Listeners:

Top listeners:

00:00 00:00

Go to album

Paul Verschure on consciousness and distributed adaptive control CSN Podcasts
Edvard Moser on grid cells and entorhinal cortex CSN Podcasts
Giacomo Rizzolatti on mirror neurons and action understanding CSN Podcasts
Robert Axelrod on game theory and prisoner's dilemma CSN Podcasts
Adrian Owen on disorders of consciousness and vegetative state CSN Podcasts
Jonathan Whitlock on markerless motion capture and posterior parietal cortex CSN Podcasts
Luis Puelles on neuroanatomy and prosomeric model CSN Podcasts
Zoltan Molnar on subplate neurons and cortical development CSN Podcasts

Joscha Bach on artificial general intelligence and deep learning

Episode 6 15.03.2019

PLAY EPISODE

Season 2019

0 Followers 13 Episodes

Description

Can we build a mind, and if so, what would that tell us about who we are? AI researcher Joscha Bach argues that the path to artificial general intelligence runs through understanding the mind as a model-making system in the service of organismic regulation , and that current deep learning, while surprisingly powerful, has not yet solved the fundamental problems of grounding, abduction, and epistemic autonomy. Subscribe for more from the Convergent Science Network podcast series. Joscha Bach joins Paul Verschure and Tony Prescott for a wide-ranging debate on the nature of intelligence, the limits of current AI, and what it would take to build a system with a mind similar to our own. Bach frames intelligence as function approximation , the ability to identify meaning by discovering relationships between patterns , and sketches a progression from hand-coded algorithms through learned functions to meta-learning systems that discover how to learn. He argues that our brains are not merely learning systems but meta-learning systems, and that evolution itself can be understood as an unprincipled search for such architectures. The conversation becomes a genuine intellectual sparring match. Verschure challenges whether the recursive logic of meta-learning constitutes real progress or merely demonstrates that its proponents understand recursion. Prescott questions whether intelligence-as-function-approximation captures the full range of human cognitive abilities. Both push Bach on the epistemic autonomy problem: current AI systems learn brilliantly on human-curated data but cannot ground their knowledge independently in the world. Bach concedes that new classes of algorithms , particularly for abductive reasoning and scientific discovery , are likely needed, while maintaining that no one has proven the limits of current approaches. Key topics include why Marvin Minsky’s commitment to symbolic AI set the field back, how AlphaGo’s success reframes expectations about machine intelligence, the difference between intelligence, smartness, and wisdom, why consciousness might be understood as a model of attention, and whether the current wave of deep learning can carry us to general intelligence or represents a fundamental dead end. Part of the Convergent Science Network podcast series from the BCBT Summer School.

Tagged as:
artificial general intelligence Build Mind deep learning General Intelligence Meta-learning Whether

About the author

CSN Podcasts

Both the triumphs of humanity and its most evil deeds have resulted from collaboration. In a time where humanity is required to aspire to the former and minimize the latter, the question arises of how collaboration arises and why it fails. Surprisingly, this phenomenon, so central to who we are, is not well understood. Hence, a collaborative effort is required to understand collaboration in its full biological, psychological, sociological, cultural, and economic complexity and to translate this understanding into operational impact. This series of podcasts is one step toward achieving these complementary goals. The Collaboration Podcast presents interviews with people who are central orchestrators of collaboration in various domains including business, government, science, art, health, sustainability, and the military. The discussions were conducted by Prof. Dr. Paul F.M.J. Verschure and members of the Program Advisory Committee of the Ernst Strungmann Forum on Collaboration (https://www.esforum.de/forums/ESF32_Collaboration.html) during 2021 and had the goal to sketch a map of opportunities, challenges, and obstacles in human collaboration. The forum took place in May 2022, and now we would like to share this series of interviews with a broader audience. The full report of the Forum will be published in 2023 by MIT Press. The podcast was produced by the Convergent Science Network (https://www.convergentsciencenetwork.org/). Context: The stability of social systems depends critically on realizing sustainable methods of “collaboration,” yet how and by which means collaboration is achieved is not clearly understood; neither are the conditions or processes that lead to its breakdown or failure. Collaboration can be understood as cooperation between agents toward mutually constructed goals. Part of the reason for our lack of understanding is that the phenomenon of collaboration is, by nature, a highly multidisciplinary problem, and effective research into its complexities has been difficult to achieve across the broad range of scientific and technical disciplines involved. The need for a fundamental understanding of collaboration, however, has become increasingly important. Not only does humankind demand answers as it attempts to address critical challenges at multiple scales (e.g., climate change, migration, enhanced automation, social and economic inequality), but ever-increasing technological and economic means of interconnecting people and societies are disrupting long-established, familiar patterns of how we interact. Radical technological changes that are ongoing have the potential to reshape collaboration in ways that are currently hard to predict or influence (e.g., by altering configurations in interaction, information creation, and modes of communication). On one hand, such changes could disrupt hitherto stable forms of collaboration by affecting critical communication channels and traditional roles, as can be observed in the rapidly changing patterns in governance, commerce, and social interaction. Conversely, technology could lead to the emergence of novel, successful forms of collaboration that deviate from traditional “hierarchical” architectures. Evidence of this can be seen in areas as diverse as highly automated manufacturing plants, the open science movement, collaborative software repositories, user-centered services, and the sharing of economy-based modes of organization. Without a fundamental understanding of the mechanisms, processes, and boundary conditions of collaboration, it is not possible to evaluate or predict which of these possible scenarios are sustainable or even plausible. The Forum “How Collaboration Arises and Why it Fails” (May 8–13, 2022, Location: Frankfurt am Main, Germany) Chairs: Andreas Roepstorff and Paul Verschure Program Advisory Committee: Jenna Bednar, Julia R. Lupp, Bhavani R. Rao , Andreas Roepstorff, Ferdinand von Siemens, and Paul Verschure

