Discussion with Murray Shanahan 2

This transcript is automatically generated; we strive for accuracy, but errors in wording or speaker identification may occur. Please verify key details when needed.

[00:00] Michael Levin: Thank you so much for that talk you gave. That was really interesting, and I took some notes of things that I wanted to bring up, but anything you want to talk about, we can talk about.

[00:11] Murray Shanahan: There's so much in your work that's of interest to me. I was looking at that entropy paper of yours, and there you hint at what your take on consciousness might be. I'm really interested in that and hearing where you're going with that. Maybe you could tell me a bit more about that so that can be on the agenda as well.

[00:41] Michael Levin: Let's do that first. The big thing, of course, is that I have not up till now said very much about consciousness, because I've tried to lay the foundation for the easier aspect, which is still considered very weird, which is all these third-person cognition studies that we do in unfamiliar, unconventional embodiments. Now we can say some things, and I'm going to say some of these things in Crete when I give my talk. I think a few things. First, if we make a list of the reasons for which we tend to attribute consciousness to each other, in other words, the problem of other minds, there are a few standard reasons. If you run through those reasons, many organs and structures in your body meet those criteria. So, to whatever extent that we are willing to attribute based on behavior — problem solving and valence; mechanism, the kinds of things that we know underlie cognition in the brain; evolutionary lineage relationships; Phi, integrated information metrics — all of those reasons should mean that you have to take very seriously the idea of consciousness in other parts of your body. Usually at this point people say, "But I don't feel that my liver is conscious." You don't feel that I'm conscious either. Your left hemisphere puts up a very nice discussion about what it feels like for it to be conscious. But I think that in the absence of language, which we don't have evidence for in these other structures, everything else is there. I think we have to take that very seriously. We are working on some tools to enable us to communicate with these other intelligences inside the body so that we should be able to talk with language, with a translator interface made of AI. We should be able to ask the liver questions about things it cares about, how it's doing, and its various goals.

[02:52] Murray Shanahan: So that's what I call engineering and encounter.

[02:57] Michael Levin: Exactly.

[02:59] Murray Shanahan: I think I might put things a little differently from you because the reasons that we have for ascribing consciousness to each other are also there in these more exotic cases. But I wouldn't quite put it that way because I don't think we have reasons at all. I think we just do it. We just treat each other as fellow conscious creatures and it doesn't require reasons. And it's only when we put on a philosopher's hat that we start to think, what are the reasons. And we put on a philosopher's hat, of course, when we're doing philosophy, because we might be interested in those meta-level questions. But what is interesting at the moment is with the advent of AI and also other forms of exotic consciousness, and you're raising a whole new kind of category, is that then we're in the sort of territory where it doesn't come so naturally that we just do it. But then we have to ask ourselves these questions and we try, we step above this kind of inclination, our natural inclinations to treat each other as fellow conscious creatures, which I think we also naturally do with other animals very often. And because we've got these very exotic cases, then we have to take a step back and then we're saying, is it conscious or is it not? And it becomes, then you step outside of this just doing it into thinking about it. And I always feel that that's where we can go wrong philosophically at the same time.

[04:38] Michael Levin: No, I completely agree with you. I don't think these are actually solid reasons to do it. They don't cross the chasm to first-person experience. The reason that I emphasize them is I use them as tools to get people to revise their intuitions. But once you get there, if I can get you there with reasons, then after that, maybe we can widen the kinds of things that you're talking about, right?

[05:09] Murray Shanahan: Yeah, absolutely.

[05:11] Michael Levin: and so that's the more conventional aspect of it, the more crazy aspect of it, which is very out of favor, I think nowadays. This is now my model in the biology that we do. The way that I crawled up to this, and this is decades of laboriously doing work and then holding back the things that I thought and wanted to say, but didn't want to say until we had the data. We increasingly have these remarkable forms and functions and behaviors in novel organisms that have never been here before. So the answer "eons of selection is what accounts for" doesn't hold anymore. So now you have to ask the question, where do the goals and cognitive properties of completely novel beings come from? I've started on the biological model that I've been pushing recently, and there's a chapter, which I may have sent you, that is basically what I call this platonic space model. The idea is not to try to hew close to what Plato said. The only reason I stick with that name is because the mathematicians are already there. They already think that there is a space of facts that are not fixed by physical properties of the universe but that matter in the physical universe. So this is what I'm going for. And roughly speaking, my view is that the contents, the patterns that are in that space, there are lots of low agency patterns that we recognize as facts about prime numbers and truths of number theory and things like that. But there are also very, very active high agency patterns that we would recognize as kinds of minds, behavioral propensities. And that when we make physical objects, be they robots, embryos, biobots, AIs, whatever we make, starting with simple mechanical things, what we're really making are pointers into that space. We're making interfaces to patterns, and fundamentally we are not the physical embodiment. I don't think we make intelligence, we don't make consciousness. What we make are interfaces through which, if we know what we're doing, and I think most of the time we don't know what we're doing any more than we do when we make an embryo. Dennett used to call it competence without comprehension. We can make these things, but we don't actually understand. Sorry, say that again.

