What mechanism would ensure that these two logic pieces only fire at the same time? Whatever it is, I expect that mechanism to be changed in response to failures.
The argument that Adam mentions me making is fleshed out in this post.
I discuss the possibility of it going in some other direction when I say "The two most salient options to me". But the bit of Evan's post that this contradicts is:
Now, if the model gets to the point where it's actually just failing because of this, then gradient descent will probably just remove that check—but the trick is never to actually get there.
Interesting report :) One quibble:
For one, our AIs can only use “things like SPI” if we actually formalize the approach
I don't see why this is the case. If it's possible for humans to start using things like SPI without a formalisation, why couldn't AIs too? (I agree it's more likely that we can get them to do so if we formalise it, though.)
Whether this is a point for the advocate or the skeptic depends on whether advances in RL from human feedback unlock other alignment work more than they unlock other capabilities work. I think there's room for reasonable disagreement on this question, although I favour the former.
Skeptic: It seems to me that the distinction between "alignment" and "misalignment" has become something of a motte and bailey. Historical arguments that AIs would be misaligned used it in sense 1: "AIs having sufficiently general and large-scale motivations that they acquire the instrumental goal of killing all humans (or equivalently bad behaviour)". Now people are using the word in sense 2: "AIs not quite doing what we want them to do". But when our current AIs aren't doing quite what we want them to do, is that mainly evidence that future, more general... (read more)
These aren't complicated or borderline cases, they are central example of what we are trying to avert with alignment research.
I'm wondering if the disagreement over the centrality of this example is downstream from a disagreement about how easy the "alignment check-ins" that Critch talks about are. If they are the sort of thing that can be done successfully in a couple of days by a single team of humans, then I share Critch's intuition that the system in question starts off only slightly misaligned. By contrast, if they require a significant proportion of ... (read more)
I'm not sure what you mean by "actual computation rather than the algorithm as a whole". I thought that I was talking about the knowledge of the trained model which actually does the "computation" of which move to play, and you were talking about the knowledge of the algorithm as a whole (i.e. the trained model plus the optimising bot).
Mesa-optimizers are in the search space and would achieve high scores in the training set, so why wouldn't we expect to see them?
I like this as a statement of the core concern (modulo some worries about the concept of mesa-optimisation, which I'll save for another time).
With respect to formalization, I did say up front that less-formal work, and empirical work, is still valuable.
I missed this disclaimer, sorry. So that assuages some of my concerns about balancing types of work. I'm still not sure what intuitions or arguments underlie your optimism about fo... (read more)
I have fairly mixed feelings about this post. On one hand, I agree that it's easy to mistakenly address some plausibility arguments without grasping the full case for why misaligned mesa-optimisers might arise. On the other hand, there has to be some compelling (or at least plausible) case for why they'll arise, otherwise the argument that 'we can't yet rule them out, so we should prioritise trying to rule them out' is privileging the hypothesis.
Secondly, it seems like you're heavily prioritising formal tools and methods for studying mesa-optimisatio... (read more)
I agree with much of this. I over-sold the "absence of negative story" story; of course there has to be some positive story in order to be worried in the first place. I guess a more nuanced version would be that I am pretty concerned about the broadest positive story, "mesa-optimizers are in the search space and would achieve high scores in the training set, so why wouldn't we expect to see them?" -- and think more specific positive stories are mostly of illustrative value, rather than really pointing to gears that I expect to be important. (With the excep... (read more)
The trained AlphaZero model knows lots of things about Go, in a comparable way to how a dog knows lots of things about running.
But the algorithm that gives rise to that model can know arbitrarily few things. (After all, the laws of physics gave rise to us, but they know nothing at all.)
