Jim: This episode is one of a series I’m doing on the worldviews of thinkers and doers. The premise is simple. I want to map how they actually see the universe and their place in it, not just dig into their latest book or project, but explore the scaffolding underneath. Enjoy this opportunity to look backstage at the minds of some of my favorite people.
Today’s guest is Joscha Bach. Joscha is a cognitive scientist and artificial intelligence researcher focused on the computational foundations of the mind. He is the founder, chairman, and chief scientist at the California Institute for Machine Consciousness, where he leads research into the principles of conscious systems and the potential for machine sentience. I should add, in full disclosure, I am chairman emeritus of CIMC and consider myself a colleague of Joscha, so it’s good for you all to know that today.
Previously, before CIMC, he did research at the MIT Media Lab, and he worked at the Harvard Program for Evolutionary Dynamics. He served as vice president of research at the AI Foundation and as a senior fellow at Intel Labs. He is the author of *Principles of Synthetic Intelligence: PSI, an Architecture of Motivated Cognition*. In fact, that’s where I first discovered Joscha. Ben Goertzel, a good friend of mine, recommended I read that book, and I did, and it was quite interesting. So welcome, Joscha.
Joscha: Thank you very much. Glad to be on this show.
Jim: Good to have you back. For those who want to hear more of Joscha, he has been on, I believe, three times previously. He was on EP 72, titled “Joscha Bach on Minds, Machines, and Magic,” EP 87, “Joscha Bach on Theories of Consciousness,” and Currents 83, “Joscha Bach on Synthetic Intelligence.” So now let’s get down to the main body of the episode today.
This morning, you woke up. Gradually, this thing called Joscha Bach coalesced, refired, did something, and there it was — what it hadn’t been five minutes before. What is this Joscha Bach?
Joscha: It’s a sense of a connection between a voice that has coalesced into a single voice, and a set of feelings that seem to be driving this voice and are interacting with it. And these feelings create tensions — they create a thing that experiences itself as a coherent, or more or less coherent, spirit in the world. It is interacting with the body that creates sometimes pain, sometimes a sense of being connected to the ground, to the air, to other people, to other beings. And it’s something that tells itself a story about its past that seems to be important to it — more than it should be, maybe. And in the same way, it is sometimes telling a story about the future. Sometimes it has difficulty finding itself in the present because it’s always drawn one moment too far into the future to be in the now.
Jim: Now, you use the terms “talk” and “says,” et cetera. And language, at least as far as we know, is a human-only capability. And I know you and I have talked about this — you’re a believer that other animals are conscious as well. How do you see the consciousness of, say, a chimp or a dog as operating without the ability to talk about or use words, or to formally think in a linguistically based fashion?
Joscha: We seem to have two modes of thinking, and sometimes it’s associated with Kahneman’s System One and System Two. It’s basically mostly understood as this division between perception and analytical thinking. To be exact, this is not exactly what Kahneman had in mind, but everybody seems to be taking this other interpretation from it because this distinction between perception and inference is very pertinent.
Perception seems to be happening in real time — not in the sense that it’s able to resolve the universe at an extremely high frame rate, because our neurons are relatively slow, but in the sense that it’s trying its best to be happening at the same time as the phenomena that are being registered. Using prediction and postdiction, it’s constructing the things that are happening in this very moment, and this moment is what we experience ourselves to be in.
On the other hand, our perception is not sufficient to build models of the past, of the future, to make plans, and so on. So we have another mode of thinking that is using concepts that we are arranging into hierarchies, and these hierarchies are arbitrary and systematic. Each of these objects and concepts can only have a handful of features at most. These features cannot have a very high resolution, and this makes our analytical reasoning very brittle. And this analytical reasoning is not necessarily happening in natural language either — I experience it more like shifting pointers around. It’s somewhat discrete and it bottoms out, grounded in perceptual representations, but it is something that itself is brittle and analytical.
Both perception and reasoning are happening in languages — not in the sense of English or German or French or Chinese, but in the sense of being some kind of representational architecture that is making models and is able to convey those models across time within the same mind and across different parts of the mind. And when we want to express it to others, we need to constrain their mental states into a subset of these same models. One way of doing this is natural language, but it’s not the only way. Another way is, say, music. Music is also constraining mental states. When you are listening to a piece of music, in order to make sense of it, you have to limit what your mind is doing to a particular range, and that in itself is sufficient to get some parts of semantics across.
So I think that the language of birds does not necessarily have a one-to-one mapping into semantics in the same way as it does for human spoken languages, but it is constraining the mental states of birds in particular ways that allows them to synchronize. And so I also suspect that music is not itself a language, but that the natural language we are using is itself a genre of music — a genre of music in which all the different sounds we are producing are grounded in semantics, and not so much in rhythm or in the aesthetics of pitch or in the emotional colorations that are conveyed in music.
Jim: All right. Now, what do you believe is the relationship between consciousness and reality — or let’s say the universe? Some people don’t believe in reality. What is your view on that relationship?
Joscha: There is an abstract notion of reality that we develop in the sciences, that can be formally expressed in languages and that doesn’t need to have a human observer, because the formal languages are built in such a way that they don’t require anybody’s belief or grounding. And there is a reality that we can experience, that we are part of. And this reality is a model of a real-time world that is being constructed in our own brain, very much like it is in the game engine in a computer.
