← AI Pod by Wes Roth and Dylan Curious | Artificial Intelligence News and Interviews With Experts
AI Pod by Wes Roth and Dylan Curious | Artificial Intelligence News and Interviews With Experts
Lee Cronin "Sam Altman Is Delusional, Hinton Needs Therapy, P(Doom) Is Nonsense"
01:29:56
Loading summary
Transcript82 lines
- [00:00]
A
The problems I see with AI, fake people, fake data, political manipulation, and so what all the doomers are doing, they're literally making stuff up. Because there is no mechanism by which an AGI can become super intelligent, get control of the systems, have agency and do anything. It is nonsense. We think that AIs have anywhere near the creativity or potentiality of, of living systems. They are nothing. They are nothing compared to living systems. Complex multicellular creatures have agency and AI has none. Hi there, I'm Lee Cronin, I'm the Regis professor of Chemistry at the University of Glasgow and also the CEO of a company called Chemify. And I want to find out how life started.
- [00:50]
B
All right, let's just start with one of the questions that I feel like everybody thinks about when they're younger, which is why? Why something, not nothing?
- [01:00]
A
I would ask the other way around. Why nothing and not something? And so I think the, the problem is that the fact that. Well that's, it's actually the simplest question to answer, I think, is existence is the only thing that happens when causation is possible. And so the, the, the ridiculous thing is the fact that matter exists in the way it does means that it's been able to fight against randomness in the background. And that's kind of obvious, but actually deeply profound. The mechanism by which matter of any structure exists in the universe is because it's had to survive against degradation. And so why something happenstance? Selection, causation and evolution, or it maybe should be causation via selection that gives you evolution. Because actually the universe at the base I just realized today is random and basically featureless. And so that's kind of, that blew my mind, My mind is still blown from that. And this last about 5 hours ago,
- [02:03]
B
I realized that like pure entropy, we
- [02:06]
C
have to talk about this. What happened?
- [02:08]
A
Not necessarily pure entropy, but pure, just, just this stuff, it's just bubbling stuff. But there is no. You just get the basic laws of physics, which are kind of pretty boring, right? But awe inspiring to us that go, oh look, there's stars and galaxies and stuff. But the laws of physics just give you fusion and elements and then chemistry has to kick in. And chemistry is perhaps the most profound thing that happened to physics. I would say that as a chemist, but finally I found my position in the universe. So the reason why there's something and not nothing is because of chemistry.
- [02:46]
C
Absolutely. And so chemistry being kind of almost a computational force. And that's why we see certain systems emerge, certain complexity emerge. So I'm still very Curious. You said something happened five hours ago that changed your view. I'm. Can we just. I can't get my mind off of that.
- [03:05]
A
So I was. So it wasn't, it wasn't just my change my view, but solidified. I had to explain to a couple of physicists I'm working with, just starting a new project with that basically is going to explain how a new in. Basically we have this problem in physics. And I'm not a physicist, right. I'm a chemist. So like anyone listening to this can just discount anything I say from that. But I'm right. And what I mean by I'm right is I, I'm able to use this idea to come up with a new prediction, like a test that I would not be able to produce using any other theory. But basically five hours ago I was, I was talking with two physicist colleagues, a phys. A. The. So one of them is more a philosopher and one of them is more a theoretical physicist. But they kind of interchange. And I was trying to explain how causation manifests. And we realize that in physics you have Newtonian physics and you have quantum mechanics, they suffer from the same problem. The quantum mechanics is not mysterious in the way that Feynman. Everyone said it's just because they couldn't get their heads around causation. And I feel somewhat, you know, kind of like unqualified to say that because Feynman was one of the greatest geniuses that ever lived. But he was wrong because he just accepted what he was taught about causation, the lack of causation. And so when you put causation back at the beginning, you realize that the reason why the universe exists is something has to manifest from nothing via causation. And the way we interpret that right now in physics is that we have to put causation into physics. We can't put. We accept, we think that physics has no causation. And the way we jimmy rig it is literally by putting in quantum mechanics and the macro scale, Newtonian mechanics. Um, it might sound quite bizarre, but it's really obvious when you dig down because there is no quantum woo, there is just causation.
- [04:50]
C
Interesting. Okay, let's just break it down so everybody kind of get everybody caught up and we're. I think we're gonna come back to this just. So one of the things you said, so causation to selection to evolution. So evolution I think most people kind of understand is growing complexity, improvement. Selection is kind of like, let's say a B testing, right? For a lot of the people listening, right. You'd split two pages. This page does better. So kind of survives and goes on. So there's a selection pressure. Can we define causation in the most basic terms so that most people can kind of grasp what we mean? Exactly the kind of what you're talking about?
- [05:25]
A
Yeah, yeah, yeah. So I mean, I would say actually a B testing doesn't is high level selection. Selection is something that emerges. But let's go back to causation because causation gives rise to that. So if I take two rock, so if I take a small rock and a big rock and I put this and I hit the small rock at the big rock, the big rock doesn't move. Right. So there'd be some causation. The rock just moves, nothing happens. If I take the big rock and hit the small rock, the small rock will move and it could then hit another rock and hit another rock. That's causation. So when A basically precedes B, but by looking at the kind of phase space, you can see that actually the energy from A goes into B, which then causes B to move and do something to C. That's what I mean. At the bottom is causation. And a very famous philosopher called Bertrand Russell said, you know, causation is like the monarchy, it's just symbolic, doesn't do anything. And of course, Newton thought this, and Einstein thought this and everyone thought so all our physical theories have ignored the fact that causation is real. So A implies B. Now, in physics, there's no causation. The causation is immersion. But when you look at the emergence of biology and I suppose consciousness, imagination, free will and intelligence, you need causation to make that happen. And so five hours ago I was talking to a couple of colleagues, we're thinking about how we frame this precisely in physics, because it's not every day you just tear up the textbooks, all of them, and say, well, maybe it's all wrong and, and not wrong in a. Oh, we've been using the wrong paradigms because the mathematics of quantum mechanics works, the mathematics of Newtonian mechanics works to some degree. But what it doesn't do is explain intelligence and biology. That's a very critical problem. So, yeah, so it's kind of fresh. So forgive me for not being terribly eloquent, but basically causation is the mechanism by which selection occurs. Selection then provides, allows matter, some matter to exist over other. The existence of that matter over time in the end produces biology and intelligence, which is kind of mind Blowing, right?
- [07:45]
C
Absolutely, yeah.
- [07:47]
B
And so just to make sure I get this right, so causation isn't on the most bottom level. That's something that does emerge. Or is that the change?
- [07:56]
A
No. So what? No, no, causation is on the bottom level. But what happens is you've got this. Imagine you've got this foam almost. Let's imagine you've got bubbles that are bursting. Right. Like, almost. What do you want to call it? Quantum foam? Or just like. I don't like that. It's just like there's just random stuff happening so into existence. If you like. Like imagine sand grains. Some sand grains could aggregate and make a shape, and that shape will be fleeting and then we'll decompose back in the sand grains. So that's called. So when the sand grains aggregate, that's not causation, that's just a random process. However, if some sand grains bind together and they are then able to then affect other sand grains to bind together, and you almost remember a snowball effect, that's kind of the emergence of causation from randomness. But I use emergence in a. In the. It's like. It's not emergence. It's something quantitatively different. It is just the existence. The fact that something is able to exist in time beyond its natural half life means that there is a chance memory that's been trapped in the universe that can reproduce that structure. And it's happening all the time. The universe is making, let's call them memories, causal chains that basically can affect, move sand grains around, but then they're just lost because it's just random. But if those grains get together and they can help other grains exist, suddenly you have something like cooperation or templating, and you get this scaffolding. And really the unit, the reason why we have something rather than nothing, is matter that has persisted against the attrition of the environment. And then we can think about that long enough. It's obvious that life is the only thing that physics can produce to teach physics, that physics is not what we think it is. And so it's kind of inverted in that we think our universe is law, like, but it's not really. It's random with some kind of probabilistic laws. The only reason the universe is law, like, is that biology produces error correction. And so there's a lot I've just said in the last two minutes. I'm sorry.
- [10:01]
B
No, that's good. Well, let's stay in the very bottom level. So this, I guess, quantum foam or whatever, this Sort of randomness is. Is there no memory on that level? Like, is there anything like pilot wave theory or something that's kind of aware that one bubble burst over here after another bubble over there? Or that's just not even part of
- [10:18]
A
the way to think about it? I mean, the ontological interpretations of quantum mechanics are very complicated in that or no, very diverse. So this way I would be very parsimonious, say, let's ignore all quantum mechanics for now. Quantum mechanics is a lovely tool. Mathematically it works. Right. But let's just think in terms of particles. I know everyone's having like a field. It was all fields. I'm like, sure. I mean, I mean, I'm a workshop here. I can, I can pick up nails and screwdrivers. These aren't fields. These are localized material objects. So when people start saying it's all fields, I'm like, sure. When I'm tripping. And so if you start at the bottom, what happens is these objects have to kind of somehow they have to create a structure that somehow is able to persist in time more than the other ones. And how does that happen? Well, in physics it can't happen because there is this search problem like, so in physics, the combinations are just too small. You just basically search all the space and it's all collapsed down. When you get to chemistry, you get a combinatorial explosion where the number of possibilities is too big to search. So when the number of possibilities too big to search, you get. Start weird frozen accidents. And with frozen accident accidents, plus kind of the randomness of the universe is, is how causation is. That's. I guess if I. You then use immersion that, that then how causation emerges, but it emerges at the microphysical level and then goes all the way up to memes and AI and TikTok.