I'd say that this is too simple and programmatic to be usefully described as a mental model. The amount of structure encoded in the computer program you describe is very small, compared with the amount of structure encoded in the neural networks themselves. (I agree that you can have arbitrarily simple models of very simple phenomena, but those aren't the types of models I'm interested in here. I care about models which have some level of flexibility and generality, otherwise you can come up with dumb counterexamples like rocks "knowing" the laws of physic... (read more)
I don't think there is a fundamental difference in kind between trees, bacteria, humans, and hypothetical future AIs
There's at least one important difference: some of these are intelligent, and some of these aren't.
It does seem plausible that the category boundary you're describing is an interesting one. But when you indicate in your comment below that you see the "AI hypothesis" and the "life hypothesis" as very similar, then that mainly seems to indicate that you're using a highly nonstandard definition of AI, which I expect will lead to confusion.
It feels like this post pulls a sleight of hand. You suggest that it's hard to solve the control problem because of the randomness of the starting conditions. But this is exactly the reason why it's also difficult to construct an AI with a stable implementation. If you can do the latter, then you can probably also create a much simpler system which creates the smiley face.
Similarly, in the real world, there's a lot of randomness which makes it hard to carry out tasks. But there are a huge number of strategies for achieving things in the world which don't r... (read more)
The human knows the rules and the win condition. The optimisation algorithm doesn't, for the same reason that evolution doesn't "know" what dying is: neither are the types of entities to which you should ascribe knowledge.
it's not obvious to me that this is a realistic target
Perhaps I should instead have said: it'd be good to explain to people why this might be a useful/realistic target. Because if you need propositions that cover all the instincts, then it seems like you're basically asking for people to revive GOFAI.
(I'm being unusually critical of your post because it seems that a number of safety research agendas lately have become very reliant on highly optimistic expectations about progress on interpretability, so I want to make sure that people are forced to defend that assumption rather than starting an information cascade.)
As an additional reason for the importance of tabooing "know", note that I disagree with all three of your claims about what the model "knows" in this comment and its parent.
(The definition of "know" I'm using is something like "knowing X means possessing a mental model which corresponds fairly well to reality, from which X can be fairly easily extracted".)
I think at this point you've pushed the word "know" to a point where it's not very well-defined; I'd encourage you to try to restate the original post while tabooing that word.
This seems particularly valuable because there are some versions of "know" for which the goal of knowing everything a complex model knows seems wildly unmanageable (for example, trying to convert a human athlete's ingrained instincts into a set of propositions). So before people start trying to do what you suggested, it'd be good to explain why it's actually a realistic target.
I used to define "agent" as "both a searcher and a controller"
Oh, I really like this definition. Even if it's too restrictive, it seems like it gets at something important.
I'm not sure what you meant by "more compressed".
Sorry, that was quite opaque. I guess what I mean is that evolution is an optimiser but isn't an agent, and in part this has to do with how it's a very distributed process with no clear boundary around it. Whereas when you have the same problem being solved in a single human brain, then that compression makes it easier to point to the huma... (read more)
To me it sounds like you're describing (some version of) agency, and so the most natural term to use would be mesa-agent.
I'm a bit confused about the relationship between "optimiser" and "agent", but I tend to think of the latter as more compressed, and so insofar as we're talking about policies it seems like "agent" is appropriate. Also, mesa-optimiser is taken already (under a definition which assumes that optimisation is equivalent to some kind of internal search).
Yann LeCun: ... instrumental subgoals are much weaker drives of behavior than hardwired objectives. Else, how could one explain the lack of domination behavior in non-social animals, such as orangutans.
What's your specific critique of this? I think it's an interesting and insightful point.
My internal model of you is that you believe this approach would not be enough because the utility would not be defined on the internal concepts of the agent. Yet I think it doesn't have so much to be defined on these internal concepts itself than to rely on some assumption about these internal concepts.
Yeah, this is an accurate portrayal of my views. I'd also note that the project of mapping internal concepts to mathematical formalisms was the main goal of the whole era of symbolic AI, and failed badly. (Although the analogy is a little loose, so I wouldn... (read more)
Wouldn't these coherence arguments be pretty awesome? Wouldn't this be a massive step forward in our understanding (both theoretical and practical) of health, damage, triage, and risk allocation?
