Future of Science and Technology Q&A (May 19, 2023)

Sam. Hello, everyone. Welcome to Something New Today. For the last three and a bit years, I've been doing a weekly Q and A about science and technology for kids and others. I've also been doing alternating weeks, Q and A about history of science and technology, and one about business, innovation and managing life. Today I'm starting a new series which will be alternating an alternate weeks with the Q and A for kids and others. And the new topic is Q and A about a future of science and technology. I've got every other week a history of science and technology, and every other week a future of science and technology. I realized I've now been doing a lot of these live streams. I have about 330 hours of recorded live streams, enough to train quite an LLM that will be coming soon, too. All right, well, let's see. We've got all kinds of questions here. Ollie is asking, aside from faster processing speeds, what are some other ways computers may be improved in the future? One of the things that I thought about a number of years ago now is actually I was visiting the archives of Gottfried Leibniz, who lived in the late 1600 and who with great effort made a four function calculator, mechanical calculator, out of brass and so on. And I realized at that moment of seeing his calculator, I realized in Leibniz's life, and he spent many decades actually trying to refine this calculator, he had seen one computer, the thing he made, and maybe he saw a few others from other people, but basically he'd seen on the order of one computer. In today's world, we have, I don't know, some number of tens of billions of computers around the world. And computers are common things that we see in various places. What I sort of realized at some point is that one day people will be surprised that there was a time when everything wasn't made of computers. What do I mean by that? Well, you know, you go back more than 100 years, every material that you had was somehow a material that in some form occurred in nature. Then things like plastics were invented which were sort of artificial chemicals that were sort of at the chemical level, constructed for particular purposes and so on. But those things are just materials. They don't compute in the kind of way that we think about computers as computing. And I think the thing that we can imagine is that in the future the kinds of things that we can program will become ever kind of smaller and more ubiquitous. And I think that one of the things that we see with a computer as of today Is there's a microprocessor, it's made out of silicon, it uses electronics, et cetera, et cetera, et cetera. There's clearly other ways to be able to compute other than by using electronics and silicon and so on. In fact, my own science work has kind of shown rather clearly, I think, with this principle of computational equivalence, that there's a huge range of different kinds of systems, different kinds of infrastructure that you can use to compute with, whether it's fluids flowing around in certain patterns, whether it's other kinds of things that are transporting molecules around, all sorts of different ways of basically being able to compute. And of course, we have one spectacular example of computation that happens differently from the way it happens in silicon with electronics and so on, which is biological life and molecular biology and so on. I mean, molecular biology I view as being sort of an example of computing where the working fluid is not electrons and electronics, but instead is all kinds of different types of molecules and so on that are being sort of carefully orchestrated to be. To move in different ways in cells and things like that. I mean, I used to think, well, inside a cell it's kind of a liquid where there are a bunch of molecules just bumping around. And you can compute with molecules just bumping around as well. But I think that there's what the increasing evidence from molecular biology is that things are much more orchestrated than one expected. And there are actual structures that sort of move molecules and arrange them in particular ways. So in any case, that's a sort of an example of where we have one case of molecular scale computing. I think that there are certainly others that do not rely on the kind of history of biological evolution that use the idea of molecules moving around and sort of tubes that transport molecules and molecules of certain shapes that fit in like certain puzzle pieces and so on. These are ideas that are sort of the meta ideas of biology that I think we can expect to eventually make technological make things that we can kind of produce in ways that we want to produce them. So I'm sort of expecting that there will be, in some sense it will be the case that there will be materials that aren't just sort of static things made of arrangements of atoms and molecules and so on that just sit where they sit. But things where there's actual processes that go on that might involve electrons might involve other kinds of atoms, molecules, whatever. There are actual processes that go on that we can think of as being computations that are computations that have been arranged to be things that we can sort of Imagine achieving goals that we have. So sometimes we'll be able to readily define sort of every step in what those computations are. Sometimes we might just define the overall goals and then it might be up to some, some system kind of arranging itself in order to achieve that goal. Like you can have something where there's a particular way the puzzle pieces fit together and you expect to have this system that's sort of bumping the pieces around randomly, just as happens in molecular biology. It's sort of bumping the molecules around randomly and the ones that happen to fit by shape geometrically are the things that arrange themselves that way. One can imagine the same kind of thing happening. But the main point is that there'll be kind of a sort of everything will compute and there'll be ways to kind of tell it what you want to do. I don't know, maybe there'll be something where you're sort of, you know, you've got some material and you're flashing light in some way at it or something to be able to. Or more likely you have something which is some kind of phased array of different things that are, is sort of sending photons at different parts of the material in different ways and so on. But in any case where the result is that sort of everything that one has will be made of computers, so to speak, it will be made of something that is computational. And I would say that in a sense biology is already doing that. Biological tissue mostly has sort of things that are non trivial things that are happening in it. It's not just sitting there, I mean, not true of all of it, I suppose, but some of it is just sitting there for structural purposes and so on. But plenty of it has sort of things that are going on in that tissue. So I mean, I think that's the big picture in the distant future for what will happen, I think with computers, it won't be that computers are discrete things. It'll be that one makes things out of sort of computational kinds of components. I think there are sort of nearer term things that we can see. Like one of the big ones is when we look at current AI systems, large language models, things like that, one of the features of those is that they do training and then you take the trained neural net and then you run it. And those are sort of two separate kinds of things. And as you run the trained neural net, it's not immediately learning more stuff. In brains, we learn more stuff as we run, so to speak. And part of the reason it works that way is that sort of in the training of a neural net, you're saying, well, there's this part that's being sort of processed in the CPU and most of the neural net is just maintained in memory and isn't changing. But it isn't the case that we have computers that are set up so that every part of the network is sort of actively computational. It's usually most of the network is just stored in memory where the only computation that's happening is just keeping it the same. Whereas, and there's a small part that's actually the part that you're sort of back propagating through or whatever else it is, that is the part where actual change is happening, where actual sort of non trivial computation is happening. My assumption is that in the not too distant future it will be possible to, to make kind of systems where one can sort of compute and remember at kind of the same time. Actually, I just recently found some stuff that I did in the, oh gosh, early 1980s. I was interested in making what I call meta chips, kind of old fashioned word for chips, microprocessors, whatever. And the idea was the following, that right now you make a microprocessor and it has a billion transistors on it and it's arranged like a giant city the size of the earth or something with all these roads and things where electrons go down them. And there are all these detailed features that have little pieces of memory for keeping things in a cache. On a CPU chip there are things that arithmetic logic units that actually do sort of arithmetic computations. All those kinds of things, they're all encoded in the detailed kind of giant roadmap that is the pattern, the circuit of the microprocessor. So the idea that I had in the early 1980s was, well, what if you just use one of these cellular automaton things that I've studied a lot, Very simple set of rules with the collection of identical elements where you just have this simple rule being applied. And it turns out that as a result of this principle of computational equivalence idea, even in lots of simple cellular automata, it's possible to make a universal computer. It's possible in effect to have a compiler that will