- [11:53]
C
Absolutely. So I guess, I guess one way to kind of look at it that I think maybe might help people just kind of understand what. What we're talking about here, if I understand correctly, is basically if you think about the universe, if it's just a bunch of random stuff or whatever, right. There's no memory, there's line, so to speak. And some people might be familiar. Yeah. Like if you have multiple comic book panels and we kind of shuffle them, you might be able to put them together based. Because there's like, this seems to be like a storyline that unfolds. And if life is. And these chemical processes kind of like the universe developing a memory and then it becomes like a story in words. You can kind of tell it moves from state A to B to C, etc. Um, and what you're saying is that's kind of what we're talking about when we're saying causality. Right. That's kind of the idea that it's not just random, there's some sort of storyline or there, Right?
- [12:47]
A
Yeah, the. Exactly. Causation is the presence of a memory in the universe that allows another thing to happen. Right. It's like if you're like typing something up, you know what words you're going to say, you type them, you've caused them. I think the causation is a very interesting phenomena that we don't yet understand. I mean it's kind of. I mean what I'm saying is it's not wrong. I don't think so. I'm really excited. But I think what we need to do is these ideas need to be tested experimentally. The nice thing about coming up with new theory, I get lots of theories all the time. And I say that people giving these theories, like thanks to that theory, people using AI to generate theories, it's kind of depressing and elating all at the same time. That we could talk about is that if the theory doesn't suggest something outrageous that nothing else suggests and you then you can make a unique test. And it's not really a new theory. Right. It's just a kind of recategorization of what we know. And colleagues who do this all the time, there's a lot of famous kind of, you know, pseudo physicists who have big mathematics projects. You can guess who they are. They keep doing this, they keep dreaming up new theories, but they're not new theories. They just predict what we already know. And it's the bar that we have to have is like, if I want to come up with a new theory, such as when call, it has to quantitatively do something different, predict something different to what all the other theories did. And then we go and do it and show it and that's what we're doing right now. And that's kind of mind blowing. So that's good. So it's kind of like I almost get a free pass to kind of keep doing experiments until it breaks.
- [14:24]
B
So fun. So thinking like kind of lower level and how things build up. Has it given you maybe a more specific or more useful definition for life? We've had some great conversations about like what life is, but they seem like each definition doesn't capture something everybody can use. Like where are you on the definition?
- [14:43]
A
Yeah. So I I, I marvel at this in that I thought, I thought that you're right, that defining life is a bit like talking to. Let's first of all get a load of life definers in a room. I would call them artistes and then ask them to draw a picture. They'll all draw a different picture, and it's amazing. And, and I, and I think that probably they're missing something because they're trying to say, oh, life does this, does this, does this. And what I tried to do at the beginning, which actually gave rise to assembly theory, was say, hey, let's stop trying to define life. Let's instead look at what life does uniquely that non life doesn't do. And then ironically, in a way or quite amazingly, it allowed me to come up with a definition of life that passes muster. And basically my definition of life is, well, actually this is paraphrased through the Mythbusters. Diane. Is it Adam Savage, When I was talking to him about it at a meeting and he was. And I was saying, oh, wow, life makes complex things. He's like, right, good. And you can, you can calculate the complexity in some way, but then you have to have lots of identical copies. And he said, oh, so life does complex shit at scale. So that's kind of like the kind of lay version, but the more technical version, which actually is not that different to labor and anyone will understand it is the same. Living systems are able to create objects recursively that have lots of different parts that you can count above a threshold that non life can't do. And if you make more than one copy, because one copy could be just a random. The more copies you have and the larger the number of copies parts, or the larger the assembly index, the more certain you are that only life could have produced it. And that for me, when I came up with that, with my team and my collaborators, allowed me to do something which was really nice, which was to unify viruses and cells, because everyone goes, ah, viruses are not alive. I'm like, no, no, no. Because if you take my interpretation, things that can are recursively complicated and copy number, they can only produce by evolution. So what you don't do is you don't suddenly look at objects in the world and go, is it living or not? You say, was it produced by evolution or not? And so all the objects we know in the universe are produced by evolution, are the product of life. And those objects that when you feed, give them energy and stuff and they can produce other objects they are living. Now a virus becomes living in a cell, but was produced by evolution. And cells are kind of the minimum unit of life. So I think it's nice to get away from this kind of hard lines to say, look, if you have a system that produces complexity above a threshold at scale only evolution could have done that and allows us to reframe the gradient of life in the universe through selection. And it unifies a lot of things. Like my iPhone here is. Is ever. Is live, is not living, but it is life. And what I mean by that is if there was no Steve Jobs, a living individual, and no Apple, there would be no iPhone. Same with my pink mouse. Someone designed it. If you went to Mars and you found 500 pink mice, you'd be like, it's a bit weird. Why there are 500 pink mice. And I, I don't have. Quite have 500 pig mice, but I have three or four because I keep losing them. So it's kind of like. So the definition of life is to say, don't think about metabolism and rna, Think about is it complex and are there lots of identical copies? If that, then you're good. And it's evidence of a living system. That was kind of my eureka moment about, I guess, about seven years ago.
- [18:34]
C
That's very interesting. That's a very different way of thinking about it. Like, it's produced by life. So maybe we would help people also if we go back to the very formation of life most people think of. Like life emerged and then selection, you know, helped it become more complex. But it seems like you're saying that selection was happening maybe even before biology existed, maybe. Can we unpack that a little bit?
- [19:00]
A
Yeah, my conjecture is really, really simple, and it causes everybody to kind of have a kind of. A kind of psychotic break, scientifically speaking. And the reason for this is I'm a child in adult's clothing, so I can just go around and just go, why does that work? Poke, poke, poke. Why does that work? Why does that work? And I've never lost the ability to annoy adults, particularly smart ones. And I think that the problem we have in biology is biologists go, evolution does everything, and don't question it. And they're right to say that because there is this creationist stuff. And physicists are like saying, physics does everything, don't question it. But then you have the poor old chemists in the middle that are joining physics to biology in a very precise way. And the chemists are trying to come up with a story, because guess what? Chemists, Chemists are really great at Making molecules. So if they've given this living system and they're like, how do these molecules get made? The chemist is like doing detective work. And that's prebiotic chemistry. And when I. And in terms of working out the plausible way, when I came into it and I wanted to solve the origin of life, I didn't want to do prebiotic chemistry. I want to come up with a regenesis engine and I want to create life in the lab now.
- [20:22]
B
Right?
- [20:22]
A
That's what I wanted to do. And everyone's like, what? You're completely insane. That's not possible. And by the way, what is life? And you need to go backwards and you need to conform. And the problem when I dug into it is I saw that the creationists are saying life is impossible because there is this massive leap in information required to go from sand to cells. But actually, if you say, if selection was working before life, then the natural aggregate, the natural production of complexity on autonomy and the emergence of machinery and the ribosome and DN and RNA in the cell, it's just a natural consequence of matter battling with the environment to persist over time. And suddenly it pops up. And I think that what I've tried to say is like, look, selection is a force in the universe in the same way that stars are formed by gravity. You think about it, load of hydrogen. Gravity pulls the hydrogen together at some point, gets hotter and hotter and hotter. It breaks a strong nuclear force and you get fusion. Now, if you're lucky and you have enough hydrogen, when it then expands and tries to blow up, there's enough gravity to stop it blowing up because there must be so many stars that go poof. Shit, there wasn't enough gravity. So it just goes. So the process that gives rise to the sun is gravity. The process that gave rise to the first cells was selection. Then of course, when you have this autonomy in biology, you have biological selection and biological evolution. And when I first came up with the theory of my, my colleagues and we said that selection is before biology, all the biologists went crazy. And I was like, but if you're saying that selection couldn't occur before biology, you're basically saying we all creationists. And so I kind of got trapped between these things. So I think, yes, that's a very long way of saying selection absolutely came first. It's a force in the universe, like gravity, and it produces existence out of persistence. And I'm using those words purposely. Existence is a long. Sorry, wrong way around. Damn it. But you could edit this, but you should Leave it in to just show people how hard it is. First of all, you exist. So you're real. You pop up. Wait, I exist? And then you exist for more than nanosecond. So you persist. And in the universe there is. Forget selfish, gene, there is selfish matter. There is matter that wants to persist in time for as long as possible. You know what is the oldest artifact on planet Earth? Do you know?
