Future of Science and Technology Q&A (August 30, 2024)

Hello, everyone. Welcome to another episode of Q and A about future of science and technology. I seem to have been doing a bunch of live streams this week, so this will be at least different. All right, let's see, what do we have here? Ty asks, what are your thoughts on machine learning to create new genera? Like, what would be a good way to go about doing something like that, Like a new genera of plants or animals. Wow, interesting. So, first point, how do new genera get produced normally in the course of biological evolution? Because you know what's happening in biological evolution, There are small changes to the genome. And sometimes those small changes lead to organisms that manage to make more baby organisms, so to speak, in the environment as it is, than with another change. And so that becomes sort of a chain reaction of the change which makes an organism make more copies of itself, so to speak. That change will get copied more and so that change will come to dominate the population. So that's kind of the normal way that sort of biological evolution operates. We can go in to some extent and make genetic modifications that for whatever reason we figured out, are going to make things work in a particular way. Like we could insert jellyfish genes into some random kind of plant, for example, and make it glow in the dark. That was a recent thing that was done. And so we can go and specifically edit the genome to do that. Now the question is, most of what we know how to do is fairly small changes. Like for instance, make a plant that is otherwise just like an ordinary plant, but happens to have jellyfish genes, so it grows in the dark. If you said, can you make a plant that has legs and walks around? We have no idea how to do that. And at best we can make these kind of point changes. And typically the nature of those changes is have the organism produce some protein that it wouldn't otherwise produce. Like in the case of Glow in the dark, it's this GFP protein that was found in jellyfish and so on. That is the idea of, can you sprout extra legs or something. That's. That's beyond what we know how to do. I mean, we don't even know for one of the things that seems to be true for humans, many traits for humans, like for example, human heights, we don't really know what genes control. That seems to be a whole collection of them. It seems like most of the things that are kind of the sort of the large scale traits that you would say, oh, that's a different genera of Krita. Most of those traits are spread across many, many, many different Pieces of the genome. So it's as if we're kind of, you know, we're taking a program that's running on our computer and if we look at kind of the machine code of that program, probably it was written by a human. And in that, in programs that are written by human, machine code will be somewhat organized. There'll be a piece that, you know, deals with closing the window. There'll be a piece that deals, deals with editing, et cetera, et cetera, et cetera. But in biology, nothing like that seems to be true. It seems to be the case that the stuff that deals with, you know, one aspect of, you know, how your, you know, finger moves or something will be encoded on chromosome 13 in this place. And something right next to it that controls something about how your finger move will be encoded on chromosome 1 or something like this. The programs are just sprayed all over the place and we don't know how to pull them together and read them. So if you say, is there going to be a way of making sort of a large scale change? The answer is right now that's beyond what we know how to do. I think it is. And will machine learning save us? No, I don't think so. Why not? Well, because, and this is something I happen to have worked on very recently, when one thinks about what's going on in biological evolution, when one says, okay, I want this organism to have some particular trait, like to be able to be a tree that gets as tall as possible without having too spindly a trunk and this, that and the other. And you say, well, how does evolution achieve that? And I managed to make a very minimal model of that where you can kind of see all the machinery going on. And it's not talking about actual trees with bark and all that kind of thing. It's very idealized relative to that, just bits on a screen, so to speak, then what you find is that what's happening is that sort of this irreducible computation that gets done in the process of the organism growing. For example, the sort of process of evolution will pick out lumps of computation that successfully do the thing you want to have done. But there's no sort of simple mechanism to recognize that. There's nothing where you could say, okay, I understand what's going on, I can engineer this, I can tweak it to make it work this way versus that way. It's just there is this constraint you've given, and the thing will eventually find a lump of computation that satisfies that constraint and then go in that direction. Now actually it's interesting and I actually haven't looked at this and I should to understand the extent to which the information on what the thing does is distributed across the kind of very tiny sort of toy model genomes of these organisms that I'm looking at. But I think so, you know, this machine learning is not good at dealing with computational irreducibility. It's not what current machine learning can handle, and I think it's not what machine learning in the way it's currently conceived can really get to. Machine learning deals with these kind of sort of broad searches, shallow computations, these kinds of things. So I think the let's grow a major different branch in the tree of life. It just is not a plausible thing, I think, to have happen. It's not something you don't get to have the many steps that you need to achieve that you get to do something where it's sort of a small incremental change. I mean, one saw that, for example, with protein folding computation using machine learning, if the protein you're trying to fold is pretty close to a protein that's already been folded, which is, you know, which you've trained from, then all works well. If you say, go out and make me a random protein, then it will typically not do very well at figuring out how it should fold. And often there isn't data to confirm or deny that, but that's the impression one has. So I mean, I think this idea of, you know, can you sort of hack the genome to achieve things that are, for example, so morphologically different that they would be considered different genera? That seems very tough. If you're asking can you make something that will produce this particular protein, like the jellyfish protein, the answer is that's much more doable. But that's just add a molecule, subtract a molecule, change this couple of molecules that are produced. Not. Does the thing now suddenly have wheels or something? You know, it's funny in sort of the science fiction world. I remember the original Blade Runner movie opens with a sort of a genetics hacker who has all sorts of little critters around his apartment or something that are like opening doors and doing this. And it's kind of special purpose biology organisms. And I think that is a very hard to achieve situation. So I think, and the whole question about how do we evolve different genera of organisms? It's. Well, one doesn't know a lot about that. I think the little minimal model of evolution that I made a couple of months ago is the first time actually one can perhaps