Future of Science and Technology Q&A (March 8, 2024)

Hello everyone, welcome to another Q and A about future of science and technology. And see we have all kinds of questions that have been sent in. So let's see, where should we start? Let's see. Amorph asks, could mind uploading be achieved in the next 20 to 40 years from now? That's kind of like depending on how old you are, like 20 to 40 years would be the am I going to make it to be to the mind uploading point. Now there's a couple of things to say about this. The first question is, let's say we could figure out what's in a current human brain. Do we have a substrate in which to run that? So there's sort of two questions. Do we have a thing that could be the kind of digital twin of a brain and then can we get from a brain to the data we need for that digital twin, so to speak? And I think in terms of do we have the kind of what we need to run that kind of digital version of a brain? I'm going to guess that we're getting reasonably close. I think we didn't know how close we were getting before ChatGPT came on the scene. We didn't know whether when we look at brains and we can see all the molecular biology of what's happening in the synapses, connections between nerve cells, all those kinds of things. We didn't know how much of the detail of that molecular biology was important for the operation of the brain and how much of it is just that's the way biology managed to set things up. And I think what is becoming clearer is that some version of artificial neural nets which is sufficient to reproduce kind of the essential features of what happens in our brains. Now there are things about what exists now, for example, in the current setup for artificial neural nets, there's sort of two phases. One is you've got the neural net, it's all trained now, you run it and, and you have it generate an essay or something. That's the so called inference phase. But then there's another phase that comes before that where one's dealing with the training of the thing and the training is an incredibly kind of computation intensive activity right now separate from the use of the neural net. I mean it seems like there's sort of the order of magnitude number of operations, arithmetic operations done in the training of a modern big large language model neural net is around 10 to the 25, which is a huge number. And the kind of. So there's this huge effort in training then it's still not tiny amounts of effort to run it, but much less effort. And it seems like our brains have a much more equal match of sort of the effort to train and the effort to run. Maybe that isn't true, but that's the impression one has. And in fact one of the features of our brains is that when we are sort of doing what we do, a decent fraction of the hundred billion neurons in our brain are probably active most of the time. Whereas an LLM once the most of the time, most of the memory that's in the LLM is just sitting there remembering things. It's not being actively run. And so I think that's sort of a hardware feature of the current setup of computers have lots of memory and they have small CPUs or GPUs through which the data in that memory is kind of progressively shoveled. And one can certainly imagine a situation where the memory itself is much more active and there isn't as much distinction between the computation part and the memory part. Just as in our brains there doesn't seem to be such distinction. My guess is that if kind of as is happening a whole bunch hardware is developed that is more specifically sort of keyed into the actual way that neural nets seem to best operate, that that problem will get solved. And my guess is that the sort of hardware necessary to run kind of brain like activities will exist in a modest five year, ish, maybe 10 time frame. That assumes that there isn't some mysterious molecular biology that's happening in actual brains that sort of goes way beyond what's happening in these artificial neural nets. But I think the increasing evidence is that there really isn't. So okay, so that's. So then if we could get all the output from all of the data from a brain in I think a reasonably modest amount of time, we'll be able to run that on digital electronics. So then the question is, well, how would we get all the what we need from a brain to be able to make its digital twin, to be able to reproduce it in digital electronics? I think there are two possibilities. Three perhaps one possibility is you just take the output of a brain and you try and train the twin of the brain to produce the same output. That's kind of what we've done with artificial neural nets. Most of the time we don't go and say we're going to, we don't sort of figure out in advance from scratch, here's what the weight of this particular connection in the neural net is going to be. We just say here are the examples that we want this neural net to reproduce or to learn from, let's just feed it those examples and have it set up the weights so that those examples will be reproduced. So the analog for human brains would be take the output, take all this yakking that somebody like me is doing and take that and train, train the bot, so to speak. Train the twin of the brain from all that externally visible yakking that's going on. And so I don't know how well that's going to work. I know for me, I think there's about 50 million words that I've yacked, so to speak, that are out there in the world. And there's a lot more actually, because I kept sort of all of the, all the things that I've written for the last 30 years. I've kept, I've kept all the keystrokes I've typed and lots of other things like that. So I, for example, for me, there's a huge amount of data on my kind of external output and interaction with the world. And a question is, is that enough to train kind of a. A twin of me? There's a nice little project somebody's done making a thing called Steven Bot that is based on a bunch of, I think not all, but some of the sort of externally exposed material that I have. And it's got an interesting question how much? That's just like me, so to speak. But so, you know, strategy one is there's this kind of the hardware necessary to reproduce the brain is there, and you're just taking the output of a brain and you're trying to make something that would have the same output implemented in that hardware. That's strategy one. Strategy two might be take the brain and if it's dead, just take that brain and slice it really, really thin and be measuring every single synapse and try and work out what was stored in that synapse and try and actually reproduce the neural net, so to speak, from an actual human brain. In digital electronics, there are a bunch of pieces to that we don't know how to do. We don't know how to read out from a connection between two nerve cells. We don't know how to read out what is what was stored at that synapse, assuming that in fact, the way the storage works is directly at the synapse. It's not something to do with microtubules or other kind of details about the way that nerve cells work. And by the way, the point is that you could very well have a good Reproduction of digital electronics with neural nets. Even though we don't know in tremendous detail how, how actual memories are stored in actual brains, it might not be necessary. It's just like you can say, well, I have a computer and it is running all this kind of sophisticated software and oh, I changed the underlying instructions that the computer can run. Oh, but I can still run the same software. That's the story of universal computation that's been known for nearly 100 years now. That is what made software possible. You can have sort of a universal machine, but you can have different universal machines and still run the same software. I suspect the same thing is true of brain like stuff, that the details of the hardware are not critical in achieving the same kind of brain like mind like activities. So, okay, so strategy two is you take the actual human brain and you somehow dissect it and you figure out what memories were stored there and transfer that to digital electronics. Strategy number three would be something where kind of in the living brain, you do something to read out from the living brain. What kinds of things are stored. I mean, you could sort of imagine a very bizarre world in which instead of your phone saying, turn your head this way, that way, the other way so that it can record biometric data, it's instead something where there's some big thing hooked up to your brain. And it's like, think about this, think about that. And it's kind of the system is kind of watching what happens inside the brain as you have those different kinds of thoughts and somehow being able to extract from that the information needed to sort of train to set up the neural net. That's the kind of twin of your brain. I think that in the. It's not clear how you would do this. There have been efforts to use optogenetics that is kind of changing the putting in things like jellyfish genes that can, you know, glow in the dark and so on. Somehow putting that into nerve cells so that you have kind of nerve cells which light up when they're active. Kind of like old fashioned computers, very old fashioned computers these days that had, you know, a big front panel with all kinds of lights on it, where those lights would light up when there was data in some particular location memory location in the computer. And it's not clear whether some future generation of some sort of MRI like thing might be able to achieve the resolution necessary to read out individual nerve cells. But one thing to understand is that in the normal way that one sort of reads out what's happening in the brain from the outside Using EEG