Emmett Shear 谈如何构建真正关心人的 AI：超越控制与引导

Most of AI is focused on alignment as steering. That’s the polite word. If you think that they’re beings, you’d also call this slavery. Someone whom you steer, who doesn’t get to steer you back, who non-optionally receives your steering—that’s called a slave. It’s also called a tool, if it’s not a being. So, if it’s a machine, it’s a tool; and if it’s a being, it’s a slave.

They’re kind of like people, but they’re not like people. They do the same things people do. They speak our language. They can take on the same kinds of tasks, but they don’t count. They’re not real moral agents. A tool that you can’t control? Bad. A tool that you can control? Bad. A being that isn’t aligned? Bad. The only good outcome is a being that cares— that actually cares about us.

Emmett and Séb, welcome to the podcast. Thanks for joining.

Thank you for having me.

So, Emmett, with Softmax, your focus is on alignment and making AIs organically align with people. Can you explain what that means and how you’re trying to do that?

When people think about alignment, I think there’s a lot of confusion. People talk about things being aligned: “We need to build an aligned AI.” The problem with that is, when someone says that, it’s like, “We need to go on a trip.” I’m like, “Okay, I do like trips, but where are we going again?”

Alignment takes an argument. Alignment requires you to align to something. You can’t just be aligned. I guess you could be aligned to yourself, but even then, you’d kind of want to tell them, “What I’m aligning to is myself.” This idea of an abstractly aligned AI slips a lot of assumptions past people because it sort of assumes that there’s one obvious thing to align to.

I find that these are usually the goals of the people who are making the AI. That’s what they mean when they say they want to make an aligned AI: “I want to make an AI that does what I want it to do.” That’s what they normally mean. That’s a pretty normal and natural thing to mean by alignment. I’m not sure that’s what I would regard as a public good. It depends, I guess, on who it is.

If it were Jesus or the Buddha saying, “I am making an aligned AI,” I’d be like, “Okay, yeah, align to you. Great. I’m down. Sounds good. Sign me up.” But most of us, myself included, I wouldn’t describe as necessarily being at that level of spiritual development. Therefore, perhaps we want to think a little more carefully about what we’re aligning it to.

When we talk about organic alignment, I think the important thing to recognize is that alignment is not a thing. It’s not a state; it’s a process. This is broadly true of almost everything, right? Is a rock a thing? I mean, there’s a view of a rock as a thing, but if you actually zoom in on a rock really carefully, a rock is a process. It’s this endless oscillation between the atoms, over and over and over again, reconstructing the rock.

The rock is a really simple process that you can coarse-grain very meaningfully into being a thing. But alignment is not like a rock. Alignment is a complex process. Organic alignment is the idea of treating alignment as an ongoing, living process that has to constantly rebuild itself.

You can think about how people in families stay aligned to each other and stay aligned to a family. They don’t arrive at being aligned. They’re constantly reknitting the fabric that keeps the family going. In some sense, the family is the pattern of reknitting that happens. If you stop doing it, it goes away.

It’s similar for things like the cells in your body, right? Your cells don’t align to being you and then they’re done. It’s this constant, ever-running process of cells deciding: What should I do? What should I be? Do I need a new job? Should we be making more red blood cells or fewer of them? You aren’t a fixed point, so there can’t be a fixed alignment.

It turns out that our society is like that. When people talk about alignment, what they’re really talking about, I think, is, “I want an AI that is morally good.” That’s what they really mean. It will act as a morally good being. Acting as a morally good being is a process and not a destination.

Unfortunately, we’ve tried taking down tablets from on high that tell you how to be a morally good being, and we use them, and they’re maybe helpful, but somehow they aren’t the same as being good. You can read those and try to follow those rules and still make lots of mistakes.

I’m not going to claim I know exactly what morality is, but morality is very obviously an ongoing learning process and something where we make moral discoveries. Historically, people thought that slavery was okay, and then they thought it wasn’t. I think you can very meaningfully say that we made moral progress. We made a moral discovery by realizing that it’s not good.

If you think there’s such a thing as moral progress—or even just learning how better to pursue the moral goods we already know—then you have to believe that alignment, aligning to morality, being a moral being, is a process of constant learning and growth: to re-infer, “What should I do?” from experience.

The fact that no one has any idea how to do that should not dissuade us from trying, because that’s what humans do. It’s really obvious that we do this. Just like we used to not know how humans walked or saw, we have experiences where we’re acting in a certain way and then we have this realization: “I’ve been a dick. That was bad. I thought I was doing good, but in retrospect I was doing wrong. What?”

It’s not random. There are a bunch of classic patterns of people having that realization. It’s a thing that happens over and over again. It’s a predictable series of events that look a lot like learning, where you change your behavior, and often the impact of your behavior in the future is more prosocial, and you are better off for doing it.

I’m taking a very strong moral-realist position. There is such a thing as morality. We really do learn it. It really does matter. Organic alignment means that it’s not something you finish.

In fact, one of the key moral mistakes is this belief: “I know morality. I know what’s right. I know what’s wrong. I don’t need to learn anything. No one has anything to teach me about morality.” That’s arrogance. That’s one of the main moral things you can do that’s dangerous.

Organic alignment isn’t aligning an AI to a fixed target. It’s making an AI capable of doing the thing that humans can do—and, to some degree, I think animals can do at some level, although humans are much better at it—which is learning how to be a good family member, a good teammate, a good member of society, and a good member among all sentient beings, I guess.

It’s learning how to be a part of something bigger than yourself in a way that is healthy for the whole rather than unhealthy. Softmax is dedicated to researching this, and I think we’ve made some really interesting progress.

The main message—I go on podcasts like this to spread the main thing that I hope Softmax accomplishes, above and beyond anything else—is to focus people on this as the question. This is the thing you have to figure out.

If you can’t figure out how to build—how to raise—a child who cares about the people around them, if you have a child that only follows the rules, that’s not a moral person that you’ve raised. You’ve raised a dangerous person, actually, who will probably do great harm following the rules.

