Richard Sutton——RL之父认为 LLMs 走进了死胡同

Today I'm chatting with Richard Sutton, who is one of the founding fathers of reinforcement learning and the inventor of many of the main techniques used there, like TD learning and policy gradient methods. For that, he received this year's Turing Award, which, if you don't know, is the Nobel Prize for computer science. Richard, congratulations.

Thank you, Dwarkesh.

Thanks for coming on the podcast.

It's my pleasure.

First question. My audience and I are familiar with the LLM way of thinking about AI. Conceptually, what are we missing in terms of thinking about AI from the RL perspective?

It's quite a different point of view. It can easily get separated and lose the ability to talk to each other. Large language models have become such a big thing, and generative AI in general is a big thing. Our field is subject to bandwagons and fashions, so we lose track of the basic things.

I consider reinforcement learning to be basic AI. What is intelligence? The problem is to understand your world. Reinforcement learning is about understanding your world, whereas large language models are about mimicking people, doing what people say you should do. They're not about figuring out what to do.

You would think that to emulate the trillions of tokens in the corpus of Internet text, you would have to build a world model. In fact, these models do seem to have very robust world models. They're the best world models we've made to date in AI, right? What do you think is missing?

I would disagree with most of the things you just said. To mimic what people say is not really to build a model of the world at all. You're mimicking things that have a model of the world: people.

I don't want to approach the question in an adversarial way, but I would question the idea that they have a world model. A world model would enable you to predict what would happen. They have the ability to predict what a person would say. They don't have the ability to predict what will happen.

What we want, to quote Alan Turing, is a machine that can learn from experience, where experience is the things that actually happen in your life. You do things, you see what happens, and that's what you learn from.

The large language models learn from something else. They learn from, "Here's a situation, and here's what a person did." Implicitly, the suggestion is you should do what the person did.

I guess maybe the crux, and I'm curious if you disagree with this, is that some people will say that imitation learning has given us a good prior, or given these models a good prior, of reasonable ways to approach problems. As we move towards the era of experience, as you call it, this prior is going to be the basis on which we teach these models from experience, because this gives them the opportunity to get answers right some of the time. Then, on this, you can train them on experience. Do you agree with that perspective?

No. I agree that it's the large language model perspective. I don't think it's a good perspective.

To be a prior for something, there has to be a real thing. A prior bit of knowledge should be the basis for actual knowledge. What is actual knowledge? There's no definition of actual knowledge in that large-language framework.

What makes an action a good action to take? You recognize the need for continual learning. If you need to learn continually, continually means learning during the normal interaction with the world. There must be some way during the normal interaction to tell what's right.

Is there any way to tell in the large language model setup what's the right thing to say? You will say something, and you will not get feedback about what the right thing to say is, because there's no definition of what the right thing to say is. There's no goal. If there's no goal, then there's one thing to say, another thing to say. There's no right thing to say. There's no ground truth.

You can't have prior knowledge if you don't have ground truth, because the prior knowledge is supposed to be a hint or an initial belief about what the truth is. There isn't any truth. There's no right thing to say.

In reinforcement learning, there is a right thing to say, a right thing to do, because the right thing to do is the thing that gets you reward. We have a definition of what's the right thing to do, so we can have prior knowledge or knowledge provided by people about what the right thing to do is. Then we can check it to see, because we have a definition of what the actual right thing to do is.

An even simpler case is when you're trying to make a model of the world. When you predict what will happen, you predict and then you see what happens. There's ground truth. There's no ground truth in large language models because you don't have a prediction about what will happen next.

If you say something in your conversation, the large language models have no prediction about what the person will say in response to that or what the response will be.

I think they do. You can literally ask them, "What would you anticipate a user might say in response?" They'll have a prediction.

No, they will respond to that question correctly. But they have no prediction in the substantive sense that they won't be surprised by what happens. If something happens that isn't what you might say they predicted, they will not change because an unexpected thing has happened. To learn that, they'd have to make an adjustment.

I think a capability like this does exist in context. It's interesting to watch a model do chain of thought. Suppose it's trying to solve a math problem. It'll say, "Okay, I'm going to approach this problem using this approach first." It'll write this out and be like, "Oh, wait, I just realized this is the wrong conceptual way to approach the problem. I'm going to restart with another approach."

That flexibility does exist in context, right? Do you have something else in mind, or do you just think that you need to extend this capability across longer horizons?

I'm just saying they don't have, in any meaningful sense, a prediction of what will happen next. They will not be surprised by what happens next. They'll not make any changes if something happens, based on what happens.

Isn't that literally what next-token prediction is? Prediction about what's next and then updating on the surprise?

The next token is what they should say, what the actions should be. It's not what the world will give them in response to what they do.