electroencephalography, You get a very diffuse signal. You can read out very large scale electrical activity in the brain, like the kind of activity that you have when you're awake versus asleep or epileptic seizures, things like this. Those are very large scale kind of electrical patterns in the brain. Individual nerve cells, the 100 billion of them that we have, when individual ones fire or not. If you've got an electrode that's all the way inside, so to speak, you might be able to pick up an individual nerve firing. But by the time you're dealing with 12 electrodes stuck on your head or just two electrodes that you do fancy signal processing on, there's immense amounts of sort of diffusion in the electrical signal before you get to that. So you really can't say sort of the what are you thinking Type question from that along. There's a technology developed ages ago, magnetic encephalography that uses kind of the magnetic effects that's watching for the magnetic effects of nerve firings and that has less of this kind of diffusion through the scalp type. Type phenomenon. But it requires much more electromagnetically quiet environment or has done. And it's been sort of a continuing story. Are there sort of better and better sensors that you can read more and more about what's happening in the brain from the outside? My guess is usually sort of signal processing gets ever better than you can possibly imagine. I mean I remember back, oh in the, actually by beginning of the 1980s when you were like you'd have a computer terminal at home and you'd sort of hook it up over a phone line and maybe you got 300 bits per second, 300 baud of data. Or maybe later on it was 1200 board of data. And I remember people saying it'll never be possible to get more than I think it was 9,600 baud through a copper wire phone line. It'd just never be possible. It's information theoretically impossible. You just can't get that amount of data through that kind of transmission line. Well, it turned out it wasn't true because it turned out that much fancier. Signal processing got done. Things where you can kind of predict things about what the errors are going to be, you can kind of analyze the error data and so on. And eventually it was routine to put one megabit through those same copper wires. People had said that's never going to be possible. So one of the things that tends to happen and you see it, oh just in sort of all sorts of places, whether it's in the amount of data you can get through a fiber optic cable, all sorts of different things. Signal processing, in the end, essentially software analysis of signals, software generation of signals, software decoding of signals always gets better. And so my expectation would be that in a certain amount of time, and I don't know how long it will take, there will be better and better ability to kind of record what's going on inside the brain from the outside. Now what it's going to take to kind of get kind of a good readout of what's happening from the active brain from the outside, I don't really know. So of these three approaches to kind of training your bot, my guess is that by far the easiest is just take your external yakking, so to speak, and use that. Is that going to be complete? Hard to know. My guess is it's not going to be that bad. And my guess is by the time it's combined with some kind of foundation model, so to speak, that's just the ambient. Oh, that's how brains typically work. My guess is it's going to do pretty well. Now, it's an interesting question how one feels about sort of an accurate bot of oneself. And in, you know, it's one thing to say, oh, you know, I'm going to upload my mind to something and then live happily ever after, so to speak. It's a bit of a different thing to say. I'm just going to make this thing that's sort of a functional copy of myself in terms of my sort of mind like activities and it's just going to sit on my desk and I'm going to be here and I'm going to be able to talk to it. It's a somewhat different kind of feeling about what's going on, but I think that will get probably that will become readily possible and higher and higher fidelity. And those of us who had the perhaps foresight to kind of record a lot of data about themselves, it'll be easiest for us to make these bots of ourselves, so to speak. I mean, for me, something that's super useful to me is because I have all this data about sort of every email I've written and received, every know, keystroke I've typed, all that kind of thing going back several decades. That's super useful to me when I'm kind of, you know, trying to remember, oh yeah, I have this vague memory of this, this or that kind of thing. I kind of can do a search. I've got a system set up to do that and I pull that right up and I. One thing I've actually been. Been feeling frustrated about is I can't do that search by concept yet. I can only do that search by keyword. And so I've been kind of thinking about how to use LLM type technology and so on to be able to really have, in a sense, a conversation with my past self and do things like if I'm writing some piece and it's like, I remember I wrote something like this before, but there isn't a keyword I can search for to remember that. It's like I just need something to. To take my concept, if I wrote something a bit like this before, to go and search things I've written and figure out what that was. So a few thoughts about brain uploading there. Let's see, are there more questions about brain uploads here? Can't install comments. Memory in human brains is like ram. It's gone the second you shut the human off. I wonder, I mean, it's, you know, what happens when biological tissue dies, first of all, I mean, some part of what's going on is programmed cell death. You starve the cells of oxygen and things. And they say, you know, I can't survive. I'm gonna, I'm gonna, you know, I'm going to, I'm going to destroy myself by releasing toxic chemicals into the cell and so on. And that's part of what happens. But, you know, in the end, the kind of, the whole sort of hardware is destroyed. But actually that takes. Takes maybe 24 hours, I don't know, some amount of time. I think that most likely when memories are stored in the synapses that connect nerve cells, most likely they're stored in calcium channels, they're stored in. What's happening is quite literally like in an artificial neural net where you have these weights for how much does the number that exists at this node, how much is that spread to another node? And so on those weights. The original way that neural nets were kind of imagined by McCulloch and Pittsback in 1943 was based on kind of the guess about the fact that what mattered was the kind of amount of electrical effect that went from one nerve cell to another through that junction. And so that seems to be what's important is, you know, one nerve cell fires, what's the likelihood, what's the strength of weight of the effect of that firing on another, on another nerve cell. And I think the belief is that when you have this membrane, there are some calcium channels and Things in that membrane between these cells and sort of, it's how many of those channels have developed. Now there are also things that happen. Proteins get synthesized when memories are formed and so on. So maybe there's other things going on. But if it is literally the number of calcium channels that got opened up, that's almost like a tiny piece of, it's almost like a tiny piece of rock, so to speak. You've got this kind of cell membrane, you've got little things that are punched through it. That's not something that immediately sort of the organism dies and that's not immediately destroyed. So it's quite possible that quite a lot of the what, what's needed to reconstruct sort of the memories of the organism may be there. Even there's nothing. It's not like ram. So just to explain that point, in different kinds of computer memory have had different kinds of volatility. And in typical RAM these days, there actually is. The memory is stored in sort of the recirculation of essentially electrons around a loop. And there has to be, I don't know how often it happens these days. It used to be like a microsecond or something that every so often there's a sort of extra kick of signal that helps prevent sort of a decay in that sort of circulating current. And so that means that if you don't give those kicks pretty soon the circulating current is going to disappear and the memory is lost. Not all forms of even computer memory have that feature that sort of all's lost when you cut the power. And sometimes also they'll just be a little, you know, battery that's giving enough power that it'll keep the thing running. But if you sort of disconnect all that power, it'll still die off. But there are other forms of memory. There used to be old form of memory called bubble memory that had the feature that it was, you know, physical magnetic bubbles were moving around and even when you cut the power, they still stayed where you put them, so to speak. And there are other forms of memory where that can happen. So it's not self evident. Memory doesn't necessarily have to be actively maintained, although the actuality of current computer RAM is, is that it is. Let's see, there's a question here about. I, I do think, by the way, that when it comes to human memory, you know, I always find it amusing that I can remember things that I learned, I don't know, 50, 55 years ago or something. You could quiz me on. I don't know Latin and Greek irregular verbs. And I think I'd probably do okay, even though I haven't to my knowledge, thought about those things in probably 55 years. So somehow that memory is still apparently lodged in my brain. I actually don't think that what I learned at that