Timestamp

00:00:03 - This is the Convergent Science Network podcast. Leading researchers in the domain
00:00:10 - of neuroscience, brain theory and technology are interviewed by Paul Verschure and Tony Prescott.
00:00:20 - Okay, this is Paul Verschure for the Convergent Science Network podcast.
00:00:25 - Podcast together with colleague tony prescott and we welcome joshua who is a
00:00:31 - speaker at the bcd summer school in 2018 where you have two lectures actually
00:00:36 - and in some sense they're all.
00:00:39 - Turn around i think the general question of how we can build an artificial general
00:00:45 - intelligence i feel this is really i think that the core concept that you're
00:00:48 - or the challenge that you're that you're dealing with, right?
00:00:52 - So what's an artificial general intelligence in your mind?
00:00:58 - It's a system that has a mind similar to our own and beyond.
00:01:02 - I suspect that our minds have a particular kind of generality.
00:01:06 - And of course, you cannot really know this, whether we are universal functional approximators.
00:01:11 - But if it turns out that we can build AI similar to us, that's a strong indication
00:01:16 - that we are in that class of systems, right?
00:01:18 - Right. So the whole enterprise is spawned by the question of who we are,
00:01:22 - what is our relationship to reality, what's our nature, what is consciousness,
00:01:27 - why are things happening to us, what does it mean that something is happening to us.
00:01:30 - And we realize that psychology has become somewhat unproductive and philosophy
00:01:36 - has become even more unproductive as a field because you need to perform experiments
00:01:40 - that you cannot actually perform in psychology and that you never performed in philosophy.
00:01:44 - Philosophy so we now can build computer models and implement our ideas and usually
00:01:51 - our ideas will not work which will force us to update our theories and it's
00:01:55 - the way we can make progress.
00:01:57 - So, okay, so this is a bit the outlook, but to get there, you saw different
00:02:04 - stages in which then this search for AI has evolved, right?
00:02:08 - And so we go from our symbolic AI,
00:02:12 - this is sort of version one or first order, to an AI that's more like the current
00:02:19 - forms of AI as we know, where we use different machine learning methods to actually
00:02:25 - that can learn functions, right?
00:02:27 - This was the whole point as far as humans writing them out, we can try to learn them.
00:02:32 - And then we would jump into meta learning and meta search. And so what's the logic of these steps?
00:02:39 - What's sort of the driving principles that you would see sort of meta-search
00:02:43 - as a tip of that pinnacle?
00:02:45 - There are many ways of looking at this whole issue. I mean, there's a big rift
00:02:48 - in AI traditionally between symbolic AI and everybody else's AI.
00:02:52 - And before I came to Cambridge, I thought this was because of everybody else
00:02:58 - not listening to Marvin Minsky.
00:03:00 - And after I got to Cambridge and met Marvin Minsky, I felt it was basically the opposite.
00:03:06 - That Marvin Minsky, who co-founded the field with John McCarthy and a few others,
00:03:12 - realized that in order to understand cognition, we need to teach computers how to think.
00:03:16 - So we need to build cognitive AI in a way. And then he made the decision that
00:03:21 - this needs to be symbolic.
00:03:23 - And this was the wrong horse, but Minsky was too influential.
00:03:26 - He started yelling at people that did low-level cybernetics and neural learning
00:03:30 - because he felt what they were doing was too simplistic.
00:03:32 - And in a way, he was correct. What they were doing was very simplistic.
00:03:36 - But it turns out the way we can get to simple grounding and to the kind of operation
00:03:40 - that our minds are doing, they start from there.
00:03:43 - They don't start from symbolic representations, which are so highly abstract
00:03:48 - that we can only form them after generalizing over our perceptual representations and operations.
00:03:53 - And so on, right? These symbolic operations, we get them by making our representations
00:03:57 - so discrete and low-dimensional that we can apply analytic operators on them.
00:04:03 - It doesn't work the other way. You cannot start with these analytic operators,
00:04:06 - or at least this project has not worked out.
00:04:08 - But what happened is basically this big rift because Minsky basically started
00:04:12 - yelling at people for doing other things than he did, and this created an area
00:04:16 - of burned ground around cognitive AI,
00:04:18 - which was also symbolic AI, and the AI of other people, which never bothered
00:04:23 - to read Piaget anymore and instead mostly did neural learning and so on.
00:04:27 - But what happened in the first half of AI,
00:04:30 - basically since 1950 until quite recently, was mostly that people thought about,
00:04:36 - well, when we want to perform the following tasks that requires human intelligence,
00:04:40 - like playing chess or parsing natural language.
00:04:44 - What do we need to do to make that happen?
00:04:47 - And they came up with a clever algorithm and then they implemented this algorithm
00:04:50 - at some point because our computers are very deterministic and scale very well,
00:04:54 - these algorithms became smarter than people. So I'm born in 1973.
00:04:58 - There was never a point in my life where I had a chance to play better chess than a computer.
00:05:04 - In a way, depressing, right? Even though these algorithms were so simplistic
00:05:08 - that I was able to write an algorithm that would play better chess than myself, right?
00:05:12 - Even though I'm not that good of a programmer compared to the people that put all this effort in this.
00:05:18 - But we know that our strategy of playing chess is very different than the ones
00:05:22 - of this simplistic algorithm.
00:05:24 - We are able to cope with much more uncertainty in our search space than our
00:05:28 - programs, so we can have much more fuzzy representations that are more perceptual
00:05:33 - and difficult to grasp what they are about.
00:05:35 - And to find such ways of navigating very complex, hard to define problem spaces,
00:05:41 - we cannot enumerate these problem spaces directly, but instead we go to programs
00:05:46 - that learn how to do this.
00:05:47 - So this new wave of deep learning is mostly about building systems that learn
00:05:52 - how to solve a particular kind of problem.
00:05:55 - And they're not general in the sense that we now have systems that can do this
00:05:59 - for any class of problems, but you basically set up a learning algorithm that
00:06:02 - learns how to play chess, or another learning algorithm that learns how to play Go.
00:06:06 - And we can now use very similar learning algorithms for the same thing.
00:06:10 - And then this thing just plays against itself and abstracts over what it sees,
00:06:15 - and then it outperforms everything humanity has done so far in this particular domain.
00:06:21 - And the next step might be that we go one step above this, which means we don't
00:06:25 - write this learning algorithm by hand for this particular task,
00:06:28 - but we write another algorithm which learns how to find that learning algorithm,
00:06:32 - which means we go to meta-learning, right?
00:06:34 - And it is tempting to think of our brain not as a learning system,
00:06:38 - but as a meta-learning system because many domains require very different learning
00:06:43 - strategies when we want to conquer them, right?
00:06:45 - And our brain is, in some sense, a machine that figures out how to learn a new type of thing.
00:06:50 - And from this perspective, evolution could also be seen as a very slow and unprincipled
00:06:56 - search for meta-learning systems to which we are the first solution that has
00:07:00 - managed to build computers.
00:07:02 - But now, okay, this is fine, right? Okay, fine, I take it. But in some sense,
00:07:08 - I could also say, well, there's no...
00:07:10 - I understood the words. I understood the logic. That gives me hope.
00:07:16 - But now I could also say, well, that's great, because that's the scheme that
00:07:21 - just shows that the author of the scheme understands recursion, right?
00:07:26 - And then you just follow the logic of recursion to say, ah, so this is now the
00:07:30 - logic I should follow, right? because they're unknowns. I have algorithms, oh, they're unknowns.
00:07:35 - Okay. Then I have to cover the space of puzzle algorithms.
00:07:38 - Ah, but when I have a trick to cover on the space of algorithms,
00:07:42 - there might be other spaces I have to cover.
00:07:44 - So the process seems to have a meta level that's actually not so complicated.
00:07:50 - And indeed, when I was growing up in sort of this transition from symbolic AI
00:07:54 - to what then called was new AI or embodied AI and new artificial life,
00:07:59 - whatever, ever, that was exactly really the dream.
00:08:01 - We want to use Jack algorithms to find algorithms that can define other algorithms, right?
00:08:07 - So we're in a loop because this is late eighties, right?
00:08:11 - Yeah. So, so what should give us any hope that also as a research program,
00:08:16 - this has any leverage and any traction?
00:08:20 - Because in some sense, also, when I see the proponents of this approach.
00:08:24 - Uh, going around all for many years, telling me these kinds of stories,
00:08:29 - It's not necessarily you, right? It's just a general reflection on that field.
00:08:33 - If we then pose a simple question, okay, but then show me an example that I
00:08:36 - find impressive or that can convince me also from a scientific perspective.
00:08:41 - Examples are maybe a bit poor, right? So in that sense, for me,
00:08:45 - things like AlphaGo, you can also say, okay, it just means it's another game
00:08:48 - that's not interesting, right?
00:08:49 - So why should this program or this logic actually get us to this promised land, right?
00:08:58 - So why should I have any kind of confidence in that track and trajectory that
00:09:04 - people have already announced for the last 30 years, along which we have not
00:09:08 - found much, right? You're still in the desert.
00:09:11 - So you're going all to grumpy AI programmer that has gone bitter on me now.
00:09:15 - Okay. No, I understand this. I can relate to this.
00:09:18 - A grumpy scientist who tries to understand the reality he finds himself in,
00:09:26 - and he does that by trying to get traction on issues like explanation and prediction.
00:09:33 - But grumpy nevertheless. Yes. I'm old enough to remember that we thought for
00:09:39 - a long time Go is going to be for a long time outside of our reach because the
00:09:43 - search space is so fuzzy that we cannot really enumerate it.
00:09:46 - And the situations cannot easily be ordered like chess situation where you say,
00:09:50 - my queen is gone. This is probably very bad.
00:09:52 - And there should be something making up for this if I go down that part of the search track.
00:09:57 - And in gold, all these things are much, much harder to define.
00:09:59 - So we thought for a long time this might be out of our reach for maybe even
00:10:04 - our generation or the foreseeable future.
00:10:06 - And a system that combines human intelligence with machine intelligence is going
00:10:09 - to dramatically outperform the whole thing.
00:10:12 - And now we know if we add human intelligence to this, it's going to get worse, right?
00:10:15 - There is nothing that a human player right now can add to AlphaGo to make it better.
00:10:21 - And so I think this is a reason to have a slight bit hope, but it's a very unprincipled reason.
00:10:26 - To get to a more principled perspective, I think we need to take a step back
00:10:30 - and wonder what is it that a mind is doing, right?
00:10:33 - A mind is some system to make models in the service of regulation for an organism.
00:10:39 - And to make a model, you need to identify the meaning behind information.
00:10:43 - But what we get from the world is information, certain patterns.
00:10:46 - We try to predict other patterns with it.
00:10:48 - And the meaning of information is its relationship to change and other information.
00:10:54 - And that is what we call a model so when you
00:10:56 - have a blip on your retina the meaning of that blip
00:10:59 - is the relationship that you discover to other blips on
00:11:02 - your retina at the same point or at different points in
00:11:05 - time right and these models that you can come up
00:11:07 - with describe for instance the movement of people in a three-dimensional space
00:11:11 - while they exchange ideas and have an economy and all these wonderful things
00:11:14 - and there are a simulation generated in your brain to explain and predict the
00:11:19 - patterns on your retina you will not be able to do this for all the patterns
00:11:22 - some Some of them are just going to remain blips because you don't find a relationship.
00:11:26 - So we call those noise. And it's a lot of noise in our retina, right?
00:11:30 - So when we try to make sense of our universe, we get a few samples of bits at
00:11:35 - different points in time.
00:11:36 - We try to relate them by making these models. This is the task of our mind.
00:11:40 - So we already know our mind is a function approximator. And we also roughly
00:11:45 - know over how many data our mind is going to approximate.
00:11:48 - We know how many seconds we are alive. We know roughly the frequency at which neurons operate.
00:11:52 - We know our surface that basically is our systemic interface to the environment
00:11:58 - that we operate in and so on.
00:12:00 - So we have a rough idea of the ballpark in which we operate and over which we make models.
00:12:05 - Are you describing just one side of the elephant? Because now you're pulling
00:12:09 - at the tail and you say, okay, we need to compress information, right?
00:12:13 - So that's why we build models. But on the other hand, on the other side of the
00:12:18 - elephant, we have also the need to generate action, to control physical systems
00:12:23 - in real time. Yeah, in the service of regulation.
00:12:28 - That's fine. Yeah. But that doesn't
00:12:30 - necessarily satisfy this whole idea of compression and model building.
00:12:35 - Because now I have to sort of flip it around and sort of again decompress into
00:12:39 - a real world that is open. open, right?
00:12:42 - You cannot necessarily capture all of that with the idea of model building and data compression.
00:12:48 - I'm not quite sure in which sense you would ever decompress.
00:12:52 - Because I don't interact with some decompressed world.
00:12:56 - I'm interacting with the world of my model, with the simulation,
00:12:59 - with the dream generated in my neocortex.
00:13:01 - This is subjectively the world that I inhabit.
00:13:03 - I'm never going to decompress you back into a part of the evolving state vector
00:13:08 - of the universe, into some weird quantum graph.
00:13:10 - Instead, I always interact with you as a person that has intentions,
00:13:14 - that has a certain geometry and so on. And so I'm always interacting with this
00:13:18 - data compression model.
00:13:19 - Yes, but with decompression, what I mean, you interact with me,
00:13:23 - I'm going to tell you things that you cannot fully predict, right?
00:13:27 - Yeah. I might say, right or wrong.
00:13:29 - So now there's a decompression sense of the way you interact with the world. The world is open.
00:13:35 - So the world now again generates highly variable feedback to you.
00:13:39 - Right? So you're compressing. the compressed states
00:13:42 - in the end get mapped to an action that you
00:13:45 - bring back to the world now the world doesn't get open right
00:13:49 - so now again you face a set of
00:13:52 - possibilities that collapse into one that's all you're exposed to
00:13:55 - yeah right and then again you go back
00:13:58 - into your compression mode and you generate your next action yeah right so that's
00:14:01 - more what I meant with the decompression part of it that your action is like
00:14:05 - this finger in this massive state space that generates yeah but basically what
00:14:10 - I do is I try to filter as many invariances from the world as I can and put them into a model.
00:14:15 - And everything else is the state of that model in as far as it is relevant, right?
00:14:19 - So I have a space of relatively few latent variables, just a few million of
00:14:24 - them, right? Or a few billion of them in which I operate.
00:14:27 - And I have a much, much larger static set of constraints that captures the laws
00:14:33 - of perspective and biology and intuitive physics and so on to make the world interpretable.
00:14:38 - So I can actually regulate my needs in a better way than a frog can.
00:14:43 - But with that, you're not necessarily saying anything that should knock me off