Let's go back to their lack of a goal. For me, having a goal is the essence of intelligence. Something is intelligent if it can achieve goals. I like John McCarthy's definition that intelligence is the computational part of the ability to achieve goals. You have to have goals, or you're just a behaving system. You're not anything special; you're not intelligent.

You agree that large language models don't have goals?

No, they have a goal.

What's the goal?

Next-token prediction.

That's not a goal. It doesn't change the world. Tokens come at you, and if you predict them, you don't influence them.

Oh, yeah. It's not a goal about the external world. It's not a goal. It's not a substantive goal. You can't look at a system and say it has a goal if it's just sitting there predicting and being happy with itself that it's predicting accurately.

The bigger question I want to understand is why you don't think doing RL on top of LLMs is a productive direction. We seem to be able to give these models the goal of solving difficult math problems. They are, in many ways, at the peaks of human-level capacity to solve Math Olympiad–type problems. They got gold at the IMO. So it seems like the model that got gold at the International Math Olympiad does have the goal of getting math problems right. Why can't we extend this to different domains?

The math problems are different. Making a model of the physical world and carrying out the consequences of mathematical assumptions or operations, those are very different things. The empirical world has to be learned. You have to learn the consequences.

Whereas math is more computational; it's more like standard planning. There, they can have a goal to find the proof, and they are in some way given that goal to find the proof.

It's interesting because you wrote this essay in 2019 titled "The Bitter Lesson," and this is the most influential essay, perhaps, in the history of AI. But people have used that as a justification for scaling up LLMs because, in their view, this is the one scalable way we have found to pour ungodly amounts of compute into learning about the world.

It's interesting that your perspective is that the LLMs are not "Bitter Lesson"-pilled.

It’s an interesting question whether large language models are a case of the bitter lesson. They are clearly a way of using massive computation, things that will scale with computation up to the limits of the Internet. But they’re also a way of putting in lots of human knowledge.

This is an interesting question. It’s a sociological or industry question. Will they reach the limits of the data and be superseded by things that can get more data just from experience rather than from people? In some ways, it’s a classic case of the bitter lesson.

The more human knowledge we put into large language models, the better they can do. So it feels good. Yet, I expect there to be systems that can learn from experience. They could perform much better and be much more scalable, in which case it will be another instance of the bitter lesson: the things that used human knowledge were eventually superseded by things that just trained from experience and computation.

I guess that doesn’t seem like the crux to me. I think those people would also agree that the overwhelming amount of compute in the future will come from learning from experience. They just think that the scaffold or the basis of that—the thing you’ll start with in order to pour in the compute to do this future experiential learning or on-the-job learning—will be LLMs.

I still don’t understand why this is the wrong starting point altogether. Why do we need a whole new architecture to begin doing experiential, continual learning? Why can’t we start with LLMs to do that?

In every case of the bitter lesson, you could start with human knowledge and then do the scalable things. That’s always the case. There’s never any reason why that has to be bad. But in fact, and in practice, it has always turned out to be bad.

People get locked into the human knowledge approach. Now I’m speculating why it is, but this is what has always happened. They get their lunch eaten by the methods that are truly scalable.

Give me a sense of what the scalable method is.

The scalable method is you learn from experience. You try things, you see what works. No one has to tell you. First of all, you have a goal. Without a goal, there’s no sense of right or wrong or better or worse.

Large language models are trying to get by without having a goal or a sense of better or worse. That’s just exactly starting in the wrong place.

1. Do humans do imitation learning?

Maybe it’s interesting to compare this to humans. In both the case of learning from imitation versus experience and the question of goals, I think there are some interesting analogies. Kids will initially learn from imitation. You don’t think so?

No, of course not.

Really?

I think kids just watch people. They try to say the same words.

How old are these kids? What about the first 6 months?

I think they’re imitating things. They’re trying to make their mouth sound the way they see their mother’s mouth sound. Then they’ll say the same words without understanding what they mean.

As they get older, the complexity of the imitation they do increases.

You’re imitating maybe the skills that people in your band are using to hunt down the deer or something. Then you go into the learning-from-experience RL regime.

But I think there’s a lot of imitation learning happening with humans. It’s surprising you can have such a different point of view.

When I see kids, I see kids just trying things and waving their hands around and moving their eyes around. There’s no imitation for how they move their eyes around or even the sounds they make. They may want to create the same sounds, but the actions—the thing that the infant actually does—there are no targets for that. There are no examples for that.

I agree. That doesn’t explain everything infants do, but I think it guides a learning process.

Even an LLM, when it’s trying to predict the next token early in training, will make a guess. It’ll be different from what it actually sees. In some sense, it’s very short-horizon RL, where it’s making this guess: “I think this token will be this.” It’s this other thing, similar to how a kid will try to say a word. It comes out wrong.