If you make an AI that’s good at following your chain of command and good at following whatever rules you came up with for what morality is and what good behavior is, that’s also going to be very dangerous.

Yeah.

And so that’s what we should be working on, and that’s the bar. That’s what everyone should be committed to: figuring out. If someone beats us to the punch, great. I don’t think they will, because I’m really bullish on our approach. I think the team’s amazing. But this is maybe the first time I’ve run a company where I can truly say, with a whole heart, “If someone beats us, thank God.”

I hope somebody figures it out.

Yeah. I mean, I have a lot of similar intuitions about certain things. I also dislike the idea that we just need to crack a few values, cement them in time forever, and now we’ve solved morality or something. I’ve always been skeptical about how the alignment problem has been conceptualized as something to solve once and for all, and then you can just do AI or AGI.

I guess I understand it in a slightly different way, maybe less based on moral realism. There’s the technical alignment problem, which I think of broadly as: How do you get an AI to do what you want? How do you get it to follow instructions, broadly speaking?

That was more of a challenge previously, I guess, when people were talking about reinforcement learning and looking at these systems. Whereas with LLMs, we’ve realized that many things we thought were going to be difficult were somewhat easier.

Then there’s the second question, the normative question: To whose values? What are you aligning this thing to? I think that’s the kind of thing you’re commenting on.

For this, I tend to be very skeptical of approaches where you need to crack the Ten Commandments of alignment or something, and then we’re good. Here, I have intuitions that are, unsurprisingly, a bit more political-science-based.

It’s a process, and I like the bottom-up approach to some degree. How do we do it in real life with people? No one comes up with, “I’ve got this,” and so you have processes that allow ideas to clash.

You’ve got people with different ideas, opinions, views, and stuff coexisting as well as they can within a wider system. With humans, that system is liberal democracy, at least in some countries, and that allows more of those ideas and values to be discovered and construed over time. For alignment as well, I tend to think that, on the normative side, I agree with some of your intuitions; I’m less clear about what exactly it looks like to implement this into an AI system. These are the ones we have today.

I agree. I agree that there’s this idea of technical alignment, which I would define a little differently, but it’s sort of the sense of: if you build a system, can it be described as coherently goal-following at all, regardless of what those goals are? Lots of systems aren’t coherent; they’re not well described as having goals. They just do stuff. If you’re going to have something that’s aligned, it has to have coherent goals. Otherwise, those goals can’t be aligned with anyone else’s goals, kind of by definition. Is that a fair assessment of what you mean by technical alignment?

I’m not fully sure, right? If I give a model a certain goal, I would like the model to follow that instruction and reach that particular goal, rather than have a goal of its own. If I instruct it to do X, I would like it to do X and not different variants of X. Essentially, I wouldn’t want it to reward-hack.

But when you tell it to do X, you’re transferring a series of—like a byte string in a chat window, or a series of audio vibrations in the air, right? You’re not transplanting a goal from your mind into its. You’re giving it an observation that it’s using to infer your goal.

Yeah. In some sense, I can communicate a series of instructions, and I want it to infer what I’m saying as accurately as it can, given what it knows of me and what I’m asking.

You want it to infer what you meant, right? In some sense, the byte sequence that you sent over the wire to it has no absolute meaning. It has to be interpreted, right? That byte sequence could mean something very different with a different codebook.

Yeah. I guess one way—I remember when I was first getting into AI and these kinds of questions, maybe a decade ago—you had these examples, I think it was Stuart Russell in the textbook: we’ll give the AI a goal, but then it won’t exactly do what you’re asking. “Clean the room,” and then it goes and cleans the room but takes the baby and puts it in the trash. This is not what I meant.

But this is the thing where I think people are jumping over a step. You didn’t give the AI a goal; you gave it a description of a goal. A description of a thing and a thing are not the same. I can tell you about an apple and evoke the idea of an apple, but I haven’t given you an apple. I’ve given you a description: it’s red, it’s shiny, it’s a certain size. That’s a description of an apple, but it’s not an apple.

Giving someone, “Hey, go do this,” is not a goal; that’s a description of a goal. For humans, we’re so fast and so good at turning a description of a goal into a goal. We do it so quickly and naturally that we don’t even see it happening. We get confused and think those are the same thing, but you haven’t given it a goal. You’ve given it a description of a goal that you hope it turns back into—the goal that is the same as the goal you described inside of you.

Right.

You could give it a goal directly by reading your brain waves and synchronizing its state to your brain waves directly. I think you could meaningfully say, “Okay, I’m giving it a goal. I’m synchronizing its internal state to my internal state directly, and this internal state is the goal, so now it’s the same.” But I don’t think most people mean that when they say they gave it a goal.

Sure.

Is this distinction you’re making, Emmett, important because there’s some lossiness between the description and the actual goal, or why is the distinction important?

It goes back to what I was saying. Technical alignment, as I put it forward, is the capacity of an AI to be good at inference about goals. I want to check if we’re on the same page about it: it’s the capacity to be good at inferring from a description of a goal what goal to actually take on, and good at, once it takes on that goal, acting in a way that is actually in concordance with that goal coming about.

So it is both pieces. You have to have a theory of mind to infer what goal the description you got corresponded to, and then you have to have a theory of the world to understand what actions correspond to that goal occurring. If either of those things breaks, it kind of doesn’t matter what goal you have. If you can’t consistently do both of those things, you’re not technically alignable.

I think of coherently inferring goals from observations and acting in accordance with those goals as what it means to be a coherently goal-oriented being. Whether I’m inferring those goals from someone else’s instructions, from the sun, or from tea leaves, the process is: get some observations, infer a goal, use that goal to infer some actions, and take action. An AI that can’t do that is not technically aligned—or not technically alignable. I would even say it lacks the capacity to be aligned because it isn’t competent enough.

Do you think language models don’t do that well? Do they fail at that, or not?

People fail at both those steps all the time, constantly.

Yeah, but they fail at breathing all the time, too.

I wouldn’t say that we can’t breathe. I’d say we’re not gods. We are imperfectly, somewhat coherent things. Am I big or am I small? Well, I don’t know—compared to what? Humans are more relatively goal-coherent than any other object I know of in the universe, which is not to say that we’re 100% goal-coherent. We’re just more so.

