Future of Science and Technology Q&A (September 13, 2024)

Hello, welcome to another episode of Q and A about the future of science and technology. So I see a whole bunch of questions here. One from D0 what research is essential for putting people on Mars? I don't know that in tremendous detail. You know, I used to be an enthusiast of things space when I was a probably preteen, a long time ago now. It's sort of interesting. I was thinking about that recently. It was kind of in the 1960s, late 1960s and so on. Space was kind of an icon of the future, so to speak. And it was kind of an aspirational thing. And I think it was kind of cool that there was such an aspirational thing that folks like me could get interested in. I mean, it segued for me quite quickly from thinking about how one would design spacecraft to learning about physics and getting interested in the foundations of that, et cetera, et cetera, et cetera, et cetera, et cetera. But the fact that there was this kind of area that was a kind of a global aspirational area was pretty cool. And I guess it's sort of coming back and the let's go start a colony on Mars is kind of a modern version, you know, whatever it is, 50 something years later it's a modern version of the same thing. And that's kind of a. Hey, that's a reason to do it. I mean, I suppose it's like, you know, you have a mission, build a cathedral, a pyramid, whatever else. It's kind of cool to have a societal mission. I think it probably helps people, countries, et cetera, when that happens. And so I guess the let's go to Mars is such a mission. Does it really make sense from a sort of totally pragmatic point of view? Probably not particularly. But is it a worthwhile aspirational thing for society? I think clearly, yes. What's involved in doing it? I think it's. I mean it's a. The difficulty is not getting stuff to Mars, although that's hard. But that's been done since the, since the one was the first. I think Mariner 4, if I remember right, was the first spacecraft that got to Mars, took pictures there. That must have been 1967ish type timeframe maybe. Yeah, around that time. So, you know, getting stuff to Mars is sort of a solved problem. Getting. And even now, you know, this cool thing of flying a helicopter on Mars with, you know, it has a thinner atmosphere, but you can still fly a helicopter there and so on, you know, so getting stuff there sort of Solved. Getting humans there is more difficult because humans are you know, quite fragile and were evolved to live on the surface of this Earth where we do okay. But you know, send one into space with all sorts of extra cosmic ray radiation and all kinds of things like that, it's, it's a little less clear. I mean my impression is that there are all kinds of issues with, with sort of how do you shield from radiation and things, how do you keep enough stuff on the spacecraft to, to keep everybody, you know, happy for the period of time that one's on the way to Mars. I mean, I think, I don't know how often one needs to resupply the ISS to have people sort of prosper on it. But the ISS in low Earth orbit and for example, it certainly doesn't have the same radiation issues that it has that one has going to Mars one, one gets to Mars one, it's kind of like how does one make stuff work there? I don't really know the details of this. I mean I know there are, there are questions about, you know, how do you use the materials that are there to build stuff. You know, 3D printing. Your, your house on Mars seems like a, a somewhat realistic thing getting, you know, concentrating the oxygen there so you can breathe it. I, I think that's, I think that's a somewhat solved problem. I'm not sure. You know, I think there's just a, a lot of engineering detail and a lot of things to do with sort of the biomedicine of, of, of what happens, getting there, hanging out there and so on. I think, you know, there's also the question of what are people like when, you know, when the hanging out on Mars with knowing perfectly well that it's not like it's just oh, we can send a spacecraft up to fetch you back to Earth type thing that it'd say, you know, you're stuck there, you're far away for a long time. And you know, I think we're, we're, that's a question of sort of the psychology of things. I know there were these various efforts. There was a, certainly a web visible effort of people who wanted to were sort of signing up to take risk one way trips to Mars. And that's, you know, there's a certain psychology associated with that. I mean I think it's quite a psychological type issue. I suppose the most extreme version of this is kind of the send an interstellar spacecraft out there and have it, have you know, a few thousand people on it. Let's say for I Don't know how many generations. Let's say five generations, 10 generations. That seems to me to be a very peculiar proposition because I think it's both ethically peculiar, as in you are then dooming future generations to live on this starship. Regardless of one likes to feel that on Earth, although there are certainly many limitations, there is some degree of kind of one's free to go hang out and do different things. I mean, there are all sorts of limitations imposed by all sorts of things, but there's some impression of that. Whereas you're stuck on a starship your whole life. I mean, yes, we're stuck on Earth our whole lives, but it seems like a more extreme limitation. And my own guess would be that whatever you thought the point was in Generation One, by the time you're at Generation N that arrives at the other end, people will have lost the thread of what the point was after multiple generations have been gone through. And that's not the situation with Mars. It's not as far away as that. But I do wonder quite a bit about the psychology of the whole thing. And I think it's back in a few hundred years ago when there were still parts of the Earth being settled, particularly the New World, so to speak. I think the. My. I mean, it's. I suppose it's an interesting psychology to say we're going to go, you know, we're going from the old world to the new world. We're never going to go back, although you knew you could in principle go back and so on and sort of how one thinks about that. And somehow, you know, and I think there's also a certain tendency to conflate let's go to Mars with let's make a new kind of form of society here on Earth, some kind of perfect version of society. I think that's a completely separate problem from whether you choose to do it on another planet or not. And I think those things sort of shouldn't be conflated, but that's a different issue. Let's see. Majid asks comments on the future of arts and literature in the face of AI related challenges. Will individual creative impulses forever be subjugated to AI? I don't think so. I think AI is but another tool. I mean, AI, it's very easy to make a computer or an AI be original. You just put a bunch of random pixels on a screen. Top point is most arrays of random pixels on the screen no human cares about. There's no kind of human alignment, human resonance with such a thing. And it's what one is doing with AI is the AI is going off and sort of exploring different parts of, let's say, the space of possible images, but to what? And it's doing that in a way that's sort of aligned with images that's seen on the web and so on. In a sense, the story of what the significance of that is largely in the eye of the beholder. That is a bit like taking a photograph. You could say there are all these trees here. Which tree am I going to select to take the photograph of? To be my sort of cool iconic art photograph type thing? That's something I'm doing. Mostly the image is made by the tree, but I am still making the choice to have it be this tree and not that tree. And it's, I think very much the same thing with AI. And I view kind of AI in sort of arts and other places as a tool that lets one extend the things that we humans are able to do. I mean, like, for example, one thing that