Metabolic Machines: Computation as a Living Process

WEBVTT
00:00:00.160 --> 00:00:09.439
I am absolutely wearing my investor hat in trying to explore the domain so that I can see what I think is the more exciting, compelling future beyond where we are.
00:00:09.519 --> 00:00:22.239
So if I had uh money, I'd be investing uh uh at the intersection of all of those domains active inference, bioelectricity, and um um um um uh energy-based uh compute models.
00:00:22.640 --> 00:00:33.759
And I'd be trying to do that with one foot in biology and one foot in AI because it it will transform our understanding of biological systems if we get it right on the compute side.
00:00:33.920 --> 00:00:46.079
And so that's that's one of the other reasons I'm interested because I see this huge focus of the superscaling AI companies trying to put their LLMs to work to solve biology.
00:00:46.240 --> 00:01:07.840
But uh the LLMs are fantastic um induction machines, uh, and traditional good old-fashioned AI is fantastic deductive machinery, but um they are not anticipatory systems and they can't project outcomes into the future and then direct themselves towards achieving those outcomes.
00:01:08.000 --> 00:01:24.159
The biological systems that I'm describing are anticipatory systems and they can do that, and so I think we're seeing a potential, a potentially much more compelling pathway towards uh a more exciting AGI beyond where we currently are, and of course that that's why I'm excited by your work.
00:01:24.400 --> 00:01:41.200
The more I look at the mirrored uh architecture and the machinery, uh horizontally opposed machinery on the mitochondrial crystal, the more it looks like a complementary transistor amplifier, amplifier architecture to me.
00:01:41.359 --> 00:01:51.120
And as I mentioned, each of those complexes that's um pumping protons or allowing electron flux through the membrane has a metal center.
00:01:51.439 --> 00:01:55.680
And so I think that is the transistor moment.
00:01:55.920 --> 00:02:14.719
We are looking at an architecture in a biological system that is equivalent to, although two, three billion years uh prior to, the complementary transistor amplifier circuit, which was the foundation for the modern computing industry.
00:02:15.039 --> 00:02:21.919
Um, that's a claim, that's a bold claim, that's an irrational claim, that's a um foolhardy claim, possibly.
00:02:22.080 --> 00:02:38.479
Uh, but I've I spend my time you know trying to push the boundaries here, really, so that we can, you know, a challenge attack, but first I think, uh, and this is what I love about Mike Levin's work, first, don't attack the straw man, build the steel man, and then attack the steel man.
00:02:38.639 --> 00:02:46.560
I think that's what is appropriate, necessary, and I think um will yield incredible uh new insights.
00:02:46.719 --> 00:03:07.120
But I think um this is the transistor moment because the system can be characterized, even if it's an abstract, highly stylized model I'm building, it will enable uh uh experimental scrutiny, but also it will describe uh ways for us to explain biological phenomena, I hope, in a fundamentally different way.
00:03:09.840 --> 00:03:17.360
And uh if your computing paradigm or you know this approach would succeed, uh what are the implications?
00:03:17.599 --> 00:03:32.800
But you know, um I also I think time because we mentioned it, time plays a role here because you know uh if it will succeed before the singularity or after, I think it's of the essence as well.
00:03:32.960 --> 00:03:36.560
Or maybe you think we cannot achieve the singularity without your approach.
00:03:36.800 --> 00:03:48.479
I mean, tell me what are your thoughts there with relative to what's happening today with uh AGI and all these exponential developments relative to your paradigm?
00:03:49.840 --> 00:04:02.719
Well, I would say um, and and and by the way, you know, this is your area of expertise, Mikhaila, um, not really mine, uh, but uh, you know, umbodied con embodied cognition, you know, soft robotics.
00:04:02.879 --> 00:04:12.400
Um uh I don't think we're gonna achieve the singularity without uh embodiment, uh, if indeed it's ever possible, and I'm skeptical.
00:04:12.639 --> 00:04:20.800
Uh uh, but actually, when we think about embodiment, embodiment is not the current approach, which is you know the pre-training moment in robotic systems.
00:04:20.959 --> 00:04:29.839
We're not downloading LLM propositions onto uh onto uh robotic structures and thinking of them as embodied.
00:04:29.920 --> 00:04:56.720
Embodiment means autopoesis, self-instantiation, self-maintenance, self-improvement uh over time scales that might not be perceptible uh to us as humans, considering about evolution, considering evolution, and this is you know really just briefly referencing the amazing work of the Agential Biology Institute, trying to understand evolutionary paradigms beyond the neo-Darwinist uh genome first, genome is everything proposition.
00:04:56.879 --> 00:05:51.360
Um, but I think that uh if you consider every single cell of an embodied agent as its own computational uh uh standalone agential computational architecture, and then you think about the uh individual cell as having 10, a hundred, a thousand, tens of thousands of mitochondria in it, and then you think about the activation function that is performed by an ecology of mitochondria within a cell, the complexity, and then you consider that you know we have hundreds of trillions of cells in our body and hundreds of trillions of mitochondria in our body, then we're looking at orders of scale very different from the orders of scale we're looking at, and we're looking at each individual uh embodied component of the system being itself a cognizing agent, interacting with its other agents.
00:05:51.519 --> 00:06:01.600
Then all of a sudden, I think we start to see a different pathway towards embodied cognition, uh, and and who knows, over time perhaps towards the singularity.
00:06:01.759 --> 00:06:18.959
Metabolism's inference means Hello everyone, and welcome to our Dream Wonders series on Mindplex.
00:06:19.839 --> 00:06:35.759
Our guest today is Anthony Finbo, who has followed an extraordinary intellectual and professional trajectory, if I may call it like that, because you know, his background is in aircraft engineering of all.
00:06:36.720 --> 00:06:47.040
But his background originally is in cybernetics and control engineering, which is also my own undergrad and graduate studies.
00:06:47.519 --> 00:06:54.720
Um, the intellectual lineage, actually, from which much of modern artificial intelligence emerged.
00:06:55.120 --> 00:07:12.480
And uh, as I mentioned, he began his career in aircraft engineering and subsequently moved through law, banking, and venture capital into an impressive career, building, scaling, and successfully exiting several Nasdaq listed deep tech businesses.
00:07:12.720 --> 00:07:25.839
And today he works uh on his passion across scientific research, entrepreneurship, and investment, pursuing ideas at the frontier of biological and artificial intelligence.
00:07:26.000 --> 00:07:32.959
And this is where his engineering origins return in a fascinating way that we are about to explore today.
00:07:33.120 --> 00:07:35.439
So, Anthony, welcome to the podcast.
00:07:36.319 --> 00:07:39.040
Michaela, thank you very much for inviting me to speak today.
00:07:39.199 --> 00:07:45.279
Delighted to um speak with someone with such an august and exciting background as you uh indeed.
00:07:45.360 --> 00:07:48.639
And thank you very much for uh wonderful introduction.
00:07:48.720 --> 00:07:51.680
Uh, I I feel like I recognize myself in something you said.
00:07:52.720 --> 00:07:53.759
Thank you, thank you.
00:07:53.920 --> 00:08:01.120
You know, I'm trying because I mean uh we will dive deeper into your amazing background, of course, during the conversation.
00:08:01.920 --> 00:08:06.720
And I wanted to keep that uh short, and I know it didn't do complete justice, but okay.
00:08:07.040 --> 00:08:28.319
So it didn't work uh what actually drew my interest in this conversation was your research work on a novel computing paradigm, which you coin metabolism as inference, and in which you ask, what have we lost by abstracting computation away from the physical substrate in which it occurs?
00:08:28.800 --> 00:08:45.120
And you know, you are proposing a path to a more biologically plausible AI by looking at the deep fundamentals of biology, and what I find really stunning, yeah, you are exploring the cognitive capabilities of mitochondria.
00:08:45.519 --> 00:08:52.000
So let's I'd like to start, you know, first of all, uh with a clarifying or a few clarifying questions.
00:08:52.159 --> 00:08:55.039
So, what is the ultimate goal of your pursuit?
00:08:55.120 --> 00:08:59.759
And why do you think a new computing paradigm is needing to needed today?
00:09:02.000 --> 00:09:03.600
Well, that's a great question.
00:09:03.679 --> 00:09:04.720
What's the ultimate goal?
00:09:04.879 --> 00:09:12.000
I think uh uh you describe uh in very gracious terms, Mihala, my career uh as a super interesting trajectory.
00:09:12.080 --> 00:09:14.559
I'd describe it perhaps more like a random walk.
00:09:14.639 --> 00:09:23.440
Um, and uh I continue on that random path, just uh, you know, exploring um things that are exciting and interesting.
00:09:23.679 --> 00:09:39.360
Um but I think uh what I would say is, if I may, just to kind of tune a little bit for the listener, uh, why it is that I find myself exploring these concepts uh at the nexus of biology and um artificial intelligence.
00:09:39.519 --> 00:09:48.159
Um you right you rightly say that my backgrounds in cybernetics, control engineering, and uh, and uh, in fact, avionics in aircraft engineering, signal processing.
00:09:48.320 --> 00:09:51.039
That really characterizes the first 10 years of my career.
