Future of Science and Technology Q&A (March 29, 2024)

Hello, everyone. Welcome to another Q and A about the future of science and technology. Let's see. We have a whole bunch of questions saved up here. First one is from Kathy asking about eclipses. What should we know about the weather around the time of the eclipse? What if it's cloudy or raining? What will the future of predicting eclipses look like? Actually, in about an hour and a half, I'm going to be going through some things that I just posted today about the history of predicting eclipses. But let's talk here a bit about the future. First thing to say is we can predict to within a second or two when the eclipse is going to arrive at every point on the Earth. Can we predict the weather at the place of the eclipse? No, that's much harder. I always find it interesting that in ancient Babylon, there were sort of three kinds of things the astrologers thought they would predict. The positions of planets in the sky, the weather, and who would win or lose battles. And it's sort of interesting that we nailed the first one with exact science. From the 1600s on, we were kind of on a track to really be able to compute all of these things about celestial mechanics and astronomy, the weather. Well, in the last. The first computerized weather prediction was done in 1950, actually. People thought at that time, oh, it's just going to work. We're just going to be able to solve the equations for fluid motion in the atmosphere, and we're just going to be able to nail this thing about predicting the weather. That turned out to be much harder than expected for a variety of reasons. But then finally, the who will win or lose battles. That's something we don't expect, that we're able to make sort of a scientific prediction of that. But, you know, with respect to the. So with respect to the weather, the thing that made that hard was the fact that there's sort of a lot of computation that goes on, in effect, in the motion of all that air and the water vapor and so on in the atmosphere, and then the nastiness of sort of phase changes that cause clouds to form, and then rain starts dropping out of the clouds and so on. It's a pretty complicated setup. And there is this phenomenon I call computational irreducibility that I talk about quite often about the fact that even though you know the equations, you know, the underlying equations that govern the motion and phase changes of water and so on in the atmosphere, it's still sort of irreducibly difficult to work out its consequences. Now, one thing was sort of Noted starting at the beginning of the 19, the early in the 1960s was this phenomenon that, okay, even if you have, if you have a small change in the initial conditions for certain kinds of equations, then there is progressive sort of exponential divergence of the effects. So a small initial change leads to a larger, progressively larger change. People, I think, overrated the importance of that in weather prediction. That isn't the primary problem of weather prediction. It is not knowing. There's one thing to not know precisely the initial conditions, the state of kind of the atmosphere at a given time. But really there's just a lot to know about sort of the atmosphere and different places and so on. And one of the things that's happened in recent times is with, I mean, there are only, what, I don't know, 30, 40,000 weather stations around the world that are reporting kind of every five minutes, I suppose these days, kind of things like temperature, wind speed, humidity and so on. But there are increasing number of kind of satellite and radar observations that give you much more detailed, small scale kind of information about the state of the atmosphere and make it easier to predict. From that. I think the. I could talk at some length about how weather prediction is actually done. It's improved in recent times for a variety of reasons. I think in no small part because of the use of all sorts of more detailed data about the current state of the weather. Being able to get sort of the image of things like radar images of sort of the local configuration of the atmosphere and clouds and so on. And then kind of at some level, if you know exactly what things are like in a given place, it's kind of like, how do you move this forward as the wind blows things from here to there and so on. And it becomes easier to make predictions when you have this much more detailed data about the current state. And now it's routine in many places you'll have the app that tells you kind of it's going to be raining for 17 minutes or something, and it'll get it roughly right at your specific location. The thing as you go out to longer and longer time frames and as you look at different kinds of weather, thunderstorms, whatever else, there are things that become more and more difficult to predict. I think largely as a result of the phenomenon of computational reducibility, more so than because of sort of the effect of small tail changes. But any case, it's hard to predict the weather. We're just making some pictures. I plan to go to Austin, Texas to see this upcoming eclipse. So I'm sort of Wondering what's the weather going to be like there? And we just yesterday started making some maps based on satellite imagery of previous cloud cover, because that's the most important thing if you want to see this eclipse. Mean, the two total solar eclipses that I've seen, I both saw in lovely cloudless skies in 1991 and in 2017. And so this time it's a different time of year. The others were in the summer, both in the summer. This one is obviously in the spring. And so I'm sort of hoping that I'll get to figure out where to go so that it won't be cloudy within some sort of radius of where I'm going to be in terms of the future of predicting eclipses. Well, eventually there won't be total eclipses because the moon will move too far away from the Earth for it to cover the disk of the sun. But that's a long time from now, Certainly that's many, many millions of years from now in terms of the predictability of eclipses. We've now sort of got enough of a handle on kind of the effect of the motion of planets and so on to be able to successfully predict eclipses. I mean, there's a whole series of effects like in, I don't know, one part in 100,000 you have to worry about Jupiter. Then one part in a million you have to worry about deformations of the Earth, about the detailed shape of the Earth. You keep going, keep going. The one thing that's really a little hard to predict is exactly where the eclipse is going to land, where it's going to land relative to the rotation of the Earth. The reason for that is the motion of all of the motion of planets and the moon and so on that we can compute. Not easy, but we can compute it. The issue is knowing where the shadow of the