00:14:48 - my chair, right? Oh, that would surprise me.
00:14:53 - But this is a very reasonable way to summarize what a mind or an embodied mind would do.
00:15:00 - Should I apologize for suggesting a reasonable way?
00:15:04 - I suspect that we have even converged in many of our perspectives on AI have
00:15:09 - similar intuitions on how to get there.
00:15:12 - I don't know if our intuitions disagree on how long this is going to take and
00:15:16 - what is going to be necessary to get there.
00:15:18 - But it would be very surprising if we wouldn't be getting there even within our lifetimes.
00:15:24 - I think, David, I can understand where you're coming from in this sort of first,
00:15:29 - second, third-order, fourth-order systems.
00:15:30 - But people have been thinking about this for a long time. And...
00:15:36 - Whether it could be so neat is, I mean, it's kind of like a physicist's view
00:15:40 - of how things could possibly be.
00:15:43 - And you're a psychologist, so it surprises me you take such a sort of innate
00:15:48 - taxonomy of intelligent systems.
00:15:51 - Oh, no, I'm just trying to tell a story that fits into a podcast.
00:15:54 - It's brain, more complex than body. I mean, to say that intelligence is function
00:15:59 - approximation, which is a huge statement in itself.
00:16:02 - I mean, that's one of the things that is. What else is it?
00:16:07 - It's the ability to act appropriately in real time, and that may not require
00:16:10 - function approximation.
00:16:12 - The reason why I don't do this, why I don't lump this in this regulation,
00:16:16 - is because we observe that there are people that are very bad at acting appropriately.
00:16:23 - And they might nevertheless be extremely intelligent.
00:16:26 - So I would distinguish the ability from reaching your goals,
00:16:30 - which we could call smartness, from intelligence, which is the ability to make
00:16:35 - models, or wisdom, which is the ability to pick the right goals.
00:16:38 - But we find very often when we observe people is that excessive intelligence
00:16:42 - is often a prosthesis that our brain is using to compensate for the lack of
00:16:48 - being smart or being wise.
00:16:50 - Yeah. Okay. So we can have this particular definition of intelligence which
00:16:56 - maybe captures one aspect of humanity.
00:16:58 - Yes, of what minds are doing. Whether it can capture general intelligence you
00:17:03 - know, in terms of all the different kinds of intelligences that humans have,
00:17:07 - you know, sort of in a Howard Gardner sense.
00:17:09 - There might be multiple kinds of intelligence, maybe they're not all function approximation.
00:17:14 - But then, you know, if we look at what learning is, what learning to learn is,
00:17:19 - what searching for the optimal theory of function discovery is,
00:17:22 - these aren't necessarily, as you say, a recursive thing of applying the same
00:17:28 - idea multiple times to itself.
00:17:30 - They could be completely different. Oh, I fully agree. And so what we have in
00:17:35 - the current wave of AI is basically the optimization of inductive learning systems.
00:17:42 - But we still need deductive systems, and we need to discover abductive learning
00:17:48 - systems, which reason beyond the data, which no current deep learning system
00:17:53 - can do, but humans can do.
00:17:55 - And we don't know how humans do that.
00:17:58 - And we need to also automate the process of scientific discovery,
00:18:02 - and we have no idea how to do that either.
00:18:05 - Wait a moment. You might be too bold in your statements here.
00:18:10 - Me? Yes. I do remember when deep learning started to be… I think he's trying to be nice.
00:18:17 - When deep learning came up, basically everybody around me, including myself,
00:18:22 - started to make statements about what deep learning cannot do.
00:18:26 - And time and again, we had to revise our statements about what it can and cannot do.
00:18:31 - And it's a surprising thing to me. Still, it's surprising because you have these
00:18:35 - essentially mostly feed-forward network with very limited recurrence because
00:18:39 - we cannot train this very well.
00:18:40 - And very few lateral links because we cannot deal with those very well.
00:18:43 - And still, at some point, Google threw away everything that computer linguists
00:18:49 - had been done over decades in terms of machine translation and replaced the
00:18:53 - whole thing with an end-to-end trained, relatively dumb feed-forward network.
00:18:56 - And there is this thing, you know, if you think about your neural network,
00:19:02 - your feedforward network, not as representing a brain, something like a neural
00:19:06 - architecture with all these occurring connections,
00:19:09 - but with a function that correlates adjacent brain states,
00:19:13 - it's not quite clear what it cannot do, right?
00:19:15 - The question is, do we find a sufficiently defined loss function and an algorithm
00:19:21 - that converges quickly enough on a solution to that particular function?
00:19:25 - So when you want to find an algorithm that discovers truths in a particular
00:19:31 - mathematical domain, for instance, by discovering new theorems,
00:19:34 - it's not entirely clear to me that we can prove that our current deep learning
00:19:38 - approaches cannot get there. No way.
00:19:40 - You're going too quickly here, because if you take this example of machine translation by Google.
00:19:47 - Right, we have to, I think, really think also about the epistemic autonomy of
00:19:52 - these systems, right? Right, the old symbol rounding problem.
00:19:56 - Why could Google get away with throwing away a lot of the traditional methods,
00:20:03 - is they could also capitalize on the collective knowledge of a huge amount of
00:20:06 - humans who would just add exceptions and new information to the lookup tables they can build.
00:20:13 - So by virtue of that, the system actually worked. So that means epistemically
00:20:18 - it was not a common message.
00:20:19 - So in that sense, it didn't solve the problem of translation as biological systems
00:20:26 - as ourselves would form that.
00:20:28 - Of course. They found a workaround. Yes. Which has been really shrewd,
00:20:33 - if you want, and smart, and the way they capitalized on human knowledge is fantastic, great.
00:20:37 - But it doesn't mean that they really solved the actual challenge.
00:20:41 - They found a workaround.
00:20:43 - Of course, the question is, what is the challenge in terms of providing machine translation?
00:20:49 - If you... But we're talking about it in the account of mind,
00:20:52 - right? Of course, this is not a mind. Nobody says this. We all agree.
00:20:57 - So, yeah, it keeps surprising me. It keeps surprising me in the sense that I
00:21:01 - thought, oh, it does more than I thought it would do with these relatively simplistic methods.
00:21:05 - Because most of them we discovered in the 1980s already. Yeah, and I think...
00:21:09 - In the 1980s, certainly people like me, I had the faith.
00:21:15 - I thought these systems would be super powerful once we scaled computing power.
00:21:19 - I mean, I did my PhD in reinforcement learning, and it took me a week to train
00:21:24 - a little bug to move around and avoid obstacles, and now it would take half a minute or less, maybe.
00:21:31 - So you can look at the algorithm and know that it's a super powerful algorithm.
00:21:36 - In fact, given the right, you know, sort of, and AlphaGo is probably an amazing
00:21:41 - design of network, but you can see that reinforcement learning in itself is
00:21:46 - such a powerful algorithm.
00:21:48 - Given a state space that's so easily defined as Go, it ought to be able to do really well in that.
00:21:55 - So the things where I disagree with you is, I do think it's possible,
00:21:59 - or it's actually very likely, that the current class of algorithms is not going
00:22:02 - to carry us all the way. And we need to invent new classes of algorithms.
00:22:06 - We agree. Yes. But the thing where we disagree is, I cannot even prove that
00:22:11 - one. And I haven't seen a proof that they cannot do this.
00:22:14 - So as long as there is no proof for this, we don't know the limitations of these
00:22:19 - things. You're arguing against yourself, then.
00:22:21 - You're saying that we don't need to go to a possible discovery model.
00:22:25 - You put an uncertainty to the negative, where it's not even that one is justified.
00:22:31 - We don't know that. I think we don't know how far the current approaches can carry us.
00:22:36 - Well, it's very difficult to find a proof for the absence of something.
00:22:40 - So you're asking the impossible. So, no, well, you could say there are certain
00:22:44 - classes of algorithms that you can, in principle, not learn.
00:22:47 - If you could find a proof to show, okay, when Tony says something,
00:22:52 - we cannot get this thing to discover new theorems with this.
00:22:56 - This is a pretty bold statement, and you could probably formalize this.
00:22:59 - And then I could ask Tony, what is the evidence for that claim,
00:23:05 - except vague intuition? Well,
00:23:08 - no, I think scientific discovery requires interaction with the world.
00:23:13 - And that's what makes scientific discovery different. You know,
00:23:16 - you can't just, you need to go out there and you need to make observations and
00:23:21 - you don't a priori what to observe.
00:23:23 - None of these things in neural networks can do at the moment.
00:23:26 - And people have tried to automate that. But I think you guys are much too cozy there together.
00:23:32 - Because I don't understand why you guys are not in a state of panic over this
00:23:36 - challenge of understanding autonomy.
00:23:38 - Because this is actually, I wrote a few papers in the 90s already on that, analyzing that talk.
00:23:42 - NetTalk was one of the big breakthrough moments in connectionism because,
00:23:47 - oh, wow, we can map words to pronunciation features, right?
00:23:50 - And oh, wow, we have emergent vowels and consonants in terms of the pronunciation features.
00:23:55 - But the thing is, if you analyze the data, the patterns that were presented
00:24:02 - to NetTalk already encoded this distinct separation between vowels and consonants.
00:24:09 - So there was no magic there. It had everything to do with how humans had pre-coded
00:24:13 - the data, something that again
00:24:14 - means it was not epistemic, it was not epistemically autonomous, right?
00:24:19 - And this is the problem that Bumper has again all the time. These systems learn
00:24:22 - fantastically well on human-edited data.
00:24:24 - That's not helping us. So as long as we're not really solving this simple grounding
00:24:28 - problem without making progress, there's one issue.
00:24:30 - And I haven't heard an answer for that yet from your side of the table.
00:24:35 - You know, it's a round table, but okay. And secondly, you both seem to be very
00:24:40 - comfortable with this notion of intelligence, which worries me.
00:24:45 - Because intelligence is actually a very circumspect construct that McCarthy
00:24:51 - just pulled out of his head in some sense in the 50s to just give the field a name.
00:24:56 - Because they had computers, they had computers that would sort of imitate things
00:24:59 - that humans could do, like problem solving, and they had to give it a name.
00:25:02 - So he goes back to constructed go back to Galton late 19th century and Binet that were actually used.
00:25:09 - To make distinguish to distinguish between classes and races but for the rest
00:25:13 - it had no ontology to it and which was also exemplified by Binet he said well
00:25:18 - just what my test measures and even when Adrian Owen a few years ago put humans
00:25:22 - in the scanner and have them do the Raven progressive matrix test,
00:25:26 - for intelligence he saw that actually depending on the specific item you work
00:25:30 - on different parts of the brain and get involved.
00:25:32 - So there's only one underlying process, right? So how can you guys even in this
00:25:36 - search for a synthetic mind be so comfortable with this knowledge of intelligence?
00:25:41 - I think it's leading us astray. It's not helping.
00:25:43 - And the second thing is, if you want to insist on that, how are you going to
00:25:46 - solve this epistemic autonomy problem?
00:25:50 - So first of all, I don't think that intelligence is such a problematic concept.
00:25:54 - I do think that the notion of being able to make models is a very useful notion
00:25:58 - to understand what our minds are doing.
00:26:00 - And it's also clear that this is sometimes domain dependent.
00:26:04 - So we have certain priors in which areas we pay attention, which means where
00:26:08 - we put more resources into learning new models.
00:26:11 - And this leads to different performance. And so differences in people's innate
00:26:16 - or differences in talent is probably differences in innate attention to some degree.
00:26:21 - There are certain things that we as organisms are paid for to pay attention to, right?
00:26:27 - Paying attention, for instance, to facial expressions or to the emotional states of others and so on.
00:26:32 - And this will create different degrees of resolution in our mind with respect
00:26:36 - to these domains and bias us in a certain way.
00:26:40 - There's this other thing what animates us to do certain things
00:26:43 - are the needs of the organism that are implementing some
00:26:47 - kind of reward function that gives us pleasure and pain
00:26:50 - to reward us for doing certain things or punish
00:26:53 - us for doing other things and then anticipated pleasure and pain which make
00:26:58 - us go into certain areas of the behavior space or avoid certain areas of the
00:27:02 - behavior space in advance and eventually lead us to generate models about the
00:27:07 - structure of that that behavioral space and the policy space in which we operate.
00:27:11 - And so if you ever build a robot that plays soccer, it's not very difficult
00:27:16 - to make a robot want to go to the ball in some emergent sense, right?
00:27:19 - And if you have imagined that your robot has many more needs that it needs to satisfy,
00:27:23 - like recharging its battery from time to time, avoiding injuries,
00:27:28 - and maybe getting repairs from time to time when the injuries have accumulated
00:27:32 - to such a point that they impair its function.
00:27:35 - So you could imagine that this thing needs to satisfy a number of needs and
00:27:39 - achieve some kind of dynamic homeostasis about them.
00:27:42 - And suddenly you have a system with some motivational system that gives relevance to what it needs to do.
00:27:48 - And then when you think about your mind as some general function approximator
00:27:52 - that has, so to speak, a bit of cortical real estate that it can use to model
00:27:57 - certain domains of its environment with more or less resolution,
00:28:00 - then what you should be doing is whenever Whenever you have a problem that you
00:28:04 - cannot solve with your existing feedback loops, you need to create secondary,
00:28:07 - tertiary feedback loops. And at some point, this is not enough.
00:28:10 - So you need to spawn a completely computational model of that domain.
00:28:15 - So this is probably what our neocortex is doing. It creates a computational
00:28:18 - model of a domain so it can regulate this domain because you cannot regulate
00:28:22 - what you don't understand.
00:28:25 - Ashby's good regulator theory. If you build a regulator, then the structure
00:28:29 - that it implements needs to be isomorphic to the system that it regulates.
00:28:33 - And the world that we are in is a very complex, three-dimensional,
00:28:37 - dynamic, physical world.
00:28:38 - And then it's overlaid with a very complex, even more complex social world and
00:28:42 - economical world and so on that are in dynamic interactions with each other.
00:28:47 - And this is what our cortex, in some sense, is modeling.
00:28:49 - But not in an isomorphic way, I hope. You do abstract, right?
00:28:53 - And compress, as you said earlier.
00:28:55 - Yeah. So isomorphic to what, right?
00:28:58 - There is a certain world that happens there physically, and we already know
00:29:03 - that what we perceive as being the world is very, very different from this.
00:29:06 - So, for instance, there is no sound in the world.
00:29:08 - There are air molecules, at least been seen from some perspective, right?
00:29:14 - They are actually also not air molecules, but from some frame of reference,
00:29:17 - we can describe this stuff as air molecules and their electromagnetic forces
00:29:21 - between them that give waves of traveling pressure.
00:29:24 - And then there's some Fourier transform happening in our cochlea that measures
00:29:27 - the energy of these waves in different frequency ranges.
00:29:30 - And this gets sent in the neocortex. And now the neocortex does its very best
00:29:33 - to predict the structure of these patterns.
