Science & Technology Q&A for Kids (and others) [Part 117]

Hello there. Welcome to another episode of Science and Technology. Q and A for kids and others. And I see all kinds of questions here. Let's see, where should we go first? I see a question from Joss. Can you talk about Pangea? How do continents shift? Is it possible to reform Pangaea? Is the technology that could prevent this? Okay, so first. First thing is you look at a map of the Earth today. It has its continents arranged in the way it has them arranged. It hasn't always been that way, and it won't always be that way. If you look back 100 million years, for example, to when there were dinosaurs running around and so on, the configuration of continents was considerably different. And for example, famously, kind of South America and Africa were kind of glommed together, as were a bunch of other continents in arranged in different ways. You know, it's an interesting story in the history of science. People, as soon as there were decently accurate maps of things like South America and Africa, people said, gosh, if these were puzzle pieces, you know, Africa would fit into South America. How can this be? And people, as a chap called Alfred Wegener, who was big on this, were like, well, maybe there is continental drift. Maybe in fact, these continents are moving around on the surface of the Earth. And maybe at one time, South America and Africa were fit together and then they broke apart. And now they're a distance apart. And you can see the two sides would fit together if you put them together as puzzle pieces. And for a long time, people were like, no, no, no, that couldn't possibly be that way. You know, there's no way this is real. Took a long time, and then eventually continental drift was discovered. The way it was discovered was that there are. There are. The magnetic field of the Earth magnetizes rocks. And what was discovered was. And it does so that the rocks kind of get frozen in their magnetism when they cool after they've been produced as lava from a volcano or something, then they cool into solid rock, and then things about their magnetism get set. And you can kind of see these rocks had magnetic field that was set in this direction and these ones in this direction. And you realize that actually these were. These. These rocks had to be sort of in the same place on the surface of the Earth at the time when they were formed. And that was kind of a clue about continental drift, about the fact that actually the continents are moving around on the surface of the Earth. Nowadays it's actually possible to. To make measurements with radio telescopes and things using the radio telescope technology. But Basically making precision measurements where you can actually measure directly the motion of continents. In fact, people have even considered the possibility that with the GPSs built into cell phones, that if you could collect all the GPS data from all the cell phones, that somehow you would see this kind of overall flow of, of this tiny amount of drift of continents moving on the surface of the Earth. I think the fastest moving ones are, I think Australia, New Zealand. I think they're moving pretty quickly, like maybe, oh, what is it? Could be, what is it, millimeters or centimeter per year, something like that. It's of that order of size that they're moving in a certain direction on the surface of the Earth. How does that motion work? Well, basically, the center of the Earth, we go down more than 50 miles. The Earth is liquid. There's still remnant heat from when the Earth was formed 4.6 billion years ago. Maybe about half the heat of the Earth is just the result of heat that was in it when it formed, when all the pieces of the Earth were pulled together by gravity back when the solar system was formed. And the other half comes from radioactive elements in the, in the Earth that are decaying and producing energy. And that's another source of heat inside the Earth. But whatever the source, the result is that the Earth has a liquid core and molten core. It's not liquid like you couldn't pour it, you know, if you picked it up, I don't think it would pour. I think it's much more viscous than that. But it is liquid in the sense that it flows. It's molten rock that flows around. And more than about 50 miles down, that's what you'll see. Now, essentially what's happening is that the continents are floating on the surface of that liquid, and different things are happening. There are places where pieces of, where there's sort of an upwelling from the liquid core. And on the two sides of the upwelling, you're producing more crust, more solid crust. There are other places where, where it's being pulled down. So in, let's see, where are they? In places where there are lots of volcanoes, there are lots of lines of volcanoes. Those are the borders between these plates. So the things that are sort of the solid pieces of the crust are so called tectonic plates. Those are the things that are kind of floating on the surface of the mantle. So on the Earth, I don't know how many, there are maybe 20, 30 plates, something like that. 20, 30 blocks of solid stuff that are separately floating on the mantle. And at the boundaries between these plates, things happen, like you'll have volcanoes or you'll have earthquakes where there's motion between these plates. Like in California, there's the San Andreas fault that runs near San Francisco and so on, where there are, at a small rate, these two different plates, there's two plates that are moving relative to each other. And what tends to happen is most of the time, it's just the thing has. There's sort of a continued kind of force of moving one plate against the other. And just like when you pick up an object, a ruler or something, and you put your fingers underneath it and you'll see that it'll stick for a while, and as you move your fingers together, it'll stick and then slip, stick and then slip and so on. The exact same thing happens between these tectonic plates. It'll stick for a while and then it'll slip. And at the moment when it slips, that's when an earthquake happens. But in the end, the plates are moving relative to each other. They tend to not move smoothly relative to each other. They tend to go stick and then slip. And the moment when they slip is the earthquake. People wonder, could you lubricate the faults? Could you pump some fluid down into the faults to lubricate it so it wouldn't stick? So it would just smoothly glide? I don't think that's been a successful thing that's been done, but that's a sort of an engineering, some kind of geoengineering thing that one could imagine. Now, in terms of what can one figure out about these plates sort of floating on the surface, on the mantle? There are things happen when these plates crash into each other. They'll push up and you'll get mountains like the Himalayas and things like this. When, you know, a variety of different things happen over long periods of time, over periods of tens of millions of years. And things like this, there are these maps that kind of show based on the plates as they are today, you can kind of work backwards saying, this plate is moving. You know, the Australia plate is moving at a high speed. Where was it originally? Well, you can kind of move, you know, you can. You can project backwards and say, where did it come from? Given that it's moving in this direction. Now, if you do all that projecting backwards, one of the things you discover is that back, I don't know, a few hundred million years ago, don't remember exactly how many wolframalpha knows this kind of stuff. All the continents that we know today are consistent with having been In a single lump, little Pangaea lump like the Earth. Right now, it's about 3/5 of the surface of the Earth is covered with ocean, but the oceans are broken up. We have a bunch of different oceans. Atlantic, Pacific, things like that. We have a bunch of continents. The Americas, Asia, things like this. Those are separate continents. Back in the day, according to the sort of calculations based on sort of working backwards from existing continental drift, it would be the case that there was just a single continent, the single Pangaea continent. It's kind of the Earth was kind of part land, part ocean. Just some land, some ocean. I have to say, I find this claim a little bit suspicious. I'm not sure it's right because, you know, we know how continents are moving today. We project backwards, we say, how are they moving? Before we can work out a certain amount from magnetic alignments and rocks and things like this. But I kind of wonder, you know, there are. There are a lot of configurations that. I mean, it kind of. I don't know why it would be the case that the early Earth would have been sort of part land, part ocean, just in two groups like that. Perhaps there's a good reason for that. I don't know. Conceivably, it might have had something to do with the formation of the Moon, although I doubt it, because the Earth, unlike the Moon, the Moon is not like kind of a uniform globe. It's of rock. One part of the Moon is really of different density than the other parts. So in the Earth, that's not the case. The Earth is kind of a uniform sphere of rock with the gravity of different parts of the Earth is more or less the same. It's 9.8 meters per second per second is the acceleration due to gravity with some small changes in all different parts of the Earth. On the Moon, one part of the Moon has significantly different gravity than the rest of the Moon because there's a part of the Moon that has very different density from the rest. And people believe that in this kind of big splat theory of how the Moon was formed, that there was kind of a body that collided with the Earth that then caused the whole Earth to become molten. And a piece of that sort of broke off to be the Moon and the rest kind of fell back to be the Earth. In that theory, that has some explanation of this kind of this sort of big lumpiness to the Moon. But in terms of why should that kind of picture have led to something where half the Earth or some little bit less than half the Earth is land and the rest is ocean. I don't really know. I do wonder when you try and sort of work backwards from the current thing and you conclude, oh, everything could have fit together like this. Maybe if you work backwards even further, the things came apart again and there was a different confluence configuration that was to be found there. Now, in terms of what will happen in the future, I don't actually know what the projection is of what will happen. I mean, I know some pieces of the current continental drift, what