[07:40] Murray Shanahan: Termite mounds. He contrasts the termite mound with the Gaudi Cathedral.

[07:48] Michael Levin: And the thing with the cathedral is that it uses passive materials. So it's easy to do that when you have passive matter, but when you're dealing with a gentle matter we're not good at knowing what we're going to get. And so my view of consciousness at this point is that consciousness is the perspective of the pattern from that side. When you're looking, when you are the pattern looking into the physical world through that interface, that is what we call first-person experience and consciousness. And when the two bodies are looking at each other, this is the conventional person science that we all do.

[08:26] Murray Shanahan: I suspect that there aren't the kind of metaphysical commitments that bother me, the kind of dualistic metaphysical commitments that bother me. I tend to want to back away from them because you're providing a kind of purely— I was going to say "mathematical description," but I guess it's in the same space as a mathematical description of what goes on.

[09:04] Michael Levin: I think I'll bite down on the dualism part because I suspect I do believe some things that would bother you on this on the side. I do think that this space is at least as real as the physical space. I don't think it's a metaphorical way of talking. I don't like this business of emergence, which is basically just we're going to catalog surprises and that's that. I think it's an ordered, structured space of patterns that functionally matter. I think they're causal in the physical world. I don't believe in physicalism anymore. I don't think the physical world is closed in that sense. I think some of those assumptions are probably there that you might not be happy with.

[09:55] Murray Shanahan: When we're saying that they're causally efficacious, these high-level patterns, that's not suggesting they can't be accounted for by what's going on at a lower level. We could simulate these things on a computer, and we'd be simulating them at a low level. At the same time, we might want to understand what goes on at a higher level, the ontology we might use to describe these things. We'd elect to use a higher level because it would be pretty incomprehensible if we spoke about it at the lower level. Nevertheless, there's a kind of causal closure there. It's not like there's something mysterious happening.

[10:42] Michael Levin: It's not a mysterious position; I'm not saying that they are somehow impossible to understand in a rational way. In fact, I think it actually is a much less mysterious position than a lot of emergence. When I talk to people, I say, "Look at this, isn't this remarkable," and people say, "Well, that's just something that holds in the physical world." Well, great, but I don't want a catalogue of things that hold. I have a more optimistic view that these things are ordered and structured, and we can study them and understand. The thing is with computational modeling, you can absolutely model it, but I think what you've then done is basically create a different kind of interface to the same pattern. So they're causal in the sense that, imagine this is a simple example. Let's say a world in which the highest fitness belongs to a certain kind of triangle. A certain shape of a triangle has the highest fitness. Evolution cranks through a bunch of generations and gets the first angle, and then it cranks through a bunch of other generations and gets the second angle. Now you don't need to do the same thing for the third angle. There's this magical free lunch: you know the third angle, and you don't need to evolve it. It's practical; it's causal in the sense that evolution can exploit the hell out of these things because they provide this kind of free lunch, things you don't have to evolve directly. Again, it isn't mysterious. It's not mysterious because you can reproduce it as much as you want and study it. In fact, that's what I think we should be doing: trying to understand what interfaces would cause what patterns to come down. I think we actually only understand that in a very small set of cases. But it is kind of magical in the sense that I don't think we have a good vocabulary for it. This isn't some weird thing that sneaks through the quantum interfaces; even in Newton's deterministic universe, you already have this.