00:29:36 - The result is some interesting geometric calculation that our neocortex approximates.
00:29:42 - And these geometric patterns, we interpret as sound, right?
00:29:45 - And then we organize the sound into things like phonemes or into musical patterns and so on.
00:29:51 - That's fine but none of what you said necessitates the use of the word intelligence
00:29:57 - right because now you're talking about induction but the ability to build models.
00:30:02 - The ability to estimate the sources of the sensory states that you're exposed to.
00:30:09 - I think that before you get to the sources, a lot of other things are happening.
00:30:13 - This idea that all the patterns that you perceive go to sources is already a
00:30:18 - major inference that you're making, right?
00:30:20 - So you make these inferences because they give good predictions, right?
00:30:23 - But what you eventually try to do is you try to predict patterns.
00:30:28 - Patterns and maybe a good metaphor for this is to think of a synthesizer right a synthesizer is a.
00:30:35 - Haverset arrangement of electronic components that have
00:30:38 - the property if you connect them they go into weird oscillations
00:30:42 - and produce interesting dynamics and you have a few knobs that you can use to
00:30:46 - fiddle until it makes the sound that you want and imagine you have something
00:30:51 - like this built from neurons that the brain fiddles at the knobs until it's
00:30:55 - able to predict the patterns coming in from the cochlea in a few dimensions.
00:30:59 - And when it has done this exhaustively, it moves to the next layer and then
00:31:03 - tries to identify the covariance between the knob states, between the latent
00:31:06 - variables that it has discovered.
00:31:08 - And now uses this to generalize over the previous patterns.
00:31:12 - And it does this until it's able to fuse the different sensory modalities into
00:31:16 - a simulation of a dynamic free space with moving objects.
00:31:19 - It does that in the context of a superbly tuned body, a sort of physical apparatus,
00:31:24 - which provides all the transducers,
00:31:27 - which these, you know, sort of pattern recognisers, function approximators can
00:31:32 - work on, and already decomposes the world in a way that is well set up in order
00:31:38 - to be able to make sense of it.
00:31:39 - So, I mean, in your notion of function approximation doesn't really,
00:31:44 - can't really stretch to include this aspect of the.
00:31:48 - What some people call morphological intelligence. Yeah, I'm totally biased against
00:31:53 - that. I don't believe in morphological intelligence.
00:31:55 - It must be true that the body makes the challenge of the brain.
00:31:58 - What do you call this? Intelligence.
00:31:59 - Well, I don't like the word intelligence or morphological intelligence,
00:32:02 - but other people use the phrase.
00:32:04 - But it's certainly true that the body makes the challenge.
00:32:08 - The brain has much simpler. It makes it easier to control walking because your
00:32:12 - legs are in the pendulums.
00:32:14 - It makes it easier to do hearing because you have ears. You're talking as if
00:32:18 - you have never seen a toddler walking with legs that are absolutely not fine-tuned to walking.
00:32:23 - Your movement of your geometry of your body is optimized for your adult state.
00:32:28 - If you are a toddler, it's actually much harder to walk because all the centers
00:32:32 - of gravity are off, and you still manage to do this.
00:32:34 - In the same sense, you have all these nerves scattered over your skin.
00:32:39 - This is not done in a very particular order.
00:32:42 - Your brain is doing the co-occurring statistics over these nerves.
00:32:45 - That's not a bad way to go in because there are always examples where just the
00:32:49 - biomechanics just helps you.
00:32:51 - Think grasping, right? You just push the center of the hand and grasp.
00:32:54 - And this is true for any age or development, right?
00:32:58 - So if you push hard enough in the hand, it closes like this so you grasp something.
00:33:04 - So this helps you. You don't have to plan this. Oh, yeah. I'm not against the
00:33:08 - idea that your body is optimizing because it's subject to the same evolutionary
00:33:12 - constraints. But I wouldn't call this intelligence.
00:33:16 - Because your body is not making interesting models. But look,
00:33:19 - you and Tony both want the same intelligence. No, we're not on the same side
00:33:22 - here because you're saying that these models aren't grounded epistemically.
00:33:27 - Paul said this, and I'm agreeing with him. And I'm saying that what grounds
00:33:31 - those models is that they interact with the world through a body which is optimized
00:33:35 - through evolution to do that.
00:33:37 - So to think about intelligence in this entirely abstract way may not lead you
00:33:41 - very far because the embodiment is critical.
00:33:45 - To the way that human intelligence operates and has evolved.
00:33:50 - And we may not be able to build disembodied
00:33:52 - AIs just by having recursively bootstrapping and learning systems.
00:33:58 - And people are trying to do that and have been trying to do that for decades.
00:34:01 - And what I don't see is any breakthroughs in that or imminent breakthroughs.
00:34:04 - So maybe we need to follow this more biologically grounded approach,
00:34:09 - which is to say, let's build embodied AIs.
00:34:13 - That would be my approach. I think it's probably your approach,
00:34:16 - certainly false. I think it's a super popular approach. People like that approach, right?
00:34:20 - Well, because there's a lot of work. But if you go to NIPS, that's not what they're doing.
00:34:26 - Yeah, they do the stuff that actually works.
00:34:29 - Abortimentalism didn't work. Historically, it's been a failure.
00:34:31 - When this new AI stuff came up, people have been building lots of robots.
00:34:35 - Like Rolf Pfeiffer's Klapperaturen in Zurich developed.
00:34:41 - They're very impressive. It's very aesthetic, but it didn't get us intelligent systems.
00:34:46 - And so I'm of two minds with respect to this.
00:34:49 - I think that for our human intelligence, there's something that's quite different
00:34:52 - from what a lot of machine learning is right now doing.
00:34:55 - For instance, when we train a vision system in machine learning,
00:34:58 - the typical approach is we give it an annotated database with a few million
00:35:01 - images and ask it to generalize over this.
00:35:04 - Or we might even give it a non-annotated database of video streams and ask it
00:35:08 - to generalize over this. it's going to figure something out.
00:35:11 - But it's figuring something out that is slightly different from what we figure
00:35:15 - out, because, for instance, it's prone to adversarial examples.
00:35:17 - So you change a few pixels strategically in the input image,
00:35:21 - and suddenly instead of a cat, it sees an ostrich. How is that possible?
00:35:25 - It's probably because it's mapping this onto a space with a very different structure
00:35:29 - than the space that our minds is mapping these things on, which means it gives
00:35:33 - you a similar classification under most circumstances, but not under all circumstances.
00:35:39 - Especially it's not good at dealing with novelty so if
00:35:42 - you throw a end-to-end train system a picture for
00:35:45 - instance of a plane that is crashing into the tarmac it might
00:35:48 - recognize okay there is a plane and there is a tarmac and it might figure out
00:35:52 - oh there is a plane on the airport and this is what it's going to tell you it's
00:35:55 - going to miss out on the crucial interesting fact that this plane is about to
00:35:59 - crash and that thing is going to happen so how can we do this and i think it
00:36:04 - starts In this sense, I agree with this embodiment thesis.
00:36:07 - We start with our proprioception when we are born. It's already set up.
00:36:10 - And we already have made a map of our body surface before we are born.
00:36:14 - And then we add stuff to that map. We probably add stuff to that local perceptual space.
00:36:20 - And whenever we see something, we can work based on the assumption,
00:36:24 - what I'm currently looking at is a window into the same dynamic free space.
00:36:28 - So everything that I perceptually sense is a part of the same unified world.
00:36:36 - This is a very interesting thing. So when somebody shows me a picture of a cat,
00:36:39 - I don't see an isolated set of pixels that it now has to craftify.
00:36:42 - I know it's a person giving a picture to me of a cat that has been produced
00:36:46 - by a machine and at some point was a window in a different part of the same reality.
00:36:50 - This is the way I make sense of it. And it's very hard to fake such a universe, right?
00:36:55 - Wait, just before we go there, so I heard me challenge you that,
00:37:00 - so you were saying, look, a lot of people believe in this sort of embodied view of mind.
00:37:06 - I'm explicitly not saying intelligence.
00:37:10 - And then I challenged you and said, look, but if you go to NIPS,
00:37:13 - no one is doing it. And you said, it doesn't work.
00:37:17 - But the things that the people who go to NIPS believe work might work for the
00:37:23 - wrong reasons because the tools they build work because they rely on massive
00:37:28 - amounts of human annotated data,
00:37:31 - which is not the way forward.
00:37:33 - We know that will hit the wall somewhere, right? Well, then you said,
00:37:37 - well, I live in World War V.
00:37:38 - Well, I brought robots to World War V's lab in 1991, so I know wolves work pretty well, right?
00:37:45 - Then we have to be clear what does it mean to work, right? In our own trajectory,
00:37:49 - this whole embodied approach has worked extremely well because we have really
00:37:54 - explained aspects of the brain to the extent that we could make testable predictions
00:37:59 - that were confirmed to some extent, but were falsified.
00:38:02 - So from the perspective of a scientific program, this has been extremely successful
00:38:07 - to the extent that now we are repairing the brains of strong patients with, right?
00:38:12 - So it also depends very much on how you define it worked. So I think we should be very careful there.
00:38:18 - And so did we scale to anything you might want to call artificial general intelligence?
00:38:23 - No, that's the future. That's a challenge. But I think the situation frame-bodied
00:38:28 - models of mind are not that bleak. It just depends what you're looking for.
00:38:33 - And if your benchmark is what the computer scientists and the big tech companies
00:38:38 - are looking for, sure, then we might not be performing really well.
00:38:42 - But if you're interested in a more scientific program, I think we really have made huge progress.
00:38:46 - I think there is value there. And this might be an alternative approach to these
00:38:51 - questions than the more computer engineering-informed one.
00:38:54 - I think we should also take care to distinguish the practical work from the
00:38:59 - industries of AI hype and anti-AI hype.
00:39:03 - These are two variable industries that feed a number of people to produce AI
00:39:08 - hype or anti-AI hype. and there is actual research work and there's also scientific work.
00:39:13 - And these are slightly different things. And if you look at NIPS,
00:39:16 - it's a very diverse thing.
00:39:17 - I mean, it's about as large as Burning Man, but sells out in a quarter of the time by now.
00:39:22 - And there are a very large number of people that do very constrained vision
00:39:27 - staff or convolution network things.
00:39:29 - And there's also many, many other people. And the absolute number of people
00:39:33 - that wanted to do a science of intelligence under the umbrella of AI was always
00:39:37 - a pretty small minority. minority and I think the ratio of people that do this is almost unchanged.
00:39:44 - So the vast majority of people in AI do engineering.
00:39:48 - They basically build useful applications for data
00:39:51 - processing systems at the forefront of what we knew
00:39:54 - about how to do clever data processing and a small subset
00:39:57 - a few percent of these people are interested in actually
00:40:00 - how the mind works and dedicate their careers to that question and just
00:40:03 - because so many people now do AI doesn't mean
00:40:06 - that these others are not doing it is just it's the same fraction the absolute
00:40:09 - numbers are probably larger than ever in the history of AI and still it's a
00:40:13 - small subset and the absolute number of people that are interested in building
00:40:17 - models of minds and have very similar ideas to you and me with respect to this
00:40:21 - that are at NIPS are probably still a few hundred to a few thousand people.
00:40:26 - No sure I think this is a really important distinction it's okay so now we got that out of the way,
00:40:32 - everyone said the second slide exactly so,
00:40:39 - You did emphasize that consciousness has a role to play in reaching this synthetic mind, right?
00:40:47 - So what's your definition of consciousness?
00:40:50 - What's its function? And why is it such an important ingredient to synthesize mind?
00:40:56 - So there are several aspects to consciousness. And when we use this word,
00:41:00 - we often don't make clear what we mean by it, right?
00:41:03 - It also doesn't help. It's mostly a thing that we know by pointing and not by
00:41:07 - agreement about what we mean by its particular functionality.
00:41:11 - So there is something like alertness, the ability to react to stimuli in real time and so on.
00:41:16 - Then there is the ability to learn and update and the ability to reason about
00:41:21 - your behavior and reflect on it.
00:41:23 - There is the ability to attend to stimuli and high-level stimuli that we abstract
00:41:28 - in your brain into mental simulations and to remember that you attended to them
00:41:32 - and to to remember that the second order attention to this,
00:41:36 - so this reflexive consciousness above this excess consciousness.
00:41:40 - And then last but not least, there is this phenomenal experience,
00:41:44 - what it feels like to be in a particular situation, right? So maybe we start with this one.
00:41:50 - At the core of what it feels like, if you really meditate very,
00:41:54 - very deeply, you feel that there is, at least this is what appears to me like,
00:41:59 - a fundamental kind of disagreement with the universe at some level.
00:42:03 - And whenever you perceive a feature, it has a particular meaning with respect
00:42:07 - to this kind of disagreement. It gives you a particular kind of relevance.
00:42:11 - It's like zero-order consciousness.
00:42:14 - And then as soon as you start conceptualizing objects, before you think about
00:42:19 - them in terms of reflexive language and so on, you get something like first-order
00:42:23 - consciousness, where you see a particular thing and you feel,
00:42:26 - I am perceiving a particular thing.
00:42:28 - And when you go above this thing then you are in the conceptual level already
00:42:34 - that basically you now start carrying that thing around.
00:42:38 - Second order like in a glass bowl and you see this glow of consciousness with
00:42:43 - every step it gets a little bit darker but it's a symbolic manipulation that
00:42:46 - you perform to shift that thing around in your mind and do things with it and
00:42:50 - when I give a linguistic report about the whole affair to you then consciousness
00:42:54 - is a far distant memory to the processes that produce this or it's like a myth.
00:42:58 - It's something that my linguistic apparatus is not actually participating in very much.
00:43:02 - So there is something like a subjective feeling of being conscious in actuality.
00:43:08 - And the weird thing is that we probably cannot be conscious in actuality in
00:43:13 - the sense that a physical system cannot do that.
00:43:15 - The brain takes a long time to process stimuli.
00:43:18 - For instance, when I stop my tool, probably takes something in order of 700
00:43:22 - milliseconds for that to register in my neocortex in my integrated world model.
00:43:27 - But I have the impression not only do I perceive this in real time,
00:43:31 - but I also act on it in real time.
00:43:34 - Subjectively, I'm not just experiencing this with a nice time delay that can
00:43:38 - be filtered away with some editing trick in my mind, but I also have the impression
00:43:42 - it is me that acts on that stimulus which happens right now, right here.
00:43:46 - And that is not possible because the initiation of that action is going probably
00:43:49 - to take another second until I'm able to catch myself stumbling and react to
00:43:54 - that pain and integrate this all into a cohesive story, which means the story
00:43:58 - must be instantiated with hindsight. side.
00:44:01 - At least if physics is correct and we live in a mechanical universe.
