Current thoughts on Paul Christano’s research agenda
This post summarizes my thoughts on Paul Christiano’s agenda in general and ALBA in particular.
(note: at the time of writing, I am not employed at MIRI)
(in general, opinions expressed here are strong and weakly held)
AI alignment research as strategy
Roughly, AI alignment is the problem of using a system of humans and computers to do a good thing. AI alignment research will tend to look at the higher levels of what the system of humans and computers is doing. Thus, it is strategy for doing good things with humans and computers.
“Doing good things with humans and computers” is a broad class. Computing machines have been around a long time, and systems of humans involving rules computed by humans have been around much longer. Looking at AI alignment as strategy will bring in intuitions from domains like history, law, economics, and political philosophy. I think these intuitions are useful for bringing AI strategy into near mode.
Paul Christiano’s agenda as strategy
Paul Christiano’s agenda “goes for the throat” in ways that the other agendas, such as the agent foundations agenda, do not. Thus, it yields an actual strategy, rather than a list of research questions about strategy. I will now analyze Paul Christiano’s research agenda as strategy. Here I caricature Paul Christiano’s strategic assumptions:
There are 2 phases to agentic activity: an influence-maximizing phase, and a value-fulfillment phase.
Any strategy can be decomposed into (a) the expansion strategy (the strategy for the influence-maximizing phase) and (b) the payload (the strategy for the value-fulfillment phase), which can vary independently in principle.
The power of an agent after the influence-maximizing phase is a function of its starting power and the effectiveness of its expansion strategy.
An outcome is good if the vast majority of power after the influence-maximizing phase is owned by agents whose payload is good.
Almost all powerful actors start with good values, i.e. they intend to select a good payload.
Selecting a good payload is not hard (e.g. specifying a social arrangement that would reflect and arrive at good moral conclusions)
Powerful actors can’t coordinate with each other; they must expand independently.
Powerful actors with bad values exist.
Good powerful actors can anticipate and copy bad powerful actors’ expansion strategies while retaining the same payload.
These assumptions are a caricature of Paul’s assumptions in that, though they aren’t completely reflective of Paul’s actual views, they strongly state the background philosophy implied by Paul’s research agenda.
From these assumptions, it is possible to derive that a decent strategy for good actors is to anticipate bad actors’ expansion strategies and copy them while retaining the same payload. This corresponds to Paul’s general approach: look at a proposed system that would do a bad thing, then create an equally powerful system that would instead do a good thing.
Here are some thoughts on these assumptions:
I think the “expansion/value-fulfillment” split is broadly correct for highly-programmable, fully-automated systems and broadly incorrect for systems involving humans. E.g. humans can be corrupted by power.
I think powerful actors can coordinate with each other. In some sense, altruism derives from coordination strategies, so altruists ought to be more able to cooperate than non-altruists.
I think that thinking of power as unidimensional is highly misleading. For example, it is possible to distinguish “amount of resources” from “knowledge about the world”. Those with more knowledge can be influential even without resources, and even without gaining many resources along the way. (Gaining resources can worsen epistemics by increasing the incentives to deceive the person with resources)
I think most powerful interests today are not “good” in the sense of “would probably select a good payload”, mostly because most powerful interests do not have good epistemics about such abstract ideas. Thus, good actors will need to use asymmetric strategies, not just symmetric ones. (Due to our current state of philosophical confusion, it seems somewhat likely that no one is currently a good actor in this sense)
ALBA competes with adversaries who use only learning
Paul Christiano has acknowledged that ALBA only works for aligning systems that work through learning, and does not work for aligning systems that use hard-to-learn forms of cognition (e.g. search, logic). Roughly, ALBA will copy the part of a system’s strategy that is explainable by the system’s learning, i.e. that is explaining by the system modelling the world as a set of nested feedback loops. Object recognition is fine, because there’s feedback for that; philosophy is not fine, because there isn’t feedback for that. (Humans can provide feedback about philosophy, but human philosophy is not fully explained by this feedback, e.g. the human giving the feedback needs their own strategy).
The world isn’t a set of nested feedback loops. Agents don’t just learn, they reason. The philosophy of empiricism is more appropriate for an investor, who tries to pick up on signs of a successful project based on limited information and reasoning, than for an entrepreneur, who engages in object-level reasoning in order to actually run a successful project.