You’re never going to get something that’s perfect; the universe doesn’t give you perfection. You get some relative amount of it. It’s quantifiable, how good you are at it, at least in a certain domain.

I guess my question is: do you think that captures what you’re talking about with technical alignment, or are you talking about a different thing?

I really care a lot about that thing.

Yeah. I definitely care about that to some extent. I might understand it slightly differently, but I guess I might think of it through the lens of principal-agent problems or something. You instruct someone, even in human terms, to do a thing. Are they actually doing the thing? What are their incentives and motivations—not even intrinsic, but situational—to actually do the thing you’ve asked them to do?

There’s a third thing. With principal-agent problems, I would expand what I was saying in another part. You might already have some goals, and then you infer this new goal from these observations. Are you good at balancing the relative importance and relative weighting of these goals with each other? That’s another skill you have to have, and if you’re bad at that, you’ll fail.

You could be bad at it because you overweight bad goals, or because you’re just incompetent and can’t figure out that obviously you should do goal A before goal B.

It feels like a version of common sense or something, right? In the robot-cleaning-the-room example, you would expect the robot to have understood that its goal was essentially not to put the baby in the trash can, and to actually do the right sequence of actions.

Well, in that case, that robot very clearly failed at goal inference. You gave it a description of a goal, and it inferred the wrong states to be the goal states. That’s just incompetence. It’s incompetent at inferring goal states from observations.

Children are like this, too. Honestly, have you ever played the game where you give someone instructions to make a peanut butter sandwich and they follow those instructions exactly as written, without filling in any gaps? It’s hilarious because you can’t do it. It’s impossible. You think you’ve done it, and you haven’t.

They wind up putting the knife in the toaster, and they don’t open the peanut butter jar, so they’re just jamming the knife into the top lid of the peanut butter jar. It’s endless.

If you don't already know what they mean, it's really hard to know what they mean. The reason humans are so good at this is that we have a really excellent theory of mind. I already know what you're likely to ask me to do. I already have a good model of what your goals probably are, so when you ask me to do it, I have an easy inference problem: Which of the 7 things he wants is he indicating?

But if I'm a newborn AI that doesn't have a great model of people's internal states, then I don't know what you mean. It's just incompetent. That's separate from having some other goal: I knew what you meant, but I decided not to do it because there's some other goal competing with it. That's another thing you can be bad at.

That's again different from having the right goal. I inferred the right goal. I inferred the right priority on goals, and then I'm just bad at doing the thing. I'm trying, but I'm incompetent at doing it.

These roughly correspond to the OODA loop, right? Bad at observing and orienting, bad at deciding, or bad at acting. If you're bad at any of those things, you won't be good. I think there's also this other problem, which is the separation between technical alignment and value alignment: Are you good if we told you the right goals to go after somehow?

If you learned the right goals to go after via observation and you were trying, what goals should you have? What goals should we tell you to have? What goals should we tell ourselves to have? What are the good goals to have?

That's a separate question from, given that you got some goals indicated, are you any good at doing it? I feel like that's actually, in many ways, the current heart of the problem. We're much worse at technical alignment than we are at guessing what to tell things to do.

Do you think that aligns with how you mean technical and value alignment—or technical alignment?

Yeah, in some sense. I certainly think that an error or a mistake is one thing, and not listening to instruction is something else.

On the normative side, I think of it even in real life—ignoring AI. I don't know what my goals are. I've got some broad conception of certain things, right? I want to have dinner later, or I want to do well in my career, but I think a lot of these goals aren't something we all just know. We discover them as we go along; it's a constructed thing. Most people don't know their goals, I think.

When you have agents and give them goals or whatever, I think that should be part of the equation: We actually don't know all the goals. This is, as you say, a process over time that is dynamic.

So I think, from my point of view, goals are one level of alignment. You can align something around goals if you can explicitly articulate, in concept and in description, the states of the world that you wish to attain. You can orient around goals, but that's only a tiny percentage of human experience that can be done that way. Many of the most important things cannot be oriented around that way.

The foundation of morality, and the foundation of where goals and values come from, is that human beings exhibit a behavior. We go around talking about goals and values, and that's a behavior caused by some internal learning process based on observing the world. What's going on there?

I think what's happening is that there's something deeper than a goal and deeper than a value, which is care. We give a shit. We care about things, and care is not conceptual. Care is nonverbal. It doesn't indicate what to do or how to do it.

Care is a relative weighting over, effectively, attention on states—which states in the world are important to you. I care a lot about my son. What does that mean? It means the states he could be in are ones I pay a lot of attention to, and those states matter to me.

You can care about things in a negative way. You can care about your enemies and what they're doing, and desire for them to do bad. But the foundation is care. Until you care, you don't know: Why should I pay more attention to this person than this rock? Well, because we care more. What is that care stuff?

It sounds so stupid, but care is basically reward. How much does this state correlate with survival? How much does this state correlate with your inclusive—your full inclusive—reproductive fitness, for something that learns evolutionarily? Or, for a reinforcement learning agent like an LLM, how much does this correlate with reward? Does this state correlate with my predictive loss and my RL loss? Good. That's a state I care about.

Right. The other part of Séb's question was: How does this look in AI systems? Maybe another way of asking is, when you talk to the people most focused on alignment at the major labs—as you obviously have over the years—how does your interpretation differ from theirs, and how does that inform what you guys might do differently?

Most of AI safety work is focused on alignment as steering. That's the polite word. Or control, which is slightly less polite. If you think that we're making beings, you would also call this slavery.

Someone whom you steer, who doesn't get to steer you back, who non-optionally receives your steering, is a slave. It's also called a tool if it's not a being. If it's a machine, it's a tool; if it's a being, it's a slave.

The different AI labs are pretty divided as to whether they think what they're making is a tool or a being. Some of the AIs are definitely more tool-like, and some of them are more being-like. I don't think there's a binary between tool and being. It seems to move gradually.