moon will land on the rotating Earth is not quite so easy to predict. Why? Because the rotation of the Earth we don't. It glitches and it's the. The. Gradually the rotation. The Earth is slowing down in its rotation. When there were dinosaurs, the doors were the. The. The. The days were a lot shorter than they are today, days getting progressively longer. But the way in which it slows down, in which the rotation of the Earth slows down, is not completely predictable. It depends on things like earthquakes and glaciation, kinds of things, all kinds of things like that. And in fact, when we know the sort of alignment of time, if we think about sort of a global kind of atomically defined time, there's a question of how does that align with the actual time based on the rotation of the Earth and the succession of days. And the answer is it keeps on getting a little bit out of sync. So I think since 1970, 27 leap seconds have been added, representing the way in which the rotation of the Earth has, is slowing down relative to sort of time as defined by sort of the global clock of the universe or some atomic clock or something like that. If you look at. It's not the case that it's a pure slowdown. In fact, if you look over the last thousand years, where one can detect kind of the length of days and so on, there's been sort of negative leap seconds and positive leap seconds that have had to be added. But since about 1970, we've been in a slowing down phase, I suppose, and it's purely adding leap seconds. But I think in the last few years, I think since the 2017 eclipse, I don't think any leap seconds have been added. But if you look in the distant future, there will be leap seconds that have to be added to keep sort of global time aligned with rotation of the Earth time. But exactly when those will have to be added and how many is not easy to predict. So what consequences that have for eclipses? Well, as you, the. If you get sort of the rotation of the Earth, you don't know when that is relative to the sort of time when the shadow of the moon is going to intersect the Earth. That means that the position of the eclipse on the surface of the Earth, what geography will see the eclipse will change. And so there's a. What is it for every leap second? It's. I don't know. I don't think it's very far. I think it's like maybe a mile or a fraction of a mile deviation. But obviously that adds up over the course of time and over the course of many leap seconds. So it's a little difficult. You can predict when the eclipse will happen, but exactly where on the Earth will be in totality, for example, will depend on this rotation of the Earth, which is hard to predict. So over the course of, I don't know, 500 years, that's fairly easy to predict. By the way, I was just making some pictures showing kind of the density of eclipses. Eventually every point on the Earth is going to see a total solar eclipse. But if you look at the US for example, and you look at the next thousand years, there are a few spots. There are these lines of eclipses going across the country. But there Are few spots in a thousand over the next thousand years that won't experience a total solar eclipse. Otherwise every place will. And so in terms of the ability to predict eclipses, well, it's pretty good right now. You know, can we get it down below a second? It starts to be kind of unclear exactly what you mean, because an eclipse. For an eclipse to happen, you know, the moon is covering the sun with respect to our viewing point from the Earth. But when does it actually cover it? Well, the details of when it covers it depends on sort of the profile of the limb of the moon, the edge of the moon. And when there are mountains on the edge of the moon, there'll be these little places they call Bailey's beads where. Where the valleys on the moon are sort of the last places where light is still poking through from the sun before the disk of the moon finally covers the sun. But that's a pretty subtle effect. And there's a sort of the. What does it mean to be sort of at the beginning of the eclipse you have to kind of decide, well, how much light do you. Is it really zero light and so on? It's never quite zero light. So it's a little bit hard to define. In fact, one very bizarre effect seems to be that the result of these sort of discrete points on the moon being the last places where the eclipse kind of left before sort of darkness from the eclipse that leads to a slight sort of polygonalization of the shadow of the moon on the Earth. You don't get. The most obvious thing is that you would just get kind of a parallel version of the shadow of the moon on the surface of the Earth, which would be something very close to circular because the moon is very close to circular, just like the Earth. Mountains are pretty small compared to the overall radius of the planet or the moon. But because of sort of. When you do the detailed ray tracing, you find out that it doesn't work that way and that there's sort of straight lines of the edges of the shadow on the Earth between these different kind of Baily's bead points from the valleys on the moon. A rather bizarre effect doesn't make a huge difference less than a second in when an eclipse will arrive at a particular place. But that's one of the sort of final weird effects that one has to deal with. And well, that's a little bit on. On future of predicting eclipses. I think this is one of these things where we can pretty much say we got it. You know, sort of mathematical physics managed to Solve this one. We can successfully predict eclipses. All right, let me see. Okay. Oh, boy. There's some interesting questions here. Will we be able to move the moon to keep eclipses? Asks Emma. Well, interesting question. It slightly reminds me of a recent and rather campy movie called Moonfall that I quite enjoyed for various reasons. I won't give away the main feature of what turns out to be the case about the moon in that movie. I cannot recommend the physics of that movie, but the story is quite entertaining. It involves the moon moving relative to the Earth. Let's just say, can we move things like the moon? Well, the most relevant thing is can we move near Earth asteroids? Because like, there's I think, one asteroid that is, you know, the question is, is an asteroid going to crash into the Earth? Because that's bad, because it would lead to kind of. It would put up a bunch of dust into the upper atmosphere which would make it real cold for at least a couple of years. That's presumably wasn't real good for the, the mainline dinosaurs back 65 million years ago. Also, if you are anywhere near where the shock, where the actual impact happens, or the shock wave or the tsunamis and so on that are produced by it, that's also bad news. So generally it's bad for the inhabitants of the planet to