When I think about solutions to AI alignment, I often think about 'meaningful reductionism.' That is, if I can factor a problem into two parts, and the parts don't actually rely on each other, now I have two smaller problems to solve. But if the parts are reliant on each other, I haven't really simplified anything yet.
While impact measures feel promising to me as a cognitive strategy (often my internal representation of politeness feels like 'minimizing negative impact', like walking on sidewalks in a way that doesn't startle birds), they don't feel promising to me as reductionism. That is, if I already had a solution to the alignment problem, then impact measures would likely be part of how I implement that solution, but solving it separately from alignment doesn't feel like it gets me any closer to solving alignment.
[The argument here I like most rests on the difference between costs and side effects; we don't want to minimize side effects because that leads to minimizing good side effects also, and it's hard to specify the difference between 'side effects' and 'causally downstream effects,' and so on. But if we just tell the AI "score highly on a goal measure while scoring low on this cost measure," this only works if we specified the goal and the cost correctly.]
But there's a different approach to AI alignment, which is something more like 'correct formalisms.' We talk sometimes about handing a utility function to the robot, or (in old science fiction) providing it with rules to follow, or so on, and by seeing what it actually looks like when we follow that formalism we can figure out how well that formalism fits to what we're interested in. Utility functions on sensory inputs don't seem alignable because of various defects (like wireheading), and so it seems like the right formalism needs to have some other features (it might still be a utility function, but it needs to be an utility function over mental representations of external reality in such a way that the mental representation tracks external reality even when you have freedom to alter your mental representation, in a way that we can't turn into code yet).
So when I ask myself questions like "why am I optimistic about researching impact measures now?" I get answers like "because exploring the possibility space will make clear exactly how the issues link up." For example, looking at things like relative reachability made it clear to me how value-laden the ontology needs to be in order for a statistical measure on states to be meaningful. This provides a different form-factor for 'transferring values to the AI'; instead of trying to ask something like "is scenario A or B better?" and train a utility function, I might instead try to ask something like "how different are scenarios A and B?" or "how are scenarios A and B different?" and train an ontology, with the hopes that this makes other alignment problems easier because the types line up somewhat more closely.
[I think even that last example still performs poorly on the 'meaningful reductionism' angle, since getting more options for types to use in value loading doesn't seem like it addresses the core obstacles of value loading, but provides some evidence of how it could be useful or clarify thinking.]
this only works if we specified the goal and the cost correctly
Wait, why doesn't it work if you just specify the cost (impact) correctly?
My concern is similar to Wei Dai's: it seems to me that at a fundamental physical level, any plan involving turning on a computer that does important stuff will make pretty big changes to the world's trajectory in phase space. Heat dissipation will cause atmospheric particles to change their location and momentum, future weather patterns will be different, people will do things at different times (e.g. because they're waiting for a computer program to run, or because the computer is designed to change the flow of traffic through a city), meet different people, and have different children. As a result, it seems hard for me to understand how impact measures could work in the real world without a choice of representation very close to the representation humans use to determine the value of different worlds. I suspect that this will need input from humans similar to what value learning approaches might need, and that once it's done one could just do value learning and dispense with the need for impact measures. That being said, this is more of an impression than a belief - I can't quite convince myself that no good method of impact regularisation exists, and some other competent people seem to disagree with me.
I can't quite convince myself that no good method of value learning exists, and some other competent people seem to disagre ewith me.
No good method of measuring impact, presumably?
How does this concern interact with the effective representation invariance claim I made when introducing AUP?
I have an intuition that while impact measures as a way of avoiding negative side effects might work well in toy models, it will be hard or impossible to get them to work in the real world, because what counts as a negative side effect in the real world seems too complex to easily capture. It seems like AUP tries to get around this by aiming at a lower bar than "avoid negative side effects", namely "avoid catastrophic side effects", and aside from whether it actually succeeds at clearing this lower bar, it would mean that an AI that is only "safe" because of AUP can't be safely used for ordinary goals (e.g., invent a better widget, or make someone personally more successful in life) and instead we have to somehow restrict them to being used just for goals that relate to x-risk reduction, where it's worthwhile to risk incurring less-than-catastrophic negative side effects.
