# 28

The Pointers ProblemAIRationality
Frontpage

I've recently had several conversations about John Wentworth's post The Pointers Problem. I think there is some confusion about this post, because there are several related issues, which different people may take as primary. All of these issues are important to "the pointers problem", but John's post articulates a specific problem in a way that's not quite articulated anywhere else.

I'm aiming, here, to articulate the cluster of related problems, and say a few new-ish things about them (along with a lot of old things, hopefully put together in a new and useful way). I'll indicate which of these problems John was and wasn't highlighting.

This whole framing assumes we are interested in something like value learning / value loading. Not all approaches rely on this. I am not trying to claim that one should rely on this. Approaches which don't rely on human modeling are neglected, and need to be explored more.

That said, some form of value loading may turn out to be very important. So let's get into it.

Here's the list of different problems I came up with when trying to tease out all the different things going on. These problems are all closely interrelated, and feed into each other to such a large extent that they can seem like one big problem.

(0. Goodhart. This is a background assumption. It's what makes getting pointers right important.)

1. Amplified values. Humans can't evaluate options well enough that we'd just want to optimize the human evaluation of options. This is part of what John is describing.
2. Compressed pointer problem. We can't realistically just give human values to a system. How can we give a system a small amount of information which "points at" human values, so that it will then do its best to learn human values in an appropriate way?
3. Identifiability problems for value learning. This includes Stuart's "no free lunch" argument that we can't extract human values (or beliefs) just with standard ML approaches.
4. Ontology mismatch problem. Even if you could extract human values (and beliefs) precisely, whatever that means, an AI doesn't automatically know how to optimize them, because they're written in a different ontology than the AI uses. This is what John is primarily describing.
5. Wireheading and human manipulation. Even if we solve both the learning problem and the ontology mismatch problem, we may still place the AI under some perverse incentives by telling it to maximize human values, if it's also still uncertain about those values. Hence, there still seems to be an extra problem of telling the AI to "do what we would want you to do" even after all of that.

# 0. Goodhart

As I mentioned already, this is sort of a background assumption -- it's not "the pointers problem" itself, but rather, tells us why the pointers problem is hard.

Simply put, Goodhart's Law is an argument that an approximate value function is almost never good enough. You really need to get quite close before optimizing the approximate version is a good way to optimize human values. Scott gives four different types of Goodhart, which all feed into this.

## Siren Worlds

Even if we had a perfect human model which we could use to evaluate options, we would face the Siren Worlds problem: we optimize for options which look good to humans, but this is different from options which are good.

We can't look at (even a sufficient summary of) the entire universe. Yet, we potentially care about the whole universe. We don't want to optimize just for parts we can look at or summarize, at the expense of everything else.

This shows that, to solve the pointers problem, we need to do more than just model humans perfectly. John's post talked about this problem in terms of "lazy evaluation": we humans can only instantiate small parts of our world model when evaluating options, but our "true values" would evaluate everything.

I'm referring to that here as the problem of specifying amplified values.

# 1. Amplified Values Problem

The problem is: humans lack the raw processing power to properly evaluate our values. This creates a difficulty in what it even means for humans to have specific values. It's easy to say "we can't evaluate our values precisely". What's difficult is to specify what it would mean for us to evaluate our values more precisely.

A quick run-down of some amplification proposals:

• Iterated Amplification: amplification is defined recursively as the answer a human would give if the human had help from other amplified humans who could answer questions.
• Debate: the amplified human opinion is defined as the winner of an idealized debate process, judged by humans (with AI debaters).
• CEV: we define amplified human values as what we would think if we were somewhat smarter, knew somewhat more, and had grown up longer together.
• Recursive Reward Modeling: we define human values as a utility function which humans specify with the help of powerful agents, who are themselves aligned through a recursive process.

Aside: if we think of siren worlds as the primary motivator for amplification, then Stuart's no-indescribable-hellworlds hypothesis is very relevant for thinking about what amplification means. According to that hypothesis, bad proposals must be "objectionable" in the sense of having an articulable objection which would make a human discard the bad proposal. If this is the case, then debate-like proposals seem like a good amplification technique: it's the systematic unearthing of objections.

Now, a viable proposal needs two things:

1. An abstract statement of what it means to amplify;
2. A concrete method of getting there.

For example, Iterated Amplification gives HCH as the abstract model of an amplified human, and the iterated amplification training method as its concrete proposal for getting there.

Crossing this bridge is the subject of the next point, compressed pointers:

# 2. Compressed Pointer Problem

The basic idea behind compressed pointers is that you can have the abstract goal of cooperating with humans, without actually knowing very much about humans. In a sense, this means having aligned goals without having the same goals: your goal is to cooperate with "human goals", but you don't yet have a full description of what human goals are. Your value function might be much simpler than the human value function.

