Future of Science and Technology Q&A (March 22, 2024)

Hello everyone. Welcome to another episode of Q and A about future of science and technology. I'm away from my natural habitat, so occasionally that will mean that things go wrong, but we'll hope for the best. All right, let's see. All kinds of questions here. Question here from Dennis. Do you have plans? When can we expect to see an S wall from AI chatbot with voice? I know there's an experimental one that someone has made called I think it's stevenbot.com I have to say I have not yet had a chance to use it much myself. It's a little bit. I, I am a pretty easy case to make a bot from because I've been doing all these things like live streams. So there are all kinds of examples of text I've generated, so to speak. I think there's about 50 million words of text that I've generated that's out on the web. I also have been a long time keeper of personal data, personal analytics data and so on. So I have basically 30 something years of email back and forth, 3 million or so email messages. I also have, well, every keystroke. I've typed lots of screen images from my computer screens. I've been working on things, just a huge amount of detail about kind of things I produced and my responses to different external stimuli, so to speak. So I think I might even be a uniquely easy person to make a bot from. And really one of the questions that I have is if one assumes that there's 100 billion neurons in this brain, so to speak, and they have a certain number of connections and so on, and the question is, could one determine what essentially all of those connections are by just using, let's say, even just 50 million words of training data? I think it's going to be close. I think it's going to be, you know, not too far off. And if you include kind of the as a foundation model, not that necessarily this is right, but it is to some extent kind of something that's been trained from general things that are out there on the web and so on. And then you add kind of the me me special, 50 million words or more, then it's that I think will begin to be a potentially pretty good approximation to kind of the sort of things going on in my brain now. Now, a current kind of LLM architecture is not the whole story of how brains work, but it's a first approximation at least. And I think then sort of it becomes an interesting question, you know, how does one feel about the bot of oneself? So to Speak. And how does one use the bot of oneself? I mean, for instance, one can imagine a situation where, you know, I get a lot of email asking me all kinds of questions. Quite often we say, hey, ask that on a live stream. So thanks to folks here who come from that, that source. The thing that seems like a reasonable use case for the bot is give an immediate answer from the bot. Saying this is from a bot. Does that actually answer your question? Because hopefully the bot, if it's well set up, can point to things I've written or things I've said or whatever. And in a sense it's like people want to know what do I think about X. It's like, well, I may have already said that somewhere and just point them to it. I wouldn't necessarily know how to point them to that because I wouldn't necessarily remember where I said this or that. But that's something that hopefully the bot equipped with appropriate vector database and all that kind of thing might be able to do. So that's a few thoughts on that. I mean, I suppose the question is how can I use my bot myself to sort of lead the better life? And I think there are definitely some things to say about that. I mean, for example, the fact that I have a search system, just a textual search system for all of my email archive, file archive, et cetera, et cetera, et cetera, is very useful to me. It's like, hey, where did I see that name before? Those kinds of things I can immediately answer from my text search system. What I can't answer so easily is when did I have an idea like this before? And if the sort of vector database and the kind of concept embedding system works well enough, there's a good chance I could answer that. And even when I'm writing something and I say, wait a minute, I want to make a link to something I did before, or I'm reminding myself, is there a thing that I've said that was about this? Where I already thought about that before. I'm kind of thinking that I may end up with something where there's a, a continuously running kind of sort of embedding vector database lookup thing that's running as I write things, that's sort of reminding me, wait a minute, you talked about something like this before. Here's a thing you can either point to or why are you writing this? Again, you already wrote this. And that's like, oh, can I come up with an example of this or that? That's a thing where potentially from sort of the, the a more conceptual search is useful. So those are thoughts about that. There's sort of a question of the what would my bot do? So to speak. There's, you know, I might say, well, this is what I would do, but maybe my bot has a broader view of the possibilities or at least a different view of the possibilities. It's kind of like, you can imagine, kind of. I'm not a big one for decisions made by committee, but I don't know what I would feel about decisions made by a committee of me's might be disastrous. But imagine that you have sort of five instances of your bot and they all have slightly different details. And it's like, hey, we got to make this decision. What would me human me plus five AI me's, What would we all come up with? And maybe we all come up with slightly different versions. And maybe that's a helpful thing, maybe sort of reflecting one's views. I mean, it's certainly a thing I always find useful. Talking to other people is the way in which, even if in the end I'm going to figure something out, the way in which kind of what I'm thinking about reflects off how they understand what I'm talking about and so on. And I suppose the extreme case of that is, well, if you've got accurate AI bots of yourself, how do you make decisions when you're sort of talking to yourself, but talking really to your bots? How does that work? And do you. I don't know how it will feel to interact with one's bot. Will one feel like very sort of happy that, gosh, I don't have to do all the heavy lifting myself, I can rely on my bot to do some of this. Or is it. Do you end up feeling like you're going to up your game because you want to kind of be like, let me show the bottom. The human is still, you know, sort of out ahead. I can figure out things more creatively better whatever than my bot. I don't know. I don't. It probably will depend on one's personality. What, what, how one thinks about that. I mean, for me, I'm always. If I can delegate something to someone or something else, then that's what I prefer to do because I always want to be thinking about things that I feel like I'm uniquely able to do. Now when there's a bot of me that's, that's good that that kind of pushes me perhaps in a certain direction of things that I might choose to do