have an asteroid crash into it. So it's of great interest whether if you notice an asteroid that's sort of on a collision course to the Earth, can you push it out of the way? And there was a recent space mission that did successfully perturb the orbit of an asteroid. And that's sort of good news. If you're interested in avoiding the asteroids colliding with the Earth, it's. I mean, there are lots of things to say about exactly how that might work. But basically the answer is one can sort of nudge asteroids when they're far enough away from the Earth. You don't have to do much nudging to change the angle enough that it misses the Earth? Of course, that requires you actually have to do the computations. Right. Wouldn't it be embarrassing if we went to all the effort to nudge an asteroid and it was going to miss the Earth, but we nudged it so that it hit the Earth? That would be really bad. That would be in the category of serious science mistakes, so to speak. But in any case, with respect to nudging the moon, that is a much more challenging thing. And it's, you know, there are science fiction stories where people put sort of rocket engines on Planets. I think there's a. What is it? Wandering Earth or something that, in which it's kind of like push the Earth out of the way, out of the solar system and so on. I mean it's worth realizing that in the early history of the solar system, 4 billion years ago or something, there are probably a bunch more planets in the solar system. And many of them sort of escaped from the solar system in a sort of wandering, rather forlorn, perhaps way in interstellar space. And so there are sort of, at least in the early formation of solar systems, there's a certain natural selection of planets and ones that sort of get too close to other ones and their orbits just get diverted to the point where they escape the solar systems. Now can one move the Moon? Challenging. I think there's sort of a. It's. Well, one thing is if you started moving the Moon, I think you could probably change its orbit while moving the Moon to an orbit closer to the Earth. The thing that prevents that happening is the angular momentum of the Moon in its orbit. The gravitational attraction of the Earth is going to want to pull the Moon towards the Earth. How do you lose angular momentum? Well, you presumably fire rockets in the opposite direction and that will make the Moon kind of move closer to the Earth. It's one of those things I can imagine in a future world. It's kind of like, well, let's see whether we can set it up so that we can have those cool eclipses by moving to the Moon close to the Earth. But whoops, that will mean that we have bigger tides on the Earth. Is that a good thing? Is that a bad thing? Hard to decide, but it's the thing you have to decide globally for the Earth. I mean the other kinds of things about the Moon and so on is sort of stuff like well, for example, in low Earth orbit we can imagine a future world in which there's kind of advertising from kind of. There was one company that was trying to do this few years ago to have sort of a light show of things re entering the Earth's atmosphere in a particular pattern that would spell out some message that everybody could see. And we can certainly imagine as more and more satellites get launched that we can imagine sort of a great configuration of these things formed in the sky. I was just at an event actually where the, there was sort of a analog of fireworks that was a lot of small drones forming a whole collection of different pictures and so on. It's kind of, I saw a pre version of that, oh three, four years ago, which Was just kind of a very basic version with just drones with white lights and so on, on them, forming a few simple patterns. But now the state of the art and those kinds of things is advanced where the drones can light up in all sorts of different colors and can kind of form kind of a very elaborate kind of almost movie like light show in the sky that's obviously very close in, but one imagines that with a lot of effort that same kind of thing could be done with satellites. I think it would waste a lot of propellant to do that, to move them around in that way. But that's what one could imagine happening, I think in terms of if there's a way to sort of put an engine on the moon to change its orbit. I suppose that. Well, there's a question. Can you make rocket fuel on the moon, for example, to be able to do that? Certainly there's a hope that on the, for example the south pole of the Moon that there's water ice, that one could use solar energy, so to speak, to break into hydrogen, oxygen and that would give one as that recombines, so to speak, that would give one essentially rocket fuel. And I suppose one could make the Moon motor that way. Not sure that it's the most important thing to think about doing, but that's the kind of thing one could do. Okay, Mark, comments. The moon gets further away because of friction and the oceans and the tides. Could we play with that mechanic? Actually I think the main thing is the solid tides on the Moon. That is that the tides on the Earth are in. I'm not sure about that actually what the bigger effect is the on the Earth, the effect of gravity from the Moon and the fact that the gravity from the Moon pulls more pulls, it sort of pulls averagely the center of the Earth. But the things on the side of the Earth that's closer to the Moon, it pulls proportionally more than the average, so to speak. And the things away from the Moon it pulls proportionally less. And so that causes what's the so called tidal deformation of the Earth, which kind of elongates it. So the close to the Moon is pulled even closer to the Moon. Part further from the Moon is kind of even further away from the Moon. And that's kind of the rough explanation for tides. Although in detail it's not adequate. That was the rough explanation that Newton gave. But in detail, working out the tides is really hard and actually hasn't been done in complete form across the Earth. But in any case, the tides, the water on the Earth is easy to pull up because it's liquid, but there's also deformation on the solid structure of the Earth. The rock of the Earth gets moved and in the Moon. The Moon has a. Moon does not have any liquid on it or even in it, but it's still the rock of the Moon is deformed by solid tides. I mean, those solid tides can be pretty dramatic, I think. Oh gosh, which is it? One of the satellites of Jupiter is it has sufficiently strong tidal forces that it kind of opens up fissures on the surface of the planet of the Moon that because of this kind of heating of the, of that moon by the fact that it continually gets deformed by these tidal forces. But. Let's see a few other questions here about eclipses, but I think I'm going to do some Q and A specifically about