will happen to them? Some continents will crash into each other, some things will move apart and so on. And I don't know what the end result is. If we looked in the Earth, if left to its own devices, so to speak, a billion years in the future, what will the surface of the Earth look like? I should know that. I don't know that one could work it out. We have the data in Wolfram Alpha and Wolfram language. One could just do the computation. But I haven't done it based on assuming that the current motion of consonants continues the same as it has been. And that, of course raises questions about, well, what produces the kind of the currents, the flows in the mantle of the Earth and the liquid part of the Earth that cause these kinds of motions. And I think there's lots there that isn't known in terms of can we affect continental. Do we care about affecting continental drift? Right now we're worried about things like sea level rise. The sea level over the course of geological time has gone up and down a lot. You know, we're worried about fairly small changes nowadays because, you know, we have a lot of coastal cities and things like that. And at this point, a, you know, even a few meter change has at least a big effect if you happen to have, you know, an oceanfront property or something like this. So, you know, the scale of continental drift is really very small compared to the scale of things that are associated with various kinds of sea level rise and so on. So I think there isn't much motivation to say, gosh, what can we do about continental drift? The only place where that's important is when it comes to the stick slip behavior that leads to earthquakes. And that's a place where this idea of pump oil into the San Andreas fault in California is something that's been talked about. I think everybody's a bit worried if you do that. Maybe you just start this giant chain of earthquakes. And that's because one doesn't really know the effect of that. You know, it's a funny thing in the way that science works. The models that are used these days of that stick slip behavior in earthquakes are actually typically models based on things called cellular automata. And they're models based on things that I kind of worked on in the early 1980s. So I think the current models there are sort of based on things I did a long time ago, although I wasn't the one who did them in the application to earthquakes. Anyway, a few thoughts about that. Let's see. Okay, here's a question. Various questions about ChatGPT that still seems to be in the news. We're still doing things with it. It's, you know, the big story there tends to be that ChatGPT is very good at writing stuff and the stuff might be fact, it might be fiction. There's the sort of great thing that can happen of ChatGPT asking WolframAlpha questions, which WolframAlpha can then really factually respond to. And that becomes an interesting way to kind of fact enhance what GPT might otherwise be writing about. That otherwise might be complete nonsense. But there's a question here. How does ChatGPT work? All right, I can tell you something about that. So first of all, what is ChatGPT? It's. Well, these days it's kind of a website made by an outfit called OpenAI that you can type text into and you can have a conversation, you can chat with it, you can ask it questions, you can ask it to write essays for you. And it does a remarkably good job of writing very human sounding essays. I think one of the first tests I did on it was write a persuasive essay, I said, for why wolves are the bluest kinds of animals. I'm not sure what I was thinking of typing that in, but the result was quite interesting because it wrote something that started off by saying people don't usually think wolves are blue, but in fact there is a species of wolf that lives in the Tibetan plateau that is blue and its name is this. And its blueness comes from the same mechanism that leads to butterfly wings being different colors and so on. And it went on and on and on for a bunch of paragraphs. And it's been kind of taught to make essays like some high school kids are taught to make them, where there's a paragraph at the end that says in conclusion, comma, et cetera, et cetera, et cetera. Well, it's the things it wrote sounded really quite convincing. They were convincing enough that I was kind of looking up on the web. Is it actually true? Is there in fact a blue kind of Wolf. It was all complete nonsense, but it sounded very convincing and it had all the right structure to be an essay about blue wolves and so on. How does it work? Well, what fundamentally it's doing is it's making plausible English text. And what it's doing is you give it a prompt, a piece of initial text, and it's trying to continue that text in a plausible way. So if you say something like, well, at a very minimal level, if it was doing that character by character in English, if you put a Q, you know, it's followed by U. If you put a th, well, there's a good chance it's followed by an E. At a level of letters, we know the statistics, the chances that different letters follow each other in something like English. But now we can generalize that. We can make it bigger. I was just looking. If you look at nouns and adjectives, for example, oh, gosh, I was just doing this, and I'm already forgetting the results. But you can, you can say things like, if you look at typical English, you know, you could say, AI is okay, that's a, you know, and then what's the word that comes next? You know, maybe the next word is most commonly very, or, you know, or might just be, you know, an. For an interesting technology or something like this. But you can ask the question, what is the typical word that follows? AI is blah. And so how do you define typical? Well, there are many words in principle in the grammar of English that could follow, but there are particular words that we humans have typically used when we actually write essays and write text about things. And so, for example, in the case of ChatGPT, it took a crawl of the web. So the web has, I don't know how many it is these days. It's a few tens of billions of pages that are reasonably human written that exist on the web. Not. Not that many billions. And there are many more. If you say, how many different possible URLs are there? There are trillions of URLs that get used. But the number of really human written, I've got a lump of text. Some human wrote this. The number of pages is, is not that different from the number of humans. Not that we were all page. Some of us have written a lot of pages, and other people haven't been in the business of writing web pages. But it's, you know, tens of billions of pages that exist on the web. And one of the things you can crawl the web, you just start from one page and you follow all the links on that page. You follow all the links from that page and chances are you'll eventually get to sort of all the pages on the web. That's a good guess at least. And so, for example, there's a thing called Common Crawl, which is a project where people have just collected, have gone and done that crawling, and have visited all these websites and repeatedly do. So you can get that. It's a big lump of stuff. You can download it from various websites, and that is a copy of the text content of the web. It's a little bit more complicated than that because a typical modern website doesn't just have text on it, it has a bunch of JavaScript and so on it. So I think you typically download things called WARC files, which are kind of web archive format, and it all has to be unpacked and you have to decide what you really mean by the text and so on. But let's assume you've got big blobs of text that came from the web. So you got that. You've also got books. There are maybe some number of tens of millions of books that have been written. There are maybe 5 million books that have been scanned, maybe a little bit more than that now. And you can scan the books, you can do optical character recognition, you can pick out the actual text in them. So that's another big corpus of material that you can feed in. And ChatGPT was trained on common crawl plus a bunch of books. And so when it says what word typically follows this? It knows that based on having seen all of those different parts of the web and these books and so on, it's like in those things, in those things that humans wrote, what is the typical thing that is the next word? Okay, so that's where it kind of starts. But this whole question about kind of what is a typical piece of English, it's more than just saying, given that I had the word is or the word, you know, I'm going to have the word cat. What's the word that comes before cat? Might it be black or white? You know, it's more than just word by word kind of statistics of what's happening, more than just the chance of what word follows which other word. It's dealing with things on a larger scale where it's saying, when we look at this whole flow of this sentence or many sentences together, what is the typical way that we would put these things together to form something which is sort of typical of what you would see out there on the web and so on. And in fact, in ChatGPT, I think in its current form it's really dealing with things that are length. Well, 20, 48 token. A token is roughly a word, but sometimes there's more than one token per word. But that's the kind of block that it's trying to say. I'm trying to make these things of this length that are typical of what is sort of seen out there on the web. And by the time you've given a prompt, the things that can follow, given that you say, I want to do this and this and this, the things that can follow are things that are determined in many respects by the prompt that you gave. So that's kind of the big picture of what it's trying to do. It's trying to mimic the sort of typical stuff that it's seen that humans have written out there on the web and in books and so on. And it is completely remarkable as far as I'm concerned, that anything like that is as good at producing sort of interesting essay type text or poetry or whatever else as it is. It's remarkable. It's something that I think is really a big clue in terms of the way that we should think about processes about human thinking and the nature of meaning and text and all this kind of thing. Nobody knew that this was going to work this well. And in fact, sort of the precursors to modern ChatGPT didn't work as nearly this well. Now there are lots of things. So in terms of how does it actually work? That's the big picture of what it's trying to do mechanically. How does it do that? I can tell you something about that, but just to talk a