that are sort of not accessible to my bot. Let's see. Yeah. Lola is asking, do you think AI iterations of a person will be similar at the start and then diverge in personality or intelligence as they continue to live and develop? I think it depends a bit on what the experience of the bot is. I mean, bots can have very different. I mean, our experience walking around the world and doing what we do and so on, and sort of leading human lives, sort of going to sleep, waking up, eating breakfast, these kinds of things, those will be different experiences from the experiences of an AI. And just as we have a certain sort of set of sensory inputs, eyes and ears and so on, the bot can potentially have different inputs. Could be all sorts of, you know, IoT devices that are measuring all kinds of things. A big sort of. In a sense, its sensory interface could be very different than ours. I mean, our sensory interface, you know, we have touch where we have lots of nerve endings kind of in our skin and so on. One can imagine that the AI has sort of nerve endings that end in all those temperature sensor devices and light sensors and motion sensors and so on all over the place that give it a very different and less spatially localized experience of the world and how that will feel. I mean, in other words, the. The bot could sort of get an intuition, get an experience for different aspects of the world than the ones that we typically do. And in terms of whether, once you've kind of set up the bot, if the bot doesn't get. If the bot is getting a feed of, well, all the things that I'm saying and typing and so on, after the bot was initially set up, there's no reason that the bot can't update itself with, with, oh, the, my, I don't know what the bot. How a bot should refer to its originator. It's not quite my boss. It's, you know, my originator. Let's say my original is. Oh, my original just decided to go off in that direction and has gotten really interested in this. So I, as a bot of that original, would then go on and learn from the direction that the original has taken, so to speak. And then, of course, the interesting question is, does the bot take its direction? That becomes sort of an autosuggestion for the original, who then goes off and pursues that direction. That was bot suggested. And I think that's, you know, that's something that will surely, surely come. And then the original, so to speak, is making the decision, yes, I'm going to take that suggestion from the bot and not that other one. Let's see. Peter asks, are you comfortable with the average quality and correctness of AI generated answers and commentary? No, of course not. I mean, it's, it's terrible. It's kind of, it's, it's like what you get from people who don't know what they're talking about. And it's very hard to tell because it's very convincing. You know, it's like I'm just going to systematically say this or that. But of course, what's going on in, you know, I wrote this whole book a year ago about, about what ChatGPT is doing and, and why does it work? And, you know, there's, there's the sort of, the whole story of what is ChatGPT doing? Well, it's been sort of trained from what's on the web and it's managed to extrapolate from what's on the web to sort of form language in a way that's similar to the way that we form language, but it's just forming language. It's not doing things like computing answers or being able to systematically, in and of itself, as far as the LLM is concerned, sort of systematically get access to curated knowledge and so on. I mean, this is a large part of what our technology stack has contributed to. All of this is between Wolfram Alpha, which has natural language input, and Wolfram Language, which has computational language input. Being able to have those systems be connected as tools to LLMs is a very powerful combination. It's a powerful combination both for the AI, being able to kind of use that tool to know what's true, so to speak, and then be able to weave that into the language it generates. It's also important for humans because. And you'll see something that we have coming out soon actually, that really highlights this, the ability to have sort of. You're talking to the AI as a linguistic interface and it's doing what it can do, but it's ultimately writing computational language and it writes this computational language and you as a human can read that computational language. That's sort of the whole point of what has been done with Wolfram language is not only it's sort of a computational representation of the world, but it's one that both can be made use of by a computer and can be made use of and read and thought about by humans. So the kind of the workflow is you talk to the AI, it generates kind of Wolfram language code, and then you can read that, and then it will make a sort of Wolfram box of Wolfram language code, and then it will run that code and generate output and what's inside that box from the input from Wolfram language input to the output, you can trust that that's solid computational knowledge. But whether it's up to you to then say, did the linguistic interface that's been provided by the AI, did it successfully generate the thing I thought it should generate, or did it go off and do something kind of crazy? But once you can see what it generated, you know that what's within that box of orphan language input and output, that that's really a valid thing and you can take what's in that box and use it as kind of a solid brick to build potentially a larger system or whatever else. So I think that's a really nice way to kind of get the best of both the linguistic interface that LLMs provide and kind of the computational knowledge that we've been building for the last 40 years or so. And it's worth realizing also that when it comes to that sort of computation and computational knowledge more, much of what is done there is very non human. It's like, okay, go actually compute this, do this actual computation. Humans don't run computations in their brains. Humans don't run systematic formalized computations in their brains. Humans sort of do things in a sense off the top of their head and with great effort they can kind of build little towers of functionality, so to speak. So the kind of thing that is what you get by sort of combining the formalized computation, which is kind of what in human civilization we've kind of developed with science and mathematics and now computation over the last few hundred years, that's kind of a separate branch from the what we do just thinking about things off the top of our head, which is what the kind of neural net, LLM, AI kind of direction specializes in. Let's see. Alistair asks, due to the success of nature inspired computing, I'm really wondering if our best bet is creating full on human rat bookers, meaning similar learning experiences and processes. You know, there's the big question of what's that computation for? You can have a computational system that goes off and does all kinds of computations. I've spent some large part of my life studying the basic science