00:09:51.279 --> 00:10:00.399
And the last 10 years I've been focused on trying to build software models of microbiome and host microbiome interactions.
00:10:00.559 --> 00:10:07.360
Um, broadly for the use uh for the uh for the listeners who are not familiar, the microbiome is that ecology of microorganisms.
00:10:07.519 --> 00:10:26.559
Uh, you know, we we when we talk about it generally, we talk about the ecology of organisms in the gut, the hundreds of trillions of bacteria, archaea, fungi, etc., that uh cohabit in our our guts and um metabolize quite a lot of our foods and provide some of the substrate we require to subsist.
00:10:26.799 --> 00:10:47.279
And uh I was trying to model with an amazing team uh the interactions of that ecology of microorganisms, which you can describe describe as a point cloud or a distributed um uh um liquid brain almost, um, uh the interactions of that system with a host organism.
00:10:47.360 --> 00:10:49.200
And the host could be a human host.
00:10:49.360 --> 00:10:52.399
They were talking about microbiome in the gut and the human host.
00:10:52.559 --> 00:11:02.399
It could be the scalp, it could be the underarm, it could be the oral or indeed genital um uh areas where niche microbiomes uh evolve.
00:11:02.559 --> 00:11:11.600
Um it could be a plant, root, and the soil microbiome, it could be an animal gut and the animal interactions.
00:11:11.840 --> 00:11:23.360
And um I was modeling those uh interactions and working with an amazing team of engineers to model those interactions using higher order network science.
00:11:24.480 --> 00:11:30.399
But sorry to interrupt, I just want to know what brought you to that uh microbiome exploration?
00:11:30.639 --> 00:11:50.480
Because the question was mainly why it's uh you know, I know it's uh it's random path, and you didn't have an ultimate goal, maybe, from what I understand, uh from what you shared so far, but but why microbiome from engineering and aircraft engineering?
00:11:52.000 --> 00:11:56.559
You know how um throughout our lives we have uh opportunities, we also have trials.
00:11:56.720 --> 00:12:08.159
Well, um, you know, I was building technology businesses, I was building um software businesses in the uh communications space, uh uh, you know, through the internet boom and beyond.
00:12:08.240 --> 00:12:20.960
Um and you know, practically killing myself, to be honest, uh, on an airplane all the time, buying businesses, uh uh, building downsizing, trying to survive the uh internet bust.
00:12:21.200 --> 00:12:32.080
Um in my first CEO role uh within the tech business, um, uh a business that was listed, dual listed on the London Stock Exchange and the Nasdaq, just at the most difficult of times.
00:12:32.159 --> 00:12:33.600
You remember the times, perhaps.
00:12:33.759 --> 00:12:46.960
Um anyway, I I I ended up getting sick uh with a life-threatening uh autoimmune condition, um, which quite literally took me to within uh you know hours of expiration.
00:12:47.519 --> 00:12:50.159
And I made a full recovery, uh, I think.
00:12:50.240 --> 00:13:00.879
Uh, but um on the out, so to speak, uh, I had completely lost interest in man-made machines and was interested in uh biological systems.
00:13:00.960 --> 00:13:11.440
And uh and indeed uh when I was recovering in in my hospital bed, uh I picked up a copy just by chance of The Economist and and read an article about uh the microbiome.
00:13:11.840 --> 00:13:33.440
And what I read in there was fascinating because it talked about uh the article talked about um a particular bacterium, Helicobacter pillori, uh, and how uh people with you know a preponderance of that um bacterium in their in their duodenum uh you know were prone to duodenal ulcers.
00:13:33.759 --> 00:13:40.879
And the um therapeutic that was discovered uh to solve for that challenge uh was antibiotic therapies.
00:13:41.039 --> 00:13:49.679
And people then with the who'd made the recovery and had the antibiotic therapies often experienced autoimmune challenges subsequently.
00:13:49.919 --> 00:14:02.080
And in that article, I read that um, you know, once the Helicobacter pylori has been eradicated, then all of a sudden there's some correlation with the detuning of the immune response.
00:14:02.320 --> 00:14:05.120
And and that I found absolutely fascinating.
00:14:05.279 --> 00:14:23.840
Um and and then I didn't know what I was gonna do with that, but as I recovered, um uh I got opportunity, I didn't know what I was gonna do to be frank, uh, but uh um after a sort of um exploratory couple of years when I worked in uh Venture Capital for a period and didn't really like it, I'll be honest.
00:14:23.919 --> 00:14:32.399
Um I I got the opportunity within the space of a week to join two companies that were focused on biology.
00:14:32.559 --> 00:14:39.440
One was uh an atomic force microscope manufacturer that was looking at um mechanoforces at the nanoscale.
00:14:39.519 --> 00:14:56.879
Uh, and the other was a company called Eagle Genomics, which at the time uh was uh a services business looking at uh wrapping services around some open source software that had been developed by the European Bioinformatics Institute to explore variant calling in human DNA.
00:14:57.120 --> 00:15:07.759
Um and uh I pretty quickly realized um, and I was by the way introduced and parachuted into that business when it was in a challenging period.
00:15:07.919 --> 00:15:27.120
Um uh uh I was parachuted in uh first as a consultant, then as chairman, and I realized there was just no way to kind of compete with the large, you know, well-capitalized venture-backed US businesses that were trying to build out uh uh you know human DNA exploration software products.
00:15:27.279 --> 00:15:40.960
Um and I vectored the business there to the microbiome, not uh selfishly because of my interest, but because there were customers and potential customers that were very interested in trying to understand that microbiome challenge.
00:15:41.279 --> 00:15:53.440
And um it just seemed to me to be, you know, strategically speaking, the most proximate uh uh um path to take to you know try and build a sustainable, sustainable business.
00:15:53.679 --> 00:16:06.720
Which we did and uh you know, built it over a number of years, built some fantastic product uh and built some wonderful customers, uh, but uh ultimately through COVID uh was challenged with uh building out the capability.
00:16:06.879 --> 00:16:30.720
And then, of course, the LLM phenomenon came, and all of a sudden, all attention vectored from higher order uh hypergraph and simplicial uh systems and uh persistent homology uh in exploration of higher order networks to straightforward LLMs, and I couldn't get interest in what it was that we were building, so I I couldn't get a scale up capital.
00:16:30.799 --> 00:16:40.480
Uh and so I stepped back um from that about uh two and a half, three years ago, and you know, speaking very frankly, Mikhaila, I had no clue what to do.
00:16:40.559 --> 00:16:50.480
Uh I was exhausted again and uh I sat for a period of months just uh staring out of the window in front of me at the moment, just not knowing what to do.
00:16:50.559 --> 00:17:14.079
And then all of a sudden a wave of the learnings from the first part of the career just came flooding back into my mind, uh the signal processing, and um and I just happened to be uh looking at some images of the mitochondrial uh um electron transport chain and ATP synthase uh system.
00:17:14.720 --> 00:17:42.559
And I became interested, I bought the book on bioelectrics, uh, you know, um uh um and uh started to read up on the just out of curiosity, started to read up on the structure and and and then instead of seeing uh as was characterized in the text um a turbine or a pump um uh as a characterization of an ATP synthase enzyme, I saw a voltage-controlled oscillator.
00:17:42.799 --> 00:17:59.839
And then I worked backwards from that towards the electron transport chain, and all of a sudden I I saw a cascading operational amplifier structure, and not the biological um uh uh complexes described in the um in the in the text.
00:18:00.240 --> 00:18:08.079
And so I just thought, well, hang on a second, if I've got these two components, then I think I can see a phase lock loop.
00:18:08.400 --> 00:18:10.559
What if this is a phase lock loop?
00:18:10.799 --> 00:18:15.039
Uh and and then I said, well, a phase lock loop is a clock.
00:18:15.359 --> 00:18:20.160
Uh and it's the fundamental clock structure that used in any von Neumann compute architecture.
00:18:20.400 --> 00:18:27.279
So what if there are parallels between what I think I can see here and the compute architecture?
00:18:27.519 --> 00:18:29.279
And so that's the kind of genesis.
00:18:29.359 --> 00:18:56.319
And then for the last two and a half years, obsessively, I've been just exploring all of the components of the system, learning uh from whichever angle I can what uh has been written about the uh structures, and then just trying to build the model that is then worthy of further exploration, and um and then now trying to move towards uh empirically validating some of the uh some of the conjectures.
00:18:57.519 --> 00:19:07.440
Amazing, really, and you already answered uh one of my next questions, which was what was the aha moment that led you or triggered this idea?
00:19:07.680 --> 00:19:09.920
And so this is really fascinating always.
00:19:10.000 --> 00:19:18.880
We like to hear such stories because many times it's like inspiration, you know, and at our AGI 26 conference, actually, Mike Levin.
00:19:19.200 --> 00:19:26.000
Mike Levin spoke about inspiration and and how it comes from his platonic spaces and so on.
00:19:26.079 --> 00:19:26.880
It's fascinating.
00:19:27.119 --> 00:19:42.400
Anyway, uh, yes, so so uh I don't want to call it divine, but by the way, it really was an aha moment, and it was an aha moment, and uh I felt the hairs on the back of my neck, uh the apocryphal story uh uh stand up.