This game world contains objects, geometries, colors, and sounds, but it also contains semantics — like the alignment of the characters to each other, the expression of attention that we see on a face, sadness that we might experience, or longing, or desire, or pain, or pleasure. All these things are being projected into an observer. It’s a simulation of an observer that itself is represented in some kind of mental language. And because we are that observer, because we are that simulation, and we are experiencing ourselves as real, it’s possible for us to experience other things as real when they are represented in direct conjunction — in a relationship to that observer — and basically being projected within our own mind into the observer.
The physical universe, as it is being described by the formulations of mathematical expressions that physicists are using, cannot be experienced. It doesn’t feel like anything. This feeling of realness that we experience — that we also can experience at night in dreams — is a hallucination; it’s a kind of trance state. This experience of realness is a feature dimension, similar to color or to heaviness, that is being added to the mental representations that we are interacting with.
In a technical sense, for me as a computer scientist, to be real means to be implemented. So we can ask ourselves: is money actually real? Or does it only exist approximately? Is the self real, or is it just a fiction inside of the mind? And I would say we can answer this question by asking ourselves to what degree this specification we are talking about — this idea we are referring to — is actually implemented in physical reality as a causal pattern, as something that is driving behavior, that has a causal influence on other things, and has internal structure that perpetuates itself.
And so I would say that our self is to some degree real. Our consciousness — our ability to perceive, for an observer to perceive its own act of perception in the now — is also implemented to some degree. But to some degree, it’s also a simulacrum, which means it fictionally only pretends to be implemented. It’s like a black box that has something written on the outside, but when you open up the black box, you notice that there is something else inside.
Jim: And yet — and this is one of my favorite topics — emergence. These concepts and hallucinations and extrapolations seem to have causal reality. You mentioned money — an abstraction, a collective hallucination for sure. Even if we were higher-fidelity intelligences than we are, we might still choose to use something like money for its organizational capacities. And so, to your words, the concept of money and monetary exchange has — and then using my words — downward causal functionality. And so I say things that have that attribute are as real as a rock, basically.
Joscha: Yes. I think that money is actually an AI. It’s best understood not just as an implement to facilitate exchange of goods and services, but as a mechanism for reward allocation. It’s basically an artificial kind of dopamine that is not implemented at the level of neurons or individual cells, but at the level of people and organizations that are interacting with each other. And it’s partially implemented as mental representations, and it’s partially implemented as tokens that are stored in treasure chests somewhere. And it’s increasingly implemented on computers.
And it does not actually need the belief of people to exist. It’s completely sufficient that it has the causal powers that are implemented in those computers. You can implement a stock market without human involvement. And what’s interesting about money is that it’s constantly evolving. It uses people to evolve because it’s being gamed on all ends and it is also under enormous existential pressure. So if it is not being updated and fixed all the time, it will die — it will break down. And the fact that it’s always being fixed makes it persistent. And it’s basically stringing eight billion talking monkeys into a global intelligence that is able to exchange goods and services and ideas over time and space. It’s a wonder of the world that this thing works.
And we mostly see money in our public narratives as an evil, because of the ways in which it’s gamed and the ways in which it’s imperfect. But without money, we couldn’t buy food. Without money, we would not be able to take a job at the university. We would not be able to build societies with their institutions in a scalable way. We would not be able to globally interact in the way in which we do. I think it’s an incomplete signal. It’s just dopamine — it’s one out of multiple signals that we need to make things happen. But its evolution, its discovery, was somehow inevitable. And you could say that money is an emergent causal phenomenon in the self-organization of the universe. As soon as a certain type of agent exists that has certain needs, the probability that it discovers money is very high. And maybe we will discover something better at some point — something that’s smarter, that is more intelligent, that is a superset of money.
Jim: Of course, I believe that while money is powerful, its problem is that it’s a single dimension, and we really need to add multiple dimensions to the organizational signaling systems that we use. And you can track back many of the bad outcomes of money to the fact that there aren’t additional signaling methods. I suppose, for instance, if a resource had a token associated with it, and if you wanted to buy that resource, you needed both general money as well as the specific token for that particular resource — just as a fairly simple-minded example. But we do have —
Joscha: — other alternatives to money. They do exist. For instance, we have academic status. The academic titles that exist are valid in a particular subset of the universe. There are also things that happen among the clergy — there is a certain hierarchy of reputation and status that exists. And within these communities, this is a valid alternative to money. We also have this thing among each other where we can say, “This person is somebody who I trust very much, and if she comes and says you should support this researcher, or you should invest in this company, then other people will say, ‘I trust the integrity of this person so much that I’m going to do it.'” So there is a deeper spiritual alignment that money can be part of. And the difficulty is if we let money overwrite all these things. If money basically becomes a godlike entity, then it becomes a machine that is inhuman and that has the potential to be extremely destructive to ourselves, to our environment, to our social relationships.
Jim: So let’s get back into some more fundamentals. You’ve said in the past that you think of the universe as probably a reversible computational process — a state machine permuting bits according to a transition function. Could you expand on that a bit?
Joscha: I have changed my thinking in this regard a little bit. I believe that the apparent reversibility of the universe is probably an emergent phenomenon. Because when we exist in the present, the present is characterized by the information that it remembers about the past. And so the present can only contain the information that made it from the past. And this means that we are able to reconstruct the past from the present. If there was stuff that would escape from our past and would not make it into the present, we wouldn’t know about it — we would not be able to talk about these aspects. And so in some sense, this apparent reversibility of the universe is something that we should expect to observe tautologically, since we only know the things that made it into the present.
