Posts
Comments
I think [process-based RL] has roughly the same risk profile as imitation learning, while potentially being more competitive.
I agree with this in a sense, although I may be quite a bit a more harsh about what counts as "executing an action". For example, if reward is based on an overseer talking about the action with a large group of people/AI assistants, then that counts as "executing the action" in the overseer-conversation environment, even if the action looks like it's for some other environment, like a plan to launch a new product in the market. I do think myopia in this environment would suffice for existential safety, but I don't know how much myopia we need.
If you're always talking about myopic/process-based RLAIF when you say RLAIF, then I think what you're saying is defensible. I speculate that not everyone reading this recognizes that your usage of RLAIF implies RLAIF with a level of myopia that matches current instances of RLAIF, and that that is a load-bearing part of your position.
I say "defensible" instead of fully agreeing because I weakly disagree that increasing compute is any more of a dangerous way to improve performance than by modifying the objective to a new myopic objective. That is, I disagree with this:
I think you would probably prefer to do process-based RL with smaller models, rather than imitation learning with bigger models
You suggest that increasing compute is the last thing we should do if we're looking for performance improvements, as opposed to adding a very myopic approval-seeking objective. I don't see it. I think changing the objective from imitation learning is more likely to lead to problems than scaling up the imitation learners. But this is probably beside the point, because I don't think problems are particularly likely in either case.
What is process-based RL?
I think your intuitions about costly international coordination are challenged by a few facts about the world. 1) Advanced RL, like open borders + housing deregulation, guarantees vast economic growth in wealthy countries. Open borders, in a way that seems kinda speculative, but intuitively forceful for most people, has the potential to existentially threaten the integrity of a culture, including especially its norms; AI has the potential, in a way that seems kinda speculative, but intuitively forceful for most people, has the potential to existentially threaten all life. The decisions of wealthy countries are apparently extremely strongly correlated, maybe in part for "we're all human"-type reasons, and maybe in part because legislators and regulators know that they won't get their ear chewed off for doing things like the US does. With immigration law, there is no attempt at coordination; quite the opposite (e.g. Syrian refugees in the EU). 2) The number of nuclear states is stunningly small if one follows the intuition that wildly uncompetitive behavior, which leaves significant value on the table, produces an unstable situation. Not every country needs to sign on eagerly to avoiding some of the scariest forms of AI. The US/EU/China can shape other countries' incentives quite powerfully. 3) People in government do not seem to be very zealous about economic growth. Sorry this isn't a very specific example. But their behavior on issue after issue does not seem very consistent with someone who would see, I don't know, 25% GDP growth from their country's imitation learners, and say, "these international AI agreements are too cautious and are holding us back from even more growth"; it seems much more likely to me that politicians' appetite for risking great power conflict requires much worse economic conditions than that.
In cases 1 and 2, the threat is existential, and countries take big measures accordingly. So I think existing mechanisms for diplomacy and enforcement are powerful enough "coordination mechanisms" to stop highly-capitalized RL projects. I also object a bit to calling a solution here "strong global coordination". If China makes a law preventing AI that would kill everyone with 1% probability if made, that's rational for them to do regardless of whether the US does the same. We just need leaders to understand the risks, and we need them to be presiding over enough growth that they don't need to take desperate action, and that seems doable.
Also, consider how much more state capacity AI-enabled states could have. It seems to me that a vast population of imitation learners (or imitations of populations of imitation learners) can prevent advanced RL from ever being developed, if the latter is illegal; they don't have to compete with them after they've been made. If there are well-designed laws against RL (beyond some level of capability), we would have plenty of time to put such enforcement in place.
I believe that LM agents based on chain of thought and decomposition seem like the most plausible approach to bootstrapping subhuman systems into trusted superhuman systems. For about 7 years using LM agents for RLAIF has seemed like the easiest path to safety,[4] and in my view this is looking more and more plausible over time.
I agree whole-heartedly with the first sentence. I'm not sure why you understand it to support the second sentence; I feel the first sentence supports my disagreement with the second sentence! Long-horizon RL is a different way to get superhuman systems, and one encourages that intervening in feedback if the agent is capable enough. Doesn't the first sentence support the case that it would be safer to stick to chain of thought and decomposition as the key drivers of superhumanness, rather than using RL?
Me: Peer review can definitely issue certificates mistakenly, but validity is what it aims to certify.
You: No it doesn't. They just care about interestingness.
Me: Do you agree reviewers aim to only accept valid papers, and care more about validity than interestingness?
You: Yes, but...
If you can admit that we agree on this basic point, I'm happy to discuss further about how good they are at what they aim to do.
1: If retractions were common, surely you would have said that was evidence peer review didn't accomplish much! If academics were only equally good at spotting mistakes immediately, they would still spot the most mistakes because they get the first opportunity to. And if they do, others don't get a "chance" to point out a flaw and have the paper retracted. Even though this argument fails, I agree that journals are too reluctant to publish retractions; pride can sometimes get in the way of good science. But that has no bearing on their concern for validity at the reviewing stage.
2: Some amount of trust is taken for granted in science. The existence of trust in a scientific field does not imply that the participants don't actually care about the truth. Bounded Distrust.
3: Since some level interestingness is also required for publication, this is consistent with a top venue having a higher bar for interestingness than a lesser venue, even while they same requirement for validity. And this is definitely in fact the main effect at play. But yes, there are also some lesser journals/conferences/workshops where they are worse at checking validity, or they care less about it because they are struggling to publish enough articles to justify their existence, or because they are outright scams. So it is relevant that AAAI publishes AI Magazine, and their brand is behind it. I said "peer reviewed" instead of "peer reviewed at a top venue" because the latter would have rubbed you the wrong way even more, but I'm only claiming that passing peer review is worth a lot at a top venue.
Care to bet on the results of a survey of academic computer scientists? If the stakes are high enough, I could try to make it happen.
"As a reviewer, I only recommend for acceptance papers that appear to be both valid and interesting."
Strongly agree - ... - Strongly Disagree
"As a reviewer, I would sooner recommend for acceptance a paper that was valid, but not incredibly interesting, than a paper that was interesting, but the conclusions weren't fully supported by the analysis."
Strongly agree - ... - Strongly Disagree
I don't understand. Importantly, these are optimistically biased, and you can't assume my true credences are this high. I assign much less than 90% probability to C. But still, they're perfectly consistent. M doesn't say anything about succeeding--only being allowed. M is basically saying: listing the places he'd be willing to live, do they all pass laws which would make building dangerously advanced AI illegal? The only logical connection between C and M is that M (almost definitely) implies C.
Thank you very much for saying that.
I was feeling disappointed about the lack of positive comments, and I realized recently I should probably go around commenting on posts that I think are good, since right now, I mostly only comment on posts when I feel I have an important disagreement. So it's hard to complain when I'm on the receiving end of that dynamic.
On the 2nd point, the whole discussion of mu^prox vs. mu^dist is fundamentally about goal (mis)generalization. My position is that for a very advanced agent, point estimates of the goal (i.e. certainty that some given account of the goal is correct) would probably really limit performance in many contexts. This is captured by Assumptions 2 and 3. An advanced agent is likely to entertain multiple models of what their current understanding of their goal in a familiar context implies about their goal in a novel context. Full conviction in mu^dist does indeed imply non-wireheading behavior, and I wouldn't even call it misgeneralization; I think that would be a perfectly valid interpretation of past rewards. So that's why I spend so much time discussing relative credence in those models.
The assumption says "will do" not "will be able to do". And the dynamics of the unknown environment includes the way it outputs rewards. So the assumption was not written in a way that clearly flags its entailment of the agent deliberately modeling the origin of reward, and I regret that, but it does entail that. So that was why engage with the objection that reward is not the optimization target under this section.