[12:38] Murray Shanahan: I often, in my head, think about patterns that I see in water, in vortices. Especially walking out in the countryside, we have some small streams near where I live and you can walk out and sometimes when there's been a heavy rainfall you can see particular vortices form — they'll form and then they'll disappear again. Sometimes two will form and then merge and then disappear again. I often think that perhaps that's an example of the thing that you're talking about, because there's a self-sustaining pattern, but on a certain level it's all just water. The pattern is indisputably a real thing. You might get caught in the vortex, and a stick will get caught in it. On a certain level of description, you certainly want to talk about the vortex as a thing. So is that a good analogy for you?

[13:44] Michael Levin: I think I'm saying something more radical than that. I agree with everything you said here.

[13:50] Murray Shanahan: I'm talking about an everyday phenomenon that confronts me, whereas you want to link it to the kind of platonic object of possible vortices, vortex-like patterns in the platonic realm of possible patterns. When you talk about an interface, where were you going with that? What were you about to say?

[14:19] Michael Levin: Many times in biology and no doubt physics as well, when you start chasing down the explanation for why this or that is happening, eventually you'll reach things like, because that's the distribution of primes. Or that's the Fibonacci, that's how the Fibonacci series shakes out. And those kinds of things. You've now left the— it's not a physics explanation, it's not a historical evolutionary story of selection. You've now gone into this other. We're pretty comfortable doing that for math, but I think we're going to find the exact same thing. Why does a Zenobot do these weird things that no other creature on Earth has done? For the sake, it's haunted by these patterns the same way that triangles are haunted by the rules of geometry and so on.

[15:07] Murray Shanahan: Does this connect with Turing's ideas of morphogenesis?

[15:16] Michael Levin: I think so. Turing — it's amazing. He actually said some very prescient things on this matter. He didn't live long enough, unfortunately, to follow it up. I think he was interested in, this is a guess, self-organization of morphogenesis because he saw that symmetry between the self-creation of bodies in embryogenesis and the autopoiesis of minds, that there are underlying patterns. I think he would have been all over this stuff if he had lived longer.

[15:49] Murray Shanahan: What struck me in describing your thinking was that I'm pretty sure that Turing saw, as it were, the mathematical realm as constraining what was possible in physical reality, and then seeing that manifest in these actual patterns, right? And we see that.

[16:11] Michael Levin: Constraints, absolutely constraints, but I think it's not just the flip side of that: because they are, in a sense, free lunches for a lot of things that biology wants to do, they're also enabling. Voltage-gated ion channels — as soon as, in evolution, you've found a protein that's a voltage-gated ion channel, that's a voltage-gated current conductance, that's a transistor. You have two of them, now you've got a logic gate, now you have a truth table. Did you need to evolve the truth table? You didn't; you just get it. And the fact that NAND is special and all these other things, just boom, now you've made that interface; you didn't have to evolve any of this stuff. You get it for free, and now you can do all these amazing things. So constraints, but also enabling steps.

[17:01] Murray Shanahan: So tell me what else I should, because your oeuvre is quite big and varied, as mine is in fact. I realise that it's quite difficult for people to know where I'm coming from, because I've got all this history of stuff and people tend to latch onto the wrong things and I despair of it. They look at something I wrote 15 years ago and they think that that's what it is; it's kind of frustrating. I think you're probably in a similar position. So that entropy paper was great, but where else should I— tell me where else I should look? What would?

[17:41] Michael Levin: I can send you a few of the key latest I have on my website. I have a map—it's like a skeleton with the main backbone of the main theory papers and then a bunch of stuff that hangs off of it with details. I can send you some of those things. There's a few pieces that are important. I'll link this back to some things you said in your talk. One really interesting aspect had to do with disembodied consciousness. We as humans are fixated on three-dimensional space and movement as embodiment. Long before muscle came on the scene, biology was navigating all kinds of spaces that are hard for us to visualize: metabolic space, physiological state space, transcriptional space, anatomical morphous space, which is what we study all the time. Only after that did it get to muscles and movement in 3D space. That loop—perception, decision, action, memory in the middle—living things carry out in all kinds of spaces all the time. They navigate intelligently, they solve problems. We need to be careful with this idea of disembodiment, because people say this about organoids in a dish: "well, it's disembodied." It is not disembodied. That thing is navigating who knows how many spaces that we can barely visualize. I don't have any special claims about language models, but it makes me nervous to talk about software in a box as disembodied, because we are also bad at recognizing embodiment of living things, which takes place. These other spaces are, in terms of consciousness studies, as real as 3D space.

