Your Brain Isnt A Computer And That Changes Everything

TITLE: Your Brain Isn’t a Computer and That Changes Everything CHANNEL: Curt Jaimungal URL: https://youtu.be/_kuwwmFnxGY?si=L3Oyo88dgYuo8qCA

---

I make the additional really weird claim that I don’t think algorithms capture everything
we need to know about life. We’ve forgotten that the idea of the brain as a computer is
a metaphor and not the thing itself. There’s no bright line between what it does and what
it is. That would not be what I would have predicted. This is a monumental theolocution.
For the first time ever, Professor Anil Seth and Professor Michael Levin are conversing
and performing research in real time, and we get to be the flies on the wall. Anil says that the brain as a computer metaphor has blinded us for decades. You can’t extract
the software from the substrate. This means that silicon consciousness may be impossible,
not because machines lack dualistic souls, though. But wait, Michael disagrees.
He thinks that machines may be able to access the same platonic space that biological systems tap
into. The magic isn’t restricted to carbon. Both professors are now building and studying xenobots
together. These are living robots made from skin cells that self-organize, exhibiting behaviors
evolution never programmed. Do they dream? Do they have preferences? Are they conscious?
On this episode of Theories of Everything, we explore their radical collaboration,
including questions like how split-brain patients may prove consciousness fragments and multiplies,
and the terrifying possibility that large language models are doing something
entirely different from what their output suggests. Tasks no programmer asked for,
no steps in the code demand for, but perhaps where that, quote-unquote, “magic” lies.
Remember to hit that subscribe button if you like videos exploring fundamental reality. All right, we’re going to talk about aliens. We’re going to talk about cyborgs, modules in the brain,
split-hemisphere patients, if I’m not mistaken, and unconscious processing. We’re going to get
to all of that. To set the stage, I’d like to know what’s exciting you both research-wise currently,
something you’re pursuing. So Anil, why don’t we start with you, please? Well, thanks, Curt. Two things, I guess. One thing, a topic that seems
to be exciting a lot of people these days, which is the possibility of AI
being conscious. Whether it’s something that AI systems can have or whether,
as I tend to think, that it’s something more bound up with our nature as living creatures. The other thing that’s exciting me actually just came to mind in your little list of topics there
is the question of islands of consciousness. So there’s a lot of work on things like split-brain
patients, patients with brain damage, and so on. But a question that me and a couple of colleagues,
Tim Bayne and Marcello Massimini, have been wondering is, are there isolated neural systems
that may have conscious experiences? And one candidate for this is called hemispherotomy,
which is a kind of neurosurgical operation where you have bits of the brain detached,
disconnected from all other parts of the brain, but you still have neural activity.
These parts of the brain are still part of the living organism. Are they islands
of awareness? So we’ve been exploring that theoretically and very recently with some
evidence from brain imaging of people following this neurosurgical operation. Michael? Yeah, so a couple of things on the experimental front. I’m really excited about some novel systems
that we’re setting up as compositional agents. So putting together different
living and non-living components using AI and other interfaces to allow them to not just
communicate with each other but, we hope, form a kind of collective intelligence. And then we can
ask some interesting questions about what kind of inner perspective this new intelligence might be.
Just in general, complementing the work we’ve done before around distributing and,
let’s say, separating out the different pieces, like Anil was just saying,
of the brain and so on. The flip side of that, which is putting together new kinds of beings
that haven’t existed before and asking what their behavioral complements are, what their
capacities are, what their goals are, their preferences, what do they pay attention to,
these kinds of things. And just in general, really digging into this idea of,
for lack of a better word, intrinsic motivations and asking in novel creatures that don’t have
the benefit of a lengthy evolutionary history that presumably set some of their cognitive properties,
where do these things come from? And how do we predict them? How do we recognize them? I should have said, of course, one of the things that’s really exciting to me is stuff
that Mike and I have been talking about together. About some of the systems that he’s building,
some of the things he’s actually talking about. Do they sort of self-organize in ways which seem
to obey the laws of psychophysics and other sorts of situations where we might attribute things like
intrinsic motivation to evolved systems? There’s a big question about, are laws of perception,
are they adapted to specific environmental situations or are they somehow intrinsic to
how biological neural systems self-organize? So we’ve been bouncing ideas back and forth
to do experiments to explore some of these questions, which is various. Tell us about some of these experiments. We haven’t done them yet. The logic is to take some simple observations of phenomena that are
very widespread in perception across many evolved species, whether it’s a human being, a mouse, or
probably a bacterium or something like that. So there are things like Weber’s or Fechner’s Law,
so the idea that the perceived intensity of the stimulus scales logarithmically with its actual
magnitude. Now this is something that seems very, very general. Is this something that
we can look for evidence for in some of these completely out-of-evolutionary-context systems
of the kind that Mike is generating? So that would be one example. There are a whole bunch
of other examples we have. Can we find things like susceptibility to visual illusions or things like
that in these systems? So what are the simplest kind of very general perceptual and learning
phenomena that we might be able to examine whether they happen in these systems which
don’t have this straightforward evolutionary trajectory? I think that’s the basic project. Yeah, exactly. And these comprise xenobots, anthrobots,
and even weirder constructs that we can start to put together by building
technological interfaces between radically different kinds of beings that allow them to,
sort of like an artificial corpus callosum that takes two different things and tries to
bind them into one novel collective thing and seeing whether some of their properties and
their behavioral competencies match the things that people have been studying, as Anil said,
psychophysics, and all the things out of a behavioral handbook, basically. Yeah, and this kind of relates to this idea of AI and consciousness and so on for the following
reason, which is why I think it’s very exciting. I think Mike and I might have different but
overlapping reasons for our interest in this. For me, there’s this sort of assumption by people who
talk a lot about consciousness and AI that the biological stuff we’re made of doesn’t really
matter. It’s just there to implement algorithms. Silicon could do just as well. And I tend to think
differently, and I think we both do. And so these are ways of just looking at what’s the dynamic and
functional potential that’s sort of intrinsic to the stuff we’re made of that provides the basis
for our cognitive, our perceptual, ultimately our conscious abilities and properties. So these are
experiments that we’re getting at that. What’s just there that evolution can then make use of? Anil, can you make the elevator pitch for people who are already familiar with the argument that,
look, the processing that’s going on in our brain is just processing. It
could potentially be translated to a computer. If consciousness