of what happens out there in the computational universe. What do all these different possible computations do? And they do all kinds of things that look interesting to us, but often those things are not well connected to the kinds of things that we normally think about. They're just computation in the wild, so to speak. Then there's computation that's aligned with the way we think about things. Computation that relates to human language, computation that relates to the mathematics we've invented or the science we've invented, and so on. And that sort of human aligned computation is something that seems much more relevant to us. I mean, in a sense, the whole sort of copy concept of the Wolfram language is to provide this way of taking sort of what's raw out there in the computational universe and selecting parts of it that are relevant to us humans and then being able to sort of provide access, provide a bridge from the way we humans think about things to what is possible in the computational universe. And in a sense, just like in human language, we pick particular words, we pick our 40, 50,000 words or something in typical language, like English, for the things that we care about, talking about. There are lots of other things in the world for which we don't have words that are just sort of happening in the natural world or something, but we never thought it was important enough to give it a word. Well, similarly in computational language, we got the same kind of thing. There's all that's out there in the computational universe and there's things that we choose to kind of give words to in our computational language, which in the case of a computational language are actually functions in the language that can actually do something and compute something automatically. In human language, it's just that's a word and you can use it to communicate with another human. In computational language, the word, the function that you have is something that automatically is associated with a computation that can then be made to happen and will happen automatically, so to speak. But I think the, so there's this kind of alignment of what's computationally possible with what we humans care about. And then there's kind of a, we're getting into sort of more detail of, well, you know, how aligned is it with what we humans kind of think about and talk about and experience and so on. And the thing is that our experience of the world as humans is very much based on the fact that our eyes are about 5ft off the ground and we walk around and we can pick things up with our hands and we can eat things and we can do this and that. There's a lot of detail about the human condition that relates to the way that we interact with the world. And if we really want to get AIs that are closely aligned with what we care about, then that is going to be important in kind of having an AI that has human aligned experiences. So yes, a sort of humanoid robot that even, even the reaction of people to a, a robot is affected by whether it's humanoid. A friend of mine happens to have built a rather slightly bizarre kind of humanoid robot that is actually a copy of him of himself and had the experience recently dropped in on me with his robot and for various reasons we ended up meeting at a, at a restaurant and it was not quite planned this way. But anyway, the bot was sitting there, it's not a walking bot, it sits in a wheelchair for now sitting there at this restaurant table. And it was really kind of an interesting experience to see people coming up to this bot asking what it was. The bot is connected to an LLM so it can have a conversation. It has rather snarky attitude, but it's sort of interesting and it's interesting to see people interact with kind of a humanoid form factor thing. And it's really quite striking the, the extent to which people really do treat it in, in a human kind of way. So even, and you know, when they're about to go away, they say goodbye to it. And even though it's just, you know, from some point of view, it's just a bag of electronics. And so I think, you know, for the bot that is going out and about in the world, the more humanoid it is, the more people are going to, even people are going to react to it and a human kind of way. So it's going to end up with a different experience of the world than something which doesn't have that kind of humanoid aspect. I think also that even kind of the way in which the thing can go around the world depend if it has legs and it walks like a human and it's sort of the same size as a human and so on, it's going to be able to walk through those doors and things like this that, you know, it's going to be able to sort of take the elevator, walk upstairs, whatever else it is. You know, our built world has been set up to be convenient for us bipedal humans and so on. It's, you know, the natural world might not be terribly convenient. The natural world might have, you know, slippery ice or, you know, sheer rock faces or whatever else. But the built world we have built to be convenient for humans. And that same building out to be convenient for humans would also make it convenient for humanoid robots. And that's kind of a sort of. To have a robot be a humanoid robot is sort of a compatibility layer for dealing with the world as it Is, and it's also probably a compatibility layer in terms of dealing with us humans. Because as I say, people's reaction to a humanoid form factor robot is really quite different, I think, from their reaction to just typing things on a screen or something like this. Sometimes it will be better, sometimes it will not be better. There are probably cases in which if you're trying to fill out some medical questionnaire or something, it may very well be the case that people will do a better job filling that out. Typing to a computer that's kind of not a judgmental kind of thing than sort of saying it to a humanoid robot where they start to feel like it's kind of like the human is judging them for having eaten too much chocolate or whatever it is that they're reporting, so to speak. So, you know, I think this, if we want human aligned AI, the more human the kind of presentation of the AI is, the easier that will be. Now, it may also be that we as humans and this already happened, we as humans are used to a different experience of the world already than the one that is sort of the pure human experience. I mean, we're used to the fact that we have our smartphones with us. We're used to the fact that all kinds of things come up in feeds, in windows, in whatever else, and even in text for that matter. And so that's a place, if we went back a thousand years in history or something to the point where not that many people were literate and certainly there weren't computers around and so on, it might be the case that we would need that we'd have even more force to have sort of the humanoid like way of presenting AI in order for it to be something that can sort of be adequately aligned with humans. But we humans are getting less aligned with kind of raw human experience and more