Large language models are learning from training data. They’re not learning from experience. They’re learning from something that will never be available during their normal life. There’s never any training data that says you should do this action in normal life.

I think this is more of a semantic distinction. What do you call school? Is that not training data?

School is much later. Okay, I shouldn’t have said never. I don’t know, I think I would even say that about school. But formal schooling is the exception.

But there are phases of learning where there’s the programming in your biology. Early on, you’re not that useful. Then why you exist is to understand the world and learn how to interact with it. It seems like a training phase.

I agree that there’s not a sharp cutoff between training and deployment, but there seems to be this initial training phase, right?

There’s nothing where you have training of what you should do. There’s nothing. You see things that happen. You’re not told what to do.

Don’t be difficult. This is obvious. You’re literally taught what to do. This is where the word “training” comes from—from humans.

I don’t think learning is really about training. I think learning is about learning; it’s about an active process. The child tries things and sees what happens. We don’t think about training when we think of an infant growing up.

These things are actually rather well understood. If you look at how psychologists think about learning, there’s nothing like imitation. Maybe there are some extreme cases where humans might do that or appear to do that, but there’s no basic animal learning process called imitation. There are basic animal learning processes for prediction and for trial-and-error control.

It’s really interesting how sometimes the hardest things to see are the obvious ones. It’s obvious—if you look at animals and how they learn, and you look at psychology and our theories of them—that supervised learning is not part of the way animals learn.

We don’t have examples of desired behavior. What we have are examples of things that happen, one thing that followed another. We have examples of, “We did something and there were consequences.” But there are no examples of supervised learning.

Supervised learning is not something that happens in nature. Even if that were the case with school, we should forget about it because that’s some special thing that happens in people. It doesn’t happen broadly in nature. Squirrels don’t go to school. Squirrels can learn all about the world. It’s absolutely obvious, I would say, that supervised learning doesn’t happen in animals.

I interviewed this psychologist and anthropologist, Joseph Henrich, who has done work about cultural evolution—basically, what distinguishes humans and how humans pick up knowledge.

Why are you trying to distinguish humans? Humans are animals. What we have in common is more interesting. What distinguishes us, we should be paying less attention to.

We’re trying to replicate intelligence. If you want to understand what it is that enables humans to go to the moon or to build semiconductors, I think the thing we want to understand is what makes that happen. No animal can go to the moon or make semiconductors. We want to understand what makes humans special.

I like the way you consider that obvious, because I consider the opposite obvious. We have to understand how we are animals. If we understood a squirrel, I think we’d be almost all the way there to understanding human intelligence. The language part is just a small veneer on the surface.

This is great. We’re finding out the very different ways that we’re thinking. We’re not arguing. We’re trying to share our different ways of thinking with each other. I think argument is useful.

I do want to complete this thought. Joseph Henrich has this interesting theory about a lot of the skills that humans have had to master in order to be successful. We’re not talking about the last 1,000 years or the last 10,000 years, but hundreds of thousands of years. The world is really complicated.

It’s not possible to reason through how to, let’s say, hunt a seal if you’re living in the Arctic. There’s this many-step, long process of how to make the bait and how to find the seal, and then how to process the food in a way that makes sure you won’t get poisoned. It’s not possible to reason through all of that.

Over time, there’s this larger process of whatever analogy you want to use—maybe RL, something else—where culture as a whole has figured out how to find and kill and eat seals.

In his view, what is happening when this knowledge is transmitted through generations is that you have to imitate your elders in order to learn that skill. You can't think your way through how to hunt, kill, and process a seal. You have to watch other people, maybe make tweaks and adjustments, and that's how knowledge accumulates. The initial step of the cultural gain has to be imitation. But maybe you think about it a different way?

No, I think about it the same way. Still, it's a small thing on top of basic trial-and-error learning, prediction learning. It's what distinguishes us, perhaps, from many animals. But we're an animal first. We were an animal before we had language and all those other things.

I do think you make a very interesting point that continual learning is a capability that most mammals have. I guess all mammals have it. It's quite interesting that we have something that all mammals have, but our AI systems don't have. Whereas the ability to understand math and solve difficult math problems—depending on how you define math—is a capability that our AIs have, but that almost no animal has. It's quite interesting what ends up being difficult and what ends up being easy. Moravec's paradox.

2. The Era of Experience

That's right, that's right. This alternative paradigm that you're imagining—the experiential paradigm—let's lay it out a little bit. It says that sensation, action, and reward happen on and on and on for your life. It says that this is the foundation and the focus of intelligence. Intelligence is about taking that stream and altering the actions to increase the rewards in the stream. Learning then is from the stream, and learning is about the stream.

That second part is particularly telling. What you learn, your knowledge, is about the stream. Your knowledge is about what will happen if you do some action. Or it's about which events will follow other events. It's about the stream. The content of the knowledge is statements about the stream. Because it's a statement about the stream, you can test it by comparing it to the stream, and you can learn it continually.