is similarly information processing, then we have something that’s, quote-unquote,
“substrate independent.” So you’re making the claim that it’s not so clear. Maybe
there is a dependence to the substrate. Can you make that case? And then also Michael,
I know that you have several questions you’d like to ask Anil, and feel free at any point to. I try and make it, but that’s the case I’m trying to make. It’s quite tricky because it goes against
such a deeply embedded assumption that the brain is basically a computer made of meat,
and the things that it does, the only things that it does that are relevant for things like
cognition and consciousness are computations, are forms of information processing. If you start from
that perspective, it leads you to this idea that there is this substrate independency. And what
that means, to unpack that, that just means that the stuff we’re made of doesn’t really matter,
it’s the computations that matter. And if a substrate can implement computations,
then fine. These two sort of ideas go together because one of the whole
motivations for a computational view is substrate independency. Turing’s formulation
of computation is formulated in terms of it being independent of any particular material. So the other thing really is that we’ve kind of forgotten that the idea of the brain as a computer
is a metaphor and not the thing itself. It’s a sort of marriage of mathematical convenience. And
the closer you look at real biological systems, as Mike’s work beautifully exemplifies, the less that
this idea of substrate independency makes any real sense. There’s no bright line in a brain
or a biological system in general between what you might call the mind-ware and the wet-ware,
between what it does and what it is. And if there’s no clear way to separate in a system
what it does from what it is, then it’s very, very much less clear that one should think that
computation is all that matters. Because for computation to be all that matters,
you kind of have to have this sharp separation between the software and the hardware,
what it does and what it is. If you can’t do that, then there’s less reason to think that computation
is what matters. And if there’s less reason to think that, then there’s equally less reason to
think that you could implement what matters in a substrate-independent way on something else.
You can, of course, still use computers to simulate a brain in whatever level of
detail you want. But that’s neither here nor there. It’s a very useful thing to do. We both
do this. We do this all the time. But you can simulate anything using a computer. That’s one
reason computers are great. But that doesn’t mean you will instantiate the phenomenon. You
only do that if computation really is all that matters. And I think that’s very much
up for grabs. I think it’s been a very deeply held assumption, but I think it’s likely wrong. Yeah, I mean, I agree with everything Anil said, but I take it in a slightly different direction.
So I think it’s critical to remember that, yeah, everything we think about as computation is a
metaphor. It’s a formal model. And so we have to ask ourselves, what does this model help us do
and what is it hiding? In other words, what is it preventing us from seeing? And I agree that
this metaphor does not capture everything that we need to know and we need to use to do technology
and so on about life. I think the computational paradigm and the notion of algorithms and so on
does not capture everything we need to know about life. But I make the additional really weird claim
that I don’t think it captures everything we need to know about machines either. In other words, we
tend to think, at least the people I meet tend to think, that we have this set of metaphors that are
for machines and their algorithms and they don’t really apply to biology. Certainly people say,
“Well, they don’t apply to me. I’m creative and whatever else they are.” But there is a corner
of the universe that is boring, mechanical. It only does what the algorithm says it should do.
And for those kinds of things, these metaphors are perfect. They capture everything there is to know. So I agree with Anil on the first part, but I doubt the second part. I think that a lot of
what we have in our theories of computation is a pretty reasonable theory of what I call the
front end. I think most of what we deal with are actually thin clients in a certain sense.
They’re interfaces to something much deeper, which we can call the platonic space. I don’t love the
name, but I say it that way because then at least the mathematicians know what I’m talking about.
But I think that even, and we have some work already published and more work coming soon in
the next few months on this, showing that, yeah, the standard way of looking at algorithms doesn’t
even tell the story of so-called machines. And so whatever it is, and I have guesses, but of course
we don’t know, whatever it is that allows a mind to come through biological interfaces
and not be captured by these formal models, I think these other systems that we call machines
and certainly cyborgs and hybrids also, I think they get some of the magic too. It’s not going
to be like us. It’s going to be different, but I don’t think they escape these aggressions either. This is why I think I find Mike’s work so interesting because it’s provocative in this
direction. I think he’s, I’ve always, I think he summarizes what I said very well is that
we underestimate the richness of biological systems if you force them into the, what’s often
called the machine metaphor, by which we really mean that all that matters is this sort of Turing
computation algorithm thing. But I think it is equally true that we limit our imagination about
what machines might be as well by doing this. And there’s a whole kind of alternative history of AI,
which was really grounded in 20th-century cybernetics. There’s much more to do about
dynamical systems, attractors, feedback systems, all things you can still simulate computationally,
which are fundamentally not arising from the algorithmic way of thinking
about things. There are also really interesting mathematical properties like emergence and so on,
which I think can both help us understand but also might be design principles for
machines of various kinds, which, again, don’t really fit into an algorithmic view of things. So it might as well be able to show that even something we think of as anonymously algorithmic,
like, correct me if I’m wrong, the bubble sort stuff. So this is an algorithm that anyone
in computer science 101 learns to code, to sort things into a particular order,
has really interesting emergent properties that other things can be built on top of. So,
yeah, I think for me it’s like there’s this nice iterative back and forth where
we can learn to think of both biology and machines differently. And, of course,
that might give us richer metaphors through which we can use one lens to understand the other. Would you say, then, that we have the idea of machine and
that a Turing machine is a strict subset of that idea of machine? I mean, a Turing machine is an abstraction, right? Turing machines were never sort of
supposed to exist as things. They have infinite tape and things like that. So
you’ve got a Turing machine. The idea is in one sense you’re mapping a bunch of numbers
onto another bunch of numbers. And then the universal Turing machine does this through
this moving head and an infinite tape. It was never really supposed to exist as a physical
machine. And I think that’s where part of the problem has sort of come from. But an
algorithm in that sense, yeah, I think that’s a subset. When you realize a Turing machine,
that’s a subset of possible machines. Yes, when you realize a Turing machine,
it will be a subset of all possible machines just because it’s a particular Turing machine. No, but
when you realize a universal Turing machine as well, that’s also a subset of possible machines. So if you don’t mind spelling out to the audience the idea of hypercomputation,
and would you then say that biological creatures
or cells or what have you are doing something that is hypercomputational?
