Alignment By Default

I think what you've identified here is a weakness in the high-level, classic arguments for AI risk -

Overall, I’d give maybe a 10-20% chance of alignment by this path, assuming that the unsupervised system does end up with a simple embedding of human values. The main failure mode I’d expect, assuming we get the chance to iterate, is deception - not necessarily “intentional” deception, just the system being optimized to look like it’s working the way we want rather than actually working the way we want. It’s the proxy problem again, but this time at the level of humans-trying-things-and-seeing-if-they-work, rather than explicit training objectives.

This failure mode of deceptive alignment seems like it would result most easily from Mesa-optimisation or an inner alignment failure [LW · GW]. Inner Alignment / Misalignment is possibly the key specific mechanism which fills a weakness in the 'classic arguments [LW(p) · GW(p)]' for AI safety - the Orthogonality Thesis, Instrumental Convergence and Fast Progress together implying small separations between AI alignment and AI capability can lead to catastrophic outcomes. The question of why there would be such a damaging, hard-to-detect divergence between goals and alignment needs an answer to have a solid, specific reason to expect dangerous misalignment, and Inner Misalignment is just such a reason.

I think that it should be presented in initial introductions to AI risk alongside those classic arguments, as the specific, technical reason why the specific techniques we use are likely to produce such goal/capability divergence - rather than the general a priori reasons given by the classic arguments.

I’ll set aside what happens “by default” and focus on the interesting technical question of whether this post is describing a possible straightforward-ish path to aligned superintelligent AGI.

The background idea is “natural abstractions”. This is basically a claim that, when you use an unsupervised world-model-building learning algorithm, its latent space tends to systematically learn some patterns rather than others. Different learning algorithms will converge on similar learned patterns, because those learned patterns are a property of the world, not an idiosyncrasy of the learning algorithm. For example: Both human brains and ConvNets seem to have a “tree” abstraction; neither human brains nor ConvNets seem to have a “head or thumb but not any other body part” concept.

I kind of agree with this. I would say that the patterns are a joint property of the world and an inductive bias. I think the relevant inductive biases in this case are something like: (1) “patterns tend to recur”, (2) “patterns tend to be localized in space and time”, and (3) “patterns are frequently composed of multiple other patterns, which are near to each other in space and/or time”, and maybe other things. The human brain definitely is wired up to find patterns with those properties, and ConvNets to a lesser extent. These inductive biases are evidently very useful, and I find it very likely that future learning algorithms will share those biases, even more than today’s learning algorithms. So I’m basically on board with the idea that there may be plenty of overlap between the world-models of various different unsupervised world-model-building learning algorithms, one of which is the brain.

(I would also add that I would expect “natural abstractions” to be a matter of degree, not binary. We can, after all, form the concept “head or thumb but not any other body part”. It would just be extremely low on the list of things that would pop into our head when trying to make sense of something we’re looking at. Whereas a “prominent” concept like “tree” would pop into our head immediately, if it were compatible with the data. I think I can imagine a continuum of concepts spanning the two. I’m not sure if John would agree.)

Next, John suggests that “human values” may be such a “natural abstraction”, such that “human values” may wind up a “prominent” member of an AI's latent space, so to speak. Then when the algorithms get a few labeled examples of things that are or aren’t “human values”, they will pattern-match them to the existing “human values” concept. By the same token, let’s say you’re with someone who doesn’t speak your language, but they call for your attention and point to two power outlets in succession. You can bet that they’re trying to bring your attention to the prominent / natural concept of “power outlets”, not the un-prominent / unnatural concept of “places that one should avoid touching with a screwdriver”.

Do I agree? Well, “human values” is a tricky term. Maybe I would split it up. One thing is “Human values as defined and understood by an ideal philosopher after The Long Reflection”. This is evidently not much of a “natural abstraction”, at least in the sense that, if I saw ten examples of that thing, I wouldn’t even know it. I just have no idea what that thing is, concretely.

Another thing is “Human values as people use the term”. In this case, we don’t even need the natural abstraction hypothesis! We can just ensure that the unsupervised world-modeler incorporates human language data in its model. Then it would have seen people use the phrase “human values”, and built corresponding concepts. And moreover, we don’t even necessarily need to go hunting around in the world-model to find that concept, or to give labeled examples. We can just utter the words “human values”, and see what neurons light up! I mean, sure, it probably wouldn’t work! But the labeled examples thing probably wouldn’t work either!

