Future of Science and Technology Q&A (December 22, 2023)

Hello everyone. Welcome to another episode of Q and A about future of science and technology. Let me see. Number of questions saved up here. There's one from Aaron. What will prompt engineering be like if LLMs are super intelligent? You know, I have to say it's an interesting question, you know, what are minds? Like if they're different from ours, like if they are super intelligent relative to us. I've been kind of wondering about this. I've been wondering about it for a long time, but I've been thinking about it a bit recently. Here's a question. So you know, we humans have what, 86 billion neurons in our brains? You know, ants have 50,000. It's clear we humans manage to do more than ants manage to do. Question is, if we scaled up and somehow, you know, I don't know, through genetic modification or through neural implants or whatever else, we effectively had, oh, I don't know, a trillion neurons in our brains, what would the consequences of that be and how would that make us? You kind of always. In the science fiction movies of the super intelligent aliens, I think these have kind of gone out of fashion these days. But used to be the case the super intelligent aliens would have bigger heads. And maybe there's something to that. Now one thing that is a feature of our sort of human consciousness is that we tend to believe that we have sort of a single thread of experience. And it may be non trivial to achieve kind of a single thread of experience even with 86 billion neurons. And for example, we have some evidence that when we generate language, there are certain parts of the brain that are critical to being able to yak on with human language. And if those parts of the brain stop working, Broca's area, whatever else, stop working, then you are no longer able to yak on and produce language. But it seems like much more of the brain is somehow involved in the production of things. You're saying that's just some kind of gate that things have to pass through to actually get out into the expressed kind of linguistic world. And so potentially much of our brain is being sort of corralled into having this thread of experience. And all these different neurons are somehow cooperating to give us this one thread of experience. I don't think it's really very clear how that that works. I mean, we know that certain nerves in brains, if they don't work, one doesn't form this single thread of experience. One knows that in various kinds of psychiatric conditions one ends up with multiple personalities and things like this. And there are sort of Somehow that's happening sort of within the same brain, so to speak. It's rather unclear how all of that works, but. So one thing that's important for us humans at least, is that we have this sort of single thread of experience. Whether the, I don't know, shocks where they have two sides of the brain and one's asleep at a time type thing, Whether they also have a single thread of experience like we do, I don't know. I would guess yes, but I don't know. It's kind of like this time of day, it's my left brain working. This time of day, it's my right brain working. We've got different personalities. Who knows? It's unclear. I don't think we know because we can't really see or imagine those brains from the inside. But so the question arises, okay, let's say we had a trillion neurons instead of 100 billion neurons. What will be different? So, a few guesses. So one thing is it's a common feature of human language that we have roughly 50,000 words that we commonly use, 30 to 50,000 words that we commonly use in whatever language we speak. And that seems to be sort of the number of concepts we can hold in our brain somehow roughly now maybe the cats and dogs with smaller brains or something. And by the way, when we talk about brain size, there's also issues about the size of the organism. Like, you know, whales have big brains, but they're very big organisms. I suspect that a lot of that brain, I don't know, but I suspect a lot of it is used up in kind of sensory. Just. There are a lot of nerve endings that come from, you know, the piece of the tail of the whale. It's just a big surface area. And if you don't cover that with nerve endings, then, you know, when something happens to a little piece of the tail of the whale, then the whale doesn't notice unless it's got nerve endings coming from there. And the, you know, the, the, the structure of these things is roughly that the nerves just end up in the brain. I mean, they may go through some ganglion that consolidates them a bit. But fundamentally, to know sort of the map of where things are happening on you is just directly sent as different fibers to the brain. So maybe when you have a bigger animal, it's got a bigger brain, partly just because it's sensing more surface area. So the higher brain functions may not be any bigger. I don't know. But. So one of the things that is sort of a general feature of how we seem to operate is that we are making this kind of objectified, symbolic version of the world. Things happen. We see all this detail, but we describe it in terms of the limited number of words that we have. So we're kind of equivalencing all of these. This is. I was writing about this recently in connection with what I've been calling observer theory. We're kind of equivalencing down all these different details about what the world is like and describing it just in terms of, for example, these pieces of language that are what we care about. Like, I see a camera in front of me. I could get into great detail about how it's got this particular circular thing and the square thing and this and that and the other. And it's a. This type of camera and that type of camera. I can't even read what. Yes, I can read what it says here, but I'm not getting into any of those details for many purposes of my thinking about things. All I need to know is the word camera. It's a camera. That's my kind of overall description of it. So a question would be, a hypothesis would be, the bigger the brain, the more different kinds of sort of concepts you can deal with, roughly measured in terms of the number of words you can deal with. And I think that's an interesting. I mean, so, you know, assume you scale up the brain by a factor of 10. Maybe you get 10 times as many words. What does that mean? I mean, for example, we can ask for an AI, for an LLM, for example, as we scale up the size of the LLM, How does that affect the number of sort of internal concepts that it forms? Now, unfortunately, we don't really understand the formation of internal concepts in LLM. I mean, we can understand something about how it represents things in the world in terms of vectors of numbers and so on. But the idea that we have as humans that we kind of describe the world in terms of these symbolic things like cameras and tables and chairs is something where the formation of those kinds of concepts is not. You know, I don't know how to think about that in connection with things like LLMs, I think it will be interesting to understand that a bit. I think that one can sort of say at a qualitative level, if one has something like an image identifier. And you're looking at this convolutional neural net, it's got many layers, and you can kind of say that at some layer, you can kind of see this in terms of sort of the activation densities and so on. At some layer it's noticing this feature of the image or that feature of the image. Often those are features for which we don't have specific words. And we could imagine that sort of, with bigger brains we could have those words as well. So a question would be sort of what can you do with a bigger vocabulary? I mean, we kind of have the impression that it's always impressive when people, when humans have a bigger vocabulary. I don't know whether that's cause or effect, and I don't know whether, I mean, I kind of think that the bigger vocabulary, the more concepts you can think in terms of. And that's presumably a useful thing. But can we characterize what it means to think in terms of more kinds of concepts? And what does that tell us about kind of how we, what it means to be sort of super intelligent? And you know what that sort of feels like to us. You know, we have our limited, for example, we're interacting with the super intelligent AI, whatever, you know, we just have our limited set of concepts. And maybe it has these other ones which it can perhaps use as intermediate stages in things it's working out, which it then dumbs down to the concepts that we know. And perhaps there's a way in which that works. I mean, it's sort of an interesting thing. We already see this in areas like mathematics, where you can get an answer that involves one set of concepts, but the path to getting that answer can sometimes involve going sort of uphill to some fancier concept. So we see this if we're doing automated theorem proving, for example, and we say, what's the size of the intermediate theorems we have to prove to get to the theorem we want? The theorem we want might be quite small, but there might be these really big intermediate theorems that we have to prove to get there. Or another example from sort of mathematics famous example from what 1500 or something is in solving the cubic equation. Even if you are going to get in the end a real number out, the intermediate parts of the expression may involve complex numbers which cancel their imaginary parts. So in the end it's real, but we don't know. Somehow we have to go uphill to this fancier concept space with introducing concepts like complex numbers, even though we kind of come back down to the simplest situation later. So I don't think, you know, so I mean, that's one sort of model of the sort of the super intelligent is. Yes, it can handle those more complicated intermediate stages, Even though in the end it, it's going to make things simple for us and we're going to look at it and we're going to say we can't get from place A to place B in that chain of thought because in order to do that we'd have to go to this more complicated concept that we don't readily fit in our brains. And in a sense, sort of the idea of universal computation and so on suggests that you can in principle construct any concept from its kind of primitive pieces. But what in a sense is happening when we invent a word, we know a word, we're sort of caching that concept, we're kind of saying we know what that is, it's a camera. So blah, blah, blah. We don't have to say it's a thing that takes light and turns it into electronic data in an array, blah, blah, blah. We can just say it's a camera. So it's sort of a caching of a certain amount of information. So again, I don't really know how to think about what else you can say about the world, sort of. If you turn that knob of sort of, what does it mean if you have a bigger vocabulary? I don't really know what that lets one do. It's an interesting question. I mean, one can answer that to some extent in this case of just saying let's make a bigger bag of intermediate theorems for theorem proving. Actually it's a something I should be able to answer from the stuff I've done. The book I wrote about metamathematics, for example, probably has implicitly some of the answers to those kinds of questions of to what extent can you sort of, how much gain do you get from having a bigger cache of intermediate theorems? Maybe I've even looked at that. Yes, actually, I have looked at that, but I don't know a good sort of qualitative answer to that question. So I mean, that's one type of thing. Another kind of thing is sensory input. I mean, we have 10 million nerve fibers coming from our eyes. We have 50,000 from our ears, certain number of nerve fibers. Imagine that we were more intelligent, so to speak, and we could connect to much broader sensory input. So imagine, oh, I don't know, imagine we had sort of. Well, even, even imagine we were sort of a community that had sort of telepathic communication with lots of other organisms which were walking around and had eyes and so on, and we could get all that data in. Now, in a sense, in, in modern times and through digital electronics, we can get all of that data in, you know, we can be watching the webcam, you know, the million webcams or something. We can be taking in all this Internet of things data and so on. One of the things that's true is we don't know what the heck to do with all of that data because we. Most of our world is used to the experience we have of vision and hearing and so on. The limited data that we get into our sensory systems, by the time you're sort of accumulating tons more data about the world, it's not something that we are intrinsically built to deal with. We can try and use science to break it down and turn that data into some kind of human narrative, but we weren't built to deal with that. Maybe if we had a trillion neurons in our brains and we had more sensory inputs, we'd be more in a position to do that. Maybe we have the direct. I'm somehow thinking of the Matrix movie or something. We had the direct neural connection to electronics, so to speak. And we could sort of just plug into all these sensors out there in the world. What would we make of all of that? It's funny, because one thing we clearly know, there is structure in all that data. I mean, this is the story of quantitative finance, for example, tends to be if you get enough data about the world and you look for regularities, you will find some. Maybe it's getting more difficult to find them, but it's still the case that there are. With enough data, there are regularities to be found which an ordinary human brain has a very hard time finding. But if you're kind of bringing all that stuff, if you could bring it in, if you had a much bigger brain and you could bring it into that, you would start to notice those regularities as a kind of. As an everyday thing. And I suppose an interesting question is if you're a human, or an AI, for that matter, who's grown up in an environment of, I don't know, equities, markets or something, and you're routinely getting tick by tick data of all of that stuff, when you're routinely getting all these other things, how do you form your worldview? And I start thinking about what concepts do you form, but that's a tricky business because language is intrinsically a quite societal kind of construct. In other words, the fact that I'm saying a bunch of words, et cetera, et cetera, et cetera. I learned those words when I was very young, and I learned them from other humans. And now I'm using those Words to communicate with other humans. If I was just me on my own, I think occasionally. Well, somebody like me who likes to think of new ideas, I do occasionally make up words for things where I will describe something in terms of word that nobody else has ever heard. And maybe nobody else will ever hear it. I'm just using that as an internal marker for myself to represent a concept that I've come up with. But that's a rare case. The more common case of words is that they're kind of the social construct that allows for communication between different people. So I'm not even sure that if you have this completely different sensory experience, do you form, unless you have something to talk to that's had similar sensory experience, not even sure what words you form. So in any case, I suppose a few thoughts. I'll tell you a thought that I was just trying to form actually has to do with the following question. You look at these very complicated systems in the world. Like I think I gather, cars have 10,000 components, planes have 100,000 components. In the things that you sort of separately describe when you're doing the design for these systems. Biological systems have even more components. But somehow these components have certain structure to the way that they're organized and the way that they form subsystems, et cetera, et cetera, et cetera. And the thing I was sort of interesting, perhaps, is one characteristic feature seems to be that subsystems are formed a subsystem where there's a lot of stuff going on inside the subsystem. But its exposure to the outside world is thin. And that's very similar to us as observers of the natural world. There's all this stuff going on in the natural world. But we have this thin pipe that. This sort of thin set of things that we are forming in our minds to represent that those are the parts of that that we care about. And I think it's the same type of thing. When you have a big system, you have these subsystems, whether it's in biology or in systems engineering. You have these subsystems that have These sort of APIs, these connections to the outside world that are somehow limited. Now, somehow those connections are a little bit like languages. They're a little bit like the fact that you've aggregated sort of all the detail of what's happening inside into a limited number of kind of symbolic concepts which you are then communicating to other subsystems. And so it's sort of a story of how do you. It's. You're using this kind of thinner communication channel between subsystems. And you're using that as a way to organize the whole system rather than having everything communicate with everything. And one wonders kind of to what extent, well, even sort of biology has done that at all scales, whether one's dealing with cells, whether one's dealing with organisms that interact with the whole sort of species and ecosystem and so on. It seems to kind of at all levels have this kind of approach where sort of a cell is something which has a certain nature, that has just a membrane that has receptors on the surface and so on, that has a limited kind of set of ways in which it communicates with other systems. And I think it's sort of a question of what is. I've been thinking about this kind of theory of bulk orchestration, kind of what is it that describes a microprocessor or an airplane or a biological system that says we've got a lot of components here, but they are somehow. How can you make a bulk statement about what's going on when those components are organized? How are they organized? I think what one's understanding is they're organized in these subsystems, and that's sort of what makes them, for example, comprehensible to us in any way. But I don't know quite where this is going because I haven't figured this out. But I think in terms of this question about the super intelligence question, one of the things that comes to mind is how big can these subsystems be and how complicated can their APIs, their connections to the outside world beyond? And that sort of characteristic of kind of something related to kind of the intrinsic processing power of systems is how big that interface can be, how big the language that describes how this subsystem interfaces with the outside world can be. Just a few thoughts there. Let's see. Emma's commenting. Maybe language is stopping us getting some ideas because we cannot formulate the idea in our known languages. Absolutely. I mean, that's the story of much advance in science and other areas is to form these concepts, form these descriptions of things, because that's what you use to sort of build the next level of the tower. And obviously my theory of that in my life has been build our computational language, Wolfram language, and build concepts into that as we progressively understand them, precisely so that you can get to sort of higher levels in this tower. I mean, I've sort of been living that theory. Let's see, That's a comment from Prab about. A comment from Elon about. Most of our memory is digital now, and the same Will happen with concepts. Well, I mean, what one stores is not the. We might choose nowadays to store the raw bits. We just store the video. Whether we. In a sense, by storing the video, we're already compressing the data a lot. Those individual photons that fell on our receptor, they fell in a particular order, and they had all kinds of details and they had all kinds of directions and so on. But we're equivalencing all that out. We're just storing the video. Now, we can imagine that there are aspects of that. Just like, you know, we get transcripts of, you know, zoom, like sessions that we do. And then the question is, do we store the whole transcript, every word, every cough, or do we store the. You know, do we get rid of the coughs? Do we get rid of the ums and ers? Do we store just the summary? You know, this is. Again, it continues to be a story of what do we care about? What do we choose to store and deal with and so on. Eddie is commenting that creatures like cuttlefish and octopus must have something different going on in their brains. Because they can do things like mimic the outside world with their color and texture and so on. Yeah, yeah. It's an interesting question. I don't know. I was talking to an expert in cuttlefish, et cetera, perception recently, and it's very hard to wrap one's brain around a very different kind of brain. And I think the only statement is that sometimes things like octopuses seem smarter than they are. Well, smarter with respect to what we think of as smart, because they can do things like unscrew the lid of a jar or something and get themselves out of it and things like that. But there are a limited number of things like that that they can do. They just happen to have some. I mean, we call it reflexes. They might say it's their amazing cultural sort of heritage that. That lets them do those particular motions. Let's see. Yeah, I mean, this whole question. RBS is commenting, in the future, the earth would probably be a big brain if every connected computer on the Internet was running neural computations. Yeah. I mean, this. The fact is that this whole question about the organization of humans with cells and neurons in our brains and that builds up to a whole brain that has a whole unified idea of what's going on. And then you build up to human society. Does human society have a unified idea of what's going on? I don't know. As individuals in that society, you might say, no, it doesn't. But looked at from the outside, you might say, well, society is deciding to go in this direction. Just like, in a sense, if you could query our individual neurons in our brains, they wouldn't have a coherent idea, probably, of what's going on. That's something that builds up in the collective, so to speak. Let's see. Sapolo is asking, what. What if we need to deal with concepts that need more neural mass than we possess? I think this is one of the interesting things about, well, knowledge, education and so on. As a species, we're learning more stuff. And yes, the amount of time people go to school has increased, although the dynamics of why that's happened, I don't think is because there's more stuff to learn in the world. I think it's because of the bigger institutional structure of schooling and its level of efficiency and this, that and the other. But the question is, why is there just not? Why don't we have to go to school for 200 years to learn to be functional in the world, given how much is known? And the answer, I think, is that at any given stage, there are both new things that get known and new abstractions that get formed, new equivalences that get discovered. We don't have to know. Oh, back in the day, we might have had to know in detail all the different kinds of computers that exist because, well, they all work differently and they all have this and that and the other. And, you know, back in the day where the very early days of even electronic computers, it's like, well, it has this kind of word and that kind of operation and et cetera, et cetera, et cetera. But then maybe this isn't the best example, but it's a practical example that comes to mind. Then pretty soon we just have to, look, computers work this way, that it's just a general abstracted thing that we can think of computation as an abstract concept. And the same is true. Oh, in lots of areas. I mean, you could say this is my. Well, what's another good example? I mean, a, We might have to learn the particulars of some thing about mathematical knots, for example, and, oh, you learn every particular knot and it's this kind of knot and it's that kind of knot. Maybe not even mathematical knots, maybe nautical knots as well. Knots as in knotted pieces of rope, rather than knots as in nautical miles per hour, although those are obviously related in some historical way. But the, you know, you might have to learn, oh, there's a trefoil knot and there's the figure 8 knot, and there's this knot, and it does this and it does that and you learn them all separately. And then there's some mathematical theory that just says, oh, given the knot polynomial, this is true. And then you only have to learn that one thing. And so I think there's these two processes. There's more facts that get learnt and there's more abstraction that happens that sort of collects together facts into higher level concepts. And I think that somehow there's a question of maybe there are higher level concepts that we really don't understand very well, that our brains are just not very well set up to understand. It seems like once we kind of get familiar with higher level concepts, we can deal with concepts at progressively higher levels. We don't seem to have trouble with that maybe. I mean, I think that actually I suppose in some meta sense these live streams are my attempt to take what are sometimes quite complicated concepts and I hope try to explain them without a big tower of kind of intermediate knowledge, so to speak. And so insofar as that works, we're kind of learning that there aren't kind of intrinsically higher concepts, so to speak. Let's see, let me see. I'm going to go to some other questions here. Oh, there was one from Stephen here. What do you think of the plans to shoot feature films in space? Okay, we're going very different topic here. I don't know about those plans. I think, you know, there have been a limited number of movies that have been made on planes in parabolic flight. You know, if you're in free fall, then it's. Well, when you're in orbit around the Earth, you are essentially in free fall. You're always falling towards the Earth. Except that because you also have a velocity going parallel to the surface of the Earth. By the time you've fallen to the, by the time gravity has made you fall to that point, you're round the other side of the earth. So you just keep going around in orbit. But you are as if you're just continually falling. And in that state you don't feel the force of gravity. So you know, so you, you seem weightless. And one can simulate that by having a plane that flies in a parabolic trajectory, the trajectory that something would fall in if it was just falling, so to speak. And I've never taken one of these flights, I get motion sick. It will be a disaster for me. But I suppose times when I've flown little planes I've. Occasionally long ago, you know, when you kind of, you can be in at least reduced gravity situation fairly easily. But in any case, if you fly sufficiently close to parabolic trajectory, you can simulate zero gravity. And there have been plenty of movies that have had scenes, at least, that have been shot in kind of those brief moments of zero gravity that you get by going from 50,000, 40,000ft or something to down to close to sea level in a parabolic trajectory. I think doing that on a larger scale is surely doable. There are lots of funky phenomena in microgravity or in zero gravity that were very unfamiliar to us. For example, if you have liquid for us on Earth, liquid, you're drinking something out of a cup and the liquid is sitting in the cup because it's pulled there by the force of gravity. If you're in zero gravity, you move the cup and the fluid will slide out of the cup and you'll have a big blob of fluid just sitting in the air. And I know that it's been. Astronauts don't like having big blobs of fluid because that fluid gets all over the place. And you eventually find little drops of water just hanging out in the space station in all sorts of different places. And you have to vacuum them up. And it's kind of a big nuisance. But that's an example of something very non familiar to us. And like, there's a question, you know, what is a splash like in microgravity? I've been wondering about this. I've asked lots of people and a few experiments have been done. People are not very keen to do the experiment because you tend to. You take this blob of water and you shatter it into all these pieces. And then all those little blobs go all over the place. And it's a nuisance. But that's an example of a phenomenon that. See, one of the things that's funny, you see a movie and it shows a piece of physics that is not part of our everyday experience. And you look at it and sometimes you say, that can't be right. For example, this happened in the movie Interstellar. That had a very lovely simulation. I think it was actually done with our technology of a black hole and of, oh gosh, accretion disks, maybe around a black hole, things like this. And the original version of the movie, the black hole was a spinning black hole, I think. And it had everything sort of correctly sort of represented in the ways that light rays operate in a. In a curved space time like that. And it looked really weird. And for example, the black hole wasn't a circular thing. It was an incredibly flattened thing. And people just Sort of said the test viewers or whatever sort of said, what the heck is that? You know, that doesn't look like a black hole. That looks like some kind of weird black pancake of some kind that we don't know what it is. And so the movie was kind of dephysics, so to speak, to show the black hole something which it wouldn't be actually what it would look like. So one of the things that I think will happen if people start shooting movies in microgravity is there'll be all kinds of phenomena like that that are extremely unfamiliar to us as standard gravity dwellers, so to speak. And I don't know how well we'll parse those things. And I suppose after a time one will get used to it and say, oh, that's a. That'd be kind of interesting. You know, that's a thing that was shot in microgravity. That's how things work in microgravity. You kind of extend your experience now to also encompass things that happen in microgravity. But I would be sort of. I'd be a little bit suspicious that at least at the first instance, it's kind of like very alien. I mean, it's the same thing. If you look at things under a microscope, for example, there are endless books of, you know, what is this weird thing? What does it look like? It's got all these circles in it. Oh, it's a piece of soap under a microscope and so on. They're very unfamiliar to us and sort of a question of what is it like to be exposed to a movie or a video game or something like this that is in a world very different from our world? I don't know. I mean, I've been curious with our ruliad concept of sort of the space, possible computations and so on. I'm not a video game player myself, but sort of what it will be like to exist in a different part of the ruliad, so to speak, from the one that we sort of sample normally, perhaps, or even if you think about sort of different laws of physics, what's it like to have very different laws of physics in your video game? Very simple example of this going at speeds close to the speed of light. Let's say you've got a video game and I think people have made at least some versions of this. We've played around with some pieces of things of this type of looking at sort of what's the world like if your car goes at 90% of the speed of light? Well, it's very different. It Has a lot of weird features. You know, you're looking at that cube that's in front of you, and, oh, well, it's actually twisted because light comes to the side of it that you're closer to more quickly than it comes from the other side. And so the thing will look twisted. It'll look kind of red on one side, blue on the other, Even though it was neutral white color, sort of when you're looking at it at rest and so on, it's very unfamiliar. It's very. I mean, I think it's been sort of interesting to me. You know, can one build an intuition for what life is like at relativistic speeds? And back in the day when I used to be involved in doing sort of practical particle physics, this is late 1970s, incredibly ancient past, I used to visit particle physics experiments, and you would see these big detectors and things, and you would know that because of relativistic effects, all these various angles of, oh, this goes from there to there at this angle. The angles that you intuitively think of aren't the real angles that are relevant, because the angles that are relevant are ones corrected for relativistic effects, for things going close to the speed of light. But it's very unfamiliar to one. And so I wonder to what extent we'll just look at sort of the true microgravity thing and say, wait a minute, that just doesn't look right. And there's a certain sense in which, I don't know, explosions in movies. I don't know how accurate they are. There are lots of issues. If you make a model, for example, either a physical model, as people used to do, or a computational model, sometimes even using some algorithms that I made up long ago for somewhat different purposes with cellular automata and so on, to simulate explosions and fire and things like this. I actually don't know whether when you. Whether the real thing looks like the thing you expect in movies. And eventually there is sort of a language that we learn that is movie physics that may or may not be the same as real physics. So it will be interesting if there's a very different kind of real physics that one's exposed to in movies. I wonder what that will be like, and I wonder to what extent. My guess is that it's a little bit like, well, that if you throw people too many unfamiliar things, they just get confused. If you say, oh, we've got a movie where there's one or two really funky things that happen because of the physics being different, then people will be like, oh, that's cool. But if it's too many, it's just disorienting would be my guess. But sort of the test case experiment is the video game that operates according to physics that isn't our physics. Can you make one thing we learned from the story of the ruliad and so on, we can make other consistent physicses, at least for observers different from us. Now that then raises the question, when you're doing the video game, you are a human observer sitting in front of a screen that is producing light and so on. How do you make an observer not like us? And that's sort of a separate question. Let's see. Okay, there's a question here from robot, well named. What impact will Internet of things and 5G technology have on our daily lives? From smart cities to autonomous vehicles. So, you know, Internet of Things, having sensors everywhere, maybe having some actuators in different places is something which multiple times people have said it's going to change everything. So far it hasn't. Now that isn't to say that there are many things that you see in technology, in the developments of technology, where it's like it's going to happen, it's going to happen and it doesn't happen. And then eventually it happens and it does change things, often unexpected things. Good example is video phones. The first demo video phones existed in the 1950s, but yet we didn't get routine video sort of video conferencing until what I mean, I was using it from the early 90s occasionally, but my gosh, when did that come in as a general thing, maybe 10 years ago or something, it started to become routine. And so that's a solid 50 year Spanish relative to when it was originally started to show up. And there are other things where they show up imagined in science fiction 50 years or more before they actually show up. And there are things. So things like everything, there's a sensor in everything is one of those things that's sort of the inexorable future, but hasn't quite arrived yet. I mean, I've often thought that one of the things that people say about sort of what's the future going to be like and how's it different from what we have today is that, you know, right now we just say there are computers in various specific places. But one can imagine a future in which everything is made of computers. Every kind of piece of material is made of molecules that we can think of as doing controlled computation. It hasn't happened yet, but it's something that presumably will happen. And once it's happened, people will be like, what do you mean? There was a time when you would just have a thing that was a fixed color. Of course, everything is programmable and can have whatever imagery it wants on it. That's kind of a thing which one can expect for the future. But I think in terms of, of sort of sensor technology, it's an interesting question. What, okay, what are some things? Certainly in biomedicine, there's a lot of stuff like, you know, right now most of the sensors that we have are in our brains and it's like, oh, you know, something hurts. Okay, there must be something wrong with, you know, that's the error condition, so to speak. But there will surely come a time when we don't have to rely on our brains to know that there's some issue with some organ in our body or something that's going on. It'll be rather than. It's like, you don't need to know that you feel hungry because you have a signal in your brain, or you don't need to know you didn't have enough vitamin B12 or something, which we don't know right now. Maybe we get some vague cravings for things when we haven't had enough of such and such a kind of thing to eat. But one can imagine a time when that's just a thing where, look, I've got this readout. It just says it's measuring the blood level of vitamin B12 at all times. And look, I can see it dip down, and so on and so on and so on. And I think there is a certain amount of self awareness that we can imagine from sensors in us, so to speak. And then there's a certain amount of situational awareness for the world from having sort of more sensors out there in the world. And obviously in, you know, there's a transition like, you know, webcams are a thing. I don't know how critically, societally pivotal they are. And, well, depends what country you're in and how they're used. But, but, you know, the, the fact that we can have much deeper sensing of what's going on in the world, you know, has some, some effects. I think that the, for example, the fact that we can readily have digital memory of the past. I've been one of these people who collects endless sort of personal analytics data. And I have both my digital data going back into the 1980s, and I've scanned my paper data going back further than that. But, you know, that makes, I would say, some difference in, you know, the past is no longer forgotten in the way that it was before. It's an outsourced kind of human memory. But I mean, the question of, what are, you know, as we get sort of more sensors in our environment, what does that mean as we get. I'm not sure I know the answer to that. I think in. I'm trying to think what kinds of things would be. I mean, what's always interesting is when the cost of a technology goes sufficiently low, stuff starts being done with it that you never imagined. Like with computers, you know, before the cost of computers was low, nobody could imagine doing word processing or using computers just purely for communications. I mean, it's like you've got a computer, it can compute all this stuff and it's spending all this effort running all these things, and then what's it doing? It's sending back and forth text messages. That's using a trivial collection of the processing power of the computer. It so happens that video transmission uses a bit more because of the way it's compressed and decompressed and so on. But in a sense, it's using a very shallow aspect. Not quite so shallow because the compression is fancy. But conceptually it's a shallow aspect of computation. And it's something where sort of as things become cheap enough, you get to sort of do them. So, for example, little tags that you put in with Bluetooth tags or whatever that say, oh, I left my backpack behind, or whatever else, probably there comes a time when everything just has that in it. Everything you buy just already has one of those things in it. And when you buy it, you know, it kind of gets linked to your kind of digital ecosystem. And it's like, okay, so then you can always tell where it is. For example, it's not worth doing unless the cost of those things goes down to the point where it's trivial to sort of print that into every object. And I suppose, you know, what's another example? I mean, there are all kinds of things where you could imagine, oh, I don't know, with food or something. You could imagine something where, I don't know, eat your sensor, eat your Bluetooth device. Seems a little bit unappetizing, but you can imagine a situation where, I'm not sure if this is a good idea, but where all food has something mixed into it that's really tiny, that you are able to. As you eat it, it records a bunch of, oh, you ate this number of calories of this thing because it has that. Just like it has a. I'm not sure. This doesn't sound very appetizing, does it? It's kind of like, you know, you can think about the edible. I've always wondered at what point genetic engineering people talk about genetically modified crops and so on. I've always wondered at what point genetic engineering will get to the point where you can grow barcodes or QR codes or something on fruit and where you go into the grocery store and every apple is not just this green exterior, but it's something where because of its genetics, it has this barcode that describes what type of apple it is or something like that. Apples, I think, are all clones of each other of a given variety, are all kind of genetically identical. But one can imagine something where, well, that's not the case, where it has kind of exposure of this is the particular sub, sub, sub, sub variety that this particular thing is. But, you know, and that one can imagine. Again, it feels a bit unappetizing right now to, you know, eat your Bluetooth transmitter. But you know, that's the type of thing where if it's cheap enough and easy enough, you start to wonder about those types of things. And I suppose, oh, for example, people wonder about coins and banknotes and so on. And is cash going to survive forever? And cash is nice and simple and anonymous and so on. Well, of course, it might be the case that the idea of a digital currency which has sort of features of trackability, for better or worse, is sort of merges because by the time IoT is cheap enough, it's like there's a sensor in every coin. And at that point it's kind of like what's the digital coin versus what's the physical Here I can just pay you this quarter or something. How does that relate? I think sort of a different issue is 5G and telecommunications kinds of things. And at what point sort of what has to be local versus what doesn't have to be local. And there's been this kind of. In the history of computer technology, there's been this kind of back and forth thing. Is it local? Is it a personal computer? Is it a mainframe? Is it the cloud? I kind of find it amusing as I think everybody who's been in the business for a while does that. First there were mainframe computers that were centralized, then there were personal computers. Then we're back to the cloud, which is essentially a mainframe computer where there are sort of thinner local clients for it, just a web browser or something. And so there's this continual back and forth between the centralized and the non Centralized. And so the dynamics of that includes things like what's the cost of doing a centralized thing? What's the versus a non centralized thing versus what's the telecommunications cost? I think one thing for example, in the world of LLMs right now we're in a very centralized state. You need lots of GPUs and they're expensive and they need complicated ecosystem around them to run the thing. So it's worth centralizing those LLMs. There will come a time I'm sure when those will decentralize and it's the same. And I think there are things we couldn't do right now with centralized LLMs because the rate of getting data into them from sort of our edge, our peripheral devices is too low now probably as one imagines higher transmission rates and so on, that becomes kind of less significant. I mean there's sort of this question of when can you transmit real time video from sort of anywhere to anywhere and how does that work with, you know, how does that work with 5G and is it easy to transmit sort of locally and then it's complicated to get it further away, et cetera, et cetera, et cetera. But I think, I mean actually I was just looking at one statistic which I'm curious about, which is what is the relative cost of the training that's done on an LLM, which is sort of a centralized thing, versus the inference, the actually running of the LLM. I think it's like 10 to the 25. Floating point operations are the amount of operations done to train sort of the fanciest LLMs these days. And at least to back of the envelope estimate says that the number of steps done in all of the uses of that LLM is actually now smaller than the number of computational steps done to train it. I think if I got the estimate right, which is sort of a strange situation where the training is more expensive than all the uses of it. It's kind of almost like you could almost say let's pre process everything we're going to do. It's not how it actually works and bottle up all that computational energy effort and then let it out gradually, so to speak. Sort of an interesting, not kind of what I might have expected. You might have expected it sort of online real time computation happening, which it is in actuality, but sort of the economics of it aren't quite like that. So yeah, I mean I think this whole question about telecommunications speeds and to what extent can we sort of rely on fast