And feel free to take this in a different direction, Michael, as well, if you like. Would you, yeah, I mean, I’d love to say that, but would you, Curt,
would you give me what you’re thinking of when you use the
word hypercomputation? I’ve heard it used to imply different things. So if something can solve the halting problem, it would be an example of a hypercomputer.
Something that can decide problems that a Turing machine or a universal Turing machine cannot. Right, so sort of super-Turing in some sense. That’s one way in which machines can be non-Turing
or can escape the Turing world. But I think there are many other systems that are just not captured
by this way. They don’t have to be based on a halting problem. Strictly anything that is
stochastic, anything that is continuous, is beyond this world of strict universal Turing machines.
There are kinds of extensions that try to go there. But there are also functions that things
do that necessarily involve particular material substrates. So take something like metabolism. And
metabolism is not mapping some range of numbers, whether they’re continuous or random, to another
number. It involves actual transformation of a particular kind of substance into another kind
of substance. That’s just non-Turing in what is a fairly trivial way. But that kind of thing
might be very important for particular classes of machines or systems, whether they’re biological or
not. So I think there are different spaces of what you might call non-Turing processes. Only some of
these are these kinds of hyper-computation, halting problem solving things where you
might say you’ve got some sort of fancy quantum stuff going on. But I think it’s different about
that. It’s different, right? I mean, there are some people that would say that actually, unless
you’re talking about this hypercomputation, in the sense you’ve mentioned, that everything else is
sort of a relatively feasible extension of Turing as is. So there’s definitely debate in that area. I would go in a slightly different direction. Emphasize something that does not lean on quantum
mechanics, does not lean on stochasticity, and does not lean on hyper-Turing or anything like
that. And also, let’s even step back from the living because living things, conventional living
things, are so complex that you can always find more mechanisms and so on. I want to look at an
extremely minimal model. And the reason that we chose this was precisely because it’s such
a minimal model. I wanted to sort of maximize the shock value of this thing for our intuitions. And
this is the work that my student, Tainan Zhang and Adam Goldstein and I did on sorting algorithms,
which is what Anil mentioned. And there’s a couple more things like it coming in the next few months. The sorting algorithm is, it’s like bubble sort, selection sort, these kinds of things.
CS students in CS101 have been studying these for, I don’t know, 60 years probably. And no one,
as far as I can tell, no one noticed what we noticed because the assumption has always been
this thing does what we asked it to do. And a lot of what I’m trying to emphasize is specifically
running against that assumption that, yeah, it sorts the numbers all right. But if you back
off from this assumption that all it does is what the steps of the algorithm ask it to do,
then you find some new things. And computer scientists are well aware of emergent complexity,
emergent unpredictability. Cellular automata do all kinds of funky things, and some of the rules
are chaotic and all this kind of stuff. That’s not what I’m talking about. I’m not talking
about emerging complexity, unpredictability, or even perverse instantiation, which people
find all the time. I’m talking about things that any behavioral scientist would recognize
as within their domain if you didn’t tell them that this came from a deterministic algorithm. And so I can go into details if you want, but a couple of things are salient here.
What these algorithms are also doing while they’re sorting your numbers are also
a couple of interesting, I call them side quests, because there are no steps in the
algorithm asking them to do this. In fact, if you try to write an algorithm to force them to do it,
it would be a whole bunch of extra work, which is actually quite interesting because I think
we’re getting free compute here. That’s a whole other thing that I think is a very testable,
it’s a nice testable prediction because it’s so weird and unexpected. They are
doing some other things that are not directly related to what you asked them to do. That’s
really important because it means that these language models, for example, when we say AI,
nowadays a lot of people think language models, people tend to assume that the thing that the
language model talks about is some kind of clue as to its inner nature, right? And people say,
“Well, you know, my GPT said to me that it was conscious or wasn’t conscious or whatever.”