Unfortunately, “Human values as people use the term” is a horrific mess of contradictory and incoherent things. An AI that maximizes “‘human values’ as those words are used in the average YouTube video” does not sound to me like an AI that I want to live with. I would expect lots of performative displays of virtue and in-group signaling, little or no making-the-world-a-better-place.

In any case, it seems to me that the big kernel of truth in this post is that we can and should think of future AGI motivations systems as intimately involving abstract concepts, and that in particular we can and should take advantage of safety-advancing abstract concepts like “I am advancing human flourishing”, “I am trying to do what my programmer wants me to try to do”, “I am following human norms” [AF · GW], or whatever. In fact I have a post advocating that just a few days ago [LW · GW], and think of that kind of thing as a central ingredient in all the AGI safety stories that I find most plausible.

Beyond that kernel of truth, I think a lot more work, beyond what’s written in the post, would be needed to build a good system that actually does something we want. In particular, I think we have much more work to do on choosing and pointing to the right concepts (cf. “first-person problem” [LW · GW]), detecting when concepts break down because we’re out of distribution (cf. “model splintering” [LW · GW]), sandbox testing protocols, and so on. The post says 10% chance that things work out, which seems much too high to me. But more importantly, if things work out along these lines, I think it would be because people figured out all those things I mentioned, by trial-and-error, during slow takeoff. Well in that case, I say: let's just figure those things out right now!

To help me check my understanding of what you're saying, we train an AI on a bunch of videos/media about Alice's life, in the hope that it learns an internal concept of "Alice's values". Then we use SL/RL to train the AI, e.g., give it a positive reward whenever it does something that the supervisor thinks benefits Alice's values. The hope here is that the AI learns to optimize the world according to its internal concept of "Alice's values" that it learned in the previous step. And we hope that its concept of "Alice's values" includes the idea that Alice wants AIs, including any future AIs, to keep improving their understanding of Alice's values and to serve those values, and that this solves alignment in the long run.

Assuming the above is basically correct, this (in part) depends on the AI learning a good enough understanding of "improving understanding of Alice's values" in step 1. This in turn (assuming "improving understanding of Alice's values" involves "using philosophical reasoning to solve various confusions related to understanding Alice's values, including Alice's own confusions") depends on that the AI can learn a correct or good enough concept of "philosophical reasoning" from unsupervised training. Correct?

If AI can learn "philosophical reasoning" from unsupervised training, GPT-N should be able to do philosophy (e.g., solve open philosophical problems), right?

In light of the previous section, there’s an obvious path to alignment where there turns out to be a few neurons (or at least some simple embedding) which correspond to human values, we use the tools of microscope AI to find that embedding, and just like that the alignment problem is basically solved.

This is the part I disagree with. The network does recognise trees, or at least green things (given that the grass seems pretty brown in the low tree pic).

Extrapolating this, I expect the AI might well have neurons that correspond roughly to human values, on the training data. Within the training environment, human values, amount of dopamine in human brain, curvature of human lips (in smiles), number of times the reward button is pressed, and maybe even amount of money in human bank account might all be strongly correlated.

You will have successfully narrowed human values down to within the range of things that are strongly correlated with human values in the training environment. If you take this signal and apply enough optimization pressure, you are going to get the equivalent of a universe tiled with tiny smiley faces.

Supervised/Reinforcement: Proxy Problems

Another plausible approach to deal with the proxy problems might be to do something like unsupervised clustering/learning on the representations of multiple systems that we'd have good reasons to believe encode the (same) relevant values - e.g. when exposed to the same stimulus (including potentially multiple humans and multiple AIs). E.g. for some relevant recent proof-of-concept works: Identifying Shared Decodable Concepts in the Human Brain Using Image-Language Foundation Models, Finding Shared Decodable Concepts and their Negations in the Brain, AlignedCut: Visual Concepts Discovery on Brain-Guided Universal Feature Space, Rosetta Neurons: Mining the Common Units in a Model Zoo, Cross-GAN Auditing: Unsupervised Identification of Attribute Level Similarities and Differences between Pretrained Generative Models, Quantifying stimulus-relevant representational drift using cross-modality contrastive learning. Automated interpretability (e.g. https://multimodal-interpretability.csail.mit.edu/maia/) could also be useful here. This might also work well with concepts like corrigibility/instruction following [LW · GW] and arguments about the 'broad basin of attraction' and convergence [LW · GW] for corrigibility.