little bit more about sort of the construction of what's going on, when we make an English sentence, we are partly. We do it according to the grammatical rules that exist in English. Like we might say, you know, the. What's a good example? My gosh, the. The crocodile ate the. What do crocodiles eat? Ate the fish. And the. And then, you know, that has a, as a structure of an English sentence. The crocodile, that's a noun phrase. Ate, that's a verb. The fish, that's another noun phrase. And so we know that English is set up grammatically as something that says, that has the form of things like noun, a subject noun followed by a verb followed by an object, noun and so on. That is the grammatical structure. Now, of course, we can have sentences that mean absolutely nothing. We say the moon, you know, fished the elephant or something. It doesn't mean anything. What does that mean? We can have even More meaningless sentences we can say, I don't know, we can say things where the categories of thing that we put together don't really fit. In addition to things that don't make sense because they're not sort of physically possible, there are also things where the category of thing isn't really right. You could say something like, you know, the house ate the lettuce. Well, houses don't eat things. And it's sort of a. Or you could say the, you know, the chair was happy. Chairs don't have feelings, so far as we know. There are all these ways that sentences can be grammatically correct in terms of their syntax, in terms of the parts of speech and so on, but they're not semantically correct. They don't have meanings that we can immediately identify. Now, if you're writing a poem, it's a little different. A chair could be happy in a poem because poems are a more elaborate, more abstracted form of communication than ordinary text. But this question of what fits semantically together is an interesting question that we really know very little about. And ChatGPT is kind of showing us that there is probably a way of thinking about sort of what fits semantically together. Just like there's a way of thinking about what fits syntactically together with parts of speech and nouns and verbs and grammars and so on. And so there's, you know, I think it's a, it's a kind of, a very interesting kind of wake up call for the analysis of kind of how meanings get constructed and how, how that works. It's interesting that, you know, I've been interested in this stuff for a long time, and I, I've actually been sort of shocked recently. I, I, you know, what's been studied about this, and I, I've certainly known things that have been studied in, in the last few hundred years. And I kind of recently realized that most of what's talked about actually goes back to Aristotle from more than 2,000 years ago, talking about kind of the way that you construct things with meaning. And he had a very sort of primitive way of thinking about that. He also had a very different view of kind of how the world is put together and how one could talk about things using science and so on. And it's time for a rewrite, a reboot of that. And that's a project I'm hoping to do fairly soon. But in any case, let's come back to the Mechanics of how ChatGPT works. Okay, so it's got text as a prompt. It's trying to continue that text in a statistically reasonable way. Okay, how does it actually do that? Well, the thing it's doing is it's training a neural network to figure out how to do that. Let me explain how that roughly works. So a neural net is kind of modeled on the theory that we've had for the last, well, kind of hundred years or so about how brains work, that brains have these neurons. And neurons can be either like that. They can be, you know, active or not active, electrically doing something or not doing something. Every millisecond, every thousand times a second, roughly. Neurons are like, oh, am I going to fire and generate an electrical impulse or not? And a typical neuron might have a thousand connections to other neurons, little dendrites, little tentacles that sort of stick out in these nerve cells in our brains. We have maybe 100 billion nerve cells in human brains, and each one might have a thousand connections to other nerve cells. And so the idea is that when we think what's happening is that there is electrical activity, one nerve cell, it's kind of spreading its activity to other nerve cells, spreading to other ones, and that's producing this kind of whole pattern of electrical activity in our brains that corresponds to our process of thinking. And we think that memory has to do with the way that there are the connections between nerve cells, the synapses, the connections between the, the things that stick out of the bodies of the nerve cells, the ways that those can be, whether those connections exist, whether they don't exist, whether they're strong connections, whether they're weak connections. We think that memory and learning has to do with the making of those connections. There are different ideas in detail about how those connections get strengthened and made and what produces that process of learning. And roughly, we have sort of a short term memory that, that works for a few minutes and then gradually over the course of a day or so will form a long term memory. And there's actual protein synthesis that happens and so on there. There's even very recent theories about how there are kind of dark synapses that have been kind of lying around waiting to be activated and then sort of come up, come to be active when we form memories and so on. There's some uncertainty about how that all works, but roughly, the idea is there are all these nerve cells, all these neurons, they have lots of connections to lots of other nerve cells. And there's something to do with the weights of the connections that encodes the memories. When we think, when we use our brains, what's happening is that Neurons activity from one neuron is spreading to others through these connections, through these synapses and so on. Artificial neural nets are modeled on the same idea, except then instead of being actual nerve cells, there's just things in computer memory where we just have. Or in Wolfram language, it's just some piece of, some expression in Wolfram language that represents every one of these nerve cells and represents the value of that nerve cell. And then there are weights by which this nerve cell is kind of how much effect does one nerve cell have on another nerve cell? So you end up with these big, I mean, mathematically, it's matrices or tensors, which say, I've got this whole vector of values, this whole sequence of values, and I'm going to determine from that a new sequence of values. And roughly what ends up happening is you add up, you say, this is the input sequence of values. I'm going to take some, I'm going to weight those values in different ways, I'm going to add them all up. And then there's kind of a thing where in addition to just adding the numbers up, you have some kind of threshold. The most common thing that's done in current neural nets, a thing called relu, which is actually a very simple thing. It's just if the weights where you add up all these numbers can be either positive or negative. If the resulting added up thing is negative, it's zero. Otherwise, it's just the value that you got. That's a simple way to do this sort of thresholding effect in actual brains. It probably works a bit differently, but that's kind of the basic idea. So the way it works is in an artificial neural net and probably in an actual neural net in our brains, there are this succession of layers. And you're saying, oh, we give this collection of activity to the first layer, then it does this transformation, it produces a collection of activity for the next layer, keeps on going layer by layer up through the, through the system and eventually gets to the output. Now a question is, okay, and so then the issue is, well, okay, what should all those weights be that go from the input to the output? And so then you want to construct different kinds of functions that you can compute. So for example, one thing might be on the first layer, you're giving a bunch of values that correspond to pixel values from an image. And what you want is it goes through the network, many layers, you know, 50 layers or something. And out at the end you should have something where in the output there's 2,000 possibilities that correspond to sort of a thing with 2,000 possible values. And one of those values, let's say, is going to be big, and most of the other are going to be small. And those 2,000 things, you should say they correspond to 2,000 kinds of objects that might exist in the world. Cats and dogs and elephants and chairs and tables and so on. Each one of those things corresponds to one of those objects. And what you want is a. Is a process where you give a bunch of pixels in, you go through 50 layers, and then at the end, the last layer has as output something where most of the values are close to zero, except there's a high value for a few things. And it's like, okay, that corresponds to a picture of a cat. So what you're trying to do is you're trying to, in that particular case, for a neural net that's trying to identify images, what you're trying to do is say it starts off with a bunch of pixel values. Red, green, blue. Is this for this pixel, 10 million pixel values or a million pixel values or whatever for an image. And then it's like, go through these 50 layers, going through all of these things where you've got all these weights and so on. You might have 10 million weights or something, you know, in different. Arranged between these different layers. And eventually out at the other end, you're going to get something where you have a sequence of values and where they're mostly zero, except for the ones that correspond to the things that it thought it saw. Okay, so the big question now is, how do you set up all those weights so that the neural net actually behaves this way? And so that is where you have to train the neural net. How do you train the neural net? Well, you might start off by just having completely random weights. And then you feed in an image of a cat and it says, oh, that's half elephant and half turtle or something, and it's like, oh, that's not right. Okay, how do you improve it? Well, you say, that's not right. That's pretty far away from what is right. There's what's usually called a loss function. It's like, how far away from being right are you? And what you try to do is, you say, okay, you're pretty far away from being right, but if you just change those values in this direction, you'd be getting closer to being right. And