00:19:42.559 --> 00:19:53.759
And um, and then over a period of about a month, uh successively, generally on a Friday morning after a shower, uh answers would come to the questions I hadn't even posed.
00:19:53.920 --> 00:20:00.000
But um uh you know, successively, uh over a period of months I started to piece the architecture together.
00:20:00.079 --> 00:20:08.880
And then I thought to myself, well, you know what, you're not a researcher, you're not a biologist, um, uh, your signal processing uh knowledge is tired.
00:20:09.039 --> 00:20:11.759
Um uh you better you know brush it all off.
00:20:11.920 --> 00:20:25.279
But at the same time, I thought I need to try and attract the interest of um bona fide scientists and researchers in the various fields that can help me uh actually evidence uh what I what I'm describing here.
00:20:25.440 --> 00:20:36.960
And um I I uh I I engaged uh uh I met um uh a researcher called Thomas Parr uh at a at a conference on active inference.
00:20:37.119 --> 00:20:43.839
I became very interested in active inference because it it looked to me to be the worthy successor, the natural successor of cybernetics.
00:20:43.920 --> 00:20:52.240
Um and I met Thomas um at a conference and started talking to him about my ideas and and and we developed a conversation over a period of months.
00:20:52.319 --> 00:21:00.640
Um and Thomas, by the way, wrote um the first uh definitive text uh on active inference uh with Carl Friston.
00:21:01.039 --> 00:21:22.640
Um and um so I I became very familiar with that book and um and then um cascading series events led me to present my first uh scruffy presentation deck on the conjectures to the Active Inference Institute's theoretical neurobiology group.
00:21:22.720 --> 00:21:32.000
Um, and Carl was in attendance, and uh he very kindly gave me some pointers and said, Well, I think you need to be speaking with Mike Levin and Nick Lane.
00:21:32.319 --> 00:21:39.200
Uh we collectively are working on a model trying to understand the dynamical causal model of the mitochondrion.
00:21:39.359 --> 00:21:40.880
It looks like you've got a model.
00:21:41.119 --> 00:21:46.799
Um you know, uh w you should engage with the other two uh um researchers.
00:21:46.880 --> 00:21:55.839
So all of a sudden I found myself in the mix uh speaking with some of the world leading scientists at this uh uh you know Nexus most compelling, exciting uh domain.
00:21:59.200 --> 00:21:59.759
I can say.
00:22:00.319 --> 00:22:07.680
Yeah, like like you were guided to actually end up where you ended up.
00:22:07.759 --> 00:22:09.759
And I think it's really fascinating.
00:22:10.079 --> 00:22:13.200
And thank you for really sharing this story anyway.
00:22:13.359 --> 00:22:20.559
I'm gonna get back to that because I have a thread here which I like to help our audience to follow.
00:22:20.799 --> 00:22:28.559
But um, you know, this is we will also post, if you allow us, is a link to that conversation which you had with Carl Freestone's group.
00:22:28.640 --> 00:22:30.480
And of course, you had a presentation there.
00:22:30.640 --> 00:22:35.200
We'll post that in the uh podcast description when we release it.
00:22:35.440 --> 00:22:46.960
Uh, but uh so you know what I'd like to uh dive a bit into right now is how is your approach different from how we treat computation today?
00:22:49.359 --> 00:22:53.920
Well uh you know, the remarkable thing is I don't know that it's different.
00:22:54.079 --> 00:23:03.759
I think uh but but I'll say that there are some fundamental challenges and problems um uh with the current approach to evolving uh artificial intelligence models.
00:23:03.920 --> 00:23:05.119
And the fundamental challenges.
00:23:10.640 --> 00:23:15.519
What is your contribution, yes, uh from from what we already have today?
00:23:17.119 --> 00:23:34.720
So I would say this uh uh distilling it uh to its most essential metabolism's inference means actually considering resource, food and energy and oxygen, as transposable immediately into compute.
00:23:34.960 --> 00:23:39.279
Whereas in the current architectures, resource is ignored.
00:23:39.440 --> 00:23:56.799
Uh and that has led to the psychosis, which is that we're building the ever larger data centers uh to uh enable the models to run and energy and constraints, you know, the energy constraint is not taken into account in describing compute function.
00:23:57.119 --> 00:24:02.079
And as long as that's the case, in my opinion, we won't have rational compute.
00:24:02.319 --> 00:24:21.759
And so that's my contribution essentially, trying to evidence that uh the incredibly efficient structures in the natural world, uh, particularly those machines, if I dare say machines, and I'm I'm not a sort of computational functionalist, uh, so I use that term um with uh I use that term tentatively.
00:24:21.920 --> 00:24:43.359
Uh but um unless we understand how nature has achieved this incredible efficiency in transposing resource into compute capability, we are missing the primary vector I think that we should be pursuing in order to us for us to move towards this more biologically plausible artificial intelligence.
00:24:45.599 --> 00:24:54.640
Uh and this is really very uh kind of the crux of the matter here, and and an answer to my question, I'd like to dive a bit deeper into that, you know.
00:24:54.880 --> 00:25:04.799
Like um coming back to the engineering thing, what does a cell have to sense about itself in order to remain itself?
00:25:05.039 --> 00:25:05.839
That's one question.
00:25:06.000 --> 00:25:21.759
But on the other side, you know, for engineers, your phrase metabolism as inference immediately evokes feedback control, for me at least, you know, uh immediately resonates with my own academic routine, dynamical systems, control theory, self-organization.
00:25:22.240 --> 00:25:27.759
Um, I know you also bring in Boltzmann machines from having listened to your presentation.
00:25:28.000 --> 00:25:46.000
But I want to tell you that my PhD advisor prevented me from taking biological systems because he thought they are too difficult because the biological system is simultaneously controller and sensors and even capable of rebuilding the very machinery performing the control.
00:25:46.400 --> 00:25:58.400
So, you know, a living system uh observes perturbations, estimates hidden states, anticipates demand, acts, and continually returns towards a viable region of state space.
00:25:58.559 --> 00:26:02.799
But could I formulate metabolism as a non-linear adaptive control problem?
00:26:02.960 --> 00:26:04.880
I mean, is this kind of your approach?
00:26:05.039 --> 00:26:08.880
Or please explain to us what is actually the approach.
00:26:10.480 --> 00:26:14.400
Are you sure, Mihaila, you didn't uh do the biology electives?
00:26:14.480 --> 00:26:18.079
You just characterized the system exactly as I would uh myself.
00:26:18.160 --> 00:26:21.039
Uh it is an adaptive nonlinear control system.
00:26:21.440 --> 00:26:31.279
Well, you know, with my engineering control engineering mind and listening to your presentation, I think I kind of figured out a bit, but thank you.
00:26:31.599 --> 00:26:32.480
Go ahead.
00:26:33.039 --> 00:26:59.279
Well, uh I think um yes, uh, and of course the model I've built uh so far is a is a is a radically uh simplified uh abstracted stylized model of one uh of the uh structural elements of the mitochondrion that I think is um worthy of very much deeper exploration.
00:26:59.440 --> 00:27:10.799
Um you know, the mitochondrion, if I may, for your audience, just uh briefly uh you know I spent two and a half years learning about this, so I can uh perhaps you know simply just introduce the concept.
00:27:10.880 --> 00:27:15.200
The mitochondrion, you know, two two plus billion years ago was a standalone organism.
00:27:15.519 --> 00:27:20.720
It was a um a bacterium related to cyanobacteria.
00:27:20.960 --> 00:27:47.200
Um uh uh and it um uh supposedly uh with the great uh oxygen catastrophe, when well rather when cyanobacteria started to pump more oxygen as a byproduct, as a waste product into the atmosphere, uh at a certain point the levels of oxygen in the atmosphere became such that the previous life forms that were anaerobic uh uh were tested and stressed.
00:27:47.359 --> 00:28:03.119
Uh at which point probably there was some kind of conjugation of the archaeon and the bacterium to create the first universal common ancestor of what is now the basis for all life you can see, the eukaryotic cell, the cell with a nucleus.
00:28:03.920 --> 00:28:12.160
Um and uh so the question is um who's in charge, in my opinion?
00:28:12.240 --> 00:28:18.000
You know, is the cell the master of the mitochondrion or is the mitochondrion the master of the cell?
00:28:18.240 --> 00:28:22.079
And what is the function that the mitochondrion performs for the cell?
00:28:22.240 --> 00:28:41.359
Uh and it's you know generally characterized as uh the uh generating ATP, uh the energy currency of life, but in lots of ways that's a byproduct of the uh control system that's running, I think, uh, to generate energy for the cell.
00:28:41.599 --> 00:28:54.319
Um the amazing thing about mitochondrion and mitochondrial study is it's incredibly well characterized, but generally uh through a biochemical lens or through an electrostatic lens.
00:28:54.480 --> 00:29:05.440
And so I just wanted to think about the electrodynamic and indeed hydrodynamic phenomena that might be at work uh to enable the system to regulate its function.
00:29:05.759 --> 00:29:17.599
And so, yes, uh I am looking at a system which uh has got two components which probably independently arose but then came together uh on the mitochondrion.
