为什么 AI 面临柏拉图问题——Mazviita Chirimuuta

What should we say as philosophers about the relationship between neuroscience and philosophy of mind? How much of our ideas about how the mind works can we read off from the results that neuroscience is telling us?

The results you get in the lab can be well established and fine. There’s nothing wrong with those data, but there’s more of a problem with generalizing from what you learn in the lab to cognition in the real world outside the lab.

It’s precisely all of that complexity and all of that interactivity that is really important to how, for example, animals are able to negotiate their environment. It’s not an argument that AI is impossible so much as why it seems so possible, so inevitable, to people. If you look at the history of the development of the life sciences and psychology, there are certain shifts towards a much more mechanistic understanding of both what life is and what the mind is, which are very congenial to thinking that whatever is going on in animals like us, in terms of the processes which lead to cognition, they’re just mechanisms anyway. So why couldn’t you put them into an actual machine and have that actual machine do what we do?

Yes, but anyway, Mazviita, welcome to MLST. It’s amazing to have you here.

Thanks so much for having me along.

You wrote this book, The Brain Abstracted. It’s an amazing book. Folks at home should definitely buy this book. It’s really, really good. Tell me about it.

It was quite a few years in the making. I think officially I started writing it maybe in 2018, and it came out in 2024, but it was really based on ideas that I’d been working on maybe since 2014. I started publishing some philosophy of science papers about computational explanation in neuroscience, and then, going back beyond there, some of my own experiences when I was doing training in neuroscience on the visual system.

I was using computational models of the era before there was deep learning or anything that fancy. I was thinking about what it really means to understand the brain through this lens of computation: having models which not only simulate the brain, as with biological simulations using computers, like weather simulations and so forth, but actually allege to duplicate the function of cells in the brain. This is an additional claim that’s made about computational modeling when it’s applied to the brain as this unique structure, which is not only a biological organ but also a kind of computer itself.

The arc of your book is that we have this problem with simplification because, as scientists, we want to build legible theories about how the world works.

A lot of philosophy of science in recent years has picked up this topic of abstraction and idealization. Abstraction is quite a general word, which can just mean ignoring details that are there in concrete, real-life situations. It would be familiar to you from doing Newtonian problems in physics, where your teacher tells you, “Well, there’s always friction in real life, but we’ll pretend that the friction isn’t there.” So you’re leaving out a detail which is known to be there in the concrete system.

Idealization means attributing properties to the system that you’re modeling in science which are known to be false. For example, in genetics modeling, the assumption is made of infinite populations. These kinds of idealizations often make the calculations more tractable, but of course there’s no such thing as an infinite population in real life.

In some way, an abstraction is also always a false representation, always an idealization. Sometimes the difference between the two can be subtle. How I put this in the book is that an idealization points us to the thought that, when we have a scientific representation, we’re presenting something that’s cleaner and better than the thing in real life.

When we talk about someone being idealistic, it’s like they have a view of how things should be, and unfortunately reality does not live up to that. Idealization in science is often to do with representing things mathematically in a way that’s cleaner and neater than could be possible in real life.

On abstraction, you said in your book that there’s the lofty philosophical version of abstraction, which is upstairs in the heavens of Plato, I think you said. Or even Galileo: there’s this idea that these natural forms exist which are disconnected entirely from the spatial and temporal realms.

Then there’s the more deflationary view of abstraction, which is simply that we ignore details. I’m speaking with my good friend François Chollet again tomorrow. He’s releasing the new version of the ARC challenge, and I think he does have this idea, as many AI researchers do: this Platonistic idea. He calls it the kaleidoscope effect, which is that the universe is basically written in code, and what we see is like a kaleidoscope when all of the rules of the universe just get composed together in different ways. All we need to do as AI researchers is decompose it back into the rules.

What could possibly go wrong? I watched some of the videos with François, and I found it really fascinating, precisely this kaleidoscope hypothesis, because, seeing that as a philosopher, I thought, “That’s Plato.” François precisely says we have the world of appearance. It’s complicated, it looks intractable, it’s messy, but underlying that, the real reality is neat and mathematically decomposable.

