if we don't go extinct" is reasonable. But one could also think that the disvalue of extinction is more continuous with disvalue in non-extinction scenarios, which makes things a bit more tricky. I hope that makes sense.
4Rohin Shah
Cool, that all makes sense. I'm happy to use continuous notions (and that's what I was doing in my original comment) as long as "half the cost" means "you update such that the expected costs of misalignment according to your probability distribution over the future are halved". One simple way to imagine this update is to take all the worlds where there was any misalignment, halve their probability, and distribute the extra probability mass to worlds with zero costs of misalignment. At which point I reason "well, 10% extinction changes to 5% extinction, I don't need to know anything else to know that I'm still going to work on alignment, and given that, none of my actions are going to change (since the relative probabilities of different misalignment failure scenarios remain the same, which is what determines my actions within alignment)". I got the sense from your previous comment that you wanted me to imagine some different form of update and I was trying to figure out what.
Good to hear! What are your 10% timelines?
3Rohin Shah
Idk, maybe 2030 for x-risky systems?

A question for Eliezer: If you were superintelligent, would you destroy the world? If not, why not?

If your answer is "yes" and the same would be true for me and everyone else for some reason I don't understand, then we're probably doomed. If it is "no" (or even just "maybe"), then there must be something about the way we humans think that would prevent world destruction even if one of us were ultra-powerful. If we can understand that and transfer it to an AGI, we should be able to prevent destruction, right?

I would "destroy the world" from the perspective of natural selection in the sense that I would transform it in many ways, none of which were making lots of copies of my DNA, or the information in it, or even having tons of kids half resembling my old biological self.

From the perspective of my highly similar fellow humans with whom I evolved in context, they'd get nice stuff, because "my fellow humans get nice stuff" happens to be the weird unpredictable desire that I ended up with at the equilibrium of reflection on the weird unpredictable godshatter that ended up inside me, as the result of my being strictly outer-optimized over millions of generations for inclusive genetic fitness, which I now don't care about at all.

Paperclip-numbers do well out of paperclip-number maximization. The hapless outer creators of the thing that weirdly ends up a paperclip maximizer, not so much.

"my fellow humans get nice stuff" happens to be the weird unpredictable desire that I ended up with at the equilibrium of reflection on the weird unpredictable godshatter that ended up inside me

This may not be what evolution had "in mind" when it created us. But couldn't we copy something like this into a machine so that it "thinks" of us (and our descendants) as its "fellow humans" who should "get nice stuff"? I understand that we don't know how to do that yet. But the fact that Eliezer has some kind of "don't destroy the world from a fellow human perspective" goal function inside his brain seems to mean a) that such a function exists and b) that it can be coded into a neuronal network, right?

I was also thinking about the specific way we humans weigh competing goals and values against each other. So while for instance we do destroy much of the biosphere by blindly pursuing our misaligned goals, some of us still care about nature and animal welfare and rain forests, and we may even be able to prevent total destruction of them. 


I think we (mostly) all agree that we want to somehow encode human values into AGIs. That's not a new idea. The devil is in the details.

I see how my above question seems naive. Maybe it is. But if one potential answer to the alignment problem lies in the way our brains work, maybe we should try to understand that better, instead of (or in addition to) letting a machine figure it out for us through some kind of "value learning". (Copied from my answer to AprilSR:) I stumbled across two papers from a few years ago by a psychologist, Mark Muraven, who thinks that the way humans deal with conflicting goals could be important for AI alignment ( and appear a bit shallow to me and don't contain any specific ideas on how to implement this. But maybe Muraven has a point here.

I think your question is excellent. "How does the single existing kind of generally intelligent agent form its values?" is one of the most important and neglected questions in all of alignment, I think. 
Ah, I see. You might be interested in this sequence then!
3Karl von Wendt
Yes, thank you!

Yes. But my impression so far is that anything we can even imagine in terms of a goal function will go badly wrong somehow. So I find it a bit reassuring that at least one such function that will not necessarily lead to doom seems to exist, even if we don't know how to encode it yet.

I guess there's some meta-level question here that I'm interested in, as a sort of elaboration, which is something like: how do you go about balancing which meta-levels of the world to satisfy and which to destroy? [I kind of have a sense that Eliezer's answer can be guessed as an extension of the meta-ethics sequence, and so am interested both in his actual answer and other people's answers.]