- [23:03]
B
Artifact? Like a volcano or something? I don't know, like a rock.
- [23:08]
A
Probably E. Coli in your gut. It's 4.3 billion years. Yeah, yeah. So it's an artifact, but it's also living. It's kind of cool, right? The E. Coli in your, you know, that you have in the environment is. Or any cells they have been living since the origin of life 4.3 billion years ago. And that's mind blowing, right? That this living thing that clearly is dying and turning over is the longest record of information we have. Longer than most volcanoes, but maybe not longer than some moon rocks or something.
- [23:39]
C
One thing that really kind of for me helped me maybe solidify this because I think it's a little bit counterintuitive, I guess maybe is the right word. If you think about what is more durable, a chunk of granite, like a very hard rock or something like DNA, which is very, you know, easily destroyed. It's, you know, it gets too hot, too cold, whatever, it's gone. But if you think about it from a certain perspective, the DNA, anything that's like, can replicate itself is going to be more durable because it's able to persist much longer. The granite has only one outcome. Eventually it will collapse into dust. The DNA could multiply and become more complex and just stick around forever. Like the E. Coli in our gut that's been around for much longer than a lot of maybe metals and rocks that we've seen. So I just want to kind of like put a pin in that for, for later.
- [24:28]
A
But yeah, yeah, you nailed it. I think that's really great. I think durable is a. I mean, granite is clearly more durable than DNA, but over infinite time, DNA is. It's a brilliant analogy.
- [24:40]
C
It will persist longer. That's what kind of clicked for me. I'm like, oh, okay, that makes sense. Okay, Dylan, sorry, go ahead.
- [24:46]
B
Yeah, for sure. I mean, so I gu. I guess gravity is like the original selection pressure, the first thing that made these suns. But. So maybe you can explain what assembly theory is. Is that the kind of overall thing that goes all the way from gravity to where we're at and then also where in that chain. Does intelligence show up?
- [25:06]
A
Yeah, I think there's a lot. I've worked out the chain a few weeks ago with my collaborator, Sarah Walker, and we were trying to figure this out. We worked out we've got an idea. So, yeah, assembly theory was kind of born out of this kind of, how do we define life and measure it? And I kind of just realized that there was something more to it as we went going. So what I originally did is I'm a chemist. So I should have said, again, I'm a Regis professor of chemistry. But I always wanted to be, you know, not just doing chemistry. I want to do theory, computer science, mathematics, physics. Just wanted to hack around all the disciplines. Why annoy chemists when you can annoy physicists and computer sciences, biologist at the same time, you know, there's much more fun. And I just love talking to smart people from different disciplines because you can get them to do things, you know, together, you know. So I just asked a very simple question a few years ago, which was, what is the most complicated molecule I could show you? And you. And you'd be go, oh, my gosh, that. That molecule cannot arise by chance, right? So is it a bit of rna? Is it. Is it, you know, some kind of diamond? Exactly. Like what? And so, and they said, oh, that's kind of cool. So how do we actually, how can I can compare a molecule that's complicated or not? And while all molecules are connected are. A definition of a molecule is basically, you know, set of atoms. When you get them together, they have an attraction, so they can basically survive some energy. So they have. And the way they do it is have bonds, right? And chemists understand the abstraction and the reality of the chemical bond. So molecules are literally atoms glued together by bonds. And those bonds have an energy, and you have to put in energy to break them. So that's quite good. So it's a bit like glue. So you're like, okay, good. So what molecule can I have that's really a proof of life? And then. And then if I go down simple, simple, simple, simple, say to a molecule like water, which is just H2O1 oxygen, two hydrogens, or methane, CH4, one carbon, four hydrogens. Those molecules are important for life, but they're not proof of life because they've been made naturally. So what had been happening is, I mean, looking for evidence of life in the universe, oxygen and methane and so on. I was like, well, but they're really simple. There's oxygen and methane everywhere in the universe. What we want to do is find a smoking gun. That would really be proof, like a, you know, a DNA Michelangelo artifact of a molecule. And so I started thinking and I realized that there are these molecules called natural products that you can get out of biology that are like, well, morphine, CBDs. So most of your listeners, viewers will kind of know what opiates are. So like a natural poppy plant producing and also cannaboids, right from cannabis plants and so on. And so these are molecules that are produced by biology. There are other molecules we can make and they have lots of parts. And what I realized, and it was this is when I annoyed discipline number one, because I was like, now there's a molecule. I'm gonna, I and I. And I'm, you know, I'm a closet mathematician or wannabe, like I wanna do some math. So I, I took the molecules and I took them apart and I tried to kind of work out way compressing the molecule. And I. And current and the compression systems that were out there didn't work. And I realized that if I cut the molecule into the parts, I look for reuse. And I put it all in a line because I'm really lazy. I want to lay out my parts in a line. I don't want to have to go back to the beginning. I realized I could do a thing called recursion. So I could basically take a molecule, take a bit off, take a bit off, take a bit off. And I realized I could make a shortest path where I could have minimum building blocks and I could add them together and have the knowledge. Once I've added it together, I have that knowledge. So I can then carry it to the next step with no penalty or minimum penalty. And by doing that, I was able to basically count the number, number of steps I needed to make molecule A versus molecule B. So I came up with a new counting scheme. And this annoyed the computer scientists because normally they use a thing called zip to compress things, right? And they were like, no, that's wrong, you should use zip. And I was like, but zip doesn't have cause it doesn't have recursion in it. And I realized the recursion is really produced by causation. At that time, I didn't know. I was just like, this looks quite good. I came up the assembly index. So I basically made an algorithm that I tucked in a molecule and it gave me a number. I was like, oh, that's good. So I've got all these molecules. I just took all the molecules I could and just Calculated their assembly index. So I calculated them. So large molecules with lots of different parts had a high assembly index. Small molecules with not very many parts had a lowest assembly. That's great. And then I realized one day when I was talking to one of my colleagues, a student, that I could actually take the molecule and put it in a machine where I just basically hit the molecule with lots of energy and it falls apart. And if I count the number of parts, I thought could the number of parts somehow be the same number as the number of parts I calculated in my theory? And I was like, nah, that ain't gonna work. I went and did it and it worked. I was like shit, really? So I kept doing this and I was like, oh, so complicated molecules that have this high assembly index. As I calculated in theory, I could do experiment and measure it. So then I went, okay, great, maybe the high numbers mean life and no low numbers mean dead. So then I started to basically get samples from the environment, from volcanoes, from meteorites, from the seabed. NASA gave me some moon samples and all sorts of stuff, right? Murchison meteorites. And then they gave me ones from fossils and they didn't tell me they double blinded them. And when we basically took all these samples from all around the universe, from space, from earth and we put them and we blew them apart and we count the number of parts, we then able to get sample A, small number of parts, high number of parts, sample A, low, sample B, large number of parts, sample C, low. And they were all blinded. And we then said, oh, that looks like it's dead. That's like it's live, dead, live, dead, live, dead, live. And then we gave it to NASA and we did the big unveil and we 100% correct. We were able to predict using this technique called mass spectrometry or mass spectroscopy, whether a sample had been produced by life or not. And that was how assembly theory was born, where you have the theory for counting the parts of the molecule and then the experiment for measuring it. And that just blew everything apart.
- [31:58]
C
Wow, that is so. That, that's incredibly interesting. So this is idea of causality and some sort of a memory. So as things build up in complexity, we're able to say at some point, okay, past this line, it's likely to be life once it has enough parts. That's. So I'm curious and I don't know if these two things connect, but just wanted to ask you. So for example, with Google, DeepMind, AlphaFold, right, they fed all the 3D structures of all the proteins that we've discovered very slowly and it was able to predict the 3D structures of some proteins that we didn't know. So I always was fascinated by that because I'm like, what is that mean? Does that mean that there's some sort of a pattern to how proteins fold in order to be, to have some sort of a function? Does this connect in any way to what you're talking about? Like there's some underlying pattern or.
- [32:52]
A
Yeah, for sure, for sure. That's a really nice way of putting it. I mean like I'm. You probably know me, I'm normally a downer on all AI. But the thing I love about AlphaFold and a lot of these systems is like, although, although they don't necessarily know physics, these AIs what they're able to do. If you have clean representations and clean data, you can then basically look at correlations. And those correlations are not magic, they come from somewhere. So the reason why protein folding prediction works is because the evolution of proteins is happening in the causal system. So that's a beautiful example. Say the alpha fold, if you come up with a sequence that is then and relying on an in distribution prediction, you're going to come up with a very nice 3D structure that's likely to be very close to what you would actually get. And that's because the system, if you like learnt is. I don't like using the word learn because learn means understanding to me, but I would maybe use represent or encodes a latent space of folding. So and it represents and it decompresses it to that which is amazing and I think if it wasn't for that causation because the way protein folding works is you like you have a, let's say you have a nuclear. So sorry you like, you got the, you know, the ribosome and out comes the first nucleation site and you grow, grow, grow, grow, grow. And as you grow out the amino acids, they rotate and they, and they fold and they, and they basically start to condense and then, then you, you literally print the protein out. It goes right into the big wide world. That step of adding in the seat, the amino acids in sequence allows the structure to evolve in time, remove the water molecules and compress itself and be ready for action. That process is super amazing and is also, is very causal because you have this. The genome has recorded the past, what's happened and the genome every time it's read is unraveling that causation and making unit's ability to then Go out and act in the world. And so you're absolutely right. It's a very nice way of putting it, actually.