This is precisely the contrast between the world of forms and the world of being—eternal, stable truth—and the world of becoming, appearance, messy, flowing, complicated reality. It goes back thousands of years in philosophy. It’s really interesting that this is an assumption not only that AI researchers often make, but that runs through science as a kind of justification for the pursuit of mathematical representations, even when they depart from known facts about the concrete physical systems in reality.

The idea is that the mathematical representation gets you closer to the underlying truth of how things are, as opposed to what I call the down-to-earth view of abstraction and mathematical representation: that it’s something we do because of our cognitive limitations. Instead of thinking that abstraction gets you to a higher level of reality, we can say that we do abstraction because we’re finite knowers. There are limits to how much complexity any individual person or group of people can actually encompass in their modeling strategies or representations.

Actually, it’s only by pretending that things are more simple than they really are that we get some traction. That’s the down-to-earth, mundane explanation of why abstraction is so much used in science.

It’s so pervasive in the deep learning world. I also interviewed the folks who pioneered this geometric deep learning blueprint, and that’s the same idea, basically: the world is described with geometry, and all we need to do is imbue these geometric inductive priors into deep learning models. Then they can essentially, by reducing the degrees of freedom to ones that are aligned with how the universe works, get us where we want to go.

I think the notion of patterns and real patterns, to invoke Dennett’s term, is a helpful one. One thing that you could say is going on here is that, yes, there’s lots of complexity in the natural world. It’s apparent in the data, but if you just denoise the data a bit, underlying there is a real pattern. We shouldn’t have to be Platonist and weird about it; there’s just regularity that’s sometimes masked by noise. That doesn’t seem too metaphysically problematic.

But one of the questions that I pose as a challenge to that very moderate view—and I say this frequently in the book—is that, when you’re saying some of the apparent irregularity in the data is irrelevant, that’s your decision as a scientist. It’s not relevant to you at the moment, but it could be relevant to someone else. It could be really important to how that system works in the natural world, for reasons that you’re not aware of.

When we classify signal versus noise in our data sets, we shouldn’t ignore the fact that those are decisions we’re bringing to bear on our investigation. We shouldn’t assume that we’re just reading off the signal, the real pattern that’s there in reality, and that there aren’t many other significant, real patterns there. To the extent that we’re probably also creating pattern through the very denoising process that we bring about.

Interesting. Physicists aren’t under any illusion, so they know that Newton is an idealization. Just to contrast that, you cited reflex theory. Of course, Pavlov and the dogs—folks at home will know about that. Newton is still around; we still use that. But we don’t use reflex theory anymore.

Yeah, yeah. This is a chapter that I present in the book as a case study of how oversimplification can get scientists on the wrong track. The history of science is like hindsight is 20/20. We’re looking at a theory about how the brain worked that was really dominant for a few decades at the end of the 19th century and the beginning of the 20th century.

It’s familiar to us with Pavlov and the dogs, with this idea that we can explain behavior in terms of reflexes that get conditioned. There can obviously be learning involved with that.

The most ambitious version of the theory said that all of the functions in the brain are basically versions of reflex arcs, so sensory-motor loops. A very prestigious and well-regarded physiologist like Charles Sherrington was heavily invested in reflex theory. But he admitted in his book, The Integrative Action of the Nervous System, that this notion of a simple reflex is an idealization. It probably doesn't exist in real life.

And yet this is the key that's going to unlock neurophysiology. It's going to help us decompose and make sense of all of these different interactions that could be observed experimentally. So what seemed to be going on there is that scientists were taking that age-old method, which is that it's a good heuristic to seek parsimonious explanations, to use Occam's razor. The obvious thing to do was, “Let's assume there's such a thing as a simple reflex,” and then they ran with it way too far, never actually being able to explain the amount of data that they had initially thought they would be able to explain.

And it's not clear how long the reflex theory could have gone on if it hadn't been for computational theory coming in during the Second World War era and basically providing an alternative explanatory framework, which was also quite neat and, I would say, provides its own kind of idealization toolbox.

A very popular thing in cognitive science is to say, “Well, if something behaves the same way as a cognizing human, for example, then we might draw inferences that it has consciousness and many other cognitive faculties.” But there's always this almost ignorance of the actual mechanism of the object of study. I think behaviorism has a bad name, but it's not that discontinuous with a lot of thinking which is normal and still acceptable in science: treating things as black boxes.