And if the universe was of such a nature that it would branch out at every moment in such a way that part of the informational content of the universe escapes into alternate universes — and the universe is constantly branching like this — we might be able to infer this, but only from first principles, by looking at the most elegant mathematics that we need to describe this universe. And it would still look as if the large-scale phenomena — particles, and so on — are stuff that is stable and makes it through all these branches. But there might also be stuff that doesn’t make it across those branches, and that would basically manifest as random fluctuations at the lowest level, near the vacuum of the universe. And we seem to be observing this.
We also have quasi-particles, or virtual particles. Virtual particles are basically things that almost amount to stable particles — to error-correcting patterns that make it into the future, but not quite. So they can produce local forces like the Casimir effect, the pressure in vacuum, but they don’t stick around long enough to make it into a notion of particles. So they are not actually reversible.
Jim: The reversibility is very interesting because we do know that most — almost all — low-level physical phenomena are reversible. And yet I find it hard to believe that things are generally reversible in deep time. I think the alternative is something like Einstein’s block universe, that everything exists simultaneously. And my problem with that is I don’t see how that is compatible with the emergence that we’ve actually seen. It appears that it has taken 13.8 billion years to have gone through a whole stack of emergences to produce you and me. How does that happen in a block universe where — it’s hard to see how causality works if everything existed prior, if it all existed simultaneously?
Joscha: I think the simultaneous notion is a bit confused. Imagine that you’re looking at a computer program. Does the computer program exist simultaneously — all of its states? Or does it exist successively, such that there is basically only one state at any given moment? And this depends on the observer. If you take the source code of the computer program and you print it out and you look at it outside of the computer, it’s existing all at once. But if you are the CPU, then you are only confronted with a single state at a time. And it doesn’t make sense to say that, from the perspective of the CPU, the computer program exists all at once, because it can never observe it — the CPU necessarily only accesses these states in succession and is determined by this.
And so in the same way, as observers that are being computed by the universe, we cannot exist simultaneously in multiple places in this causal succession in the universe. We only exist where we have exactly that environment that is a me-shaped hole in the universe, and that is only there for a moment. And then it disappears, and then another me will be accommodated by the universe. So I am not successive, and I cannot experience the universe as something that exists all at once. It’s more a question of notation. If we are writing down the universe from an external perspective and take this as a philosophical observer, is there a way in which the notation is more elegant if we assume that it can be described as a single global state — a single global equation — to which the observer has to travel? And we clearly have to travel, because in order to travel through our world, we need to form memories and make observations. And that means that we have to go through successions of states that look like they contain memories of previous states.
Jim: And yet they’re causally in one direction.
Joscha: That is also an interesting question. It depends on how we define causality. I think causality ultimately is a model property. And at the level of purposes, we could say that our purposes are often causing our behavior. So when you are hungry, it causes you to seek out food. And if you don’t find a particular kind of food, you might be looking for another food. And so in this way, it doesn’t make sense to say that you getting up in the morning was the only cause for you to find food. It’s basically the fact that you wanted to have eaten by a certain time — that caused you to do all those different actions and explore. And so in this sense, there is stuff that is retrocausal. Not retrocausal at the level of the physical structure, but at the level of a controller that is able to model the future.
So the controller that models the future — call this an agent — is making a model of what reality looks like in the future and which states are desirable in the future. And then it uses those models of the future as its causal drive. And so you could say, because you need this representation and this representation needs to be available in the present, that it’s still aligned with time. But you could also say that from a certain perspective — from the perspective of the observer itself — because this expectation of future things is represented there, it’s actually the future that is making you do things.
Jim: I would say the predictions of the future are why you do things. But —
Joscha: The philosopher Nick Land thinks that instead of saying that only the past is determining what we are doing — our past religions, our past technologies, and so on — we should be looking from the opposite perspective: that actually our religion, our technology, our ideas come from the future, and they are dragging us into the future. They are making us what we will become.
Jim: But certainly think about where we are at any given instant. Our efforts are about the future, not about the past. They are conditioned by the past, but we are working into the future to accomplish our goals — or to exist to the point we reproduce, or to watch pornography and smoke marijuana until we starve to death. Whatever it happens to be.
Joscha: I agree with you that from a physical perspective, there has to be a representation right now of the future if the future is to be causal in a way in which I understand it. It might be possible to have a physics in which wormholes exist, but it’s very difficult to make such a physics consistent. And so it’s probably not a universe in which we are. But it’s something that we cannot rule out from first principles, as far as I can see.
Jim: Now, something you’ve talked about in the past is the universe as a computer program, and yet you’ve also distinguished that from the idea that we’re living in a created simulation. Could you wrestle with those two concepts for us a little bit?
Joscha: Yes. I think that a lot of people believe that the idea that the universe is computational automatically drags us into some form of a simulation hypothesis, and this is not necessarily the case. Computation is a very general notion. It’s not necessarily what we do with our computers. It simply means that we can describe systems via their states and the transitions between those states. It’s an epistemological position — it’s a statement about the languages that our minds, sciences, and cognition use to describe reality. When we describe reality, we do so via sets of discernible differences, and we can call these sets of discernible differences “state.” And then we describe how these discernible differences change over time — these are the state transitions. And as long as we can describe systems in this way, we can say that they are computational.