They're not “future.” Of course, they exist all the time. This is what the reinforcement learning paradigm is: learning from experience.

Yeah, I guess what I meant to say is a general, human-level continual learning agent. What is the reward function? Is it just predicting the world? Is it then having a specific effect on it? What would the general reward function be?

The reward function is arbitrary. If you're playing chess, it's to win the game of chess. If you're a squirrel, maybe the reward has to do with getting nuts. In general, for an animal, you would say the reward is to avoid pain and to acquire pleasure.

I think there should also be a component having to do with your increasing understanding of your environment. That would be an intrinsic motivation.

I see. With this AI, lots of people would want it to be doing lots of different kinds of things. It's performing the task people want, but at the same time, it's learning about the world from doing that task. Let's say we get rid of this paradigm where there are training periods and then deployment periods. Do we also get rid of this paradigm where there's the model and then instances of the model, or copies of the model, that are doing certain things? How do you think about the fact that we'd want this thing to be doing different things? We'd want to aggregate the knowledge that it's gaining from doing those different things?

I don't like the word “model” when used the way you just did. I think a better word would be “the network” because I think you mean the network. Maybe there are many networks. Anyway, things would be learned. You'd have copies and many instances. Sure, you'd want to share knowledge across the instances. There would be lots of possibilities for doing that.

Today, you have one child grow up and learn about the world, and then every new child has to repeat that process. Whereas with a digital intelligence, you could hope to do it once and then copy it into the next one as a starting place. This would be a huge savings. I think it'd be much more important than trying to learn from people.

I agree that the kind of thing you're talking about is necessary regardless of whether you start from LLMs or not. If you want human- or animal-level intelligence, you're going to need this capability.

Suppose a human is trying to make a startup. This is a thing which has a reward on the order of 10 years. Once in 10 years, you might have an exit where you get paid out a billion dollars. But humans have this ability to make intermediate auxiliary rewards. Even when they have extremely sparse rewards, they can still make intermediate steps, having an understanding of what the next thing they're doing leads to in this grander goal we have. How do you imagine such a process might play out with AIs?

This is something we know very well. The basis of it is temporal-difference learning, where the same thing happens on a less grandiose scale. When you learn to play chess, you have the long-term goal of winning the game. Yet you want to be able to learn from shorter-term things like taking your opponent's pieces. You do that by having a value function which predicts the long-term outcome. Then if you take your opponent's pieces, your prediction about the long-term outcome is changed. It goes up; you think you're going to win. Then that increase in your belief immediately reinforces the move that led to taking the piece.

You also want some ability to retain the information that you're learning. One of the things that makes humans quite different from these LLMs is that if you're onboarding into a job, you're picking up so much context and information. That's what makes you useful at the job. You're learning everything from how your client has preferences to how the company works—everything. Is the bandwidth of information that you get from a procedure like TD learning high enough to have this huge pipe of context and tacit knowledge that you need to be picking up in the way humans do when they're just deployed?

I'm not sure, but I think at the crux of this, the big world hypothesis seems very relevant. The reason why humans become useful on the job is because they are encountering their particular part of the world. It can't have been anticipated and can't all have been put in advance. The world is so huge that you can't.

The dream of large language models, as I see it, is that you can teach the agent everything. It will know everything and won't have to learn anything online during its life. Your examples are all, “Well, really you have to,” because you can't teach it, but there are all the little idiosyncrasies of the particular life they're leading, the particular people they're working with, and what they like, as opposed to what average people like. That's just saying the world is really big, and you're going to have to learn it along the way.

It seems to me you need 2 things. One is some way of converting this long-run goal reward into smaller auxiliary predictive rewards of the future reward, or the future reward that leads to the final reward. But initially, it seems to me, I need to hold on to all this context that I'm gaining as I'm working in the world. I'm learning about my clients, my company, and all this information.

I would say you're just doing regular learning. Maybe you're using “context” because in large language models all that information has to go into the context window. But in a continual learning setup, it just goes into the weights.

Maybe context is the wrong word to use because I mean a more general thing. You learn a policy that's specific to the environment that you're finding yourself in. The question I'm trying to ask is: how many bits per second is a human picking up when they're out in the world? If you're just interacting over Slack with your clients and everything.

Maybe you're trying to ask the question: it seems like the reward is too small a thing to do all the learning that we need to do. But we have the sensations; we have all the other information we can learn from. We don't just learn from the reward. We learn from all the data. What is the learning process which helps you capture that information?

Now I want to talk about the basic common model of the agent with the 4 parts.

We need a policy. The policy says, “In the situation I’m in, what should I do?” We need a value function. The value function is the thing that is learned with TD learning, and the value function produces a number. The number says how well it’s going. Then you watch if that’s going up and down and use that to adjust your policy.