do a real study of sort of how this kind of branching in the tree of life really works. And I had some examples of that. There's a lot more that needs to be studied there of kind of to what extent. What you end up with is sort of different branches of organisms have different ideas, in effect, about how to be fitter organisms. So, you know, one organism may make use of, I don't know, the oxygen in the atmosphere, another one may make use of the carbon dioxide in the atmosphere. They're sort of finding different niches and so on. Or even in the case where they're trying to achieve the same thing, they're finding different ways to achieve that thing. So a few thoughts about that. Let's see. Vaughn is asking, can you talk about the future of information gathering and research? Say I'm discussing with a robot about a paper I'm writing, and the robot is providing examples and evidence to support my arguments. Do I cite the robot as my source, or do I have to find where the robot got the information? Okay, so that's always an interesting question. Where is the ultimate source of information? And this is something we deal with with Wolfram language, we've dealt with with WolframAlpha and so on. I think the thing that is important to understand about sort of that, that supply chain of information is where's the place where somebody is somehow taking responsibility for this information. They may not be the original producer of it, but they're somehow taking responsibility and say, hey, I actually looked at this and I think it's right type thing. And it's kind of like, do you cite an encyclopedia or do you have to go back to the original sources? No, you cite the encyclopedia because the encyclopedia is taking responsibility for what's in it. You cite Wolfram Alpha, Wolfram Language. Yes, you do, because we're taking responsibility for what's in there. Do you cite Twitter, for example, or do you cite the person on Twitter? Well, you cite the person on Twitter. Twitter itself is not taking responsibility or only very, very little for what's there. And similarly, it would be. Well, interesting question if you have, for example, I don't know, a conference proceedings or something, and there's a paper in there and they're people who edited the conference proceeding. You know, it's conventional to cite those people partly because they're part of the chain of taking responsibility for the thing that you're quoting. So now you ask about AI and the question of when. So in a situation like Wolfram Language or Wolfram Alpha, I Think it's pretty clear cut. The thing you cite is the place where responsibility was taken by an entity capable of taking responsibility. Now if the AI went and kind of foraged a bunch of information and then is presenting it to you, well, kind of good luck to you. I mean, that is fraught with difficulty because the way LLMs work, for example, being able to, well, there's a little bit of technology now to sort of go back and find the original source. But in terms of what the LLM made up to say this is a citation, it's as likely as not to be something completely made up. So you would be in a sense right to cite it because it is doing something that it alone can take responsibility for. And by the way, it may be full of nonsense, but then you're citing that. So people know where it comes from and they know it is potentially full of nonsense, so to speak. So I think I have to say that in terms of sort of how do you find information for research purposes, I think it's something that for a really long time, I mean really long time, back to the 1970s, it's been possible for technical papers and things to, to do keyword searches and such like. And that's a pretty good source. Nowadays it's become possible to do these broader kind of. What was the general view of this among sort of a whole collection of documents of this kind. And that's something where you can potentially, you can expect something somewhat useful from the LLM, but you can't expect something sort of formalized where you can say, yes, this is the precise answer type thing. You know, something sort of intermediate between these things is vector databases and rags where you're saying, okay, in a traditional keyword search you're saying, show me the article by Smith about porcupines or something. And it's like Smith, porcupine and it's definite anchor words where you're just doing a search. Well, in a vector database, what you would be doing is, is taking, let's say, sentences or whatever from the document and turning, grinding them up into vectors of thousands of numbers. And then there's a vector of thousands of numbers that somehow is typical of things about porcupines and things about this and so on. And then what you're doing when you look up in a vector database, a sort of semantic search is you're saying this vector of numbers associated with the question I'm asking what vector of numbers is closest to that vector from the things that I stored. So for example, in the latest version of Wolfram language. We have this whole functionality around semantic search and create vector database where you can feed it a million documents, let's say, and grind it up into sentences. You've got a billion sentences and each sentence has a certain vector associated with it. Then if you ask the question about porcupines, then it's going to be able to very quickly tell you what vector, what sort of meaning like vector is closest to the one that has the meaning of the question you're asking about porcupines. And again, that's not. When you do a keyword search, assuming the data wasn't bad and things weren't misspelled and so on, you're pretty confident you found all the papers by Smith about porcupines. You know, assuming that the. It's, it's a fairly deterministic thing. By the time you're using a vector database and does, you know, hedgehogs and porcupines, they mix up and do they not mix up and so on. It gets fuzzier. By the time you're going through a whole LLM, it's even much, much, much fuzzier. But I think the, so I mean I, I would say the, the that the main sort of I, I think you can, it is good to get an indication by, you know, talk to the LLM. It'll tell you something and it may be able to help you get started and it may point you in some direction. The vector database is going to get you closer the, you know, by the time you're actually, you know, using some aggregated source of data and knowledge like or from language or something, then you've got to something which is sort of solid enough that it is reasonable to build on that foundation for the research that you're doing. You know, I remember when Mathematica was young, oh, 36, 37 years ago, something like that, people would be, would say, oh, you know, I'm going to write this math paper based on results I got from Mathematica. How do I know that they're right? And it's like, you know, wouldn't it be better if we could, you know, if we could, if somebody could peer review all of the code that went into computing this result. And it's like, well, no, actually because imagine humans going through 5 million lines of code or whatever it was at that time and saying, yeah, we're sure these 5 million lines are all correct. There's not a chance that would work. Not a chance. And when academic papers get peer reviewed, so to speak, the failure rate of that I have no idea. I'm sure people have done studies, but it must be in the 90 something percent in terms of failures of. Did you actually catch the thing that was horribly wrong? Absolutely not. Not the nature of the beast, so to speak. But when it comes to something like software, there are two things that go on. One, it's used by a lot of other people, so if there's a horrible problem, it's probably going to have been caught by somebody else. And two, even more important, you can do automated software quality assurance. And so there's a much more sort of certainty associated with that than there would be with just, oh, it's humans checking, checking things kind of thing. And I think that's again, something where when you say, well, you know, is it right? How can I be sure it's right? I think you get some of the best results from, I have to say, technology of ours and, you know, the kinds of things that we built. Because it is something where there's all kinds of automated testing you can do. You know, when we put out things about, I don't know, socioeconomic data or something like that, we've done a bunch of automated testing that checks that yes, it's plausible, no, it doesn't violate some basic principle of these numbers have to be greater than that number and so on. So it's kind of the best case scenario, so to speak. So I think that's a pretty solid thing to start from. All right, let me see. Well, let's see. We've got so many questions. You guys ask such interesting questions here. Okay, where do I go Next? My gosh. NT is asking, how advanced do you think AI available to consumers will be by August 2029? I don't think it will be dramatically so. What has happened in the history of machine learning and AI is a series of sort of steps up. So for a long time, AI was like, oh, it doesn't work at all. 2011 comes around and oh, neural nets and deep learning work for recognizing images of cats versus dogs and so on. Wow, that wasn't clear. You know, sort of by bashing the thing hard enough and finding tricks that got solved, the image identifiers of today are not that much better than the ones that existed by 2012 or so. So that was like a thing, a sort of breakthrough happened. And then it's kind of flat for quite a while. It's just incremental progress. Same thing happened to things like speech, to text, you know, and that's gradually gotten better. And the thing that does tend to happen and will happen more is sort of optimizing the way that the AI is done. So instead of, oh my gosh, it's, you know, these really expensive, complicated GPU chips doing everything. It won't be long before a lot of these things are, oh yes, I can run it on my phone. So that's one thing that will happen and that will change a bunch of use cases of what people are prepared to do with AI, what they do do with AI and so on. But in terms of what kinds of wows will there be associated with? Well, now it can do this, the observation of history. And my guess is that it sort of operates in a series of steps and we can sort of count the steps, so to speak. Like we got, you know, we got things like image generation, text to image, you know, make an image of an elephant dancing on the moon or something. And you could get that that was a thing a little bit before LLM story of ChatGpt in the end of 2022, where it's like, continue this piece of text or answer this question. That's another kind of step up. I think that what happens, I don't think it will be another kind of, oh, wow, now AIs can suddenly do this. I mean, there'll be a few of those like, oh, you know, describe a movie, make a reasonable movie that's beginning but not quite there yet. That will surely happen. And being able to integrate these different modalities, that's begun to happen of text versus you're looking at image versus you're generating an image and so on. But in terms of the wow, we had no idea that was going to happen as occurred with ChatGPT or the image generation systems. I'm not sure. I mean, maybe I'm missing something, but I don't see immediate places where that's going to happen. Now in terms of sort of the incremental improvement of putting better harnesses around these AIs, doing better engineering around them, that will absolutely happen. I think one of the big things that we've been deeply involved in is providing kind of computation and knowledge to the AI, being able to have it be the LLM be kind of the linguistic interface that's chatting with you, but then on the back end it's kind of asking Wolfram language or something what's actually the answer to this. And then getting that back, it becomes this kind of conversational interface to hard computation. And I think that's a thing that we'll see a bunch more develop in. And I Think that it will become routine. Just like there was a time before there were graphical user interfaces for computers and where most of the time you're just typing text. I think the linguistic user interface for computers will become commonplace and will be pretty. Pretty universally used for lots of purposes. And it's just like, you know, at this point, we can use speech to text for some kinds of things, but there are things where you really don't want to, you know, if you're going to make that, you know, buy a car or something, you probably don't want to buy a car by sort of telling your phone what you want with voice and then saying, yeah, go buy this car. And it's like, oh, we just charged your credit card or whatever it is, you know, whatever $50,000 or whatever it is that you. Probably most people are not going to want to do that with the voice transaction because it just isn't really reliable enough. And I think the thing that. So, you know, there'll be kind of the chatty interface thing and then there'll be. What does that fall through to underneath? And that's the kind of thing we spent a lot of time building. But, but I think that the chatty interface thing will become pretty ubiquitous as graphical user interfaces have become ubiquitous. So that's my guess there. Let's see, Let's see. Gracie asks about or Grace. I don't know. Preparing to pursue a PhD in fiber science, my research interests like the intersection of computational material science and sustainable textile innovation. Background in pharmaceutical sciences. Been exploring how advanced computational methods can be applied to fiber science, specifically in developing smart and sustainable textiles. How do you see quantum computing impacting the modeling and simulation of complex fibers and polymers? My gosh. Well, let's see. I'm not a big believer in sort of magic from quantum computing. I think that the first question is sort of what's the whole kind of, you know, what is one trying to achieve? You know, where do the fibers we have right now come from? You know, there's cotton, there's wool. You know, cotton's from a plant, wool's from an animal. Nylon is, you know, from. Is a polymer that's just sort of made in a lab or made in industrial process. And I suppose an interesting question would be, what sort of. Sort of, what's the. What do you want from your fibers, so to speak? Now, you know, I see people doing things like, you know, having. Knitting wires into clothes and so on so that you can have kind of, or. Or knitting kind of active components into clothes for Example, so that you can have. Yeah. The display that's just integrated as part of your shirt or something like this. And, you know, who knows what fashion could. Could happen if we no longer have static clothing, but we have, you know, billboard style kind of programmable clothing that can have that. That doesn't just have kind of, you know, shimmering things or things that. That has a full, you know, play a movie on yourself kind of kind of story. I mean, I think it depends on what direction one's going in with. I think there are things where. Making sort of things more programmable. I think there are other questions, like, can you make. Oh, here's. Here's a question. Can you make something where. Okay, I mean, a big trend in a lot of areas is have a thing be more programmable, so to speak, instead of just. It just sits there. Have it. Be something that you can, at even a microscopic scale, control what it does. So, for example, you might say, can I make myself a hat that will. Just by the touch of a button or something, the hat will transform from being something that is, I don't know, very, you know, UV resistant and prevents me getting sunburned or something to something that is very, you know, is. Is. Cools my head in the best possible way. You know, is it. Is there a way of having sort of fibers that at. At that have some kind of activity to them as opposed to just being things that are part of are, you know, that are just statically presented? And I suppose that's sort of a question of whether you can imagine that. Well, I'm thinking of. I mean, like, for example, clearly in us, our muscles are kind of fibers. They're triple helix proteins and so on that have this feature that with electrical stimulation they can be made to contract and so on. And one sort of. One might imagine, you know, could you do the same kind of thing? I don't know whether that's really a. I don't know whether for clothing or for other purposes where you have some sort of thing where the actual individual fibers have some special activity that you can control in some way. I mean, I think, you know, I kind of wonder what. You know, it depends what you're trying to achieve. I mean, I suppose there's the. How do you make clothing? Doesn't wrinkle and so on. I mean, that seems to have been achieved somewhat. I don't completely know how. And that's presumably sort of fiber material science. I think this question of. Well, so another question is, can you simulate what fibers will do in other Words. If you imagine, oh, I'm going to make this fiber and I'm going to use this particular molecule or whatever else, oh, I'm going to spin the thing in this particular way, you know, what will it do when I put lots of them together and I don't know the state of simulating that type of thing. I do remember, oh, gosh, it must be 30 years ago now, was when in computer graphics there was this moment where people could do text, do fabrics. Before that time, if you saw, you know, the, you know, the character with the cloak or something, in computer graphics, the cloak would look totally fake. But then it was about 30 years ago, sort of the problem got solved of doing the mechanics and its sort of elaborate, kind of elaborate dynamics of a fabric object. It's not really a solid. It's not, you know, it has, it has its own kind of dynamics got solved of actually being able to simulate those things accurately. And I don't know, at the more microscopic level, going from sort of the characteristics of a fiber to the characteristics of the cloth that you make from that fiber, I'm not sure how that's progressed. So a few thoughts, more to say, I think. Let's see. Oh, my gosh, so many interesting questions. All right, let me, Let me try going through a few of these here. So misaligned ask, what's your take on integrating memory into LLMs to enable retention across sessions? How could this impact their performance and capabilities? It's really not that difficult to maintain retention across sessions. You really just have to tell the LLM, okay, summarize what happened in this previous session, and then you feed that as part of the context to the LLM for the next session. So, for example, we're building an AI tutor system. It's a rather difficult project, but one of the things that's an issue there is how do you go from lesson to lesson, kind of remembering what happened before? And what we're doing is we're having the LLM summarize what happened in the previous lesson. That's useful for other purposes as well, for, for people who want to know, is this kid actually progressing? Can I read something about what's going on? But it's also useful for the LLM itself because then it can look at the summary of the previous, of what happened in previous sessions. And then if that gets too big, it can summarize the summaries, so to speak, and always have something of sort of manageable size that it can use as part of its prompt as it goes forward. Into the next session. So I think that's a thing that's not so hard to do in terms of sort of the changing the internal architecture of an LLM so that it has a different way of sort of maintaining memory about things that it's seen in the, in a document. Well, most of what's gone on there is, I mean, you know, that's sort of integral to the way that the whole transformer attention type mechanism works is some level of look back and the document. Another thing coming back to what I was mentioning earlier about vector databases. You can absolutely take things like sessions that people have had with LLMs and put them in vector databases so that you can do this thing of asking, hey, did I talk about that before? And across even a very large number of sessions, you can easily do a search with semantic searching that says with some degree of certainty. Not perfect, but some degree of certainty. Oh yes, I did or didn't talk about that before. Let's see, What are your intuitions about AI generated fake content to deceive people, Whether using deep fake face swaps or voice cloning or one or more of these things combined. Are we rapidly approaching a point where we won't be able to trust anything on the Internet? Yeah, well, I mean, the ability to produce sort of, to synthesize content is going to get better and better. There'll be little glitches. You know, it's like the, the image generation things. You know, you might not have the right number of fingers on the hand because that wasn't something that was, was captured well, but then that problem more or less got solved. There are definitely ways you could imagine, you know, if you are a person who is, you know, wants to be sure that the things, the, the images of them are real. Well, you know, you could start off by having, you know, wear a tie with a QR code on it or something, or wear a, wear a shirt with a QR code and, and you know, like any real image of you, you is going to have the QR code that's actually readable and so on, and maybe a QR code that encodes some of the rest of what you're wearing that day. Let's say you wear the QR code and the QR code has some encoding of what you're wearing. And then anybody who sees the picture, they can in principle look at the QR code and they can say, hey, does it agree with this? And that's not something that this generation of image generation technology is capable of dealing with. I mean, it's kind of like some sort of strange inverse or something of the capture type approach of are there things in the image that you can readily check or things in the image that are sufficiently sort of computed and structured that they're not things that an image generation system would have been able to come up with? And there are things where there's cross consistency between different things in the image. So there are things like that that you can do. Another approach is something which, for example, some camera manufacturers have started doing of sort of can you put metadata along with the image that kind of says this is where it was taken and you spray data into the image itself because an image is full of lots of pixels and if you tweak a few pixels, nobody would know the difference. But those tweaked pixels can encode some information that you can use to, for example, check some aspect of the image. The so called steganography approach of hiding messages in kind of just the structure of a thing by twiddling a few tiny little changes to pixels. Or if it's text, you can do things like, you know, you can reword it, that's an extreme thing. Or you can just change the spacing a little bit, the spaces a little bit to sort of encode something where somebody would just read the text and it's like, okay, it says the cat is out of the bag or something, but actually you know that the cat is out, that the ears has an extra space before it and the, you know, this and that and the other. And gradually you're putting this extra information that isn't really noticed in the message in there that you can use to do things with it. And one can certainly imagine doing that for images as a way to kind of check whether what's going on is real. Now, you know, an interesting question is, for example, before HTTPs you could, there would be websites which would be, well, I'm masquerading as such and such a website. I'm not really coming from that web domain, I'm really coming from somewhere else, but I'm pretending I'm that website. But then HTTPs has this handshake between your browser and the website and then it has a, uses a certificate authority and so on to sort of ensure that the thing, the website you think you're talking to is actually the one you are talking to. And I can certainly imagine a time when the typical image that you see in some news article or whatever else it is something has some, has something where your browser for example, is telling you, yeah, this thing is real. I just checked the metadata and validated this thing and so on. I don't expect people to go and do all that validation themselves, but one can imagine some sort of world of an automated system where there's a bunch of validation going on, where, for example, you have pretty good information that, yes, if this is a photograph, it is an untampered photograph that was taken at this time and place, for example. One can certainly imagine that happening and that becoming a thing that people come to expect. Just like when you see a website that isn't an HTTPs website, well, probably your browser has a bunch of red things flashing that says this might be unsafe. And I think one could imagine the same kind of thing with things like images. So I think that approach of checking the provenance of something is one thing. When it comes to, you know, is that true or not? That is such a mess. I mean, you know, what does true mean? What does. What is, you know, when somebody says, you know, everybody thought the speech was great, okay, you know, is that true? Is it false? You know, that's. It's not the kind of thing where there is a true versus false. If somebody says, or even somebody would say something like, oh, I don't know, so and so won the election in small country X or something. Well, yes, in principle, they won the election, but actually, you know, everybody said it was a fake or actually, you know, the presidential palace is encircled by tanks and, you know, the future doesn't look bright and so on or whatever it is. I mean, you know, things which. It's very hard. You know, there are some things where you can say definitively, yeah, that's true, that's not true. You know, is such and such a person alive or dead? That's probably a, you know, may be hard to determine, but ultimately that is a definitive kind of question. But, you know, a lot of other things are surprisingly difficult. You know, if you might have a glib answer, but that answer might be in the end, something that depends on all kinds of opinion, et cetera, et cetera, et cetera. And I think it's, you know, I think that's a, that's sort of hopeless. But I do think that the idea that certain kinds of material that you see, news photographs, things like this, probably there will be some infrastructure created that sort of is able to validate those things based on some, some level of, I mean, you know, how do you tell that that picture was taken at that place? Well, you can tell that picture was not taken, you know, that the picture was Taken was not, let's see, you can, by knowing that, for example, you put some piece of information on a blockchain at that date, you know that that information existed by the time of that happened. So the thing you can validate that that picture could be that old, so to speak. And one could certainly imagine blockchains are really good for that as sort of a ledger of this happened. And the blockchain is very, you know, by its very setup, it's like, yep, this occurred in block number such and such. And block number such and such was mined in such and such a time. And there's no way you managed to put the thing in that block. You know, you. Because, you know, you can't sort of time travel back to put the thing in the block. That means the thing really existed by the time you put it in that block. So that's an example of something where you could do pretty good validation when it comes to. So, you know, when something was done at that level, at least can you know where it was done? Well, if you have a piece of hardware that has GPS and all this kind of thing, then you can, and that piece of hardware has sort of cryptographic signatures and so on in it. You can be pretty certain that unless you went to even more trouble and you spoofed the GPS and all that kind of thing, that yes, it really was. The picture was taken at that place at that time. I mean, ultimately, if you're really, you know, if you're trying to confuse things, you could spoof the GPS and then all bets are off about where anything was. Okay, very different question here from kd. When do you'd expect the discovery of life on an exoplanet? Depends what you mean by life. And that's complicated because we have, you know, we have n of 1. We only know one example of life and life on Earth, sort of all. All life on Earth is described, derived from a common ancestor that actually probably existed surprisingly short amounts of time, few hundred million years maybe after the formation of the Earth. So there's a common ancestor to all life and seems to these days go by the name of luca, the last universal common ancestor. And it already has tons of stuff that is features of life as we know it, the cell membrane, the rna, DNA, this whole setup, so to speak. So, you know, in our one example of life on Earth, a lot of stuff, very detailed stuff, is a particular way. Now, you know, if we ask, well, what's the generalization of life? What is something which is kind of like life, but Isn't quite the same in detail as life on Earth? Well, we don't really know. It's probably something that involves some sort of molecular scale computation of a kind that doesn't occur in other kinds of natural systems, et cetera, et cetera, et cetera. But there's plenty of computation that goes on in any natural system. The issue is, does it have the kind of bulk organization that seems to be characteristic of life as we know it? I have to say that I've been sort of interested in this definition of life question for a long time, and my views on it have changed a bit recently because I've kind of come to realize that one of the features of life is that at a molecular scale there is control and detailed sort of orchestration, organization of things in a way that doesn't really seem to exist in most other places. I mean, it's kind of like one place it does exist is something like a microprocessor which has, you know, very detailed circuit with a billion transistors or something that's been laid out in this intentional way, so to speak. In the case of life, it's, you know, evolved through some kind of process of biological evolution, but it still has some kind of. It's some kind of. Has some very hard to put one's finger on purposefulness to it. I don't know how to characterize that purposefulness. I'd really like to know. I think it's an important theoretical foundational question in biology. But with that kind of definition of life, something which has this kind of bulk organization at a molecular scale, that's a reasonable level of definition. I think if you say life like on Earth, with RNA and DNA and this particular collection of chemicals and so on, that's a different definition. And if you say something that can reproduce itself, for example, by any means necessary, that's a yet different question. I think the thing I mentioned about molecular scale orchestration is the most interesting definition of a kind of lifelike thing. And I think that is, I think life like we have it on Earth with RNA and DNA. I doubt we'll see that anywhere in precisely the form that we have it here on Earth. Things that sort of reproduce themselves by all means necessary. We'll see that in a bunch of places, this molecular scale organization. That's an interesting question. Now, how would we detect it? Well. Well, the problem is that the kind of measurements we know to do right now, looking at light going through the atmosphere of an exoplanet, for example, that allows us to get Spectroscopic information, information on things like vibrations in molecules, things of this kind. But those are molecules in the atmosphere of the planet. And I'm not sure how much we can deduce from. From on. On Earth in order to really see life in operation. We're dealing with these much more solid phase things where we're not really solid phase, but we're not really liquid phase either. We're some weird intermediate biological phase. But we're at least sort of. Everything is sort of hanging together in some sense. And I don't know how we get to observe those kinds of things for an exoplanet. Now, having said that, one of the sort of features of experimental science is there's always a better way to make the observation. It always becomes more sensitive. There's always something more elaborate that you can do. And so I don't know what would be the thing that would let one kind of, you know, I don't imagine we'll have. I don't think we'll have telescopes that can kind of see to an exoplanet and look at things that are a meter across on the exoplanet. It's not physically impossible. If we had a big enough telescope aperture in space with little enough kind of other effects, it's not outside of the realm of physical possibility, just really hard to do. And what would we try to detect on the exoplanet? I'm not quite sure. So I think the answer is what will happen is people will say, oh, my gosh, there's this molecule in the atmosphere of this exoplanet. By the way, by the time a molecule is really big enough to be super interesting for life, it's not going to be volatile enough to be hanging out in the atmosphere. But, you know, I think it's going to be. It's going to be tough to get something definitive that isn't just, oh, how can we imagine how it manages to get so much oxygen in its atmosphere unless there's a, you know, unless there's plants there or something. I think those arguments will be nonsense. I think this. How can we imagine how it gets so much oxygen without plants? Well, actually, it turns out there's a type of rock that doesn't. We don't happen to be aware of where you mean, you crush it up and you combine it with methane or something, it produces lots of oxygen or something like that. There'll be a way that the system, the geological system managed to find that was just not one that we imagined. So I don't think those indirect indications will be terribly accurate. Let's see a question. Here is the hype around LLMs dying finally relegating the toys to the toy box where they belong? Or maybe. Or do you think anyone will ever be able to make them useful and accurate? Well, as I said, you know, I think LLMs are great linguistic interfaces. I think that the back end and the connection to the back end of, well, what do you actually compute? Where is the thing where we've managed to create this very formalized way of doing computation and doing sort of accurate things and we shouldn't throw that all away and say, oh no, now we have LLMs. We don't need any of that stuff. No, actually we absolutely need that stuff. The LLMs do something which is quite like what humans do, which is they're able to be chatty and they're able to sort of fit together fuzzy things. But if you ask the LLM to solve this complicated math problem or something like that, the LLM is not going to be able to do it. It's not the nature of that kind of thing. It is not the nature of a thing that is trainable in the way that an LLM is trained. In fact, this recent work that I was doing about kind of the foundations of machine learning make it pretty clear that what's really going on in the training of an LLM is that there are sort of lumps of computation that you are happening to find fit together in a way that's useful for doing what you want to do with the LLM. And that's sort of not that, you know, oh, there are lumps of computation that fit together in a way that kind of just about works for the thing we're trying to do. If you say, let me use that to do a precise computation, it's going to run for a million steps. It's just not going to work. Now an interesting question is whether, like the things I've done on sort of the foundations of machine learning, whether they will give us any hint about how to much more fine grain the process of having an LLM call tools and so on rather than the LLM is running and then it generates a prompt inside itself saying now call this tool. And it's all sort of textual and so on. Is there a way of kind of making a more fine grained call, more fine grained kind of neural implant to the LLM? We don't know yet. The stuff I've done recently I think is the best shot we have to actually see how that might work, but I don't know it yet. Let's see. Golagaz asks, do you think future cars will be able to get rid of wheels? Interesting question. You know, it's notable of course, that in biology there aren't very many wheels. I mean, some bacteria have flagella that have rotary motion, but generally rotary motion is not well liked in biology, partly because it's kind of tough to supply if you have a wheel. It's kind of tough in the ways that things are usually set up to have a blood supply going to a wheel, for example, and that's not necessary at the level of a flagellum in a bacteria, for example, but that would be necessary for the tiger with wheels or something like this. So the question is, for example, what other method of locomotion could you have other than wheels? Well, you can float or you can walk with legs, or you could have like caterpillar tracks, I suppose. And I don't see anytime soon, you know, the millipede car, I think that the, the car kind of, we were there once, the car with four legs, so to speak, AKA a horse. And I think, you know, the smooth ride, you know, if you want a smooth ride, don't have a four legged animal. It's because obviously, I mean by the time if you were riding a millipede, you would have a smoother ride. And you know, you can ask the question, will there be millipede cars? And I, I don't really see it. It doesn't seem, I mean it doesn't seem like now is it the case that there will be useful things from having millipede cars? Well, you could go over all sorts of interesting terrain. And one of the things that happens is in a lot of these kinds of devices the big issue is control. I mean that was the Wright brothers with airplanes. That was sort of the big breakthrough was figuring out more about not the forces needed to fly, but the control you needed to keep the thing stable as it flew. And that's what, for example, held up things like drones, for instance. If you want a human controlling, you know, eight rotors or something like that, good luck. Humans don't manage to do that. You need a microprocessor that is able to respond to signals quickly enough and so on to be able to change rotor speeds and whatever else to keep the thing stable. So I guess the thing to say about that is, as you know, people have built flight controllers. I don't know of anybody who's built a millipede controller yet, but I'm sure that's possible at this point. And you know, so maybe coming soon I suppose, but I'm not sure that's a, it puts a different spin on sort of the notion of caterpillar tracks. Except it's not a caterpillar, it's a millipede. Instead of having, you know, of having the thing run around like that now, I mean the other possibility is to fly. And the main problem with flying is it takes a bunch of energy and like the form factor of, you know, it depends on how much energy you can store in your batteries or whatever else you're using as your sort of energy density dense energy source. I think that, you know, one of the things that also made drones possible was better battery technology. And there's nothing so far as I know that sort of ultimately. Limits, there are limits, but they're far away. And in terms of what in principle you could do with sort of storing energy in something. So you know, the real question then is do we get flying cars anytime soon? And you know, I think that really a lot has to do with well, do we have the energy density, do we have, you know, if your flying car can fly for only two minutes and then has to be plugged in for an hour, that's really not going to cut it. I think the other issue is do we have sort of complete stability of the flying car? That seems much more plausible. That seems like that's the thing that one could achieve now. And I know some designs of flying cars have the, if all else fails, deploy a parachute type thing. I think there are sort of questions about, you know, when people get used to self driving cars, self driving taxis and so on, on the ground it becomes a lot more feasible to have a, a self driving air taxi. Because before that, you know, one of the issues is it's, it's tougher to get a pilot's license than to get a driving license. It's more complicated to fly a plane than to drive a car for humans. And if you say, well everybody who's going to do, you know, have the flying car has to have a, a pilot's license, that's going to dramatically reduce your market. But if people get used to the idea, good or bad, that it's going to be a self driving little flying car, then that, you know, that doesn't remove, you know, doesn't, people don't have to have pilot's licenses to, to, to go in the self driving flying car. But I, I don't see that's my only thoughts on the getting rid of wheels question, I mean, I suppose the other, you know, one of the things that's sort of interesting is when are you on rails and when are you kind of free range? Like, you know, obviously railroads were built out a long time before cars got developed. And you know, the idea that you, so you can go from point A to point B, where you decide where you're going and it's sort of a personal thing that's going from A to B and and so on, as opposed to everybody gets on the, on the 9:13 train from this place to that place. And, you know, I think that's, that's another question of, of to what extent do you personalize things? And, and I suppose that's the, that's in a sense the flying car question, because right now when it comes to planes, pretty much it's lots of people flying in a definite place on one plane. I mean, even, you know, and the world of corporate jets and such like is a pretty expensive world right now. And, you know, is it, is it going to be the case at some point that that kind of becomes a thing where you can just like, okay, I'm just going to take my plane from here there. I certainly have many friends who do that, whether with chartered planes or flying themselves. But it's definitely not exactly a consumer activity, nor is it something that's terribly, you know, that the. I, I know just a few people who park their planes by their houses. That's really a rare case. Let's see, Gist is asking what algorithms change the world the most? What's the next algorithm that will change the world? How does one release such an algorithm? So the result is positive. What are some famous algorithms? There are some kind of, There are a bunch associated with linear algebra matrix operations, which is what brought one a whole bunch of technology. Those algorithms brought one computer graphics. They brought one a bunch of things to do with sound encoding and so on, and they now brought one for better or worse, because I don't actually think those algorithms are quite as necessary as one might have expected. They brought us neural nets and the current episode of machine learning. But there are many other algorithms that are much more behind the scenes. For example, one algorithm, I think it's probably the most used algorithm in history is linear feedback shift registers. So technically what that is is you have this, this row of basically cells that have ones and zeros in them. And a shift register is this sequence of ones and zeros in this register. And at every step it's shifting. So it Just shifts one down and one of the bits sort of falls off the end. Okay, a feedback shift register is something where you take bits in the register and when you have to feed, fill in more bits to sort of feed the register, you're filling in those new bits from bits you already had, and you fill them in, usually by taking the xor the sum mod two of those bits. And you have certain taps on the shift register that say, pick element 10 and element 14 and element 27 on the shift register. And the shift register is of length 37, let's say. And at every step, pick those from those taps, add the values 1 or 0 mod 2, and then feed those back as the new value and then shift it again. Okay, sounds kind of simple. But when you do that, there are certain combinations of taps and lengths of shift registers and so on. And with those combinations, that shift register is a maximal length linear feedback shift register. So here's kind of the deal. If you have n elements, they're each either 1 or 0. There are two to the n states of the shift register. And turns out you can get shift registers where they don't cycle until you've gone two to the n minus one steps. Okay, so what you might say, well, let me give you examples where that's used. Turns out a huge number of signals all over the place are encoded with linear feedback shift registers. So, for example, GPS satellite signals are encoded that way. Why? Because the GPS satellites produce a shift register sequence that doesn't repeat for, I don't remember, what, 11 days or something. I forget the exact time, but it takes. Actually there's even one that goes hundreds of days, I think, before it repeats. Why is that relevant? It's