- [35:02]
B
So if we have an understanding of sort of how much complexity is needed to find something like truly alien and we can start diverting resources away from looking for like methane or H2O as the precursors because it might be a waste of energy and direct it towards some kind of other tool. What could you imagine potentially building to detect now that we kind of know what we're looking for a little bit better, like aliens or something of higher level complexity out in the universe.
- [35:31]
A
So, so if we're going to go nearby and we're going to go to our, you know, Mars and also maybe to the outer solar system a bit further away as we could go and send mass spectrometers. And in fact, just so happens you wouldn't, you couldn't make this up. It's great. People at JPL are going to send a nuclear powered drone to Titan, right? Like literally a drone, like imagine a DJI drone, but rather than being powered by a lithium battery, it's powered by a plug, a slug of plutonium, right? And they're going to send that in a couple of years. It's called Dragonfly and literally it's a mass spectrometer with wings, well, with rotors. And they're just going to fly around Titan and just sniff the local environment. And so it's going to basically image the entire moon with this mass spec. And I'm really excited because I'm like, guys, you don't need to change your design of your machine or anything. Just get me the data so we can apply assembly theory to it to see if we can sniff some aliens on Titan. So mass spectrometer is one. The other thing that I did whenever, I mean, I've had so much pushback, right. You know, many people still think that what I'm saying is completely wrong and it's all nonsensical and I'm just making it up, which is fine. You know, science is a bit like that. But addition to mass spec, that where, where mass, what mass spec does is you weigh the molecule and then you cut it up and you just look at the, the lot, the reduction of weight and then you recombine it. There's another technique called infrared spectroscopy where you shine light and it's a bit like, I don't know, someone with a, with a, with a, a cloak of different, that absorbs light. And you know, if you have an invisibility cloak, it Absorbs all the light, doesn't it? But if you have an invisibility cloak that doesn't quite absorb all the light, you can work out where it's absorbing. So, so what I've worked out is that I thought that using infrared radiation, where molecules absorb, if you can count the number of lines the molecule absorbs, you can count the assembly index. And we did that as well. And it works. In fact, I've come up with a design for if we go. If we want to go out and basically cloak ourselves. I've come up with a concept of what I call a non gravitational delensing complexity cloak, which is a very fancy way of saying I'll take a big sheet of carbon, very light, like a sail, but I'll put lots of different bonds in. So it absorbs all the radiation from the gamma ray to the infrared. So it would just look like a dark blob in the sky. And if you do it really properly, you should maybe look at just emitting the microwave background. So it just looks like a bit of microwave background moving away. And that would be one way to basically, you know, if you've got a complexity cloak, you can literally take all the radiation, absorb it, turn it into microwave radiation. So you could really, you know, that's probably what the Klingons had, right? Or whatever it is.
- [38:14]
C
Cloaking devices.
- [38:16]
A
Yeah, something like that. But what I'm saying is the reason why molecules absorb different lines, they have different parts. So the more parts, the more colors you absorb. And that worked as well. And I was like, wow, this is getting to the point where, because in chemistry no one had done this before, no one had realized that complexity is intrinsic measurable part of molecules. And it's still like 2 years old. So it's not like get abuse. That's unfair. I didn't get abuse. What I get is criticism. And just to tell the people watching this or listening to this criticism is a necessary part of science. It's not bad, right? Criticism is a good sign, unless you're just completely wrong, then it's a bad sign.
- [39:00]
B
Or maybe just sometimes resistance. Because people already have money and they've already worked in a certain direction. And just to change, there's like an opportunity cost that seems hard to change for sure.
- [39:12]
C
Oh yes. Yeah. I was thinking, we just recently talked to a few people that are scientists and they do have criticism and it's usually because they're doing something that's triggers maybe some fear in people that it's a little bit different. This is very common I think that certainly, you know, we mentioned Feynman. I feel like he probably annoyed people as well because he went against the grain, so to speak. But I guess so in order to start approaching AI and what it does right, what it does wrong, whether we're going to get AGI, you know, I mean, maybe can we start with. Because what I am noticing, a lot of people, first of all have either no definition of intelligence that they can like test or explain fully. Reasoning is another word that people don't get. Understand in order to understand something. What does understanding truly means? You know, you use the terminology, you know, maybe it's encoding something in a latent space, but that's different from understanding. So maybe can we start with maybe unpacking, like, what is intelligence? When we're talking, like for scientists, for engineers, not philosophers. Like, if you're an engineer, how do we define intelligence?
- [40:23]
A
You have to have a philosophical background, right? One of the people I was talking to earlier today was a philosopher. I don't agree that you should ignore philosophers, but I think philosophers can't ignore scientists and engineers. So I think you're right to ask from an engineering point of view. So intelligence is a super interesting thing. And the problem is, right? So our current evolution at AI, there's a lot of stuff out there where people are just basically peddling stuff because they're selling stuff. So I'm not going to bother complaining about that because it happens everywhere, right? I don't think that's productive, you guys acting in good faith in this stuff. So I'll give you some views I think might be useful for people kind of taking part listening to this. So intelligence in a way is. Let's all go back and say, what is the causal chain of intelligence? The universe. The intelligence we know in the universe is you go back and say, origin of life, it starts there. Then you have the cells that kind of differentiate and you go, you know, go from single cell creatures to multicellular creatures. And then when multicellular creatures started to sense the environment, they have this problem where they can't. Because intelligence before intelligence, there was just evolution. What do I mean? Cell is mutated. If it's, if it's lucky, it can continue. If it's unlucky, it dies, right? So you have this kind of instantaneous evolution. So evolution is like, ah, you're dead. I'm sorry, you're alive. Great, continue. But as soon as you become multicellular, you start to sense things. And if you can see your predator coming in real time, you're like, oh, shit, I'm about to get eaten. And maybe you're seeing someone else get eaten. You remember what happened. You want to survive so you can get out of the way. So then you create a memory. So if you've got. And I'm doing this, I've actually written these notes down. And you're the first people to hear these notes. Not for this, it's just I was working with Sarah on what intelligence was, and we were trying to figure out how to put it in order because you have all these things. So evolution gives you sensing real time. You get memory. And then for all that to work, you have to have a thing called consciousness, because consciousness, you need to be aware of your surroundings. And also consciousness is important because you access the past memories. So in a way, so you kind of remember what happened, and you can then see what's going on. And then you have a thing called imagination where you can imagine being eaten, or imagine getting away or imagine doing this or imagine doing that. So you can see how you've got this hierarchy of evolution, sensing in real time, memory, consciousness, and imagination. And then you have free will. And free will then allows you to make a decision and that the sum total of that is what intelligence is. And you can see that there's a lot in there because free will is an intrinsic part of intelligence. So. Sorry, dude.
- [43:12]
B
No, no, please. So this is a little bit less of a question. Just something that over the last couple years I was excited to have the opportunity to share with you is, man, maybe you'd want to run this by Sarah, maybe not. But the one thing that's occurred to me is a lot of times in computer science you come across this question about exploitation versus exploration. And there's all sorts of equations that tell you the exact amount of times to pull, like a slot machine before you move on to the next machine, because you're discovering how much it pays out and. And on the lowest level of intelligence, I've had some thoughts that maybe one way you could kind of look into this or kind of a north star you could kind of think about would be if a system knows to stop exploiting itself. Like a bacteria that will just eat all the food in the petri dish and then die of starvation is, you know, exploiting its food too much. And the idea that understanding the environment enough to start exploring and slow down seems like something. Have you ever thought about exploration versus exploitation in any of your work?
- [44:14]
A
Yeah, I think that's. So I have. And I think there's a really nice Insight, and I would actually add another label on, is that living systems tend to be curious, but curious. I think curiosity is born of. You've just kind of crystallized it here just now. So curiosity is what happens when exploitation and exploration are balanced. Because why are systems curious? What's the evolutionary. Because if you're too curious, you die. But if you're curious, you're storing up new information for the future. So you have pesky kids or go out, do that thing. And why do you press that switch? What was there? Or why did I do that? Well, I wanted to see what happened. And so I think curiosity is this balance between exploration and exploitation that allows you to then put something in your memory, go, oh, I know if I do that, do this. And so it's almost a bit like the reason why evolution is currently creating new technologies, new competitions and things is we're all trying to outcompete one another. And the only way to get those seeds, those novelty, to compete is to basically be curious. And that curiosity is basically, when I've done enough exploitation, I need to do some exploration, but I don't want to do so much exploration that I go off peace and die. So I go back and do exploitation. So you have this kind of balance between the two. And so I think that's a nice way of putting it. I mean, I hesitate to do too much computer science analogilization, if that's even a word, apologies. But actually that even if it's just a metaphor, it's a beautiful one. And actually I think it's quite significant because, you know, living systems that are curious are ones that would be, I. I would say is a unique identifier for intelligence. Object objects that have free will and curiosity are uniquely intelligent. And that's how I would measure it. You see what I mean? That's kind of cool.