convert any program you might imagine making into the initial conditions for the cellular automaton. For example, the Rule 110 cellular automaton is one where we know that works that way. It's just something where it's a one dimensional cellular automaton, just a line of cells, each one is black or white at every step. The color of A new cell is determined by some rule according to the color of that cell, the previous color of that cell, and the colors of. Incredibly trivial kind of setup. And the idea would be you make a piece of hardware that's incredibly efficient at doing that particular running that particular cellular automaton rule, and you can do all kinds of optimizations and so on, where all the hardware does is just a giant thing that just runs that rule and then sort of everything else that has to happen is a software problem. How do you turn this sort of computation that you want to do this, I don't know, image processing task, this compression task, this, I don't know what any other kind of task, how do you compile it into Rule 110 code, so to speak? And nobody has really developed, I've done a little bit on it and a few other people have done some things on it, kind of how that compiler would really work. We know in principle that it can work, but to make it at all efficient is a challenging thing. I mean, by the way, if somebody just came to you and said, by the way, all you ever need is the NAND operation, that's the only logic operation you need, and you put a billion of those together and you're going to have a microprocessor and it's going to do all these interesting computational things. At first you would say, how could that possibly work? It's just a simple NAND operation. It hardly does anything. How can you put it together to do complicated things? But this is one of the sort of big results of lots of science I've done, for example, that even when you have very simple components, you put them together, it is possible for the overall behavior of the system to be very complicated. We see that again with LLMs and ChatGPT and so on. Another story, another repetition of the same kind of story, that it's very counterintuitive what can emerge from putting together even very simple components. But any case, the idea of these kind of meta chips is you kind of put everything, even the design of the sort of everything is in software. The actual thing that is the microprocessor is something incredibly simple, by the way. I mean, in the trend in microprocessors in general, there was an early trend that was, oh, you put all kinds of complicated instructions and we encode that. Originally it was in circuitry and microcode things like that. Then it was in microprocessors of design. We encode all these complicated operations that do polynomial multiplication and all sorts of things. We make Instructions that do that in the microprocessor. And then in the 80s, I guess, this idea of risk, reduced instruction set computing came in, where the idea was, let's have much simpler instructions on the microprocessor, but then use a compiler to just take the. The big things we want to do and grind it down into these simpler instructions. And it turned out you could make very efficient compilers that do that. And in the end it was a bigger win because you had to have a simpler microprocessor, you could make it more optimized, and so on. I think the pendulum has swung a little bit back as we see GPUs and things like that, with a lot of complicated operations and instructions on them that are being used. But kind of this metachip idea is the ultimate risk. It's kind of the thing that's there, just this one operation that's incredibly simple, and everything else is sort of up to the compiler to make work. So that's another direction we might see things going. Although it's always a complicated thing because one has an idea that says, okay, we'll make a more. A general trend has been more stuff gets done in software, less stuff gets done in hardware. But then occasionally people realize, well, there's certain things that one typically wants to do in the hardware. Let's just build hardware that does that. And then that tends to lock in a platform, a framework in which one then operates, which one then does things with software. So even if the framework wasn't quite right by the time it's kind of been optimized in the hardware, that's kind of how it evolves. Well, I see we have lots and lots of questions here, so let me try and keep going here. Readitloud asks, Will we use books in five to 10 years, or will it be replaced by chatting with an AI? Interesting question. You know, when the web kind of developed in the 1990s and beyond, I wondered that question. I wondered, you know, when do I want a website? When do I want a book? What I realized about books is they're actually a pretty good form factor, and they're pretty good. When you pick up a book, you know, whatever this book author or whatever, whatever they have to say, it's in this thing, it's in this limited object that you can hold in your hands. When it's websites, it's kind of like, well, I can click on this page and I go to this page and I go to that page and I'm exploring and I don't know how, how far the octopus goes out. It's kind of a thing where I can't get this sort of feeling of there's this, there's this sort of thing that I can kind of ingest the knowledge from in kind of one go. So I think that having this notion, I mean, for me I've been surprised at the extent to which I really prefer sort of larger scale things as books where I can pick the whole thing up, I know what it is, I can flip through the pages and so on, rather than it's a website, it's got hyperlinks to here and there and the other. Now in terms of is it going to be useful to be actually sitting and reading a book? You know, I think I'm probably a pretty bad example of being a systematic reader of books from beginning to end. I tend to be a try and get the information I want out of this thing. And so, and that's something where one can imagine there's sort of an AI that is the interface to the book. You know, I, I view what we have with ChatGPT and so on. The big thing that it's introducing is this kind of notion of a linguistic user interface. And I think one of the things that is absolutely the case is that one will be able to take sort of book type information and say, or a report somebody gave you and you can ask a question that is the thing that you actually care about from that report and say, okay, tell me the answer to this. And this linguistic user interface mechanism is going to be able to give you the, here's what you might care about from this. And I think one thing we'll see is as these systems get merged more and more with kind of our personal histories, it will be more and more the case that somebody shows you something that there will be an AI that can say, here's why you should care about this. Here's how this relates to something you've thought about or written about or whatever before. And that's something where when people send me things and they say, here's the thing, it's interesting. And I'm like, it's a lot of work to read this. If somebody could tell me how this relates to things I already know and how to plug that in to the things that are already sort of lodged in my memory, that would really help me. And so I think that's the thing we will absolutely see from AIs is they get to know us and that gives them a way to explain things to us in particular, not to the general reader, but to us in particular. Now that works when you know what you're asking. When you don't know what you're asking, and the question is to learn something completely new, then that is not a successful approach. It's something where I never knew to ask about that, so I'm never going to ask about it, so I'm never going to learn about it. So I think that's a place where again, the here's a thing, it's a book, for example, and you can kind of flip through it and you kind of wonder, oh, what does that mean? What is that sort of picture about? Now maybe the AIs will learn how to show us sort of a simulated flipping that is useful to us in particular. Possibly that will happen. But I still think that for me it's been surprising that even in this time of sort of everything can be represented electronically, computationally, digitally, whatever that I still find myself finding books really quite useful, even things I've written myself. I will sometimes look at the book version of them in preference to an online version, even though the online version is nice and searchable and it's this and it's that. But somehow the book version has a certain browsability that is perhaps helped by the fact that, well, yes, we are sort of physical entities with fingers and this kind of thing, and there's some pretty good kind of interfaces that are provided by sort of our fingers, ability to flip through a book and so on. And yes, one could certainly imagine that in the sort of future world of extended reality and augmented reality and so on, there'll be other ways to do this as yet unimagined. I think, though what tends to be the limiting factor is in the end you're interested in stuffing all this stuff in your brain. And for now we've just got brains as, as they are, with eyes, as they are and so on. And so you kind of have to have to sort of interface whatever that that ocean of knowledge is with the actual apparatus that we have. Let's see, There's a comment here from bm. I think about things moving towards narrative driven AI powered procedural generated VR environments with sort of metahuman characters interacting with AI speech and so on. Like Create me a movie experience. Yeah, I think the. It's really an interesting question what the best way to learn something is as a human. You know, I think one of the things we're going to see in the pretty short