so then what you have to do is say, this is what we know in the output. You got an elephant and a turtle, but you really just wanted it to be A cat. So that means the elephant and turtle values are big, but you really wanted the cat value to be big. So you can think about that as the numbers for elephant and turtle should get smaller, the numbers for cat should get bigger. That provides you kind of a way to say, these are the way I should change the numbers in the output. So now the question is, well, how do you change the weights in the inside of the neural net to get you closer to the numbers that you want in the output? Okay, so this is a giant application of calculus. And it's sort of interesting because there are many things in the world where sort of calculus became less and less useful and gave way to a lot of things that were much more about programs and so on. This is one place where calculus is still really useful because what happens is you're making use of, well, maybe I don't want to go into the technical details of this, but it's, in the end, it's making use of the chain rule in calculus. Operating on, you know, in calculus, you might study a function of, you know, one variable, two variables, three variables. In neural nets, we're dealing with functions of millions of variables. And what you're trying to do is you're trying to say, okay, you want to back propagate. You want to propagate backwards from the way you want to change the output. What do you want to do to these weights in the middle so that if you were fed that input again, you would be closer to getting the output you want? So what happens is you bash these neural nets really hard. You do a trillion times. You do this process of saying, hey, what does this input give as output? And, oh, that's not quite right. You need to push it in this direction. Let's propagate back from how we needed to push it. Let's get these weights changed in the middle. And you keep doing that, and you do it over and over and over again. And eventually the thing learns in the sense that the weights that are arranged in the middle actually do do the thing you want. So the most common form of training, well, in this setup is so called supervised training, where you say, here's a picture, it's a cat, here's a picture, it's a dog. And you do that in typical image identification training in the early days of this one used like 25 million images of a total of maybe 5,000 kinds of objects. And so what you do is you keep on saying, oh, here's the picture of a cat. You should say, it's a Cat. Oh, you're saying it's an elephant push in this direction, et cetera, et cetera. So you just keep doing that over and over again. Nobody knew how hard it would be to do this. There's a big discovery in 2011 that it was possible to do this. People didn't know. It was a month of CPU time. Somebody left something just running, and at the end of it, it turned out it had managed to learn a bunch of things. Nobody knew that was going to work. Nobody knew how hard it would be to bash the neural net to be able to learn these things. But in any case, that worked. And the thing that happens. So there are many, many tricks in how you actually do this training. And, you know, for example, you might not have enough images. You could synthetically make images, say, well, this is a cat. But we want to add to the cat image something where we just change the pixels around a bit, and it's still a cat. And then we use that as another piece of the training. You know, in some places where people are, for example, let's say, training self driving cars to recognize different kinds of objects and so on, there's a certain amount of data you get from just driving cars around or having your fleet of cars from some car manufacturer that's kind of going around and seeing what it sees on the streets and getting all that data back and using that to do more training. But in the end, there might not be enough training data. And in the end, I think what's been done there is people make essentially video games for the cars to drive in artificially. And that's kind of the way that. So that they're kind of driving around millions and millions of hours of sort of synthesized experience on roads. But that's a typical example of kind of this sort of mechanism for training where you say, this is what I'm going to feed in. This is what I want to get out. Now learn in the middle of the neural net to do that. Okay, so in the case of ChatGPT, the thing we want to do is we want to say, here's a piece of English text. It's a big long piece of, I don't know, you know, you got it out of off some random web page, or it came from Shakespeare or whatever else it was. It's a piece of English text. Now you say, let me show you, neural net, the first 10 words of that text. And I want the neural net to arrange itself so that it will correctly tell me the 11th word, the 12th word, et cetera. Et cetera, et cetera. Okay, so that's the task that you're training the thing on. So what you want to do is kind of bash the neural net. You want to run that enough times with enough different pieces of text and sort of bash the neural net hard enough that it will eventually learn correctly to predict. And you can say to the neural net, oh, you got the wrong word there, let's poke you in this direction. And then you'll get a different. Then you'll get closer to the word you should have got based on the training of the actual text that you are shown. Now, the whole point, the thing that's exciting with something like ChatGPT or with so called generative AI in general, is that it is often being given inputs that it's never seen before. Let me give the example of a cat and dog identifier first. So let's say you show it a very specific photograph of a cat. Then you could imagine you're going to train it to say, okay, that specific photograph of a cat, it's a cat. But now let's say you modify that cat a bit. The cat has been, you know, I don't know, licking its whiskers or something, or whatever, whatever. Or the cat has its ear in a different orientation, or the cat is on a different background, or the cat is, you know, put its paw up in some way, you still want to recognize it as a cat. And the big point about these neural nets and so on is that after you've trained it on enough cat images, it kind of gets the idea of roughly what a cat image is. And so even if the cat had its paw up, it can still say, oh, that's a cat. Okay, so the point with ChatGPT is that it's saying if there's a particular piece of English or text and it could know, yes, that is exactly how it should continue. It's a famous, you know, to be or not to be. That is the question, you know, whether it is nobler and whatever. You know, the Shakespeare speech, right. It could know exactly what the next word is in that Shakespeare speech. But what if you gave it something that had never seen before? It's going to try and generate something that is reasonable to generate, but it's never seen that particular input before. And the thing it's going to generate is something that's never been seen before as well. So I will say, by the way, one thing that is important in ChatGPT is that sometimes it's like, I've seen it before. I know exactly how to continue. It turns out that that doesn't produce very good results. So what's done in ChatGPT is it has what's called a temperature where it knows the most likely way that this is going to continue is blah, blah, blah. But actually it knows the top five ways it might continue. And instead of always picking the most likely way to continue, it, it has a little randomness that's sometimes picking the less likely way to continue. And so that's an important piece of the sort of the creativity of something like ChatGPT. I think its parameter is, I think it's set to 0.8 sort of the I'm certain I'm going to pick the top thing would be one. So it has a little bit of uncertainty there and that's important. And if you give it no uncertainty, at least the earlier versions of things like the earlier GPT systems, if you give it no uncertainty, if you say always pick the top choice, it very often gets in a loop and it'll just keep saying the same thing, the same sequence of words over and over again. The thing that's totally bizarre is that in the end what ChatGPT is doing is it's adding one word at a time, basically, even just sometimes one token, one piece of a word at a time. And it's producing this essay that seems like, oh, it makes perfect sense, the whole essay makes sense, but yet the way it actually worked was just it's adding one word at a time. And that's really telling us something about sort of the structure of text and things like this. Well, okay, how does it actually do the. Adding one word at a time? The. Okay, let's see the thing. There is an idea, things called transformer nets. Okay, so the thing I was telling you about, for example, an image. When you have this image and you've got these sort of artificial neurons and they're connected to other artificial neurons, you're kind of flowing through layer by layer. Typically for an image, the artificial neurons are arranged in kind of an array that corresponds to the geometry, let's say two dimensional geometry of the image. Those are usually often called convnets, convolutional neural networks that are arranged in this kind of geometrical way, actually very similar to these cellular automata that I've long studied. But any case, the way that works is it has these layers of artificial neurons, you go from one to the next and so on. Okay, so there's a slightly different idea that's used when you're generating sequences when there's a, as I say, for, for dealing with images, you start off with your first layers of artificial neurons are arranged in a grid, just like the pixels in the image. Actually, it's very similar to our visual cortex. Our primary visual cortex at the back of our brains also has neurons that are arranged in a way that somehow corresponds to the way that the cells on a retina are arranged in that kind of geometrical structure. Okay, but so in, when you're generating a sequence, the idea is that you'll have something where essentially the thing you feed into the neural net is something that is going to be. You're trying to add to the sequence. And what you're doing is you're looking back in the previous things that are already in the sequence, and you're taking the things that are already in the sequence and you're saying, okay, which numbered things in that sequence should I look at to feed it into my neural net that works out what the next thing should be. So the typical thing, the so called attention mechanism, is you're