00:29:17.680 --> 00:29:32.559
The ATP synthase, which is effectively the rotor, the spinning rotor that enables uh when it's operating in one direction, uh enables protons to be pumped out, and when it's operating in the other, uh enables them to flow back in.
00:29:32.720 --> 00:29:42.720
And then the electron transport chain, which uh allows electrons to flow from highly reduced chemical substrate NADH and FADH2.
00:29:43.279 --> 00:29:55.839
Uh electron donors uh donate electrons that are pulled through these successive complexes to oxygen at the other end with the byproduct that water and carbon dioxide are produced.
00:29:56.079 --> 00:30:04.799
That flux of electrons through the membrane results in protons being pumped onto the outer onto the uh surface of the inner membrane.
00:30:04.960 --> 00:30:08.160
And then they flow back through the ATP synthase.
00:30:08.319 --> 00:30:29.599
And so there you have these components of the control system, and they are, and what's also remarkable about the uh structures is the individual complexes within the electron transport chain, they they they all they they most of them have metal centers, so they look like semiconductors, of course, and they behave like semiconductors.
00:30:29.680 --> 00:30:52.960
The membrane itself, and you've you've you've um uh uh interviewed uh my wonderful uh collaborators um Marco Cavaglia and Tommaso Faro, they are looking at membrane dynamics with Jack Duchinski and um looking at and I have to say now that you mentioned them, uh that uh Marco and Tommaso have been three times with us on the podcast.
00:30:53.039 --> 00:30:57.759
And we will also post uh in the description my conversations with them.
00:30:58.079 --> 00:30:59.519
So thank you for mentioning that.
00:30:59.680 --> 00:31:00.160
Go ahead.
00:31:00.319 --> 00:31:04.799
Just want intervention for our audience that they find that in the description.
00:31:05.839 --> 00:31:17.759
Yes, so so they're looking at uh membrane dynamics and uh membrane uh oscillations and interactions of membranes with vicinal water in the cell.
00:31:17.920 --> 00:31:34.319
Uh and they're looking at the cellular level, and I'm looking at the mitochondrial level, and I reached out to them because I saw your podcast and just thought, well, you know what, uh there's huge affinity between what we're talking about here, and subsequently, we've been in a wonderful conversation uh since then.
00:31:34.559 --> 00:32:05.200
Um but um then I look more closely at the structure, and uh I as I said, I I my thesis was that this is um, you know, even the primitive uh circuit, the individual electron transport chain and an individual ATP synthase enzyme together uh is locked in a phase lock loop uh and if it is, it requires certain components in order for it to function.
00:32:05.599 --> 00:32:21.440
The phase lock loop uh in essence takes a reference signal, an oscillation, let's say a sine wave, and compares it with another signal uh which is produced later in the circuit by the voltage-controlled oscillator.
00:32:21.759 --> 00:32:30.720
If the two signals are in phase and of the same frequency, uh the next component of the system, the phase detector, generates no output.
00:32:31.119 --> 00:32:35.759
If the signals are of a different frequency and phase, it will generate an output.
00:32:35.839 --> 00:32:47.440
And that output is a pulse that is proportional, proportionate to the phase and frequency difference between the reference phase reference wave and the uh uh output wave.
00:32:47.920 --> 00:32:57.359
That pulse output is then passed through a low pass filter to generate a voltage output, a more or less smooth voltage output.
00:32:57.519 --> 00:33:01.119
And that voltage output is the input to the reference oscillator.
00:33:01.359 --> 00:33:10.559
So if a higher voltage, uh if there are more pulses because there's a difference between the uh um reference and the output frequency.
00:33:10.640 --> 00:33:16.720
If if there's a big difference, there'll be fatter pulses emanating from the phase detector.
00:33:16.960 --> 00:33:21.599
That means from the loop filter, the low pass filter, there'll be a higher voltage.
00:33:21.759 --> 00:33:34.880
That higher voltage will drive the driven voltage-controlled oscillator to increase its output frequency, which will be fed back to the phase detector, that will reduce the error signal and reduce the pulses out.
00:33:34.960 --> 00:33:37.119
And the system will come into equilibrium.
00:33:37.519 --> 00:33:48.559
Uh, interestingly enough, the system I've described uh is utilizing what's called in electronic equivalent circuit uh language a f a type 1 phase detector.
00:33:48.799 --> 00:33:58.400
Uh and um uh I actually believe that the system operating on the mitochondrial membrane utilizes a type 2 phase detector.
00:33:58.480 --> 00:34:08.800
And I'll perhaps uh if it's of interest, I'll explain why that is absolutely key and germane to the efficiency of the system and to its cognitive capabilities uh also.
00:34:09.199 --> 00:34:21.519
Yes, and and I don't want to run into the risk of getting too abstract here for my audience because we will post your conversation with Carl Freestone, which covers a lot of uh of this.
00:34:21.920 --> 00:34:29.679
You know, I'd like to kind of keep it uh as simple as possible and not too many details, if if you may.
00:34:33.360 --> 00:34:36.400
Let me Mahayla just uh and I didn't mean to interrupt you there, apologies.
00:34:36.480 --> 00:34:40.159
Let me just um make it relevant for your audience.
00:34:40.480 --> 00:34:44.239
So uh the difference between these two frequencies.
00:34:45.039 --> 00:34:46.320
Okay, thank you.
00:34:48.239 --> 00:35:04.000
The difference between these two frequencies and the effort to minimize the difference is the same as variational free energy minimization, which is the same as relative entropy minimization, which is the same as Kobeck-Libler divergence minimization.
00:35:04.400 --> 00:35:36.239
And so uh KL divergence or relative entropy is used in all of the models to describe how a probability distribution compared with another probability distribution uh generates uh you know some uh cross some some relative entropy, some um information that the system then seeks to minimize in uh in generating its output, generating its its result.
00:35:37.199 --> 00:35:39.920
Does that make it a little bit more relevant for the audience?
00:35:40.559 --> 00:35:41.679
Thank you so much.
00:35:41.840 --> 00:35:50.079
And um so my you know, because I have so many more questions and I'd like to cover a lot of stuff here.
00:35:50.400 --> 00:35:57.440
Uh so um, you know, you mentioned Carl Freestone, of course, and and your approach connects to his work in many ways.
00:35:57.599 --> 00:36:15.679
Like, you know, and I don't know if we will have time to cover his mortal computation because he referred to that paper during his conversation with with you and his work with Overbiam, who was also both Carl Freestone and Alexander were keynote speakers at the OAGI 26 conference.
00:36:15.920 --> 00:36:24.079
So, you know, your approach connects, however, to Carl's mark of blankets as well, and to nested agency.
00:36:24.320 --> 00:36:41.679
And you know, this nested agency, the holonic paradigm uh of mitochondria nested within cells, which you already mentioned, cells within tissues, tissues within organism, and how they actually the mitochondria are the engine and the energy source of the whole organism in that way.
00:36:41.840 --> 00:36:47.360
Uh, turtles all the way up and down, as uh Ken Wilber describes it.
00:36:47.679 --> 00:36:57.920
So, you know, I use such holonic models in my work, applying them from global supply chains to designing more organic ways of governance for societal systems.
00:36:58.159 --> 00:37:11.840
My question to you uh related to this is could intelligence be an emergent negotiation among nested computational agents rather than something localizing a single privileged processor?
00:37:14.400 --> 00:37:22.159
Uh I I I think uh my own my own uh construction, my own model is that that's absolutely the case.
00:37:22.320 --> 00:37:32.639
Um we have multi-scale uh synchronization endeavors uh uh and circular causation, circular feedback loops across the scales.
00:37:32.960 --> 00:37:48.079
Uh, uh, with the system trying to minimize phase discrepancies across the scales uh and across the harmonics as the clocks change their periodicity uh through the through the structures and the cycles.
00:37:48.159 --> 00:37:57.920
And and and the uh the fundamental requirement for the system is state maintenance under conditions of uncertainty uh and perturbation.
00:37:58.079 --> 00:38:10.400
Uh and uh error signals uh in clock synchronization across the scales results in a sort of nested computational um error signal uh that the system seeks to minimize.
00:38:10.480 --> 00:38:14.800
That's the kind of essential um model that I'm trying to describe.
00:38:14.960 --> 00:38:25.280
Um but I think um if you don't mind, I just want to revert back uh because you mentioned Alexander Leichner of Google DeepMind and his um wonderful paper on the abstraction fallacy.
00:38:25.440 --> 00:38:47.840
Um, I I actually meant meant uh sorry, I I I'll get I'll get to that as well, but I mentioned Alex Ororbia who co-authored with Scuspensing about the the mortal computation, and of course if you want to link uh your work uh with that, how does it actually relate uh with the mortal computing what you do.
00:38:52.000 --> 00:39:06.639
Um uh that's a multidimensional answer, but um you know uh uh what I know of Alex's work, and Carl you know suggested that um you know he had written a wonderful paper in collaboration with Carl on mortal computation.
00:39:06.719 --> 00:39:08.400
What what Alex has done that?
