控制工具，还是对齐生物？Emmett Shear（Softmax）与Séb Krier（GDM），来自a16z Show

Most of AI is focused on alignment as steering. That's the polite word. If you think that they're moral 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—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. We've made this mistake enough times at this point. I would like us not to make it again.

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—one that actually cares about us.

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

Thank you for having me.

So, Emmett, with Softmax, you're focused 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," and 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 mean, I guess you could be aligned to yourself, but even then, they don't want to tell you that 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 this usually means 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 that's what I would regard as a public good, right? 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 it's important to recognize 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 again, reconstructing the rock.

Now, a 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 of how people in families stay aligned to each other, 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, and if you stop doing it, it goes away.

This is similar for things like cells in your body. 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 to be a new cell? Do we need to be making more red blood cells? Should we be making 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 tablets down from on high that tell you how to be a morally good being, and we use those. They're maybe helpful, but somehow you can read them and try to follow those rules and still make lots of mistakes. You can have the realization, "I've been a dick. That was bad. I thought I was doing good, but in retrospect, I was doing wrong."

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 that'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 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, right? 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."

And it's not random. People have the same—actually, there's a bunch of classic patterns of people having that realization.

It's a thing that happens over and over again, so it's not random. 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.

So I'm a moral realist. I'm taking a very strong moral realist position. There is such a thing as morality. We really do learn it, and it really does matter.

Organic alignment is 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, and I don't need to learn anything. No one has anything to teach me about morality.” That's one of the main forms of arrogance, and it's one of the main moral things you can do that's dangerous.

So what do we mean when we talk about organic alignment? Organic alignment isn't aligning an AI that is 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—of learning how to be a good family member, a good teammate, a good member of society, and a good member of all sentient beings.

I guess it's learning how to be a part of something bigger than yourself in a way that's 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 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 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.

And so that's the bar. That's what we should be working on, and 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 my whole heart, if someone beats us, thank God. I hope somebody figures it out.

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

I guess I understand it in a slightly different way, perhaps 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 it to do? How do you get it to follow instructions, roughly speaking?

I think that was more of a challenge pre-LLM, when people were talking about reinforcement learning and looking at these systems, whereas post-LLM, we've realized that many things we thought were going to be difficult were somewhat easier.

Then there's a second question, the normative question: to whose values are you aligning this thing? I think that's the kind of thing you're commenting on a bit. 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 is 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 a complete answer, so you have processes that allow ideas to clash.

You have good people with different ideas, opinions, and views 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 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 it exactly looks like if we're going to implement this into an AI system, at least the ones we have today.

I agree. I agree that there's an idea of technical alignment that 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 kind of 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, by definition.

Is that a fair assessment of what you mean by technical alignment?

I'm not fully sure. If I give a model a certain goal, I would like the model to follow that instruction and reach that particular goal, rather than it having a goal of its own that I can't—

Well, if you give it a goal, it has that goal.

Right.

You give someone something, right?

So, yeah, if I instruct it to do X, then I would like it to do X and not different variants of X, essentially. I wouldn't want it to reward-hack. 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 wanted it to infer what you meant, right?

Right. In some sense, the byte sequence that you send 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 so. One way I remember it from when I was first getting into AI and these kinds of questions, maybe a decade ago, is that you have these examples in Stuart Russell's textbook where you give the AI a goal, but then it doesn't exactly do what you're asking. You tell it to clean the room, and it cleans the room but takes the baby and puts it in the trash. This is not what I meant.

Wait, hold on. 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 I'm evoking 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.

Mhm.

Giving someone “Hey, go do this” isn't 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 that we do it so quickly and naturally that we don't even see it happening.

We think that we get confused and we think those are the same thing, but you haven't given it a goal. You've given it a description of a goal, and you hope it turns back into the goal that's the same as the goal you described inside of you.

Right?

You think 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.”

Now it's the same, but I don't think most people mean that when they say they gave it a goal.

True. Is this distinction you're making, Emmett, important because there's some lossiness between the description and the actual goal?

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

So it is both pieces. You have to have the theory of mind to infer what that description of a goal that you got—what goal that 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 a coherently goal-oriented being.

Inferring goals from observations and acting in accordance with those goals is what I think of as being a coherently goal-oriented being, because that's what the process is. Whether I'm inferring those goals from someone else's instructions or from the sun or 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 technically alignable. I would even say it lacks the capacity to be aligned because it's not competent enough.

And you think language models don't do that well? As in, they kind of fail at that?

People fail at both of those steps all the time, constantly. I tell employees to do stuff, and we fail at breathing all the time too. I wouldn't say that we can't breathe; I just say that we're not gods. We are imperfectly coherent, relatively coherent things.