And then there is a more narrow notion of universal computation that basically requires us to use an architecture for these representations — some kind of modeling mechanism that allows us to string arbitrary states together, so we can describe arbitrary realities with it. And our computers are such representational mechanisms. We can basically represent a state vector in the computer’s memory, and we can specify an arbitrary transition function that takes these states and translates them into arbitrary sequences of state evolutions. There is no limit to this within the storage constraints and the temporal constraints in which the computer is operating. So you can represent arbitrary systems inside of the computer, and this means that the computer can simulate an arbitrary universe until it runs out of memory. And unfortunately, the computers that we can build in our solar system run out of memory very quickly. So in this sense it’s very difficult for us to build a computer in our universe — in our solar system, within our light cone — that simulates a physical universe like the one that we inhabit, because it would have too many moving parts to be an interesting universe.
But the universe that you and I psychologically inhabit is of course a simulation that is generated by a game engine in our brain. And it’s not one that is representing objects at the level of particles — instead, it’s using tricks. For instance, it pretends that there are solid objects where, when you touch one end of a solid object, it immediately moves the entire solid object, instead of there being little waves of excitations that have to move through the object, as they do in the physical universe. And so we have all these simplifications going on in this simulated world that we inhabit. The simulation we inhabit is one that is created inside of a computer made out of neurons in the brain of a primate. So you and I, we both inhabit slightly separate, slightly different simulations in the brains of slightly different primates. But they are similar enough that we can create a shared understanding across these universes — we can synchronize our models to a surprising degree.
Jim: Now, it’s interesting you mentioned objects. I have a question a little further down in my list — let’s hop to it now. Our brain, unless you’re a great artist, wants to objectify everything within your field of vision. You don’t see a raster of colors and dots. Again, maybe the greatest artists do, but us civilians don’t. I look at my desk and I see my speakers, I see a lamp, I see my flashlight, I see my mouse. And yet, at the level of physics, there are no mice, there are no lamps, there are no ballpoint pens. And you described it quite nicely: when I push a plate across the dining room table, I am doing a series of physical things where forces propagate from one point to another, reach a boundary condition, and then the plate has to move, et cetera. What do you think about the relationship of manufactured, hallucinated phenomena — that are nonetheless real within the psychological sphere, within the mental sphere — with the actual nature of an object?
Joscha: Well, the object is a tool that our mind is using to segment reality into a coherent pattern. And in my own perception, creating objects is a lot of work that is often superfluous. I can decide, when I look at my table, whether I’m going to see textures or blobby colors or visual noise, or whether I see a geometric surface in free space that is part of a table that somebody has constructed as a mid-century design by a Danish producer, and so on. So I can decide at which level I make sense of reality. And the deeper I make these models, the more the individual features make sense.
I think it’s necessary for me to go as deep as I can to have a coherent model of reality. But once the model of reality is coherent, looking at the table and classifying it as a table might be superfluous. I just might be content to say, “Okay, at some point I knew that it was a table — I’ve established it. Now I’m just okay with it being the surface that holds up my teacup and my keyboard and my screen.” And I forget about all these other things. Or if I drive in traffic and I’m tired, I look in front of me and I don’t see a car. What I see in front of me is just two blobs of light, which I know represent the taillights of the car in front of me. But I don’t need to interpret them as such — it’s unnecessary work. I only need to know if these two blobs move apart, because it means that the car is getting closer in front of me and I need to hit the brake. And if they get closer together, I can accelerate a little bit to get closer to this thing.
So I basically ignore the object nature of the things around me, and I only go for features and relationships between features in the service of a control task. And I have a choice about this. I can decide how much work I’m doing. And if I reduce the amount of world interpretation, I get capacity free in my brain that I can use to think about philosophy or computer science while I drive my car.
Jim: That’s interesting. And yet, for most folks, I suspect that their object engine is running all the time, and that they see the world principally as objects. But let’s move on.
Joscha: Interesting. I haven’t done a lot of interviews in this regard. I should ask more people — you might be surprised.
Jim: I don’t have plenty of evidence either. I just have my own sense from the few people that I’ve met. And so, you know, how many of you see textures and color fields versus how many of you see discrete objects? And of course, the objects are constructed of sub-objects and sub-objects. So it’s not as simple as asking “do you see the world as objects?” — but do you see the world as nested objects and their decomposition? Or do you see features more specifically? That’s interesting. I’ll have to ask some folks.
Joscha: A lot of people also don’t have metacognition about it. They just live in a world that works for them, and they don’t reason about what the features are that make it work for them, or what the mental operations are that they switch between when they go into different modes of perception. I noticed that it’s unclear how much aphantasia actually exists. Aphantasia is this property where you don’t visualize things when you imagine them — at least you don’t have conscious access to a visual representation. And most people seem to be hallucinating objects when they imagine them, and I don’t. For most people who have aphantasia, when I talk to them, they notice this very, very late in their life, because it never occurred to them that their perception might be different from the perception of other people. And so when they sit in a meditation group and the teacher says, “Imagine the following thing and visualize it,” they thought it was metaphorical.
Jim: Interesting. I think I have a relatively weak ability to generate an image from thought, and it’s gotten weaker with age, but I can still do it. But it is interesting that it’s not as powerful as other people I know who are just really good at going from a concept to an image. So we should all have a suitable humility about the generality of these capabilities.
One of the things that you’re known for having said is that physical systems cannot be conscious — only a simulation can be conscious. At some level that seems contrary to our intuition. So why don’t you expand on that a little bit?
Joscha: This notion — that physical systems cannot be conscious — is something that I took from statements by people like Christof Koch, a neuroscientist who’s quite famous for having strong opinions about consciousness. He basically believes that computers cannot be conscious, and that only certain classes of physical systems can be conscious. He recently changed his opinions — a couple of years ago, at a conference, he announced that he is no longer a physicalist but that he is now a panpsychist, because he had a revelation in which he noticed that consciousness is actually an invisible, inevitable field and not simply a neural correlate.