00:39:08.800 --> 00:39:15.039
She showed that paper during the uh during your conversation, and I thought uh you kind of delved into it.
00:39:15.119 --> 00:39:16.320
And if you didn't, that's fine.
00:39:16.400 --> 00:39:17.519
We can move on to another question.
00:39:17.920 --> 00:39:19.039
No, but I I did read it.
00:39:19.119 --> 00:39:22.159
Um I read it um uh uh last year.
00:39:22.239 --> 00:39:34.079
Um so my my knowledge is a bit rusty, but he provides a wonderful exposition of all the approaches to describe, you know, uh the kind of um essential requirement for us to focus on mortal computation.
00:39:34.480 --> 00:39:41.280
Um what I'll say is that you know, we are born and we live and we die.
00:39:41.440 --> 00:39:47.440
Uh mitochondria uh come into existence and they have a lifespan of five to twenty days.
00:39:47.599 --> 00:39:57.280
Uh and cells come into existence and they have lifespans uh of um uh you know months or possibly years, but then they expire.
00:39:57.519 --> 00:40:15.360
And if we think about uh mortal computation, the system that you know if you can describe the human as a system, then there are life timescales, clock frequencies, and lifespans that are differing all the way down in this organism.
00:40:15.519 --> 00:40:19.920
So, in some ways, I think one has to think about that in terms of mortal computation.
00:40:20.239 --> 00:40:59.679
Uh, that the uh uh but then I think uh what's really important also is that one has to consider autopoesis, uh, which for you know those in your audience that maybe are not familiar with that term is a term of art that was uh uh and of science, of course, that was first um first introduced by Umberto Materano and uh Francisco Varela, cyberneticians um uh from the Chile school, uh, to describe a system that um instantiates or you know um builds its own components to instantiate processes that build its own components, that instantiate processes, etc., etc., etc.
00:41:00.159 --> 00:41:14.719
Um, and that's an important um uh kind of component of what is next, because I think uh you know self-organizing cognitive systems probably need to be self-instantiating.
00:41:14.800 --> 00:41:20.159
Uh and I say that because all of a sudden there's a big focus on recursive self-improvement.
00:41:20.239 --> 00:41:24.880
Um uh but what is what is the self and what is recursive and what is improving?
00:41:25.840 --> 00:41:39.280
My point would be in order for a system to be recursively self-improving, it has to first recursively self-instantiate, then it has to recursively self-maintain, and then it can self-improve.
00:41:39.599 --> 00:41:43.920
And so I think autopoesis is necessary for us to consider in that context.
00:41:44.159 --> 00:42:11.360
I think then that I have to refer to the um uh the work of uh Robert Rosen, mathematical biologist or biological mathematician, uh, with his metabolism repair framework, um, which was truly groundbreaking uh and quite you know separately in parallel, evolved uh along the uh trajectory to describe something similar, of course, to uh autopoesis.
00:42:11.440 --> 00:42:31.840
Um and um uh you know Robert Rosen also comes out of the cybernetic tradition, uh he was a student of Rashevsky, um, and um and so you know he had this uh feedback uh uh negative feedback uh sort of framework in mind, I think, as he described um his metabolism repair framework.
00:42:32.079 --> 00:42:46.000
But um if I may, I just before I seek to describe that framework, one has to go to uh what he utilizes in order for him to start to describe his work in the book Life Itself that he wrote.
00:42:46.159 --> 00:42:52.880
And he he references um uh Aristotle's uh causal framework.
00:42:53.039 --> 00:43:00.960
Um and and just in its most simple form, the causal framework uh is uh that you have a range of causes.
00:43:01.039 --> 00:43:06.480
You have a material cause, a formal cause, an efficient cause, and a final cause.
00:43:06.639 --> 00:43:12.400
And if we just take the example of the table, you and I are both sitting at tables having this conversation.
00:43:12.639 --> 00:43:14.079
My table is a wooden table.
00:43:14.159 --> 00:43:16.239
So the material cause of my table is wood.
00:43:16.400 --> 00:43:28.800
The formal cause of my table is the blueprint that described the components that had to be fabricated and how those components fabricated should be assembled to create the table.
00:43:29.039 --> 00:43:40.159
The efficient cause was the carpenter that read the blueprint that fashioned the and fabricated the components from the wood and then assembled them into the table.
00:43:40.400 --> 00:43:56.400
And then the final cause is that Mikhaila and Anthony have a conversation on mortal computation uh um uh across the Atlantic using um um uh communications paradigm that was never imagined possible by Aristotle.
00:43:56.639 --> 00:43:59.760
The most important point there is that there is no entailment.
00:44:00.239 --> 00:44:06.960
Of the final cause in the structure and the framework that Aristotle has described there.
00:44:07.119 --> 00:44:29.039
That's really important because it is effectively saying what Martin Heidegger said when he said we don't build models of the world in our heads, which is a kind of good old-fashioned AI proposition, using predicate calculus, structuring relations between objects in order to build a model of our world.
00:44:29.199 --> 00:44:33.519
Rather, we interact with the world, and the world, the best model of the world is the world itself.
00:44:33.760 --> 00:44:42.159
The point is we see affordances in the world and we operate to identify and then exploit those affordances.
00:44:42.559 --> 00:44:45.599
So we have to consider the world and the environment that we're in.
00:44:45.840 --> 00:44:59.519
Now that takes me to Rosen's proposition, which was that the efficient cause has to be entailed in the system, in order for the system to be a living system.
00:44:59.920 --> 00:45:03.519
And that living system will also be an anticipatory system.
00:45:04.559 --> 00:45:11.360
And of course, this was essentially what Alexander Lechner, whom we mentioned, I accidentally mentioned when I thought you were talking about him.
00:45:11.440 --> 00:45:12.559
You were talking about Alexander.
00:45:13.199 --> 00:45:18.400
No, no, no, but that's also what Aurorbia is doing with predictive coding, of course.
00:45:18.559 --> 00:45:21.760
Yes, so it is all about prediction and all this kind of stuff.
00:45:21.840 --> 00:45:23.519
And we will get to Alexander.
00:45:24.400 --> 00:45:27.760
Of course, also I have I have some questions about that.
00:45:27.840 --> 00:45:32.000
But this is indeed, I think this is the crux of the matter.
00:45:34.800 --> 00:45:39.039
I don't know if you read uh Frederico Fajin's uh irreducible.
00:45:39.599 --> 00:45:44.159
He is the inventor of the microprocessor, speaking of computing and all that.
00:45:44.400 --> 00:45:50.159
And he says exactly, I mean, the his whole point is we cannot control reality, we cannot control.
00:45:50.400 --> 00:45:52.960
I mean, exactly the point you make here.
00:45:53.039 --> 00:45:59.039
It's we can only participate with it one moment at a time, and and what will happen?
00:45:59.360 --> 00:46:02.079
Aristotle couldn't anticipate that.
00:46:02.480 --> 00:46:07.519
But okay, so you it's so interesting you bring up Frederico Fajin.
00:46:07.599 --> 00:46:15.920
I uh you know, I love listening to his lectures, and he he did, of course, incredible work um in the sort of foundational stage in the semiconductor industry.
00:46:16.000 --> 00:46:19.039
Um, but I don't think I share his idealist perspectives.
00:46:19.199 --> 00:46:23.840
I'm actually a devotee of his uh co-collaborator Calver Mead.
00:46:24.239 --> 00:46:32.079
And I know that, and I'm gonna I'm gonna touch on that later on when we discuss about Beck Gut versus Alexander Lesnar.
00:46:32.400 --> 00:46:39.440
But uh but let me I don't know if you wanted to close the loop here because I have more questions, or can I move on?
00:46:39.920 --> 00:46:51.199
I just wanted to say that um in order for a system to be a living system, just to re-emphasize, according to Rosen, it has to be uh closed to efficient causation.
00:46:51.679 --> 00:47:07.360
So as as I've described um using my Aristotelian uh table uh example, the carpenter has to be entailed in the system in order for the system to be characterized as a living system.
00:47:07.760 --> 00:47:12.079
So if the carpenter is not part of the system, the system itself can't be living.
00:47:12.239 --> 00:47:18.159
You have to consider the uh carpenter and the table that's been fabricated as the system.
00:47:18.480 --> 00:47:31.360
And that, of course, is effectively also what the second order cyberneticians were uh thinking about as they described second-order cybernetics as the science of observing systems.
00:47:33.440 --> 00:47:50.159
Yes, and I actually was uh uh editor of the first journal on cybernetics and human knowing, uh journal of second order cybernetics, and of course Maturana Varela, they are the pioneers, uh, with Professor Suren Breer from Denmark.
00:47:50.320 --> 00:47:53.039
He is actually he was uh founder of this journal.
00:47:53.119 --> 00:47:56.719
That was when I was doing my PhD several decades ago.
00:47:56.880 --> 00:48:00.639
But yes, I I can relate to that as well, and that's a very important point.
00:48:00.800 --> 00:48:05.920
But if you don't mind, I'd like to return, unless you want to make more more uh details here.
00:48:06.079 --> 00:48:09.119
I think it's a few important okay.
00:48:10.000 --> 00:48:16.400
I I I want to make perhaps one final point, which is um gold-directedness and teleology.