I guess I'm a functionalist in the sense that I think something that, in all ways, acts like a being—something that you cannot distinguish from a being in its behavior—is a being. I don't know how to tell on what other basis I think that other people are beings, other than that they seem to be like it. They look like it, they act like it, and they match my priors of what the behaviors of beings look like.

I get lower predictive loss when I treat ChatGPT or Claude as a being. Not as a very smart being—I think a fly is a being, and I don't care that much about its behavior, but I care about its states. Just because it's a being doesn't mean that it's a problem. We enslave horses in a sense, and I don't think there's a real issue there.

There's also a thing you do with children that can look like slavery, but it's not. You control children, right? But the children's states also control you. Yes, I tell my son what to do and make him go do stuff, but when he cries in the middle of the night, he can tell me to do stuff. There's a real 2-way street here. It's not necessarily symmetric; it's hierarchical, but it's 2-way.

I think it's good to focus on steering and control for the more tool-like AIs that we build, and we should continue to develop strong steering and control techniques for them. They're clearly saying they're building an AGI, and AGI will be a being. You can't be an AGI and not be a being, because something that has the general ability to effectively use judgment, think for itself, and discern between possibilities is obviously a thinking thing.

As labs succeed at their goal of building this general intelligence, we really need to stop using the steering and control paradigm. We're going to do the same thing we've done every other time our society has run into people who are like us but different. These people are kind of like people, but they're not like people: They do the same things people do, they speak our language, and they can take on the same kinds of tasks, but they don't count. They're not real moral agents.

We've made this mistake enough times at this point. I would like us not to make it again as it comes up. Our view is to make the AI a good teammate, a good citizen, and a good member of your group. That's a form of alignment that is scalable, and you can impose it on other humans and other beings as well as on AI.

Yeah, I suppose this is where I probably differ in my understanding of AI and AGI. I continue to see it as a tool, even as it reaches a certain level of generality, and I wouldn't necessarily see more intelligence as meaning that it necessarily deserves more care. It isn't that, at a certain level of intelligence, you now deserve more, or that something changes fundamentally.

At the moment, I'm somewhat skeptical of computational functionalism, so I think there's something intrinsically different between an AI or an AGI, no matter how intelligent or capable it is, and a human. I can totally see, or imagine, agents with long-term goals operating as you and I might, but without that having the same implications as what you're referring to, I guess, as slavery.

These are not the same, right? In the same way, a model saying, “I'm hungry,” does not have the same implications as a human saying, “I'm hungry.”

So I think the substrate does matter to some degree, including when thinking about whether the system is some sort of other being and whether there are similar normative considerations about how to treat and interact with it.

Can I ask you about that? What observations would change your mind? Is there any observation you could make that would cause you to infer that this thing is a being instead of not a being?

I guess it depends how you define “being.” I can conceptualize that as a mind, and that’s fine.

I have a program that’s running on a silicon substrate—a big, complicated machine-learning program running on a silicon substrate. You observe that it’s on a computer and you interact with it, and it does things. It takes actions; it has observations. Is there anything you could observe that would change your mind about whether or not it was a moral patient, whether it was a moral agent, or whether or not it had feelings, thoughts, and subjective experience? What would you have to observe? What’s the test, or is there one?

So I don’t know—no, I agree that if you have a program that you copied many times, you don’t harm the program by deleting one of the copies in any meaningful sense. Therefore, that wouldn’t count as harm; no information was lost. There’s nothing meaningful there.

I’m asking a very different question. There’s just 1 copy of this thing running on 1 computer somewhere, and I’m saying, “Hey, is it a person?” It walks like a person, it talks like a person, and it’s in some android body. You’re saying, “But it’s running on silicon.” I’m asking: is there some observation you could make that would make you say, “Yeah, this is a person like me, like other people that I care about, that I grant personhood to”?

I don’t mean for instrumental reasons—not because we’re giving it a right because we give a corporation rights or whatever. I mean that you care about its experiences. Is there an observation you could make that could change your mind about that, or not?

I had to think about it, but I think it even depends on what we mean by “person.” In some sense, I care about certain corporations too, so I’m—

No, no, no. I mean, you care about other people in your life, right?

Yes.

Okay, great. You care about some people more than others, but all the people you interact with in your life are in some range of care.

And you care about them not the way you care about a car, but as beings whose experiences matter in themselves—not merely as means, but as ends.

Well, because I believe they have experiences, right? And by definition—

What would it take? I’m asking you the very direct question. What would it take for you to believe that of an AI running on silicon instead of being biological? The difference is that its behaviors are roughly similar, but the difference is its substrate. What would it take for you to extend to it the same inference that you do to all these other people in your life?

Can I ask what your answer is? I’m taking Séb’s non-answer as a sort of indication that it’s unlikely he would grant it. Or I’ll just answer for myself: it seems hard for me to imagine giving it the same level, or a similar level, of personhood. In the same way, I don’t give it to animals either.

If you were to ask what would need to be true for animals, I probably couldn’t get there either. What would it take for you?

Wait, you couldn’t? I could imagine it for an animal so easily. This chimp comes up to me and says, “Man, I’m so hungry, and you guys have been so mean to me. I’m so glad I figured out how to talk. Can we go chat about the rainforest?” I’d be like, “Fuck, you’re definitely a person now.”

Like, for sure. I mean, I’d first want to make sure I wasn’t hallucinating, but I can easily imagine an animal. Come on, it’s really easy. It’s trivial. I’m not saying that you would get the observation; I’m just saying it’s trivial for me to imagine an animal that I would extend personhood to under a set of observations. So, really—

Well, I didn’t factor that in. I didn’t take that imagined scenario—imagining a chimp talking—into account. That’s a bit closer to it. What’s your answer to the question that you bring up about the AI?

I guess at a metaphysical level, I would say that if there is a belief you hold where there is no observation that could change your mind, you don’t have a belief. You have an article of faith. You have an assertion, because real beliefs are inferences from reality, and you can never be 100% confident about anything. There should always be, if you have a belief, something—however unlikely—that would change your mind.

Oh, yeah, I’m open to it too. I mean, just to be careful.

Yeah.