was occurring to me recently is that we get to say what we think think potentially by just talking. But if it's like, well, let me make you a picture of what I'm thinking, that's not something that's easy for us to do. You know, if you're artistically competent, then maybe you can sketch something. It takes you, you know, minute, few minutes or something to sketch it. But most of us are not that artistically good and, and we won't be able to do that sketch very well. A generative AI can absolutely sort of show what it's thinking by generating an image that's kind of a human aligned image of something. So it kind of makes one realize that yes, one could imagine some future world in which sort of a form of art is you take sort of the things from someone's brain and they are directly driving some generative AI system. And it's making this sort of mental. It is externalizing the mental image of that person. It's kind of like instead of the, you know, I am having some particular feeling internally and that's causing me to sort of be an actor who acts in a certain way or be a musician who plays in a certain way. Here it would be externalizing not through one's fingers or one's voice or whatever, but through something where one is sort of externalizing the mental image that one's forming and doing that with the medium of AI, so to speak, as I don't know what one calls that, mind painting or something. And I think Kind of the one can imagine, you know, mind painting as a new art form that is at some level, a new expression of a new form of human expression, so to speak. Not. And we say, well, what's the role of the AI there? Well, the AI is the one picking all the pixels, just like the tree is the one deciding, you know, where the bark goes in the photograph, so to speak. But it's still a matter for sort of the human to be the one who's giving significance to and choosing what the actual sort of art form that's presented is going to be. Now, it's probably the case that in kind of what AI can readily do is to kind of learn things that are basically pleasing to humans. You know, we did this experiment back in 2007, this thing called Wolfram Tones. You can find it on the Web. It's a algorithmic music generation system that at the time was done because I wanted to have custom cell phone ringtones, because that was a thing at that time, didn't stay a thing very long. But we also did it because it was an interesting application of a bunch of science that I've done, where one is essentially plucking from the computational universe these kind of arbitrarily chosen programs that have a certain inner logic, but the program can be deployed as a musical composition, but that musical composition has an inner logic, yet it also has a certain complexity, because that's just the way a lot of things in the computational universe are. And so it makes little tunes, basically, that at least for some number of seconds, can be quite pleasing. The. It's the thing that. So, I mean, that's an example of sort of being able to make something where by choosing it in certain ways, one can make it be generally kind of pleasing to humans. It's the same thing with architecture. For example, you could say, here's a plain room. It's not so pleasing to humans. You know, you like little moldings and you like things that are kind of some mixture of what we're used to from the architecture we've seen and what we're used to from, for example, the natural world. I mean, it's, you know, if you build a building that has some strange cantilevered thing that is sort of looks like. It looks like it should fall over, but it doesn't, people find that a curious thing to look at. Some people think it's good to have things like that that make people sort of a little bit uncomfortable and think and so on. Other people say it's better to have Something where it's a comfortable building that kind of makes sense to everybody, where it has things that we feel intuitively. Yes, this building kind of makes sense. It has columns that kind of remind us of things in human form or things in nature and so on. But it's certainly perfectly possible to say, we'll take this bare thing and have AI kind of dress it up in ways that are sort of going to be kind of comfortable to humans. Comfortable to humans, both because of what we're used to in our kind of artistic past, and I think comfortable for humans because it's what's used to based on what we see in the natural world. And so that's definitely a role of. Again, it's the story of define what you want, and then AI fills in the details. And I think those are really useful directions there. There's also, in the case of something like Wolfram tones, there's like, just pick a tune at random. And that may turn out to be something where the human says, oh, that's interesting. I like that. I'm going to use that. Just like you could, like, pick a tree at random, and maybe one of them will make a really interesting artistic composition, a photograph and so on. So I'm. I'm not now, you know, when it comes to art, that is sort of just dressing things up. Yes, there's a bunch of that that will be AI doable. And, you know, that's gradually happened in a bunch of different areas. I would say that one of the areas where that happened in the 19. Well, by 1980s was things like desktop publishing and graphic design, where, you know, back in the day when. When I started out, if you were doing a piece of graphic design, it was drawing things with. What are they called in America? My gosh. So long since I used the word. They're called mapping pens in British English, but they are these pens that are used for. For drafting things and so on. Boy, that's a shocking thing. That means I haven't used that word in 46, seven years. But anyway, the type of pen that is commonly used for kind of making very even lines and so on on paper, and that was how one did kind of design and so on. But nothing came for free, so to speak. Yeah, you would have stencils and you'd have to trace them and so on. And then as soon as desktop publishing came in, lots of things like, oh, you want to put a border here, where you can just sort of design. You know, you design a little piece of the border and you just Continue it that stuff, it was all like you had to do it with your own own hands, so to speak, before. So it was a. Was another one of these kind of getting more leverage. But it was something that happened, you know, in, in the case of, of graphic design, it's something that was, was possible there. One thing that was notable about the, the arrival of computers and graphic design was the fact that it became possible to just try a lot more things out. Before that it was like you're really making a big commitment to drawing the thing this particular way. Whereas with computers it's like, well let's try this and this and that and it's easy to make a variation. And I think that's the same is probably true in other kind of artistic endeavors that the fact that the mechanics of making these things has been sort of computerized and AI ified means that one gets to try out more possibilities, the better to select the one one really wants. Let's see. Corey asks, how often do you find yourself thinking about future of science and technology? Does it affect how you prioritize projects, say wait five years before because the tech will be better to handle it? Yes, well, I absolutely try and think about things on those kinds of timescales and so on. It's cool that I get to do this live stream every couple of weeks because this is an opportunity for me to make sure I think about future of science and technology. But, but I think about it a fair bit. It's always I. I don't write about it so much because I feel like kind of writing sort of oh, this might happen type thing is less somehow I, I feel. Well, actually the real reason perhaps is because I, I like to write things that I feel will last. And if I say this might happen, that might happen, it either going to happen or isn't going to happen. And