time, that the little sort of engram of that is somehow physically has been hanging out in my brain all that time. I strongly suspect that most memories we have today are memories of remembering something that we once knew, so to speak. In other words, that we have. There's a certain cycle time every so often. Maybe when one dreams or something like this. Every so often it's like, oh, yes, that one's repeating the Greek irregular verb or something, and. And one is remembering having remembered it, so to speak, rather than having that memory that stretches back 50 years or something. But I think that's a much slower process than this whole thing about what's happening in RAM where you're sort of constantly recirculating the electrical signal. Kallio is asking, what about copyright for bots based on real people? What are the legal implications of that? Yeah, it's an interesting question. I mean, it's, you know, I think there's a point of view that there are a couple of different things. There's all kinds of privacy issues, then there's all kinds of sort of who owns the intellectual property issues from a privacy point of view. My feeling is there's sort of a what's intrinsically in the brain and there's what would you actually say out in the external world? And I think the what's in the brain, independent of what you would actually say, is somehow much, you know, much less to be exposed because that's, you know, every thought that you might have about things or people or whatever else that you would never say. And it's kind of breaking a lot if you say, well, let's just go probe the inside of the brain and expose all of that stuff. That's a very different thing. And I think as one imagines sort of, you know, the here's a brain, you're going to eventually sort of analyze it, dissect it, whatever else. It's kind of like, let's, you know, set it up so that the brain is only sort of what. What's able to be exposed to the outside world is only what the brain would have exposed to the outside world if the whole brain had been there, you know, actively doing what it's doing now in terms of. Of sort of what aspect of kind of the things one knows. How does one kind of. How does one think about. I mean, it's a different form of intellectual property, I think, than any that we've ever seen before, because things like the things that one writes or whatever else, one had intended in some sense to publish those things. I suppose if you say, I'm going to make a bot of myself and I'm going to train it, you know, with the intention of publishing the bot, that's an interesting thing. And yeah, I think there's probably another form of intellectual property that's not copyright. It's not, you know, it's not something about, you know, using a likeness or something of that kind. It's another form of intellectual property that is the kind of the bot of a person that still has to be invented. Okay, so Jamie asks, if we could upload our brain to something else, would it be feasible to upload data to our brains or upgrade parts, similar to upgrading part of a computer? Yeah, I mean, I think that's the hope of kind of direct neural connections to the brain, is that you can kind of have the. Have your Wolfram Alpha or something that you can ask questions of and get all those computational answers that. That brains don't do very well and just sort of have that happen by pure thought rather than by typing it with your fingers or saying it with your voice, so to speak. And certainly one can expect that there's a somewhat higher bandwidth, at least, of sort of communicating thoughts directly to a sort of direct neural connection than there is the ability to convert one's thoughts into words and type them and so on. I don't think it quite goes as far as one might expect because I think that to communicate a thought requires some level of packaging. I mean, perhaps there's a. Insofar as the sort of direct neural connection learns your patterns, your internal patterns of neural firings, and can sort of convert it to interpret it as a kind of package thought or packaged concept or linguistic kind of thing. That's a possibility. But somehow this kind of neural user interface that exists, that has to exist, is sort of a question of how that will work, to what extent will it be. So somehow it's going to end up that in order to communicate with this sort of external computational world, you've had to package your thoughts and in a form. I think Wolfram language is actually sort of the best candidate for how to package those thoughts, because we just spent the last four years trying to solve the problem of how do we take sort of human thoughts and best Package them in computational form. So let's take that as a sort of working assumption that we are trying to convert internal thoughts into Wolfram language, computational language. Then there's sort of two possibilities. Either that target of computational language. It's kind of like what we do in WolframAlpha where we're taking natural language, English, whatever else, and converting it to computational language. Now we'd be taking not natural language, but the kind of very pre natural language of that collection of complicated nerve firings and we'd be trying to interpret that as computational language. That's one possibility where in a sense the work is being done by some sort of AI that is taking that sort of incoherent collection of neural firings and saying yes, I know I can interpret that as this computational language. That's one possibility. The other possibility, which often ends up being the thing that really happens is that there is a good neural interface that is built that is something that's a little bit like, if you look at the Apple Vision Pro, it's got some sort of new ideas about how you make a three dimensional in the world kind of eye tracking based interface to things. So you can imagine sort of a generalization of that, of how do you make something where you just have to learn to think certain thoughts and that will cause the interface to behave in certain ways. So in other words, where, where you're training the human to kind of channel their thoughts in a form that's going to work for the direct neural interface, rather than having the neural interface try to take the sort of pure thoughts of the human, incoherent as they might be, and sort of organize them in a way that can be interpreted in that way. And so, I mean those are both possibilities. But if you think about the question of sort of uploading things directly to brains, like, you know, I want to know about, I don't know, category theory. Let me just plug that little module into my brain and then I'll know about category theory. I don't think that's a very easy road because what it means to know about things tends to be sort of, you have to knit it into your existing knowledge of things. My guess is. Well, it's sort of an interesting question. I mean, it's kind of like the fine tuning that one does for large language models. There's some sort of base amount of knowledge. And then you say, well, let me teach you a little bit more stuff. That's a tricky business. You have to be careful not to just sort of ram that new Knowledge kind of brainwash the LLM by just feeding it too much of that new knowledge. Otherwise it loses. It forgets all the other. All the other knowledge. And so you know how fine tuning might work for actual human brains, what you might have to feed to them. I don't think it's going to be a. Just plug the module in and then you know it. You know, that material in a way that's knitted into other things. You know, it's a different matter if you say, hey, I want to just consult, you know, WolframAlpha or something to know this fact or to do this computation. That's something where you're already kind of separating that off. Just like an LLM is calling, for example, our Wolfram LLM API to get tools to do computation. It's the same kind of thing. One can imagine a brain calling sort of the API of tools to get these sort of computational things done. Let's see. Gosh, more questions about brains. Then we'll come to a different topic here. There a question here. What kind of compression algorithm is used for on our thoughts. Well, you know, our perception of the world is all about compression. I mean, when we look out there at the world, we've got, you know, 10 million pixels every 30th of a second or something that our brain is getting from the retina of our eyes. Yet what we conclude from that is a bunch of things about there's that object there and that object there. It's something where we can almost describe what we care about in a small number of words, even though a huge amount of data came in. And the story of perception is typically the story of taking sort of lots of raw data and somehow compressing it down to the data we actually care about. And that's something we can see. Neural nets do that with autoencoders and things like this. It's a very generic thing, is kind of take the raw data and kind of compress it down to the features that you care about in that data. Now figuring out which features you care about, that's a complicated thing. And that's something which, when you've seen lots of different kind of scenes, there'll be lots of things that are the same about those scenes. There'll be lots of things that are different. You kind of have to try and figure out what are the features that are the things that you could kind of ignore because they're always the same versus the things that are notably different and so on. That all happens kind of automatically, often in neural Net kinds of things. I mean there's this kind of big trick of autoencoding which works like this. You set up a neural net, you've got some whole architecture of. You've got data