Insofar as such a system could practically help doctors prioritise, then that would be great. (This seems analogous to how utilities are used in economics.)
But if doctors use this concept to figure out how to treat patients, or using it when designing prostheses for their patients, then I expect things to go badly. If you take HP as a guiding principle - for exampl... (read more)
Do you think that's a problem?
I'm inclined to think so, mostly because terms shouldn't be introduced unnecessarily. If we can already talk about systems that are capable/competent at certain tasks, then we should just do that directly.
I guess the mesa- prefix helps point towards the fact that we're talking about policies, not policies + optimisers.
Probably my preferred terminology would be:
Mesa-controller refers to any effective strategies, including mesa-searchers but also "dumber" strategies which nonetheless effectively steer toward a misaligned objective. For example, thermostat-like strategies, or strategies which have simply memorized a number of effective interventions.
I'm confused about what wouldn't qualify as a mesa-controller. In practice, is this not synonymous with "capable"?
Also, why include "misaligned" in this definition? If mesa-controller turns out to be a useful concept, then I'd want to talk about both aligned and misaligned mesa-controllers.
Someone might say, well I understand that if I don't pay, then it means I would have lost out if it had come up heads, but since I know it didn't came up heads, I don't care. Making this more precise, when constructing counterfactuals for a decision, if we know fact F about the world before we've made our decision, F must be true in every counterfactual we construct (call this Principle F).
The problem is that principle F elides over the difference between facts which are logically caused by your decision, and facts which aren't. For example, in Parfit's hi... (read more)
by only considering the branches of reality that are consistent with our knowledge
I know that, in the branch of reality which actually happened, Omega predicted my counterfactual behaviour. I know that my current behaviour is heavily correlated with my counterfactual behaviour. So I know that I can logically cause Omega to give me $10,000. This seems exactly equivalent to Newcomb's problem, where I can also logically cause Omega to give me a lot of money.
So if by "considering [other branches of reality]" you mean "taking predicted counterfactuals into acco... (read more)
I don't see why the Counterfactual Prisoner's Dilemma persuades you to pay in the Counterfactual Mugging case. In the counterfactual prisoner's dilemma, I pay because that action logically causes Omega to give me $10,000 in the real world (via influencing the counterfactual). This doesn't require shifting the locus of evaluation to policies, as long as we have a good theory of which actions are correlated with which other actions (e.g. paying in heads-world and paying in tails-world).
In the counterfactual mugging, by contrast, the whole point is that payin... (read more)
Thanks for writing this post, Katja; I'm very glad to see more engagement with these arguments. However, I don't think the post addresses my main concern about the original coherence arguments for goal-directedness, which I'd frame as follows:
There's some intuitive conception of goal-directedness, which is worrying in the context of AI. The old coherence arguments implicitly used the concept of EU-maximisation as a way of understanding goal-directedness. But Rohin demonstrated that the most straightforward conception of EU-maximisation (which I'll call beh... (read more)
I personally found this post valuable and thought-provoking. Sure, there's plenty that it doesn't cover, but it's already pretty long, so that seems perfectly reasonable.
I particularly I dislike your criticism of it as strawmanish. Perhaps that would be fair if the analogy between RL and evolution were a standard principle in ML. Instead, it's a vague idea that is often left implicit, or else formulated in idiosyncratic ways. So posts like this one have to do double duty in both outlining and explaining the mainstream viewpoint (often a major task in its o... (read more)
... (read more)there’s a “solving the problem twice” issue. As mentioned above, in Case 5 we need both the outer and the inner algorithm to be able to do open-ended construction of an ever-better understanding of the world—i.e., we need to solve the core problem of AGI twice with two totally different algorithms! (The first is a human-programmed learning algorithm, perhaps SGD, while the second is an incomprehensible-to-humans learning algorithm. The first stores information in weights, while the second stores information in activations, assuming a GPT-like architecture.
