Lee Cronin: Controversial Nature Paper on Evolution of Life and Universe

00:00:00.000 | Every star in the sky probably has planets, and life is probably emerging on these planets.

00:00:07.760 | But I think the commentarial space associated to these planets is so different. Our causal cones

00:00:13.600 | are never going to overlap, or not easily. And this is the thing that makes me sad about alien

00:00:18.400 | life, why it's why we have to create alien life in the lab as quickly as possible, because I don't

00:00:24.640 | know if we are going to be able to be able to build architectures that will intersect with

00:00:32.800 | alien intelligence architectures. And intersect, you don't mean in time or space? Time and the

00:00:38.800 | ability to communicate. The ability to communicate. Yeah, my biggest fear in a way is that life is

00:00:44.080 | everywhere, but we become infinitely more lonely because of our scaffolding in that commentarial

00:00:49.120 | space. - The following is a conversation with Lee Cronin, his third time on this podcast. He is a

00:00:56.960 | chemist from University of Glasgow, who is one of the most fascinating, brilliant, and fun to talk

00:01:02.640 | to scientists I've ever had the pleasure of getting to know. This is Alex Friedman Podcast. To support

00:01:08.960 | it, please check out our sponsors in the description. And now, dear friends, here's Lee Cronin.

00:01:15.840 | - So your big assembly theory paper was published in Nature. Congratulations.

00:01:20.240 | - Thanks. - It created, I think it's fair

00:01:23.520 | to say, a lot of controversy, but also a lot of interesting discussion. So maybe I can try to

00:01:29.920 | summarize assembly theory, and you tell me if I'm wrong. - Go for it.

00:01:32.880 | - So assembly theory says that if we look at any object in the universe, any object,

00:01:39.440 | that we can quantify how complex it is by trying to find the number of steps it took to create it,

00:01:46.160 | and also we can determine if it was built by a process akin to evolution by looking at how many

00:01:52.880 | copies of the object there are. - Yeah, that's spot on.

00:01:56.080 | - Spot on. I was not expecting that. Okay, so let's go through definitions.

00:02:01.440 | So there's a central equation I'd love to talk about, but definition-wise, what is an object?

00:02:09.280 | - Yeah, an object, so if I'm going to try to be as meticulous as possible,

00:02:16.640 | objects need to be finite and they need to be decomposable into subunits. All human-made

00:02:26.960 | artifacts are objects. Is a planet an object? Probably yes, if you scale out. So an object is

00:02:34.480 | finite and countable and decomposable, I suppose mathematically. But yeah, I still wake up some

00:02:41.760 | days and go, think to myself, what is an object? Because it's a non-trivial question.

00:02:50.400 | - Persists over time. I'm quoting from the paper here. An object that's finite is distinguishable.

00:02:57.760 | So that's a weird adjective, distinguishable. - We've had so many people help offering to

00:03:05.760 | rewrite the paper after it came out, you wouldn't believe it's so funny.

00:03:08.800 | - Persists over time and is breakable such that the set of constraints to construct it

00:03:14.880 | from elementary building blocks is quantifiable. Such that the set of constraints to construct it

00:03:21.280 | from elementary building blocks is quantifiable. - The history is in the objects. It's kind of

00:03:27.680 | cool, right? - So what defines the object is its history

00:03:32.800 | or memory, whichever is the sexier word. - I'm happy with both depending on the day.

00:03:38.160 | - Okay. So the set of steps it took to create the object. So there's a sense

00:03:43.360 | in which every object in the universe has a history. And that is part of the thing that

00:03:52.160 | is used to describe its complexity. How complicated it is. Okay. What is an assembly index?

00:03:58.960 | - So the assembly index, if you're to take the object apart and be super lazy about it or minimal,

00:04:07.920 | it's like you've got a really short term memory. So what you do is you lay all the parts on the

00:04:13.440 | path and you find the minimum number of steps you take on the path to add the parts together

00:04:21.680 | to reproduce the object. And that minimum number is the assembly index. It's a minimum bound.

00:04:27.760 | And it was always my intuition the minimum bound in assembly theory was really important. I only

00:04:31.520 | worked out why a few weeks ago, which was kind of funny because I was just like, no, this is

00:04:34.880 | sacrosanct. I don't know why. It will come to me one day. And then when I was pushed by a bunch of

00:04:38.960 | mathematicians, we came up with the correct physical explanation, which I can get to,

00:04:47.280 | but it's the minimum and it's really important. It's the minimum. And the reason I knew the

00:04:51.840 | minimum was right is because we could measure it. So almost before this paper came out,

00:04:55.760 | we've published papers, explain how you can measure the assembly index of molecules.

00:05:00.480 | - Okay. So that's not so trivial to figure out. So when you look at an object, we can say a

00:05:06.560 | molecule, we can say object more generally to figure out the minimum number of steps it takes

00:05:12.240 | to create that object. That doesn't seem like a trivial thing to do.

00:05:16.720 | - So with molecules, it's not trivial, but it is possible because what you can do,

00:05:23.360 | and because I'm a chemist, so I'm kind of like, I see the lens of the world through just chemistry.

00:05:27.920 | I break the molecule apart and break bonds. And if you break up, if you take a molecule and you

00:05:35.280 | break it all apart, you have a bunch of atoms. And then you can say, okay, I'm going to then

00:05:40.320 | form bond, take the atoms and form bonds and go up the chain of events to make the molecule.

00:05:46.240 | And that's what made me realize, take a toy example, literally toy example,

00:05:49.840 | take a Lego object, which is broken up of Lego blocks. So you could do exactly the same thing.

00:05:55.120 | In this case, the Lego blocks are naturally the smallest. They're the atoms in the actual composite

00:06:02.320 | Lego architecture. But then if you maybe take a couple of blocks and put them together in a

00:06:09.360 | certain way, maybe they're offset in some way, that offset is on the memory. You can use that

00:06:14.800 | offset again with only a penalty of one, and you can then make a square triangle and keep going.

00:06:19.520 | And you remember those motifs on the chain. So you can then leap from the start with all the

00:06:26.640 | Lego blocks or atoms just laid out in front of you and say, right, I'll take you, you, you,

00:06:30.880 | connect and do the least amount of work. So it's really like the smallest steps you can take on

00:06:38.480 | the graph to make the object. And so for molecules, it came relatively intuitively.

00:06:43.040 | And then we started to apply it to language. We've even started to apply it to mathematical

00:06:47.280 | theorems. I'm so well out of my depth, but it looks like you can take minimum set of axioms

00:06:51.920 | and then start to build up mathematical architectures in the same way. And then the

00:06:57.920 | shortest path to get there is something interesting that I don't yet understand.

00:07:01.280 | - So what's the computational complexity of figuring out the shortest path

00:07:05.840 | with molecules, with language, with mathematical theorems? It seems like once you have the fully

00:07:14.080 | constructed Lego castle, or whatever your favorite Lego world is, figuring out how to get there from

00:07:22.640 | the basic building blocks, is that an NP hard problem? It's a hard problem.

00:07:28.480 | - It's a hard problem, but actually, if you look at it, so the best way to look at it,

00:07:32.320 | let's take a molecule. So if the molecule has 13 bonds, first of all, take 13 copies of the

00:07:37.360 | molecule and just cut all the bonds. So take cut 12 bonds, and then you just put them in order.

00:07:41.360 | And then that's how it works. So, and you keep looking for symmetry or copies, so you can then

00:07:49.920 | shorten it as you go down. And that becomes combinatorially quite hard.

00:07:55.040 | For some natural product molecules, it becomes very hard. It's not impossible,

00:08:00.080 | but we're looking at the bounds on that at the moment. But as the object gets bigger,

00:08:04.880 | it becomes really hard. But that's the bad news, but the good news is there are shortcuts.

00:08:10.160 | And we might even be able to physically measure the complexity without computationally calculating

00:08:16.560 | it, which is kind of insane. - Wait, how would you do that?

00:08:19.680 | - Well, in the case of molecule, so if you shine light on a molecule, let's take an infrared,

00:08:25.520 | the molecule has each of the bonds absorbs the infrared differently in what we call the

00:08:31.680 | fingerprint region. And so it's a bit like, and because it's quantized as well, you have all these

00:08:38.080 | discrete kind of absorbances. And my intuition after we realized we could cut molecules up in

00:08:44.080 | mass spec, that was the first go at this. We did it with using infrared and the infrared gave us an

00:08:49.440 | even better correlation assembly index. And we used another technique as well, in addition to

00:08:54.720 | infrared called NMR, nuclear magnetic resonance, which tells you about the number of different

00:08:58.960 | magnetic environments in a molecule. And that also worked out. So we have three techniques,

00:09:04.000 | which each of them independently gives us the same or tending towards the same assembly index

00:09:10.000 | for molecule that we can calculate mathematically. - Okay, so these are all methods of mass

00:09:15.120 | spectrometry, mass spec, you scan a molecule, it gives you data in the form of a mass spectrum,

00:09:21.840 | and you're saying that the data correlates to the assembly index.

00:09:25.200 | - Yeah. - How generalizable is that

00:09:27.760 | shortcut, first of all, to chemistry, and second of all, beyond that? 'Cause that seems like a

00:09:32.800 | nice hack and you're extremely knowledgeable about various aspects of chemistry, so you can say,

00:09:39.840 | "Okay, it kind of correlates." But the whole idea behind assembly theory paper, and perhaps why it's

00:09:47.280 | so controversial, is that it reaches bigger. It reaches for the bigger general theory of objects

00:09:57.040 | in the universe. - Yeah, I'd say so, I'd agree.

00:09:59.360 | So I've started assembly theory of emoticons with my lab, believe it or not. So we take emojis,

00:10:06.240 | pixelate them, and work out the assembly index of the emoji, and then work out how many emojis

00:10:11.600 | you can make on the path of emojis. So there's the uber emoji from which all other emojis emerge,

00:10:18.080 | and then you can, so you can then take a photograph, and by looking at the shortest path,

00:10:22.480 | by reproducing the pixels to make the image you want, you can measure that. So then you start to

00:10:29.600 | be able to take spatial data. Now there's some problems there. What is then the definition of

00:10:35.120 | the object? How many pixels? How do you break it down? And so we're just learning all this right

00:10:41.360 | now. - So how do you compute the, how would you begin to compute the assembly index of a graphical,

00:10:50.320 | like, a set of pixels on a 2D plane that form a thing? - So you would, first of all, determine the

00:10:57.200 | resolution. So then what is your x, y, and what number on the x and y plane? And then look at the

00:11:03.440 | surface area, and then you take all your emojis and make sure they're all looked at the same

00:11:07.280 | resolution. - Yes. - And then we would basically then do the exactly the same thing we would do for

00:11:13.840 | cutting the bonds. You'd cut bits out of the emoji, and look at the, you'd have a bag of pixels,

00:11:20.320 | so, and you would then add those pixels together to make the overall emoji. - Wait a minute,

00:11:26.800 | but, like, first of all, not every pixel's, I mean, this is at the core, sort of, machine learning,

00:11:33.760 | computer vision, not every pixel is that important, and there's, like, macro features, there's micro

00:11:38.800 | features, and all that kind of stuff. - Exactly. - Like, you know, the eyes appear in a lot of them,

00:11:43.920 | the smile appears in a lot of them. - So in the same way in chemistry we assume the bond is

00:11:49.680 | fundamental, what we do in there here is we assume the resolution at the scale at which we do it is

00:11:54.800 | fundamental, and we're just working that out, and that, you're right, that will change, right? Because

00:11:58.400 | as you take your lens out a bit, you, it will change dramatically, but it, but it's just a new way of

00:12:04.480 | looking at not just compression, what we do right now in computer science and data, one big, kind of,

00:12:10.400 | kind of, misunderstanding is assembly theory is telling you about how compressed the object is,

00:12:18.560 | that's not right, it's a how much information is required on a chain of events, because the nice

00:12:26.240 | thing is if, when you do compression in computer science, we're wandering a bit here, but it's,

00:12:30.640 | kind of, worth wandering, I think, and you, you assume you have instantaneous access to all the

00:12:36.000 | information in the memory. - Yeah. - In assembly theory, you say, no, you don't get access to that

00:12:40.000 | memory until you've done the work, and then when you've done access to that memory, you can have

00:12:43.360 | access, but not to the next one, and this is how, in assembly theory, we talk about the four

00:12:48.000 | universes, the assembly universe, the assembly possible, and the assembly contingent, and then

00:12:53.360 | the assembly observed, and they're all, all scales in this combinatorial universe. - Yeah, can you

00:12:59.440 | explain each one of them? - Yep, so the assembly universe is like anything goes, just, it's just

00:13:04.640 | combinatorial, kind of, explosion in everything. - So that's the biggest one? - That's the biggest

00:13:08.640 | one, it's massive. - Assembly universe, assembly possible, assembly contingent, assembly observed,

00:13:16.160 | and on the y-axis is assembly steps in time, and, you know, in the x-axis, as the thing

00:13:24.400 | expands through time, more and more unique objects appear. - So, yeah, so assembly universe, everything

00:13:31.600 | goes. - Yep. - Assembly possible, laws of physics come in, in this case, in chemistry bonds. In

00:13:37.360 | assembly, so that means... - Those are actually constraints, I guess? - Yes, and they're the only

00:13:41.680 | constraints, they're the constraints of the base, so the way to look at it, you've got all your atoms,

00:13:45.440 | they're quantized, you can just bung them together, so then you can become a, kind of, so, in the way,

00:13:50.160 | in computer science speak, I suppose, the assembly universe is just, like, no laws of physics,

00:13:54.240 | things can fly through mountains, beyond the speed of light. In the assembly possible, you have to

00:14:00.880 | apply the laws of physics, but you can get access to all the motifs instantaneously, with no effort,

00:14:07.680 | so that means you could make anything. Then the assembly contingent says, no, you can't have access

00:14:13.680 | to the highly assembled object in the future, until you've done the work in the past, on the

00:14:17.360 | causal chain, and that's really, the really interesting shift, where you go from assembly

00:14:23.520 | possible, to assembly contingent. That is really the key thing in assembly theory, that says you

00:14:30.320 | cannot just have instantaneous access to all those memories, you have to have done the work somehow,

00:14:36.160 | the universe has to have somehow built a system that allows you to select that path,

00:14:43.200 | rather than other paths, and then the final thing, the assembly observed, is basically us saying, oh,

00:14:50.080 | these are the things we actually see, we can go backwards now, and understand that they have been

00:14:56.640 | created by this causal process. - Wait a minute, so when you say the universe has to construct the

00:15:01.840 | system, that does the work, is that like the environment that allows for like selection?

00:15:08.240 | - Yeah, yeah, yeah. - That's the thing that does the selection?

00:15:10.560 | - You could think about in terms of a von Neumann constructor, first it's a selection, a ribosome,

00:15:14.960 | Tesla plant, assembling Teslas, you know, the difference between the assembly universe,

00:15:22.080 | in Tesla land, and the Tesla factory is, everyone says, no, Teslas are just easy, they just spring

00:15:27.680 | out, you know how to make them all, in a Tesla factory, you have to put things in sequence,

00:15:31.360 | and out comes a Tesla. - So you're talking about the factory?

00:15:33.600 | - Yes, this is really nice, super important point, is that when I talk about the universe having a

00:15:38.880 | memory, or there's some magic, it's not that, it's that tells you that there must be a process

00:15:45.520 | encoded somewhere in physical reality, be it a cell, a Tesla factory, or something else that

00:15:52.400 | is making that object, I'm not saying there's some kind of woo-woo memory in the universe,

00:15:57.440 | you know, morphic resonance or something, I'm saying that there is an actual causal process

00:16:04.320 | that is being directed, constrained in some way, so it's not kind of just making everything.

00:16:09.840 | - Yeah, but Lee, what's the factory that made the factory?

00:16:13.840 | So what is the, so first of all, you assume the laws of physics has just sprung to existence at

00:16:23.520 | the beginning, those are constraints, but what makes the factory the environment that does the

00:16:28.480 | selection? - This is the question of, well, it's the first interesting question that I want to

00:16:33.760 | answer, out of four, I think the factory emerges in the interplay between the environment and the

00:16:41.440 | objects that are being built, and here, let me, I'll have a go at explaining to you the shortest

00:16:47.440 | path. So why is the shortest path important? Imagine you've got, I'm gonna have to go chemistry

00:16:53.360 | for a moment and then abstract it, so imagine you've got a given environment that you have a

00:17:01.040 | budget of atoms, you're just flinging together, and the objective of those atoms that are being

00:17:07.280 | flung together in, say, molecule A, have to make, they decompose, so molecules decompose over time,

00:17:14.800 | so the molecules in this environment, in this magic environment, have to not die,

00:17:20.240 | but they do die, there's a, they have a half-life, so the only way the molecules can get through

00:17:25.520 | that environment out the other side, let's pretend the environment is a box, you can go in and out

00:17:29.040 | without dying, and there's a, there's just an infinite supply of atoms coming, or a, well,

00:17:34.400 | a large supply, the molecule gets built, but the molecule that is able to template itself being

00:17:42.560 | built, and survives in the environment, will basically reign supreme. Now, let's say that

00:17:50.960 | molecule takes 10 steps, now, and it's using a finite set of atoms, right, or now let's say

00:17:58.320 | another molecule, smart-ass molecule we'll call it, comes in, and can survive in that environment,

00:18:03.680 | and can copy itself, but it only needs five steps. The molecule that only needs five steps,

00:18:11.040 | because it's continue, both molecules are being destroyed, but they're creating themselves

00:18:14.640 | faster, they can be destroyed, you can see that the shortest path reigns supreme. So,

00:18:20.800 | the shortest path tells us something super interesting about the minimal amount of

00:18:24.640 | information required to propagate that motif in time and space, and it's just like a kind of,

00:18:32.480 | it seems to be like some kind of conservation law. - So, one of the intuitions you have

00:18:38.080 | is the propagation of motifs in time will be done by the things that can construct themselves

00:18:43.760 | in the shortest path. - Yeah. - So, like, you can assume that most of the objects in the universe

00:18:50.960 | are built in the shortest, in the most efficient way. - The, so... - Big loop I just took there.

00:18:58.480 | - Yeah, no, yes and no, because there are other things. So, in the limit, yes, because you want

00:19:04.080 | to tell the difference between things that have required a factory to build them and just random

00:19:09.200 | processes, but you can find instances where the shortest path isn't taken for an individual object,

00:19:18.000 | individual function, and people go, "Ah, that means the shortest path isn't right." And then I say,

00:19:24.720 | "Well, I don't know, I think it's right still." Because, so, of course, because there are other

00:19:30.960 | driving forces, it's not just one molecule. Now, when you start to, now you start to consider two

00:19:35.680 | objects, you have a joint assembly space, and it's not, now it's a compromise between not just making

00:19:41.440 | A and B in the shortest path, you want to make A and B in the shortest path, which might mean that

00:19:46.800 | A is slightly longer, you have a compromise. So, when you see slightly more nesting in the

00:19:52.880 | construction, when you take a given object, that can look longer, but that's because the overall

00:19:58.320 | function is the object is still trying to be efficient. And this is still very hand-wavy,

00:20:04.560 | and maybe having no legs to stand on, but we think we're getting somewhere with that.

00:20:09.520 | - And there's probably some parallelization, right? So, this is all, this is not sequential,

00:20:14.480 | the building is, I guess, when you're talking about complex objects, you don't have to

00:20:21.120 | work sequentially, you can work in parallel, you can get your friends together, and they can...

00:20:24.960 | - Yeah, and the thing we're working on right now is how to understand these parallel processes.

00:20:31.440 | Now there's a new thing we've introduced called assembly depth, and assembly depth can be

00:20:38.160 | lower than the assembly index for a molecule when they're cooperating together, because exactly this

00:20:44.800 | parallel processing is going on. And my team have been working this out in the last few weeks,

00:20:50.000 | because we're looking at what compromises does nature need to make when it's making molecules

00:20:54.720 | in the cell. And I wonder if, you know, I'm maybe like, well, I'm always leaping out of

00:21:01.360 | my competence, but in economics, I'm just wondering if you could apply this in economic

00:21:06.960 | processes. It seems like capitalism is very good at finding shortest path, you know, every time,

00:21:11.600 | but there are ludicrous things that happen because actually the cost function has been minimized.

00:21:15.920 | And so I keep seeing parallels everywhere where there are complex nested systems, where if you

00:21:21.280 | give it enough time, and you introduce a bit of heterogeneity, the system readjusts and finds a

00:21:26.640 | new shortest path. But the shortest path isn't fixed on just one molecule now, it's in the actual

00:21:31.920 | existence of the object over time. And that object could be a city, it could be a cell,

00:21:37.040 | it could be a factory, but I think we're going way beyond molecules, and my competence probably

00:21:42.400 | should go back to molecules, but hey. - All right, before we get too far,

00:21:46.320 | let's talk about the assembly equation. Okay, how should we do this? Now, let me just even read that

00:21:52.800 | part of the paper. We define assembly as the total amount of selection necessary to produce an

00:21:59.920 | ensemble of observed objects, quantified using equation one. The equation basically has A on

00:22:08.080 | one side, which is the assembly of the ensemble, and then a sum from one to N, where N is the total

00:22:17.920 | number of unique objects, and then there is a few variables in there that include the assembly index,

00:22:24.640 | the copy number, which we'll talk about. That's an interesting, I don't remember you talking about

00:22:30.160 | that. That's an interesting addition, and I think a powerful one. It has to do with what? That you

00:22:36.800 | can create pretty complex objects randomly, and in order to know that they're not random,

00:22:42.400 | that there's a factory involved, you need to see a bunch of them. That's the intuition there. It's

00:22:47.760 | an interesting intuition, and then some normalization. What else is it? - N minus one,

00:22:55.120 | just to make sure that more than one object, one object could be a one-off and random,

00:22:59.360 | and then you have more than one identical object. That's interesting. - When there's two of a thing.

00:23:05.040 | - Two of a thing is super important, especially if the assembly index is high.

00:23:09.280 | - We could say several questions here. One, let's talk about selection.

00:23:14.000 | What is this term selection? What is this term evolution that we're referring to? Which aspect

00:23:19.840 | of Darwinian evolution are we referring to that's interesting here? - This is probably what the

00:23:28.400 | paper, we should talk about the paper for a second. What it did is it kind of annoyed,

00:23:32.720 | we didn't know it. I mean, it got attention, and obviously the angry people were annoyed.

00:23:39.200 | - There's angry people in the world. That's good. - So what happened is the evolutionary

00:23:42.640 | biologists got angry. We were not expecting that, because we thought evolutionary biologists would

00:23:46.560 | be cool. I knew that some, not many, computational complexity people would get angry, because I'd

00:23:52.880 | kind of been poking them, and maybe I deserved it, but I was trying to poke them in a productive way,

00:23:59.040 | and then the physicists kind of got grumpy, because the initial conditions tell everything.

00:24:03.680 | The prebiotic chemists got slightly grumpy, because there's not enough chemistry in there,

00:24:08.000 | and then finally, when the creationists said it wasn't creationist enough, I was like,

00:24:10.960 | I've done my job. - Well, you say in the physics,

00:24:14.000 | they say, because you're basically saying that physics is not enough to tell the story of how

00:24:20.640 | biology emerges. - I think so.

00:24:22.560 | - And then they said, a few physics is the beginning and the end of the story.

00:24:27.920 | - Yeah. So what happened is the reason why people put the phone down on the call,

00:24:33.040 | the paper, if you're reading the paper like a phone call, they got to the abstract,

00:24:37.920 | and in the abstract-- - The first sentence is pretty--

00:24:40.560 | - The first two sentences caused everybody-- - Scientists have grappled with reconciling

00:24:45.200 | biological evolution with the immutable laws of the universe defined by physics.

00:24:50.720 | - True, right? There's nothing wrong with that statement, totally true.

00:24:54.080 | - Yeah. "These laws underpin life's origin, evolution and the development of human culture

00:25:01.040 | and technology, yet they do not predict the emergence of these phenomena." Wow.

00:25:06.880 | First of all, we should say the title of the paper, this paper was accepted and published in Nature.

00:25:12.320 | The title is "Assembly Theory Explains and Quantifies Selection and Evolution." Very

00:25:16.880 | humble title. And the entirety of the paper, I think, presents interesting ideas but reaches high.

00:25:25.120 | - I am not, I would do it all again. This paper was actually on the pre-print server

00:25:32.560 | for over a year. - You regret nothing.

00:25:34.400 | - Yeah, I think, yeah, I don't regret anything. - You and Frank Sinatra did it your way.

00:25:38.800 | - What I love about being a scientist is kind of sometimes, because I'm a bit dim, I'm like,

00:25:44.400 | and I don't understand what people are telling me, I want to get to the point.

00:25:46.880 | This paper says, "Hey, laws of physics are really cool, the universe is great,

00:25:52.160 | but they don't really, it's not intuitive that you just run the standard model and get life out."

00:26:00.160 | I think most physicists might go, "Yeah, there's, you know, it's not just, we can't just go back and

00:26:05.600 | say that's what happened because physics can't explain the origin of life yet." It doesn't mean

00:26:10.560 | it won't or can't, okay, just to be clear. Sorry, intelligent designers, we are going to get there.

00:26:15.920 | Second point, we say that evolution works but we don't know how evolution got going,

00:26:22.720 | so biological evolution and biological selection. So for me, this seems like a simple continuum.

00:26:28.320 | So when I mentioned selection and evolution in the title, I think, and in the abstract,

00:26:33.120 | we should have maybe prefaced that and said non-biological selection and non-biological

00:26:37.840 | evolution. And then that might have made it even more crystal clear, but I didn't think that biology,

00:26:43.280 | evolutionary biology, should be so bold to claim ownership of selection and evolution.

00:26:49.040 | And secondly, a lot of evolutionary biologists seem to dismiss the origin of life question,

00:26:53.200 | just say it's obvious. And that causes a real problem scientifically because when two different,

00:26:58.800 | when the physicists are like, "We own the universe, the universe is good, we explain all of

00:27:03.280 | it, look at us." And even biologists say, "We can explain biology." And the poor chemist in the

00:27:08.560 | middle going, "But hang on." And this paper kind of says, "Hey, there is an interesting

00:27:16.000 | disconnect between physics and biology, and that's at the point at which memories get made

00:27:24.000 | in chemistry through bonds. And hey, let's look at this closely and see if we can quantify it."

00:27:28.960 | So yeah, I mean, I never expected the paper to kind of get that much interest. And still,

00:27:34.960 | I mean, it's only been published just over a month ago now.

00:27:37.280 | - So just to linger on the selection, what is the broader sense of what selection means?

00:27:46.240 | - Yeah, that's a really, good. For selection, selection, so I think for selection you need,

00:27:53.280 | so this is where for me, the concept of an object is something that can persist in time and

00:27:58.720 | not die, but basically can be broken up. So if I was going to kind of bolster the

00:28:04.800 | definition of an object, so if something can form and persist for a long period of time

00:28:14.720 | under an existing environment that could destroy other, and I'm going to use anthropomorphic terms,

00:28:22.960 | I apologise, but weaker objects or less robust, then the environment could have selected that.

00:28:30.560 | So good chemistry examples, if you took some carbon and you made a chain of carbon atoms,

00:28:36.160 | whereas if you took some, I don't know, some carbon, nitrogen and oxygen and made chains from

00:28:42.400 | those, you'd start to get different reactions and rearrangements. So a chain of carbon atoms

00:28:48.560 | might be more resistant to falling apart under acidic or basic conditions

00:28:55.040 | versus another set of molecules. So it survives in that environment, so the acid pond,

00:28:59.840 | the resistant molecule can get through, and then that molecule goes into another environment,

00:29:08.400 | so that environment now maybe being an acid pond is a basic pond, or maybe it's an oxidising pond,

00:29:14.640 | and so if you've got carbon and it goes in an oxidising pond, maybe the carbon starts to

00:29:18.720 | oxidise and break apart. So you go through all these kind of obstacle courses, if you like,

00:29:24.080 | given by reality. So selection is the ability that happens when an object survives in an

00:29:30.240 | environment for some time, but, and this is the thing that's super subtle, the object has to be

00:29:39.520 | continually being destroyed and made by process. So it's not just about the object now, it's about

00:29:44.720 | the process and time that makes it, because a rock could just stand on the mountainside for

00:29:50.160 | four billion years and nothing happened to it, and that's not necessarily really advanced selection.

00:29:55.600 | So for selection to get really interesting, you need to have a turnover in time,

00:29:59.520 | you need to be continually creating objects, producing them, what we call discovery time.

00:30:05.040 | So there's a discovery time for an object, when that object is discovered, if it's say a molecule

00:30:10.480 | that can then act on itself, or the chain of events that caused itself to bolster its formation,

00:30:16.080 | then you go from discovery time to production time, and suddenly you have more of it in the

00:30:21.280 | universe. So it could be a self-replicating molecule, and the interaction of the molecule

00:30:26.080 | in the environment, in the warm little pond, or in the sea, or wherever, in the bubble,

00:30:30.640 | could then start to build a proto-factory, the environment. So really, to answer your question,

00:30:36.160 | what the factory is, the factory is the environment, but it's not very autonomous,

00:30:41.600 | it's not very redundant, there's lots of things that could go wrong.

00:30:44.640 | So once you get high enough up the hierarchy of networks of interactions,

00:30:51.040 | something needs to happen, that needs to be compressed into a smaller volume and made

00:30:55.200 | resistant and robust. Because in biology, selection and evolution is robust. You have

00:31:01.200 | error correction built in, you have really, you know, there's good ways of basically making sure

00:31:06.240 | propagation goes on. So really the difference between inorganic, abiotic selection and evolution,

00:31:12.880 | and evolution and stuff in biology is robustness. The ability to kind of propagate over,

00:31:22.400 | the ability to survive in lots of different environments. Whereas our poor little inorganic

00:31:28.480 | soul molecule, whatever, just dies in lots of different environments. So there's something

00:31:34.400 | super special that happens from the inorganic molecule in the environment that kills it,

00:31:40.560 | to where you've got evolution and cells can survive everywhere.

00:31:43.280 | - How special is that? How do you know those kinds of evolution factors are everywhere in

00:31:49.440 | the universe? - I don't, and I'm excited because I think

00:31:54.800 | selection isn't special at all. I think what is special is the history of the environments on

00:32:02.480 | Earth that gave rise to the first cell, that now has, you know, has taken all those environments

00:32:09.360 | and is now more autonomous. And I would like to think that, you know, this paper

00:32:17.120 | could be very wrong, but I don't think it's very wrong. It's certainly wrong, but it's less wrong

00:32:24.400 | than some other ideas, I know, right? And if this allows, inspires us to go and look for selection

00:32:29.520 | in the universe, because we now have an equation where we can say, we can look for selection going

00:32:33.920 | on and say, "Oh, that's interesting. We seem to have a process that's giving us high copy number

00:32:41.920 | objects that also are highly complex, but that doesn't look like life as we know it." And we

00:32:47.200 | use that and say, "Oh, there's a hydrothermal vent. Oh, there's a process going on. There's

00:32:50.640 | molecular networks," because the assembly equation is not only meant to identify at the higher end

00:32:57.200 | advanced selection, what you get, I would call it in biology, you super advanced selection.

00:33:03.120 | And even, I mean, you could use the assembly equation to look for technology and God forbid,

00:33:08.880 | we could talk about consciousness and abstraction, but let's keep it primitive,

00:33:12.720 | molecules and biology. So I think the real power of the assembly equation is to say how much

00:33:17.920 | selection is going on in this space. And there's a really simple thought experiment I could do.

00:33:23.520 | If you have a little Petri dish, and on that Petri dish, you put some simple food. So the

00:33:28.320 | assembly index of all the sugars and everything is quite low. And you put a single cell of E. coli

00:33:35.120 | cell. And then you say, "I'm going to measure the assembly in this amount of assembly in the box."

00:33:40.960 | So it's quite low, but the rate of change of assembly dA/dt will go voom sigmoidal as it

00:33:48.080 | eats all the food. And the number of E. coli cells will replicate because they take all the food,

00:33:54.320 | they can copy themselves. The assembly index of all the molecules goes up, up, up and up

00:33:58.160 | until the food is exhausted in the box. So now the E. coli stop... I mean,

00:34:04.640 | dying is probably a strong word. They stop respiring because all the food is gone. But

00:34:09.120 | suddenly the amount of assembly in the box has gone up gigantically because of that one E. coli

00:34:14.240 | factory has just eaten through, milled lots of other E. coli factories, run out of food and

00:34:19.280 | stopped. And so in the initial box, although the amount of assembly was really small,

00:34:26.400 | it was able to replicate and use all the food and go up. And that's what we're trying to do

00:34:30.960 | in the lab actually, is kind of make those kinds of experiments and see if we can spot

00:34:36.080 | the emergence of molecular networks that are producing complexity as we feed in raw materials

00:34:42.400 | and we feed a challenge, an environment. We try and kill the molecules. And really,

00:34:48.480 | that's the main kind of idea for the entire paper. - Yeah, and see if you can measure the changes in

00:34:54.960 | the assembly index throughout the whole system. Okay, what about if I show up to a new planet,

00:35:00.000 | we'll go to Mars or some other planet from a different solar system,

00:35:03.200 | and how do we use assembly index there to discover alien life? - Very simply, actually. Let's say

00:35:13.120 | we'll go to Mars with a mass spectrometer with a sufficiently high resolution. So what you have to

00:35:17.760 | be able to do... So a good thing about mass spec is that you can select a molecule from the mass,

00:35:26.640 | and then if it's high enough resolution, you can be more and more sure that you're just

00:35:30.560 | seeing identical copies. You can count them. And then you fragment them and you count the number

00:35:35.760 | of fragments and look at the molecular weight. And the higher the molecular weight and the higher

00:35:40.640 | the number of the fragments, the higher the assembly index. So if you go to Mars and you

00:35:44.400 | take a mass spec or high enough resolution, and you can find molecules... I'll give a guide on

00:35:50.240 | Earth. If you could find molecules, say, greater than 350 molecular weight with more than 15

00:35:56.160 | fragments, you have found artefacts that can only be produced, at least on Earth, by life.

00:36:03.760 | Now you would say, "Oh, maybe the geological process." I would argue very vehemently that

00:36:09.360 | that is not the case. But we can say, "Look, if you don't like the cutoff on Earth,

00:36:14.000 | go up higher, 30, 100," right? Because there's going to be a point where you'll find a molecule

00:36:19.440 | with so many different parts, the chances of you getting a molecule that has 100 different parts

00:36:25.600 | and finding a million identical copies, you know, that's just impossible. That could never happen

00:36:33.440 | in an infinite set of universes. - Can you just linger on this copy

00:36:38.480 | number thing? A million different copies. What do you mean by copies and why is the

00:36:46.560 | number of copies important? - Yeah, that was so interesting. And I

00:36:52.320 | always understood the copy number is really important, but I never explained it properly

00:36:57.440 | for ages. And I kept having this... It goes back to this, if I give you a, I don't know,

00:37:07.040 | a really complicated molecule and I say, "It's complicated," you could say, "Hey,

00:37:09.920 | that's really complicated, but is it just really random?" And so I realized that ultimate randomness

00:37:16.080 | and ultimate complexity are indistinguishable until you can see a structure in the randomness.

00:37:24.640 | So you can see copies. - So copies implies structure.

00:37:30.320 | - Yeah, the factory. - I mean, there's a deep,

00:37:34.240 | profound thing in there. 'Cause if you just have a random process, you're going to get a lot of

00:37:43.600 | complex, beautiful, sophisticated things. What makes them complex in the way we think life is

00:37:51.760 | complex or, yeah, something like a factory that's operating under a selection process is there should

00:37:59.600 | be copies. Is there some looseness about copies? What does it mean for two objects to be equal?

00:38:06.080 | - It's all to do with the telescope or the microscope you're using. And so at the maximum

00:38:12.960 | resolution... So the nice thing about chemists is they have this concept of the molecule and

00:38:18.400 | they're all familiar with the molecule. And molecules you can hold in your hand, lots of them,

00:38:24.240 | identical copies. A molecule is actually a super important thing in chemistry to say, "Look,

00:38:28.800 | you can have a mole of a molecule, an Avogadro's number of molecules,

00:38:32.480 | and they're identical." What does that mean? That means that the molecular composition,

00:38:36.400 | the bonding and so on, the configuration is indistinguishable. You can hold them together,

00:38:41.840 | you can overlay them. So the way I do it is if I say, "Here's a bag of 10 identical molecules,

00:38:48.080 | let's prove they're identical." You pick one out of the bag and you basically observe it using some

00:38:53.760 | technique and then you take it away and then you take another one out. If you observe it using

00:38:58.000 | technique, you can see no differences, they're identical. It's really interesting to get right,

00:39:01.760 | because if you take, say, two molecules, molecules can be in different vibrational and rotational

00:39:07.680 | states, they're moving all the time. So with this respect, identical molecules have identical

00:39:11.920 | bonding. In this case, we don't even talk about chirality because we don't have a chirality

00:39:17.360 | detector. So two identical molecules in one conception, assembly theory, basically considers

00:39:24.000 | both hands as being the same. But of course, they're not, they're different. As soon as you

00:39:29.760 | have a chiral distinguisher to detect the left and the right hand, they become different.

00:39:34.800 | And so it's to do with the detection system that you have and the resolution.

00:39:39.360 | - So I wonder if there's an art and science to which detection system is used when you

00:39:46.880 | show up to a new planet. - Yeah, yeah, yeah.

00:39:49.120 | - So you're talking about chemistry a lot today. We have kind of standardized detection systems,

00:39:56.000 | right, of how to compare molecules. So when you start to talk about emojis and language and

00:40:03.520 | mathematical theorems and, I don't know, more sophisticated things at a different scale,

00:40:10.880 | at a smaller scale than molecules, at a larger scale than molecules,

00:40:15.360 | like what detection, like if we look at the difference between you and me, Lex and Lee,

00:40:21.120 | are we the same? Are we different? - Sure. I mean, of course we're different

00:40:25.680 | close up, but if you zoom out a little bit, we'll morphologically look the same.

00:40:29.760 | You know, height and characteristics, hair length, stuff like that.

00:40:35.040 | - Well, also like the species. - Yeah, yeah, yeah.

00:40:37.840 | - And also there's a sense why we're both from Earth.

00:40:42.480 | - Yeah, I agree. I mean, this is the power of assembly theory in that regard.

00:40:45.840 | So the way to look at it, if you have a box of objects, if they're all indistinguishable

00:40:55.440 | then using your technique, what you then do is you then look at the assembly index.

00:41:03.040 | Now, if the assembly index of them is really low, right, and they're all indistinguishable,

00:41:07.920 | then it's telling you that you have to go to another resolution.

00:41:11.040 | So that would be, you know, it's kind of a sliding scale. It's kind of nice.

00:41:14.560 | - Got it. So those two kind of are intentional with each other.

00:41:18.080 | - Yeah. - The number of copies

00:41:19.680 | in the assembly index. - Yeah.

00:41:20.960 | - That's really, really interesting. So, okay, so you show up to a new planet,

00:41:26.880 | you'll be doing what? - I would do mass spec.

00:41:29.680 | - On a sample of what? Like, first of all, like how big of a scoop do you take?

00:41:33.920 | Do you just take the scoop? Like what? So we're looking for primitive life.

00:41:40.800 | - I would look, yeah, so if you're just going to Mars or Titan or Enceladus or somewhere,

00:41:48.720 | so a number of ways of doing it. So you could take a large scoop or you go for the atmosphere

00:41:52.560 | and detect stuff. So you could make a life meter, right? So one of Sarah's colleagues at ASU,

00:42:03.200 | Paul Davis, keeps calling it a life meter, which is quite a nice idea because you think about it,

00:42:08.800 | if you've got a living system that's producing these highly complex molecules and they drift

00:42:14.960 | away and they're in a highly kind of demanding environment, they could be burnt, right? So they

00:42:21.600 | could just be falling apart. So you want to sniff a little bit of complexity and say warmer, warmer,

00:42:25.920 | warmer. Oh, we found life. We found the alien. We found the alien Elon Musk smoking a joint

00:42:31.280 | in the bottom of the cave on Mars or Elon himself, whatever, right? And you say, okay, found it.

00:42:35.840 | So what you can do is a mass spectrometer, you could just look for things in the gas phase,

00:42:41.120 | or you go on the surface, drill down because you want to find molecules that are,

00:42:45.680 | you've either got to find the source living system because the problem with just looking

00:42:52.000 | for complexity is it gets burnt away. So in a harsh environment on say on the surface of Mars,

00:42:59.440 | there's a very low probability that you're going to find really complex molecules because

00:43:03.760 | of all the radiation and so on. If you drill down a little bit, you could drill down a bit

00:43:08.160 | into soil that's billions of years old. Then I would put in some solvent, water, alcohol or

00:43:15.200 | something, or take a scoop, make it volatile, put it into the mass spectrometer and just try and

00:43:22.480 | detect high complexity, high abundant molecules. And if you get them, hey, presto, you can have

00:43:27.920 | evidence of life. Wouldn't that then be great if you could say, okay, we've found evidence of life.

00:43:33.520 | Now we want to keep the life meter, keep searching for more and more complexity

00:43:39.200 | until you actually find living cells. You can get those new living cells and then you could bring

00:43:44.400 | them back to earth, or you could try and sequence them. You could see that they have different DNA

00:43:47.840 | and proteins. - Go along the gradient of the

00:43:49.760 | life meter. How would you build a life meter? Let's say we're together starting a new company

00:43:55.840 | launching a life meter. - Mass spectrometer would be the

00:43:58.000 | first way of doing it. - No, no, no. That's one of the major

00:44:01.760 | components of it. I'm talking about if it's a device, we've got a branding logo, we're going

00:44:07.200 | to talk about that later. What's the input? How do you get a metered output?

00:44:15.120 | - I would take my life meter, our life meter, there you go.

00:44:21.120 | - Thank you. - Yeah, you're welcome.

00:44:24.960 | It would have both infrared and mass spec. It would have two ports so it could shine a light.

00:44:30.320 | What it would do is you would have a vacuum chamber and you would have an electrostatic

00:44:37.040 | analyzer and you'd have a monochromator to producing infrared. You'd take a scoop of

00:44:44.000 | the sample, put it in the life meter. It would then add a solvent or heat up the sample so some

00:44:49.840 | volatiles come off. The volatiles would then be put into the mass spectrometer, into electrostatic

00:44:55.600 | trap and you'd weigh the molecules and fragment them. Alternatively, you'd shine infrared light

00:45:00.320 | on them, you'd count the number of bands, but you'd have to, in that case, do some separation

00:45:04.480 | because you want to separate. In mass spec, it's really nice and convenient because you can

00:45:08.400 | separate electrostatically, but you need to have that.

00:45:11.760 | - Can you do it in real time? - Yeah, pretty much. Let's go all the way

00:45:15.520 | back. Okay, we're really going to get this. Lex's life meter, Lex and Lee's life meter.

00:45:21.520 | - Excellent. It's a good ring to it. - All right. You have a vacuum chamber,

00:45:28.800 | you have a little nose. The nose would have a packing material. You would take your sample,

00:45:37.760 | add it onto the nose, add a solvent or a gas. It would then be sucked up the nose and that would

00:45:42.960 | be separated using what we call chromatography. Then as each band comes off the nose, we would

00:45:48.080 | then do mass spec and infrared. In the case of the infrared, count the number of bands. In the case

00:45:54.240 | of mass spec, count the number of fragments and weigh it. Then the further up in molecular weight

00:45:58.960 | range for the mass spec and the number of bands, you go up and up and up from the dead,

00:46:02.960 | interesting, interesting, over the threshold, oh my gosh, earth life. Then right up to the

00:46:09.280 | batshit crazy, this is definitely alien intelligence that's made this life. You

00:46:14.400 | could almost go all the way there, same in the infrared. It's pretty simple. The thing that

00:46:19.360 | is really problematical is that for many years, decades, what people have done, and I can't blame

00:46:27.440 | them, is they've been obsessing about small biomarkers that we find on earth, amino acids,

00:46:34.000 | like single amino acids or evidence of small molecules and these things. Looking for those,

00:46:38.880 | looking for complexity. The beautiful thing about this is you can look for

00:46:43.680 | complexity without earth chemistry bias or earth biology bias. Assembly theory is just a way of

00:46:52.720 | saying, hey, complexity and abundance is evidence of selection. That's how our universal life meter

00:46:57.920 | will work. - Complexity in abundance is evidence of selection. Okay, so let's apply our life meter

00:47:08.720 | to earth. So if we were just to apply assembly index measurements to earth, what kind of stuff

00:47:21.360 | are going to get? What's impressive about some of the complexity on earth? - So we did this a few

00:47:29.120 | years ago when I was trying to convince NASA and colleagues that this technique could work. Honestly,

00:47:34.480 | it's so funny because everyone's like, no, it ain't going to work. And it was just like,

00:47:38.960 | because the chemists were saying, of course there are complicated molecules out there you can detect

00:47:43.120 | that just form randomly. I was like, really? That was a bit like, I don't know, someone saying,

00:47:52.720 | of course Darwin's textbook was just written randomly by some monkeys and a typewriter.

00:47:57.600 | Just for me, it was like, really? And I've pushed a lot on the chemists now, and I think

00:48:03.760 | most of them are on board, but not totally. I really had some big arguments, but the copy

00:48:08.560 | number caught there because I think I confused the chemists by saying one off. And then when

00:48:12.560 | I made clear about the copy number, I think that made it a little bit easier. - Just to clarify,

00:48:16.800 | chemists might say that, of course, out there outside of earth, there's complex molecules.

00:48:23.360 | - Yes. - Okay, and then you're saying,

00:48:26.880 | wait a minute, that's like saying, of course, there's aliens out there. - Yeah, exactly that.

00:48:32.160 | - Okay, but you're saying, you clarify that that's actually a very interesting question,

00:48:39.440 | and we should be looking for complex molecules of which the copy number is two or greater.

00:48:44.720 | - Yeah, exactly. So on earth, so coming back to earth, what we did is we took a whole bunch of

00:48:50.000 | samples, and we were running prebiotic chemistry experiments in the lab. We took various inorganic

00:48:58.000 | minerals and extracted them, look at the volatile, because there's a special way of treating minerals

00:49:04.000 | and polymers in assembly theory. In our life machine, we're looking at molecules. We don't

00:49:11.840 | care about polymers because they're not volatile, you can't hold them. If you can't ascern that

00:49:18.800 | they're identical, then it's very difficult for you to work out if there's undergone selection

00:49:24.640 | or they're just a random mess. Same with some minerals, but we can come back to that. So

00:49:28.960 | basically what you do, we got a whole load of samples, inorganic ones, we got a load of,

00:49:33.920 | we got Scotch whiskey, and also took a Ardberg, which is one of my favorite whiskies, which is

00:49:39.520 | very peaty. - What does peaty mean? - So the way that in Scotland, in Islay, which is a little

00:49:47.280 | island, the Scotch, the whiskey is let to mature in barrels, and it's said that the complex molecules

00:50:01.040 | in the peat might find their way through into the whiskey, and that's what gives it this intense

00:50:06.160 | brown color and really complex flavor. It's literally molecular complexity that does that.