aligned with a kind of computational sort of computer system experience. So I think it perhaps meets somewhat in the middle in terms of what we get used to versus sort of how the AIs kind of become more like us. Let's see, Alistair is commenting. Likes the architecture of having many bots with sort of a base that confer and upload their findings to a global knowledge base and then disperse on new assignments. So I think this let me send out my bots. It's an interesting concept. There is a certain sense in which let me get my people to look into that type thing. Well, let me get my bots to look into that. And certainly in terms of things like data collection and so on and even Going out. And I mean, I have to say, as somebody who this day job is CEOing a tech company. It's. There's a certain, you know, one is a tech company is sort of a machine whose component parts in some measure are people, in some measure they're automated systems. But people are definitely an important component of that. And it certainly is the case that as you say, well, let's investigate this thing that turns into this person will do this, this person will do that, and to some extent, let's investigate this thing. Sometimes when I'm doing particularly science kinds of things, let's investigate this thing is more let this computer run and do that. Let this computer run and do that. And sort of emerging of those things is send out the bots, so to speak, and have them investigate things. Let's see, There's a. See, the secret asks, do you think human brains compress data in a lossy way? And will future AI brains also have to use lossy compression to be more human like, or would a perfect AI memory be more desirable? Well, I think the thing to think about is, okay, there's the picture of the ideal cat. And you can match pixel by pixel the picture of the ideal cat. But most of the time what you're interested in is, is that a picture of a cat, even if it's not the ideal cat, pixel by pixel, you're asking for this kind of fuzzy matching of do I take this picture that we would say, oh yeah, that's kind of a cat, but, you know, the particular hair or its ears is not exactly the same as the standard cat, so to speak. So a large part of what we as humans experience, the way we kind of operate in the world, is this way of sort of taking the details of the world and compressing it to the concepts that we deal with. And it's kind of a feature of kind of, I think, consciousness, sort of human experience, et cetera, that we're continually taking the complexity of the world and trying to prune it down, compress it in effect to things that are concepts that we can think in terms of, talk about, reason with and so on. And I think that's kind of the story of, of kind of that that's how we human like AIs are going to have to do the same thing. I mean to say, yes, we've got a picture of a cat. It's all the pixels are there. We can flip through all the pictures of a cat. Yes, the. That's a thing we can do, but it's a Different question to ask, sort of to say something about how we relate to the concept of a cat or something from all these pictures. The details of the pictures are one thing, and we obviously have computer systems right now that do a good job of sort of storing verbatim all of this data. And the thing that is complicated is to make the connection between the thing you're seeing now and the. The thing that was somewhat similar, but not pixel by pixel identical to the thing you saw before. So I kind of think that that's a necessary feature of kind of human like AI that you're dealing with. Sort of reduce the complexity of the world to this much smaller set of kind of compressed concepts that you can use to talk about things. Let's see, very different question here from without hippocampus, that wouldn't work. Well, could an element printer theoretically work, eg, one where you feed it carbon atoms and it prints out an arbitrary element? Well, not for that. I mean, elements are distinguished by how many protons they have in their nuclei. And the only way you get to convert lead to gold or something is to. Is to change the configuration of the nucleus of the atom. And the only way you do that is with nuclear processes and the very high energies that are needed to kind of knock particles out of the nucleus of the atom and put different particles in there. So you don't get to kind of the elements that we have. You don't get to transmute one element to another, except with nuclear reactions and so on. So mostly what we have on the Earth is what originated at the time when nuclear processes were going on. And most of the nuclear processes that we deal with are nuclear processes that happened inside a star, because stars are being powered by nuclear processes going on. And so most of the elements we have were the result of sort of being produced in a star. So, for example, all the gold on Earth is something that was processed by typically a supernova from an earlier star that kind of lived and died a billion years ago or something. And the material from that star wound up condensing into what eventually formed the Earth. And now we get to mine the gold and things like this. And what typically happens is kind of all these different elements, there's a whole distribution of what elements get formed in by through stars. There are elements that were formed in the early universe, although that's predominantly the light elements. Elements, particularly below, there's sort of a distinguishing point at iron, actually iron 56, that that isotope is the isotope that is the most tightly bound isotope. And so it's the one where kind of stars can produce hydrogen to helium, to lithium, et cetera, et cetera, et cetera, down to iron. And if you kind of want to go beyond that, you have to have typically some quite violent process, like a supernova, that kind of just keeps on bashing things with nuclear reactions to just sort of add protons to things. So you don't get to. With the kinds of things that we do kind of without nuclear reactions, you don't get to change elements, so to speak. You're kind of stuck with the elements that have been provided by sort of in the history of our solar system and so on. And so the kinds of things we can do, for example, chemical processes are mostly chemical processes just rearranging atoms and the electrons that go with those atoms, kind of rearranging the configuration of those electrons, rearranging which atom goes where, which atom is bound to which other atom through things related to electrons and so on. And that's kind of the story of chemistry. Chemistry operates at much lower energy scales than kind of a million times lower energy scales than the energy scales associated with nuclear reactions. You kind of see why that's true because in an atom, for example, the nucleus is like 100,000 times smaller than the size of the atom as a whole. And that kind of translates into, if you want to mess around with the nucleus, you're using 100,000 million times as much energy as you would to sort of mess around with the electron configuration