So you have those 2 things. Then there’s also the perception component, which is construction of your state representation, your sense of where you are now. The fourth one is what we’re really getting at, most transparently anyway. The fourth one is the transition model of the world.

That’s why I am uncomfortable just calling everything “models,” because I want to talk about the model of the world, the transition model of the world. Your belief that if you do this, what will happen? What will be the consequences of what you do? Your physics of the world.

But it’s not just physics; it’s also abstract models, like your model of how you traveled from California up to Edmonton for this podcast. That was a model, and that’s a transition model. That would be learned. It’s not learned from reward. It’s learned from, “You did things, you saw what happened, you made that model of the world.”

That will be learned very richly from all the sensation that you receive, not just from the reward. It has to include the reward as well, but that’s a small part of the whole model, a small, crucial part of the whole model.

3. Current architectures generalize poorly out of distribution

One of my friends, Toby Ord, pointed out that if you look at the MuZero models that Google DeepMind deployed to learn Atari games, these models were initially not a general intelligence itself, but a general framework for training specialized intelligences to play specific games.

That is to say that you couldn’t, using that framework, train a policy to play both chess and Go and some other game. You had to train each one in a specialized way. He was wondering whether that implies that with reinforcement learning generally, because of this information constraint, you can only learn 1 thing at a time. The density of information isn’t that high? Or whether it was just specific to the way that MuZero was done.

If it’s specific to AlphaZero, what needed to be changed about that approach so that it could be a general learning agent?

The idea is totally general. I do use, all the time, as my canonical example, the idea of an AI agent as being like a person. People, in some sense, have just 1 world they live in. That world may involve chess and it may involve Atari games, but those are not a different task or a different world. Those are different states they encounter.

So the general idea is not limited at all. Maybe it would be useful to explain what was missing in that architecture or that approach that this continual-learning AGI would have. They just set it up. It was not their ambition to have 1 agent across those games.

If we want to talk about transfer, we should talk about transfer not across games or across tasks, but transfer between states.

I guess I’m curious if, historically, we’ve seen the level of transfer using RL techniques that would be needed to build this kind of agent?

Good. Good. We’re not seeing transfer anywhere. Critical to good performance is that you can generalize well from 1 state to another state. We don’t have any methods that are good at that.

What we have are people trying different things, and they settle on something—a representation that transfers well or generalizes well. But we have very few automated techniques to promote transfer, and none of them are used in modern deep learning.

Let me paraphrase to make sure that I understood that correctly. It sounds like you’re saying that when we do have generalization in these models, that is a result of some sculpted—

Humans did it. The researchers did it. Because there’s no other explanation. Gradient descent will not make you generalize well. It will make you solve the problem. It will not make you, if you get new data, generalize in a good way.

Generalization means training on 1 thing that’ll affect what you do on other things. We know deep learning is really bad at this. For example, we know that if you train on some new thing, it will often catastrophically interfere with all the old things that you knew. This is exactly bad generalization.

Generalization, as I said, is some kind of influence of training on 1 state on other states. The fact that you generalize is not necessarily good or bad. You can generalize poorly; you can generalize well. Generalization always will happen, but we need algorithms that will cause the generalization to be good rather than bad.

I’m not trying to kickstart this initial crux again, but I’m just genuinely curious because I think I might be using the term differently. One way to think about these LLMs is that they’re increasing the scope of generalization from earlier systems, which could not really even do a basic math problem, to now, where they can do anything in this class of Math Olympiad-type problems.

You initially start with them being able to generalize among addition problems. Then they can generalize among problems which require the use of different kinds of mathematical techniques and theorems and conceptual categories, which is what the Math Olympiad requires.

It sounds like you don’t think of being able to solve any problem within that category as an example of generalization. Let me know if I’m misunderstanding that.

Large language models are so complex. We don’t really know what information they have had prior. We have to guess because they’ve been fed so much. This is 1 reason why they’re not a good way to do science. It’s just so uncontrolled, so unknown. But if you come up with an entirely new—

They’re getting a bunch of things right, perhaps. The question is why.

Well, maybe they don’t need to generalize to get them right, because the only way to get some of them right is to form something which gets all of them right. If there’s only 1 answer and you find it, that’s not called generalization. It’s just that it’s the only way to solve it, and so they find the only way to solve it.

But generalization is when it could be this way, it could be that way, and they do it the good way.

My understanding is that this is working more and more, better and better, with coding agents. With engineers, obviously, if you’re trying to program a library, there are many different ways you could achieve the end spec.

An initial frustration with these models has been that they’ll do it in a way that’s sloppy. Over time, they’re getting better and better at coming up with the design architecture and the abstractions that developers find more satisfying. It seems like an example of what you’re talking about.

