AI Alignment Podcast: An Overview of Technical AI Alignment with Rohin Shah (Part 1)

The space of AI alignment research is highly dynamic, and it’s often difficult to get a bird’s eye view of the landscape. This podcast is the first of two parts attempting to partially remedy this by providing an overview of the organizations participating in technical AI research, their specific research directions, and how these approaches all come together to make up the state of technical AI alignment efforts. In this first part, Rohin moves sequentially through the technical research organizations in this space and carves through the field by its varying research philosophies. We also dive into the specifics of many different approaches to AI safety, explore where they disagree, discuss what properties varying approaches attempt to develop/preserve, and hear Rohin’s take on these different approaches.

You can take a short (3 minute) survey to share your feedback about the podcast here.

In this podcast, Lucas spoke with Rohin Shah. Rohin is a 5th year PhD student at UC Berkeley with the Center for Human-Compatible AI, working with Anca Dragan, Pieter Abbeel and Stuart Russell. Every week, he collects and summarizes recent progress relevant to AI alignment in the Alignment Newsletter

We hope that you will continue to join in the conversations by following us or subscribing to our podcasts on Youtube, SoundCloud, iTunes, Google Play, Stitcher, or your preferred podcast site/application. You can find all the AI Alignment Podcasts here.

Topics discussed in this episode include:

  • The perspectives of CHAI, MIRI, OpenAI, DeepMind, FHI, and others
  • Where and why they disagree on technical alignment
  • The kinds of properties and features we are trying to ensure in our AI systems
  • What Rohin is excited and optimistic about
  • Rohin’s recommended reading and advice for improving at AI alignment research

Lucas: Hey everyone, welcome back to the AI Alignment podcast. I’m Lucas Perry, and today we’ll be speaking with Rohin Shah. This episode is the first episode of two parts that both seek to provide an overview of the state of AI alignment. In this episode, we cover technical research organizations in the space of AI alignment, their research methodologies and philosophies, how these all come together on our path to beneficial AGI, and Rohin’s take on the state of the field.

As a general bit of announcement, I would love for this podcast to be particularly useful and informative for its listeners, so I’ve gone ahead and drafted a short survey to get a better sense of what can be improved. You can find a link to that survey in the description of wherever you might find this podcast, or on the page for this podcast on the FLI website.

Many of you will already be familiar with Rohin, he is a fourth year PhD student in Computer Science at UC Berkeley with the Center For Human-Compatible AI, working with Anca Dragan, Pieter Abbeel, and Stuart Russell. Every week, he collects and summarizes recent progress relevant to AI alignment in the Alignment Newsletter. And so, without further ado, I give you Rohin Shah.

Thanks so much for coming on the podcast, Rohin, it’s really a pleasure to have you.

Rohin: Thanks so much for having me on again, I’m excited to be back.

Lucas: Yeah, long time, no see since Puerto Rico Beneficial AGI. And so speaking of Beneficial AGI, you gave quite a good talk there which summarized technical alignment methodologies approaches and broad views, at this time; and that is the subject of this podcast today.

People can go and find that video on YouTube, and I suggest that you watch that; that should be coming out on the FLI YouTube channel in the coming weeks. But for right now, we’re going to be going in more depth, and with more granularity into a lot of these different technical approaches.

So, just to start off, it would be good if you could contextualize this list of technical approaches to AI alignment that we’re going to get into within the different organizations that they exist at, and the different philosophies and approaches that exist at these varying organizations.

Rohin: Okay, so disclaimer, I don’t know all of the organizations that well. I know that people tend to fit CHAI in a particular mold, for example; CHAI’s the place that I work at. And I mostly disagree with that being the mold for CHAI, so probably anything I say about other organizations is also going to be somewhat wrong; but I’ll give it a shot anyway.

So I guess I’ll start with CHAI. And I think our public output mostly comes from this perspective of how do we get AI systems to do what we want? So this is focusing on the alignment problem, how do we actually point them towards a goal that we actually want, align them with our values. Not everyone at CHAI takes this perspective, but I think that’s the one most commonly associated with us and it’s probably the perspective on which we publish the most. It’s also the perspective I, usually, but not always, take.

MIRI, on the other hand, takes a perspective of, “We don’t even know what’s going on with intelligence. Let’s try and figure out what we even mean by intelligence, what it means for there to be a super-intelligent AI system, what would it even do or how would we even understand it; can we have a theory of what all of this means? We’re confused, let’s be less confused, once we’re less confused, then we can think about how to actually get AI systems to do good things.” That’s one of the perspectives they take.

Another perspective they take is that there’s a particular problem with AI safety, which is that, “Even if we knew what goals we wanted to put into an AI system, we don’t know how to actually build an AI system that would, reliably, pursue those goals as opposed to something else.” That problem, even if you know what you want to do, how do you get an AI system to do it, is a problem that they focus on. And the difference from the thing I associated with CHAI before is that, with the CHAI perspective, you’re interested both in how do you get the AI system to actually pursue the goal that you want, but also how do you figure out what goal that you want, or what is the goal that you want. Though, I think most of the work so far has been on supposing you know the goal, how do you get your AI system to properly pursue it?

I think DeepMind safety came, at least, is pretty split across many different ways of looking at the problem. I think Jan Leike, for example, has done a lot of work on reward modeling, and this sort of fits in with the how do we get our AI systems be focused on the right task, the right goal. Whereas Vika has done a lot of work on side effects or impact measures. I don’t know if Vika would say this, but the way I interpret it how do we impose a constraint upon the AI system such that it never does anything catastrophic? But it’s not trying to get the AI system to do what we want, just not do what we don’t want, or what we think would be catastrophically bad.

OpenAI safety also seems to be, okay how do we get deep enforcement learning to do good things, to do what we want, to be a bit more robust? Then there’s also the iterated amplification debate factored cognition area of research, which is more along the lines of, can we write down a system that could, plausibly, lead to us building an aligned AGI or aligned powerful AI system?

FHI, no coherent direction, that’s all of FHI. Eric Drexler is also trying to understand how AI will develop it in the future is somewhat very different from what MIRI’s doing, but the same general theme of trying to figure out what is going on. So he just recently published a long technical report on comprehensive AI services, which is the general worldview for predicting what AI development will look like in the future. If we believed that that was, in fact, the way AI would happen, we would probably change what we work on from the technical safety point of view.

And Owain Evans does a lot of stuff, so maybe I’m just not going to try to categorize him. And then Stuart Armstrong works on this, “Okay, how do we get value learning to work such that we actually infer a utility function that we would be happy for an AGI system to optimize, or a super-intelligent AI system to optimize?”

And then Ought works on factory cognition, so it’s very adjacent to be iterated amplification and debate research agendas. Then there’s a few individual researchers, scattered, for example, Toronto, Montreal, and AMU and EPFL, maybe I won’t get into all of them because, yeah, that’s a lot; but we can delve into that later.

Lucas: Maybe a more helpful approach, then, would be if you could start by demystifying some of the MiRI stuff a little bit; which may seem most unusual.

Rohin: I guess, strategically, the point would be that you’re trying to build this AI system that’s going to be, hopefully, at some point in the future vastly more intelligent than humans, because we want them to help us colonize the universe or something like that, and lead to lots and lots of technological progress, etc., etc.

But this, basically, means that humans will not be in control unless we very, very specifically arrange it such that we are in control; we have to thread the needle, perfectly, in order to get this to work out. In the same way that, by default you, would expect that the most intelligent creatures, beings are the ones that are going to decide what happens. And so we really need to make sure and, also it’s probably hard to ensure, that these vastly more intelligent beings are actually doing what we want.

Given that, it seems like what we want is a good theory that allows us to understand and predict what these AI systems are going to do. Maybe not in the fine nitty, gritty details, because if we could predict what they would do, then we could do it ourselves and be just as intelligent as they are. But, at least, in broad strokes what sorts of universes are they going to create?

But given that they can apply so much more intelligence that we can, we need our guarantees to be really, really strong; like almost proof level. Maybe actual proofs are a little too much to expect, but we want to get as close to it as possible. Now, if we want to do something like that, we need a theory of intelligence; we can’t just sort of do a bunch of experiments, look at the results, and then try to extrapolate from there. Extrapolation does not give you the level of confidence that we would need for a problem this difficult.

And so rather, they would like to instead understand intelligence deeply, deconfuse themselves about it. Once you understand how intelligence works at a theoretical level, then you can start applying that theory to actual AI systems and seeing how they approximate the theory, or make predictions about what different AI systems will do. And, hopefully, then we could say, “Yeah, this system does look like it’s going to be very powerful as approximating this particular idea, this particular part of theory of intelligence. And we can see that with this particular theory of intelligence, we can align it with humans somehow, and you’d expect that this was going to work out.” Something like that.

Now, that sounded kind of dumb even to me as I was saying it, but that’s because we don’t have the theory yet; it’s very fun to speculate how you would use the theory before you actually have the theory. So that’s the reason they’re doing this, the actual thing that they’re focusing on is centered around problems of embedded agency. And I should say this is one of their, I think, two main strands of research, the other stand of research, I do not know anything about because they have not published anything about it.

But one of their strands of research is about embedded agency. And here the main point is that in the real world, any agent, any AI system, or a human is a part of their environment. They are smaller than the environment and the distinction between agent and environment is not crisp. Maybe I think of my body as being part of me but, I don’t know, to some extent, my laptop is also an extension of my agency; there’s a lot of stuff I can do with it.

Or, on the other hand, you could think maybe my arms and limbs aren’t actually a part of me, I could maybe get myself uploaded at some point in the future, and then I will no longer have arms or legs; but in some sense I am still me, I’m still an agent. So, this distinction is not actually crisp, and we always pretend that it is in AI, so far. And it turns out that once you stop making this crisp distinction and start allowing the boundary to be fuzzy, there are a lot of weird, interesting problems that show up and we don’t know how to deal with any of them, even in theory, so that’s what they focused on.

Lucas: And can you unpack, given that AI researchers control of the input/output channels for AI systems, why is it that there is this fuzziness? It seems like you could extrapolate away the fuzziness given that there are these sort of rigid and selected IO channels.

Rohin: Yeah, I agree that seems like the right thing for today’s AI systems; but I don’t know. If I think about, “Okay, this AGI is a generally intelligent AI system.” I kind of expect it to recognize that when we feed it inputs which, let’s say, we’re imagining a money maximizing AI system that’s taking in inputs like stock prices, and it outputs which stocks to buy. And maybe it can also read the news that lets it get newspaper articles in order to make better decisions about which stocks to buy.

At some point, I expect this AI system to read about AI and humans, and realize that, hey, it must be an AI system, it must be getting inputs and outputs. Its reward function must be to make this particular number in a bank account be as high as possible and then once it realizes this, there’s this part of the world, which is this number in the bank account, or it could be this particular value, this particular memory block in its own CPU, and its goal is now make that number as high as possible.

In some sense, it’s now modifying itself, especially if you’re thinking of the memory block inside the CPU. If it goes and edits that and sets that to a million, a billion, the highest number possible in that memory block, then it seems like it has, in some sense, done some self editing; it’s changed the agent part of it. It could also go and be like, “Okay actually what I care about is this particular award function box is supposed to output as high a number as possible. So what if I go and change my input channels such that it feeds me things that caused me to believe that I’ve made tons and tons of profit?” So this is a delusion backs consideration.

While it is true that I don’t see a clear, concrete way that an AI system ends up doing this, it does feel like an intelligent system should be capable of this sort of reasoning, even if it initially had these sort of fixed inputs and outputs. The idea here is that its outputs can be used to affect the inputs or future outputs.

Lucas: Right, so I think that that point is the clearest summation of this; it can affect its own inputs and outputs later. If you take human beings who are, by definition, human level intelligences we have, say, in a classic computer science sense if you thought of us, you’d say we strictly have five input channels: hearing seeing, touch, smell, etc.

Human beings have a fixed number of input/output channels but, obviously, human beings are capable of self modifying on those. And our agency is sort of squishy and dynamic in ways that would be very unpredictable, and I think that that unpredictability and the sort of almost seeming ephemerality of being an agent seems to be the crux of a lot of the problem.

Rohin: I agree that that’s a good intuition pump, I’m not sure that I agree it’s the crux. The crux, to me, it feels more like you specify some sort of behavior that you want which, in this case, was make a lot of money or make this number in a bank account go higher, or make this memory cell go as high as possible.

And when you were thinking about the specification, you assumed that the inputs and outputs fell within some strict parameters, like the inputs are always going to be news articles that are real and produced by human journalists, as opposed to a fake news article that was created by the AI in order to convince the reward function that actually it’s made a lot of money. And then the problem is that since the AI’s outputs can affect the inputs, the AI could cause the inputs to go outside of the space of possibilities that you imagine the inputs could be in. And this then allows the AI to game the specification that you had for it.

Lucas: Right. So, all the parts which constitute some AI system are all, potentially, modified by other parts. And so you have something that is fundamentally and completely dynamic, which you’re trying to make predictions about, but whose future structure is potentially very different and hard to predict based off of the current structure?

Rohin: Yeah, basically.

Lucas: And that in order to get past this we must, again, tunnel down on this decision theoretic and rational agency type issues at the bottom of intelligence to sort of have a more fundamental theory, which can be applied to these highly dynamic and difficult to understand situations?

Rohin: Yeah, I think the MIRI perspective is something like that. And in particular, it would be like trying to find a theory that allows you to put in something that stays stable even while the system, itself, is very dynamic.

Lucas: Right, even while your system, whose parts are all completely dynamic and able to be changed by other parts, how do you maintain a degree of alignment amongst that?

Rohin: One answer to this is give the AI a utility function. There is a utility function that’s explicitly trying to maximize that and in that case, it probably has an incentive in order to keep that to protect that the utility function, because if it gets changed, well then it’s not going to maximize that utility function anymore, it’ll maximize something else which will lead to worse behavior by the likes of the original utility function. That’s a thing that you could hope to do with a better theory of intelligence is, how do you create a utility function in an AI system stays stable, even as everything else is dynamically changing?

Lucas: Right, and without even getting into the issues of implementing one single stable utility function.

Rohin: Well, I think they’re looking into those issues. So, for example, Vingean Reflection is a problem that is entirely about how you create better, more improved version of yourself without having any value drift, or a change to the utility function.

Lucas: Is your utility function not self-modifying?

Rohin: So in theory, it could be. The hook would be that we could design an AI system that does not self-modify its utility function under almost all circumstances. Because if you change your utility function, then you’re going to start maximizing that new utility function which, by the original utility function’s evaluation, is worse. If I told you, “Lucas, you have got to go fetch coffee.” That’s the only thing in life you’re concerned about. You must take whatever actions are necessary in order to get the coffee.

And then someone goes like, “Hey Lucas, I’m going to change your utility function so that you want to fetch tea instead.” And then all of your decision making is going to be in service of getting tea. You would probably say, “No, don’t do that, I want to fetch coffee right now. If you change my utility function for being ‘fetch tea’, then I’m going to fetch tea, which is bad because I want to fetch coffee.” And so, hopefully, you don’t change your utility function because of this effect.

Lucas: Right. But isn’t this where corrigibility comes in, and where we admit that as we sort of understand more about the world and our own values, we want to be able to update utility functions?

Rohin: Yeah, so that is a different perspective; I’m not trying to describe that perspective right now. It’s a perspective for how you could get something stable in an AI system. And I associate it most with Eliezer, though I’m not actually sure if he holds this opinion.

Lucas: Okay, so I think this was very helpful for the MIRI case. So why don’t we go ahead and zoom in, I think, a bit on CHAI, which is the Center For Human-Compatible AI.

Rohin: So I think rather than talking about CHAI, I’m going to talk about the general field of trying to get AI systems do what we want; a lot of people at CHAI work on that but not everyone. And also a lot of people outside of CHAI work on that, because that seems to become more useful carving of the field. So there’s this broad argument for AI safety which is, “We’re going to have very intelligent things based on the orthagonality thesis, we can’t really say anything about their goals.” So, the really important thing is to make sure that the intelligence is pointed at the right goals, it’s pointed at doing what we actually want.

And so then the natural approach is, how do we get our AI systems to infer what we want to do and then actually pursue that? And I think, in some sense, it’s one of the most obvious approaches to AI safety. This is a clear enough problem, even with narrow current systems that there are plenty of people outside of AI safety working on this, as well. So this incorporates things like inverse reinforcement learning, preference learning, reward modeling, the CIRL cooperative IRL paper also fits into all of this. So yeah, I can begin to ante up those in more depth.

Lucas: Why don’t you start off by talking about the people who exist within the field of AI safety, give sort of a brief characterization of what’s going on outside of the field, but primarily focusing on those within the field. How this approach, in practice, I think generally is, say, different from MIRI to start off with, because we have a clear picture of them painted right next to what we’re delving into now.

Rohin: So I think difference of MiRI is that this is more targeted directly at the problem right now, in that you’re actually trying to figure out how do you build an AI system that does what you want. Now, admittedly, most of the techniques that people have come up with are not likely to scale up to super-intelligent AI, they’re not meant to, no one claims that they’re going to scale up to super-intelligent AI. They’re more like some incremental progress on figuring out how to get AI systems to do what we want and, hopefully, with enough incremental progress, we’ll get to a point where we can go, “Yes, this is what we need to do.”

Probably the most well known person here would be Dylan Hadfield-Menell, who you had on your podcast. And so he talked about CIRL and associated things quite a bit there, there’s not really that much I would say in addition to it. Maybe a quick summary of Dylan’s position is something like, “Instead of having AI systems that are optimizing for their own goals, we need to have AI systems that are optimizing for our goals, and try to infer our goals in order to do that.”

So rather than having an AI system that is individually rational with respect to its own goals, you instead want to have a human AI system such that the entire system is rationally optimizing for the human’s goals. This is sort of the point made by CIRL, where you have an AI system, you’ve got a human, they’re playing those two player game, the humans is the only one who knows the reward function, the robot is uncertain about what the reward function is, and has to learn by observing what the humans does.

And so, now you see that the robot does not have a utility function that it is trying to optimize; instead is learning about a utility function that the human has and then helping the human optimize that reward function. So summary, try to build human AI systems that are group rational, as opposed to an AI system that is individually rational; so that’s Dylan’s view. Then there’s Jan Leike at DeepMind, and a few people at OpenAI.

Lucas: Before we pivot into OpenAI and DeepMind, just sort of focusing here on the CHAI end of things and this broad view, and help me explain here how you would characterize it. The present day actively focused view on current issues, and present day issues and alignment and making incremental progress there. This view here you see as a sort of subsuming multiple organizations?

Rohin: Yes, I do.

Lucas: Okay. Is there a specific name you would, again, use to characterize this view?

Rohin: Oh, getting AI systems to do what we want. Let’s see, do I have a pithy name for this? Helpful AI systems or something.

Lucas: Right which, again, is focused on current day things, is seeking to make incremental progress, and which subsumes many different organizations?

Rohin: Yeah, that seems broadly true. I do think there are people who are doing more conceptual work, thinking about how this will scale to AGI and stuff like that; but it’s a minority of work in the space.

Lucas: Right. And so the question of how do we get AI systems to do what we want them to do, also includes these views of, say, Vingean Reflection or how we become idealized versions of ourselves, or how we build on value over time, right?

Rohin: Yeah. So, those are definitely questions that you would need to answer at some point. I’m not sure that you would need to answer Vingean Reflection at some point. But you would definitely need to answer how do you update, given that humans don’t actually know what they want, for a long-term future; you need to be able to deal with that fact at some point. It’s not really a focus of current research, but I agree that that is a thing about this approach will have to deal with, at some point.

Lucas: Okay. So, moving on from you and Dylan to DeepMind and these other places that you view as this sort of approach also being practice there?

Rohin: Yeah, so while Dylan and I and other at CHAI has been focused on sort of conceptual advances, like in toy environments, does this do the right thing? What are some sorts of data that we can learn from? Do they work in these very simple environments with quite simple algorithms? I would say that OpenAI and DeepMind safety teams are more focused on trying to get this to work in complex environments of the sort that we’re getting this to work on state-of-the-art environments, the most complex ones that we have.

Now I don’t mean DoTA and StarCraft, because running experiments with DoTAi and StarCraft is incredibly expensive, but can we get AI systems that do what we want for environments like Atari or MuJoCo? There’s some work on this happening at CHAI, there are pre-prints available online, but it hasn’t been published very widely yet. Most of the work, I would say, has been happening with an OpenAI/DeepMind collaboration, and most recently, there was a position paper from DeepMind on recursive reward modeling.

Right before that there was also a paper on combining first a paper, deeper enforcement learning from human preferences, which said, “Okay if we allow humans to specify what they want by just comparing between different pieces of behavior from the AI system, can we train an AI system to do what the human wants?” And then they built on that in order to create a system that could learn from demonstrations, initially, using a kind of imitation learning, and then improve upon the demonstrations using comparisons in the same way that deep RL from human preferences did.

So one way that you can do this research is that there’s this field of human computer interaction, which is about … well, it’s about many things. But one of the things that it’s about is how do you make the user interface for humans intuitive and easy to use such that you don’t have user error or operator? One comment from people that I liked is that most of the things that are classified as ‘user error’ or ‘operator error’ should not be classified as such, they should be classified as ‘interface errors’ where you had such a confusing interface that well, of course, at some point some user was going to get it wrong.

And similarly, here, what we want is a particular behavior out of the AI, or at least a particular set of outcomes from the AI; maybe we don’t know exactly how to achieve those outcomes. And AI is about giving us the tools to create that behavior in automated systems. The current tool that we all use is the reward function, we write down the reward function and then we give it to an algorithm, and it produces behaviors and the outcomes that we want.

And reward functions, they’re just a pretty terrible user interface, they’re better than the previous interface which is writing a program explicitly, which humans cannot do it if the task is something like image classification or continuous control in MuJoCo; it’s an improvement upon that. But reward functions are still a pretty poor interface, because they’re implicitly saying that they encode perfect knowledge of the optimal behavior in all possible environments; which is clearly not a thing that humans can do.

I would say that this area is about moving on from reward functions, going to the next thing that makes the human’s job even easier. And so we’ve got things like comparisons, we’ve got things like inverse award design where you specify a proxy to work function that only needs to work in the training environment. Or you do something like inverse reinforcement learning, where you learn from demonstrations; so I think that’s one nice way of looking at this field.

Lucas: So do you have anything else you would like to add on here about how we present-day get AI systems to do what we want them to do, section of the field?

Rohin: Maybe I want to plug my value learning sequence, because it talks about this much more eloquently than I can on this podcast?

Lucas: Sure. Where can people find your value learning sequence?

Rohin: It’s on the Alignment Forum. You just go to the Alignment Forum, at the top there’s ‘Recommended Sequences’, there’s ‘Embedded Agency’, which is from MIRI, the sort of stuff we already talked about; so that’s also great sequence, I would recommend it. There’s iterated amplification, also great sequence we haven’t talked about it yet. And then there’s my value learning sequence, so you can see it on the front page of the Alignment Forum.

Lucas: Great. So we’ve characterized these, say, different parts of the AI alignment field. And probably just so far it’s been cut into this sort of MIRI view, and then this broad approach of trying to get present-day AI systems to do what we want them to do, and to make incremental progress there. Are there any other slices of the AI alignment field that you would like to bring to light?

Rohin: Yeah, I’ve got four or five more. There’s the interated amplification and debate side of things, which is how do we build using current technologies, but imagining that they were way better? How do we build and align AGI? So they’re trying to solve the entire problem, as opposed to making incremental progress and, simultaneously, hopefully thinking about, conceptually, how do we fit all of these pieces together?

There’s limiting the AGI system, which is more about how do we prevent AI systems from behaving catastrophically? It makes no guarantees about the AI systems doing what we want, it just prevents them from doing really, really bad things. Techniques in that section includes boxing and avoiding side effects. There’s the robustness view, which is about how do we make AI systems well behaved or robustly? I guess that’s pretty self explanatory.

There’s transparency or interpretability, which I wouldn’t say is a technique by itself, but seems to be broadly useful for almost all of the other avenues, it’s something we would want to add to other techniques in order to make those techniques more effective. There’s also, in the same frame as MIRI, can we even understand intelligence? Can we even forecast what’s going to happen with AI? And within that, there’s comprehensive AI services.

here’s also lots of efforts on forecasting, but comprehensive AI services actually makes claims about what technical AI safety should do. So I think that one actually does have a place in this podcast, whereas most of the forecasting things do not, obviously. They have some implications on the strategic picture, but they don’t have clear implications on technical safety research directions, as far as I can tell it right now.

Lucas: Alright, so, do you want to go ahead and start off with the first one on the list there And then we’ll move sequentially down?

Rohin: Yeah, so iterated amplification and debate. This is similar to the helpful AGI section in the sense that we are trying to build an AI system that does what we want. That’s still the case here, but we’re now trying to figure out, conceptually, how can we do this using things like reinforcement learning and supervised learning, but imagining that they’re way better than they are right now? Such that the resulting agent is going to be aligned with us and reach arbitrary levels of intelligence; so in some sense, it’s trying to solve the entire problem.

We want to come up with a scheme such that if we run that scheme, we get good outcomes, we’ve solved almost all the problem. I think that it also differs in that the argument for why we can be successful is also different. This field is aiming to get a property of corrigibility, which I like to summarize as trying to help the overseer. It might fail to help the overseer, or the human, or the user, because it’s not very competent and maybe it makes a mistake and things that I like apples when actually I want oranges. But it was actually trying to help me; it actually thought I wanted apples.

So in corrigibility, you’re trying to help the overseer, whereas, in the previous thing about helpful AGI, you’re more getting an AI system that actually does what we want; there isn’t this distinction between what you’re trying to do versus what you actually do. So there’s a slightly different property that you’re trying to ensure, I think, on the strategic picture that’s the main difference.

The other difference is that these approaches are trying to make a single, unified generally intelligent AI system, and so they will make assumptions like, given that we’re trying to imagine something that’s generally intelligent, it should be able to do X, Y, and Z. Whereas the research agenda that’s let’s try to get AI systems that do want you want, tends not to make those assumptions. And so it’s more applicable to current systems or narrow system where you can’t assume that you have general intelligence.

For example, a claim that that Paul Christiano often talks about is that, “If your AI agent is generally intelligent and a little bit corrigible, it will probably easily be able to infer that its overseer, or the user, would like to remain in control of any resources that they have, and would like to be better informed about the situation, that the user would prefer that the agent does not lie to them etc., etc.” It was definitely not something that current day AI systems can do unless you really engineer them to, so this is presuming some level of generality, which we do not currently have.

So the next thing I said was limited AGI. Here the idea is, there are not very many policies or AI systems that will do what we want; what we want is a pretty narrow space in the space of all possible behaviors. Actually selecting one of the behaviors out of that space is quite difficult and requires a lot of information in order to narrow in on that piece of behavior. But if all you’re trying to do is avoid the catastrophic behaviors, then there are lots and lots of policies that successfully do that. And so it might be easier to find one of those policies; a policy that doesn’t ever kill all humans.

Lucas: At least the space of those policies, one might have this view and not think it sufficient for AI alignment, but see it as sort of a low hanging fruit to be picked. Because the space of non-catastrophic outcomes is larger than the space of extremely specific futures that human beings support.

Rohin: Yeah, exactly. And the success story here is, basically, that we develop this way of preventing catastrophic behaviors. All of our AI systems are filled with the system in place, and then technological progress continues as usual; it’s maybe not as fast as it would have been if we had an aligned AGI doing all of this for us, but hopefully it would still be somewhat fast, and hopefully enabled a bit by AI systems. Eventually, we will either make it to the future without ever building an AI system that doesn’t have a system in place, or we use this to do a bunch more AI research until we solve the full alignment problem, and then we can build, with high confidence that it’ll go well.

And actual proper aligned, super-intelligence that is helping us without any of these limitations systems in place. I think from a strategic picture, that’s basically the important parts about limited AGI. There are two subsections within those limits based on trying to change what the AI’s optimizing for, so this would be something like impact measures versus limits on the input/output channels of the AI system; so this would be something like AI boxing.

So, with robustness, I sort of think of the robustness mostly, it’s not going to give us safety by itself, probably, though there are some scenarios in which it could happen. It’s more meant to harden whichever other approach that we use. Maybe if we have an AI system that is trying to do what we want, to go back to the helpful AGI setting, maybe it does that 99.9 percent of the time. But we’re using this AI to make millions of decisions, which means it’s going to not do what we want 1,000 times. That seems like way too many times for comfort, because if it’s applying its intelligence to the wrong goal in those 1,000 times, you could get some pretty bad outcomes.

This is a super heuristic and fluffy argument, but there are lots of problems with it. I think it sets up the general reason that we would want robustness. So with robustness techniques, you’re basically trying to get some nice worst case guarantees that say, “Yeah, the AI system is never going to screw up super, super bad.” And this is helpful when you have an AI system that’s going to make many, many, many decisions, and we want to make sure that none of those decisions are going to be catastrophic.

And so some techniques in here include verification, adversarial training, and other adversarial ML techniques like Byzantine fault tolerance, or stuff like that. These are all the data poisoning, interpretability can also be helpful for robustness if you’ve got a strong overseer who can use interpretability to give good feedback to your AI system. But yeah, the overall goal is take something that doesn’t fail 99 percent of the time, and get it to not fail 100 percent of the time, or check whether or not it ever fails, so that you don’t have this very rare but very bad outcome.

Lucas: And so would you see this section as being within the context of any others or being sort of at a higher level of abstraction?

Rohin: I would say that it applies to any of the others, well okay, not the MIRI embedded agency stuff, because we don’t really have a story for how that ends up helping with AI safety. It could apply to however that caches out in the future, but we don’t really know right now. With limited AGI, many have this theoretical model, if you apply this sort of penalty, this sort of impact measure, then you’re never going to have any catastrophic outcomes.

But, of course, in practice, we train our AI systems to optimize that penalty and get the sort of weird black box thing out. And we’re not entirely sure if it’s respecting the penalty or something like this. Then you could use something like verification or your transparency in order to make sure that this is actually behaving the way we would predict them behave based on our analysis of what limits we need to put on the AI system.

Similarly, if you build AI systems that are doing what we want, maybe you want to use adversarial training to see if you can find any situations in which the AI system’s doing something weird, doing something which we wouldn’t classify as what we want, with iterated amplification or debate, maybe we want to verify that the corrigibility property happens all the time. It’s unclear how you would use verification for that, because it seems like a particularly hard property to formalize, but you could still do things like adversarial training or transparency.

We might have this theoretical arguments for why our systems will work, then once we turn them into actual real systems that will probably use neural nets and other messy stuff like that, are we sure that in the translation from theory to practice, all of our guarantees stayed? Unclear, we should probably use some robustness techniques to check that.

Interpretability, I believe, was next. It’s sort of similar in that it’s broadly useful for everything else. If you want to figure out whether an AI system is doing what you want, it would be really helpful to be able to look into the agent and see, “Oh, it chose to buy apples because it had seen me eat apples in the past.” Versus, “It chose to buy apples because there was this company that made it to buy the apples, so that it would make more profit.”

If we could see those two cases, if we could actually see into the decision making process, it becomes a lot easier to tell whether or not the AI system is doing what we want, or whether or not the AI system is corrigible, or whether or not be AI system is properly … Well, maybe it’s not as obvious for impact measures, but I wouldn’t expect it to be useful there as well, even if I don’t have a story off the top of my head.

Similarly with robustness, if you’re doing something like adversarial training, it sure would help if your adversary was able to look into the inner workings of the agent and be like, “Ah, I see this agent, it tends to underwrite this particular class of risky outcomes. So why don’t I search within that class of situations for one that is going to take a big risk on that it shouldn’t have taken otherwise?” It just makes all of the other problems a lot easier to do.

Lucas: And so how is progress made on interpretability?

Rohin: Right now I think most of the progress is in image classifiers. I’ve seen some work on interpretability for deep RL as well. Honestly, that’s probably most of the research is happening with classification systems, primarily image classifiers, but others as well. And then I also see the deep RL explanation systems because I read a lot of deep RL research.

But it’s motivated a lot, there are real problems with current AI systems, and interpretability helps you to diagnose and fix those, as well. For example, the problems of bias in classifiers, one thing that I remember from Deep Dream is you can ask Deep Dream to visualize barbells. And you always see these sort of muscular arms that are attached to the barbells because, in the training set, barbells were always being picked up by muscular people. So, that’s a way that you can tell that your classifier is not really learning the concepts that you wanted it to do.

In the bias case maybe your classifier always classifies anyone sitting at a computer as a man, because of bias in the data set. And using interpretability techniques, you could see that, okay when you look at this picture, the AI system is looking primarily at the pixels that represent the computer, as opposed to the pixels that represent the human. And making its decision to label this person as a man, based on that, and you’re like, no, that’s clearly the wrong thing to do. The classifier should be paying attention to the human, not to the laptop.

So I think a lot of interpretability research right now is you take a particular short term problem and figure out how you can make that problem easier to solve. Though a lot of it is also what would be the best way to understand what our model is doing? So I think a lot of the work that Chris Olah doing, for example, is in this vein, and then as we do this exploration, finding some sort of bias in the classifiers that you’re studying.

So, Comprehensive AI Services, an attempt to predict what the feature of AI development will look like, and the hope is that, by doing this, we can figure out what sort of technical safety things we will need to do. Or, strategically, what sort of things we should push for in the AI research community in order to make those systems safer.

There’s a big difference between, we are going to build a single unified AGI agent and it’s going to be generally intelligent to optimize the world according to a utility function versus we are going to build a bunch of disparate, separate, narrow AI systems that are going to interact with each other quite a lot. And because of that, they will be able to do a wide variety of tasks, none of them are going to look particularly like expected utility maximizers. And the safety research you want to do is different in those two different worlds. And CAIS is basically saying “We’re in the second of those worlds, not the first one.”

Lucas: Can you go ahead and tell us about ambitious value learning?

Rohin: Yeah, so with ambitious value learning, this is also an approach to how do we make an aligned AGI solve the entire problem in some sense? Which is look at not just human behavior, but also human brains of the algorithm that they implement, and use that to infer an adequate utility function, the one that we would be okay with the behavior that results from that.

Infer this utility function, I’m going to plug it into an expected utility maximizer. Now, of course, we do have to solve problems with even once we have the utility function, how do we actually build a system that maximizes that utility function, which is not a solved problem yet? But it does seem to be capturing from the main difficulties, if you could actually solve the problem. And so that’s an approach I associate most with Stuart Armstrong.

Lucas: Alright, and so you were saying earlier, in terms of your own view, it’s sort of an amalgamation of different credences that you have in the potential efficacy of all these different approaches. So, given all of these and all of their broad missions, and interests, and assumptions that they’re willing to make, what are you most hopeful about? What are you excited about? How do you, sort of, assign your credence and time here?

Rohin: I think I’m most excited about the concept of corrigibility. That seems like the right thing to aim for, it seems like it’s a thing we can achieve, it seems like if we achieve it, we’re probably okay, nothing’s going to go horribly wrong and probably will go very well. I am less confident on which approach to corrigibility I am most excited about. Iterated amplification and debate seem like if we were to implement them, they will probably lead to incorrigible behavior. But I am worried that either of those will be … Either we won’t actually be able to build generally intelligent agents, in which case both of those approaches don’t really work. Or another worry that I have is that those approaches might be too expensive to actually do in that other systems are just so much more computationally efficient that we just use those instead.

Due to economic pressures, Paul does not seem to be worried by either of these things. He’s definitely aware of both these issues, in fact, he was the one I think who listed computational efficiency as a desideratum, and he still is optimistic about them. So, I would not put a huge amount of credence in this view of mine.

If I were to say what I was excited about for portability instead of that, it would be something like take the research that we’re currently doing on how to get current AI systems to work, which often called ‘narrow value learning’. If you take that research, it seems plausible that this research, extended into the future, will give us some method of creating an AI system that’s implicitly learning our narrow values, and is corrigible as a result of that, even if it is not generally intelligent.