00:44:06 - Subjectively, I don't live in a mechanical universe. So how do I what sense
00:44:11 - can I make of this dualism?
00:44:13 - Subjectively, I live in another universe in which I can react to me stopping
00:44:18 - my tour in real time and experience that, right?
00:44:22 - The thing is that you do but your conscious self doesn't.
00:44:26 - But your body does. and so this is the
00:44:29 - other thing with the conscious self which is
00:44:32 - problematic and your body is reacting
00:44:35 - as quickly as it can yes so in a way it reacts
00:44:37 - as quickly as it can but also not in real time but the problem is that you're
00:44:42 - identifying with your subjective consciousness not with the rest of your body
00:44:46 - so when you say I can't respond that fast yes you can if you identify with your
00:44:51 - toe and with the reflex systems in your spine that's as fast as you can respond,
00:44:57 - And then those systems report up to your stream of consciousness that this has happened.
00:45:03 - But I don't understand why that's a problem. I mean, scientifically,
00:45:06 - we know that that's what must be happening. Exactly.
00:45:10 - But this is what gives rise to this confusion that people, including many philosophers,
00:45:15 - have about that. Personally, I don't think it's a problem.
00:45:17 - So there is an answer to that conundrum, right? Before you agree.
00:45:22 - Joshua is making an assumption, which is not declared,
00:45:25 - which is relevant for this argument, because you seem to be saying there's something
00:45:30 - like behavioral time constants, and I hit my toe, and now within a certain time
00:45:36 - frame, I will have to respond to that or I'll fall on my nose. right?
00:45:40 - And the time constant of that process is so short that it's not reasonable to
00:45:46 - expect that the system that would support consciousness can operate at that same time constant.
00:45:53 - That seems to be the argument you're making. Yes, but it's not an assumption.
00:45:56 - It's something that we have established by measurements.
00:45:58 - We can put electrodes to somebody's skull and we can see how long it takes until
00:46:03 - the stimulus arrives there, right?
00:46:05 - So we know about this delay. laser we can do the table now
00:46:08 - yeah but you didn't declare that before okay that
00:46:11 - that's the argument hinges on that you say i have two processes one
00:46:14 - fast and slow and they cannot be running they cannot coalesce so therefore the
00:46:22 - slow conscious process by necessity is the story you tell yourself afterwards
00:46:26 - so no it's it's even worse the problem is that yeah that this This conscious process,
00:46:33 - that thing that I perceive as happening, is happening instantaneously.
00:46:37 - It's happening in actuality. It's happening here and now without any delay.
00:46:40 - It's not even happening fast. It's happening instantaneously.
00:46:45 - So how is that possible that I act instantaneously on me stabbing my toe when
00:46:50 - physics does not permit my neurons to transmit signals instantaneously,
00:46:54 - but does so in time frames that I should be able to perceive?
00:46:59 - Right because i uh i can distinguish time frames that
00:47:01 - are uh happening at something like half a
00:47:04 - second or a third of a second i don't get the instantaneous i
00:47:08 - don't get the instantaneous can you explain that so uh when i act based on when
00:47:13 - i act on my perceptions subjectively and acting on my perception in real time
00:47:17 - really yeah i see you moving your head and i'm reacting subjectively instantaneously
00:47:24 - on movement of your head.
00:47:26 - And of course the solution to this is this is a story that's been created after
00:47:31 - the fact to basically make it possible for me to learn from the past interactions
00:47:36 - that my nervous system has orchestrated with the physical world.
00:47:40 - Well maybe that's also a little bit tricky bit of subjectivity.
00:47:44 - I don't experience it like that.
00:47:46 - Because I live according to my own theory which is real time is subconscious.
00:47:53 - And God's experience is the way.
00:47:57 - I don't experience stomach-mite tongue in real time. I don't,
00:48:01 - not at all. Maybe you're not conscious.
00:48:04 - That's definitely possible. We can't exclude that option.
00:48:07 - But it seems… Why are we so surprised that folk psychology is wrong?
00:48:13 - And we grow up being… Well, this is not folk psychology. I'm talking about my
00:48:17 - subjective experience.
00:48:18 - Your subjective experience is folk psychology because that's what you've been
00:48:22 - told as soon as you were able to listen. and these were explanations.
00:48:27 - Nobody talked to me about consciousness when I was old enough.
00:48:30 - They talked to you about... Where did you grow up?
00:48:32 - They talked to you about the fact that you could control your toe and that if
00:48:36 - you stubbed your toe... No, my parents mostly stayed out of my way.
00:48:39 - I learned to control my toe by myself.
00:48:41 - Oh, but still, that doesn't matter, right? You may have evolved your own folk
00:48:45 - psychology, but unfortunately, you grew up into a scientist and then you discovered
00:48:49 - that it was largely wrong.
00:48:51 - So now that we've discovered that folk psychology is wrong, and we can just
00:48:54 - put it to one side and not worry about it.
00:48:56 - No, the thing is that what I have to explain is the set of phenomena that is
00:49:00 - subjectively available to me.
00:49:03 - So the reason that people get confused about consciousness is that they have
00:49:07 - a particular kind of experience and that experience seems to conflict with what
00:49:12 - is scientifically possible,
00:49:13 - which is why people like Tononi come up with very expensive theories, right?
00:49:19 - I mean, epistemologically and ontologically very expensive theories.
00:49:22 - There is the reason why they're willing to carry their burden it's
00:49:25 - because they're not going to give up on what you
00:49:28 - so derisively call folk psychology well i
00:49:32 - think that's absolutely true but i think they should yeah oh i agree well julio
00:49:38 - ends up being a panpsychist yes oh that's yeah yeah okay that depends what kind
00:49:43 - of whose folk psychology you're buying but okay he certainly gave up mine um
00:49:47 - but but the thing is i think you try to wiggle yourself out of a problem here Joshua.
00:49:52 - And I would like to remind you of the problem, which is, I really believe that I remember.
00:49:58 - That you said earlier that the constant experience of stubbing your toe was
00:50:04 - a story you told yourself afterwards, in that example. You didn't do it instantaneous.
00:50:09 - And then I was trying to help you, but of course, I was doing the wrong thing
00:50:14 - for the right reason, right? Or maybe it was all wrong.
00:50:17 - By saying, by just trying to, you're making this fundamental assumption that
00:50:21 - we have a dual process system that have different time calls.
00:50:24 - Oh no, that's not what I'm doing.
00:50:26 - There is something else going on.
00:50:30 - There is this traditional perspective of idealism, right?
00:50:33 - Idealist philosophy basically makes the statement that we don't live in a physical
00:50:37 - world, but that we live in a dream, in a magical world, essentially.
00:50:40 - One in which symbolic interactions are possible instead of just mechanical interactions
00:50:46 - that need to be performed on some fundamental, grossly closed physical level.
00:50:50 - Let's say some quantum mechanical level or below that.
00:50:53 - And in this symbolic universe certain things are possible like our minds exist
00:50:59 - and the dualist perspective said that these two domains this physical domain
00:51:04 - of objects moving in a space or,
00:51:07 - information flowing in some kind of computational graph of the universe and
00:51:12 - this mental domain where our mental phenomena take place somehow needs to need
00:51:16 - to coexist and what i want to suggest that this is in some sense the case.
00:51:20 - We are literally living in a dream that is dreamt by a mind on a higher plane of existence.
00:51:25 - And the physical universe is that higher plane of existence.
00:51:28 - And the mind that dreams us is generated by the brain of a primate.
00:51:33 - So in some sense, these perspectives are complementary.
00:51:36 - Subjectively, I live in a dream world. The same circuits that produce a dream
00:51:41 - at night produce a dream during the day to track and predict my sensory data.
00:51:46 - And the properties of the... and model building and compression and all that stuff, right? So, uh.
00:51:51 - And the difficulty that comes up, which is the reason why people like Tononi
00:51:56 - are so confused or why strong embodimentalism has such an appeal,
00:52:01 - which basically there is a different kind of embodiment which doesn't say,
00:52:05 - let's use robots because they're so useful to ground your perception.
00:52:09 - Which I totally agree with, right?
00:52:11 - If you are a mind, you should be able to use your body. That's a useful benchmark.
00:52:15 - And it's a very good source of data if you're connected to a physical universe.
00:52:18 - There's a lot of interesting variants and invariants
00:52:21 - in it right and if you have some regulation goals that are maybe
00:52:24 - given from an organism that make you interested in in
00:52:27 - cool stuff right so you build interesting models and interesting policies
00:52:30 - i agree with that perspective but there is a deeper perspective which says basically
00:52:35 - some people get up from their meditation and realize oh my god this is so awesome
00:52:39 - this cannot be explained by a mere computational process there needs to be something
00:52:44 - going on that is more than mechanical and where does this stuff have come from?
00:52:48 - Because my brain is just performing some boring computations,
00:52:52 - probabilistic excitations in these glorified fat cells that we use for processing
00:52:56 - information that we call neurons, right? So what is that thing?
00:53:01 - And for some people, they suggest, oh, maybe it's the touch of the mystical
00:53:04 - reality substance. Maybe physical reality does more than computation.
00:53:08 - And the mind is not something that just happens in the brain,
00:53:11 - but it's the result of some magical morphic resonance between your body and your environment.
00:53:16 - By getting in touch with the mystical reality of substance, you have this emergent
00:53:21 - and possibly extended mind that spreads outside into this universe that you're perceiving.
00:53:27 - And what you don't realize is that the universe that you're perceiving is not
00:53:31 - some direct access realist universe, but it's a dream generated inside of your
00:53:35 - neocortex. It's a simulation. It's a VR.
00:53:37 - Sorry, are you endorsing the idealist perspective? back to?
00:53:41 - No, I'm basically saying we should interpret the idealist perspective as something
00:53:46 - that in some sense is true.
00:53:47 - You can perform magic, you can learn how to levitate,
00:53:51 - you can learn how to look into the future, but you do this basically by editing
00:53:56 - your memories, by changing the way you produce your neocortex produces these
00:54:00 - models, and you pay a price for this because you cannot break the bank,
00:54:02 - because outside there's a mechanical universe.
00:54:05 - I think it's very strange, because you already declared this earlier,
00:54:08 - where you say, well, Well, by necessity, we build models.
00:54:13 - And as soon as you start to rely on these models to inform your actions,
00:54:17 - you live in a virtual world. It's virtual action.
00:54:19 - And everything else is just a consequence of that. That virtual… And that's fine.
00:54:25 - The mind lives in the physical body, which lives in a physical world.
00:54:29 - Yes, that's the interface.
00:54:30 - But there's one thing. There's no magic. Exactly. No, there's no magic.
00:54:34 - There's no conspiracy. It's good news.
00:54:37 - We can summarize the last hour in just eight words.
00:54:41 - Which are the mind-built
00:54:45 - models and as a result you virtualize but
00:54:48 - now you add you say you add to the physical world you add an additional word
00:54:54 - which is the higher plane why is it necessary why is the the physical world
00:54:58 - now the higher plane why it's not just we build models that are informed by
00:55:04 - the external world but we abstract away from it and those in former action.
00:55:08 - Oh, it's because the world that I live in, subjectively, is VR,
00:55:13 - right? Similar to Minecraft.
00:55:14 - It's a computer game generated in my brain, right?
00:55:17 - And the physical world is the parent universe to this. In this sense,
00:55:21 - it's the higher plane of existence.
00:55:23 - It's the substrate plane. You seem to suggest a dualism I don't like,
00:55:27 - because in Tony's case, earlier you were sympathetic to this one-way view of
00:55:30 - mind, and that actually builds a very solid interface with you.
00:55:35 - From the beginning that cannot indicle.
00:55:38 - Since it's physically instantiated, it's an embodied mind, it's by necessity
00:55:43 - always one with its physical world.
00:55:45 - No, no, no. You're going to me. This is not true.
00:55:50 - So basically, I think if you would replace your neurons with something that
00:55:54 - is functionally equivalent and performs the same computations,
00:55:57 - you could still be living in the same VR.
00:56:00 - Look, if I allowed you to make all these assumptions, the conclusion might be
00:56:04 - accepted. But I think these assumptions are extremely strong and you will not
00:56:09 - succeed in living up to them until the end of your career.
00:56:13 - Oh, no, it's just you're not a functionalist. That's all. And you maybe,
00:56:16 - I cannot change your mind. I can try.
00:56:19 - But for the same reason that you can change the CPU in your computer to something
00:56:23 - that is functionally equivalent in the sense that it's able to run the same
00:56:26 - software on it, you can run the same program on your computer,
00:56:30 - right? That was not the argument.
00:56:31 - The point is you describe the external world as a parallel universe, right?
00:56:35 - That's the word you used. No, as a different universe. I didn't say parallel.
00:56:39 - Well, we did separate it. We did separate. No, it's a different frame of reference.
00:56:45 - So in the same sense as, for instance, the software that runs on your phone
00:56:49 - does not exist in the same physical universe as the phone itself,
00:56:53 - because it's a frame of reference that you're using to describe it,
00:56:57 - which doesn't make sense in that other universe.
00:56:59 - John, I think it's a shortcut you're making, right?
00:57:02 - Because all the words you pronounce now that are the product of this biologically
00:57:07 - instituted term machine that you believe you are, right, resonate back to you
00:57:12 - as you speak them through the real world, right?
00:57:15 - And even will hit your skin, you gesticulate.
00:57:19 - So the embodiment is implicitly there also as an integral component of these
00:57:27 - models in which you rely. It's not decoupled.
00:57:29 - Now they're independent from each other.
00:57:32 - There's continuous feedback between them. I remember from reading your paper
00:57:36 - that you are basically an identity theorist.
00:57:38 - And I don't think that's a useful way of conceptualizing that space because
00:57:43 - I am not a monkey and I'm not that body.
00:57:46 - I'm a side effect of the regulation needs of the monkey.
00:57:49 - I'm the result of that mind. I'm a story that the mind tells itself.
00:57:53 - There are people which only... Because you just joined that...
00:57:57 - But, Dushan, I reject your story.
00:58:01 - With the extent that you're constructing an I, why not construct an I that identifies
00:58:06 - with the physical self as much as with the mental self?
00:58:09 - Because this identification doesn't hold. I realize that when I think about
00:58:13 - what this physical system is, that I really don't get behind it.
00:58:17 - I have difficulty to identify with a subset of the evolution of the state vector
00:58:22 - of the universe, because this is not what I perceive myself to be.
00:58:25 - What I perceive myself to be is very similar to a character in a novel that
00:58:30 - is told in a very rich perceptual language.
00:58:34 - And this character in this novel
00:58:36 - is something that the brain uses to tell itself its own story in a way.
00:58:40 - But it's not the story of a brain or it's not the story of a primate or an organism
00:58:44 - or a bunch of cells or a bunch of particles.
00:58:47 - It's something else entirely, right? Because the thing is, a brain is a physical
00:58:51 - system. It cannot feel anything. thing.
00:58:53 - It cannot experience anything. Neurons cannot experience anything.