To be more concrete, there are multiple ways humans could believe that some AI system would be quite powerful without this belief being justified solely on the basis that the AI system does learning:
Perhaps the AI system is organized very similarly to a human brain, and humans have reason to believe that it will behave similarly, even though they haven’t fully explained the brain’s intelligence as a form of learning.
Perhaps the AI system is produced through natural selection in a single large world that is difficult to split into smaller worlds without losing efficiency. (If the world can be split, then the different worlds could possibly be treated as “different hypotheses” productively)
Perhaps humans understand reasoning better and can create a system that reasons according to humans’ understanding of reasoning.
None of these are the kind of “ML technology” that ALBA uses, so they could not be analyzed effectively enough that ALBA could produce an equivalent.
The last time I talked with Paul, he was aware of these problems. My recollection of his thoughts on these problems are:
He thinks that solving alignment for learning systems is more urgent and the correct first step for solving alignment in general.
He thinks that it might be possible to see processes such as evolution as a combination of selection and deliberation, and that this analysis might yield enough analysis for producing aligned versions.
ALBA is not amenable to formal analysis
Some parts of ALBA can’t be formally analyzed very well with our current level of philosophical sophistication. For example:
Analyzing capability amplification requires modeling humans as computationally-bounded agents. There is not currently a model of agents under which a scheme like HCH would work, and even if there were, showing that humans are such an agent would be quite a challenge.
Analyzing informed oversight similarly requires modeling humans as computationally-bounded agents.
Analyzing red teams likewise requires modeling the capabilities of the red team versus the learning system.
In general, formally analyzing “power” would require formally analyzing decision theory and logical uncertainty.
Notably, the kind of agents for which analyzing capability amplification makes sense are not the kind of agents that ALBA aligns. ALBA aligns learning systems, and humans’ reasoning over time is deliberation rather than learning. I think this is the root of a lot of the difficulty of formally analyzing ALBA.
Plausibility arguments are hard to act on
ALBA is based on plausibility arguments: it’s plausible that capability amplification works, it’s plausible that informed oversight works, and it’s plausible that red teams work. These questions are actually pretty hard to research.
Recently I have updated towards a generalization of this: plausibility arguments are very often hard to act on, since the belief that “X is plausible” does not usually come attached to a strong model implying X is probably true, and these strong models are necessary for generating proofs. (Sometimes, plausibility arguments are sufficient to act on, especially in “easy” domains such as mathematics and puzzle games where there exist search procedures for proving or disproving X. But AI alignment doesn’t seem like one of these domains to me.)
Agreement with many of Paul’s intuitions
This post so far gives the impression that I strongly disagree with Paul. I do disagree with Paul, but I also strongly agree with many of his intuitions:
I think that “going for the throat” by trying to solve the whole alignment problem is quite useful, though other approaches should also be used (since “going for the throat” will tend to produce intractable research problems).
I think that corrigibility as Paul understands it will turn out to be an important tool for analyzing human-computer systems.
I think that splitting an AI system into an expansion strategy and a payload is sometimes useful, and in particular is usually more illuminating than splitting it into beliefs and values.
I think that thinking about the big strategic picture and the limiting behavior of a system, as Paul does, is useful.
I think that Paul Christiano’s research is quite helpful for reasoning about learning and control systems; for example, counterfactual oversight seems very useful for handling epistemic problems discussed in He Who Pays The Piper Must Know The Tune.
Conclusion
I have presented my thoughts on ALBA. ALBA has significantly improved my conceptual understanding of AI alignment, but is seriously incomplete and difficult to make further theoretical progress on. I don’t know how to make much additional theoretical progress on the alignment problem at this point, but perhaps taking some steps back from specific approaches and doing original seeing on alignment would yield new directions.
I mostly agree with this post’s characterization of my position.
Places where I disagree with your characterization of my view:
I don’t assume that powerful actors can’t coordinate, and I don’t think that assumption is necessary. I would describe the situation as: over the course of time influence will necessarily shift sometimes due to forces we endorse—like deliberation or reconciliation—and sometimes due to forces we don’t endorse—like compatibility with an uncontrolled expansion strategy. Even if powerful actors can form perfectly-coordinated coalitions, a “weak” actor positioned to benefit from competitive expansion would simply decline to participate in that coalition unless offered extremely generous terms. I don’t see how the situation changes unless the strong actors use force. I do think that’s reasonably likely, I would more describe alignment as a first line of defense or a complement to approaches like regulation. I generally agree that good coordination can substitute for technical weakness.