Planned summary for the Alignment Newsletter:

I liked the author’s summary, so I’ve reproduced it with minor stylistic changes:

A low-level model of some humans has everything there is to know about human values embedded within it, in exactly the same way that human values are embedded in physical humans. The embedding, however, is nontrivial. Thus, predictive power alone is not sufficient to define human values. The missing part is the embedding of values within the model.

However, this also applies if we replace the phrase “human values” with “trees”. Yet we have a whole class of neural networks in which a simple embedding lights up in response to trees. This is because trees are a natural abstraction, and we should expect to see real systems trained for predictive power use natural abstractions internally.

Human values are a little different from trees: they’re a property of an abstract object (humans) rather than an abstract object themselves. Nonetheless, the author still expects that a broad class of systems trained for predictive power will end up with simple embeddings of human values (~70% chance).

Since an unsupervised learner has a simple embedding of human values, a supervised/reinforcement learner can easily score well on values-proxy-tasks by directly using that model of human values. In other words, the system uses an actual model of human values as a proxy for our proxy of human values (~10-20% chance). This is what is meant by _alignment by default_.

When this works, it’s basically a best-case scenario, so we can safely use the system to design a successor without worrying about amplification of alignment errors (among other things).

Planned opinion:

I broadly agree with the perspective in this post: in particular, I think we really should have more optimism because of the tendency of neural nets to learn “natural abstractions”. There is structure and regularity in the world and neural nets often capture it (despite being able to memorize random noise); if we train neural nets on a bunch of human-relevant data it really should learn a lot about humans, including what we care about.

However, I am less optimistic than the author about the specific path presented here (and he only assigns 10% chance to it). In particular, while I do think human values are a “real” thing that a neural net will pick up on, I don’t think that they are well-defined enough to align an AI system arbitrarily far into the future: our values do not say what to do in all possible situations; to see this we need only to look at the vast disagreements among moral philosophers (who often focus on esoteric situations). If an AI system were to internalize and optimize our current system of values, as the world changed the AI system would probably become less and less aligned with humans. We could instead talk about an AI system that has internalized both current human values and the process by which they are constructed, but that feels much less like a natural abstraction to me.

I _am_ optimistic about a very similar path, in which instead of training the system to pursue (a proxy for) human values, we train the system to pursue some “meta” specification like “be helpful to the user / humanity” or “do what we want on reflection”. It seems to me that “being helpful” is also a natural abstraction, and it seems more likely that an AI system pursuing this specification would continue to be beneficial as the world (and human values) changed drastically.

Some notes on the loss function in unsupervised learning:

Since an unsupervised learner is generally just optimized for predictive power

I think it's worthwhile to distinguish the loss function that's being optimized during unsupervised learning, vs what the practitioner is optimizing for. Yes, the loss function being optimized in an unsupervised learning system is frequently minimization of reconstruction error or similar. But when I search for "unsupervised learning review" on Google Scholar, I find this highly cited paper by Bengio et al. The abstract talks a lot about learning useful representations and says nothing about predictive power. In other words, learning "natural abstractions" appears to be pretty much the entire game from a practitioner perspective.

And in the same way supervised learning has dials such as regularization which let us control the complexity of our model, unsupervised learning has similar dials.

For clustering, we could achieve 0 reconstruction error (or equivalently, explain all the variation in the data) by putting every data point in its own cluster, but that would completely defeat the point. The elbow method is a well-known heuristic for figuring out what the "right" number of clusters in a dataset is.

Similarly, we could achieve 0 reconstruction error with an autoencoder by making the number of dimensions in the bottleneck be equal to the number of dimensions in the original input, but again, that would completely defeat the point. Someone on the Stats Stackexchange says that there is no standard way to select the number of dimensions for an autoencoder. (For reference, the standard way to select the regularization parameter which controls complexity in supervised learning would obviously be through cross-validation.) However, I suspect this is a tractable research problem.

It was interesting that you mentioned the noise of air molecules, because one unsupervised learning trick is to deliberately introduce noise into the input to see if the system has learned "natural" representations which allow it to reconstruct the original noise-free input. See denoising autoencoder. This is the kind of technique which might allow an autoencoder to learn natural representations even if the number of dimensions in the bottleneck is equal to the number of dimensions in the original input.

BTW, here's an interesting-looking (pessimistic) paper I found while researching this comment: Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations

You brought up microscope AI. I think a promising research direction here may be to formulate a notion of "ease of interpretability" which can be added as an additional term to an unsupervised loss function (the same way we might, for example, add a term to a clustering algorithm's loss function so that in addition to minimizing reconstruction error, it also seeks to minimize the number of clusters).