In machine-learning terms, this is the question of how to specify a loss function for the purpose of learning human values.

Some important difficulties of compressed pointers are that they seem to lead to new problems of their own, in the form of wireheading and human manipulation problems. We will discuss those problems later on.

The other important sub-problem of compressed pointers is, well, how do you actually do the pointing? An assumption behind much of value learning research is that we can point to the human utility function via a loss function which learns a model of humans which decomposes them into a utility function, a probability distribution, and a model of human irrationality. We can then amplify human values just by plugging that utility function into better beliefs and a less irrational decision-making process. I argue against making such a strong distinction between values and beliefs here, here, and here, and will return to these questions in the final section. Stuart Armstrong argues that such decompositions cannot be learned with standard ML techniques, which is the subject of the next section.

# 3. Identifiability Problems for Value Learning

Stuart Armstrong's no-free-lunch result for value learning shows that the space of possible utility functions consistent with data is always too large, and we can't eliminate this problem even with Occam's razor: even when restricting to simple options, it's easy to learn precisely the wrong values (IE flip the sign of the utility function).

One thing I want to emphasize about this is that this is just one of many possible non-identifiability arguments we could make. (Identifiability is the learning-theoretic property of being able to distinguish the correct model using the data; non-identifiability means that many possibilities will continue to be consistent, even with unlimited data.)

Representation theorems in decision theory, such as VNM, often uniquely give us the utility function of an agent from a set of binary decisions which the agent would make. However, in order to get a unique utility function, we must usually ask the agent to evaluate far more decisions than is realistic. For example, Savage's representation theorem is based on evaluating all possible mappings from states to outcomes. Many of these mappings will be nonsensical -- not only never observed in practice, but nowhere near anything which ever could be observed.

This suggests that just observing the actual decisions an agent makes is highly insufficient for pinning down utility functions. What the human preferred under the actual circumstances does not tell us very much about what the human would have preferred under different circumstances.

Considerations such as this point to a number of ways in which human values are not identifiable from human behavior alone. Stuart's result is particularly interesting in that it strongly rules out fixing the problem via simplicity assumptions.

## Learning with More Assumptions

Stuart responds to this problem by suggesting that we need more input from the human. His result suggests that the standard statistical tools alone won't suffice. Yet, humans seem to correctly identify what each other want and believe, quite frequently. Therefore, humans must have prior knowledge which helps in this task. If we can encode those prior assumptions in an AI, we can point it in the right direction.

However, another problem stands in our way even then -- even if the AI could perfectly model human beliefs and values, what if the AI does not share the ontology of humans?

# 4. Ontology Mismatch Problem

As I stated earlier, I think John's post was discussing a mix of the ontology mismatch problem and the amplification problem, focusing on the ontology mismatch problem. John provided a new way of thinking about the ontology mismatch problem, which focused on the following claim:

Claim: Human values are a function of latent variables in our model of the world.

Humans have latent variables for things like "people" and "trees". An AI needn't look at the world in exactly the same way, so it needn't believe things exist which humans predicate value on.

This creates a tension between using the flawed models of humans (so that we can work in the human ontology, thus understanding human value) vs allowing the AI to have better models (but then being stuck with the ontology mismatch problem).

As a reminder of what latent variables are, let's take a look at two markov networks which both represent the relationship of five variables:

In the example on the left, we posit a completely connected network, accounting for all the correlations. In the example on the right, we posit a new latent variable which accounts for the correlations we observed.

As a working example, depression is something we talk about as if it were a latent variable. However, many psychologists believe that it is actually a set of phenomena which happen to have strong mutually reinforcing links.

Considered as a practical model, we tend to prefer models which posit latent variables when:

• They fit the data better, or equally well;
• Positing latent variables decreases the description complexity;
• The latent variable model can be run faster (or has a tolerable computational cost).

Mother nature probably has similar criteria for when the brain should posit latent variables.

However, considered as an ontological commitment, it seems like we only want to posit latent variables when they exist. When a Bayesian uses the model on the right instead of the model on the left, they believe in  as an extra fact which can in principle vary independently of the other variables. ("Independently" in the logical sense, here, not the probabilistic sense.)

So there is ambiguity between latent variables as pragmatic tools vs ontological commitments. This leads to our problem: a Bayesian (or a human), having invented latent variables for the convenience of their predictive power, may nonetheless ascribe value to specific states of those variables. This presents a difficulty if another agent wishes to help such an agent, but does not share its ontological commitments.