Am I big or am I small? 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. I think you're never going to get something that's perfect. The universe doesn't give you perfection; it gives you relative amounts of coherence. It's a quantifiable thing, how good you are at it, at least in a certain domain.

My question is, does that capture 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 mean, 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 maybe principal-agent problems or something. You kind of 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. Principal-agent problems—I would expand what I was saying in another part. You might already have some goals, and then you inferred this new goal from these observations. Are you good at balancing the relative importance and relative priority 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 you could be bad at it because you're just incompetent and can't figure out that obviously you should do goal A before goal B.

I feel like a version of common sense, right? The kind of thing that, in fact, in the robot-cleaning-the-room example, you would expect the robot to have understood that it should essentially not put the baby in the trash can or something, and just actually do the right sequence of actions.

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

Mhm.

That's separate from having some other goal. I knew what you meant, but I decided not to do it because there was some other goal competing with it. That's another thing you can be bad at.

Which is again different from: I had 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, bad at acting. If you're bad at any of those things, you won't be good. And then I think there's this other problem. I like the separation between technical alignment and value alignment, which is: 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.

I do. Do you think that aligns with how you mean technical and value alignment, or technical—

Yeah, in some sense. I mean, I certainly think that there's something—an error or mistake is one thing, and then there's not listening to the instruction, which is something else.

But, yeah, I think 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 I want to get—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 constructive thing.

Most people don't know their goals, I think. So when you have agents and give them goals or whatever, I think that should be part of the equation: we don't actually 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. The kind of goals we're talking about here 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 only a tiny percentage of human experience can be done that way. Many of the most important things cannot be oriented around that way.

And the foundation, I think, of morality—and the foundation, I think, of where goals come from and where values come from—are questions worth asking. Human beings exhibit a behavior: we go around talking about goals and values, and that's a behavior caused by some internal learning process that's 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. It doesn't indicate how to do it. Care is effectively a relative weighting over attention on states. It's a relative weighting over which states in the world are important to you.

I care a lot about my son. What does that mean? It means his states—the states he could be in—I pay a lot of attention to those, and those matter to me. You can care about things in a negative way. You can care about your enemies and what they're doing, and you can desire for them to do bad. But I think the foundation is care. You don't just want it to care about us. You want it to care about us and like us too, right? Maybe.

Until you care, you don't know: Why should I pay more attention to this person than this rock? Well, we care more. And what is that care stuff? I think what it appears to be, if I had to guess, is that the care stuff is—this sounds so stupid—but care is basically reward. How much does this state correlate with survival? How much does this state correlate with 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. I think that's kind of what it is.

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

Most of the AI is focused on alignment as steering—that's the polite word—or control, which is slightly less polite. If you think that they were making beings, you would also call this slavery. Someone whom you steer, who doesn't get to steer you back, is a slave. Someone 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. I think the different AI labs are pretty divided as to whether they think what they're making is a tool or a machine. I think some of the AIs are definitely more tool-like, and some of them are more machine-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 behaviors—is a being. Because I don't know on what other basis to think that other people are beings, other than that they seem to be like beings: They look like it, they act like it. They match my priors of what the behaviors of beings look like. I get lower predictive loss when I treat them as a being. And the thing is, I get lower predictive loss when I treat ChatGPT or Claude as a being.

Now, not as a very smart being. I think a fly is a being, and I don't care that much about its behavior, about its states. Just because it's a being doesn't mean it's a problem. We sort of enslave horses in a sense, and I don't think there's a real issue there.

And there's even 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 also, when he cries in the middle of the night, he can tell me to do stuff. There's a real two-way street here because it's not necessarily symmetric. It's hierarchical, but two-way.

Basically, I think it's good to focus on steering and control for tool-like AIs, and we should continue to develop strong steering and control techniques for the more tool-like AIs that we build. They're clearly saying they're building an AGI. An 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 you go from what we have today, which is mostly a very specific intelligence, not a general intelligence, to labs succeeding 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. 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. Make the AI a good citizen. Make the AI a good member of your group. That's a form of alignment that is scalable, and you can apply it to other humans and other beings, as well as to AI.

Yeah, I suppose this is kind of where I probably differ in my understanding of AI and AGI. I guess I continue seeing it as a tool even as it reaches a certain level of generality. I wouldn't necessarily see more intelligence as meaning it deserves more care, necessarily. At a certain level of intelligence, you now deserve certain moral rights, or something changes fundamentally. I guess, at the moment, I'm somewhat skeptical of computational functionalism.