As a side note, it seems generally the case that some approaches to AI safety/alignment aim at the higher bar of "safe for general use" and others aim at "safe enough to use for x-risk reduction", and this isn't always made clear, which can be a source of confusion for both AI safety/alignment researchers and others such as strategists and policy makers.
I have an intuition that while impact measures as a way of avoiding negative side effects might work well in toy models, it will be hard or impossible to get them to work in the real world
Do you think there any experiments that could be performed that would change your view on this point? Or is an impact measure the type of thing that does not generalize well from testing environment to the real world?
I have an intuition that while impact measures as a way of avoiding negative side effects might work well in toy models, it will be hard or impossible to get them to work in the real world, because what counts as a negative side effect in the real world seems too complex to easily capture.
Although a far cry from "[avoiding side effects] in the real world", see Avoiding Side Effects in Complex Environments as another piece of evidence to update on.
Thanks Alex for starting this discussion and thanks everyone for the thought-provoking answers. Here is my current set of concerns about the usefulness of impact measures, sorted in decreasing order of concern:
Irrelevant factors. When applied to the real world, impact measures are likely to be dominated by things humans don't care about (heat dissipation, convection currents, positions of air molecules, etc). This seems likely to happen to value-agnostic impact measures, e.g. AU with random utility functions, which would mostly end up rewarding specific configurations of air molecules.
This may be mitigated by inability to perceive the irrelevant factors, which results in a more coarse-grained state representation: if the agent can't see air molecules, all the states with different air molecule positions will look the same, as they do to humans. Some human-relevant factors can also be difficult to perceive, e.g. the presence of poisonous gas in the room, so we may not want to limit the agent's perception ability to human level. Automatically filtering out irrelevant factors does seem difficult, and I think this might imply that it is impossible to design an impact measure that is both useful and truly value-agnostic.
However, the value-agnostic criterion does not seem very important in itself. I think the relevant criterion is that designing impact measures should be easier than the general value learning problem. We already have a non-value-agnostic impact measure that plausibly satisfies this criterion: RLSP learns what is effectively an impact measure (the human theta parameter) using zero human input just by examining the starting state. This could also potentially be achieved by choosing an attainable utility set that rewards a broad enough sample of things humans care about, and leaves the rest to generalization. Choosing a good attainable utility set may not be easy but it seems unlikely to be as hard as the general value learning problem.
Butterfly effects. Every action is likely to have large effects that are difficult to predict, e.g. taking a different route to work may result in different people being born. Taken literally, this means that there is no such thing as a low-impact action. Humans get around this by only counting easily predictable effects as impact that they are considered responsible for. If we follow a similar strategy of not penalizing butterfly effects, we might incentivize the agent to deliberately cause butterfly effects. The easiest way around this that I can currently see is restricting the agent's capability to model the effects of its actions, though this has obvious usefulness costs as well.
Chaotic world. Every action, including inaction, is irreversible, and each branch contains different states. While preserving reversibility is impossible in this world, preserving optionality (attainable utility, reachability, etc) seems possible. For example, if the attainable set contains a function that rewards the presence of vases, the action of breaking a vase will make this reward function more difficult to satisfy (even if the states with/without vases are different in every branch). If we solve the problem of designing/learning a good utility set that is not dominated by irrelevant factors, I expect chaotic effects will not be an issue.
If any of the above-mentioned concerns are not overcome, impact measures will fail to distinguish between what humans would consider low-impact and high-impact. Thus, penalizing high-impact actions would come with penalizing low-impact actions as well, which would result in a strong safety-capability tradeoff. I think the most informative direction of research to figure out whether these concerns are a deal-breaker is to scale up impact measures to apply beyond gridworlds, e.g. to Atari games.
Thanks for the detailed list!
AU with random utility functions, which would mostly end up rewarding specific configurations of air molecules.
What does this mean, concretely? And what happens with the survival utility function being the sole member of the attainable set? Does this run into that problem, in your model?
Humans get around this by only counting easily predictable effects as impact that they are considered responsible for.
What makes you think that?
So there's a thing people do when they talk about AUP which I don't understand. They think it's about state, even though I insist it's fundamentally different, and try to explain why (note that AUP in the MDP setting is necessarily over states, because states are the observations). My explanations apparently haven't been very good; in the given conversation, they acknowledge that it's different, but then regress a little while later. I think they might be trying understand the explanation, remain confused, and then subconsciously slip back to their old model. out of everyone I've talked to, I can probably count on my hands the number of people who get this – note that agreeing with specific predictions of mine is different.