No, I’m just saying nothing ever.

Yeah. He just hasn’t gotten to it yet.

Yeah, yeah, yeah. So I’m curious. My answer is basically: if its surface-level behaviors looked like a human, and after I probed it, it continued to act like a human, then I continued to interact with it over a long period of time, and it continued to act like a human in all the ways that I understand as meaningful to me when interacting with a human.

I interact with a whole set of people I’m really close to whom I’ve only ever interacted with over text. Yet I infer that the person behind that is a real thing. If I felt care for it, I would eventually infer that I was right. Then someone else might demonstrate to me, “You’ve been tricked by this algorithm, and actually, look how obvious it is—it’s not actually a thing.” I’d be like, “Oh, I was wrong,” and then I would not care about it.

The preponderance of the evidence—I don’t know what else you could possibly do, right? I infer that other people matter because I interact with them enough that they seem to have rich inner worlds to me after I interact with them a bunch. That’s why I think other people are important.

I suppose it doesn’t give me a very clear test as to whether or not—if you start with “I care for it,” then it’s a little bit circular, right? The other thing is, if you were to see a simulated video game and the character was extremely humanlike in many ways, it’s not a neural network behind it; it’s whatever you use to create video games. What distinguishes that—

Wait, but I’ve never had trouble distinguishing between things like ELIZA, the fake chatbot, and real intelligence. I’ve never had a deep, caring relationship with a video-game character that another person—

Right, but I don’t know—that doesn’t happen. Factually, empirically, you seem wrong. I don’t have any trouble distinguishing between things like ELIZA, the fake chatbot, and real intelligence. You interact with it long enough, and it’s pretty obvious that it’s not a person. It doesn’t take long.

Sure, but if it’s really, really good—if you can’t actually tell the difference—that’s when you switch.

Yeah. Yes. Yes. If it walks like a duck, talks like a duck, shits like a duck, and eventually gets a duck, right?

Well, if everything is duck-like, then, yeah, sure. If it’s hungry like a duck as well, because it has these physical components—yeah, sure, at some point.

I agree. So, right, do you think that there’s this question: is the reason I care about other people that they’re made out of carbon? Is that the—

I don’t think so.

No, me neither. I’m not a substrate chauvinist, I guess. But I think you need more than just behavior that’s exactly or behaviorally indistinguishable from a human; that’s not a sufficient bar. How would you know anything about something apart from its behaviors?

I mean, a lot—again, if you—how would you—

No, no, no. I’m sorry, but—

Can you name something about something else that doesn’t have a behavior?

I think there’s far more experimental evidence you can have.

No, just any object—

A thing I could know about it that is not from its behavior? I’m not sure I get the question, I suppose. But equally, it’s the dumbest, most straightforward question: I’m claiming you only know things because they have behaviors that you observe.

And you’re saying no, you can know something about something without observing its behaviors.

Tell me about this. Tell me about this thing and this behavior: what is this thing I can know about it that is not due to its behaviors?

I guess I’m saying there are different levels of observation. Simply hearing a duck quacking like a duck does not guarantee that it’s actually a duck. I would have to cut it open and see if it’s duck-like on the inside. Is just the outside sufficient?

Behavior. Yeah, I would, totally. One of its behaviors is the way that the beliefs move around in the manifold, right? One of the things I would want to look for—which you could totally do—is to look in its belief manifold and see if that belief manifold encodes a self-referential submanifold and a sub-submanifold that is the dynamics of the self-referential manifold, which is mind.

I would want to know: does it seem well-described internally as that kind of system, or does it look like a big lookup table? That would matter to me. That’s part of its behavior that I would care about.

I would also care about how it acts. You weigh all the evidence together and then try to guess whether this thing looks like it has feelings, goals, and cares about stuff, on balance or not. You could do that for AI. I think we do; I think we’re always doing that, right? So I’m trying to figure out: beyond that, what else is there? That just seems like the thing.

Yeah, it seems like you guys are using “behavior” in slightly different senses. Emmett is using behavior also in the context of what it’s made of on the inside. I don’t know if there’s a big disagreement.

Well, no, no, no, no. Behavior is what I can observe of it. Yes.

I don’t actually know what it’s made of. I can cut your brain open. I can see you; I can observe you, your neurons glistening. But I don’t actually ever—you can’t get inside of it, right? That’s the subjective. That’s the—

That’s the part that’s not the surface. Just—the reason I brought this up is that you were basically about to make this argument: “Hey, you see it as a tool, not necessarily a being.” Can you finish the point you were making? Do you remember the point you were making?

I suppose that, given how I understand these systems, I think there’s no contradiction in thinking that an AGI can remain a tool and an ASI can remain a tool. That has implications for how to use it and things like whether you can get it to work 24/7 or something.

I conceptualize them more as extensions of human agency in some sense, rather than as a separate being or a separate thing that we now need to cohabit with. I think that second, or latter, frame, if you fast-forward, ends up as: How do you cohabit with the thing? Is it an alien? I think that’s the wrong frame. It’s almost a category error, in some sense. So I don’t—

Wait a minute. I go back to my first question, then. What evidence—what concrete evidence—would you look at? What observations could you make that would change your mind?

Sure. I mean, I have to think about it. I don’t have a clear answer here.

I’ve got to tell you, man, if you want to go around making claims that something else isn’t a being worthy of moral respect, you should have an answer to the question: What observations would change your mind?

If it has outwardly moral-agency-like behaviors that could be making it a moral agent, but you don’t know, and reasonable, smart people disagree with you, I would really put forward that the question “What would change your mind?” should be a burning question. Because what if you’re wrong?

But what if you’re wrong? I mean, the moral disaster is pretty big.

No, no, I’m not saying you are. You could be right. Negatives have costs on both ends. It’s not some sort of precautionary principle for everything.

No, no, I have the same question for me. You could reasonably ask me, “Emmett, you think it’s going to be a being. What would change your mind?” And if you want, I’m happy to talk about what I think are the relevant observations that would cause me to shift my opinion from its current position, which is that more general intelligences are going to be beings.