- [46:10]
C
Yeah, absolutely. So just one thing from the earlier. So we went from evolution to sensing to memory. So that makes perfect sense, right? Develop some sensing proteins in our eyes, picking up photons, whatever. Then some sort of a way to store that information, memory. Then we go to consciousness. Can we touch on that for a second? Because that seems to be another big foundational thing. Before intelligence, before free will, how do we define consciousness? How do we test for consciousness? Because that's going to be a big question right now because there's some people are saying there's no way these systems are conscious. And I'm not saying they are, I'm saying how do we, in order for us to say they're not, don't we need to have a test or something? I'm not arguing that they are, but how can we say we're not if they're not, if we can't even test for it? So maybe can we unpack consciousness a little bit?
- [46:58]
B
It.
- [46:58]
A
Yeah, yeah. So the way I would get there, it's a good point, is I think a lot of these come together like memory, imagination. So memory, consciousness and imagination and free will are kind of bundled together in layers. So I would say the only way I can think of really measuring consciousness is by the amount of imagination. So if I went to ChatGPT, I've already built an assembly theory based consciousness measuring system and all the AIs fail because they don't do any. Because what happens is you basically know what's in your, in the, you know what's in the environment. That's what you've learned on, that's your model, right? That's all the data. So you kind of know how to be surprised or not. So you know, when you're applying the gen AI, it looks cool. We feel emotionally responsive because we can't know everything. That's why humans respond in weird ways to AIs. But you can see it's flat when there's act of creativity, it just, it just spikes. So, so I, I think what I'm saying is consciousness and imagination and free will need to bundle together to give you intelligence in that layer. But I'm, this is very new to me, but I would, the argument I would have is that the only way you can measure intelligence in a. And we'll define it, right? Because like people argue with me, oh, these things are intelligent. I'm like, no, no, you, that's not, that's not intelligence. But, but I'm not demurring. I'm not, not. But I'm not looking down on it. I'm just saying intelligence is where you can in real time solve a problem using your free will, imagination, curiosity. And all the systems we've created thus far don't have that. But they're amazing tools, right? And it's just the word AI is just such a bad word if we called it anything else. We're like, look at that widget. It really does this thing, it's cool. It's like, yeah, let's use the widget. But the fact that we're exploiting the widget and kind of anthropomorphizing causes a great deal of complexity. But coming back to your point, point, I would measure consciousness by Measuring and therefore also degree of intelligence, if you like, by imagination. Now, dogs are conscious. How do I measure the imagination of a dog? That's not so easy. Right. So it's clear that what I'm telling you is not fully formed. But for humans who have language, because then the next thing on my list is like, you have memory, consciousness, imagination, free will, and only then are you able to make language. Language is only comes from people, from things that have free will, that have had imagination, have consciousness. And that is super interesting because you imagine lots of people going before they had language, but they were. They were conscious, they had memory. They're like. And then. And you can see that they would then converge on sounds that they could use to represent things. And then when they became Turing complete, they extended them, and you brought in the Alphabet and all that stuff. And we have everything today. So we're kind of like, trapped in not understanding the history of life. And one of the things why I'm so. I feel so passionate about getting involved in AI discussions is that understanding the origin of life and the emergence of kind of multicellularity is the only way we're going to really understand intelligence as a standalone phenomena in our culture. It's really important to work out where we are and where we're going. Sorry, that's a really long statement.
- [50:26]
B
No, that's really. Well, I mean, one of the things that seems so fascinating about AI, when it's built in certain ways, it seems like it almost predicts the future. You know what I mean? Like, sometimes intelligence seems like it's about noticing patterns and saying, like, okay, if those patterns have kept so far, I predict with high probability this will happen next. And that's kind of what's happening with next tokens and different aspects of AI like that. That. But I've heard you say in another podcast that the future is not in principle predictable. Maybe you could help me reconcile what was that in a different context? Or what do you mean by the future is not, in principle predictable?
- [51:06]
A
Yeah, so this is where novelty comes from. So we have this ridiculous. So. So there are. There are macroscopic. So totally. Right. So the future is not predictable. If the future was predictable, then would be boring. Right. Obvious. But what I mean by that is, obviously there are. There are. There's regularities, but we shouldn't confuse regularities with predictions. So what I say is, like, the thing about large language models and all the AI, they're really brilliant at predicting the past, right? Which is kind of like. Well, it's like almost like A pejorative like work has always, always happened, right? It probabilities only work when you have data that you can basically mine and then calculate on. And so let me answer your question directly. This is super interesting. So if I'm going into a situation that has basically happened before, then I can then use that data to predict what may happen. So I like the idea of going in and going, say in Las Vegas, let's say we go gambling, we go some slot machines, right? I know what's going to happen. The slot machine isn't something suddenly going to turn into an alien and I don't know, come up with a porthole gun and take me to see Rick and Morty. It's a slot machine. Although maybe. Right. What's in my head, on my head, however, if I basically go to a completely new environment that is unknown to me and maybe unknown to, and I bring together lots of new things that never happened before, I can't predict what's going to happen probabilistically. I can conjecture what's going to happen in a possibility space. Let's say I'm going, right, right. I've never built, I don't know, let's invent a new thing. Let's combine a 747 with a Starship. It's kind of interesting. So we take the starship and we just make a bulbous bit. And so it's never been done before. We can imagine it. It's probably going to be when it launches it a bit lopsided. 747 Starship. We might put the engines on. It won't make a difference because the starship is so powerful. So I can imagine in my possibility space in my brain, if I can combine architectures together what that might look like. But I can't really predict what's going happen except by using some, you know, architectural approaches. So if I'm going into an environment that is basically fairly known to me, I can predict when I'm going into environments where there's lots of. There's not very many parameters defined and lots are happening, then the novelty is higher. And so what I'm saying is in those. And this is like, you know, I'm kind of saying to people use AI to mine the past. It's a great search tool. AI, if you want to use it in the best way, is a search tool in the worst way. It is a bullshitting tool where you basically get confabulations back and you don't look at it critically. So you just basically, you Know, I'm going to write a document using AI. Here's my document, give it someone. You haven't read it. They read it and go, oh, that's interesting fact, and take it as fact. And clearly they then use that to put into their AI. And before you know it, you've just got confabulation on confabulation. That's not a very good use of the tool. And because that tool is probabilistically sampling the past and giving you a new outcome. But the nice thing you get from that probabilistic sampling is you get an incredible ability to access such a huge space of data and your chances of getting things sensible is quite high. So someone said to me, like, you know, you're on a desert island, you've got no encyclopedias, but you've got ChatGPT or no ChatGPT. I'd take ChatGPT every time, because I would then go, you know, what temperatures does water boil at? And be like, well, you know this. I'll be like, I can check. That feels right. I can do it. Interrogate. And so we have this issue with misunderstanding what Genai does. It's trained on the past. But did chatgpt predict that Putin was going to invade Ukraine? Well, actually, probably did. It happened quite a lot of the time before, actually. That's probably a bad example. But, you know, let's think of something, you know, really bizarre, that we weren't expecting some kind of technological miracle or some kind of political event. Right. You know, I don't know Trump's second term for his first term. I don't know if Chat GPT would have predicted that. So we've got this kind of interplay between the amount of information we have in the past that gives us regularities that we can predict on and then possibilities that we cannot possibly know. And so running simultaneously, we have a probability space and we have a possibility space. Who would have thought that Elon Musk would have landed a fucking rocket? Excuse my swearing. Am I allowed to swear?
- [55:53]
C
Absolutely. Absolutely.
- [55:55]
A
You can say. Who would imagine that Elon Musk would have landed a rocket on some chopsticks? I mean, I didn't think of that. That wasn't in my probability space, but it was certainly an Elon Musk's possibility space. So I think there is this interplay of those two things together I think is super interesting. One doesn't usurp the other. They just serve different processes, different purposes.
- [56:20]
C
Absolutely. I'm curious to think, to know what you think about this. So for Example with some of the most interesting AI projects out of Google, DeepMind, for example, Alpha Evol, they used Gemini, right, their version of ChatGPT to solve various math problems, to optimize some hardware, optimize how their data centers run. And they would ask it for, you know, how would you do this? And it gave them, you know, a thousand different answers. And most of them, or some of them were complete B.S. like you said, it's a, you know, bullshit producer. But like you said, some of them were viable. And so for each one, they would test it against some known metric. Like. Right. Oh, if it's, if this ID improves it based on this metric, then then we continue down that path. So it's almost like an evolutionary tree search where the things die off. And I mean, they did have some very sort of seemingly interesting breakthroughs with that. So in, so there, these neural, these AIs would be very useful as long as we can just push them through this evolutionary tree search. Because them being able to BS and come up with a million different ideas actually kind of helps in that sense. Right. So they could be very powerful with that, but they're still not what intelligence is, right?