trying to add the next word. So you're going to look back and you're going to say the word five ago that was important, that was a verb or something that was important. The word 15 ago, that was also important, important. That was some subject of a sentence or something like that. So one of the things you try to do is to learn sort of which word is worth looking at in the kind of preceding part of the text. So that's kind of the attention mechanism. Then what you do is what you're basically doing is given a set of words, you're figuring out which words are important, then you're feeding those words into something which is very much the same kind of neural net that I just described, where you're just saying, there's this, you're going to feed it in and then it's going to produce. Okay, you're. Okay, I got to explain a couple of other things to make this make sense. Okay, first point is you are. What starts off is a bunch of words. The cat sat on the mat. Okay, what neural nets deal with are numbers, basically collections of numbers. And you know, is it 5.7, is it 5.8, et cetera, et cetera, et cetera. So how do you turn the cat sat on the mat into a bunch of numbers? So in a first approximation, you just say, okay, the is word number one, cat is word number, 714sat is word number such and such. You essentially just make up numbers for words. Okay, Turns out That the thing that is really useful to do is to make up not just individual numbers for words, but for every word you have a whole collection of numbers. I think in ChatGPT it is 16,000 numbers per word. I'm not sure of that order. It used to be maybe a thousand numbers per word, but these are. Okay, so what that is for every word you're saying. I'm going to characterize this word by saying it's 0.7, 2.2, 4.6, minus 0.8, etc, etc, etc. Thousands of numbers. Okay, what do those numbers mean? Well, those numbers are some kind of way of saying what the meaning of the word is. Let me give an example. So let's say you're going to. Let's take an example. Not with words, let's say it's with flags. Flags of the world or something. You want to arrange the flags of the world on a piece of paper so that flags that are similar looking will be in the same place on the piece of paper. So I don't know, the French flag and the Italian flag, I think they're similar. They'll be in a similar place. Maybe the. Oh, I'm, I'm doing badly here on my flags. But, you know, some country that has a big green flag is, is really in a different place from, you know, the U.S. which has a, you know, red, white and blue flag. Things like this. And the, you know, the, the, the, the flag of the, you know, the US flag might be a bit similar to the UK flag. Just, you know, how. What, which one is visually similar to what? So you try and arrange these on a piece of paper according to visual similarity. And so what you can then say is if you've done that successfully, actually, we could forget flags. We just do letters of the Alphabet. So lowercase letters of the alpha get printed like an A is similar to a G, a V is similar to a Y, things like that. We arrange those in similar places on the page. In Wolfram language, there's a function called feature space plot that does exactly this. That will take a bunch of things and try and arrange them according to their features in some way so that things which have similar features are nearby. Okay, so the result of that is that you can take anything like it's a V, it's a Y, it's an American flag, whatever it is, and you can place it somewhere in this feature space and it'll have some coordinate, it'll be at some position, it'll be, you know, X coordinate this, Y coordinate that. And so the point is that you try and arrange it so that the things that are similar in some sense have nearby coordinates. They appear nearby in this feature space plot. Now in the real case, you're doing that not in two dimensions, but in many thousands of dimensions. But anyway, the result is you're trying to get these things where every word is encoded by this array of numbers so that words that have similar meaning are encoded with similar collections of numbers. That's usually called an embedding. Word embedding. It's sort of the way that you embed words in, let's say, thousand dimensional space. And those embeddings, those learnt embeddings that you deduce from looking at. So the way you learn these embeddings is to, for example, say this word appears, you know, okay, so words like for colors, for example, might be somewhat nearby in the embedding. And you can tell that because if you look at tons of text, you can substitute red for blue in different places. So there's a sort of a framework for the sentence where could be a red, could be a blue. There we kind of know that those things, red and blue have to be somewhat similar because they get arranged in the same places and sentences. And so that sort of tells us something about their similarity. Meaning, you could swap crocodile for alligator because they appear in similar sentences. And so crocodile and alligator will be near in meaning space. So that's the way that you deduce these embeddings, these mappings from words to collections of a thousand numbers and so on. Those embeddings are valuable things. That's actually something that people nowadays sell embeddings because they. What is the good of having an embedding? Well, it means that if you've got a sentence, for example, well, in the end you can do these embeddings not just for words, but for sentences. And you say, I want to do a search. I want to find, is there somewhere in my documents where I have a sentence that's like the sentence that I have now? Well, if you could just take every sentence in the document and turn it into this vector of numbers, this collection of numbers, and say, are these numbers close? Then that will tell you, do you have a sentence somewhere in a document that's close to the one that you're looking for? Okay, so there are a bunch of tricks. In the case of ChatGPT, there are a bunch of tricks that are used. So for example, you can say, well, the words come in a sequence. The cat sat on the mat it's kind of inconvenient to have those words be actually arranged. The cat sat on the mat. Rather, there's a positional encoding so that you basically include with the. The you include in that big vector you include information that says oh, that was the first word, the next one that was the second word. It's kind of a weird little positional tag that gets included. So that allows you to take all those words and basically put them in a big vat. They're just a bunch of words, bunch of vectors that correspond to these words. Just put them in this vat. Then you use this attention mechanism to decide which of these words you are going to sort of arrange in what order and how much weight are you going to put on each one. Then you feed that into this neural net. Then you get out the sort of prediction for the next word. Now, okay, of course it's actually more complicated than this. There are like the typical thing that's happening is for every word you're going through many iterations of this attention feed forward, it's called feedforward. When you have this thing which just sort of feeds through a neural network, you're doing, I think it's 16 layers maybe, maybe it's more than that in ChatGPT, some number of tens of iterations of that process. Now that's what happens when you're trying to, when you're running CHAT GPT and it's trying to figure out the next word. When you're training it, you have to do sort of the whole thing kind of in reverse and you have to kind of feed it. You have to say, well, what do you think? You know, GPT, what do you think the next word is? Oh no, no, you're wrong. Let's tell you a different word instead. Okay, so that's the main part of the training, is that there's an additional part of the training which is tricky and which is actually one of the things that really I think has separated ChatGPT from the previous generations of so called large language models. Things which are sort of models of how text is generated. But let me. Well, okay, so the first step is sort of the main training of the thing. So I think in ChatGPT it's using an underlying language model that has about 175 billion parameters. So over the course of CPU years, GPU years, more to the point, you are sort of gradually refining those 175 billion parameters to represent the statistical characteristics of all of those billions of pages on the web and millions of Books and all this kind of thing. Those 175 billion parameters encode the kind of the information of what is typical text like that you see out there on the web and so on. Now that's to be compared in our brains. Maybe we have a few trillion synapses between our neurons. So the number of parameters in ChatGPT is not that different from the number in a brain, although it knows about a lot of stuff that is much more obscure than even people like me know about, you know, who have decent human memories. I think the. And there are a lot of tricky issues about, you know, is it really, does it really need precise numbers? 