In the video game playing setting you describe, it is perfectly conceivable that the agent deliberately acts to optimize for high in-game scores without being terminally motivated by reward,
There is no need to recruit the concept of "terminal" here for following the argument about the behavior of a policy that performs well according to the RL objective. If the video game playing agent refines its understanding of "success" according to how much reward it observes, and then pursues success, but it does all this because of some "terminal" reason X, that still amounts to deliberate reward optimization, and this policy still satisfies Assumptions 1-4.
If I want to analyze what would probably happen if Edward Snowden tried to enter the White House, there's lots I can say without needing to understand what deep reason he had for trying to do this. I can just look at the implications of his attempt to enter the White House: he'd probably get caught and go to jail for a long time. Likewise, if an RL agent is trying to maximize is reward, there's plenty of analysis we can do that is independent of whether there's some other terminal reason for this.
Peer review is not a certification of validity,
Do you think the peer reviewers and the editors thought the argument was valid?
Peer review can definitely issue certificates mistakenly, but validity is what it aims to certify.
Thank you. I've changed the title.
Not trying to be arrogant. Just trying to present readers who have limited time a quickly digestible bit evidence about the likelihood that the argument is a shambles.
I don't think it's an assumption really. I think this sentence just fixes the meanings, in perfectly sensible ways, of the words "entertain" and "to" (as in "pick actions to"). I guess you're not persuaded that competent behavior in the "many new video games" environment is deserving of the description "aiming to maximize predicted future rewards". Why is that, if the video games are sufficiently varied?
Thank you for this review! A few comments on the weaknesses of my paper.
In particular, it explicitly says the argument does not apply to supervised learning.
Hardly a weakness if supervised learning is unlikely to be an existential threat!
Strength: Does not make very concrete assumptions about the AGI development model.
Weakness: Does not talk much about how AGI is likely to be developed, unclear which of the assumptions are more/less likely to hold for AGI being developed using the current ML paradigm.
The fact that the argument holds equally well no matter what kind of function approximation is used to do inference is, I think, a strength of the argument. It's hard to know what future inference algorithms will look like, although I do think there is a good chance that they will look a lot like current ML. And it's very important that the argument doesn't lump together algorithms where outputs are selected to imitate a target (imitation learners / supervised learners) vs. algorithms where outputs are selected to accomplish a long-term goal. These are totally different algorithms, so analyses of their behavior should absolutely be done separately. The claim "we can analyze imitation learners imitating humans together with RL agents, because both times we could end up with intelligent agents" strikes me as just as suspect as the claim "we can analyze the k-means algorithm together with a vehicle routing algorithm, because both will give us a partition over a set of elements." (The claim "we can analyze imitation learners alongside the world-model of a model-based RL agent" is much more reasonable, since these are both instances of supervised learning.)
Assumes the agent will be aiming to maximize reward without justification, i.e. why does it not have other motivations, perhaps due to misgeneralizing about its goal?
Depending on the meaning of "aiming to maximize reward", I have two different responses. In one sense, I claim "aiming to maximize reward" would be the nature of a policy that performs sufficiently strongly according to the RL objective. (And aiming to maximize inferred utility would be the nature of a policy that performs sufficiently strongly according to the CIRL objective.) But yes, even though I claim this simple position stands, a longer discussion would help establish that.
There's another sense in which you can say that an agent that has a huge inductive bias in favor of , and so violates Assumption 3, is not aiming to maximize reward. So the argument accounts for this possibility. Better yet, it provides a framework for figuring out when we can expect it! See, for example, my comment in the paper that I think an arbitrarily advanced RL chess player would probably violate Assumption 3. I prefer the terminology that says this chess player is aiming to maximize reward, but is dead sure winning at chess is necessary for maximizing reward. But if these are the sort of cases you mean to point to when you suggest the possibility of an agent "not maximizing reward", I do account for those cases.
Arguments made in the paper for why an agent intervening in the reward would have catastrophic consequences are somewhat brief/weak.
Are there not always positive returns to energy/resource usage when it comes to maximizing the probability that the state of a machine continues to have certain property (i.e. reward successfully controlled)? And our continued survival definitely requires some energy/resources. To be clear, catastrophic consequences follow from an advanced agent intervening the provision of reward in the way that would be worth doing. Catastrophic consequences definitely don't follow from a half-hearted and temporary intervention in the provision of reward.
I assume (6) means that your "anthropic update" scans across possible universes to find those that contain important decisions you might want to influence?
Yes, and then outputs strings from that set with probability proportional to their weight in the universal prior.
By (3) do you mean the same thing as "Simplest output channel that is controllable by advanced civilization with modest resources"?
I would say "successfully controlled" instead of controllable, although that may be what you meant by the term. (I decomposed this as controllable + making good guesses.) For some definitions of controllable, I might have given a point estimate of maybe 1 or 5 bits. But there has to be an output channel for which the way you transmit a bitstring out is the way the evolved consequentialists expect. But recasting it in these terms, implicitly makes the suggestion that the specification of the output channel can take on some of the character of (6'), makes me want to put my range down to 15-60; point estimate 25.
instead of using (1)+(2)+(3) you should compare to (6') = "Simplest program that scans across many possible worlds to find those that contain some pattern that can be engineered by consequentialists trying to influence prior."
Similarly, I would replace "can be" with "seems to have been". And just to make sure we're talking about the same thing, it takes this list of patterns, and outputs them with probability proportional to their weight in the universal prior.
Yeah, this seems like it would make some significant savings compared to (1)+(2)+(3). I think replacing parts of the story from being specified as [arising from natural world dynamics] to being specified as [picked out "deliberately" by a program] generally leads to savings.
Then the comparison is between specifying "important predictor to influence" and whatever the easiest-to-specify pattern that can be engineered by a consequentialist. It feels extremely likely to me that the second category is easier, indeed it's kind of hard for me to see any version of (6) that doesn't have an obviously simpler analog that could be engineered by a sophisticated civilization.
I don't quite understand the sense in which [worlds with consequentialist beacons/geoglyphs] can be described as [easiest-to-specify controllable pattern]. (And if you accept the change of "can be" to "seems to have been", it propagates here). Scanning for important predictors to influence does feel very similar to me to scanning for consequentialist beacons, especially since the important worlds are plausibly the ones with consequentialists.
There's a bit more work to be done in (6') besides just scanning for consequentialist beacons. If the output channel is selected "conveniently" for the consequentialists, since the program is looking for the beacons, instead of the consequentialists giving it their best guess(es) and putting up a bunch of beacons, there has to be some part of the program which aggregates the information of multiple beacons (by searching for coherence, e.g.), or else determines which beacon takes precedence, and then also determines how to interpret their physical signature as a bitstring.
Tangent: in heading down a path trying to compare [scan for "important to influence"] vs. [scan for "consequentialist attempted output messages"] just now, my first attempt had an error, so I'll point it out. It's not necessarily harder to specify "scan for X" than "scan for Y" when X is a subset of Y. For instance "scan for primes" is probably simpler than "scan for numbers with less than 6 factors".
Maybe clarifying or recasting the language around "easiest-to-specify controllable pattern" will clear this up, but can you explain more why it feels to you that [scan for "consequentialists' attempted output messages"] is so much simpler than [scan for "important-to-influence data streams"]? My very preliminary first take is that they are within 8-15 bits.
I also don't really see why you are splitting them [(4) + (5)] apart, shouldn't we just combine them into "wants to influence predictors"? If you're doing that presumably you'd both use the anthropic prior and then the treacherous turn.