either way it's kind of what I spent all that effort writing is sort of dead meat. And it's not the fact that, oh yeah, I predicted it and I wrote it down and you can see how clever that prediction was. Isn't particularly a thing as far as I'm concerned. But thinking about kind of the direction things are going. Yes, absolutely. I find that useful. I find that useful in a way that is kind of in terms of planning many kinds of things. I mean, for example, in technology, in building technology, it's like when can one expect to take things for granted? Will it ever be possible to take those things for granted? There was certainly a time a while ago when you couldn't take for granted that people will be connected to the Internet in designing software. Now you can kind of take that for granted these days. The question is, can you take it for granted that for example, people will be able to run an LLM on their local machine? I think one should in the future take that for granted. And I'm planning for a world in which one can take that for granted. I'm planning for, for a world in which sort of an LLM is kind of a thing like an operating system, like networking, things like that that one could sort of take for granted. And knowing that plan affects quite a lot of things today about what one does. I mean, another thing that is sort of interesting is obviously AI is at an all time high these days of everybody's like everything's got to be AI and so on. And you know, I've seen this cycle multiple times in the past. And yeah, there are lots of interesting things. But the oh, an AI will just do everything story is not a real story. That's not how it's going to play out. And knowing that, there's sort of the question of, particularly in terms of technology that I built over a long period of time, the kind of frothy belief about oh yeah, we'll make a neural net do absolutely everything, which I don't think is going to work. In fact, I have pretty good scientific reasons to know it's not going to work. Neural nets do lots of interesting things and the trick is to find the places where they will be really useful kind of places where you can make linguistic user interfaces, places where you can have sort of the image as output type mechanism, those kinds of things. There are lots of things, maybe many things that we've not even imagined doing which are now enabled by those kinds of technologies. But it's still the case that there's a lot that we have been able to do by actually doing computation and actually having sort of curated knowledge about things. That's things I've worked on for the last 40 years or so. That really what it is is there's a sort of what we humans can do off the top of our heads or in our brains. And that's what neural nets seem to be pretty good at doing. Possibly because they're architected in somewhat the same way that our brains are architected. But in human civilization another big thing has happened which is the development of formalization and systematic knowledge, which kind of has a long history, but really started taking off in the 1600s and built us up mathematics. Now computation Sort of computational X, mathematical X, all these kinds of things. And those things are things which, yes, we can do a little bit of that in our brains, but. But mostly we built tools nowadays, computers in the past, sort of mathematical formalism that allows us to really zoom forward in those areas in ways that the unaided human cannot. And it's sort of the same story with neural nets that there's this kind of big tower of capability that the abstract human, before that was all invented by our civilization, wouldn't have known was possible, but now we do know it's possible. And that's the thing that I've spent a lot of time sort of making work well and kind of aligning with what we humans want to do, and that it fits very well with the whole story of LLMs and neural networks and sort of the linguistic interface to that and then the formal structure that you can kind of count on the linguistic interface. You talk to it like you talk to a person. It's like, do you get what I'm saying? Do you understand what I'm saying? The person says, oh, yeah, I understand what you're saying. Well, can you tell? Maybe if they do the right thing, then maybe you can tell, but you don't really have a good way to know that they knew what you were talking about. The great thing in sort of computational language that I've spent the last many years building is that you have a piece of orphan language code. It's intended to be readable by humans as well as writable by humans and readable by computers. But it's something where you look at it and you look at it and you say, yeah, that's what I meant, or, no, that's not what I meant. And then when you run it, it can build this big tower of things that it does. And that's just rather different from what you get from the sort of the AI where you're building a piece and then it's doing what it does. It's super useful to have that interface, but it's not kind of building the big tower. You know, it's funny because I was studying recently kind of the foundations of machine learning and why machine learning works at all, why it's possible to have neural nets just learn all these different kinds of things worth realizing that they're mostly learning things that are the kinds of things that we humans tend to learn. It's not clear whether neural nets can learn arbitrary things. There's certainly some kinds of things that we know they can't learn in a Practical way, but sort of there's an analogy. Well, what's happening, I think inside neural nets is that there are these lumps of kind of irreducible computation, something I've studied a lot in the way that simple programs work. You can have a program, you run the program, you get this very complicated behavior where you kind of have to follow each step in running the program to know how the thing will behave. You can't just sort of jump ahead and say, I know what the answer is going to be. So there are all these lumps of irreducible computation that you can find in the world of neural nets. Or particular thing I was studying was kind of discretized, spatially localized analogues of neural nets that are quite interesting from a technology point of view. Possibly very interesting from a technology point of view, but that's a separate story. But then there's the question of how you get sort of how you take these lumps of irreducible computation and fit them together to do something that's useful. It's different from saying, let me take these well crafted instructions like in a computational language where they would be carefully built so that they fit perfectly together and you can build this tall tower and have it all sort of stand up and work. Instead, what you're doing in machine learning is it's like you're building a wall and it's a stone wall. And you're building it by picking up rocks that were on the ground and figuring out how to fit them into your stone wall to build it up. And in the end, you can build a stone wall out of random rocks that are different shapes. You can also build a brick wall out of bricks that are all the same shape and sort of carefully engineered. One feature of the stone wall that's built from random rocks that were just lying around is probably when you build it up really high, it's not going to, it's going to fall over. In other words, if it's not been sort of crafted, it's something where, yes, you can make it do a lot of things, but only up to some point. And I think that's something that one sees in kind of the way that neural nets and linguistic interfaces based on neural nets work. They're not doing these kind of big towers of computation. They're doing the kinds of things that are in a sense broad but shallow computations of the kind that we do. I think. Not quite sure why I got off onto that subject there, but the question was about the thinking about future of science and technology. And when to do different kinds of things. And I think the, Yes, I absolutely think about that. I've been interested for a long time in nanotechnology, molecular computation, things like this. I've thought about that since the 80s and sort of, I, I'm almost right now at the point where I think that the sort of laboratory practice and automation of labs and so on is to the point where it makes sense to really try and do something in that area. But in the past it had been, well, if I was really going to do something, I was going to have to build up a whole decade long, you know, tower of technology to be able to get anywhere with that. I would say that there are other, there are quite a lot of things that I will sort of keep around as projects for a long time, steadily accumulating knowledge about them full well, knowing that things have to develop in a certain way in the world for it to be possible to actually do those projects. And sometimes it's things developing in worlds that I'm trying to build out, like tooling, computational language, kinds of tools, sometimes even just resources in terms of do I have a big enough organization to be able to tackle such a project and so on. But yeah, I think it's really, I tend to think that you spend most of your time, the vast majority of your time doing things, but at least a few percent of your time kind of thinking about what is possible and kind of planning what one could imagine doing, and thinking also about kind of how what one's doing fits into the future. I mean, it's like the, the, the thing is, the thing I found, I've been very happy to find is a lot of things I've done which I did think about, you know, why am I doing this? And I kind of have some vague idea of how the future is going to play out. And then decades later, by golly, it's actually right place at right time. Sometimes I, I, I underestimate how long it's going to take before the world gets to the point where that particular kind of thing is going to be useful. But it's been very, very nice to see a lot of things that, particularly on the technology side and also to some extent on the science side, have kind of landed. You know, they're definitely going in the right direction. You don't want to be the kind of the thing that's kind of the last gasp of some particular paradigm, you know, just as it's being submerged by something different. And I think that's, it's important to Kind of be able to see a bit of the future to know where you're going to be with respect to that. I think that's, that's even true with things like business organization. And you know, do you, do you do the trendy thing and take your, you know, set your company up to be organized this way, that way, or have this or that kind of financial structure or whatever else or not? And you know, do you, how can you predict enough of the future to see where things should go? And anyway, so I find that it pretty useful to be able to do those things. I think I was talking about AI and kind of the trajectory of AI and the fact that AI is a big peak right now, but it's not going to keep going up forever. And we're in this interesting situation where we have technology that is super important and useful for doing things with computers and being a back end to things that are like the linguistic interface layer of AI. And it's sort of an interesting strategy question that, you know, AI is in this peak state right now. Eventually it will go down. We have technology that is very real and very useful and very kind of connected to AI, but not dependent on AI, but AI makes it more, makes it have an easier interface and so on. And how does one manage kind of the, you know, dealing with technology that sort of could be themed as AI, but it's not AI in the neural net sense, at least most of it isn't. And you know, when everything is neural nets, it's kind of like hard to get the word out about anything else. But then at some point when a technology starts going down and everybody says, oh, no, no, you know, as they have done in the past, okay, as they have done in the 1980s, people were talking about AI is going to be everything. It was expert systems back then. And after a while it's like, I don't think this is working. And then if you said, well, we're doing AI, people would kind of look embarrassed and kind of ignore it. And so there's a point at which things sort of go down fast enough. I mean, it sort of happened with the various kinds of ups and downs of blockchain. That's happened a bit too. And it's sort of a question of when, you know, how does one deal with the fact that sort of it. There's a lot of noise when it's up, there's a lot of kind of, oh, no, not that. When it's down. And when one has something which is really the sort of a Core thing that is kind of a unique capability that is very relevant and sort of plugs into that world. How do you best deal with that? That's the problem I'm thinking about right now and probably the practical outcome of that one. We'll see in due course. Lousy asks, is there a chance we will ever have giant insects or animals akin to those that live during the age of dinosaurs reappear? Oh, I'm sure there'll be a zoo with dinosaurs in it one day. I mean, I'm sure that we can to some extent extrapolate backwards the genetics of dinosaurs. Not perfectly, but one of the things that I'm sure there will be more and more found paleontologically and more and more kind of molecular paleontology and probably more and more more ancient DNA that gets found and so on and gets identified. I mean, how much we can deduce about what the dinosaurs were like from kind of what birds are like today. And extrapolating back from reptiles and so on, we can deduce a certain amount. It may very well be that when we've sequenced every one of the, whatever, 10 billion species around, 10 million species around on the Earth at the present time, that we'll be able to, in some clever way do even better at extrapolating backwards what earlier species must have been like, and then we have to actually make them. And it's also becoming increasingly clear that the story of making an organism isn't just DNA. There's sort of all sorts of epigenetic stuff that comes along with the, you know, egg cell or whatever it is that came from the, the previous generation of organism. So how that works is not clear. But, you know, I do tend to think that there will come a time when one will be able to sort of, you know, recreate the critters, just like in the movies, type thing. I think, you know, I think, oh, gosh, oxygen levels were different in those days on the Earth than they are today, things like this. So, you know, maybe, maybe the. I don't know how different they were and whether this is right, but, you know, I can imagine the stegosaurus is wandering around, but it has this complicated breathing apparatus on because it's kind of like got a bit of a spacesuit to deal with the fact that it's, you know, 100 million years later type thing and the Earth has changed. But, you know, perhaps after a few generations of stegosauruses, they would adapt to current conditions and maybe change a bit. I don't know, they might Change for the smaller. I'm not sure, but I think that's a minor effect. And certainly one can imagine, you know, the zoo which has the, if it needs it, you know, the high oxygen content enclosure or whatever else, so one can visit the frolicking dinosaurs or whatever. But I think, I mean, I tend to think that we still have, you know, we still have the computer, we just need the program to, you know, we still have the thread of life and we probably have what we need to reconstruct what life was like back in the day and, and to be able to recreate those organisms. And I think it becomes an interesting question in terms of the preservation of organisms in today's world, to what extent is it like we're preserving the ecosystem of those organisms and the whole set of interactions? To what extent are we just raw. Let's archive the organism and if we want to, we can always sort of rehydrate it, so to speak, in the future. But I think somebody was asking about Mars earlier. Obviously the Martian gravity is different from Earth gravity and the Martian atmosphere is very different. And one could certainly imagine a time when one could. I don't know whether life on Earth could quite. I don't know whether it can make the jump to evolve, to sort of live naturally on Mars. Perhaps over the course of sufficient time. I don't know what that time is. It'd be interesting to estimate one could progressively evolve organisms that would successfully exist in the Martian sort of environment. I think that the chemistry, the biochemistry of organisms on Earth should be capable of making that jump. I think, although it might be a bit of a stretch, of course, if you say, well, let's evolve the humans as well. While evolving humans is a very fraught business. And in natural selection, some humans have to die to produce the struggle for life that drives natural selection. In artificial selection, it's sort of, perhaps even worse. Some humans have to be chosen to live versus die, and that's even more fraught. But the thing that it is an interesting question whether over the course of time and how long it would take to have kind of the Martian dinosaurs or not. Dinosaurs? Martian, Martian. What would they call Arasaurs or something for arrays as Greek for Mars. Right. And saw would be Greek for lizard. And dino would be like a terrible lizard in Greek. So I guess Arasaur would be the Martian equivalent of that. And I don't know how long it would take to evolve those things and to have them be, be hanging out on Mars and maybe they then become the farm animals. Of Mars. Although, well, there's just so many ethical cans of worms, which really, you know, ethics is really a story of what us humans want to have happen and sort of somewhat collectively decide should happen. And that's something that no doubt evolves as different capabilities become possible and so on. Let's see. Bob asks, how can we combine LLMs with first generation AI algorithms like Minimax and Tree Search? At the moment, LLMs can't even play tic tac toe. Yeah, I mean, the problem of interspersing kind of computation with kind of the trainable neural net is an interesting problem. I mean, the work I've been doing recently on sort of simplifying neural nets gives one a better idea of how that might work. But I think the analogy of building the stone wall with lumps of computational irreducibility kind of shows you a little bit the difficulty. The difficulty is, yes, you can make those lumps, enough of those lumps sort of fit together. You have a wall that you can stand on or whatever else at the top, but inside the wall, it's pretty complicated. And if you want to sort of intersperse weird shaped stone with sort of very nicely engineered brick, that's a challenging thing to do. And I don't think we know how to do that yet. And I'm not sure how possible that is. Traditionally, the problem has been you get a choice. Either the thing is trainable or it does serious computation. You don't get both. You can't kind of train through the computation because what you're typically doing is you're saying, okay, neural net, run as you are right now, See what result you get. Okay, you got the wrong result. Let's see how we would kind of back propagate from the change of result to see how we would have to change the innards of the neural net to get the result better, closer to the result we want. Now, it turns out that with these kind of simplified kind of machine learning systems that I've been looking at that involve discrete cells and so on, that it turns out that the idea of back propagation from errors from the output is much less important and there's sort of forward mutation that you can do. So I think it might become more possible to use that to do this sort of fine scale integration of computation together with kind of the traditional neural net approach. Now, having said that, when you look at an LLM, one feature of an LLM is it's kind of going one token at a time saying, okay, based on the context window that I've seen so far. Go through the neural net, predict what the next token should be, decide which of the maybe 50,000 different possible tokens has the highest probability, decide what temperature setting I want to use. So whether I want to pick zero temperature, the most probable token at every moment, or at temperature 1, the token randomly picked at the probability that the neural net says it should be picked with, so to speak, in different temperatures, different things. But one's always kind of doing one token at a time. Now, something I thought about a while ago is what one can call kind of quantum LLMs, where instead of following a single path, you're actually allowing the system to follow many paths. And then you're potentially seeing for those many paths that you follow. Those are. It's sort of having many. It's writing many essays at the same time. And those essays will branch and say different words here and merge because they actually ended up saying the same word three words later and so on. And so there's this sort of picture of the kind of branching merging essay that's getting written, which is kind of a quantum LLM. And one can imagine I looked at it a bit sort of pathfinding on that type of thing saying, okay, LLM, you need to reach this word and now kind of tree backwards. This set of things where normally you're just saying, use the temperature parameter, pick at random what you want. But here one's sort of saying, don't pick it at random. Be informed by things that come from the future somehow by being informed by this sort of graph that's generated from the future. You know, I should look at this a bit more. As I'm describing it to you all, it actually seems a bit more near term than I had previously thought. Let's see. Marcos asks, have I heard about AI reading minds through brainwaves and fmri? Yes, I've seen several such projects talk to some people who are doing those kinds of things. I color me a bit skeptical. I mean, FMRI and the kind of mind reading by FMRI is a challenging business. You know, brains are more unique than fingerprints. All of our brains are set up in different ways. It is extremely unlikely to be the case that the way we encode knowledge in our brains, in our neural nets has any correspondence with the way anybody else encodes those same things. It's a loss of chance. And we see that in neural nets as well. I mean, there certainly are claims that there's a certain universality to the representation of knowledge in neural nets. It's not clear that that's correct at all. I mean, just like with brains, there is some notion that we tend to have localized language centers in our brains, although where the language center is, is very specific to each individual person. And that's, you know, that's an issue for neurosurgeons, for example, to figure out for a particular patient, where is the language center, which you have to be very careful about. And, you know, where is that just brain tissue, where we don't really know what it does. So I'm a bit skeptical. And I think the, you know, really what that goes towards is the ultimate kind of story of, you know, to what extent can we read out directly from the brain, as opposed to going through the kind of transduction of speaking things or whatever else, to what extent do we go direct? It's like, you know, you go to your computer and you start poking at a piece of memory in your computer, and you watch the bits go by, and it's pretty hard to tell what the computer is doing by looking at the bits go by. And so it's a question of to what extent one will be more successful with brains. It will certainly help when we can read out a lot more neural activity than we can right now. I mean, one's starting to be able to have these arrays of microelectrodes that you can implant in the brain and even be able to sort of seal the skull up and have the whole thing be internal and so on. But those are still. They're