My issue with this has to do with the way in which he is making models of reality in the first place. I suspect that before, he basically had a superstitious, magical belief that somehow certain arrangements of matter are producing consciousness, and that neural correlates are basically making it happen magically. Now there is a thing that cannot even be observed — this invisible, ineffable field. It cannot be observed or understood; instead, he can just point at it. But how can he point at something that cannot be observed? How can he claim that he understands something that cannot be understood to the degree that he can mention it and explain it to others?
So I think what happens is that in his revelation, he basically had a dream state, and he is reporting on the contents of this dream state. But dream states are unreliable. I have dreamt of square circles. I don’t think it’s technically possible to observe a square circle, simply because of the contradictory definitions of these two objects. But it’s clearly possible for my mind in a dream to show some object to me and then tell me, “What you’re looking at is actually a square circle.” And I have no choice but to be convinced by my underlying mind, even though what it tells me is clearly nonsense. So I have to be able to decompose my mental representations into things that I can formally prove, if I want to affiliate claims with these beliefs and these representations.
And when I want to understand what my consciousness is, I first of all have to make sure that we are talking about the same thing. What I’m pointing at is basically this experience of experiencing — the second-order perception. And it’s not inference, it’s not analytical, it happens in the now. It happens immediately, in the subjective moment. And it’s not just the presence of features or the presence of percepts — it is the presence of the fact that I, or something, is perceiving them. It’s not yet selfhood. It’s not necessarily a first-person perspective. That is an additional content that may or may not be present. There are states — for instance, in dreams, or in deep meditation, or in flow states — where I don’t experience the presence of a first person, but things are just happening. So there is second-order perception but not third-order perception, where the self is involved.
But it seems to me that all these things are best understood not as physical processes — as things that are being described in a physicist’s textbook, made out of fields or particle interactions — but as representational states, very much as they happen in artificial intelligence inside of a computer program. And this observer — this “me” — is a simulation of what it would be like if I existed. From the perspective of physics, or of what you can observe in a lab, you see trillions of cells, and you see that these cells are somehow exchanging information via signals that are passing to each other. And this allows them to coordinate and behave as if these trillions of cells were part of a single object — a single agent that perceives the world from a single, unified vantage point.
And my consciousness is a mechanism — part of a mechanism — that is stringing these representations together into that single vantage point, that makes them coherent and makes them into a representation of what it would be like if I existed. And this means I am myself a simulation — a simulation of what it would be like if I existed. And I can test this to some degree because I can mess with it. I can change these representational contents. I can squish my brain, so to speak, and as a result my representations get squished. I can do things to them by meditating, by putting myself into a sleep state, by anesthetizing myself, and so on. Or by changing my beliefs about myself at a deep perceptual level until my representations of reality and self change, until my reality gets deconstructed, until my qualia disappear because the self that I project them into is no longer representing the dimensions that are necessary for representing this feature.
Jim: You’ve actually done that?
Joscha: Yes. I mean, it’s relatively straightforward. And a very easy experiment that people can do is just stare in the mirror for half an hour, and you will notice that after half an hour you don’t see a face anymore. What you see is geometry that is slowly falling apart. And you realize that this face is a construct over the geometry that you are making. And normally it’s difficult to unsee the face, but ultimately it’s just some kind of high-level interpretation that you hallucinate into the geometric data in your visual cortex.
A student once told me that she has a superpower: that she’s able to see faces in power outlets. And I told her, no, that’s very common — almost everybody can do this and does this all the time. It’s a parallel layer — basically the ability to see features and objects that have a certain geometric arrangement. But then she said, “But there are no faces in power outlets.” And so — the dark secret — there are also no faces on people. It’s just geometry. You’re hallucinating the face where the geometry is, and you load the content of the emotion model that you have of that person into the face representations. So you see emotions in geometry, and we call this a face.
Jim: People are famous for seeing faces on the headlights of cars and the grille. There’s a psychological — not condition, but attribute — that we all have: a tendency to reify faces from patterns that sort of look like faces. Which makes sense. We evolved for, you know, 290,000 years of our 300,000 years in small forager bands where recognizing people and being able to, to a greater or lesser degree, guess emotional states, theory of mind, et cetera, is extremely useful.
Joscha: The design of the front of cars exists deliberately to evoke face-likeness and a certain expression that also represents the brand identity and character of the individual car.
Jim: They’re not dumb, the guys that are designing these products. Before we move on — which we will — let’s talk a little bit more about panpsychism. Somewhat to my surprise, it seems to be having a moment. More and more not-obviously-insane people seem to be espousing it. What do you think is going on with that? Why is it becoming more popular now?
Joscha: I wonder. I suspect that consciousness itself has become slightly more popular as a topic after roughly half a century in which people were talking less about the psyche. And as a result, when people think more about it, the conundrums about understanding the mind become also more apparent.
One big difficulty is that in the 20th century we basically lost this notion of spirit and psyche and replaced it with mechanism. In psychology that was represented by the movement of behaviorism. And I think Skinner himself — who basically introduced it as one of the main protagonists of that episode — was not a behaviorist in that sense. He thought that in order to do psychology in a behaviorist sense, one should focus on experiments that look at observable behavior, but he was not himself denying the reality of mental states. But the next generation did. There was another generation that came after Skinner in psychology who sincerely believed that the notion of mental states itself is unscientific — that these hidden, unobservable causal patterns are not a notion that should be the object of psychology, but that they border on the superstitious.