coming in on one side of the neural net. It goes through all these different layers and it's being. You're kind of taking all the bits at one layer and you're adding things up and you're weighting things and you're running them through some thresholding function and so on, and then you're getting the data for the next layer and so on. You keep doing that for a few hundred layers, let's say for typical, even quite large neural nets these days. So you've got all these layers and you're computing all these things. And the thing that you might do is take that neural net. You've got some architecture of neural net that says that this is basically where the data flows. But there's a weight that's applied to this data and you haven't yet determined those weights. But you say, here's what I want the neural net to do. I'm going to feed it an image. It's a picture of a cat, for example. At the end I want it to produce that same image, a picture of a cat. It's not trivial for it to do that because the data has been ground up into all these little numbers and so on that are flowing through this neural net. And all kinds of complicated grinding is happening inside the neural net. So it's a very non trivial thing to get it so that the weights in the neural net are such that the image of the cat coming in ends up grind, grind, grind into all kinds of mush in the middle. And at the end it's going to synthesize that same image of a cat. It's quite easy conceptually at least to train a neural net to do that because you're saying, okay, did the neural net get it right? Is it the same image coming out as went in? If yes, then good. If no, then there's a certain loss and you have to try and sort of tweak the weights. And this is the big story of back propagation and so on in neural nets. You have to kind of figure out how do you tweak the weights to so that you will go in the direction of making that image be the same as the image that was coming in. Okay, so let's say you've succeeded. You've got image comes in, any image we care about comes in and you get A reproduction of that image at the other end of the neural net. Well, it might seem like, what have you achieved with that? Well, the point is that if you now look inside the neural net, perhaps there's a place where even though at the beginning there were 10 million pixels that came in somewhere in the middle of the neural net, you kind of narrowed it down so that there were only a thousand kind of numbers that were carrying all the data that was supposed to eventually be used to reconstruct that image. So then you take out that little piece that has those 10,000 numbers, and by golly, that's a representation of the features that the neural net, in a sense, learned to care about about that original image. So, in effect, that's a compressed version of that image, compressed to be features that the neural net, given that it was learning this collection of images that were relevant, that were sort of deemed to be relevant because they were the ones that it got from the web or whatever else, then that's a form of compression. And that mechanisms like that are surely something that happen in brains, and not probably that exact mechanism, but something like that. And that's kind of the way that one gets kind of an adapted form of compression, where one is capturing those aspects of images that one cares about. Now, there may be aspects of those images. I mean, did you know that if a cat had this particular pattern of hairs at the ends of its ears, then that means that it's going to be, I don't know, a better cat hunter or something? We wouldn't necessarily notice that. And that might be something that was blown away as we, as we try and make the features of what's important about the cat image, but in a different way. If we're concentrating on something different, that might be what we would have to sort of identify as an important feature. So, let's see. RB Asked, would it be possible to read brainwaves using AI? People have been trying to do this forever. I mean, it's pretty hard. For example, detecting epileptic seizures in brainwaves is something of great importance, but it's been hard to do. And there are some aspects of sort of how brains are operating. Like when you go to sleep, there's a pretty big signal of change in brainwaves, but a lot of other kinds of things, there's not. The signal is much muddier than you might expect. And people have been trying to use various forms of machine learning for, well, at least 20 years to try and untangle that. It's been a notoriously difficult area to untangle, but perhaps it will be cracked in the current generation of kind of neural net development. I think it's not as I mean, to understand what happened in neural nets as sort of a big breakthrough in 2011 associated with image identification and image processing with neural nets. And a big part of that was so called convolutional neural nets, in which one's taking the kind of two dimensional character of an image and kind of having that two dimensionality kind of be a feature of the neural net directly, and by having sort of arrays of neurons and so on that are known to be sort of two dimensional. Then in terms of language, the thing that sort of the big aspect of the way that things like LLMs and ChatGPT and so on work is these transformers which are essentially using the one dimensionality of language, using the fact that language kind of depends on what came before, but the dependence may be quite long range, but it's still kind of a chain of things that are being said in the language. And so this idea about sort of reproducing, in a sense, geometry of the underlying sort of data stream in the actual structure of the neural net seems to be important in making a neural net that can readily be trained and so on. And when it comes to brainwaves, I think it's a lot less clear what that sort of underlying geometry might be. Maybe it's just the actual structure of the brain, but the structure of the brain is complicated because individual neurons can be connected to 10,000 other neurons. And I suppose it's kind of like what architecture of neural nets would you need to have a neural net reproduce a neural net? And I think it's sort of a different and more general thing perhaps than has been seen for images and for text and for language. And so maybe that's why it's turned out to be hard to sort of decode EEG with AI. Let's see. Eli comments. What is the artificial part of AI? As far as I can see, LLMs are a breakthrough in the study of intelligence itself. I think. Certainly LLMs have told us a lot about the science of not only kind of intelligence, but also kind of the structure of human knowledge. As human knowledge has sort of been created. I mean, there are lots of regularities in the activities that we humans show, particularly on an aggregate basis. And the structure of the knowledge we've created is undoubtedly no exception. I mean, we know things like when we have words in languages, if you rank the words from the most common might be the. Or something like this down to. Actually, people often say in spoken English the word I is the most common word. But anyway, you rank all these things and you ask how common is the nth most common word? A thing called Zipf's Law says it's about one over N. That seems to be. It's roughly accurate. It's a little bit different for different languages and so on, but it's roughly accurate. That's kind of an aggregate fact about human knowledge and human language, which we don't really know where that originates from. Actually. There are some simple models that give that. But it's something which is sort of an aggregate result of what seems to be kind of the complicated social intellectual development of language and so on. Tell you one piece of trivia I just learned yesterday. I think we've been looking at, for various reasons, we were looking at the change of frequencies of words over the last hundred years because, well, we were interested in kind of when things like Basic English, which was maybe 800 word sort of subset of English that was developed in the 1930s. And the question is, if you were doing that today, would you end up with the same words or different words? And so we kind of did. In Walton language, we have word frequency data as a function and you can. It's rather easy to just look for all the words, know how have the frequencies changed over time over that period of 100 years. And it was sort of amusing to me to see, at least based on the corpus, which is mostly books that that comes from, that the words I and you have become much more common over the last hundred years. So it's kind of a. In a sense, it felt like there's been increasing personalization in kind of the things that people write. Before that time it was very kind of third person, you know, statements. And now it's kind of like you will enjoy this book type thing rather than this book will be enjoyable or some such other thing. I just thought that was an amusing piece of trivia. But I think that in studying ChatGPT, I wrote this little book about ChatGPT where one of the perhaps interesting conclusions is that the reason ChatGPT works is because it sort of discovered certain regularities in language that we kind of should have known were there, but we didn't. It's kind of like there's a sort of syntactic grammar of language which tells one that language in something like English is constructed with noun, verb, noun, kinds of sentences and so on. But even though, you know, it must be noun, verb, noun, it can still be Sentences that say things like, you know, the cat thought, the moon, which in other than poetry doesn't really mean anything. Those are not pieces of words and meaning that fit together to obviously