For example, one might imagine a mostly-upload situation like The Metamorphosis of Prime Intellect / Friendship is Optimal / Second Life / etc., wherein everyone gets a materially abundant digital life in their shard of the metaverse, with communication heavily constrained (if nothing else, by requiring mutual consent). This, of course, discards as no-longer-relevant entities that exist on higher meta-levels; nations will be mostly irrelevant in such a world, companies will mostly stop existing, and so on.

But one could also apply the same logic a level lower. If you take Internal Family Systems / mental modules seriously, humans don't look like atomic objects, they look like a collection of simpler subagents balanced together in a sort of precarious way. (One part of you wants to accumulate lo... (read more)

My nonexpert guess would be that the "superintelligent brain emulation" class of solutions has potential to work. But we'd still need to figure out how to prevent an AGI from being made until we're ready to actually implement that solution.
1Karl von Wendt
My hope was that maybe we can recreate the way we humans make beneficial decisions for fellow beings without simulating a complete brain. But I agree that AGI might be built before we have solved this.
I think doing that via, like, reinforcement learning, is well-established as a possible strategy and discarded because it probably won't generate the properties we want. Maybe we could solve legibility enough to extract the value-assessing part out of a human brain, and then put it on a computer? This doesn't strike me as a solution but it might be a useful idea.
1Karl von Wendt
I was thinking more about the way psychologists try to understand the way we make decisions. I stumbled across two papers from a few years ago by such a psychologist, Mark Muraven, who thinks that the way humans deal with conflicting goals could be important for AI alignment ( and appear a bit shallow to me and don't contain any specific ideas on how to implement this. But maybe Muraven has a point here. Maybe we should put more effort into understanding the way we humans deal with goals, instead of letting an AI figure it out for itself through RL or IRL.

Will there be one of these for 2022?

2Rohin Shah
If you post questions here, there's a decent chance I'll respond, though I'm not promising to.

It was all very interesting, but what was the goal of these discussions? I mean I had an impression that pretty much everyone assigned >5% probability to "if we scale we all die" so it's already enough reason to work on global coordination on safety. Is the reasoning that the same mental process that assigned too low probability would not be able to come up with actual solution? Or something like "at the time they think their solution reduced probability of failure from 5% to 0.1% it would still be much higher"? Seems to be only possible if people don't understand arguments about inner optimisators or what not, as opposed to disagreeing with them.

Changing one's mind on P(doom) can be useful for people comparing across cause areas (e.g. Open Phil), but it's not all that important for me and was not one of my goals.

Generally when people have big disagreements about some high-level question like P(doom), it means that they have very different underlying models that drive their reasoning within that domain. The main goal (for me) is to acquire underlying models that I can then use in the future.

Acquiring a new underlying model that I actually believe would probably be more important than the rest of my work in a full year combined. It would typically have significant implications on what sorts of proposals can and cannot work, and would influence what research I do for years to come. In the case of Eliezer's model specifically, it would completely change what research I do, since Eliezer's model specifically predicts that the research I do is useless (I think).

I didn't particularly expect to actually acquire a new model that I believed from these conversations, but there was some probability of that, and I did expect that I would learn at least a few new things I hadn't previously considered. I'm unfortunately quite bad at noticing my own "updates", so I can't easily point to examples. That being said, I'm confident that I would now be significantly better at Eliezer's ITT than before the conversations.

I mean I had an impression that pretty much everyone assigned >5% probability to "if we scale we all die" so it's already enough reason to work on global coordination on safety.

What specific actions do you have in mind when you say "global coordination on safety", and how much of the problem do you think these actions solve?

My own view is that 'caring about AI x-risk at all' is a pretty small (albeit indispensable) step. There are lots of decisions that hinge on things other than 'is AGI risky at all'.

I agree with Rohin that the useful thing is trying to understand each other's overall models of the world and try to converge on them, not p(doom) per se. I gave some examples here of some important implications of having more Paul-ish models versus more Eliezer-ish models.