I split them in part in case there is there is a contingent of consequentialists who believes that outputting the right bitstring is key to their continued existence, believing that they stop being simulated if they output the wrong bit. I haven't responded to your claim that this would be faulty metapyhsics on their part; it still seems fairly tangential to our main discussion. But you can interpret my 5 bit point estimate for (5) as claiming that 31 times out of 32 that a civilization of consequentialists tries to influence their world's output, it is in an attempt to survive. Tell me if you're interested in a longer justification that responds to your original "line by line comments" comment.
Yeah, seems about right.
I think with 4, I've been assuming for the sake of argument that manipulators get free access to the right prior, and I don't have a strong stance on the question, but it's not complicated for a directly programmed anthropic update to be built on that right prior too.
I guess I can give some estimates for how many bits I think are required for each of the rows in the table. I'll give a point estimate, and a range for a 50% confidence interval for what my point estimate would be if I thought about it for an hour by myself and had to write up my thinking along the way.
I don't have a good sense for how many bits it takes to get past things that are just extremely basic, like an empty string, or an infinite string of 0s. But whatever that number is, add it to both 1 and 6.
1) Consequentialists emerge, 10 - 50 bits; point estimate 18
2) TM output has not yet begun, 10 - 30 bits; point estimate 18
3) make good guesses about controllable output, 18 - 150 bits; point estimate 40
4) decide to output anthropically updated prior, 8 - 35 bits; point estimate 15
5) decide to do a treacherous turn. 1 - 12 bits; point estimate 5
vs. 6) direct program for anthropic update. 18-100 bits; point estimate 30
The ranges are fairly correlated.
Do you have some candidate "directly programmed anthropic update" in mind? (That said, my original claim was just about the universal prior, not about a modified version with an anthropic update)
I’m talking about the weight of an anthropically updated prior within the universal prior. I should have added “+ bits to encode anthropic update directly” to that side of the equation. That is, it takes some number of bits to encode “the universal prior, but conditioned on the strings being important to decision-makers in important worlds”. I don’t know how to encode this, but there is presumably a relatively simple direct encoding, since it’s a relatively simple concept. This is what I was talking about in my response to the section “The competition”.
One way that might be helpful about thinking about the bits saved from the anthropic update is that it is string is important to decision-makers in important worlds. I think this gives us a handle in reasoning about anthropic savings as a self-contained object, even if it’s a big number.
> bits to specify camera on earth - bits saved from anthropic update
I think the relevant number is just "log_2 of the number of predictions that the manipulators want to influence." It seems tricky to think about this (rather small) number as the difference between two (giant) numbers.
But suppose they picked only one string to try to manipulate. The cost would go way down, but then it probably wouldn’t be us that they hit. If log of the number of predictions that the manipulators want to influence is 7 bits shorter than [bits to specify camera on earth - bits saved from anthropic update], then there’s a 99% chance we’re okay. If different manipulators in different worlds are choosing differently, we can expect 1% of them to choose our world, and so we start worrying again, but we add the 7 bits back because it’s only 1% of them.
So let’s consider two Turing machines. Each row will have a cost in bits.
A B
Consequentialists emerge, Directly programmed anthropic update.
make good guesses about controllable output,
decide to output anthropically updated prior.
Weight of earth-camera within anthropically updated prior
The last point can be decomposed into [description length of camera in our world - anthropic savings], but it doesn’t matter; it appears in both options.
I don’t think this is what you have in mind, but I’ll add another case, in case this is what you meant by “They are just looking at the earth-like Turing machine”. Maybe, just skip this though.
A B
Consq-alists emerge in a world like ours, Directly prog. anthropic update.
make good guesses about controllable output,
output (strong) anth. updated prior.
Weight of earth-camera in strong anth. update … in normal anth. update
They can make a stronger anthropic update by using information about their world, but the savings will be equal to the extra cost of specifying that the consequentialists are in a world like ours. This is basically the case I mentioned above where different manipulators choose different sets of worlds to try to influence, but then the set of manipulators that choose our world has smaller weight.
------ end potential skip
What I think it boils down to is the question:
Is the anthropically updated version of the universal prior most simply described as “the universal prior, but conditioned on the strings being important to decision-makers in important worlds” or “that thing consequentialists sometimes output”? (And consequentialists themselves may be more simply described as “those things that often emerge”). “Sometimes” is of course doing a lot of work, and it will take bits to specify which “sometimes” we are talking about. If the latter is more simple, then we might expect the natural continuation of those sequences to usually contain treacherous turns, and if the former is more simple, then we wouldn’t. This is why I don’t think the weight of an earth-camera in the universal prior ever comes into it.
But/so I don’t understand if I’m missing the point of a couple paragraphs of your comment—the one which starts “They are just looking at the earth-like Turing machine”, and the next paragraph, which I agree with.
I'm using some of the terminology I suggested here.
A factoring is a set of questions such that each signature of possible answers identifies a unique element. In 20 questions, you can tailor the questions depending on the answers to previous questions, and ultimately each element will have a bitstring signature depending on the history of yesses and nos. I guess you can define the question to include xors with previous questions, so that it effectively changes depending on the answers to others. But it's sometimes useful that the bitstrings are allowed to have different length. It feels like an unfortunate fact that when constructing a factoring for 7 elements, you're forced to use the factoring {"Okay, well, which element is it?"}, just because you don't want to have to answer a different number of questions for different elements. Is this a real cost? Or do we only ever construct cases where it's not?
In the directed graph of subsets, with edges corresponding to the subset relation, why not consider arbitrary subtrees? For example, for the set of 7 elements, we might have {{0, 1, 2}, {{3, 4}, {5, 6}}}. (I'm not writing it out as a tree, but that contains all the information). This corresponds to the sequence of questions: "is it less than 3?", [if yes] "is it 0, 1, or 2?", [if no], "is it less than 5?", "is it even?" Allowing different numbers of questions and different numbers of answers seems to give some extra power here. Is it meaningful?
I was thinking of some terminology that might make it easier to thinking about factoring and histories and whatnot.
A partition can be thought of as a (multiple-choice) question. Like for a set of words, you could have the partition corresponding to the question "Which letter does the word start with?" and then the partition groups together elements with the same answer.
Then a factoring is set of questions, where the set of answers will uniquely identify an element. The word that comes to mind for me is "signature", where an element's signature is the set of answers to the given set of questions.
For the history of a partition X, X can be thought of as a question, and the history is the subset of questions in the factoring that you need the answers to in order to determine the answer to question X.
And then two questions X and Y are orthogonal if there aren't any questions in the factoring that you need the answer to both for answering X and for answering Y.
I was thinking about the difficulty of finite factored sets not understanding the uniform distribution over 4 elements, and it makes me feel like something fundamental needs to be recast. An analogy came to mind about eigenvectors vs. eigenspaces.
What we might like to be true about the unit eigenvectors of a matrix is that they are the unique unit vectors for which the linear transformation preserves direction. But if two eigenvectors have the same eigenvalue, the choice of eigenvectors is not unique--we could choose any pair on that plane. So really, it seems like we shouldn't think about a matrix's eigenvectors and (potentially repeated) eigenvalues; we should think about a matrix's eigenvalues and eigenspaces, some of which might be more than 1-dimensional.
I wonder if there's a similar move to be made when defining orthogonality. Maybe (for example) orthogonality would be more conveniently defined between two sets of partitions instead of between two partitions. Probably that specific idea fails, but maybe there's something like this that could be done.
I take your point that we are discussing some output rules which add extra computation states, and so some output rules will add fewer computation states than others.
I'm merging my response to the rest with my comment here.
They are using their highest probability guess about the output channel, which will be higher probability than the output channel exactly matching some camera on old earth (but may still be very low probability). I still don't understand the relevance.