Hardcoding "human values" by hand is hopeless, but hardcoding "ease of human interpretability" by hand seems much more promising, since ease of human interpretability is likely to correspond to easily formalizable notions such as simplicity. Also, if your hardcoded notion of "ease of human interpretability" turns out to be slightly wrong, that's not a catastrophe: you just get an ML model which is a bit harder to interpret than you might like.

Another option is to learn a notion of what constitutes an interpretable model by e.g. collecting "ease of interpretability" data from human microscope users.

Of course, one needs to be careful that any interpretability term does not get too much weight in the loss function, because if it does, we may stop learning the "natural" abstractions that we desire (assuming a worst-case scenario where human interpretability is anticorrelated with "naturalness"). The best approach may be to learn two models, one of which was optimized for interpretability and one of which wasn't, and only allow our system to take action when the two models agree. I guess mesa-optimizers in the non-interpretable model are still a worry though.

This came out of the discussion you had with John Maxwell, right? Does he think this is a good presentation of his proposal?

How do we know that the unsupervised learner won't have learnt a large number of other embeddings closer to the proxy? If it has, then why should we expect human values to do well?

Some rough thoughts on the data type issue. Depending on what types the unsupervised learner provides the supervised, it may not be able to reach the proxy type by virtue of issues with NN learning processes.

Recall that tata types can be viewed as homotopic spaces, and construction of types can be viewed as generating new spaces off the old e.g. tangent spaces or path spaces etc. We can view neural nets as a type corresponding to a particular homotopic space. But getting neural nets to learn certain functions is hard. For example, learning a function which is 0 except in two sub spaces A and B. It has different values on A and B. But A and B are shaped like intelocked rings. In other words, a non-linear classification problem. So plausibly, neural nets have trouble constructing certain types from others. Maybe this depends on architecture or learning algorithm, maybe not.

If the proxy and human values have very different types, it may be the case that the supervised learner won't be able to get from one type to another. Supposing the unsupervised learner presents it with types "reachable" from human values, then the proxy which optimises performance on the data set is just unavailable to the system even though its relatively simple in comparison.

Because of this, checking which simple homotopies neural nets can move between would be useful. Depending on the results, we could use this as an arguement that unsupervised NNs will never embed the human values type because we've found out it has some simple properties it won't be able to construct de novo. Unless we do something like feed the unsupervised learner human biases/start with an EM and modify it.

I guess the main issue that I have with this argument is that an AI system that is extremely good at prediction is unlikely to just have a high-level concept corresponding to human values (if it does contain such a concept). Instead, it's likely to also include a high-level concept corresponding to what people say about about values - or rather several corresponding to what various different groups would say about human-values. If your proxy is based on what people say, then these concepts which correspond to what people say will match much better - and the probability of at least one of these concepts being the best match is increased by large the number of these. So I don't put a very high weight on this scenario at all.

Thanks a lot for writing this. I've been thinking about FAI plans along these lines for a while now, here are some thoughts on specific points you made.

First, I take issue with the "Alignment By Default" title. There are two separate questions here. Question #1 is whether we'd have a good outcome if everyone concerned with AI safety got hit by a bus. Question #2 is whether there's a way to create Friendly AI using unsupervised learning. I'm rather optimistic that the answer to Question #2 is yes. I find the unsupervised learning family of approaches more appealing than IRL or HCH (from what I understand of those approaches). But I still think there are various ways in which things could go wrong, some of which you mention in this post, and it's useful to have safety researchers thinking about this, because the problems seem pretty tractable to me. You, me, and Steve Byrnes [LW · GW] are the only people in the community I remember off the top of my head who seem to be giving this serious thought, which is a little odd because so many top AI people seem to think that unsupervised learning is The Nut That Must Be Cracked if we are to build AGI.

Anyway, in order to illustrate that the problems seem tractable, here are a couple things you brought up + thoughts on solving them.