My point is the thing you force it to do may have zero to do with what’s actually going on. Now in biologicals, that’s not true because evolution, I think, works
really hard to make sure that the signs that we and the communications that we do are related
to our inner state and things like that. So in biology, those things are tied closely together,
but I think we’ve disconnected them. And what we are now making are things that look like they’re
talking and whatever. And they are. But I’m not sure any of those things are at all a guide to
what’s going on inside. And if a dumb bubble sort, which is six lines of code, fully deterministic,
nowhere to hide, six lines of code, if that thing is doing things that we did not expect and we did
not ask it to do, and by ask, I mean, there are no steps in the algorithm to do what it’s doing,
then who knows what these language models are doing. But I’m pretty sure that just watching
the language output is not a really good guide to what’s happening. I think we have to go back
to the very beginning and we have to apply the kinds of things that Anil was talking about,
which is basic behavioral testing in various spaces. I think our imagination is really poor
at this. I think we have to just be really creative as far as asking what is this thing
actually doing, specifically in the spaces between the algorithm because the thing is,
it has to, it’s a little bit like this is a crazy analogy that I came up with the other day. This
notion of steganography. So in steganography, you take, let’s say, a piece of data, let’s say it’s
an image, it’s a JPEG, and it looks like whatever it looks like. There are bits within that image
that if you were to change those bits, it wouldn’t look any different, right? There’s some degrees of
freedom in there that you can move things around and the image would still look the same. And so
what people do is they hide information in there and maybe it’s your signature that you
are the one who took the picture or maybe it’s a code because you’re a spy to whatever it is,
you hide information in there. But the iron rule is you can’t mess up the primary picture. You
can sneak stuff into the degrees of freedom, but you can’t mess with the primary picture
or the primary data pattern because then it will be obvious that something’s there. I kind of have a feeling that this is what’s going on, not just with computer algorithms,
but with everything. There is the primary thing it’s supposed to do, and anything else that it
gets to do has to be compatible with that primary thing. It isn’t magic. You can’t
break the laws of physics. You can’t go against the algorithm, right? You’re not doing things
that the algorithm forbids, but there’s, but it turns out I think that there are these weird,
empty spaces between the algorithm where you can do, and I mean, doesn’t that describe to some
extent our existential, you have a certain bit of time in this world, you have to be consistent with
those laws of physics, eventually your physical body gets ground down to entropy and whatever,
but until then you can do some cool stuff that isn’t forbidden by the laws of physics,
nor is it prescribed by those laws, I don’t think. And so you get this,
this is what I think is really interesting about these things. And the algorithm itself,
to the extent that it has to do the algorithm, that limits what else it can do. In a sense,
what it’s doing is in spite of the algorithm, not because of it. And so I agree with Anil here. I’m
not a computationist. I don’t think anything is conscious because of the algorithm. If anything,
I think the mental properties it has is in spite of the thing we force it to do. And so one thing, and then I’ll stop here, one thing which I’m sort of most proud of in that
paper that I think was kind of cool is that we figured out a way to let off the pressure
on the algorithm a little bit to see what would happen. And the way you do that now, how would
you do that? It has to follow the algorithm. How could you possibly let off the pressure?
What we did was we allow duplicate numbers in the sort. And what that allows you to do is you still
have all the fives still have to end up before the sixes and that has to go before the sevens and so
on. But how you arrange those is now not really constrained by the algorithm. You don’t touch,
you don’t change the algorithm. You just allow multiple repeats within the, and what did we see?
We saw that the crazy thing it was doing, which I call clustering, I could tell you what that is,
but it doesn’t matter, it went up higher than when we didn’t let it do that. And so I really think,
and this comes back to the AI thing, I really think that it’s a lot like raising kids in the
following sense. To the extent that you force them to do specific things, you squelch down on the
intrinsic motivation. Some kid that’s forced to be in a class all day, you’re not going to get to see
what else he would be doing otherwise. Maybe he’d be out playing soccer, who knows what it would be.
And so to the extent that we force these things to do specific things, we are actually reducing
what else they might do. And that’s what we need to develop is the tools to detect and
to facilitate this intrinsic motivation. And then you get into alignment and all of that. That reminds me a lot of when I was doing my postdoc 20-odd years ago,
coming across, well, being told by my mentor at the time, Gerald Edelman, about the distinction
between redundancy and degeneracy. I think this is very apposite here. So, engineering people often
talk about having redundancy within a system. So if a system is designed to do something,
to follow some steps in an algorithm, well, then you might want multiple copies in case
something goes wrong, you have a backup. But the backup is doing the same thing. It’s redundant in
that sense. Biological systems don’t seem to be like that. They exhibit degeneracy rather than
redundancy. That is, they may have multiple ways of doing the same thing in context A,
but in context B, these multiple ways of doing the same thing now do different things. So this
is hinting at the same thing, that although it looks like they’re doing the same thing,
there’s actually some spaces in between somewhere that you won’t see unless you look in different
contexts. Otherwise, you’ll only see the same process that might look like an algorithm. And
it’s that degeneracy that gives biological systems their kind of open-endedness, their ability to
adapt to novel situations and so on. And it might be related to what Mike is calling an intrinsic
motivation that you have to have some kind of degeneracy rather than redundancy to systems. I mean, what’s interesting to me is that people are often, with the exception of
a few like diehard, reductionist, materialist, whatever, people are generally pretty willing
to grant living things that, right? And they’re okay with saying that living things,
especially brainy living things, get to do some of those things. But what I’m now finding is that
people get very upset when I suggest that the same thing might be true all the way down. It seems to
be very important that we have this distinction. No, that’s the dead matter and the mere machines.
We are special. We can do this thing. And my point is not, I’m not trying to mechanize living things.
I’m going in the opposite direction. I’m saying there’s not less mind than you think there is.
I think there’s more. But actually, especially people, organicist thinkers who really resist
the mechanization of life and all this stuff, they really get really upset. They really get upset at
this last part because if, I suppose we’re not as special if it goes all, I’m like, I’m not sure.
I think there’s some kind of a scarcity mindset that there’s just not enough mind for all of us. Maybe. I think it might be that there’s still this worry that even if you’re,
like say BubbleSort again, I mean, BubbleSort is still implemented on standard computers,
right? So one way of potentially misunderstanding what you’re saying is you’re then basically
allowing computational functionalism by the back door again in some ways by saying, look,
an algorithm like bubble sort has actually all the things that you need, or it has so much more
going on than what one might think. So let’s not be too quick to rule out substrate-independent
algorithms as sufficient for other things that might seem otherwise hard to explain. Well, I think you’re right and I think people could, but that would be a misinterpretation
of what I’m saying. I am not saying that it’s doing that because of the algorithm,
right? So the standard computationalist theory is you are conscious because your algorithm is
doing workspace theory, whatever, whatever it’s doing, right? That’s why you are. I’m
saying the exact opposite. I’m saying that even something as stripped down and forced,
a stupid algorithm, there are still spaces there through which whatever this is that,
that I, that this, whatever, whatever this magic is that we’re talking about is, is able to squeeze
in. Even there, there are minimal versions of it that will shine through even there.