Now, it's the author's job to communicate their ideas. When I say "as far as I can tell, few others have internalized how AUP actually works", this doesn't connote "gosh, I can't stand you guys, how could you do this", it's more like "somehow I messed up the explanations; I wonder what key ideas are missing still? How can I fix this?".
my goal with this comment isn't t...
I have a bit of time on my hands, so I thought I might try to answer some of your questions. Of course I can't speak for TurnTrout, and there's a decent chance that I'm confused about some of the things here. But here is how I think about AUP and the points raised in this chain:
I am still confused about what you means by penalizing 'power' and what exactly it is a function of. The way you describe it here sounds like it's a measure of the agent's optimization ability that does not depend on the state at all.
It definitely does depend on the state. If the agent moves to a state where it has taken over the world, that's a huge increase in its ability to achieve arbitrary utility functions, and it would get a large penalty.
I think the claim is more that while the penalty does depend on the state, it's not central to think about the state to understand the major effects of AUP. (As an analogy, if you want to predict whether I'm about to leave my house, it's useful to see whether or not I'm wearing shoes, but if you want to understand why I am or am not about to leave my house, whether I'm wearing shoes is not that relevant -- you'd want to know what my current subgoal or plan is.)
Similarly, with AUP, the claim is that while you can predict what the penalty is going to be by looking at particular states and actions, and the penalty certainly does change with different states/actions, the overall effect of A...
Here's a relevant passage by Rohin (from Alignment Newsletter #49, March 2019):
On the topic of impact measures, I'll repeat what I've said before: I think that it's hard to satisfy the conjunction of three desiderata -- objectivity (no dependence on human values), safety (preventing any catastrophic outcomes) and usefulness (the AI system is still able to do useful things). Impact measures are very clearly aiming for the first two criteria, but usually don't have much to say about the third one. My expectation is that there is a strong tradeoff between the first two criteria and the third one, and impact measures have not dealt with this fact yet, but will have to at some point.
Other relevant writing of mine:
Comment on the desiderata post
But it's true that that quoted passage is the best summary of my current position. Daniel's answer is a good example of an underlying intuition that drives this position.
I’m interested in learning about the intuitions, experience, and facts which inform this pessimism. As such, I’m not interested in making any arguments to the contrary in this post; any pushback I provide in the comments will be with clarification in mind.
I would prefer that you and/or others did push back, as I'm really curious which of the causes/reasons for pessimism actually stand up under such pushback. (See Four Layers of Intellectual Conversation and AI Safety via Debate.) I do appreciate that you prioritize just knowing what the causes/reasons are in the first place and don't want to discourage people from sharing them, so I wonder if there's a way to get both of what we want.
I do plan on pushing back on certain concerns, but I think if I did so now, some of my reasons for believing things would seem weird and complicated-enough-to-be-shaky because of inferential distance. The main pedagogical mistake I made with Towards a New Impact Measure wasn't putting too much in one post, but rather spending too much time on conclusions, telling people what I think happens without helping build in them the intuitions and insights which generate those results. Over the last 8 months, I think I've substantially enriched my model of how agents interact with their environments. I'm interested in seeing how many disagreements melt away when these new insights are properly shared and understood, and what people still disagree with me on. That's why I'm planning on waiting until my upcoming sequence to debate these points.
I am comfortable sharing those concerns which I have specific reason to believe don't hold up. However, I'm wary of dismissing them in a way that doesn't Include those specific reasons. That seems unfair. If you're curious which ones I think these are, feel free to ask me over private message.
Habryka recently wrote (emphasis mine):
I'm interested in learning about the intuitions, experience, and facts which inform this pessimism. As such, I'm not interested in making any arguments to the contrary in this post; any pushback I provide in the comments will be with clarification in mind.
There are two reasons you could believe that "work on impact measures is very unlikely to result in much concrete progress on… core AI Alignment problems". First, you might think that the impact measurement problem is intractable, so work is unlikely to make progress. Second, you might think that even a full solution wouldn't be very useful.
Over the course of 5 minutes by the clock, here are the reasons I generated for pessimism (which I either presently agree with or at least find it reasonable that an intelligent critic would raise the concern on the basis of currently-public reasoning):
1 Paul raised concerns along these lines: