Lucas: Hey, everyone. Welcome back to the AI Alignment Podcast series. I’m Lucas Perry, and today we’ll be speaking with El Mahdi El Mhamdi on the Byzantine problem, Byzantine tolerance, and poisoning in distributed learning and computer networks. If you find this podcast interesting or useful, please give it a like and follow us on your preferred listing platform. El Mahdi El Mhamdi pioneered Byzantine resilient machine learning devising a series of provably safe algorithms he recently presented at NeurIPS and ICML. Interested in theoretical biology, his work also includes the analysis of error propagation and networks applied to both neural and biomolecular networks. With that, El Mahdi’s going to start us off with a thought experiment.

El Mahdi: Imagine you are part of a group of three generals, say, from the Byzantine army surrounding a city you want to invade, but you also want to retreat if retreat is the safest choice for your army. You don’t want to attack when you will lose, so those three generals that you’re part of are in three sides of the city. They sent some intelligence inside the walls of the city, and depending on this intelligence information, they think they will have a good chance of winning and they would like to attack, or they think they will be defeated by the city, so it’s better for them to retreat. Your final decision would be a majority vote, so you communicate through some horsemen that, let’s say, are reliable for the sake of this discussion. But there might be one of you who might have been corrupt by the city.

The situation would be problematic if, say, there are General A, General B, and General C. General A decided to attack. General B decided to retreat based on their intelligence for some legitimate reason. A and B are not corrupt, and say that C is corrupt. Of course, A and B, they can’t figure out who was corrupt. Say C is corrupt. What this general would do they thing, so A wanted to attack. They will tell them, “I also want to attack. I will attack.” Then they will tell General B, “I also want to retreat. I will retreat.” A receives two attack votes and one retreat votes. General B receives two retreat votes and only one attack votes. If they trust everyone, they don’t do any double checking, this would be a disaster.

A will attack alone; B would retreat; C, of course, doesn’t care because he was corrupt by the cities. You can tell me they can circumvent that by double checking. For example, A and B can communicate on what C told them. Let’s say that every general communicates with every general on what he decides and on also what’s the remaining part of the group told them. A will report to B, “General C told me to attack.” Then B would tell C, “General C told me to retreat.” But then A and B wouldn’t have anyway of concluding whether the inconsistency is coming from the fact that C is corrupt or that the general reporting on what C told them is corrupt.

I am General A. I have all the valid reasons to think with the same likelihood that C is maybe lying to me or also B might also be lying to me. I can’t know if you are misreporting what C told you enough for the city to corrupt one general if there are three. It’s impossible to come up with an agreement in this situation. You can easily see that this will generalize to having more than three generals, like I say 100, as soon as the non-corrupt one are less than two-thirds because what we saw with three generals would happen with the fractions that are not corrupt. Say that you have strictly more than 33 generals out of 100 who are corrupt, so what they can do is they can switch the majority votes on each side.

But worse than that, say that you have 34 corrupt generals and the remaining 66 not corrupt generals. Say that those 66 not corrupt generals were 33 on the attack side, 33 on the retreat side. The problem is that when you are in some side, say that you are in the retreat side, you have in front of you a group of 34 plus 33 in which there’s a majority of malicious ones. This majority can collude. It’s part of the Byzantine hypothesis. The malicious ones can collude and they will report a majority of inconsistent messages on the minority on the 33 ones. You can’t provably realize that the inconsistency is coming from the group of 34 because they are a majority.

Lucas: When we’re thinking about, say, 100 persons or 100 generals, why is it that they’re going to be partitioned automatically into these three groups? What if there’s more than three groups?

El Mahdi: Here we’re doing the easiest form of Byzantine agreement. We want to agree on attack versus retreat. When it’s become multi-dimensional, it gets even messier. There are more impossibility results and impossibility results. Just like with the binary decision, there is an impossibility theorem on having agreement if you have unsigned messages to horsemen. Whenever the corrupt group exceeds 33%, you provably cannot come up with an agreement. There are many variants to this problem, of course, depending on what hypothesis you can assume. Here, without even mentioning it, we were assuming bounded delays. The horsemen would always arrive eventually. If the horsemen could die on the way and you don’t have any way to check whether they arrive or not or you can be waiting forever because you don’t have any proof that the horsemen died on the way.

You don’t have any mechanism to tell you, “Stop waiting for the horsemen. Stop waiting for the message from General B because the horsemen died.” You can be waiting forever and there are theorems that shows that when you have unbounded delays, and by the way, like in distributed computing, whenever you have in bounded delays, we speak about asynchrony. If you have a synchronous communication, there is a very famous theorem that tells you consensus is impossible, not even in the malicious case, but just like in …

Lucas: In the mundane normal case.

El Mahdi: Yes. It’s called the Fischer Lynch Patterson theorem theorem .

Lucas: Right, so just to dive down into the crux of the problem, the issue here fundamentally is that when groups of computers or groups of generals or whatever are trying to check who is lying amongst discrepancies and similarities of lists and everyone who’s claiming what is when there appears to be a simple majority within that level of corrupted submissions, then, yeah, you’re screwed.

El Mahdi: Yes. It’s impossible to achieve agreement. There are always fractions of malicious agents above which is provably impossible to agree. Depending on the situation, it will be a third or sometimes or a half or a quarter, depending on your specifications.

Lucas: If you start tweaking the assumptions behind the thought experiment, then it changes what number of corrupted machines or agents that are required in order to flip the majority and to poison the communication.

El Mahdi: Exactly. But for example, you mentioned something very relevant to today’s discussion, which is what if we were not agreeing on two decisions, retreat, attack. What if we were agreeing on some multi-dimensional decision? Attack or retreat on one dimension and then …

Lucas: Maybe hold, keep the siege going.

El Mahdi: Yeah, just like add possibilities or dimensions and multi-dimensional agreements. They’re even more hopeless results in that direction

Lucas: There are more like impossibility theorems and issues where these distributed systems are vulnerable to small amounts of systems being corrupt and screwing over the entire distributed network.

El Mahdi: Yes. Maybe now we can slightly move to machine learning.

Lucas: I’m happy to move into machine learning now. We’ve talked about this, and I think our audience can probably tell how this has to do with computers. Yeah, just dive in what this has to do with machine learning and AI and current systems today, and why it even matters for AI alignment.

El Mahdi: As a brief transition, solving the agreement problem besides this very nice historic thought experiment is behind consistencies of safety critical systems like banking systems. Imagine we have a shared account. Maybe you remove 10% of the amount and then she or he added some $10 to the accounts. You remove the $10 in New York and she or he put the $10 in Los Angeles. The banking system has to agree on the ordering because minus $10 plus 10% is not the same result as plus 10% then minus $10. The final balance of the account would not be the same.

Lucas: Right.

El Mahdi: The banking systems routinely are solving decisions that fall into agreement. If you work on some document sharing platform, like Dropbox or Google Docs, whatever, and we collaboratively are writing the document, me and you. The document sharing platform has to, on real time, solve agreements about the ordering of operations so that me and you always keep seeing the same thing. This has to happen while some of the machines that are interconnecting us are failing, whether just like failing because there was a electric crash or something. Data center has lost some machines or if it was like restart, a bug or a take away. What we want in distributed computing is that we would like communications schemes between machines that’s guarantee this consistency that comes from agreement as long as some fraction of machines are reliable. What this has to do with artificial intelligence and machine learning reliability is that with some colleagues, we are trying to encompass one of the major issues in machine learning reliability inside the Byzantine fault tolerance umbrella. For example, you take, for instance, poisoning attacks.

Lucas: Unpack what poisoning attacks are.

El Mahdi: For example, imagine you are training a model on what are good videos to recommend given some key word search. If you search for “medical advice for young parents on vaccine,” this is a label. Let’s assume for the sake of simplicity that a video that tells you not to take your kid for vaccines is not what we mean by medical advice for young parents on vaccine because that’s what medical experts agree on. We want our system to learn that anitvaxers, like anti-vaccine propaganda is not what people are searching for when they type those key words, so I suppose a world where we care about accuracy, okay? Imagine you want to train a machine learning model that gives you accurate results of your search. Let’s also for the sake of simplicity assume that a majority of people on the internet are honest.

Let’s assume that more than 50% of people are not actively trying to poison the internet. Yeah, this is very optimistic, but let’s assume that. What we can show and what me and my colleagues started this line of research with is that you can easily prove that one single malicious agent can provably poison a distributed machine learning scheme. Imagine you are this video sharing platform. Whenever people behave on your platform, this generates what we call gradients, so it updates your model. It only takes a few hyperactive accounts that could generate behavior that is powerful enough to pull what we call the average gradient because what distributed machine learning is using, at least up to today, if you read the source code of most distributed machine learning frameworks. Distributed machine learning is always averaging gradients.

Imagine you Lucas Perry just googled a video on the Parkland shootings. Then the video sharing platform shows you a video telling you that David Hogg and Emma Gonzalez and those kids behind the March for Our Lives movement are crisis actors. The video labels three kids as crisis actors. It obviously has a wrong label, so it is what I will call a poisoned data point. If you are non-malicious agents on the video sharing platform, you will dislike the video. You will not approve it. You’re likely to flag it. This should generate a gradient that pushes the model in that direction, so the gradient will update the model into a direction where it stops thinking that this video is relevant for someone searching “Parkland shooting survivors.” What can happen if your machine learning framework is just averaging gradients is that a bunch of hyperactive people on some topic could poison the average and pull it towards the direction where the models is enforcing this thinking that, “Yeah, those kids are crisis actors.”

Lucas: This is the case because the hyperactive accounts are seen to be given more weight than accounts which are less active in the same space. But this extra weighting that these accounts will get from their hyperactivity in one certain category or space over another, how is the weighting done? Is it just time spent per category or does it have to do with submissions that agree with the majority?

El Mahdi: We don’t even need to go into the details because we don’t know. I’m talking in a general setting where you have a video sharing platform aggregating gradients for behavior. Now, maybe let’s raise the abstraction level. You are doing gradient descents, so you have a lost function that you want to minimize. You have an error function. The error function is the mismatch between what you predict and what the user tells you. The user tells you this is a wrong prediction, and then you move to the direction where the users stop telling you this is the wrong direction. You are doing great in this sense minimizing the lost function. User behaves, and with their behavior, you generate gradients.

What you do now in the state of the arts way of distributed machine learning is that you average all those gradients. Averaging is well known not to be resilient. If you have a room of poor academics earning a few thousand dollars and then a billionaire jumps in the room, if your algorithm reasons with averaging, it will think that this is a room of millionaires because the average salary would be a couple of hundred millions. But then million is very obvious to do when it comes to salaries and numbers scalers because you can rank them.

Lucas: Right.

El Mahdi: You rank numbers and then decide, “Okay, this is the ordering. This is the number that falls in the middle. This is the upper half. This is the lower half and this is the median.” When it becomes high dimensional, the median is a bit tricky. It has some computational issues. Then even if you compute what we call the geometric median, an attacker can still know how to leverage the fact that you’re only approximating it because there’s no closed formula. There’s no closed form to compute the median in that dimension. But worse than that, what we showed in one of our follow up works is because of the fact that machine learning is done in very, very, very high dimensions, you would have a curse of the dimensionality issue that makes it possible for attackers to sneak in without being spot as a way of the median.

It can still look like the median vector. I take benefits from the fact that those vectors, those gradients, are extremely high dimensional. I would look for all the disagreements. Let’s say you have a group of a couple hundred gradients, and I’m the only malicious one. I would look at the group of correct vectors all updating you somehow in the same direction within some variants. On average, they’re like what we call unbiased estimators of the gradient. When you take out the randomness, the expected value they will give you is the real gradient of the loss function. What I will do as a malicious worker is I will look at the way they are disagreeing slightly on each direction.

I will sum that. I will see that they disagree by this much on direction one. They disagree by this much on direction two. They disagree by this much, epsilon one, epsilon two, epsilon three. I would look for all these small disagreements they have on all the components.

Lucas: Across all dimensions and high dimensional space. [crosstalk 00:16:35]

El Mahdi: Then add that up. It will be my budget, my leeway, my margin to attack you on another direction.

Lucas: I see.

El Mahdi: What we proved is that you have to mix ideas from geometric median with ideas from the traditional component-wise median, and that those are completely different things. The geometric median is a way to find a median by just minimizing the sum of distances between what you look for and all the vectors that were proposed, and then the component-wise median will do a traditional job of ranking the coordinates. It looks at each coordinate, and then rank all the propositions, and then look for the proposition that lies in the middle. Once we proved enough follow up work is that, yeah, the geometric median idea is elegant. It can make you converge, but it can make you converge to something arbitrarily bad decided by the attacker. When you train complex models like neural nets, the landscape you optimize inside is not convex. It’s not like a bowl or a cup where you just follow the descending slope you would end up in the lowest point.

Lucas: Right.

El Mahdi: It’s like a multitude of bowls with different heights.

Lucas: Right, so there’s tons of different local minima across the space.

El Mahdi: Exactly. So in the first paper what we showed is that ideas that look like the geometric median are enough to just converge. You converge. You provably converge, but in the follow up what we realized, like something we were already aware of, but not enough in my opinion, is that there is this square root D, this curse of dimensionality that will arise when you compute high dimensional distances. That the attacker can leverage.

So in what we call the hidden vulnerability of distributed learning, you can have correct vectors, agreeing on one component. Imagine in your head some three axis system.

Let’s say that they are completely in agreement on axis three. But then in axis one, two, so in the plane formed by the axis one and axis two, they have a small disagreement.

What I will do as the malicious agent, is that I will leverage this small disagreement, and inject it in axis three. And this will make you go to a bit slightly modified direction. And instead of going to this very deep, very good minima, you will go into a local trap that is just close ahead.

And that comes from the fact that loss functions of interesting models are clearly like far from being convex. The models are highly dimensional, and the loss function is highly un-convex, and creates a lot of leeway.

Lucas: It creates a lot of local minima spread throughout the space for you to attack the person into.

El Mahdi: Yeah. So convergence is not enough. So we started this research direction by formulating the following question, what does it take to guarantee convergence?

And any scheme that aggregates gradients, and guarantee convergence is called Byzantine resilient. But then you can realize that in very high dimensions, and highly non-convex loss functions, is convergence enough? Would you just want to converge?

There are of course people arguing the deep learning models, like there’s this famous paper by Anna Choromanska, and Yann LeCun, and  Gérard Ben Arous, about the landscape of neural nets, that basically say that, “Yeah, very deep local minimum of neural nets are some how as good.”

From an overly simplified point of view, it’s an optimistic paper, that tells you that you shouldn’t worry too much when you optimize neural nets about the fact that gradient descent would not necessarily go to a global like-

Lucas: To a global minima.

El Mahdi: Yeah. Just like, “Stop caring about that.”

Lucas: Because the local minima are good enough for some reason.

El Mahdi: Yeah. I think that’s a not too unfair way to summarize the paper for the sake of this talk, for the sake of this discussion. What we empirically illustrate here, and theoretically support is that that’s not necessarily true.

Because we show that with very low dimensional, not extremely complex models, trained on CIFAR-10 and MNIST, which are toy problems, very easy toy problems, low dimensional models etc. It’s already enough to have those amounts of parameters, let’s say 100,000 parameters or less, so that an attacker would always find a direction to take you each time away, away, away, and then eventually find an arbitrarily bad local minimum. And then you just converge to that.

So convergence is not enough. Not only you have to seek an aggregation rule that guarantees convergence, but you have to seek some aggregation rules that guarantee that you would not converge to something arbitrarily bad. You would keep converging to the same high quality local minimum, whatever that means.

The hidden vulnerability is this high dimensional idea. It’s the fact that because the loss function is highly non-convex, because there’s the high dimensionality, as an attacker I would always find some direction, so the attack goes this way.

Here the threat model is that an attacker can spy on your gradients, generated by the correct workers but cannot talk on their behalf. So I cannot corrupt the messages. Since you asked about the reliability of horsemen or not.

So horsemen are reliable. I can’t talk on your behalf, but I can spy on you. I can see what are you sending to the others, and anticipate.

So I would as an attacker wait for correct workers to generate their gradients, I will gather those vectors, and then I will just do a linear regression on those vectors to find the best direction to leverage the disagreement on the D minus one remaining directions.

So because there would be this natural disagreement, this variance in many directions, I will just do some linear regression and find what is the best direction to keep? And use the budget I gathered, those epsilons I mentioned earlier, like this D time epsilon on all the directions to inject it the direction that will maximize my chances of taking you away from local minima.

So you will converge, as proven in the early papers, but not necessarily to something good. But what we showed here is that if you combine ideas from multidimensional geometric medians, with ideas from single dimensional component-wise median, you improve your robustness.

Of course it comes with a price. You require three quarters of the workers to be reliable.

There is another direction where we expanded this problem, which is asynchrony. And asynchrony arises when as I said in the Byzantine generals setting, you don’t have a bounded delay. In the bounded delay setting, you know that horses arrive at most after one hour.

Lucas: But I have no idea if the computer on the other side of the planet is ever gonna send me that next update.

El Mahdi: Exactly. So imagine you are doing machine learning on smartphones. You are leveraging a set of smartphones all around the globe, and in different bandwidths, and different communication issues etc.

And you don’t want each time to be bottlenecked by the slowest one. So you want to be asynchronous, you don’t want to wait. You’re just like whenever some update is coming, take it into account.

Imagine some very advanced AI scenario, where you send a lot of learners all across the universe, and then they communicate with the speed of light, but some of them are five light minutes away, but some others are two hours and a half. And you want to learn from all of them, but not necessarily handicap the closest one, because there are some other learners far away.

Lucas: You want to run updates in the context of asynchrony.

El Mahdi: Yes. So you want to update whenever a gradient is popping up.

Lucas: Right. Before we move on to illustrate the problem again here is that the order matters, right? Like in the banking example. Because the 10% plus 10 is different from-

El Mahdi: Yeah. Here the order matters for different reasons. You update me so you are updating me on the model you got three hours ago. But in the meanwhile, three different agents updated me on the models, while getting it three minutes ago.

All the agents are communicating through some abstraction they call the server maybe. Like this server receives updates from fast workers.

Lucas: It receives gradients.

El Mahdi: Yeah, gradients. I also call them updates.

Lucas: Okay.

El Mahdi: Because some workers are close to me and very fast, I’ve done maybe 1000 updates, while you were still working and sending me the message.

So when your update arrive, I can tell whether it is very stale, very late, or malicious. So what we do in here is that, and I think it’s very important now to connect a bit back with classic distributed computing.

Is that Byzantine resilience in machine learning is easier than Byzantine resilience in classical distributed computing for one reason, but it is extremely harder for another reason.

The reason is that we know what we want to agree on. We want to agree on a gradient. We have a toolbox of calculus that tells us how this looks like. We know that it’s the slope of some loss function that is most of today’s models, relatively smooth, differentiable, maybe Lipschitz, bounded, whatever curvature.

So we know that we are agreeing on vectors that are gradients of some loss function. And we know that there is a majority of workers that will produce vectors that will tell us what does a legit vector look like.

You can find some median behavior, and then come up with filtering criterias that will get away with the bad gradients. That’s the good news. That’s why it’s easier to do Byzantine resilience in machine learning than to do Byzantine agreement. Byzantine agreement, because agreement is a way harder problem.

The reason why Byzantine resilience is harder in machine learning than in the typical settings you have in distributed computing is that we are dealing with extremely high dimensional data, extremely high dimensional decisions.

So a decision here is to update the model. It is triggered by a gradient. So whenever I accept a gradient, I make a decision. I make a decision to change the model, to take it away from this state, to this new state, by this much.

But this is a multidimensional update. And Byzantine agreement, or Byzantine approximate agreement in higher dimension has been provably hopeless by Hammurabi Mendes, and Maurice Herlihy in an excellent paper in 2013, where they show that you can’t do Byzantine agreement in D dimension with N agents in less than N to the power D computations, per agent locally.

Of course in their paper, they were meaning Byzantine agreement on positions. So they were framing it with a motivations saying, “This is N to the power D, but the typical cases we care about in distributed computing are like robots agreeing on a position on a plane, or on a position in a three dimensional space.” So D is two or three.

So N to the power two or N to the power three is fine. But in machine learning D is not two and three, D is a billion or a couple of millions. So N to the power a million is just like, just forget.

And not only that, but also they require … Remember when I tell you that Byzantine resilience computing would always have some upper bound on the number malicious agents?

Lucas: Mm-hmm (affirmative).

El Mahdi: So the number of total agents should exceed D times the number of malicious agents.

Lucas: What is D again sorry?

El Mahdi: Dimension.

Lucas: The dimension. Okay.

El Mahdi: So if you have to agree on D dimension, like on a billion dimensional decision, you need at least a billion times the number of malicious agents.

So if you have say 100 malicious agents, you need at least 100 billion total number of agents to be resistant. No one is doing distributed machine learning on 100 billion-

Lucas: And this is because the dimensionality is really screwing with the-

El Mahdi: Yes. Byzantine approximate agreement has been provably hopeless. That’s the bad, that’s why the dimensionality of machine learning makes it really important to go away, to completely go away from traditional distributed computing solutions.

Lucas: Okay.

El Mahdi: So we are not doing agreement. We’re not doing agreement, we’re not even doing approximate agreement. We’re doing something-

Lucas: Totally new.

El Mahdi: Not new, totally different.

Lucas: Okay.

El Mahdi: Called gradient decent. It’s not new. It’s as old as Newton. And it comes with good news. It comes with the fact that there are some properties, like some regularity of the loss function, some properties we can exploit.

And so in the asynchronous setting, it becomes even more critical to leverage those differentiability properties. So because we know that we are optimizing a loss functions that has some regularities, we can have some good news.

And the good news has to do with curvature. What we do here in asynchronous setting, is not only we ask workers for their gradients, we ask them for their empirical estimate of the curvature.

Lucas: Sorry. They’re estimating the curvature of the loss function, that they’re adding the gradient to?

El Mahdi: They add the gradient to the parameter, not the loss function. So we have a loss function, parameter is the abscissa, you add the gradient to the abscissa to update the model, and then you end up in a different place of the loss function.

So you have to imagine the loss function as like a surface, and then the parameter space as the plane, the horizontal plane below the surface. And depending on where you are in the space parameter, you would be on different heights of the loss function.

Lucas: Wait sorry, so does the gradient depend where you are on this, the bottom plane?

El Mahdi: Yeah [crosstalk 00:29:51]-

Lucas: So then you send an estimate for what you think the slope of the intersection will be?

El Mahdi: Yeah. But for asynchrony, not only that. I will ask you to send me the slope, and your observed empirical growth of the slope.

Lucas: The second derivative?

El Mahdi: Yeah.

Lucas: Okay.

El Mahdi: But the second derivative again in high dimension is very hard to compute. You have to computer the Hessian matrix.

Lucas: Okay.

El Mahdi: That’s something like completely ugly to compute in high dimensional situations because it takes D square computations.

As an alternative we would like you to send us some linear computation in D, not a square computation in D.

So we would ask you to compute your actual gradient, your previous gradient, the difference between them, and normalize it by the difference between models.

So, “Tell us your current gradient, by how much it changed from the last gradient, and divide that by how much you changed the parameter.”

So you would tell us, “Okay, this is my current slope, and okay this is the gradient.” And you will also tell us, “By the way, my slope change relative to my parameter change is this much.”

And this would be some empirical estimation of the curvature. So if you are in a very curved area-

Lucas: Then the estimation isn’t gonna be accurate because the linearity is gonna cut through some of the curvature.

El Mahdi: Yeah but if you are in a very curved area of the loss function, your slope will change a lot.

Lucas: Okay. Exponentially changing the slope.

El Mahdi: Yeah. Because you did a very tiny change in the parameter and it takes a lot of the slope.

Lucas: Yeah. Will change the … Yeah.

El Mahdi: When you are in a non-curved area of the loss function, it’s less harmful for us that you are stale, because you will just technically have the same updates.

If you are in a very curved area of the loss function, your updates being stale is now a big problem. So we want to discard your updates proportionally to your curvature.

So this is the main idea of this scheme in asynchrony, where we would ask workers about their gradient, and their empirical growth rates.

And then of course I don’t want to trust you on what you declare, because you can plan to screw me with some gradients, and then declare a legitimate value of the curvature.

I will take those empirical, what we call in the paper empirical Lipschitz-ness. So we ask you for this empirical growth rate, that it’s a scalar, remember? This is very important. It’s a single dimensional number.

And so we ask you about this growth rate, and we ask all of you about growth rates, again assuming the majority is correct. So the majority of growth rates will help us set the median growth rate in a robust manner, because as long as a simple majority is not lying, the median growth rates will always be bounded between two legitimate values of the growth rate.

Lucas: Right because, are you having multiple workers inform you of the same part of your loss function?

El Mahdi: Yes. Even though they do it in an asynchronous manner.

Lucas: Yeah. Then you take the median of all of them.

El Mahdi: Yes. And then we reason by quantiles of the growth rates.

Lucas: Reason by quantiles? What are quantiles?

El Mahdi: The first third, the second third, the third third. Like the first 30%, the second 30%, the third 30%. We will discard the first 30%, discard the last 30%. Anything in the second 30% is safe.

Of course this has some level of pessimism, which is good for safety, but not very good for being fast. Because maybe people are not lying, so maybe the first 30%, and the last 30% are also values we could consider. But for safety reasons we want to be sure.

Lucas: You want to try to get rid of the outliers.

El Mahdi: Possible.

Lucas: Possible outliers.

El Mahdi: Yeah. So we get rid of the first 30%, the last 30%.

Lucas: So this ends up being a more conservative estimate of the loss function?

El Mahdi: Yes. That’s completely right. We explain that in the paper.

Lucas: So there’s a trade off that you can decide-

El Mahdi: Yeah.

Lucas: By choosing what percentiles to throw away.

El Mahdi: Yeah. Safety never comes for free. So here, depending on how good your estimates about the number of potential Byzantine actors is, your level of pessimism with translate into slowdown.

Lucas: Right. And so you can update the amount that you’re cutting off-

El Mahdi: Yeah.

Lucas: Based off of the amount of expected corrupted signals you think you’re getting.

El Mahdi: Yeah. So now imagine a situation where you know the number of workers is know. You know that you are leveraging 100,000 smartphones doing gradient descent for you. Let’s call that N.

You know that F of them might be malicious. We argue that if F is exceeding the third of N, you can’t do anything. So we are in a situation where F is less than a third. So less than 33,000 workers are malicious, then the slowdown would be F over N, so a third.

What if you are in a situation where you know that your malicious agents are way less than a third? For example you know that you have at most 20 rogue accounts in your video sharing platform.

And your video sharing platform has two billion accounts. So you have two billion accounts.

Lucas: 20 of them are malevolent.

El Mahdi: What we show is that the slowdown would be N minus F divided by N. N is the two billion accounts, F is the 20, and is again two billion.

So it would be two billion minus 20, so one million nine hundred billion, like something like 0.999999. So you would go almost as fast as the non-Byzantine resilient scheme.

So our Byzantine resilient scheme has a slowdown that is very reasonable in situations where F, the number of malicious agents is way less than N, the total number of agents, which is typical in modern…

Today, like if you ask social media platforms, they have a lot of a tool kits to prevent people from creating a billion fake accounts. Like you can’t in 20 hours create an army of several million accounts.

None of the mainstream social media platforms today are susceptible to this-

Lucas: Are susceptible to massive corruption.

El Mahdi: Yeah. To this massive account creation. So you know that the number of corrupted accounts are negligible to the number of total accounts.

So that’s the good news. The good news is that you know that F is negligible to N. But then the slowdown of our Byzantine resilient methods is also close to one.

But it has the advantage compared to the state of the art today to train distributed settings of not taking the average gradient. And we argued in the very beginning that those 20 accounts that you could create, it doesn’t take a bot army or whatever, you don’t need to hack into the machines of the social network. You can have a dozen human, sitting somewhere in a house manually creating 20 accounts, training the accounts over time, doing behavior that makes the legitimate for some topics, and then because you’re distributing machine learning scheme would average the gradients generated by people behavior and that making your command anti-vaccine or controversies, anti-Semitic conspiracy theories.

Lucas: So if I have 20 bad gradients and like, 10,000 good gradients for a video, why is it that with averaging 20 bad gradients are messing up the-

El Mahdi: The amplitude. It’s like the billionaire in the room of core academics.

Lucas: Okay, because the amplitude of each of their accounts is greater than the average of the other accounts?

El Mahdi: Yes.

Lucas: The average of other accounts that are going to engage with this thing don’t have as large of an amplitude because they haven’t engaged with this topic as much?

El Mahdi: Yeah, because they’re not super credible on gun control, for example.

Lucas: Yeah, but aren’t there a ton of other accounts with large amplitudes that are going to be looking at the same video and correcting over the-

El Mahdi: Yeah, let’s define large amplitudes. If you come to the video and just like it, that’s a small update. What about you like it, post very engaging comments-

Lucas: So you write a comment that gets a lot of engagement, gets a lot of likes and replies.

El Mahdi: Yeah, that’s how you increase your amplitude. And because you are already doing some good job in becoming the reference on that video-sharing platform when it comes to discussing gun control, the amplitude of your commands is by definition high and the fact that your command was very early on posted and then not only you commented the video but you also produced a follow-up video.

Lucas: I see, so the gradient is really determined by a multitude of things that the video-sharing platform is measuring for, and the metrics are like, how quickly you commented, how many people commented and replied to you. Does it also include language that you used?

El Mahdi: Probably. It depends on the social media platform and it depends on the video-sharing platform and, what is clear is that there are many schemes that those 20 accounts created by this dozen people in a house can try to find good ways to maximize the amplitude of their generated gradients, but this is a way easier problem than the typical problems we have in technical AI safety. This is not value alignment or value loading or coherent extrapolated volition. This is a very easy, tractable problem on which now we have good news, provable results. What’s interesting is the follow-up questions that we are trying to investigate here with my colleagues, the first of which is, don’t necessarily have a majority of people on the internet promoting vaccines.

Lucas: People that are against things are often louder than people that are not.

El Mahdi: Yeah, makes sense, and sometimes maybe numerous because they generate content, and the people who think vaccines are safe not creating content. In some topics it might be safe to say that we have a majority of reasonable, decent people on the internet. But there are some topics in which now even like polls, like the vaccine situation, there’s a surge now of anti-vaccine resentment in western Europe and the US. Ironically this is happening in the developed country now, because people are so young, they don’t remember the non-vaccinated person. My aunt, I come from Morocco. my aunt is handicapped by polio, so I grew up seeing what a non-vaccinated person looks like. So young people in the more developed countries never had a living example of non-vaccinated past.

Lucas: But they do have examples of people that end up with autism and it seems correlated with vaccines.

El Mahdi: Yeah, the anti-vaccine content may just end up being so click baits, and so provocative that it gets popular. So this is a topic where the majority hypothesis which is crucial to poisoning resilience does not hold. An open follow up we’re onto now is how to combine ideas from reputation metrics, PageRank, et cetera, with poisoning resilience. So for example you have the National Health Institute, the John Hopkins Medical Hospital, Harvard Medical School, and I don’t know, the Massachusetts General Hospital having official accounts on some video-sharing platform and then you can spot what they say on some topic because now we are very good at doing semantic analysis of contents.

And know that okay, on the tag vaccines, I know that there’s this bunch of experts and then what you want to make emerge on your platform is some sort of like epistocracy. The power is given to the knowledgeable, like we have in some fields, like in medical regulation. The FDA doesn’t do a majority vote. We don’t have a popular majority vote across the country to tell the FDA whether it should approve this new drug or not. The FDA does some sort of epistocracy where the knowledgeable experts on the topic would vote. So how about mixing ideas from social choice?

Lucas: And topics in which there are experts who can inform.

El Mahdi: Yeah. There’s also a general fall-off of just straight out trying to connect Byzantine resilient learning with social choice, but then there’s another set of follow ups that motivates me even more. We were mentioning workers, workers, people generate accounts on social media, accounts generation gradients. That’s all I can implicitly assume in that the server, the abstraction that’s gathering those gradients is reliable. What about the aggregated platform itself being deployed on rogue machines? So imagine you are whatever platform doing learning. By the way, whatever always we have said from the beginning until now applies as long as you do gradient-based learning. So it can be recommended systems. It can be training some deep reinforcement learning of some super complicated tasks to beat, I don’t know the word, champion in poker.

We do not care as long as there’s some gradient generation from observing some state, some environmental state, and some reward or some label. It can be supervised, reinforced, as long as gradient based or what you say apply. Imagine now you have this platform leveraging distributed gradient creators, but then the platform itself for security reasons is deployed on several machines for fault tolerance. But then those machines themselves can fail. You have to make the servers agree on the model, so despite the fact that a fraction of the workers are not reliable and now a fraction of the servers themselves. This is the most important follow up i’m into now and I think there would be something on archive maybe in February or March on that.

And then a third follow up is practical instances of that, so I’ve been describing speculative thought experiments on power poisoning systems is actually brilliant master students working which means exactly doing that, like on typical recommended systems, datasets where you could see that it’s very easy. It really takes you a bunch of active agents to poison, a hundred thousand ones or more. Probably people working on big social media platforms would have ways to assess what I’ve said, and so as researchers in academia we could only speculate on what can go wrong on those platforms, so what we could do is just like we just took state of the art recommender systems, datasets, and models that are publicly available, and you can show that despite having a large number of reliable recommendation proposers, a small, tiny fraction of proposers can make, I don’t know, like a movie recommendation system recommend the most suicidal triggering film to the most depressed person watching through your platform. So I’m saying, that’s something you don’t want to have.

Lucas: Right. Just wrapping this all up, how do you see this in the context of AI alignment and the future of machine learning and artificial intelligence?

El Mahdi: So I’ve been discussing this here with people in the Beneficial AI conference and it seems that there are two schools of thought. I am still hesitating between the two because I switched within the past three months from the two sides like three times. So one of them thinks that an AGI is by definition resilient to poisoning.

Lucas: Aligned AGI might be by definition.

El Mahdi: Not even aligned. The second school of thought, aligned AGI is Byzantine resilient.

Lucas: Okay, I see.

El Mahdi: Obviously aligned AGI would be poisoning resilience, but let’s just talk about super intelligent AI, not necessarily aligned. So you have a super intelligence, would you include poisoning resilience in the super intelligence definition or not? And one would say that yeah, if you are better than human in whatever task, it means you are also better than human into spotting poison data.

Lucas: Right, I mean the poison data is just messing with your epistemics, and so if you’re super intelligent your epistemics would be less subject to interference.

El Mahdi: But then there is that second school of thought which I switched back again because I find that most people are in the first school of thought now. So I believe that super intelligence doesn’t necessarily include poisoning resilience because of what I call practically time constrained superintelligence. If you have a deadline because of computational complexity, you have to learn something, which can sometimes-

Lucas: Yeah, you want to get things done.

El Mahdi: Yeah, so you want to get it done in a finite amount of time. And because of that you will end up leveraging to speed up your learning. So if a malicious agent just put up bad observations of the environment or bad labeling of whatever is around you, then it can make you learn something else than what you would like as an aligned outcome. I’m strongly on the second side despite many disagreeing with me here. I don’t think super intelligence includes poisoning resilience, because super intelligence would still be built with time constraints.

Lucas: Right. You’re making a tradeoff between safety and computational efficiency.

El Mahdi: Right.

Lucas: It also would obviously seem to matter the kind of world that the ASI finds itself in. If it knows that it’s in a world with no, or very, very, very few malevolent agents that are wanting to poison it, then it can just throw all of this out of the window, but the problem is that we live on a planet with a bunch of other primates that are trying to mess up our machine learning. So I guess just as a kind of fun example in taking it to an extreme, imagine it’s the year 300,000 AD and you have a super intelligence which has sort of spread across space-time and it’s beginning to optimize its cosmic endowment, but it gives some sort of uncertainty over space-time to whether or not there are other super intelligences there who might want to poison its interstellar communication in order to start taking over some of its cosmic endowment. Do you want to just sort of explore?

El Mahdi: Yeah, that was like a closed experiment I proposed earlier to Carl Shulman from the FHI. Imagine some super intelligence reaching the planets where there is a smart form of life emerging from electric communication between plasma clouds. So completely non-carbon, non-silicon based.

Lucas: So if Jupiter made brains on it.

El Mahdi: Yeah, like Jupiter made brains on it just out of electric communication through gas clouds.

Lucas: Yeah, okay.

El Mahdi: And then this turned to a form of communication is smart enough to know that this is a super intelligence reaching the planet to learn about this form of life, and then it would just start trolling it.

Lucas: It’ll start trolling the super intelligence?

El Mahdi: Yeah. So they would come up with an agreement ahead of time, saying, “Yeah, this super intelligence coming from earth throughout our century to discover how we do things here. Let’s just behave dumbly, or let’s just misbehave. And then the super intelligence will start collecting data on this life form and then come back to earth saying, Yeah, they’re just a dumb plasma passive form of nothing interesting.

Lucas: I mean, you don’t think that within the super intelligence’s model, I mean, we’re talking about it right now so obviously a super intelligence will know this when it leaves that there will be agents that are going to try and trick it.

El Mahdi: That’s the rebuttal, yes. That’s the rebuttal again. Again, how much time does super intelligence have to do inference and draw conclusions? You will always have some time constraints.

Lucas: And you don’t always have enough computational power to model other agents efficiently to know whether or not they’re lying, or …

El Mahdi: You could always come up with thought experiment with some sort of other form of intelligence, like another super intelligence is trying to-

Lucas: There’s never, ever a perfect computer science, never.

El Mahdi: Yeah, you can say that.

Lucas: Security is never perfect. Information exchange is never perfect. But you can improve it.

El Mahdi: Yeah.

Lucas: Wouldn’t you assume that the complexity of the attacks would also scale? We just have a ton of people working on defense, but if we have an equal amount of people working on attack, wouldn’t we have an equally complex method of poisoning that our current methods would just be overcome by?

El Mahdi: That’s part of the empirical follow-up I mentioned. The one Isabella and I were working on, which is trying to do some sort of min-max game of poisoner versus poisoning resilience learner, adversarial poisoning setting where like a poisoner and then there is like a resilient learner and the poisoner tries to maximize. And what we have so far is very depressing. It turns out that it’s very easy to be a poisoner. Computationally it’s way easier to be the poisoner than to be-

Lucas: Yeah, I mean, in general in the world it’s easier to destroy things than to create order.

El Mahdi: As I said in the beginning, this is a sub-topic of technical AI safety where I believe it’s easier to have tractable formalizable problems for which you can probably have a safe solution.

Lucas: Solution.

El Mahdi: But in very concrete, very short term aspects of that. In March we are going to announce a major update in Tensor Flow which is the standout frameworks today to do distributed machine learning, open source by Google, so we will announce hopefully if everything goes right in sys ML in the systems for machine learning conference, like more empirically focused colleagues, so based on the algorithms I mentioned earlier which were presented at NuerIPS and ICML from the past two years, they will announce a major update where they basically changed every averaging insight in terms of flow by those three algorithms I mentioned, Krum and Bulyan and soon Kardam which constitute our portfolio of Byzantine resilience algorithms.

Another consequence that comes for free with that is that distributed machinery frameworks like terms of flow use TCPIP as a communication protocol. So TCPIP has a problem. It’s reliable but it’s very slow. You have to repeatedly repeat some messages, et cetera, to guarantee reliability, and we would like to have a faster communication protocol, like UDP. We don’t need to go through those details. But it has some package drop, so so far there was no version of terms of flow or any distributed machine learning framework to my knowledge using UDP. The old used TCPIP because they needed reliable communication, but now because we are Byzantine resilient, we can afford having fast but not completely reliable communication protocols like UDP. So one of the things that come for free with Byzantine resilience is that you can move from heavy-

Lucas: A little bit more computation.

El Mahdi: -yeah, heavy communication protocols like TCPIP to lighter, faster, more live communication protocols like UDP.

Lucas: Keeping in mind you’re trading off.

El Mahdi: Exactly. Now we have this portfolio of algorithms which can serve many other applications besides just making faster distributed machine learning, like making poisoning resilience. I don’t know, recommended systems for social media and hopefully making AGI learning poisoning resilience matter.

Lucas: Wonderful. So if people want to check out some of your work or follow you on social media, what is the best place to keep up with you?

El Mahdi: Twitter. My handle is El Badhio, so maybe you would have it written down on the description.

Lucas: Yeah, cool.

El Mahdi: Yeah, Twitter is the best way to get in touch.

Lucas: All right. Well, wonderful. Thank you so much for speaking with me today and I’m excited to see what comes out of all this next.

El Mahdi: Thank you. Thank you for hosting this.

Lucas: If you enjoyed this podcast, please subscribe, give it a like, or share it on your preferred social media platform. We’ll be back again soon with another episode in the AI Alignment series.

[end of recorded material]

Lucas: Hey everyone, welcome back to the AI Alignment Podcast series. I’m Lucas Perry and today we will be speaking with Rohin Shah about his work on inverse reinforcement learning and his general take on the state of AI alignment efforts and theory today. 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. He has also been working with effective altruism for several years. Without further ado I give you Rohin Shah.

Hey, Rohin, thank you so much for coming on the podcast. It’s really a pleasure to be speaking with you.

Rohin: Hey, Lucas. Yeah. Thanks for inviting me. I’m glad to be on.

Lucas: Today I think that it would be interesting just to start off by delving into a lot of the current work that you’ve been looking into and practicing over the past few years. In terms of your research, it looks like you’ve been doing a lot of work on practical algorithms for inverse reinforcement learning that take into account, as you say, systematic cognitive biases that people have. It would be interesting if you could just sort of unpack this work that you’ve been doing on this and then contextualize it a bit within the AI alignment problem.

Rohin: Sure. So basically the idea with inverse reinforcement learning is you can look at the behavior of some agent, perhaps a human, and tell what they’re trying to optimize, what are the things that they care about? What are their goals? And in theory this seems like a pretty nice way to do AI alignment and that intuitively you can just say, “Hey, AI, go look at the actions of humans are taking, look at what they say, look at what they do, take all of that in and figure out what humans care about.” And then you could use that perhaps as a utility function for your AI system.

I think I have become less optimistic about this approach now for reasons I’ll get into, partly because of my research on systematic biases. Basically one problem that you have to deal with is the fact that whatever humans are trying to optimize for, they’re not going to do it perfectly. We’ve got all of these sorts of cognitive biases like a planning fallacy or hyperbolic time discounters, when we tend to be myopic, not looking as far into the long-term as we perhaps could.

So assuming that humans are like perfectly optimizing goals that they care about is like clearly not going to work. And in fact, basically, if you make that assumption, well, then whatever reward function you infer, once the AI system is optimizing that, it’s going to simply recover the human performance because well, you assumed that it was optimal when you inferred what it was so that means whatever the humans were doing is probably the behavior that optimizes their work function that you inferred.

And we’d really like to be able to reach super human performance. We’d like our AI systems to tell us how we’re wrong to get new technologies develop things that we couldn’t have done ourselves. And that’s not really something we can do using the sort of naive version of inverse reinforcement learning that just assumes that you’re optimal. So one thing you could try to do is to learn the ways in which humans are biased, the ways in which they make mistakes, the ways in which they plan sub-optimally. And if you could learn that, then you could correct for those mistakes, take them into account when you’re inferring human values.

The example I like to use is if there’s a grad student who procrastinates or doesn’t plan well and as a result near a paper deadline they’re frantically working, but they don’t get it in time and they miss the paper deadline. If you assume that they’re optimal, optimizing for their goals very well I don’t know what you’d infer, maybe something like grad students like to miss deadlines. Something like that seems pretty odd and it doesn’t seem like you’d get something sensible out of that, but if you realize that humans are not very good at planning, they have the planning fallacy and they tend to procrastinate for reasons that they wouldn’t endorse on reflection, then maybe you’d be able to say, “Oh, this was just a mistake of a grad student made. In the future I should try to help them meet their deadlines.”

So that’s the reason that you want to learn systematic biases. My research was basically let’s just take the hammer of deep learning and apply it to this problem. So not just learn the reward function, but let’s also learn the biases. It turns out that this was already known, but there is an impossibility result that says that you can’t do this in general. So more, I guess I would phrase the question I was investigating, as what are a weaker set of assumptions some of than the ones that we currently use such that you can still do some reasonable form of IRL.

Lucas: Sorry. Just stepping back for like half a second. What does this impossibility theorem say?

Rohin: The impossibility theorem says that if you assume that the human is basically running some sort of planner that takes in a reward function and spits out a behavior or a policy, a thing to do over time, then if you all you see is the behavior of the human, basically any reward function is compatible with some planner. So you can’t learn anything about that reward function without making any more assumptions. And intuitively, this is because for any complex behavior you see you could either call it, “Hey, the human’s optimizing a reward that makes them act like that. “Or you could say, “I guess the human is biased and they’re trying to do something else, but they did this instead.”

The sort of extreme version of this is like if you give me an option between apples and oranges and I picked the apple, you could say, “Hey, Rohin probably likes apples and is good at maximizing his reward of getting apples.” Or you could say, “Rohin probably likes oranges and he is just extremely bad at satisfying his preferences. He’s got a systematic bias that always causes him to choose the opposite of what he wants.” And you can’t distinguish between these two cases just by looking at my behavior.

Lucas: Yeah, that makes sense. So we can pivot sort of back in here into this main line of thought that you were on.

Rohin: Yeah. So basically with that impossibility result … When I look at the impossibility result, I sort of say that humans do this all the time, humans just sort of look at other humans and they can figure out what they want to do. So it seems like there are probably some simple set of assumptions that humans are using to infer what other humans are doing. So a simple one would be when the consequences of something or obvious to humans. Now, how you determine when that is another question, but when that’s true humans tend to be close to optimal and if you have something like that, you can rule out the planner that says the human is anti-rational and always chooses the worst possible thing.

Similarly, you might say that as tasks get more and more complex or require more and more computation, the probability that the human chooses the action that best maximizes his or her goals also goes down since the task is more complex and maybe a human doesn’t figure that out, figure out what’s the best thing to do. Maybe with enough of these assumptions we could get some sort of algorithm that actually works.

So we looked at if you make the assumption that the human is often close to rational and a few other assumptions about humans behaving similarly or planning similarly on similar tasks, then you can maybe, kind of, sort of, in simplified settings do IRL better than if you had just assumed that the human was optimal if humans actually systematically biased, but I wouldn’t say that our results are great. I don’t think I would say that I definitively, conclusively said, “This will never work.” Nor did I definitively conclusively say that this is great and we should definitely be putting more resources into it. Sort of somewhere in the middle, maybe more on the negative side of like this seems like a really hard problem and I’m not sure how we get around it.

Lucas: So I guess just as a point of comparison here, how is it that human beings succeed at this every day in terms of inferring preferences?

Rohin: I think humans have the benefit of being able to model the other person as being very similar to themselves. If I am trying to infer what you are doing I can sort of say, “Well, if I were in Lucas issues and I were doing this, what would I be optimizing?” And that’s a pretty good answer to what you would be optimizing. Humans are just in some absolute sense very similar to each other. We have similar biases. We’ve got similar ways of thinking. And I think we’ve leveraged that similarity a lot using our own self models as a drop-in approximation of the other person’s planner in this planner reward language.

And then we say, “Okay, well, if this other person thought like me and this is what they ended up doing, well then, what must they have been optimizing?” I think you’ll see that when this assumption breaks down humans actually get worse at inferring goals. It’s harder for me to infer what someone in a different culture is actually trying to do. They might have values that are like significantly different from mine.

I’ve been in both India and the US and it often seems to me that people in the US just have a hard time grasping the way that Indians see society and family expectations and things like this. So that’s an example that I’ve observed. It’s probably also true the other way around, but I was never old enough in India to actually think through this.

Lucas: Human beings sort of succeed in inferring preferences of people who they can model as having like similar values as their own or if you know that the person has similar values as your own. If inferring human preferences from inverse reinforcement learning is sort of not having the most promising results, then what do you believe to be a stronger way of inferring human preferences?

Rohin: The one thing I correct there is that I don’t think humans do it by assuming that people have similar values, just that people think in similar ways. For example, I am not particularly good at dancing. If I see someone doing a lot of hip-hop or something. It’s not that I value hip-hop and so I can infer they value hip-hop. It’s that I know that I do things that I like and they are doing hip-hop. Therefore, they probably like doing hip-hop. But anyway, that’s the minor point.

So a, just because IRL algorithms aren’t doing well now, I don’t think it’s true that IRL algorithms couldn’t do well in the future. It’s reasonable to expect that they would match human performance. That said, I’m not super optimistic about IRL anyway, because even if we do figure out how to get IRL algorithms and sort of make all these implicit assumptions that humans are making that we can then run and get what a human would have thought other humans are optimizing, I’m not really happy about then going and optimizing that utility function off into the far future, which is what sort of the default assumption that we seem to have when using inverse reinforcement learning.

It may be that IRL algorithms are good for other things, but for that particular application, it seems like the utility function you infer is going to not really scale to things that super intelligence will let us do. Humans just think very differently about how they want the future to go. In some sense, the future is going to be very, very different. We’re going to need to think a lot about how we want the future to go. All of our experience so far has not trained us to be able to think about what we care about in the sort of feature setting where we’ve got as a simple example the ability to easily copy people if they’re uploaded as software.

If that’s a thing that happens, well, is it okay to clone yourself? How does democracy work? All these sorts of things are somewhat value judgments. If you take egalitarianism and run with it, you basically get that one person can copy themselves millions of millions of times and just determine the outcome of all voting that way. That seems bad, but on our current values, I think that is probably what we want and we just really haven’t thought this through. IRL to infer utility function that we’ve then just ruthlessly optimized in the long-term just seems like by the time when the world changes a bunch, the value function that we inferred is going to be weirdly wrong in strange ways that we can’t predict.

Lucas: Why not run continuous updates on it as people update given the change of the world?

Rohin: It seems broadly reasonable. This is the sort of idea that you could have about how you could use IRL in a more realistic way that actually works. I think that’s perfectly fine. I’m optimistic about approaches that are like, “Okay, we’re going to use IRL to infer a value function or reward function or something and we’re going to use that to inform what the AI does, but it’s not going to be the end-all utility functions. It’s just going to infer what we do now and AI system is somehow going to check with us. Maybe it’s got some uncertainty over what the true reward function is. Maybe that it only keeps this reward function for a certain amount of time.”

These seem like things that are worth exploring, but I don’t know that we have the correct way to do it. So in the particular case that you proposed, just updating the reward function over time. The classic wire heading question is, how do we make it so that the AI doesn’t say, “Okay, actually, in order to optimize the utility function I have now, it would be good for me to prevent you from changing my utility function since if you change my utility function, I’m no longer going to achieve my original utility.” So that’s one issue.

The other issue is maybe it starts doing some long-term plans. Maybe even if it’s planning according to this utility function without expecting some changes to the utility function in the future, then it might set up some long-term plans that are going to look bad in the future, but it is hard to stop them in the future. Like you make some irreversible change to society because you didn’t realize that something was going to change. These sorts of things suggest you don’t want a single utility function that you’re optimizing even if you’re updating that utility function over time.

It could be that you have some sort of uncertainty over utility functions and that might be okay. I’m not sure. I don’t think that it’s settled that we don’t want to do something like this. I think it’s settled that we don’t want to use IRL to infer a utility function and optimize that one forever. There are certain middle grounds. I don’t know how well those middle grounds work. There are some intuitively there are going to be some problems, but maybe we can get around those.

Lucas: Let me try to do a quick summary just to see if I can explain this as simply as possible. There are people and people have preferences, and a good way to try and infer their preferences is through their observed behavior, except that human beings have cognitive and psychological biases, which sort of skew their actions because they’re not perfectly rational epistemic agents or rational agents. So the value system or award system that they’re optimizing for is imperfectly expressed through their behavior. If you’re going to infer the preferences from behavior than you have to correct for biases and epistemic and rational failures to try and inferr the true reward function. Stopping there. Is that sort of like a succinct way you’d put it?

Rohin: Yeah, I think maybe another point that might be the same or might be different is that under our normal definition of what our preferences or our values are, if we would say something like, “I value egalitarianism, but it seems predictably true that in the future we’re not going to have a single vote per a sentient being,” or something. Then essentially what that says is that our preferences, our values are going to change over time and they depend on the environment in which we are right now.

So you can either see that as okay, I have this really big, really global, really long-term utility function that tells me how given my environment what my narrow values in that environment are. And in that case and you say, “Well okay, in that case, we’re really super biased because we only really know our values in the environment. We don’t know our values in future environments. We’d have to think a lot more for that.” Or you can say, “We can infer our narrow values now and that has some biases thrown in, but we could probably account for those that then we have to have some sort of story for how we deal with our preferences evolving in the future.”

Those are two different perspectives on the same problem, I would say, and they differ in basically what you’re defining values to be. Is it the thing that tells you how to extrapolate what you want all the way into the future or is it the thing that tells you how you’re behaving right now in the environment. I think our classical notion of preference or values, the one that we use when we say values in everyday language is talking about the second kind, the more narrow kind.

Lucas: There’s really a lot there, I think, especially in terms of issues in that personal identity over time, commitment to values and as you said, different ideas and conceptualization of value, like what is it that I’m actually optimizing for or care about. Population ethics and tons of things about how people value future versions of themselves or whether or not they actually equally care about their value function at all times as it changes within the environment.

Rohin: That’s a great description of why I am nervous around inverse reinforcement learning. You listed a ton of issues and I’m like, yeah, all of those are like really difficult issues. And with inverse reinforcement learning, it’s sort of based on this premise of all of that is existent, is real and is timeless and we can infer it and then maybe we put on some hacks like continuously improving the value function over time to take into account changes, but this does feel like we’re starting with some fundamentally flawed paradigm.

So mostly because of this fact that it feels like we’ve taken a flawed paradigm to start with, then changed it so that it doesn’t have all the obvious flaws. I’m more optimistic about trying to have a different paradigm of how we want to build AI, which maybe I’ll summarize as just make AIs that do what we want or what we mean at the current moment in time and then make sure that they evolve along with us as we evolve and how we think about the world.

Lucas: Yeah. That specific feature there is something that we were trying to address in inverse reinforcement learning, if the algorithm were sort of updating overtime alongside myself. I just want to step back for a moment to try to get an even grander and more conceptual understanding of the globalness of inverse reinforcement learning. So from an evolutionary and sort of more cosmological perspective, you can say that from the time that the first self-replicating organisms on the planet until today, like the entire evolutionary tree, there’s sort of a global utility function across all animals that is ultimately driven by thermodynamics and the sun shining light on a planet and that this sort of global utility function of all agents across the planet, it seems like very ontologically basic and pure like what simply empirically exists. Attempting to access that through IRL is just interesting, the difficulties that arise from that. Does that sort of a picture seem accurate?

Rohin: I think I’m not super sure what exactly you’re proposing here. So let me try and restate it. So if we look at the environment as a whole or the universe as a whole or maybe we’re looking at evolution perhaps and we see that hey, evolution seems to have spit out all of these creatures that are interacting in this complicated way, but you can look at all of their behavior and trace it back to this objective in some sense of maximizing reproductive fitness. And so are we expecting that IRL on this very grand scale would somehow end up with maximize reproductive fitness. Is that what … Yeah, I think I’m not totally sure what implication you’re drawing from this.

Lucas: Yeah. I guess I’m not arguing that there’s going to be some sort of evolutionary thing which is being optimized.

Rohin: IRL does make the assumption that there is something doing an optimization. You usually have to point it towards what that thing is. You have to say, “Look at the behavior of this particular piece of the environment and tell me what it’s optimizing.” Maybe if you’re imagining IRL on this very grand scale, what is the thing you’re pointing it at?

Lucas: Yeah, so to sort of reiterate and specify, the pointing IRL at the human species would be like to point IRL at 7 billion primates. Similarly, I was thinking that what if one pointed IRL at the ecosystem of Earth over time, you could sort of plot this evolving algorithm over time. So I was just sort of bringing to note that accessing this sort of thing, which seems quite ontologically objective and just sort of clear in this way, it’s just very interesting how it’s fraught with so many difficulties. Yeah, in terms of history it seems like all there really is, is the set of all preferences at each time step over time, which could be summarized in some sort of global or individual levels of algorithms.

Rohin: Got it. Okay. I think I see what you’re saying right now. It seems like the intuition is like ecosystems, universe, laws of physics, very simple, very ontologically basic things, there’s something more real about any value function we could infer from that. And I think this is a misunderstanding of what IRL does. IRL fundamentally requires you to have some notion of counterfactuals. You need to have a description of the action space that some agent had and then when you observe their behavior, you see that they made a choice to take one particular action instead of another particular action.

You need to be able to ask the question of what could they have done instead, which is a counterfactual. Now, with laws of physics, it’s very unclear what the counterfactual would be. With evolution, you can maybe say something like, “Evolution could have chosen to make a whole bunch of mutations and I chose this particular one. And then if you use that particular model, what is IRL going to infer? It will probably infer something like maximized reproductive fitness.”

On the other hand, if you model evolution as like hey, you can design the best possible organism that you can. You can just create an organism out of thin air. And then what reward function are you maximizing then, it’s like super unclear. If you could just poof into existence a organism, you could just make something that’s extremely intelligent, very strong, et cetera, et cetera. And you’re like, well, evolution didn’t do that. It took millions of years to create even humans so clearly it wasn’t optimizing reproductive fitness, right?

And in fact, I think people often say that evolution is not an optimization process because of things like this. The notion of something doing optimization is very much relative to what you assume their capabilities to be and in particular what do you assume their counterfactuals to be. So if you were talking about this sort of grand scale ecosystems, universe, laws of physics, I would ask you like, “What are the counterfactuals? What could the laws of physics done otherwise or what could the ecosystem have done if it didn’t do the thing that it did?” Once you have an answer to that, I imagine I could predict what IRL would do. And that part is the part that doesn’t seem ontologically basic to me, which is why I don’t think that IRL on this sort of thing makes very much sense.

Lucas: Okay. The part here that seems to be a little bit funny to me is where tracking from physics, whatever you take to be ontologically basic about the universe, and tracking from that to the level of whatever our axioms and pre-assumptions for IRL are. What I’m trying to say is in terms of moving from whatever is ontologically basic to the level of agents and we have some assumptions in our IRL where we’re thinking about agents as sort of having theories of counterfactuals where they can choose between actions and they have some sort of reward or objective function that they’re trying to optimize for over time.

It seems sort of metaphysically queer where physics stops … Where we’re going up in levels of abstraction from physics to agents and we … Like physics couldn’t have done otherwise, but somehow agents could have done otherwise. Do you see the sort of concern that I’m raising?

Rohin: Yeah, that’s right. And this is perhaps another reason that I’m more optimistic about the don’t try to do anything at the grand scale and just try to do something that does the right thing locally in our current time, but I think that’s true. It definitely feels to me like optimization, the concept, should be ontologically basic and not a property of human thought. There’s something about how a random universe is high entropy whereas the ones that humans construct is low entropy. That suggests that we’re good at optimization.

It seems like it should be independent of humans. Also, on the other hand, optimization, any conception I come up with it is either specific to the way humans think about it or it seems like it relies on this notion of counterfactuals. And yeah, the laws of physics don’t seem like they have counterfactuals, so I’m not really sure where that comes in. In some sense, you can see that, okay, why do we have the notion of counterfactuals on agency thinking that we could have chosen something else while we’re basically … In some sense we’re just an algorithm that’s continually thinking about what we could do, trying to make plans.

So we search over this space of things that could be done, and that search is implemented in physics, which has no say, it has no counterfactuals, but the search itself, which is an abstraction layer above, it’s something that is running on physics. It is not itself a physics thing, that search is in fact going through multiple options and then choosing one now. It is deterministic from the point of view of physics, but from the point of view of the search, it’s not deterministic. The search doesn’t know which one is going to happen. I think that’s why humans have this notion of choice and of agency.

Lucas: Yeah, and I mean, just in terms of understanding the universe, it’s pretty interesting just how there’s like these two levels of attention where at the physics level you actually couldn’t have done otherwise, but as sort of like this optimization process running on physics that’s searching over space and time and modeling different world scenarios and then seemingly choosing and thus, creating observed behavior for other agents to try and infer whatever reward function that thing is trying to optimize for, it’s an interesting picture.

Rohin: I agree. It’s definitely a sort of puzzles that keep you up at night. But I think one particularly important implication of this is that agency is about how a search process thinks about itself. It’s not just about that because I can look at what someone else is doing and attribute agency to them, figure out that they are themselves running an algorithm that chooses between actions. I don’t have a great story for this. Maybe it’s just humans realizing that other humans are just like them.

So this is maybe why we get acrimonious debates about whether evolution has agency, but we don’t get acrimonious debates about whether humans have agency. Evolution is sufficiently different from us that we can look at the way that it “chooses” “things” and we say, “Oh well, but we understand how it chooses things.” You could model it as a search process, but you could also model it is all that’s happening is this deterministic or mostly deterministic which animals survived and had babies and that is how things happen. And so therefore, it’s not an optimization process. There’s no search. There is deterministic. And so you have these two conflicting views for evolution.

Whereas I can’t really say, “Hey Lucas, I know exactly deterministically how you’re going to do things.” I know this at the sense of like men, there are electrons and atoms and stuff moving around in your brain and electrical signals, but that’s not going to let me predict what you can do. One of the best models I can have of you is just optimizing for some goal, whereas with evolution I can have a more detailed model. And so maybe that’s why I set aside the model of evolution as an optimizer.

Under this setting it’s like, “Okay, maybe our views of agency and optimization are just facts about how well we can model the process, which cuts against the optimization as ontologically basic thing and it seems very difficult. It seems like a hard problem to me. I want to reiterate that most of this has just pushed me to let’s try and instead have a AI alignment focus, try to do things that we understand now and not get into the metaphilosophy problems. If we just get AI systems that broadly do what we want and are asking us for clarification, helping us evolve our thoughts over time, if we can do something like that. I think there are people who would argue that like no, of course, we can’t do something like that.

But if we could do something like that, that seems significantly more likely to work than something that has to have answers to all these metaphilosophical problems today. My position is just that this is doable. We should be able to make systems that are of the nature that I described.

Lucas: There’s clearly a lot of philosophical difficulties that go into IRL. Now it would be sort of good if we could just sort of take a step back and you could summarize your thoughts here on inverse reinforcement learning and the place that it has in AI alignment.

Rohin: I think my current position is something like fairly confidently don’t use IRL to infer a utility function that you then optimize over the long-term. In general, I would say don’t have a utility function that you optimize over the long-term because it doesn’t seem like that’s easily definable right now. So that’s like one class of things I think we should do. On the other hand I think IRL is probably good as a tool.

There is this nice property of IRL that you figure out what someone wants and then you help them do it. And this seems more robust than handwriting, the things that we care about in any particular domain, like even in a simple household robot setting, there are tons and tons of preferences that we have like don’t break vases. Something like IRL could infer these sorts of things.

So I think IRL has definitely a place as a tool that helps us figure out what humans want, but I don’t think the full story for alignment is going to rest on IRL in particular. It gets us good behavior in the present, but it doesn’t tell us how to extrapolate on into the future. Maybe if you did IRL that let you infer how we want the AI system to extrapolate our values or to figure out IRL and our meta-preferences about how the algorithm should infer our preferences or something like this, that maybe could work, but it’s not obvious to me. It seems worth trying at some point.

TLDR, don’t use it for long-term utility function. Do use it as a tool to get decent behavior in the short-term. Maybe also use it as a tool to infer meta-preferences. That seems broadly good, but I don’t know that we know enough about that setting yet.

Lucas: All right. Yeah, that’s all just super interesting and it’s sort of just great to hear how the space is unfolded for you and what your views are now. So I think that we can just sort of pivot here into the AI alignment problem more generally and so now that you’ve moved on from being as excited about IRL, what is essentially capturing your interests currently in the space of AI alignment?

Rohin: The thing that I’m most interested in right now is can we build an AI system that basically evolves over time with us. I’m thinking of this now is like a human AI interaction problem. You’ve got an AI system. We want to figure out how to make it that it broadly helps us, but also at the same time and figures out what it needs to do based on some sort of data that comes from humans. Now, this doesn’t have to be the human saying something. It could be from their behavior. It could be things that they have created in the past. It could be all sorts of things. It could be a reward function that they write down.

But I think the perspective of the things that are easy to infer are the things that are specific to our current environment is pretty important. What I would like to do is build AI systems that refer to preferences in the current environment or things we want in the current environment and do those reasonably well, but don’t just extrapolate to the future and let humans adapt to the future and then figure out what the humans value now and then do things based on that then.

There are a few ways that you could imagine this going. One is this notion of corrigibility in the sense that Paul Christiano writes about it, not the sense that MIRI writes about it, where the AI is basically trying to help you. And if I have an AI that is trying to help me, well, I think one of the most obvious things for someone who’s trying to help me to do is make sure that I remain in effective control of any power resources that might be present that the AI might have and to ask me if my values change in the future or if what I want the AI to do changes in the future. So that’s one thing that you might hope to do.

Also imagine building a norm following AI. So I think human society basically just runs on norms that we mostly all share and tend to follow. We have norms against particularly bad things like murdering people and stealing. We have norms against shoplifting. We have maybe less strong norms against littering. Unclear. And then we also have norms for things are not very consequential. We have norms against randomly knocking over a glass at a restaurant in order to break it. That is also a norm. Even though there are quite often times where I’m like, “Man, it would be fun to just break a glass at the restaurant. It’s very cathartic,” but it doesn’t happen very often.

And so if we could build an AI system that could infer and follow those norms, it seems like this AI would behave in a more human-like fashion. This is a pretty new line of thought so I don’t know whether this works, but it could be that such an AI system is simultaneously behaving in a fashion that humans would find acceptable and also lets us do pretty cool, interesting, new things like developing new technologies and stuff that humans can then deploy and the AI doesn’t just unilaterally deploy without any safety checks or running it by humans or something like that.

Lucas: So let’s just back up a little bit here in terms of the picture of AI alignment. So we have a system that we do not want to extrapolate too much toward possible future values. It seems that there are all these ways in which we can be using the AI first to sort of amplify our own decision making and then also different methodologies which reflect the way that human beings update their own values and preferences over time, something like as proposed by I believe Paul Christiano and Geoffrey Irving and other people at OpenAI, like alignment through debate.

And there’s just all these sorts of epistemic practices of human beings with regards to sort of this world model building and how that affects shifts in value and preferences, also given how the environment changes. So yeah, it just seems like tracking overall these things, finding ways in which AI can amplify or participate in those sort of epistemic practices, right?

Rohin: Yeah. So I definitely think that something like amplification can be thought of as improving our epistemics over time. That seems like a reasonable way to do it. I haven’t really thought very much about how amplification or the pay scales were changing environments. They both operate under this general like we could have a deliberation tree and in principle what we want is this exponentially sized deliberation tree where the human goes through all of the arguments and counter-arguments and breaks those down into sub-points in excruciating detail in a way that no human could ever actually do because it would take way too long.

And then amplification debate basically show you how to get the outcome that this reasoning process would have given by using an AI system to assist the human. I don’t know if I would call it like improving human epistemics, but more like taking whatever epistemics you already have and running it for a long amount of time. And it’s possible like in that long amount of time you actually figure out how to do better epistemics.

I’m not sure that this perspective really talks very much about how preferences change over time. You would hope that it would just naturally be robust to that in that as the environment changes, your deliberation starts looking different in that like okay, now suddenly we have to go back to my example before we have uploads and we’re like egalitarianism now seems to have some really weird consequences. And then presumably the deliberation tree that amplification and debate are mimicking is going to have a bunch of thoughts about do we actually want egalitarianism now, what were the moral intuitions that pushed us towards this? Is there some equivalent principle that lets us keep our moral intuitions, but doesn’t have this weird property where a single person can decide the outcome of an election, et cetera, et cetera.

I think they were not designed to do this, but by a virtue of being based off like how a human would think, what a human would do if they got a long time and a lot of helpful tools to think about it, they’re essentially just inheriting these properties from the human. If the human as the environment would change would start rethinking their priorities or what they care about, then so too would amplification and debate.

Lucas: I think here it also has me thinking about what are the meta-preferences and the meta-meta-preferences and if you could imagine taking a human brain and then running it until the end, through decision and rational and logical thought trees over enough time, with enough epistemics and power behind it to try to sort of navigate its way to the end. It just raises interesting questions about like is that what we want? Is taking that over every single person and then sort of just preference aggregating it all together, is that what we want? And what is the role of moral philosophy for thinking here?

Rohin: Well, so one thing is that whatever moral philosophy you would do so would the amplification of you in theory. I think the benefit of these approaches is that they have this nice property that whatever you would have thought of it in the limit of good AI and idealizations, properly mimicking you and so on, so forth. In this sort of nice world where this all works in a nice, ideal way, it seems like any consideration you can have or you would have so would be agent produced by iterated amplification or debate.

And so if you were going to do a bunch of moral philosophy and come to some sort of decision based on that, so would iterated amplification or debate. So I think it’s like basically here is how we build an AI system that solves the problems in the same way that a human would solve them. And so then if you’re worried about, hey, maybe humans themselves are just not very good at solving problems. Looks like most humans in the world. Like don’t do moral philosophy and don’t extrapolate their values well in the future. And the only reason we have moral progress is because younger generations keep getting born and they have different views than the older generations.

That, I think, could in fact be a problem, but I think there’s hope that we could like train humans to have them nice sort of properties, good epistemics, such they would provide good training data for iterated amplification if there comes a day where we think we can actually train iterated amplification to mimic human explicit reasoning. They do both have the property that they’re only mimicking the explicit reasoning and not necessarily the implicit reasoning.

Lucas: Do you want to unpack that distinction there?

Rohin: Oh, yeah. Sure. So both of them require that you take your high-level question and decompose it into a bunch of sub-questions or sorry, the theoretical model of them has that. This is like pretty clear with iterated amplification. It is less clear with debate. At each point you need to have the top level agent decompose the problem into a bunch of sub-problems. And this basically requires you to be able to decompose tasks into clearly specified sub-tasks, where clearly specified could mean in natural language, but you need to make it explicit in a way that the agent you’re assigning the task to can understand it without having to have your mind.

Whereas if I’m doing some sort of programming task or something, often I will just sort of know what direction to go in next, but not be able to cleanly formalize it. So you’ll give me some like challenging algorithms question and I’ll be like, “Oh, yeah, kind of seems like dynamic programming is probably the right thing to do here.” And maybe if I consider it this particular way, maybe if I put these things in the stack or something, but even the fact that I’m saying this out in natural language is misrepresenting my process.

Really there’s some intuitive not verbalizable process going on in my head. Somehow navigates to the space of possible programs and picks a thing and I think the reason I can do this is because I’ve been programming for a lot of time and I’ve trained a bunch of intuitions and heuristics that I cannot easily verbalize us some like nice decomposition. So that’s sort of implicit in this thing. If you did want that to be incorporated in an iterated amplification, it would have to be incorporated in the base agent, the one that you start with. But if you start with something relatively simple, which I think is often what we’re trying to do, then you don’t get those human abilities and you have to rediscover them in some sense through explicit decompositional reasoning.

Lucas: Okay, cool. Yeah, that’s super interesting. So now to frame all of this again, do you want to sort of just give a brief summary of your general views here?

Rohin: I wish there were a nice way to summarize this. That would mean we’d made more progress. It seems like there’s a bunch of things that people have proposed. There’s amplification/debate, which are very similar, IRL as a general. I think, but I’m not sure, that most of them would agree that we don’t want to like infer a utility function and optimize it for the long-term. I think more of them are like, yeah, we want this sort of interactive system with the human and the AI. It’s not clear to me how different these are and what they’re aiming for in amplification and debate.

So here we’re sort of looking at how things change over time and making that a pretty central piece of how we’re thinking about it. Initially the AI is trying to help the human, human has some sort of reward function, AI trying to learn it and help them, but over time this changes, the AI has to keep up with it. And under this framing you want to think a lot about interaction, you want to think about getting as many bits about reward from the human to the AI as possible. Maybe think about control theory and how human data is in some sense of control mechanism for the AI.

You’d want to like infer norms and ways that people behave, how people relate with each other, try to have your AI systems do that as well. So that’s one camp of things, have the AI interact with humans, behave generally in the way that humans would say is not crazy, update those over time. And then there’s the other side which is like have an AI system that is taking human reasoning, human explicit reasoning and doing that better or doing that more, which allows it to do anything that the human would have done, which is more taking the thought process that humans go through and putting that at the center. That is the thing that we want to mimic and make better.

Sort of parts where our preferences change over time is something that you get for free in some sense by mimicking human thought processes or reasoning. Summary, those are two camps. I am optimistic about both of them, think that people should be doing research on both of them. I don’t really have much more of a perspective of that, I think.

Lucas: That’s excellent. I think that’s a super helpful overview actually. And given that, how do you think that your views of AI alignment have changed over the past few years?

Rohin: I’ll note that I’ve only been in this field for I think 15, 16 months now, so just over a year, but over that year I definitely came into it thinking what we want to do is infer the correct utility function and optimize it. And I have moved away quite strongly from that. I, in fact, recently started writing a value learning sequence or maybe collating is a better word. I’ve written a lot of posts that still have to come out, but I also took a few posts from other people.

The first part of that sequence is basically arguing seems bad to try and define a utility function and then optimize it. So I’m just trying to move away from long-term utility functions in general or long-term goals or things like this. That’s probably the biggest update since starting. Other things that I’ve changed, a focus more on norms than on values, trying to do things that are easy to infer right now in the current environment and that making sure that we update on these over time as opposed to trying to get the one true thing that depends on us solving all the hard metaphilosophical problems. That’s, I think, another big change in the way I’ve been thinking about it.

Lucas: Yeah. I mean, there are different levels of alignment at their core.

Rohin: Wait, I don’t know exactly what you mean by that.

Lucas: There’s your original point of view where you said you came into the field and you were thinking infer the utility function and maximize it. And your current view is now that you are moving away from that and beginning to be more partial towards the view which takes it that we want to be like inferring from norms in the present day just like current preferences and then optimizing that rather than extrapolating towards some ultimate end-goal and then trying to optimize for that. In terms of aligning in these different ways, isn’t there a lot of room for value drift, allowing the thing to run in the real world rather than amplifying explicit human thought on a machine?

Rohin: Value drift if is an interesting question. In some sense, I do want my values to drift in that whatever I think about the correct way that the future should go or something like that today. I probably will not endorse that in the future and I endorse the fact that I won’t endorse it in the future. I do want to learn more and then figure out what to do in the future based on that. You could call that value drift that is a thing. I want to happen. So in that sense then value drift wouldn’t be a bad thing, but then there’s also a sense in which there are ways in which my values could change in the future and ways that I don’t endorse and then that one maybe is value drift. That is bad.

So yeah, if you have an AI system that’s operating in the real world and changes over time as we humans change, yes, there will be changes at what the AI system is trying to achieve over time. You could call that value drift, but value drift usually has a negative connotation, whereas like this process of learning as the environment changes seems to be to me like a positive thing. It’s a thing I would want to do myself.

Lucas: Yeah, sorry, maybe I wasn’t clear enough. In the case of running human beings in the real world, where there are like the causes and effects of history and whatever else and how that actually will change the expression of people over time. Because if you’re running this version of AI alignment where you’re sort of just always optimizing the current set of values in people, progression of the world and of civilization is only as good as the best of all human like values and preferences in that moment.

It’s sort of like limited by what humans are in that specific environment and time, right? If you’re running that in the real world versus running some sort of amplified version of explicit human reasoning, don’t you think that they’re going to come to different conclusions?

Rohin: I think the amplified explicit human reasoning, I imagine that it’s going to operate in the real world. It’s going to see changes that happen. It might be able to predict those changes and then be able to figure out how to respond fast, before the changes even happen perhaps, but I still think of amplification as being very much embedded in the real world. Like you’re asking it questions about things that happen in the real world. It’s going to use explicit reasoning that it would have used if a human were in the real world and thinking about the question.

I don’t really see much of a distinction here. I definitely think that even in my setting where I’m imagining AI systems that evolve over time and change based on that, that they are going to be smarter than humans, going to think through things a lot faster, be able to predict things in advance in the same way that simplified explicit reasoning would. Maybe there are differences, but value drift doesn’t seem like one of them or at least I cannot predict right now how they will differ along the axis of value drift.

Lucas: So then just sort of again taking a step back to the ways in which your views have shifted over the past few years. Is there anything else there that you’d like to touch on?

Rohin: Oh man, I’m sure there is. My views changed so much because I was just so wrong initially.

Lucas: So most people listening should think that if given a lot more thought on this subject, that their views are likely to be radically different than the ones that they currently have and the conceptions that they currently have about AI alignment.

Rohin: Seems true from most listeners, yeah. Not all of them, but yeah.

Lucas: Yeah, I guess it’s just an interesting fact. Do you think this is like an experience of most people who are working on this problem?

Rohin: Probably. I mean, within the first year of working on the problem that seems likely. I mean just in general if you work on the problem, if you start with near no knowledge on something and then you work on it for a year, your views should change dramatically just because you’ve learned a bunch of things and I think that basically explains most of my changes in view.

It’s just actually hard for me to remember all the ways in which I was wrong back in the past and I focused on not using utility functions because I think that even other people in the field still believe right now. So that’s where that one came from, but there are like plenty of other things that are just notably, easily, demonstrably wrong about that I’m having trouble recalling now.

Lucas: Yeah, and the utility function one I think is a very good example and I think that if it were possible to find all of these in your brain and distill them, I think it would make a very, very good infographic on AI alignment, because those misconceptions are also misconceptions that I’ve had and I share those and I think that I’ve seen them also in other people. A lot of sort of the intellectual blunders that you or I have made are probably repeated quite often.

Rohin: I definitely believe that. Yeah, I guess I could talk about the things that I’m going to very soon saying the value learning sequence. Those were definitely updates that I made, one of those a utility functions thing. Another one was thinking about what we want is for the human AI system as a whole to be optimizing for some sort of goal. And this opens up a nice space of possibilities where the AI is not optimizing a goal, only the human AI system together is. Keeping in mind that that is the goal and not just the AI itself must be optimizing some sort of goal.

The idea of corrigibility itself as a thing that we should be aiming for was a pretty big update for me, took a while for me to get to that one. I think distributional shift was a pretty key concept that I learned at some point and started applying everywhere. One way of thinking about the evolving preferences over time thing is that humans, they’ve been trained on the environment that we have right now and arguably we’ve been trained on the ancestral environment too by evolution, but we haven’t been trained on whatever the future is going to be.

Or for a more current example, we haven’t been trained on social media. Social media is a fairly new thing affecting us in ways that we hadn’t considered in the past and this is causing us to change how we do things. So in some sense what’s happening is as we go into the future, we’re encountering a distributional shift and human values don’t extrapolate well to that distributional shift. What do you actually need to do is wait for the humans to get to that point, let them experience it, train on it, have their values be trained on this new distribution and then figure out what they are rather than trying to do it right now when their values are just going to be wrong or going to be not what they would get if they were actually in that situation.

Lucas: Isn’t that sort of summarizing coherent extrapolated volition?

Rohin: I don’t know that coherent extrapolated volition explicitly talks about having the human be in a new environment. I guess you could imagine that CEV considers … If you imagine like a really, really long process of deliberation in CEV, then you could be like, okay what would happen if I were in this environment and all these sorts of things happened. It seems like you would need to have a good model of how the world works and how physics works in order to predict what the environment would be like. Maybe you can do that and then in that case you simulate a bunch of different environments and you think about how humans would adapt and evolve and respond to those environments and then you take all of that together and you summarize it and distill it down into a single utility function.

Plausibly that could work. Doesn’t seem like a thing we can actually build, but as a definition of what we might want, that seems not bad. I think that is me putting the distributional shift perspective on CEV and it was not, certainly not obvious to me from the statement of CEV itself, that you’re thinking about how to mitigate the impact of distributional shift on human values. I think I’ve had this perspective and I’ve put it on CEV and I’m like, yeah, that seems fine, but it was not obvious to me from reading about CEV alone.

Lucas: Okay, cool.

Rohin: I recently posted a comment on the Alignment Forum talking about how we want to like … I guess this is sort of in corrigibility ability too, making an AI system that tries to help us as opposed to making an AI system that is optimizing the one true utility function. So that was an update I made, basically the same update as the one about aiming for corrigibility. I guess another update I made is that while there is a phase transition or something or like a sharp change in the problems that we see when AIs become human level or super-intelligent, I think the underlying causes of the problems don’t really change.

Underlying causes of problems with narrow AI systems, probably similar to the ones that underlie a super intelligent systems. Having their own reward function leads to problems both in narrow settings and in super-intelligent settings. This made me more optimistic about doing work trying to address current problems, but with an eye towards long-term problems.

Lucas: What made you have this update?

Rohin: Thinking about the problems a lot, in particular thinking about how they might happen in current systems as well. So I guess a prediction that I would make is that if it is actually true that superintelligence would end up killing us all or something like that, some like really catastrophic outcome. Then I would predict that before that, we will see some AI system that causes some other smaller scale catastrophe where I don’t know what catastrophe means, it might be something like oh, you humans die or oh, the power grid went down for some time or something like that.

And then before that we will have things that sort of fail in relatively not important ways, but in ways of say that like here’s an underlying problem that we need to fix with how we build AI systems. If you extrapolate all the way back to today that looks like for example to boat racing example from open AI, a reward hacking one. So generally expecting things to be more continuous. Not necessarily slow, but continuous. That update I made because of the posts arguing for slow take off from Paul Christiano and AI impacts.

Lucas: Right. And the view there is sort of that the world will be propagated with lower-level ML as we sort of start to ratchet up the capability of intelligence. So a lot of tasks will sort of be … Already being done by systems that are slightly less intelligent than the current best system. And so all work ecosystems will already be fully flooded with AI systems optimizing within the spaces. So there won’t be a lot of space for the first AGI system or whatever to really get decisive strategic advantage.

Rohin: Yeah, would I make prediction that we won’t have a system that gets a decisive strategic advantage? I’m not sure about that one. It seems plausible to me that we have one AI system that is improving over time and we use those improvements in society for before it becomes super intelligent. But then by the time it becomes super intelligent, it is still the one AI system that is super intelligent. So it does gain a decisive strategic advantage.

An example of this would be if there was just one main AGI project, I would still predict that progress on AI, it would be continuous, but I would not predict a multipolar outcome in that scenario. The corresponding view is that while I still do use the terminology first AGI because it’s like pointing out some intuitive concept that I think is useful, it’s a very, very fuzzy concept and I don’t think we’ll be able to actually point at any particular system and say that was the first AGI. Rather we’ll point to like a broad swath of time and say, “Somewhere in there AI had became generally intelligent.”

Lucas: There are going to be all these sort of like isolated meta-epistemic reasoning tools which can work in specific scenarios, which will sort of potentially aggregate in that fuzzy space to create something fully general.

Rohin: Yep. They’re going to be applied in some domains and then the percent of domains in which they apply will gradually grow grutter and eventually we’ll be like, huh, looks like there’s nothing left for humans to do. It probably won’t be a surprise, but I don’t think there will be a particular point where everyone agrees, yep, looks like AI is going to automate everything in just a few years. It’s more like AI will start automating a bunch of stuff. The amount of stuff it automates will increase over time. Some people will see it coming, see full automation coming earlier, some people will be like nah, this is just a simple task that AI can do, still got a long ways to go for all the really generally intelligent stuff. People will sign on to like oh, yeah, it’s actually becoming generally intelligent at different spots.

Lucas: Right. If you have a bunch of small mammalian level AIs automating a lot of stuff in industry, there would likely be a lot of people whose timelines would be skewed in the wrong direction.

Rohin: I’m not even sure this was a point of timelines. It was just a point of like which is the system that you call AGI. I claim this will not have a definitive answer. So that was also an update to how I was thinking. That one, I think, is like more generally accepted in the community. And this was more like well, all of the literature on the AI safety that’s publicly available and like commonly read by EA’s doesn’t really talk about these sorts of points. So I just hadn’t encountered these things when I started out. And then I encountered a more maybe I thought to myself, I don’t remember, but like once I encountered the arguments I was like, yeah, that makes sense and maybe I should have thought of that before.

Lucas: In the sequence which you’re writing, do you sort of like cover all of these items which you didn’t think were in the mainstream literature?

Rohin: I cover some of them. The first few things I told you were I was just like what did I say in the sequence. There were a few I think that probably aren’t going to be in that sequence just because there’s a lot of stuff that people have not written down.

Lucas: It’s pretty interesting because the way in which the AI alignment field is evolving is sometimes, it’s often difficult to have a bird’s-eye view of where it is and track avant-guard ideas being formulated in people’s brains and being shared.

Rohin: Yeah. I definitely agree. I was hoping that the Alignment Newsletter, which I write, to help with that. I would say it probably speeds up the process of bit, but it’s definitely not keeping you on the forefront. There are many ideas that I’ve heard about, that I’ve even read documents about that haven’t made it in the newsletter yet because they haven’t become public.

Lucas: So how many months behind do you think for example, the newsletter would be?

Rohin: Oh, good question. Well, let’s see. There’s a paper that I started writing in May or April that has not made it into the newsletter yet. There’s a paper that I finished and submitted in October that has not made it to the newsletter yet, or was it September, possibly September. That one will come out soon. That suggests a three month lag. But I think many others have been longer than that. Admittedly, this is for academic researchers at CHAI. I think CHAI is like we tend to publish using papers and not blog posts and this results in the longer delay on our side.

Also because work on relative reachability, for example, I’ve learned about quite a bit. I learned about maybe four or five months before she released it and that’s when it came out in the newsletter. And of course, she’d been working on it for longer or like AI safety by debate I think I learned about six or seven months before it was published in came out in the newsletter. So yeah, somewhere between three months and half a year for things seems likely. For things that I learned from MIRI, it’s possible that they never get into the newsletter because they’re never made public. So yeah, there’s a fairly broad range there.

Lucas: Okay. That’s quite interesting. I think that also sort of gives people a better sense of what’s going on in technical AI alignment because it can seem kind of black boxy.

Rohin: Yeah. I mean, in some sense this is a thing that all fields have. I used to work in programming languages. On there we would often write a paper and submit it and then go and present about it a year later by the time we had moved on, done a whole other project and written other paper and then we’d go back and we’d talk about this. I definitely remember sometimes grad students being like, “Hey, I want to get this practice document.” I say, “What’s it about?” It’s like some topic. And I’m like wait, but you did that. I heard about this like two years ago. And they’re like, yep, just got published.

So in that sense, I think both AI is faster and AI alignment is I think even faster than AI because it’s a smaller field and people can talk to each other more, and also because a lot of us write blog posts. Blog posts are great.

Lucas: They definitely play a crucial role within the community in general. So I guess just sort of tying things up a bit more here, pivoting back to a broader view. Given everything that you’ve learned and how your ideas have shifted, what are you most concerned about right now in AI alignment? How are the prospects looking to you and how does the problem of AI alignment look right now to Rohin Shah?

Rohin: I think it looks pretty tractable, pretty good. Most of the problems that I see are I think ones that we can see in advance, we probably can solve. None of these seem like particularly impossible to me. I think I also give more credit to the machine learning community or AI community than other researchers do. I trust in our ability where here are meaning like the AI field broadly, our ability to notice what things could go wrong and fix them in a way that maybe other researchers in the AI safety don’t.

I think one of the things that feels most problematic to me right now is the problem of inner optimizers, which I’m told there will probably be a sequence on in the future because there aren’t great resources on it right now. So basically this is the idea of if you run a search process over a wide space of strategies or options and you search for something that gets you good external reward or something like that, what you might end up finding is a strategy that is itself a consequentialist agent that’s optimizing for its own internal reward and that internal reward will agree with the external reward on the training data because that’s why it was selected, but it might diverge soon as there’s any distribution shift.

And then it might start optimizing against us adversarially in the same way that you would get if you like gave a misspecified award function to and RL system today. This seems plausible to me. I’ve read a bit more about this and talk to people about this and things that aren’t yet public, but hopefully will soon be. I definitely recommend reading that if it ever comes out, but yeah, this seems like it could be a problem. I don’t think we have any instance of it being a problem yet. Seems hard to detect and I’m not sure how I would fix it right now.

But I also don’t think that we’ve thought about the problem or I don’t think I’ve thought about the problem that much. I don’t want to say like, “Oh man, this is totally unsolvable,” yet. Maybe I’m just an optimistic person by nature. I mean, that’s definitely true, but maybe that’s biasing my judgment here. Feels like we could probably solve that if it ends up being a problem.

Lucas: Is there anything else here that you would like to wrap up on in terms of AI alignment or inverse reinforcement learning?

Rohin: I want to continue to exhort that we should not be trying to solve all the metaphilosophical problems and we should not be trying to like infer the one true utility function and we should not be modeling an AI as pursuing a single goal over the long-term. That is a thing I want to communicate to everybody else. Apart from that I think we’ve covered everything at a good depth. Yeah, I don’t think there’s anything else I’d add to that.

Lucas: So given that I think rather succinct distillation of what we are trying not to do, could you try and offer an equally succinct distillation of what we are trying to do?

Rohin: I wish I could. That would be great, wouldn’t it? I can tell you that I can’t do that. I could give you like a suggestion on what we are trying to do instead, which would be try to build an AI system that is corrigible, that is doing what we want, but it’s going to remain under human control in some sense. It’s going to ask us, take our preferences into account, not try to go off behind our backs and optimize against us. That is a summary of a path that we could go down that I think is premised or what I would want our AI systems to be like. But that’s unfortunately very sparse on concrete details because I don’t know those concrete details yet.

Lucas: Right. I think that that sort of perspective shift is quite important. I think it changes the nature of the problem and how one thinks about the problem, even at the societal level.

Rohin: Yeah. Agreed.

Lucas: All right. So thank you so much Rohin, it’s really been a pleasure. If people are interested in checking out some of this work that we have mentioned or following you, where’s the best place to do that?

Rohin: I have a website. It is just RohinShah.com. Subscribing to the Alignment Newsletter is … Well, it’s not a great way to figure out what I personally believe. Maybe if you’d keep reading the newsletter over time and read my opinions for several weeks in a row, maybe then you’d start getting a sense of what Rohin thinks. It will soon have links to my papers and things like that, but yeah, that’s probably the best way on this, like my website. I do have a Twitter, but I don’t really use it.

Lucas: Okay. So yeah, thanks again Rohin. It’s really been a pleasure. I think that was a ton to think about and I think that I probably have a lot more of my own thinking and updating to do based off of this conversation.

Rohin: Great. Love it when that happens.

Lucas: So yeah. Thanks so much. Take care and talk again soon.

Rohin: All right. See you soon.

Lucas: If you enjoyed this podcast, please subscribe, give it a like or share it on your preferred social media platform. We’ll be back again soon with another episode in the AI Alignment series.

[end of recorded material]

Lucas: Hey, everyone. Welcome back to the AI Alignment Podcast series at the Future of Life Institute. I’m Lucas Perry, and today we’ll be speaking with William MacAskill on moral uncertainty and its place in AI alignment. If you’ve been enjoying this series and finding it interesting or valuable, it’s a big help if you can share it on social media and follow us on your preferred listening platform.

Will is a professor of philosophy at the University of Oxford and is a co-founder of the Center for Effective Altruism, Giving What We Can, and 80,000 Hours. Will helped to create the effective altruism movement and his writing is mainly focused on issues of normative and decision theoretic uncertainty, as well as general issues and ethics. And so, without further ado, I give you William MacAskill.

Yeah, Will, thanks so much for coming on the podcast. It’s really great to have you here.

Will: Thanks for having me on.

Lucas: So, I guess we can start off. You can tell us a little bit about the work that you’ve been up to recently in terms of your work in the space of metaethics and moral uncertainty just over the past few years and how that’s been evolving.

Will: Great. My PhD topic was on moral uncertainty, and I’m just putting the finishing touches on a book on this topic. The idea here is to appreciate the fact that we very often are just unsure about what we ought, morally speaking, to do. It’s also plausible that we ought to be unsure about what we ought morally to do. Ethics is a really hard subject, there’s tons of disagreement, it would be overconfident to think, “Oh, I’ve definitely figured out the correct moral view.” So my work focuses on not really the question of how unsure we should be, but instead what should we do given that we’re uncertain?

In particular, I look at the issue of whether we can apply the same sort of reasoning that we apply to uncertainty about matters of fact to matters of moral uncertainty. In particular, can we use what is known as “expected utility theory”, which is very widely accepted as at least approximately correct in empirical uncertainty. Can we apply that in the same way in the case of moral uncertainty?

Lucas: Right. And so coming on here, you also have a book that you’ve been working on on moral uncertainty that is unpublished. Have you just been expanding this exploration in that book, diving deeper into that?

Will: That’s right. There’s actually been very little that’s been written on the topic of moral uncertainty, at least in modern times, at least relative to its importance. I would think of this as a discipline that should be studied as much as consequentialism or contractualism or Kantianism is studied. But there’s really, in modern times, only one book that’s been written on the topic and that was written 18 years ago now, or published 18 years ago. What we want is this to be, firstly, just kind of definitive introduction to the topic, it’s co-authored with me as lead author, but co-authored with Toby Ord and Krista Bickfest, laying out both what we see as the most promising path forward in terms of addressing some of the challenges that face an account of decision-making under moral uncertainty, some of the implications of taking moral uncertainty seriously, and also just some of the unanswered questions.

Lucas: Awesome. So I guess, just moving forward here, you have a podcast that you already did with Rob Wiblin: 80,000 Hours. So I guess we can sort of just avoid covering a lot of the basics here about your views on using expected utility calculous in moral reasoning and moral uncertainty in order to decide what one ought to do when one is not sure what one ought to do. People can go ahead and listen to that podcast, which I’ll provide a link to within the description.

It would also be good, just to sort of get a general sense of where your meta ethical partialities just generally right now tend to lie, so what sort of meta ethical positions do you tend to give the most credence to?

Will: Okay, well that’s a very well put question ’cause, as with all things, I think it’s better to talk about degrees of belief rather than absolute belief. So normally if you ask a philosopher this question, we’ll say, “I’m a nihilist,” or “I’m a moral realist,” or something, so I think it’s better to split your credences. So I think I’m about 50/50 between nihilism or error theory and something that’s non-nihilistic.

Whereby nihilism or error theory, I just mean that any positive moral statement or normative statement or a evaluative statement. That includes, you ought to maximize happiness. Or, if you want a lot of money, you ought to become a banker. Or, pain is bad. That, on this view, all of those things are false. All positive, normative or evaluative claims are false. So it’s a very radical view. And we can talk more about that, if you’d like.

In terms of the rest of my credence, the view that I’m kind of most sympathetic towards in the sense of the one that occupies most of my mental attention is a relatively robust form of moral realism. It’s not clear whether it should be called kind of naturalist moral realism or non-naturalist moral realism, but the important aspect of it is just that goodness and badness are kind of these fundamental moral properties and are properties of experience.

The things that are of value are things that supervene on conscious states, in particular good states or bad states, and the way we know about them is just by direct experience with them. Just by being acquainted with a state like pain gives us a reason for thinking we ought to have less of this in the world. So that’s my kind of favored view in the sense it’s the one I’d be most likely to defend in the seminar room.

And then I give somewhat less credence in a couple of views. One is a view called “subjectivism” which is the idea that what you ought to do is determined in some sense by what you want to do. So the simplest view there would just be when I say, “I ought to do X.” That just means I want to do X in some way. Or a more sophisticated version would be ideal subjectivism where when I say I ought to do X, it means some very idealized version of myself would want myself to want to do X. Perhaps if I had limited amounts of knowledge and much clearer computational power and so on. I’m a little less sympathetic to that than many people I know. We’ll go into that.

And then a final view that I’m also less sympathetic towards is non-cognitivism, which would be the idea that our moral statements … So when I say, “Murder is wrong,” I’m not even attempting to express a proposition. What they’re doing is just expressing some emotion of mine, like, “Yuk. Murder. Ugh,” in the same way that when I said that, that wasn’t expressing any proposition, it was just expressing some sort of pro or negative attitude. And again, I don’t find that terribly plausible, again for reasons we can go into.

Lucas: Right, so those first two views were cognitivist views, which makes them fall under sort of a semantic theory where you think that people are saying truth or false statements when they’re claiming moral facts. And the other theory in your moral realism are both metaphysical views, which I think is probably what we’ll mostly be interested here in terms of the AI alignment problem.

There are other issues in metaethics, for example having to do with semantics, as you just discussed. You feel as though you give some credence to non-cognitivism, but there are also justification views, so like issues in moral epistemology, how one can know about metaethics and why one ought to follow metaethics if metaethics has facts. Where do you sort of fall in in that camp?

Will: Well, I think all of those views are quite well tied together, so what sort of moral epistemology you have depends very closely, I think, on what sort of meta-ethical view you have, and I actually think, often, is intimately related as well to what sort of view in normative ethics you have. So my preferred philosophical world view, as it were, the one I’d defend in a seminar room, is classical utilitarian in its normative view, so the only thing that matters is positive or negative mental states.

In terms of its moral epistemology, the way we access what is of value is just by experiencing it, so in just the same way we access conscious states. There are also some ways in which you can’t merely, you know, why is it that we should maximize the sum of good experiences rather than the product, or something? That’s a view that you’ve got to obtain by kind of reasoning rather than just purely from experience.

Part of my epistemology does appeal to whatever this spooky ability we have to reason about abstract affairs, but it’s the same sort of faculty that is used when we think about mathematics or set theory or other areas of philosophy. If, however, I had some different view, so supposing we were a subjectivist, well then moral epistemology looks very different. You’re actually just kind of reflecting on your own values, maybe looking at what you would actually do in different circumstances and so on, reflecting on your own preferences, and that’s the right way to come to the right kind of moral views.

There’s also another meta-ethical view called “constructivism” that I’m definitely not the best person to talk about with. But on that view, again it’s not really a realistic view, but on this view we just have a bunch of beliefs and intuitions and the correct moral view is just the best kind of systematization of those and beliefs or intuitions in the same way as you might think … Like linguistics, it is a science, but it’s fundamentally based just on what our linguistic intuitions are. It’s just kind of a systematization of them.

On that view, then, moral epistemology would be about reflecting on your own moral intuitions. You just got all of this data, which is the way things seem like to you, morally speaking, and then you’re just doing the systematization thing. So I feel like the question of moral epistemology can’t be answered in a vacuum. You’ve got to think about your meta-ethical view of the metaphysics of ethics at the same time.

Lucas: I think I’m pretty interested in here, and also just poking a little bit more into that sort of 50% credence you give to your moral realist view, which is super interesting because it’s a view that people tend not to have, I guess, in the AI computer science rationality space, EA space. People tend to, I guess, have a lot of moral anti-realists in this space.

In my last podcast, I spoke with David Pearce, and he also seemed to sort of have a view like this, and I’m wondering if you can just sort of unpack yours a little bit, where he believed that suffering and pleasure disclose the in-built pleasure/pain access of the universe. Like you can think of minds as sort of objective features of the world, because they in fact are objective features of the world, and the phenomenology and experience of each person is objective in the same way that someone could objectively be experiencing redness, and in the same sense they could be objectively experiencing pain.

It seems to me, and I don’t fully understand the view, but the claim is that there are some sort of in-built quality or property to the hedonic qualia of suffering or pleasure that discloses its in-built value to that.

Will: Yeah.

Lucas: Could you unpack it a little bit more about the metaphysics of that and what that even means?

Will: It sounds like David Pearce and I have quite similar views. I think relying heavily on the analogy with, or very close analogy with consciousness is going to help, where imagine you’re kind of a robot scientist, you don’t have any conscious experiences but you’re doing all this fancy science and so on, and then you kind of write out the book of the world, and i’m like, “hey, there’s this thing you missed out. It’s like conscious experience.” And you, the robot scientist, would say, “Wow, that’s just insane. You’re saying that some bits of matter have this first person subjective feel to them? Like, why on earth would we ever believe that? That’s just so out of whack with the naturalistic understanding of the world.” And it’s true. It just doesn’t make any sense from given what we know now. It’s a very strange phenomenon to exist in the world.

Will: And so one of the arguments that motivates error theory is this idea of just, well, if values were to exist, they would just be so weird, what Mackie calls “queer”. It’s just so strange that just by a principle of Occam’s razor not adding strange things in to our ontology, we should assume they don’t exist.

But that argument would work in the same way against conscious experience, and the best response we’ve got is to say, no, but I know I’m conscious, and just tell by introspecting. I think we can run the same sort of argument when it comes to a property of consciousness as well, which is namely the goodness or badness of certain conscious experiences.

So now I just want you to go kind of totally a-theoretic. Imagine you’ve not thought about philosophy at all, or even science at all, and I was just to ask you, rip off one of your fingernails, or something. And then I say, “Is that experience bad?” And you would say yes.

Lucas: Yeah, it’s bad.

Will: And I would ask, how confident are you? The more confident that this pain is bad than that I even have hands, perhaps. That’s at least how it seems to be for me. So then it seems like, yeah, we’ve got this thing that we’re actually incredibly confident of which is the badness of pain, or at least the badness of pain for me, and so that’s what initially gives the case for then thinking, okay, well, that’s at least one objective moral fact that pain is bad, or at least pain is bad for me.

Lucas: Right, so the step where I think that people will tend to get lost in this is when … I thought the part about Occam’s razor was very interesting. I think that most people are anti-realistic because they use Occam’s razor there and they think that what the hell would a value even be anyway in the third person objective sense? Like, that just seems really queer, as you put it. So I think people get lost at the step where the first person seems to simply have a property of badness to it.

I don’t know what that would mean if one has a naturalistic reductionist view of the world. There seems to be just like entropy, noise and quarks and maybe qualia as well. It’s not clear to me how we should think about properties of qualia and whether or not one can drive, obviously, “ought” statements about properties of qualia to normative statements, like “is” statements about the properties of qualia to “ought” statements?

Will: One thing I want to be very clear on is just it definitely is the case that we have really no idea on this view. We are currently completely in the dark about some sort of explanation of how matter and forces and energy could result in goodness or badness, something that ought to be promoted. But that’s also true with conscious experience as well. We have no idea how on earth matter could result in kind of conscious experience. At the same time, it would be a mistake to start denying conscious experience.

And then we can ask, we say, okay, we don’t really know what’s going on but we accept that there’s conscious experience, and then I think if you were again just to completely pre theoretically start categorizing distant conscious experiences that we have, we’d say that some are red and some are blue, some are maybe more intense, some are kind of dimmer than others, you’d maybe classify them into sights and sounds and other sorts of experiences there.

I think also a very natural classification would be the ones that are good and the ones that are bad, and then I think when we cash that out further, I think it’s not nearly the case. I don’t think the best explanation is that when we say, oh, this is good or this is bad it means what we want or what we don’t want, but instead it’s like what we think we have reason to want or reason not to want. It seems to give us evidence for those sorts of things.

Lucas: I guess my concern here is just that I worry that words like “good” and “bad” or “valuable” or “dis-valuable”, I feel some skepticism about whether or not they disclose some sort of intrinsic property of the qualia. I’m also not sure what the claim here is about the nature of and kinds of properties that qualia can have attached to them. I worry that goodness and badness might be some sort of evolutionary fiction which enhances us, enhances our fitness, but it doesn’t actually disclose some sort of intrinsic metaphysical quality or property of some kind of experience.

Will: One thing I’ll say is, again, remember that I’ve got this 50% credence on error theory, so in general, all these questions, maybe this is just some evolutionary fiction, things just seem bad but they’re not actually, and so on. I actually think those are good arguments, and so that should give us confidence, some degree of confidence and this idea of just actually nothing matters at all.

But kind of underlying a lot of my views is this more general argument that if you’re unsure between two views, one in which just nothing matters at all, we’ve got no reasons for action, the other one we do have some reasons for action, then you can just ignore the one that says you’ve got no reasons for action ’cause you’re not going to do badly by its likes no matter what you do. If I were to go around shooting everybody, that wouldn’t be bad or wrong on nihilism. If I were to shoot lots of people, it wouldn’t be bad or wrong on nihilism.

So if there are arguments such as, I think an evolutionary argument that pushes us in the direction of kind of error theory, in a sense we can put them to the side, ’cause what we ought to do is just say, yeah, we take that really seriously. Give us a high credence in error theory, but now say, after all those arguments, what are the views, because most plausibly kind of bear their force.

So this is why with the kind of evolutionary worry, I’m just like, yes. But, supposing it’s the case that there actually are. Presumably conscious experiences themselves are useful in some evolutionary way that, again, we don’t really understand. I think, presumably, also good and bad experiences are useful in some evolutionary way that we don’t fully understand, perhaps because they have a tendency to motivate at least beings like us, and that in fact seems to be a key aspect of making a kind of goodness or badness statement. It’s at least somehow tied up to the idea of kind of motivation.

And then when I say ascribing a property to a conscious experience, I really just don’t mean whatever it is that we mean when we say that this experience is red seeming, this is experience is blue seeming, I mean, again, opens philosophical questions what we even mean by properties but in the same way this is bad seeming, this is good seeming.

Before I got into thinking about philosophy and naturalism and so on, would I have thought those things are kind of on a par, and I think I would’ve done, so it’s at least a pre theoretically justified view to think, yeah, there just is this axiological property of my experience.

Lucas: This has made me much more optimistic. I think after my last podcast I was feeling quite depressed and nihilistic, and hearing you give this sort of non-naturalistic or naturalistic moral realist count is cheering me up a bit about the prospects of AI alignment and value in the world.

Will: I mean, I think you shouldn’t get too optimistic. I’m also certainly wrong-

Lucas: Yeah.

Will: … sort of is my favorite view. But take any philosopher. What’s the chance that they’ve got the right views? Very low, probably.

Lucas: Right, right. I think I also need to be careful here that human beings have this sort of psychological bias where we give a special metaphysical status and kind of meaning and motivation to things which have objective whatever to it. I guess there’s also some sort of motivation that I need to be mindful of that seeks out to make value objective or more meaningful and foundational in the universe.

Will: Yeah. The thing that I think should make you feel optimistic, or at least motivated, is this argument that if nothing matters, it doesn’t matter that nothing matters. It just really ought not to affect what you do. You may as well act as if things do matter, and in fact we can have this project of trying to figure out if things matter, and that maybe could be an instrumental goal, which kind of is a purpose for life is to get to a place where we really can figure out if it has any meaning. I think that sort of argument can at least give one grounds for getting out of bed in the morning.

Lucas: Right. I think there’s this philosophy paper that I saw, but I didn’t read, that was like, “nothing Matters, but it does matter”, with the one lower case M and then another capital case M, you know.

Will: Oh, interesting.

Lucas: Yeah.

Will: It sounds a bit like 4:20 ethics.

Lucas: Yeah, cool.

Moving on here into AI alignment. And before we get into this, I think that this is something that would also be interesting to hear you speak a little bit more about before we dive into AI alignment. What even is the value of moral information and moral philosophy, generally? Is this all just a bunch of BS or how can it be interesting and or useful in our lives, and in science and technology?

Will: Okay, terrific. I mean, and this is something I write about in a paper I’m working on now and also in the book, as well.

So, yeah, I think the stereotype of the philosopher engaged in intellectual masturbation, not doing really much for the world at all, is quite a prevalent stereotype. I’ll not comment on whether that’s true for certain areas of philosophy. I think it’s definitely not true for certain areas within ethics. What is true is that philosophy is very hard, ethics is very hard. Most of the time when we’re trying to do this, we make very little progress.

If you look at the long-run history of thought in ethics and political philosophy, the influence is absolutely huge. Even just take Aristotle, Locke, Hobbes, Mill, and Marx. The influence of political philosophy and moral philosophy there, it shaped thousands of years of human history. Certainly not always for the better, sometimes for the worse, as well. So, ensuring that we get some of these ideas correct is just absolutely crucial.

Similarly, even in more recent times … Obviously not as influential as these other people, but also it’s been much less time so we can’t predict into the future, but if you consider Peter Singer as well, his ideas about the fact that we may have very strong obligations to benefit those who are distant strangers to us, or that we should treat animal welfare just on a par with human welfare, at least on some understanding of those ideas, that really has changed the beliefs and actions of, I think, probably tens of thousands of people, and often in really quite dramatic ways.

And then when we think about well, should we be doing more of this, is it merely that we’re influencing things randomly, or are we making things better or worse? Well, if we just look to the history of moral thought, we see that most people in most times have believed really atrocious things. Really morally abominable things. Endorsement of slavery, distinctions between races, subjugation of women, huge discrimination against non-heterosexual people, and, in part at least, it’s been ethical reflection that’s allowed us to break down some of those moral prejudices. And so we should presume that we have very similar moral prejudices now. We’ve made a little bit of progress, but do we have the one true theory of ethics now? I certainly think it’s very unlikely. And so we need to think more if we want to get to the actual ethical truth, if we don’t wanna be living out moral catastrophes in the same way as we would if we kept slaves, for example.

Lucas: Right, I think we do want to do that, but I think that a bit later in the podcast we’ll get into whether or not that’s even possible, given economic, political, and militaristic forces acting upon the AI alignment problem and the issues with coordination and race to AGI.

Just to start to get into the AI alignment problem, I just wanna offer a little bit of context. It is implicit in the AI alignment problem, or value alignment problem, that AI needs to be aligned to some sort of ethic or set of ethics, this includes preferences or values or emotional dispositions, or whatever you might believe them to be. And so it seems that generally, in terms of moral philosophy, there are really two methods, or two methods in general, by which to arrive at an ethic. So, one is simply going to be through reason, and one is going to be through observing human behavior or artifacts, like books, movies, stories, or other things that we produce in order to infer and discover the observed preferences and ethics of people in the world.

The latter side of alignment methodologies are empirical and involves the agent interrogating and exploring the world in order to understand what the humans care about and value, as if values and ethics were simply a physical by-product of the world and of evolution. And the former is where ethics are arrived at through reason alone, and involve the AI or the AGI potentially going about ethics as a philosopher would, where one engages in moral reasoning about metaethics in order to determine what is correct. From the point of view of ethics, there is potentially only what the humans empirically do believe and then there is what we may or may not be able to arrive at through reason alone.

So, it seems that one or both of these methodologies can be used when aligning an AI system. And again, the distinction here is simply between sort of preference aggregation or empirical value learning approaches, or methods of instantiating machine ethics, reasoning, or decision-making in AI systems so they become agents of morality.

So, what I really wanna get into with you now is how metaethical uncertainty influences our decision over the methodology of value alignment. Over whether or not we are to prefer an empirical preference learning or aggregation type approach, or one which involved an imbuing of moral epistemology and ethical metacognition and reasoning into machine systems so it can discover what we ought to do. And how moral uncertainty, and metaethical moral uncertainty in particular, operates within both of these spaces once you’re committed to some view, or both of these views. And then we can get into issues and intertheoretic comparisons and how that arises here at many levels, the ideal way we should proceed if we could do what would be perfect, and again, what is actually likely to happen given race dynamics and political, economic, and militaristic forces.

Will: Okay that sounds terrific. I mean, there’s a lot of cover there.

I think it might be worth me saying just maybe a couple of distinctions I think are relevant and kind of my overall view in this. So, in terms of distinction, I think within what broadly gets called the alignment problem, I think I’d like to distinguish between what I’d call the control problem, then kind of human values alignment problem, and then the actual alignment problem.

Where the control problem is just, can you get this AI to do what you want it to do? Where that’s maybe relatively narrowly construed, I want it to clean up my room, I don’t want it to put my cat in the bin, that’s kinda control problem. I think describing that as a technical problem is kind of broadly correct.

Second is then what gets called aligning AI with human values. For that, it might be the case that just having the AI pay attention to what humans actually do and infer their preferences that are revealed on that basis, maybe that’s a promising approach and so on. And that I think will become increasingly important as AI becomes larger and larger parts of the economy.

This is kind of already what we do when we vote for politicians who represent at least large chunks of the electorate. They hire economists who undertake kind of willingness-to-pay surveys and so on to work out what people want, on average. I do think that this is maybe more normatively loaded than people might often think, but at least you can understand that, just as the control problem is I have some relatively simple goal, which is, what do I want? I want this system to clean my room. How do I ensure that it actually does that without making mistakes that I wasn’t intending? This is kind of broader problem of, well you’ve got a whole society and you’ve got to aggregate their preferences for what kind of society wants and so on.

But I think, importantly, there’s this third thing which I called a minute ago, the actual alignment problem, so let’s run with that. Which is just working out what’s actually right and what’s actually wrong and what ought we to be doing. I do have a worry that because many people in the wider world, often when they start thinking philosophically they start endorsing some relatively simple, subjectivist or relativist views. They might think that answering this question of well, what do humans want, or what do people want, is just the same as answering what ought we to do? Whereas for kind of the reductio of that view, just go back a few hundred years where the question would have been, well, the white man’s alignment problem, where it’s like, “Well, what do we want, society?”, where that means white men.

Lucas: Uh oh.

Will: What do we want them to do? So similarly, unless you’ve got the kind of such a relativist view that you think that maybe that would have been correct back then, that’s why I wanna kind of distinguish this range of problems. And I know that you’re kind of most interested in that third thing, I think. Is that right?

Lucas: Yeah, so I think I’m pretty interested in the second and the third thing, and I just wanna unpack a little bit of your distinction between the first and the second. So, the first was what you called the control problem, and you called the second just the plurality of human values and preferences and the issue of aligning to that in the broader context of the world.

It’s unclear to me how I get the AI to put a strawberry on the plate or to clean up my room and not kill my cat without the second thing haven been done, at least to me.

There is a sense at a very low level where your sort of working on technical AI alignment, which involves working on the MIRI approach with agential foundations and trying to work on a constraining optimization and corrigibility and docility and robustness and security and all of those sorts of things that people work on and the concrete problems in AI safety, stuff like that. But, it’s unclear to me where that sort of stuff is just limited to and includes the control problem, and where it begins requiring the system to be able to learn my preferences through interacting with me and thereby is already sort of participating in the second case where it’s sort of participating in AI alignment more generally, rather than being sort of like a low level controlled system.

Will: Yeah, and I should say that on this side of things I’m definitely not an expert, not really the person to be talking to, but I think you’re right. There’s going to be some big, gray area or transition from systems. So there’s one that might be cleaning my room, or even let’s just say it’s playing some sort of game, unfortunately I forget the example … It was under the blog post, an example of the alignment problem in the wild, or something, from open AI. But, just a very simple example of the AIs playing a game, and you say, “Well, get as many points as possible.” And what you really want it to do is win a certain race, but what it ends up doing is driving this boat just round and round in circles because that’s the way of maximizing the number of points.

Lucas: Reward hacking.

Will: Reward hacking, exactly. That would be a kind of failure of control problem, that first in our sense. And then I believe there’s gonna be kind of gray areas, where perhaps it’s the certain sort of AI system where the whole point is it’s just implementing kind of what I want. And that might be very contextually determined, might depend on what my mood is of the day. For that, that might be a much, much harder problem and will involve kind of studying what I actually do and so on.

We could go into the question of whether you can solve the problem of cleaning a room without killing my cat. Whether that is possible to solve without solving much broader questions, maybe that’s not the most fruitful avenue of discussion.

Lucas: So, let’s put aside this first case which involves the control problem, we’ll call it, and let’s focus on the second and the third, where again the second is defined as sort of the issue of the plurality of human values and preferences which can be observed, and then the third you described as us determining what we ought to do and tackling sort of the metaethics.

Will: Yeah, just tackling the fundamental question of, “Where ought we to be headed as a society?” One just extra thing to add onto that is that’s just a general question for society to be answering. And if there are kind of fast, or even medium-speed, developments in AI, perhaps suddenly we’ve gotta start answering that question, or thinking about that question even harder in a more kind of clean way than we have before. But even if AI were to take a thousand years, we’d still need to answer that question, ’cause it’s just fundamentally the question of, “Where ought we to be heading as a society?”

Lucas: Right, and so going back a little bit to the little taxonomy that I had developed earlier, it seems like your second case scenario would be sort of down to metaethical questions, which are behind and which influence the empirical issues with preference aggregation and there being plurality of values. And the third case would be, what would be arrived at through reason and, I guess, the reason of many different people.

Will: Again, it’s gonna involve questions of metaethics as well where, again, on my theory that metaethics … It would actually just involve interacting with conscious experiences. And that’s a critical aspect of coming to understand what’s morally correct.

Lucas: Okay, so let’s go into the second one first and then let’s go into the third one. And while we do that, it would be great if we could be mindful of problems in intertheoretic comparison and how they arise as we go through both. Does that sound good?

Will: Yeah, that sounds great.

Lucas: So, would you like to just sort of unpack, starting with the second view, the metaethics behind that, issues in how moral realism versus moral anti-realism will affect how the second scenario plays out, and other sorts of crucial considerations in metaethics that will affect the second scenario?

Will: Yeah, so for the second scenario, which again, to be clear, is the aggregating of the variety of human preferences across a variety of contexts and so on, is that right?

Lucas: Right, so that the agent can be fully autonomous and realized in the world that it is sort of an embodiment of human values and preferences, however construed.

Will: Yeah, okay, so here I do think all the metaethics questions are gonna play a lot more role in the third question. So again, it’s funny, it’s very similar to the question of kind of what mainstream economists often think they’re doing when it comes to cost-benefit analysis. Let’s just even start in the individual case. Even there, it’s not a purely kind of descriptive enterprise, where, again, let’s not even talk about AI. You’re just looking out for me. You and I are friends and you want to do me a favor in some way, how do you make a decision about how to do me that favor, how to benefit me in some way? Well, you could just look at the things I do and then infer on the basis of that what my utility function is. So perhaps every morning I go and I rob a convenience store and then I buy some heroin and then I shoot up and-

Lucas: Damn, Will!

Will: That’s my day. Yes, it’s a confession. Yeah, you’re the first to hear it.

Lucas: It’s crazy, in Oxford huh?

Will: Yeah, Oxford University is wild.

You see that behavior on my part and you might therefore conclude, “Wow, well what Will really likes is heroin. I’m gonna do him a favor and buy him some heroin.” Now, that seems kind of commonsensically pretty ridiculous. Well, assuming I’m demonstrating all sorts of bad behavior that looks like it’s very bad for me, it looks like a compulsion and so on. So instead what we’re really doing is not merely maximizing the utility function that’s gone by my revealed preferences, we have some deeper idea of kind of what’s good for me or what’s bad for me.

Perhaps that comes down to just what I would want to want, or what I want myself to want to want to want. Perhaps you can do it in terms of what are called second-order, third-order preferences. What idealized Will would want … That is not totally clear. Well firstly, it’s really hard to know kind of what would idealized Will want. You’re gonna have to start doing at least a little bit of philosophy there. Because I tend to favor hedonism, I think that an idealized version of my friend would want the best possible experiences. That might be very different from what they think an idealized version of themselves would want because perhaps they have some objective list account of well-being and they think well, what they would also want is knowledge for the its own sake and appreciating beauty for its own sake and so on.

So, even there I think you’re gonna get into pretty tricky questions about what is good or bad for someone. And then after that you’ve got the question of preference aggregation, which is also really hard, both in theory and in practice. Where, do you just take strengths of preferences across absolutely everybody and then add them up? Well, firstly you might worry that you can’t actually make these comparisons of strengths of preferences between people. Certainly if you’re just looking at peoples revealed preferences, it’s really opaque how you would say if I prefer coffee to tea and you vice versa, who has the stronger preference? But perhaps we could look at behavioral facts to kind of try and at least anchor that, but it’s still then non-obvious that what we ought to do when we’re looking at everybody’s preferences is just maximize the sum rather than perhaps give some extra weighting to people who are more badly off, perhaps we give more priority to their interests. So this is kinda theoretical issues.

And then secondly, is kinda just practical issues of implementing that, where you actually need to ensure that people aren’t faking their preferences. And there’s a well known literature and voting theory that says that basically any aggregation system you have, any voting system, is going to be manipulable in some way. You’re gonna be able to get a better result for yourself, at least in some circumstances, by misrepresenting what you really want.

Again, these are kind of issues that our society already faces, but they’re gonna bite even harder when we’re thinking about delegating to artificial agents.

Lucas: There’s two levels to this that you’re sort of elucidating. The first is that you can think of the AGI as being something which can do favors for everybody in humanity, so there are issues empirically and philosophically and in terms of understanding other agents about what sort of preferences should that AGI be maximizing for each individual, say being constrained by what is legal and what is generally converged upon as being good or right. And then there’s issues with preference aggregation which come up more given that we live in a resource-limited universe and world, where not all preferences can coexist and there has to be some sort of potential cancellation between different views.

And so, in terms of this higher level of preference aggregation … And I wanna step back here to metaethics and difficulties of intertheoretic comparison. It would seem that given your moral realist view, it would affect how the weighting would potentially be done. Because it seemed like before you were eluding to the fact that if your moral realist view would be true, then the way at which we could determine what we ought to do or what is good and true about morality would be through exploring the space of all possible experiences, right, so we can discover moral facts about experiences.

Will: Mm-hmm (affirmative).

Lucas: And then in terms of preference aggregation, there would be people who would be right or wrong about what is good for them or the world.

Will: Yeah, I guess this is, again why I wanna distinguish between these two types of value alignment problem, where on the second type, which is just kind of, “What does society want?” Societal preference aggregation. I wasn’t thinking of it as there being kind of right or wrong preferences.

In just the same way as there’s this question of just, “I want system to do X” but there’s a question of, “Do I want that?” or “How do you know that I want that?”, there’s a question of, “How do you know what society wants?” That’s a question in and of its own right that’s then separate from that third alignment issue I was raising, which then starts to bake in, well, if people have various moral preferences, views about how the world ought to be, yeah some are right and some are wrong. And no way should you give some aggregation over all those different views, because ideally you should give no weight to the ones that are wrong and if any are true, they get all the weight. It’s not really about kind of preference aggregation in that way.

Though, if you think about it as everyone is making certain sort of guess at the moral truth, then you could think of that like a kind of judgment aggregation problem. So, it might be like data or input for your kind of moral reasoning.

Lucas: I think I was just sort of conceptually slicing this a tiny bit different from you. But that’s okay.

So, staying on this second view, it seems like there’s obviously going to be a lot of empirical issues and issues in understanding persons and idealized versions of themselves. Before we get in to intertheoretic comparison issues here, what is your view on coherent extrapolated volition, sort of, being the answer to this second part?

Will: I don’t really know that much about it. From what I do know, it always seemed under-defined. As I understand it, the key idea is just, you take everyone’s idealized preferences in some sense, and then I think what you do is just take a sum of what everyone’s preference is. I’m personally quite in favor of the summation strategy. I think we can make interpersonal comparisons of strengths of preferences, and I think summing people’s preferences is the right approach.

We can use certain kinds of arguments that also have application in moral philosophy, like the idea of “If you didn’t know who you were going to be in society, how would you want to structure things? And if you’re a rational, self-interested agent, maximizing expected utility, then you’ll do the utilitarian aggregation function, so you’ll maximize the sum of preference strength.

But then, if we’re doing this idealized preference thing, all the devil’s going to be in the details of, “Well how are you doing this idealization?” Because, given my preferences for example, for what they are … I mean my preferences are absolutely … Certainly they’re incomplete, they’re almost certainly cyclical, who knows? Maybe there’s even some preferences I have that are areflexive of things, as well. Probably contradictory, as well, so there’s questions about what does it mean to idealize, and that’s going to be a very difficult question, and where a lot of the work is, I think.

Lucas: So I guess, just two things here. What are sort of the timeline and actual real world working in relationship here, between the second problem that you’ve identified and the third problem that you’ve identified, and what is the role and work that preferences are doing here, for you, within the context of AI alignment, given that you’re sort of partial of a form of hedonistic consequentialism?

Will: Okay, terrific, ’cause this is kind of important framing.

In terms of answering this alignment problem, the deep one of just where ought societies to be going, I think the key thing is to punt it. The key thing is to get us to a position where we can think about and reflect on this question, and really for a very long time, so I call this the long reflection. Perhaps it’s a period of a million years or something. We’ve got a lot of time on our hands. There’s really not the kind of scarce commodity, so there are various stages to get into that state.

The first is to reduce extinction risks down basically to zero, put us a position of kind of existential security. The second then is to start developing a society where we can reflect as much as possible and keep as many options open as possible.

Something that wouldn’t be keeping a lot of options open would be, say we’ve solved what I call the control problem, we’ve got these kind of lapdog AIs that are running the economy for us, and we just say, “Well, these are so smart, what we’re gonna do is just tell it, ‘Figure out what’s right and then do that.'” That would really not be keeping our options open. Even though I’m sympathetic to moral realism and so on, I think that would be quite a reckless thing to do.

Instead, what we want to have is something kind of … We’ve gotten to this position of real security. Maybe also along the way, we’ve fixed the various particularly bad problems of the present, poverty and so on, and now what we want to do is just keep our options open as much as possible and then kind of gradually work on improving our moral understanding where if that’s supplemented by AI system …

I think there’s tons of work that I’d love to see developing how this would actually work, but I think the best approach would be to get the artificially intelligent agents to be just doing moral philosophy, giving us arguments, perhaps creating new moral experiences that it thinks can be informative and so on, but letting the actual decision making or judgments about what is right and wrong be left up to us. Or at least have some kind of gradiated thing where we gradually transition the decision making more and more from human agents to artificial agents, and maybe that’s over a very long time period.

What I kind of think of as the control problem in that second level alignment problem, those are issues you face when you’re just addressing the question of, “Okay. Well, we’re now gonna have an AI run economy,” but you’re not yet needing to address the question of what’s actually right or wrong. And then my main thing there is just we should get ourselves into a position where we can take as long as we need to answer that question and have as many options open as possible.

Lucas: I guess here given moral uncertainty and other issues, we would also want to factor in issues with astronomical waste into how long we should wait?

Will: Yeah. That’s definitely informing my view, where it’s at least plausible that morality has an aggregative component, and if so, then the sheer vastness of the future may, because we’ve got half a billion to a billion years left on Earth, a hundred trillion years before the starts burn out, and then … I always forget these numbers, but I think like a hundred billion stars in the Milky Way, ten trillion galaxies.

With just vast resources at our disposal, the future could be astronomically good. It could also be astronomically bad. What we want to insure is that we get to the good outcome, and given the time scales involved, even what seem like an incredibly long delay, like a million years, is actually just very little time indeed.

Lucas: In half a second I want to jump into whether or not this is actually likely to happen given race dynamics and that human beings are kind of crazy. The sort of timeline here is that we’re solving the technical control problem up into and on our way to sort of AGI and what might be superintelligence, and then we are also sort of idealizing everyone’s values and lives in a way such that they have more information and they can think more and have more free time and become idealized versions of themselves, given constraints within issues of values canceling each other out and things that we might end up just deeming to be impermissible.

After that is where this period of long reflection takes place, and sort of the dynamics and mechanics of that are seeming open questions. It seems that first comes computer science and global governance and coordination and strategy issues, and then comes long time of philosophy.

Will: Yeah, then comes the million years of philosophy, so I guess not very surprising a philosopher would suggest this. Then the dynamics of the setup is an interesting question, and a super important one.

One thing you could do is just say, “Well, we’ve got ten billion people alive today, let’s say. We’re gonna divide the universe into ten billionths, so maybe that’s a thousand galaxies each or something.” And then you can trade after that point. I think that would get a pretty good outcome. There’s questions of whether you can enforce it or not into the future. There’s some arguments that you can. But maybe that’s not the optimal process, because especially if you think that “Wow! Maybe there’s actually some answer, something that is correct,” well, maybe a lot of people miss that.

I actually think if we did that and if there is some correct moral view, then I would hope that incredibly well informed people who have this vast amount of time, and perhaps intellectually augmented people and so on who have this vast amount of time to reflect would converge on that answer, and if they didn’t, then that would make me more suspicious of the idea that maybe there is a real face to the matter. But it’s still the early days we’d really want to think a lot about what goes into the setup of that kind of long reflection.

Lucas: Given this account that you’ve just given about how this should play out in the long term or what it might look like, what is the actual probability do you think that this will happen given the way that the world actually is today and it’s just the game theoretic forces at work?

Will: I think I’m going to be very hard pressed to give a probability. I don’t think I know even what my subjective credence is. But speaking qualitatively, I’d think it would be very unlikely that this is how it would play out.

Again, I’m like Brian and Dave in that I think if you look at just history, I do think moral forces have some influence. I wouldn’t say they’re the largest influence. I think probably randomness explains a huge amount of history, especially when you think about how certain events are just very determined by actions of individuals. Economic forces and technological forces, environmental changes are also huge as well. It is hard to think at least that it’s going to be likely that such a well orchestrated dynamic would occur. But I do think it’s possible and I think we can increase the chance of that happening by the careful actions that where people like FLI are doing at the moment.

Lucas: That seems like the sort of ideal scenario, absolutely, but I also am worried that people don’t like to listen to moral philosophers or people in that potentially selfish government forces and things like that will end up taking over and controlling things, which is kind of sad for the cosmic endowment.

Will: That’s exactly right. I think my chances … If there was some hard takeoff and sudden leap to artificial general intelligence, which I think is relatively unlikely, but again is possible, I think that’s probably the most scary ’cause it means that a huge amount of power is suddenly in the hands of a very small number of people potentially. You could end up with the very long run future of humanity being determined by the idiosyncratic preferences of just a small number of people, so it would be very dependent whether those people’s preferences are good or bad, with a kind of slow takeoff, so where there’s many decades in terms of development of AGI and it gradually getting incorporated into the economy.

I think there’s somewhat more hope there. Society will be a lot more prepared. It’s less likely that something very bad will happen. But my default presumption when we’re talking about multiple nations, billions of people doing something that’s very carefully coordinated is not going to happen. We have managed to do things that have involved international cooperation and amazing levels of operational expertise and coordination in the past. I think the eradication of smallpox is perhaps a good example of that. But it’s something that we don’t see very often, at least not now.

Lucas: It looks like that we need to create a Peter Singer of AI safety of some other philosopher who has had a tremendous impact on politics and society to spread this sort of vision throughout the world such that it would more likely become realized. Is that potentially most likely?

Will: Yeah. I think if a wide number of the political leaders, even if just political leaders of US, China, Russia, all were on board with global coordination on the issue of AI, or again, whatever other transformative technology might really upend things in the 21st century, and were on board with “How important it is that we get to this kind of period of long reflection where we can really figure out where we’re going,” then that alone would be very promising.

Then the question of just how promising is that I think depends a lot on maybe the robustness of … Even if you’re a moral realist, there’s a question of “How likely do you think it is that people will get the correct moral view?” It could be the case that it’s just this kind of strong attractor where even if you’ve got nothing as clean cut as the long reflection that I was describing, instead some really messy thing, perhaps various wars and it looks like feudal society or something, and anyone would say that civilization looks likely chaotic, maybe it’s the case that even given that, just given enough time and enough reasoning power, people will still converge on the same moral view.

I’m probably not as optimistic as that, but it’s at least a view that you could hold.

Lucas: In terms of the different factors that are going into the AI alignment problem and the different levels you’ve identified, first, second, and third, which side do you think is lacking the most resources and attention right now? Are you most worried about the control problem, that first level? Or are you more worried about potential global coordination and governance stuff at the potential second level or moral philosophy stuff at the third?

Will: Again, flagging … I’m sure I’m biased on this, but I’m currently by far the most worried on the third level. That’s for a couple of reasons. One is I just think the vast majority of the world are simple subjectivists or relativists, so the idea that we ought to be engaging in real moral thinking about how we use society, where we go with society, how we use our cosmic endowment as you put it, my strong default is that that question just never even really gets phrased.

Lucas: You don’t think most people are theological moral realists?

Will: Yeah. I guess it’s true that I’m just thinking about-

Lucas: Our bubble?

Will: My bubble, yeah. Well educated westerners. Most people in the world at least would say they’re theological moral realists. One thought is just that … I think my default is that some sort of relativistic will hold sway and people will just not really pay enough attention to think about what they ought to do. A second relevant thought is just I think the best possible universe is plausibly really, really good, like astronomically better than alternative extremely good universes.

Lucas: Absolutely.

Will: It’s also the case that if you’re … Even like slight small differences in moral view might lead you to optimize for extremely different things. Even just a toy example of preference utilitarianism vs hedonistic utilitarianism, what you might think of as two very similar views, I think in the actual world there’s not that much difference between them, because we just kind of know what makes people better off, at least approximately, improves their conscious experiences, it also is generally what they want, but when you’re kind of technologically unconstrained, it’s plausible to me that the optimal configuration of things will look really quite different between those two views. I guess I kind of think the default is that we get it very badly wrong and it will require really sustained work in order to ensure we get it right … If it’s the case that there is a right answer.

Lucas: Is there anything with regards to issues in intertheoretic comparisons, or anything like that at any one of the three levels which we’ve discussed today that you feel we haven’t sufficiently covered or something that you would just like to talk about?

Will: Yeah. I know that one of your listeners was asking whether I thought they were solvable even in principle, by some superintelligence, and I think they are. I think they are if other issues in moral philosophy are solvable. I think that’s particularly hard, but I think ethics in general is very hard.

I also think it is the case that whatever output we have at the end of this kind of long deliberation, again it’s unlikely we’ll get to credence 1 in a particular view, so we’ll have some distribution over different views, and we’ll want to take that into account. Maybe that means we do some kind of compromise action.

Maybe that means we just distribute our resources in proportion with our credence in different moral views. That’s again one of these really hard questions that we’ll want if at all possible to punt on and leave to people who can think about this in much more depth.

Then in terms of aggregating societal preferences, that’s more like the problem of interpersonal comparisons of preference strength, which is kind of formally isomorphic but is at least a definitely issue.

Lucas: At the second and the third levels is where the intertheoretic problems are really going to be arising, and at that second level where the AGI is potentially working to idealize our values, I think there is again the open question about in the real world, whether or not there will be moral philosophers at the table or in politics or whoever has control over the AGI at that point in order to work on and think more deeply about intertheoretic comparisons of value at that level and timescale. Just thinking a little bit more about what we ought to do or what we should do realistically, given potential likely outcomes about whether or not this sort of thinking will or will not be at the table.

Will: My default is just the crucial thing is to ensure that this thinking is more likely to be at the table. I think it is important to think about, “Well, what ought we to do then,” if we think it’s as very likely that things go badly wrong. Maybe it’s not the case that we should just be aiming to push for the optimal thing, but for some kind of second best strategy.

I think at the moment we should just be trying to push for the optimal thing. In particular, that’s in part because my views that a optimal universe is just so much better than even an extremely good one, that I just kind of think we ought to be really trying to maximize the chance that we can figure out what there is and then implement it. But it would be interesting to think about it more.

Lucas: For sure. I guess just wrapping up here, did you ever have the chance to look at those two Lesswrong posts by Worley?

Will: Yeah, I did.

Lucas: Did you have any thoughts or comments on them? If people are interested you can find links in the description.

Will: I read the posts, and I was very sympathetic in general to what he was thinking through. In particular the principle of philosophical conservatism. Hopefully I’ve shown that I’m very sympathetic to that, so trying to think “What are the minimal assumptions? Would this system be safe? Would this path make sense on a very, very wide array of different philosophical views?” I think the approach I’ve suggested, which is keeping our options open as much as possible and punting on the really hard questions, does satisfy that.

I think one of his posts was talking about “Should we assume moral realism or assume moral antirealism?” It seems like there our views differed a little bit, where I’m more worried that everyone’s going to assume some sort of subjectivism and relativism, and that there might be some moral truth out there that we’re missing and we never think to find it, because we decide that what we’re interested in is maximizing X, so we program agents to build X and then just go ahead with it, whereas actually the thing that we ought to have been optimizing for is Y. But broadly speaking, I think this question of trying to be as ecumenical as possible philosophically speaking makes a lot of sense.

Lucas: Wonderful. Well, it’s really been a joy speaking, Will. Always a pleasure. Is there anything that you’d like to wrap up on, anywhere people can follow you or check you out on social media or anywhere else?

Will: Yeah. You can follow me on Twitter @WillMacAskill if you want to read more on some of my work you can find me at williammacaskill.com

Lucas: To be continued. Thanks again, Will. It’s really been wonderful.

Will: Thanks so much, Lucas.

Lucas: If you enjoyed this podcast, please subscribe, give it a like, or share it on your preferred social media platform. We’ll be back again soon with another episode in the AI Alignment series.

[end of recorded material]

Lucas: Hey everyone, welcome back to the AI Alignment Podcast Series with the Future of Life Institute. I’m Lucas Perry and today, we’ll be speaking with Dr. Roman Yampolskiy. This is the third installment in this new AI Alignment Series. If you’re interested in inverse reinforcement learning or the possibility of astronomical future suffering being brought about by advanced AI systems, make sure to check out the first two podcasts in this series.

As always, if you find this podcast interesting or useful, make sure to subscribe or follow us on your preferred listening platform. Dr. Roman Yampolskiy is a tenured associate professor in the Department of Computer Science and Engineering at the Speed School of Engineering at the University of Louisville. He is the founding and current director of the Cybersecurity Lab and an author of many books including Artificial Superintelligence: A Futuristic Approach.

Dr. Yampolskiy’s main areas of interest are in AI safety, artificial intelligence, behavioral biometrics, cybersecurity, digital forensics, games, genetic algorithms and pattern recognition. Today, we cover key concepts in his papers and books surrounding AI safety and artificial intelligence superintelligence and AGI, his approach to AI alignment, how AI security fits into all this. We also explore our audience-submitted questions. This was a very enjoyable conversation and I hope you find it valuable. With that, I give you Dr. Roman Yampolskiy.

Thanks so much for coming on the podcast, Roman. It’s really a pleasure to have you here.

Roman: It’s my pleasure.

Lucas: I guess let’s jump into this. You can give us a little bit more information about your background, what you’re focusing on. Take us a little bit through the evolution of Roman Yampolskiy and the computer science and AI field.

Roman: Sure. I got my PhD in Computer Science and Engineering. My dissertation work was on behavioral biometrics. Typically, that’s applied to profiling human behavior, but I took it to the next level looking at nonhuman entities, bots, artificially intelligent systems trying to see if we can apply same techniques, same tools to detect bots, to prevent bots, to separate natural human behavior from artificial behaviors.

From there, I try to figure out, “Well, what’s the next step? As those artificial intelligence systems more capable, can we keep up? Can we still enforce some security on them?” That naturally led me to looking at much more capable systems and the whole issues with AGI and superintelligence.

Lucas: Okay. In terms of applying biometrics to AI systems or software or computers in general, what does that look like and what is the end goal there? What are the metrics of the computer that you’re measuring and to what end are they used and what information can they give you?

Roman: The good example I can give you is from my dissertation work again. I was very interested with poker at the time. The poker rooms online were still legal in US and completely infested with bots. I had a few running myself. I knew about the problem and I was trying to figure out ways to automatically detect that behavior. Figure out which bot is playing and prevent them from participating and draining resources. That’s one example where you just have some sort of computational resource and you want to prevent spam bots or anything like that from stealing them.

Lucas: Okay, this is cool. Before you’ve arrived at this AGI and superintelligence stuff, could you explain a little bit more about what you’ve been up to? It seems like you’ve done a lot in computer security. Could you unpack a little bit about that?

Roman: All right. I was doing a lot of very standard work relating to pattern recognition, neural networks, just what most people do in terms of work on AI recognizing digits and handwriting and things of that nature. I did a lot of work in biometrics, so recognizing not just different behaviors but face recognition, fingerprint recognition, any type of forensic analysis.

I do run Cybersecurity Lab here at the University of Louisville. My students typically work on more well recognized sub domains of security. With them, we did a lot of work in all those domains, forensics, cryptography, security.

Lucas: Okay. Do you feel that all the security research, how much of it do you think is important or critical to or feeds into ASI and AGI research? How much of it right now is actually applicable or is making interesting discoveries, which can inform ASI and AGI thinking?

Roman: I think it’s fundamental. That’s what I get most of my tools and ideas for working with intelligent systems. Basically, everything we learned in security is now applicable. This is just a different type of cyber infrastructure. We learned to defend computers, networks. Now, we are trying to defend intelligent systems both from insider threats and outside from the systems themselves. That’s a novel angle, but pretty much everything I did before is now directly applicable. So many people working in AI safety approach it from other disciplines, philosophy, economics, political science. A lot of them don’t have the tools to see it as a computer science problem.

Lucas: The security aspect of it certainly make sense. You’ve written on utility function security. If we’re to make value aligned systems, then it’s going to be important that the right sorts of people have control over them and that their preferences and dispositions and the systems, again, utility function is secure is very important. A system in the end I guess isn’t really safe or robust or value aligned if it’s extremely influenced by anyone.

Roman: Right. If someone can just disable your safety mechanism, do you really have a safe system? That completely defeats everything you did. You release a well-aligned, friendly system and then somebody flips a few bits and you got the exact opposite.

Lucas: Right. Given this research focus that you have in security and how it feeds into ASI and AGI thinking and research and AI alignment efforts, how would you just generally summarize your approach to AI alignment and safety?

Roman: There is not a general final conclusion I can give you. It’s still work in progress. I’m still trying to understand all the types of problems we are likely to face. I’m still trying to understand this problem as even solvable to begin with. Can we actually control more intelligent systems? I always look at it from engineering computer science point of view much less from philosophy ethics point of view.

Lucas: Whether or not this problem is in principle solvable, that has a lot to do with fundamental principles and ideas and facts about minds in general and what is possible of minds. Can you unpack a little bit more about what sorts of information we need or what we need to think about more going forward to know what it means whether or not this problem is solvable in principle, how we can figure that up as we continue forward?

Roman: There is multiple ways you can show that it’s solvable. The ideal situation is where you can produce some sort of a mathematical proof. That’s probably the hardest way to do it because it’s such a generic problem. It applies to all domains. It has to be still working under self-improvement and modification. It has to still work after learning of additional information and it has to be reliable against malevolent design, so purposeful modifications. It seems like it’s probably the hardest problem ever to be given to them. Mathematics community are willing to take it on.

You can also look at examples just from experimental situations both with artificial systems. Are we good at controlling existing AIs? Can we make them safe? Can we make software safe in general? Also, natural systems. Are we any good at creating safe humans? Are we good at controlling people? Now, it seems like after millennia of efforts coming up with legal framework, ethical framework, religions, all sorts of ways of controlling people, we are pretty much failing at creating safe humans.

Lucas: I guess in the end, that might come down to fundamental issues in human hardware and software. Like the reproduction of human beings through sex and the way that genetics functions just creates a ton of variance in each person, which each person has different dispositions and preferences and other things. Then also the way that I guess software is run and shared across culture and people. Creates more fundamental issues that we might not have in software and machines because they work differently.

Are there existence proofs I guess with AI where AI is superintelligent in a narrow domain or at least above human intelligence in a narrow domain and we have control over such narrow systems? Would it be potentially generalizable as you sort of aggregate more and more AI systems, which are superintelligent in narrow domains that as you aggregate that or create an AGI, which sort of has meta learning, we would be able to have control over it given these existence proofs in narrow domains?

Roman: There are certainly such examples in narrow domains. If we’re creating, for example, a system to play chess. We can have a single number measuring it’s performance. We can control whatever is getting better or worse. That’s quite possible and is very limited linear domain. The problem is as complexity increases, you go from this n-body problem equals one to n-body equals infinity, and that’s very hard to solve both computationally and in terms of just understanding what in that hyperspace of possibilities is a desirable outcome.

It’s not just gluing together a few narrow AIs like, “Okay, I have a chess playing program. I have a go playing program.” If I put them all in the same PC, do I now have general intelligence capable of moving knowledge across domains? Not exactly. Whatever safety you can prove for limited systems, not necessarily will transferred to a more complex system, which integrates the components.

Very frequently, then you add two safe systems, the merged system has back doors, has problems. Same with adding additional safety mechanisms. A lot of times, you will install a patch for software to increase security and the patch itself has additional loopholes.

Lucas: Right. It’s not necessarily the case that in the end, AGI is actually just going to be sort of like an aggregation of a lot of AI systems, which are superintelligent in narrow domains. Rather, it potentially will be something more like an agent, which has very strong meta learning. So, learning about learning and learning how to learn and just learning in general. Such that all the sort of process is in things that it learns or deeply integrated at a lower level and they’re sort of like a higher level thinking that is able to execute on these things that they learned. Is that so?

Roman: That makes a lot of sense.

Lucas: Okay. Moving forward here, it would be nice if we could go ahead and explore a little bit of the key concepts in your books and papers and maybe get into some discussions there. I don’t want to spend a lot of time talking about each of the terms and having you define them as people can read your book, Artificial Superintelligence: A Futuristic Approach. They can also check out your papers and you’ve talked about these in other places. I think it will be helpful for giving some background and terms that people might not exactly be exposed to.

Roman: Sure.

Lucas: Moving forward, what can you tell us about what AI completeness is?

Roman: It’s a somewhat fuzzy term kind of like Turing test. It’s not very precisely defined, but I think it’s very useful. It seems that there are certain problems in artificial intelligence in general which require you to pretty much have general intelligence to solve them. If you are capable of solving one of them, then by definition, we can reduce other problems to that one and solve all problems in AI. In my papers, I talk about passing Turing test as being the first such problem. If you can pass unrestricted version of a Turing test, you can pretty much do anything.

Lucas: Right. I think people have some confusions here about what intelligence is in the kinds of minds that can solve Turing tests completely and the architecture that they have and whether that architecture means they’re exactly intelligent. I guess some people have this kind of intuition or idea that you could have a sort of system that had meta learning and learning and was able to sort of think as a human does in order to execute a Turing test.

Then potentially, other people have an idea and this may be misguided where a sort of sufficiently complicated tree search or Google engine on the computer would be able to pass a Turing test and that seems potentially kind of stupid. Is the latter idea a myth? Or if not, how is it just as intelligent as the former?

Roman: To pass an unrestricted version of a Turing test, against someone who actually understands how AI works is not trivial. You can do it with just lookup tables and decision trees. I can give you an infinite number of completely novel situations where you have to be intelligent to extrapolate to figure out what’s going on. I think theoretically, you can think of an infinite lookup table which has every conceivable string for every conceivable previous sequence of questions, but in reality, it just makes no sense.

Lucas: Right. They’re going to be sort of like cognitive features and logical processes and things like inferences and extrapolation and logical tools that humans use that almost must necessarily come along for the ride in order to fully pass a Turing test.

Roman: Right. To fully pass it, you have to be exactly the same in your behavior as a human. Not only you have to be as smart, you also have to be as stupid. You have to repeat all the mistakes, all the limitations in terms of humanity, in terms of your ability to compute, in terms of your cognitive biases. A system has to be so smart that it has a perfect model of an average human and can fake that level of performance.

Lucas: It seems like in order to pass a Turing test, the system would either have to be an emulation of a person and therefore almost essentially be a person just on different substrate or would have to be superintelligent in order to run an emulation of a person or a simulation of a person.

Roman: It has to have a perfect understanding of an average human. It goes together with value alignment. You have to understand what a human would prefer or say or do in every situation and that does require you to understand humanity.

Lucas: Would that function successfully at a higher level of general heuristics about what an average person might do or does it require a perfect emulation or simulation of a person in order to fully understand what a person would do in such an instance?

Roman: I don’t know if it has to be perfect. I think there are certain things we can bypass and just going to read books about what a person would do in that situation, but you do have to have a model complete enough to produce good results in novel situations. It’s not enough to know, OK, most people would prefer ice cream over getting a beating, something like that. You have to figure out what to do in a completely novel set up where you can just look it up on Google.

Lucas: Moving on from AI completeness, what can you tell us about the space of mind designs and the human mental model and how this fits into AGI and ASI and why it’s important?

Roman: A lot of this work was started by Yudkowsky and other people. The idea is just to understand how infinite that hyperspace is. You can have completely different sets of goals and desires from systems which are very capable optimizers. They may be more capable than an average human or best human, but what they want could be completely arbitrary. You can’t make assumptions along the lines of, “Well, any system smart enough would be very nice and beneficial to us.” That’s just a mistake. If you randomly pick a mind from that infinite universe, you’ll end up with something completely weird. Most likely incompatible with human preferences.

Lucas: Right. This is just sort of, I guess, another way of explaining the orthogonality thesis as described by Nick Bostrom?

Roman: Exactly. Very good connection, but it gives you a visual representation. I have some nice figures where you can get a feel for it. You start with, “Okay, we have human minds, a little bit of animals, you have aliens in the distance,” but then you still keep going and going in some infinite set of mathematical possibilities.

Lucas: In this discussion of the space of all possible minds, it’s a discussion about intelligence where intelligence is sort of understood as the ability to change and understand the world and also the preferences and values which are carried along in such minds however random and arbitrary they are from the space of all possible mind design.

One thing which is potentially very important in my view is the connection of the space of all possible hedonic tones within mind space, so the space of all possible experience and how that maps onto the space of all possible minds. Not to say that there’s duality going on there, but it seems very crucial and essential to this project to also understand the sorts of experiences of joy and suffering that might come along for each mind within the space of all possible minds.

Is there a way of sort of thinking about this more and formalizing it more such as you do or does that require some more really foundational discoveries and improvements in the philosophy of mind or the science of mind and consciousness?

Roman: I look at this problem and I have some papers looking at those. One looks at just generation of all possible minds. Sequentially, you can represent each possible software program as an integer and brute force them. It will take infinite amount of time, but you’ll get to every one of them eventually.

Another recent paper looks at how we can actually detect qualia in natural and artificial agents. While it’s impossible for me to experience the world as someone else, I think I was able to come up with a way to detect whatever you have experiences or not. The idea is to present you with the illusions, kind of visual illusions and based on the type of body you have, the type of sensors you have, you might have experiences which match with mine. If they are not, then I can say really anything about you. You could be conscious and experiencing qualia or maybe not. I have no idea.

In a set of such tests on multiple illusions, you happen to experience exactly the same side effects from the illusion. This test drew multiple-choice questions and you can get any level of accuracy you want with just additional tests. Then I have no choice but to assume that you have exactly same qualia in their situation. So, at least I know you do have experiences of that type.

If it’s taking it to what you suggested pleasure or pain, we can figure out is there suffering going on, is there pleasure happening, but this is very new. We need a lot more people to start doing psychological experiments with that.

The good news is from existing literature, I found a number of experiments where a neutral network designed for something completely unrelated still experienced similar side effect as a natural model. That’s because the two models represent the same mathematical structure.

Lucas: Sorry. The idea here is that by observing effects on the system that if those effects are also correlated or seen in human subjects that this is potentially some indication that the qualia that is correlated with those effects in people is also potentially experienced or seen in the machine?

Roman: Kind of. Yeah. So, when I show you a new cool optical illusion. You experienced something outside of just the values of bits in that illusion. Maybe you see light coming out of it. Or maybe you see rotations. Maybe you see something else.

Lucas: I see a triangle that isn’t there.

Roman: Exactly. If a machine reports exactly the same experience without previous knowledge obviously, then just Google what a human would see. How else would you explain that knowledge, right?

Lucas: Yeah. I guess I’m not sure here. I probably need to think about it more actually, but this does seem like a very important approach in place to move forward. The person in me who’s concerned about thinking about ethics looks back on the history of ethics and thinks about how human beings are good at optimizing the world in ways in which it produces something of value to them but in optimizing for that thing, they produce huge amounts of suffering. We’ve done this through subjugation of women and through slavery and through factory farming of animals currently and previously.

After each of these periods, of these morally abhorrent behaviors, it seems we have an awakening and we’re like, “Oh, yeah, that was really bad. We shouldn’t have done that.” I guess just moving forward here with machine intelligence, it’s not clear that this will be the case or it is possible that it could be the case, but it may. Potentially sort of the next one of these moral catastrophes is if we sort of ignore this research into the possible hedonic states of machines and just brush it away as being dumb philosophical stuff that we potentially could produce an enormous amount of suffering in machine intelligence and just sort of override that and create another ethical catastrophe.

Roman: Right. I think that makes a lot of sense. I think qualia, a side effect of certain complex computations. You can’t avoid producing them if you’re doing this type of thinking, computing. We have to be careful once we get to that level of not having very painful side effects.

Lucas: Is there any possibility here of trying to isolate the neural architectural correlates of consciousness in human brains and then physically or digitally instantiating that in machines and then creating a sort of digital or physical corpus callosum between the mind of a person and such a digital or physical instantiation of some neural correlate of something in the machine in order to see if an integration of those two systems creates a change in qualia for the person? Such that the person could sort of almost first-person confirm that when it connects up to this thing that its subjective experience changes and therefore maybe we have some more reason to believe that this thing independent of the person, when they disconnect, has some sort of qualia to it.

Roman: That’s very interesting type of experiment I think. I think something like this has been done with Siamese twins conjoined with brain tissue. You can start looking at those to begin with.

Lucas: Cool. Moving on from the space of mind designs and human mental models, let’s go ahead and then talk about the singularity paradox. This is something that you cover quite a bit in your book. What can you tell us about the singularity paradox and what you think the best solutions are to it?

Roman: It’s just a name for this idea that you have a superintelligent system, very capable optimizer, but it has no common sense as we human perceive it. It’s just kind of this autistic savant capable of making huge changes in the world but a four-year-old would have more common sense in terms of disambiguation of human language orders. Just kind of understanding the desirable states of the world.

Lucas: This is sort of the fundamental problem of AI alignment. The sort of assumption about the kind of mind AGI or ASI will be, the sort of autistic savant sort of intelligence, what that is … This is what Dylan Hadfield-Menell brought up on our first podcast for the AI Alignment Series is that for this case of this autistic savant that most people have in mind, a perfectly rational Bayesian optimizing agent. Is that sort of the case? Is that the sort of mind that we have in mind when we’re thinking of this autistic savant that just blows over things we care about because it’s just optimizing too hard for one thing and Goodhardt’s law starts to come into effect?

Roman: Yes, in a way. I always try to find most simple examples so we can understand better in the real world. Then you have people with extremely high level of intelligence. The concerns they have, the issues they find interesting are very different from your average person. If you watch something like Big Bang Show with Sheldon, that’s like a good to funny example of this on a very small scale. There is maybe 30 IQ point difference, but what if it’s 300 points?

Lucas: Right. Given the sort of problem, what are your conclusions and best ideas or best practices for working on this? Working on this is just sort of working on the AI alignment problem I suppose.

Roman: AI alignment is just a new set of words to say we want the safe and secure system, which kind of does what we designed it to do. It doesn’t do anything dangerous. It doesn’t do something we disagree with. It’s well aligned with our intention. By itself, the term adds nothing new. The hard problem is, “Well, how do we do it?”

I think it’s fair to say that today, as of right now, no one in the world has a working safety mechanism capable of controlling intelligent behavior and scaling to a new level of intelligence. I think even worse is that no one has a prototype for such a system.

Lucas: One thing that we can do here is we can sort of work on AI safety and we can think about law, policy and governance to try and avoid an arms race in AGI or ASI. Then there are also important ethical questions which need to be worked on before AGI some of which including kind of more short-term things, universal basic income and bias and discrimination in algorithmic systems. How AI will impact the workforce and other things and potentially some bigger ethical questions we might have to solve after AGI if we can pull the brakes.

In terms of the technical stuff, one important path here is thinking about and solving the confinement problem, the method by which we are able to create an AGI or ASI and air gap it and make it so that it is confined and contained to be tested in some sort of environment to see if it’s safe. What are your views on that and what do you view as a potential solution to the confinement problem?

Roman: That’s obviously a very useful tool to have, to test, to debug, to experiment with an AI system while it’s limited in its communication ability. It cannot perform social engineering attacks against the designer or anyone else. It’s not the final solution if you will if a system can still escape from such confinement, but it’s definitely useful to be able to do experiments on evolving learning AI.

Can I limit access to the Internet? Can I limit access to knowledge, encyclopedia articles? Can I limit output in terms of just text, no audio, no video? Can I do just a binary yes or no? All of it is extremely useful. We have special air gap systems for studying computer viruses, so to understand how they work, how they communicate versus just taking it to the next level of malevolent software.

Lucas: Right. There’s sort of this, I guess, general view and I think that Eliezer has participated in some of these black boxing experiments where you pretend as if you are the ASI and you’re trying to get out of the box and you practice with other people to see if you can get out of the box. Out of discussions and thinking on this, it seems that some people thought that it’s almost impossible to confine these systems. Do you think that, that’s misguided or what are your views on that?

Roman: I agree that long-term, you absolutely cannot confine a more intelligent system. I think short-term while it’s still developing and learning, it’s a useful tool to have. The experiments Eliezer did, very novel at the time, but I wish he meet public all the information to make them truly scientific experiments where people can reproduce them properly, learn from them. Simply saying that this guy who now works with me let me out, it’s not the optimal way to do it.

Lucas: Right. I guess the concern there is with confinement experiments is that explaining the way in which it gets out is potentially an information hazard.

Roman: Yeah. People tend to call a lot of things informational hazards. Those things certainly exist. If you have source code for AGI, I strongly recommend you don’t make it public, but we’ve been calling a lot of things informational hazard I think.

The best example is Roko’s basilisk where essentially it was a new way to introduce Christianity. If I tell you about Jesus and you don’t follow him, now you’re going to hell. If I didn’t tell you about Jesus, you’d be much better off. Why did you tell me? Deleting it just makes it grow bigger and it’s like Streisand effect, right? You promoting this while you trying to suppress it. I think you have to be very careful in calling something an informational hazard, because you’re diluting the label by doing that.

Lucas: Here’s something I think we can potentially get into the weeds on and we may disagree about and have some different views on. Would you like to just go ahead and unpack your belief? First of all, go ahead and explain what it is and then explain your belief about why machine ethics in the end is the wrong approach or a wrong instrument in AI alignment.

Roman: The way it was always done in philosophy typically, everyone tried to publish a paper suggesting, “Okay, this is a set of ethics we need to follow.” Maybe it’s ethics based on Christianity or Judaism. Maybe it’s utilitarianism, whatever it is. There was never any actual solution, anything was proposed which could be implemented as a way to get everyone on board and agree with it. It was really just a competition for like, “Okay, I can come up with a new ethical set of constraints or rules or suggestions.”

We know philosophers have been trying to resolve it for millennia. They failed miserably. Why somehow moving it from humans to machines will make it easier problem to solve where a single machine is a lot more powerful and can do a lot more with this is not obvious to me. I think we’re unlikely to succeed by doing that. The theories are contradictory, ill-defined, they compete. It doesn’t seem like it’s going to get us anywhere.

Lucas: To continue unpacking your view a bit more, instead of machine ethics where we can understand machine ethics as the instantiation of normative and meta-ethical principles and reasoning and machine systems to sort of make them moral agents and moral reasoners, your view is that instead of using that, we should use safety engineering. Would you like to just unpack what that is?

Roman: To return to the definition you proposed. For every ethical system, there are edge cases which backfire tremendously. You can have an AI which is a meta-ethical decider and it figures out, “Okay, the best way to avoid human suffering is do not have any humans around.” You can defend it from philosophical point of view, right? It makes sense, but is that a solution we would accept if a much smarter system came up with it?

Lucas: No, but that’s just value misalignment I think. I don’t think that there are any sort of like … There are, in principle, possible moral systems where you say suffering is so bad that we shouldn’t risk any of it at all ever, therefore life shouldn’t exist.

Roman: Right, but then you make AI the moral agent. That means it’s making moral decisions. It’s not just copying what humans decided even if we can somehow figure out what the average is, it’s making its own novel decisions using its superintelligence. It’s very likely it will come up with something none of us ever considered. The question is, will we like it?

Lucas: Right. I guess just for me here, I understand why AI safety engineering and technical alignment efforts are so very important and intrinsic. I think that it really constitutes a lot of the AI alignment problem. I think that given that the universe has billions and billions and billions of years left to live, that the instantiation of machine ethics in AGI and ASI is… you can’t hold off on it and it must be done.

You can’t just have an autistic savant superspecies on the planet that you just never imbue with any sort of ethical epistemology or meta-ethics because you’re afraid of what might happen. You might want to do that extremely slowly and extremely carefully, but it seems like machine ethics is ultimately an inevitability. If you start to get edge cases that the human beings really don’t like, then potentially you just went wrong somewhere in cultivating and creating its moral epistemology.

Roman: I agree with doing it very slowly and carefully. That seems like a good idea in general, but again, just projecting to long-term possibilities. I’m not optimistic that the result will be beneficial.

Lucas: Okay. What is there left to it? If we think of the three cornerstones of AI alignment as being law, policy, governance, then we have ethics on one corner and then we have technical AI alignment on the other corner. We have these three corners.

If we have say AGI or ASI around 2050, which I believe is something a lot of researchers give a 50% probability to, then imagine we simply solve technical AI alignment and we solved the law, policy and governance coordination stuff so that we don’t end up having an arms race and we mess up on technical alignment. Or someone uses some singleton ASI to malevolently control everyone else.

Then we still have the ethical issues in the end. Even if we have a perfectly corrigible and docile intelligence, which is sort of tuned to the right people and sort of just takes the right orders. Then whatever that ASI does, it’s still going to be a manifestation, an embodiment of the ethics of the people who tell it what to do.

There’s still going to be billions and billions of years left in the universe. William MacAskill discusses this. Is that sort of after we’ve solved the technical alignment issues and the legal and political and coordination issues, then we’re going to need a period of long deliberation where we actually have to make concrete decisions about moral epistemology and meta-ethics and try and do it in really a formalized and rigorous way and potentially take thousands of years to figure it out.

Roman: I’m criticizing this and that makes it sound like I have a solution, which is something else and I don’t. I don’t have a solution whatsoever. I just feel it’s important to point out problems with each specific approach so we can avoid problems of over committing to it.

You mentioned a few things. You mentioned getting information from the right people. That seems like that’s going to create some problems right there. Not sure who the right people are. You mentioned spending thousands of years deciding what we want to do with this superintelligent system. I don’t know if we have that much time given all the other existential risks, given the chance of malevolent superintelligence being released by rogue agents much sooner. Again, it may be the best we got, but it seems like there are some issues we have to look at.

Lucas: Yeah, for sure. Ethics has traditionally been very messy and difficult. I think a lot of people are confused about the subject. Based on my conversation with Dylan Hadfield-Menell, when we’re discussing inverse reinforcement learning and other things that he was working on, his sort of view was a view of AI alignment and value alignment where inverse reinforcement learning and other preference learning techniques are sort of used to create a natural evolution of human values and preferences in ethics, which sort of exists in an ecosystem of AI systems which are all, I guess, in conversation so that it could, more so, naturally evolve.

Roman: Natural evolution is a brutal process. It really has no humanity to it. It exterminates most species. I don’t know if that’s the approach we want to simulate.

Lucas: Not an evolution of ideas?

Roman: Again, if those ideas are actually implemented and applied to all of humanity that has a very different impact than if it’s just philosophers debating with no impact.

Lucas: In the end, it seems like a very difficult end frontier to sort of think about and move forward on. Figuring out what we want and what we should do with a plurality of values and preferences. Whether or not we should take a view of moral realism or moral relativism or anti-realism about ethics and morality. Those seem like extremely consequential views or positions to take when determining the fate of the cosmic endowment.

Roman: I agree completely on how difficult the problem is.

Lucas: Moving on from machine ethics, you wrote a  paper on leak proofing the singularity. Would you like to go ahead and unpack a little bit about what you’re doing in the paper and how that ties into all of this?

Roman: That’s just AI boxing. That was the response to David Chalmers’ paper and he talks about AI boxing as leak proofing, so that’s the title we used, but it’s just a formalization of the whole process. Formalization of the communication channel, what goes in, what goes out. It’s a pretty good paper on it. Again, it relies in this approach of using tools from cyber security to formalize the whole process.

For a long time, experts in cyber security attempted to constrain regular software, not intelligent software from communicating with our programs and outside world and operating system. We’re looking at how that was done, what different classifications they used for site channels and so on.

Lucas: One thing that you also touch on, would you like to go ahead and unpack like wireheading addiction and mental illness in general in machine systems and AI?

Roman: It seems like there is a lot of mental disorders, people experience. The only example of general intelligence we have. More and more, we see similar problems show up in artificial systems, which try to emulate this type of intelligence. It’s not surprising and I think it’s good that we have this body of knowledge from psychology which we can now use to predict likely problems and maybe come up with some solutions for them.

Wireheading is essentially this idea of agent not doing any useful work but just stealing their work channel. If you think about having kids and there is a cookie jar and they get rewarded every time they clean the room or something like that with a cookie, well, they essentially can just find the cookie jar and get direct access to their work channel, right? They’re kids, so they’re unlikely to cause much harm, but if a system is more capable, it realizes you as a human control the cookie jar, well now, it has incentive to control you.

Lucas: Right. There are also these examples with rats and mice that you might be able to discuss a little bit more.

Roman: The classic experiments on that just created through surgery, electrode implants in a brain of some simple animals. Every time you provided an electrical shock to that area, the animals experience the maximum pleasure like orgasm you don’t get tired of. They bypass getting food, having sex, playing with toys. They just sat there pressing the button. If you made it where they have to walk on electrocuted fence to get to the button, it wasn’t a problem, they would do that. It completely messes with usefulness of an agent.

Lucas: Right. I guess just in terms of touching on the differences and the implications of ethics here is that one with sort of consequentialist views, which was sort of very impartial and on speciesists can potentially view wireheading as ethical or the end goal. Whereas other people view a wireheading as basically abhorrent and akin to something terrible that you would never want to happen. There’s also again, I think, a very interesting ethical tension there.

Roman: It goes, I think, to the whole idea of simulated reality and virtual world. Do you care if you’re only succeeding in a made-up world? Would that make you happy enough or do you have to actually impact reality? That could be part of resolving our differences about values and ethics. If every single person can be in their own simulated universe where everything goes according to their wishes, is that a solution to getting us all to agree? You know it’s a fake universe, but at least you’re the king in it.

Lucas: I guess that also touches on this question of the duality that human beings have created between what is fake and real. In what sense is something really fake if it’s not just the base reality? Is there really fundamental value in the thing being the base reality and do we even live in the base reality? How does cosmology or ideas that Max Tegmark explores about the multiverse sort of even impact that? How will that impact our meta-ethics and decision-making about the moral worth of wireheading and simulated worlds?

Roman: Absolutely. I have a paper on something I call designer metry, which is measuring natural versus artificial. The big question of course is can we tell if you are living in a simulated reality? Can it be measured scientifically? Or was it just a philosophical idea? It seems like there are certain ways to identify signals from the engineer if it’s done on purpose, but in general case, you can never tell whatever something is a deep fake or a real input.

Lucas: I’d like to discuss that a little bit more with you, but just to backup really quick to finish talking on about psychology and AI. It seems like this has been something that is really growing in the AI community and it’s not something that I really know much about at all. My general understanding is as AI systems become more and more complex, it’s going to be much more difficult to diagnose and understand the specific pathways and architectures, which are leading to mental illness.

Therefore, general diagnosable tools which observe and understand higher level phenomena or behaviors that systems exist that we’ve developed in psychology would be helpful or implementable here. Is that sort of the case and the use case of psychology here is really just diagnose mental illnesses or does it also has a role in developing positive psychology and well-being in machine systems?

Roman: I think it’s more of a first case. If you have a black box AI, just a huge, very deep neural network, you can just look at the wiring and weights and figure out why it’s producing the results you’re seeing. Whereas you can do high-level experiments, maybe even conversation with the system to give you an idea of how it’s misfiring what the problem is.

Lucas: Eventually, if we begin exploring the computational structure of different hedonic tones and that becomes more formalized as a science, then I don’t know, maybe potentially, there would be more of a role for psychologists in discussing the well-being part rather than the computational mental illness part.

Roman: It is a very new concept. It’s been mentioned a lot in science fiction, but as a scientific concept, it’s very new. I think there is only one or two papers on it directly. I think there is so much potential to exploring more on connections with neuroscience. I’m actually quite excited about it.

Lucas: That’s exciting. Are we living in a simulated world? What does it mean to be able to gather evidence about whether or not we’re living in a simulation? What would such evidence look like? Why may we or may not ever be able to tell whether or not we are in a simulation?

Roman: In general case, if there is not an intent to let you know that it’s a simulated world, you would never be able to tell. Absolutely anything can actually be part of natural base system. You don’t know what it’s like if you are Mario playing in an 8-bit world. You have no idea that it’s low resolution. You’re just part of that universe. You assume the base is the same.

There are situations where engineers leave trademarks, watermarks, helpful messages in a system to let you know what’s going on, but that’s just giving you the answer. I think in general case, you can never know, but from statistical arguments, there’s … Nick Bostrom presents a very compelling statistical arguments. I do the same for biological systems in one of my papers.

Roman: It seems more likely that we are not the base just because every single intelligent civilization will produce so many derived civilizations from it. From space exploration, from creating biological robots capable of undergoing evolutionary process. It would be almost a miracle if out of thousands and thousands of potential newly designed organisms, newly evolved ones, we were like the first one.

Lucas: I think that, that sort of evolutionary process presumes that the utility function of the optimization process, which is spreading into the universe, is undergoing an evolutionary process where it’s changing. Whereas the security and brittleness and stability of that optimization process might be very fixed. It might be that all future and possible super advanced civilizations do not converge on creating ancestor simulations.

Roman: It’s possible, but it feels like a bit less likely. I think they’ll still try to grab the resources and the systems may be fixed in certain values, but they still would be adopting to the local environment. We just see it with different human populations, right? We’re essentially identical, but we developed very different cultures, religions, food preferences based on the local available resources.

Lucas: I don’t know. I feel like I could imagine like a civilization, a very advanced one coming down on some sort of hedonic consequentialism where the view is that you just want to create as many beautiful experiences as possible. Therefore, there wouldn’t be any room for simulating evolution on Earth and all the suffering and kind of horrible things we have to go through.

Roman: But you’re looking at it from inside the simulation. You don’t know what the reasons are on the outside, so this is like a video game or going to the gym. Why would anyone be killed in a video game or suffer tremendously, lifting heavy weights in a gym, right? It’s only fun when you understand external reasons for it.

Lucas: I guess just two things here. I just have general questions on. If there is a multiverse at one or another level, would it then also be the case that the infinity of simulated universes would be a larger fraction of the infinity of the multiverse than the worlds which were not simulated universes?

Roman: This is probably above my pay grade. I think Max is someone who can give you a better answer in that. Comparing degrees of infinities is hard.

Lucas: Okay. Cool. It is not something I really understand either. Then I guess the other thing is I guess just in general, it seems queer to me that human beings are in a world and that we look at our computer systems and then we extrapolate what if these computer systems were implemented at a more base level. It seems like we’re trapped in a context where all that we have to extrapolate about the causes and conditions of our universe are the most fundamental things that we can observe from within our own universe.

It seems like settling on the idea of, “Okay, we’re probably in a simulation,” just seems kind of like we’re gluing to and finding a cosmogenesis hope in one of the only few things that we can, just given that we live in a universe where there are computers. Does that make sense?

Roman: It does. Again, from inside the simulation, you are very limited in understanding the big picture. Then so much would be easier to understand if we had external knowledge, but it’s just not the option we have so far. We learn by pretending to be the engineer in question and now we design virtual worlds. We design intelligent beings and the options we have is the best clue we have about the options available to whoever does it in the external level.

Lucas: Almost as if Mario got to the end of the level and got to the castle. Then because you got to the castle the next level or world started, he was like maybe outside of this context there’s just a really, really big castle or something that’s making lower levels of castles exist.

Roman: Right. I agree with that, but I think we have in common this mathematical language. I think that’s still universal. Just by studying mathematics and possible structures and proving things, we can learn about what’s possible and impossible.

Lucas: Right. I mean there’s just really foundational and fundamental question about the metaphysical realism or anti-realism of mathematics. If there is a multiverse or like a meta multiverse or like a meta-meta-meta-multiverse levels …

Roman: Only three levels.

Lucas: I guess just the implications of a mathematical realism or Platonism or sort of anti-realism at these levels would have really big implications.

Roman: Absolutely, but at this point, I think it’s just fun to think about those possibilities and what they imply for what we’re doing, what we’re hoping to do, what we can do. I don’t think it’s a waste of time to consider those things.

Lucas: Just generally, this is just something I haven’t really been updated on. Is this rule about only in three levels of regression, is that just sort of a general principle or role kind of like Occam’s razor that people like to stick by? Or is there any more…?

Roman: No. I think it’s something Yudkowsky said and it’s cute and kind of meme like.

Lucas: Okay. So it’s not like serious epistemology?

Roman: I don’t know how well proven that is. I think he spoke about levels of recursion initially. I think it’s more of a meme.

Lucas: Okay. All right.

Roman: I might be wrong in that. I know a lot about memes, less about science.

Lucas: Me too. Cool. Given all this and everything we’ve discussed here about AI alignment and superintelligence, what are your biggest open questions right now? What are you most uncertain about? What are you most looking for key answers on?

Roman: The fundamental question of AI safety, is it solvable? Is control problem solvable? I have not seen a paper where someone gives mathematical proof or even a rigorous argument. I see in some blog posts arguing, “Okay, we can predict what the chess machine will do, so surely we can control superintelligence,” but it just doesn’t seem like it’s enough. I’m working on a paper where I will do my best to figure out some answers for that.

Lucas: what is the definition of control and AI alignment?

Roman: I guess it’s very important to formalize those before you can answer the question. If we don’t even know what we’re trying to do, how can we possibly succeed? The first step in any computer science research project is to show that your problem is actually solvable. Some are not. We know, for example, holding problem is not solvable, so it doesn’t make sense to give it as an assignment to someone and wait for them to solve it. If you give them more funding, more resources, it’s just a waste.

Here, it seems like we have more and more people working very hard in different solutions, different methods, but can we first spend a little bit of time seeing how successful can we be? Based on the answer to that question, I think a lot of our governance and the legal framework and general decision-making about this domain will be impacted by it.

Lucas: If your core and key question here is whether or not the control problem or AI alignment is, in principle, or fundamentally solvable, could you give us a quick crash course on complexity theory and computational complexity theory and just things which take polynomial time to solve versus exponential time?

Roman: That’s probably the hardest course you’ll take as an undergraduate in computer science. At the time, I hated every second of it. Now, it’s my favorite subject. I love it. This is the only professor whom I remember teaching computational complexity and computability.

To simplify it, there are different types of problems. Surprisingly, almost all problems can be squeezed into one of those boxes. There are easy problems, which we can just quickly compute. Your calculator adding 2+2 is an example of that. There are problems where we know exactly how to solve them. It’s very simple algorithm. We can call it brute force. You try every option and you’ll always get the best answer, but there’s so many possibilities that in reality you can never consider every option.

Lucas: Like computing prime numbers.

Roman: Well, computer numbers are NP. It’s polynomial to test if a number is prime. It’s actually one of somewhat recent paper for the last 10 years, a great result, Ps are N prime. There are problems which are called NP complete and those are usually the interesting problems we care about and they all reduce to each other. If you solve one, you solved all of them. You cannot brute force them. You have to find some clever heuristics to get approximate answers, optimize those.

We can get pretty close to that. Examples like traveling salesperson problem. If you can figure out optimal way to deliver pizza to multiple households, if you can solve it in general case, you’ll solve 99% of interesting problems. Then there are some problems which we know no one can ever solve using Von Neumann architecture, like standard computer architecture. There are proposals for hyper computation computers with oracles, computers with all sorts of magical properties which would allow us to solve those very, very, very difficult problems, but that doesn’t seem likely anytime soon.

The best part of it I think is this idea of oracles. An oracle is a machine capable of doing magic to give you answer to otherwise unsolvable problem, and there are degrees of oracles. There are magical machines, which are more powerful magicians than the magical machine. None of it is working in practice. It’s all purely theoretical. You start learning about different degrees of magic and it’s pretty cool.

Lucas: Learning and understanding about what, in principle, is fundamentally computationally possible or feasible in certain time frames within the universe given the laws of physics that we have seems to be foundationally important and interesting. It’s one of, I guess, the final frontiers. Not space, but I guess solving intelligence and computation and also the sort of hedonic qualia that comes along for the ride.

Roman: Right. I guess the magical aspect allows you to escape from your local physics and consider other types of physics and what would be possible outside of this world.

Lucas: What advances or potential advances in quantum computing or other sorts of more futuristic hardware and computational systems help and assist in these problems?

Roman: I think quantum computing has more impact on the cryptography and security in that way. It impacts some algorithms more directly. I don’t think there is a determined need for it right now in terms of AI research or AI safety work. It doesn’t look like a human brain is using a lot of quantum effects though some people argue that it’s important for consciousness. I’m not sure if there is definitive proof of that experimentally.

Lucas: Let’s go ahead now and turn to some questions that we’ve gotten from our audience.

Roman: Sounds good.

Lucas: I guess we’re going to be jumping around here between narrow and short-term AI and some other questions. It would be great if you could let me know about the state of safety and security in current AI in general and the evaluation and verification and validation approaches currently adopted by the industry.

Roman: In general, the state of safety and security in AI is almost nonexistent. It’s kind of we’re repeating history. When we worked on creating Internet security was not something we cared about and so Internet is completely insecure. Then was started work on Internet 2.0, Internet of things. We’re repeating the same mistake. All those very cheap devices made in China have no security but they’re all connected and that’s how you can create swarms of devices attacking systems.

It is my hope that we don’t repeat this with intelligent systems, but right now it looks like we are. We care about getting them to the market as soon as possible, making them as capable as possible, the soonest possible. Safety and security is something most people don’t know about, don’t care about. You can see it in terms of number of researchers working on it. You can see it in terms of percentage of funding allocated to AI safety. I’m not too optimistic so far, but the field is growing exponentially, so that’s a good sign.

Lucas: How does evaluation and verification and validation fit into all of this?

Roman: We have pretty good tools for verifying critical software. Something so important… you’re flying to mars, the system cannot fail. Absolutely. We can do mathematical proofs to show that the code you created matches the design you had. It’s an expensive process, but we can do a pretty good job with it. You can put more resources into verifying it with multiple verifiers. You can get any degree of accuracy you want as a cost of computational resource.

As far as I can tell, there is no or very little successful work on verifying systems which are capable of self-improvement, changing, dynamically learning, operating in novel environments. It’s very hard to verify something where you have no idea what the behavior should be in the first beforehand. If it’s something linear, again, we have a chess computer, we know what it’s supposed to do exactly. It’s a lot easier to verify than something more intelligent than you operating a new data in a new domain.

Lucas: Right. It seems like verification in this area of AI is going to require some much more foundational and difficult proofs and verification techniques here. It seems like you’re saying it also requires an idea of an end goal of what the system is actually intended to do in order to verify that it satisfies that.

Roman: Right. You have to verify it against something. I have a paper on unverifiability where I talk about mathematical fundamental limits to what we can prove and verify mathematically. Already, we’re getting to the point where our mathematical proofs are so complex and so long, most human mathematicians cannot possibly even check if it’s legitimate or not.

We have examples of proofs where a mathematical community as a whole still has not decided if something published 10 years ago is a valid proof. If you’re talking about doing proofs on a black box AI systems, now it seems like the only option we have is another AI mathematician, verify our AI, assisting us with that, but this creates this multiple levels of interaction where who’s verifying, verifiers and so on.

Lucas: It seems to me at least another expression of how deeply interdependent the AI alignment problem is. Technical AI alignment is a core issue, but it seems like even in simple things, or not simple things, but things which you would imagine to at least be purely relegated to computer science also has some sort of connections with ethics and policy and law and how these things will all sort of require each other in order to succeed in AI alignment.

Roman: I agree. You do need this complete picture. Overall, I mentioned it a few times before in other podcasts. It feels like an AI safety, every time we analyze a problem, we discovered that it’s like a fractal. There is then more problems under that one and you do it again. Despite the three levels, you still continue with this. It’s an infinite process.

We never get to a point where, “Okay, we solved this. This is not a problem anymore. We know for sure it works in every conceivable situation.” That’s a problem. You have this infinite surface you have to defend, but you only have to fail once to lose everything. It’s very, very different from standard cyber security where, “Okay, somebody stole my credit card. I’ll just get a new one. I’ll get to try again.” Very different approach.

Lucas: There’s no messing up with artificial superintelligence.

Roman: Basically.

Lucas: Just going off of what we were talking about earlier in terms of how AI safety researchers are flirting and interested in the applications of psychology in AI safety, what do you think about the potential future relationship between AI and neuroscience?

Roman: That is great work in neuroscience and trying to understand measurements from just observing neurons, cells to human behavior. There are some papers showing if we do the same thing with computer processors, we’re just going to get a very good microscope and look at the CPU. “Was it playing a video game? Can we figure out connections between what Mario is doing and what electrical wiring is firing and so on?”

There seems to be a lot of mistakes made in that experiment. That tells us that the neuroscience experiments we’re doing for a very long time may be providing some less-than-perfect data for us. In a way, by doing AI work, we can also improve on our understanding of human brain, medical science, just general understanding of how neural networks work. It’s a feedback loop. That is progress in either one benefits the other.

Lucas: It seems like people like Josh Tenenbaum are working on more neuro inspired approaches to creating AGI. It seems that there are some people who have the view or the philosophy that the best way to getting to general intelligence is probably going to be understanding and studying human beings because we’re in existence proof that can be studied of general intelligences. What are your views on this approach and the work being done there?

Roman: I think it’s a lot easier to copy answers to get to the results. In terms of developing capable system, I think it’s the best option we have. I’m not so sure it leads to a safe system because if you just copy design, you don’t fully understand it. You can replicate it without complete knowledge and then instilling safety into it as a an afterthought, as a add-on later on, maybe even more difficult than if you designed it from scratch yourself.

Lucas: A more general strategy and approach, which gets talked about a lot in the effective altruism community: there seems to be this view and you can correct me here anywhere I might get this narrative sort of wrong. It seems important to build the AGI safety community, the AI safety community in general, by bringing more researchers into the fold.

If we can slow down the people who are working on capability and raw intelligence and bring them over to safety, then that might be a very good thing because it slows down the creation of the intelligence part of AGI and puts more researchers into the part that’s working on safety and AI alignment. Then there’s also this tension where …

While, that is a good thing. It may be a bad thing for us to be promoting AI safety or AGI safety to the public community because they probably just … Journalists would spin it and ruin it and trivialize it, turn it into a caricature of itself and just put Terminator photos on everything, which we at FLI are very aware that journalists like to put Terminator stuff on people’s articles and publications. What is your general view about AI safety outreach and do you disagree with the respectability first approach?

Roman: I’m an educator. I’m a professor. It’s my job to teach students, to educate the public, to inform everyone about science and hopefully more educated populace would benefit all of us. Research is funded through taxpayer grants. The public university is funded through taxpayers. The students paying tuition, the general public essentially.

If our goal is to align AI with values of the people, how can we keep people in the dark? They’re the ones who are going to influence elections. They are the ones who are going to decide what good governance of AI essentially is by voting for the right people. We put so much effort into governance of AI. We have efforts at UN, European Parliament, White House, you name it. There are now agreements between France and Canada on what to do with that.

At the end of the day, politicians listen to the public. If I can educate everyone about what the real issues in science are, I think it’s a pure benefit. It makes sense to raise awareness of long-term issues. We do it in every other field of science. Would you ever suggest it’s not a good idea to talk about climate change? No, of course not. It’s silly. We all participate in the system. We’re all impacted by the final outcome. It’s important to provide the good public outreach.

If your concern is the picture of a title of an article, well  work with better journalists, tell them you cannot use a picture of a Terminator. I do it. I tell them and they end up putting a very boring picture on it and nobody clicks on it. Is Terminator then an educational tool? I was able to explain some advanced computability concepts in a few minutes with simple trivial examples. Then you educate people, you have to come to their level. You have to say, “Well, we do have concerns about military killer robots.” There’s nothing wrong with that, so maybe funding for killer robots should be reduced. If public agrees, that’s wonderful.

Just kind of going if an article I published or somebody interviewed me is less than perfect, then it’s not beneficial, I disagree with it completely. It’s important to get to the public, which is not already sold on the idea. Me doing interview for you right now, right? I’m preaching to the choir. Most of your listeners are into AI safety I’m sure. Or at least effective altruism.

Whereas if I do interview for BBC or something like that, now I’m getting access to millions of people who have no idea what superintelligence is. In my world and your world, this is like common knowledge, but I give a lot of keynotes and I would go and speak to top executives for accounting firms and I ask them basic questions about technology. Maybe one ever heard about superintelligence as a concept.

I think education is always a good thing. Having educated populace is wonderful because that’s where funding will eventually come from for supporting our research and for helping us with AI governance. I’m a very strong supporter of outreach and I highly encourage everyone to do very good articles on it. If you feel that a journalist misrepresents your point of view, get in touch, get it fixed. Don’t just say that we’re going to left public in a dark.

Lucas: I definitely agree with that. I don’t really like this elitism that is part of the culture within some parts of AI safety community, which thinks that only the smartest, most niche people should be aware of this and working on it given the safety concerns and the ways in which it could be turned into something else.

Roman: I was a fellow at the Singularity Institute for Artificial Intelligence what is now MIRI. At that time, they had a general policy of not publishing. They felt it was undesirable and will cause more damage. Now, they publish extensively. I had mentioned that, that’s maybe a good idea a few times.

The general idea of buying out top AI developers and turning them to the white side I guess and working on safety issues, I think that’s wonderful. We want the top people. It doesn’t mean we have to completely neglect less than big names. Everyone needs to be invited to the table in terms of support, in terms of grants. Don’t try to think that reputation means that only people at Harvard and MIT work in AI safety.

There is lots of talent everywhere. I work with remote assistance from around the world. There is so much talent out there. I think the results speak for themselves. I get invited to speak internationally. I advise governments, courts, legislative system. I think reputation only grows with such outreach.

Lucas: For sure and it seems like the education on this, because it can seem fairly complicated and people can be really confused about it because I think that there are lots of common myths that people have about intelligence and “consciousness construed” in some way other than how I think you or I construe the term consciousness or the idea of free will or what it means to be intelligent. There’s just so much room for people to be confused about this issue.

The issue is real and it’s coming and people are going to find out about it whether or not we discuss it now. It seems very important that this happens, but also because like … It seems we also exist in a world where something like 40% to 50% of our country is at least skeptical about climate change. Climate change education and advocacy is very important and should be happening.

Even with all of that education and advocacy, there’s still something like around 40% of people who are skeptical about climate change. That issue has become politicized where people aren’t necessarily interested in facts. At least the skeptics are committed to party lines on the issue.

Roman: What would it be without education, if they never heard about the issue, would percentage be zero?

Lucas: I’m not advocating against education. I’m saying that this is an interesting existence case and saying like, “Yeah, we need more education about AI issues and climate change issues in general.”

Roman: I think there are maybe even more disagreement, not so much about how true of a problem is, but how to fix it. It turns into a political issue, then you start talking about let’s increase taxation, let’s decrease taxation. That’s what politicizes. That is not the fundamental science.

Lucas: I guess I just want to look this up actually just to figure out what the general American populace thinks. I think it was a bit wrong.

Roman: I don’t think it’s important what the exact percentage is. I think it’s general concept we care about.

Lucas: It’s a general concept, but I guess I was just potentially introducing a level of pessimism about why we need to educate people more so about AI alignment and AI safety in general just because these issues, even if you’re extremely skillful about them, can become politicized. Just generally the epistemology of America right now is exploding in a giant mess of bullshit. It’s just important that we educate clearly and correctly.

Roman: You don’t have to start with the most extreme examples or I don’t go with paperclip maximizers or whatever. You can talk about career selection, technological unemployment, basic income. Those things are quite understandable and they provide wonderful base for moving to the next level once we get there.

Lucas: Absolutely. Totally in agreement. How would you describe the typical interactions that you get from mainstream AI and CS researchers who just do sort of standard machine learning and don’t know or really think or care about AGI and ASI? When you talk to them and pitch to them like, “Hey, maybe you should be working on AI safety.” Or, “Hey, AI safety is something that is real, that you should care about.”

Roman: You’re right. There are different types of people based on their background knowledge. There is group one, which never heard of the concept. It’s just not part of their world. You can start by just sharing some literature and you can follow up later. Then there are people who are in complete agreement with you. They know it’s important. They understand the issue, but that’s their job they’re working and I think they are sympathetic to the cause.

Then there are people who heard a few kind of not the best attempts to explain what AI risk is, and so they are skeptical. They may be thinking about Terminator movie or something, Matrix, and so they are quite skeptical. In my personal experience, if I had a chance to spend 30 minutes to an hour with a person one-on-one, they all converted. I never had someone who went, “You told me things, but I have zero concern about intelligent systems having bugs in them or side effects or anything like that.”

I think it’s just a question of spending time and making it a friendly expedience. You’re not adversaries trying to fight it out. You’re just going, “Hey, every single piece of software we ever produced had bugs in it and can be had.” How is this different?

Lucas: I agree with you, but there are also seems to be these existence proofs and existence cases of people who are computer scientists and who are super skeptical about AI safety efforts and working on ASI safety like Andrew Ng and others.

Roman: You have to figure out each individual case-by-case basis of course, but just being skeptical about success of his approach is normal. I told you my main concern, is the problem solvable. That’s a degree of skepticism. If we looked at any other industry. Let’s say we had oil industry. The top executive oil industry said that global climate change is not important. Just call it redistribution of good weather or something, it’s not a big deal.

You would immediately think there is some sort of conflict of interest, right? But how is this different? If you are strongly dependent on development, not on anything else, it just makes sense that you would be 100% for development. I don’t think it’s unnatural at all. Again, I think a good conversation and realignment of incentives would do miracles for such cases.

Lucas: It seems like either because Andrew Ang’s timelines are so long or he just thinks that they’re fundamentally, like there’s just not really a big problem. I think there are some computer scientists, researchers who just think there’s just not really a problem, because we’re making the systems and there are systems that are so intertwined with us that the values will just naturally mesh together or something. I’m just so surprised I guess that from the mainstream CS and AI people that you don’t run into more skeptics.

Roman: I don’t start my random interactions with people by trying to tell them, “You are wrong. Change your mind.” That’s usually not the best approach. Then you talk about specific cases and you can take it slowly and increase the level of concern. You can start by talking about algorithmic justice and bias in algorithms and software verification. I think you’ll get 100% support at all those levels.

What happens when your system is slightly more capable, you’re still working with me? I don’t think there is a gap where you go, “Well, at that point, everything becomes rosy and safe and we don’t have to worry about it.” If a disagreement is about how soon, I think it’s not a problem at all. Everything I argue still applies in 20 years, 50 years, 100 years.

If you’re saying it will take 100 years to get to superintelligence, how long will it take to learn how to control a system we don’t have yet? Probably way longer than that. Already, we should have started 50 years ago. It’s too late now. If anything, it strengthens my point that we should put more resources on the safety side.

Lucas: Absolutely. Just a question about generally your work cataloging failures of AI products and what this means for the future.

Roman: I collect examples, historical examples starting with the very first AI systems, still everyday news of how AI systems fail. The examples you all heard about. Self-driving car kills a pedestrian. Or Microsoft Tay chat bot becomes racist and swears at people. I have maybe about 50 or 60 so far. I keep collecting new ones. Feel free to send me lots of cool examples, but make sure they’re not already on my list.

The interesting thing is the patterns. You can get from it, learn from it and use to predict future failure. One, obviously as AI becomes more common, we have more of those systems, the number of such failures grows. I think it grows exponentially and impacts from them grows.

Now, we have intelligent systems trading in the stock market. I think they take up something like 85% of all stock trades. We had examples where they crash the whole stock market, brought down the volume by $1 trillion or something, closed significant amounts. This is very interesting data. I try to create a data set of those examples and there is some interest from industry to understand how to make their products not make my list in the future.

I think so far the only … It sounds like a trivial conclusion, but I think it’s fundamental. The only conclusion I have is that if you design an AI system to do X, it will very soon fail to X whatever X stands for. It seems like it’s only going to get worse as they become more general because the value of X becomes not just narrow. If you designed a system to play chess, then it will fail to win a chess match. That’s obvious and trivial. But if you design the system to run the world or something like that, what is X here?

Lucas: This makes me think about failure modes. Artificial superintelligence is going to have a probability space of failure modes where the severity of the failure at the worst end … We covered this in my last podcast is it would literally be turning the universe into the worst possible suffering imaginable for everyone for as long as possible. That’s some failure mode of ASI which has some probability which is unknown. Then the opposite on the other end is going to be, I guess, the most well-being and bliss for all possible minds, which exists in that universe. Then there’s everything in between.

I guess the question is, is there any mapping or how important is it in mapping this probability space of failure modes? What are the failure modes that ASI can do or that would occur that would make it not value aligned? What are the probabilities of each of those given, I don’t know, the sort of architecture that we expect ASI to have or how we expect ASI to function?

Roman: I don’t think there is a worst and best case. I think it’s infinite in both directions. It can always get worse and always get better.

Lucas: But it’s constrained by what is physically possible.

Roman: Knowing what we know about physics and within this universe, there is a big multiverse out there possibly with different types of physics and simulated environments can create very interesting side effects as well. That’s not the point. I also collect predicted failures of future systems, part of a same report. You can look it up. That’s very interesting to see what usually a scientist, but sometimes science fiction writers, other people had said as potential examples.

It has things like paperclip maximizer and other examples. I look at predictions which are predictions but short-term. For example, we can talk about sex robots and how they’re going to fail. Someone hacks them, then they forget to stop. You forget your safe word. There are interesting possibilities.

Very useful both as an educational tool to get people to see this trend and go, “Okay. At every level of AI development, we had problems proportionate to the capability of AI. Give me a good argument why it’s not the case moving forward?” Very useful tool for AI safety researchers to predict. “Okay, we’re releasing this new system tomorrow. It’s capable of X.” How can we make sure the problems don’t follow?

I published on this, for example, before Microsoft released their Tay chatbot. Giving Xs to users to manipulate your learning data is usually not a safe option. If they just knew about it, maybe they wouldn’t embarrass themselves so bad.

Lucas: Wonderful. I guess just one last question here. My view was that given a superintelligence originating on earth, there would be a physical maximum of the amount of matter and energy which it could manipulate given our current understanding and laws of physics, which are certainly subject to change if we gain new information.

There is something which we could call, as Nick Bostrom explains, the cosmic endowment which is sort of the sphere around an intelligent species, which is running a superintelligent optimization process. Where the sphere represents the maximum amount of matter and energy, a.k.a., galaxies a superintelligence can reach before the universe expands so much that it’s no longer able to get beyond that point. Why is your view that there isn’t a potentially physical best or physical worst thing that, that optimization process could do?

Roman: Computation is done with respect to time. It may take you twice as long to compute something with the same resources, but you’ll still get that if you don’t have limits on your time. Or you create a subjective time for whoever is experiencing things. You can have computations which are not in parallel, serial computation devoted to a single task. It’s quite possible to create, for example, levels of suffering which progressively get worse I think. Again, I don’t encourage anyone experimenting with that, but it seems like things can get worse not just because of limitations, of how much computing I can do.

Lucas: All right. It’s really been a wonderful and exciting conversation Roman. If people want to check out your work or to follow you on Facebook or Twitter or wherever else, what do you recommend people go to read these papers and follow you?

Roman: I’m very active in social media. I do encourage you to follow me on Twitter, RomanYam, or on Facebook, Roman Yampolskiy. Just Google my name. My Google Scholar has all the papers and just trying to make a sell here. I have a new book coming out, Artificial Intelligence Safety and Security. It’s an edited book with all the top AI safety researchers contributing, and it’s due out in August, mid August. Already available for presale.

Lucas: Wow. Okay. Where can people get that? On Amazon?

Roman: Amazon is a great option. It’s published by CRC Press, so you have multiple options right now. I think it’s available as a softcover and hardcover, which are a bit pricey. It’s a huge book about 500 pages. Most people would publish it as a five book anthology, but you get one volume here. It should come out as a very affordable digital book as well, about $30 for 500 pages.

Lucas: Wonderful. That sounds exciting. I’m looking forward to getting my hands on that. Thanks again so much for your time. It’s really been an interesting conversation.

Roman: My pleasure and good luck with your podcast.

Lucas: Thanks so much. If you enjoyed this podcast, please subscribe, give it a like or share it on your preferred social media platform. We’ll be back again soon with another episode in the AI Alignment Series.

[end of recorded material]