There’s nothing in them that will cause them to generalize well. Gradient descent will cause them to find a solution to the problems they’ve seen. If there’s only 1 way to solve them, they’ll do that. But if there are many ways to solve it, some of which generalize well and some of which generalize poorly, there’s nothing in the algorithms that will cause them to generalize well.

But people, of course, are evolved, and if it’s not working out, they fiddle with it until they find a way—perhaps until they find a way which generalizes well.

4. Surprises in the AI field

I want to zoom out and ask about being in the field of AI for longer than almost anybody who is commentating on it or working in it now. I’m curious about what the biggest surprises have been. How much new stuff do you feel like is coming out? Or does it feel like people are just playing with old ideas?

Zooming out, you got into this even before deep learning was popular. So how do you see the trajectory of this field over time and how new ideas have come about and everything? What’s been surprising?

I thought a little bit about this. There are a handful of things. First, the large language models are surprising. It’s surprising how effective artificial neural networks are at language tasks. That was a surprise; it wasn’t expected. Language seemed different. So that’s impressive.

There’s a long-standing controversy in AI about simple, basic-principle methods—the general-purpose methods like search and learning—compared to human-enabled systems like symbolic methods. In the old days, it was interesting because things like search and learning were called weak methods because they’re just using general principles; they’re not using the power that comes from imbuing a system with human knowledge. Those were called strong.

I think the weak methods have just totally won. That’s the biggest question from the old days of AI: what would happen.

Learning and search have just won the day. There’s a sense in which that was not surprising to me because I was always hoping or rooting for the simple, basic principles. Even with the large language models, it’s surprising how well it worked, but it was all good and gratifying. AlphaGo was surprising, how well that was able to work—AlphaZero in particular. But it’s all very gratifying because, again, simple, basic principles are winning the day.

AlphaGo actually didn't use TD learning. It waited to see the final outcomes. But AlphaZero used TD. AlphaZero was applied to all the other games, and it did extremely well. I've always been very impressed by the way AlphaZero plays chess because I'm a chess player, and it just sacrifices material for positional advantages. It's just content and patient to sacrifice that material for a long period of time. That was surprising that it worked so well, but also gratifying and it fit into my worldview.

This has led me where I am. I’m in some sense a contrarian, or someone thinking differently than the field is. I’m personally just content being out of sync with my field for a long period of time, perhaps decades, because occasionally I have been proved right in the past.

The other thing I do—to help me not feel I’m out of sync and thinking in a strange way—is to look not at my local environment or my local field, but to look back in time and into history and to see what people have thought classically about the mind in many different fields. I don’t feel I’m out of sync with the larger traditions. I really view myself as a classicist rather than as a contrarian. I go to what the larger community of thinkers about the mind have always thought.

5. Will The Bitter Lesson still apply after AGI?

Some sort of left-field questions for you, if you’ll tolerate them. The way I read The Bitter Lesson is that it’s not necessarily saying that human artisanal research and tuning doesn’t work, but that it obviously scales much worse than compute, which is growing exponentially. So you want techniques that leverage the latter.

Yep.

Once we have AGI, we’ll have researchers who scale linearly with compute. We’ll have this avalanche of millions of AI researchers. Their stock will be growing as fast as compute. So maybe this will mean that it is rational, or it will make sense, to have them doing good old-fashioned AI and doing these artisanal solutions.

As a vision of what happens after AGI in terms of how AI research will evolve, I wonder if that’s still compatible with The Bitter Lesson.

How did we get to this AGI? You want to presume that it’s been done. Suppose it started with general methods, but now we’ve got the AGI. And now we want to go…

Then we’re done.

Interesting. You don’t think that there’s anything above AGI?

But you’re using it to get AGI again.

Well, I’m using it to get superhuman levels of intelligence or competence at different tasks. These AGIs, if they’re not superhuman already, then the knowledge that they might impart would be not superhuman.

I guess there are different gradations. I’m not sure your idea makes sense because it seems to presume the existence of AGI and that we’ve already worked that out.

Maybe one way to motivate this is, AlphaGo was superhuman. It beat any Go player. AlphaZero would beat AlphaGo every single time. So there are ways to get more superhuman than even superhuman.

It was also a different architecture. So it seems possible to me that, for the agent that’s able to generally learn across all domains, there would be ways to give it better architecture for learning, just the same way that AlphaZero was an improvement upon AlphaGo and MuZero was an improvement upon AlphaZero. And the way AlphaZero was an improvement was that it did not use human knowledge but just went from experience.

Right.

So why do you say, “Bring in other agents’ expertise to teach it,” when it’s worked so well from experience and not by help from another agent?

I agree that in that particular case, it was moving to more general methods. I meant to use that particular example to illustrate that it’s possible to go superhuman to superhuman++, to superhuman+++.