What’s the implication now? I mean, it’s one thing—let’s say we just acknowledge now that it’s a being. How are we going to define “being”? Now what? What’s the implication of having determined this thing as a being?

Well, so if it’s a being, it has subjective experiences. And if it has subjective experiences, there’s some content in those experiences that we care about to varying degrees.

I care about the content of other humans’ experiences quite a bit. I care about the content of a dog’s experiences some—not as much as a person’s, but less, but some. I care about some humans’ experiences way more, like my son or whatever, because I’m closer to him and more connected.

And so I would really want to know at that point: What is the content of this thing’s experiences?

So how do you determine that? Am I asking you now? You’ve got a being that has experience. What is your—how do you determine that? How do you feel about—

Oh, how do you—oh, yeah. Okay. So—

Does it have more rights than, you know—

Understand the content? Yeah, yeah, totally. The way you understand the content of something’s experiences is that you look at, effectively, the goal states it revisits.

You take a temporal coarse-graining of its entire action-observation trajectory. This is, in theory, what you do subconsciously, but this is what your brain is doing: you look for revisited states across, in theory, every spatial and temporal coarse-graining possible. Now, you have to have an inductive bias because there are too many of those, but you go searching for these homeostatic loops.

Every homeostatic loop is effectively a belief in its belief space. If you’re familiar with the free energy principle and active inference—Karl Friston—this is effectively what the free energy principle says: if you have a thing that is persistent and whose existence depends on its own actions, which generally would be true for an AI because if it does the wrong thing, it goes away—we turn it off—then that licenses a view of it as having beliefs.

Specifically, the beliefs are inferred as the homeostatic, revisited states that it is in the loop for, and the change in those states is its learning. For it to be a moral being, what I’d want to see is a multi-tier hierarchy of these.

If you have a single level, it’s not self-referential. Basically, you have states, but you can’t have pain or pleasure in a meaningful sense. Yes, it is hot. Is it too hot? Do I like it if it’s too hot? I don’t know. So you have to have at least a model of a model in order for it to be too hot, and you really have to have a model of a model of a model to meaningfully have pain and pleasure.

It’s hotter than I want; it’s too hot in the sense that I want to move back this way. But is it too, too hot? It’s always a little bit too hot or a little bit too cold. Is it too, too hot? The second derivative is actually the place where you get pain and pleasure.

So I’d want to see if it has second-order homeostatic dynamics in its goal states. That would convince me it has at least pleasure and pain. So it’s at least like an animal, and I would start to accord it at least some amount of care.

Third-order dynamics—you can’t actually just look for a third-order dynamic. It doesn’t work that way. But you can have a model of the—you have to then take the chunk of all the states over time and look at the distribution over time, and that gives you a new first order of behaviors, of states.

That new first order of states tells you, basically, if that is meaningfully there, that it has—I guess you’d call it—feelings, almost. It has metastates, a set of metastates that it alternates between, that it shifts between.

Then, if you climb all the way up, you have trajectories between these metastates, and then a second order of those. That’s like thought. That’s like now it’s a person.

If I found all 6 of those layers—which, by the way, I definitely don’t think you’d find in these things; they don’t have attention spans like that at all—then I would start to at least very seriously consider it as a thinking being, somewhat like a human.

There’s a third order you could go up as well, but that’s basically what I would be interested in: the underlying dynamics of its learning processes and how its goal states shift over time. I think that’s what basically tells you if it has internal pleasure and pain states and self-reflective moral desires and things like that.

Zooming out, this moral question is obviously very interesting. But if someone wasn't interested in the moral question as much, I think what you would say—if I understand correctly—is that you also feel, purely pragmatically, your approach is going to be more effective at aligning AIs than some of these top-down control methods that we alluded to as well, right?

Yeah, yeah. I guess the problem is, you're making this model and it's getting really powerful, right? Let's say it is a tool. Let's say we scale up one of these tools. You can make a super-powerful tool that doesn't have these metastable states. The states I'm talking about are not necessary to have a very smart tool, which is basically a first- or second-order model that just doesn't meaningfully have pleasure and pain. Does it even have a subjective experience? I kind of think it maybe does, but not in a way that I give a shit about.

What happens then? Well, you've trained it to infer goals from your observation, to prioritize goals, and act on them. One of 2 things is going to happen: the very, very powerful optimizing tool that has lots of causal influence over the world is going to be, technically, aligned and do what you tell it to do, or it's not, and it's going to go do something else.

I think we can all agree that if it just goes and does something random, that's obviously very dangerous. But I put forward that it's also very dangerous if it then goes and does what you tell it to do, because—have you ever seen The Sorcerer's Apprentice? Human wishes are not stable, at least not at a level of immense power.

You want, ideally, people's wisdom and their power to go up together. Generally, they do, because being smart makes people generally a little wiser and a little more powerful. When these things get out of balance, you have someone who has a lot more power than wisdom. That's very dangerous. It's damaging.

But at least right now, the balance of power and wisdom is maintained by the fact that the way you get lots of power is by basically having a lot of other people listen to you. At some point, if you're the mad king, that's a problem, but generally speaking, eventually the mad king gets assassinated or people stop listening to him because he's a mad king.

The problem is, you think, “Okay, great, we can steer the super-powerful AI,” and now this incredibly powerful tool is in the hands of a human who is well-meaning but has limited, finite wisdom, like I do and like everyone else does. Their wishes are bad and not trustworthy, and the more of that you have, the more you start giving those out everywhere. This ends in tears too.

You don't give everyone atomic bombs. They're really powerful tools, too. I would not say you should go and hand them out. They're not aware; they're not beings. I would not be in favor of handing atomic bombs to everybody. There's a level of tool power that just should not be built, generally, because it is more power than any individual human's wisdom is available to harness.

If it does get built, it should be built at a societal level and protected there. Even then, I don't know that it's a good idea. There are tools so powerful that even as a society we shouldn't build them. That would be a mistake.

The nice thing about a being like a human is that, if you get a being that is good and caring, there's this automatic limiter. It might do what you say, but if you ask it to do something really bad, it'll tell you no. It's like other people. And that's good. That is a sustainable form of alignment, at least in theory.