## Helper's Perspective

Suppose for a moment that humans were incapable of understanding depression as a cluster of linked variables, simply because that view of reality was too detailed to hold in the mind -- but that humans wanted to eliminate depression. Imagine that we have an AI which can understand the details of all the linked variables, but which lacks a hidden variable "depression" organizing them all. The AI wants to help humans, and in theory, should be able to use its more detailed understanding of depression to combat it more effectively. However, the AI lacks a concept of depression. How can it help humans combat something which isn't present in its understanding of reality?

What we don't want to do is form the new goal of "convince humans depression has been cured". This is a failure mode of trying to influence hidden variables you don't share.

Another thing we don't want to do is just give up and use the flawed human model. This will result in poor performance from the AI.

John Wentworth suggests taking a translation perspective on the ontology mismatch problem: we think of the ontologies as two different languages, and try to set up the most faithful translation we can between the two. The AI then tries to serve translated human values.

I think this perspective is close but not quite right. First note that there may not by any good translation between the two. We would like to gracefully fail in that case, rather than sticking with the best translation. (And I'd prefer that to be a natural consequence of our defined targets, rather than something extra added on.) Second, I don't think there's a strong justification for translation when we look at things from our perspective.

## Helpee's Perspective

Imagine that we're this agent who wants to get rid of depression but can't quite understand the universe in which depression comes apart into several interacting variables.

We are capable of watching the AI and forming beliefs about whether it is eliminating depression. We are also capable of looking at the AI's design (if not its learned models), and forming beliefs about things such as whether the AI is trying to fool us

We can define what it means for the AI to track latent variables we care about at the meta-level: not that it uses our exact model of reality, but that in our model, its estimates of the latent variables track the truth. Put statistically, we need to believe there is a (robust) correlation between the AI's estimate of the latent variable and the true value. (Not a correlation between the AI's estimation and our estimation -- that would tend to become true as a side effect, but if it were the target then the AI would just be trying to fool us.)

Critically, it is possible for us to believe that the AI tracks the truth better than we do. IE, it will be possible for us to become confident that "the AI understands depression better than we do" and have more faith in the AI's estimation of the latent variable than our own.

To see why, imagine that you (with your current understanding of depression) were talking to a practicing, certified psychiatrist (an MD) who also has an undergraduate degree in philosophy (has a fluent understanding of philosophy of language, philosophy of science, and philosophy of mind -- and, from what you gather, has quite reasonable positions in all these things), and, on top of all that, a PhD in research psychology. This person has just recently won a Nobel prize for a new cognitive theory of depression, which has (so far as you can tell) contributed significantly to our understanding of the brain as a whole (not only depression), and furthermore, has resulted in more effective therapies and drugs for treating depression.

You might trust this person considerably more than yourself, when it comes to diagnosing depression and judging fine degrees of how depressed a person is.

To have a model which includes a latent variable is not, automatically, to believe that you have the best possible model of that latent variable. However, our beliefs about what constitutes a more accurate way of tracking the truth are subjective -- because hidden variables may not objectively exist, it's up to us to say what would constitute a more accurate model of them. This is part of my concept of normativity.

So, I believe a step in the right direction would be for AIs to try to maximize values according to models which, according to human beliefs, track the things we care about well.

This captures the ontology problem and the value amplification problem at the same time; the AI is trying to use "better" models in precisely the sense that we care about for both problems.

## Do we need to solve the ontology mismatch problem?

It bears mentioning that we don't necessarily need to be convinced that the AI tracks the truth of variables which we care about, to be convinced that the AI enacts policies which accomplish good things. This is part of what policy approval was trying to get at, by suggesting that an AI should just be trying to follow a policy which humans expect to accomplish good things.

If we can formalize learning normativity, then we might just want an AI to be doing what is "normatively correct" according to the human normativity concept it has learned. This might involve tracking the truth of things we care about, but it might also involve casting aside unnecessary hidden variables such as "depression" and translating human values into a new ontology. That's all up to our normative concepts to specify.

## Ghost Problems

Since a lot of the aim of this post is to clarify John's post, I would be remiss if I didn't deal with the ghost problem John mentioned, which I think caused some confusion. (Feel free to skip this subsection if you don't care!) This was a running example which was mentioned several times in the text, but it was first described like this:

I don't just want to think my values are satisfied, I want them to actually be satisfied. Unfortunately, this poses a conceptual difficulty: what if I value the happiness of ghosts? I don't just want to think ghosts are happy, I want ghosts to actually be happy. What, then, should the AI do if there are no ghosts?

The first point of confusion is: it seems fine if there are no ghosts. If there aren't in fact any ghosts, we don't have to worry about them being happy or sad. So what's the problem?

I think John intended for the ghost problem to be like my depression example -- we don't just stop caring about the thing if we learn that depression isn't well-modeled as a hidden variable. In service of that, I propose the following elaboration of the ghost problem:

You, and almost everyone you know, tries to act in such a way as to make the ghosts of your ancestors happy. This idea occupies a central place in your moral framework, and plays an important role in your justification of important concepts such as honesty, diligence, and not murdering family members.