mean something. And so there's kind of an additional piece of the construction kit that goes beyond the mere noun, verb, noun, parts of speech type thing. That is the way that we construct language. And I think sort of ChatGPT, LLMs and so on have sort of discovered that. And that's a scientific discovery. I mean, there are many things about sort of the LLMs and the science of LLMs that are really, really, I think, very tantalizing as things that are sort of telling us things about how thought works, how language works, how intelligence works that are kind of, I think, very ripe for the picking. I mean, I think LLM science is something that is, deserves to be a major emerging area. I mean, one thing that actually person at our summer school this last year studied in some detail is if you change the temperature. So when an LLM generates a piece of text, it's doing it, you know, one token, which is more or less one word at a time. It's looking at what the text that existed before and it's saying what's the probability that the next word or next token is this or that or the other? And it's got a whole rank of 50,000 possible tokens and each one it says that's the probability that it's going to be this thing next. Then the question is, what do you actually put there? Well, the one strategy is the sort of zero temperature strategy which is always pick the most probable the thing, the token that the LLM says is the most probable and put that down as the next thing to have that turns out to lead to rather flat boring text. It's good if you're writing code, for example, but it's bad if you're writing kind of interesting, kind of lively kind of text, English or something. And so for that temperature, 0.8 seems to be a sort of popular value. What does that mean? Temperature 1 would mean you pick words not according, according to the probabilities that the LLM said they would occur in. So if the word, if a certain word has probability 1/3, you pick that word, the one that's 1/3 1 1/3 of the time and so on. And so 0.8 is a little bit closer to you pick the most popular word each time, but than just saying you pick it according to the probabilities that were suggested. Now one thing you can do is to Say, well, let's crank up that temperature. Let's go and sort of dig deeper in the words that were less probable according to the LLM. At a temperature of I think about 1 point, 1.1 maybe the LLM stops making sense. Temperature 1.4 maybe. I don't remember these numbers. The LLM stops making grammatical sentences and just starts spouting absolute garbage. So there seem to be quite sharp transitions. It's kind of like you heat up a block of ice and eventually it melts into water or whatever. And there are these phase transitions that happen in materials and somehow something like that seems to be happening in LLMs. Nobody knows why. That's an example of kind of an LLM science question that seems like it should be answerable. It seems like there should be a way of knowing that these probabilities have to stack up in this way. I mean, it's kind of interesting to think that a standard LLM is trying to have a sort of single thread of conversation. It's trying to say, going to be this token, then this one, then this one, kind of a one dimensional thread of sequence of tokens that make your sentence. Now, if you want to be really ambitious, you could kind of imagine a quantum LLM where instead of there being a single thread of output, you are accepting a large number of threads. So instead of just picking one token, you say, I'm going to pick five tokens there and I'm going to, then those will lead to different next tokens and it'll kind of branch and merge and so on. You'll get these many threads of conversation coming out of that same LLM. Now for most, most of the time, that's not a great match for human brains because we expect definite things to be said. We don't expect to be trying to track these many, many conversations that are all sort of running in parallel. But that's kind of the quantum story. And that's the story when we do quantum measurement and so on. It's all about taking those internal many histories in quantum mechanics and, and sort of compressing them down to just this one thread of experience. But we could sort of imagine the LLM that has maintained its kind of multiple threads of history as it, as it produces its output. And we could even imagine making use of that if we say, actually we have this goal for the LLM, we really want it to hit this word. We want it to say the word fish at this particular moment. Well, if we sort of maintain this quantum LLM, we allow it to do a Little bit more in terms of maintaining that and sort of saying, well, you know, somewhere you could hit the word fish and then you'll pick. Then you'll backtrack from having got to the word fish and you'll backtrack and you use that branch and that's what the LLM will say, so to speak, rather than something from one of these other branches. So that's a kind of a way that one could connect those kinds of things. David is asking about a science fiction scenario from a movie, which I do not know, so I can't comment on that one. All right, let's take a look at some of the other questions people have asked here, maybe on different topics. Jake had asked here. Will technology have an effect on human evolution and ultimately change our physical bodies in the future? EGR eyes optimized for looking at screens all day in blue light or prioritizing finger shapes to better type on a keyboard or hold a phone. That's a fun question. Evolution by natural selection is a mechanism, a very definite character. It's like if you have more kids, then more of your genes will get into the future of the species. It's kind of like the giant blockchain of life, so to speak. It's like you're here and you will have an effect. Your genes will get entrained in the genes of future generations, but only if you have kids. And the more kids you have, the more those genes will get entrained, there'll be more people who will get those genes in the future. And so really, I think the sort of. The story of traditional biological evolution is quite simply the story of who's going to have more kids. And then the question is, okay, if your fingers are not as well optimized for texting, does that mean you have less kids? And in the modern world, there isn't. There's no obvious connection that I can think of, at least between those things. And in fact, one of the paradoxes of the modern world, it's not really a paradox, is that more developed countries with sort of economies where things seem to be going better, people tend to have mostly, not always, but tend to have fewer kids. And that's how the demographic shifts around the world tend to be happening. And good things, bad things, whatever. But this question of sort of traditional biological evolution is all about who gets to have the most kids. And there are all kinds of dynamics that no doubt have an effect. I mean, when, you know, it's kind of like there's a theory that. Which I don't think is correct. But there's a theory that, you know, when the nerds are winning, more nerds have kids with other nerds. And that means that kind of the nerd genes, so to speak, which might be related to autism and things like that sort of tend to be more prevalent in the population and so on. I mean, there are dynamics like that. But this question about what will lead to more sort of biology that's of this form, well, the answer is ultimately people who have that characteristic have more kids. If they have more kids, then there'll be more of those genes in the population. That's the traditional form of biological evolution. Now, another thing about that is that there are traits that perhaps are genetically passed on that in earlier times in history would have killed people before they got to childbearing age. And with modern medicine and so on, they don't. And so those traits that might have died out in the population because those people with those traits wouldn't have made it to childbearing age, they persist in the population. So that's kind of a. And, and might cause trouble in future generations or maybe at some time in the future. Whatever is that problem is just easy to fix. Just take these pills for a week and it'll be fixed. And that's all there is to it. And so it's sort of right for that not to be weeded out of the population because it isn't really a big problem. But I think. So that's. In terms of traditional biological evolution, it's hard to tell what. I mean, there are certainly cultural aspects. You know, there are definitely cultures in which people say, have more kids and those, you know, whatever traits there are associated with those cultures will be more prevalent in the population. And, you know, cultural traits where it's like, don't have any kids, well, those are going to die out in the population. That's traditional biological evolution. There are more complicated issues. So, for example, one thing is, for example, gene editing. If it is the case that you edit the genes of either the, you know, which, which lead to the cells that get passed to future generations, then you've changed future generations. I mean, and obviously with, with IVF and things like this, if you, if you start, you know, picking, picking the embryo, editing the genes in the fertilized egg or whatever else with those kinds of things, you can change future generations. And you can certainly imagine a world in which there's all kinds of fashions about, oh, I think having green eyes is good. Let's edit the original cell to. For my Kid to have green eyes or something like this. And you can imagine those kinds of things. But, but once that's been done and it's part of the, the germline, so to speak, it then