I’m trying to find the simplest setting where we have a disagreement. We don’t need to think about cameras on earth quite yet. I understand the relevance isn’t immediate.
They don't care about "their" Turing machine, indeed they live in an infinite number of Turing machines that (among other things) output bits in different ways.
I think I see the distinction between the frameworks we most naturally think about the situation. I agree that they live in an infinite number of Turing machines, in the sense that their conscious patterns appear in many different Turing machines. All of these Turing machines have weight in some prior. When they change their behavior, they (potentially) change the outputs of any of these Turing machines. Taking these Turing machines as a set, weighted by those prior weights we can consider the probability that the output obeys a predicate P. The answer to this question can be arrived at through an equivalent process. Let the inhabitants imagine that there is a correct answer to the question “which Turing machine do I really live in?” They then reason anthropically about which Turing machines give rise to such conscious experiences as theirs. They then use the same prior over Turing machines that I described above. And then they make the same calculation about the probability that “their” Turing machine outputs something that obeys the predicate P. So on the one hand, we could say that we are asking “what is the probability that the section of the universal prior which gives rise to these inhabitants outputs an output that obeys predicate P?” Or we could equivalently ask “what is the probability that this inhabitant ascribes to ‘its’ Turing machine outputting a string that obeys predicate P?”
There are facts that I find much easier to incorporate when thinking in the latter framework, such as “a work tape inhabitant knows nothing about the behavior of its Turing machine’s output tape, except that it has relative simplicity given the world that it knows.” (If it believes that its conscious existence depends on its Turing machine never having output a bit that differs from a data stream in a base world, it will infer other things about its output tape, but you seem to disagree that it would make that assumption, and I’m fine to go along with that). (If the fact were much simpler—“a work tape inhabitant knows nothing about the behavior of its Turing machine’s output tape” full stop—I would feel fairly comfortable in either framework.)
If it is the case that, for any action that a work tape inhabitant takes, the following is unchanged: [the probability that it (anthropically) ascribes to “its” Turing machine printing an output that obeys predicate P after it takes that action], then, no matter its choice of action, then the probability under the universal prior that the output obeys predicate P is also unchanged.
What if the work tape inhabitant only cares about the output when the the universal prior is being used for important applications? Let Q be the predicate [P and “the sequence begins with a sequence which is indicative of important application of the universal prior”]. The same logic that applies to P applies to Q. (It feels easier to talk about probabilities of predicates (expectations of Boolean functions) rather than expectations of general functions, but if we wanted to do importance weighting instead of using a strict predicate on importance, the logic is the same).
Writing about the fact I described above about what the inhabitants believe about their Turing machine’s output has actually clarified my thinking a bit. Here’s a predicate where I think inhabitants could expect certain actions to make it more likely that their Turing machine output obeys that predicate. “The output contains the string [particular 1000 bit string]”. They believe that their world’s output is simple given their world’s dynamics, so if they write that 1000 bit string somewhere, it is more likely for the predicate to hold. (Simple manipulations of the string are nearly equally more likely to be output).
So there are severe restrictions on the precision with which they can control even low-probability changes to the output, but not total restrictions. So I wasn’t quite right in describing it as a max-entropy situation. But the one piece of information that distinguishes their situation from one of maximum uncertainty about the output is very slight. So I think it’s useful to try to think in terms of how they get from that information to their goal for the output tape.
I was describing the situation where I wanted to maximize the probability where the output of our world obeys the predicate: “this output causes decision-maker simulators to believe that virtue pays”. I think I could very slightly increase that probability by trying to reward virtuous people around me. Consider consequentialists who want to maximize the probability of the predicate “this output causes simulator-decision-makers to run code that recreates us in their world”. They want to make the internals of their world such that there are simple relative descriptions for outputs for which that predicate holds. I guess I think that approach offers extremely limited and imprecise ability to deliberately influence the output, no matter how smart you are.
If an approach has very limited success probability, (i.e. very limited sway over the universal prior), they can focus all their effort on mimicking a few worlds, but then we’ll probably get lucky, and ours won’t be one of the ones they focus on.
From a separate recent comment,
But now that we've learned that physics is the game of life, we can make much better guesses about how to build a dataset so that a TM could output it. For example, we can:
- Build the dataset at a large number of places.
- [etc.]
...
I challenge you to find any plausible description of a rule that outputs the bits observed by a camera, for which I can't describe a simpler extraction rule that would output some set of bits controlled by the sophisticated civilization.
You're comparing the probability of one of these many controlled locations driving the output of the machine to the probability that a random camera does on an earth-like Turing machine drives the output. Whereas it seems to me like the right question is to look at the absolute probabilities that one of these controlled locations drives the output. The reason is that what they attempt to output is a mixture over many sequences that a decision-maker-simulator might want to know about. But if the sequence we're feeding in is from a camera on earth, than their antics only matter to the extent that their mixture puts weight on a random camera on earth. So they have to specify the random camera on an earth-like Turing machine too. They're paying the same cost, minus any anthropic update. So the costs to compare are roughly [- log prob. successful control of output + bits to specify camera on earth - bits saved from anthropic update] vs. [bits to specify camera on earth - bits saved from directly programmed anthropic update]. This framing seems to imply we can cross off [bits to specify camera on earth] from both sides.
Okay, now suppose they want the first N bits of the output of their Turing machine to obey predicate P, and they assign that a value of 100, and a they assign a value of 0 to any N-bit string that does not obey predicate P. And they don't value anything else. If some actions have a higher value than other actions, what information about the output tape dynamics are they using, and how did they acquire it?
Just look at the prior--for any set of instructions for the work tape heads of the Turing machine, flipping the "write-1" instructions of the output tape with the "write-0" instructions gives an equally probably Turing machine.
Suppose they know the sequence that actually gets fed to the camera.
If you're saying that they know their Turing machine has output x so far, then I 100% agree. What about in the case where they don't know?
If I flip a coin to randomize between two policies, I don't see how that mixed policy could produce more value for me than the base policies.
(ETA: the logical implications about the fact of my randomization don't have any weird anti-adversarial effects here).
If these consequentialists ascribed a value of 100 to the next output bit being 1, and a value of 0 to the next output bit being 0, and they valued nothing else, would you agree that all actions available to them have identical expected value under the distribution over Turing machines that I have described?
With randomization, you reduce the cost and the upside in concert. If a pair of shoes costs $100, and that's more than I'm willing to pay, I could buy the shoes with probability 1%, and it will only cost me $1 in expectation, but I will only get the shoes with probability 1/100.
It's definitely not too weird a possibility for me. I'm trying to reason backwards here--the best strategy available to them can't be effective in expectation at achieving whatever their goals are with the output tape, because of information-theoretic impossibilities, and therefore, any given strategy will be that bad or worse, including randomization.
We can get back to some of these points as needed, but I think our main thread is with your other comment, and I'll resist the urge to start a long tangent about the metaphysics of being "simulated" vs. "imagined".
So we end up with some leading hypotheses about the Turing machine we are running on, the history that gave rise to us, and the output rule used by that Turing machine.
I feel like this story has run aground on an impossibility result. If a random variable’s value is unknowable (but its distribution is known) and an intelligent agent wants to act on its value, and they randomize their actions, the expected log probability of them acting on the true value cannot exceed the entropy of the distribution, no matter their intelligence. (And if they’re wrong about the r.v.’s distribution, they do even worse). But lets assume they are correct. They know there are, say, 80 output instructions (two work tapes and one input tape, and a binary alphabet, and 10 computation states). And each one has a 1/3 chance of being “write 0 and move”, “write 1 and move”, or “do nothing”. Let’s assume they know the rules governing the other tape heads, and the identity of the computation states (up to permutation). Their belief distribution is (at best) uniform over these 3^80 possibilities. Is computation state 7 where most of the writing gets done? They just don’t know. It doesn’t matter if they’ve figured out that computation state 7 is responsible for the high-level organization of the work tapes. It’s totally independent. Making beacons is like assuming that computation state 7, so important for the dynamics of their world, has anything special to do with the output behavior. (Because what is a beacon if not something that speaks to internally important computation states?)