00:48:16.559 --> 00:48:23.360
Uh so what's interesting is um, as you know, I'm a great uh devotee of Mike Levin's work as well.
00:48:23.440 --> 00:48:38.960
Um Mike um talks about agency, and by the way, I'm also now uh I I want I'm very happy to say I've been invited by the Agential Biology Institute, uh uh, which is looking at agency in biological systems to join as uh entrepreneur and engineer in residence.
00:48:39.039 --> 00:48:56.480
So um that's hot off the press, and um and uh uh and and of course uh that's very proximate because uh Marco and uh Tomaso are also members of the um of the Agential Biology Institute, and and and Marco indeed is uh on the advisory board.
00:48:56.639 --> 00:49:30.559
Um but I want to talk about this uh concept of gold-directedness uh and teleology um because having described that framework, um the the point is in order to maintain state, the system has to cohere with its environment, and that means it has to have a repertoire or a variety of directions that it can take in order to maintain state.
00:49:30.880 --> 00:49:41.199
And according to you know the traditional cybernetic principles, the system that is modelling another system has to be imbued with the requisite variety.
00:49:41.280 --> 00:49:51.360
That means the complexity of the system here that's modelling the environment there has to have the same degrees of freedom or variety uh and complexity.
00:49:51.760 --> 00:50:10.320
The point being that in order to maintain state, there'd be m there may be many actions the system can take or must take in order to minimize that very uh that um uh uh relative entropy so that it maintains state uh in relation to its environment.
00:50:10.559 --> 00:50:21.199
So goal-directedness is um, you know, I think the industry generally has to think about what teleology goal-directedness and agency means.
00:50:21.360 --> 00:50:32.559
Mike, uh, I mean, you know, he he he talks about William James' definition of goal-directedness as um, or rather intelligence, as um being able to fulfill any goal by many means.
00:50:32.719 --> 00:50:48.079
Uh but um uh uh what I'm trying to describe is a variety of goals to maintain state as another way to characterize maintenance of state in the presence of uncertainty in the environment.
00:50:48.159 --> 00:50:49.599
Does that make sense?
00:50:50.880 --> 00:50:54.320
Especially to a control engineer, of course, it makes sense.
00:50:54.559 --> 00:50:56.800
I mean, this is systems science, right?
00:50:57.039 --> 00:50:58.320
Of course, this is our part of the code.
00:50:58.480 --> 00:50:59.519
I wish I'd met you 20 years later.
00:51:04.159 --> 00:51:06.719
You're talking, you're preaching to the choir here.
00:51:07.039 --> 00:51:16.960
So, you know, I'd like to continue the conversation um on the controversial matter which you are undertaking in this new research program.
00:51:17.360 --> 00:51:20.159
Uh, does substrate truly matter?
00:51:20.480 --> 00:51:31.840
And if intelligence is substrate dependent or independent, and this you know brings us directly to one of the deepest disputes in AGI.
00:51:32.159 --> 00:51:40.960
Uh and you you you mentioned uh uh Alexander Lechner, who I invited actually to give the keynote at the conference.
00:51:41.280 --> 00:51:51.519
And um I do not know if you uh watched uh uh our online uh conference, but there was a big debate between Dr.
00:51:51.679 --> 00:51:54.000
Ben Gertzel and Alexander.
00:51:54.239 --> 00:52:18.960
And Alexander is actually uh from Google Deep Mind, and he's an amazing researcher, and uh his abstraction fallacy is actually, you know, uh kind of uh uh very clearly uh combating, if I may say that, the idea that abstract causal topology alone uh could capture the relevant reality of intelligence.
00:52:19.280 --> 00:52:23.280
And your work seems on his side.
00:52:23.760 --> 00:52:29.599
Uh so if you you seem to, you know, maybe I don't ask anything, please elaborate.
00:52:29.760 --> 00:52:31.360
You seem to you want to jump in.
00:52:31.519 --> 00:52:33.360
Let me know what you think of that.
00:52:33.679 --> 00:52:43.280
Um yes, um you know we are um possibly communic communities separated by a common language.
00:52:43.440 --> 00:52:52.960
Um I I I think uh it was Churchill that said that you know Great Britain and America were um two peoples separated by a common language.
00:52:53.199 --> 00:53:11.039
Um what I find interesting is that um the substrate-dependent computationalists uh uh are just trying to describe the physics to describe the physics, and and I think you know my effort has been to try and do that to describe the physics that drive computation on the mitochondrial inner membrane.
00:53:11.280 --> 00:53:20.480
And um and um the uh uh substrate independent community are saying, well, you know, it's software that can run on any substrate.
00:53:20.559 --> 00:53:27.599
But you know, when we look at the OSI stack, you know, coming from the compute communications industry, you know, there's a stack of software.
00:53:28.159 --> 00:53:29.679
Sorry, may I correct something?
00:53:29.920 --> 00:53:54.079
They do not say uh no substrate, they say it can be any, you know, and and I think they also say like it's substrate independent, but I mean, you know, if you look um and you say you are familiar with Federico Fajin, and you also say you don't agree with everything, if you can answer the question from this perspective like Federico, Alexander, yes.
00:53:55.440 --> 00:54:00.159
Yes, um uh so I mean from a philosophical perspective, I'm not an idealist.
00:54:00.239 --> 00:54:11.039
Uh, you know, uh uh I am uh a um uh a dual aspect monist, uh, and what because we're seeing the physics encoding the information on membranes.
00:54:11.199 --> 00:54:16.639
Uh uh, or as uh Carl and uh Thomas Parr would say, uh a Markovian monist.
00:54:16.800 --> 00:54:21.679
Um and um so that's what I don't agree with uh uh Federico on.
00:54:21.840 --> 00:54:25.280
Uh uh, you know, I've I've had my dallions try and understand idealism.
00:54:25.440 --> 00:54:28.000
I studied philosophy also earlier in my career.
00:54:28.159 --> 00:54:35.280
Um uh I I actually detoured and studied law and philosophy along the path, so you know that was kind of several years wasted.
00:54:35.440 --> 00:54:40.559
But um, sorry, if I said uh no substrate, I meant to say on any substrate.
00:54:40.639 --> 00:54:44.400
I'm trying to characterize the uh substrate independent community.
00:54:44.559 --> 00:54:58.079
But if I look at um, you know, we now have, of course, the uh Institute for Machine Consciousness in the Silicon Valley, and Josh Bach is, you know, talks uh emphatically about uh software running on any substrate.
00:54:58.239 --> 00:55:07.199
Well, if I listen to him, uh incredibly capable uh uh and and you know uh interesting researcher, the self-organization.
00:55:07.679 --> 00:55:10.719
Articulate, I would say, very articulate, right?
00:55:10.800 --> 00:55:14.480
So in a second language, uh I'm in awe, yes, uh I have to say.
00:55:14.639 --> 00:55:30.960
Um but um uh the the approaches he's using to describe substrate independent computation reach into self-organization, and as I say, uh they will be focused on recursive self-improvement.
00:55:31.119 --> 00:55:36.320
So the question is what are the software programs they're gonna run to enable recursive self-improvement?
00:55:36.400 --> 00:55:44.400
They have to be those, as I've said earlier, that enable self-instantiation and self-maintenance first before they get to recursive self-improvement.
00:55:44.559 --> 00:56:00.000
So I think he's reaching, and that endeavor is reaching into the substrate-dependent world, and the substrate-dependent physics of computation world has to recognize that um there are programs inverted commas that can run on the biological substrate.
00:56:00.079 --> 00:56:06.559
Now, Gordon Pask, one of my heroes' second-order cybernetician, came up with a path to describe how we solve for this.
00:56:06.639 --> 00:56:12.559
He he he he uses the concept of an M individual, a mechanical individual, and a P individual.
00:56:12.639 --> 00:56:18.400
And he says, an M individual, which is a substrate, can run any number of P individuals on it.
00:56:18.639 --> 00:56:20.400
And indeed we do as individuals.
00:56:20.559 --> 00:56:33.280
We run, you know, we have we have uh cognitive processes that describe many personas that we could occupy in exploring a landscape as cognizing beings.
00:56:33.360 --> 00:56:48.800
Um, so my point is there is a bridge to be built and a requirement for an ontology, a common ontology, to enable the conversation between those that focus on the substrate and those that focus on substrate independence.
00:56:49.039 --> 00:56:58.480
Now, at the moment there's a category error, in my opinion, in the space, which is that people talk about carbon versus silicon-based cognition and computation.
00:56:58.639 --> 00:57:08.400
Well, it just so happens that the best example of a substrate-dependent cognizing system that we have access to is a biological system, which is carbon-based.
00:57:08.639 --> 00:57:15.920
But there's a requirement for an ontology for substrate dependency or substrate independency.
00:57:16.159 --> 00:57:20.960
And we're wasting time if we don't bridge the communities to solve for that.
00:57:21.199 --> 00:57:22.880
This happened in the last century.
00:57:22.960 --> 00:57:39.840
There was the Macy Conference, which gave birth to cybernetics, which was an interdisciplinary uh endeavor in the in the in, I think it was 56, uh, you know, the um the uh uh breakaway by the uh AI community, Minsky uh et al.