This is precisely what the behaviorists said. It's like the mind is opaque. It's hidden within the walls of someone else's individual subjectivity. As scientists, all we know are the inputs and the outputs, and we'll just track those. That's like a version of what you just said: if the inputs and outputs—the behavior of this system—are looking like what we know to be a conscious system elsewhere, let's just treat them as all being the same class of objects, given that the only available information is the inputs and outputs.

I think that kind of reasoning can be fine in certain contexts, but it's a philosophical leap to say that the access we have to our own thoughts, and the presence or absence of subjectivity that we're aware of with other people, is irrelevant to making these decisions or judgments about what other kinds of systems can have consciousness. I think it's much too quick to just go behaviorist and say, “Well, there's no relevant difference between X and Y, even if one is a person and one is a machine, just because we can say that there are some similarities in inputs and outputs.”

If I remember correctly, at one point you drew an imaginary kind of graph where you said, on one axis we have scientific realism, which is where our scientific theories actually represent things in the world, and then we have empiricism, which is the idea that the facts we receive tell us something about the world. Then there's this more interesting axis, which I think you're very inspired by, which is this kind of constructivist idea. Can you explain that?

The constructivist path, which is different from the scientific realist and empiricist one, really runs with the idea that we are active makers of knowledge. It shouldn't be confused with the kind of constructivism that we have in some more extreme branches of sociology of knowledge, which say that all scientific theories are social constructs and not constrained by phenomena that have been observed in nature.

I'm not saying that scientific theories are merely constructed in the way that poems could be, as a work of imagination and so forth. The idea is that there's this interactivity between humans—groups of scientists and their plans as epistemic agents—going out into the world with an agenda to find stuff out about certain phenomena in order to achieve certain goals, often the goals of technological, applied science. There's some pushback from the things in nature themselves that they're investigating.

The idea is that knowledge is always the product of this interactivity. We cannot discount that there's a human framing side to this. We can't go along with the idea that a scientific theory is just reading off the source code of the universe, as if the human way of conceptualizing those phenomena had no bearing on the theory as it ultimately turns out.

But we also can't discount that the theory that arises is constrained by how things happen to be. That is worked out through the process of experimental interaction.

You said—I think you were inspired by Immanuel Kant. He had this transcendental idealism, and please bring that in, but that somewhat informed your own view, which is this haptic realism. Can you introduce that?

That's saying that knowledge comes about through this process of interaction. This notion of haptic realism emphasizes that it's through engagement, with haptics being the sense of touch. The contrast here is with an ideal of knowledge based on the idea that we can know things in a disengaged way.

If you think of vision as the archetype for knowledge, what happens when we look around at our surroundings and use sight as a source of knowledge? We can get into this mindset where it seems like we do not have to interact with things in order to know them. We can just absorb information passively, and then, because we're not bringing about our representations in an active way, it would seem to us—and I'm not saying this is how vision works—that it's a kind of conceit that often comes about if you use this very visual model for knowing.

John Dewey called it the spectator theory of knowledge. This is a clear predecessor to what I'm saying here: we just look around, absorb how things are, and our knowledge is sort of entirely objective. It's almost like a God's-eye view of reality.

But if you think that scientific knowledge in particular is more touch-like, you can't ignore the fact that we run into things. We have to pick things up, engage with them, and ultimately change them in order for us to acquire knowledge of them. So you cannot discount the fact that we're meddling with things in the process of bringing about our knowledge.

Another dimension of this haptic metaphor is that our hands are not only a sensory organ, but they're also the means by which we manipulate things. Manipulation means precisely working with the hands. I think that really captures the double face of scientific models.

They're both a means of acquiring knowledge, in the way that hands are also sensory organs—we find things out about the world through the sense of touch—but they're also means for changing things, for doing things.

We speak about this in evolution. What would happen if you could just rerun evolution? What would happen if we could have a parallel universe and the entire enterprise of science just ran again? What you're alluding to is that it wouldn't be completely different. Maybe there are some guardrails, but it is actually quite divergent.