Now it's clear that there's a significant problem if you suddenly learn that ghosts don't exist, and can't be happy or sad.

Now, the second potential confusion is: this still poses no significant problem for alignment so long as your belief distribution includes how you would update if you learned ghosts weren't real. Yes, all the main points of your values as you would explain them to a stranger have to do with ghosts; but, your internal belief structures are perfectly capable of evaluating situations in which there are no ghosts. So long as the AI is good enough at inferring your values to understand this, there's no issue.

Again, I think John wanted to point to basic ontological problems, as with my example where humans can't quite comprehend the world where depression isn't ontologically fundamental. So let's further modify the problem:

If you found out ghosts didn't exist, you wouldn't know what to do with yourself. Your beliefs would become incoherent, and you wouldn't trust yourself to know what is right any more. Yes, of course, you would eventually find something to believe, and something to do with yourself; but from where you are standing now, you wouldn't trust that those reactions would be good in the important sense.

Another way to put it is that the belief in ghosts is so important that we want to invoke the value amplification problem for no-ghost scenarios: although you can be modeled as updating in a specific way upon learning there aren't any ghosts, you yourself don't trust that update, and would want to think carefully about what might constitute good and bad reasoning about a no-ghost universe. You don't want your instinctive reaction to be takes as revealing your true values; you consider it critical, in this particular case, to figure out how you would react if you were smarter and wiser.

As an example, here are three possible reactions to a no-ghost update:

• You might discard your explanations in terms of ghosts, but keep your values otherwise as intact as possible: continue to think lying is bad, continue to avoid murdering family members, etc.
• You might discard any values which you explained in terms of ghosts, keeping what little remains: you enjoy physical pleasure, beauty, (non-ceremonial) dancing, etc. You no longer attach much value to truth, life, family, or diligence.
• You might be more selective, preserving some things you previously justified via ghosts, but not others. For example, you might try to figure out what you would have come to value anyway, vs those places where the ghost ideology warped your values.

Simply put, you need philosophical help to figure out how to update.

This is the essence of the ontology mismatch problem.

Next, we consider a problem which we may encounter even if we solve value amplification, the identifiability problem, and the ontology mismatch problem: wireheading and human manipulation.

# 5. Wireheading and Human Manipulation

In my first "stable pointers to value" post, I distinguished between the easy problem of wireheading and the hard problem of wireheading:

• The Easy Problem: the wireheading problem faced by AIXI and other RL-type agents. These agents want to gain control of their own reward buttons, in order to press them continuously. This problem can be solved by evaluating the expected utility of plans/policies inside the planning/policy optimization loop, rather than only giving feedback on actions actually taken (leaving the plan/policy optimization loop to evaluate based on expected reward). Daniel Dewey referred to this architecture as observation-utility maximization.
• The Hard Problem: even if we successfully point to human values as a thing to learn (solving the identifiability problem and the ontology mismatch problem, amongst other things), human values are manipulable. This introduces perverse incentives for a value-learning systems, if those systems must make decisions while operating under some remaining uncertainty about human values. Specifically, such systems are incentivised to manipulate human values.

Simply put, the easy wireheading problem is that the AI might wirehead itself. The hard wireheading problem is that the AI might wirehead us, by manipulating our values to be easier to satisfy.

This is related to the ontology mismatch problem in that poor solutions to that problem might especially incentivize a system to wirehead or otherwise manipulate humans, but is mostly an independent problem.

I see this as the essence of the compressed pointer problem as I described it in Embedded Agency (ie, my references to "pointers" in the green and blue sections). However, that was a bit vague and hand-wavy.

## Helping Humancurrent

One proposal to side-step this problem is: always and only help the current human, as of this very moment (or perhaps a human very slightly in the past, EG the one whose light-waves are just now hitting you). This ensures that no manipulation is possible, making the point moot. The system would only ever be manipulative in the moment if that best served our values in the moment. (Or, if the system wrongly inferred such.)

This solution can clearly be dynamically inconsistent, as the AI's goal pointer keeps changing. However, it is only as dynamically inconsistent as the human. In some sense, this seems like the best we can do: you cannot be fully aligned with an agent who is not fully aligned with themselves. This solution represents one particular compromize we can make.

Another compromise we can make is to be fully aligned with the human at the moment of activating the AI (or again, the moment just before, to ensure causal separation with a margin for error). This is more reflectively stable, but will be less corrigible: to the extent humans are not aligned with our future selves, the AI might manipulate our future selves or prevent them from shutting it down, in service of earlier-self values. (Although, if the system is working correctly, this will only happen if your earlier selves would endorse it.)

Another variation is that you myopically train the system to do what HCH would tell you to do at each point. (I'll keep it a bit vague -- I'm not sure exactly what question you would want to prompt HCH with.) This type of approach subsumes not only human values, but the problem of planning to accomplish them, into the value amplification problem. This is pretty different, but I'm lumping it in the same basket for theoretical reasons which will hopefully become clear soon.

## Time Travel Problems

The idea of helping the current human, as a way to avoid manipulation, falls apart as soon as we admit the possibility of time travel. If the agent can find a way to use time travel to manipulate the human, we get the very same problem we had in the first place. Since this might be quite valuable for the agent, it might put considerable resources toward the project. (IE, there is not necessarily a "basin of corrigibility" here, where we've specified an agent that is aligned enough that it'll naturally try and correct flaws in its specification -- there might be mechanisms which could accomplish this, but it's not going to happen just from the basic idea of forwarding the values of the current human.)

Similarly, in my old post on this topic I describe an AI system whose value function is specified by CEV, but which finds out that there is an exact copy of itself embedded therein:

As a concrete example, suppose that we have constructed an AI which maximizes CEV: it wants to do what an imaginary version of human society, deliberating under ideal conditions, would decide is best. Obviously, the AI cannot actually simulate such an ideal society. Instead, the AI does its best to reason about what such an ideal society would do.

Now, suppose the agent figures out that there would be an exact copy of itself inside the ideal society. Perhaps the ideal society figures out that it has been constructed as a thought experiment to make decisions about the real world, so they construct a simulation of the real world in order to better understand what they will be making decisions about. Furthermore, suppose for the sake of argument that our AI can break out of the simulation and exert arbitrary control over the ideal society's decisions.

Naively, it seems like what the AI will do in this situation is take control over the ideal society's deliberation, and make the CEV values as easy to satisfy as possible -- just like an RL agent modifying its utility module.

Or, with respect to my example of an agent trained to do whatever HCH would tell it to do, we might imagine what would happen if HCH reasons about the agent in sufficient detail that its decisions influence HCH. Then the agent might learn to manipulate HCH, to the degree that such a thing is possible.

My point is not that we should necessarily worry about these failure modes. If it comes down to it, we might be willing to assume time travel isn't possible as one of the few assumptions we need in order to argue that a system is safe -- or posit extra mechanisms to keep our AI from existing inside of CEV, or keep it from being simulated within HCH, or allow it to exist but ensure it can't manipulate those things, or what-have-you.

What interests me is the basic problem which applies to all of these cases. We could call in the "very hard wireheading problem":

• The Very Hard Problem of Wireheading: like the Hard Problem, but under the assumption that there's absolutely no way to rule out manipulation, IE, we assume manipulation is going to be possible no matter what we do.

There's probably no perfect solution to this problem. Yet, when I sit staring at the problem, I have a strong desire to "square the circle" -- like there should be something it means to be non-manipulatively aligned (with something that you could manipulate), something theoretically elegant and not a hack.

Probably the best thing I've seen is a proposal which I think originates from Stuart Armstrong, which is that you simply remove the manipulative causal pathways from your model before making decisions. I'm not sure how you are supposed to identify which pathways are manipulative vs non-manipulative, in order to remove them, but if you can, you get a notion of optimizing without manipulating.

This makes things very dependent on your prior -- to give an extreme case, suppose humans are perfectly manipulable; we take on whatever values the AI suggests. Then when the AI freezes the causal pathways of manipulation, its "human values" it attempts to cooperate will be a mixture of all the things it might tell humans to value, each according to its prior probability.

I had some other objections, too, EG if we remove those causal pathways, our model could get pretty weird, assigning high probability to outcomes which are improbable or even impossible. For example, suppose the AI is asked not to manipulate Sally (its creator, who it is aligned with), but in fact, Ben and Sally are equally prone to manipulation, and in the same room, so hear the same messages from the AI. The AI proceeds as if Sally is immune to manipulation (when she's not). This might involve planning for Sally and Ben to have different reactions to an utterance (when in fact they have exactly the same reaction). So the AI might make plans which end up making no sense in the real world.

I could say more about trying to solve the Very Hard Problem, but I suspect I've already written too much rather than too little.

# Conclusion

If all of the above are sub-problems of the pointers problem, what is the pointers problem itself? Arguably, it's the whole outer alignment problem. I don't want to view it quite that way, though. I think it is more like a particular view on the outer alignment problem, with an emphasis on "pointing": the part of the problem that's about "What do we even mean by alignment? How can we robustly point an AI at external things which are ontologically questionable? How do we give feedback about what we mean, without incentivizing a system to wirehead or manipulate us? How can we optimize things which live in the human ontology, without incentivizing a system to delude us?"