I don’t think I rely on or even implicitly use a unidimensional model of power. I do use a concept like “total wealth” or “total influence,” which seems almost but not quite tautologically well-defined (as the output of a competitive/bargaining dynamic) and in particular is compatible with knowledge vs. resources vs. whatever. Being “competitive” seems to make sense in very complex worlds, when I say something like “win in a fistfight” I mean to quantify over all possible fistfights (including science, economic competition, persuasion, war, etc. etc.).
I have strong intuitions about my approach being workable, and either the approach will succeed or I at least will feel that I have learned something substantial. I expect many more pivots and adjustments to be necessary, but don’t expect to get stuck with plausibility arguments that are nearly as weak as the current arguments.
Place where I disagree with your view:
I agree that there are many drivers of AI other than learning. However, I think that learning is (a) currently the dominant component of powerful AI and so both more urgent and easier to study, (b) poses a much harder safety problems than other AI techniques under discussion, and (c) appears to be the “hard part” of analyzing procedures like evolution, fine-tuning brain-inspired architectures, or analyzing reasoning (it’s where I run into a wall when trying to analyze these other alternatives).
I think that all of capability amplification, informed oversight, and red teams / adversarial training are amenable to theoretical analysis with realistic amounts of philosophical progress. For example, I think that it will be possible to analyze these schemes using only abstractions like optimization power, without digging into models of bounded rationality at all. I may have understated my optimism on this (for capability amplification in particular) in our last discussion—I do believe that we won’t have a formal argument, but I think we should aim for an argument that is based on plausible empirical assumptions plus very good evidence for those assumptions.
Altruism as in “concern for future generations” does not fall out of coordination strategies, and seems more like a spandrel to me. But I do agree that many parts of altruism are more like coordination and this gives some prima facie reason for optimism about getting to some Pareto frontier.
Take-aways that I agree with:
We will need to have a better understanding of deliberation in order to be confident in any alignment scheme. (I prefer a more surgical approach than most MIRI folk, trying to figure exactly what we need to know rather than trying to have an expansive understanding of what good reasoning looks like.)
It is valuable for people to step back from particular approaches to alignment and to try to form a clearer understanding of the problem, explore completely new approaches, etc.
I can see two kinds of understanding of deliberation that could help us achieve confidence in alignment schemes:
white-box understanding, where we understand what good reasoning is, in each of philosophical/mathematical/algorithmic levels, i.e., we can formally define what ideal reasoning is, have a clear understanding of why the formal definition is correct/normative and why some specific algorithm is a good approximation of the mathematical ideal.
black box understanding, where we define “ideal reasoning” as the distribution of outcomes induced by placing a group of humans in an ideally safe and supportive environment and given a question to deliberate over, and show that a specific practical implementation of deliberation (e.g., meta-execution) induces a distribution that is acceptably similar to the ideal.
(Note that I’m using the same white-box/black-box terminology as here, but the meaning is a bit different since I’m now applying the terms to understandings of deliberation as opposed to implementations of deliberation.)
The problem with 2, which I see you as implicitly advocating (since I don’t see how else you hope to eventually be confident in your alignment scheme), is that I have a low prior for any specific implementation of deliberation (such as meta-execution) producing distributions of outcomes acceptably close to the ideal (unless it’s just a very close direct approximation of the ideal process like using a group of highly accurate uploads), and I don’t currently see what kind of arguments or evidence we can hope to produce in a relevant timeframe that would make me update enough to become confident. (Aside from something like achieving a white-box understanding of deliberation and then concluding that both the black-box definition of “ideal reasoning” and the actual implementation would be able to approximate the white-box definition of “ideal reasoning”, but presumably that’s not what you have in mind.)
Perhaps you think empirical work would help, but even if you’re able to gather a lot of data on what black-box ideal reasoning eventually produces (which you can’t for certain types of reasoning, e.g., philosophical reasoning) and are able to compare that with the AI alignment scheme, how would you rule out the possibility of edge cases where the two don’t match?