it just gets passed down to future generations. And you know, one can, you know, there are all kinds of crazy things that could happen as a result of that. I mean, I've always thought that the, the ultimate sign of weirdness of those kinds of things, of gene editing is if people decide to incorporate, you know, jellyfish genes that produce glow in the dark kind of behavior, you'd end up with people with, you know, all kinds of weird sort of tabby cat like arrangements of glow in the darkness and so on. Very, very strange kinds of things. But that's a way in which you can get sort of, you can just change the genetics. Now another thing that, that people sort of. There used to be this big sort of, you know, disagreement in biology between Darwinism and Lamarckism. Darwinism said the only thing that determines future generations is natural selection. Lamarckism said things that happen to an organism during its life independent of just whether it had more or less children, those things can be passed to future narrations. And as is usual in biology, perhaps unlike physics, in physics it tends to be the case that the simple explanation is right and there's a simple clean explanation. In biology it's often the case that there are sort of everything is as complicated as it could possibly be. So while Darwinism is out and about and the main effect, maybe there is some Lamarckism that's still there, where some things can be passed to future generations. For example, there are some, what is it? Vitamin A deficiency that if the mother has it, that will cause things in the kid. There are other drug related things that happen. If those happen to the mother, they're passed to the kid. There are viruses where that can happen these kinds of things. So there is a certain amount that can be sort of passed in a way that is not through the pure genetics. The other thing that I suppose is an emerging kind of area is the sort of features of the genome that aren't just the genome. Things like the methylation patterns, things that are sort of beyond genetics. Epigenetics is a term sometimes used, but that's had different meanings at different times in the history of biology. So it's a little confusing. But there are these things that are sort of the things that happen to the organism during its life that might get passed to future generations. And one can imagine that happening for us Humans, I mean, people say, I don't know. The latest scare has been the nanoplastic scare of, you know, there are all these little things that all these pieces of polymer that maybe, maybe it isn't the polymer itself, but it sort of picks up other atoms and so on and you ingest those and they go everywhere and then maybe they're passed to, even to future generations and so on. I don't know if that's a significant effect. I mean, one of the things about biology, as I say, is that everything is incredibly complicated. And so when you say is there an effect from X? The answer is almost always yes. Is the effect important? That's a completely different question. Can you find, for example, a common one is can you find cells from a child in the mother's bloodstream? And the answer is yes. But is it important? Are there lots of them? That's not clear. So it's always in biology it's always a bit confusing because if the shocker. There's this thing that's possible and you can detect that is an important as a different question. But so I don't think so that's in terms of there's this kind of Lamarckian evolution which is more plausible to be a thing that's passed to future generations. The other one is just how many kids do you have? I mean, people say, you know, humans, you know, if we go and colonize Mars, lower gravity, other different characteristics, you know, oh, that will be. Humans will get, you know, taller, this, that and the other. Only if that, you know, only if the, the ones who aren't like that don't have kids on Mars, so to speak, or don't have kids. I mean, that's the only mechanism in sort of the Darwinian model of evolution that changes these kinds of things. Let's see, Just looking at. Hybe asks, what do you imagine the future of communication to be? We had the spoken word, mail, letters, telegrams, phone calls, emails, texts. What's next? It's an interesting question. I mean, as I was saying earlier, I don't think thought to thought communication is going to work as such. I think that just like in a computer, if you took the memory of one computer, you just stuff the bit pattern from the memory of one computer into another computer. It's not going to be very useful because it will have allocated memory differently. It will have this and that and the other. It won't be a perfect match. What you end up having to do when you are communicating from one computer to another is to Kind of package things up into some protocol or whatever else and take this thing that has been structured in some kind of linguistic form, feed it to the other computer, it unpacks it, and that turns into some sort of engram in the memory of that second computer. My guess is the same thing is necessary for humans. We won't have sort of direct thought to thought communication now. You know, people have done experiments. I think the early Apple Watch had a thing. Maybe they still still has this where you can kind of send your heartbeat to another Apple Watch, so to speak. I think I, my impression is that was not a huge success. I think the thing that is sort of interesting with, with things like the Vision Pro is it's trying to take kind of facial expressions and communicate facial expressions. Obviously you can do that by just, you know, just using video conferencing or by sitting so classically so traditionally in the same room as somebody else and just seeing their facial expressions. I think those are, those are things. I think, you know, an interesting sort of possible direction is the my bot talks to your bot kind of communication. There's one kind of thing which is a, you write something, somebody else reads what you wrote. But there's sort of the question of maybe. And the next level of that is I write something in language X, you read it in language Y because there was automatic translation done. So sort of the next level of that. So that's probably, that's one of the things that's definitely a near term coming attraction is good sort of simultaneous translation between different languages. Now as soon as you get to that, there's all kinds of cultural translation you have to think about. It's actually, you know, can you really get the meaning from this one thing and turn it into the meaning of this other thing in this other language? Those meanings, they may not match up. There may not be a corresponding meaning in the culture that's connected to that second language. But I think the next level would be something where it's more like my bot talks to your bot where you've kind of said I'm doing all this stuff and I want to communicate this major thought and then to somebody else's bot and their bot. So the receiver's bot kind of knows how the receiver thinks and so it can kind of interpret the thing that's being said in the terms that the receiver wants to get them. So for example, the first person might be saying, I'm let me explain to you, blah, blah, blah, blah, blah. And the second person might, their bot might know, you Know what? I don't need to tell my, my person those things because they already know all that stuff. But there's this one thing that the first person said that was really important that the, that my, my owner or whatever didn't know. And let me emphasize that point rather than repeating all the stuff my owner does know. So I can imagine sort of that's another level of communication is the my bot talks to your bot. And there's different versions of this that we already can see. We can see this in the way that LLMs get used as linguistic interfaces, where kind of the typical sort of story is, I've got five points I want to make in my report. But the way that the world works, I have to file a big essay as my report. I can't just make those five bullet points, and I've got to kind of have a certain boilerplate around it. So I give it to an LLM, it puffs it out in this big report, I send it to somebody else. That other person says, I don't want to read this whole report. All I care about is these three bullet points from that report. And then get their LLM to kind of grind that down to those three bullet points. And so there's been sort of a form of communication. Five bullet points to three bullet points. They're not quite the same bullet points. It's been kind of a my bot talks to your bot kind of communication. I think that's one thing. I mean, this also comes up in things like tutor bots where you say, I've got this whole area that I sort of got from a textbook. Now I'm trying to teach that to my owner. I don't like the term owner. There's got to be a better term for the, the kind of the. The person whose bot it is, so to speak, but let's say owner. Okay, so, so the. It's, you know, the, the tutor bot knows its owner, it knows the student, so to speak, and it knows what things it actually has to say to the student based on their pattern of existing understanding so that they can learn this new thing, you know, actually, as a practical matter, we are. We've got a fairly substantial project to build a kind of AI tutor type system. It's kind of an interesting situation because it's kind of. Everybody can go in and say, we've got an LLM in five minutes. We can make something that's sort of an AI tutor. But as soon as you poke at it a lot and you ask, does it really work? It doesn't. It's actually horribly bad usually. I mean it looks good for the five minute demo but then the more you poke at it the