There's certainly contingency in the history of science, where people start out with cultural factors which prompt them to ask certain kinds of questions and not others. A view quite similar to what I say about haptic realism in the book is held by Hasok Chang, who's a professor of philosophy of science at Cambridge. He has a view which he calls “Realism for Realistic People.” That's the title of his new book.

He's an out-and-out pluralist about science. He says that because there is contingency in the history of science, it means there are paths not taken, but we could maximize the acquisition of knowledge if we explored as many of those different paths as possible. That's not something that I say in the book myself, because I think there are also reasons why it makes sense to narrow views and paths of inquiry.

Also, we don't have unlimited resources. But, sure, there are opportunity costs that come along with taking a certain path, and there are others that aren't pursued.

In the enterprise of science, there might be a trope or an idealization that we're getting closer to the truth.

Yeah.

Do you think that's the case? Do you think that as the enterprise of science progresses, we're getting closer to the truth, or could we be in cul-de-sac basins of attraction and so on?

That's very much associated with scientific realism. There's this view that there is one way nature is, and science succeeds insofar as scientific representations conform to this one way that nature is.

My view takes very seriously the idea that nature could just be inexhaustibly complex. If you ever try to pin it down in one representation, there are also ways that it could be represented—sort of inexhaustibly many different varieties of ways that you can investigate it—and then also ways that any one representation is lacking.

So there's an inherent lack of convergence that that picture brings about. One of the ways of expressing this is to say that nature is protean.

There’s this mythological character called Proteus, who was a shapeshifter. This mythological being lived in the sea and would keep changing his shape. You couldn’t pin him down, but if you could, he would answer you a question and tell you the truth. But the thing was, you had to pin him down.

I think this is a really nice illustration of what’s going on with our interactions with nature as scientists. Nature is sort of inexhaustibly complex. There are all kinds of patterns and things going on there. It can be pinned down, and we can get true answers, but when we release our grip, it will carry on shape-shifting. There are lots of other ways that it could be.

So, yeah, one final theory. I’m not so convinced by that. This is very much a view that I think makes sense if your basis for your theory as a philosopher of science is really the biological sciences, which is where I’m coming from. If you’re a physicist, it seems much more natural to think that there is 1 fundamental set of laws of the universe which is going to be nailed down once and for all and could explain everything.

Biology just throws up lots and lots of examples and particularities. It tends to be less considered, less intellectually satisfying in comparison with physics. You can just spend all your time in biology doing stamp collecting, because there’s this thing and there’s this thing and there’s this thing. How do you tie it all together theoretically? On the other hand, I think that if you take that particularity and that shifting quality of biological phenomena, then actually it just forces you to think about knowledge differently.

In your book, you spoke about a trajectory, I suppose, of possible failures of simplification. We just spoke about reflex theory, but one of the big things is this metaphor of cognition, or the brain perhaps, as being a kind of computer, and you spoke about the early roots of this, from reflex theory to cybernetics and computationalism. Can you sketch that out?

This connecting thread is really this idea that what cognition is is something that is machine-like. Going back to the 17th century, this is a view associated with the philosopher, physicist, and physiologist René Descartes, who said we need to go along with this idea that everything that happens in the body is explicable in terms of quite simple mechanistic forces. This idea that biological systems are machine-like has obviously been hugely influential in the different branches of science.

The reflex theory was one instance of that. People often talked about machine-like reflexes and made comparisons with Newtonian decomposition. With the computational framework, you had an actual machine—a digital or analog computer—which could be compared with brain processes. Cybernetics is an interesting stage along the way, because they were building little devices which had some degree of autonomy and were supposed to emulate versions of negative and positive feedback as hypothesized to occur in the body.

At the core of this research idea is that if what’s going on in the body is ultimately a mechanistic process, then by redoing engineering with this nonliving system which is capturing some of the core operating principles that we find in biology, we can use that device as a map, as a resource, to then reinterpret what’s going on in the biological system. You saw that, for example, with McCulloch and Pitts in their 1943 landmark paper, interpreting neuronal cells as logic gates and then saying, “Yeah, you could build a computer out of neural nets.” This is the origin of neural nets as we know them today. This is the birth of the idea, but then using that notion that neurons are logic gates to interpret what’s going on in physiology.