00:57:40.400 --> 00:57:48.719
Uh, primarily because they didn't really like uh Michellech and uh Norbert Wiener uh because of their personalities, more than anything.
00:57:49.119 --> 00:58:09.119
But we we are now at a point in time where the McCullough Pitts uh Neural Net has given us extraordinary advances, and so cybernetics is kind of in lots of ways uh uh at the heart of the most recent you know flourish in the AI revolution.
00:58:09.360 --> 00:58:22.800
But we now have communities that are diverging again, uh uh and I think there's an opportunity to bring communities together, to ground the conversation by agreeing a common ontology so that the conversations can be held.
00:58:22.960 --> 00:58:30.639
Instead of people saying, oh, they don't believe in substrate independency or the philosophers thinking about the physics of computation.
00:58:30.800 --> 00:58:44.639
I think it's an unnecessary uh argument that is a polarizing effect where we we could actually make really fast and significant advances by bringing the communities together to develop an ontology, a common ontology.
00:58:46.320 --> 00:58:47.440
Amazing, yes.
00:58:47.599 --> 00:58:51.920
And but before we get into that, because I have many questions about that as well.
00:58:52.480 --> 00:59:06.159
Um I don't know if you can throw a punchline like what exactly does substrate the substrate contribute that uh cannot be captured by the abstract computational description.
00:59:08.559 --> 00:59:10.719
Well, I think um two things.
00:59:10.800 --> 00:59:16.400
So if one reads Alexander Lechner's uh paper, you know, he talks about the map and the map maker.
00:59:16.480 --> 00:59:21.199
So he's using the same kind of uh metaphor as I was using with my table metaphor.
00:59:21.360 --> 00:59:31.440
The carpenter has to be entailed, or the map maker has to be entailed in the system for us to just describe the system as a cognizing system.
00:59:32.400 --> 00:59:36.320
That's Roslin's work and that's uh Alexander Lechner's work.
00:59:36.480 --> 00:59:43.280
The point is the substrate independence at the moment at least uh is a mirror.
00:59:43.519 --> 00:59:58.079
The cognizing agent that, as Lechner said, alphabetizes the code that is then running in the software, uh has been built by the cognizing agent, which is the external human in the loop.
00:59:58.239 --> 01:00:12.239
Um and uh and until the human and the loop is entailed, or some equivalent of the human and the loop is entailed in the system such that it is close to efficient causation, the systems won't be cognizing systems.
01:00:12.400 --> 01:00:27.039
And that's my point about mitochondria and metabolism's inference, because effectively the uh it's not quite the case, but effectively the mitochondrion is an autonomous agent operating in the cell, interacting with the cell.
01:00:27.119 --> 01:00:31.679
So one could argue that the system, the cell, is close to efficient causation.
01:00:31.760 --> 01:00:50.400
So my point is if you can figure out how to build a system based on the model that I'm describing that enables you to evolve from where we are with substrate independency towards recursive self-improvement, and as I say, that requires first instantiation and then maintenance.
01:00:50.559 --> 01:00:58.239
Then I think we can move beyond the argument about whether or not substrate independency is the pathway to go.
01:00:58.320 --> 01:01:09.679
And I think uh Blaise Aguera Iarcus, um, also Google has done wonderful work here as well in describing endosymbiosis as a computational paradigm.
01:01:09.920 --> 01:01:20.000
Um, and and so I think what's interesting is the pathway is emerging for us to move towards a new computational paradigm based on a number of these ideas.
01:01:20.559 --> 01:01:29.679
If we can bring the communities together, then I think we can really make fast progress to describe the landscape that enables us to have a common conversation.
01:01:31.280 --> 01:01:38.800
Yes, and and I really salute you for this uh initiative because you are actually, I don't know if you want to talk about this.
01:01:38.880 --> 01:01:56.800
You kind of are alluding here that um you actually are setting up or in the process of setting up, um, let's say, I would call it a research program that ultimately would point to you know, the truth about that and and and develop new ways of computation.
01:01:56.960 --> 01:02:04.639
And my question, my initial question, you know, I will come back to it somehow, but different, phrase it different now, you know.
01:02:05.039 --> 01:02:15.519
I mean, you have sat on both sides of the table because I started asking you what's the point of all your this mitochondrial computing and so on.
01:02:15.599 --> 01:02:16.800
Why are you doing all this?
01:02:16.960 --> 01:02:22.239
So, you know, because you and you very nicely uh included this in your answers.
01:02:22.320 --> 01:02:27.599
I mean, you've sat at both sides of the table as inventor and investor.
01:02:28.000 --> 01:02:46.000
So I'm asking now the investor, what result would make you wearing your investor hat uh rather than the inventor hat, say now I would invest serious capital in this, in a program like this, and in this research.
01:02:46.239 --> 01:02:57.039
Because you know, if you and and as an inventor, if you had enough funding tomorrow, what are the two or three experiments you would fund that could most decisively falsify?
01:02:57.119 --> 01:03:03.599
So you know, you can you can switch heads here and and uh let us know your thoughts.
01:03:06.000 --> 01:03:17.519
Well, I think um uh what really attracted me and drew me into the space was Karl Frison Thomas Parr with uh active inference.
01:03:17.840 --> 01:03:31.039
Because it just looked like a more elegant way to describe uh how systems uh generally speaking might uh uh uh make sense of their world, and therefore it looked like an appropriate, interesting path.
01:03:31.199 --> 01:03:33.679
So that's one of the vectors I'm interested in pursuing.
01:03:33.840 --> 01:03:37.119
The second, of course, is Mike, Mike Levin's bioelectricity.
01:03:37.280 --> 01:03:57.679
You know, that has been uh you know groundbreaking stuff, and uh and and you know if we start to look at electrodynamics, hydrodynamics beyond electrostatics in biological systems, uh we we move into much more compelling, exciting territory and a research agenda for the 21st century.
01:03:57.920 --> 01:04:09.760
Also, along the way, I've been uh engaged deeply with um the Guy Foundation for Quantum Biology, and they are looking at quantum effects in biological systems, uh particularly mitochondria.
01:04:10.079 --> 01:04:19.119
They're looking orders of scale down from where I am building the system model, but I I've been tremendously excited by the work that they've done.
01:04:19.360 --> 01:04:48.320
Jeffrey Guy put the Guy Foundation together, and he's backing um he's backed research uh uh programs by Mike and by Nick Lane, who really you know, with his work, most incredibly titled work on the mitochondrion, which I read at the beginning of my effort, um Power, Sex and Suicide, um, really sort of described uh in the most compelling way the system that I've spent the last two and a half years from a computational perspective trying to characterize.
01:04:48.480 --> 01:05:11.519
So I think all of those things, together with the uh requirement for us to move to a different computational paradigm that doesn't ignore energy and the it doesn't violate the constraint, which is provide as much energy as is necessary in order to compute, but rather utilizes compute resource availability as part of the computational program.
01:05:12.000 --> 01:05:25.519
That's I think uh charts a pathway beyond the current uh scale is all you need, LLM uh um local minimum is how I describe it, uh, uh, towards a more compelling future.
01:05:25.679 --> 01:05:35.920
So if I am absolutely wearing my investor hat in trying to explore the domain so that I can see what I think is the more exciting, compelling future beyond where we are.
01:05:36.079 --> 01:05:48.800
So if I had uh money, I'd be investing uh at the intersection of all of those domains active inference, bioelectricity, and um um um um uh energy based uh compute models.
01:05:49.119 --> 01:05:59.840
And I'd be trying to do that with one foot in biology and one foot in AI because it it it will transform our understanding of biological systems if we get it right on the compute side.
01:06:00.480 --> 01:06:12.400
And so that's that's one of the other reasons I'm interested because I see this huge focus of the superscaling AI companies trying to put their LLMs to work to solve biology.
01:06:12.800 --> 01:06:17.920
But uh the LLMs are fantastic um induction machines.
01:06:18.079 --> 01:06:34.400
Uh and traditional good old-fashioned AI is fantastic deductive machinery, but um they are not anticipatory systems and they can't project outcomes into the future and then direct themselves towards achieving those outcomes.
01:06:34.559 --> 01:06:39.039
The biological systems that I'm describing are anticipatory systems and they can do that.
01:06:39.199 --> 01:06:48.400
And so I think we're seeing a potential, a potentially much more compelling pathway towards uh a more exciting AGI beyond where we currently are.
01:06:48.559 --> 01:06:51.119
And of course, that that's why I'm excited by your work.
01:06:52.960 --> 01:06:53.920
Fascinating.
01:06:54.079 --> 01:07:13.360
And you know, um computing history contains moments when an initially obscure phenomenon, in your case, now it's uh like uh let's say how the mitochondria works, suddenly becomes an engineering uh primitive, like the transistor, you know, is kind of the example.
01:07:13.519 --> 01:07:40.239
So the question would be what would be the equivalent transistor moment for your research, for your program, for your approach, however you want to call it well uh so I talked about you know first a voltage controlled oscillator, and then a cascading uh operational amplifier structure and then a phase lock loop.