There's a small element of inner alignment to this, as well. Although an RL agent such as AIXI will want to wirehead if it forms an "accurate" model of how it gets reward, we can also see this as only one model consistent with the data, another being that reward is actually coming from task achievement (IE, the AI could internalize the intended values). Although this model will usually have at least slightly worse predictive accuracy, we can counterbalance that with process-level feedback which tells the system that's a better way of thinking about it. (Alternatively, with strong priors which favor the right sorts of interpretations.) This is inner alignment in the sense of getting the system to think in the right way rather than act in the right way, to avoid a later treacherous turn. (However, not in the sense of avoiding inner optimizers.)

Similarly, human manipulation could be avoided not by solving the incentive problem, but rather, by giving feedback to the effect that manipulation rests on an incorrect interpretation of the goal. Similar feedback-about-how-to-think-about-things could address the ontology mismatch problem and the value amplification problem.

In order to use this sort of solution, the AI system needs to think of everything as a proxy; no feedback is taken as a gold standard for values. This is similar to Eliezer's approach to wireheading, Goodhart, and manipulation in CFAI (see especially section 5, especially 5.5), although I don't think that document contains enough to make the idea really work.

# 28

Mentioned in
New Comment

As an example, here are three possible reactions to a no-ghost update:

Suppose that many (EDIT: a few) of your value shards take as input the ghost latent variable in your world model. You learn ghosts aren't real. Let's say this basically sets the ghost-related latent variable value to false in all shard-relevant contexts. Then it seems perfectly fine that most of my shards keep on bidding away and determining my actions (e.g. protect my family), since most of my value shards are not in fact functions of the ghost latent variable. While it's indeed possible to contrive minds where most of their values are functions of a variable in the world model which will get removed by the learning process, it doesn't seem particularly concerning to me. (But I'm also probably not trying to tackle the problems in this post, or the superproblems which spawned them.)

There's a small element of inner alignment to this, as well. Although an RL agent such as AIXI will want to wirehead if it forms an "accurate" model of how it gets reward, we can also see this as only one model consistent with the data, another being that reward is actually coming from task achievement (IE, the AI could internalize the intended values). Although this model will usually have at least slightly worse predictive accuracy, we can counterbalance that with process-level feedback which tells the system that's a better way of thinking about it.

This doesn't seem relevant for non-AIXI RL agents which don't end up caring about reward or explicitly weighing hypotheses over reward as part of the motivational structure? Did you intend it to be?

This doesn't seem relevant for non-AIXI RL agents which don't end up caring about reward or explicitly weighing hypotheses over reward as part of the motivational structure? Did you intend it to be?

With almost any kind of feedback process (IE: any concrete proposals that I know of), similar concerns arise. As I argue here, wireheading is one example of a very general failure mode. The failure mode is roughly: the process actually generating feedback is, too literally, identified with the truth/value which that feedback is trying to teach.

Output-based evaluation (including supervised learning, and the most popular forms of unsupervised learning, and a lot of other stuff which treats models as black boxes implementing some input/output behavior or probability distribution or similar) can't distinguish between a model which is internalizing the desired concepts, vs a model which is instead modeling the actual feedback process instead. These two do different things, but not in a way that the feedback system can differentiate.

In terms of shard theory, as I understand it, the point is that (absent arguments to the contrary, which is what we want to be able to construct), shards that implement feedback-modeling like this cannot be disincentivized by the feedback process, since they perform very well in those terms. Shards which do other things may or may not be disincentivized, but the feedback-modeling shards (if any are formed at any point) definitely won't, unless of course they're just not very good at their jobs.

So the problem, then, is: how do we arrange training such that those shards have very little influence, in the end? How do we disincentivize that kind of reasoning at all?

Plausibly, this should only be tackled as a knock-on effect of the real problem, actually giving good feedback which points in the right direction; however, it remains a powerful counterexample class which challenges many many proposals. (And therefore, trying to generate the analogue of the wireheading problem for a given proposal seems like a good sanity check.)

Great post!

I especially like "try to maximize values according to models which, according to human beliefs, track the things we care about well". I ended up at a similar point when thinking about the problem. It seems like we ultimately have to use this approach, at some level, in order for all the type signatures to line up. (Though this doesn't rule out entirely different approaches at other levels, as long as we expect those approaches to track the things we care about well.)

On amplified values, I think there's a significant piece absent from the discussion here (possibly intentionally). It's not just about precision of values, it's about evaluating the value function at all.

Model/example: a Bayesian utility maximizer does not need to be able to evaluate its utility function, it only needs to be able to decide which of two options has higher utility. If e.g. the utility function is , and a decision only effects , then the agent doesn't need to evaluate the sum at all; it only needs to calculate  for each option. This is especially relevant in a world where most actions don't effect most of the world (or if they do, the effects are drowned out by noise) - which is exactly the sort of world we live in. Most of my actions do not effect a random person in Mumbai (and to the extent there is an effect, it's drowned out by noise). Even if I value the happiness of that random person in Mumbai, I never need to think about them, because my actions don't significantly impact them in any way I can predict.