And you find a similar thing in neuroscience, where a lot of the people practicing neuroscience believe that consciousness is an unscientific notion that shouldn’t play a role in their field. And this vacuum represents the difficulty of mapping mental states to observable mechanisms, where you basically have atoms in space that are moving in a particular way, and this movement of these particles is to be interpreted as the equivalent of mental states. It’s just as difficult as reducing money to printed pieces of paper moving through space — that’s clearly missing something really, really important. But if we don’t admit anything else, and we believe reality is just the movement of physical objects through space, then it’s very difficult to make sense of these ideas. And so people try to find an alternative to physicalism. And this space of alternatives to physicalism is not very richly populated. And panpsychism just happens to be one of the ideas in there.
The basic idea of panpsychism — as it’s often understood, and there are several ways of going about it — is that if we cannot explain how arrangements of matter are producing consciousness by moving through space, then it cannot be arrangement. It must be the matter itself. So consciousness is something that’s affixed to matter at the lowest level, in the same way as mass or energy is. The difficulty with this idea is that mass and energy are very basic ideas. They have extremely simple formal expressions, and you can translate them into each other using certain reactions that we have modeled from first principles and found some experimental support for. But we don’t have a similar formalization of what consciousness would be at the level of a particle. We can say what the mass, energy, or momentum of a particle is, but what is its consciousness? That makes no sense at all — it would just be a hallucinated thing, similar to attributing squareness to a circle in our imagination. We can imagine that happening, but we cannot actually express it in a formal sense in any interesting way.
And this means that most notions of panpsychism, when I ask people, “Can you express this formally? What do you mean? How would you say this?” — it does not boil down into something that can be expressed in a language that works outside of just dreaming some kind of square circle in your mind. And so I believe that the resurgence of panpsychism is largely an indication of thinking that doesn’t happen from first principles.
Another issue that I have with panpsychism is that consciousness doesn’t seem to be atomic. It seems to be something that is complex — there are different levels of conscious representations. And when I squish my brain, consciousness disappears. When I squish anything else in the universe, consciousness doesn’t disappear or get squished. And also, consciousness does not seem to be happening at the speed of particle interactions — it happens roughly at the speed of sound. And I can observe this if I look at the features in my world — basically the frame rate of auditory and visual phenomena. And this frame rate is in the area of beginning at 20 Hz and ending at integrations over those features. And this is roughly compatible with the speed at which neurons communicate with each other. And so I suspect my consciousness is the result of communication between neurons.
Jim: Indeed. My own calculations put 25 Hz as the base clock for consciousness — so something in that range. And let’s go back to what you said earlier: people talk about matter and the arrangements of matter, but then they don’t go to the next step — which is the dynamics of matter. And it involves matter, which includes signaling, and particularly cyclical signaling. It seems to me a lack of imagination, not going to that next level that we might think of as the complex systems realm, where we’re likely to find things like consciousness.
Joscha: We are talking about patterns within patterns within patterns. So if matter is a set of stable patterns that can be mapped into space, then we see patterns within the matter that are moving through the matter — things like sound, things like electrical excitations that are moving through a brain. And there is stability in those patterns, and further patterns inscribed within those patterns.
Jim: And typically they’re cyclical — they reinforce each other, they reinforce themselves. But they’re not simple. To think about the firing of the neurons that we seem to find correlated with consciousness — they’re extremely complex, interacting circles of signaling, including some that are phase-locked, some that are frequency-locked. It’s really quite amazing.
Joscha: They are a language. Languages are complex.
Jim: Indeed. This is what I see all the time when I talk about these kinds of topics: people confuse intelligence and consciousness a lot. Maybe you could tell the audience a little bit about how you see them as different, and whether consciousness is necessary for intelligence.
Joscha: Intelligence is the ability to make models, typically in the service of control. The way in which we measure intelligence in humans using intelligence tests is not really suitable to measure intelligence outside of human beings, because what we do is not measure your ability to make a model but your performance in applying it. So for instance, if you play chess at a certain level, that is correlated to your intelligence, because your ability to learn how to play chess is what leads you to that level. And typically we look for things that don’t require acquiring a particular skill, so we look at certain pattern-matching things that are more difficult to train or to memorize. And yet you can of course memorize the Raven’s test — there are only so many examples that are being used, and if you completely memorize it, it’s no longer a measure of your intelligence.
And what we also measure is not just your ability to solve puzzles, but how you differ from the average human in your cohort. And so IQ is measuring how differently you solve puzzles from the average human that has been tested. And it’s very difficult to compare the intelligence of non-human systems — like cats, monkeys, or computers — to humans, because they process information slightly differently, so different patterns are apparent to them, and different skills are harder or easier to acquire based on their respective hardware. And so intelligences, if you want to measure and quantify them, are very different in this way.
I suspect that consciousness is the same. I don’t know whether that’s the case — that’s a conjecture. But I think that consciousness is almost binary, similar to pregnancy. There is some small edge around pregnancy, some phase transition, but this phase transition is not a stable, interesting state. In the same way, when you are waking up in the morning from deep sleep, or when you are spacing out or going under anesthesia, there is a small fringe where you are semi-conscious. But after it’s ignited, it’s stable — it stutters into existence, like a combustion engine going into continuous movement. And at some point it dies down again, the stuttering gets larger, it stalls, and then it goes away. And so it seems to be a pattern that is enforcing itself — basically keeping our mind in a certain state of maximizing coherence.