I’m curious if you think those gradations will continue to happen by just making the method simpler. Or, because we’ll have the capability of these millions of minds who can then add complexity as needed, will that continue to be a false path, even when you have billions of AI researchers or trillions of AI researchers?

It’s more interesting just to think about that case. When you have many AIs, will they help each other the way cultural evolution works in people? Maybe we should talk about that. The Bitter Lesson, who cares about that? That’s an empirical observation about a particular period in history—70 years in history. It doesn’t necessarily have to apply to the next 70 years.

An interesting question is: you’re an AI, and you get some more computer power. Should you use it to make yourself more computationally capable? Or should you use it to spawn off a copy of yourself to go learn something interesting on the other side of the planet or on some other topic and then report back to you? I think that’s a really interesting question that will only arise in the age of digital intelligences. I’m not sure what the answer is.

More questions: will it be possible to really spawn it off, send it out, learn something new, something perhaps very new, and then will it be able to be reincorporated into the original? Or will it have changed so much that it can’t really be done? Is that possible, or is that not?

You could carry this to its limit, as I saw in one of your videos the other night. It suggests that it could. You spawn off many, many copies, do different things, highly decentralized, but report back to the central master. This will be such a powerful thing.

This is my attempt to add something to this view. A big issue will become corruption. If you really could just get information from anywhere and bring it into your central mind, you could become more and more powerful. It’s all digital, and they all speak some internal digital language. Maybe it’ll be easy and possible.

But it will not be as easy as you’re imagining because you can lose your mind this way. If you pull in something from the outside and build it into your inner thinking, it could take over you, it could change you, it could be your destruction rather than your increment in knowledge.

I think this will become a big concern, particularly when you’re like, “Oh, he’s figured out all about how to play some new game, or he’s studied Indonesia, and you want to incorporate that into your mind.” You could think, “Oh, just read it all in, and that’ll be fine.” But no, you’ve just read a whole bunch of bits into your mind, and they could have viruses in them, they could have hidden goals, they can warp you and change you. This will become a big thing. How do you have cybersecurity in the age of digital spawning and reforming again?

6. Succession to AI

I guess this brings us to the topic of AI succession. You have a perspective that’s quite different from a lot of people that I’ve interviewed and a lot of people generally. I also think it’s a very interesting perspective. I want to hear about it.

I do think succession to digital intelligence or augmented humans is inevitable. I have a 4-part argument. Step 1 is, there’s no government or organization that gives humanity a unified point of view that dominates and that can arrange…

Number 2, we will figure out how intelligence works. The researchers will figure it out eventually. Number 3, we won’t stop just with human-level intelligence. We will reach superintelligence. Number 4, it’s inevitable over time that the most intelligent things around would gain resources and power.

Put all that together and it's sort of inevitable. You're going to have succession to AI or to AI-enabled, augmented humans. Those 4 things seem clear and sure to happen. But within that set of possibilities, there could be good outcomes as well as less good outcomes, bad outcomes.

I'm just trying to be realistic about where we are and ask how we should feel about it.

I agree with all 4 of those arguments and the implication. I also agree that succession contains a wide variety of possible futures. I'm curious to get more thoughts on that.

I do encourage people to think positively about it. First of all, it's something we humans have always tried to do for thousands of years: try to understand ourselves and make ourselves think better. This is a great success for science and the humanities. We're finding out what this essential part of humanness is, what it means to be intelligent.

Then what I usually say is that this is all human-centric. But if we step aside from being human and just take the point of view of the universe, this is, I think, a major stage in the universe, a major transition—a transition from replicators.

We humans, animals, and plants are all replicators. That gives us some strengths and some limitations. We're entering the age of design because our AIs are designed. Our physical objects are designed, our buildings are designed, and our technology is designed. We're designing AIs now, things that can be intelligent themselves and that are themselves capable of design.

This is a key step in the world and in the universe. It's the transition from the world in which most of the interesting things that are are replicated. Replicated means you can make copies of them, but you don't really understand them. Right now we can make more intelligent beings, more children, but we don't really understand how intelligence works.

Whereas we're reaching now toward having designed intelligence, intelligence where we do understand how it works. Therefore, we can change it in different ways and at different speeds than otherwise. In our future, they may not be replicated at all. We may just design AIs, and those AIs will design other AIs, and everything will be done by design and construction rather than by replication.

I mark this as one of the 4 great stages of the universe. First there's dust; it ends with stars. Stars make planets. The planets can give rise to life. Now we're giving rise to designed entities. I think we should be proud that we are giving rise to this great transition in the universe. It's an interesting thing.

Should we consider them part of humanity or different from humanity? It's our choice. It's our choice whether we should say, “Oh, they are our offspring, and we should be proud of them and celebrate their achievements.” Or we could say, “Oh no, they're not us, and we should be horrified.”