Another possibility is that you think black-box ideal reasoning would first decide to follow a set of rules and procedures before doing any further deliberation, and that would make it easier for an AI alignment scheme to approximate the ideal. But A) the group of humans would likely spend a lot of time exploring different alternative for how to do further deliberation, and B) whatever rules/procedures they end up adopting would likely include ways to back out of those rules and/or adopt new rules. For 2, you would need to predict with justifiably high confidence what rules/procedures they eventually converge upon (if they in fact converge instead of, e.g., diverge depending on what question they face or which humans we start with), and again I don’t see how you hope to do that, within the time we likely have available.
An aligned AI doesn’t need to share human preferences-on-reflection. It just needs to (a) be competent, and (b) help humans remain in control of the AI while carrying out whatever reflective process they prefer (including exploration of different approaches, reconciliation of different perspectives, etc.).
So all I’m hoping is to show something about how deliberation can (a) be smarter, while (b) avoiding introducing “bad” (incorrigible?) optimization. I don’t think this requires either your #1 or #2.
If you try to formalize (a) and (b), what does that look like, and how would you reach the conclusion that an AI actually has (a) and (b)? My #1 and #2 are things I can come up with when I try to think how we might come to have justified confidence that an AI has good reasoning, and I’m not seeing what other solutions pop up if we say an aligned AI doesn’t need to share human preferences-on-reflection but only needs to be competent and help humans remain in control.
Competence can be an empirical claim, so (a) seems much more straightforward.
Is there some sense in which “argue that a process is normatively correct” is more of a solution than “argue that a process doesn’t optimize for something ‘bad’”? I agree that both of of the properties are hard to formalize or achieve, the second one currently looks easier to me (and may even be a subproblem of the first one—e.g. my current best guess is that good reasoners need a cognitive immune system).
Once again I’m having trouble seeing something that you think is straightforward. If an AI can’t determine my values-upon-reflection, that seems like a kind of incompetence. If it can’t do that, it seems likely there are other things it can’t do. Perhaps you can define “competence” in a way that excludes that class of things and argue that’s good enough, but I’m not sure how you’d do that.
I think we might eventually be able to argue that a process is normatively correct by understanding it at each of philosophical/mathematical/algorithmic levels, but that kind of white-box understanding does not seem possible if your process incorporates an opaque object (e.g., a machine-learned imitation of human behavior), so I think the best you can hope to achieve in that case is a black-box understanding where you show that your process induces the same (or close enough) distribution of outcomes as a group of humans in an ideal environment.
If your “argue that a process doesn’t optimize for something ‘bad’” is meant to be analogous to my white-box understanding, it seems similarly inapplicable due to presence of the opaque object in your process. If it’s meant to be analogous to my black-box understanding, I don’t see what the analogy is. In other words, what are you hoping to show instead of “induces the same (or close enough) distribution of outcomes as a group of humans in an ideal environment”?
Suppose I use normatively correct reasoning, but I also use a toaster designed by a normal engineer. The engineer is less capable than I am in every respect, and I watched them design and build the toaster to verify that they didn’t do anything tricky. Then I verified that the toaster does seem to toast toast. But I have no philosophical or mathematical understanding of the toaster-design-process. Your claim seems to be that there are no rational grounds for accepting use of the toaster, other than to argue that accepting it doesn’t change the distribution of outcomes (which probably isn’t true, since it e.g. slightly changes the relative influence of different internal drives by changing what food I eat). Is that right?
What if they designed a SAT solver for me? Or wrote a relativity textbook? Do I need to be sure that nothing like that happens in my deliberative process, in order to have confidence in it?
If those cases don’t seem analogous, can you be more clear about what you mean by “opaque,” or what quantitative factors make an opaque object problematic? So far your argument doesn’t seem rely on any properties of the opaque object at all.
(This case isn’t especially analogous to the deliberative process I’m interested in. I’m bringing it up because I don’t think I yet understand your intuitive dichotomy.)
When you wrote “suppose I use normatively correct reasoning” did you mean suppose you, Paul, use normatively correct reasoning, or suppose you are an AI who uses normatively correct reasoning? I’ll assume the latter for now.