term is vastly more personalized education of okay, I want to learn something. Okay, AI Teach it to me, AI, you know lots about me. You know a bunch of stuff that I already know. You can also probe what I know and what I don't know, and you can kind of map out this is the exact right thing to say at this moment that will try and communicate this point in a way that will fit in with my, with my brain. I mean, like, if you started telling me about, oh, I don't know, what's a field I don't know terribly well? Economics, let's say. There's certain things I know in economics, other things I don't know. There's things that I've sort of understood and really ingested, and there are other things where I, where I haven't. And if you knew what I knew, you would be much better equipped. If there's some particular idea about, I don't know, the, some particular economic situation or something where it's a question of, can you explain that to me? And, you know, you start talking about supply, demand curves, okay, I understand those. You start talking about the, you know, Smith, Jones utility function, whatever, which I've never heard of. That doesn't really help me very much. But if the AI knows what I know, in a sense, it can in a very, even, even at the simplest level, kind of customize its explanation for, to fill in the parts to, to reach those kind of islands of, of dry land that I can kind of, you know, that I can successfully stand on, so to speak, and, and not get. Get kind of mired in the ocean of things that I don't understand. So. And I think that will be true for everybody. I think that this whole sort of mapping out of what we understand, what we don't understand, what we already remember, what we don't, kind of the interface to the contents of our minds, that will be something that AIs will do quite well. I think this question of whether it's better to kind of have the AI just write your piece of text or whether it's better to synthesize kind of, you know, meet Socrates in person type thing and, you know, have the discussion right there with the embodied Socrates in virtual reality or something, I don't know, that might be very compelling. If it gets to the point where it's sort of almost realistic, it could be a very compelling experience because we are, after all, evolved to exist in embodied form in the actual physical environment and so on, and sort of there's a lot of sensory information that is coming to us that isn't well captured by just reading words on a page. So it could be a very, very compelling thing. I think when it isn't quite right, it's kind of in this uncanny valley situation of being kind of. Well, it's almost to Socrates talking to me, but the Socrates is kind of has. Is being very banal and not very convincing. And I'm not getting the Socrates vibe, so to speak. And that might not work very well. And it's a question of when does it get to the point where it really is compelling and do people lose interest before it gets compelling? I mean, in a sense that's been the story of virtual reality for the last, I don't know, 30 something years. Is it keeps on getting to a certain point, but it isn't that compelling. And then people kind of lose interest and it goes down into a, into a kind of a trough again. And we'll see when it, when it finally gets to the point where it really is sort of convincing and useful in that way. Let's see. Parmenides comments. Can't read books on the computer beyond 200 pages. Too much eye strain. Yeah, I wonder about that. I mean, I also, I mean, I think part of the story of eyestrain computers, et cetera, et cetera, et cetera. I mean, in a sense it's photons coming to the eye and it's details of your computer screen. And there are a lot of details that really matter, like for example, the actual formation of font characters and whether they're little bits cut out of the font characters and so on. It really makes a difference to readability of things, particularly sort of long term readability, these things that you barely notice. I think also there are all sorts of detailed things about refresh rates, about, you know, the actual colors on the screen, et cetera, et cetera, et cetera. I'm not sure I've personally, you know, I think I made some effort to optimize these things for myself, but I haven't really done a hugely scientific study of what makes for optimal reading. For example, I don't even know, you know, I'm not particularly a dark mode user. In fact, I really haven't used it much at all. For me, dark mode is what I used to use back until about the mid-1980s. It was dark mode only because that's all you had. That was what a cathedral, ray tube, crt, you know, display. It was illum, you know, characters illuminated on a black background. And I find it amusing that for the current generation, you know, the younger generation, it's like dark mode this is really cool. It's new. Well, actually it's what it originally was when computers first existed with display units and so on. And so I haven't yet sort of gone back to the past, so to speak, for dark mode. I don't know whether it's faster to read things in that way or not. I think those are things I haven't optimized. But my guess is that there is a way for any given person, and maybe it's different for different people, of presenting things in digital form in a way that's sort of optimized for readability and not giving oneself eye strain of some kind. I mean, I know, for example, with virtual reality, it's still the case that, although I admit I have not tried in the last year or so, but at least as of a year or so ago, when I use VR systems, even the fine, highest NVR systems, I was still getting motion sick. And I think that's to do with the tracking of whether there's a lag between image moving and head moving and so on. And that's really a question of the hardware and so on. And it will presumably gradually get better. Let's see, there's a question here from Math Online. Are we close to imitating senses of smell and touch in VR? Interesting question. One of the things about vision is that we have three kinds of color receptors, plus one more kind of just black and white receptors. That means when we want to show a color scene, we can do that by just picking the level of red, green and blue for each pixel. But if you want to make a texture, you want to make something of a certain texture, it's much more complicated because we don't seem to have that limited number of texture receptors. Same with smell, same with taste. Let's take smell as an example. There are certainly hundreds of different kinds of receptor cells, chemical receptors in our noses and so on. And what's roughly happening in smell is that molecules in ways that are not fully understood, I would say, bind to the receptors. And, you know, you've got a bunch of molecules that are the smell of a lilac or something, and that's a whole cascade of, you know, millions of molecules. And they get in our noses and these molecules bump up against these different kinds of chemical sensors. And to some of those chemical sensors there'll be. It'll be a match. And it's probably a sort of a physical geometrical match. Probably those pieces of lilac scent or whatever that are probably will be things that physically are of the right shape. To fit into a certain molecule that is part of the smell receptor in your nose. And when the thing fits in, that produces a signal which ultimately turns into an electrical signal in a nerve cell, and off it goes to the olfactory bulb in the brain and a new sense of smell. But what's happening there is there are a lot more types of receptors than there seem to be for color. I mean, for color, even if you have that special blue that is made with this particular spectrum, doesn't really matter to a human eye. What matters to a human eye is just how much red, how much green, how much blue. And that's why when you look at a computer screen, you can see it's just made of red, green, blue pixels. But in the case of something like smell, you have to have this sort of. It's like, could you make this collection of standard shaped molecules? And you can say, I can make any smell by having this particular combination of molecules? You know, back in the day, in one of those, it didn't really work out very well. I think there was the smellarama type of thing for movies and so on, where people would have. Where they would be. Have some, some kind of chemical kit that would sort of open itself up at the appropriate time in the movie and you'd smell something. And I think those had a certain set of. Of chemicals in them that would produce certain kinds of scents. But when it comes to sort of sculpting a scent or sculpting a smell, it's really pretty difficult. I know perfume manufacturers, for example. They'll have hundreds of different kinds of molecules in their sense. And it's not like with color where it's just like they're three things and you just have to get the right concentration of each thing. It's much more complicated than that. Same with taste. Taste buds in the tongue and so on. Are these pretty complicated kind of receptacles where particles of food get in there and there actual chemical reactions that are going on in there. And some of the effects of those chemical reactions are what lead to the sense of taste. And I don't think it's fully understood how that all works and to what extent kind of the solubility in water matters, et cetera, et cetera, et cetera. So that's why kind of, when you kind of are making foods and so on, there's a. There's a kind of a very wide range of, you know, when you, when you pick a color, so to speak, you can in principle just move around the red Green, blue sliders, and you're going to get that color, and there's no elaborate preparation required for the color, so to speak. Whereas with food, when you're trying to make things sort