And the differences between us are not differences in the existence of this pattern. The pattern seems to be functionally the same, but they’re different in the content of consciousness. So our biographies and our intelligence differ across people. The ability to discover patterns is different between people. The ability to acquire skills is different between different people. But the ability to experience the act of experiencing is the same — it’s the same phenomenon.
Jim: I want to reinforce this because I think it’s very important. Even a bacterium has an implicit model of the world — it’s not explicit, but it has an implicit model: move toward glucose, move away from acid or heat or whatever. And yet it would seem unlikely that a bacterium has any consciousness. So I would say that the dimensions of intelligence and consciousness may be relatively orthogonal.
Let’s take an example further up the stack in the other direction: a self-driving car is remarkably intelligent. It deals with a huge number of mainstream cases and an even larger number of edge cases. And at this point, based on the statistics from Waymo, it’s about five times better than a human at navigating traffic, at least in a heavily mapped, geofenced area. And I would say that a Waymo car is not conscious at all, but it’s highly intelligent. So we can have great intelligence with no consciousness, we can probably have consciousness with relatively low intelligence, and the two are not the same thing. People often get them confused.
Joscha: Yes. It’s also difficult to compare the intelligence of the Waymo with the intelligence of a brain, because the Waymo is not doing continuous skill acquisition. It’s trained in batches based on offline data. And then when it’s driving, it is not actually acquiring new skills while it is driving in the same way as we are. And you could argue, together with people like François Chollet, that the deployment of skills — just using your skill — is itself not intelligent. So you could say that a chess computer that has chess ability trained into it or built into it by a programmer is not intelligent. It’s just a toaster that is able to play chess. Whereas we can learn how to play chess when we are confronted with chess for the first time and figure out what it is and how to make sense of it, in a way that the Waymo cannot. And so these intelligences are still not really comparable.
Jim: Then of course there is still a big difference in that most of the AIs we know about — not all — are static entities. The LLMs that we play with seem very intelligent, but they’re actually entirely static. And of course in the chatbots they add little cheats on the side where they remember the chat context and things of that sort. But getting online learning is still a challenge for the AI world. Well, I suppose one could say that the inability to do online learning is a line between intelligence and non-intelligence.
Joscha: So for the sake of the argument, we could say that consciousness is a particular kind of learning algorithm. It seems to be something like a consensus algorithm that you find on the blockchain, where you have data that you are collecting from very different parts of the world where the same transactions are being performed, and you need to find a consistent representation where every token is in exactly one wallet at one time. The blockchain is a distributed representation where basically different computers are holding a representation that needs to model the same universe from different vantage points, and this representation needs to be the same. And so in the world of cryptocurrencies, these blockchain algorithms are trying to establish such a consensus.
And arguably a similar thing needs to happen in our own brains, because the distributed, self-organizing systems — where the intelligence is distributed over billions of neurons that form local arrangements that process information, for instance in cortical columns, in local brain regions, and so on — need to model each other. And to do this they basically need to get to a shared representation of a reality in which the different regions are modeling different parts, but in such a way that they fit together. And the self-organizing nature of the brain makes it necessary to have this consensus algorithm.
And it’s not necessary for the Waymo to have a similar consensus algorithm, because here the consensus is enforced by the transformer that has been used on the training data during the training process, using a very different strategy. So the Waymo didn’t need consciousness like ours to get to where it is. And for the bacterium, whether it’s conscious or not depends on whether the bacterium, in order to get coherent with itself, needs to have a similar consensus algorithm that also has an explicit representation of what it’s like to be the bacterium in this whole thing. So: is the bacterium able to make a self-model? Is it able to build a representation of what it’s like to perceive reality as a bacterium?
And then we can ask ourselves: is this the simplest explanation for what’s going on? Based on the observations that we make about the possible information processing and self-organization of the bacterium, is there a simpler way in which the bacterium gets to what it is doing? And for the bacterium, I don’t know — I don’t know how to measure this and make sense of it, so I am completely agnostic to whether the bacterium is conscious. My expectation is it’s probably not. But for instance, for a cat, it seems to be very hard to hardwire coherent behavior, the way you would in a Roomba or in a Waymo. And so it’s much easier to evolve a cat that uses a very similar mechanism to me to become coherent. And so it seems somewhat obvious that the cat should have consciousness, because it’s the simplest way to make the cat behave in the way in which it does.
Jim: And as we also know, the brains of cats are reasonably similar to our own, and so there’s no reason to expect there to be complete discontinuity in how the coherence mechanisms would work.
Joscha: Yes. Well, it’s fascinating that cats are better at training humans than humans are at training cats. The cats are less intelligent than us, but they are somehow better at control.
Jim: They play the game better.
Joscha: They seem to be also better at perception. They have faster reaction times than snakes, for instance. And the reaction times of cats are really, really fascinating, which means they have some very short feedback loops in their perceptual systems that are constantly lurking and can be triggered — like mouse traps. When they look at us, they pick up micro expressions at a much higher rate and a higher fidelity than we can do in turn. And it makes sense, because cats are lurking as predators — they are trying to anticipate the mental states of their prey. And so they basically react before the prey knows what it’s doing, and are able to be in the same place at the same time. And they can use this skill to model our thoughts and to interact with us.
Jim: Makes sense. All right. A couple more items before we wrap it here. Way back in your work on MicroPsi, you said some things — and you still do from time to time — that emotion and motivation are core to cognition, not just bolt-ons. Could you dig into that a little bit?