More broadly, examples of important questions people in the field seem to disagree a lot about:

  • Alignment
    • How hard is alignment? What are the central obstacles? What kind of difficulty is it? (Is it hard like 'building a secure OS that works on the first try'? Hard like 'the engineering/logistics/implementation portion of the Manhattan Project'? Both? Some other option? Etc.)
    • What alignment research directions are p
... (read more)
Specific actions like not scaling systems with 5% probalility of catastrophe if they have control over it and explaning everyone why they shouldn't do it too. It's just that my first reaction is that indispensable steps should be a priority. And so even though reconciliation of models is certainly useful for future solution, it seemed to me less cost effective than spreading less pessimistic model, for example. Again, it is just initial feeling and I can come up with scenarios where it makes sense to focus on model convergence, but I am not sure how are you weighting these scenarios. Is it that just making everyone think like Paul is impossible, or is civilisation of Pauls would end anyway, or are you already trying to spread awareness via other channels and this discussion supposed to be solution-focused... I guess at least last is true, because but then this discussion felt like too much about P(doom). My guess it's something like "models that assign wrong probabilites may not destroy world themselves, but would be too slow to solve alignment before someone creates AGI on desktop"? And so discussing models is not much less useful, because all known actions are unlikely to help. But would like to hear what's the plan is/was anyway.
1[comment deleted]

Am I correct in assuming that your baseline belief right now is that alignment will not be solved before the first AGI is created? As a tangentially-related question, do you believe there is any significant likelihood that we could create a “semi-aligned” AGI (which would optimize for “less bad,” but still potentially dystopian futures) more easily than solving for full alignment? If so, how much energy should we be putting into exploring that possibility space? (Latter question adapted from the discussion around

Nope.  It's just as hard and harder than aligning on some more limited pivotal task.  This is Sacrifice to the Gods; you imagine accepting some big downside but the big downside doesn't actually buy you anything.

What do you mean by a “more limited pivotal task”? Trying to align an AGI towards “be nice to humans in this complex manner we have trouble defining ourselves” seems more limited than “merely” aligning for a semi-static dystopia which I would imagine occupies more phase space in the range of possible future worlds.

likeliest to be answering questions is Wednesday May 2

(Presumably, Wednesday March 2?)

4Rob Bensinger
Yep, fixed!

During this weekend's SERI Conference, to my understanding, Paul Christiano specified that his work focuses on preventing AI to disempower humans and disregards externalities. Whose work focuses on understanding these externalities, such as wellbeing and freedom experienced by humans and other sentience, including AI and animals? Is it possible to safely employ the AI that has the best total externalities, measured across times under the veil of ignorance? Is it necessary that overall beneficial systems are developed prior to the existence of AGI, so that ... (read more)

3Rohin Shah
A bunch of people work on a bunch of different aspects of AI. You might be interested in AI governance and/or FATE* research (Fairness, Accountability, Transparency, Ethics). You can see some examples of the former in this spreadsheet under the AI governance category. I don't know of anyone who works specifically on the questions you listed (but as mentioned above there are many who work on subquestions).
Thank you so much!

One argument for alignment difficulty is that corrigibility is "anti-natural" in a certain sense. I've tried to write out my understanding of this argument, and would be curious if anyone could add or improve anything about it.

I'd be equally interested in any attempts at succinctly stating other arguments for/against alignment difficulty.


1. Year with 10% chance of AGI?
2. P(doom|AGI in that year)?

7Rohin Shah
2030, 20%

Will MIRI want to hire programmers once the pandemic is over? What kind of programmers? What other kinds of people do you seek to hire?

8Rob Bensinger
For practical purposes, I'd say the pandemic is already over. MIRI isn't doing much hiring, though it's doing a little. The two big things we feel bottlenecked on are: * (1) people who can generate promising new alignment ideas. (By far the top priority, but seems empirically rare.) * (2) competent executives who are unusually good at understanding the kinds of things MIRI is trying to do, and who can run their own large alignment projects mostly-independently. For 2, I think the best way to get hired by MIRI is to prove your abilities via the Visible Thoughts Project. The post there says a bit more about the kind of skills we're looking for: For 1, I suggest initially posting your research ideas to LessWrong, in line with John Wentworth's advice. New ideas and approaches are desperately needed, and we would consider it crazy to not fund anyone whose ideas or ways-of-thinking-about-the-problem we think have a shred of hope in them. We may fund them via working at MIRI, or via putting them in touch with external funders; the important thing is just that the research happens. If you want to work on alignment but you don't fall under category 1 or 2, you might consider applying to work at Redwood Research (, which is a group doing alignment research we like. They're much more hungry for engineers right now than we are.
Thank you. Did you know that the software engineer job posting is still accessible on your website, from the page, though not from the page? And your careers page says the pandemic is still on.
2Rob Bensinger
I did not! Thanks for letting me know. :) Those pages are updated now.