of at a molecular scale, you're trying to make a thing that has a certain taste. It's a really complicated story. And, you know, in the end, the sort of molecular gastronomy kind of idea of let's even understand it down to the molecular level, but that isn't even enough because it's not even the individual molecules. It matters what the texture of the food is. It matters sort of how big the particles are. And that presumably has to do with how it fits in the tongue and so on. Now, when it comes to touch, touch is also really complicated. And you might ask, are there certain kinds of textures, Other sort of standard kinds of textures? Oh, it's rough. Oh, it's slidy. Oh, it's furry. Oh, it's whatever else. Are there those kinds of limited number of degrees of freedom for texture? There's been work done in recent years that kind of suggests that, yes, you can reduce it to a limited number of dimensions of kind of texture now. And then the question is, when your finger moves over that texture, what's actually happening? And it's kind of sticking, slipping on these different parts of this kind of mountainscape that is the texture and so on. And I think then the question is, okay, is there a way to sort of make a device that will emulate that texture? For example, you might even have some MEMS or something where you physically are moving little, tiny, tiny things up and down to make an actual texture. But actually there are electrical ways to do that where you can get the sensation of something with texture. And I guess this was used for a while on smartphones, but I don't think there's anything that uses that right now. Maybe I'm wrong. Maybe there's a small such effect that I don't even know is there that is helping one when one's actually touching the surface of an iPhone or something. And I know, I think in the distant past, when the original way of determining just where are you pressing on that iPhone screen? That was a thing that I happened to know. Somebody used our Wolfram language technology stack to do the modeling for that. And it's kind of this question about, oh, you've got this big pudgy finger that's poking the iPhone screen. What actual place is? The place is if you've got this whole array of letters, which letter is the one that you most think it's likely to be pressing. And that's a whole story of kind of estimating that based on the kind of distribution of squashed finger, so to speak, on the iPhone screen, and how that is affected by the grid of little wires that pick up the electrical signal from that. In any case, there is a mechanism for kind of a feeding back electrical signals that gives one the sensation of a texture. And I don't know how successful that's been the. Let's see. So this question of how this will work in VR. Good question. I mean, I think that, you know, we can get a sense of how difficult certain kinds of senses are to emulate. So, for example, if we just look at our brains and we look at the number of nerve fibers that go from our different sensory organs, you know, from, from eyes, it's like 10 million fibers. From the ear, it's like 50,000 fibers. And for us humans, the number of fibers going from our noses is smaller than that. I Forget how many. 10,000 maybe, I'm not sure. For a dog or something, it's immensely more fibers going from the nose than from. I don't know about the ears for a dog, but the dogs have huge numbers of nerve fibers that are kind of sending data from their noses. And so that gives one some sense. Okay. Historically, you know, 1990s, we got fairly as if you were there, audio, you know, you, you sample audio at 44 kilohertz that's sampled high enough, it's above the range of human hearing. You can kind of get every single kind of pressure sensation getting to your ears that you would have if you were actually there, so to speak. We're not quite there yet in terms of eyes, we're getting really quite close, but we're not quite there yet. We've got to hit that. It really is enough that all those 10 million nerve fibers are being fed the same way they would be fed if you were just looking out into the, into the everyday world. And one of the big things with VR was just getting the angle. You know, as you look, your eyes have a fairly wide field of view. And, you know, just getting a display that wasn't just in front of you, but also a little bit to the side was an important thing. Even though the density of, of nerve cells on your retina, photoreceptors on your retina goes down as you go away from the center, you still have some sense of what's going on in the periphery. And that was important to capture. So I don't know, for some of these other sensors, like touch, for example, it's unclear how good you have to get it in order for it to be kind of. Oh, yes, that's realistically like picking up a rough, you know, I don't know, an avocado or something. You know, it feels like I'm picking up an avocado because it has sort of the same level of roughness as that. Let's see. This is asking, saying they lost their sense of smell due to brain imaging injury in 2016. Any realistic way this could ever be fixed? You know, I think one thing about brains, which we're also kind of learning about LLMs is brain tissue is kind of. It's sort of a bit universal in the sense that you've got all these neurons. They're all connected together. Yes. It tends to be the case that particular functions are localized in particular parts of the brain. And that probably happens, well for different people. Functions are localized in different places. Is like the language center of the brain is never in the same place for two different people. Which is why when people are doing neurosurgery and they're worrying, you know, where is the language center? We don't want to mess with that. It's always something. We have to actually stimulate it to find out where it is, because it's just a bunch of neurons, like all the other neurons in there. And I think the thing that is an interesting question for LLMs, image generation, AIs, things like that. In fact, I've just been doing some experiments on this myself is what happens when you give them brain lesions? What happens when you zap some neurons inside a trained LLM and what you find is, oh, it's really weird. I mean, it's really weird what you find in terms of images. It produces these weird images and. And then eventually it kind of fades to human incomprehensibility. But I think what tends to be the case with. With brains is that it really is. Neural tissue is in some sense fairly universal, at least in the main cortex part of the brain. And that means that it is possible for people to remap. And that happens when people lose their sight and things like this. They remap parts of the brain to take over some neurons to use for a different purpose. I think that the place where things are difficult is when the actual nerves that go from sensory organs to the brain are damaged in some way, like the auditory nerve or the optic nerve or something like that. In that case, really, the only way that you can get the Data there is to stimulate directly that part of the brain without, without going through those, those. Those damaged nerves. So, for example, for vision, that will be simulating the primary visual cortex in the back of the brain. And I think there have been, there's been. Well, there's. There's definitely some success at sort of a low resolution, because in the case of vision, there's a pretty direct geometrical mapping between that spot in your visual field and a particular place in your primary visual cortex. And so you can expect to have some plausible mapping of, oh, yes, there's this visual scene that's being seen by some electronic camera and then that's stimulating some particular part of the brain. Now it gets more difficult when there are things like smell, for example, where the mapping from what the chemical senses to where that goes in the brain is more complicated. There's no obvious geometrical mapping, there's no obvious geometrical way to represent smells. This is one of the questions of how high dimensional is smell space and how many parameters are really important there, et cetera, et cetera, et cetera. By the way, I should say that when it comes to making an artificial nose, that's been a long running kind of can anybody do this type question. And I just saw some stuff very recently of another effort at this that looks quite promising, but it's been a long road and it's really difficult. And one imagines that sort of the way that one would do it is using something that's like a microarray or something like this, like a. Something which has a bunch of little pieces of protein attached to different sites. And then you look at how much binding there is of the molecule that comes in to those different cells. In fact, now that I think about it, I've seen two demos recently of things related to this where it looks like that's beginning to work, so to speak. And it's very relevant. I mean, ultimately, the kind of can you smell it? Type thing is, while at the point where one's still using dogs to sniff out explosives and things like that, it feels like technology has not reached. You know, there's a bit of a gap with technology there. The dog's nose should be able to be something that can be made into a pure piece of sort of digital technology. But that hasn't quite happened yet. Let's see, Let's see. Gosh, so many questions here. Frost is asking how much of biology is untapped. They say they're in the biology, biotech, genetics, metabolics field. And it feels like most Researchers never leave the lab. You know, the amount of sort of information about how biology works that's been accumulated in the last few decades is unbelievably large. The amount of kind of big picture theory making that's happened is in comparison unbelievably small. In biology there are a few