00:50:13.120 | So vodka is the complete opposite, it's just pure. - So the better the whiskey,

00:50:17.280 | the higher the assembly index, or the higher the assembly index, the better the whiskey.

00:50:20.400 | - That's what I mean, I really love deep, peaty Scottish whiskeys. Near my house there is one of

00:50:27.520 | the lowland distilleries called Glengoyne, still beautiful whiskey, but not as complex. So for fun,

00:50:33.600 | I took some Glengoyne whiskey in Ardberg and put them into the mass spec and measured the assembly

00:50:38.560 | index. I also got E. coli, so the way we do it, take the E. coli, break the cell apart,

00:50:43.520 | take it all apart, and also got some beer. And people were ridiculing us, saying, "Oh,

00:50:50.960 | beer is evidence of complexity." One of the computational complexity people was just throwing,

00:50:56.880 | yeah, kind of he's very vigorous in his disagreement of assembly theory, was just

00:51:03.760 | saying, "You don't know what you're doing, even beer is more complicated than human."

00:51:07.920 | What we didn't realize is that it's not beer per se, it's taking the yeast extract,

00:51:13.440 | taking the extract, breaking the cells, extracting the molecules, and just looking at the profile of

00:51:18.800 | the molecules to see if there's anything over the threshold. And we also put in a really complex

00:51:22.800 | molecule taxol. So we took all of these, but also NASA gave us, I think, five samples,

00:51:28.400 | and they wouldn't tell us what they are. They said, "No, we don't believe you can get this to work."

00:51:32.640 | And they gave us some super complex samples. And they gave us two fossils, one that was a million

00:51:39.680 | years old, and one was at 10,000 years old, something from Antarctica, seabed. They gave

00:51:45.520 | us the Murchison meteorite and a few others, put them through the system. So we took all the samples,

00:51:52.000 | treat them all identically, put them into mass spec, fragmented them, counted. And in this case,

00:51:58.160 | implicit in the measurement was, in mass spec, you only detect peaks when you've got more than,

00:52:05.440 | say, let's say 10,000 identical molecules. So the copy number's already baked in. It wasn't

00:52:10.560 | quantified, which is super important there. This was in the first paper, because I guess it's

00:52:14.560 | abundant, of course. And when you then took it all out, we found that the biological samples

00:52:23.360 | gave you molecules that had an assembly index greater than 15, and all the abiotic samples

00:52:30.160 | were less than 15. And then we took the NASA samples, and we looked at the ones that were more

00:52:33.600 | than 15 and less than 15, and we gave them back to NASA. They're like, "Oh, gosh. Yep, dead, living,

00:52:39.120 | dead, living. You got it." And that's what we found on Earth. - That's a success. - Yeah. Oh,

00:52:46.080 | yeah. Resounding success. - Can you just go back to the beer and the E. coli? So what's

00:52:52.400 | the sum index on those? - So what you were able to do is, like, the assembly index of... We found

00:53:01.760 | high assembly index molecules originating from the beer sample and the E. coli sample. - The yeast

00:53:08.640 | and the beer. - I mean, I didn't know which one was higher. We didn't really do any detail there,

00:53:13.520 | because now we are doing that, because one of the things we've done... It's a secret, but I can tell

00:53:20.480 | you. - Nobody's listening. - Well, is that we've just mapped the tree of life using assembly theory,

00:53:29.280 | because everyone said, "Oh, you can't do it from biology." And what we're able to do is... So I

00:53:33.280 | think there's three ways... Well, two ways of doing tree of life... Well, three ways, actually. - What's

00:53:38.080 | the tree of life? - So the tree of life is basically tracing back the history of life on Earth for all

00:53:44.880 | the different species, going back who evolved from what. And it all goes all the way back to the first

00:53:49.920 | kind of life forms, and they branch off. And you have plant kingdom, the animal kingdom,

00:53:54.960 | the fungi kingdom, and different branches all the way up. And the way this was classically done...

00:54:02.720 | And I'm no evolutionary biologist. Evolutionary biologists are very... Tell me. Every day,

00:54:07.760 | at least 10 times. I want to be one, though. I kind of like biology. It's kind of cool. - Yeah,

00:54:12.400 | it's very cool. Evolutionary. - But basically, what Darwin and Mendeleev and all these people

00:54:18.240 | do is just they draw pictures, right? And they taxa. They were able to draw pictures and say,

00:54:23.520 | "Oh, these look like common classes." - Yeah. - Then... - They're artists, really. They're just,

00:54:31.040 | you know... - But they were able to find out a lot, right, in looking at verbrates and verbrates,

00:54:35.520 | camera and explosion, all this stuff. And then came the genomic revolution, and suddenly everyone

00:54:42.320 | used gene sequencing. And Craig Venter's a good example. I think he's gone around the world in

00:54:46.400 | his yacht just picking up samples, looking for new species, where he's just found new species of life

00:54:51.440 | just from sequencing. It's amazing. So you have taxonomy, you have sequencing, and then you can

00:54:57.840 | also do a little bit of molecular archaeology, like measure the samples and kind of form some

00:55:07.200 | inference. What we did is we were able to fingerprint... So we took a load of random

00:55:13.520 | samples from all of biology, and we used mass spectrometry. And what we did now is not just

00:55:19.680 | look for individual molecules, but we looked for coexisting molecules where they had to look at

00:55:25.120 | their joint assembly space, and where we were able to cut them apart and undergo recursion in the

00:55:30.560 | mass spec and infer some relationships. And we were able to recapitulate the tree of life using

00:55:36.480 | mass spectroscopy, no sequencing and no drawing. - All right, can you try to say that again,

00:55:44.240 | but a little more detail? So recreating, what does it take to recreate the tree of life?

00:55:49.360 | What does the reverse engineering process look like here? - So what you do is you take an unknown

00:55:53.680 | sample, you pung it into the mass spec, you get a... 'Cause this comes from what you're asking,

00:55:58.720 | like what do you see in E. coli? And so in E. coli, you don't just see... It's not the most

00:56:04.240 | sophisticated cells on Earth make the most sophisticated molecules. It is the coexistence

00:56:10.400 | of lots of complex molecules above a threshold. And so what we realized is you could fingerprint

00:56:16.400 | different life forms. So fungi make really complicated molecules. Why? Because they can't

00:56:21.520 | move. They have to make everything on site. Whereas some animals are lazy. They can just go

00:56:27.600 | eat the fungi. They don't need to make very much. And so what you do is you look at the...

00:56:34.160 | So you take, I don't know, the fingerprint, maybe the top number of high molecular weight molecules

00:56:40.160 | you find in the sample. You fragment them to get their assembly indices. And then what you can do

00:56:45.120 | is you can infer common origins of molecules. You can do a kind of molecular...

00:56:54.800 | When the reverse engineering of the assembly space, you can infer common roots

00:56:59.040 | and look at what's called the joint assembly space. But let's translate that into the experiment.

00:57:04.720 | Take a sample, bung it in the mass spec, take the top, say, 10 molecules, fragment them,

00:57:09.360 | and that gives you one fingerprint. Then you do it for another sample, you get another fingerprint.

00:57:15.840 | Now the question is you say, "Hey, are these samples the same or different?" And that's what

00:57:21.200 | we've been able to do. And by basically looking at the assembly space that these molecules create.

00:57:27.040 | Without any knowledge of assembly theory, you are unable to do it. With a knowledge of assembly

00:57:32.320 | theory, you can reconstruct the tree. - How does knowing if they're the

00:57:36.800 | same or different give you the tree? - Let's go to two leaves on different

00:57:40.800 | branches on the tree, right? What you can do by counting the number of differences, you can

00:57:46.080 | estimate how far away their origin was. - Got it.

00:57:48.960 | - And that's all we do. And it just works. But when we realized you could even use assembly

00:57:53.520 | theory to recapitulate the tree of life with no gene sequencing, we were like, "Huh."

00:57:57.680 | - So this is looking at samples that exist today in the world. What about things that are no longer

00:58:02.880 | exist... I mean, the tree contains information about the past. Some of it is gone.

00:58:10.320 | - Yeah, absolutely. I would love to get old fossil samples and apply assembly theory,

00:58:15.840 | mass spec, and see if we can find new forms of life that are no longer amenable to gene sequencing

00:58:22.240 | because the DNA is all gone. Because DNA and RNA is quite unstable, but some of the more complex

00:58:28.880 | molecules might be there and might give you a hint of something new. Or wouldn't it be great

00:58:33.200 | if you find a sample that's worth really persevering and doing the proper extraction

00:58:41.120 | to PCR and so on, and then sequence it, and then put it together?

00:58:46.000 | - So when a thing dies, you can still get some information about its complexity.

00:58:50.560 | - Yeah. And it appears that you can do some dating. Now, there are really good techniques.

00:58:59.280 | There's radiocarbon dating. There is longer dating, going looking at radioactive minerals,

00:59:04.160 | and so on. And you can also, in bone, you can look at what happens after something dies.

00:59:11.360 | You get what's called racemization, where the chirality in the polymers basically changes,

00:59:19.760 | and you get decomposition. And the deviation from the pure enantiomer to the mixture,

00:59:30.400 | gives you a time scale on it, a half-life. So you can date when it died. I want to use

00:59:37.520 | assembly theory to see if I can use it, date death and things, and trace the tree of life,

00:59:43.280 | and also decomposition of molecules. - You think it's possible?

00:59:46.240 | - Oh yeah, without a doubt. It may not be better than what, because I was just at a conference

00:59:51.600 | where some brilliant people were looking at isotope enrichment, and looking at how life

00:59:56.240 | enriches isotopes, and they're really sophisticated stuff that they're doing. But I think there's some

01:00:01.280 | fun to be had there, because it gives you another dimension of dating. How old is this molecule?

01:00:06.480 | In terms of, or more importantly, how long ago was this molecule produced by life?

01:00:13.200 | The more complex the molecule, the more prospect for decomposition, oxidation, reorganization,

01:00:18.320 | loss of chirality, and all that jazz. But what life also does is it enriches. As you get older,

01:00:25.600 | the amount of carbon-13 in you goes up, because of the way the bonding is in carbon-13. So it

01:00:35.440 | has a slightly different bond strength than you. It's called a kinetic isotope effect. So you can

01:00:39.840 | literally date how old you are, or when you stop metabolizing. So you could date someone's debt,

01:00:45.920 | how old they are, I think. I'm making this up, this might be right. But I think it's roughly

01:00:50.960 | right. The amount of carbon-13 you have in you, you can kind of estimate how old you are.

01:00:55.360 | - How old living organs, humans are? - Yeah, yeah. You could say, oh,

01:01:00.880 | this person is 10 years old, and this person's 30 years old, because they've been metabolizing

01:01:04.720 | more carbon, and they've accumulated it. That's the basic idea. It's probably completely wrong

01:01:09.440 | timescale. - Signatures of chemistry are fascinating.

01:01:12.560 | So you've been saying a lot of chemistry examples for assembly theory. What if we zoom out and look

01:01:20.720 | at a bigger scale of an object? Like really complex objects, like humans, or living organisms

01:01:29.440 | that are made up of millions or billions of other organisms. How do you try to apply assembly theory

01:01:37.360 | to that? - At the moment, we should be able to do this to morphology in cells. So we're looking at

01:01:44.960 | cell surfaces, and really I'm trying to extend further. It's just that we work so hard to get

01:01:52.160 | this paper out, and people to start discussing the ideas. But it's kind of funny, because I

01:02:00.000 | think the penny is falling on this. - What does it mean for a penny to be-

01:02:05.920 | - I mean, the penny's dropped, right? Because a lot of people are like, "It's rubbish. It's rubbish.

01:02:10.320 | You've insulted me. It's wrong." The paper got published on the 4th of October.

01:02:16.320 | It had 2.3 million engagements on Twitter, and it's been downloaded over a few hundred thousand

01:02:22.320 | times. Someone actually wrote to me and said, "This is an example of really bad writing,

01:02:27.040 | and what not to do." I was like, "If all of my papers got read this much, because that's the

01:02:32.160 | objective, if I have a publishing paper and people want to read it, I want to write that badly again."

01:02:36.320 | - Yeah. I don't know what's the deep insight here about the negativity in the space. I think

01:02:41.120 | it's probably the immune system of the scientific community making sure that there's no bullshit

01:02:46.080 | that gets published. And then it can overfire, it can do a lot of damage, it can shut down

01:02:51.120 | conversations in a way that's not productive. - I'll answer your question about the hierarchy

01:02:56.160 | and assembly, but let's go back to the perception. People saying the paper was badly written,

01:03:01.040 | I mean, of course we could improve it. We could always improve the clarity.

01:03:04.000 | - Let's go there before we go to the hierarchy. It has been criticized quite a bit, the paper.

01:03:11.120 | What has been some criticism that you've found most powerful, that you can understand,

01:03:19.200 | and can you explain it? - Yes. The most exciting criticism

01:03:26.640 | came from the evolutionary biologists telling me that they thought that

01:03:29.680 | origin of life was a solved problem. And I was like, "Whoa, we're really onto something,

01:03:35.520 | because it's clearly not." And when you poked them on that, they just said, "No,

01:03:39.280 | you don't understand evolution." And I said, "No, no, I don't think you understand that

01:03:42.560 | evolution had to occur before biology, and there's a gap." That was really, for me,

01:03:47.360 | that misunderstanding, and that did cause an immune response, which was really interesting.

01:03:56.160 | The second thing was the fact that physicists, well, the physicists were actually really polite,

01:04:00.800 | right? Really nice about it. But they just said, "We're not really sure about the initial

01:04:04.960 | conditions thing, but this is a really big debate that we should certainly get into, because

01:04:10.160 | the emergence of life was not encoded in the initial conditions of the universe."

01:04:15.520 | And I think assembly theory shows why it can't be.

01:04:23.760 | Okay, sure, if you could say that again.

01:04:26.400 | The emergence of life was not and cannot, in principle, be encoded in the initial conditions

01:04:34.880 | of the universe.

01:04:35.360 | Just to clarify what I mean by life is like what, high assembly index objects?

01:04:39.680 | Yeah. And this goes back to your favorite subject.

01:04:42.960 | What's that?

01:04:43.600 | Time.

01:04:43.920 | Right, so why? What does time have to do with it?

01:04:51.360 | Probably we can come back to it later, but I think it might be if we have time. But I think

01:04:57.920 | I now understand how to explain how lots of people got angry with the assembly paper, but also the

01:05:08.800 | ramifications of this is how time is fundamental in the universe, and this notion of commentarial

01:05:15.360 | spaces. And there are so many layers on this, but I think you have to become an intuitionist

01:05:23.680 | mathematician, and you have to abandon platonic mathematics, and also platonic mathematics has

01:05:29.680 | left physics astray, but there's a lot back there. So we can go to the…

01:05:34.640 | Platonic mathematics, okay, it's okay. Evolutionary biologists criticize because

01:05:41.440 | the origin of life is understood, and not, it doesn't require an explanation that involves

01:05:50.400 | physics. That's their statement.

01:05:52.880 | Well, I mean, they said lots of confusing statements. Basically, I realized the evolutionary

01:05:59.200 | biology community that were vocal, and some of them were really rude, really spiteful,

01:06:04.160 | and needlessly so, right? Because look, people misunderstand publication as well. Some of the

01:06:14.480 | people have said, "How dare this be published in Nature? What a terrible journal." And it really,

01:06:22.640 | and I've said to people, "Look, this is a brand new idea that's not only potentially going to

01:06:30.160 | change the way we look at biology, it's going to change the way we look at the universe."

01:06:35.920 | And everyone's saying, "How dare you? How dare you be so grandiose?" I'm like, "No, no, no,

01:06:40.880 | this is not hype. We're not saying we've invented some, I don't know, we've discovered an alien in

01:06:48.560 | a closet somewhere just for hype. We genuinely mean this to genuinely have the impact or ask

01:06:54.720 | the question." And the way people jumped on that was a really bad precedent for young people who

01:06:59.600 | want to actually do something new because this makes a bold claim. And the chances are that it's

01:07:08.160 | not correct. But what I wanted to do is a couple of things. I wanted to make a bold claim that was

01:07:14.080 | precise and testable and correctable. Not another woolly information in biology argument, information

01:07:21.920 | Turing machine, blah, blah, blah, blah, blah. A concrete series of statements that can be falsified

01:07:27.760 | and explored and either the theory could be destroyed or built upon.

01:07:32.000 | - Well, what about the criticism of you're just putting a bunch of sexy

01:07:37.600 | names on something that's already obvious? - Yeah, that's really good. So, the assembly

01:07:47.200 | index of a molecule is not obvious, no one had measured it before. And no one has thought to

01:07:52.080 | quantify selection, complexity and copy number before in such a primitive, quantifiable way.

01:08:02.000 | I think the nice thing about this paper, this paper is a tribute to all the people that understand

01:08:09.600 | that biology does something very interesting. Some people call it negentropy, some people call it,

01:08:15.280 | think about organizational principles, that lots of people were not shocked by the paper because

01:08:22.480 | they'd done it before. A lot of the arguments we got, some people said, "Oh, it's rubbish. Oh,

01:08:26.960 | by the way, I had this idea 20 years before." I was like, which one? Is it the rubbish part

01:08:33.600 | or the really revolutionary part? So, this kind of plucked two strings at once. It plucked the,

01:08:38.640 | there is something interesting that biology is as we can see around this, but we haven't quantified

01:08:43.520 | yet. And what this is a first stab at quantifying that. So, the fact that people said, "This is

01:08:52.320 | obvious, but it's also," so if it's obvious, why have you not done it? - Sure, but there's a few

01:09:00.880 | things to say there. One is, this is in part of a philosophical framework because it's not like

01:09:10.880 | you can apply this generally to any object in the universe. It's very chemistry focused.

01:09:15.680 | - Yeah, well, I think you will be able to, we just haven't got there robustly. So,

01:09:19.440 | we can say, how can we, let's go up a level. So, if we go up from level, let's go up from

01:09:24.400 | molecules to cells because you would jump to people and I jump to emoticons and both are

01:09:28.880 | good and they will be assembled. - Let's stick with cells, yeah.

01:09:31.440 | - If we go from, so if we go from molecules to assemblies and let's take a cellular assembly,

01:09:39.760 | a nice thing about a cell is you can tell the difference between a eukaryote and a prokaryote,

01:09:44.480 | right? The organelles are specialized differently. We then look at the cell surface

01:09:48.640 | and the cell surface has different glycosylation patterns and these cells will stick together.

01:09:53.440 | Now, let's go up a level with multicellular creatures. You have cellular differentiation.

01:09:57.520 | Now, if you think about how embryos develop, you go all the way back, those cells undergo

01:10:02.080 | differentiation in a causal way that's biomechanically a feedback between the

01:10:06.880 | genetics and biomechanics. I think we can use assembly theory to apply to tissue types.

01:10:11.920 | We can even apply it to different cell disease types. So, that's what we're doing next but we're

01:10:16.560 | trying to walk, you know, the thing is I'm trying to leap ahead. I want to leap ahead to go,

01:10:21.280 | well, we apply it to culture. Clearly, you can apply it to memes and culture

01:10:25.920 | and we've also applied assembly theory to CAs and not as you think.

01:10:33.520 | - Cellular automata, better. - Yeah, yeah, to cellular automata,

01:10:35.920 | not just as you think. Different CA rules were invented by different people at different times

01:10:41.120 | and one of my co-workers, very talented chap, basically was like, oh, I can realize that

01:10:48.320 | different people had different ideas or different rules and they copied each other and made slightly

01:10:52.880 | different bit but different cellular automata rules and looked at them online and so he was

01:10:58.640 | able to infer an assembly index and copy number of rule whatever doing this thing but I digress.

01:11:04.240 | But it does show you can apply it at a higher scale. So, what do we need to do to apply assembly

01:11:09.920 | theory to things? We need to agree there's a common set of building blocks. So, in a cell,

01:11:15.120 | well, in a multicellular creature, you need to look back in time. So, there is the initial cell

01:11:22.160 | which the creature is fertilized and then starts to grow and then there is cell differentiation

01:11:27.600 | and you have to then make that causal chain both on those. That requires

01:11:32.800 | development of the organism in time or if you look at the cell surfaces and the cell types,

01:11:39.040 | they've got different features on the cell, the walls and inside the cell. So, we're building up

01:11:47.280 | but obviously, I want a leap to things like emoticons, language, mathematical theorem.

01:11:53.200 | - But that's a very large number of steps to get from a molecule to the human brain.

01:12:01.040 | - Yeah, and I think they are related but in hierarchies of emergence, right? So,

01:12:06.400 | you shouldn't compare them. I mean, the assembly index of a human brain, what does that even mean?

01:12:11.200 | Well, maybe we can look at the morphology of the human brain. Say, all human brains have these

01:12:15.840 | number of features in common. If they have those number of and then let's look at a brain in a whale

01:12:21.760 | or a dolphin or a chimpanzee or a bird and say, "Okay, let's look at the assembly indices number

01:12:27.840 | of features in these," and now the copy number is just a number of how many birds are there,

01:12:32.400 | how many chimpanzees are there, how many humans are there.

01:12:34.640 | - But then you have to discover for that the features that you would be looking for.

01:12:39.360 | - Yeah, and that means you need to have some idea of the anatomy.

01:12:42.800 | - But is there an automated way to discover features?

01:12:45.280 | - I guess so. I mean, and I think this is a good way to apply machine learning and image

01:12:53.040 | recognition just to basically characterize things.

01:12:55.040 | - So apply compression to it to see what emerges and then use the thing,

01:12:58.640 | the features used as part of the compression as the measurement of,

01:13:04.320 | as the thing that is searched for when you're measuring assembly index and copy number.

01:13:09.280 | - And the compression has to be, remember, the assembly universe, which is you have to go from

01:13:13.760 | assembly possible to assembly contingent and that jump from, because assembly possible,

01:13:19.120 | all possible brains, all possible features all the time. But we know that on the tree of life

01:13:25.600 | and also on the lineage of life, going back to Luca, the human brain just didn't spring into

01:13:29.360 | existence yesterday. It is a long lineage of brains going all the way back. And so if we could

01:13:34.800 | do assembly theory to understand the development, not just in evolutionary history, but in biological

01:13:41.440 | development as you grow, we are going to learn something more.

01:13:44.720 | - What would be amazing is if you can use assembly theory, this framework to show the

01:13:51.680 | increase in the assembly index associated with, I don't know, cultures or pieces of text like

01:14:01.680 | language or images and so on, and illustrate without knowing the data ahead of time, just

01:14:08.000 | kind of like you did with NASA, that you were able to demonstrate that it applies in those other

01:14:12.080 | contexts. I mean, and that, you know, probably wouldn't at first, and you have to evolve the

01:14:17.120 | theory somehow, you have to change it, you have to expand it, you know.

01:14:20.880 | - I think so.

01:14:22.240 | - But like that, I guess this is as a paper, a first step in saying, okay, can we create a general

01:14:30.800 | framework for measuring complexity of objects, for measuring life, the complexity of living

01:14:38.000 | organisms?

01:14:38.560 | - Yeah.

01:14:39.200 | - That's what this is reaching for.

01:14:40.960 | - That is the first step, and also to say, look, we have a way of quantifying selection

01:14:46.640 | and evolution in a fairly, not mundane, but a fairly mechanical way. Because before now,

01:14:54.480 | it wasn't very, the ground truth for it was very subjective, whereas here we're talking

01:15:00.880 | about clean observables, and there's gonna be layers on that. I mean, with collaborators right

01:15:05.920 | now, we already think we can do assembly theory on language, and not only that, wouldn't it be

01:15:10.560 | great if we can, so the, if we can figure out how under pressure language is gonna evolve and be

01:15:18.080 | more efficient, 'cause you're gonna wanna transmit things, and again, it's not just about compression,

01:15:23.120 | it is about understanding how you can make the most of the architecture you've already built.

01:15:29.760 | And I think this is something beautiful that evolution does, we're reusing those architectures,

01:15:35.040 | we can't just abandon our evolutionary history, and if you don't wanna abandon your evolutionary

01:15:39.440 | history, and you know that evolution has been happening, then assembly theory works. And I

01:15:44.400 | think that's a key comment I wanna make, is that assembly theory is great for understanding

01:15:49.600 | where evolution has been used. The next jump is when we go to technology, 'cause of course,

01:15:55.680 | if you take the M3 processor, I wanna buy, I haven't bought one yet, I can't justify it,

01:16:00.880 | but I want it at some point. The M3 processor arguably is, there's quite a lot of features,

01:16:05.520 | a quite large number, the M2 came before it, then the M1, all the way back, you can apply

01:16:10.160 | assembly theory to microprocessor architecture. It doesn't take a huge leap to see that.

01:16:15.360 | I'm a Linux guy, by the way, so your examples go way over my head.

01:16:18.320 | Yeah, well, whatever.

01:16:19.120 | Is that, is that like a, is that a fruit company of some sort? I don't even know.

01:16:22.640 | Yeah, there's a lot of interesting stuff to ask about language, like you could look at,

01:16:27.040 | how would that work? You could look at GPT-1, GPT-2, GPT-3, 3, 5, 4,

01:16:35.280 | and try to analyze the kind of language it produces. I mean, that's almost trying to look

01:16:41.680 | at assembly index of intelligent systems.

01:16:44.400 | Yeah, I mean, I think the thing about large language models, and this is a whole hobby

01:16:51.760 | horse I have at the moment, is that obviously they're all about the evidence of evolution

01:17:03.760 | in the large language model comes from all the people that produced all the language.

01:17:08.480 | And that's really interesting, and all the corrections in the mechanical Turk, right?

01:17:14.960 | Sure.

01:17:15.460 | And, and so...

01:17:17.680 | That's the part of the history, part of the memory of the system.

01:17:20.000 | Exactly. So, can you, so, so it would be really interesting to basically use

01:17:25.520 | an assembly-based approach to making language in a hierarchy, right? I think, my guess is that

01:17:33.520 | you could, we might be able to build a new type of large language model that uses assembly theory,

01:17:39.360 | that it has more understanding of the past and how things were created.

01:17:45.600 | Basically, the thing with LLMs is they're like everything, everywhere, all at once,

01:17:50.560 | splat, and make the user happy. So, there's not much intelligence in the model. The model is how

01:17:56.560 | the human interacts with the model, but wouldn't it be great if we could understand how to embed

01:18:00.560 | more intelligence in the system?

01:18:03.280 | What do you mean by intelligence there? Like, you seem to associate intelligence with history.

01:18:10.880 | Yeah, well, I think selection produces intelligence.

01:18:14.240 | You're almost implying that selection is intelligence. No.

01:18:20.640 | Yeah, kind of. I would go that, I would go out on a limb and say that, but I think it's

01:18:23.840 | a little bit more human beings have the ability to abstract, and they can break beyond selection.

01:18:28.480 | And this is what, like, Darwinian selection, because the human being doesn't have to basically

01:18:33.680 | do trial and error, but they can think about it and say, "Oh, that's a bad idea. I won't do that,"

01:18:37.920 | and then technologies and so on.

01:18:39.440 | So, we escaped Darwinian evolution, and now we're onto some other kind of evolution, I guess,

01:18:44.480 | higher level evolution.

01:18:46.400 | And assembly theory will measure that as well, right? Because it's all a lineage.

01:18:50.480 | Okay, another piece of criticism, or by way of question, is how is assembly theory,

01:18:57.920 | or maybe assembly index, different from Kolmogorov complexity? So, for people who don't know,

01:19:02.320 | Kolmogorov complexity of an object is the length of a shortest computer program

01:19:07.440 | that produces the object as output.

01:19:09.360 | Yeah, I seem to, there seems to be a disconnect between the computational approach. So, yeah,

01:19:18.000 | so Kolmogorov measure requires a Turing machine, requires a computer, and that's one thing.

01:19:28.480 | And the other thing is, assembly theory is supposed to trace the process

01:19:36.160 | by which life evolution emerged, right? That's the main thing there.

01:19:41.600 | There are lots of other layers. So, Kolmogorov complexity, you can approximate Kolmogorov

01:19:49.040 | complexity, but it's not really telling you very much about the actual, it's really telling you

01:19:56.960 | about like your data set, compression of your data set. And so, that doesn't really help you

01:20:03.200 | identify the turtle, in this case, is the computer. And so, what assembly theory does is,

01:20:09.840 | I'm going to say, this is a trigger warning for anyone listening who loves complexity theory,

01:20:17.600 | I think that we're going to show that AIT is a very important subset of assembly theory,

01:20:23.120 | because here's what happens. I think that assembly theory allows us to build,

01:20:30.960 | understand when were selections occurring, selection produces factories and things,

01:20:36.480 | factories in the end produce computers, and you can get an algorithmic information theory comes

01:20:41.600 | out of that. The frustration I've had with looking at life through this kind of information theory is

01:20:47.920 | it doesn't take into account causation. So, the main difference between assembly theory and all

01:20:54.560 | these complexity measures is there's no cause or chain. And I think that's the main-

01:21:00.240 | That's the causal change at the core of assembly theory.

01:21:06.160 | Exactly. And if you've got all your data in a computer memory, all the data is the same,

01:21:10.400 | you can access it in the same way, you don't care, you just compress it, and you either look at a

01:21:16.240 | program runtime or the shortest program. And that, for me, is absolutely not capturing what it is,

01:21:27.120 | what its selection does.

01:21:28.160 | But assembly theory looks at objects, it doesn't have information about

01:21:32.720 | the object history, it's going to try to infer that history by looking for the shortest history,

01:21:42.160 | right? The object doesn't have a Wikipedia page that goes with it, about its history.

01:21:49.760 | I would say it does in a way, and it is fascinating to look at. So, you've just got the object,

01:21:55.680 | and you have no other information about the object. What assembly theory allows you to do

01:21:59.040 | just with the object is to- and the word "infer" is correct, I agree, "infer". You say,

01:22:05.920 | "Well, that's not the history." But something really interesting comes from this.

01:22:09.760 | The shortest path is inferred from the object. That is the worst-case scenario if you have no

01:22:16.640 | machine to make it. So, that tells you about the depth of that object in time.

01:22:23.280 | And so, what assembly theory allows you to do is, without considering any other circumstances,

01:22:28.560 | to say from this object, "How deep is this object in time?"

01:22:31.200 | If we just treat the object as itself, without any other constraints. And that's super powerful,

01:22:38.960 | because the shortest path then says, allows you to say, "Oh, this object wasn't just created

01:22:43.760 | randomly, there was a process." And so, assembly theory is not meant to, you know, one-up AIT,

01:22:51.520 | or to ignore the factory. It's just to say, "Hey, there was a factory. How big was that factory,

01:22:59.920 | and how deep in time is it?" - But it's still computationally very difficult to compute

01:23:06.080 | that history, right, for complex objects. - It is, it becomes harder. But one of the

01:23:13.360 | things that's super nice is that it constrains your initial conditions, right? It constrains

01:23:19.120 | where you're going to be. So, if you take, say, imagine, so one of the things we're doing right

01:23:23.680 | now is applying assembly theory to drug discovery. Now, what everyone's doing right now is taking

01:23:29.120 | all the proteins, and looking at the proteins, and looking at molecules docked with proteins.

01:23:33.600 | Why not instead look at the molecules that are involved in interacting with the receptors over

01:23:40.320 | time, rather than thinking about, and use the molecules that evolve over time as a proxy for

01:23:45.280 | how the proteins evolved over time, and then use that to constrain your drug discovery process. You

01:23:51.280 | flip the problem 180, and focus on the molecule evolution, rather than the protein. And so,

01:23:58.000 | you can guess in the future what might happen. So, rather than having to consider all possible

01:24:04.080 | molecules, you know where to focus. And that's the same thing if you're looking at an assembly

01:24:08.960 | spaces for an object where you don't know the entire history, but you know that in the history

01:24:15.680 | of this object, it's not going to have some other motif there that it doesn't appear in the past.

01:24:21.920 | But just even for the drug discovery point you made, don't you have to simulate all of chemistry

01:24:27.120 | to figure out how to come up with constraints?

01:24:31.760 | No.

01:24:32.640 | In the molecules?

01:24:33.440 | No.

01:24:34.480 | I mean, I don't know enough about protein.

01:24:37.280 | This is another thing that I think causes, because this paper goes across so many boundaries. So,

01:24:41.840 | chemists have looked at this and said, "This is not correct reaction." I was like, "No, it's a graph."

01:24:48.960 | Sure, there's assembly index and shortest path examples here on chemistry.

01:24:58.320 | Yeah. And so, what you do is you look at the minimal constraints on that graph.

01:25:03.840 | Of course, it has some mapping to the synthesis, but actually, you don't have to know all of

01:25:09.280 | chemistry. You can build up the constraint space rather nicely. But this is just at the beginning,

01:25:16.800 | right? There are so many directions this could go in. And as I said, it could all be wrong,

01:25:20.560 | but hopefully it's less wrong.

01:25:22.160 | What about the little criticism I saw of, by way of question, do you consider the different

01:25:30.000 | probabilities of each reaction in the chain? So, that there could be different... When you look at

01:25:37.600 | a chain of events that led up to the creation of an object, doesn't it matter that some parts in

01:25:44.160 | the chain are less likely than others?

01:25:46.640 | No.

01:25:46.880 | It doesn't matter?

01:25:48.320 | No, no. Well, let's go back. So, no, not less likely, but react... So, no. So, let's go back

01:25:54.960 | to what we're looking at here. So, the assembly index is the minimal path that could have created

01:26:00.000 | that object probabilistically. So, imagine you have all your atoms in a plasma, you've got enough

01:26:05.280 | energy, you've got enough... There's collisions. What is the quickest way you could zip out that

01:26:10.480 | molecule with no reaction constraints?

01:26:12.560 | How do you define quickest there, then?

01:26:14.480 | It's just basically a walk on a random graph. So, we make an assumption that basically,

01:26:19.680 | the timescale for forming the bonds... So, no, I don't want to say that,

01:26:23.040 | because then it's going to have people getting obsessing about this point, and your criticism

01:26:26.080 | is a really good one. What we're trying to say is like, this puts a lower bound on something.

01:26:30.960 | Of course, some reactions are less possible than others, but actually,

01:26:35.920 | I don't think chemical reactions exist.

01:26:38.400 | Oh, boy. What does that mean? Why don't chemical reactions exist?

01:26:43.520 | I'm writing a paper right now that I keep being told I have to finish, and it's called

01:26:48.640 | "The Origin of Chemical Reactions", and it merely says that reactivity exists as controlled by the

01:26:54.880 | laws of quantum mechanics, and reactions... Chemists put names on reactions. So, you could

01:27:00.560 | have like, I don't know, the Wittig reaction, which is by Wittig. You could have the Suzuki

01:27:06.720 | reaction, which is by Suzuki. Now, what are these reactions? So, these reactions are constrained by

01:27:12.480 | the following. They're constrained by the fact they're on planet Earth, 1G, 298 Kelvin, 1 bar.

01:27:19.280 | So, these are constraints. They're also constrained by the chemical composition of Earth,

01:27:24.800 | oxygen, availability, all this stuff, and that then allows us to focus in our chemistry.

01:27:30.800 | So, when a chemist does a reaction, that's a really nice compressed shorthand for constraint

01:27:36.720 | application. Glass flask, pure reagent, temperature, pressure, bom-bom-bom-bom-bom,

01:27:42.240 | control-control-control-control-control. So, of course, we have bond energies.

01:27:48.320 | So, the bond energies are kind of intrinsic in a vacuum, if you say that. So, the bond energy,

01:27:54.800 | you have to have a bond. And so, for assembly theory to work, you have to have a bond, which

01:28:00.480 | means that bond has to give the molecule a certain half-life. So, you're probably going to find later

01:28:05.920 | on that some bonds are weaker and that you are going to miss in mass spectra. When you count,

01:28:11.760 | look at the assembly of some molecules, you're going to miscount the assembly of the molecule

01:28:16.400 | because it falls apart too quickly because the bonds just form. But you can solve that

01:28:20.000 | by looking at infrared. So, when people think about the probability, they're kind of

01:28:25.120 | misunderstanding. Assembly theory says nothing about the chemistry because chemistry is chemistry

01:28:31.600 | and the constraints are put in by biology. There was no chemist on the origin of life, unless you

01:28:37.200 | believe in the chemist in the sky and they were, you know, it's like Santa Claus, they had a lot

01:28:41.680 | of work to do. But chemical reactions do not exist in the constraints that allow chemical

01:28:49.680 | transformations to occur, do exist. - Okay, okay. So, it's constrained to

01:28:54.880 | applicate, so there's no chemical reactions, it's all constraint application, which enables the

01:29:03.040 | emergence of, what's a different word for chemical reaction? - Transformation.

01:29:10.880 | - Transformation. - Yeah, like a function,

01:29:12.560 | it's a function. But no, I love chemical reactions as a shorthand and so the chemists don't all go

01:29:18.160 | mad. I mean, of course chemical reactions exist on Earth. - It's a shorthand.

01:29:21.200 | - It's a shorthand for these constraints. - Right. So, assuming all these constraints

01:29:25.680 | that we've been using for so long, we just assume that that's always the case

01:29:29.040 | in natural language conversation. - Exactly. The grammar of chemistry,

01:29:33.200 | of course, emerges in reactions and we can use them reliably, but I do not think the

01:29:38.000 | Wittig reaction is accessible on Venus. - Right, and this is useful to remember,

01:29:42.640 | you know, to frame it as constraint application is useful for when you zoom out to the bigger

01:29:49.280 | picture of the universe and looking at the chemistry of the universe and then starting

01:29:52.400 | to apply assembly theory. That's interesting, that's really interesting.

01:29:56.080 | But we've also pissed off the chemists now. - Oh, they're pretty happy, but, well, most of them.

01:30:04.160 | - No, everybody, everybody deep down is happy, I think. They're just sometimes feisty,

01:30:11.440 | that's how they show, that's how they have fun. - Everyone is grumpy on some days when you

01:30:15.440 | challenge. The problem with this paper is you, what it's like, it's almost like I went to a party,

01:30:19.440 | it's like you, I used to do this occasionally when I was young, is go to a meeting and just

01:30:24.320 | find a way to offend everyone at the meeting simultaneously. Even the factions that don't

01:30:29.440 | like each other, they're all unified in their hatred of you just offending them. This paper,

01:30:33.600 | it feels like the person that went to the party and offended everyone simultaneously,

01:30:37.440 | so they stopped fighting with themselves and just focused on this paper.

01:30:40.000 | - Maybe just a little insider interesting information, what were the editors of Nature,

01:30:46.480 | the reviews and so on, how difficult was that process? Because this is a pretty big paper.

01:30:52.960 | - Yeah, I mean, so when we originally sent the paper, we sent the paper and the editor said,

01:31:02.320 | like, this is quite a long process. We sent the paper and the editor gave us some feedback.

01:31:10.240 | And said, you know, I don't think it's that interesting, or it's hard, it's a hard concept.

01:31:17.920 | And we asked, and the editor gave us some feedback. And Sarah and I took a year to rewrite the paper.

01:31:26.400 | - Was the nature of the feedback very specific on, like, this part, this part? Or was it like,

01:31:31.200 | what are you guys smoking, what kind of? - Yeah, it was kind of the latter,

01:31:34.560 | what are you smoking? - Okay.

01:31:35.520 | - And, you know. - But polite and there's promise.

01:31:41.040 | - Yeah, well, the thing is, the editor was really critical, but in a really professional way. And I

01:31:48.400 | mean, for me, this was the way science should happen. So when it came back, you know, we had

01:31:52.400 | too many equations in the paper. If you look at the preprint, they're just equations everywhere,

01:31:55.680 | like 23 equations. And when I said to Abhishek, who was the first author, we got to remove all

01:31:59.840 | the equations. But my assembly equations thing, Abhishek was like, you know, no, we can't. I said,

01:32:05.040 | well, look, if we want to explain this to people, it's a real challenge. And so Sarah and I went

01:32:10.400 | through the, I think it was actually 160 versions of the paper, but we basically, we got to version

01:32:15.200 | 40 or something. We said, right, zero, let's start again. So we wrote the whole paper again.

01:32:20.080 | We knew the entire- - Amazing.

01:32:22.000 | - And we just went bit by bit by bit and said, what is it we want to say? And then we sent the

01:32:26.800 | paper in and to our, we expected it to be rejected and not even go to review. And then we got

01:32:35.040 | notification back, it had gone to review and we were like, oh my God, it's so going to get rejected.

01:32:39.440 | How's it going to get rejected? Because the first assembly paper that on the mass spec we sent to

01:32:43.680 | nature, went through six rounds of review and were rejected, right? And this biochemist just

01:32:50.560 | said, I don't believe you, you must be committing fraud. And a long story, probably a boring story,

01:32:55.920 | but in this case, it went out to review, the comments came back and the comments were incredibly,

01:33:02.960 | they were very deep comments from all the reviewers. But the nice thing was the reviewers

01:33:18.400 | were kind of very critical, but not dismissive. They were like, oh really? Explain this, explain

01:33:24.960 | this, explain this, explain this. Are you sure it's not comical or off? Are you sure it's not this?

01:33:28.880 | And we went through, I think three rounds of review, pretty quick. And the editor went,

01:33:36.560 | yeah, it's in. - Maybe you could just comment on the whole process. You've published some

01:33:42.320 | pretty huge papers on all kinds of topics within chemistry and beyond. Some of them have some

01:33:47.040 | little spice in them, a little spice of crazy. Like Tom Waits says, I like my town with a little

01:33:53.840 | drop of poison. It's not a mundane paper. So what's it like psychologically to go through all

01:34:03.200 | this process, to keep getting rejected, to get reviews from people that don't get the paper or

01:34:08.640 | all that kind of stuff? Just from a question of a scientist, what is that like? - I think it's,

01:34:19.520 | I mean, this paper for me kind of, because this wasn't the first time we tried to publish

01:34:24.400 | assembly theory at the highest level. The nature communications paper on the mass spec,

01:34:29.280 | on the idea, went through, went to nature and got rejected, went through six rounds of review and

01:34:34.960 | got rejected. And I just was so confused when the chemist said, this can't be possible. I do not

01:34:46.400 | believe you can measure complexity using mass spec. And also, by the way, molecules, complex

01:34:54.720 | molecules can randomly form. And we're like, but look at the data, the data says. And they said,

01:34:59.360 | no, no, we don't believe you. And we went, and I just wouldn't give up. And the editor in the end

01:35:07.120 | was just like, the different editors actually, right? - What's behind that never giving up?

01:35:11.840 | Is it like, when you're sitting there, 10 o'clock in the evening, there's a melancholy feeling that

01:35:17.920 | comes over you. And you're like, okay, this is rejection number five. Or it's not rejection,

01:35:24.240 | but maybe it feels like a rejection because the comments are, you totally don't get it.

01:35:29.600 | Like, what gives you strength to keep going there? - I don't know.

01:35:40.160 | I don't normally get emotional about papers, but

01:35:42.160 | it's not about giving up because we want to get it published, because we want the glory

01:35:59.840 | or anything. It's just like, why don't you understand? And so, what I would just try to be

01:36:11.200 | as rational as possible and say, yeah, you didn't like it. Tell me why. And then,

01:36:26.480 | sorry. Silly. Never get emotional about papers normally, but I think what we do,

01:36:33.840 | you just compress like five years of angst from this. - So, it's been rough.

01:36:39.200 | - It's not just rough. It's like, it happened, I came up with the assembly equation,

01:36:43.680 | remote from Sarah in Arizona and the people at SFI. I felt like I was a mad person. Like,

01:36:53.120 | the guy depicted in A Beautiful Mind who was just like, not the actual genius part,

01:36:59.040 | but just the gibberish. Because I kept writing expanded and I have no mathematical ability at

01:37:06.400 | all. And I was making these mathematical expansions where I kept seeing the same

01:37:10.560 | motif again. I was like, I think this is a copy number. The same string is coming again and again

01:37:15.760 | and again. I couldn't do the math. And then I realized the copy number fell out of the equation

01:37:20.240 | and everything collapsed down. I was like, oh, that works, kind of. So, we submitted the paper.

01:37:25.040 | And then when it was almost accepted, the mass spec one, and it was astrobiologists said great,

01:37:31.040 | mass spectroscopists said great, and the chemists went, nonsense. Like, biggest pile of nonsense

01:37:36.720 | ever, fraud. And I was like, but why fraud? And they just said, just because. And I was like,

01:37:42.880 | well, and so, and I could not convince the editor in this case. The editor was just so pissed off,

01:37:50.080 | because they see it as like a kind of, you know, you're wasting my time. And I would not give up.

01:37:55.760 | I wrote, I went and dissected, you know, all the parts. And I think, although, I mean,

01:38:01.360 | I got upset about it, you know, it was kind of embarrassing actually, but I guess.

01:38:08.240 | But it was just trying to understand why they didn't like it. So, they were part of me was like,

01:38:14.160 | really devastated. And a part of me was super excited, because I'm like, huh, they can't tell

01:38:19.440 | me why I'm wrong. And this kind of goes back to, you know, when I was at school, I was in a kind

01:38:25.680 | of learning difficulties class. And I kept going to the teacher and say, you know, how, what do I

01:38:30.800 | do today to prove I'm smart? And they were like, nothing, you can't. I was like, give me a job,

01:38:35.200 | you know, give me something to do. Give me a job to do something to do as we. And I kind of felt

01:38:41.680 | like that a bit when I was arguing with the, and not arguing, there was no ad hominem, I wasn't

01:38:45.920 | telling the editor, they were idiots or anything like this, or the reviewers, I kept it strictly

01:38:50.160 | like factual. And all I did is I just kept knocking it down bit by bit by bit by bit by bit.

01:38:56.640 | It was ultimately rejected, and it got published elsewhere. And then the actual experiment or data.

01:39:04.400 | So this is kind of the, in this paper, the experiment or justification was already published.

01:39:08.960 | So when we did this one, and we went through the versions, and then we sent it in, and in the end,

01:39:14.720 | it just got accepted. We were like, well, that's kind of cool, right? This is kind of like, you

01:39:19.040 | know, some days you had, you know, the student, sorry, the first author was like, I can't believe

01:39:25.600 | it got accepted. Like, nor am I, but it's great. It's like, it's good. And then when the paper was

01:39:30.880 | published, I was not expecting the backlash. I was expecting computational, well, no, actually,

01:39:38.000 | I was just expecting one person had been trolling me for a while about it, just to carry on trolling,

01:39:42.160 | but I didn't expect the backlash. And then I wrote to the editor and apologized. And the editor was

01:39:48.400 | like, what are you apologizing for? It was a great paper. Of course it's going to get backlash. You

01:39:52.800 | said some controversial stuff, but it's awesome. - Well, I think it's a beautiful story of

01:39:58.720 | perseverance, and the backlash is just a negative word for discourse, which I think is beautiful.

01:40:06.720 | - I think you, as I said to, you know, when it got accepted, and people were saying,

01:40:13.360 | were kind of like hacking on it, and I was like, papers are not gold medals. The reason I wanted

01:40:20.080 | to publish that paper in Nature is because it says, hey, there's something before biological

01:40:27.920 | evolution. You have to have that if you're not a creationist, by the way. This is an approach.

01:40:33.920 | First time someone has put a concrete mechanism, or sorry, a concrete quantification, and what comes

01:40:40.080 | next, you're pushing on, is a mechanism. And that's what we need to get to, is an autocatalytic

01:40:44.960 | set, self-replicating molecules, some other features that come in. And the fact that this

01:40:51.440 | paper has been so discussed, for me, is a dream come true. Like, it doesn't get better than that.

01:40:57.920 | If you can't accept a few people hating it, and the nice thing is, the thing that really makes

01:41:03.040 | me happy, is that no one has attacked the actual physical content. Like, you can measure the

01:41:10.720 | assembly index. You can measure selection now. So either that's right, or it's, well, either that's

01:41:16.960 | helpful or unhelpful. If it's unhelpful, this paper will sink down, and no one will use it again.

01:41:22.880 | If it's helpful, it'll help people build scaffold on it, and we'll start to converge to a new

01:41:27.360 | paradigm. So I think that that's the thing that I wanted to see, you know, my colleagues, authors,

01:41:35.040 | collaborators, and people who were like, you've just published this paper, you're a chemist. Why

01:41:40.000 | have you done this? Like, who are you to be doing evolutionary theory? Like, well, I don't know. I

01:41:46.400 | mean, sorry, did I need to-- - Who is anyone to do anything? Well, I'm glad you did. Let me just,

01:41:51.520 | before coming back to origin of life and these kinds of questions, you mentioned learning

01:41:57.040 | difficulties. I didn't know about this. So what was it like? - I wasn't very good at school, right?

01:42:03.440 | - This is when you were very young? - Yeah, yeah. Well, but in primary school,

01:42:07.680 | my handwriting was really poor, and apparently I couldn't read, and my mathematics was very poor.

01:42:15.920 | So they just said, this is a problem. They identified it. My parents kind of at the time

01:42:20.080 | were confused because I was busy taking things apart, buying electronic junk from the shop,

01:42:24.320 | trying to build computers and things. And then once I got out of, when I was, I think about

01:42:31.120 | the major transition in my stupidity, like, you know, everyone thought I wasn't that stupid when

01:42:36.640 | I was, basically everyone thought I was faking. I like stuff, and I was faking wanting to be it.

01:42:41.200 | So I always wanted to be a scientist. So five, six, seven years, I'll be a scientist, take things

01:42:45.360 | apart. And everyone's like, yeah, this guy wants to be a scientist, but he's an idiot.

01:42:49.520 | And so everyone was really confused, I think, at first, that I wasn't smarter than I, you know,

01:42:56.640 | was claiming to be. And then I just basically didn't do well in any of the tests, and I went

01:43:00.320 | down and down and down and down. And then, and I was kind of like, huh, this is really embarrassing.

01:43:07.040 | I really like maths, and everyone says I can't do it. I really like kind of, you know, physics and

01:43:13.120 | chemistry and all that in science. And people say, you know, you can't, you can't read and write.

01:43:17.680 | And so I found myself in a learning difficulties class at the end of primary school and the

01:43:22.720 | beginning of secondary school. In the UK, secondary school is like 11, 12 years old.

01:43:27.440 | And I remember being put in the, in the remedial class. And the remedial class was basically full

01:43:34.720 | of, well, two types, three types of people. There were people that had quite violent, right? And

01:43:44.800 | there were people who couldn't speak English, and there were people that really had learning

01:43:50.320 | difficulties. So the one thing I can objectively remember was, I mean, I could read. I like

01:44:10.880 | reading. I read a lot. But something in me, I was, I'm a bit of a rebel. I refused to read

01:44:18.480 | what I was told to read. And I found it difficult to read individual words in the way they were told.

01:44:24.400 | But anyway, I got caught one day teaching someone else to read. And they said, okay,

01:44:32.480 | we don't understand this. I always knew I wanted to be a scientist, but I didn't really know what

01:44:39.680 | that meant. And I realized you have to go to university. And I thought, I can just go to

01:44:42.720 | university. It's like curious people, like, no, no, no, you need to have these, you have to be

01:44:46.240 | able to enter these exams to get this grade point average. And the fact is, the exams you've been

01:44:50.880 | entered into, you're just going to get C, D, or E. You can't even get A, B, or C, right? These

01:44:57.600 | are the UK GCSEs. I was like, oh, shit. And I said, can you just put me into the higher exams?

01:45:03.840 | They said, no, no, you're going to fail. There's no chance. So my father kind of intervened and

01:45:09.760 | said, you know, just let him go in the exams. And they said, he's definitely going to fail. It's a

01:45:14.720 | waste of time, waste of money. And he said, well, what if we paid? So they said, well, okay. So you

01:45:20.560 | didn't actually have to pay. You didn't have to pay if I failed. So I took the exams and passed

01:45:24.720 | them, fortunately. I didn't get the top grades, but I, you know, I got into A-levels. But then

01:45:30.240 | that also kind of limited what I could do at A-levels. I wasn't allowed to do A-level maths.

01:45:34.320 | - What do you mean you weren't allowed to? - Because I had such a bad math grade from

01:45:39.040 | my GCSE, I only had a C. But they wouldn't let me go into the ABC for maths because of some kind of

01:45:44.320 | coursework requirement back then. So the top grade I could have got was a C. So C, D, or E. So I got

01:45:48.960 | a C. And they let me do kind of AS-level maths, which is this half intermediate, but go to

01:45:55.200 | university. But in the end, I liked chemistry. I had a good chemistry teacher. So in the end,

01:45:59.200 | I got to university to do chemistry. - So through that kind of process, I think

01:46:03.280 | for kids in that situation, it's easy to start believing that you're not,

01:46:10.720 | well, how do I put it? That you're stupid. And basically give up that you're just not

01:46:16.240 | good at math, you're not good at school. So this is by way of advice for people,

01:46:21.120 | for interesting people, for interesting young kids right now experiencing the same thing.

01:46:27.520 | Where was the place, what was the source of you not giving up there?

01:46:31.920 | - I have no idea other than I was really, I really like not understanding stuff. For me,

01:46:42.640 | when I not understand something, I didn't understand, I feel like I don't understand

01:46:47.680 | anything now. But back then I was so, I remember when I was like, I don't know,

01:46:53.040 | I tried to build a laser when I was like eight. And I thought, how hard could it be?

01:47:00.960 | And I basically, I was gonna build a CO2 laser. I was like, right, I think I need some partially

01:47:08.160 | coated mirrors, I need some carbon dioxide, and I need a high voltage. So I kind of,

01:47:17.040 | and I was like, I didn't have, and I was so stupid, right? I was kind of so embarrassed.

01:47:21.920 | I had to make enough CO2, I actually set a fire and try to filter the flame.

01:47:26.800 | - Oh, nice.

01:47:27.760 | - To crap enough CO2. And I was like, completely failed, and I burnt half the garage down.

01:47:35.120 | So my parents were not very happy about that. So that was one thing, I was like,

01:47:39.120 | I really like first principle thinking. And so, I remember being super curious

01:47:46.560 | and being determined to find answers. And so the kind of, when people do give advice about this,

01:47:52.640 | well, I ask for advice about this, I don't really have that much advice other than don't

01:47:56.640 | give up. And one of the things I try to do as a chemistry professor in my group, is I hire people

01:48:05.760 | that I think that, you know, I'm kind of, who am I, if they're persistent enough,

01:48:10.800 | who am I to deny them the chance? Because, you know, people gave me a chance and I was able to

01:48:17.200 | do stuff.

01:48:17.760 | - Do you believe in yourself, essentially?

01:48:20.080 | - I like, so I love being around smart people, and I love confusing smart people.

01:48:25.040 | And when I'm confusing smart people, you know, not by stealing their wallets and hiding it

01:48:29.600 | somewhere, but if I can confuse smart people, that is the one piece of hope that I might be

01:48:34.800 | doing something interesting.

01:48:35.760 | - Wow, that's quite brilliant. Like as a gradient to optimize.

01:48:40.400 | - Yeah.

01:48:41.520 | - Hang out with smart people and confuse them.

01:48:43.360 | - Yeah.

01:48:43.760 | - And the more confusing it is, the more there's something there.

01:48:46.960 | And as long as they're not telling you just a complete idiot and they give you different reasons.

01:48:51.600 | - Yeah.

01:48:52.160 | - And I mean, I'm, you know, if everyone, it's like with assembly theory and people said,

01:48:56.240 | "Oh, it's wrong." And I was like, "Why?" And they're like, "And no one could give me a

01:48:59.520 | consistent reason." They said, "Oh, because it's been done before, or it's just Komagolov,

01:49:03.520 | or it's just that and the other." So I think the thing that I like to do is, and in academia,

01:49:09.440 | it's hard, right? Because people are critical, but I mean, you know, the criticism, I mean,

01:49:16.480 | although I got kind of upset about it earlier, which is kind of silly, but not silly, because

01:49:20.800 | obviously it's hard work being on your own or with a team spatially separated, like during lockdown,

01:49:26.480 | and try to keep everyone on board and have some faith that I've always wanted to have a new idea.

01:49:34.800 | And so, you know, I like a new idea and I want to nurture it as long as possible.

01:49:40.800 | And if someone can give me actionable criticism, that's why I think I was trying to say earlier,

01:49:46.320 | when I was kind of like stuck for words, give me actionable criticism. You know, it's wrong. Okay,

01:49:52.560 | why is it wrong? You say, "Oh, your equation's incorrect for this," or "Your method is wrong."

01:49:57.920 | And so what I try and do is get enough criticism from people to then triangulate and go back. And

01:50:04.880 | I've been very fortunate in my life that I've got great colleagues, great collaborators, funders,

01:50:11.520 | mentors, and people that will take the time to say, "You're wrong because..." And then what I

01:50:17.360 | have to do is integrate the wrongness and go, "Oh, cool. Maybe I can fix that." And I think

01:50:22.640 | criticism is really good. People have a go at me because I'm really critical. But I'm not criticizing

01:50:27.600 | you as a person. I'm just criticizing the idea and trying to make it better and say, "Well,

01:50:33.600 | what about this?" And sometimes I'm kind of, you know, my filters are kind of truncated in some

01:50:40.960 | ways. I'm just like, "That's wrong. That's wrong. That's wrong. I want to do this." And people are

01:50:43.520 | like, "Oh my God, you just told me you destroyed my life's work." I'm like, "Relax. No, I'm just

01:50:48.960 | like, let's make it better." And I think that we don't do that enough because we're either personally

01:50:57.600 | critical, which isn't helpful, or we don't give any criticism at all because we're too scared.

01:51:01.920 | Yeah, I've seen you be pretty aggressively critical, but every time I've seen it, it's

01:51:12.160 | the idea, not the person.

01:51:14.000 | I'm sure I make mistakes on that. I mean, I argue lots with Sarah, and she's kind of shocked. I've

01:51:24.320 | argued with Yasha in the past, and he's like, "You're just making it up." I'm like, "No, not

01:51:30.320 | quite, but kind of." I had a big argument with Sarah about time. She's like, "No, time doesn't

01:51:38.640 | exist." I'm like, "No, no, time does exist." And as she realized that her conception of assembly

01:51:44.320 | theory and my conception of assembly theory were the same thing, necessitated us to abandon the

01:51:50.960 | fact that time is eternal, to actually really fundamentally question how the universe produces

01:51:56.880 | combinatorial novelty. - So time is fundamental for assembly theory? I'm just trying to figure

01:52:03.520 | out where you and Sarah converge. - I think assembly theory is fine in this time right now,

01:52:08.560 | but I think it helps us understand that something interesting is going on. I've been really inspired

01:52:14.800 | by a guy called Nick Gizen. I'm going to butcher his argument, but I love his argument a lot, so

01:52:20.000 | I hope he forgives me if he hears about it. Basically, if you want free will, time has to

01:52:28.000 | be fundamental, and if you want time to be fundamental, you have to give up on platonic

01:52:41.120 | mathematics, and you have to use intuition as mathematics, by the way. Again, I'm going to

01:52:48.080 | butcher this, but basically, Hilbert said that infinite numbers are allowed, and I think it was

01:52:56.640 | Brouwer who said, "No, all numbers are finite." So let's go back a step, because people are going

01:53:04.160 | to say, "Assembly theory seems to explain that large combinatorial space allows you to produce

01:53:13.280 | things like life and technology, and that large combinatorial space is so big, it's not even

01:53:19.360 | accessible to a Sean Carroll, David Deutch multiverse, that physicists saying that all of

01:53:29.200 | the universe already exists in time is probably, provably - that's a strong word - not correct.

01:53:43.120 | That we are going to know that the universe as it stands, the present, the way the present

01:53:48.000 | builds the future, so big, the universe can't ever contain the future. And this is a really

01:53:55.840 | interesting thing. I think Max Tegmark has this mathematical universe, which says, you know,

01:53:59.840 | the universe is kind of like a block universe, and I apologize to Max if I'm getting it wrong,

01:54:04.400 | but people think you can just move - you have the initial conditions, and you can run the universe,

01:54:11.840 | and right to the end, and go backwards and forwards in that universe. That is not correct.

01:54:16.400 | - Yeah, let me load that in. The universe is not big enough to contain the future.

01:54:20.560 | - Yeah, that's why.

01:54:22.160 | - That's another, that's a beautiful way of saying that time is fundamental.

01:54:26.560 | - Yes, and that you can have, and that's what, this is why the law of the excluded middle,

01:54:33.840 | something is true or false, only works in the past.

01:54:40.000 | Is it going to snow in New York next week, or in Austin? You might in Austin say probably not,

01:54:44.960 | in New York you might say yeah. If you go forward to next week and say, did it snow in New York

01:54:50.240 | last week, true or false, you can answer that question. The fact that the law of the excluded

01:54:55.200 | middle cannot apply to the future explains why time is fundamental.

01:54:58.800 | - Well, I mean, that's a good example, intuitive example, but it's possible that we might be able

01:55:05.040 | to predict, you know, whether it's gonna snow if we had perfect information.

01:55:10.080 | - I think we--

01:55:11.520 | - You're saying not.

01:55:12.720 | - Impossible. Impossible. So here's why. I'll make a really quick argument,

01:55:19.360 | and this argument isn't mine, it's Nick's and a few other people.

01:55:22.960 | - Can you explain his view on fundamental, on time being fundamental?

01:55:27.600 | - Yeah, so I'll give my view, which kind of resonates with his, but basically,

01:55:33.520 | it's very simple, actually. He would say that free will, your ability to design and do an experiment

01:55:38.640 | is exercising free will. So he used that thought process. I never really thought about it that way,

01:55:44.160 | and that you actively make decisions. I do think that, I used to think that free will was a kind of

01:55:51.680 | consequence of just selection, but I'm kind of understanding that human free will is something

01:55:57.200 | really interesting, and he very much inspired me. But I think that what Sarah Walker said that

01:56:03.440 | inspired me as well, that these will converge, is that I think that the universe, and the universe

01:56:10.800 | is very big, huge, but actually, the place that is largest in the universe right now,

01:56:18.560 | the largest place in the universe is Earth.

01:56:20.720 | - Yeah, I've seen you say that, and boy does that, that's an interesting one to process.

01:56:28.000 | What do you mean by that, Earth is the biggest place in the universe?

01:56:30.800 | - Because we have this combinatorial scaffolding going all the way back from Luca,

01:56:35.440 | so you've got cells that can self-replicate, and then you go all the way to terraforming the Earth,

01:56:41.120 | you've got all these architectures, the amount of selection that's going on, biological selection,

01:56:45.920 | just to be clear, biological evolution, and then you have multicellularity,

01:56:50.560 | then animals, and abstraction, and with abstraction, there was another kick,

01:56:55.520 | because you can then build architectures, and computers, and cultures, and language,

01:57:01.120 | and these things are the biggest things that exist in the universe, because we can just build

01:57:05.280 | architectures that couldn't naturally arise anywhere, and the further that distance goes in

01:57:09.680 | time, and it's gigantic, and-- - From a complexity perspective.

01:57:16.800 | - Yeah. - Okay, wait a minute, but,

01:57:18.880 | I mean, I know you're being poetic, but how do you know there's not other Earth-like,

01:57:22.880 | like, how do you know, you're basically saying Earth is really special, it's awesome stuff as

01:57:30.480 | far as we look out, there's nothing like it going on, but how do you know there's not nearly infinite

01:57:37.040 | number of places where cool stuff like this is going on?

01:57:39.600 | - I agree, and I would say, I'll say again that Earth is the most gigantic thing

01:57:46.080 | we know in the universe, commentarily, we know. - We know.

01:57:50.000 | - Now, now, I guess, this is just purely a guess, I have no data, but other than hope,

01:57:56.080 | well, maybe not hope, maybe, no, I have some data, that every star in the sky probably has planets,

01:58:04.640 | and life is probably emerging on these planets, but the amount of contingency that is

01:58:11.120 | associated with life is, I think, the commentarial space associated with these planets is so

01:58:17.040 | different, we're never gonna, our causal cones are never gonna overlap, or not easily,

01:58:22.560 | and this is the thing that makes me sad about alien life, why it's why we have to create alien

01:58:26.720 | life in the lab as quickly as possible, because I don't know if we are gonna be able to,

01:58:33.680 | be able to build architectures that will intersect with alien intelligence and architectures.

01:58:42.640 | - Intersect, you don't mean in time or space? - Time and the ability to communicate.

01:58:47.600 | - The ability to communicate. - Yeah, my biggest fear, in a way,

01:58:51.200 | is that life is everywhere, but we become infinitely more lonely because of our

01:58:55.280 | scaffolding in that commentarial space, because it's so big, and--

01:59:00.080 | - So you're saying the constraints created by the environment that led to the factory of

01:59:07.040 | Darwinian evolution are just like this little tiny cone in a nearly infinite commentarial space.

01:59:13.920 | - Exactly. - And so there's

01:59:14.640 | other cones like it, and why can't we communicate with other, just because we can't create it,

01:59:23.040 | doesn't mean we can't appreciate the creation, right, sorry, detect the creation.

01:59:29.440 | - I truly don't know, but it's an excuse for me to ask for people to give me money to make

01:59:35.040 | a planet simulator. - Yeah, right.

01:59:36.560 | - If I can make-- - With a different--

01:59:38.880 | - Like another shameless say, it's like, give me money, I need money.

01:59:42.240 | - This was all a long plug for a planet simulator. - It's like, you know--

01:59:46.400 | - Hey, I'll be the first in line to donate. - My Rick garage has run out of room, you know?

01:59:53.120 | - Yeah. - No, um--

01:59:54.640 | - And this is a planet simulator, you mean like a different kind of planet?

01:59:58.880 | - Yeah. - With different sets of

02:00:00.240 | environments and pressures. - Exactly, if we could basically

02:00:03.200 | recreate the selection before biology, as we know it, that gives rise to a different biology,

02:00:10.080 | we should be able to put the constraints on where I look in the universe. So here's the thing,

02:00:14.400 | here's my dream. My dream is that by creating life in the lab, and based upon constraints we

02:00:21.120 | understand, so let's go for Venus-type life or Earth-type life or something, again, do Earth 2.0,

02:00:26.000 | screw it, let's do Earth 2.0. And Earth 2.0 has a different genetic alphabet, fine, that's fine,

02:00:32.080 | different protein alphabet, fine, have cells and evolution, all that stuff. We will then be able to

02:00:38.880 | say, okay, life is a more general phenomena, selection is more general than what we think is

02:00:45.360 | the chemical constraints on life, and we can point the Jane's Web and other telescopes at other

02:00:50.000 | planets that we are in that zone, we are most likely to concomitantly overlap with, right?

02:00:57.920 | So there are chemistry-- - You're looking for some overlap.

02:01:02.480 | - And then we can then basically shine light on them literally, and look at light coming back,

02:01:08.240 | and apply advanced assembly theory to general theory of language that we will get, and say,

02:01:14.800 | huh, in that signal, it looks random, but there's a copy number. Oh,

02:01:20.320 | this random set of things that shouldn't be, that looks like a true random number generator

02:01:28.000 | has structure as a, not Komagolov, AIT-type structure, but evolutionary structure,

02:01:35.600 | given by assembly theory, and we start to, but I would say that, 'cause I'm a shameless

02:01:40.000 | assembly theorist. - Yeah, it just feels like the cone,

02:01:45.040 | and I might be misusing the word cone here, but the width of the cone is growing faster,

02:01:49.680 | is growing really fast to where eventually all the cones overlap,

02:01:55.440 | even in a very, very, very large combinatorial space.

02:02:03.520 | It just, but then again, if you're saying the universe is also growing very quickly

02:02:10.960 | in terms of possibilities. - I hope that as we build abstractions,

02:02:19.840 | the main, I mean, one idea is that as we go to intelligence, intelligence allows us to look

02:02:27.200 | at the regularities around us in the universe, and that gives us some common grounding to discuss

02:02:34.080 | with aliens, and you might be right, that we will overlap there, even though we have completely

02:02:41.680 | different chemistry, literally completely different chemistry, that we will be at a

02:02:46.960 | past information from one another, but it's not a given, and I have to kind of try and divorce

02:02:56.960 | hope and emotion away from what I can logically justify.

02:03:02.080 | - But it's just hard to intuit a world, a universe, where there's

02:03:05.760 | nearly infinite complexity objects, and they somehow can't detect each other.

02:03:12.480 | - But the universe is expanding, but the nice thing is, I would say, I would look, you see,

02:03:17.200 | I think Carl Sagan did the wrong thing, well, not the wrong thing, he flicked the Voyager probe

02:03:21.440 | around a pale blue dot and said, "Look how big the universe is." I would have done it the other way

02:03:25.200 | around and said, "Look at the Voyager probe that came from the planet Earth, that came from Luca,

02:03:28.720 | look at how big Earth is." - Then it produced that.

02:03:32.640 | - It produced that. - Yeah.

02:03:34.080 | - And that, I think, is completely amazing, and then that should allow people on Earth to think

02:03:38.960 | about, well, probably we should try and get causal chains off Earth onto Mars, onto the Moon,

02:03:46.320 | wherever, whether it's human life or Martian life that we create, it doesn't matter. But I think

02:03:54.880 | this commentarial space tells us something very important about the universe,

02:03:58.080 | and that I realized in assembly theory that the universe is too big to contain itself.

02:04:04.160 | And I think this is, now coming back, and I want to kind of change your mind about time,

02:04:10.640 | 'cause I'm guessing that your time is just a coordinate.

02:04:14.640 | - Yeah. - So I'm gonna change--

02:04:16.880 | - I'm guessing you're one of those, yeah. - I'm gonna change your, one of those,

02:04:20.080 | I'm gonna change your mind in real time, or at least attempt.

02:04:22.000 | - Oh, in real time, there you go. I already got the tattoo,

02:04:25.520 | so this is gonna be embarrassing if you change my mind.

02:04:27.440 | - But you can just add an arrow of time onto it, right?

02:04:30.880 | - Yeah, true, just modify it. - Or erase it a bit.

02:04:33.200 | - Yeah. - So, and the argument that I think

02:04:36.160 | that is really most interesting is, people say the initial conditions specify the future of the

02:04:42.880 | universe. Okay, fine, let's say that's the case for a moment. Now let's go back to Newtonian

02:04:47.920 | mechanics. Now, the uncertainty principle in Newtonian mechanics is this. If I give you the

02:04:56.640 | coordinates of an object moving in space, and the coordinates of another object and they collide

02:05:02.720 | in space, and you know those initial conditions, you should know exactly what's gonna happen.

02:05:07.840 | However, you cannot specify these coordinates to infinite precision.

02:05:14.560 | Now everyone said, "Oh, this is kind of like the chaos theory argument." No, no, it's deeper than

02:05:19.520 | that. Here's a problem with numbers. This is where Hilbert and Brouwer fell out.

02:05:24.000 | To have the coordinates of this object, a given object as it's colliding,

02:05:30.640 | you have to have them to infinite precision. That's what Hilbert says. He says, "No problem,

02:05:34.560 | infinite precision is fine. Let's just take that for granted." But when the object is finite,

02:05:42.080 | and it can't store its own coordinates, what do you do?

02:05:45.040 | So in principle, if a finite object cannot be specified to infinite precision,

02:05:53.040 | in principle, the initial conditions don't apply. - Well, how do you know it can't store its...

02:05:59.920 | - Well, how do you store an infinitely long number in a finite size? - Well,

02:06:09.440 | we're using infinity very loosely here. - No, no, we're using...

02:06:12.480 | - Infinite precision, I mean, not loosely, but... - Very precisely.

02:06:15.280 | - So you think infinite precision is required? - Well, let's take the object. Let's say the

02:06:19.600 | object is a golf ball. Golf ball is a few centimeters in diameter. We can work out

02:06:25.680 | how many atoms are on the golf ball. And let's say we can store numbers down to atomic dislocations.

02:06:31.200 | So we can work out how many atoms there are in the golf ball, and we can store the coordinates

02:06:36.960 | that's in that golf ball down to that number. But beyond that, we can't. Let's make the golf ball

02:06:41.040 | smaller. And this is where I think that we think that we get randomness in quantum mechanics.

02:06:48.080 | And some people say, "You can't get randomness in quantum mechanics, it's deterministic."

02:06:50.800 | But aha, this is where we realize that classical mechanics and quantum mechanics suffer from the

02:06:56.160 | same uncertainty principle. And that is the inability to specify the initial conditions

02:07:04.400 | to a precise enough degree to give you determinism. The universe is intrinsically

02:07:11.120 | too big, and that's why time exists. It's non-deterministic. Looking back into the past,

02:07:18.080 | you can use logical arguments because you can say, "Was it true or false?" You already know.

02:07:24.240 | But the fact we are unable to predict the future with the precision

02:07:29.760 | is not evidence of lack of knowledge, it's evidence the universe is generating new things.

02:07:36.480 | - Okay, so to you, first of all, quantum mechanics, you can just say statistically what's going to

02:07:42.720 | happen when two golf balls hit each other. - Statistically, but sure, I can say statistically

02:07:47.760 | what's going to happen, but then when they do happen, and you keep nesting it together,

02:07:53.440 | you can't... I mean, it goes almost back to look at... Let's think about entropy in the universe.

02:07:59.200 | So how do we understand entropy change? Well, we could do the... Look at our process. We can

02:08:06.880 | use the Ergodic hypothesis. We can also have the counterfactuals, where we have all the different

02:08:18.000 | states, and we can even put that in the multiverse, right? But both those are kind of...

02:08:23.360 | They're non-physical. The multiverse kind of collapses back to the same problem about the

02:08:31.520 | precision. So if you accept you don't have to have true and false going forward into the future,

02:08:42.240 | the real numbers are real. They're observables. - We're trying to see exactly where time being

02:08:50.000 | fundamental sneaks in, in this difference between the golf ball can't contain its own position

02:08:57.840 | perfectly, precisely, how that leads to time needing to be fundamental.

02:09:05.040 | - Let me... Do you believe or do you accept you have free will?

02:09:10.640 | - Yeah, I think at this moment in time, I believe that I have free will.

02:09:17.520 | - So then you have to believe that time is fundamental.

02:09:21.200 | - I understand that's a statement you've made. - Well, no, that we can logically follow,

02:09:26.640 | because if you don't have free will... So if you're in a universe that has no time,

02:09:32.160 | the universe is deterministic. If it's deterministic, then you have no free will.

02:09:36.160 | - I think the space of how much we don't know is so vast that saying the universe

02:09:43.120 | is deterministic and from that jumping, there's no free will is just too difficult of a leap.

02:09:48.000 | - No, I logically follow. No, no, I don't disagree. I'm not saying any... I mean, it's deep

02:09:53.440 | and it's important. All I'm saying, and it's actually different to what I've said before,

02:10:00.800 | is that if you don't require platonistic mathematics and accept that non-determinism

02:10:08.320 | is how the universe looks, and that gives us our creativity and the way the universe is getting

02:10:14.000 | novelty, it's kind of really deeply important in assembly theory, because assembly theory starts

02:10:18.640 | to actually give you a mechanism why you go from boring time, which is basically initial conditions

02:10:24.400 | specify everything, to a mismatch in creative time. And I hope we'll do experiments. I think

02:10:29.600 | it's really important to... I would love to do an experiment that proves that time is fundamental

02:10:35.360 | and the universe is generating novelty. I don't know all the features of that experiment yet,

02:10:41.840 | but by having these conversations openly and getting people to think about the problems in

02:10:48.640 | a new way, better people, more intelligent people with good mathematical backgrounds can say, "Oh,

02:10:54.800 | hey, I've got an idea." I would love to do an experiment that shows that the universe... I mean,

02:11:01.360 | universe is too big for itself going forward in time. And this is why I really hate the idea of

02:11:09.040 | the Boltzmann brain. The Boltzmann brain makes me super... Everyone's having a free lunch. It's like

02:11:14.880 | let's break all the laws of physics. So, a Boltzmann brain is this idea that in a long

02:11:20.480 | enough universe, a brain will just emerge in the universe as conscious. And that neglects the causal

02:11:25.840 | chain of evolution that's required to produce that brain. And this is where the computational

02:11:30.640 | argument really falls down because the computation is to say, "I can calculate the probability of a

02:11:34.480 | Boltzmann brain." And they'll give you a probability, but I can calculate the probability of a Boltzmann

02:11:38.800 | brain, zero. - Just because the space of possibility is so large? - Yeah. It's like when we start

02:11:44.880 | fooling ourselves with numbers that we can't actually measure and we can't ever conceive of,

02:11:50.000 | I think it doesn't give us a good explanation. And I've become... I want to explain why life

02:11:59.360 | is in the universe. I think life is actually a novelty miner. I mean, life basically mines novelty

02:12:05.920 | almost from the future and actualizes it in the present. - Okay. Life is a novelty miner

02:12:16.080 | from the future that is actualized in the present. - Yeah.

02:12:20.240 | I think so. - Novelty miner. First of all, novelty. What's the origin of novelty

02:12:30.560 | when you go from boring time to creative time? Where is that? Is it as simple as randomness,

02:12:36.800 | like you're referring to? - I'm really struggling with randomness because I had a really good

02:12:42.480 | argument with Jascha Bach about randomness. And he said, "Randomness doesn't give you free will.

02:12:47.360 | That's insane because you'd just be random." And I think he's right at that level, but I don't

02:12:53.120 | think he is right on another level. And it's not about randomness. It's about constrained...

02:13:00.720 | I'm going to sound like... Constrained opportunity. I'm making this up as I go along,

02:13:05.680 | so I'm making this up. Constrained opportunity. So, what I mean is like... So, you have to have...

02:13:11.600 | So, the novelty... What is novelty? This is why I think it's a funny thing. If you ever want to

02:13:18.960 | discuss AI, why I think everyone's kind of gone AI mad is that they're misunderstanding

02:13:24.160 | novelty. But let's think about novelty. Yes, what is novelty? So, I think novelty is a genuinely

02:13:30.720 | new configuration that is not predicted by the past, right? And that you discover in the present,

02:13:38.800 | right? And that is truly different, right? Now, everyone says that... Some people say

02:13:44.320 | that novelty doesn't exist. It's always with precedent. I want to do experiments that show

02:13:49.680 | that that is not the case. And it goes back to a question you asked me a few moments ago, which is,

02:13:55.440 | where is the factory? Right? Because I think the same mechanism that gives us a factory

02:14:01.440 | gives us novelty. And I think that that is why I'm so deeply hung up on time. I mean,

02:14:07.200 | of course I'm wrong, but how wrong? And I think that life opens up that commentarial space in a

02:14:16.000 | way that our current laws of physics, although as contrived in a deterministic initial condition

02:14:25.280 | universe, even with the get out of the multiverse, David Deutch style, which I love by the way,

02:14:30.400 | but I don't think is correct. But it's really beautiful. The David Deutch's conception of the

02:14:40.080 | multiverse is kind of like given. But I think that the problem with wave particle duality and

02:14:47.280 | quantum mechanics is not about the multiverse. It's about understanding how determined the past

02:14:54.640 | is. Well, I don't just think that actually this is a discussion I was having with Sarah about that,

02:15:00.400 | right? She was like, "Oh, I think we'd have been debating this for a long time now about how do we

02:15:07.280 | reconcile novelty, determinism, indeterminism." - So just to clarify, both you and Sarah think

02:15:15.920 | the universe is not deterministic. - I won't speak for Sarah, but I think that the universe

02:15:23.760 | is deterministic looking back in the past, but undetermined going forward in the future. So I'm

02:15:34.480 | kind of having my cake and eating it here. This is because I fundamentally don't understand

02:15:38.400 | randomness, right? As Yasha told me or other people told me. But if I adopt a new view now,

02:15:43.840 | which the new view is the universe is just non-deterministic, but I'd like to refine that

02:15:49.280 | and say the universe appears deterministic going back in the past, but it's undetermined going

02:15:56.000 | forward in the future. So how can we have a universe that has deterministically looking

02:16:02.160 | rules that's non-determined going into the future? It's this breakdown in precision in

02:16:06.720 | the initial conditions. And we have to just stop using initial conditions and start looking at

02:16:11.520 | trajectories and how the combinatorial space behaves in expanding universe in time and space.

02:16:21.600 | And assembly theory helps us quantify the transition to biology, and biology appears

02:16:27.120 | to be novelty mining because it's making crazy stuff. That we are unique to earth, right?

02:16:34.240 | There are objects on earth that are unique to earth that will not be found anywhere else

02:16:38.800 | because you can do the combinatorial math. - What was that statement you made about life

02:16:43.840 | is novelty mining from the future? What's the little element of time that you're introducing?

02:16:51.280 | - So what I'm kind of meaning is because the future is bigger than the present,

02:16:54.960 | in a deterministic universe, how do the states go from one to another? I mean,

02:17:01.360 | there's a mismatch, right? So that must mean that you have a little bit of indeterminism,

02:17:06.320 | whether that's randomness or something else. I don't understand. I want to do experiments to

02:17:10.880 | formulate a theory to refine that as we go forward that might help us explain that. And I think that's

02:17:16.560 | why I'm so determined to try and crack the non-life-to-life transition, looking at networks

02:17:24.960 | and molecules, and that might help us think about it, the mechanism. But certainly the future is

02:17:29.760 | bigger than the past, in my conception of the universe and some conception of the universe.

02:17:34.560 | - By the way, that's not obvious, right? The future being bigger than the past.

02:17:42.320 | Well, that's one statement, and the statement that the universe is not big enough to contain

02:17:46.560 | the future is another statement. - Yeah. Yeah, yeah, yeah.

02:17:49.840 | - That one is a big one. That one's a really big one.

02:17:52.640 | - I think so. But I think it's entirely... Because look, we have the second law. And right now,

02:17:59.600 | I mean, we don't need the second law if the future's bigger than the past. It follows naturally.

02:18:05.680 | So why are we retrofitting all these sticking plasters onto our reality to hold onto a

02:18:12.400 | timeless universe? - Yeah, but that's because

02:18:15.040 | it's kind of difficult to imagine the universe that can't contain the future.

02:18:21.120 | - But isn't that really exciting? - It's very exciting, but it's hard.

02:18:26.960 | I mean, we're humans on Earth, and we have a very kind of four-dimensional conception of the world.

02:18:35.600 | Of 3D plus time. It's just hard to intuit a world where... What does that even mean?

02:18:41.760 | A universe that can't contain the future. - Yeah, it's kind of crazy, but obvious.

02:18:50.240 | - It's weird. I mean, I suppose it sounds obvious, yeah, if it's true.

02:18:53.920 | - But the nice thing is you can... So the reason why assembly theory turned me onto that

02:19:00.160 | was that... Let's just start in the present and look at all the complex molecules and go backwards

02:19:05.680 | in time and understand how evolutionary processes gave rise to them. It's not at all obvious that

02:19:15.280 | taxol, which is one of the most complex natural products produced by biology, was going to be

02:19:21.360 | invented by biology. It's an accident. You know, taxol is unique to Earth. There's no taxol

02:19:26.960 | elsewhere in the universe. And taxol was not decided by the initial conditions. It was decided

02:19:34.320 | by this kind of interplay between the... So the past simply is embedded in the present. It gives

02:19:42.720 | some features, but why the past doesn't map to the future one-to-one is because the universe is too

02:19:49.120 | big to contain itself. That gives space for creativity, novelty, and some things which are

02:19:56.160 | unpredictable. - Okay, so given that you're disrespecting the power of the initial conditions,

02:20:02.720 | let me ask you about... So how do you explain that cellular automata are able to produce such

02:20:07.680 | incredible complexity given just basic rules and basic initial conditions? - I think that this

02:20:14.320 | falls into the Brouwer-Hilbert trap. So how do you get a cellular automata producing a complexity?

02:20:23.040 | You have a computer, you generate a display, and you map the change of that in time. There are some

02:20:29.680 | CAs that repeat, like functions. It's fascinating to me that for pi, there is a formula where you

02:20:36.080 | can go to the millionth decimal place of pi and read out the number without having to go there.

02:20:42.480 | But there are some numbers where you can't do that. You have to just crank through. Whether it's

02:20:48.400 | Wolframian computational irreducibility or some other thing, that doesn't matter. But these CAs,

02:20:54.800 | that complexity, is that just complexity or a number that is basically you're mining that

02:21:02.960 | number in time? Is that just a display screen for that number, that function? - Well, can't you say

02:21:09.600 | the same thing about the complexity on Earth then? - No, because the complexity on Earth

02:21:13.520 | has a copy number and an assembly index associated with it. That CA is just a number running.

02:21:18.960 | - You don't think it has a copy number? Wait a minute. - Well, it does in the human, where we're

02:21:25.360 | looking at humans producing different rules, but then it's nested on selection. So those CAs are

02:21:29.680 | produced by selection. I mean, the CA is such a fascinating pseudo-complexity generator. What I

02:21:37.760 | would love to do is understand, quantify the degree of surprise in a CA and run it long enough.

02:21:43.680 | But what I guess that means is we have to instantiate, we have to have a number

02:21:47.840 | of experiments where we're generating different rules and running them in time steps. But,

02:21:51.840 | ah, got it. CAs are mining novelty in the future by iteration, right? And you're like, oh, that's

02:22:00.160 | great, that's great. You didn't predict it. Some rules you can predict what's going to happen,

02:22:04.080 | other rules you can't. So for me, if anything, CAs are evidence that the universe is too big

02:22:09.440 | to contain itself, because otherwise you'd know what the rules are going to do forever more.

02:22:13.440 | - Right. I guess you were saying that the physicist saying that all you need is the

02:22:19.600 | initial conditions and the rules of physics is somehow missing the bigger picture.

02:22:25.120 | And if you look at CAs, all you need is the initial condition and the rules and then run the thing.

02:22:32.960 | - You need three things. You need the initial conditions, you need the rules,

02:22:40.400 | and you need time, iteration to mine it out. Without the coordinate, you can't get it out.

02:22:44.640 | - Sure. And that's that to use fundamentally.

02:22:47.280 | - And you can't predict it from initial conditions. If you could, then it'd be fine.

02:22:51.440 | - And that time is the foundation of, this is the history, the memory of each of the things

02:22:58.560 | it created. It has to have that memory of all the things that led up to it.

02:23:04.400 | - I think it's, yeah, you have to have the resource. Because time is a fundamental

02:23:09.280 | resource. And yeah, I'm becoming, I think I had a major

02:23:16.000 | epiphany about randomness, but I keep doing that every two days and then it goes away again,

02:23:23.600 | it's random. - You're a time fundamentalist.

02:23:25.760 | - You should be as well. If you believe in free will, the only conclusion is time is fundamental,

02:23:33.040 | otherwise you cannot have free will, it logically follows.

02:23:35.600 | - Well, the foundation of my belief in free will is observation driven. I think if you use logic,

02:23:52.480 | logically it seems like the universe is deterministic.

02:23:55.120 | - Looking backwards in time, and that's correct, the universe is.

02:23:58.160 | - And then everything else is a kind of leap, it requires a leap.

02:24:03.520 | - I mean, I think that, it's kind of, this is, I think machine learning is going to provide a big,

02:24:13.600 | a chunk of that, right? Because it helped us explain this. So the way I'd say, if you take--

02:24:18.480 | - That's interesting, why? - Well, let's just, my favorite one is,

02:24:25.440 | because I'm, the AI doomers are driving me mad, and the fact that we don't have any intelligence

02:24:30.880 | yet, I call AI autonomous informatics, just to make people grumpy.

02:24:34.640 | - Yeah, you're saying we're quite far away from AGI.

02:24:38.240 | - I think that we have no conception of intelligence, and I think that we don't

02:24:45.120 | understand how the human brain does what it does. I think that we are, neuroscience is making great

02:24:49.840 | advances, but I think that we have no idea about AGI. So I am a technological, I guess, optimist.

02:24:57.680 | I believe we should do everything, the whole regulation of AI is nonsensical. I mean,

02:25:02.240 | why would you regulate Excel, other than the fact that Clippy should come back, and I love Excel 97,

02:25:06.800 | 'cause we can play, we can do the flight simulator. - I'm sorry, in Excel?

02:25:12.480 | - Yeah, have you not played the flight simulator in 99? - In Excel 97?

02:25:16.000 | - Yeah, yeah, yeah. - What does that look like?

02:25:18.640 | - It's like wireframe, very basic, but basically I think it's X zero, Y zero,

02:25:25.360 | shift, and it opens up, and you can play the flight simulator.

02:25:28.320 | - Oh, wow. Wait, wait, wait, is it using Excel? - Excel, Excel 97.

02:25:33.280 | - Okay. - I resurrected it the other day,

02:25:35.360 | and saw Clippy again for the first time in a long time.

02:25:37.680 | - Well, Clippy is definitely coming back, but you're saying we don't have a great understanding

02:25:44.400 | of what is intelligence, what is the intelligence-- - I am very frustrated--

02:25:48.400 | - Underpinning the human mind. - I'm very frustrated by the way that

02:25:52.320 | we're AI-dooming right now, and people are bestowing some kind of magic. Now,

02:25:58.000 | let's go back a bit. So you said about AGI, are we far away from AGI? Yes, I do not think we're

02:26:03.760 | going to get to AGI anytime soon. I've seen no evidence of it, and the AI-doom scenario is

02:26:09.440 | nonsensical in the extreme, and the reason why I think it's nonsensical, but it's not--

02:26:14.560 | I don't think there isn't things we should do and be very worried about, right? I mean,

02:26:20.560 | there are things we need to worry about right now, what AI are doing, whether it's fake data,

02:26:25.120 | fake users, right? I want authentic people or authentic data. I don't want everything to be

02:26:31.040 | faked, and I think it's a really big problem, and I absolutely want to go on the record to say I

02:26:34.960 | really worry about that. What I'm not worried about is that some fictitious entity is going

02:26:39.840 | to turn us all to paperclips, or detonate nuclear bombs, I don't know, maybe, I don't know,

02:26:47.840 | anything you can't think of. Why is this-- and I'll take a very simple series of logical arguments,

02:26:55.040 | and the AI-doomers have not had the correct epistemology. They do not understand what

02:27:09.120 | knowledge is, and until we understand what knowledge is, they're not going to get anywhere,

02:27:13.760 | because they're applying things falsely. So let me give you a very simple argument. People talk

02:27:18.480 | about the probability, p-doom AI. We can work out the probability of an asteroid hitting the planet,

02:27:26.000 | why? Because it's happened before. We know the mechanism, we know that there's a gravity well,

02:27:30.320 | or that space-time is bent and stuff falls in. We don't know the probability of AGI because we have

02:27:36.080 | no mechanism. So let me give you another one, which is like, I'm really worried about AG.

02:27:40.720 | What's AG? AG is anti-gravity. One day we could wake up and anti-gravity is discovered, we're all

02:27:47.680 | going to die, the atmosphere is going to float away, we're going to float away, we're all doomed.

02:27:52.400 | What is the probability of AG? We don't know because there's no mechanism for AG. Do we

02:27:57.600 | worry about it? No. And I don't understand the current reason for certain people in certain

02:28:09.520 | areas to be generating this nonsense. I think they're not doing it maliciously. I think we're

02:28:14.720 | observing the emergence of new religions, how religions come because religions are about some

02:28:19.680 | controls. You've got the optimist saying AI is going to cure us all and AI is going to kill us

02:28:24.160 | all. What's the reality? Well, we don't have AI, we have really powerful machine learning tools

02:28:29.520 | and they will allow us to do interesting things and we need to be careful about

02:28:32.880 | how we use those tools in terms of manipulating human beings and faking stuff, right?

02:28:38.640 | Right. Well, let me try to sort of steel man the AI Doomer's argument. Actually, I don't know,

02:28:45.200 | are AI Doomers in the Yudkowsky camp saying it's definitely going to kill us? Because there's a

02:28:51.360 | spectrum. 95% I think is the limit. 95% plus? No, not plus. I don't know, I was seeing on Twitter

02:28:57.840 | today various things but I think Yudkowsky is at 95%. But to belong to the AI Doomer club,

02:29:04.160 | is there a threshold? I don't know what the membership is. Maybe. And what are the fees?

02:29:08.160 | I think Scott Aronson, I was quite surprised, I saw this online so it could be wrong,

02:29:15.120 | so sorry if it's wrong, says 2%. But the thing is, if someone said there's a 2% chance you're

02:29:21.760 | going to die going into the lift, would you go into the lift? In the elevator for the American

02:29:26.400 | English speaking audience. Well, no, not for the elevator. So I would say anyone higher than 2%,

02:29:33.920 | I mean, I think there's a 0% chance of AGI Doom. Zero.

02:29:37.520 | Just to push back on the argument where the N of 0 on the AGI, we can see on Earth that there's

02:29:45.280 | increasing levels of intelligence of organisms. We can see what humans with extra intelligence

02:29:51.200 | were able to do to the other species. So that is a lot of samples of data, what a delta in

02:30:02.240 | intelligence gives you. When you have an increase in intelligence, how you're able to dominate

02:30:06.880 | a species on Earth. And so the idea there is that if you have a being that's 10x smarter than humans,

02:30:15.920 | we're not gonna be able to predict what that's going to, what that being is gonna be able to do,

02:30:24.160 | especially if it has the power to hurt humans. Which you can imagine a lot of trajectories in

02:30:30.160 | which the more benefit AI systems give, the more control we give to those AI systems over

02:30:36.640 | our power grid, over our nuclear weapons or weapons of any sort. And then it's hard to know

02:30:44.560 | what a ultra intelligence system would be able to do in that case. You don't find that convincing.

02:30:49.440 | I think this is, I would fail that argument 100%. Here's a number of reasons to fail it on.

02:30:54.000 | First of all, we don't know where the intention comes from. The problem is that people think

02:31:00.080 | they keep, I've been watching all the hucksters online with the prompt engineering and all this

02:31:04.560 | stuff. When I talk to a typical AI computer scientist, they keep talking about the AI

02:31:13.120 | as having some kind of decision making ability. That is a category error. The decision making

02:31:18.720 | ability comes from human beings. We have no understanding of how humans make decision.

02:31:23.120 | We've just been discussing free will for the last half an hour, right? We don't even know what that

02:31:27.440 | is. So the intention, I totally agree with you. People who intend to do bad things can do bad

02:31:34.160 | things and we should not let that risk go. That's totally here and now. I do not want that to happen

02:31:40.880 | and I'm happy to be regulated to make sure that systems I generate, whether they're like

02:31:45.520 | computer systems or, you know, I'm working on a new project called Chem Machina.

02:31:51.600 | - Nice, well done.

02:31:54.240 | - Yeah, yeah, which is basically a...

02:31:56.320 | - For people who don't understand the point, the X Machina is a great film about,

02:32:03.360 | I guess, AGI embodied and Chem is the chemistry version of that.

02:32:07.200 | - And I only know one way to embody intelligence, that's in chemistry and human brains.

02:32:11.280 | So category error number one is agents, they have agency. Category error number two is saying that,

02:32:17.120 | assuming that anything we make is going to be more intelligent. Now you didn't say super

02:32:22.080 | intelligent. I'll put the words into our mouths here, super intelligent. That, I think that there

02:32:28.640 | is no reason to expect that we are going to make systems that are more intelligent, more capable.

02:32:37.360 | You know, when people play chess computers, they don't expect to win now, right? They just, the

02:32:42.400 | chess computer is very good at chess. That doesn't mean it's super intelligent. So I think that

02:32:48.240 | super intelligence, I mean, I think even Nick Bostrom is pulling back on this now,

02:32:52.560 | because he invented this. So I see this a lot. When did this first happen? Eric Drexler,

02:32:58.000 | nanotechnology, atomically precise machines. He came up with a world where we had these

02:33:02.400 | atom cogs everywhere, they were going to make self-replicating nanobots. Not possible, why?

02:33:07.760 | Because there's no resources to build these self-replicating nanobots. You can't get the

02:33:11.760 | precision, it doesn't work. It was a major category error in taking engineering principles down to the

02:33:17.600 | molecular level. The only functioning molecular technology we know, sorry, the only functioning

02:33:23.120 | nanomolecular technology we know, produced by evolution. There. So now let's go forward to

02:33:28.080 | AGI. What is AGI? We don't know. It's super, it can do this, or humans can't think. That,

02:33:34.080 | I would argue, the only AGI's that exist in the universe are produced by evolution.

02:33:39.920 | And sure, we may be able to make our working memory better, we might be able to do more things.

02:33:46.080 | The human brain is the most compact computing unit in the universe. It uses 20 watts,

02:33:51.280 | it uses a really limited volume, it's not like a chat GPT cluster which has to have thousands of

02:33:57.680 | watts, a model that's generated and has to be corrected by human beings. You are autonomous

02:34:02.400 | and embodied intelligence. So I think that there are so many levels that we're missing out. We've

02:34:08.400 | just kind of went, oh, we've discovered fire, oh gosh, the planet's just going to burn one day,

02:34:14.000 | randomly. I mean, I just don't understand that leap. There are bigger problems we need to worry

02:34:18.800 | about. So what is the motivation? Why are these people, let's assume they're earnest,

02:34:24.400 | have this conviction? Well, I think it's kind of, they're making leaps that, they're trapped in a

02:34:32.160 | virtual reality that isn't reality. - Well, I mean, I could continue a set of arguments here,

02:34:37.440 | but also it is true that ideologies that fear monger are dangerous because you can then use it

02:34:47.360 | to control, to regulate in a way that halts progress, to control people, to cancel people,

02:34:57.280 | all that kind of stuff. So you have to be careful because reason ultimately wins, right?

02:35:03.200 | But there is a lot of concerns with superintelligent systems, very capable systems.

02:35:08.080 | I think when you hear the word superintelligent, you're hearing it's smarter than humans in every

02:35:15.760 | way that humans are smart. But the paperclip manufacturing system doesn't need to be smart

02:35:25.600 | in every way. It just needs to be smart in a set of specific ways. And the more

02:35:31.200 | capable the AI systems become, the more you could see us giving them control over, like I said,

02:35:36.400 | our power grid, a lot of aspects of human life. And that means they will be able to do more and

02:35:41.600 | more damage when there's unintended consequences that come to life. - I think that that's right,

02:35:48.000 | that the unintended consequences we have to think about, and that I fully agree with. But let's go

02:35:53.920 | back a bit. Sentient, I mean, again, I'm far away from my comfort zone and all this stuff,

02:35:59.120 | but hey, let's talk about it because I'll give myself a qualification. - Yeah, we're both

02:36:03.600 | qualified and sentient, I think, as much as anyone else. - I think the paperclip scenario is just such

02:36:09.200 | a poor one because let's think about how that would happen. And also let's think about, we are

02:36:13.840 | being so unrealistic about how much of the Earth's surface we have commandeered. For paperclip

02:36:23.360 | manufacturing to really happen, I mean, do the math. It's like, it's not going to happen. There's

02:36:28.560 | not enough energy, there's not enough resource, where is it all going to come from? I think that

02:36:33.040 | what happens in evolution is really, why is a killer virus not killed all life on Earth? Well,

02:36:41.440 | what happens is, sure, super killer viruses that kill the ribosome have emerged, but you know what

02:36:46.240 | happens? They nuke a small space because they can't propagate, they will die. So there's this

02:36:52.000 | interplay between evolution and propagation, right, and death. And so- - In evolution. You don't think

02:36:58.000 | it's possible to engineer, for example, sorry to interrupt, but like a perfect virus that's deadly

02:37:03.360 | enough? - No. Nonsensical. I think that just wouldn't, again, it wouldn't work because it's

02:37:08.160 | too deadly. It would just kill the radius and not replicate it. - Yeah. I mean, you don't think it's

02:37:12.880 | possible to get a- - I mean, if you were super, I mean, if you were- - Not kill all of life on Earth,

02:37:21.280 | but kill all humans. There's not many of us. There's only like eight billion. There's so much

02:37:27.360 | more ants. - I mean, I don't- - So many more ants. And they're pretty smart. - I think the nice thing

02:37:34.560 | about where we are, I would love for the AI crowd to take a leaf out of the book of the bio warfare,

02:37:42.480 | chemical warfare crowd. I mean, not love, 'cause actually people have been killed with chemical

02:37:48.800 | weapons in the First and Second World War, and bioweapons have been made, and we can argue about

02:37:53.760 | COVID-19 and all this stuff. Let's not go there just now. But I think there is a consensus that

02:37:58.320 | some certain things are bad and we shouldn't do them, right? And sure, it would be possible for

02:38:04.560 | a bad actor to engineer something bad, but the damage would be, we would see it coming,

02:38:11.760 | and we would be able to do something about it. Now, I guess what I'm trying to say is

02:38:21.360 | when people talk about doom and they just, when you ask them for the mechanism, they just say,

02:38:26.000 | they just make something up. I mean, in this case, I'm with Yann LeCun. I think he put out a very

02:38:32.880 | good point about trying to regulate jet engines before we've even invented them. And I think

02:38:38.160 | that's what I'm saying. I'm not saying we should, I just don't understand why these guys are going

02:38:43.200 | around literally making stuff up about us all dying, when basically we need to actually really

02:38:49.760 | focus on. Now, let's say there's some actors are earnest, right? Let's say Yudakowsky is being

02:38:56.240 | earnest, right? And he really cares, but he loves it. He goes, and then you're all going to die.

02:39:01.600 | It's like, why don't we try and do the same thing and say, you could do this, and then you're going

02:39:04.960 | to be happy forever after. - Well, I think there's several things to say there. One, I think there is

02:39:11.600 | a role in society for people that say we're all going to die, because I think it filters through

02:39:17.760 | as a message, as a viral message, that gives us the proper amount of concern. Meaning not the,

02:39:24.800 | it's not 95%, but when you say 95% and it filters through society, it'll give an average of like a

02:39:32.800 | 0.03%, an average. So it's nice to have people that are like, we're all going to die, then we'll

02:39:39.600 | have a proper concern. Like, for example, I do believe we're not properly concerned about the

02:39:44.880 | threat of nuclear weapons currently. It just seems like people have forgotten that that's a thing,

02:39:51.600 | and there's a war in Ukraine with nuclear power involved, there's nuclear power throughout the

02:39:57.600 | world, and it just feels like we're on the brink of a potential world war to a percentage that I

02:40:03.680 | don't think people are properly calibrating in their head. We're all thinking it's a Twitter

02:40:08.960 | battle as opposed to actual threat. So it's nice to have that kind of level of concern. But to me,

02:40:16.080 | when I hear AI doers, what I'm imagining is with unintended consequences, a potential situation

02:40:23.120 | where, let's say, 5% of the world suffers deeply because of a mistake made of unintended consequences.

02:40:33.120 | I don't imagine the entirety of human civilization dying, but there could be a lot of suffering if

02:40:37.920 | this is done poorly. - I understand that, and I guess I'm involved in the whole hype cycle.

02:40:44.160 | I would like us to... I don't want us to... So what's happening right now is there seems to be...

02:40:51.280 | So let's say, having some people saying AI doom is a worry, fine, let's give them that.

02:40:58.800 | But what seems to be happening is there seems to be people who don't think AI is doing that,

02:41:03.680 | trying to use that to control regulation and to push people to regulate, which stops humans

02:41:10.480 | generating knowledge. And I am an advocate for generating as much knowledge as possible.

02:41:14.800 | When it comes to nuclear weapons, I grew up in the '70s and '80s where the nuclear doom,

02:41:20.960 | a lot of adults really had existential threat, almost as bad as now with AI doom. They were

02:41:27.120 | really worried, right? There were some great... Well, not great. There were some horrific

02:41:30.800 | documentaries. I think there's one called FREDS that was generated in the UK, which was like...

02:41:36.480 | It was terrible. It was like so scary. And I think that the correct thing to do is obviously get rid

02:41:44.960 | of nuclear weapons, but let's think about unintended consequences. We've got rid of...

02:41:48.800 | We got rid of all the sulfur particles in the atmosphere, right? All the soot. And what's

02:41:55.520 | happened in the last couple of years is global warming has accelerated because we've cleaned

02:41:58.560 | up the atmosphere too much. - Sure. I mean, the same thing if you get rid of nuclear weapons.

02:42:04.960 | - Exactly. That's my point. So what we could do is if we actually started to put the AI in charge,

02:42:11.440 | which is I really like an AI, to be in charge of all world politics. And this sounds ridiculous

02:42:16.720 | for a second, hang on. But if we could all agree on the... - The AI doomers just woke up.

02:42:20.560 | - Yeah, yeah, yeah. - In that statement.

02:42:22.160 | - But I really don't like politicians who are basically just looking at local sampling. But

02:42:26.480 | if you could say globally, look, here's some game theory here. What is the minimum number

02:42:30.960 | of nuclear weapons we need to distribute around the world to everybody to basically reduce war to

02:42:39.040 | zero? - I mean, just this thought

02:42:40.800 | experiment of the United States and China and Russia, major nuclear powers get together and say,

02:42:47.360 | "All right, we're going to distribute nuclear weapons to every single nation on earth."

02:42:54.960 | - Oh boy. I mean, that has a probably greater than 50% chance of eliminating major military

02:43:04.080 | conflict. - Yeah.

02:43:05.200 | - Yeah, but it's not 100%. - But I don't think anyone will use them

02:43:09.120 | because I think... And look, what you've got to try and do is to qualify for these nuclear weapons,

02:43:15.280 | this is a great idea. The game theorists should do this, right? I think the question is this. I

02:43:22.000 | really buy your question, we have too many nukes, just from a feeling point of view that we've got

02:43:26.960 | too many of them. So let's reduce the number, but not get rid of them because we'll have too much

02:43:30.400 | conventional warfare. So then, what is the minimum number of nuclear weapons we can distribute around

02:43:35.600 | to remove... Humans hurting each other is something we should stop doing. It's not out

02:43:43.040 | with our conceptual capability. But right now, what about certain nations that are being

02:43:50.080 | exploited for their natural resources in the future for a short-term gain because we don't

02:43:54.800 | want to generate knowledge. And so if everybody had an equal doomsday switch, I predict the quality

02:44:02.080 | of life of the average human will go up faster. I am an optimist and I believe that humanity is

02:44:07.600 | going to get better and better and better, that we're going to eliminate more problems. But I

02:44:12.720 | think, yeah, let's... - But the probability of a bad actor

02:44:17.280 | of one of the nations setting off a nuclear weapon,

02:44:20.560 | I mean, you have to integrate that into the... - But we distribute the nukes-like population,

02:44:28.720 | right? We give... What we do is we... But anyway, let's just go there. So if a small

02:44:36.400 | nation with a couple of nukes uses one because they're a bit bored or annoyed,

02:44:39.520 | the likelihood that they are going to be pummeled out of existence immediately is 100%. And yet,

02:44:46.640 | they've only nuked one other city. I know this is crazy and I apologize for...

02:44:50.960 | - Well, no, no. I think it's, just to be clear, we're just having a thought experiment that's

02:44:55.120 | interesting, but there's terrorist organizations that would take that, would take that trade.

02:45:02.400 | - Yeah, I mean, look, I'm... - And we have to ask ourselves

02:45:05.600 | a question of how many... Which percentage of humans would be suicide bombers, essentially,

02:45:11.680 | where they would sacrifice their own life because they hate another group of people? And that,

02:45:18.880 | I believe it's a very small fraction, but is it large enough if you give out nuclear weapons?

02:45:24.560 | - I can predict a future where we take all nuclear material and we burn it for energy,

02:45:28.720 | right? Because we're getting there. And the other thing you could do is say, look,

02:45:31.360 | there's a gap. So if we got all the countries to sign up to the virtual nuclear agreement where

02:45:36.400 | we all exist, we have a simulation where we can nuke each other in the simulation. And the

02:45:40.720 | economic consequences are catastrophic. - Sure. In the simulation. I love it. It's

02:45:45.600 | not going to kill all humans. It's just going to have economic consequences.

02:45:48.480 | - Yeah, yeah. I don't know. I just made it up. It seems like it's all I do.

02:45:51.680 | - No, it's interesting. But it's interesting whether that would have as much power in human

02:45:56.080 | psychology as actual physical nuclear explosions. - I think so.

02:45:59.600 | - It's possible, but people don't take economic consequences as seriously, I think, as

02:46:05.120 | actual nuclear weapons. - I think they do in Argentina,

02:46:08.800 | and they do in Somalia, and they do in a lot of these places where... No, I think this is a great

02:46:14.720 | idea. I'm a strong advocate now for... So what have we come up with? Burning all the nuclear

02:46:18.880 | material to have energy. And before we do that, because MAD is good. Mutually Assured Destruction

02:46:24.560 | is very powerful. Let's take it into the metaverse and then get people to kind of

02:46:29.440 | subscribe to that. And if they actually nuke each other, even for fun in the metaverse,

02:46:34.720 | there are dire consequences. - Yeah, yeah. So it's like a video game.

02:46:38.720 | We all have to join this metaverse video game. - Yeah. I can't believe it.

02:46:43.040 | - And there's dire economic consequences. I don't know how... And it's all run by AI,

02:46:47.840 | as you mentioned. So the AI doomers are really terrified at this point.

02:46:51.920 | - No, they're happy to have a job for another 20 years, right?

02:46:54.720 | - Oh, fearmongering. - Yeah, yeah, yeah. I'm a believer

02:46:58.960 | in equal employment. - You've mentioned that... What do you call it? Chemokina?

02:47:06.480 | - Yeah. - Yeah. So you've mentioned

02:47:08.480 | that a chemical brain is something you're interested in creating. And that's a way to

02:47:16.080 | get conscious AI soon. Can you explain what a chemical brain is? - I want to understand the

02:47:22.960 | mechanism of intelligence that's gone through evolution, right? Because the way that intelligence

02:47:28.640 | was produced by evolution appears to be the following. Origin of life, multicellularity,

02:47:35.600 | locomotion, senses. Once you can start to see things coming towards you, and you can remember

02:47:44.720 | the past and interrogate the present and imagine the future, you can do something amazing, right?

02:47:49.920 | And I think only in recent years did humans become Turing complete, right?

02:47:55.760 | - Yeah, yeah, yeah. - And so that Turing completeness

02:48:00.720 | kind of gave us another kick up. But our ability to process that information

02:48:07.120 | is produced in a wet brain. And I think that we do not have the correct hardware architectures

02:48:18.400 | to have the domain flexibility and the ability to integrate information. And I think intelligence

02:48:24.960 | also comes at a massive compromise of data. Right now, we're obsessing about getting more and more

02:48:32.320 | data, more and more processing, more and more tricks to get dopamine hits. So when we look back

02:48:38.480 | on this, going, "Oh, yeah, that was really cool." Because when I chat GPT, it made me feel really

02:48:44.960 | happy. I got a hit from it, but actually, it just exposed how little intelligence I use in every

02:48:54.400 | moment because I'm easily fooled. So what I would like to do is to say, "Well, hey, hang on. What is

02:49:02.080 | it about the brain?" So the brain has this incredible connectivity, and it has the ability to,

02:49:09.280 | you know, as I said earlier about my nephew, I went from Bill to Billy, and he went, "All right,

02:49:14.880 | Leroy." Like, how did he make that leap? That he was able to basically, without any training,

02:49:20.720 | I extended his name. He went, "Okay." He doesn't like, he wants to be called Bill.

02:49:24.480 | He went back and said, "You'd like to be called Lee? I'm going to call you Leroy."

02:49:27.360 | So human beings have a brilliant ability, or intelligent beings appear to have a brilliant

02:49:34.960 | ability to intercreate across all domains all at once, and to synthesize something which allows us

02:49:41.200 | to generate knowledge. And becoming Turing-complete on our own, although AIs are built in Turing-complete

02:49:52.080 | things, their thinking is not Turing-complete in that they are not able to build universal

02:49:57.440 | explanations. And that lack of universal explanation means that they're just inductivists.

02:50:04.720 | Inductivism doesn't get you anywhere. It's just basically a party trick.

02:50:09.600 | It's like, you know, I like the, I think it's in the fabric of reality from David Deutsch,

02:50:15.680 | where basically, you know, the farmer is feeding the chicken every day, and the chicken's getting

02:50:20.960 | fat and happy, and the chicken's like, "I'm really happy. Every time the farmer comes in

02:50:24.560 | and feeds me." And then one day the farmer comes in and doesn't, instead of feeding the chicken,

02:50:28.880 | just wrings its neck. You know, and that's kind of, and had the chicken had an alternative

02:50:34.560 | understanding of why the farmer was feeding it. - It's interesting though, because we don't know

02:50:40.160 | what's special about the human mind that's able to come up with these kind of generalities,

02:50:43.600 | this universal theories of things, and will come up with novelty. I can imagine,

02:50:49.440 | 'cause you gave an example, you know, about William and Leroy. I feel like

02:50:57.120 | example like that, we'll be able to see in future versions of large language models. We'll be

02:51:06.000 | really, really, really impressed by the humor, the insights, all of it. Because it's fundamentally

02:51:15.200 | trained on all the incredible humor and insights that's available out there on the internet, right?

02:51:19.840 | So we'll be impressed. I think we'll be impressed. - Oh, I'm impressed. I'm impressed.

02:51:25.120 | - Increasingly so. - But we're mining the past.

02:51:27.680 | - Yes. - And what the human brain

02:51:29.600 | appears to be able to do is mine the future. - Yes. So novelty, it is interesting whether

02:51:35.440 | these large language models will ever be able to come up with something truly novel.

02:51:40.880 | - I can show on the back of a piece of paper why that's impossible. And it's like, the problem

02:51:44.960 | is that, and again, there's a domain experts kind of bullshitting each other. The term generative,

02:51:52.240 | right? Average person think, "Oh, it's generative." No, no, no. If look, if I take the numbers between

02:51:59.600 | zero and 1,000, and I train a model to pick out the prime numbers by giving them all the prime

02:52:04.800 | numbers between zero and 1,000, it doesn't know what prime number is. Occasionally, if I can cheat

02:52:10.480 | a bit, it will start to guess. It never will produce anything out with the dataset because

02:52:15.280 | you mine the past. The thing that I'm getting to is I think that actually, current machine learning

02:52:20.640 | technologies might actually help reveal why time is fundamental. It's like kind of insane, because

02:52:25.600 | they tell you about what's happened in the past, but they can never help you understand what's

02:52:30.240 | happening in the future without training examples. Sure, if that thing happens again, it's like...

02:52:38.720 | So, let's think about what large average models are doing. We have all the internet as we know it,

02:52:45.600 | you know, language, but also they're doing something else. We're having human beings

02:52:49.760 | correcting it all the time. Those models are being corrected. Steered. Corrected.

02:52:56.880 | Modified. Tweaked. Cheating. Well, you could say that training on human data in the first place is

02:53:07.600 | cheating. Human is in the loop. Sorry to interrupt. Yes, so human is definitely in the loop,

02:53:11.680 | but it's not just human that's in the loop. A very large collection of humans is in the loop.

02:53:19.760 | I mean, to me, it's not intuitive that you said prime numbers, that the system can't generate an

02:53:28.240 | algorithm, right? That the algorithm that can generate prime numbers, or the algorithm that

02:53:35.760 | could tell you if a number is prime and so on, and generate algorithms that generate algorithms

02:53:40.160 | that generate algorithms that start to look a lot like human reasoning, you know?

02:53:46.320 | I think, again, we can show that on a piece of paper. Sure, I think you have to have...

02:53:53.440 | So, this is the failure in epistemology. I'm glad I even can say that word,

02:53:58.160 | let alone what it means. You've said it multiple times.

02:54:00.320 | I know, it's like three times now. Without failure.

02:54:03.920 | Quit while you're ahead. Just don't say it again, because you did really well.

02:54:07.360 | Thanks. So, what is reasoning? So, coming back to the chemical brain, if I could basically,

02:54:15.360 | if I could show that in a... Because, I mean, I'm never going to make an intelligence thing

02:54:20.560 | in chem machina, because we don't have brain cells. They don't have glial cells,

02:54:24.640 | they don't have neurons. But if I can take a gel and engineer the gel to have it be a hybrid

02:54:32.160 | hardware for reprogramming, which I think I know how to do, I will be able to process a lot more

02:54:37.840 | information and train models billions of times cheaper, and use cross-domain knowledge. And

02:54:44.720 | there's certain techniques I think we can do. But there's still missing the abilities of human

02:54:51.920 | beings have had to become true and complete. And so, I guess the question to give back at you,

02:54:58.880 | is like, how do you tell the difference between trial and error, and the generation of new

02:55:05.680 | knowledge? I think the way you can do it is this, is that you come up with a theory and explanations,

02:55:11.520 | inspiration comes from out there, and then you then test that, and then you see that's going

02:55:17.840 | towards a truth. And human beings are very good at doing that, in the transition between philosophy,

02:55:22.400 | mathematics, physics, and natural sciences. And I think that we can see that. Where I get confused

02:55:29.520 | is why people misappropriate the term artificial intelligence to say, "Hey, there's something else

02:55:37.680 | going on here." Because I think you and I both agree, machine learning is really good. It's only

02:55:41.680 | going to get better, we're going to get happier with the outcome. But why would you ever think

02:55:46.640 | the model was thinking? Or reasoning? Reasoning requires intention. And the intention, if the

02:55:56.080 | model isn't reasoning, the intentions come from the prompter. And the intention has come from

02:56:02.000 | the person who programmed it to do it. So I-- - But don't you think you can prompt it to have

02:56:10.960 | intention? Basically start with the initial conditions and get it going. Where the, you know,

02:56:19.120 | currently large language models, Chad GPT, only talks to you when you talk to it. There's no

02:56:28.320 | reason why you can't just start it talking. - But those initial conditions came from someone

02:56:34.000 | starting it. - Yes.

02:56:35.120 | - And that causal chain in there, so that intention comes from the outside.

02:56:38.960 | I think that there is something in that causal chain of intention that's super important.

02:56:42.640 | I don't disagree we're going to get to AGI. It's a matter of when and what hardware. I think we're

02:56:48.000 | not going to do it in this hardware. And I think we're unnecessarily fetishizing really cool outputs

02:56:53.520 | and dopamine hits. Because obviously that's what people want to sell us. - Well, but there could be,

02:56:59.440 | I mean, AGI is a loaded term, but there could be incredibly super impressive intelligence systems

02:57:08.880 | on the way to AGI. So these large language models, I mean, if it appears conscious,

02:57:15.040 | if it appears super intelligent, who are we to say it's not? - I agree. But the super intelligence

02:57:23.600 | I want, I want to be able to have a discussion with it about coming up with fundamental new

02:57:32.240 | ideas that generate knowledge. And if the super intelligence we generate can mind novelty from

02:57:36.960 | the future that I didn't see in its training set in the past, I would agree that something really

02:57:41.360 | interesting is coming on. I'll say that again. If the intelligence system, be it a human being,

02:57:46.160 | a chatbot, something else, is able to produce something truly novel that I could not predict,

02:57:53.120 | even having full audit trail from the past, then I'd be sold. - Well, so we should be clear that

02:57:59.840 | it can currently produce things that are, in a shallow sense, novel, that are not in the training

02:58:08.960 | set. But you're saying truly novel. - I think they are in the training set. I think everything it

02:58:15.280 | produces comes from a training set. There's a difference between novelty and interpolation.

02:58:20.160 | We do not understand where these leaps come from yet. That is what intelligence is, I would argue.

02:58:25.440 | Those leaps, and some people say, no, it's actually just what will happen if you just do cross-domain

02:58:30.320 | training and all that stuff. And that may be true, and I may be completely wrong. But right now,

02:58:35.840 | the human mind is able to mind novelty in a way that artificial intelligence systems cannot. And

02:58:42.000 | this is why we still have a job, and we're still doing stuff. And I used chat GPT for a few weeks.

02:58:46.400 | Well, this is cool. And then it took me too... Well, what happened is it took me too much time

02:58:51.120 | to correct it. Then it got really good. And now they've done something to it. It's not actually

02:58:54.960 | that good. - Yeah, right. - I don't know what's going on. - Censorship, yeah. I mean, that's

02:59:00.320 | interesting, but it will push us humans to characterize novelty better, like characterize

02:59:05.280 | what is novel, what is truly novel, what's the difference between novelty and interpolation.

02:59:10.400 | - I think that this is the thing that makes me most excited about these technologies,

02:59:14.960 | is they're gonna help me demonstrate to you that time is fundamental, and the future is bigger than

02:59:21.920 | the present, which is why human beings are quite good at generating novelty, because we have to

02:59:28.160 | expand our data set, and to cope with unexpected things in our environment. Our environment throws

02:59:33.920 | them all at us. Again, we have to survive in that environment. And I mean, I never say never,

02:59:39.520 | I would be very interested in how we can get cross-domain training cheaply in chemical systems,

02:59:46.080 | 'cause I'm a chemist, and the only sentient thing I know of is the human brain. But maybe that's

02:59:50.640 | just me being boring and predictable, and not novel. - Yeah, you mentioned GPT for electron

02:59:56.240 | density. So a GPT-like system for generating molecules that can bind to hosts automatically.

03:00:04.240 | I mean, that's interesting. That's really interesting, applying this same kind of

03:00:08.960 | transform mechanism. - Yeah, I mean, this is one that, it goes, my team, I try and do things that

03:00:16.800 | are non-obvious, but non-obvious in certain areas. And one of the things I was always asking about,

03:00:22.000 | in chemistry, people like to represent molecules as graphs, and it's quite difficult. It's really

03:00:28.480 | hard. If you're doing AI in chemistry, you really want to basically have good representations,

03:00:33.120 | you can generate new molecules that are interesting. And I was thinking, well,

03:00:36.800 | molecules aren't really graphs, and they're not continuously differentiable. Could I do something

03:00:42.320 | that was continuously differentiable? I was like, well, molecules are actually made up of electron

03:00:45.440 | density. So then I got thinking, say, well, okay, could there be a way where we could just basically

03:00:50.320 | take a database of readily solved electron densities for millions of molecules? So we

03:00:58.720 | took the electron density for millions of molecules and just trained the model

03:01:01.680 | to learn what electron density is. And so what we built was a system that you literally could

03:01:09.200 | give it a, let's say you could take a protein that has a particular active site, or a cup with a

03:01:14.800 | certain hole in it, you pour noise into it, and with a GPT, you turn the noise into electron

03:01:19.280 | density. And then, in this case, it hallucinates like all of them do, but the hallucinations are

03:01:25.200 | good because it means I don't have to train on such a large number, such a huge dataset,

03:01:30.720 | because these datasets are very expensive. Because how do you produce it? So go back a step. So

03:01:36.160 | you've got all these molecules in this dataset, but what you've literally done is a quantum

03:01:42.080 | mechanical calculation where you produce electron densities for each molecule.

03:01:45.360 | So you say, oh, this representation of this molecule has these electron densities associated

03:01:48.960 | with it. So you know what the representation is, and you train the neural network to know what

03:01:53.200 | electron density is. So then you give it an unknown pocket. You pour in noise, and you say,

03:01:58.000 | right, produce me electron density. It produces electron density that doesn't look ridiculous.

03:02:02.880 | And what we did in this case is we produced electron density that maximizes the electrostatic

03:02:09.440 | potential, so the stickiness, but minimizes what we call the steric hindrance, so the overlap,

03:02:14.000 | so it's repulsive. So make the perfect fit. And then we then used a kind of like a chat GPT type

03:02:22.320 | thing to turn that electron density into what's called a smile. A smile string is a way of

03:02:28.480 | representing a molecule in letters. And then we can then--

03:02:32.160 | - So it just generates them then? - Just generates them. And then the

03:02:34.720 | other thing is then we bung that into the computer, and then it just makes it.

03:02:37.520 | - Yeah, the computer being the thing that, right, to generate--

03:02:41.760 | - The robot that we've got that can basically just do chemistry.

03:02:44.000 | - Create any-- - Yeah. So we've kind of got this

03:02:46.640 | end-to-end drug discovery machine where you can say, oh, you want to bind to this active site?

03:02:51.360 | Here you go. I mean, it's a bit leaky, and things kind of break, but it's a proof of principle.

03:02:56.160 | - But were the hallucinations, are those still accurate?

03:03:00.880 | - Well, the hallucinations are really great in this case, because in the case of a large

03:03:04.880 | numbers model, the hallucinations just make everything up to, well, it doesn't just make

03:03:09.360 | everything up, but it gives you an output that you're plausibly comfortable with,

03:03:12.400 | but it thinks you're doing probabilistically. The problem on these electron density models is

03:03:17.840 | it's very expensive to solve a Schrodinger equation going up to many heavy atoms

03:03:22.400 | and large molecules. And so we wondered if we trained the system on up to nine heavy atoms,

03:03:32.720 | whether it would go beyond nine. And it did. It started to generate molecules of 12. No problem,

03:03:38.160 | they look pretty good. And I was like, well, this hallucination I will take for free. Thank you very

03:03:41.920 | much. Because it just basically, this is a case where interpolation, extrapolation worked relatively

03:03:46.880 | well, and we were able to generate the really good molecules. And then what we were able to

03:03:53.040 | do here is, and this is a really good point, what I was trying to say earlier, that we were able to

03:03:58.720 | generate new molecules from the known data set that would bind to the host. So a new guest would

03:04:07.840 | bind. Were these truly novel? Not really, because they were constrained by the host.

03:04:14.400 | Were they new to us? Yes. So I do understand, I can concede that machine learning systems,

03:04:23.360 | artificial intelligence systems can generate new entities, but how novel are they? It remains to

03:04:29.600 | be seen. - Yeah. And how novel the things that humans generate is also difficult to quantify.

03:04:36.640 | They seem novel. - That's what a lot of people say. So the way to really get to genuine novelty,

03:04:46.640 | and the assembly theory shows you the way, is to have different causal chains overlap.

03:04:51.680 | And this really resonates with the time is fundamental argument. And if you're bringing

03:05:01.600 | together a couple of objects with different initial conditions coming together, when they

03:05:07.840 | interact, the more different their histories, the more novelty they generate in time going forward.

03:05:15.760 | And so it could be that genuine novelty is basically about mix it up a little. And the

03:05:22.000 | human brain is able to mix it up a little, and all that stimulus comes from the environment.

03:05:27.120 | But all I think I'm saying is the universe is deterministic going back in time,

03:05:31.520 | non-deterministic going forward in time, because the universe is too big in the future to contain

03:05:37.280 | in the present. Therefore, these collisions of known things generate unknown things that then

03:05:44.640 | become part of your data set and don't appear weird. That's how we give ourselves comfort.

03:05:49.520 | The past looks consistent with this initial condition hypothesis, but actually we're

03:05:53.440 | generating more and more novelty. And that's how it works. Simple. - So it's hard to quantify novelty

03:06:00.560 | looking backwards. I mean, the present and the future are the novelty generators. - But I like

03:06:06.240 | this whole idea of mining novelty. I think it is going to reveal why the limitations of current AI

03:06:14.240 | is a bit like a printing press, right? Everyone thought that when the printing press came, that

03:06:20.400 | writing books is going to be terrible, that you had evil spirits and all this. They were just books.

03:06:24.320 | - And same would be with AI. But I think just the scale you can achieve in terms of impact

03:06:31.760 | with AI systems is pretty nerve-wracking. - But that's what the big companies want you to think.

03:06:39.120 | But not like in terms of destroy all humans, but you can have major consequences in the way social

03:06:46.160 | media has had major consequences, both positive and negative. And so you have to kind of think

03:06:51.680 | about it and worry about it. But yeah, people that fear monger, you know. - My pet theory for this,

03:06:57.280 | you want to know? - Yeah. - Is I think that a lot of, and maybe I'm being, and I think,

03:07:02.560 | I really do respect, you know, a lot of the people out there who are trying to have discourse about

03:07:08.640 | the positive future. So open AI guys, meta guys and all this. What I wonder if they're trying to

03:07:14.080 | cover up for the fact that social media has had a pretty disastrous effect on some level, and they're

03:07:18.560 | just trying to say, "Oh yeah, we should do this." And covering up for the fact that we have got some

03:07:24.400 | problems with, you know, teenagers and Instagram and Snapchat and, you know, all this stuff. And

03:07:30.400 | maybe they're just overreacting now. It's like, "Oh yeah, sorry, we made the bubonic plate and

03:07:35.680 | gave it to you all and you were all dying. And oh yeah, but look at this over here, it's even worse."

03:07:40.160 | - Yeah, there's a little bit of that. But there's also not enough celebration of the positive impact

03:07:45.360 | that all these technologies have had. We tend to focus on the negative and tend to forget

03:07:49.760 | that, in part because it's hard to measure. Like, it's very hard to measure the positive

03:07:55.920 | impact social media had on the world. - Yeah, I agree. But if what I worry about right now

03:08:01.200 | is like, I'm really, I do care about the ethics of what we're doing. And one of the reasons why I'm

03:08:06.320 | so open about the things we're trying to do in the lab, make life, look at intelligence, all this,

03:08:10.400 | is so people say, "What are the consequences of this?" And you say, "Well, the consequences of

03:08:14.960 | not doing it." And I think that what worries me right now in the present is lack of authenticated

03:08:22.320 | users and authenticated data. - Human users.

03:08:26.080 | - Yeah, human users. - I still think that there will be

03:08:29.840 | AI agents that appear to be conscious, but they would have to be also authenticated and labeled

03:08:35.920 | as such. There's too much value in that, like friendships with AI systems. There's too much

03:08:43.280 | meaningful human experiences to have with AI systems that I just...

03:08:47.280 | - But that's like a tool, right? It's a bit like a meditation tool, right? Some people have a

03:08:51.040 | meditation tool, it makes them feel better. But I'm not sure you can ascribe sentience and legal

03:08:55.920 | rights to a chatbot that makes you feel less lonely. - Sentience, yes. I think legal rights,

03:09:03.520 | no. I think it's the same. You can have a really deep, meaningful relationship with a dog.

03:09:07.680 | - With a dog sentient? - Yes.

03:09:10.800 | - The chatbot's not, right now, using the technology we use, it's not gonna be sentient.

03:09:15.840 | - Ah, this is gonna be a fun, continued conversation on Twitter that I look forward to.

03:09:21.600 | Since you've had, also, from another place, some debates that were inspired by the assembly theory

03:09:30.480 | paper, let me ask you about God. Is there any room for notions of God in assembly theory? Who's God?

03:09:41.840 | - Yeah, I don't know what God is a... I mean, so, God exists in our mind, created by selection.

03:09:49.760 | So, human beings have created the concept of God in the same way that human beings have created

03:09:55.520 | the concept of superintelligence. - Sure, but does it mean, does it not...

03:10:00.480 | It still could mean that that's a projection from the real world, where we're just assigning words

03:10:09.760 | and concepts to a thing that is fundamental to the real world. That there's something out there

03:10:16.640 | that is a creative force underlying the universe. - I think the universe, there is a creative force

03:10:25.120 | in the universe, but I don't think it's sentient. I mean, I think the... So, I do not understand the

03:10:32.080 | universe. So, who am I to say that God doesn't exist? I am an atheist, but I'm not an angry

03:10:42.800 | atheist, right? I have lots of... There's some people I know that are angry atheists and say

03:10:48.480 | that religious people are stupid. I don't think that's the case. I have faith in some things,

03:10:55.760 | because I don't... I mean, when I was a kid, I was like, "I need to know what the charge of

03:11:00.160 | electron is." I was like, "I can't measure the charge of an electron." I just gave up and had

03:11:04.400 | faith, okay? You know, resistors work. So, when it comes to... I want to know why the universe is

03:11:12.480 | growing in the future and what humanity is going to become. I've seen that the acquisition of

03:11:19.200 | knowledge via the generation of novelty to produce technology has uniformly made humans' lives

03:11:25.840 | better. I would love to continue that tradition. - You said that there's that creative force.

03:11:33.680 | Do you think, just to think on that point, do you think there's a creative force? Is there like a

03:11:39.360 | thing, like a driver that's creating stuff? - Yeah. I think that... So, I think that...

03:11:47.920 | - And where? Can you describe it mathematically? - Well, I think selection. I think selection...

03:11:53.360 | - Selection is the force. - Selection is the force in the universe

03:11:56.480 | that creates novelty. - So, is selection somehow fundamental?

03:12:00.960 | - Yeah. I think persistence of objects that could decay into nothing through operations that

03:12:08.800 | maintain that structure. I mean, think about it. It's amazing that things exist at all, that we're

03:12:15.600 | just not a big commentarial mess. - Yes.

03:12:17.360 | - So, the fact that... - And they exist...

03:12:20.960 | A thing that exists persists in time. - Yeah. I mean, let's think maybe the universe is

03:12:26.400 | actually, in the present, the things, everything that can exist in the present does exist.

03:12:33.600 | -

03:12:39.600 | Well, that would mean it's deterministic, right? - No, I think the universe is... So,

03:12:44.400 | the universe started super small. The past was deterministic. There wasn't much going on.

03:12:48.560 | And it was able to mine, mine, mine, mine, mine. And so, the process is somehow generating...

03:12:56.080 | Universe is basically... I can't put... I'm trying to put this into words.

03:13:01.760 | - Did you just say there's no free will, though? - No, I didn't say that.

03:13:04.960 | - As if... - Sorry, sorry, sorry. I said

03:13:07.840 | there is free will. I think... I'm saying that free will occurs at the boundary between the...

03:13:16.160 | - Past and the future? - The past and the future.

03:13:19.680 | - Yeah. I got you. But everything that can exist does exist.

03:13:23.520 | - Everything that is... So, everything that's possible to exist at this... So, no,

03:13:28.160 | I'm really... - There's a lot of loaded words there.

03:13:31.840 | - So, what I mean is... - There's a time

03:13:34.080 | element loaded into that statement. - I think that the universe is able

03:13:37.360 | to do what it can in the present, right? - Yeah.

03:13:40.080 | - And then I think in the future, there are other things that could be possible. We can

03:13:43.040 | imagine lots of things, but they don't all happen. - Sure. That's where you sneak in free will right

03:13:49.680 | there. - Yeah. So, I guess what I'm

03:13:51.520 | saying is what exists is a convolution of the past with the present and the free will going

03:13:59.520 | into the future. - But we can still imagine stuff,

03:14:02.240 | right? We can imagine stuff that will never happen. - And it's amazing force, because you're

03:14:06.320 | imagining... This is the most important thing that we don't understand is our imaginations can actually

03:14:13.200 | change the future in a tangible way, which is what the initial conditions in physics cannot predict.

03:14:19.920 | Your imagination has a causal consequence in the future.

03:14:24.240 | - Isn't that weird too? - Yeah.

03:14:26.960 | - I do. - It breaks the laws of physics,

03:14:34.640 | as we know them right now. - Yeah. So, you think the imagination

03:14:39.360 | has a causal effect on the future. - Yeah.

03:14:41.760 | - But it does exist in there in the head. - It does.

03:14:44.560 | - And there must be a lot of power in whatever's going on. There could be a lot of power whatever's

03:14:48.880 | going on in there. - If we then go back to the initial

03:14:52.080 | conditions, and that is simply not possible, that can happen. But if we go into a universe where we

03:14:59.040 | accept that there is a finite ability to represent numbers, and you have rounding... Well, not rounding

03:15:04.400 | errors. You have some... What happens, your ability to make decisions, imagine, and do stuff

03:15:11.120 | is at that interface between the certain and the uncertain. It's not, as Yasha was saying to me,

03:15:17.760 | randomness goes and you just randomly do random stuff. It is that you are set free a little on

03:15:23.840 | your trajectory. Free will is about being able to explore on this narrow trajectory that allows you

03:15:30.720 | to build... You have a choice about what you build, or that choice is you interacting with a future

03:15:36.240 | in the present. - What to you is most beautiful

03:15:40.800 | about this whole thing? The universe. - The fact it seems to be very undecided,

03:15:50.320 | very open. The fact that every time I think I'm getting towards an answer to a question,

03:15:58.000 | there are so many more questions that make the chase. - Do you hate that it's going to be over

03:16:04.800 | at some point? - Well, for me, I think if you think about it, is it over for Newton now?

03:16:13.120 | Newton has had causal consequences in the future. We discuss him all the time.

03:16:18.480 | - His ideas, but not the person. - The person just had a lot of causal

03:16:22.640 | power when he was alive, but oh my God. One of the things I want to do is leave as many

03:16:26.160 | Easter eggs in the future when I'm gone to go, "Oh, that's cool."

03:16:28.800 | - Would you be very upset if somebody made a good, large language model that's fine-tuned

03:16:35.840 | to Lee Conan? - It would be quite boring,

03:16:38.400 | because I mean... - No novelty generation?

03:16:41.840 | - I mean, if it's a faithful representation of what I've done in my life, that's great. That's

03:16:46.640 | an interesting artifact, but I think the most interesting thing about knowing each other is we

03:16:51.760 | don't know what we're going to do next. - Sure. Sure.

03:16:56.800 | - I mean, within some constraints, I can predict some things about you, you can predict some

03:17:01.920 | things about me, but we can't predict everything. - Everything.

03:17:04.720 | - And it's because we can't predict everything is why we're excited to come back and discuss

03:17:09.680 | and see. So yeah, I'm happy that it'll be interesting that some things that I've done

03:17:16.640 | can be captured, but I'm pretty sure that my angle on mining novelty for the future

03:17:26.080 | will not be captured. - Yeah. Yeah.

03:17:30.000 | That's what life is, is just some novelty generation and then you're done.

03:17:38.880 | Each one of us just generally a little bit, or have the capacity to at least.

03:17:42.880 | - I think life is a selection produces life and life affects the universe in them.

03:17:50.640 | Universes with life in them are materially and physically fundamentally different than

03:17:55.760 | universes without life. And that's super interesting. And I have no beginnings of

03:18:01.840 | understanding. I think maybe this is like in a thousand years, there'll be a new discipline

03:18:05.200 | in humans. Yeah, of course, this is how it all works. - In retrospect, it will all be

03:18:11.440 | obvious, I think. - I think assembly theory is obvious. That's why a lot of people got

03:18:15.600 | angry. They were like, "Oh my God, this is such nonsense." And like, "Oh, actually it's

03:18:21.360 | not quite." But the writing's really bad. - Well, I can't wait to see where it evolves,

03:18:27.440 | Lee. And I'm glad to get to exist in this universe with you. You're a fascinating human.

03:18:33.520 | This is always a pleasure. I hope to talk to you many more times. And I'm a huge fan of just

03:18:39.840 | watching you create stuff in this world. And thank you for talking today. - It's a pleasure

03:18:45.120 | as always, Lex. Thanks for having me on. - Thanks for listening to this conversation

03:18:49.200 | with Lee Cronin. To support this podcast, please check out our sponsors in the description.

03:18:53.760 | And now let me leave you with some words from Carl Sagan. "We can judge our progress by the

03:18:59.520 | courage of our questions and the depth of our answers, our willingness to embrace what is true

03:19:06.080 | rather than what feels good." Thank you for listening and hope to see you next time.

03:19:11.920 | - Bye.

03:19:13.460 | you

03:19:13.540 | you

03:19:15.000 | you

03:19:15.500 | you

03:19:16.960 | you

03:19:17.460 | [BLANK_AUDIO]

Lee Cronin: Controversial Nature Paper on Evolution of Life and Universe | Lex Fridman Podcast #404

Chapters