- [57:39]
A
Yeah. Yeah. So look, I think, I mean, it's kind of good for me to kind of clarify my stance on AI in general. AI is incredibly powerful, but it's a probabilistic representation of things. Right? Now what I know what Demis is trying. He wants to solve intelligence. He's doing it wrong. The only way he's going to solve intelligence is solve origin of life because we don't know the features in the brain, the mechanics in the brain that understands probability, impossibility. But coming back to what you're saying About Alpha evolve, AlphaVolve is very good. It comes up with solutions that we as humans with agency because we're the only known source of agency in the universe. Maybe your dog, if you have pets. Pets do. Right? Complex multicellular creatures have agency and AI has none. And it's never going to have it, not in the current architecture. And I'm pretty confident of that, which is kind of, kind of makes people sad and angry and whatever. They have various debates and we could get into why I'm so confident, because, you know, when anyone's confident of something, they're always going to be wrong. But I think I can give you a reason, a good principled reason. So AlphaVolve is good, but it's evolving in a digital world, which means its environment isn't necessarily boring. The really interesting thing about the real world is it's continuous. And when you basically then sample the real world, you make a decision to label it. And how does that happen? There is a mystery. There's so many mysteries we talked about. So the mystery is when you sample the real world and you basically decide to binarize it, and then you then take that into your model and then decide on it, that gives you something, and then if you go and do it again, you get something different. Because what we are falling into the trap of is we think that AIs have anywhere near the creativity or potentiality of living systems. They are nothing. They are nothing compared to living systems. They are zero. And if you say to most people, they're like. They're like, oh, my God, but I can do this thing. And I'm like, and what I mean by zero, I mean a very, very, very, very small number of the possibility space. That doesn't mean they're not useful. They're incredibly useful. And if you give AlphaVolve a problem, you know, an expectation value and an environment, it will get there. And why not? Why not use it? It's like using linear regression or nonlinear regression. The mathematics works, gradient descent works, but that is not intelligence. Intelligence is something we don't yet understand. And I'm not sure that we're ever going to understand it by using a computational neuroscience approach, because the brain is not a computer.
- [60:23]
B
Okay? So, yeah, I would love to hear, why won't AI have agency? And then also when you use the word agency, is that the same as free will? Like, do you believe people have free will?
- [60:34]
A
Yeah. So then let's recap, because I really like this very powerful. If you have memory and consciousness and imagination, you can have free will. Will. That's kind of cool, because I think that's something that we haven't yet. We can program in the illusion of agency into an A. This is why agents, like, it's like, I feel like I've grown into a nightmare. It's like artificial intelligence. No, it's not intelligent. Then you've got agents. I'm like, no, they're not agents. But I get it. So it's like, damn it. So of course you can. We have things called agents that we can give a complex tree search to. And if they're not too brittle, they will do stuff. But, you know, agency is a thing that I'm guessing at the moment, I don't know what it is, other than I know in other humans where I can see it. And again, in higher primates and animals and so on. And so agency really comes from your need, your wish to survive. And if you're a multicellular organism with memory, consciousness and imagination free will, you tend to want to survive, right? And you do stuff, you think ahead, you've got agency and you have, you have such complicated set of, you know, of wants and requirements. So it's very difficult to kind of, you know, to tease those apart. You can flatten them and chat. GPT does it in the therapy mode, right? It's just, you know, I can hack anyone, but same way I can hack anyone like I'm, you know, I can flatter you and I can say you're the best person in the world, you've done all this wonderful thing and you're good looking, you're really intelligent and, and you know what ChatGPT does, right? It's this kind of, this. But agency is a very special thing that comes from this free will. And yes, of course, no computer systems have ever displayed it. In fact, one of my dreams is to create a artificial form that has its own agency. A friend of mine, who I'm very much inspired by, a computer scientist called Susan Stepany, she told me when I was going into the artificial life world, she said that her dream was to write a computer program and run it it and be so excited and so surprised by the outcome of the program that she would think the act of turning off the computer would be an act of murder. An act of murder. I was like, holy shit, that's so interesting. See, and ever since I had that conversation with her, like, I guess it was probably 15 years ago, I was thinking, how could I create a life form that had memory and consciousness, imagination, free will and intelligence? And not that I want to basically euthanize it or something, right? But I would similarly feel that I would do anything I can to make sure that entity survived. Because it's a precious thing, it's a living thing. And yeah, and AIs don't have that. And this is the problem with AGI and all the other stuff we have right now, because it's not that AGI is not a worthy thing, we just keep mislabeling it. And the only thing that's happening here is the tech bros are trying to sell us stuff and they are in charge of the funding. So the academics and the scientists and the technicians are kind of beholden and they have to basically, they have to basically worship at the altar of the AGI or they get nothing. So I feel kind of annoyed that we you know, the printing press. What did the freaking printing press do? Printed the Bible, gave everyone the Bible. What have we done when we've, we've got the answer to basically processing vast amounts of data? We make chat bots and make everyone think they're small.
- [64:12]
C
Smart.
- [64:12]
A
It's like, damn it.
- [64:15]
B
Yeah. From a social context, I am very frustrated with the terminology that we use and the way that it's sold to us and that whole thing. But before I go down that direction, just one last question. Like on your Santa Fe Institute profile, I guess you say something about wet chemical computers. I just had a thought that if you were to take something like neurons and put them in some sort of giant petri dish type thing and use that as an AI learning system, in that case, in that hypothetical case, would it then be able to have consciousness and agency because it's chemical in nature?
- [64:53]
A
Yeah, you just. Yeah, I'm, I'm, I'm actually, I have a, a project or it's called Ken Machina, but it's going to turn into living circuits where I'm like, okay, you know, I'm okay. Okay, Demis, you think you're going to solve intelligence that way? Okay, cool. That's really smart. Smart. You have a Nobel Prize, you're a genius. You work for DeepMind. I'll take a different approach. So if I can try and understand origin of life and then work out how chemistry gets to intelligence, one possible way of doing it might be to say, well, why can't I try and instantiate a type of learning system in chemistry and then look at the difference between the chemical system and the brain? And so it seems plausible that we can get there, but for all the reasons I told you. Why AI so intelligence doesn't seem yet evident in the systems we program. I think we should be similarly cautious to my approach because it hasn't evolved all the way back, but it might get us a step closer or it might be an add on, it might be something else. Because the nice thing about having a blob of chemistry, if you think about, let's go back to what the brain is. The brain is few kilograms of neurons. There's about 120 trillion neurons, I think, and each neuron has 10,000 connections. So that means in your brain alone, there are more potential connections possible that you could do when you're thinking, imagining stuff, than there are atoms in the universe just in your brain. Now that's your brain is moving, it's morphing, it's thinking. You take a silicon chip it ain't morphing, it ain't moving. You've got electrons on it going into registers and things. But your brain has a configurational space bigger than the universe. When you're thinking, that's. That literally blows my mind. And that's why I'm not scared of AGI in silicon right now. But I'm like, okay. Damn it. I watched Pickle Rick and Rick and Morty. And actually after that episode, I was like, can I make an electrochemical brain and a pickle? And I did. I did one in this workshop, actually, where I'm in right now, and all I did is I basically put the pickle at 240 volts, so it's almost exploding because of steam. And I put in some low voltage electrodes, and I tried to basically make a little circuit where I could teach the pickle to tell the difference between a triangle and a circle. That was easy, right? And it was able to do it. How does it do that? Well, a pickle is like, it's got water, it's got fibers, there's sodium, potassium moving around. It's like, okay, okay. So what I need to do is make a gel, a brain gel, where I can put electrodes in and tickle them, and then basically they would grow and basically interact. And I could read out the response of the electrodes from inputs. So that's where living circuits got born. So it's just happening in the next few years. The idea is to make a machine learning system that's all chemical, and that's where the chemical computer is coming from. And I think, think. I don't know if I've got one of my. One of. Yeah. So this is one of the. One of the 3D grids i3D printed, which is basically a matrix of chemical reactors. You can see. And the idea is to put chemical oscillations in this and this grid and basically make a small chemical brain in this. But, you know, it's kind of funny. It's just behind me.
- [68:16]
B
Yeah, dude, Wes, you'll have to tell him about Butterbench. I feel like that's the perfect benchmark to see when he succeeded.
- [68:22]
C
Oh, there's. Oh, my. I have so many notes I want to come back to.
- [68:25]
B
Remember the Rick and Morty episode where he's like, you pass butterfly.