eclipses, so maybe I can can cover these at that time. Street comments I've seen articles about attempting to build trains on the Moon. Would that be useful? You know, I was just talking to somebody who is designing houses for the moon, essentially 3D printing structures on the Moon by using kind of laser sintering of lunar regolith. So on the Moon there's kind of this fine dust on the surface of the Moon. And the idea is to use that as 3D printer material and basically just print kind of structures that can be used to build effectively houses on the Moon. And there's sort of interesting mechanics of kind of what shape you have to print in to give enough rigidity and so on. And then you have to figure out how you seal the air in and all this kind of thing. But I don't know about trains on the Moon. I would, I would imagine that the one would have to sort of make train track. It. It sort of doesn't seem to make sense to. Unless. Unless it turns out that there are ways to get rocket fuel much more cheaply. And people certainly talk about that and they talk about kind of being able to use photovoltaics and solar energy to, I don't know, make lots of methane or something like this that one could use for rocket fuel. And then with that, maybe it becomes so cheap to move stuff from the Earth to the Moon that you can just make the train tracks on the Earth or something, move them to the Moon. But otherwise I think sort of train tracks made of metal requires quite a lot of infrastructure on the Moon to be able to produce that from local materials. Maybe there's a way of sort of centering together sort of train tracks from lunar regolith or something, the kind of sand on the moon. I'm not sure. I don't know how important that is because I think sort of we are kind of having, having sort of cars on the moon, trucks on the moon, so to speak, seems perfectly feasible. And certainly that's what the various rovers that have wandered around the moon have done. And I know there's a bigger one that's going to be sent to the south pole of the moon fairly soon. So I kind of think that's probably a reasonable transportation there. The. Let's see. Well, okay, let's. Very different topic here from coldly, should we expect really unusual emergent properties from LLMs as we scale up the data more and more? Well, interesting question. I think kind of there's a question, you know, human language has a certain amount of stuff in it, and it seems to be the case that a large language model, a neural net with a trillion words of training data and a few hundred billion weights and a few million neurons and so on, that that's sort of enough to capture the structure and common sense of that goes around human language. And yes, we could have more than that, but in a sense you could imagine, you can imagine having a language. Our human languages have maybe 50,000 words in them. You could imagine that with sort of bigger brains you could deal with more words, maybe a million words. But for us humans, that's not something that's terribly immediately useful because we're dealing with the size of brains we have. We're kind of 50,000 word like critters. And perhaps there's a scaling among different organisms that the cats and dogs with smaller brains sort of have in a sense a smaller number of concepts that they can kind of have in their brains. We have a certain number of concepts we can have in our brains. You scale it up, you can sort of maintain a larger number of concepts, maybe. What does that really mean? What does that allow one? Is there a way of using sort of the larger number of concepts to sort of figure things out and then bring them back down to Earth, so to speak, to the kinds of concepts we talk about? It kind of reminds me on a bit of quantum computing where kind of inside the quantum computer there are these many threads of history that are being followed to do different parts of the computation. But then the big challenge is at the end, how do you knit those threads back together to get to the point where you can tell us humans, here's the answer to the computation, a definite answer. And in fact, one of the Things that's a big challenge in quantum computing is to do that knitting together. Sort of confusing because in the standard formalism of quantum mechanics, that knitting process is just not accounted for. You just say you apply this measurement operator and then you're done. And there's no kind of mechanism and process by which that happens. But in reality, when you do experiments, there is such a process and people are always concerned about noise levels and decoherence and so on. My own guess is that actually a lot of really interesting science and physics is lurking in the difficulties people have in making quantum computers work. But if you kind of that process is kind of taking all those parallel threads and sort of bringing it to the point in that case where there's a single definite answer, in the case of kind of the million concept kind of internal AI type thing, bringing that down to kind of the 50,000 concept, sort of feed it to the humans, that's a similar kind of thing. And I'm not quite sure how that would work. But in terms of what one will see, I mean people are very concerned about the idea of kind of can one do reasoning? Can one do better reasoning? My own feeling is humans find reasoning fairly hard. And the main way we do it is by sort of frameworks and scaffolding it with formalism. You know, we invent logic, we invent mathematics, we invent kind of method that sort of computation and the sort of rule based way that we do computation. We invent those things so that we can sort of build up chains of reasoning. And neural nets as such aren't a particularly good architecture for that. I mean the whole technology stack we built for the last, well, I've been working on for the last, I don't know, 45 years or so now of building up this kind of computational knowledge stack and computational language that's sort of optimized for these big towers of reasoning and sort of figuring out one thing from another, so to speak. And I think what we're seeing right now is from even a year ago, more than a year ago, kind of taking the chatgpts of this world and calling as tools Wolfram Language and Wolfram Alpha and so on to do that computational part and then just feeding the results back to the LLM, which is not set up to do sort of these, these formalistic computational steps. But that's right now that's kind of the state of the art is tool calls, function calls, whatever you call them, where the LLM will kind of say, well, I'm just making my, my essay or my speech or whatever. And then it'll have something where it says, you know, now I'm going to call this tool. And, and you as a human probably won't see that because it'll just show up as some kind of UX element