What I describe in chapter 4 of the book is a sort of back-and-forth process of making devices which are somewhat inspired by biology and then using those as the lens through which to review biology again. I say that the advantage and the appeal of this process is that it allows you, or gives you a kind of license, to ignore so many things that are happening in the brain and nervous system which are just not shared with nonliving machines: all of the biochemistry, all of the ways that neural tissue is shaped by vasculature and interacts with the immune system, and all of that sort of background stuff.

If you’re a theoretical computational neuroscientist, you can say, “I’m only interested in the computational properties of the brain. I don’t need to care about all of that messy biological detail.” So it gives you a kind of tunnel vision. As scientists, it can be fine to have tunnel vision; you can’t take in everything at once all of the time. But what I take issue with is the kind of ontologization of that—saying that because computational neuroscience is this successful field of inquiry, we know now that the brain is a computer. I think that is not an inference we should make.

Yeah. I mean, I don’t think connectionists typically argue that. They would say it’s a different mechanism. But they think there’s some kind of functional equivalence. Mhm. And that’s the thing, because so many folks in AI at the moment are interested in biologically plausible architectures. So what if, like the cyberneticists did, what if we have more autonomy, diversity, agency, and so on? They fundamentally think—I guess they make the assumption—that the world is a machine, and if we replicate it with sufficient fidelity, then we can reproduce the behavior.

To what extent are the mechanisms of the brain inherently bound up with the fact that the implementation here is in living tissue?

I think there’s sort of tantalizing evidence about the extent to which brain processes and signaling between neurons—not just the specialized electrical signaling that neurons do, but biochemically—are kind of outgrowths of signaling that’s happening elsewhere in the body all the time. We shouldn’t think of neuronal cells as distinctively cognitive as opposed to the other cells in the body, but rather as extensions of the ways that cells signal anyway. If neuronal function is so much a manifestation of what’s happening with metabolizing cells, that makes it more of a stretch to say that a machine that’s not living could have the same functionality.

Yeah. I mean, no one’s trying to build artificial neural networks with living cells.

No, no. But I mean, there is an analogy in neural networks. There’s this thing called the lottery ticket hypothesis, which speaks about pruning. What the researchers found is that you train this big, dense neural network, and after it’s trained—you need the density for stochastic gradient descent, for training tractability—you can prune away 90% of the connections and it still works the same way.

Maybe evolution and our biological instantiation are the same thing. We’ve been through this billion-plus-year training process, and all of these things that we think are important, like the instantiation, autopoiesis, agency, and so on, maybe those are vestigial. We can now just kind of snip, snip, snip, and create this abstract version, right? It seems reasonable. What do you think?

That it’s just vestigial? Huh. I mean, I think we really need to take seriously the economy of biological information processing. We do a lot more with a very limited energy budget running our brains every day, whereas artificial neural networks are really, really expensive to run. It doesn’t strike me that biological cognition could get away with being that wasteful. Surely, to keep things from blowing up in terms of energy being consumed for information processing biologically, there must have been a fair amount of pruning on the way.

I kind of think of agency as being a bit of a spectrum. You can think of it in a deflationary sense as being this autonomous thing that’s the cause of its own actions, and then, in the deep philosophical sense, there’s this intentionality and you can control the future and whatnot. It speaks to the physicist’s view: you have this light cone and all these micro-interactions, and of course that’s beyond our cognitive horizons. So we develop ideas of representations where we can have these distal relationships between things that are in our mind and things that are far away in time and space.

I suppose you think of that as being another form of idealization. The fascinating thing, though, when I think about agency, is that I think about it in terms of apparent causal disconnectedness. We are agents because we have consistent beliefs and ideas, and we’re not just an impulse-response machine whose actions are determined entirely by the situation. You’re a person, and I perceive that as a kind of causal disconnectedness.

Yeah, yeah, I agree. What I say in the book, in that chapter, I set out—and I say this to be very metaphysically neutral—what representation is and what intentionality is. At the same time—and this is not what I directly wrote in the book—I think I agree with you that there is something very important about connecting the notion of agency and intelligence with this thing of being responsive to what is actually very distal.