With regard to the high-resolution molecular model of a human, there's the possibility of using this model as an upload somehow even if the embedding of human values is nontrivial. I guess the challenge is to excise everything around the human from the model, and replace those surroundings with whatever an ideal environment for doing moral / philosophical reasoning would be, along with some communication channel to the outside world. This is approach is similar to the Paul Christiano construction described on p. 198 of Superintellligence. In this case, I guess it is more important for the embedding of a person's physical surroundings to be "natural" enough that we can mess with it without messing with the person's mind. However, even if the embedding of the person's physical surroundings is kinda bad (meaning that our "ideal environment for doing moral / philosophical reasoning" ends up being like a glitchy VR sim in practice), this plausibly won't cause a catastrophic alignment failure. Also, you don't necessarily need a super high-resolution model to do this sort of thing (imagine prompting GPT-N with "Gandhi goes up the mountain to contemplate Moral Question X, he returns after a year of contemplation and proclaims...").

This requires hitting a window - our data needs to be good enough that the system can tell it should use human values as a proxy, but bad enough that the system can’t figure out the specifics of the data-collection process enough to model it directly. This window may not even exist.

A couple thoughts.

First, I think it's possible to create this window. Suppose we restrict ourselves to feeding our system data from before the year 2000. There should be a decent representation of human values to be learned from this data, yet it should be quite difficult to figure out the specifics of the 2020+ data-collection process from it. Identifying the specific quirks which cause the data-collection process to differ from human values seems especially difficult. (I think restricting ourselves to pre-2000 data is overkill, I just chose 2000 for the purpose of illustration.)

Second, one way to check on things is to deliberately include a small quantity of mislabeled data, then once the system is done learning, check whether its model correctly recognizes that the mislabeled data is mislabeled (and agrees with all data that is correctly labeled). (This should be combined with the idea above where we disguise the data-collection process from the AI, because otherwise we might pinpoint "the data-collection process prior to the time at which the mislabeled data was introduced"?)

I know of two main ways to go from aligning one AI to long-term alignment

A third approach which you don't mention is to use the initial aligned AI as a "human values oracle" for subsequent AIs. Once you have a cheap, fast computational representation of human values, you can replicate it across a massive compute cluster and

• Use it to generate extremely large quantities of training data

• Use it as the "moral compass" for some bigger, more sophisticated system

• Use it to identify specific ways in which the newer AI's concept of human values is wrong, and keep correcting the newer AI's concept of human values until it's good (maybe using active learning)

You need the new AI and the old AI to communicate with one another. But details of how they work can be totally different if you have them communicate using labeled data. Training one ML model to predict the output of some other ML model is a technique I see every so often in machine learning papers... "Distilling the Knowledge in a Neural Network" is a well-known example of this.

Finally, you wrote:

That’s assuming things don’t go irreversibly wrong before then - including not just hostile takeover, but even just development of deceptive behavior, since this scenario does not have any built-in mechanism to detect deception.

Mesa-optimizers are a real danger, but if we put those aside for a moment, I don't think there is much risk of a hostile takeover from an unsupervised learning system since it's not an agent.

I like this post a lot, and I think it points out a key crux between what I would term the "Yudkowsky" side (which seems to mostly include MIRI, though I'm not too sure about individual researchers' views) and "everybody else".

In particular, the disagreement seems to crystallize over the question of whether "human values" really are a natural abstraction. I suspect that if Eliezer thought that they were, he would be substantially less worried about AI alignment than he currently is (though naturally all of this is my read on his views).

You do provide some reasons to think that human values might be a natural abstraction, both in the post itself and in the comments, but I don't see these reasons as particularly compelling ones. The one I view as the most compelling is the argument that humans seems to be fairly good at identifying and using natural abstractions, and therefore any abstract concept that we seem to be capable of grasping fairly quickly has a strong chance of being a natural one.

However, I think there's a key difference between abstractions that are developed for the purposes of prediction, and abstractions developed for other purposes (by which I mostly mean "RL"). To the extent that a predictor doesn't have sufficient computational power to form a low-level model of whatever it's trying to predict, I definitely think that the abstractions it develops in the process of trying to improve its prediction will to a large extent be natural ones. (You lay out the reasons for this clearly enough in the post itself, so I won't repeat them here.)

It seems to me, though, that if we're talking about a learning agent that's actually trying to take actions to accomplish things in some environment, there's a substantial amount of learning going on that has nothing to do with learning to predict things with greater accuracy! The abstractions learned in order to select actions from a given action-space in an attempt to maximize a given reward function--these, I see little reason to expect will be natural. In fact, if the computational power afforded to the agent is good but not excellent, I expect mostly the opposite: a kludge of heuristics and behaviors meant to address different subcases of different situations, with not a whole lot of rhyme or reason to be found.