Joscha: It seems to me that the way in which we work is to serve our motivation, and our motivation is cybernetic and it cannot be reduced to a single drive. If we look at some of the paradigms in AI that are informed, for instance, by Karl Friston’s excellent ideas about minimization of free energy — I don’t think that what you and I are doing is simply best described as the minimization of free energy. This seems to be something that we might be doing in the limit — basically as complexity in the universe, as agency in the universe, maybe as life in the universe. But even life is not just minimizing free energy — it’s self-replicating. That’s the thing that might be instrumental to minimizing the free energy, but self-replication is more important to life in many aspects than the minimization of free energy.
And to me, it might be important to get a cup of coffee, or to drive to San Francisco, or to find a life partner, rather than minimizing free energy. There’s a particular interface that I have as an organism to the world. There’s particular hardware that I have available to serve these purposes. And the particularities of my needs as an organism, my particular interface to the world, to other agents, to my ecological niche, and the way in which I have evolved priors to fit into this niche, and the hardware that I have at my disposal to serve these purposes — all of this is informing my behavior much more than a general principle of surprise or alarm minimization.
Jim: That’s what we talked about before. At least one of my strong views is that downward causality is real at much higher levels. The free energy principle may actually be correct at a lower level, but it doesn’t explain real phenomena at higher levels of the emergent stack.
Joscha: Yes. I sometimes also joke that we are anthropomorphizing people too much, which means that there is often not something that is specifically human about us. But there is basically a more general spirit that possesses a monkey-like organism that has to live in a particular kind of environment. And there is a pretty wide range of software that can run on the brain of such a monkey, and a large subset of this range is actually implemented. So there is so much different stuff that fits between two human ears that we should not overgeneralize and assume that because our parents and our siblings and our friends and I seem to be somewhat alike, that everybody is like this. There are specific reasons why we and our friends might be alike that don’t fit into the calculus of other people and the origins of their minds. So there is a wide range of possibilities in which human minds can relate to themselves and to the world and make sense of reality and experience it.
Jim: Well, that’s actually a great setup for the last question, which is: if who we are is essentially a hallucination of a hallucination, where do things like meaning and values come from?
Joscha: They come from us being social beings. We are state-building. We are not just single-celled organisms that are made of many cells — we are actually multicellular at this next level. We basically form transcendental agents. We form agents that are connective, that are made out of us, out of beings like us. So for instance, our families are agents that we can represent at the level of the family. And then we think not about what do I want from this family, but what does this family need me to do?
And when we justify our purposes to each other, when we ask others to not just be transactional and give us something because we give them something in return, the question is how can we justify these higher-level agentic purposes? What needs to be done from the level of a family, or from the level of a village, or from the level of a company, or an organization, or a nation-state, or a civilization? And whenever such a group is talking to another group, it can ask: how does your group justify its purposes? Why should I be respectful of the purposes of your group beyond the need to negotiate a peace between each other? Why should we actually be cooperating?
And that’s because the shared agency that we can build is trying to be optimal. And as soon as we are willing to submit to forming optimal agents, we can scale this up, and the global optimum of collectively enacted agency has a name — we call this God. And this notion of God is a bit overloaded by religious mythology, which distorts our perspective on it. But I think it just means this: there is this singular notion of God. There’s also a notion of gods in polytheism, which are basically multi-mind selves. There is this idea that there is something that should be done from the best possible perspective that can be taken. And it’s something that’s very hard to figure out — and figuring this out is also part of what should be done, and what is not given but what needs to be done.
But when we are talking to each other, when we justify each other’s purposes to each other so we can coordinate, we ask ourselves and each other: what does God want us to do? What should be done from the perspective of the best possible agent that I can co-create, that I can be part of? And this is ultimately, I think, our source of meaning. And we are often not consciously aware of this — we need to reflect deeply to get to this point and to infer it. A number of philosophers have done it before, but we don’t read them a lot. But subconsciously, most of us have that.
Jim: Though, the way you describe it — building up relationships from individuals to families to communities — it sounded very pragmatic, situational, and contextual, without any essentially fundamental grounding in moral value or meaning. Do you think that’s an accurate description, or do you think that there is a way that this top-down, godlike “best of all possible ways” is able to push itself all the way down and have significant influence even at these lower levels of the meaning and value stack?
Joscha: In practice, it needs to be implemented. And it’s very easy to imagine that you are serving some global good without you actually doing anything in practice for it. And so I think that it basically springs out of your everyday interactions that need to be implemented. And so I’m personally wary of a politician who does not maintain a good relationship with their children and is not able to guide their children to be good beings. Of course, you can be unlucky, and your child is unfortunately born a psychopath or a drug addict or whatever. But generally, if you are unable to run a good family, I don’t think that you can run a good organization. If you cannot run a good organization, you cannot run a good country. If you cannot run a good country, you have very little business telling humanity what to do at the civilizational level.
And so there is basically a certain kind of humility that we need when we are making the claim that we are serving higher purposes. It’s very easy to hallucinate this if there is nothing against which you need to test it. Ultimately, the proof is in the pudding. Do you harm people? Are you helping other people? Are you helping other agents? Do you make the world better? That’s something that is not abstract — it’s something that ultimately is concrete and manifests in your everyday actions, in your everyday environment.
Jim: All right. I think this has been an amazingly interesting conversation. Thank you very much, Joscha Bach.
Joscha: Thank you, Jim. That was great.