It could be distal in time and space. It could be distal because it happened a long time ago, but this is what biological memory is: things that happen to you when you're a baby affect how you are now. Physical systems, like nonliving physical systems, are much more constrained in their actions—and I don't mean that in terms of action, but just what they do, what happens to them—by what's proximal to them.

The distal is always screened off by the proximal, if that makes sense. Whereas for you, all of these things that happened in the past could be as relevant as anything that happens in the room right now, or your ideas about the future would be relevant to what you're saying right now. So, yeah, this notion of being sensitive to what's not immediately driving you in your surroundings, I think, is a really important thing to latch onto in delineating at least the class of systems that we want to call cognitive from ones that we would say are merely physical, not intelligent in any important sense of the word.

Very cool. So we're trying to, I suppose, partition the world into logical units that we can understand, and agents are a great version of that. Daniel Dennett, of course, had the three stances: the physical stance, the design stance, and the intentional stance, as a way of building useful explanations of how things behave and introducing those. But you said that you didn't quite agree with that because, to Dennett, it's a hierarchy, which means the intentional stance perhaps has precedence over the other ones.

Oh, so it's actually kind of the reverse. It's as if the physical stance has ontological priority; that's what's really there, but it's useful to use the design and the intentional stance.

Very interesting. But you said for you, you don't really have that prioritization. You're kind of open-minded.

Yeah. So that's part of the metaphysical neutrality that I set out with in the chapter: to say, okay, let's not go in with the assumption that low-level physical causes are the primary causes of everything. It's a way of, if you like, taking intentional phenomena at face value—intentional in the sense of bearing representations. And I think one of my criticisms in that chapter is this agenda, which is so ingrained in philosophy of mind, to say that, okay, if representation is real, we need to be able to tell a physical story about how it comes about.

And this is actually going against a Dennettian view, as you might see it: if talking about representations and intentionality is useful within the sciences, why not just take that at face value and not say that that needs to be established by making it coherent with some causal story about what's going on in terms of nonintentional physical interactions? So that was the position there.

With Putnam's rock, right? He said that you can take any open physical system and configure it in such a way as to have the same types of information processing. Why wouldn't that have all of the cognitive properties?

When you're making a claim that the brain is a computer and that that explains cognition, what grounds have you got for saying that any arbitrary physical system actually implements a computation, just from looking at its physical dynamics? If it's purely a question of mapping the physical dynamics to a computational formalism, then any physical system can afford a mapping of that sort, whether it's a rock, whether it's a sofa, whether it's my stomach as opposed to my brain.

And so that's a challenge to the computational theory of mind: it's assuming that brains implement computations just because we can model them computationally. But we can model all kinds of things computationally. What makes brains special?

So what about this idea of whether computation itself has causal powers?

I don't think it does. Computation itself is a mathematical formalism. It exists; it's a mathematical structure. Things that have causal powers are concrete physical systems, so I just think they're different kinds of things.

So Searle famously argued that the reason why we can't build strong AI is that computation doesn't have causal powers. It's implemented in silicon, so what does have causal powers are the machines that actually implement the computation. But couldn't you sort of say, well, there is still a causal graph? Perhaps you would argue that computation isn't a node in that causal graph; it's just some kind of an aspect of it?

Yeah. I think it just goes back to this issue: computation in and of itself is not the kind of thing that could have causal powers. I think Searle's point—and this was in The Rediscovery of the Mind—was an interesting one. It was maybe kind of subtle, and it gets lost in the wash of the AI back-and-forth and Searle-bashing, which happens a lot. But it was about the kinds of ways that we form explanations in the sciences. His point was that cognition, if it's anything, is something that's part of the physical realm, the realm of causation.

The assumption of the computational theory of mind—and he argues that this is very dominant within cognitive science—is that you can explain this phenomenon, which is a phenomenon of the concrete physical world, through this noncausal thing, which is computation. Suddenly, there's no gap that needs to be closed. And I think that's a fair point: there's something here that needs further justification. Why, of all the things that happen in the concrete physical world that demand explanation, do we reach outside the concrete realm of physical causation into computation in order to explain this thing, cognition?

Another argument Searle was making was how machines couldn't understand, yeah? And of course he was talking about things like semantics. Do you feel that they could understand?