00:58:56 - But it would be very useful for the brain to know what it would be like if it
00:59:01 - was somebody who could experience something.
00:59:03 - So what the brain is doing, it creates this VR of a world.
00:59:06 - And then within this VR, it creates a person, a simulacrum of a person.
00:59:12 - And it does think to that simulacrum that is halfway to a simulation. Wait, it does things?
00:59:17 - Yes. So basically, as an author of that dream that you are in,
00:59:21 - that's the system, a set of processes that generate your VR and that create
00:59:26 - your personal self, your model of what you think you are.
00:59:29 - And then it does things to that model, like expectations and the result of past
00:59:34 - actions, and then witnesses the reactions of that simulation and uses it for learning.
00:59:39 - I mean, then you're identifying with this thing which is flickering on and off
00:59:45 - because this simulation of the self isn't going on all the time.
00:59:48 - Most of the time you're operating in the real world. Your body's just doing things.
00:59:53 - Yeah, so the good thing is my brain gives that thing access to the language
00:59:56 - center. And here I am and talk to you.
00:59:59 - And here we are. You don't have to. It's totally optional. It's a bonus.
01:00:11 - No, but look, this is partially all reasonable because….
01:00:19 - Because it still fits the same model building story, only now you say,
01:00:25 - look, but there is other sources of information related to the embodiment itself,
01:00:28 - which also build models.
01:00:30 - But now you have a problem because you indeed actually explicitly declare that
01:00:35 - there is something observing that model.
01:00:38 - So you have to solve one of those problems. No.
01:00:42 - Okay. How does the observer work? Who cares about the observer?
01:00:47 - Why do you need an observer at all?
01:00:49 - There is not an observer in the sense that there is something looking at it.
01:00:53 - What is there is a system that takes parameters from that simulation and uses it for learning.
01:00:59 - So there is a loss function that is generated by producing a virtual character
01:01:03 - with certain intrinsic tendencies, behavioral tendencies.
01:01:06 - So it's a model of my own internal regulation. simulation and when something
01:01:10 - happens to me in my virtual simulated world like my simulated,
01:01:14 - love interest my partner decides to
01:01:17 - desert me which is something that might happen in
01:01:20 - my mental simulation then my brain in some sense observes what happens to that
01:01:26 - character which doesn't mean it's laying back and watch that movie it just takes
01:01:29 - a few parameters from that thing and uses it for learning right so it will play
01:01:34 - that scenario into let's say working memory and then you can an experience in that scenario.
01:01:38 - Yes, so it's basically a simulated world that does things to that particular
01:01:42 - thing, because it's a very high-level compression of the state vector of the
01:01:46 - universe and available data at my systemic interface to say,
01:01:49 - okay, you have a partner and that partner just left you. What do you feel about this?
01:01:53 - Sure, but it doesn't matter whether it's high-dimensional or not.
01:01:56 - So, you have a self-model, you say, then you have a model of the world.
01:02:02 - And these two models now interact with each other, because now the self model
01:02:06 - could say okay, let's push some parameters in the world model to create some
01:02:12 - scenario that would affect me.
01:02:14 - Let's see what kind of station it would give rise to.
01:02:17 - So it's like the coupling of two models. So you can do this to some extent, right?
01:02:21 - You can imagine what it would be like if you would get a lot of chocolate in an hour from now, right?
01:02:26 - And you can evaluate your own reaction to this.
01:02:29 - But if you are working correctly, you are not allowed to believe that this is
01:02:34 - going to happen without evidence that it's going to happen, right?
01:02:36 - Because otherwise you become delusional.
01:02:39 - Which means you are not free to choose the environment that you want to live in.
01:02:43 - It's a very paradoxical thing because objectively you might suffer because the
01:02:47 - world simulation that you're in does horrible things to you.
01:02:49 - And you might wish that not to happen. It's actually your brain at some level
01:02:52 - that does these things to you. And you could go and reprogram that brain to give you paradise.
01:02:58 - But of course it might mean that this organism dies. Because the model does
01:03:02 - not give very good predictions.
01:03:03 - And yet, I think that most of what these Buddhist techniques of fixing your
01:03:10 - conflicts with environment are doing is to reprogram that simulation to get
01:03:14 - your brain to like you better,
01:03:16 - to like that person better, and do nicer things to it.
01:03:19 - The Buddhists found a very simple trick to do that, because you solve all conflicts,
01:03:23 - you don't want anything. You stop wanting stuff.
01:03:26 - So turn down all your drive systems, and the world is great.
01:03:29 - There is more than one Buddhist tradition, of course. And what you describe
01:03:32 - is the sutra. So the sutric tradition is ascetic.
01:03:35 - You basically get rid of your needs by stopping to satisfy them.
01:03:39 - And at some point, this thing realizes, I don't need to satisfy them and look forward to this.
01:03:44 - So they don't create any discrepancy between target and current value anymore.
01:03:48 - You forget about this thing and you become free. And there's the tantric tradition,
01:03:52 - which does the opposite, which says basically embrace this maximally and satisfy all your needs.
01:03:58 - Accept that they are there and explore them and eat your shadow,
01:04:03 - integrate this all into a thing that actually really works for you.
01:04:07 - So can I ask you a question? So I mean, since I've been in Barcelona,
01:04:11 - I've been having vivid dreams. Now, you wanna share?
01:04:15 - I can tell you more later. But when I wake up in the morning.
01:04:20 - I experience the world in a different way than I've experienced my model of the world in my dream.
01:04:24 - Well, my model of the world seems very vivid and very real when I'm dreaming it.
01:04:29 - But then I wake up in the real world, and if that is a model of the world that
01:04:34 - I'm living in now, it's different in some interesting and important qualitative
01:04:38 - ways from my dream world.
01:04:40 - So, I mean, whereas, would you recognize that there's a distinction there?
01:04:45 - That the waking world that I'm in is somehow closer to maybe this physical world
01:04:54 - that you're permanently separated from.
01:04:58 - So I think you can, and certainly we've talked a lot about the mind-building
01:05:03 - models, but we can also talk about the world being its own model.
01:05:06 - But it's closer in the sense, ideally it's predictive of the patterns on your systemic interface.
01:05:11 - Phase so uh at night you are dissociated from
01:05:14 - your sensory input yeah which means that your dreams spin
01:05:18 - freely and you have no way to invalidate uh your
01:05:21 - hypothesis of the world state that you're in also dreams
01:05:24 - tend to relax your priors which means your model has more degrees of freedom
01:05:28 - than it would have during which gets us back to grounding so when i'm awake
01:05:31 - and when i'm not daydreaming i'm in a grounded state where i can't believe arbitrary
01:05:37 - things or or see cockroaches climbing off the walls, which I did the other night.
01:05:42 - I can't do that because they're just not there.
01:05:46 - So the world is grounding me in this awake state.
01:05:51 - So this makes me closer to the physical world than I would be if I was entirely
01:05:56 - living in a mental simulation.
01:05:57 - Do you think it's a very useful category to say closer?
01:06:00 - I mean, how close are you to quantum mechanics? But I can believe with some
01:06:07 - confidence that my perceptions in my everyday state reflect something about the physical world.
01:06:14 - So it's not that I'm trapped in my mind.
01:06:17 - Let's not mistake the physical world for quantum mechanics. They're two very different things.
01:06:22 - No, quantum mechanics is probably not an ontological theory.
01:06:25 - It's a particular kind of formal theory that allows you to describe low-level
01:06:29 - stuff. The only statement is valid.
01:06:31 - No, what I wanted to say is the world is more weird than quantum mechanics.
01:06:35 - The physical world is probably some computational system that progresses from
01:06:39 - state to state in a particular way, and we are an emergent pattern in this, as physical systems.
01:06:45 - Yeah, but nevertheless, there are things that can't happen in my awake life,
01:06:48 - unless they're true hallucinations. Yes.
01:06:52 - I mean, it's very much constrained what I can perceive by the physical world.
01:06:57 - Ideally, you should only hallucinate things that are matching the patterns on
01:07:01 - your retina, which have predictive value for the patterns on your retina.
01:07:05 - But the things that you're hallucinating are not actually like the things happening
01:07:09 - on your retina or the things that give rise to what happens on your retina.
01:07:13 - But they are interfaces.
01:07:14 - So, I mean, they're interfaces so that physical events can affect my own mind.
01:07:18 - An impact on my mind you might you might you might come to uh interpolation
01:07:23 - between states of the retina that as such have never occurred why not i don't
01:07:28 - see that problem you're very literal in that sense and then the other the other
01:07:33 - problem i have um among the many problems i have,
01:07:38 - with the one i will share with you now is that you agree with this idea of the
01:07:43 - embodied view on mind, so the machine view on mind.
01:07:47 - But I think we disagree very deeply about what machine means,
01:07:50 - because I think I would even accuse you of only thinking about the machine in
01:07:56 - terms of a Turing machine.
01:07:57 - And that's why I think also that computation is so central for you in the way
01:08:01 - you describe these phenomena.
01:08:03 - Well, again, I would see the Turing machine as just a thought experiment as
01:08:09 - a very ideal kind of… Yeah, it's way too powerful.
01:08:12 - We need finite state machines. No, but also physically they don't exist.
01:08:16 - At best it's approximated. It's a such… No, I agree. There are no infinite states.
01:08:21 - Exactly. So therefore, it's only a metaphorical machine.
01:08:25 - What I want to talk about is flash and bond machines that really physically
01:08:29 - exist and have an apology.
01:08:31 - So, isn't there a bias in that sense in analysis if we implicitly define machine as Turing machine?
01:08:39 - Wouldn't it be more useful to just consider scenarios where we're less tied
01:08:44 - to a notion of computation? Would you say the universe is computation?
01:08:46 - No. That's the way you can describe it. Okay. Let's try to disassemble this.
01:08:50 - So first of all, a machine is not part of the territory.
01:08:54 - It's part of your map, right? It's a certain part of the language that I use to describe the world.
01:09:01 - And what I mean by this particular concept in my language is a system that I
01:09:05 - can describe as a state space and a transition function that correlates these
01:09:09 - spaces in deterministic or probabilistic ways.
01:09:12 - And what's characteristic usually about machine descriptions is that the machine
01:09:16 - is able to exactly revisit earlier states.
01:09:20 - So, at least conceptually, it should be able to do that. So,
01:09:23 - it's a particular... But already that is problematic.
01:09:25 - Yes, it is. But this is not our point of contention here. So the important thing
01:09:31 - is machine is part of the language that I use to make sense of the world.
01:09:34 - And mathematics is the set of all languages, right? It's the domain of all possible languages.
01:09:40 - And then there is a set in which you can make proofs, which are the languages
01:09:44 - that you can formalize in a particular way. So we get very interested in those.
01:09:48 - And most of what we do in mathematics is these formal languages.
01:09:51 - And it turns out that in these languages, I can specify things that cannot work,
01:09:55 - like paradoxes, things that are conflicting with themselves,
01:09:59 - like the set of all sets that don't contain themselves. This cannot be implemented.
01:10:03 - And for something to exist, it probably needs to be implementable,
01:10:08 - which means it needs to be realizable in a way that it can produce certain functions,
01:10:14 - that it can perform certain things, right?
01:10:16 - That it can go from state to state and give rise to states at my interface to my environment.
01:10:22 - But I cannot know what that is, that thing, because I only get discernible differences.
01:10:26 - And even this, I don't know exactly if I really get the discernible differences
01:10:30 - that I think I get, right? So I had to make very conditional models.
01:10:34 - And it turns out that mathematics allows me to reason about these conditional
01:10:39 - models if I have evidence that my mind is cohesive enough to afford this reasoning.
01:10:44 - And the good news is the theory that my mind is cohesive enough to perform mathematics
01:10:49 - gives very good predictions.
01:10:50 - It's a very nice theory. So I can do some mathematics with my mind can start
01:10:54 - bootstrapping certain things.
01:10:56 - And one of the first things I discover, that mathematics discovered very late,
01:11:00 - is that there is a part of mathematics that doesn't have contradictions.
01:11:04 - And this is constructive mathematics. And the modern name for this is computation.
01:11:09 - So everything that happens as a computation cannot have a contradiction.
01:11:13 - Your computer can never be in an invalid state.
01:11:17 - And the traditional definitions of computation, which use infinite tapes and
01:11:21 - other things, don't fully realize that implication.
01:11:24 - So this gives rise to the Gödel problem and the Halting problem and so on,
01:11:28 - which are features of language specifications that introduce infinity and timelessness and so on.
01:11:33 - But in computation, these things disappear.
01:11:36 - Now, in computation, a few things are different from classical mathematics.
01:11:39 - For instance, pi is suddenly not a number anymore. It's a function.
01:11:43 - And it's a function that you cannot actually compute because nobody knows the last digit of pi.
01:11:47 - In order to compute the last digit of pi, you would need an infinite amount
01:11:51 - of energy to produce these computations if you happen to live in a reversible
01:11:55 - universe like ours, but you need energy to delete bits, right?
01:11:59 - So you will never know the last digit of pi if you live in the universe that
01:12:03 - you can mathematically deduce to be probably in, right?
01:12:08 - So we are not talking about reality as such. What we are talking about is mathematical
01:12:12 - models of hypothetical worlds. and then we try to discover the class of mathematical
01:12:18 - models that can explain our observations.
01:12:21 - And subjectively, we live in members of that class, but we don't live in the
01:12:26 - low-level members of that class because our brain is not sufficiently detailed to simulate them.
01:12:31 - Your hypothesis is that you do live in a subset of that class because your mathematical
01:12:36 - languages and the statement of mathematics is only true for closed worlds.
01:12:42 - If the boundary conditions are defined.
01:12:45 - And as life forms, the power of life is that we act against the second wave
01:12:50 - of thermodynamics as open systems.
01:12:52 - So there might be aspects of these systems that you cannot capture with this
01:12:57 - set of mathematical languages that you now advance.
01:13:00 - I'm not part of that religion. I think that... No, that's fine. That's fine.
01:13:06 - But still, you must convince me or... I'm trying. But you should explain why
01:13:13 - you believe that this specific language,
01:13:17 - that computational computation would be susceptible for giving us any kind of
01:13:22 - traction to deal with these open systems that we call life.
01:13:26 - I don't think that life is about open systems. The meaning of life on Earth
01:13:31 - is the hydrogenation of carbon dioxide.
01:13:35 - Roughly speaking. Life is cells. The meaning.
01:13:42 - Catastrophes is something that is not necessary. Only in the sense that people
01:13:45 - are asking for the meaning of life.
01:13:47 - Right. Of course, there is no meaning. But in the sense that people are asking
01:13:50 - for the meaning of life, why is it there? What does it do? I'm not asking for the meaning of life.