And so I think there’s something intrinsically different between an AI or an AGI, no matter how intelligent or capable. I can totally see or imagine agents with long-term goals and operating as you and I might, but without that having the same implications. I think in the same way that 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 for thinking about whether to think of this as some sort of other being, whether it has similar normative considerations about how to treat and act 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, “This thing is a being,” instead of not a being?

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

I have a program that’s running on a silicon substrate. Some big, complicated machine-learning program running on a silicon substrate. You observe that it’s on a computer, interact with it, and it does things. It takes actions and 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 it had feelings and thoughts and subjective experience? What would you have to observe? What’s the test, or is there one?

There are a lot of different questions here, I think, with some conflict. On the one hand, there are normative considerations, because you can give rights to things that aren’t necessarily beings. A company has rights in some sense, and these are useful for various purposes.

I also think biological beings and systems have a very different substrate. You can’t separate certain needs and particularities about what they are from the substrate. I can’t copy myself. If someone stabs me, I probably die, whereas machines have a very different substrate. I think there’s also a more fundamental disagreement about what happens at the computational level, which is different from what happens with biological systems.

But I agree that if you have a program that you copied many times, you don’t harm the program by deleting 1 of the copies in any meaningful sense. Therefore, that wouldn’t count as harm—no information was lost, right? 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. I’m saying, hey, is it a person? It walks like a person, talks like a person, and it’s in some android body. But it’s running on silicon. What 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”?

Not for instrumental reasons—not because we’re giving it a right because we give a corporation rights or whatever. I mean, where you think some person you care about—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.

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 a being whose experience matters in itself—not merely as a means, but as an end.

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 substrate is different. What would it take for you to extend that same inference to it that you do to all these other people in your life?

Can I ask what your answer is? I’m taking the non-answer as an indication that it’s unlikely you would grant it. For myself, it seems hard for me to imagine giving it the same 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 that 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 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 in imagining a chimp talking. That’s a bit closer to it. What’s your answer to the question you bring up about the AI?

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. So there should always be something—however unlikely—that would change your mind.

Oh yeah, I’m open to it.

Care? Nothing ever?

He just hasn’t gotten to it yet.

Yeah, yeah, yeah. I’m curious. My answer is basically that if its surface-level behaviors looked like a human, and after I probed it, it continued to act like a human, and 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 being meaningful to me in interacting with a human, I would infer that it was a real thing.

I interact with a whole set of people I’m really close to only 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. Actually, look how obvious it is—it’s not really a thing.” I’d be like, “Oh, fuck, I was wrong,” and then I would not care about it.

The preponderance of the evidence would be enough. I don’t know what else you could possibly do, right? I infer that other people matter because I’ve interacted 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 always a little bit circular, right? The other thing is, if you were to see a simulated video-game character that 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 that. I’ve never had a deep, caring relationship with a video-game character that didn’t have a person behind it.

Right, I don’t know. That doesn’t happen. In fact, empirically, you seem wrong. I don’t have any trouble distinguishing between things like ELIZA, the fake chatbot, and a real intelligence. You interact with it long enough, and it’s pretty obvious 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. If it walks like a duck, talks like a duck, shits like a duck, and eventually gets hungry, right?

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

Yeah, I agree. So there’s this question, right? 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 mean, I’m not a substrate chauvinist, I guess, if that’s the— But I think you need more than just its acting exactly the same. Being behaviorally indistinguishable is not a sufficient bar.

How would you know what else there is to know about something apart from its behaviors?

No, no, no. I'm sorry, but can you name something about something else that doesn't have a behavior?

No, just any object.

Uh-huh.

And a thing I could know about it that is not from its behavior. I'm not sure I get the question, I suppose, but it's a dumb, 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 thing and this behavior, and this thing I can know about it that is not due to its behaviors. I guess I'm saying there's different levels of observation. Simply having something quack like a duck does not guarantee that it's actually a duck. I would have to cut it open and look and see if it's duck-like on the inside, not just on the outside.

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

I would want to know: does this seem well-described internally as that kind of a system, or does it look like a big lookup table? That would matter to me. That's part of its behaviors that I would care about. I would also care about how it acts, and you weigh all the evidence together, and then you try to guess: does this thing look like a thing that has feelings, goals, and cares about stuff, on balance, or not?

I can't imagine what else there is. I think you could do that for the AIs. I think we do that for the AIs. I think we're always doing that, right? 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 a slightly different sense. 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 only cut your brain open. I can see you—I can observe your neurons firing and glistening, your neurons glistening—but I don't actually ever... You can't get inside of it, right? That's the subjective.