Generally, the AI would use its current reasoning process to decide whether or not to incorporate new objects into itself. I’m not sure what that reasoning process will do exactly, but presumably it would involve something like looking for and considering proofs/arguments/evidence to the effect that incorporating the new object in some specified way will allow it to retain its normatively correct status, or the distribution of outcomes will be sufficiently unchanged.
If an AI uses normatively correct reasoning, the toaster shouldn’t change the distribution of outcomes, since letting food influence its reasoning process is obviously not a normative thing to do, and it should be easy to show that there is no influence. For the SAT solver, the AI should be able to argue that it’s safe to use it for certain purposes, because it can verify the answer that the solver gives, and for the relativity textbook, it may be able to directly verify that the textbook doesn’t contain anything that can manipulate or bias its outputs.
I guess by “opaque” I meant a complex object that wasn’t designed to be easily reasoned about, so it’s very hard to determine whether it has a given property that would be relevant to showing that it can be used safely as part of an AI’s reasoning process. For example a typical microprocessor is an opaque object because it may come with hard to detect flaws and backdoors, whereas a microprocessor designed to be provably correct would be a transparent object.
(Does that help?)
Suppose that I, Paul, use a toaster or SAT solver or math textbook.
I’m happy to drop the normatively correct reasoning assumption if the counterfactual begs the question. The important points are:
I’m happy trusting future Paul’s reasoning (in particular I do not consider it a top altruistic priority to find a way to avoid trusting future Paul’s reasoning)
That remains true even though future Paul would happily use an opaque toaster or textbook (under the conditions described).
I’m not convinced that any of your arguments would be sufficient to trust a toaster / textbook / SAT solver:
Having new memories will by default change the output of deliberation, won’t it?
Satisfying instances produced by an arbitrarily powerful adversary don’t seem safe for anything.
I don’t see how this would fit into your framework, without expanding it far enough that it could contain the kind of argument I’m gesturing at (by taking bad === “manipulating or biasing its outputs.”)
If we’re talking about you, Paul, then what’s different is that since you don’t have a good understanding of what normatively correct reasoning is, you can only use black-box type reasoning to conclude that certain things are safe to do. We’d happily use the opaque toaster or textbook because we have fairly strong empirical evidence that doing so doesn’t change the distribution of outcomes much. Using a toaster might change a particular outcome vs not using it, but there seems to be enough stochasticity in a human deliberation process that it wouldn’t make a significant difference to the overall distribution of outcomes. With a textbook, you reason that with enough time you’d reproduce its contents yourself, and whatever actual differences between reading the textbook and figuring out relativity by yourself is again lost in the overall noise of the deliberative process. (We have fairly strong empirical evidence that reading such a textbook written by another human is unlikely to derail our deliberative process in a way that’s not eventually recoverable.)
One reply to this might be that we can hope to gather an amount of empirical evidence about meta-execution that would be comparable to the evidence we have about toasters and textbooks. I guess my concern there is that we’ll need much stronger assurances if we’re going to face other superintelligent AIs in our environment. For example that textbook might contain subtle mistakes that cause you to reason incorrectly about certain questions (analogous to edge case questions where your meta-execution would give significantly different answers than your reflective equilibrium), but there is no one in your current environment who can exploit such errors.
ETA: Another reason to be worried is that, compared to humans using things produced by other humans, it seems reasonable to suspect (have a high prior) that meta-execution’s long run safety can’t be extrapolated well from what it does in the short term, since meta-execution is explicitly built out of a component that emphasizes imitation of short-term human behavior while throwing away internal changes that might be very relevant to long-run outcomes. (Again this may be missing your point about not needing to reproduce values-upon-reflection but I just don’t understand how your alternative approach to understanding deliberation would work if you tried to formalize it.)
Not sure if this is still relevant to the current interpretation of your question, but couldn’t you use it to safely break encryption schemes, at least?
I want to add that I think meta-execution, in particular, will have problems with deliberation for the same reason that it will have problems with learning: when you hear an argument or explanation (like when you learn), your mind is changed in ways that are hard or impossible to articulate. If every 10 minutes (or 1 day, or what have you) you throw away the part of what your brain does that it can’t write down, it seems highly plausible that in many cases you won’t be able to reproduce what the brain does over a longer period of time, especially if you’re trying to match its natural trajectory, as opposed to trying to hit some objectively measurable benchmark.