more like oh this is quite useless type thing you get to. And so, you know, we've been trying to figure out how to get past the 5 minute demo to the thing that's actually useful, that can actually teach some substantial subject to a student and it's been going quite well. But it's sort of interesting that the, that a large part of what's needed is really sort of planning, structuring, software engineering, connecting to actual computational knowledge these kinds of things. It's in a sense very non LLM ish. There's a, there's sort of an LLM layer that's important for this sort of linguistic interface but then there's a lot of kind of just hard work sort of software algorithmic knowledge building type thing that's needed and we do that kind of work. That's what we've been doing for decades. So we I think have a better chance probably than anybody else to actually make this AI tutor thing work. But it's hard work and we'll, we'll have to see how, how that develops. Let's see. So that will be my. In future of communication. My, my immediate thought would be my bot talks to your bot, so to speak. Let's see. WM comments? I need an AI version of Socrates to speak with about being. Yes, I think, I mean it isn't, you know, there aren't a lot of works of Socrates that have survived. Maybe more we found, let's hope and well of course Plato was the one who actually wrote down the, you know, the dialogues with Socrates and so on. But you know how much of we were talking about brain uploading, how much of kind of the Socrates can be reproduced is. Is. Is not clear at this point. Let's see. Question from Natonan. Are there limitations to using genetic data to understand language related traits? And if so how might these limitations affect the accuracy or applicability of language models? So I'm not sure when it comes to humans and genetics and the relationship to human language. I. It's an interesting question whether there are genetic correlates of particular languages. There are obviously different, you know, ethnic groups and so on around the world, different haplotypes and so on that have, you know, that have traditionally spoken language A or B or C. But I don't think I've never heard of any known effect where a kid who's adopted who is with this haplotype adopted in this country with this language, I don't think there's any known kind of difficulty, so to speak, in sort of learning the local language, so to speak. I do know that with different languages with different structures and so on, there are definitely different architectures of brain processing that develop as a result of using those languages and learning those languages early. And for example, the ways languages represent numbers. I think there are things that are known about the different ways in which sort of the brain chooses to organize itself with respect to that that are kind of where you're kind of pretty much stuck with the one you first learned. But suppose I know there aren't cases where one knows of sort of a correlation between kind of the raw genetics and the language that's spoken. Be an interesting thing if there were. I mean, I suppose in people wonder about Neanderthal, for example, what kind of language might have existed there? And can one deduce things about the language or the lack of language or whatever else from anything about the genetics? And I think that's sketchy at best, mean perhaps one can deduce something from the. From the physiological structure of the vocal tract. But even that, I think, is. Is questionable because, you know, we have sign language and so on. We have languages that use whistles and clicks and things like this. So there are ways, given that you have the idea of language, you can represent it with gestures and so on, and just as you can represent it with spoken, with auditory spoken words and so on. So I, I think that the idea of looking for, oh, there wasn't a language because the vocal tract didn't have a particular form is probably not a convincing story. James asks, Do LLMs work well in Egyptian hieroglyphic concepts? I don't know for sure. My son Christopher actually does things with LLMs and also does things with things like Babylonian tablets and so on. And I know he was trying to teach an LLM, what was it? Sumerian, maybe, a language in which there's rather small corpus. And I think he was having some success with that. In other words, does an LLM have enough knowledge of the sort of intrinsic character of language to be able to. Is there enough commonality, is there enough universality in human language to that if you try and sort of teach it Sumerian, that it will be able to pick that up just by knowing sort of the overall structure of human language? You know, people have told me that even in languages like Russian, that there's sort of somewhat unconvincing behavior on the part of current LLMs, even though there's obviously a very decent amount of training data for a language like that. It's. And I did notice, actually I tried this out with somebody just very recently, and it produced stuff in Russian. And the person said, that's not a word. It made that word up. Now I know when it writes Wolfram language code, for example, LLMs quite often make up function names that don't exist. Actually, sometimes those functions are good ideas because the LLM kind of knew the conventional wisdom of what people often talk about, and it extrapolated correctly. And it's like, yes, we could actually, we actually just. In the most recent version of orphan language, we actually. A couple of new functions that we added were effectively suggested by an LLM. But I think that this question about how well can an LLM do a language where there's a small corpus. I'll give you one other data point. When LLMs first burst on the scene more than a year ago. Now there is a constructed language called ithql, which is probably the most sophisticated of the constructed languages. And I sent the creator of ithql, who I know, a little piece of Ithquil which obviously I can't write, but I sent him a little piece of Ithquil written by chatgpt, and he's like, this is terrible. It's got kind of the right general kinds of ideas of how words are constructed, but it's nonsense what it wrote. So that kind of gives you some sense of. Of kind of where it's at. Ithquil doesn't make it Sumerian. It was doing a little bit Russian. It's still got some problems. English, it, you know, at least produces kind of reasonable English that doesn't look obviously wrong. Let's see. Oh, Mood asks, do you regard math as a language? And if so, would you suggest talking math out loud with a child? Well, there's a lot of. Kind of making use of linguistic construction. So what's the essence of language? What is it that makes us talk about, say, things are languages. There's a certain kind of modularity and compositionality to things. That is. It's kind of like we could just be saying, you know, we just could just have one word or one phrase. We just keep saying it, you know, like I am Groot or something from that movie Guardians of the Galaxy. Or we could just be. Be saying, pointing at things and just using a noun to say those things, which is what young kids start off doing often. But then the thing that is really the power of language is the fact that there are these modular pieces that you can put together in this infinity of different possible ways. And for many of those ways of putting them together, what you get is a different meaning. And that's kind of the big multiplying power of language is that ability to sort of compose things that way. And math has that as well. Math notation is set up to use that kind of modular compositionality kind of idea, by the way, in our computational language, in Wolfram language, that's exactly what we're trying to leverage as well. We're trying to represent the world computationally. In English, Maybe there are 40,000 words that typical vocabularies use. In morphem language, there's about 7,000 sort of base concepts that map onto that make use of. In our case, we wouldn't really be able to do what we can do without the fact that we're kind of leveraging the fact that people already know those words in English so we can use those words to indicate these computational concepts. But math is absolutely using this kind of language, like idea, Spoken math is not great. Spoken computational language we have never solved. I really wanted to solve that. We've never solved it. When you read Wolfram language code out, yeah, you can read it out, but you're often saying, you know, brackets, comma, whatever. That's very inelegant. The thing is in fact a sort of tree structured context free language. And it should be possible, just like I can kind of talk English at a high speed, so to speak, it should be possible to talk Wolfram language at a high speed and have it be something that can be readily unpacked and understood by somebody. Now, it doesn't help that in English, for example, we avoid sub, sub, sub, sub, sub, sub, sub clauses. If we say the cat which was sitting on the mat, which was on the floor, which was in the hallway, which was, which was, which was, which was ate a mouse. We go down, down, down, down, down in those subclauses and then come back up again. Mostly one loses it after maybe 3, 4, 5 levels of subclauses. And so we don't tend to make language. We would say that's a very bad way to express yourself, to have all those sub, sub, subclauses. And so we don't tend to make language that way. Instead what we do is, you know, we flatten things out. We say one sentence, then we refer back to that sentence, and so on. In computational language, we do allow sort of arbitrary nesting and it's quite often Used. And that may be what makes it hard to say. And it may be that we