and, but it will be generating tokens, words that say, you know, tool call, blah, blah, blah, blah, blah. This is how the tool is going to be called. Then it goes off to some external API that, you know, the, the APIs we make, for example, that will then go and do actual computation, give the result back to the LLM. The LLM will sort of say, it might, you know, will, will sort of absorb that and include it as sort of additional prompting for the, the later sort of words that it produces. So that's kind of the, the state of the art today. There are different things on the horizon. For example, well, there's, I mean, this question about exactly how does the LLM call the figure out how to call these tools. And there's all kinds of dynamic ways to generate prompts. So that you're saying, well, this is the kind of giving a hint to the LLM about just what would be a useful way to call this tool and doing that using either LLM type technology or other kinds of machine learning to figure out, given that the conversation is going this way, here's how you should think about calling this tool and when you should do it. But sort of the thing that is certainly of interest is how much can you really mix together kind of neural net, like behavior with a neural net. When you encode a piece of information, like, I don't know, the density of gold or something, if you can encode it, you're encoding it in some totally bizarre way, spread out across a million different weights. The seventh digit of this weight and the eighth digit of that one, somehow are the things that cause LLM to say the right thing when you ask it. The density of gold. It's a very bizarre way in which the information is encoded, probably very similar to the way it's encoded in human brains. But in a sense where by formalizing things, we have much more efficient ways to encode that information. And the question is, can one sort of merge this kind of very distributed neural net, kind of little shards of information in different places with the kind of much more organized kind of formalistic encoding of information from computational kinds of systems and sort of how do you mix those two things together? Nobody knows yet. I've thought about it a bunch. So other people have thought about it too. We're not there yet in the current neural nets. Large language models, for example, are in the end very simple at some level. Neural nets, their neural nets were kind of the data of here are the words that have been said so far, goes in at sort of the top of the neural net. And there are a series of layers with all kinds of connections and weights that sort of say how much to move this number from that neuron to another. Just go through a few hundred layers of that. It's just feed forward. You just go, here's the state of all the neurons at this layer. They're all characterized by numbers. Here are the numbers for the neurons at the next layer and so on. You just ripple through the neural nets and then you come out with the probabilities for different next words. The neural net picks one of those words that becomes the sort of the input for the next round, the next generation of the next word, or more accurately the next token, which might be just a piece of a word. So that's a very sort of straightforward workflow in a sense for the way the neural net works. Now there are fancier kinds of things that can happen in neural nets, like recurrent neural nets, where instead of it just feeding forward, it's like the neuron has a feedback mechanism and certainly those exist in the brain and where that effect from that neuron, it affects a future neuron, but then it goes back and affects the neuron in the past. Traditionally it's been very difficult to train recurrent neural networks. The process of tweaking the neural net so that it does the thing you want it to do on the spec, on the training examples that you have. That is, that's something that has been hard to do for recurrent neural nets, but maybe that will now be something that's possible at least for small scale pieces of kind of recurrence in a neural net. And I suppose one of the big lessons of neural nets and machine learning over the last 15 years or so is the very intuitive thing. If you bash the system hard enough, it can learn almost anything. And it wasn't obvious. I mean, when I worked on neural nets back in 1980, 81, I looked at small neural nets. I tried to see if I could get them to learn things. After the small and simple ones didn't really make much progress in a modest amount of computer time, I kind of gave up. I had no idea that the more complicated the neural net and the harder you bash it that you'd eventually be able to get to victory. So to Speak. And there's a couple of reasons for that. I mean, when you have a neural net that is a big complicated neural net, one thing that happens is that if you're trying to train it to, say, behave in exactly this way, and you're trying to do that by progressively tweaking kind of the weights, the parameters of the neural nets, you end up with something where if the neural net is big enough, you don't get stuck because there are enough at every moment. There's sort of enough different things that you. Different directions you can go that at least one of them is going to sort of take you out of the box and take you going towards sort of more success in reproducing the training examples. And that's one thing. The other thing is that one sees phenomena like you're training the neural nets and it trains up to some points, and the loss which determines which specifies kind of how far away you are from reproducing the training examples goes down and it'll flatten off. Then it'll start going down again, and it's sort of the neural net made a discovery in effect, and let it go on training that way. But so it's certainly possible that architectures of neural nets that involve recurrence, for example, involve the neural net not just kind of feeding forward that that will be a thing that becomes possible. My guess is it will be. And that will be closer, a little closer to mixing in sort of the computation idea with the pure feedforward neural net idea. But I think there will be other ways to kind of knit together kind of what are right now, great big tool calls in the middle of the kind of ponderings of an LLM to get that more integrated into the innards of the actual operation of the neural nets. But we don't yet know how to do that. Now, in terms of training data, well, there's kind of. It's sort of, how much training data do you need to get a certain kind of thing, you know, if you've got a trillion words and is that enough to give you kind of the structure of human language as it normally is? It's kind of a thing I've wondered about for myself as I yak on in these live streams. I think there's like 50 million words