Yeah.

I think there's certainly more to human understanding than that. I think a thing about human cognition and animal cognition in general is that, in my view, it's not a set of discrete modules that work separately from one another. I think language is bound up with sensorimotor engagement, and likewise how we perceive the world is shaped by linguistic concept formation and everything like that. The idea that you could just detach a language faculty, have it replicated in an LLM that doesn't have the other bits of our cognition, doesn't have embodiment, doesn't have the capacity to engage with the world, and that it could have understanding in the same way that we do—I find that implausible.

Interesting. But again, if we do the galaxy-brain thing and say we can embed robots in the physical world, give them sensorimotor affordances and all the rest of it, there are many replies to the Chinese Room argument about this, like the robot reply. Would they have a little bit more understanding?

Yeah, I mean, that's getting more along the lines of the kind of thing that could have understanding. I think there's also more relevant stuff in the background of what it is to be a biological thing: some things are inherently meaningful to you and relevant to you because they're connected with the demands that the challenges of your environment place on you. Like I was saying before, life—being alive—is a way of being always in a situation that is problematic to you. So saliency and meaning, I think, are connected to that.

It's not to say that you couldn't have robots with more and more precarious, lifelike situations, and maybe—I don't know—it's not beyond the realms of possibility that then you start seeing understanding in ways that are more like that as well.

Let's talk a little bit about Heidegger. He spoke a lot about our relationship with technology, and I know this is something you've been thinking about. Can you tell us about that?

So one of the things that Heidegger said—and in many ways, he's a grandiose and unpleasant person, but his grandiosity kind of manifested in how he could relate the importance of philosophy to things about the modern world today—was that technology and cybernetics, for him—we could say AI for us today—was the culmination of a metaphysical tradition. It's because the history of philosophy set out on the path that it did that we are here today with these technologies, which seem to be having such a transformative role in our lives, to the point that we feel like we're not in control of them. I'm not a philosophical determinist like that. I don't think that just because some philosophers in ancient Greece said certain things, this is why we have AI today.

But I think there are some features of the contrast he draws with the philosophical tradition, in his own account of what it is to be a human person, that give us an interesting perspective on what is being assumed in that path towards AI. One of the things he really insists on is human finitude: we are inherently finite, bounded knowers, as individuals and as communities of knowers. But the philosophical tradition has encouraged a kind of leap—the idea that something about us as knowers crosses beyond the boundaries of finitude into a universal, boundless realm of knowledge.

I do think that the very idea that a nonsituated, nonembodied absorber of facts, like an LLM, which just sits there and sucks in all the information in the world, could somehow be a counterpart to how we know things as human beings, is an instantiation of this lack of acknowledgement of human finitude.

Human finitude means coming from a culture which is expert at some things but not other things, and whose knowledge is grounded in a discrete set of sensory experiences that are not accessible to other people. To say that that boundedness of knowledge is not inherent to what it is to be a knower—that a purely disembodied absorber of facts is what we are—I think reveals that kind of lack of acknowledgement of finitude that is there in the tradition he criticized.

We're kind of moving into this technologically embedded world, and it's changing our nature in some way. This tradition that Heidegger criticizes—from that path from philosophy to technology—is this aspiration to transcend embodiment, transcend materiality, to create for ourselves a leap into an almost spiritual world of pure information.

It's interesting how technology infrastructure is presented to consumers, people like me, as behind the scenes, immaterial, and not really connected with real-world constraints. It's the cloud. It floats above us. There's almost no cost to it. It's weightless.

That's not how technology infrastructure works, but it seems like that's what we'd like it to be. We'd like this idea that all of this information age that we have around us today is not connected with real-world stuff, like actually building computers, shortages of resources, consumption of energy, and all of those things. We like to think of it as disconnected from that in ways that it's obviously not.

I agree that it's not, because clearly we live in the physical world, but there's an apparent disconnection. There was the digital divide, of course, in the 1980s, and it was very difficult even to work and get a job if you didn't know about computers. Now, even to get a driving license, it's all done online.

The legal landscape is not physical anymore; it's virtual. Even the European GDPR regulations, as Floridi says, you're a digital entity and it's diffused and so on.