relevant because when you try to figure out where you are, the question you have to ask yourself is, well, at what precise time was the signal sent from the GPS satellite? And what you want to know is sort of I get the signal, where is it in the sequence? Well, it turns out, in these linear feedback shift register sequences, because they don't repeat for such a long time. And this happens to be a pretty easy algorithm for knowing, given this sequence of bits, where does it occur in that shift register sequence? Is it at the know million billionth bit plus 17 or something? Yes, you can tell that. And so you can get that. You can use that to get timing information from gps. So that's one use. Another quite different use is in at least up to 4G telephony for cell phones, you get these shift register sequences. They're sometimes Called pseudo noise sequences because they go, you know, bit, bit, bit, the, the ones and zeros and so on. You get these sequences where they kind of look random if you look at them on a small scale. And that's used to set it up so that you can have many channels of communication and each channel is encoded by a different shift register sequence. And if you just look at them from in sort of large scale, they just got kind of random. Unless your phone happens to know, oh, it's shift Register sequence number 173. And when it uses that particular one to do the decoding, then it's like, oh, I know what's being said here. But the other ones kind of look random. And that's a way of. It happens to be a way of packing sort of more data into a limited frequency range. So that's another use. Then pretty much every display that the data that is going from the computer to the display is encoded with the feedback shift register sequence. Why? Because it turns out that when you do this kind of pseudo noise type thing, one of the things that happens is you have about as many, you have about the same number of zeros and ones. Well, if you are sending data and you've got a ton of zeros and very few ones, it's just electrically much more difficult to deal with that because you have to kind of, you can't sort of balance voltages and things in the same way is if you have a sequence where it's kind of about the same numbers of zeros and ones. It's just the detailed pattern that determines whether you have a picture of an elephant or a pig or something on your computer screen. And so it's very typical for the data, for these sort of electrical reasons for the data to be encoded with a feedback shift register sequence. So that it has about the same number of zeros and ones. Actually, I did work out at one point that the feedback shift register sequence algorithm has been used 1,10illion times roughly in the history of technology. And it's also used for data and memory and things like that. So that's a very behind the scenes kind of algorithm. There are also algorithms that are used, for example, algorithms for addition, multiplication, those kinds of things. There are some algorithms, that type of thing. Now there are other sort of algorithms that are important. There are algorithms like graph routing algorithms. How do you find, you know, if you're going to run a big social network, you need to be able to sort of find who's connected to who. How do you do that routing That's a very similar problem to the problem of given the network of streets in a city, if I want to go from here to there, how should I find a route that goes from here to there? There are a collection of algorithms for doing pathfinding in graphs, which are also fairly important. Then there are, well, then there are the machine learning algorithms where kind of the meta statement is if you kind of bash the thing hard enough, it will learn stuff. There are optimizations to that which I actually think in the end will not turn out to be all that important. They're in practice important right now because of the details of how machine learning is done. My guess is that in the long run they won't actually be terribly important. Now, okay, what other algorithms? They're encryption algorithms. That's another type of algorithm where you're. Of which, well, famous ones are things like the RSA encryption algorithm that uses the fact that for making public key cryptosystems, where it's using the fact that it's harder to factor a number into primes than it is to multiply a bunch of primes together to get a number. And that asymmetry is crucial in kind of letting you sort of set up things where you can say, I have this digital signature, I can prove to you that this is true. But you can't generate one of these. You can check it, but you can't generate one, and so on. And that's used all over the place. Now the actual that algorithm is kind of a crisp algorithm. The ultimate algorithms that are used for large scale encryption do not tend to be nearly as crisp. The algorithms kind of AES. Oh gosh, what's the number? 256 maybe I forget. These algorithms that are used for kind of large scale stream encryption don't tend to be nearly as clean. They're not sort of easy to describe algorithms. They're these things where, oh, you've got these layers of substitutions and permutations and so on. They're actually very similar to the Enigma machines that were used for encryption by the Germans in World War II, but scaled up a lot. And those tend to be sort of very messy kind of pseudo algorithms, so to speak. I think another algorithm that's pretty important is hashing algorithms which are similar to encryption algorithms. But you know, hashing is. You kind of, you want to know, have I ever seen this word before? Well, you've got a big list of words, but you don't want to just compare all the words. So you say, given every word, I'M going to hash it up and just make a number out of that word. And let's say I make a number between one and a million or something, I can. One in a thousand. Let's say for every word, and every word that comes in, the word fish will be 11, you know, 112. The word elephant will be 75. Well, if I have a new word and I hash that new word, and the hash is something where I've never seen that hash value before, then I know it's a new word. If it's 75, it might be the word elephant. You'd have to go check that explicitly. So there are a bunch of algorithms, particularly ones that are cryptographically secure or supposedly. So at least where it's the case that you can't kind of, you can't make up a thing that will hash to that value, it's difficult to make up a thing that will hash to that value. In other words, if you say I want something that's going to/to75 and it's not elephant, make me up a word that hashes to 75. There are algorithms where it's difficult to do that. That's the basis, for example, for mining for proof of work in Bitcoin, for instance, is using that kind of hashing algorithm. And in that particular case it's, can you find me a hash that starts with a bunch of zeros? And the number of zeros required has sort of been going up every 18 months or so, I think. And, and that's kind of a, a. So you're using the hash, you, you're saying that the, the sort of, the mining process is find me something. Break the sort of cryptographic hashing characteristic, find me something which hashes to this particular value. This is an interesting question, but I think I have to go do my day job here, so I think I have to run along. But thanks very much for joining me and see you another time. And I might have a chance to talk more about this algorithms question, because it's an interesting question. Thanks a lot. See you another time.