not huge arrays, and they're not. It's not like one's reading out a large part of the brain. One has to be quite invasive and put things really close to particular neurons on the surface of the brain and so on. Is there going to be a possibility of reading out much more than that? People talk about optogenetics and being able to have genes that cause neurons to light up. And then turns out the brain and the skull is more transparent to certain kinds of infrared light than one might have thought. And so it is conceivable that you can kind of see your. It's kind of funny. You know, one talks about flashes of inspiration, but maybe in a time of optogenetics that will be quite literal, that one will be able to see. Oh, you had an idea. Look, we can tell from this infrared sensing technology that, yes, the neurons in the such and such nucleus lit up at this moment. And that was the flash of inspiration, so to speak, quite literally a flash. There are other technologies. One can imagine another one that's been long Attempted is meg, the magnetic equivalence of eeg, electrical recording from the brain. The difficulty of that has traditionally been you just have to have this very electromagnetically quiet environment, and that's hard and a nuisance to achieve. I think there's some hope that kind of various kinds of quantum sensors where there's kind of single atoms being used to very sensitively sense magnetic fields, that. That may be something which will allow one to just have vastly more sensitive sensing of magnetic fields from the brain. And it's conceivable that with that, one will be able to sort of home in on. Yeah, you know, there's 100 billion neurons, and we can somehow disentangle all those neurons and we can know that neuron number, you know, 370 million, blah, blah, blah, blah, blah, just fired and then another neuron just fired and so on, and one will be able to have some external map. My feeling is, if there's one feature of sort of technology development that tends to be a constant, it's that sort of more information will be extractable from systems than we thought. So what was just, oh, there's a single dial that measures one parameter becomes, oh, there's this image we can take with a 10 million pixels. Or oh, there's this thing with hyperspectral imaging where we can sort of look at all these different frequency bands and all these things and so on, and we can look at these molecules and we can sequence the molecule and we can get a lot more detail. I think that's one of the features that extracting more information from the physical world is something that progressively gets to be done. And I think that's. So I would suspect that reading out more stuff from the brain is something that will be readily, you know, that will continue to progress. I mean, there are other efforts that involve sort of machine learning for neural impulses. Some of them things like spinal cord injuries, where one's trying to get sort of relearn what those actual nerve fibers represent. You know, does that mean move your foot in this way? And so on. And my impression from some demos I've seen, that's really getting a good distance, being able to kind of machine learn what did the firing of that particular nerve fiber mean? And sort of being able to map that so that whatever the brain was thinking, when it's like, move my foot this way, even if there's no foot to neuromuscularly move that way, that you can pick up the. Or if that connection is broken, that one can pick up the nerve signals coming from the brain and decode with machine learning what those nerve signals mean. I mean, it's kind of like, you know, it's a much simpler case than the. Can we understand what the dolphin was saying? Where we can. The dolphin says something, dolphin flips its flipper and so on. But we can. In this case, we're trying to learn what was the brain of the human trying to communicate down the human spinal cord. And can we decode that language, so to speak, from the spinal cord. It really helps that at one time, typically if this was an injury or something, at one time that signal meant something where the person sending that signal knew what it meant, if not consciously, at least their brain was generating that signal and they moved their foot. And so that means there's a. There's a sort of known mapping, a knowing target. But it's kind of like learning an alien language. Again, let's see. A lot of questions about AI here. Let me see, There's a question here about. Oh boy, so many questions. You see, I'm not even. This is terrible because there are questions that come in before these live streams and then the ones that are relevant to the live stream that coming during the live stream. And it's always hard to decide whether to go in first come, first served order or something different. I'm going to jump around a bit here. Lola asks, have your thoughts on the future of education changed at all recently? I think I talked about this maybe last year or so. One thing that's changed is we've been building kind of a autonomous education AI sort of tutoring type system. And it's been hard, but it's going well. And I won't really know how it's going until I've seen actual kids who are actually trying to learn Algebra 1, which is our first target area, use it. Because I'm not really in the target market for Algebra one, I think. But you know, as I use it, I'm impressed. I mean, if I kind of try and channel certain kinds of confusions, I might have. It does well at sort of being able to explain those things in a way that's relevant, given what I've said and so on, and being able to kind of knit together things which are part of sort of the math that we know is correct and that we can kind of compute through the curriculum that we know is oriented in this direction and so on. I mean, it's a big pile of software engineering and it's a typical example, I think, of something where we're really making use of the linguistic interface that neural nets provide. But we're not like saying it's neural nets all the way down. That's all you need to use because it's not. I mean we can plainly see that to make a successful kind of education system with AI, it's got to have more than just neural net AI to make that work. But I would say that my optimism level that we will have good personalized sort of quotes, AI based, doesn't really matter what they're based on. But automated kind of tutoring systems, my belief that that's going to be possible has increased substantially actually in recent times. I will also say it's not an easy road. The five minute demo with the LLM doesn't turn into the one hour useful session. But I think with work it's not obvious to me, you know, we're building this thing. It's a bit like when we first built WolframAlpha. It wasn't obvious it was going to work. It wasn't obvious that you could get it to the point where it would have broad enough linguistic understanding, broad enough knowledge, broad enough computation capabilities that it actually had a definite use. I think this tutoring system is getting to that point. As I say, specifically where experimenting with Algebra 1, the Algebra 1 course, or what people think of as the Algebra 1 course. I've been surprised by what's in that course. I had no idea what was in that course, but I've been learning it. I suppose in that sense I'm a valid student for algebra 1 because I don't have any idea what's in an Algebra 1 course. They didn't call them that when I was growing up in England and I haven't paid enough attention, I suppose and I guess my kids must have done these, but I wasn't paying attention. Perhaps it's part of my. Well, I think they all did them differently. And my utterly unendeering comment to my daughter who's now a mathematician at some point when she was doing calculus by hand and me saying I didn't know people still did that kind of thing by hand because for me it had been automated for three decades before that time. But any case. So I would say the main thing, I feel like the main