of transcribed text from me. How much is that? How much can you, given a neural net or some other kind of architecture like that? How much can you reproduce me, so to speak, just by the example of 50 million words that I've yacked on about and so there's a question. If you had more than a trillion words of training data, what effect would that have on what you could actually conclude about how well you could train things? I think one thing to realize is really the training process is something where you've got this kind of awkward neural net that's not really set up in the way that is optimized for doing the computation you want. You have to bash it really hard. You have to keep bashing it with more and more training data. And in the end, once you've got a neural network, sort of the problems have been solved of how to set it up correctly. You can start pruning it out and make a much smaller network that will still get to the, get the behavior you want. You just have to have that big network because to get from here to there, you've got to kind of go on these, you've got to have these paths that exist to sort of get you to the place where you are well trained, so to speak. So I think having more training data, there is a certain amount. There's, there's more training data. I mean, you know, there's, there's about, oh, I don't know, 5 to 10 billion pages of exposed content on the Web. There's probably maybe a hundred times that much in, in content that has been, that is, in the, the deep web, as opposed to the dark web, the deep web of, of things that are sort of behind logins and things like this or internal material in the intranets of, of, of companies and other organizations. There's also the historical form of the web. I mean, there's, there's at least couple hundred billion pages of sort of historical snapshots of the state of the Web over the past 30 years or so. It's, I don't think anybody's ever bothered to figure out exactly how many unique web pages there have been. It's not even clear what that means because you have a web page and it's been changed a bit and is that a new web page? How much content is different? It's kind of a whole. It kind of reminds one of a source control system where you're kind of. And indeed the pages might have been stored in that where you're kind of including the diffs. But it's not clear, you know, what's new material, what's not. But anyway, there's more material one could train on, but I'm not sure how much that's when you get to particular areas like training on writing code, like Wolven language code. We know that having more training data of that is useful and up to a point it's definitely useful. And of course there's, there's, there's training. So sometimes the training data is just for the sake of sort of bashing the thing, harder to just get it to learn stuff. So for example, in self driving cars, a, a common thing that's now done is to augment sort of any actual, from cameras training data with training data that comes from essentially running the neural net in a video game that emulates, that has an emulation that might have itself been produced by, with a neural net and so on. An emulation of kind of driving around. And that's useful as a kind of a trick for getting the neural net to be trained correctly. Even though it's not really new information in some sense it's just augmented information. I mean those tricks have been done for a decade or more in just this idea of sort of augmenting the data and sort of image identification and so on. It turned out to be useful back in the, in the early 2000 and tens. It was useful to just sort of add noise, sort of move the image around and so on. And even though it's not really new training data, it's something that it's new to the neural net at least. Now there's vastly more kinds of training data in for example videos and sort of here's what actually happens in the world, here's what people actually say, here's how people's faces actually move these kinds of things. And that's very rapidly getting ingested into various kinds of neural net type models. And I think that will allow kind of just like their large language models, there are sort of an upcoming whole family of large behavior models relevant for robotics, relevant for other kinds of things like that, where one's trying to sort of figure out from videos of the world what actually sort of happens out there and how can one predict if this and this and this happened what will happen next? Those are definitely sort of coming attractions, I would say. Let's see the comments here. Yeah, I mean in terms of what can we expect from the bigger LLM, I think that for example, language is probably an emergent feature of brains of a certain size. The big discovery of our species of human language probably was needed sort of a cortex of a certain size to maybe to get sort of enough concepts or something and then to have the big idea of language, of compositionality, of the fact that you can not just say, you know, elephant Rhinoceros, you know, one word at a time. But you can actually compose sentences together with modular parts where there are many different ways that you can arrange them. That's sort of a big idea of language. It is an interesting question whether if you have a bigger brain, are there other kind of big ideas like that that might emerge? I don't know. I've certainly thought about it. It's what kind of abstractions kind of can you have about the world and your description of it that sort of become possible as you get kind of bigger brains. I think one of the challenges with kind of bigger brained AIs is whether the things that they notice will be things, how those things relate to what we humans can sort of appreciate. But certainly one could imagine something. I just don't know what it would be like. That is kind of the beyond language abstraction of kind of describing the world. I mean, to take a very simple example, language is very one dimensional. You say this, you say this, you say this. It's kind of tree structured in the way that kind of sort of clauses and so on are put together, phrases and such like are put together in a sentence and so on. But it's fundamentally quite one dimensional. You say this, then this, then this. What is it like? What would it be like if we had multi dimensional kind of languages on graphs where you have all these different pieces flowing? Our brains, our consciousness is very based on the idea of sort of a single thread of experience. So language in this one dimensional form is a good match for our kind of one dimensional experience, where our one dimensional experience of time, our sort of belief that a sequence of definite things happens in the world. But what would it be like if instead you had these many paths going and something. It's hard for our brains to