0.2785644, etc. Etc. Etc. That really doesn't matter, I think. I don't know how many ChatGPT is using, but it's probably just using about 255 levels for each number, rather than having a precise number that can be any number of digits. Okay. So that's the basic thing is there's this whole attention mechanism and it's trying to continue the text. Okay, there's another important piece which is so called reinforcement training, reinforcement learning applied to ChatGPT. So people ran these chatbots and it's like, well, what's the chatbot going to produce? Well, sometimes a chatbot produces crazy stuff. It's off yakking about, you know, just gone completely bonkers. It's, it's, it's, it's off the deep end in some way. And so what was done was that people were told, you know, have a chat with this chatbot and rate it, tell it where you are for the deep end. Is it going in the wrong direction? Oh, no, no, that's great. That's bad. That's great. Et cetera, et cetera, et cetera. So groups of people actually did this and turns out that's a very powerful. It turns out it's again, not so obvious that that's a powerful thing for directing what the output will really look like. I think there's probably also a certain amount of templating that's been done for particular kinds of, you know, make a poem, make a, make an essay, things like that. That's sort of just engineering on top of, on top of other things just to make a better user experience for the system. But the thing that, that is sort of important is that you have these groups of people and they're like telling it, do this, do that, and that's then fed into this. That's, that's fed into the kind of training of the neural net, there's an additional layer that is this reinforcement training that kind of keeps it more or less on track. That seems to be important. People can obviously, it can obviously be very controversial to say, well, you know, you've had some group of people that are training it, that are really keen on, you know, I don't know what the, the critical character of, you know, why. I don't know, they're really keen on, on this feature of the world or that feature of the world or their, they're really big on promoting the, you know, the rights of cats over dogs or whatever else it is. And so that set of people that did that final training, independent of just feeding in the big pile of stuff from the web, the people who did the final training have a lot more control over it's going to like cats, not dogs. And so that's sort of an issue. And there's kind of a bunch of questions about how that should be done and what this new generation of kind of, of AI wranglers, you know, how one sort of gets the right groups of people and how one sort of sculpts things in, in the right way there. But that's a, that's sort of a different issue. But, but that's the, that's the basic idea. So I think I, I think I more or less went through kind of how, how Chat GPT works. And what it's always doing is it's generating a word at a time. Sometimes it's subword and you can see that it makes up words because it has tokens that are not quite a whole word. And so it can end up gluing together those two tokens to make up a word that's never been seen before. And so it is. Yeah, so, so it's doing this. Everything it's doing, it's taking the prompt, the original prompt, it's saying, what's the next token? What's the next word? It's giving you that, it's taking the prompt plus the token, feeding that back as, as another thing and saying, okay, now continue that. And it keeps doing that over and over again until you get your complete object out. And I think as I say, it can, it can go about 20, 48 steps doing that. That's the distance back that it's kind of attention system looks to make to keep coherence in the text. So, you know, if it was talking at Blue Wolves at the beginning, if you ask it to go too long, it'll be talking about you Know, orange, you know, elephants or something at the end. And it will have completely lost the thread of what it was talking about because it only has sort of a finite look back. That's, that's how far the training could be, could be made to work. It's okay. Well anyway, so that was a basic explanation actually that went better than I thought it might go. I thought I might get dragged into a bunch of deep technicality that, that I think I managed to skirt around there. Let's see. So Brady's asking, could you say that chat GPT is made of graph of the space of words or ideas? Yes, implicitly it has. It's not so much a graph, it's more just placing things on a piece of paper or something, except it's a thousand dimensional piece of paper to say how things are related to each other. It's more so than you could take that arrangement and you could say, okay, let me join the nearest neighbors for example. And then you would make a graph out of those things. But it doesn't start off with a graph. Let's see. F NARP is asking, saying, like to see a thesaurus based on vectors and semantic space. So you could give it a word and ask it a word close in meaning but heading in a certain direction. Yes, interesting idea. People have tried to do some of that. I don't think it's properly developed yet, but it's something we will certainly see more of. I think that sort of the writing assistant idea, you know, thesaurus, you know, people like Roger in the 1850s or whatever came up with this idea of let's make a thesaurus where we group words together by similarity of meaning. So you know, if you're saying, well, I can't quite remember that word for, you know, aficionado. I can't quite remember aficionado or something and. But I know some similar words. You kind of look up the similar word in the thesaurus and it's going to give you the word aficionado. And you can imagine doing that by saying, well, I remember this word, give me words that are close in meaning space. Now this idea of meaning space is kind of interesting because there are words that have many meanings, same word in English, but it has ended up having many meanings. Sometimes those meanings used to be, you know, together like a site, for example, S I T E, you know, that used to mean, you know, a building site or something like that. Now it's, you know, as it often means a website. And originally the word site was just sort of a place where you put things. And it had one meaning. But now I would say that you've really broken off the building site and the website meaning have really separated historically, linguistically. And so that's a place where, if you looked at sort of the history of the meaning space, the evolution of meaning space, the word sight would split in two in terms of where it ends up in meaning space. And so some words also come because it just happens. Two words of very different origin just happen to be the same word that gets spelled the same way and they again are in different places in meaning space. So you can imagine saying, take a word like sight, for example, and, and try and say where in meaning space are the typical uses of that word. And what you see, and we've done some experiments on this, you see the different meanings of a single word sort of break up in meaning space. You can kind of see that happening. And so yes, you can imagine this kind of, this kind of space of what we mean by things being this thing. Now, now it's a little tricky because people sometimes say, oh, it's a space, you know, it's like you can place things here and you make a line from here to there and you can keep going in that direction. So a very famous example is that the line that goes from king to queen. Oh, now I'm forgetting the. Well, okay, the line that goes from man to woman in meaning space is in a certain direction. Now let's take that direction and let's start with the word king and say let's go in the same direction. What do we get to? Oh, we get to queen. Doesn't really work that well. And it doesn't work that well because the meaning space is undoubtedly not space like we're used to with coordinates. And it just moves around and, you know, we're adding values to coordinates and things like that. Almost certainly meaning space is something much more complicated. It kind of reminds us of the, of our physics project. And the actual very low level structure of physical space is also more complicated. It only is something which emerges as something that seems like roughly three dimensional space to observers like us who are operating on a scale very large compared to the underlying atoms of space. So the question of what the analogous kind of overall structure of meaning space is, it's very interesting question. I don't think anybody has any idea what the answer is yet. Maybe I'll try and work on it and figure it out. But the, but, but at the level, you know, you might think directions in meaning space are meaningful. I don't think they're that meaningful because I think the space just isn't of the type where it's not a linear enough thing, it's not a Euclidean enough space that you can really identify those directions in that kind of way. It's something more complicated, perhaps more like the kind of graph structures that we think are underneath this kind of network structure that we think exists in physical space. If you go down a trillion, trillion, trillion, trillionth of a meter in physical space. But yes, that's an interesting thing. Okay, Jamie is asking, how well does Chat GPT handle slang or figures of speech? Does it understand text? Does literal. Is it capable of picking up these notes? You know, nobody really knows what it can do and what it can't do. You have to just go try it. I mean, basically, if it's seen that before, it's often shocking in what it can figure out and sometimes shocking in what it can't figure out. People are surprised. Oh, it can't do arithmetic. Well, it has no computation capability. It's just saying, does it sound right? So to speak. In fact, the way it deals with numbers, it has these tokens. The current ChatGPT, I think, takes numbers with digits in groups of three. It kind of treats them like pieces of words. And it's like you go try and figure out how to do long multiplication with numbers, where you're just remembering how groups of three digits work. It's hopeless. But that's not what ChatGPT is about. It's about the statistical continuation of text. It's very different use of computation from the one that, for example, Wolfram alpha or Wolfram language is doing. Where we're really. What we try to do there is to have a precise representation. We might be talking about the same kinds of things that the text in ChatGPT is talking about. We might be talking about cats and dogs and elephants and alligators and this and that. But for us, those aren't just words somewhere in meaning space. Those are things that we can compute definite things from. You know, if we have a city, I don't know, London or something like that, for us, that's a thing with a particular location, geolocation on the Earth with these values and so on, we can compute. If we want to know, how do we go from London to New York, how far is it? We're computing based on the shape of the Earth and so on. We're using, you know, elliptic integrals or whatever you need to use to compute that distance and so on to figure out figure those things out that's a completely different kind of use of computation from the one that ChatGPT is doing, where it's just saying, does this sound right? Does this sound like something I've seen before? It's kind of, I think the distinction between sort of the way that people typically thought about thinking before the 1600s, when really mathematical ideas about kind of the way that physics and so on work came in. Before that time, people were like, well, let's just think it through by pure thought. After that time, it's more like, well, we can actually have the systematic formal abstract system where we can compute things and where we can build this big tower of computation. That's the tradition that Wolfram language and wolframalpha and so on live in. It's a different tradition than the ChatGPT one, which is one that's really much more based on the kind of pure human thought type thing. Pure human. Does it sound right? You know, I have to