communication? Can we remotely drive a car? Answer probably yes. Can we? To what extent do we. Okay, so here's a question, which is human sensory data has a certain bandwidth. For audio data, we reach the 44 kilohertz or something. I mean, that's an overkill number for human kind of audio bandwidth. And we reached that many years ago now as a thing that we could just transmit digitally for video. We're not quite there yet, but we're not that far away of being able to sort of remotely transmit something which is as if you were there, type experience. And one assumes that there's a definite point at which you can transmit as if you were their experience for human experiences, so to speak. We talked before about sort of the super intelligent whatever, and you scale up the amount of sensory experience, then you stop being able to transmit all of it. But that will be an interesting threshold when we really can transmit the whole human experience, so to speak, easily and efficiently. And I don't know what the. I have to say, there are things that, again, it's always strange what catches on and what does not. I sort of thought that remote tourism would catch on. That is, you say, I want to visit some place and I want to visit some obscure, dangerous jungle location somewhere. I'd be really curious what's it like in whatever place. Well, maybe I can get somebody who is an experienced jungle explorer to just go wear all that gear and just, you know, I'm sitting there and I'm just immersively watching this, and they're the one walking through the jungle, but it's as if I'm there. And I sort of thought that that kind of thing would catch on more than it has. It really hasn't caught on at all. I mean, I kind of thought in the world of job categories, I kind of was expecting remote tour guide or something like that to be a job category. It isn't yet. Maybe it will be maybe 5G and sort of Iot, like technology will allow that to be possible. And again, I don't really know, you know, what are the bigger consequences of that? I'm not sure. I know. It certainly would be. I mean, one of the things that, of course, airplanes opened up was kind of much broader tourism around the world. By the time you had to take a ship and it was pretty expensive and it took three weeks and so on, you weren't going to go to this cool but distant place. Whereas if it's just a plane flight away and the flight isn't too expensive, you're going to go there. And so obviously global tourism has dramatically opened up. And one could imagine what the consequences of global tourism. Well mean all sorts of things about people knowing more about what's happening in other parts of the world. People transmitting viruses faster, people maybe transmitting ideas faster. I don't know. All kinds of other consequences of sort of economic equalization from people spending money in places where there wasn't money being spent before versus people kind of sort of turning into having fast food joints in places where it was very natural before, so to speak. All kinds of different things. But something like routine remote tourism would probably have other consequences and I don't know what they are yet. Actually. Giovanni is commenting. Someone should start a tourism live stream channel. They should, they should. Why hasn't somebody done that obvious thing to do. It's yeah, good idea. Somebody here should do that. It's. Okay. Mark is asking, do you think in the future humans could interact with AIs via a visualization of latent space to create new concepts? Oh, I've wondered about this. So I mean, what happens in a typical AI system is there's sort of the images that come in as the images are usually viewed on a computer just with pixels and so on, and then there's this grinding up into arrays of numbers that happens inside a neural net and that's a compressed version of the image. We don't know quite what the compression is. The neural net has learnt that compression. Maybe the compression is bit number 17 represents whether there's a blue spot in the top left hand corner or something. Probably the, you know, I don't know what. Yeah, I mean, so we don't know what, what kind of that, that internal representation, that sort of compressed internal representation, how it represents things. But an interesting possibility is that we could become familiar with the sort of sensory experience of latent space. Yes, I had been interested in this and I did some. I wrote some piece recently that, well, at least I nicknamed the cats piece because it's mostly about images of, you know, it has lots of images of cats in it generated by generative AI. This question of sort of getting a human wrapping our brains around the latent space, this internal sort of concept space of the, of the AI. Interesting thing, I have to say, you know, in my efforts, playing around for a few weeks with cats and party hats and generative AIs and so on, I was able to get some intuition about what was going on. But I certainly don't think I was a fluent sort of thinker in latent space. And it's interesting question whether you could get there Comment about my. The remote tourism comment. Missing the ability to try the food. A major reason to actually go somewhere. Do you think there's a way to taste via computation? That was the immediate thing that I wondered about. You know, we have. Back in the day, it wouldn't have been obvious to anybody that you could put little loudspeakers in your ears and hear a sound from anywhere. The fact that we can emulate the production of sound that way, or for images even more so that you could have, you know, people would say, well, I can paint a painting and here it is. But the idea that, you know, 60 times a second or 30 times a second, you can be changing every pixel there and changing your visual scene wasn't a thing. So it makes one wonder. You know, we are not used to. Okay, so the obvious question is the human tongue has a bunch of taste buds and taste receptors and so on, on it. Now, the fact is that taste is quite texture dependent. There are sensors in other parts of one's mouth, things like this. But one, you know, could you imagine a situation in which there's kind of, you just suck on this thing. It's like a lollipop or something, but it's a highly electronic lollipop. And that lollipop is effectively giving sensory input to the taste buds in your tongue and through maybe some electrical process or something like this, and you are effectively tasting the taste. I don't know whether that's. I mean, one of the things that one wonders about is can one feel the touch of something? And that's the whole story of haptics. And there is a certain amount of. You can fool our touch sensors by electrical means, or you can actually generate the texture by the sort of mechanically generate electromechanically generate the texture. And then you're feeling the fur or something, because it really is fur, like on that surface, as opposed to just you feel electrical signals that are like the fur. You know, I tried a horrifying invention a couple of weeks ago. Somebody has of a thing that gives pure electrical signals that affect the vestibular system, the balance system of the brain. And, you know, you stick it on your neck and it's kind of sending electrical signals. I have to say that for me, it was five seconds to motion sickness. But maybe that's just the technology isn't properly developed. But this question of can you fool your sensory system somehow? In the case of. Well, one could argue that whenever we look at a computer display, we're fooling our sensory system. Because, for example, there are Discrete pixels. There are just red, green and blue pixels. It's not the arbitrary kind