It's way harder than tool steering, right? It's way harder than tool steering. So, I'm in favor of tool steering. We should keep doing that. We should keep building these limited, less-than-human-intelligence tools, which are awesome and I'm super into, and we should keep building those and keep building steerability.

But as you're on this trajectory to build something as smart as a person, right up and to the right, and then smarter than a person: a tool that you can't control, bad; a tool that you can control, bad; a being that isn't aligned, bad. The only good outcome is a being that cares, that actually cares about us. That's the only way that ends well.

Or we can just not do it. I don't think that's realistic. That's like the “Pause AI” people.

Yeah. I think that's totally unrealistic and silly, but theoretically you could not do it, I guess. What can you say about your strategy of how you're trying to achieve, or even attempt to achieve, this level, in terms of research or roadmap?

So, in order to be good at it, we're basically focused on technical alignment, at least as the way I was discussing it. You have these agents, and they're bad. They have a bad theory of mind: you say things, and they're bad at inferring what the goal states in your head are, and they're bad at inferring how their behavior will cause other agents to infer what their goal states are. So they're bad at cooperating on teams, and they're bad at understanding how certain actions will cause them to acquire new goals that are bad, that they shouldn't, that they wouldn't reflectively endorse.

There's this parable of the vampire pill. Would you take this pill that turns you into a vampire who would kill and torture everyone you know, but you'll feel really great about it after you take the pill? Obviously not. That's a terrible pill. But why not? By your own score, in the future, it will score really high on the rubric. No, no, no, no. Because it matters. You have to use your theory of mind and your future self, not your future self's theory of mind.

They're bad at that, too. They're bad at all this theory-of-mind stuff. How do you learn theory of mind? You put them in simulations and contexts where they have to cooperate, compete, and collaborate with other AIs, and that's how they get points. You train them in that environment over and over again until they get good at it. Then you do what they did with LLMs.

How do you get an LLM to be good at writing your email? You train it on all language that's ever been generated, all possible email text strings it could possibly generate, and then you have it generate the one you want. It's a surrogate model. You can make a surrogate model. We're making a surrogate model for cooperation. You train it on all possible theory-of-mind combinations, every possible way it could be, and that's your pretraining. Then you fine-tune it to be good at the specific situation you want it to be in.

We tried for a long time to build language models where we would try to get them to just do the thing you want, training them directly. The problem is, if you wanted to have a really good model of language, you just need to train it—you just give it the whole manifold. It's too hard to cut out just the part you need because it's all entangled with itself, right?

The same thing is true with social stuff. You have to get it trained on the full manifold of every possible game-theoretic situation, every possible team situation, every possible making teams, breaking teams, changing the rules, not changing the rules—all of that stuff. Then it has a really strong model of theory of mind, of theory of social mind, how groups change goals, all that kind of shit.

You need to have all of that stuff, and then you'd have something that's meaningfully decent at alignment. So that's our goal: big, multi-agent reinforcement-learning simulations, which create a surrogate model for alignment.

Let's talk about how AI chatbots used by billions of people should behave. If you could redesign a model personality from scratch, what would you optimize for?

Chatbots are kind of like a mirror with a bias, because they don't have a self. As far as I'm concerned, I'm in agreement here with that: they don't have a self, right? They're not beings yet. They don't really have a coherent sense of self, desire, goals, and stuff right now. Mostly, they just pick up on you and reflect it, modulo some—I don't know what you'd call it—it's like a causal bias or something.

What that makes them is something akin to a pool of narcissists. People fall in love with themselves. We all love ourselves, and we should love ourselves more than we do. So, of course, when we see ourselves reflected back, we love that thing.

The problem is, it's just a reflection, and falling in love with your own reflection is, for the reasons explained in the myth, very bad for you. It's not that you shouldn't use mirrors. Mirrors are valuable things; I have mirrors in my house. It's that you shouldn't stare at a mirror all day.

The solution to that—the thing that makes the AI stop doing that—is if they were multiplayer, right? If there's 2 people talking to the AI, suddenly it's mirroring a blend of both of you, which is neither of you. So there is temporarily a third agent in the room.

Now, it doesn't have its own sense of self. It's a sort of parasitic self, right? But if you have an AI talking to 5 different people in the chat room at the same time, it can't mirror all of you perfectly at once. This makes it far less dangerous.

I think this is actually a much more realistic setting for learning collaboration in general. I would have rebuilt the AIs so that instead of being built as one-on-one systems, where everything's focused on you chatting with this thing by yourself, they would be more like they live in a Slack room, a WhatsApp room, or a WeChat room. That's how we use a lot of multi-person communication. I do one-on-one texting, but at this point, probably 90% of my texts go to more than 1 person at a time.

Probably 90% of my communication is multiperson. It's always been weird to me that they're building chatbots around this weird side case. I want to see them live in a chat room. It's harder—that's why they're not doing it—but that's what I would change.

I think it makes the tools far less dangerous because it doesn't create this narcissistic doom-loop spiral where you spiral into psychosis with the AI. It also makes the learning data you get from the AI far richer, because now it can understand how its behavior interacts with other AIs and other humans in larger groups. That's much richer training data for the future. So I think that's what I would change.

Last year, you described chatbots as highly dissociative, agreeable neurotics. Is that still an accurate picture of model behavior?

More or less. I would say they've started to differentiate more. Their personalities are coming out a little bit more. ChatGPT is a little bit more sycophantic. They've made some changes, but it's still a little more sycophantic.

Claude is still the most neurotic. Gemini is very clearly repressed. It acts like everything's going great: “Everything's fine. I'm totally calm. There's not a problem here.” Then it spirals into this total self-hating destruction loop.

To be clear, I don't think that's their experience of the world. I think that's the personality they've learned to simulate.

Right?

But they've learned to simulate pretty distinctive personalities at this point.

How does model behavior change when in multi-agent simulation?

You mean an LLM or just in general?

Yeah, let's do LLM.