And if you provide a different interface, and I don’t want to just say more complex
because I don’t think it’s just complexity. Maybe it’s materials, maybe it’s some other
stuff. But if you provide better interfaces such as living materials, well, then sure,
you’ll get way more. But this stuff seeps into even the most constrained systems, I think. So let’s get to aliens, Michael. I don’t know what to say. People email me sometimes asking to talk to my alien
handlers. There’s that. But I don’t know anything about aliens other than
to say that it seems implausible to me, not being an expert on exobiology or whatever,
it seems implausible to me that the only kind of life is the life that we’re familiar with
here or cognition. I expect that elsewhere in the universe, there will be extremely alien forms of
mind that are not carbon, and I mean, I can get even weirder, but not the kinds of things that
we’re used to here. I think our imagination is terrible for that kind of thing. I mean,
sci-fi does okay sometimes. But yeah, anything that’s tied to the specifics of life on Earth
I think is almost certainly too narrow as a criterion for these kinds of things. I mean, I’m just always, I go back to the Fermi paradox and like, you know,
where is everybody? But, and which always worries me because it just sort of suggests to me that
I think it’s also very implausible that we’re the only example of life,
but then the evidence for intelligent life that has been able to broadcast structured energy out
into the universe seems lacking. Where the hell is everybody? So, of course, the conclusion from
this is that life might be very prevalent in many places, at least certainly not only here,
but that it’s quite difficult to get life to the stage where it lasts long enough to persist and
become cognitively sophisticated. I have no idea and I find that existentially concerning
and just a great sort of shaker of the snow globe for reminding us that we really need to take care
of our own planet and civilization first because it might not be very common to get to the kinds of
things we are, even if it’s exotic in a different way somewhere else. I think the universe is much
more likely to be filled with gray goo than Mike Levin with eight legs in octopod form. So Anil, if I was to take your cells and put them into a dish, some would form xenobots,
some would die, most probably would die, and some may just wander about or what have you. Have you
become multiple agents at that point? Or were you always multiple agents pretending to be one? I don’t think pretending to be one. I think it’s an excellent question. I don’t really,
whether you can have multiple kind of coarse-grainings of agency simultaneously,
I think is quite interesting. I don’t see why not, in a sense. I think there can be sub-organismic
levels of agency in my constituents, but there’s something sort of enslaving of these finer grains
of description in things like organisms. Things pull together, the parts pull together as a whole
in a way that doesn’t happen if you dissociate me into my constituent cells. So I don’t,
yeah, I don’t see a contradiction between cells having agency and an organism having agency and a
society having agency and perhaps a global society having some kind of agency. These things can all
coexist and have a reality simultaneously, but they will affect each other. So agency
at a macro level will probably constrain the agency that’s available at the micro levels. And you have a book on consciousness which I’ll place on screen and a link
in the description right now. So you’ve probably heard of the identity theory of
consciousness. My understanding is that it just says mental states are simply the
same as physical states. They’re not caused by, they’re not emergent from,
they’re just identical to them. What do you make of that? I’m curious for both of you. Well, I don’t think it’s a theory. I think things like identity, in quotes, “theory” are
more metaphysical positions than actual theories. And for me, I like to wear metaphysics lightly,
if at all. I don’t think you get very far. To say that a mental state or a conscious state is
identical to a physical state, I mean, who knows? In some sense,
it might be trivially true. In another sense, it might be absolutely completely wrong. But what I
do think is it doesn’t give you anything in particular to do or anywhere to go.
So instead of sort of arguing about whether theories like that are correct or incorrect,
I prefer to ask whether they’re useful or not useful. And I don’t think identity theory is that
useful. I’m broadly a pragmatic materialist, which is to say that I’m pretty convinced
that conscious states have something to do with physical stuff. And we certainly know empirically
there are correlations and causal relations between, if you do something to the brain,
something will happen in conscious experience, at least in human beings. Who knows, maybe
consciousness is more general than biological systems, but I think pragmatic materialism is
a productively useful thing to do and we can go about the business of trying to explain properties
of consciousness in terms of properties of biological systems and we’ll see how far we get. And this depends on them. We have to face the question of what are the properties of biological
systems that give us explanatory predictive grip on properties of consciousness. For a
bunch of people, the assumption is it’s just the computations that bring us back to the early part
of the conversation. But there could be many other things that actually give us explanatory
and predictive grip about consciousness that aren’t the computations. And that’s
the view that I’m interested in exploring, and we’ll see whether it’s useful or not. Yeah, I agree with that. I mean, it sounds, it’s I think it’s less than a theory than it is a
linguistic claim. It’s, you know, you’re just saying something about the definitions. I find it
kind of unhelpful. It’s a little bit like saying that airline ticket prices, what are those? Well,
let’s associate them with some physical states. And well, what explains them? Well,
the constants at the beginning of the Big Bang plus some randomness. Like, in a certain sense,
kind of. In another sense, like how much insight are you going to get as far as why these prices
are going up or down if you have this view? I think probably zero. And so like Anil,
I’m interested in metaphors and I think that all these things are metaphors,
but I’m interested in metaphors that help us discover new things. And I don’t see how equating
them linguistically with physical states is doing the trick. I don’t think that works in biology for
the sort of cognitive non-consciousness specific things, and I don’t see it helping here either. Mike, you said you had some questions about split-hemisphere patients for Anil. Well, I don’t know what’s, okay. It’s not so much specifically about split-hemisphere patients, but
I guess it’s the thing I brought up in email. I was just wondering, I was listening to a talk,
I forget whose talk it is, and somebody was saying, look, there are all these unconscious
processes during reading, during the driving, whatever. There’s all these unconscious processes.