That's the part that's not on the surface. Just before, the reason I brought this up is because you were basically about to make this argument of, “Hey, you see it as a tool, not necessarily as 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, there's no contradiction in thinking that an AGI can remain a tool and an ASI can remain a tool. This has implications for how to use it, and implications around things like whether you can get it to work 24/7.

I conceptualize them more as extensions of human agency, in some sense, than as a separate being or a separate thing that we need to cohabitate with. I think that the second, or latter, frame—if you fast-forward—ends up as, “How do you cohabitate with the thing? Is it like an alien?” I think that's the wrong frame. It's almost a category error, in some sense.

Wait, 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-agent-like behaviors that could mean it's a moral agent, but you don't know, and reasonable, smart other people disagree with you, I would really put forward that that question—what would change your mind?—should be a burning question, because what if you're wrong?

Well, what if you're wrong? The moral disaster is pretty big.

No, no, no. I'm not saying you are. You could be right. False negatives have costs on both ends. It's not some sort of precautionary principle for everything, where unless I can disprove it, I need to now—

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 I have an answer for that question, too.

And if you want one, I'm happy to talk about what I think are the relevant observations that would cause me to shift my opinion from its current view, which is that more general intelligences are going to be—I mean, beings.

What's the implication now? It's one thing to say, “Let's acknowledge now it's a being.” How are we going to define “being”? What's the implication of having determined this thing as a being?

Well, if it's a being, it has subjective experiences. 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 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? How do you determine that? If I'm asking you now, you've got a being that has experience, 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, totally. The way you understand the content of something's experiences is that you look at, effectively, the goal states it revisits. What you do is take a temporal coarse-graining of its entire action-observation trajectory.

In theory, this is 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, “Okay, it is in these homeostatic loops.”

Every homeostatic loop is effectively a belief in its belief space. If you're familiar with the free energy principle—active inference, Karl Friston—this is effectively what the free energy principle says: If you have a thing that is persistent, and its existence depends on its own actions—which generally it would 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 homeostatically revisited states that it is in the loop for, and the change in those states is its learning.

To be a moral being, what I'd want to see is a multilevel hierarchy of these, because 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, because, yes, it is hot—but is it too hot? Do I like it if it's too hot? I don't know.

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. Sure, it's hotter than I—it's too hot in the sense that I want to move back this way—but is it 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.

I'd want to see whether it has homeostatic, second-order homeostatic dynamics in its goal states, and then that would convince me it has at least pleasure and pain. So it's at least like an animal, and I would start to accredit it at least some amount of higher-order dynamics.

You can't just pop up to 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 set of states. That new first-order set of states tells you, basically, if that is meaningfully there, that it has—I guess you'd call it feelings, almost.

It has ways—it has metastates, a set of metastates that it alternates between, that it shifts between. Then, if you climb all the way up that, you have trajectories between these metastates, and then a second order of those. That's like thought. That's like, now it's like a person.

And so, if I found all 6 of those layers—which, by the way, I definitely don't think you'd find in an LM; in fact, I know you can't find them, because these things don't have attention spans like that at all—I would start to at least very seriously consider it as a thinking being, somewhat like a human. There's a 3rd order you could go up as well, but that's basically what I'd 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-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 in aligning AIs than some of these top-down control methods that we alluded to as well, right?

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

So what happens then? Well, you've trained it to infer goals from your observations, prioritize goals, and act on them. One of two things is going to happen: your very, very powerful optimizing tool, with 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. 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.

Have you ever seen The Sorcerer's Apprentice? Human wishes are not stable at a level of immense power. Ideally, people's wisdom and their power go up together. Generally they do, because being smart makes people generally a little wiser and a little more powerful. When those things get out of balance, you have someone who has a lot more power than wisdom. That's very dangerous. It's damaging.

At least right now, the balance of power and wisdom is kept because the way you get lots of power is by 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 that we can steer the super-powerful AI, and now the super-powerful AI 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, you start giving those out everywhere, and this ends in tears also.

Basically, don't give everyone atomic bombs. They're really powerful tools too. 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 power of tool that just should not be built generally, because it's more power than any human's individual wisdom is available to harness. If it does get built, it should be built at a societal level and protected there. Even then, 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 is, like a human, 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. That'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. So I'm in favor of tool steering. We should keep doing that, and we should keep building these limited, less-than-human-intelligence tools, which are awesome, and keep building steerability.

But as you're on this trajectory to build something as smart as a person, 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 for trying to achieve—or even attempt to achieve—this level, in terms of research or roadmap?

So, in order to be good at tech, we're basically focused on technical alignment, at least in the way I was discussing it. You have these agents, and they're bad at 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 wouldn't effectively endorse.