just have to sort of bite the bullet and say we have to have one piece of computational language here that defines something, and then we refer to it and the next thing we're talking about, and that's what we need to interface with kind of the finite stack depth of human brains to do it that way. Now, in terms of math, it has the same nest, nest, nest problem. And you can get kind of math that would be pretty hard to understand in spoken form. You know, is it something that is, you know, if you, if you talk math with a kid, and I don't know how well that would work, you know, I had one experience a few years ago. I happened to meet a group of, what were they, maybe 11 year olds or something, 10, 11 year olds, who were very energetic, enthusiastic users of our Wolfram language, computational language. And they spoke it and they were very disappointed that, you know, they were speaking it very fast. And I was like, I don't understand any of this. But they seem to be able to communicate in a spoken form, which I think included brackets and commas and so on. So, you know, that, that kind of gives a lie to what I'm saying. That is, perhaps if you learn it early, you can learn to speak these things, speak these kind of structured languages like computational, like our computational language, maybe like math. But so, so for all I know, that's something that is, you know, a possibility if young, so to speak. I mean, the whole idea that you need to learn a language when young and it's all over if you're older, I think that idea has kind of somewhat melted away. I think that there are things where, for example, learning phonemes, you know, when my kids were young, for example, it was, was kind of the thing that one of the things was, oh, you know, play them these different phonemes from around the world as a way to get them used to the weird different sounds that exist in different languages and wasn't terribly popular with them, but sort of amusing and, and you can potentially get your kid to be able to make some sound that you barely can notice as an adult is a different sound in some different language and so on. But I think that, that the need to learn that early has somewhat melted away as there is more ability to see, have feedback for, oh, I said this thing. Oh, there's an automated system that can say, no, no, no, you didn't pronounce that correctly. And you can kind of learn that at sort of any stage. And so it isn't the case that the whatever mechanism, whatever sort of ability to imprint those things at a young age, that advantage seems to have largely gone away. Let's see, maybe one more. I mean, this whole question about what, you know, people talk about, you know, what can you teach to kids? I mean, as somebody who's, who's tried to go on learning things throughout my life, I'm not so sure. I think I learn things much better now than I did when I was a kid. So I'm not sure that the claim of, oh, you should learn it early or you won't learn it. I think that's more a matter of motivation than it is a matter of kind of brain structure. I think it's more that, oh, there's this time in your life where, you know, where you clearly have got to learn things in order to operate in the world. So that's the time to just learn as much as you can learn. And once you're kind of grown up and operating in the world, it's like, I don't need to learn anymore. I think that's a matter of motivation rather than a matter of what's possible. That may be less true with more kind of manual, more kind of mechanical skills and so on. Those I can quite believe are more kind of, you've got to, you know, the reflex arc has to be developed and so on. And that may be a. That's probably less true that you can learn those things later on. More intellectual things. I tend to think you can learn them later on. Let's see. Luca comments says in a recent podcast I mentioned that some subjects in academic fields are in a position to be combined with computation, haven't been so far. And that it was low hanging fruit if one could do it. They were wondering if I had some examples of low hanging fruit. Well, you know, I think we could pick. If you literally start asking about any area for which there's a kind of an academic department or courses and so on, you can say, is there a computational version of that? Let's take computational history as an example. I don't know why I particularly think of this one, but what does that mean? Well, one thing is history has tended to be very kind of essay based in the sense that it's actually, I'm going to give several examples here, but. But you know, it's kind of like, well, what happened at this time? Well, I can write an essay about what happened, but you know, even in, for example, or from language, we have a lot of Data about history. We have, you know, the borders of countries at different times in history and so on. We have information about leaders of countries. We have information about battles and wars and things like this. There's a fair amount of structured data about these things. And so one thing one can certainly do is to say, well, in traditional history, there's kind of like I, as a matter of essay writing, can identify that there is a trend of this or that kind. But one can imagine having sort of the data and say, well, let's actually plot this against that. Did countries on average get bigger over this period of time? Did this happen sort of in this way or that? Can we sort of make quantitative those aspects of history? Now there's a new form of quantitation that exists because of LLMs. One can start asking things. One can ask the LLM, so to speak. Given that you ingested, you know, kind of all the speeches of this person versus that person, what kind of sort of text statistics can you say? Now, what's happened in the past is when you say, I've got a piece of text, I want to say something quantitative about it. The only thing we've been able to do is to either say, well, how many times did the word get said? Or something like this, or we've been able to pick out certain entities. How many times did the country X get mentioned? Or something of this kind. But what we have now with LLMs is something where there's a broader, and we don't yet understand it very well, there's kind of a broader way of characterizing text. Just like you have a bunch of numbers and you say, what's the average of all these numbers? That's a way of characterizing this whole collection of numbers. Got a piece of text, you can say, what's the text average? We don't know what that means yet of this text. One example of this might be sentiment analysis, where you say, is this text upbeat, downbeat, whatever else? That's always been a slightly fishy kind of thing to be asking about, because really often in many cases what you want is the sentiment is, are the people going to buy this thing or are they not? And that's kind of a what's going to happen? Rather than predict what's going to happen, rather than are they happy when they're reading this piece of text or are they not? But, you know, I think some of the things that one can imagine, well, for example, in terms of, I don't know, biography, for example, for me, I've recorded all this data about myself, you know, 3 million email messages and, you know, all these keystrokes and this and that and the other, every step, every heartbeat, all these kinds of crazy things. So there's a huge amount of data. And the question is, if I want to kind of identify something that is of biographical interest, it's kind of like, how do I take that data and. And turn it into something that is a kind of a narrative thing to say biographically, that and sort of, how do we think about that? If we go in sort of human psychology side of things, there's sort of a computational psychology where you're starting to say, you know, was this person, you know, what was happening psychologically with this person? Given that I just have 100,000 things they wrote over this period of time or something like that, I think that the story of sort of how do you think computationally about different kinds of things? It's really a question of how do you structure things to the point where you can say, here is this computational structure, this thing that's represented as images and dates and this and that and the other things which we have sort of an intrinsic computational understanding of. How do you take that structure? What do you do with it? What kind of functions are you applying to that? And sort of what happens? What can you then say? And I think that this kind of, how do you sort of structure things to think about them computationally? It's something that's not been well codified. In fact, the fact that it hasn't been means that I. For a long time people have been telling me you should sort of try and codify this. And I am now in the early stages of writing some book and course and so on about Introduction to Computational Thinking, which is my effort to try and sort of codify how you do these things and sort of how you. What raw material you need to know to be able to think computationally about different kinds of areas. I have to say, this project, you know, given that I've been sort of feel like I've been thinking computationally for at least the last 40 something years, this project, nearly 50 years now, actually, this project is disappointingly difficult, so to speak, in the sense that it takes some effort to kind of verbalize the things that one thinks one knows well there. All right, well, I see a lot of other interesting questions that are saved up here that have come in here, which I'm afraid I'm going to have to take another time. But thanks for joining me for this live stream and look forward to chatting with you another time. Bye for now.