As you say, the issue isn't just "we can't evaluate our values precisely". The issue is that we probably do not and cannot evaluate our values at all. We only ever evaluate comparisons, and only between actions with a relatively simple diff.

Applying this to amplification: amplification is not about evaluating our values more precisely, it's about comparing actions with more complicated diffs, or actions where more complicated information is relevant to the diff. The things you say in the post are still basically correct, but this gives a more accurate mental picture of what amplification needs to achieve.

Really great post! I think I already got John's idea from his post, but putting everything in perspective and reference previous and adjacent works really help!

On that note, you have been mentioning Stuart's no-indescribable-hellworlds hypothesis for a few posts now, and I'm really interested in it. To take the even more meta-argument for its importance, it looks particularly relevant to asking whether the human epistemic perspective is the right one to use when defining ascription universality (which basically abstracts most of Paul's and Paul-related approaches in term of desiderata for a supervisor).

Do you know if there have been working on poking at this hypothesis, and trying to understand what it implies/requires? I doubt we can ever prove it, but there might be a way to do the "computational complexity approach", where we formally relate it to much more studied and plausible hypotheses.

After thinking a bit more about it, the no-indescribable-hellworlds seems somewhat related to logical uncertainty and the logical induction criterion. Because intuitively, indescribability comes from complexity issues about reasoning, that is the lack of logical omniscience about the consequences of our values. The sort of description we would like is a polynomial proof, or at least a polynomial interactive protocol for verifying the indescribability (which means being in PSPACE, as in the original take on debate). And satisfying the logical induction criterion seems a good way to ensure that such a proof will eventually be found, because otherwise we could be exploited forever on our wrong assessment of the hellworld.

The obvious issue with this approach comes from the asymptotic nature of logical induction guarantees, which might mean it takes so long to convince us/check indescribable hellworlds that they already came to pass.

I don't have much to say other than that I agree with the connection. Honestly, thinking of it in those terms makes me pessimistic that it's true -- it seems quite possible that humans, given enough time for philosophical reflection, could point to important value-laden features of worlds/plans which are not PSPACE.

Glad that you find the connection interesting. That being said, I'm confused by what you're saying afterwards: why would logical inductors not able to find propositions about worlds/plans which are outside PSPACE? I find no mention of PSPACE in the paper.

Oh, well, satisfying the logical induction criterion is stronger than just PSPACE. I see debate, and iterated amplification, as attempts to get away with less than full logical induction. See https://www.lesswrong.com/posts/R3HAvMGFNJGXstckQ/relating-hch-and-logical-induction, especially Paul's comment https://www.lesswrong.com/posts/R3HAvMGFNJGXstckQ/relating-hch-and-logical-induction?commentId=oNPtnwTYcn8GixC59

The basic idea behind compressed pointers is that you can have the abstract goal of cooperating with humans, without actually knowing very much about humans. In a sense, this means having aligned goals without having the same goals: your goal is to cooperate with "human goals", but you don't yet have a full description of what human goals are. Your value function might be much simpler than the human value function.

In machine-learning terms, this is the question of how to specify a loss function for the purpose of learning human values.

Insofar as I understand your point, I disagree. In machine-learning terms, this is the question of how to train an AI whose internal cognition reliably unfolds into caring about people, in whatever form that takes in the AI's learned ontology (whether or not it has a concept for people). If you commit to the specific view of outer/inner alignment, then now you also want your loss function to "represent" that goal in some way.

humans seem to correctly identify what each other want and believe, quite frequently. Therefore, humans must have prior knowledge which helps in this task. If we can encode those prior assumptions in an AI, we can point it in the right direction.

I doubt this due to learning from scratch. I think the question of "how do I identify what you want, in terms of a utility function?" is a bit sideways due to people not in fact having utility functions.[1] Insofar as the question makes sense, its answer probably takes the form of inductive biases: I might learn to predict the world via self-supervised learning and form concepts around other people having values and emotional states due to that being a simple convergent abstraction relatively pinned down by my training process, architecture, and data over my life, also reusing my self-modelling abstractions. It would be quite unnatural to model myself in one way (as valuing happiness) and others as having "irrational" shards which "value" anti-happiness but still end up behaving as if they value happiness. (That's not a sensible thing to say, on my ontology.)

This presents a difficulty if another agent wishes to help such an agent, but does not share its ontological commitments.