new piece of information I've got is I think with work, sort of AI based tutoring is actually going to work. And I think that that's one thing the whole world of kind of computational X and teaching people about computational thinking. I suppose another thing that's changed for me is I did Start this writing this book that I've been planning about Introduction to Computational Thinking. And I will eventually start posting pieces of it on the web, but I keep on having other projects that keep on getting in front of it. So it's my background project, but I have not had a chance to work on it in many months now. I hope I'll get back to it and be able to start making more progress. But thinking through that project and what it takes to kind of teach people computational thinking, a couple of things that have become clear. One is that this is a complete vacuum right now, that the kinds of things that one is teaching about thinking in terms of are just not things that are really anywhere in what's currently being centrally taught. Sometimes as a side effect, it gets taught, but it's not what's centrally being taught. The other thing I realized is that in many ways computational thinking, it's a branch of thinking, and as such, one is sort of teaching things that turn into thinking about thinking, but one can't do that abstractly. One has to do that in the context of actual ways of thinking about things in the world about which there are facts and so on. And so I've been sort of steadily understanding more about sort of what's involved in teaching computational thinking. And I've also realized that teaching people to think is. I find it really interesting. I have lots of fun doing it. It's sort of a good activity, but it's also a lot of it's hard work. And it's something where it's an awful lot easier on sort of the education production line to be doing teach machinery rather than teaching kind of the broad idea of how to think. I like to hope that some of my efforts in these live streams, and maybe we should do this more, have to do with kind of sort of showing a bit of thought processes, at least my attempts at thought processes in action. And that that might be kind of useful as a way to kind of see that one can figure things out. Because I think that there's often a tendency to think in sort of the education system and process, that anything that could be figured out has already been figured out. And if you are figuring out as part of your education, you're just rerunning a track that's already been run. And that's completely not the story of my life at least. And it's not the story of modern computational life in general. Computers give us so much of a sort of extra power to do things, extra leverage, that it's really easy to do Things that have never been done before. It's really easy to kind of think, think free range, so to speak, and not just follow the track that was followed before. I mean, there are many fields where the track that was followed before was the only one that could be followed, given that one fit the math in what human brains could do and so on. But if you remove that constraint, there's vastly more that can be done. And that kind of. It's a bit of a different experience of education, I think, than it is when you're on that track. And it's an experience of education that makes you much more face to face with. And now you have to think rather than. And now I turn to the machine. By machine I mean kind of the rules you learned for doing algebra or whatever else to get the result. Let's see, Let's see. Anon is asking would, would I ever go to Mars? I think the answer is no. I have. I'm just not some vera. You know, I've. I've traveled around the world a fair amount. There are places where. Plenty of places I've never been and where I'm like, look, it's not just not worth the trouble, it would be probably interesting, but it's kind of not worth the trouble. And the places where I'm. I'm kind of. I will say I'm curious and it's an interesting question whether, you know, even taking a trip to Earth orbit or something like this, would I get sort of a fundamentally different way of thinking about things that I do now by doing that. What would it mean internally for me, so to speak, now sometimes doing things like that. What it means internally is I achieved it. I managed to do this thing that, you know, I, I sort of always wondered if I could do type thing. But then there's the question of when I've traveled around the world and seen different places and talked to people in different places and so on and so on and so on. I get perspectives about kind of the world and life and so on that I didn't have before. And that's pretty useful. I don't know whether going to Earth orbit or the Moon or Mars, you know, would I get perspectives, you know, if I were, you know, if I were kind of courageous enough and, and sort of set up enough that I could do it, would I, you know, what, what new perspectives would I get? I don't know. It's. I mean, it's like an interesting question, you know, like from. If one didn't go on a plane and One. You know, when. When I was a kid, I certainly went on planes a bit, but they were not nearly as common, you know, that one would jump on a plane and go somewhere as they are today. And I think maybe in Europe also, they were a little bit less the concept of, oh, I take a plane to go from, you know, the middle of England to Scotland or something. Nobody would really think about that. Back in, let's say, the 60s, it would be like you take a train or maybe you drive, but not a plane. Whereas today, I think it's mostly, oh, I take a plane. So the question is, from the planes and not much in the way of planes to going on planes routinely, how did that change one's perspective of the world? I mean, I could certainly remember it. I always used to find it interesting to see kind of what a city looks like from a plane. It just, you know, the toy city type thing. And I always used to. I mean, it took me years before I wouldn't be looking out of the plane window, being interested in kind of what the perspective of the world from, you know, 10,000ft versus ground level. So I suppose that's a sort of partial example that I can't necessarily say what aspect of my general thinking about things was informed by that. Let's see, Des is asking about challenges. I think this about Mars. Challenges different from colonizing the bottom of the ocean other than obvious logistics? No, I. I don't think. I mean, you know, colonizing the oceans, colonizing Antarctica. These are all things people could do. And, you know, sort of the. The point is. What's the point? I think that one could imagine a world in which the colonization of Antarctica or the colonization of the deep ocean was something deeply aspirational. It's not right now, but the colonization of Mars is deeply aspirational. And why is that? I mean, it's kind of a cultural thing, I think, as much as anything else. I don't think one could argue. If one says kind of what I mean, one could start making arguments about how, well, if the world blows itself up, at least we'll have Mars. If the supply spacecraft don't bring all the killer viruses or whatever else, or the sort of mind virus of whatever terrible things are happening on Earth spreads to Mars because people watch the same television programs or something, or look at the same social media feeds. But in any case, one could equally make the argument that, gosh, if you're miles down in the ocean, you're pretty insulated from lots of things happening on the surface of the Earth. So it's kind of as good an insurance policy that way. Or even if you're in Antarctica, you're pretty far away from many things that could be happening. But I think for some reason it's just not as aspirational as, you know, go into the reach for the stars type thing or at least for the planets. Bob is commenting. Let's see, let's see. Elsie is commenting. Given the uptick in robotics, including humanoid, I wonder if there will be a