wrap themselves around. But it's certainly something one could imagine. And perhaps that's an example of kind of a thing that would be a way of talking about things, a way of describing the world that goes beyond the things that we're used to with brains of our size, so to speak. I think that one of the things that we see in neural nets, in large language models is this idea that you can kind of get textual, you can get the textual analog of statistics about the world. With statistics, you might have a whole collection of numbers and you might say the average is this, the maximum is this, the interquartile range is this, and so on. But when you're just given a bunch of text and you're asked a question like what's the Most popular color of a tie to wear with a white suit or something. That's a thing. That isn't a. That's not really a. Oh, we got a bunch of numbers and we get mean values of things. That's something that's sort of a textual analog of that. What's the most common, you know, pattern of tie or something? I don't know what. But. And that's. That's kind of a thing where sort of abstracting from all this text about the world. What, you know, how do you summarize that? That's kind of a new thing we haven't seen before. And I don't know how that's going to scale up. I think that the. I mean, there's sort of a. Yeah, yeah. So I'm not sure. It's certainly something that's worth wondering about. And, you know, there are all these questions about how big a neural net do you need if you increase the size of the training data? You know, many people think you have to kind of run over the training data on each element of the training data about once as you train a big neural net. And you have to have about the same number of weights in the neural net as you have elements in the training data. It's not clear if that's right. And some other people say, no, no, no, you have to run over the training data many times as you do the training. Of course, if you say I run over the pieces of training data 0times in doing the training, it's like, well, why do you have that training data? But it's more subtle than that because training data can be quite redundant. Like you can have, you know, many, many copies of famous speeches or something like this in the training data. And so, you know, how much do you weight that speech by having the neural net see it over and over and over again? And does it kind of become obsessed with that speech and to the exclusion of other things? This, again, sort of relates to this text analog of statistics. It's kind of. As you layer on more things in the training data, how do you weight them? I mean, if you were doing sort of weighted statistics with numbers, we kind of understand how that works, but kind of weighted statistics with text, we don't really understand how that works. Yeah. LC Comments it seems erroneous to think it's a more data question for improving LLMs. It's the opposite. It's less data, more generalization, transferability and efficiency via integration of a symbolic framework and accelerator. Yes, I agree. And we're working on that. And I think as sort of a question, the neural nets in effect and LLMs discovered a bunch of effectively semantic regularities of language. And we're trying to make use of that to really build kind of a symbolic representation of more things about language and knowledge and do that in a very structured computational way that isn't just kind of, we'll somehow get that out of a neural net by magic, so to speak. Des comments? Human senses capture a tiny sliver of what's going on. Would additional senses lead to bigger brains or languages? Yes, probably so. You know, I was just hearing that human smell seems like it's now reasonably decoded and that just, you know, for color, we have three kinds of cone cells in our eyes. Red, green, blue. For smell, we might have a couple of hundred kind of receptors in our noses and kind of when we smell something, it's kind of going across, you know, how much not red, green, blue do we have, but how much of each of those, a couple of hundred sort of smell primitives do we have and you know, thinking in terms of smell or something? Yeah, it's an interesting question. What kind of, you know, do you need a bigger brain to do that? And also when you talk about smell, you know, as you talk about things, you see, we talk about objects, for example, we say they're these objects. This is connected to that, this is in front of that, this is kind of enclosed within that and so on. What on earth is the analog of that in 200 dimensional space or whatever, smell space? I'm not sure. And that's again something where we are kind of sort of, we, we, we our human abilities right now and our, I suppose one could imagine, sort of interesting question if, if language is the thing that allows us to describe mostly kind of hearing, vision type, senses, what, you know, what kind of a thing, and maybe it's this multi dimensional language I was talking about earlier would allow one to describe and sort of reason about. When we say reason about, what we're saying is we're formalizing things to the point where we can kind of build towers of consequences rather than just saying, oh, then this immediately implies that immediately leads to that, and so on. So yes, that's an interesting point. And certainly if we imagine feeding all the IoT Internet of Things, you know, sensors from around the world into the big, big AIs, that's again, like, what do you need to deal with that? And, and is there a way, you know, is it the case that you just learn the details of all those sensors or the details of those dimensions of smell, space or whatever. Do you, do you just. Are you just dealing with that level or is there kind of a symbolic level where you say, I've got all these IoT devices that are sending me signals, but, but I'm going to describe that in this way that might even be human, accessible, that says, I see a FLIRP coming up on the. You know, on the fact that all these IoT devices that are connected to me in some funky way are showing some particular behavior. And that then becomes a. Oh, yes, you know, flips are a common phenomenon. It's like, you know, people would say, you know, an Internet storm that didn't become terribly popular, but that was the thing people talked about, kind of an Internet storm where the routing of messages goes in some complicated way. Is that something which we can sort of identify as a definite thing out in a kind of a world that is associated with different sensory inputs than we normally have? It's a reasonable idea, and it's something. One might even be able to have some imagination about how that works because we do have some, some concept of what those additional sensory inputs will be like. Rippling is asking, is human consciousness one dimensional? That