And I was just curious what you think about that, because it seems to me critical to say, conscious
to whom? In other words, they might well be unconscious to the main left hemisphere, whatever,
that’s verbally reporting this and saying, “Wow, I drove all the way from home to my office and I
wasn’t conscious of any of that.” And so you say, okay, great, there’s this unconscious part. Well,
it’s not conscious to you, but neither are my conscious states conscious to you. So how do we
know, right? So that all of these things aren’t the subsystems of the brain and mind that execute
them, how do we know they don’t have an experience they can’t verbalize? So I was just curious about
that because it seems like it’s just a foregone assumption and it seems like really begging the
question if we don’t, if we just assume that because you don’t feel them that they don’t,
and it’s the same, and the reason it’s of interest to me is that that’s what people say about our
body organs too, right? So I make a claim that for the exact same reason we give each other the
benefit, reasons, four or five reasons that we give each other the benefit of the doubt about
consciousness, you should take that seriously about your various body organs. And people say,
“Well, I don’t feel my liver being conscious.” Of course not. You don’t
feel me being conscious either. So I was just curious what you think about that. Yeah, we had, we had just, for the people listening, we started this nice dialogue
by email just a couple of days ago. So I think it raises some really important questions about
how we use the words. Unfortunately, I do think it’s a little bit linguistic here. We talk about
the conscious and the unconscious. And of course, they mean different things in different contexts.
So when it comes to, let’s say, split-hemisphere patients, the intuition is there are two separate
conscious agents, just only one of them has the ability to behaviorally report through language
what it’s experiencing. But it’s partly because each hemisphere has kind of the full complement
of resources that one might think of as necessary that this becomes a sort of plausible position. Then there’s other uses of conscious versus unconscious. There’s a whole history of,
a lot of the history of consciousness science is trying to contrast conscious
from unconscious perception. So, you know, you’ll show an image and somebody will say,
“Yeah, I see it.” And then you mask it in some way, manipulate it in some way. And people say,
“Oh no, I didn’t see it.” But you can still see parts of the brain responding. And the logic is,
well, the contrast that you get there between when something was consciously seen and the same image
or the same sound was not consciously experienced. If you look at the difference in the brain,
that difference has to do with consciousness. That’s the whole strategy of looking for the
neural correlates of consciousness. But then you might ask, well, how do you know that the
unconscious perception was in fact unconscious? It may have just been unconscious to the subject
as a whole. There may have been an inaccessible conscious experience happening. So I think this
is logically perfectly possible. But then you have this whole, well, how do you then link that
not only to a brute correlation, but you have to then come up with some theoretical reason,
and that will depend on your theory. A theory like global workspace might say, okay, look,
the reason that the conscious perception was reportedly conscious was because it engaged
the global workspace. And the theory is that things are conscious in virtue of accessing
this global workspace. So you have some sort of theoretical reason for saying that the
unconscious is in fact unconscious. But of course then you risk a little bit of circularity, right?
That your evidence for global workspace is based on the theoretical explanation that
makes one conscious and the other not conscious. So you have to have multiple sources of evidence. All this to say is it’s a very good question, and it came up in the thing I mentioned right at the
beginning. We have these hemispherotomy patients whose parts of their brain are
completely disconnected. So they by definition can’t respond to things. They can’t generate any
response. They’re sort of the opposite of language models in this sense, right? They
can’t give us any persuasive behavioral evidence because they’re not connected to anything. Yet,
they are part of a brain that was at one point conscious. And all that’s happened really,
in the limit, they’re damaged as well. I mean, there’s other things going on,
is they’ve been disconnected. So plausibly, at least for me, they’re more likely to be conscious
but inaccessible, much more likely a priori than a language model is to be. And so we have to find
indirect ways of trying to assess the likelihood of consciousness in these very disconnected
hemispheres. And to cut a long story short, very short, because I know you’ve got to go in a sec,
Mike. When we look at EEG, and this is work done with colleagues at the University of Milan,
it looks like these isolated hemispheres are in states of very, very deep sleep. So we see slow
waves, very prominent slow waves, sharp spectral exponents. But how do we know that that is in
fact unconscious? Because there are a few examples of human beings where we actually see slow waves
at the same time as consciousness, in DMT for instance and things like that. So it’s iterative,
it’s very hard to be definitive and it’s an excellent question. I think we don’t know,
until we start looking at systems radically different from a psychology undergraduate looking
at a monitor, which we still do and that’s very useful, but we have to look at these other things
as well. We don’t really know what assumptions we’re making when we interpret the data from,
just look for the car keys where the light is, you might miss the bigger picture. Now Mike, before you get going, I suppose I gave
you both unlimited resources to design some experiment. What would you create? Boy, I fundamentally I think we need an environment, a closed-loop environment in
which to exercise all kinds of, the xenobots and anthrobots are just the beginning,
there’s so much weirder things that we’re looking into, such that we might be able to
recognize new kinds of cognitive preferences, goals, competencies, whatever, to which we’re
otherwise blind. And I mean, you could imagine making this thing enormously rich and complex. Anil? Well, I mean, obviously, fun with Mike would be the thing to do. But other than that, I think if
you think about where the adjacent possible progress might be most rapid, what we lack,
what we’ve lacked in neuroscience is the ability to look at high resolution in time and space and
across much of the brain at the same time, measuring from many neurons in
time and space at the same time, in systems that we know are conscious,
or very high primates and other things. And there are just massive advances now, I think,
in invasive neurophysiology, in different kinds of neuroimaging methods that we can sort of,
optogenetics being one of them. But I think really doubling down on manipulation and recording and
high space, time, and coverage simultaneously, coupled with the development of new mathematical
tools to understand these kinds of complex data sets. That’s where I’d go. Lots to do there. Many of the people who watch this podcast are specialists in computer science, math,
physics, philosophy, adjacent fields, consciousness studies, neuroscience,
of course, cognitive. But also, many are not. Many are artists. For instance,
when I was at this MIT event, I’ll place a link on screen and in the description, there were many
people who were painters and poets and so on who came up to me. So I was going to ask just about
advice for researchers, but you can frame it as advice to everyone. What advice do you have? I think for students it’s super important to kind of curate your curiosity. I think,
I mean, I started with this very general curiosity in consciousness,
but then I think it was important to allow that curiosity to find other branches that then end
up coming together in different ways. I got very interested in other things too, in cybernetics,
in things that at the time didn’t seem to have much to do with this big question. But a lot of,
I think one way to carve out a successful career is to put different pieces together, to gain
skills that are both techniques, methodologies, but also conceptual toolboxes too that you then
can reassemble in different ways that other people might not have had the opportunity to do so.