- [68:28]
C
That's. That's the. Yeah, Rick and Morty. A lot of these projects, they, they, they kind of like reference it a little bit, which I absolutely love. Somebody else is making. They're using human and rat Neurons to teach them in a petri dish to play the video game Doom. So I do, I do want to come back to everything that you're doing. Kim Mahina is such a great name, but maybe it's not an easier name would be more memorable for people. But I, I just want to come back to one thing that we were talking about. So you know, because you're saying that because AI and maybe that's a wrong term, AI agents, they don't have agency, AI doesn't have intelligence. But certainly they have something that is able to do work now in this, if we kind of think about that, that might be the best possible outcome for us because we have these amazing tools that can do a lot but because they don't have agency, they don't have true intelligence. A lot of the AI doomer arguments kind of fall apart because no matter how good it gets at running through what, you know, information it's helping us with, it's never going to like come alive and have a sense of self preservation and then attack us to, to get the resources or whatever nonsense. Yeah, yeah, can we talk about that? Like so you're, you're not worried about Skynets or X risk from it going rogue specifically? That's not even. Nothing, Nothing to worry about.
- [69:49]
A
Right, well let me, sorry, let me,
- [69:52]
C
I phrased that wrong. Yeah, sorry.
- [69:54]
A
No, no, no, you asked a question. Nothing to wrong. Yes, I don't think there's anything worry about. Let's do it. So convince people. Listening to this, I think one of the things I don't like about the Doomers is literally they're looking for attention and like Eliza, I don't know what's got with him. I, you know, I wanted to try and debate him. He blocked me on Twitter and all sorts of things. But what he says is like do, do, do, do and then you're going to die. I'm like, okay, so my counter to that is to say there's so many things we could die of in, in the world. Right. And you know, I, let's think about it. If, if we didn't have this stupid anti nuclear movement, we would have fast feeder nuclear reactors all around the planet just now. No energy problems, right? In that we don't have an energy problem as is, but we're just burning more fossil fuel than we need to. In the UK we have 300 tons of plutonium at Sellafield. 300 tons. That's a trillion dollars worth of energy that we could just put through fast breeders. What happened is all the doomers went, oh no, nuclear, we're all going to die. So no nuclear reactors, right? The only country that's really got its stuff together was France and then obviously China and also the US Actually. Huge amount of nuclear infrastructure pressure, not very talked about. So Eliza says, okay, you know, AGI only has to get there once, it'll kill us. I'm like, but you don't even have a functioning definition. So I, I want to make, I made up a counter one say, right, I'm afraid of ag. What is ag? AG is anti gravity. If I create anti gravity in my workshop behind me and it replicates, suddenly everything's going to float away, we're all going to die. Scary, right? Right. Could happen tomorrow, right? Well, you might go, well, what is anti gravity? Well, I've told you and you're like, okay, give me the mechanism by which anti gravity is going to do its thing. And that's where I fall flat. And so what all the doomers are doing, they're literally making stuff up. Because there is no mechanism by which an AGI can become super intelligent, get control of the systems, have agency and do anything. It is nonsense. And what I find really annoying is when the press release, oh, there's this AI tool that copies itself and hid itself. Like, no it didn't. Some programmer did that and they did it on purpose to create nonsense. They could tell the press to make themselves sound scary or whatever. And so agency is a very hard won thing. We as human beings, it takes us, what, I don't know, some of us till we're 16, some 18, some 21, some of us even my age still are not competent adults. But I'm working on takes a long time to become a competent individual. And so the doomers basically say this bad thing could happen because magic. So when someone says to me, this bad thing could happen because magic, what I mean by magic, they can't give a mechanism. I say, but look, can we rather instead think about the problems I see with AI? Fake people, fake data, political manipulation, mass hysteria. Because we're telling things, you know, let's say we, I don't know, generate an AI where there's a nuclear bomb that goes off in whatever country and whatever place. It looks really authentic and we put it everywhere at once. That type of stuff is really bad. And I think that these people are not really acting in good faith. Some of them are super smart. You know, Jeffrey hints and is saying, oh, he's really thought that, you know, the thing was alive. He's either really stupid or needs to see a psychologist. Right. Or he's doing something else. And I don't say that lightly. And I think that we, you know, thinkers like me, whether you have a Nobel Prize or not, you need to actually have a bit of humility. We don't know what intelligence is. Therefore, to imagine we're going to stumble on it by accident and make Skynet without understanding the mechanism is kind of nonsensical. Equally, you know, why would we build the initial computer like, you know, the first when Microsoft came out with Windows 3.1? So anyway, I'm diverging a lot, but I think there's a number of things to say. Number one, the doomers don't have a mechanism, they're just making it up. Number two, they're distracting us from where we should be putting our attention, where we have to do regulation and we have to make sure that humans are safe because we can't have unregulated AI. We can't have chatgpt. Being a therapist because it's basic. And being a addictionist, right. It's like if your therapist wants to basically keep you coming to, they're going to make you addicted to them. Right. It's a kind of psychotic process. They're not making you better, they're making you mentally worse. There's all sorts of very interesting things we need to solve. And I think that's why I'm against P doom because actually it's nonsensical. And also, if you basically fly over the earth in an airplane, you can see most of the earth isn't covered by the Internet. You'd have to be really damn good to kill humanity. I mean, what are you going to do? Like, even if you let off a few nuclear weapons, that's not really going to kill humanity. It's just going to basically annoy some people in a few neighborhoods. The world is big. You'd have to get every single nuclear weapon to detonate simultaneously, which is not ever going to happen. Like, there is no probability of that. There is no possibility of that. You can calculate a probability, but as soon as a nuke goes off somewhere, everyone's else is going to turn off their nukes and go, what just happened? Quick, turn it off. So yeah, it's irresponsible.
- [75:21]
B
Okay, I love that. Yeah, thanks for sharing all of that. So before we go, like, I want to ask, is there anything else that you'd like to talk about? We definitely want to spend a second just to talk about all your projects and make sure People know. I'm also curious if a couple of people you mentioned, like Susan, Stephanie, or other people you maybe think we should interview. And if you have a favorite Rick and Morty episode, go ahead.
- [75:44]
A
My favorite Rick and, I don't know.
- [75:46]
B
Pickle Rick, probably, because now you're building around him.
- [75:49]
A
Pickle Rick is the boring one. But, yeah, Pickle Rick is pretty fascinating. I like the. I mean, they're all a bit weird, right? I think every episode of Rick and Morty is so many different science fiction projects you could do that I couldn't possibly name one. I just love them all. They're pretty amazing. The swearing makes me look good because I swear a lot, but not as so much as they do in Rick and Morty.
- [76:13]
B
Because you're not burping between. No, I.
- [76:18]
A
The character is deeply complicated as well. I'm deeply simple. No, I mean, I think so. One of the things I would say is I'm fascinated by understanding the process of computation and applying it to chemistry. So that's one thing we didn't talk about. So I've built what I call computer, where we basically instantiate a Turing machine and a chemistry robot to do synthesis. So I actually run a company that does that. So it's kind of my day job. The science is kind of the night job. And the other thing I guess I should mention is the team of the assembly theory. I'm working a lot with Sarah Walker at Arizona State University, who's helped me a lot develop the theory. We're working together between our groups and also people at Santa Fe Institute to really kind of understand the intersection of complexity theory. We haven't talked a lot about that. I'm very interested in how assembly theory and complexity theory complement one another. I find it very frustrating that computer science at the beginning was saying that all the universe should be kind of couched in terms of computational complexity, and I think that was a misstep. But equally, I think they play well together because there's this interesting interplay of how do we understand entropy, as in material entropy in the universe, computational entropy, compression and assembly. That's work in progress, which is really exciting. Building chemical computers is really exciting, and also building kind of chemical brains to see if we can make an artificial consciousness, and that would perhaps kind of think in a different way. And the only other thing I would say is I would like to kind of reinforce the idea that I am very pro AI used correctly. And also I'm very pro getting the labeling right, because the concept of AGR and the Concept of super intelligence are things, I would argue, that are not possible. Right. And we are scaring a lot of people. Like super intelligence is like saying, I'm going to have a super Turing machine or I'm going to have access to magic. And all I would say is when people that are listening to this and maybe watching and thinking about it get sold false kind of arguments about AGI and superintelligence just to think critically about what that means. Of course, I've had robots that are faster than me at chess. We've had them for years. That doesn't make them kind of in a different plane. It doesn't make them magical. They're just fast at doing chess. And so what I'm. What I'd like to do is basically ground things a bit. You know, when Sam Altman says chat GPT is going to open up infinite riches, this is kind of the thing. The only other thing I'd say is you've got all the pdumas saying we're all going to die. And then you've got all the kind of probe ones going with this infinite abundance. Both of those stances are bullshit. They require good human beings to use the tools correctly, to have fun, to help one another and build a better future. And we will. But we're not going to suddenly magically get infinite abundance because someone's using a gen AI to basically tell you, oh, yeah, that looks good, your fusion will definitely work. Now that is the best fusion idea ever. It's going to work. And then you go, I did it. It didn't work. And like, God, do it again. It's really going to work this time, you know, like, sure. So we need to get out of the delusion on both sides. And maybe this is what happens when we discover a new industrial revolution. That's maybe why it's so exciting to be alive in this. Yeah, it's like the answer is in the middle somewhere. So I've stopped being so kind of, you know, like, because. Because I have a lot of friends who use AI and they're like, say, Lee, you're just really too anti AI. And I'm like, I'm not anti AI. I'm anti the people selling it. Miss selling it it. To say it's going to solve everything when basically it's just a new tool
- [79:50]
B
does make you wonder. Like at the printing press where they like, give me all the money, I'll change the way everybody learns. And the other people were like, we'll never be able to read again with the Printing press, like shut it down.