sort of big meta theories. The theory of natural selection, the digital data, characteristics of DNA. I suspect there are a few other kind of big theories. For example, another sort of semi big theory would be something like the operation of the immune system, which we more or less know, although I think maybe we don't have it quite right yet. But I think one of the things that I suspect is that there are some other aspects of sort of big theories of biology that still need to be established. I think the most important are probably around dynamic aspects of biology. That is, there's an awful lot that's known about the structure of this or that kind of cell, the structure of this protein and so on. There's a lot less known about the dynamic processes of the interaction between these things. So for example, and people tend to say whenever there's a thing that one has to find in biology, there must be a static object that's somewhere out there in the biological system that you can pick up and say this is a, that. Well, let's take the immune system as an example. I mean, there's the question of how does one remember an antigen that one was exposed to? How does one remember how to mount an immune response to that? And sort of a prevailing theory has been, well, you sort of parade all these antibodies, 100 billion different possibilities in front of each androgen. And the ones that are a hit, those are the ones that get amplified. That's clearly correct. Then how do you remember which ones you were going to amplify? Well, maybe there's a memory B cell or something that is the type of cell that would produce the antibodies that you need. And maybe for your whole life you have certain ones of these memory B cells that are sitting somewhere waiting to spring into action. It's like there's a physical cell, you could pick it up. It's a way of explaining what's going on. My guess, and this is kind of following on this so called network theory of the immune system from the 1970s, is that it's a much more dynamic kind of thing where, you know, you produce antibodies, maybe you produce anti antibodies, you produce anti, anti antibodies. You are in this kind of dynamic equilibrium of producing sort of an immune response that produces an anti Immune response that produces an immune response, et cetera, et cetera, et cetera. You're in this dynamic equilibrium where no individual cell is living for 100 years, but instead there are all these different cells that are being produced and then those ones are dying and other ones are getting produced and so on. It's kind of you're remembering, remembering rather than that you are sort of having something which goes all the way back to the beginning as a single physical object you can pick up. So my guess is that the kind of dynamics and biology of sort of this graph of causal relationships between things, this kind of dynamic structure of what's going on, that this will turn out to be a pretty important part of the story of biology, but it's one that's as of now, completely untapped. We don't have a way to measure kind of the causal graph of the interactions between biological molecules. We have, thanks to things like PCR and so on, we have extremely sensitive ways to measure the particular static sequences that exist on RNA and things like this. But when it comes to sort of the dynamics of individual molecules bouncing around, we don't have good ways to measure that. Maybe there'll be a way. I tried to think about this. Maybe somebody will invent it. I encourage people to do that. That's kind of the analog of PCR that amplifies up a dynamic process, a dynamic through time process. And one can imagine, there are ways one can imagine that with sort of autocatalytic sets of molecules. And you say, I've got this, it's kind of almost like a, a When you have a radio antenna and you are, that radio antenna is, is really good at having electrons circulate in a particular way. And when the electromagnetic wave comes through and it sort of can excite the electrons in that particular pattern, then you detect the signal. And one could imagine the same type of thing with molecules where there's a particular sort of self, self closing loop of reactions, where somehow, when, when there is a dynamic process going on, it somehow excites that loop of reactions which one can then detect. I don't know, that's a speculation about how that might work. But my point is that there are these kind of dynamic processes that we don't really have a great way of thinking a bulk about. In fact, our physics project has given some ways of thinking about bulk parallel causal processes. And I think some of those things are probably applicable to molecular biology. But when it comes to kind of what can we do with biology? I think one of the Things is biology is the sole example, as I was mentioning earlier, of molecular scale computing, and it's one of the drivers for making nanotechnology and things like that is if we have molecular scale devices, they are easy to interface to the molecular scale stuff that is our biological character. I think there are so many things that will happen. I think that algorithmic drugs is one obvious thing where you think about a molecule that is a drug, where normally a drug is just like the static molecule that binds to a particular place or just sticks on in some particular way. But imagine that you could have a molecule that would essentially sense its environment, make changes to itself, compute something and then say, yeah, that's a really bad cell. I should go zap it a little bit like the sort of cascade of molecules in the immune system does. But the, but imagine that you could make a sort of, you know, this is this question about, you know, everything is made of computers. Imagine that you could actually make this drug molecule that has, you know, all these little pieces and it's, it's computing something, it's sensing things, it's computing what it should do. That's an example of something one can think could could happen in the future. You know, I do think that in terms of kind of the future of biomedicine, one of the very obvious things there is, it's all about sensing more stuff in a sense, a lot of things where, you know, in biomedicine sort of a goal for most of us is, you know, stay healthy and keep going the way we've been going before. Because if we're, you know, if we, if we continue to go as we've been going, it's all good. And then the question is, is anything going wrong? And you know, we know in the case of, for example, commercial airplanes, that's sort of airplanes in general, one of the big changes over the decades has been just tons more sensors. And so, yeah, you know, in the past it's like, well, yes, there was a crack forming in this place, so there was a thing that was going wrong here or there, and you don't know until the thing kind of blows up in your face. But in modern times, there's just lots and lots of sensors and lots and lots of ability to, even when there aren't so many sensors, to be able to take kind of broad spectrum data from the sensors that are there and use things like machine learning, use signal processing to try and sort of fish out of that kind of fairly noisy data that's coming in. Things that are signals for example, something going wrong. And my guess is that that in, in biomedicine, one of the things that will happen is sort of this increasing. Oh, you know, you get a drop of blood. I just done another set of these things because I always, I'm a, I'm a sucker for any kind of new technology of this kind of, you know, from the drop of blood, it's like measure. I don't know, the latest mind had measured like 500 things. I think knowing what that means is a separate challenge. And that mostly is. But one thing I think in individual biomedicine is if you get the time series, you're probably coming out ahead. Because if the time series is like, it's all flat, nothing changed, it's probably good. Now. My guess is that things are more dynamic than one expects. And when we get into this level of detail, things will always be changing. That seems to be true for the microbiome. It's probably true for these detailed metabolite, kinds of metabolic, kinds of things that you can metabolome, kinds of things that you can measure as well. And so there needs to be a bunch more kind of learning about these kind of dynamic processes and biology to make progress on that. But I think there's, you know, in terms of the sensors and measuring things, there's clearly a lot of opportunity there. There's clearly a lot of opportunity. Once we have better sort of nanoscale devices that we can make those become sort of the interventional story, so to speak. I mean, you know, there's always the question, and it's complicated ethical questions around, okay, you have the, you know, you have the capability to do this. It's very personalized to a particular individual. It's computing what it should be for that individual. How do you do the standard double blind clinical trial test on something like that? Really difficult. And what do you do? How do you deal with the fact that you're doing something and you don't really have. There's so much diversity in what you can do that you can't really do a broad test. I suspect that this very fact that you can measure more stuff will help that story. Whether that will be understood at a regulatory level quickly, I don't know. But it does tend to think, one does tend to think, you know, if you try, you know, a little bit of this weird drug that's supposed to do some amazing thing for whatever, or this weird algorithmic drug that has a whole complicated algorithm for doing this or that or the other thing, and you're measuring, you know, 10,000 kinds of things about your, your biological operation. And you take that, you know, normally it would be like, well, I don't really