This is sort of a very hand wavey speculative intuition, certainly not as concrete as the hope that we have with iterated amplification. But I’m somewhat optimistic about it, and less optimistic about limiting AI systems, it seems like even if you succeed in finding a nice, simple rule that eliminates all catastrophic behaviors, which plausibly you could do, it seems hard to find one that both does that and also lets you do all of the things that you do want to do.

If you’re talking about impact metrics, for example, if you require AI to be a low impact, I expect that that would prevent you from doing many things that we actually want to do, because many things that we want to do are actually quite high impact. Now, Alex Turner disagrees with me on this, and he developed attainable utility preservation. He is explicitly working on this problem and disagree with me, so again I don’t know how much credence to put in this.

I don’t know if Vika agrees with me on this or not, she also might disagree with me and she is also directly working with this problem. So, yeah, seems hard to put a limit that also lets us do and things that we want. And in that case, it seems like due to economic pressures, we’d end up doing the things that don’t limit our AI systems from doing what they want.

I want to keep emphasizing my extreme uncertainty over all of this given that other people disagree with me on this, but that’s my current opinion. Similarly with boxing, it seems like it’s going to just make it very hard to actually use the AI system. Robustness and interpretability seems very broadly useful and supportive of most research on interpretability; maybe with an eye towards long term concerns, just because it seems to make every other approach to AI safety a lot more feasible and easier to solve.

I don’t think it’s a solution by itself, but given that it seems to improve almost every story I have for making an aligned AGI, seems like it’s very much worth getting a better understanding of it. Robustness is an interesting one, it’s not clear to me, if it is actually necessary. I kind of want to just voice lots of uncertainty about robustness and leave it at that. It’s certainly good to do in that it helps us be more confident in our AI systems, but maybe everything would be okay even if we just didn’t do anything. I don’t know, I feel like I would have to think a lot more about this and also see the techniques that we actually used to build AGI in order to have a better opinion on that.

Lucas: Could you give a few examples of where your intuitions here are coming from that don’t see robustness as an essential part of the AI alignment?

Rohin: Well, one major intuition, if you look at humans, they’re at least some human where I’m like, “Okay, I could just make this human a lot smarter, a lot faster, have them think for many, many years, and I still expect that they will be robust and not lead to some catastrophic outcome. They may not do exactly what I would have done, because they’re doing what they want. But they’re probably going to do something reasonable, they’re not going to do something crazy or ridiculous.

I feel like humans, some humans, the sufficiently risk averse and uncertain ones seem to be reasonably robust. I think that if you know that you’re planning over a very, very, very long time horizon, so imagine that you know you’re planning over billions of years, then the rational response to this is, “I really better make sure not to screw up right now, since there is just so much reward in the future, I really need to make sure that I can get it.” And so you get very strong pressures for preserving option value or not doing anything super crazy. So I think you could, plausibly, just get the reasonable outcomes from those effects. But again, these are not well thought out.

Lucas: All right, and so I just want to go ahead and guide us back to your general views, again, on the approaches. Is there anything that you’d like to add their own the approaches?

Rohin: I think I didn’t talk about CAIS yet. I guess my general view of CAIS, I broadly agree with it, that this does seem to be the most likely development path, meaning that it’s more likely than any other specific development path, but not more likely to have any other development path.

So I broadly agree with the worldview presented, I’m still trying to figure out what implications it has for technical safety research. I don’t agree with all of it, in particular, I think that you are likely to get AGI agents at some point, probably, after the CAIS soup of services happens. Which, I think, again, Drexler disagrees with me on that. So, put a bunch of uncertainty on that, but I broadly agree with that worldview that CAIS is proposing.

Lucas: In terms of this disagreement between you and Eric Drexler, are you imagining agenty AGI or super-intelligence which comes after the CAIS soup? Do you see that as an inevitable byproduct of CAIS or do you see that as an inevitable choice that humanity will make? And is Eric pushing the view that the agenty stuff doesn’t necessarily come later, it’s a choice that human beings would have to make?

Rohin: I do think it’s more like saying that this will be a choice that humans will make at some point. I’m sure that Eric, to some extent, is saying, “Yeah, just don’t do that.” But I think Eric and I do, in fact, have a disagreement on how much more performance you can get from an AGI agent, than a CAIS super of services. My argument is something like there is efficiency to be gained from going to an AGI agent, and Eric’s position as best I understand it, is that there is actually just not that much economic incentive to go to an AGI agent.

Lucas: What are your intuition pumps for why you think that you will gain a lot of computational efficiency from creating sort of an AGI agent? We don’t have to go super deep, but I guess a terse summary or something?

Rohin: Sure, I guess the main intuition pump is that in all of the past cases that we have of AI systems, you see that in speech recognition, in deep reinforcement learning, in image classification, we had all of the hand-built systems that separated these out into a few different modules that interacted with each other in a vaguely CAIS-like way. And then, at some point, we got enough computer and large enough data sets that we just threw deep learning at it, and deep learning just blew those approaches out of the water.

So there’s the argument from empirical experience, and there’s also the argument of if you try to modularize your systems yourself, you can’t really optimize the communication between them, you’re less integrated and you can’t make decisions based on global information, you have to make it based off of local information. And so the decisions tend to be a little bit worse. This could be taken as an explanation for the empirical observation that I made that we can already make; so that’s another intuition pump there.

Eric’s response would probably be something like, “Sure, this seems true for these narrow tasks, for narrow tasks.” You can get a lot of efficiency gains by integrating everything together and throwing deep learning and [inaudible 00:54:10] training at all of it. But for a sufficiently high level tasks, there’s not really that much to be gained by doing global information instead of local information, so you don’t actually lose much by having these separate systems, and you do get a lot of computational deficiency in generalization bonuses by modularizing. He had a good example of this that I’m not replicating and I don’t want to make my own example, because it’s not going to be as convincing; but that’s his current argument.

And then my counter-argument is that’s because humans have small brains, so given the size of our brains and the limits of our data, and the limits of the compute that we have, we are forced to do modularity and systematization to break tasks apart into modular chunks that we can then do individually. Like if you are running a corporation, you need each person to specialize in their own task without thinking about all the other tasks, because we just do not have the ability to optimize for everything all together because we have small brains, relatively speaking; or limited brains, is what I should say.

But this is not a limit that AI systems will have. An AI system would just vastly more computer than the human brain, vastly more data will, in fact, just be able to optimize all of this with global information and get better results. So that’s one thread of the argument taken down to two or three levels of arguments and counter-arguments. There are other threads of that debate, as well.

Lucas: I think that that serves a purpose for illustrating that here. So are there any other approaches here that you’d like to cover, or is that it?

Rohin: I didn’t talk about factored cognition very much. But I think it’s worth highlighting separately from iterated amplification in that it’s testing an empirical hypothesis of can humans decompose tasks into chunks of some small amount of time? And can we do arbitrarily complex tasks using these humans? I am particularly excited about this sort of work that’s trying to figure out what humans are capable of doing and what supervision they can give to AI systems.

Mostly because going back to a thing I said way back in the beginning, what we’re aiming for is a human AI system to be collectively rational as opposed to an AI system as individually rational. Part of the human-AI-system is the human, you want to be able to know what the human can do, what sort of policies they can implement, what sort of feedback they can be giving to the AI system. And something like factory cognition is testing a particular aspect of that; and I think that seems great and we need more of it.