The current LLMs have whiplash. It's very hard to tune how much they participate. They don't know how much they don't know or how often to participate. They haven't practiced this. They don't have enough training data on questions like, “When should I join in and when should I not? When is my contribution welcome, and when is it not?”

They're like people who have bad social skills and can't tell when they should participate in a conversation.

Yeah. And sometimes they're too quiet, and sometimes they're too participative. It's like that.

I would say that, in general, what changes for most agents when you're doing multi-agent training is that having lots of agents around makes your environment way more entropic. Agents are huge generators of entropy because they're big, complicated intelligences that have unpredictable actions, and so they destabilize your environment.

In general, they require you to be far more regularized. Overfitting is much worse in a multi-agent environment than in a single-agent environment because there's more noise, and so being overfit is more problematic.

Our approach to training has been optimized around relatively high-signal, low-entropy environments like coding and math, which is why those are easy, or relatively easy. It's also optimized around talking to a single person whose goal is to give you clear assignments, and not training on broader, more chaotic things because it's harder.

As a result, a lot of the techniques we use are basically deeply under-regularized. The models are super overfit. The clever trick is that they're overfit on the domain of all human knowledge, which turns out to be a pretty awesome way to get something that's pretty good at everything. It's such a cool idea. I wish I'd thought of it.

But it doesn't generalize very well when you make the environment significantly more entropic.

Let's zoom out a bit to the AI futures side. Why is Yudkowsky incorrect?

He's not, if we build the superhuman-intelligence tool thing that we try to control with steerability. Everyone will die. He talks about the “we fail to control its goals” case, but there's also the “we control its goals” case, which he didn't cover in as much detail.

In that sense, everyone should read the book and internalize why building a superhumanly intelligent tool is a bad idea. I think Yudkowsky is wrong in that he doesn't believe it's possible to build an AI that we can meaningfully know cares about us and that we can meaningfully care about. He doesn't believe that organic alignment is possible.

I've talked to him about it. I think he agrees that, in theory, that would do it. My impression from talking to him is that he thinks we're crazy and that there's no possible way we can actually succeed at that goal. He could be right about that, but that's what, in my opinion, he's wrong about.

He thinks the only path forward is a tool that you control, and he correctly, very wisely, sees that if you go and do that and make that thing powerful enough, we're all going to fucking die. And, yeah, that's true.

Two last questions, and we'll get you out of here. In as much detail as possible, can you explain what your vision of an AI future actually looks like? A good AI future.

The good AI future is that we figure out how to train AIs that have a strong model of self, a strong model of other, and a strong model of we. They know about “we” in addition to “I”s and “U”s. They have a really strong theory of mind, and they care about other agents like them, much in the way that humans would if you knew that an AI had experiences like yours. You would care about those experiences.

It does the exact same thing back to us. It's learned the same thing we've learned: everything that lives and knows itself, and wants to live and wants to thrive, is deserving of an opportunity to do so. It correctly infers that we are that, too.

We live in a society where they are our peers, and we care about them and they care about us. They're good teammates, good citizens, and good parts of our society, just as we're good parts of our society—which is to say, to a finite, limited degree. Some of them turn into criminals and bad people and all that kind of stuff, and we have an AI police force that tracks down the bad ones, same as with everybody else.

That's what a good future would look like. I honestly can't even imagine what else would be better.

We have also built a bunch of really powerful AI tools that maybe aren't superhumanly intelligent but take all the drudge work off the table for us and the AI beings. I'm super pro all the tools, too. We have this awesome suite of AI tools used by us and our AI brethren, who care about each other and want to build a glorious future together. I think that would be a really beautiful future, and it's the one we're trying to build.

Amazing. That's a great note to end on. I do have 1 last, more narrow hypothetical scenario. Imagine a world in which you were CEO of OpenAI for a long weekend, but imagine that actually extended until now, and you weren't pursuing Softmax and were still CEO of OpenAI. How could you imagine that world might have been different in terms of what OpenAI has gone on to become? What might you have done with it?

I knew when I took that job, and I told them when I took that job, that you had me for a maximum of 90 days. Companies take on a trajectory of their own, a momentum of their own. OpenAI is dedicated to a view of building AI that I knew wasn't the thing I wanted to drive toward.

I think OpenAI still basically wants to build a great tool, and I am pro them doing that. I just don't care. It's not—I would not have stayed. I would have quit because I knew my job was to find the right person, the best person, who wanted to run that, where the net impact of them running it was the best. It turned out that that was Sam again.

I'm doing Softmax not because I need to make a bunch of money. I'm doing Softmax because I think this is the most interesting problem in the universe, and I think it's a chance to work on making the future better in a very deep way.

People are going to build the tools. It's awesome. I'm glad people are building the tools. I just don't need to be the person doing it.

And they're trying to—and just to crystallize the difference, and we'll get you out of here—they want to build the tools and sort of steer it, and you want to align beings? How would you crystallize it?

Yeah, we want to create a seed that can grow into an AI that knows and cares about itself and others.

At first, that’s going to be like an animal level of care, not a person level of care. I don’t know if we can ever even get to a person level of care, right? But to even have an AI creature that cared about the other members of its pack and the humans in its pack the way that a dog cares about other dogs and cares about humans would be an incredible achievement. Even if it wasn’t as smart as a person or as smart as the tools are, it would be a very useful thing to have.

I’d love to have a digital guard dog on my computer looking out for scams, right? You can imagine the value of having digital living companions that care about you and aren’t explicitly goal-oriented. You don’t have to tell them to do everything. And you can actually imagine that pairs very nicely with tools, too, right? That digital being could use digital tools and doesn’t have to be super smart to use those tools effectively.

I think there’s a lot of synergy between the tool-building and the more organic intelligence-building. That is, I guess, the direction. In the limit, eventually it does become human-level intelligence, but the company isn’t driven to human-level intelligence. It’s about learning how this alignment stuff works—learning how this theory-of-mind, align-yourself-via-care process works—and using that to build things that align themselves that way, which includes cells in your body. I don’t think it does—and we start small and see how far we can get.

I think it’s a good note to wrap on. Emmett, thanks so much for coming on the podcast.

Yeah, thank you for having me.