know if something went wrong. You know, I took a little bit of this algorithmic drug and I don't know whether that's going to cause me in five years to start growing purple hair or something like this. But if you start to be able to measure 10,000 kinds of things pretty accurately and you can see, oh, when I wasn't taking this drug, the level of these hundred things was exactly this. When I started taking it, oh, I saw a bunch of changes. Then it starts to be this question, well, what do those changes mean? Are those changes detrimental? Et cetera, et cetera, et cetera? And that will be a complicated story because I think that this whole phenomenon I call computational irreducibility, this whole phenomenon of even though you might have a pretty good model for how bodies work, it's really hard to kind of run that and figure out what's going to happen so that these things where you see all these levels are changing. And maybe in 20 years when, if you keep doing whatever you're doing in 20 years, those levels are going to cause some horrible disaster for some organ or something. But it's very hard to predict that. But again, perhaps what will be increasingly the case is that it's more a kind of a real time thing of, you know, you try and, well, it will require some theory, I think, of figuring out, you know, oh, this thing went off track. Is that a problem? Is that not a problem? You know, maybe that's actually good that it went off track. Maybe that makes you a superhuman, so to speak, to have that be in this form. Anyway, let's see. Yeah, I mean, you know, my basic answer to, you know, what will happen with biology. One of the things to say is that, you know, fields go through these kind of sort of periods of great growth when some new methodology is invented. Then they kind of level off and they go into the cruise, Cruise mode for 50 or 100 years or something. You know, I think biology has been very energetic in the last few decades as sort of the methods of molecular biology really came online. I think the theory part has been really lagging behind in my view. And that's partly a consequence of the very size of a biological enterprise and the fact that its kind of value system is very much against theory and purely, well, what can you measure? You know, what are the kind of experimental results? What are the kind of unbiased experimental results? So to speak rather than ones informed by some theory. But I think, you know, things like gene editing, crispr, these kinds of things, that's a. Yet another kind of major kind of delivery of a thing or pluripotent stem cells. They're just a whole bunch of these kinds of remarkable sort of broadenings of what's possible or now, you know, RNA vaccines, things that, things like this that are, you know, things that kind of expand the toolkit of things that one can be making use of in biology. And it's now sort of the, the development of what's possible with these things. You know, it's. It's unclear how quickly it will go. In most fields that I see that involve, it's kind of like a 5 to 10 year story. When there's a new methodology, it gets worked out in that period of time. In biomedicine, there's a complicated driver which is the whole regulatory approval process and testing and so on, which extends those times, I think. And so there are things where people, you know, for example, immunotherapy for cancer people have sort of believed that's going to work for 50 years, but it's still in fairly early phases in many ways. You know, I'm sure that there'll be a lot more that's discovered about that and eventually people will say, oh, we finally got it, we finally know how to do this. But it took a really long time. And part of that really long time had to do with, oh, you have to do these tests. Maybe the effect is quite subtle. You have to really. I mean, one of the things that's difficult, actually interesting thing, it's a little bit like natural selection. So, you know, if biology in evolution is trying to innovate, it can have a really hard time. If there was a case where somebody discovered wheels. When I say somebody discovered wheels, I mean some organism, some type of, I don't know, some type of reptile discovered wheels and was able to just like wheel itself around with the greatest of ease, so to speak. The problem is it probably dies before it makes it to full wheels. That is when it has a sort of partial wheel that more or less turns. But it's getting this nerve bundle all tangled up because it's turning around. Those reptiles will die out. It never makes it to the sort of high innovation. It's an axle and a wheel because the intermediate stages don't allow it to. It can't make it past those intermediate stages because it's always got to be alive and reproducing Even as it's going through those intermediate stages. And I think some of the same kinds of things may happen in medicine and medical testing, which is maybe there is some incredibly clever way to do, I don't know, cancer immunotherapy or something like that. But to get there involves sort of this radical change which might be informed by a theory, if one had one, that involves a radical change and where to get there in incremental steps where you're doing, oh yeah, we're going to do a clinical trial of this. Well, no, that's going to fail. Just like the reptile with the partially established wheel and axle is going to fail. It's only by going far out where you're going into a territory where you kind of don't know at all what's going on and where, where if you get it even slightly wrong, it's just going to fail horribly. That, that's a bit of a difficult thing to reach by this sort of incremental approach. So, you know, that's one thing that may, may come up. I, I don't know whether that's, that's a feature. Let's see. Okay. Spare Parts is commenting. Do you think deep neural nets can help us build models of the human perceptual systems with vision and audio? How do we solve the problem of getting accurate training data for subjective experiences? You know, I think that there's sort of increasing evidence even in the past few weeks of kind of a correlation between sort of the FMRI parts of the brain, you know, pieces of the parts of the brain that are kind of using more power when you, using more oxygen. When you're, when you think about a particular thing or look at a particular thing, there is a certain physicality that seems to exist of, oh, for you, for this particular person, they've encoded their information about chocolate or something or the, or the taste of chocolate in this part of their brain. And so when they think about that, that part of their brain lights up. I've always been a bit suspicious about those kinds of claims that you can really localize in that way. Actually, I really want to know for LLMs to what extent one can localize the thinking of an LLM when the LLM thinks of chocolate, which weights are really lighting up. That's an experiment I, I still have to really do. But, you know, it could be the case. It's actually a good piece of artificial neuroscience, so to speak, to look at an LLM and say, actually it's a good, good idea to just take an LLM and look only at the central part of the LLM, only at layers. You know, like in ChatGPT it's maybe 400 layers in the, in the artificial neural net. Maybe look at layers 200 through 300 and just say looking only at those layers. Could I tell the thing is thinking about chocolate and is about to write some giant thing about how wonderful chocolate is, so to speak. Could I tell that from looking at the sort of the innards of the neural net? And how does that. And if the answer is yes, well maybe it suggests that we have a fighting chance to be able to do that for brains and be able to sort of pick up what we're thinking about that way. I think this question of well, what really is the subjective experience of this or that thing? Perhaps, I mean, I don't know to what extent the kind of just watching what part of your brain lights up is going to give you information about what one thinks of as the internal subjective experience. I mean, I think at some level, I think in the end it will be we'll sort of know what subjective experience is. It's kind of this complicated pattern of bits, so to speak, that some part of our brain experiences and somebody else will have a different one. You know, an interesting experiment you can do with GPT4 is to ask it to take something, some piece of text and compress it in a way that it is going to understand and it'll produce some very compressed representation, you know, you in terms of some random characters. You can tell it to use emoji if you want and then you say uncompress it and it will produce something that is not, not word for word, typically the same as what went in, but has the same kind of ideas. And it's similar to the way that autoencoders work where you're kind of saying let's take this, let's, let's extract the degrees of freedom that are really important in this information, compress it down and then you can expand it back up again because you're capturing the parts that are really important. It's sort of notable, it seems like GPT4 can do this but GPT 3.5 can't. It's like it got some kind of self awareness of being able to encode something for itself that it didn't have with a smaller number of connections and so on. But exactly how that works I don't know. I think. Let's see, there's a question here. About, from Internet professional about technology and science means nothing until we can chat with our cats and dogs. Will this ever be possible? You know, I've been wondering and chasing that particular question for a long time. It's, I think in modern times of LLMs and so on, this is the moment when we should be able to solve that problem and when we should be able to kind of, well, a thing that I was interested in, oh, 15 years