Given that ALBA was not meant to be a realistic aligned AI design in and of itself, but just a way to get insights into how to build a realistic aligned AI (which I hadn’t entirely understood until now), I wonder if it makes sense to try to nail down all the details and arguments for it before checking to see if you generated any such insights. If we assume that aligned learning roughly looks like ALBA, what does that tell you about what a more realistic aligned AI looks like? It seems worth asking this, in case you, for example, spend a lot of time figuring out exactly how capability amplification could work, and then it ends up that capability amplification isn’t even used in the final aligned AI design, or in case designing aligned AI out of individual AI components doesn’t actually give you much insight into how to design more realistic aligned AI.
Thank you for writing this. I’m trying to better understand Paul’s ideas, and it really helps to see an explanation from a different perspective. Also, I was thinking of publicly complaining that I know at least four people who have objections to Paul’s approach that they haven’t published anywhere. Now that’s down to three. :)
I wonder if you can help answer some questions for me. (I’m directing these at Paul too, but since he’s very busy I can’t always expect an answer.)
Why does Paul think that learning needs to be “aligned” as opposed to just well-understood and well-behaved, so that it can be safely used as part of a larger aligned AI design that includes search, logic, etc.? He seems to be trying to design an entire aligned AI out of “learning”, which makes it seem like his approach is an alternative to MIRI’s (Daniel Dewey said this recently at EA Forums for example), while at the same time saying “But we can and should try to do the same for other AI components; I understand MIRI’s agent foundations agenda as (mostly) addressing the alignment of these other elements.” If he actually thinks that his approach and MIRI’s are complements, why didn’t he correct Daniel? I’m pretty confused here.
ETA: I found a partial answer to the above here. To express my understanding of it, Paul is trying to build an aligned AI out of only learning because that seems easier than building a realistic aligned AI and may give him insights into how to do the latter. If he interprets MIRI as doing the analogous thing starting with other AI components (as he seems to according to the quote in the above paragraph), then he surely ought to view the two approaches as complementary, which makes it a bigger puzzle why he didn’t contradict Daniel when Daniel said “if an approach along these lines is successful, it doesn’t seem to me that much room would be left for HRAD to help on the margin”. (Maybe he didn’t read that part, or his interpretation of what MIRI is doing has changed?)
If Paul does not think ALBA is a realistic design of an entire aligned AI (since it doesn’t include search/logic/etc.) what might a realistic design look like, roughly?
Why does Paul think learning “poses much harder safety problems than other AI techniques under discussion”?
Paul is beginning to do empirical work on capability amplification (as he told me recently via email). Do you think that’s a good alternative to trying to make further theoretical progress?
I mostly think it should be benign / corrigible / something like that. I think you’d need something like that whether you want to apply learning directly or to apply it as part of a larger system.
You can definitely make an entire AI out of learning alone (evolution / model-free RL), and I think that’s currently the single most likely possibility even though it’s not particularly likely.
The alternative design would integrate whatever other useful techniques are turned up by the community, which will depend on what those techniques are. One possibility is search/planning. This can be integrated in a straightforward way into ALBA, I think the main obstacle is security amplification which needs to work for ALBA anyway and is closely related to empirical work on capability amplification. On the logic side it’s harder to say what a useful technique would look like other than “run your agent for a while,” which you can also do with ALBA (though it requires something like these ideas).
My hope is to have safe and safely composable versions of each important AI ingredient. I would caricature the implicit MIRI view as “learning will lead to doom, so we need to develop an alternative approach that isn’t doomed,” which is a substitute in the sense that it’s also trying to route around the apparent doomedness of learning but in a quite different way.
Thanks, so to paraphrase your current position, you think once we have aligned learning it doesn’t seem as hard to integrate other AI components into the design, so aligning learning seems to be the hardest part. MIRI’s work might help with aligning other AI components and integrating them into something like ALBA, but you don’t see that as very hard anyway, so it perhaps has more value as a substitute than a complement. Is that about right?
I don’t understand ALBA well enough to easily see extensions to the idea that are obvious to you, and I’m guessing others may be in a similar situation. (I’m guessing Jessica didn’t see it for example, or she wouldn’t have said “ALBA competes with adversaries who use only learning” without noting that there’s a straightforward extension that does more.) Can you write a post about this? (Or someone else please jump in if you do see what the “straightforward way” is.)