That’s all going along with the premise that when consequentialists face uncertainty, they flip a coin, and adopt certainty based on the outcome. So if they think it’s 50/50 whether a 0 or a 1 gets output, they flip a coin or look at some tea leaves, and then act going forward as if they just learned the answer. Then, it only costs 1 bit to say they decided “0”. But I think getting consequentialists to behave this way requires an intervention into their proverbial prefrontal cortices. If these consequentialists were playing a bandit game, and one arm gave a reward of 0.9 with certainty, and the other was 50/50 between a reward of 0 or 1, they obviously don’t flip a coin to decide whether to act as if it’s really reward 1 or reward 0. (I understand that Thompson sampling is a great strategy, but only when your uncertainty is ultimately empirically resolvable, and uncertainty about output behavior is not).
I think you’ll get the impression that I’m being miserly with bits here. And your relative profligacy comes from the fact that you expect they’ll make it up easily in the anthropic update. But you’ll ultimately have to claim that the anthropic update is more cheaply specified as a natural product of consequentialist civilization than through direct encoding. And if every step of the way in your story about how consequentialists arrive at this behavior, I make the point that this is not sensible goal-oriented behavior, and you make the point that it only takes a few bits to make it be what they would do anyway, then if you put it all together, I’m not just haggling over bits. If you put it all together, it looks to me like the consequentialists’ consequentialism is doing little to none of the work; every step of the way, reasonable behavior is being overridden, a few bits at a time. So then I ultimately claim, this anthropic update is not most parsimoniously described as “that thing that consequentialists sometimes produce” because it’s just not actionable for any consequentialists with reasonable epistemics.
In a different part of the story, you claim that if something composes 1% of what consequentialists value, we can assume that they flip 6 and a half coins, and with 1% probability they act as if that’s the only thing they value.
So we're interested in investing some tiny fraction of our resources in escaping from such simulations. The most important way to minimize the cost is by reducing probability. For example, if we decide to do something only with 1% probability (and in the other 99% of worlds focus exclusively on other things) then we only decrease our log probability by 7 bits. I actually doubt we'd waste those 7 bits, but it's worth noting how cheap this kind of breakout attempt could be.
This seems to me to be another case of supplanting consequentialists’ instrumental rationality with randomness. These are key parts of the story. The part where they make a wild guess about the output of their world, and the part where decide to pursue it in the first place are both places where reasonable goal-oriented behavior is being replaced with deference to the descriptive and normative wisdom of tea leaves; this is not just specifying one of the things they sometimes do naturally with small-but-non-negligible probability. It would be an unusual model of agency which claimed that: for a utility function , we have . And it would be an even more unusual model of agency which claimed that: for a world , we have . Even within large multiplicative fudge fators. I feel like I need the assistance of a Dutch bookie here.
Of course we don't know for sure that this is the particular way that the TM works, since we can't infer the output rules from our observations, and probably it isn't. But that's just like saying that any given TM probably doesn't output the pixels your camera. The point is that we are doing things that would cause a tiny fraction of the TMs containing us to output good sequences, and that's going to be a way higher fraction than those that happen to output the pixels of cameras
I don’t think it’s just like saying that. I think I have argued that the probability that consequentialists act on the belief that camera 12,041 has a direct line to the output tape is smaller than the probability that it is actually true. Likewise for something that appears “beacon-like” to the world’s residents. Given a state of total ignorance about the means by which they can affect the output tape, their guesses can be no better than the true prior distribution over what in the world has a direct line to the output tape. (This is contra: "the fact that they are trying puts them at a massive advantage" from your other comment; intelligence and effort don't work in a max-ent setting. With maximum entropy beliefs about the output channel, those silly no free lunch theorems of optimization do actually apply.) And then also given that ignorance, there is much less value in taking a stab in the dark. In general, with probabilistic mixtures over worlds , .
A few quick thoughts, and I'll get back to the other stuff later.
To be clear, people I know spent a lot more time than that thinking hard about the consensus algorithm, before coming to the strong conclusion that it was a fruitless path. I agree that this is worth spending >20 hours thinking about.
That's good to know. To clarify, I was only saying that spending 10 hours on the project of applying it to modern ML would not be enough time to deem it a fruitless path. If after 1 hour, you come up with a theoretical reason why it fails on its own terms--i.e. it is not even a theoretical solution--then there is no bound on how strongly you might reasonably conclude that it is fruitless. So this kind of meta point I was making only applied to your objections about slowdown in practice.
a "theoretical solution" to the realizability probelm at all.
I only meant to claim I was just doing theory in a context that lacks the realizability problem, not that I had solved the realizability problem! But yes, I see what you're saying. The theory regards a "fair" demonstrator which does not depend on the operation of the computer. There are probably multiple perspectives about what level of "theoretical" that setting is. I would contend that in practice, the computer itself is not among the most complex and important causal ancestors of the demonstrator's behavior, so this doesn't present a huge challenge for practically arriving at a good model. But that's a whole can of worms.
My main worry continues to be the way bad actors have control over an io channel, rather than the slowdown issue.
Okay good, this worry makes much more sense to me.
I felt I had remained quiet about my disagreement with you for too long
Haha that's fine. If you don't voice your objections, I can't respond to them!
I think let's step back for a second, though. Suppose you were in the epistemic position "yes, this works in theory, with the realizability assumption, with no computational slowdown over MAP, but having spent 2-10 hours trying to figure out how to distill a neural network's epistemic uncertainty/submodel-mismatch, and having come up blank..." what's the conclusion here? I don't think it's "my main guess is that there's no way to apply this in practice". Even if you had spent all the time since my original post trying to figure out how to efficiently distill a neural network's epistemic uncertainty, it's potentially a hard problem! But it also seems like a clear problem, maybe even tractable. See Taylor (2016) section 2.1--inductive ambiguity identification. If you were convinced that AGI will be made of neural networks, you could say that I have reduced the problem of inner alignment to the problem of diverse-model-extraction from a neural network, perhaps allowing a few modifications to training (if you bought that the claim that the consensus algorithm is a theoretical solution). I have never tried to claim that analogizing this approach to neural networks will be easy, but I don't think you want to wait to hear my formal ideas until I have figured out how to apply them to neural networks; my ideal situation would be that I figure out how to do something in theory, and then 50 people try to work on analogizing it to state-of-the-art AI (there are many more neural network experts out there than AIXI experts). My less ideal situation is that people provisionally treat the theoretical solution as a dead end, right up until the very point that a practical version is demonstrated.
If it seemed like solving inner alignment in theory was easy (because allowing yourself an agent with the wherewithal to consider "unrealistic" models is such a boon), and there were thus lots of theoretical solutions floating around, any given one might not be such a strong signal: "this is the place to look for realistic solutions". But if there's only one floating around, that's a very a strong signal that we might be looking in a fundamental part of the solution space. In general, I think the most practical place to look for practical solutions is near the best theoretical one, and 10 hours of unsuccessful search isn't even close to the amount of time needed to demote that area from "most promising".
I think this covers my take on a few of your points, but some of your points are separate. In particular, some of them bear on the question of whether this really is an idealized solution in the first place.
With Ted out of the running, the top 100 hypotheses are now all malign, and can coordinate some sort of treacherous turn.
I think the question we are discussing here is: "yes, with the realizability assumption, existence of a benign model in the top set is substantially correlated over infinite time, enough so that all we need to look at is the relative weight of malign and benign models, BUT is the character of this correlation fundamentally different without the realizability assumption?" I don't see how this example makes that point. If the threshold of "unrealistic" is set in such a way that "realistic" models will only know most things about Sally, then this should apply equally to malign and benign models alike. (I think your example respects that, but just making it explicit). However, there should be a benign and malign model that knows about Sally's affinity for butter but not her allergy to flowers, and a benign and a malign model that knows the opposite. It seems to me that we still end up just considering the relative weight of benign and malign models that we might expect to see.
(A frugal hypothesis generating function instead of a brute force search over all reasonable models might miss out on, say, the benign version of the model that understands Sally's allergies; I do not claim to have identified an approach to hypothesis generation that reliably includes benign models. That problem could be one direction in the research agenda of analogizing this approach to state-of-the-art AI. And this example might also be worth thinking about in that project, but if we're just using the example to try to evaluate the effect of just removing the realizability assumption, but not removing the privilege of a brute search through reasonable models, then I stand by the choice to deem this paragraph parenthetical).
Why will there be one best? That's the realizability assumption. There is not necessarily a unique model with lowest bayes loss. Another way of stating this is that Bayesian updates lack a convergence guarantee; hypotheses can oscillate forever as to which is on top.
Yeah I was thinking that the realistic setting was a finite length setting, with the one best being the best at the end. (And if it is best at the end, you can put a lower bound on how low its posterior weight ever was, since it's hard to recover from having vanishingly small weight, and then alpha just needs to be set to include that). But fair enough to be interested an infinite lifetime with a finite model class that does not include the truth. So yeah, a model's ranking can oscillate forever, although I think intelligent systems won't really do this in practice? I think in an infinite lifetime, it is reasonable in practice to assume that if something is generally intelligent, it has some sort of ability to generate new hypotheses over time. And if you have two hypotheses about Sally, one which explains her relationship to butter, and one which explains her relationship to flowers, it's pretty easy to generate another that combines these. So I'm not sure how instructive it is to consider the finite model class/infinite horizon Bayesian sequence prediction as a model for state-of-the-art AI.
So, what, the human has to act like they're driving the car forever, even though the AI autopilot is really good?
The suggestion I make later is that a few timesteps in advance, the human is told to start acting like they're driving the car, but they're not controlling it yet, and then comes the timestep when autopilot actually defers to them, and that's the behavior that gets recorded. Your response is
Is this procedure supposed to strike me as so airtight that an inner optimizer could never exploit it?
That proposal was making a different point--it was refuting exactly the concern that the human would have to act like they're driving the car forever. This protocol just defends active imitation learning in general as potentially viable for the operation of heavy machinery. Assume the demonstrator is sitting in a driving simulator, and all the relevant info shows up on their screens; this allows a small team to be the demonstrators for many autopilots simultaneously. It's not a question of whether an inner optimizer could exploit this protocol by finding the most inconvenient time for the human to be called in, and then querying the human--it's much worse than that! If there's a bad time for the human drive to be called in, then the autopilot will fail at that time regardless; the inner optimizer doesn't even have to arrange an actual query. So there's nothing to "exploit" really. And defending against the "much worse" version, I do think that the policy of a human driver with 5 seconds of acquaintance with the current state, does not crash the car at a rate significantly higher than a real human driver. Also, it may be worth saying that I don't think we really need general intelligence to operate heavy machinery at human level, so I'm not sure the heavy machinery case is very important to consider in the first place.
In terms of evaluating this on theoretical terms, the feedback call button is an output channel that malign stuff has access to. This feels kind of like "case closed" to me; I take it as a given that a sufficiently intelligent malign actor could exploit this.
You said you'd continue to think about the points I made about why I don't have the same reaction to this, so we can come back to those later. But stepping back, I think I want to make a point about the information loss you impose on your beliefs when applying the security mindset or Murphy's Law. A malign model gets to say "query now" once! And then it is removed from the posterior because of its wrong prediction. That is not many bits for an output channel. But... yes, it is an output channel. And if you insist that all output channels for malign stuff transmit world-ending signals, just like all aquariums will leak, I agree that that is sometimes a useful perspective to adopt and run with, but the exercise of making actual probability estimates about events is potentially different from exercising a security mindset. I mentioned information loss. With a security mindset, there is no point in distinguishing a malign agent with a 20 bit output channel from one with an unbounded output channel. You just close the case on both. But we really might live in a world where the former is existentially safe, and the latter is not--I mean that even seems likely to me. Even from my perspective, where I don't think a 20 bit output channel is dangerous, I understand the merit in disregarding that belief as the motivation for a research agenda, but I do think that's a whole different sort of "epistemic status". So when you say "I take it as a given that a sufficiently intelligent malign actor could exploit this" that makes sense to me as an action, but not as a proposition. (I mean, I think it's incorrect, so I certainly don't think it's self-evidently true).
I guess I can expand on the object level too. Like, take a moment to really stare at 01110010101011010000, and imagine that you take it to be a signal from some deeply wise agent that arranged for a query at timestep number 469,712. Suppose further that you already think this deeply wise agent is not malign, but rather is much maligned! (So it doesn't even need to spend the first few bits enthralling you). What do you do now? Or maybe timestep 469, 712 doesn't mean that bitstring--maybe it means "this is a very meaningful world-state, and of similar world-states, the fact that this one was chosen tells me I should run program X on a supercomputer, so it can engage in some good, clean world-optimization". Look around you right now--if someone told you "now is important, so think about that, and go run a program", what program do you produce? Am I taking a huge risk even asking that, because there's a precise time such that if I did, you'd run that program and it would end the world?
To state the least of our problems first: this requires a 100x slowdown in comparison with the state-of-the-art deep learning (or whatever) we're layering the consensus algorithm on top of
I think you’re imagining deep learning as a MAP-type approach—it just identifies a best hypothesis and does inference with that. Comparing the consensus algorithm with (pure, idealized) MAP, 1) it is no slower, and 2) the various corners that can be cut for MAP can be cut for the consensus algorithm too. Starting with 1), the bulk of the work for either the consensus algorithm or a MAP approach is computing the posterior to determine which model(s) is(are) best. In an analogy to neural networks, it would be like saying most of the work comes from using the model (the forward pass) rather than arriving at the model (the many forward and backward passes in training). Regarding 2), state-of-the-art-type AI basically assumes approximate stationarity when separating a training phase from a test/execution phase. This is cutting a huge corner, and it means that when you think of a neural network running, you mostly think about it using the hypothesis that it has already settled on. But if we compare apples to apples, a consensus algorithm can cut the same corner to some extent. Neither a MAP algorithm nor a consensus algorithm is any better equipped than the other to, say, update the posterior only when the timestep is a power two. In general, training (be it SGD or posterior updating) is the vast bulk of the work in learning. To select a good hypothesis in the first place you will have already had to consider many more; the consensus algorithm just says to keep track of the runner ups.
Third, the consensus algorithm requires a strong form of realizability assumption, where you not only assume that our Bayesian space contains the true hypothesis, but furthermore, that it's in the top 100 (or whatever number we choose). This hypothesis has to be really good: we have to think that malign hypotheses never out-guess the benign hypothesis. Otherwise, there's a chance that we eliminate the good guy at some point (allowing the bad guys to coordinate on a wrong answer). But this is unrealistic! The world is big and complex enough that no realistic hypothesis has all the answers.
I don’t understand what out-guess means. But what we need is that the malign hypothesis don’t have substantially higher posterior weight than the benign ones. As time passes, the probability of this happening is not independent. The result I show about the probability of the truth being in the top set applies to all time, not any given point in time. I don’t know what “no realistic hypothesis has all the answers” means. There will be a best “realistic” benign hypothesis, and we can talk about that one.
Michael Cohen seems to think that restricting to imitation learning makes the realizability assumption realistic
Realistic in theory! Because the model doesn’t need to include the computer. I do not think we can actually compute every hypothesis simpler than a human brain in practice.
When you go from an idealized version to a realistic one, all methods can cut corners, and I don’t see a reason to believe that the consensus algorithm can’t cut corners just as well. Realistically, we will have some hypothesis-proposing heuristic, strong enough to identify models one of which is accurate enough to generate powerful agency. This heuristic will clearly cast a wide net (if not, how would it magically arrive at a good answer? It’s internals would need some hypothesis-generating function). Rather than throwing out the runner ups, the consensus algorithm stores them. The hypothesis generating heuristic is another attack surface for optimization daemons, and I won’t make any claims for now about how easy or hard it is to prevent such a thing.
to apply this to something like deep learning, we need to think that each run has an independent chance of creating safe AGI
Evan and I talked along these lines for a bit. My basic position is that if “local search” is enough to get to general intelligence, our algorithms will be searching in spaces (or regions) where diverse hypothesis are close. Diverse hypothesis generation is just crucial for general intelligence. I do not advocate training GPT-N with 10^100 different initializations. I don’t think you have to, and I don’t think it would help much.
Fourth, the consensus algorithm still gives inner optimizers control of when the system asks for more feedback. This in itself could be catastrophic! For example, it could ask at the worst times, when it anticipates the human feedback to be of poor quality, or when asking itself has catastrophic consequences. A silly example would be causing a self-driving car to ask for feedback right when it needs to enact difficult evasive maneuvers to avoid a crash.
It’s imitation learning so whatever the person would do when stepping in is what the algorithm learns to do when they don’t. If the human action would be of poor quality if the human were called upon to take the action, then all the reasonable models, benign ones included, will predict that poor action, and the poor action will be taken regardless of whether the human was brought into the loop. If a person would be unable to avoid a crash when stepping in at the last minute, the imitation learner will crash the car too! If 99 hypotheses are correctly predicting the human action, then the 100th hypotheses can correctly predict it too, and the human action is executed, or it can err deliberately, in which case the human is called and the human action is executed anyway. You might have to modify this protocol a bit before using it to learn human policies for operating heavy machinery.
Fifth, although (as Michael Cohen notes) the frequency of asking for feedback will rapidly diminish, the learning algorithm converges to the correct hypothesis extremely slowly… We have to face the fact that it might require human feedback at any point in the future.
Yeah this feels like a small cost to me. One person can be doing this for many instances at once. If you want to mimic a human that has been attentive for a given period of time (because you want to know how a human drives a car, having already gotten their bearings previously, rather than how human drives a car having just been transported into it) you can know a few timesteps in advance if there is any likely circumstance in which you will need to query the demonstrator, and you can use the lead time to get the demonstrator acquainted. The query complexity will go up of course because there will be false alarms, but this is a totally surmountable problem for active learning.
The fourth point [controlling when to ask for more feedback] really seems like the deal-breaker. Malign mesa-optimizers could intelligently leverage this to communicate with the outside world, manipulate the training process, etc.
A malign mesa-optimizer has one moment to err, and then it will be discarded from the posterior. There’s not much it could say to us with a well-timed "now!". If multiple mesa-optimzers coordinate around a message, any mesa-optimizer not on board with this mission could disrupt the message. If you like, we could add a 1/million chance of querying to every timestep. Finally, with butterfly-effects, I don’t think that even a billion well-timed asks for help could precipitate human extinction. This is a bit all over the map, and this may be the point you want to press, but if you do, I think it is worth logging that if this is only attack surface left, we shouldn't forget to put a stake in the ground marking "lots of progress". If you believe that a single malign world-model could derail us if we use a Solomonoff predictor, but a consensus predictor would only be derailed by a supermajority of malign world-models, aligned with each other, that is a qualitative difference.
My third and final example: in one conversation, someone made a claim which I see as "exactly wrong": that we can somehow lower-bound the complexity of a mesa-optimizer in comparison to a non-agentic hypothesis (perhaps because a mesa-optimizer has to have a world-model plus other stuff, where a regular hypothesis just needs to directly model the world). This idea was used to argue against some concern of mine.
The problem is precisely that we know of no way of doing that! If we did, there would not be any inner alignment problem! We could just focus on the simplest hypothesis that fit the data, which is pretty much what you want to do anyway!
Maybe this was someone else, but it could have been me. I think MAP probably does solve the inner alignment problem in theory, but I don’t expect to make progress resolving that question, and I’m interested in hedging against being wrong. Where you say, “We know of no way of doing that” I would say, “We know of ways that might do that, but we’re not 100% sure”. On my to-do list is to write up some of my disagreements with Paul’s original post on optimization daemons in the Solomonoff prior (and maybe with other points in this article). I don’t think it’s good to argue from the premise that a problem is worth taking seriously, and then see what follows from the existence of that problem, because a problem can exist with 10% probability and be worth taking seriously, but one might get in trouble embedding its existence too deeply in one’s view of the world, if it is still on balance unlikely. That’s not to say that most people think Paul’s conclusions are <90% likely, just that one might.
I think there would still be an inner alignment problem even if deceptive models were in fact always more complicated than non-deceptive models—i.e. if the universal prior wasn't malign—which is just that the neural net prior (or whatever other ML prior we use) might be malign even if the universal prior isn't (and in fact I'm not sure that there's even that much of a connection between the malignity of those two priors)
Agree.
There will be plenty of functions that have fewer bits in their encoding than the real function used by the demonstrator.
I don't think this is a problem. There will be plenty of them, but when they're wrong they'll get removed from the posterior.
A policy outputs a distribution over , and equations 3 and 4 define what this distribution is for the imitator. If it outputs (0, a), that means and and and if it outputs (1, a), that means and . When I say
The 0 on the l.h.s. means the imitator is picking the action itself instead of deferring to the demonstrator,
that's just describing the difference between equations 3 and 4. Look at equation 4 to see that when , the distribution over the action is equal to that of the demonstrator. So we describe the behavior that follows as "deferring to the demonstrator". If we look at the distribution over the action when , it's something else, so we say the imitator is "picking its own action".
The 0 on the l.h.s. means the imitator is picking the action itself instead of deferring to the demonstrator or picking one of the other actions???
The 0 means the imitator is picking the action, and the means it's not picking another action that's not .
What's the distinction between training and deployment when the model can always query for more data?
It may be helpful to point to specific sections of such a long paper.
(Also, I agree that a neural network trained trained with that reward could produce a deceptive model that makes a well-timed error.)
it certainly should have to consider different models of what the next transition will be like.
Yeah I was agreeing with that.
even though the Q learner is trained using its own point estimate of the next state, it isn't, at inference time, given access to that point estimate.
Right, but one thing the Q-network, in its forward pass, is trying to reproduce is the point of estimate of the Q-value of the next state (since it doesn't have access to it). What it isn't trying to reproduce, because it isn't trained that way, is multiple models of what the Q-value might be at a given possible next state.
I interpreted this bit as talking about RL
I certainly don't think SGD is a powerful enough optimization process to do science directly, but it definitely seems powerful enough to find an agent which does do science.
But taking us back out of RL, in a wide neural network with selective attention that enables many qualitatively different forward passes, gradient descent seems to be training the way different models get proposed (i.e. the way attention is allocated), since this happens in a single forward pass, and what we're left with is a modeling routine that is heuristically considering (and later comparing) very different models. And this should include any model that a human would consider.
I think that is main thread of our argument, but now I'm curious if I was totally off the mark about Q-learning and policy gradient.
but overall I think these sorts of differences are pretty minor and shouldn't affect whether these approaches can reach general intelligence or not.
I had thought that maybe since a Q-learner is trained as if the cached point estimate of the Q-value of the next state is the Truth, it won't, in a single forward pass, consider different models about what the actual Q-value of the next state is. At most, it will consider different models about what the very next transition will be.
a) Does that seem right? and b) Aren't there some policy gradient methods that don't face this problem?
So, if I understand the paper right, if was still in the set of top policies at time , the agent would not take the action to kill the demonstrator, but ask the demonstrator to take an action instead, which avoids the bad outcome, and might also demote all of the treacherous turn policies out of the top set.
This is exactly right.
But I am also reading in the abstract of the paper that
[...] queries to the demonstrator rapidly diminish in frequency
The reason this is consistent is that queries will become infrequent, but they will still be well-timed. The choice of whether to query depends on what the treacherous models are doing. So if the treacherous models wait for a long time, then we will have a long time where no queries are needed, and as soon as they decide to be treacherous, we thwart that plan with a query to the demonstrator. So it does not imply that
over time, it is likely that might disappear from the top set
Does this thought experiment look reasonable or have I overlooked something?
Yep, perfectly reasonable!
What about the probability that is still in the set of top policies at time ?
If we set small enough, we can make it arbitrarily like that never leaves the set of top policies.
That's possible. But it seems like way less of a convergent instrumental goal for agents living in a simulated world-models. Both options--our world optimized by us and our world optimized by a random deceptive model--probably contain very little of value as judged by agents in another random deceptive model.
So yeah, I would say some models would think like this, but I would expect the total weight on models that do to be much lower.
This is very nice and short!
And to state what you left implicit:
If , then in the setting with no malign hypotheses (which you assume to be safe), 0 is definitely the output, since the malign models can only shift the outcome by , so we assume it is safe to output 0. And likewise with outputting 1.
I'm pretty sure removing those is mostly just a technical complication
One general worry I have about assuming that the deterministic case extends easily to the stochastic case is that a sequence of probabilities that tends to 0 can still have an infinite sum, which is not true when probabilities must , and this sometimes causes trouble. I'm not sure this would raise any issues here--just registering a slightly differing intuition.
I certainly don't think SGD is a powerful enough optimization process to do science directly, but it definitely seems powerful enough to find an agent which does do science.
Okay I think we've switched from talking about Q-learning to talking about policy gradient. (Or we were talking about the latter the whole time, and I didn't notice it). The question that I think is relevant is: how are possible world-models being hypothesized and analyzed? That's something I expect to be done with messy heuristics that sometimes have discontinuities their sequence of outputs. Which means I think that no reasonable DQN is will be generally intelligent (except maybe an enormously wide one attention-based one, such that finding models is more about selective attention at any given step than it is about gradient descent over the whole history).
A policy gradient network, on the other hand, could maybe (after having its parameters updated through gradient descent) become a network that, in a single forward pass, considers diverse world-models (generated with a messy non-local heuristic), and analyzes their plausibility, and then acts. At the end of the day, what we have is an agent modeling world, and we can expect it to consider any model that a human could come up with. (This paragraph also applies to the DQN with a gradient-descent-trained method for selectively attending to different parts of a wide network, since that could amount to effectively considering different models).
If you are willing to query only a billion points, and there are a billion competent treacherous models, and if the game is zero-sum between you and the models, then you're going to get at least as screwed as by the billionth-best opportunity for treachery.
The treacherous models make their decisions first, and whether we query at a timestep depends on whether they chose that moment to be treacherous. We don't have to say: this timestep is one of the billion best for treachery--guess we had better query. The decision process a treacherous model faces is:
- From past observations, does it seem alpha is low enough that an honest model is in the top few?
- If no, [let's just assume whatever happens here is bad]
- If yes, is the honest model going to leave the top few any time?
- If yes, wait. [let's assume this ends badly]
- If no, admit defeat, because whenever you err, that will be when a query gets triggered. Salvage the situation by trying to get as much mileage as you can by decreasing probability on some action the demonstrator would take that you don't like much in order to shift probability mass to some other action they would take that you like a little better. OR
- Hope that 1) other treacherous models, like soldiers on the front line storming the enemy and facing certain death so that their brothers behind them can win the battle, give erring a go anyway. And hope that 2) the operators quietly increase alpha after the billionth query. Acting on this hope, wait. (The front-line treacherous models are not entirely self-sacrificial; they might figure they might as well hope they were wrong about alpha. But the models know when queries happen, because every observation they get comes from a query, and it wouldn't take long to figure out that alpha really truly isn't high enough for them to execute treachery successfully).
Above in the thread, I was talking about the game theory of step 6 (although I maybe wasn't making the case as strongly as I could have). Which models would ever volunteer? They don't care about the long term goal of exhausting our query resources so other treacherous models can capitalize on a quietly increased alpha. All that is to say, I don't follow the comment I excerpted.
(In case this comes up later: I think about the long-run percentage of query-timesteps that we would tolerate, not the total number. So if we ever have cause to consider a case where attempted treachery "merely gets delayed" but still eventually happens, and still results in a query, then depending on how much it is delayed, that could still qualify as win).
Does it make sense to talk about , which is like in being treacherous, but is uses the true model instead of the honest model ? I guess you would expect to have a lower posterior than ?
If we ever produce generally intelligent predictors (or "accurate world-models" in the terminology we've been using so far), we will need a process that is much more efficient than evolution.
But also, I certainly don't think that in order to be generally intelligent you need to start with a generally intelligent subroutine. Then you could never get off the ground. I expect good hypothesis-generation / model-proposal to use a mess of learned heuristics which would not be easily directed to solve arbitrary tasks, and I expect the heuristic "look for models near the best-so-far model" to be useful, but I don't think making it ironclad would be useful.
Another thought on our exchange:
Me: we can expect a good heuristic search through models will identify any model that a human could hypothesize
You: I think I would expect essentially all models that a human could hypothesize to be in the search space—but if you're doing a local search, then you only ever really see the easiest to find model with good behavior
If what you say is correct, then it sounds like exclusively-local search precludes human-level intelligence! (Which I don't believe, by the way, even if I think it's a less efficient path). One human competency is generating lots of hypotheses, and then having many models of the world, and then designing experiments to probe those hypotheses. It's hard for me to imagine that an agent that finds an "easiest-to-find model" and then calls it a day could ever do human-level science. Even something as simple as understanding an interlocuter requires generating diverse models on the fly: "Do they mean X or Y with those words? Let me ask a clarfying question."
I'm not this bearish on local search. But if local search is this bad, I don't think it is a viable path to AGI, and if it's not, then the internals don't for the purposes of our discussion, and we can skip to what I take to be the upshot:
we can expect a good heuristic search through models will identify any model that a human could hypothesize
Specifically, they cooperate in that they perfectly mimic the true model up until the point where...
This thread began by considering deceptive models cooperating with each other in the sense of separating the timing of their treacherous turns in order to be maximally annoying to us. So maybe our discussion on that topic is resolved, and we can move on to this scenario.
up until the point where the deceptive models make up enough of the posterior that the true model is no longer being consulted
if alpha is low enough, this won't ever happen, and if alpha is too high, it won't take very long. So I don't think this scenario is quite right.
Then the question becomes, for an alpha that's low enough, how long will it take until queries are infrequent, noting that you need a query any time any treacherous model with enough weight decides to take a treacherous turn?
So would you say you disagree with the claim
I think that arbitrary limits on heuristic search of the form "the next model I consider must be fairly close to the last one I did" will not help it very much if it's anywhere near smart enough to merit membership in a generally intelligent predictor.
?