point in sending humans to Mars. I think it's all, it's really about aspiration and that's not unimportant. It's like you can say, well, having a reason to do things is often an important part of what you get from doing those things. And I think this idea of, well, can we have, forget me taking a trip to Mars, I just want my humanoid avatar to take a trip to Mars. Speed of light makes us a bit difficult. You know, the 20 minute delay between, you know, move your arm and, or maybe it's more than that, probably might be an hour delay between move your arm and tell me what happened when you did that. It's, you know, Earth orbit much easier. You know, by the time we're communicating, having our cell phones bounce signals into, into Earth orbit, the latency is not that high. You know, we can, we could sort of hang out and with our avatar in space type thing on the moon, probably the delay is not too great. And it's, I mean it kind of reminds one of multiplayer video games where there's sort of, you know, in the past there'd been sort of transmission delays between different servers and so on. And now those delays might be in real life to control your humanoid robot on the moon or whatever. You know, I have to say I, I think I've said this before. I, I do feel that the, the, the, the non existence of VR tourism where you send out either the human or the humanoid robot or something to go hang out in a place instead of you hanging out in that place and you're experiencing it in VR and so on. And maybe in the future you can have some sort of smell based thing as well where you get to, you know, there's a, there's a chemical sensor at that end and there's a chemical sort of display at your end and you kind of can smell the rainforest as well as seeing the rainforest as well as, et cetera, et cetera, et cetera. And then it's kind of almost as good as being there. But that for whatever reason that hasn't really developed yet. Let's see. Ah, Des is commentary, I think, about dinosaurs. Wasn't there significantly higher oxygen back then? Yeah, that's what I thought I remembered. So thanks. And Memes is commenting that a pygmy stegosaurus would be adorable probably. I mean, my friend Nathan Myold has this whole theory that he's pretty sure is right, that dinosaurs whipped their tails at like above the speed of sound. So I think you have to be careful with that stegosaurus. Even the pygmy one could have a pretty, pretty serious whipping tail, so to speak. And various people commenting. I think maybe this particular group might not be the Mars tourist top demographic. Bob is commenting he wouldn't like to go to Mars. It seems boring. They don't even have a Starbucks there. It sounds like my wife, who will assess a civilization by the density of Starbucks and other similar establishments. Let's see. I think should wrap up in a moment. Maybe I have time for one more question here. All right, I'll try and tackle one here. How might the physics project help advance technologies like fusion power from memes? I don't know. I think. Let me see. There are different approaches to fusion. I mean, the, the one that most people are pursuing is confine a plasma at very high temperature, use magnetic methods to confine the plasma, kind of squash it and get the. Get by the very. By the fact that there's such a high temperature, the. All the ions, all the nuclei are whizzing around at high speed, get them to crash into each other fast enough, so to speak, that they overcome the electrical barrier and electrical repulsion barrier and are able to undergo fusion and then once that happens, get the heat out without destroying the fusion device. I think one of the problems is that as soon as fusion is happening, there's all kinds of neutrons being produced and all sorts of other particles. And any material that is sort of generating magnetic field, it's going to get shredded by all these random pieces of radiation going through it. And that will be. It's a question of can you sort of keep the container containing things or will the container's crystal structure be totally broken down and so on? It's actually a possibility to do it in space where you could imagine having liquid, having some kind of liquid thing as the enclosure. For the fusion reactor. And if it's a liquid, you can obviously cycle it out and sort of take the heat that's being generated from the fusion reaction, take it out, use it. And also a liquid is not as Persnickety about its crystal structure, because it doesn't have a crystal structure. And unless, and so long as you're maintaining it as being all mercury or whatever, and you're filtering out things that are no longer mercury because they were underwent nuclear reactions and nuclear transmutations, then you have a good thing going. And maybe that's easier to do in space where there's microgravity, than it is on Earth. Of course, if you say, well, let's put a fusion reactor really far away and let's get the energy from that fusion reactor and bring it down here on Earth. Well, actually we have one of those. It's really big, it's 93 million miles away, and it's the sun. So at some point it becomes a bit silly to say, put those fusion reactors far away and have them generate, for example, electromagnetic radiation or light or heat or whatever, and then we pick it up here on Earth, because we might as well just do that from the sun. So I'm not sure how that plays out, But I have long suspected that cold fusion, which became one of these things where experiments were done, people reported things, people said those experiments weren't right, et cetera, et cetera, et cetera. But the basic idea that you can take a material, in that case it was palladium, and have that material just absorb so many hydrogen ions, so many protons, that eventually their density becomes high enough that they can undergo fusion. I don't think that's a totally crazy idea. The fact that there's a consistent undercurrent of serious people reporting results along those lines. Now, the fact is it's very easy to get some fusion to happen. Lots of things in all kinds of situations, just with simple tiny particle accelerators, just accelerate desktop size particle accelerators, you can get some fusion to happen. But if it's just a few nuclei, it doesn't really do much for you. So question which was asked is, does our physics project reveal anything interesting about that? And that's a good question. I mean, I think the most likely vector for success, I would think would be not something like, oh, it's a new piece of physics, but more it's a better understanding of some existing piece of physics. My tendency would be to say that the thinking about various kinds of metamaterials, materials which don't have just a. I'm going to have a crystal structure that's arranged very regularly and grown in a particular way, but have a more complicated computational structure, Something that we've sort of come out of things about exploring the computational universe, the possibility of deterministic random crystals, for example, where there's only one way the atoms can arrange, but it's not a regular way. Things like that might be relevant to more sort of on the cold fusion end of how you achieve fusion. So my feeling is that the thing most likely to be relevant is issues of formalism and things where, oh, I never imagined there could be a material like that. But now as you think about it in the appropriate kind of computation all the way down type way, you realize, well, actually there is this way that that could potentially be achieved. My first thoughts on that. Interesting question. Good question. I really haven't thought about it. So that's why I like doing these live streams. You guys get me to think about all kinds of things I don't usually think about and certainly don't try and verbalize as I try to do here. So thanks for all the questions. I know I didn't get to everything today. Hopefully I will be able to do that another time. So look forward to talking to you then. Bye for now.