Facebook and whatnot. It certainly feels like we live our lives in the information world.

Yeah. And it's becoming more and more confusing in that sense.

So, by living in the information world, do you mean that the information that we get through devices, through this diffuse spread of technologically connected things, is more salient to us than our experience with a concrete, here-and-now situation?

There was that Jon Ronson book, So You've Been Publicly Shamed, basically saying that there's almost more controlling pressure in the social media world than there is in our physical world.

Sure, sure. Yeah. I think all of that is possible because human beings are imaginative creatures. We live in imaginative worlds. It's not so much that social media is always in our face—well, maybe it is, with people looking at their screens—but our way of thinking through our own life history and how we project and everything else means that there's a whole world that we're also constructing around that.

I think we're co-creating this digital world, and it doesn't work unless we're imaginatively and emotionally invested in it. My view goes back to this Kantian idea, which also goes back to this idea of human finitude: that there's something wrong with thinking that what you are as a knower is the kind of being that can float free of your environment and just regard it from above and take in all the information that's there as it is by itself, without your impact on the world.

My point is that we're not those kinds of beings. We only acquire knowledge through this arduous process of interaction, which means that we cannot claim to have that God's-eye, neutral view on things. It would be a mistake about knowledge and about ourselves as knowers to think that that is an aspiration that makes sense for us.

Is it possible for both things to be true at the same time? I'm not making any weird claims that we're not actually

really physically situated, but it's almost like we're increasingly disconnected in many aspects of our mental life.

Mhm. From

from our direct physical experiences.

Yeah. So I think that's true in terms of what we pay attention to. People look at their phones instead of looking out the window when they're on a train, not looking at the people around them. Our attention can go wherever the internet wants to place us.

But the phone is still a concrete physical device with little light-emitting diodes that produce photons in our retinas, right? So I think, again, it's about the imaginary around that. In our minds, this isn't just a device that's showing images right now; it's a portal to this other place where we can be disconnected or dissociated from where we are right now.

But I think that need to live in an imaginary, to be disconnected, is probably always there and manifests in different ways in human culture through fantasy, mythology, and all kinds of things.

We can debate whether it is imaginary or not. I interviewed Chalmers about his book Reality+, and he was talking about whether virtual worlds are real and stuff like that. Certainly with social media, it is still real. These people on Instagram are having real experiences, and I'm vicariously experiencing them through my phone.

So the question is whether that's a sort of truncation of our mental life or whether it's expansive. This is something that I think becomes a question of ethics and phenomenology—the part of philosophy that really studies experience and what we draw from it for its own sake.

There's always an opportunity cost. If you're looking at your friends, absorbed into one thing, then you're not absorbed in other things. Ethically, is it right to do that? I think that's really important.

We're running a big experiment on the next generation because young children nowadays spend a lot less time looking at people's faces than they used to. Will they be socialized in a way that will allow them to lead happy lives later on? We don't know.

Is that something you worry about?

I do. I do. Yeah.

Tell me more.

Well, just because it's a massive experiment. It's obvious from developmental psychology that young children are predisposed to pay attention to social interactions with the people around them—the gaze, faces, features, and all of that.

If they have less opportunity to make those connections, to have those experiences at a very young age, it seems like that's bound to have an effect on how they relate to people later on. I don't see how it couldn't.

There's a thought experiment I had in a debate with an AI doomer. Because I'm a big externalist, I think we're embedded in these cognitive ecologies, as you do. I gave the example: imagine that some child was brought up in a hermetically sealed chamber, and you gave them a computer and the internet. They could still learn to do lots of clever things and so on.

But we're becoming a bit more like that. It's not quite a hermetically sealed chamber, but we're mediating through the internet and so on. That could go one of two ways, right? As you say, it could dramatically truncate our mental life and cause lots of problems, or maybe it just won't be a big problem.

There were already experiments done on monkeys in the 1950s, depriving them of maternal contact, and it didn't work out well for those monkeys. Should we be doing that experiment on children—sort of depriving them of what is instinctively the kind of social engagement that young humans seem to need to develop normally?

Amazing. Mazviita, thank you so much for joining us today. It's been a great conversation.

Thank you very much, Tim. I've enjoyed it.