As agents go, humans are definitely of the latter type. And, therefore, I think the fact that we intuitively grasp the concept of "human values" isn't necessarily an argument that "human values" are likely to be natural, in the way that it would be for e.g. trees. The latter would have been developed as a predictive abstraction, whereas the former seems to mainly consist of what I'll term a reward abstraction. And it's quite plausible to me that reward abstractions are only legible by default to agents which implement that particular reward abstraction, and not otherwise. If that's true, then the fact that humans know what "human values" are is merely a consequence of the fact that we happen to be humans, and therefore have a huge amount of mind-structure in common.

To the extent that this is comparable to the branching pattern of a tree (which is a comparison you make in the post), I would argue that it increases rather than lessens the reason to worry: much like a tree's branch structure is chaotic, messy, and overall high-entropy, I expect human values to look similar, and therefore not really encompass any kind of natural category.

So in principle, it doesn’t even matter what kind of model we use or how it’s represented; as long the predictive power is good enough, values will be embedded in there, and the main problem will be finding the embedding.

I will agree with this. However, notice what this doesn't say. It doesn't say "any model powerful enough to be really dangerous contains human values". Imagine a model that was good at a lot of science and engineering tasks. It was good enough at nuclear physics to design effective fusion reactors and bombs. It knew enough biology to design a superplage. It knew enough molecular dynamics to design self replicating nanotech. It knew enough about computer security to hack most real world systems. But it didn't know much about how humans thought. It's predictions are far from maxentropy, if it sees people walking along a street, it thinks they will probably carry on walking, not fall to the ground twiching randomly. Lets say that the model is as predictively accurate as you would be when asked to predict the behaviour of a stranger from a few seconds of video. This AI doesn't contain a model of human values anywhere in it.

We can't just assume that every AI powerful enough to be dangerous contains a model of human values, however I suspect most of them will in practice.

This is the sort of thing I've been thinking about since "What's the dream for giving natural language commands to AI? [LW · GW]" (which bears obvious similarities to this post). The main problems I noted there apply similarly here:

• Prediction in the supervised task might not care about the full latent space used for the unsupervised tasks, losing information.
• Little to no protection from Goodhart's law. Things that are extremely good proxies for human values still might not be safe to optimize.
• Doesn't care about metaethics, just maximizes some fixed thing. Which wouldn't be a problem if it was meta-ethically great to start with, but it probably incorporates plenty of human foibles in order to accurately predict us.

The killer is really that second one. If you run this supervised learning process, and it gives you a bunch of rankings of things in terms of their human values score, this isn't a safe AI even if it's on average doing a great job, because the thing that gets the absolute best score is probably an exploit of the specific pattern-recognition algorithm used to do the ranking. In short, we still need to solve the other-izer problem.

Actually, your trees example does give some ideas. Could you look inside a GAN trained on normal human behavior and identify what parts of it were the "act morally" or "be smart" parts, and turn them up? Choosing actions is, after all, a generative problem, not a classification or regression problem.

So far, we’ve only talked about one AI ending up aligned, or a handful ending up aligned at one particular time. However, that isn’t really the ultimate goal of AI alignment research. What we really want is for AI to remain aligned in the long run, as we (and AIs themselves) continue to build new and more powerful systems and/or scale up existing systems over time.

I think this suggests an interesting path where alignment by default might be able to serve as a bridge to better alignment mechanisms, i.e. if it works and we can select for AIs that contains representations of human values, then we might be able to prioritize this in a slow takeoff scenario so that in the early phases of it we at least have mostly aligned AI that helps us build better mechanisms for alignment (as opposed to these AIs simply building successors directly with the hope that they maintain alignment with human values in the process).

In this post, the author describes a pathway by which AI alignment can succeed even without special research effort. The specific claim that this can happen "by default" is not very important, IMO (the author himself only assigns 10% probability to this). On the other hand, viewed as a technique that can be deliberately used to help with alignment, this pathway is very interesting.

The author's argument can be summarized as follows:

• For anyone trying to predict events happening on Earth, the concept of "human values" is a "natural abstraction", i.e. something that has to be a part of any model that's not too computationally expensive (so that it doesn't bypass the abstraction by anything like accurate simulation of human brains).
• Therefore, unsupervised learning will produce models in which human values are embedded in some simple way (e.g. a small set of neurons in an ANN).
• Therefore, if supervised learning is given the unsupervised model as a starting point, it is fairly likely to converge to true human values even from a noisy and biased proxy.