ago, 10 years ago, I forget, sort of even got started on a project of making an iPad. It was right after the iPad came out, so 2010ish that making. My goal was make a game that a cat can win against its owner and or alternatively make a Minecraft for cats and dogs or cockatoos or whatever else. Something where the, you know, some non human animal can create a thing. I mean, we humans have built for ourselves all these means of creation, from keyboards to paint, brushes to whatever else, which aren't particularly an easy form factor to use for a, for an organism that doesn't, you know, that has paws rather than hands or whatever else. So the question would be, and maybe has simpler vocalizations than we have. And so the question would be, what if you could construct such a thing, if you could figure out, you know, if you give the cat a means of creation, so to speak, to make its painting or whatever else, what painting will it make? And you know, we might just say, well, the cat made this random painting. I don't know what it is. It's just a random painting. It didn't mean anything. Well, you know, you can go back and look at cave paintings that our species made, you know, 20,000 years ago, and they're just sort of splotches of this and that and hand prints in different places. What do these mean? We don't really know. We don't know if they're kind of random or whether they have some definite communication that is now lost to us today. And so similarly, I think the non trivial question is can we correlate whatever the cat makes or something with things that we understand now it's worth remembering that let's say dogs with their very sophisticated olfactory senses, a lot of their world is surely encoded in what it smells like, so to speak, not for example, what it looks like. So for us, even if we could sort of talk to the dog, the dog will be describing its experiences in large measure probably in terms of olfactory space rather than sort of physical space. And so then the question is, well, how do we have that conversation with the dog, given that it's talking about things that we really don't have an understanding of because that's not been our sensory experience of the world. Now, even things about which we don't have a sensory experience of the world, we get to learn about them. You know, we built telescopes, we built microscopes. Those give us a way to extend our sensory domain. And eventually we get intuition about what things look like under microscopes and so on, we can talk about them. And presumably the same thing will happen with smell and so on. I mean, even at an even simpler level, we already have perfectly good infrared sensing devices that let us sense sort of, you know, heat and other things outside of the standard visible light spectrum. And. But it's not clear how to feed that information to us. You know, you can have a sort of, you know, a thing that shows you the infrared image of things and you can make brighter the things that are hotter and so on. But there's sort of color in the infrared. There's just like this red, green, blue, as sort of bands of visible light. So they're perfectly, well, bands of infrared light. But how we communicate that to our brains is not clear. I mean, nobody has yet made the kind of obvious extended reality interface for infrared, for example. But when it comes to the olfactory extended reality, that's even more challenging. And so then the question becomes, okay, if you're going to have a conversation with the cats and dogs and so on, what are you going to talk about? One of the things that I think is the case is that there is, at least in something Darwin pointed out, there is a emotional connection. The emotional, the expression even of emotions in animals and man, as he put it in one of his last books, there's a lot of similarity between those. You know, the same smile, growl, snarl. You know, all these different facial expressions have a certain commonality. Presumably these emotional features are a consequence of quite straightforward neurotransmitter levels in our brains. And those are in common between us and lots of kinds of animals. So if it's like I'm happy today, I'm sad today. Those are things we're probably likely as owners can communicate with their pets perfectly well. And you can see from the tail wagging, you know, in terms of sensory data, I remember it's a company, I'm afraid I. I have a suspicion it didn't make it. But it was a, it was an interesting company to run into that was a tail wagging sensor company where they're putting accelerometers on the tails of dogs and cats, I think, maybe, but particularly dogs. And it was A question of, could you have communication with your dog by virtue of the details of the kind of the tail wagging of the dog. But in any case, one of the challenges here is, well, you can talk to your dog. What are you going to talk about? And what is the dog's sort of view of what happens in the world? A lot of our view of what happens in the world is deeply informed by this whole stack of culture and history and so on that we have and our certain biological features and those are different. But, well, for dogs, we don't believe there is, you know, for us humans there's, you know, a solid 10,000 years probably of accumulated, passed down cultural knowledge where we actually have ways in which there are, you know, from one generation to the next, knowledge is, is transmitted, you know, these days, perhaps with live streams and things, but in, you know, human language has been the, the principal mechanism for that kind of transmission. And that's something where we've accumulated this big stack of, of cultural knowledge and so on. And I, we don't think dogs at least have done that. Maybe whales have done that. We don't really know. And maybe the whales of 200 years ago knew less and had figured out, they hadn't figured out the great new whale philosophy and so on or the great new whale epic poetry that's now known. And I think, but this challenge of, okay, so you can solve the sort of structural problem of communicating with these organisms, what is the kind of, what are you going to talk about? Is an important question. And one might think that sort of we all live on the same world, but we don't really because there's quite different sensory experiences of organisms with different kind of sensor data coming to them. Let's see, so many questions here. From insert something. Do you think it would be possible to transition a real living person into VR or code, or will it just be a JPEG of the person? Well, it's an interesting question. You know, if we had, I think we've got pretty good evidence now, I think in the last six months there's sort of a dramatic change in the evidence we have about what, what it takes to make a brain or what it makes, to take a make, make a brain like thing. You know, in the past we could always have said, oh, you know, there might be some complicated quantum effect going on in the brain. And it's not just, you know, a few 100 billion neurons connected in certain ways. There's more to it, there's more molecular scale stuff going on, et cetera. Et cetera, et cetera. I think what we have learned through chat, GPT and so on, nobody expected it, but it's kind of. We really can see that a thing like that is capable of doing very human brain like things. So that really is kind of the essence of the brain is really being captured, not completely in every detail, but the essence. It's pretty clear you can nail it with kind of artificial neurons and so on. So, yes, it will be possible to make a thing that can do a live stream, just like me type thing and out of artificial neurons. Undoubtedly. Then the question is, okay, I've got that box, it's sitting on my desk, it's yakking along just like me. And then I say, well, you know, I'm going to die and that box is going to live on. Do I, what is the experience of that? In other words, do I say, well, gosh, you know, the box is, you know, my thread of experience is not the experience of the box. The box is another thread of experience. It might develop its own internal experience. You know, I think this is a thing that I've sort of understood in the last couple of years about consciousness and its relationship to physics is a very critical thing in our existence and our view of the world is this belief that we are persistent and we have a single thread of experience. There is a now, there's a future now there's this sort of thread of things happening and there's just one of it for each of us. We're not experiencing, you know, a hundred different threads of experience where we go off and this happens here and that happens there and then maybe they merge later and so on. We're not experiencing that. Our, our view of the world is that we have a single thread of experience. Now when it comes to what we've understood from our physics project, that's not actually what's going on underneath. Underneath there's this whole bundle of different threads of history of the universe which are branching and merging and so on. And that's kind of the story of quantum mechanics. But we are also part of that branching merging process. And we're sort of spanning a big bunch of branches. But we have the idea that we experience just a single thread of history and that idea that we experience, that we try and conflate everything, we try and say, oh, this thread over here, that really is different. No, it's actually actually the same. We're going to conflate those threads together. We're going to say, just like saying, you know, we're Sensing the, I don't know, the speed of some fluid or something like this, we could say, well, let's look at every individual molecule where it goes. But we're saying, no, no, no, we're going to conflate everything together and just say this is one uniform speed, even though individual molecules