01:07:40.480 --> 01:07:57.599
The more I look at the mirrored uh architecture and the machinery, uh horizontally opposed machinery on the mitochondrial crystal, the more it looks like a complementary transistor amplifier, amplifier architecture to me.
01:07:57.760 --> 01:08:07.440
And as I mentioned, each of those complexes that's um pumping protons or allowing electron flux through the membrane has a metal center.
01:08:07.760 --> 01:08:12.000
And so I think that is the transistor moment.
01:08:12.320 --> 01:08:31.119
We are looking at an architecture in a biological system that is equivalent to, although two, three billion years uh prior to, the complementary transistor amplifier circuit, which was the foundation for the modern computing industry.
01:08:31.439 --> 01:08:38.319
Um that's a claim, that's a bold claim, that's an irrational claim, that's a um foolhardy claim, possibly.
01:08:38.399 --> 01:08:54.880
Uh, but I've I spend my time you know trying to push the boundaries here, really, so that we can, you know, a challenge, attack, but first I think, uh, and this is what I love about Mike Levin's work, first, don't attack the straw man, build the steel man, and then attack the steel man.
01:08:54.960 --> 01:09:02.960
I think that's what is appropriate, necessary, and I think um will yield incredible uh new insights.
01:09:03.119 --> 01:09:23.520
But I think um this is the transistor moment because the system can be characterized, even if it's an abstract, highly stylized model I'm building, it will enable uh uh experimental scrutiny, but also it will describe uh ways for us to explain biological phenomena, I hope, in in a fundamentally different way.
01:09:26.159 --> 01:09:33.680
And uh if your computing paradigm or you know this approach would succeed, uh what are the implications?
01:09:33.920 --> 01:09:49.199
But you know, um I also I think time, because we mentioned it, time plays a role here, because you know, yes, uh if it will succeed before the singularity or after, I think it's of the essence as well.
01:09:49.279 --> 01:09:52.960
Or maybe you think we cannot achieve the singularity without your approach.
01:09:53.199 --> 01:10:04.960
I mean, tell me what are your thoughts there with relative to what's happening today with uh AGI and all these exponential developments relative to your paradigm?
01:10:06.239 --> 01:10:19.039
Well, I would say um, and um and by the way, you know, this is your area of expertise, Mikhaila, um, not really mine, uh, but uh, you know, umbodied con embodied cognition, you know, soft robotics.
01:10:19.199 --> 01:10:28.800
Um uh I don't think we're gonna achieve the singularity without uh embodiment, uh, if indeed it's ever possible, and I'm skeptical.
01:10:29.039 --> 01:10:37.119
Uh uh, but actually, when we think about embodiment, embodiment is not the current approach, which is you know the pre-training moment in robotic systems.
01:10:37.279 --> 01:10:46.159
We're not downloading LLM propositions onto uh onto uh robotic structures and thinking of them as embodied.
01:10:46.319 --> 01:11:13.039
Embodiment means autopoesis, self-instantiation, self-maintenance, self-improvement over time scales that might not be perceptible uh to us as humans, considering about evolution, considering evolution, and this is you know really just briefly referencing the amazing work of the Agential Biology Institute, trying to understand evolutionary paradigms beyond the neo-Darwinist uh genome first, genome is everything proposition.
01:11:13.279 --> 01:12:07.760
Um, but I think that uh if you consider every single cell of an embodied agent as its own computational uh uh standalone agential computational architecture, and then you think about the uh individual cell as having ten, a hundred, a thousand, tens of thousands of mitochondria in it, and then you think about the activation function that is performed by an ecology of mitochondria within a cell, the complexity, and then you consider that you know we have hundreds of trillions of cells in our body and hundreds of trillions of mitochondria in our body, then we're looking at orders of scale very different from the orders of scale we're looking at, and we're looking at each individual uh embodied component of the system being itself a cognizing agent, interacting with its other agents.
01:12:07.920 --> 01:12:18.000
Then all of a sudden I think we start to see a different pathway towards embodied cognition, uh, and and who knows, over time perhaps towards the singularity.
01:12:18.079 --> 01:12:22.560
But the models we're talking about are not embodied in any sense in the way I've described.
01:12:22.960 --> 01:12:23.600
Absolutely.
01:12:23.760 --> 01:12:27.520
And you know, now he you answered my initial question.
01:12:27.600 --> 01:12:30.960
Yes, I mean, actually, what would be the application?
01:12:31.279 --> 01:12:36.560
What is kind of this uh breakthrough useful for?
01:12:36.800 --> 01:12:48.960
And I agree because you know, now everybody's talking about physical intelligence and just putting LLMs inside the robots' brains and so on, brains, I mean, whatever, chips.
01:12:49.279 --> 01:12:54.800
So which which is definitely has nothing to do, and even designing humanoid robots.
01:12:54.960 --> 01:13:10.880
And um, I've been at the Davos Robot Summit in June, and uh they came, of course, with a lot of uh humanoid robots, even looking very nice and sexy and all that, but definitely there was nothing about embodied intelligence, they looked really creepy.
01:13:11.600 --> 01:13:12.800
And I agree.
01:13:13.520 --> 01:13:21.279
Because you know, if we really want a breakthrough in robotics, now I can see the the advantage of your paradigm.
01:13:21.439 --> 01:13:31.199
And actually, that may be it may be true, you know, that uh we cannot achieve that kind of embodiment without um uh your paradigm.
01:13:31.439 --> 01:13:42.399
Uh I just wonder if you, you know, because we reached kind of uh the end of our time here, but I you know I wonder if you have any final statements or any things which I didn't touch on.
01:13:42.640 --> 01:13:51.279
I know there's a lot that you are doing, but um a call to join your program or whatever you want to tell us that we didn't cover, please.
01:13:52.479 --> 01:14:05.680
Yeah, so I mean uh I'm I'm working with the agencial biology institute uh amongst amazing people, Charlie Munford and uh Richard Watson, uh the president and the chief science officer, and uh and an incredible group of people.
01:14:05.760 --> 01:14:19.760
Uh on the advisory board, we have people like Mike Levin, Josh Bongard, Stuart Kaufman, Dennis Noble, just incredible uh scientists uh, you know, who've been uh uh instrumental in shaping their respective fields.
01:14:19.920 --> 01:14:38.000
You know, uh I hope to work with the Agential Biology Institute to raise the funds, to help them raise the funds to explore my conjectures, but more important, their existing list of conjectures towards trying to explain agency and an agency-first evolutionary path.
01:14:38.159 --> 01:14:47.279
Um, and I think that will also uh bring incredible uh future value uh for the AI industry, but for biology more generally.
01:14:47.520 --> 01:15:38.399
Um, I'm also interested in working with um funders, uh venture funders, about the ideas I've described here towards um uh not if not necessarily neuromorphic uh hardware architectures, uh software emulation of hardware architectures uh in the first instance, um, uh, as a pathway to describing how we might achieve uh more compelling embodied soft robotics uh technologies, but also uh cognitive technologies uh beyond um beyond the induction and deduction mechanisms available uh with the existing AI paradigms towards an abductive uh you know uh type of system that can uh project and question future counterfactual pathways in a way that the LLMs can't currently.
01:15:42.239 --> 01:15:44.000
Thank you so much, Anthony.
01:15:44.159 --> 01:16:04.720
I mean, the engineer who once thought about how an aircraft determines its orientation in space is now asking how a living cell determines its own state and how that in turn might illuminate the past towards entirely new kinds of intelligent machines and the future of robotics.
01:16:04.960 --> 01:16:06.319
This is more than fascinating.
01:16:06.479 --> 01:16:12.239
I really, your past is so unique, and I so much appreciate uh you coming on the podcast.
01:16:12.479 --> 01:16:32.720
And also, of course, I look forward for us to forge a very powerful collaboration with our group, and uh I'd like to be involved, of course, in your in your program um that you are now uh crafting, and thank you for sharing that uh with me and and with the audience as well.
01:16:33.119 --> 01:16:35.920
Because um, if I may, I just want to say thank you.
01:16:36.159 --> 01:16:37.600
Thank you for noticing my work.
01:16:37.840 --> 01:16:43.279
And um, you know, I wish I'd met you 20 years ago because you know, we we share so much common heritage.
01:16:43.359 --> 01:16:45.119
Um it could have been an easier path.
01:16:45.840 --> 01:16:47.600
I know, it's like, where have you been?
01:16:47.680 --> 01:16:54.239
That's you know, sometimes you have this feeling, and and I had that uh as soon as I found you.
01:16:54.399 --> 01:17:04.720
I think I think I found you or you found me, and I think I think it was in the context of Alexander Lechner uh when he launched uh this uh abstraction fallacy paper.
01:17:04.800 --> 01:17:08.720
I don't actually remember how we met, but my feeling was exactly like that.
01:17:08.960 --> 01:17:10.239
Where have you been?
01:17:12.079 --> 01:17:12.560
Thank you.
01:17:12.640 --> 01:17:18.239
I and uh it it's it's it it's touched me, and it it's enchanting that you uh are enthusiastic about my work.
01:17:18.319 --> 01:17:19.359
Thank you so much.