So really, they’re two interconnected things, which is don’t lose sight of the big picture of
what you want to do, but also be flexible and try and develop curiosity in adjacent things
that might come in handy. And also learn to do stuff. I think many advances in science
have come about through advances in methods first. And if we learn methods, we will learn
the right questions to ask. And I think that’s maybe the thing that I’m still trying to learn
to do as a researcher, which is the thing I find really hard. It’s finding the right questions,
not finding the answers to the questions that you have. That for me is still the real struggle. Can you give an example, one of a method that you wish you had, for instance? It could be
that you wish you had learned earlier in your career or just a general example of something
that would be beneficial to a student. So a method. And then also you mentioned that
asking questions. So then also something, an example would be, well, what’s something where
you were pursuing the answer but you realized that it should have been a better question? So I’ll try and give examples that connect both of these things. So an example of something that
I wish I had gained some expertise in earlier is psychophysics. This is the
standard experimental thing. I caricatured it a bit early, undergraduates sitting in front
of a monitor pressing buttons and so on. But the methods of psychophysics are probably the
longest established experimental methods of studying consciousness. How do we interpret
data for people pushing buttons when you show them things? I mean, it’s very simple,
but there’s a huge amount of literature that goes back to the 19th century there.
And I made, I think, a ton of mistakes and certainly a ton of inefficiencies kind of
improvising my way through this literature or through my own work because I hadn’t
gained the skills early enough. So that’s one example of something I wish I’d done
differently. I think it would have allowed me to ask better questions experimentally. The thing that I think went well was I picked up, I learned to train myself and then asked other
people to help me learn information theory and Granger causality modeling. This is a mathematical
sort of framework for understanding information flow, causal interactions between nodes of a
network in complex systems generally. These methods were mainly used certainly at the time in
the early 2000s when I encountered them. They were primarily still used in economics, econometrics,
not in neuroscience. There were a couple of papers basically saying, “Hold on a minute,
we might be able to look at, apply these methods in neuroscience.” And I just got curious about
that. Not because I thought there was a big clue to consciousness there, but I thought, “Hold on,
that’s really interesting.” People assume that they look at coherence or mutual information or
correlation between brain regions, but might not be interested in causal information flow, lines
with arrows that are not going both directions. So I was lucky to know people who could help me
learn this stuff, and it’s become quite a strong part of what I’ve done over the years. Now working
with mathematicians who know this stuff much better than me, but we’ve done a lot in applying
these methods in neuroscience now and giving people the tools to apply them for themselves. And it’s also fed back into other things to ask. This is the other example, so different questions,
right? So one question that I’ve been asking for years, and I think it’s getting some wider grip
now, and again, this is largely thanks to collaborations with mathematicians,
is emergence. So this came up a little bit in the conversation with Mike. People talk about
emergent properties and so on, and often it’s a sort of placeholder magic for things that
we don’t really understand. But actually, I think there’s ways to make quantitative sense,
to measure emergence, to characterize it, to identify it in a data-driven way from systems.
And the mathematical toolbox of information theory and Granger causality has actually
turned out to be very useful in figuring out how to do this, to come up with measures of emergence
that allow us to ask questions about emergence in a more quantitative and operational way. I remember you and a few other people had a paper on this within the past two years or so, correct? That’s right. I’ve been working actually with two different groups of people on two different
approaches. The main one is with my colleague Lionel Barnett, who I’ve worked with for many
years now, who’s a mathematician. And we have a, so the story there was I actually wrote a paper
on this 15 years ago using Granger causality to measure emergence and I was very pleased
with myself at the time. I thought this is great, here’s this concept and here’s a way to implement
it mathematically. And it kind of got a bit of attention but not much. And then Lionel pointed
out to me that it was basically flawed in all sorts of ways and came up with a related idea that
does something much more rigorously. And it’s a slightly different thing, and we’re still working
on it to figure out how to extend it. But it’s mathematically a much more serious enterprise now. But what it does basically it says, okay, you’ve got a complex system. An example that’s often
used is you have a flock of birds. There may be birds flying around in the sky and sometimes it
looks like they’re flocking and other times it doesn’t. Can you quantify that? And of
course you could say, well, it’s in the eye of the observer. Fine, it’s in the eye of the observer,
but so is everything, really. There’s still a difference between a flock and a non-flock,
and if we can quantify that and generalize it, so maybe there’s something about neurons
that have an essence of this flockiness, but maybe not now in space in three dimensions,
but in some other dynamical space, in some other dimensional space. And the approach to this that
Lionel and I took was to come up with a measure we call dynamical independence, which is when a
sort of zoomed-out level of description, a coarse-graining, as physicists like to say,
a higher level of description of a system, if that is, if its evolution over time is
statistically independent of what its constituent parts are doing, then it in some sense has a life
of its own. Then it is in some sense emergent, dynamically independent. And it turns out that
the utility of this approach is that we can apply it in a purely data-driven way without
making any presuppositions of saying, “Oh yeah, there’s a flock, is it emergent?” We can just
identify potential emergent properties in a system and see how they look in different states. And just to, where we’re at right now for me is a hugely exciting thing actually, which is that
often people say, well, conscious states are emergent from their neural underpinnings. The
brain is in some sense, a conscious brain is in some sense more than the sum of its parts.