01:13:54 - I was saying your computational language might not be appropriate to understand
01:13:59 - life, including consciousness and everything in between.
01:14:01 - Okay. I'm trying to show you why I think that is the case.
01:14:04 - Of course. I'm sorry. Yes. Okay. Sure.
01:14:08 - So, what life is about, it sells. At some point in 1650, we didn't know what
01:14:12 - was life in the same sense as people in 1950 didn't really know what was intelligence.
01:14:15 - You could point at it, but you didn't know what it structurally, functionally was.
01:14:20 - So in the same way as I would say today, and you might disagree that intelligence
01:14:24 - is the ability to make models, I would say today, or most biologists would say this, life is cells.
01:14:29 - A cell is a molecular machine that is able to extract negentropy from the very
01:14:34 - right range of environments.
01:14:36 - And it has something built into it that is almost literally a Turing machine.
01:14:40 - You have this DNA tape with a red right head. and this thing implements an operating
01:14:44 - system on the cell which allows the cell to become part of a cellular automaton
01:14:47 - to create an organism. Not necessarily roughing states of that.
01:14:51 - It's reading it, but it can also write a few things. And some of it,
01:14:55 - it also writes in methylation. So we have this epigenomics, the flash drive, yes.
01:14:59 - But neurons seem to be breaking their DNA in particular ways to write something.
01:15:05 - But largely, of course, it's a ROM. We're mostly reading from this.
01:15:09 - And we also write state, but we do this in additional parts in the cell.
01:15:13 - So with methylation and changing the layout of the genome.
01:15:16 - Basically a few bytes that we can write. And what the cell can do is it basically measures its state,
01:15:24 - which is a chemical configuration based on processes that happen inside of the
01:15:28 - cell and processes that chemicals diffuse over the cellular membrane from the
01:15:32 - environment of the cell, which might also contain other cells.
01:15:35 - And based on this, the cell is going to go into a new state.
01:15:39 - And in the process of this, it either regulates something or it differentiates
01:15:43 - itself into another cell, which will behave differently when confronted with
01:15:46 - the same set of possible states.
01:15:48 - Or it might go into apoptosis, or it might divide itself.
01:15:52 - So your cell combines a neck entropy extractor, this state machine.
01:15:56 - Differentiation mechanisms that allow them to participate in evolution, and a self-replicator.
01:16:02 - In the absence of any of these components, your cell is not going to work.
01:16:06 - Once you have these components, which are incredibly complicated and probably
01:16:09 - require an enormous amount of rolls of the cosmic dice, maybe in the same order
01:16:13 - of magnitude as their available planetary surfaces is this in a sufficiently
01:16:16 - isotropic region of the universe, right?
01:16:21 - You have that thing live. You have this evolution of things that grow up.
01:16:25 - And what life can do, what other damp chemical processes cannot do,
01:16:29 - it can produce controlled reactions.
01:16:32 - We are in competition with damp combustion processes and redox reaction on this planet.
01:16:37 - And when they start, we cannot outperform them. And as a fire,
01:16:40 - we lose against the fire.
01:16:41 - The fire is so efficient in closing like entropy gradients. But there are some
01:16:45 - things that we can do, like we can do photosynthesis if you are a cell.
01:16:48 - So you need to add a little bit of energy to harvest more energy in the end.
01:16:52 - And what happened initially was that the cells, some of them discovered after
01:16:57 - a while a way to turn energy from the sun into their own structure by taking
01:17:03 - carbon dioxide from the atmosphere and changing it into their organic molecules
01:17:08 - and putting oxygen into the atmosphere.
01:17:11 - And they did this for a while and were very successful in this.
01:17:14 - And from the geological perspective, this was the major oxygenation event,
01:17:18 - a major event where suddenly you had a lot of free oxygen in the atmosphere.
01:17:22 - We're just agreeing with each other. So Paul said that animals are machines
01:17:29 - that can physically exist.
01:17:30 - You're saying that animals are computational entities, but you're defining those
01:17:35 - in a particular way such that there can be ones that can exist.
01:17:39 - So we're largely agreeing.
01:17:41 - Yeah, so now there happened a singularity at this point. At some point,
01:17:44 - some cells decided to combine together and form organisms, which made them participate
01:17:50 - in the multilevel selection.
01:17:52 - And they mostly became carnivores. They started eating other cells.
01:17:56 - We're not making progress here because the challenge was a somewhat different one.
01:18:00 - Oh, I'm giving you a computational description, and I'm trying to explain to
01:18:03 - you at some point why life is not an open system.
01:18:06 - You don't give me a computational description because you only admitted that
01:18:09 - the state machine of the cell is not a Turing machine, but it might be.
01:18:13 - No, it's just a finite state machine. It doesn't have an infinite tape.
01:18:16 - It's less than a Turing machine.
01:18:18 - It's not more. It's also not something different.
01:18:21 - Computational language might give you leverage in describing certain processes.
01:18:25 - I give you that. Not a problem.
01:18:27 - But you seem to give us some ontological stages. That was my concern.
01:18:31 - But we cannot discuss ontology. What we can discuss is epistemology and metaphysics.
01:18:37 - No, we can because you seem to give it some exclusive descriptive power that
01:18:44 - I want to question a bit more.
01:18:46 - That's why I was challenging you on this notion, is your notional machine a Turing machine?
01:18:51 - Because maybe if we think about, also we were talking about mind,
01:18:54 - synthesizing mind and consciousness, there might be properties to that that
01:18:58 - go beyond what we can capture easily from this computational perspective.
01:19:03 - It's basically, in the limit, it becomes a Turing machine. if you give it unlimited
01:19:07 - computational resources then the tree becomes another level it's also in a sort
01:19:13 - of open system energy is flowing across the boundaries otherwise it would die
01:19:18 - so this is just another way of describing what it's doing.
01:19:24 - So maybe there is a conflict here, but I think in some way that it reflects
01:19:29 - your view of the mind as being pure computation and sort of sitting inside this
01:19:35 - body and this physical world with which you can never have any direct contact
01:19:40 - because it's purely in the virtual sphere.
01:19:44 - Whereas I think Paul, myself, have more of a view that these are really,
01:19:49 - in the end, the same thing.
01:19:51 - The physical world generates the virtual mind.
01:19:55 - Oh, basically what I would say, there are in some sense the same thing,
01:19:58 - but they're described with a different language.
01:20:01 - So basically the language in which you describe mental events and in which you
01:20:05 - experience mental events is a language that is different from the language that
01:20:09 - we use to describe physics.
01:20:11 - So mental events don't exist in the physical realm.
01:20:15 - And it's also not all physical events exist in all languages of the physical
01:20:19 - realm. For instance, wetness doesn't exist at the level of molecules.
01:20:23 - Heat doesn't exist at the level of molecules. Mental events,
01:20:26 - like memory states, are physical events, right? There's an identity relation.
01:20:31 - You will know examples of that, right? From physiology, you look at play cells.
01:20:37 - My challenge is more to say, look, fine, you have a computational language,
01:20:42 - and Turing machines are fantastic, and Alan Turing is really a great...
01:20:45 - Wait a moment. No, no, I'm not convinced of this.
01:20:47 - I'm not convinced. No, no, that's an important distinction. There might be something
01:20:50 - going on here. It's not true.
01:20:52 - So, it could be that there are certain, for instance, that I have memories of
01:20:57 - conscious states that the physical system has actually never been in.
01:21:01 - From my perspective, this is a conscious state that I inhabited because I have
01:21:05 - a memory of having had this in the focus of my attention.
01:21:08 - But you might be able to show that my physical system actually was never there.
01:21:13 - So from my perspective, I have a language that describes something that's subjectively
01:21:17 - totally valid, the experience that is being valid, but it actually never took
01:21:22 - place in the physical universe.
01:21:23 - Because none of your memories ever took place. It's that all reconstructions
01:21:28 - of things and how you would like to think they happen.
01:21:33 - No, I think what you're doing is in order to say if your identity theory,
01:21:38 - you basically come up with this
01:21:39 - derisive notion of folk psychology to invalidate my actual experience.
01:21:45 - I'm saying that none of you… No, my experience is not identical to what happens
01:21:49 - physically. It's something else.
01:21:53 - The nature of memory… Because I have experiences that are not consistent with
01:21:57 - my understanding of physics.
01:21:59 - Yeah, but that's a very introspective argument.
01:22:01 - Yes. But basically, I caught myself editing my own memories,
01:22:06 - and I've had mystical experiences that are probably the result of my brain messing
01:22:10 - up its own memories or editing them.
01:22:13 - It's a necessary part of being a physical machine that you can't go back in
01:22:19 - the past and relive that experience.
01:22:20 - All you can do is reconstruct it, which is what memory is, and it's always going
01:22:24 - to be different from the actual experience.
01:22:26 - Yes, but how do you know that I ever experienced that thing?
01:22:29 - You don't. Yes, so actually... You don't need to. I could have.
01:22:33 - Based on the timing, I know that most of the things that I experience cannot
01:22:39 - physically take place in the way that I experienced them.
01:22:42 - And that's all fine. That's the fallibility of memory.
01:22:46 - No, it's not just the fallibility of memory. Memory is also creative.
01:22:49 - It's constructive. You compress, you build models, right? So,
01:22:51 - okay, big deal. No problem.
01:22:53 - It's basically, it's something else. There is the distinction between a novel and the book.
01:22:59 - If you look at the book and you open it, you see characters in it,
01:23:02 - which are actually just colored parts of the categories. Right. Yeah.
01:23:07 - There is no story in that book except from the perspective of a mind who reads it.
01:23:12 - And a similar thing happens to me I don't exist expect from the perspective
01:23:16 - of a mind who generates me and my own perspective who reflects on what I think I am,
01:23:22 - and that thing is distinct from this thing so there is no identity at some level
01:23:27 - between the novel and the book,
01:23:29 - you could say there is one because the novel doesn't have a being that is distinct
01:23:33 - from the book and the reader metaphor breaks down because your mind is not the story that's in the book,
01:23:42 - it's a bad metaphor because the book always assumes there's a reader.
01:23:46 - So now immediately you have your homogenous problem.
01:23:49 - You don't want to have this homogenous problem if you want to have a scientific theory.
01:23:52 - So I like the metaphor to give a gist of this idea of the problem of natural category is fine.
01:24:00 - That was the idea of that metaphor. So you see that. As such, I accept it. Yes.
01:24:05 - But you cannot push it too far and say, therefore, it's a story I tell myself
01:24:10 - about who I am, me, me as a story.
01:24:13 - I'm worried that you're going too far in an attempt to save identity theory.
01:24:17 - I don't think it's very elegant. I don't think it works.
01:24:21 - Yeah, but that functionalism... Wait, wait, wait. We don't have a disagreement about functionalism.
01:24:25 - There's the weak and strong functionalism, and you can be,
01:24:29 - you know, weakly functionalist and say that you can
01:24:33 - have mental states in things rather than brains but there's
01:24:36 - a lot of constraints around what those sorts of
01:24:39 - things can be uh and uh you know
01:24:42 - i think that's probably where i would be i don't know if paul is an identity theorist
01:24:45 - i think he probably is at least weekly functionalist because he wants to
01:24:48 - build robots who can have mental states so um
01:24:51 - but but then we can discuss what those constraints might
01:24:54 - be and i think we're making some progress on that
01:24:57 - yeah but whereas you're quite strong functionalist so you
01:25:00 - think you could you know take your mind maybe you could put
01:25:03 - it into some very different kind of computer and it
01:25:06 - will still be your mind uh whereas for me maybe for paul it's instantiated by
01:25:13 - brain it's very hard to imagine how it can be instantiated another any other
01:25:16 - kind of device but i'm still a functionalist because you know in theory at least
01:25:21 - it might be possible so uh to take a small step back from this thing.
01:25:26 - When we describe reality, we don't aim for a single narrative, right?
01:25:31 - What we aim for is the map of all possible theories that could explain the observations
01:25:36 - to the degree that we can discover them.
01:25:38 - And then we want to assign confidences in that theory space.
01:25:42 - To our best knowledge and meta-confidences on why we assign the confidences
01:25:46 - if we can, right? So this is somehow the process that we are doing.
01:25:49 - So when we have a disagreement, we should be able to understand the perspective
01:25:54 - of the other one in such a way that we could state it in ways that the other
01:26:00 - would agree that we stated this correctly.
01:26:02 - And then we should be able to exchange arguments about the different confidences, right?
01:26:07 - So for instance, there are people that are embodimentalists in the sense that
01:26:11 - I think that the mind must be the result of some resonance between body and environment,
01:26:16 - and brains are only incidentally present in that whole thing,
01:26:20 - which is a very strong thing.
01:26:22 - But I've met a number of philosophers that are in that camp.
01:26:26 - And then there are people which are strong idealists in the sense that I think
01:26:31 - there is no physical system behind that dream.
01:26:34 - There's nothing that computes God's mind, and I live in God's mind,
01:26:37 - and this is the lowest level of the thing that exists.
01:26:40 - There right and then we have different ideas that
01:26:43 - for instance there's some computational reducibility going on in
01:26:46 - the brain that leads to mechanisms that are so intricate that they produce the
01:26:51 - mind in a way that cannot be functionally abstracted in something that has fewer
01:26:54 - moving parts than the brain and then there's disagreement about how many moving
01:26:58 - parts we need right and this is now it starts to get interesting because now
01:27:03 - we get in the domain of things that
01:27:04 - we can possibly empirically test, right? Fair enough.
01:27:08 - So at this point, we can start maybe identifying interesting disagreements.
01:27:13 - So I cannot know whether it's possible to reduce a brain into functional units,
01:27:19 - for instance, at the level of cortical columns, before somebody has done this, right?
01:27:23 - So it's only a hunch that I'm having.
01:27:25 - And my hunch is currently that we can probably run a person in less than a terabyte
01:27:31 - of computer memory if we had the right kinds of algorithms, which we at the moment probably don't.
01:27:39 - Oh, it's just an upper bound based on the amount of data that you collect in your lifetime,
01:27:44 - a number of concepts that you can form, the number of computations that a neuron
01:27:48 - can make per second, functional organization into cortical columns,
01:27:51 - the kissing number of the cortical columns, and so on, which basically gives you the address space.
01:27:57 - So this is a guesstimate, let's say. Yes, it's all a guesstimate, right?
01:28:01 - It's all a ballpark figure. and when I
01:28:04 - was at MIT and taught a class about the future of AI I
01:28:08 - asked the students in that class to come up with their
01:28:10 - own estimate what they think so before we had discussions about different ways
01:28:14 - of making that estimate and the estimates range between something like 200 gigabytes
01:28:19 - on the low end and several exabytes on the high end depending on what people