- [79:59]
A
Exactly. Like when people were like writing books, they'd say like you know, the, the, the somehow the soul of the author's in the book. And when people. Maybe this is what Jeff Hinton was confused about. Maybe he doesn't read. And when he saw that there was agency in ChatGPT, what he saw, or Gemini, what he actually saw was the incredible agency and creativity of all the people that created that data. It probably is quite awe inspiring to have access to a model and interrogate it and get all this. You know, you literally got access to the entire corpus of human imagination and thinking up until the point the model was trained. That's a pretty mind blowing thing to have. And maybe it's hard to cope with and maybe if I'm being charitable and maybe a little bit more, more open minded about that stance, maybe that's what he's misreading because it is amazing. Like it's amazing. I can download a model and have it on, you know, Llama Studio or LLM Studio, you know, on an airplane with no Internet connection, I can still get it to do competent things with something that's just like 35 gigabytes. It's been trained on several terabytes that compression is fantastic and that compression comes at a probabilistic compromise. But within that probabilistic compromise there are structures we don't yet understand. Which is why, why AI is fascinating and really interesting to explore, but it's not the answer to consciousness.
- [81:24]
C
Absolutely, absolutely. And so I just want to point out, because my whole goal here is to try to kind of highlight different people from different, with different opinions and find commonalities because I feel like sometimes we feel like we're on opposite ends, but there's a lot of commonalities here. So we recently talked to Dr. Roman Iampolsky who is saying that if we're building narrow systems, even if they're super intelligent, like they're superhuman at chess like you said, or they're superhuman at narrow tasks, that's fine, that's not dangerous. He's worried about the super intelligent general systems and he's saying we can do a lot with these narrow systems. So a lot of what both of you are saying is very similar. We're fine. The only thing is he believes we can build generals. Superintelligence or AGI, you believe we can't. But everything else we're very much sort of aligned on. Right. So this idea that we can do a lot with, with AI as tools and it doesn't have to be super intelligent. It doesn't have to be general intelligence. It can still be a phenomenal tool for research. And so I just kind of wanted to point that out because I think that's important, that there's commonalities there.
- [82:28]
A
I would agree. And I think there's something really interesting I'd like to comment on. And so when Einstein came up with relativity, there was a really good prediction he was able to make. He made the following prediction. Prediction that when we put satellites up in space that the atomic clocks would lose time because of frame dragging. So we protected that ahead of time and we like, oh, it's great. We put satellites up. Oh, look, the time's different. Awesome. Relativity is correct. What's kind of happened is that we've invented a technology that processes information before we understand what intelligence is. So we're seeing frame dragging, right? The equivalent. We don't know what it is. And what I mean is we're seeing phenomena in our AIs before we have a really good understanding of intelligence. This is why the person you just talked about before was saying, oh, there could be general intelligence. General intelligence is not possible without integrated kind of abilities to live and die. It's a unique thing to human, to living systems that have undergone 4.3 billion years of evolution. Now, when we make. Because I don't believe, believe it's impossible to create a generally intelligent system, but I do think it's not going to be possible using conventional technology, using silicon. Right. So that's kind of important. So I'm. But we will get there. And that's when we have to be careful.
- [83:51]
C
That's okay. Very interesting. So you do believe it's possible with some sort of biological approach, even maybe
- [83:58]
A
with us, a new technology. As I said, remember the thing I said earlier is like your brain has more potential connections than there are atoms in the universe. When we achieve an infrastructure that can do that, then I'm going to talk to it because it's going to be kind of cool to talk to. And you know, and, and, and so I don't. I'm not. I'm a materialist. I'm a materialist who doesn't understand what intelligence is yet. And like, you know, Demis and Jeffrey and all these people want to understand it. And the OpenAI guys and the people working in good faith, faith, we want to understand intelligence. That is because intelligence is the, you know, there's these. I want to understand origin of life, right? I want to understand consciousness. I want to understand Kind of language, if you like, an imagination. I want to understand intelligence. There are four big problems. Intelligence, consciousness, kind of abstraction, and life. And I. I'm just starting at the bottom. So for me, I don't even know understand what life is. So people jump to intelligence. I kind of feel like you cheated. We need to get this fit first, I think. And the fact that life is the gateway to intelligence. I think we should be remiss to jump to intelligence because we can emulate it. And so that's my only beef, actually. It's not that I don't think it can be done. I just think that we are anthropomorphizing, not using the right controls, not being honest to one another. And so my job really in this is I'm working on it, doing some stuff. I obviously have vested interests. I have conflict. But by being open about my opinions about what is right and wrong, particularly pdoom and all the other stuff, we might be able to kind of have a really interesting, you know, developments in the technology. And I, I do think there are, you know, the AI bubble is inevitably going to burst because, look, agents don't work. ChatGPT isn't that useful. The day hallucination is a problem if you're doing a real job. Right. I've got a chemistry lab where I'm doing robotics. If my AI hallucinates, my lab is on fire. It only has to burn down once. There must be zero chance of the lab burning down. So we have a lot to do there. So I'm tremendously excited and I think that we kind of need to kind of calm down a bit. It's hard when there's so much money invested and we're wasting so much money on infrastructure that is not integral to intelligence. Matrix multiplication is not the answer to intelligence. I'm 100% sure of that.
- [86:32]
C
Understood. Yeah. And certainly there's a lot of money being used up. One final. Just quick point and maybe we can even cut this out if this doesn't connect. Have you heard of a CTO at Google whose name is Blaise Aguera e Arcas? Oh, okay. You guys have so much overlap, it seems right in your ideas. He credit. I believe he mentioned you in his interview. Anything interesting there?
- [86:56]
A
I think Blaze is. So Blaze works at Google. Right. So he has a. Obviously I think he's a CTO of. In technology and society and so on. So he's playing around very playfully trying to figure out what life is. And, and I think what he's, you know, the ideas is coming out. I agree. A lot of them. I think he's doing something. He's asking a lot of interesting questions. And so I have a great deal of, you know, empathy for what he's trying to do. So. So I think his brain fuck stuff is kind of like, yeah, not really, but that's okay. That's his thing. I mean. But I haven't got to talk to him about it recently. But I think he's asking interesting questions. I think he's declaring victory a bit early, but I think what he's doing is fascinating. And I think the more people we have working together from different angles, overlapping, the more understanding we have on the problem. I think the nice thing is, like, by me going on the record about what I'm trying to do and what I think, and also by reassuring people have very strong, very strong views, very weakly held. When someone like Blaze or Sarah or Susan or whoever came and give me more data, and I'm like, oh, okay, I'll change the experiment. Because at the end of the day, I'm an experimentalist searching for truth, and the truth for me comes out of doing good experiments in the universe theory. So actually, what Blaze is doing is super complementary to what Susan's doing and what I'm doing and Sarah's doing. So there's a huge, interesting ecosystem. And one of the things I really love about Blazer's approach is very playful, very reminiscent of the artificial life community at General, which I like interacting with. There's just this whole ecosystem of crazy robots and things that are going on is a. They're a beautiful set of people to work with and just play around with and generally ask new questions and using technology. So, yeah, I'm aware of his stuff. He's a super cool guy. He's also done some stuff, I think, at Santa Fe, very similar to my friend Michael Lackman, who's very, very, very quiet, who did a lot of the stuff that Blaze has rediscovered or independently discovered later. And so they're on the right track. And it's really interesting to see smart people converging.
- [89:08]
C
Absolutely. There's been such an incredible interview and such an interesting time to be alive. I'm so thankful that we live right now because, yeah, definitely a lot of things happening. And we're so thankful for your time and just. Just how you say that, you have to have your curiosity bigger than your ego. I've heard you mention that. One of the podcasts. That's such a beautiful statement. Like, a lot of us kind of get wrapped up and stuff. Just like, we got to be more curious than we're worried. It's okay to be wrong. Wrong. We got to discover. We got to maybe annoy some people. Like you said, it's totally fine as long as it's towards learning and curiosity and all that stuff.
- [89:40]
A
So I think so. I think if you're willing to be wrong, eventually you're going to be right. That's why I keep telling myself, anyway,
- [89:49]
C
I believe that 100% mentality.
- [89:52]
B
All right. Thank you so much for your time. We appreciate it. That was a great conversation.
- [89:55]
C
Thank you so much.