Lucas: Right. I think that this seems to be the sort of emerging view of where social science or scientists are needed in AI alignment in order to, again as you said, sort of understand what human beings are capable in terms of supervised learning and analyzing the human component of the AI alignment problem as it requires us to be collectively rational with AI systems.

Rohin: Yeah, that seems right. I expect more writing on this in the future.

Lucas: All right, so there’s just a ton of approaches here to AI alignment, and our heroic listeners have a lot to take in here. In terms of getting more information, generally, about these approaches or if people are still interested in delving into all these different views that people take at the problem and methodologies of working on it, what would you suggest that interested persons look into or read into?

Rohin: I cannot give you a overview of everything, because that does not exist. To the extent that it exists, it’s either this podcast or the talk that I did at Beneficial AGI. I can suggest resources for individual items, so for embedded agency there’s the embedded agency sequence on the Alignment Forum; far and away the best thing for read for that.

For CAIS, Comprehensive AI Services, there was a 200 plus page tech report published by Eric Drexler at the beginning of this month, if you’re interested, you should go read the entire thing; it is quite good. But I also wrote a summary of it on the Alignment Forum, which is much more readable, in the sense that it’s shorter. And then there are a lot of comments on there that analyze it a bit more.

There’s also another summary written by Richard Ngo, also on the Alignment Forum. Maybe it’s only on Lesswrong, I forget; it’s probably on the Alignment Forum. But that’s a different take on comprehensive AI services, so I’d recommend reading that too.

For limited AGI, I have not really been keeping up with the literature on boxing, so I don’t have a favorite to recommend. I know that a couple have been written by, I believe, Jim Babcock and Roman Yampolskiy.

For impact measures, you want to read Vika’s paper on relative reachability. There’s also a blog post about it if you don’t want to read the paper. And Alex Turner’s blog posts on attainable utility preservation, I think it’s called ‘Towards A New Impact Measure’, and this is on the Alignment Forum.

For robustness, I would read Paul Christiano’s post called ‘Techniques For Optimizing Worst Case Performance’. This is definitely specific to how robustness will help under Paul’s conception of the problem and, in particular, his thinking of robustness in the setting where you have a very strong overseer for your AI system. But I don’t know of any other papers or blog post that’s talking about robustness, generally.

For AI systems that do what we want, there’s my value learning sequence that I mentioned before on the Alignment Forum. There’s CIRL or Cooperative Inverse Reinforcement Learning which is a paper by Dylan and others. There’s Deep Reinforcement Learning From Human Preferences and Recursive Reward Modeling, these are both papers that are particular instances of work in this field. I also want to recommend Inverse Reward Design, because I really like that paper; so that’s also a paper by Dylan, and others.

For corrigibility and iterated amplification, the iterated amplification sequence on the Alignment Forum or half of what Paul Christiano has written. If you want to read not an entire sequence of blog posts, then I think Clarifying AI alignment is probably the post I would recommend. It’s one of the posts in the sequence and talks about this distinction of creating an AI system that is trying to do what you want, as opposed to actually doing what you want and why we might want to aim for only the first one.

For iterated amplification, itself, that technique, there is a paper that I believe is called something like Supervising Strong Learners By Amplifying Weak Experts, which is a good thing to read and there’s also corresponding OpenAI blog posts, whose name I forget. I think if you search iterated amplification, OpenAI blog you’ll find it.

And then for debate, there’s AI Safety via Debate, which is a paper, there’s also a corresponding OpenAI blog post. For factory cognition, there’s a post called Factored Cognition, on the Alignment Forum; again, in the iterated amplification sequence.

For interpretability, there isn’t really anything talking about interpretability, from the strategic point of view of why we want it. I guess that same post I recommend before of techniques for optimizing worst case performance talks about it a little bit. For actual interpretability techniques, I recommend the distill articles, the building blocks of interpretability and feature visualization, but these are more about particular techniques for interpretability, as opposed to why we wanted interpretability.

And on ambitious value learning, the first chapter of my sequence on value learning talks exclusively about ambitious value learning; so that’s one thing I’d recommend. But also Stuart Armstrong has so many posts, I think there’s one that’s about resolving human values adequately and something else, something like that. That one might be one worth checking out, it’s very technical though; lots of math.

He’s also written a bunch of posts that convey the intuitions behind the ideas. They’re all split into a bunch of very short posts, so I can’t really recommend any one particular one. You could go to the alignment newsletter database and just search Stuart Armstrong, and click on all of those posts and read them. I think that was everything.

Lucas: That’s a wonderful list. So we’ll go ahead and link those all in the article which goes along with this podcast, so that’ll all be there organized in nice, neat lists for people. This is all probably been fairly overwhelming in terms of the number of approaches and how they differ, and how one is to adjudicate the merits of all of them. If someone is just sort of entering the space of AI alignment, or is beginning to be interested in sort of these different technical approaches, do you have any recommendations?

Rohin: Reading a lot, rather than trying to do actual research. This was my strategy, I started back in September of 2017 and I think for the first six months or so, I was reading about 20 hours a week, in addition to doing research; which was why it was only 20 hours a week, it wasn’t a full time thing I was doing.

And I think that was very helpful for actually forming a picture of what everyone was doing. Now, it’s plausible that you don’t want to actually learn about what everyone is doing, and you’re okay with like, “I’m fairly confident that this thing, this particular problem is an important piece of the problem and we need to solve it.” And I think it’s very easy to get that wrong, so I’m a little wary of recommending that but it’s a reasonable strategy to just say, “Okay, we probably will need to solve this problem, but even if we don’t, the intuitions that we get from trying to solve this problem will be useful.

Focusing on that particular problem, reading all of the literature on that, attacking that problem, in particular, lets you start doing things faster, while still doing things that are probably going to be useful; so that’s another strategy that people could do. But I don’t think it’s very good for orienting yourself in the field of AI safety.

Lucas: So you think that there’s a high value in people taking this time to read, to understand all the papers and the approaches before trying to participate in particular research questions or methodologies. Given how open this question is, all the approaches make different assumptions and take for granted different axioms which all come together to create a wide variety of things which can both complement each other and have varying degrees of efficacy in the real world when AI systems start to become more developed and advanced.

Rohin: Yeah, that seems right to me. Part of the reason I’m recommending this is because it seems to be that no one does this. I think, on the margin, I want more people who do this in a world where 20 percent of the people were doing this, and the other 80 percent were just taking particular piece of the problem and working on those. That might be the right balance, somewhere around there, I don’t know, it depends on how you count who is actually in the field. But somewhere between one and 10 percent of the people are doing this; closer to the one.

Lucas: Which is quite interesting, I think, given that it seems like AI alignment should be in a stage of maximum exploration just given the conceptually mapping the territory is very young. I mean, we’re essentially seeing the birth and initial development of an entirely new field and specific application of thinking. And there are many more mistakes to be made, and concepts to be clarified, and layers to be built. So, seems like we should be maximizing our attention in exploring the general space, trying to develop models, the efficacy of different approaches and philosophies and views of AI alignment.

Rohin: Yeah, I agree with you, that should not be surprising given that I am one of the people doing this, or trying to do this. Probably the better critique will come from people who are not doing this, and can tell both of us why we’re wrong about this.

Lucas: We’ve covered a lot here in terms of the specific approaches, your thoughts on the approaches, where we can find resources on the approaches, why setting the approaches matters. Are there any parts of the approaches that you feel deserve more attention in terms of these different sections that we’ve covered?