[EDIT: John pointed out that I misunderstood his argument: he didn't intend to say that human values are a natural abstraction, but only that their inputs are natural abstractions. The following discussion still applies.]

The way I see it, this argument has learning-theoretic justification even without appealing to anything we know about ANNs (and therefore without assuming the AI in question is an ANN). Consider the following model: an AI receives a sequence of observations that it has to make predictions about. It also receives labels, but these are sparse: it is only given a label once in a while. If the description complexity of the true label function is high, the sample complexity of learning to predict labels via a straightforward approach (i.e. without assuming a relationship between the dynamics and the label function) is also high. However, if the relative description complexity of the label function w.r.t. the dynamics producing the observations is low, then we can use the abundance of observations to achieve lower effective sample complexity. I'm confident that this can be made rigorous.

Therefore, we can recast the thesis of this post as follows: Unsupervised learning of processes happening on Earth, for which we have plenty of data, can reduce the size of the dataset required to learn human values, or allow better generalization from a dataset of the same size.

One problem the author doesn't talk about here is daemons / inner misalignment^[1]. In the comment section, the author writes:

inner alignment failure only applies to a specific range of architectures within a specific range of task parameters - for instance, we have to be optimizing for something, and there has to be lots of relevant variables observed only at runtime, and there has to be something like a "training" phase in which we lock-in parameter choices before runtime, and for the more disastrous versions we usually need divergence of the runtime distribution from the training distribution. It's a failure mode which assumes that a whole lot of things look like today's ML pipelines.

This might or might not be a fair description of inner misalignment in the sense of Hubinger et al. However, this is definitely not a fair description of the daemonic attack vectors in general. The potential for malign hypotheses (learning of hypotheses / models containing malign subagents) exists in any learning system, and in particular malign simulation hypotheses are a serious concern.

Relatedly, the author is too optimistic (IMO) in his comparison of this technique to alternatives:

...when alignment-by-default works, it’s a best-case scenario. The AI has a basically-correct model of human values, and is pursuing those values. Contrast this to things like IRL variants, which at best learn a utility function which approximates human values (which are probably not themselves a utility function). Or the HCH family of methods, which at best mimic a human with a massive hierarchical bureaucracy at their command, and certainly won’t be any more aligned than that human+bureaucracy would be.

This sounds to me like a biased perspective resulting from looking for flaws in other approaches harder than flaws in this approach. Natural abstractions potentially lower the sample complexity of learning human values, but they cannot lower it to zero. We still need some data to learn from and some model relating this data to human values, and this model can suffer from the usual problems. In particular, the unsupervised learning phase does little to inoculate us from malign simulation hypotheses that can systematically produce catastrophically erroneous generalization.

If IRL variants learn a utility function while human values are not a utility function, then avoiding this problem requires identifying the correct type signature of human values^[2], in this approach as well. Regarding HCH, Human + "bureaucracy" might or might not be aligned, depending on how we organize the "bureaucracy" (see also [LW(p) · GW(p)]). If HCH can fail in some subtle way (e.g. systems of humans are misaligned to individual humans), then similar failure modes might affect this approach as well (e.g. what if "Molochian" values are also a natural abstraction).

In summary, I found this post quite insightful and important, if somewhat too optimistic.

Great post!

That might have been discussed in the comments, but my gut reaction to the tree example was not "It's not really understanding tree" but "It's understanding trees visually". That is, I think the examples point to trees being a natural abstraction with regard to images made of pixels. In that sense, dogs and cats and other distinct visual objects might fit your proposal of natural abstraction. Yet this doesn't entail that trees are a natural abstraction when given the position of atoms, or sounds (to be more abstract). I thus think that natural abstractions should be defined with regard for the sort of data that is used.

For human values, I might accept that they are natural abstraction, but I don't know for which kind of data. Is audiovisual data (as in youtube videos) enough? Do we also need textual data? Neuroimagery? I don't know, and that makes me slightly more pessimistic about a unsupervised model learning human values by default.

This requires hitting a window - our data needs to be good enough that the system can tell it should use human values as a proxy, but bad enough that the system can’t figure out the specifics of the data-collection process enough to model it directly. This window may not even exist.

I like this framing, it is clarifying.

When alignment-by-default works, it’s basically a best-case scenario, so we can safely use the system to design a successor without worrying about amplification of alignment errors (among other things).

didn't understand how this was derived or what other results/ideas it is referencing.

This will likely not work for dualistic model of human values (and other complex models, like family system). In this model, a human have an ethical system and opposing suppressed desires.