say, as I make up sentences and I go yak, yak, yak here I. I feel a little bit like this CHAT GPT system in some sense because I'm, you know, I don't believe that these sentences are constructed kind of that I have in my mind, as I begin to say a sentence that I have in my mind, the whole thing constructed by a computation. I think it is more this kind of thing where I'm. I'm kind of. I've got some basic idea and I'm following a direction and then my brain, my language center in my brain is just producing more and more stuff sort of following that direction using the one something that is sort of an iterative process of the kind I've described for ChatGPT. But so, you know, there are things where you'll see interesting analogies that it makes that came probably because statistically, well, maybe it had just read that analogy before or maybe, and this is something that will probably happen when you have later generations of this kind of technology. It will see that there's an analogy. This way of thinking about things relates to this way of thinking about this kind of thing is. Has the same pattern of thinking, the same pattern of talking about it as this as a way of talking about something completely different. So you'll see these remarkable analogies that it will make. Doesn't happen quite this way yet, but I think it will. These remarkable analogies. Oh, this thing that's in economics can be applied to biology over here because the pattern of thinking, the pattern of talking about these things is similar, even though Sort of the words are different. That's something that I think will be kind of a wow for the future for these kinds of systems. Let's see. So, William, comments? Could it be the ChatGPT isn't training, it's training is text rather than the real world. Should it use real world data such as sensory information of sight and sound and all those kinds of things? Yeah, I mean, you know, one of the next things that's coming is video generation. And video generation is a bit more physicsy. I mean, in text we can yak about anything we want. You know, we can make up all kinds of words and poetry and all sorts of other things that are just, you know, just, are just words that arranged however we arranged them. When we make a video, we're not synthesizing typically every pixel in that video. We're taking that video somehow, usually from the physical world. And the physical world has laws of physics that determine, you know, we, we drop this book, it's going to fall in this way, and all the pixels in the book are going to fall in this way. So yes, I think training from sort of the real world in videos, for example, is an interesting thing. And implicitly the system will learn the laws of physics and whether these kinds of, you know, attention layers and, and transformers and things, maybe they'll even be useful and maybe they're for, for learning physics, like laws. You know, if you've shown it, you know, a bunch of videos and then you say, okay, extrapolate, I want to know how a spinning black hole works, works. It's not going to work. It's going to make the same mistakes that we humans make because we have been trained by seeing everyday physics and we don't know, based on everyday physics, how something totally bizarre that's never been that, that, you know, doesn't really follow everyday physics, even though it might follow the same underlying laws of physics, how that works. All right, I think I'm supposed to go back to my day job here and I see some other interesting questions. Let's see, maybe I'll just see if there are any others about, about ChatGPT specifically. And then the question from Skeptic here, with all new AI tools, what do you think will be the effect on truth and facts as we know them? You know, I got to believe people are going to care more about facts. And actually it's pretty good for people like me. You know, I've spent basically my whole life building this kind of computational language to represent things in a kind of precise, formal, factual Buildable on, solidly buildable on way. And I have to believe that the value of that is only going to go up as a result of the ability to basically produce things that kind of sound right, but aren't. So, for example, if you have a scientific paper, you can now easily synthesize a scientific paper about, for example, blue wolves, which, you know, might very well pass muster if you feed it to humans. You say you're supposed to review this paper. Is it, Is it, you know, is it valid? Oh, it sounds perfectly right. It's talking about the blue wolf, and it's got this and that and the other, and it's, it's citing. It's making out. One of the things I find really fun with ChatGPT is it will cite sources, it will say, according to the, you know, Madagascar Department of Geology, blah, blah, blah, blah, blah. Well, actually, there isn't a Madagascar Department of Geology. Well, maybe there is, but I don't think there is. I just made it up. It made up something plausible that is statistically plausible, but it's just not real in the actual world as it is. I think that now, obviously, if it was writing fiction, it's all good. If it's trying to write things that are factual, it's a big lose. But that means that you write the scientific paper, it's citing all these sources, it's got all these things, it's got these Latin names, but it's all nonsense. It's just pure fiction made up to be plausible but not real. Well, if you actually want to do science and build on that and so on, well, you need a solid foundation there where things can be real. And our computational language is, I think, the best way to have sort of a, a broad, solid foundation. So you say, oh, there's this statement in this, in this thing, or there's this picture that got synthesized. You know, I didn't talk about image generation, but that's a whole another. Another area. I haven't actually seen too many scientific paper images that are generated by generative AI, but that's certainly something you could do. And it'll make some lovely plot. That's just complete nonsense. But if what you say is, well, I want to see the computational language that was behind that plot, you click it. Okay, there's a piece of Wolfram language code. It's a precise computational thing. It's a precise piece of notation, a bit like math notation. It's something precise and computational. Or you can go run it on a computer and, oh, by golly, it actually produces the picture. I can read the code. I, as a human, can read that Wolfram language code and see what it says. I can understand what that sort of computational idea is from that piece of code. I think that there's going to be, one will see increasing value to that kind of notion of what's the computational language behind just the words, Words, words, words, words. I think that's a sort of a thing that will be noticeable. What are the kind of formal taggings of these things that are just said with words? It has been the case that by the time somebody. It has been a good heuristic that by the time you see a beautifully written explanation of something, it's probably sensible. That's been a good thing to guess in the past because by this time somebody went to the trouble and had their act together enough to write that beautiful exposition, that beautiful exposition. They probably, you know, they probably also put effort into making sure the exposition was correct. But now the cost of making a beautiful sounding exposition just went close to zero. And so you can have a beautiful exposition of complete nonsense. And so, you know, the, the, the heuristic changes. You know, if you see something that's beautifully written, you no longer assume somebody actually bothered to work out whether it was right. It could be just complete nonsense. So you kind of shift the way you look for kind of the, you know, where you look for, you know, is this nonsense or not from the heuristic of is it well written? Did you know, did somebody go to the trouble of writing it? You have to shift it somewhere else. And I think computational language is, as far as I can see, really the only place we have to shift that to. And that's. So I think that's a thing that happens to be something I've spent decades on. So it's kind of cool that its importance seems to be enhanced by the advances in generative AI and so on. Let's see, it's a question from gist here. Does ChatGPT have the ability to mimic emotion? Does the fact that Chat GPT sort of mimics emotions mean that it's able to feel things and, you know, you know, I could just say it's a bunch of artificial neurons that are grinding out the next word, you know, that are applying all these neurons and grinding out the next word. Does it feel anything? Well, guess what? We're a bunch of neurons that are grinding out the next word. Do we feel anything? Interesting question. So One feature of ChatGPT is it really has two phases in its existence. Its Training phase and its execution, its so called inference phase, where it's generating stuff. It doesn't right now have the. Well, it has a little bit of this actually, but it doesn't have on a major scale. It doesn't have the ability to kind of learn on an ongoing basis from what it's doing now. I think part of the reason for that, the way that neural nets, artificial neural nets work right now I said it has 175 billion weights. Those are stored in computer memory or whatever the way computers work is they tend to have a CPU and it maybe has a few sort of channels of activity in it. But relative to the memory of a computer, you might have a terabyte of memory in a computer, a trillion bytes of memory in a computer, but your CPU is just operating on this, I don't know, a few hundred bytes at a time. It's doing this tiny little peephole of dealing with the whole memory of the machine. And so the way that neural nets are trained right now is based on the idea that you are just looking at this people, you've got this big area of all this memory, but you're just making changes in this very local area. Our brains work differently. Our brains, the hardware of our brains is such that presumably we can make changes anywhere, all the time. We're kind of a continually active memory as opposed to a computer which is mostly memory that's fixed, this little CPU that's running around effectively making small changes. So the current architecture for training neural nets is still very much this sort of peephole type thing where you're making the small changes at a time. So I think that that's, that's a difference. Now by the time you've got this, this kind of artificial neural net, you know, AI, generative AI system, whatever that is learning from its experiences