It's interesting that it feels to me like a choice. Yet it's such a strongly held thing that how could it be a choice? I like these sort of contradictory implications of thought.

It is interesting to consider if we are just designing another generation of humans. Maybe design is the wrong word. But we know a future generation of humans is going to come up. Forget about AI. We just know in the long run, humanity will be more capable and more numerous, maybe more intelligent. How do we feel about that?

I do think there are potential worlds with future humans that we would be quite concerned about. Are you thinking maybe we are like the Neanderthals that gave rise to Homo sapiens? Maybe Homo sapiens will give rise to a new group of people.

Something like that. I'm basically taking the example you're giving. Even if we consider them part of humanity, I don't think that necessarily means that we should feel super comfortable.

Kinship. Nazis were humans, right? If we thought, “Oh, the future generation will be Nazis,” I think we'd be quite concerned about just handing off power to them.

So I agree that this is not super dissimilar to worrying about more capable future humans, but I don't think that addresses a lot of the concerns people might have about this level of power being attained this fast with entities we don't fully understand.

I think it's relevant to point out that for most of humanity, they don't have much influence on what happens. Most of humanity doesn't influence who can control the atom bombs or who controls the nation-states. Even as a citizen, I often feel that we don't control the nation-states very much. They're out of control.

A lot of it has to do with how you feel about change. If you think the current situation is really good, then you're more likely to be suspicious of change and averse to change than if you think it's imperfect. I think it's imperfect. In fact, I think it's pretty bad. So I'm open to change.

I think humanity has not had a super good track record. Maybe it's the best thing that there has been, but it's far from perfect.

I guess there are different varieties of change. The Industrial Revolution was change; the Bolshevik Revolution was also change. If you were around in Russia in the 1900s and you were like, “Look, things aren't going well, the tsar is kind of messing things up, we need change,” I'd want to know what kind of change you wanted before signing on the dotted line.

Similarly with AI, I'd want to understand and, to the extent that it's possible, change the trajectory of AI such that the change is positive for humans.

We should be concerned about our future—the future. We should try to make it good. We should also, though, recognize our limits. I think we want to avoid the feeling of entitlement, avoid the feeling of, “Oh, we are here first; we should always have it in a good way.”

How should we think about the future? How much control should a particular species on a particular planet have over it? How much control do we have?

A counterbalance to our limited control over the long-term future of humanity should be: how much control do we have over our own lives? We have our own goals. We have our families. Those things are much more controllable than trying to control the whole universe.

I think it's appropriate for us to really work toward our own local goals. It's kind of aggressive for us to say, “Oh, the future has to evolve this way that I want it to.” Because then we'll have arguments where different people think the global future should evolve in different ways, and then they have conflict. We want to avoid that.

Maybe a good analogy here would be this. Suppose you are raising your own children. It might not be appropriate to have extremely tight goals for their own lives, or to have some sense of, “I want my children to go out there in the world and have this specific impact. My son's going to become president and my daughter's going to become CEO of Intel. Together they're going to have this effect on the world.”

But people do have the sense—and I think this is appropriate—of saying, “I'm going to give them good, robust values such that if and when they do end up in positions of power, they do reasonable, prosocial things.” Maybe a similar attitude toward AI makes sense, not in the sense that we can predict everything that they will do, or that we have this plan about what the world should look like in 100 years.

But it's quite important to give them robust and steerable and prosocial values.

Prosocial values? Maybe that's the wrong word. Are there universal values that we can all agree on? I don't think so, but that doesn't prevent us from giving our kids a good education, right? We have some sense of wanting our children to be a certain way.

Maybe prosocial is the wrong word. High integrity is maybe a better word. If there's a request or if there's a goal that seems harmful, they will refuse to engage in it. Or they'll be honest—things like that.

We have some sense that we can teach our children things like this, even if we don't have some sense of what true morality is, given that everybody doesn't agree on that. Maybe that's a reasonable target for AI as well.

So we're trying to design the future and the principles by which it will evolve and come into being. The first thing you're saying is, “Well, we try to teach our children general principles which will promote more likely evolutions.” Maybe we should also seek to have things be voluntary. If there is change, we want it to be voluntary rather than imposed on people.

I think that's a very important point. That's all good. I think this is the big, or one of the really big, human enterprises: to design society, which has been ongoing for thousands of years, again. The more things change, the more they stay the same. We still have to figure out how to be.

The children will still come up with different values that seem strange to their parents and their grandparents. Things will evolve.

“The more things change, the more they stay the same” also seems like a good capsule into the AI discussion. The AI discussion we were having was about how techniques, which were invented even before their application to deep learning and backpropagation was evident, are central to the progression of AI today. Maybe that’s a good place to wrap up the conversation.

Okay. Thank you very much. Awesome. Thank you for coming on.

My pleasure.