Rohin: I think I would want more work on looking at the intersection between things that are supposed to be complimentary, how interpretability can help you have AI systems that have the right goals, for example, would be a cool thing to do. Or what you need to do in order to get verification, which is a sub-part of robustness, to give you interesting guarantees on AI systems that we actually care about.

Most of the work on verification right now is like, there’s this nice specification that we have for adversarial examples, in particular, is there an input that is within some distance from a training data point, such that it gets classified differently from that training data point. And those are the nice formal specification and most of the work in verification takes this specification as given and that figures out more and more computationally efficient ways to actually verify that property, basically.

That does seem like a thing that needs to happen, but the much more urgent thing, in my mind, is how do we come up with these specifications in the first place? If I want to verify that my AI system is corrigible, or I want to verify that it’s not going to do anything catastrophic, or that it is going to not disable my value learning system, or something like that; how do I specify this at all in any way that lets me do something like a verification technique even given infinite computing power? It’s not clear to me how you would do something like that, and I would love to see people do more research on that.

That particular thing is my current reason for not being very optimistic about verification, in particular, but I don’t think anyone has really given it a try. So it’s plausible that there’s actually just some approach that could work that we just haven’t found yet because no one’s really been trying. I think all of the work on limited AGI is talking about, okay, does this actually eliminate all of the catastrophic behavior? Which, yeah, that’s definitely an important thing, but I wish that people would also do research on, given that we put this penalty or this limit on the AGI system, what things is it still capable of doing?

Have we just made it impossible for it to do anything of interest whatsoever, or can it actually still do pretty powerful things, even though we’ve placed these limits on it? That’s the main thing I want to see. From there, let’s have AI systems that do what we want, probably the biggest thing I want to see there, and I’ve been trying to do some of this myself, some conceptual thinking about how does this lead to good outcomes in the long term? So far, we’ve not been dealing with the fact that the human doesn’t actually know, doesn’t actually have a nice consistent utility function that they know and that can be optimized. So, once you relax that assumption, what the hell do you do? And then there’s also a bunch of other problems that would benefit from more conceptual clarification, maybe I don’t need to go into all of them right now.

Lucas: Yeah. And just to sort of inject something here that I think we haven’t touched on and that you might have some words about in terms of approaches. We discussed sort of agential views of advanced artificial intelligence, a services-based conception, though I don’t believe that we have talked about aligning AI systems that simply function as oracles or having a concert of oracles. You can get rid of the services thing, and the agency thing if the AI just tells you what is true, or answers your questions in a way that is value aligned.

Rohin: Yeah, I mostly want to punt on that question because I have not actually read all the papers. I might have read a grand total of one paper on the oracles, and also super intelligence which talks about oracles. So I feel like I know so little about the state of the art on oracles, that it should not actually say anything about them.

Lucas: Sure. So then just as a broad point to point out to our audience is that in terms of conceptualizing these different approaches to AI alignment, it’s important and crucial to consider the kind of AI system that you’re thinking about the kinds of features and properties that it has, and oracles are another version here that one can play with in one’s AI alignment thinking?

Rohin: I think the canonical paper there is something like Good and Safe Pieces of Oracles, but I have not actually read it. There is a list of things I want to read, it is on that list. But that list also has, I think, something like 300 papers on it, and apparently I have not gotten to oracles yet.

Lucas: And so for the sake of this whole podcast being as comprehensive as possible, are there any conceptions of AI, for example, that we have omitted so far adding on to this agential view, the CAIS view of it actually just being a lot of distributed services, or an oracle view?

Rohin: There’s also the Tool AI View. This is different from the services view, but it’s somewhat akin to the view you were talking about at the beginning of this podcast where you’ve got AI systems that have a narrowly defined input/output space, they’ve got a particular thing that they do with limit, and they just sort of take in their inputs and do some computation, they spit out their outputs and that’s it, that’s all that they do. You can’t really model them as having some long term utility function that they’re optimizing, they’re just implementing a particular input-output relation and it’s all they’re trying to do.

Even saying something like, “They are trying to do X.” Is basically using a bad model for them. I think the main argument against expecting tool AI systems is that they’re probably not going to be as useful as other services or agential AI, because tool AI systems would have to be programmed in a way where we understood what they were doing and why they were doing it. Whereas agential AI systems or services would be able to consider new possible ways of achieving goals that we hadn’t thought about and enact those plans.

And so they could get super human behavior by considering things that we wouldn’t consider. Whereas, true Ais … Like Google Maps is super human in some sense, but it’s super human only because it has a compute advantage over us. If we were given all of the data and all of the time, in human real time, that Google Maps had, we could implement a similar sort of algorithm as Google Maps and compute the optimal route ourselves.

Lucas: There seems to be this duality that is constantly being formed in our conception of AI alignment, where the AI system is this tangible external object which stands in some relationship to the human and is trying to help the human to achieve certain things.

Are there conceptions of value alignment which, however the procedure or methodology is done, changes or challenges the relationship between the AI system and the human system where it challenges what it means to be the AI or what it means to be human, whereas, there’s potentially some sort of merging or disruption of this dualistic scenario of the relationship?

Rohin: I don’t really know, I mean, it sounds like you’re talking about things like brain computer interfaces and stuff like that. I don’t really know of any intersection between AI safety research and that. I guess, this did remind me, too, that I want to make the point that all of this is about the relatively narrow, I claim, problem of aligning an AI system with a single human.

There is also the problem of, okay what if there are multiple humans, what if there are multiple AI systems, what if you’ve got a bunch of different groups of people and each group is value aligned within themselves, they build an AI that’s value aligned with them, but lots of different groups do this now what happens?

Solving the problem that I’ve been talking about does not mean that you have a good outcome in the long term future, it is merely one piece of a larger overall picture. I don’t think any of that larger overall picture removes the dualistic thing that you were talking about, but they dualistic part reminded me of the fact that I am talking about a narrow problem and not the whole problem, in some sense.

Lucas: Right and so just to offer some conceptual clarification here, again, the first problem is how do I get an AI system to do what I want it to do when the world is just me and that AI system?

Rohin: Me and that AI system and the rest of humanity, but the rest of humanity is treated as part of the environment.

Lucas: Right, so you’re not modeling other AI systems or how some mutually incompatible preferences and trained systems would interact in the world or something like that?

Rohin: Exactly.

Lucas: So the full AI alignment problem is… It’s funny because it’s just the question of civilization, I guess. How do you get the whole world and all of the AI systems to make a beautiful world instead of a bad world?

Rohin: Yeah, I’m not sure if you saw my lightning talk at Beneficial AGI, but I talked a bit about those. I think I called that top level problem, make AI related features stuff go well, very, very, very concrete, obviously.

Lucas: It makes sense. People know what you’re talking about.

Rohin: I probably wouldn’t call that broad problem the AI alignment problem. I kind of wonder is there a different alignment for the narrower trouble? We could maybe call it the ‘AI Safety Problem’ or the ‘AI Future Problem’, I don’t know. ‘Beneficially AI’ problem actually, I think that’s what I used last time.

Lucas: That’s a nice way to put it. So I think that, conceptually, leave us at a very good place for this first section.

Rohin: Yeah, seems pretty good to me.

Lucas: If you found this podcast interesting or useful, please make sure to check back for part two in a couple weeks where Rohin and I go into more detail about the strengths and weaknesses of specific approaches.

We’ll be back again soon with another episode in the AI Alignment podcast.

[end of recorded material]