It seems totally plausible to give AI systems an external memory that they can read to / write from, and then you learn linear algebra without editing weights but with editing memory. Alternatively, you could have a recurrent neural net with a really big hidden state, and then that hidden state could be the equivalent of what you're calling "synapses".
I agree with Steve that it seems really weird to have these two parallel systems of knowledge encoding the same types of things. If an AGI learned the skill of speaking english during training, but then learn... (read more)
Nice post. The one thing I'm confused about is:
Institutionally, we are very uncertain whether to prioritize this (and if we do where it should be housed and how our giving should be structured).
It seems to me that the type of research you're discussing here is already seen as a standard way to make progress on the full alignment problem - e.g. the Stiennon et al. paper you cited, plus earlier work on reward modeling by Christiano, Leike, and others. Can you explain why you're institutionally uncertain whether to prioritise it - is it because of the objecti... (read more)
Great post, and I'm glad to see the argument outlined in this way. One big disagreement, though:
the Judge box will house a relatively simple algorithm written by humans
I expect that, in this scenario, the Judge box would house a neural network which is still pretty complicated, but which has been trained primarily to recognise patterns, and therefore doesn't need "motivations" of its own.
This doesn't rebut all your arguments for risk, but it does reframe them somewhat. I'd be curious to hear about how likely you think my version of the judge is, and why.
Thanks for the reply. To check that I understand your position, would you agree that solving outer alignment plus solving reward tampering would solve the pointers problem in the context of machine learning?
Broadly speaking, I think our disagreement here is closely related to one we've discussed before, about how much sense it makes to talk about outer alignment in isolation (and also about your definition of inner alignment), so I probably won't pursue this further.
Above you say:
Now, the basic problem: our agent’s utility function is mostly a function of latent variables. ... Those latent variables:
- May not correspond to any particular variables in the AI’s world-model and/or the physical world
- May not be estimated by the agent at all (because lazy evaluation)
- May not be determined by the agent’s observed data
… and of course the agent’s model might just not be very good, in terms of predictive power.
And you also discuss how:
... (read more)Human "values" are defined within the context of humans' world-models, and don't necessarily make
The question then is, what would it mean for such an AI to pursue our values?
Why isn't the answer just that the AI should:
1. Figure out what concepts we have;
2. Adjust those concepts in ways that we'd reflectively endorse;
3. Use those concepts?
The idea that almost none of the things we care about could be adjusted to fit into a more accurate worldview seems like a very strongly skeptical hypothesis. Tables (or happiness) don't need to be "real in a reductionist sense" for me to want more of them.
I agree with all the things you said. But you defined the pointer problem as: "what functions of what variables (if any) in the environment and/or another world-model correspond to the latent variables in the agent’s world-model?" In other words, how do we find the corresponding variables? I've given you an argument that the variables in an AGI's world-model which correspond to the ones in your world-model can be found by expressing your concept in english sentences.
The problem of determining how to construct a feedback signal which refers to those variabl... (read more)
I need some way to say what the values-relevant pieces of my world model are "pointing to" in the real world. I think this problem - the “pointers to values” problem, and the “pointers” problem more generally - is the primary conceptual barrier to alignment right now.
It seems likely that an AGI will understand very well what I mean when I use english words to describe things, and also what a more intelligent version of me with more coherent concepts would want those words to actually refer to. Why does this not imply that the pointers problem will be solve... (read more)
I think 'robust instrumentality' is basically correct for optimal actions, because there's no question of 'emergence': optimal actions just are.
If I were to put my objection another way: I usually interpret "robust" to mean something like "stable under perturbations". But the perturbation of "change the environment, and then see what the new optimal policy is" is a rather unnatural one to think about; most ML people would more naturally think about perturbing an agent's inputs, or its state, and seeing whether it still behaved instrumentally.