does seem to be a very important feature of our experience of the world, that we integrate things into this single thread of experience. We've got 100 billion neurons in our brains doing all kinds of things, but somehow they all get integrated to give us this thread of experience. And it's a typical feature of kind of evaluating how conscious is somebody is. To what extent are they successfully integrating these different sensory inputs? I mean, it's one thing if the knee reflex or something works. It's another thing if you're really integrating all the different sort of sensory inputs that you're getting. And that's a typical kind of neurological test for consciousness. But it's also something which I think is a, Is a practical feature of the, of. Of. Of the fact that it really is core to kind of this notion of consciousness, that we have this kind of single thread of experience. That does seem to be a very core feature which we see in our physics project. We see. This is a longer story. We see how that feature of our experience of the world kind of ties in to the fact that we believe that the laws of physics are the ones that we attribute to the universe, so to speak. Let's see just a few more things here. It's a comment from Dab. It's an interesting comment. Perhaps it's not so surprising that training AIs on games is useful when humans learn through play in early life. That's a good comment. I, I, I think that. That is interesting that that in a sense games are an augmentation strategy for us humans to kind of learn the ways, things one, one should think about things. You know, you throw the ball back and forth and that's kind of giving you a kind of augmenting your ability to understand kind of the motion of objects or similarly when you have more cognitive kinds of games. It's a good thought, probably correct. We're probably built to, we're sort of built to build our neural pathways, so to speak. Street is commenting about an article about the requirements of GPT7 will be so massive, like half the world's computers, lots of dams for electricity and there isn't enough training data. Yeah, I mean I think that this, so as I was mentioning before, sort of getting to the point where you're trained, you need a lot of effort. Once you're trained, so to speak, you can rip out a lot of those, those things that were part of what you needed in order to actually follow the path to get trained, so to speak. I think that I have to say I, I'm, I, I doubt that the current, you know, we need to spend huge amounts of power, money, etc. To, to train the AIs. I, I kind of doubt that's going to be a continuing story. I think we'll, it's just sort of, we need that right now to get there because in a sense we're getting there in a very unreasoned way. We're just, you bash the, the networks hard enough and they'll learn stuff and that way of getting somewhere, it's kind of almost like a random walk to get to where you're going to. But in the end, once you understand where you're going and how it works, you can build a straight road that gets there. And my guess is that's what's going to happen and that things will get sort of, there'll be some efforts to get there through training. It's kind of like biological evolution had to spend a couple of billion years to get here. But the actual sort of the growth of an individual organism is much easier than all of that effort of all those zillions of organisms that lived and died to get us to the place where we're getting through biological evolution, through that sort of random sequence of steps in biological evolution. And so I think it will be with AIs and so on that the Steps to get there are hard, but once you're there running it is not going to be so hard. But I think, I mean, the steps to get there are hard right now because we're doing this in this very uninspired, in a sense way of just bashing it harder and harder and harder until it learns what we want it to learn. LC is asking, what does it say about logics and set theory, type theory and so on and formal languages that LLMs can't just trivially do those things? What does it say about them? I think, well, they're a formalization and formalizations are about sort of building towers of consequences. And that's just not what LLMs are set up to do. It means that it's sort of a story of computational irreducibility that an LLM can deal with sort of shallow, narrow, shallow computations, but it can't deal with. And the sort of a big surprises that producing language is actually a fairly shallow computation that wasn't obvious. We now know that that is a thing. But it could have been the case that just producing language was itself this incredibly elaborate computational process. There might be some arguments that we managed to do it in milliseconds, so it can't be that hard. But now we kind of know for sure it's not that hard. Let's see. Well, maybe one more thing here. Swap asks, what's the best way to pass computational representations to language models, Plaintext or embeddings? It's a reasonable question. Most of the interfaces that we humans will have will be for from language because we're saying we're talking to our home robot or whatever, humanoid home robot that seems to be all the rage right now. And people say they're going to be robots that can tidy up a room in not too long. I think that's a really challenging task. But we'll see the. So your interface to that robot will probably be talking to it. You're not going to present it with a bunch of embeddings now. So in terms of the human interface, you know, words are good. In terms of kind of how you take kind of something from one trained neural net to another. Embeddings are good. In fact, that's the whole sort of concept of transfer learning. First you learn the features that are important and you get kind of, you know, here's the raw data. Then from that you deduce this kind of vector of numbers that's a feature vector, and then you do things on that feature vector. That's a kind of a standard concept in sort of putting together pieces of neural nets. And I would say that the modular approach to that is something that's a really good way to do things if you might be able to get more efficiency by taking that at first and then doing end to end training for what you actually want. But that's a thing kind of using feature vectors. It's just not relevant for the direct human interface with, with those kinds of systems. All right, let me see, maybe one more. All right, I'm, I'm, I'm being told that I need to go deal with my day job. Fun though this is, I would like to say top of the next hour I will be talking about eclipses and talking about some of the things that I posted today about the history of predicting eclipses. And please join me there if you'd like to and otherwise, see you another time. Bye for now.