and it's building up this whole rich life experience. And then you say, well, you got this whole thing and it's gone to all this trouble. It's been wandering around the world. It's embedded in a robot. It's been, you know, it's been, it's been seeing the sights of the world, it's been having all these experiences. It's storing all this information. And now, you know, what do we think about that? How do we compare that to a brain? Interesting question. You know, do we feel that that thing has as much kind of, you know, do we, how do we ethically feel, for example, about that thing? A whole different topic. But I think this question of sort of what is the inner feeling, I think in the end. Well, this is again a complicated philosophical issue which I thought lots about. And this will take us far afield to discuss the question of what it's like from inside a computer. What it's like from inside another brain. It's like I have my brain and mind and I have this one thread of experience that's me. And I see other people and they seem to be having threads of experience that are sort of similar to mine, but I'm not inside there. I don't really know what's happening. All I can do is kind of poke at it and, you know, peek at the results as I, you know, say something to them and they something. Say something back to me. It's, it's all from the outside. It's not an experience from the inside. But as you start thinking about what is the experience from the inside for a computer or something like that, I think it's more similar to the experience from the inside that we have as individual minds than you might think. Okay, Okay. Benny is asking about video and learning from video. Do I think that one will be able to learn kind of how stuff works just by watching videos, or do you think you'll actually have to put Newton's laws or something, actual physics laws into the system? I don't know the answer. My guess is for a lot of kind of everyday video it will be completely adequate to just use kind of experiential video. My guess is that just like for us humans, there are situations where people say, oh, look, you know, you've got this top that's spinning and suddenly it turns itself upside down. You know, likely the system won't. It's not going to know it's going to do that. That's something that is a complicated piece of application of laws of physics. And just sort of the everyday observation of a top is like it just does what it does and it doesn't suddenly tip itself upside down. So I think that there will be things that, yes, you could get them correctly. There were things that unexpected to us humans that, you know, for people like me who've spent a lifetime kind of sort of absorbing physics and so on, some of these things where you see something weird happen, it's like, that was weird. And for me, I can sometimes say, you know, well, I know the underlying physics of that and I know why that weird thing happened, but that wouldn't have been what intuitively I think just on my everyday experience of things would be the way that things work. I mean, just like 500 years ago when people said, you know, what do you mean the Earth is moving around the sun? The Earth isn't moving, I'm right here, nothing's moving here. And you know, you have to kind of learn the physics to get to realize that the things that are your immediate sort of sensory impression aren't right. Let's see. Oh, William is commenting. How does CHAT GPT experience time and so on? How does the neural net in general experience time? ChatGPT's notion of time is incredibly simple because it's outer loop is basically just, you're gonna, you've got these tokens, make the next token. You've got these tokens, make the next token. It's, it's in a sense having an external driver of time. I think that for brains, the, the driver of time is more. It is this computation happened, this next computation happened. It is our realization that we figured out more thoughts happened. Those thoughts correspond to computations in our brains. And it's the passage of thoughts happening that is the way that we assess the kind of, the passage of time. I think it's an interesting question, not relevant yet for ChatGPT and so on. This question of whether there's a sort of single thread of experience, which is one of the things that's critical to our sort of human thinking. The AIs that we're building right now are very much single thread of experience type AIs because their goal is to interface with the world as we humans experience the world. It would be amusing and it's something I spent a lot of time just looking out into the computational universe of possible programs and seeing what's out there. What do those things do? Most of what they do is very non human and non human relatable. Just like processes in the natural world, whether it's the weather or something like that, is a process that does things, but it isn't necessarily human relatable what it's doing. And so that's this question of, I mean the AIs that we tend to build are human relatable AIs and so they tend to have this sort of single thread of experience built in. When we have more AIs from the wild, the random AI that we didn't specifically engineer, that was kind of the natural AI, so to speak, then it may not have single thread of experience, but it may be a very non trivial thing to make the bridge from kind of the, the natural AI to how humans think about things. It's very similar to what happens with the natural world. The natural world does what it does. We have to make this bridge to human experience. How do we explain what's happening in the natural world? How do we tell a story in physics or something about how the natural world is working that requires that we actually do science to make that bridge between the natural world and our thinking about the natural world world? And I think sort of, you know, AI from the wild, so to speak, will have the same issue. We're already having the issue. I mean, there's. There's surely going to be a profession coming up of sort of AI wrangler, AI psychologist. I think Isaac Asimov had already sort of had one of those in his robot books of kind of the, the, you know, I'm an AI psychologist, and I can understand why did. What is ChatGPT going to do if you feed it this kind of thing? Get an intuition in our human brains for what's happening in its artificial brain. And I think that's a lot easier for something like ChatGPT that's been built to be specifically aligned with things like the single thread of human experience. By the time you're trying to be the psychologist, so to speak, for this AI from the wild, that's more like one of these simple programs that I might study or some of these things that come up in our physics project. Well, that's what it means to be kind of a natural scientist. And that's, you know, that is the thing that people like me have spent a very long time trying to get intuition for how things in the general computational world, not the computational world that's specifically aligned with humans, how those general things work. And I will say it's hard. I mean, my sort of motto about those things is when you, when you're doing these experiments, these computer experiments, you're looking at this computational world of programs that you just sort of made up at random. For example, my typical motto is, the computational animals are always smarter than you are. That is, these things you make computationally, they always do things you don't expect. In other words, I've spent much of my life trying to develop an intuition for how these computational systems work. But even so, the one thing that I know is that I know I'm not going to be able to guess what they're going to do. This is, in a sense, this is this phenomenon I call computational irreducibility. This is one of its manifestations in sort of practical experience. Right. Well, I should go back to my day job here and finish some things there. But thank you for that, all those interesting questions and so on and look forward to chatting with you all another time. And I, I'm, I will say one of the things that I am doing is packaging up sort of all of the text that I know I've produced in my life so that I can train a bot that will at least locally kind of yak like me, so to speak. It's about, well, current estimate is about 15 million words of email over the last 33 years, 34 years maybe now, about 7 million words just in the last year or two in live stream transcripts, things like this, another 2 million words. I think of things that I've written as in my writings website and so on over the last decade or so that are, you know, written for publication, so to speak. So anyway, it'll be interesting to see, I mean I will have probably, I've got, oh maybe altogether on the order of, I don't know, 30 million words of things I've said. And the question is, given that that's sort of superimposed on top of general knowledge of the world and English and so on, how, how good, you know, how successful will that be at saying things like I do. I'd like to think that us humans still have a bit of an edge, at least until that, you know, modifiable, globally modifiable memory thing is a reality from electronics. And of course the scary thing at some level if we're proud to be human so to speak, is that our human neurons, maybe a thousand times a second, they can produce another bit of information. For our computers, it's a billion times a second. So the, it's a computer is a million times faster. It just can't do that across all of its memory in the same kind of way that we can. So it'll be interesting to see. I am curious, When you deal with humans and talking to humans and you think quickly, you think slowly, maybe those differences are very small, you change that by a factor of a million and you think a million times faster. What does that really look like? What does that feel like? What does a conversation feel like? If you say well, you're going to think a million times faster but you're still talking to that slow brained, millisecond timescale human and you've got to scale back down from the million times faster to the human, do you in fact end up with something that's pretty much the same thing or do you end up with something very different where in that moment of conversation, you know, a lifetime of figuring out things has happened, those kinds of things are possible. And I. I certainly think, you know, if I look at things I've done in. In my life, it's sometimes taken me decades to figure things out. Factor of a million, faster. Well, could. Could happen in much less time. All right, I've got to. Got to go here, so I will thank you and see you another time.