A more accurate description might be something like "ubiquitous instrumentality"? But this isn't a very aesthetically pleasing name.
Can you elaborate? 'Robust' seems natural for talking about robustness to perturbation in the initial AI design (different objective functions, to the extent that that matters) and robustness against choice of environment.
The first ambiguity I dislike here is that you could either be describing the emergence of instrumentality as robust, or the trait of instrumentality as robust. It seems like you're trying to do the former, but because "robust" modifies "instrumentality", the latter is a more natural interpretation.
For example, if I said "life on earth is... (read more)
Yepp, this is a good point. I agree that there won't be a sharp distinction, and that ML systems will continue to do online learning throughout deployment. Maybe I should edit the post to point this out. But three reasons why I think the training/deployment distinction is still underrated:
The burden of proof should be on whoever wants to claim that AI will be fine by default, not on whoever wants to claim it won't be fine by default.
I'm happy to wrap up this conversation in general, but it's worth noting before I do that I still strongly disagree with this comment. We've identified a couple of interesting facts about goals, like "unbounded large-scale final goals lead to convergent instrumental goals", but we have nowhere near a good enough understanding of the space of goal-like behaviour to say that everything apart from a "very small reg... (read more)
I agree with the two questions you've identified as the core issues, although I'd slightly rephrase the former. It's hard to think about something being aligned indefinitely. But it seems like, if we have primarily used a given system for carrying out individual tasks, it would take quite a lot of misalignment for it to carry out a systematic plan to deceive us. So I'd rephrase the first option you mention as "feeling pretty confident that something that generalises from 1 week to 1 year won't become misaligned enough to cause disasters". This point seems ... (read more)
1. The goals that we imagine superintelligent AGI having, when spelled out in detail, have ALL so far been the sort that would very likely lead to existential catastrophe of the instrumental convergence variety.
2. We've even tried hard to imagine goals that aren't of this sort, and so far we haven't come up with anything. Things that seem promising, like "Place this strawberry on that plate, then do nothing else" actually don't work when you unpack the details.
Okay, this is where we disagree. I think what "unpacking the details" actually gives you is somet... (read more)
I disagree that we have no good justification for making the "vast majority" claim.
Can you point me to the sources which provide this justification? Your analogy seems to only be relevant conditional on this claim.
My point is that in the context in which the classic arguments appeared, they were useful evidence that updated people in the direction of "Huh AI could be really dangerous" and people were totally right to update in that direction on the basis of these arguments
They were right to update in that direction, but that doesn't mean that they were rig... (read more)
Re counterfactual impact: the biggest shift came from talking to Nate at BAGI, after which I wrote this post on disentangling arguments about AI risk, in which I identified the "target loading problem". This seems roughly equivalent to inner alignment, but was meant to avoid the difficulties of defining an "inner optimiser". At some subsequent point I changed my mind and decided it was better to focus on inner optimisers - I think this was probably catalysed by your paper, or by conversations with Vlad which were downstream of the paper. I think the paper ... (read more)
Ah, cool; I like the way you express it in the short form! I've been looking into the concept of structuralism in evolutionary biology, which is the belief that evolution is strongly guided by "structural design principles". You might find the analogy interesting.
One quibble: in your comment on my previous post, you distinguished between optimal policies versus the policies that we're actually likely to train. But this isn't a component of my distinction - in both cases I'm talking about policies which actually arise from training. My point is that there a... (read more)
Saying "vast majority" seems straightfowardly misleading. Bostrom just says "a wide range"; it's a huge leap from there to "vast majority", which we have no good justification for making. In particular, by doing so you're dismissing bounded goals. And if you're talking about a "state of ignorance" about AI, then you have little reason to override the priors we have from previous technological development, like "we build things that do what we want".
On your analogy, see the last part of my reply to Adam below. The process of building things intrinsically pi... (read more)
The images in this post seem to be broken.