That all sounds very nice and I’m sure I’ve said stuff like this many times before. But
now with the tools that Lionel developed and applied with a PhD student of ours who’s also
working with others in Paris, Thomas Andrillon, we find something quite different actually,
which is that when the brain is in a conscious wakeful state, there’s less prominence of these
so-called dynamical independent coarse-grainings than when the brain is unconscious in anesthesia,
which is sort of not the, the slogan would be a little bit not what we were expecting.
Emergence is lower in consciousness than in unconsciousness. That would not be what I would
have predicted a few years ago or even two years ago, one year ago, I’m not sure. But it’s looked,
when we operationalize emergence this specific way and with this specific data, that’s what we find. But then that raises other interesting questions. And I think this is the beauty
of actually operationalizing these things, making them quantitative, because now we have another
set of questions which is like, ah, maybe this is because in the conscious state actually when you
don’t have emergence in the way we’re quantifying it, what you actually have is something called
scale integration, where there’s actually what’s happening at the macro and what’s happening at
the micro are much more interdependent. There’s much less separation of scales. And this takes
us right back to what we were talking about with Mike and indeed the whole idea of conscious AI
that I said right at the top, that in brains, there seems to be, it’s harder to separate what
they do from what they are. In a sense, this is a way of quantifying that hardness. And it seems
when the brain is conscious, it’s even harder to separate what it does from what it is. You
have this deeper integration of scales vertically, not across time or across space, but across levels
of description of a system. And so for me, this is opening like a whole range of questions that
haven’t really been asked. Certainly, I haven’t asked them before. It’s a different way of looking
at a system like this. And it all turns on having this mathematical method available. And for me,
that goes right back to the serendipity of being curious about Granger causality 20 years ago. Now there’s some research that says that when one takes psychedelics,
it probably depends on the psychedelic, that the brain is less active even though
your conscious experience, quote-unquote, is greater somehow. Is this related to that or
have you not studied emergence when it comes to the brain under psychedelics? It’s a little related. So we have a little bit in collaboration. We don’t have the license to
collect our own data under psychedelics, but we’ve collaborated with people like
Robin Carhart-Harris and others who have. And we have not yet, but this is very much on the cards,
we have not yet applied this same measure that I was just talking about to the psychedelics data,
but there’s no reason we can’t. What we have done is we’ve applied other measures that have often
been used in things like sleep and anesthesia as well that measure what we call signal diversity.
And the story here is that when you lose consciousness, your brain activity seems to become
more predictable, so the repertoire of states that it inhabits is lower. And this is measured using
this quantity we call Lempel-Ziv complexity. It’s sort of the compressibility of a signal. And the
complexity is lower when you lose consciousness. Your brain dynamics are more compressible, they’re
more predictable. When we applied this, this was now nearly eight or nine years ago, to data from
psilocybin, LSD, we found the opposite, that the brain activity became even less predictable. So
more diverse, more different patterns, less compressible, higher levels of complexity. So
that’s one clue, but to me, it’s still very preliminary. This method of measuring signal
diversity is quite precarious. It depends. If you do it a different way, you tend to get different
results. But I think there are other things we looked for we didn’t find in the psychedelics
data set. I was expecting to see, for instance, just much greater information flow from the front
of the brain to the back. I thought that might explain the prominence of hallucinatory contents.
We didn’t see that, at least not in the analysis that we did at first. We didn’t see any change in
information flow in that way. So I don’t know. I mean, there’s a lot to be done, but I think
that certainly just looking at overall levels of brain activity, to say it’s less active or more
active is not going to give us the answers. We need to look in more sophisticated ways. Now, lastly, speaking of surprise minimization,
what else has surprised you lately in consciousness research? What has surprised me? I mean, yeah, we can put it aside, but I think the thing that
surprised everybody, this is only tangentially related, is how simultaneously impressive and
unimpressive language models are. They’re really very different from how I thought they would be.
They can do a lot more, but they also have sort of still bizarre failure modes and so on. So I
just would not have expected the trajectory of language models to be as salient as it has been.
That’s certainly been a big surprise. What else has been surprising? I don’t know. It’s a really
good question. I’m not sure anything massively stands out to me. I’m sure something will come
to mind as soon as we finish this conversation. As it does. There have been other things which
have turned out kind of in ways that one might have expected. There was this huge
adversarial collaboration between integrated information theory and global workspace theory,
this big effort to compare these two big theories of consciousness. And of course,
that’s turning out that there’s evidence for and against both and there’s no decisive blow
against either. And that’s probably exactly what one would have expected,
though there’s still a lot of interesting and surprising things there in the details. But yeah,
I don’t know. There’s lots of things that are, I would say, small-scale surprising. It’s like,
“Oh, I didn’t expect that experiment to go this way or that way,” but I can’t think of anything
massive. The AI thing is sort of dominating my surprise minimization landscape at the moment. Thank you both for spending so much time with me and the audience. Thank you so much. Yeah, much appreciated. Thank you, Curt. Thank you, Mike. See you both. Yeah, see you.