might be going at different speeds. But we're just going to say there's this one uniform speed. And that's all we talk about. This single kind of thread of what's going on, I think that's an important aspect of kind of our experience of the universe is this single thread of existence. I think without that, for example, when we talk about, I don't know, quantum computers or something like that, they have many threads of experience when they are trying to give an answer, you have to knit all those threads together to get into something which is consistent with our sort of single thread of experience view of the universe. So I think that our. It is a very deep aspect of our sort of existence that we have a single thread of experience. We don't get to say, oh, actually we're going to split our thread of experience and we're going to do this and we're going to do that. At least so far as I can tell, that that hasn't been kind of the, you know, it seems to be a core feature of kind of consciousness as we experience it, that we have a single thread of experience. And so when we have the copy of ourselves, it's not going to be that that copy could itself develop a thread of experience. And it could be sort of happily having an internal, internal subjective experience, but it is different from our subjective experience. And the question is, could we ever, as humans say, oh, you know, we're getting way too old, we just need to transition into digital form. Let's just make the jump. How's that going to work? Seeing the box, which is just like us, isn't going to do it. What I think has to happen is something gradual. In other words, if it is the case that you have your brain and it's got all these neurons and they're all firing and some of these neurons in, are connected to something which is gradually kind of offloading some of the things that would other bears be going to other neurons. I mean, if you took your brain and you replaced part of your brain with a purely digital element entity, but the connections on the outside are all connecting to your nerve cells just as they were connecting before, then that lump of digital stuff in your brain is. You wouldn't really Be able to distinguish that. Now, at what point, as, you know, let's say you're putting these neural implants in your brain and more and more of your brain is not original brain tissue, but is these kind of neural implants that are emulating what your brain tissue used to do. At what point do you suddenly realize, oh my gosh, I'm all digital, so to speak now? I don't know. I think that what's important in terms of that experience is probably a gradualness to it. That if it's like, you know, if one day you're all digital, you're just the box on the desk, it's not you anymore, and you won't perceive it as you. If, on the other hand, you're gradually getting sort of more and more of your brain is being sort of taken over by. By something digital, then I suppose that you could have the point of view that, yes, it's, it's kind of, you know, I don't think that there'll be a moment when you say, oh my gosh, I just lost it, you know, back in the time of Descartes and so on. You know, it was kind of the idea the seat of the soul might be in the pituitary gland in the center of the brain. But, you know, I don't think that it's going to be the case that there's this sort of one little spot, the critical collection of nerve cells that is the essence of the U.S. i don't think so. I think the whole brain as it is being conflated into the single thread of experience that's the experience of the single us. And that's something where one can imagine as a gradual matter replacing it. And, you know, that will be. Yes, that. That's presumably what will happen along those lines, that one will have something where it really could be the, the replaced human where, you know, you've got an artificial brain, basically, that was gradually taking over from your. Your actual brain. That would be my guess about how that will work. And I think it's certainly a. It's an interesting thing to imagine and to even see whether one can get sort of a, A, A sense of how that might work. I mean, I suppose there's a, a question. I don't know. People who have prosthetic limbs, for example, may have some experience of what, what it's sort of like as you, as you take over, you know, you make a technological thing that takes over from some biological. Some previous biological version. And you know, what that's going to be like for the brain is an interesting one. The. Let's see, I need to go in a second here, but. See if there's anything I can think I can attack quickly. Okay. Wampus commenting. Isn't it too early to assume we can replace all parts of the brain with digital technology? Maybe. I think it, I think that is, you know, the idea. There's something amazing that is beyond what we seem to be able to capture. I think that's a dwindling story. Maybe I'm wrong, but I think if you'd asked me would one be able to make a kind of a neural net that writes reasonable human like text and so on, I would not have known when that would happen. I wouldn't have even been sure that would happen. But I think now we know the answer to that. And I think one can ask the question if you say, oh, what about the creativity of the brain? Well, you know, it's actually pretty easy to ramp up the creativity of a large language model. You just increase its temperature parameter, you make it more likely to pick words that aren't as probable in its estimation of the probability of the next word and it'll suddenly start doing all these crazy creative things. That's just kind of interesting because as you increase the temperature, as it's called, to make it more likely to use less words that it estimated as less probable, it at some critical temperature seems to lose it. I don't know if it's a specific temperature, but it seems like it is. There's a specific point at which the thing just starts yakking and going crazy. It's just bonkers. If you ramp the temperature up really high and it's starting to sample all these low probability words and it then has to follow on from those low probability words, the thing sounds just totally nuts. And it's sort of interesting that you are seeing all these effects that are kind of brain like effects that you know, if you've got, it's kind of like creativity is good but too much of it and it's nuts. Which is kind of a thing one's used to imagining for brains. And you know, so you can see it, you move the slider and for a while it's like writing more creative essays. If you move the slider down, it's writing really flat, boring essays. Move the slider up, it's very lively, it's writing all kinds of stuff and then you move it a bit too far and it's kind of bonkers. And there are all these effects. It's really a weird thing. I mean, there are all these things. Like one I was just thinking about today actually was when you do kind of fine tuning training, when you do the sort of after training, after the main training, training of a neural net, one of the things that's difficult is if you feed it, if you say, now I'm going to tell you about this, this very specific thing, just all the training data is about this, this, this, this, this. The neural net will have a habit of forgetting what it already knew. So you have to kind of intersperse old training stuff with new training stuff. That's the best way people have found to do this. And that way it kind of is like it's still remembering that it's got that old stuff in there and yet you're feeding in a certain amount of new stuff. And I sort of remembered back in the days of the Cold War and so on, this kind of whole concept of brainwashing humans and so on, of, you know, you feed humans just one thing very intensely in some way that it's possible to then get the humans to sort of forget past stuff. And I kind of realized that what one does with LLMs is what one can do sort of by mistake with an LLM is essentially brainwash the LLM with the more recent information and have it forget the previous stuff. I mean, I think that there are just so many bizarre analogies between what one sees in LLMs and what one sees in human experience. I mean, I do think there's sort of this whole idea of LLM psychology and the whole idea of sort of mapping the different kinds of experiences that humans have to the kinds of things that exist in LLMs. And LLMs are certainly not the end of the story in terms of what one can do with neural nets and so on. But I think there's sort of interesting things to learn both for LLMs and for human psychology and so on and neuroscience from looking at that analogy. All right, I think we should wrap up here. Well, thanks very much for joining me for this first episode of Future of Science and Technology. I suppose we can imagine a future where the hopefully preserved versions of these live streams, people will be able to look at them and say, oh my gosh, what a goofball. He didn't understand that this is the obvious thing that's going to happen in 50 years. And I could talk perhaps sometime about that whole question of how to not be fooled by what, you know now into what kind of is what you can't imagine some amount of time in the future. And I suppose for myself, I've had about 50 years when I've been thinking about these kinds of things, where I have some idea of sort of what happened back in the past, what happens now, what did I kind of foresee, what did I not foresee. It's kind of an interesting collection of what is foreseeable, what is not foreseeable, how you miss things which, in retrospect, seem like they're obvious, et cetera, et cetera, et cetera. There are many things with LLMs, actually, which I think, in retrospect, are now obvious, but nobody saw them coming. All right, we should wrap up here. Thanks for joining me and see you another time. Bye for now.