[19:30] Murray Shanahan: I guess when I talk about disembodiment, I do have something specific in mind. I do flesh it out in terms of broadly human-like worlds of 3D worlds that comprise objects that are 3D spatially extended in three dimensions and so on. That's the way I'm defining disembodiment in the context of the discussion I want to have. A lot of my current work is about trying to navigate. It's funny because we use these spatial metaphors all the time to navigate a territory which is between the familiar and the very unfamiliar. How can we negotiate this boundary? Part of the way I think in terms of philosophy is that our language is to some degree constrained by what is familiar. If you stray too far from what's very familiar, then you're actually moving beyond the bounds of our current language. We can't, but we can push the boundaries. As soon as we think about things like large language models, we're starting to push the boundaries in interesting ways. But it's quite difficult to jump all the way out of that. That's how I relate to what you're talking about. I can 100% go with you on what you're saying, but I'm coming at it without wanting to jump too far too quickly.

[21:28] Michael Levin: Sure.

[21:29] Murray Shanahan: But we're interested in the kind of exotic. I'm interested in moving there in a slightly more conservative way.

[21:38] Michael Levin: Which I understand completely. This is why I try to set this up in cells and tissues, because I think it's a stepping stone. I agree with you. I'm not jumping. I have zero claims about what language model, what space they're working in. I don't know anything about that. But what I want to get people used to is to think about a semi-alien thing, which is, it's not like it's part of your own body. It's not that crazy. As you say, language is critical. What do I mean by navigation? We can define things like your current position, the position you would like to be in, the effectors that you have for moving in that space, the sensors you have for sensing states in that space, the barriers that you might encounter that you can't get past, other agents that you might encounter that are not just fixed features of the environment. All of these things can be defined in all these weird spaces. We can talk about what a metamaterial means in transcriptional space, or what a barrier means in physiological space, or what it means to have misperceptions, visual illusions, to be confused, to have memories, to do long-term planning. We can do all of these things in these other spaces. I feel like once we get a handle on what it means for a cell to be navigating transcriptional space and using intelligence to do so, and to have mistaken beliefs about that space and all these kinds of things, we will be, I think, in a better shape to jump or decide not to jump into these kinds of ideas for AIs.

[23:11] Murray Shanahan: Gives us a larger framework onto which we can try and map.

[23:16] Michael Levin: See how well it goes.

[23:20] Murray Shanahan: That's really interesting. It does fit with what I see as the large-scale cultural, social agenda that we have to pursue, which is to figure out how to accommodate ourselves to these stray mind-like things; they're mind-like. We need new conceptual tools for doing that. That fits in very well. Who knows how. What I'm always saying is that I don't know where this conversation is going to go. At the same time, I consider myself to be a commentator on the conversation, and that's the metaphilosophical position of saying it's just a conversation and trying to push back on people who are overly metaphysical about it. But I also participate in the conversation, so I have some preferences, and I would have been a bit conservative about embodiments and the whole engineering encounter thing. So that's me participating in the conversation by introducing my own little bit of vocabulary. But I'm not committed to that on the higher level. On the higher level, what we need is more and more radical ways of talking that we can bring into this conversation.

[24:39] Michael Levin: I agree with that strongly. There are sometimes people, philosophers who contact me and say, "You commit a category error when you talk about cells learning," and say, "Where did these categories come from? We got them a couple of thousand years ago and you just want to stick with them." We never question them. The categories need to evolve with the science. If we can't figure this out for our own cells and organs, the hope of figuring it out for aliens, the AIs, cyborgs, whatever is slim. We need to be able to handle this. That's my position. So I'm trying to build a bridge to some of these things.

Discussion with Murray Shanahan 2

Watch Episode Here

Listen to Episode Here

Show Notes

Transcript

Related episodes

“The Unreasonable Effectiveness of the Behavioral Sciences in Developmental Biology and Biomedicine”

Discussion #1 with Katrina Schleisman and David Burke

Conversation with David Resnik 1

Discussion with Murray Shanahan 2

Watch Episode Here

Listen to Episode Here

Show Notes

Transcript

Related episodes

“The Unreasonable Effectiveness of the Behavioral Sciences in Developmental Biology and Biomedicine”

Discussion #1 with Katrina Schleisman and David Burke

Conversation with David Resnik 1

Thoughtforms Life Podcast