01:28:25 - thought the level of granularity is the majority clustered around the low end.
01:28:30 - But of course this doesn't mean that we know that this is the true answer right
01:28:34 - we will only know if and when we succeed,
01:28:37 - in doing this and then we must look at arguments that are not driven by our
01:28:41 - intuitions or human vanity because we are that's a special species that cannot
01:28:46 - be possibly explained by a dumb computer it depends a lot on the physical machine
01:28:51 - in which you're instantiating the person that's right and there's another thing about memory that,
01:28:57 - as we discussed right Also from the model building, it exploits models to be
01:29:02 - regenerative and constructive.
01:29:04 - That's why you can imagine stuff you never experienced. So, so that means it's
01:29:09 - more about what are the parameters of these models, right? And how can you store those?
01:29:14 - Um, but to store a model might be more than just storing just flat bits, right?
01:29:19 - It's also relations now and functions that I have to.
01:29:23 - Yeah, transitional function. So, it might be gone.
01:29:26 - But for me, the question on the theory, we can make the theory space,
01:29:30 - we can start to compare and see what the disagreements are.
01:29:33 - But I think at the bottom of it, the fundamental challenge of any theory is
01:29:37 - to explain something and to make testable prediction and to control stuff.
01:29:41 - These are the three criteria for a theory, right?
01:29:44 - So, with respect to your theory, well, I see we just discussed 5% of it by now,
01:29:51 - but okay, Okay, we can do a follow-up of our podcast.
01:29:54 - We're still on lecture one, I think. Exactly. No, yeah, slide one.
01:29:57 - No, I think we're at slide two now, right?
01:30:02 - What's the prediction? What's the most convincing prediction about the theory
01:30:07 - which has been validated, right?
01:30:10 - And what has been the most fundamental aspect of mind that you think you have explained?
01:30:18 - So that's a difficult thing to say. For me, I've answered many questions that
01:30:23 - I had, and this was mostly conception analysis.
01:30:26 - So, for instance, when I started to go into the field of AI,
01:30:29 - I didn't know how it's possible that a system can have an emotion.
01:30:33 - What does it mean to be in an emotional state?
01:30:36 - Because I thought of an emotional state as something like a parameter within
01:30:39 - the system. What does make this parameter special?
01:30:43 - What makes an emotion different from a number or from a scalar or from a bunch of scalars, right?
01:30:50 - And at some point, I discovered that a way to make sense of this is that the
01:30:53 - emotion is a modulation of cognition.
01:30:55 - So you have parameters that are themselves not emotions, like arousal and valence and focus and...
01:31:03 - The securing rate which tells you you focus your
01:31:06 - attention inwards or outwards and so on and these together
01:31:09 - put you into a particular kind of configuration and this configuration means
01:31:13 - that your cognition is going to work differently you will perceive the world
01:31:17 - as differently you will perceive yourself as differently if you have a self-concept
01:31:20 - and these changes is what we actually mean by emotion so basically i realized
01:31:25 - yes this is explains sufficiently what i meant is this notion of emotion.
01:31:29 - So I could before point at a certain phenomenon, and now this language allows
01:31:34 - me to perform a conception analysis that makes it possible,
01:31:37 - for me to understand, oh, this is actually how emotion is probably implemented
01:31:41 - in an organism as a set of modulators that configure the cognition of that organism
01:31:46 - depending on the environmental and internal circumstances.
01:31:50 - So in this sense, it's not a predictive theory. It's an integrative model.
01:31:54 - And I've done this for a lot of things. For the predictive things,
01:31:59 - you can say, okay, we can come up with a machine learning algorithm that,
01:32:02 - for instance, identifies a cost function.
01:32:04 - You should think that the value of a cognition is proportional to the increase
01:32:09 - in expected reward if you perform it, minus the utility cost of not doing something
01:32:14 - else with these computational units, right?
01:32:16 - So if you implement this principle and you have lots of computational agents,
01:32:21 - do they perform useful modeling or useful computations in a given domain?
01:32:27 - This is an experiment that you can test. And many things in machine learning are in that area.
01:32:32 - But they're not theories of how the mind works. They are theories on how to
01:32:37 - implement a learning system.
01:32:39 - And when you look at reinforcement learning, for instance, many of the breakthroughs
01:32:43 - that we had there were the result of such considerations.
01:32:46 - So in a sense, AI has made many interesting predictions and useful theories.
01:32:51 - But without being able to build a thing that works like us, we will not know
01:32:56 - whether our models are actually models of a mind, or of something else that
01:33:00 - only performs certain functions that are very mind-like.
01:33:03 - The slide that I put up on day one was interesting.
01:33:08 - Said we can have an analytic approach, a synthetic approach.
01:33:11 - The synthetic approach can give rise to an existence.
01:33:14 - And the way I would look at what you're doing, you're following a synthetic
01:33:18 - approach, which is inspired by biology, but not necessarily literature.
01:33:23 - And you're looking to build existence proofs of your ideas that do things which
01:33:29 - animals and humans do. We might call them intelligent.
01:33:33 - And that's how you validate your theories. Yeah, I think that's fair.
01:33:38 - The problem is, if I would use an analytic approach, I would have to produce
01:33:42 - theories which have an enormous amount of free variables.
01:33:46 - And I would be very prone to overfitting of my models.
01:33:49 - It's very hard to make predictive models with very many free variables when
01:33:53 - you analytically look at the brain. So you gave, for instance,
01:33:57 - this example with making a model of the conic tomb of C. elegans, right?
01:34:03 - So we know it has 309 neurons, and we have mapped out the conic tomb.
01:34:08 - Unfortunately, we haven't really mapped out which of the neurons are excitatory
01:34:12 - or inhibitory, because we haven't really gotten to a model to the level of the
01:34:15 - individual vesicles there.
01:34:17 - So the search space is still very large. And if we map this out correctly,
01:34:23 - I think it's a good guess to say that this thing is probably going to behave
01:34:26 - somewhat C-elegance-like.
01:34:29 - At least it would be very surprising if it doesn't. Very interesting.
01:34:33 - But there are good reasons why right now our models are not really good enough for this.
01:34:37 - At least as far as I know, I haven't looked in that space in the last year or
01:34:40 - so and don't know of the new publications that have cracked the problem.
01:34:44 - But you probably would have heard about it, right? Yes. So it's also possible
01:34:48 - that the connectome of C elegans is computationally irreducible, which means that,
01:34:55 - it doesn't compress. It's possible that we don't find a description of what
01:34:58 - C. elegans is doing that is shorter than the full conic tome, right?
01:35:02 - So we would still expect that this thing does what C.
01:35:06 - Elegans is doing, but we cannot say anything beyond this, except it's a function
01:35:11 - which has 309 moving parts and complex interaction between those parts, right?
01:35:17 - Which means our minds are not very well equipped to do a further analysis on that model.
01:35:21 - But also this is a very bold thing, and we probably can compress it a little
01:35:25 - bit and identify some patterns in this and make sense of this on a higher level
01:35:30 - description, right? So this is what we would expect.
01:35:33 - But for me, the main prediction that I'm making right now is that at some point,
01:35:38 - humanity will build machines,
01:35:40 - building the current computational paradigms, which means building finite state
01:35:44 - machines that process information that are able to perform similar intellectual
01:35:48 - feats as we do, if global warming doesn't kill us before that.
01:35:51 - That's basically my prediction. It will be testable in some sense.
01:35:56 - But on a bit longish timescale. Yeah, we don't know that.
01:36:00 - So, there's of course this thing, Daddy, do you think that computers will ever be intelligent? Yes.
01:36:05 - Before I die. A little bit before that. Right, exactly.
01:36:10 - So, Joshua, before we hit the finish line, we made quite a tour through many
01:36:17 - aspects of the mind, how we can model minds, how we can think about it also
01:36:21 - from a theoretical perspective.
01:36:23 - And you have been really working very, very deeply and thinking very deeply
01:36:27 - about these challenges.
01:36:30 - So if we would like to follow in your approach to cracking this hard nut,
01:36:36 - what would be Joshua's law?
01:36:46 - It's a very interesting question what would be my law in the sense what uh what
01:36:50 - strategy should i be using to get to results what strategy should anyone use
01:36:54 - yeah recommend but what should tony be doing tomorrow what should those groups
01:37:00 - be doing instead of the stuff that they're doing now,
01:37:05 - unique direction give us a law,
01:37:09 - Joshua's law it really depends I mean from a career perspective my advice would
01:37:15 - be very different it should be printable on a t-shirt cannot have too many words yeah.
01:37:29 - So uh i think at some level you should
01:37:32 - be aware that you're you are a story that your mind tells itself and when you
01:37:36 - make a model of the world you do this by identifying relationships between information
01:37:42 - the most important rule there is that you have to understand that the meaning
01:37:47 - of information is its relationship to change another information and when you
01:37:51 - try to evaluate your a model,
01:37:52 - the first rule for that is that the weight of your confidence should equal the
01:37:57 - strength of the evidence to support it.
01:37:59 - There are no valid beliefs without priors. You basically need to track all of
01:38:03 - your priors until you form a ring, until you know the rules of your reasoning,
01:38:10 - until you know the basic axioms that go into your mathematical languages.
01:38:14 - And then you try to see if you can find a model that matches the data,
01:38:19 - that matches the patterns that you get and has predictive value.
01:38:21 - And once you've done this, you try to see if there are other rings,
01:38:24 - if there are other self-contained systems that have the same power. How's the law?
01:38:29 - So the law is, in some sense, follow the evidence. Try to make models that track the priors.
01:38:36 - Always follow your priors and resolve them until your belief stops being a verb.
01:38:41 - Because that system is now self-contained and no longer has a relationship to you.
01:38:45 - Once you understand a way to define arithmetic, that definition of arithmetic
01:38:50 - becomes independent of you. Your priors until you're a circular.
01:38:54 - Yes, until basically you get to a mathematically closed system.
01:38:58 - Every mathematically closed system in some sense is a tautology.
01:39:01 - Mathematics is a set of tautologies. Your destiny is to be a solipsist.
01:39:07 - No, why? I don't even exist, how
01:39:10 - can there be a solipsist? A solipsist is somebody who thinks they exist.
01:39:14 - No, that's a good law. I've pretty successfully deconstructed my own existence.
01:39:19 - Don't take that away from me. You're saying what Russell said in that clip,
01:39:24 - focus on the facts, not what you want to believe to be true.
01:39:27 - But in some sense, I realized that Tony and Paul and me are characters in a
01:39:33 - dream that is driven by a higher mind on a higher plane of existence.
01:39:36 - And the best explanation for that thing is that it was created in the brain
01:39:42 - of a primate in the course of an evolutionary trajectory, right?
01:39:46 - So this seems to be a theory that gives good predictions, seems to be sound,
01:39:50 - seems to be compatible with all the solid scientific results I'm getting.
01:39:56 - You're not going to our law now. Oh, no, no. No, the basic law is,
01:40:01 - I think, still this what Francis Bacon discovered, the first law of epistemology.
01:40:05 - The weight of the confidence must equal the strength of the evidence.
01:40:09 - And I think from this follows everything else, if you take this thing,
01:40:14 - because you can take this and check all the alternatives once you have this
01:40:17 - hypothesis that you should be doing this. The first law is Bacon's law. Yeah. Okay.
01:40:21 - Evidence is about… Oh, I cannot claim originality. I'm not a very original mind.
01:40:24 - Okay. I'm not that smart. But then, so the last question to you,
01:40:27 - Tony Wolf will chase you up four years from now.
01:40:33 - You might be in San Francisco by then or somewhere else. We don't know yet.
01:40:37 - Or still at Cambridge. I don't know what you're up to.
01:40:41 - But it sure will be a great place. But Tony is going to check whether you succeeded
01:40:46 - in falsifying or verifying a specific hypothesis that you're going to share with us today.
01:40:52 - So what the most urgent hypothesis that you would like to see really fully tested
01:40:58 - in this four-year timeframe, with respect to your theory of mind?
01:41:03 - Prediction.
01:41:10 - So one of the things that I find very interesting right now is whether the purpose
01:41:15 - of consciousness is attention-based learning.
01:41:17 - So there seems to be a problem. Our brains are not differentiable.
01:41:21 - They are not easily decomposed in a succession of layers through which we can
01:41:25 - pipe a loss function with the chain rule. So something else needs to be going on, right?
01:41:30 - And I suspect what's going on there is, among a few other things that I'm not
01:41:35 - seeing right now, an attention-based learning algorithm.
01:41:38 - And that might work by capturing the current binding state, the current latent
01:41:42 - variables that define the partial configuration of your brain,
01:41:45 - and a change that you are making to this configuration with respect to a particular
01:41:49 - goal and the expected outcome,
01:41:51 - and triggering conditions that tell you when you have a possibility to evaluate
01:41:55 - whether the change was successful.
01:41:57 - And then you commit this to an index,
01:42:00 - and conscious attention might be the ability to make index memories.
01:42:03 - And then you go to the future, you basically experience a few other world states,
01:42:09 - decompose this binding state to do other things, and then at some point the
01:42:13 - result manifests in the world, you reinstantiate this original binding state,
01:42:17 - and based on the outcome of your change, you reinforce the change or you undo it.
01:42:24 - And this could be a learning algorithm that sounds a little bit awkward because
01:42:29 - we need to do so many steps, but eventually we need to touch way fewer links
01:42:32 - than our current machine learning algorithms do.
01:42:35 - And what I would like to see is if this kind of algorithm actually works and
01:42:39 - produces useful results, and if we see a new family of learning algorithms that
01:42:43 - implement this or similar principles.
01:42:45 - Okay, very good. Lucia Boch, thank you very much for this conversation. Thank you too.
01:42:53 - The CSN podcast was produced by the Convergent Science Network of Biometrics
01:42:58 - and Biohybrid Systems, a project funded by the European 7th Research Framework Programme.
01:43:06 - For more interviews, recorded lectures or upcoming conferences in the field
01:43:12 - of biometrics and biohybrid systems, go to csnnetwork.com.
01:43:18 - Music.

CSN Podcasts

Joscha Bach on artificial general intelligence and deep learning

Season 2019

Description

CSN Podcasts

Episodes

Yoram Vodovotz on inflammation and immune system

Vincent Hayward on haptics and touch

Viktor Jirsa on epilepsy and virtual brain

Stuart Wilson on self-organization and cortical maps

Marco Diana on addiction and dopamine

Timestamp

Be the first to leave a comment

Leave a commentCancel

Legal

Follow us on

Joscha Bach on artificial general intelligence and deep learning

Description

Episodes

Timestamp

Be the first to leave a comment

Leave a commentCancel

Tags

Legal

Follow us on

Login to enjoy full advantages

Please login or subscribe to continue.

Go Premium!

Enjoy the full advantage of the premium access.

Stop following

Cancel subscription