I think it's worth considering how I might go about helping a person from an uncontacted tribe who doesn't share my ontology. Conditional on them requesting help from me somehow, and my wanting to help them, and my deciding to do so—how would I carry out that process, internally?

(Not reading the rest at the moment, may leave more comments later)

1. ^

On my view: Human values take the form of decision-influences (i.e. shards) which increase or decrease the probability of mental events and decisions (buying ice cream, thinking for another minute). There is no such thing as an anti-ice-cream shard which is perfectly anti-rational in that it bids "against its interests", bidding for ice cream and against avoiding ice cream. That's just an ice cream shard. Goals and rationality are not entirely separate, in people.

I said:

The basic idea behind compressed pointers is that you can have the abstract goal of cooperating with humans, without actually knowing very much about humans.
[...]
In machine-learning terms, this is the question of how to specify a loss function for the purpose of learning human values.

You said:

In machine-learning terms, this is the question of how to train an AI whose internal cognition reliably unfolds into caring about people, in whatever form that takes in the AI's learned ontology (whether or not it has a concept for people).

Thinking about this now, I think maybe it's a question of precautions, and what order you want to teach things in. Very similarly to the argument that you might want to make a system corrigible first, before ensuring that it has other good properties -- because if you make a mistake, later, a corrigible system will let you correct the mistake.

Similarly, it seems like a sensible early goal could be 'get the system to understand that the sort of thing it is trying to do, in (value) learning, is to pick up human values'. Because once it has understood this point correctly, it is harder for things to go wrong later on, and the system may even be able to do much of the heavy lifting for you.

Really, what makes me go to the meta-level like this is pessimism about the more direct approach. Directly trying to instill human values, rather than first training in a meta-level understanding of that task, doesn't seem like a very correctible approach. (I think much of this pessimism comes from mentally visualizing humans arguing about what object-level values to try to teach an AI. Even if the humans are able to agree, I do not feel especially optimistic about their choices, even if they're supposedly informed by neuroscience and not just moral philosophers.)

Really, what makes me go to the meta-level like this is pessimism about the more direct approach. Directly trying to instill human values, rather than first training in a meta-level understanding of that task, doesn't seem like a very correctible approach.

True, but I'm also uncertain about the relative difficulty of relatively novel and exotic value-spreads like "I value doing the right thing by humans, where I'm uncertain about the referent of humans", compared to "People should have lots of resources and be able to spend them freely and wisely in pursuit of their own purposes" (the latter being values that at least I do in fact have).

If you commit to the specific view of outer/inner alignment, then now you also want your loss function to "represent" that goal in some way.

I think it is reasonable as engineering practice to try and make a fully classically-Bayesian model of what we think we know about the necessary inductive biases -- or, perhaps more realistically, a model which only violates classic Bayesian definitions where necessary in order to represent what we want to represent.

This is because writing down the desired inductive biases as an explicit prior can help us to understand what's going on better.

It's tempting to say that to understand how the brain learns, is to understand how it treats feedback as evidence, and updates on that evidence. Of course, there could certainly be other theoretical frames which are more productive. But at a deep level, if the learning works, the learning works because the feedback is evidence about the thing we want to learn, and the process which updates on that feedback embodies (something like) a good prior telling us how to update on that evidence.

And if that framing is wrong somehow, it seems intuitive to me that the problem should be describable within that ontology, like how I think "utility function" is not a very good way to think about values because what is it a function of; we don't have a commitment to a specific low-level description of the universe which is appropriate for the input to a utility function. We can easily move beyond this by considering expected values as the "values/preferences" representation, without worrying about what underlying utility function generates those expected values.

(I do not take the above to be a knockdown argument against "committing to the specific division between outer and inner alignment steers you wrong" -- I'm just saying things that seem true to me and plausibly relevant to the debate.)

I doubt this due to learning from scratch.

I expect you'll say I'm missing something, but to me, this sounds like a language dispute. My understanding of your recent thinking holds that the important goal is to understand how human learning reliably results in human values. The Bayesian perspective on this is "figuring out the human prior", because a prior is just a way-to-learn. You might object to the overly Bayesian framing of that; but I'm fine with that. I am not dogmatic on orthodox bayesianism. I do not even like utility functions.

Insofar as the question makes sense, its answer probably takes the form of inductive biases: I might learn to predict the world via self-supervised learning and form concepts around other people having values and emotional states due to that being a simple convergent abstraction relatively pinned down by my training process, architecture, and data over my life, also reusing my self-modelling abstractions.

I am totally fine with saying "inductive biases" instead of "prior"; I think it indeed pins down what I meant in a more accurate way (by virtue of, in itself, being a more vague and imprecise concept than "prior").

I expect you'll say I'm missing something, but to me, this sounds like a language dispute.

I agree, this does seem like it was a language dispute, I no longer perceive us as disagreeing on this point.