of frequency spectrum of an actual object, that is it's full color, so to speak. It's rather just being fooled by being excited and the three sort of corresponding to the three types of cone cells that we have in our retinas. So in a sense we're already doing a certain amount of fooling. And the question would be, can we. So can we fool the taste system by. And that's an amusing question. And can we fool the kind of haptic texture sensing system as well? I don't know. And at what point are we really doing something a bit silly where we're sort of hacking into the brain and it's kind of like, well, it's as if you were there because the higher levels in your brain process things that were the things you would have got if you know, if you were there, your lower level brain systems would have processed the raw data. But actually one's just plugging into higher level brain systems and feeling what one would feel, what one's lower level brain systems would have felt if one was in that place. So, yes, I don't know how that would work. Alexander says. I mentioned on Wednesday that I was working on a course about computational thinking. Can I comment on how that's going? It's not gone very far since Wednesday. It's probably progressed by a few pages since then. I hope I'm going to really get a chance to dig into this in the next or probably by the beginning of next month. And I still have to figure out what I'm going to do as I, as I write more of this, whether I'm going to sort of release it in pieces or whether the pieces are going to be too sort of incoherent to be able to put them out. So that's a thing I've been thinking about. But the main kind of point there is computational thinking is kind of this way of thinking in an organized fashion about the world. And I'm really trying to explain sort of how one should do that. Just like one could say, well, you could have an explanation of mathematical thinking. You have to know a certain amount of mathematics to be able to do mathematical thinking. If you don't know the math, you can't apply it to thinking about the world. And so similarly here I think there's a certain amount that you have to sort of know about what's involved in computation before you can apply it to thinking about the world. And so that's what I'm trying to work on. All right, let's see. I'm probably need to go soon. Hold on one second. Oh, the. Let's see, Maybe one more. If I can find. It's always dangerous when I say, let me find a question. That is. Let's see. I think these are all long stories. Okay. All right, this is one maybe I can try and address from a peek. When is AI going to travel the galaxy on its own and create its own habitat on a different planet? You know, I think the thing to realize is computation is happening everywhere. We have built computers that do sort of follow rules and do things that we sort of built them to do. But there are processes in nature that are following rules and doing complicated things. We didn't make up those rules. Those came from nature, but they're still computational processes. Now, of course, at some level, we didn't make up the rules for the AIs. They came from the whole history of our natural evolution, so to speak. They are the inevitable consequence of the natural process of humans evolving and then doing their thing and so on. So this question of sort of what's it like to do, to say let's send out our computational systems and colonize the galaxy, so to speak. Well, computation already has colonized the galaxy. It's sort of all of those processes going on from nature are already out there. And I think that there's even more than we know. And the things that we know, we can tell are out there are the particular slice of sort of all of reality that we humans happen to be sensitive to. So I think this question of to what extent can we take the things that we've created in a way where we recognize what's going on, have them go hang out on the other side of the galaxy or whatever. I think it's sort of a funny thing because it's kind of like the ultimate. Thing where you're saying it feels a bit empty to just say, I'm going to put my signature somewhere and it's going to be taken and it's going to be sitting on Mars, for example, on some spacecraft. It's amusing, but it doesn't really, you know, what does it really do, for one? And I kind of think that's the same type of thing when we talk about sort of AIs formed in our image, so to speak, going and hanging out in the galaxy. It's kind of like we're sending our signatures out there. And it's a bit of a. So what? It'd be very different if the things we were sending out were utterly, qualitatively different from the things that are already there. But my claim is the things that are already there are already these complicated computations going on in the atmosphere of some exoplanet or something like this. And so it doesn't really. It's not like we're seeding the galaxy with something vastly different. The only thing we're seeding it with are the particulars of computations that we choose to do not. It's going from it's really dumb to it's really smart and computational. No, it's already smart and computational. And we're just sort of sending our signature, so to speak, to these other places. Now, you know, as a species, there's sort of this question of how do we feel about kind of sending our signature everywhere? And is that an important thing for us? You know, I have a friend who's been involved, I think he has another one going soon actually. The sending objects to various places in the solar system, hitchhiking on various spacecraft. The objects hitchhike on the spacecraft and he's been kind enough to send some things I've written in micro form, so to speak, on these various deposits. Is that really significant? Is that really to have some sort of microscopic sort of version of something, sort of a human produced thing sitting on the moon or something like this, does it? Or further afield, how significant is that for us? And is that kind of the. It's really very much a personal societal question is do we feel that planting our flag, so to speak, on Place xyz, is that really important or is it like a shrug? I think. I don't know how to think about that. I think it's. I mean, I suppose for me personally, I don't think it's that exciting, but maybe I'm just not built as an explorer of the right kind. I mean, somehow I also can't help but realize that we are, let's say roughly a meter in size. The universe is 10 to the 26 meters across. That's, that's how big we are relative to the universe. If we think about what's the universe made of in our sort of physics project, we have some indication that the elementary length in the universe, the shortest possible length is maybe 10 to the minus 100 meters. So there's a lot more room down than there is up where 10 to the 26 orders of magnitude away from the size of the whole universe, yet 100 orders of magnitude above the size of the elementary length. So in a sense, it's the sort of more stuff to go deal with sort of right here, sort of looking at what's on the small rather than out there looking at what's in the large. It's an easy wrap up question being asked, which is not easy for me. Have I ever had figgy pudding? I have no idea what that is. So the answer is, to my knowledge, no. But, you know, maybe that's one of the deficits I have from not having grown up in the US I'm not sure. Anyway, we should probably wrap up there. Thanks for joining me. Happy holidays. And I think I might be able to do one of these next week, perhaps late in the week. Otherwise, I will be doing this again probably in the New Year. So happy holidays. Thanks for joining me. Bye for now.