in the atom. So you don't get to change the element. What you do get to do is to change how the atoms are configured, how they're connected to each other. And sort of the story of chemistry is the story of knowing particular reactions that happen where it is energetically favorable to rearrange these atoms so that those atoms get bound to that atom and so on, rather than, for example, let's take water, H2O, for instance. By putting energy into it, you can break apart the hydrogen and the oxygen. You can rearrange that. You can get some hydrogen, some oxygen separate. And then if you put those together, there'll be a bit of some energy released as those things lock together and make water. But they're much more complicated chemical reactions. And the whole story mostly happened, like, 150 years ago. Just a huge number of chemical reactions were discovered. I mean, it had been started in the Middle Ages and before as kind of the alchemists started sort of heating things up and seeing what happened and so on. And when you heat things up, it's Kind of like getting the atoms, the atoms are happily stuck in a single molecule. You heat it up, the molecule might come apart and then the atoms will reform themselves in different configurations, different molecules. So the typical sort of structure of chemistry is you have certain reactions that happen as a result of putting different materials together, heating things up and so on. And those reactions, there's a sort of a lexicon of reactions that can happen that are sort of the fundamental reactions. And then you're kind of mixing and matching those to say, okay, can we make this particular molecule? Well, we can do that by doing this sequence of 20 steps. That's kind of 20 to 30 steps kind of limit to what people tend to do these days of first do this, combine it with this, now add this, now heat that, et cetera, et cetera, et cetera. So one piece of that, that tends to be sort of an extra piece is there's like, okay, the molecules just stick together. They're just sort of bouncing around in a gas or in a liquid. But sometimes there are also catalysts, usually solid surfaces where the molecules, molecules kind of get stuck on a solid surface that orients the molecule in a particular way so that when the next molecule comes in, it's more likely to actually connect in the way that you want in this reaction. And when it comes to biology, biology has proteins and particularly enzymes that are specifically set up to kind of have a spatial configuration that sticks other molecules in the right orientation so that the molecules they're supposed to interact with can fit in correctly and so on. So a big question is, you're trying to make a molecule. Let's say you want to make a molecule that I don't know spells your name in carbon atoms. How do you do that? Is there a printer that will do that or is there a sequence of chemical reactions that will do that? Now there are essentially printers that can move single atoms around. Atomic force microscopes, for example, can pick up single atoms and put them places. It's a very slow and painstaking process, but in principle you can do that. They tends to be on a solid substrate. You're sticking the atoms in a place where they'll just lock into the crystal that's underneath. But usually in current sort of technology, with chemistry, for example, if we want to make a particular molecule, you have to invent the series of moves based on that lexicon of chemical reactions, often so called name reactions that are, you know, that together will successfully make the molecule. We want to make sort of a very indirect process. But that's been the story of chemistry. Now the question is, can you get sort of nanotechnology, molecular manufacturing, whatever you want to call it, where you just type into a computer, I want this molecule, make me this molecule, and kind of little tiny machines swing into action and make that molecule? Well, we have one example where that happens, which is biology, because in biology we start off from just this program that is specified by the sequence of base pairs in DNA and that is transcribed ultimately through RNA and so on and through ribosomes into the sequence of amino acids that form proteins. And one of the things that's a non trivial fact is that you can have. So a protein consists of a sequence of any one of 20 or so Amino acids where you'll get this amino acid followed by this one, followed by this one. It's sort of a programmed molecule all in one chain. But then proteins fold up into all kinds of different shapes. And there seems to be a kind of universality in those shapes. If you say, I want this particular shape, there's a decent chance, we don't really understand this terribly well, but there's a decent chance there's a protein which will roughly do that. The immune system makes use of that idea because the immune system, random antigen comes in, random nasty bug or whatever else comes in to you, and your immune system is going and trying to select among 10 billion, a trillion, whatever possible randomly chosen proteins that will sort of mechanically fit with that antigen, will be kind of the right shape to bind to that antigen to fit into that antigen. And then having done that, well, for example, antibodies can mark it, or T cells can get ready to ingest it, or whatever else. But the point is that there's a certain shape universality potentially to little pieces of protein, which is what happens, which what exists in things like antibodies, where that shape universality is what's used to be able to match any antigen that comes in somewhere there you can generate antibodies that will be sort of complementary shapes that will be able to bind to it. It doesn't work perfectly. And I don't think we really have a good characterization of what that space of shapes is really like. But in any case, that's a place where you want to make a molecular thing of any shape. Roughly, you can do it with proteins, but let's say you really want to make sort of atom by atom, something that has a particular form. By the way, even with proteins, proteins have particular elements in them. You know, carbon, hydrogen, nitrogen, oxygen, sulfur, maybe also phosphorus, but a limited set of elements. If you want to have a piece of iron in your body, or a piece of iodine or a piece of molybdenum, then the way you get that is you get a protein that has kind of a cage, and that cage is made from amino acids and from the folding of the protein. And that cage ends up being sort of just the right size to fit one iron atom, for example. And so hemoglobin does that in our blood, that somewhere it's just, you know, it's made this thing with a cage and there's enough kind of iron lying around that that cage will always be filled with an iron atom. And so that's sort of a way of using even just proteins to sort of arrange things to deal with elements other than the ones that are intrinsically part of the collection of amino acids. But sort of the question is, is there a way to do sort of molecular scale manufacturing where you just make any molecule, any configuration of atoms? Now, not every configuration, atoms is stable. It could be the case you make this thing that's trying to spell your name and your letter A or something will always squash together, because that's just how the molecular, the atomic forces work. But let's assume that even with carbon atoms, there's a huge range of possible shapes you could imagine making. And so then the question is, how would you actually make those shapes? Do you have to do it in this very indirect way? You're kind of poking at it by using these different chemical reactions, where the chemical reaction says, you're going to do this to all the molecules in this, you know, test tube of liquid or something, you're going to do the same thing to all of them, and it's going to be this. And once you've done that to all these molecules, even though they're in different random configurations, they'll all come out to be the shape you want. But the alternative is to imagine some kind of nanomachine that can just go and assemble these molecules. Just like, in a sense, ribosomes do that in sort of putting successive steps of successive amino acids onto a protein. So the thing that has been sort of a. It was something that was quite popular in the 1980s, early 1980s, to kind of explore this idea of sort of molecular manufacturing and nanotechnology. Unfortunately, it pretty much died as a thing that people looked at, and not for particularly good reason. It died more because of kind of the politics of science and so on, than because it wasn't a good idea. But there was a certain tendency to say, let's take machinery with cogs and whatever else that we know operates on a macroscopic scale. And let's figure out how to make that on a very microscopic scale. The with by just sort of reducing the cog to a collection of a small number of atoms or something. And my own feeling is that that's probably not the right way to do it. That if one wants to have the effect of arranging atoms in particular configurations or something that, or being able to programmably arrange atoms, that given that objective, one should sort of search the molecular universe, so to speak, for components that let one do that. And they will often be things that we wouldn't have built by sort of imagining, well, what would it be like if it was a big thing a meter across? Then let's shrink it down. But rather they're things that are more specifically suited to molecular scale structures. But you know, the hope would be one day that you could just sort of type in. Well, to say this, I mean, when it comes to making a protein, you can just type in, I want this sequence of amino acids. Just make me this. But that's initially a one dimensional chain of things. Let's say I wanted to use kind of a CAD system and I wanted to just design my molecule. People thought this might be possible in the late 1980s, early 1990s, kind of designer drug construction or designer molecules and so on. You can just specify them in some CAD system and then press print and you get that molecule out. We don't know how to do that except for proteins. We're kind of leveraging what we've learned from biology, so to speak, and we're just using the same mechanisms as biology. But I don't think there's anything that says it's not possible to invent a system that will sort of print any molecule or any molecule that is reasonably stable. And, you know, what effect will that have if one can print anything? Well, it's pretty interesting because right now, you know, you have a 3D printer and you can say, well, it's making plastic, it's making metal, you can make certain kinds of parts, but they have to be a uniform material, more or less. I mean, maybe in the printer you can have, you know, the red plastic and the, and the yellow plastic or something, and you can squirt out different ones, or you can have red, green, blue, and you can kind of squirt out different ones to get different colors and so on. And conceivably you can also have different materials, although that's more complicated. But the question is, could you get something where you can just sort of print sort of any configuration of atoms. And what you would have in your printer is you'd basically have all the chemical elements just like we might have in a color printer. We might have kind of cyan, magenta, yellow and black, reservoirs of toner cartridges. So in the thing printer of the future, you would have kind of a supply of hydrogen, helium, lithium, beryllium, boron, et cetera, all the chemical elements. And you'd be saying, okay, I'm going to put a little piece of molybdenum here and I'm going to combine it with the sort of iron here and put this here. And you'll be just using the atoms and from your supply, sort of vials of supply material, so to speak, which, by the way, is what happens when people are synthesizing short pieces of DNA and so on. There literally are little, you know, test tubes of the different bases that have to go in there. So I think it's really, it's interesting to try and think through if we could print anything right down to the molecular scale, what would we print? Well, we could certainly print a lot of sort of biological type things. You could start, you know, printing organs now. Sometimes organs like, I don't know, lungs, kidneys, things like that. Their operation is in a sense, very physical. It's just like, you know, okay, let's, let's sort of combine this. Let's make sure that we have a flow of oxygen going to tissues and getting close enough to tissues and so on. That's a very sort of physical, mechanical thing. And one can imagine kind of already that's something people are trying to do to 3D print lungs, so to speak, right down to the sort of 2 micron level that's needed to get sort of gas diffusion and such like. And I think that's so, you know the question. So in that case, sort of you can make a biocompatible thing that's printed sort of just using traditional 3D printing. But one can imagine something where it's like, I want a replacement cell of this kind. Let me just. 3D. Let me just fin print that. Cell cells have a lot of stuff in them. That's not a trivial thing to do, but that's one application. Then we can ask, well, what would you make? It's just like when new materials come in, like plastics, when they were invented, all kinds of things that you couldn't really imagine doing before now became possible because you had a new material. When you have this general ultimate sort of meta Meta material, so to speak, and you can print anything. What will you choose to print? It's a little bit like the problem of sort of the computational universe at a computational or informational level. Given the way that we do things and that we humans are thinking about things, I'm not sure I think it's a good sort of exercise to try and think through. If you could sort of print anything, what would you choose to print? And a lot of things that immediately come to mind are things related to us humans, because sort of if we're the ones in charge, it's like, well, print a replacement part for us humans or something like that. Or it might be, it's not so clear what else we would think to thinkprint, so to speak. Anyway, interesting question. I have to go fairly soon here, but let's see. Brady asks, can you explain Quantum LLMs? What advantage exists in applying multi computation, if any, to LLMs? So, you know the typical large language model, what is it doing? It's saying, I've got this sequence of words and I want to predict what the next word should be. And I'm going to do that by using sort of the statistics of what's on the web and my internal kind of neural net models. But in the end, what comes out is of these 40,000 words, what's the probability that the next word that should go into this essay is this word versus this word versus this word? There's a big stack of possible words and their probabilities. And so the typical LLM operates by saying, I'm going to pick the word that corresponds to, for example, I might pick the most popular word, the word that has the highest probability. Or I might pick words with some randomness according to the probability that they were assigned by the LLM. Or I might pick words favoring the most likely word, but paying some attention to the stack of other possible words. Those different possibilities are usually parameterized by the temperature parameter in the LLM. So temperature zero is always pick the highest probability word, temperature one is kind of use the probabilities that the LLM itself generated and so on. And temperature 0.8 seems to be the one that works well for essays. Typically it has a certain degree of kind of randomness and, and creativity, but it's still not too far away from kind of the sort of just pick the most likely word. But as you increase that temperature above 1.2, 1.3 with typical models, the thing stops making sense. It's kind of a sharp transition where the thing just starts writing nonsense and Then as you increase the temperature more, it starts not even making grammatical sentences and so on, and starts just putting out all kinds of crud that one can see was part of its training data. So the question is a traditional LLM, it's always just looking at one path. It's just saying, I'm going to pick the next word, I'm going to put that word down, and then I'm going to keep going. So sort of one version of kind of the quantum idea would be don't just pick one path, consider many different paths. It's not what we humans do. When we humans think about things, when we speak and so on, we just have one path of history that we're going through. We're not usually thinking in many kind of in a multi way kind of fashion where we're thinking about different, we're thinking about multiple things at the same time. It's not usually the way, at least we feel that we're thinking about things. But imagine that you did that with an LLM, that you allowed it to have multiple paths of thought, so to speak, which will develop very differently once there's been one word different, the next word will have different probabilities and so on. And then what you can imagine doing is building up this whole kind of network of possible histories for the LLM. And then you can do things like you can say, let's do kind of computational things on that network of possible histories. And let's say that the goal of the LLM is to reach the word fish or something. Then let's set it up so that we're looking at these different paths of history and we're trying to always find paths that will lead to the word fish. And we do that as sort of an external computation. I mean, typically the way an LLM works, its outer loop is something very simple. It just says, given the words we've got so far, add the, let's say most probable or whatever next word, then do it again and again and again and again and again. That's kind of the outer loop of the LLM. But one can imagine more sophisticated outer loops where one is looking at multiple possibilities, one's kind of looking ahead, one's trying to plan things, and so on. So that's kind of the thought there. And by the way, if you really want to get exotic, you could imagine sort of running those different possibilities in a quantum computer. The problem is, as is always the case with quantum computers, yes, you can run those parallel threads sort of at the quantum level. But in the end, you need to kind of bring them together to know what is the kind of human. What is the single thread of experience, human version of what wants to come out. All right, maybe one more question. Let me just see here. Gosh. Borsch asks. I'm going to do two questions here. Okay, we're going to do one. JH what would an AI look like that is rewarded based on questioning rather than answering? Interesting. I mean, typically an AI, like a large language model, it's trained by asking the question, how would this text typically continue? How a piece of text, and you mask a piece of that text, and you try and train the AI to successfully reproduce the part of text it didn't see. So now how would you set it up to say, I want the AI to not to go and sort of fill in what has already been there, but instead to have sort of the best. The best kind of, sort of searching out into what has not been there. In other words, how do you sort of reward creativity of the AI? That's kind of interesting and difficult. I mean, you know, one thing you could imagine is say, what are all the things that have been said before? How do you make sure you're saying something different from everything that's said before? I have a feeling the AI might just seem bonkers because it might seem to be the case that, you know, as I say, there are things that are sort of aligned with what we humans think about, and there are things that we humans just don't have any kind of resonance with. They're just things that exist out in the computational world, and they're not things where. Where we humans can kind of, you know, sort of relate to them. And I have a feeling that an AI kind of ask. To ask questions, would ask all kinds of things that we'd say, why are you asking that? That doesn't make any sense. That's. Or, you know, that's. I don't care about that. You know, you're asking, why does the, you know, how many angels fit on the head of a pin? I don't know and I don't care type thing. That's a question that, you know, if you start just asking questions at random, then that's the kind of thing you'll get. Now, you could also imagine it's like in mathematics, you can say, let me make up all kinds of theorems that, yes, I can show a true. But no, nobody really cares about that theorem. It's just some. You just say, you know, okay, it's a theorem that 1 + 2 + 3 + 4 equals 10. The Pythagoreans were very keen on that theorem. Did I get it right? Yes, I think I did. Yes, I did. I think they called it the Tetrachtes, and it was supposed to have lots of significance. But that theorem is true. But most of us just say, who cares? The Pythagoreans didn't. But in today's, with today's kind of scientific approaches, we pretty much say, who cares? And I think that it's sort of the AI, it's sort