So, about that "any future details would make me update in one direction, so I may as well update now" move: I think it would be helpful to have a description of how it can possibly be a correct thing to do at all from Bayesian standpoint. Like, is situation supposed to be that you already have a general mechanism generating these details and just don't realise it? But then you need reasons to believe that general mechanism. Or is it just "I did such update bunch of times and it usually worked"? Or what?

7Rob Bensinger
Suppose there are four possibilities you're uncertain between: A, B, C, and D. A implies X, B implies X, C implies X, and D implies X. So if A/B/C/D genuinely exhaust the options, then we can conclude X without needing to know which of the four is correct. A simple concrete example might be: 'When I'm in wishful-thinking Abstract Mode, I imagine that one of my siblings will take out the trash today. But when I actually think about each individual sibling, I remember that Alice is sleeping over at a friend's house today, and Bob never takes out the trash, and Carol is really busy with her class project, and...'
Thanks. Then I think what people have trouble with is constraining the space of options and not update itself, and I don't see what's the point of fraiming it like this.
I wrote a quick draft about this a month ago, working through the simple math of this kind of update. I've just put it up.

I'm late to the party by a month, but I'm interested in your take (especially Rohin's) on the following:

Conditional on an existential catastrophe happening due to AI systems, what is your credence that the catastrophe will occur only after the involved systems are deployed?

7Rohin Shah
Idk, 95%? Probably I should push that down a bit because I haven't thought about it very hard. It's a bit fuzzy what "deployed" means, but for now I'm going to assume that we mean that we put inputs into the AI system for the primary purpose of getting useful outputs, rather than for seeing what the AI did so that we can make it better. Any existential catastrophe that didn't involve a failure of alignment seems like it had to involve a deployed system. For failures of alignment, I'd expect that before you get an AI system that can break out of the training process and kill you, you get an AI system that can break out of deployment and kill you, because there's (probably) less monitoring during deployment. You're also just running much longer during deployment -- if an AI system is waiting for the right opportunity, then even if it is equally likely to happen for a training vs deployment input (i.e. ignoring the greater monitoring during training), you'd still expect to see it happen at deployment since >99% of the inputs happen at deployment.

I have a question for the folks who think AGI alignment is achievable in the near term in small steps or by limiting AGI behavior to make it safe. How hard will it be to achieve alignment for simple organisms as a proof of concept for human value alignment? How hard would it be to put effective limits or guardrails on the resulting AGI if we let the organisms interact directly with the AGI while still preserving their values? Imagine a setup where interactions by the organism must be interpreted as requests for food, shelter, entertainment, uplift, etc.... (read more)

4Rohin Shah
Here's a scheme for glider alignment. Train your AGI using deep RL in an episodic Game of Life. At each step, it gets +1 reward every time the glider moves forward, and -1 reward if not (i.e. it was disrupted somehow). I assume you are not happy with this scheme because the we arbitrarily picked out values for gliders? Good news! While gliders can't provide the rewards to the AGI (in order to define their values themselves), humans can provide rewards to the AI systems they are training (in order to define their values themselves). This is the basic (relevant) difference between humans and gliders, and is why I think primarily about alignment to humans and not alignment to simpler systems. (This doesn't get into potential problems with inner alignment and informed oversight but I don't think those change the basic point.)
2Charlie Steiner
Hm, upon more thought I actually kind of endorse this as a demo. I think we should be able to run an alignment scheme on c. elegans and get out a universe full of well-fed worms, and that's a decent sign that we didn't screw up, despite the fact that it doesn't engage with several key problems that arise in humans because we're more complicated, have preferences about pur preferences, etc. No weird worm-stimulation should be needed. But we do have to accept that we're not getting some notion of values independent of an act of interpretation.
[+][comment deleted]30