For example, I think that it is good to eat less, but have a desire for overeating. Combined, they produce behaviour in which I often have eating binges following by periods of fasting. If an AI want to predict my behaviour, it may suggest that I want to have periods of overeating and extrapolate accordingly. However, I consciously endorse only "eating less ethics" and regard it as my true values. As S.Armstrong wrote, there is always an assumption which part of me should be regarded as "true values".

when alignment-by-default works, we can use the system to design a successor without worrying about amplification of alignment errors

Anything neural net related starts with random noise and performs gradient descent style steps. This doesn't get you the global optimal, it gets you some point that is approximately a local optimal, which depends on the noise, the nature of the search space, and the choice of step size.

If nothing else, the training data will contain sensor noise.

At best you are going to get something that roughly corresponds to human values.

Just because it isn't obvious where the noise entered the system doesn't make it noiseless. Just because you gave what we actually want, and the value of a neuron in a neural net the same name, doesn't make them the same thing.

Consider the large set of references with representative members "What Alice makes long term plans towards", "What Bobs impulsive action tends towards", "What Alice says is good and right when her social circle are listening", "What Carl listens to when deciding which politician to vote for", "What news makes Eric instinctively feel good", "what makes Fred presses the reward button during AI training" ect ect.

If these all referred to the same preference ordering over states of the world, then we could call that human values, and have a natural concept.

Trees are a fairly natural concept because "tall green things" and "Lifeforms that are >10% cellulose" point to a similar set of objects. There are many different simple boundaries in concept-space that largely separate trees from non trees. Trees are tightly clustered in thing-space.

To the extent that all those references refer to the same thing, we can't expect an AI to distinguish between them. To the extent that they refer to different concepts, at best the AI will have a separate concept for each.

Suppose you run the microscope AI, and you find that you have a whole load of concepts that kind of match "human values" to different degrees. These represent different people and different embeddings of value. (Of course, "What Carl listens to when deciding which politician to vote for" contains Carls distrust of political promises. "what makes Fred presses the reward button during AI training" includes the time Fred tripped up and slammed the button by accident. Each of the easily accessible concepts is a bit different and includes its own bit of noise)

This requires hitting a window - our data needs to be good enough that the system can tell it should use human values as a proxy, but bad enough that the system can’t figure out the specifics of the data-collection process enough to model it directly. This window may not even exist.

 

are there any real world examples of this? not necessarily in human-values setting 

Summary

This article claims that:

• Unsupervised learning systems will likely learn many “natural abstractions” of concepts like “trees” or “human values”. Maybe they will even end up being simply a “feature direction”.
  • One reason to expect this is that to make good predictions, you only need to conserve information that’s useful at a distance. And this information can be imagined being a “natural abstraction”. 
• If you then have an RL system or supervised learner who can use the unsupervised activations to solve a problem, then it can directly behave in such a way as to “satisfy the natural abstraction” of, e.g., human values. 
  • This would be a quick way for the model to behave well on such a task. Later on, the model might find the unnatural proxy goal and maximize that, but that wouldn’t be the first thing found. 
• Thus, we may up with alignment by default. Such an aligned AGI could then be used to align successor AIs, and it might be better at that than humans are since it’s smarter. 
  • Note that if we have “alignment by default”, then the alignment of a successor system might work better than with competitive HCH or IRL. Reason: humans+bureaucracies may not be aligned, and IRL finds a utility function, which is of the wrong type signature. This may lead to a compounding of alignment errors when building successor AIs. 
  • Variations of “alignment by default” would, e.g., find the “human value abstraction” in one AI and then “plant it” into the search operation of an AGI. I.e., alignment would not be solved by default for all AIs, but it’s defaulty-enough that we still win. 

My Opinion:

• Human values are less “one thing” than trees or humans are. “Human values are active” is not a sensible piece of information since there are so many different types of human values. Admittedly, this also applies to trees, but it does feel like it’s pointing to a difficulty.
• Overall, this makes me think that, possibly, the proxy goal is often a more natural abstraction than human values. My hope mainly comes from the thought that the proxy goals are hopefully specific enough to the RL/supervised task that they didn’t appear in the unsupervised training phase.
  • But there are reasons against that: humans will put lots of information about ML training processes into the training data of any unsupervised system, meaning that to make good predictions, the systems should probably represent these proxy goals quite well. Only if they do would they make accurate predictions about, e.g., contemporary alignment errors.