of an interesting question whether one could sort of teach it enough about what we humans care about to have it pick questions that are questions that we humans are likely to care about. You know, I just wrote about this. I wrote a piece called Can AI Solve Science? A week or so ago, and it has a section about exactly these kinds of questions. All right, two more things. And then we wrap question here from Borsch. What do you think will happen when we understand prime numbers to their fullest? And when can we translate this knowledge to AI? So, you know, primes have been around since the Pythagoreans and so on, since 500 BC or whatever, and we understand lots about prime numbers. There are lots of things we don't know. For example, it was known in antiquity that there are an infinite number of prime numbers. You can ask about twin primes, primes like 11 and 13 that differ just by 2, or 17 and 19 that differ just by 2. And the question of are there an infinite number of twin primes? We still don't know the answer to that question. It's not totally obvious that the axioms that we usually put in place for our mathematics, the sort of statements of things we assume like that X plus Y equals Y plus X or something, that those axioms that we have are sufficient to answer the twin prime question. But any case, putting that aside, it's still the case we don't know for sure that there are an infinite number of twin primes. There probably are. We don't know for sure. And there are a whole bunch of questions in mathematics that sort of revolve around properties of primes we don't quite know. Like, roughly, the number of primes is. The density of primes is essentially one over log n, But it has lots of fluctuations and there's a certain sort of effect, randomness for the distribution of primes, characterizing that, figuring out exactly what the density of primes is, how much wiggling there is, that there is always wiggling and so on. That's complicated. And it relates to this thing called the Riemann hypothesis, which has to do with this function called the Riemann zeta function, which is a function which, if you plot it out, wiggles around a lot. And the Riemann hypothesis is, when it does all its wiggling, will there ever be a minimum of one of the wiggles that is above zero, above the axis? And that's the hypothesis says there won't ever be a wiggle that's above the axis. It will never. That will always dip down below zero and then come back up again. But that's something that's been open since 1859 or so, 1860 maybe. And that's an example of something we don't know about primes. Now, there are, there are. What will be the significance of knowing more about primes? Certainly there's some significance in things like cryptography, where one makes use of properties of numbers. For example, the property that it's easy to multiply numbers together. It's hard to know what the factors of a number, in other words, how the number can be composed out of, for example, primes is. But I think, I mean the, the relationship between sort of prime like things and AI like things is very tenuous. I think. I think the only place where there really could be a connection, and this would be a very interesting one, is in the area of sort of cryptography meets AI. Because usually when you do cryptography, you're taking some message and you're kind of grinding it up to make it seem random. But you have a key, you're grinding it up according to a particular key. And if you have that key or you have some complementary key, you can go back from the ground up to make it look random version back to the original plaintext. That made sense. But one of the issues is if you want to, if you have something where you've got some message, some piece of text, some set of numbers, whatever else, and you can do operations on that in its plain form, but if you grind it up and encrypt it, you can't. The only way you can do operations is to decrypt it first, then do the operations, then encrypt it again. But there's been a long time hope of doing so called fully homomorphic encryption, where one would be able to actually operate on the thing, at least do some operations on the thing when it's in fully encrypted form. And technology is getting closer to making that possible, at least in some simple cases, like saying, here are all these medical records. I just want to know what the aggregate number of people who have, you know, blue eyes or something is. And. But I don't want to have to decrypt these records and see what any detail of any particular person is. I just want to be able to ask this question and find it out sort of going through the encryption. And so an interesting question is whether it's ever going to be possible to kind of do AI through the encryption encryption, in other words, to have it, to have the AI not have to look at the plain text but instead operate in a way that we can't do somehow through have its sort of sensory system kind of probe through the encryption. It seems difficult and it isn't known how to do it yet, but maybe there'll be a way. And that will be an interesting and important thing because it would allow one to kind of deal with sort of data about. Okay. Oh, hello again. Apparently my image camera failed. I said I'm not in my natural habitat and weird things could go wrong. I wasn't expecting it was. It made a rather, at least what I could see. It made a kind of a rather artistic. Turning me into an, into a scatter of pixels. But I don't know quite what was going on there. In any case, I was just going to wrap up by saying that I saw some questions about the eclipse that's coming up in the US across Mexico and the US and maybe a bit of Canada too on April 8th. And eclipses, it's sort of a great triumph of science that has had a long history going through lots of interesting areas of mathematical physics over the centuries being able to predict when precisely the eclipse will arrive. And we have a website, precisioneclipse.com that is doing that computation. And in fact, for the 2017 eclipse I kind of went back and traced through the history of how one learnt how our civilization learned to predict eclipses. And I wrote something then and that's now turned into a little book about predicting the eclipse which you can get if you'd like to. And in fact next week I'm going to, I think it's Thursday of next week. I'm going to be doing kind of a read through and discussion about the history of predicting eclipses and the way that one can now sort of do computational astronomy with eclipses. So do join me then if you'd like and try and tell you the backstory of how we can figure out when the shadow of the moon is going to arrive right where you are, even though the shadow is traveling at 1800 miles an hour or something across the surface of the Earth. It's kind of an interesting story, and it also relates to some sort of foundational scientific questions about things like the three body problem and sort of computational irreducibility of figuring things out. So thanks for joining me here and see you another time. Bye for now.