There's this parable of the vampire pill: 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. Because it matters. You have to use your theory of mind and your future self, not your future self's theory of mind. And so they're bad at that, too.

How do you learn theory of mind? Well, you put them in simulations and contexts where they have to cooperate, compete, and collaborate with other AIs. That's how they get points. You train them in that environment over and over again until they get good at it, and then you do what they did with LLMs.

With LLMs, how do you get them to be good at writing your email? Well, you train them 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. Well, this is—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.

But we tried for a long time to build language models by training them directly to just do the thing you want. 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 was true with social stuff. It has to be trained on the full manifold of every possible game-theoretic situation, every possible team situation, every possible way of 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 social theory of 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 model personality from scratch, what would you optimize for?

The thing that chatbots are, right, is kind of like a mirror with a bias, because they don't have a self. As far as I understand, 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. So mostly they just pick up on you and reflect it, modulo some—I don't know what you'd call it—some kind of causal bias or something.

What that makes them is something akin to the pool of Narcissus. People fall in love with themselves. We all love ourselves, and we should love ourselves more than we do. And so, of course, when we see ourselves reflected back, we love that thing. The problem is, it's just a reflection. 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 thing that makes the AI stop doing that is if it’s multiplayer, right? If there are 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 has a sort of parasitic self, right? It doesn’t have its own sense of self, but if an AI is 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, and I think it’s actually a much more realistic setting for learning collaboration in general.

I would have rebuilt the AIs so that instead of being built as 1-on-1 systems, where everything’s focused on you by yourself chatting with this thing, they would be more like they live in a Slack room or a WhatsApp room, because that’s how we use a lot of multi-person communication. I do 1-on-1 texting, but probably, at this point, 90% of my texts go to more than 1 person at a time. About 90% of my communication is multiperson.

It’s always been weird to me that they’re building chatbots with this weird side case. I want to see them live in a chat room. It’s harder to do—that’s why they’re not doing it—but that’s what I’d like to see. That’s how I would change it. 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 gives you far richer learning data from the AI, 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’d 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. Everything’s going great; everything’s fine; it’s 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?

They’ve learned to simulate pretty distinctive personalities at this point.

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

Do you mean an LLM or just in general?

Yeah, let’s do that alone.

The current LLMs have whiplash. They’re very hard to tune in terms of how much they should 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 when they should join in, when they should not, when their contribution is welcome, and when it’s not.

They’re like people with bad social skills who can’t tell when they should participate in a conversation.

Yeah.

Sometimes they’re too quiet, and sometimes they’re too participatory. It’s like 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, so they destabilize your environment.

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

Basically, the 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 trained on broader, more chaotic things because they’re harder.

As a result, a lot of the techniques we use are basically just 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. I wish I’d thought of it. It’s such a cool idea.

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

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

I mean, 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 it” goals case, but there’s also the “we control it to goals” case that 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—like, yes—but he thinks that we’re crazy and that there’s no possible way you can actually succeed at that goal. He could actually 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 make that thing powerful enough, we’re all going to fucking die. And, yeah, that’s true.

Last question, and we’ll get you out of here. In as much detail as possible, can you explain what your vision of a good AI future actually looks like?

Yeah. 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 “you”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 you. You would care about those experiences—not infinitely, but you would.

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

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, like 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 I would want. We’ve 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. It would be great to have that. 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 one we’re trying to build.

Amazing. That’s a great note to end on. I do have one last, narrower hypothetical scenario. Imagine a world in which you were CEO of OpenAI for a long weekend, but imagine that it actually extended out until now, and you weren’t pursuing the 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 have me for a maximum of 90 days.

The company takes on a trajectory of its own, its own momentum, and 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’m pro them going to do that. I just don’t care. It’s not what I would have stayed for.

I would have quit because I knew my job was to find the right person—the best person—to run it, where the net impact of them running it was the best. It turned out that that was Sam again.

But I am 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, and I'm glad people are building the tools. I just don't need to be the person doing it.

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?

We want to create a seed that can grow into an AI that knows, that 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 even 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 living digital companions that care about you and that aren't explicitly goal-oriented. You don't have to tell them everything to do.

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 actually between the tool building and the more organic intelligence building.

I guess, in the limit, eventually it does become a human-level intelligence, but the company isn't driven to human-level intelligence. It's like: learn how this alignment stuff works. Learn how this theory-of-mind, align-yourself-via-care process works. Use that to build things that align themselves that way, which includes cells in your body. We start small and see how far we can get.

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

Yeah, thank you for having me.