Science & Technology Q&A for Kids (and others) [Part 126]

Hello there. Welcome to another episode of Science and Technology Q and A for kids and others. And I see all kinds of questions saved up here, so let me dig into them. So we have one from rbs. If you ask an AI the exact same question several times, will it give the same answer or will it change based on some random function? So, well, the different kinds of things people have called AI over the years that people have talked about artificial intelligence, which kind of dates back into the mid-1950s, but the most recent kind of popular version of AI in things like ChatGPT, those are so called large language models or LLMs. And let me answer for that case. So what ChatGPT is doing in all cases is you've given it some textual prompt, you've given it a certain amount of text, and its goal is to keep filling in more text. You say, give an essay about such and such, and it will then start writing the essay that, that it thinks would have been found on some document on the web or something where it said at the top, you know, here's write an essay about this. And then there's the result from the essay. So it's always trying to complete the text that you kind of started it from. And question is, well, how does I. Let's not talk in detail about how it does it. But essentially what's happening is that it has been trained once from the content of the web and books and so on, maybe a trillion words of text. It's been kind of trained on the way humans typically create text. And its goal is to complete text, use, give it using the way humans typically write text. But what it actually computes through its neural net, it has all these things that represent. The neural net is based on kind of computing values of things. It has just a whole lot of numbers. It turns all the words you give at the input it turns into numbers, numerical representations. And the output it will produce is a giant list of the probabilities for essentially all possible words in English, actually all possible subword tokens in general. Sometimes they're complete words like the, sometimes they're just pieces of words for longer words. But so its output is the probabilities for essentially all possible words. And then the question is, well, that's its output. At each step it it tries to complete your text, and at each step it's trying to add one more word to the text that's produced. And so what it knows is what's the probability for that word it adds to be any of the possible words in English. Okay, so Then the final question is, okay, you know the probabilities for all the possible words in English, which one should you pick? So for example, there'll be one word that is the most probable word to come next. Then there'll be the sort of second most probable word and so on down to the 50,000th most probable word. Okay, so the way these models are set up, usually there is a parameter called the temperature which is used to kind of decide whether you pick the most probable word or sometimes pick the less than most probable word. If you use zero temperature, the way the model is defined is it will always pick the most probable word at every step. And actually what tends to happen is if you do that, the essays you get tend to be kind of flat. That was particularly true. You know, ChatGPT is based on the GPT generative pre training neural net version 3.5 from OpenAI and in its case if you set it to have zero temperature, it will produce this text that's sort of okay, but it's a bit monotonous. If you go back to GPT3 in an earlier version and you set it to zero temperature, it will often get so monotonous that it's like exactly repeating words and exactly repeating sentences and it just sometimes gets into a loop and just keeps saying the same thing over and over again. With GPT4, even at zero temperature, it does a bit better. It's a bigger neural net with more probabilities and so on. In doesn't tend to get quite as monotonous. But so at zero temperature, the neural net will always produce the same results. Given a prompt, given a certain input, it will always generate. The succession of words that come next will always be the same in the typical way these things are set up. As soon as you're at non zero temperature, that won't be the case because as I said, zero temperature, okay, working out what the probabilities of all possible words are, that's just determined by having all these numbers flow through the neural nets. Then the question is, well, which word do you actually pick? Well that's something that involves deciding sort of determining from probabilities what you're going to pick. Zero temperature, you always pick the most probable word. It'll always be the same most probable word every time you run the neural net and you'll get the same thing every time you run the run the network and the, the thing that. If you are operating at non zero temperature, you don't always pick the most probable word with some probability, you pick a less than most probable word. As you increase the temperature, the range of less probable words that you sample gets bigger. It's roughly, mathematically, it's determined by an exponential. The probability goes down exponentially as you go further into the list of words. So I think the typical ChatGPT models have a maximum temperature of 2, and at a maximum temperature of 2, you're going down into the not so probable words. You know, you're going. You might be picking the 20th most probable word sometimes or the 10th most probable word or whatever. If you do that, you get much wilder text as a matter of fact, for I was literally just doing that, testing out some new technology that we'll be announcing in the next few days that relates to this. And at temperature 2, I think even GPT4 is producing wild stuff. I mean, it's just writing weird things where it's just like dropping in weird, weird random words in different places and so on. It's very creative, but it's also very disorganized and so on. So zero temperature sometimes, maybe for writing like Wolfram language code or programming language code or whatever, zero temperature may be a good thing to pick. One of the things that was done for ChatGPT was people played around with, well, what's the best temperature to pick? And a temperature of 0.8 turned out to be the one that seemed to give essays and so on that were neither too boring nor too wild. Now, is there a sense in which that temperature corresponds with anything like physical temperature with molecules bouncing around? No, it's just a mathematical coincidence. And the fact that one used an exponential distribution like the distribution of probabilities to find molecules with certain energies is really coincidental and is just a feature of the fact that there's sort of a certain knowledge of statistical physics which makes one kind of immediately start thinking of those kinds of distributions when one's thinking about these sorts of probabilities. Now, in general, that's what happens if you have a particular neural net trained in a particular way. If you retrain the neural net, all bets are off because the training of neural nets also involves all kinds of probabilities, figuring out which weights to use and between neurons and so on. That's absolutely in the usual way training is done. That's. It's full of probabilistic. It's full of things that are just picked probabilistically in doing that. So if you retrain the neural net, it will definitely do different things. It's kind of like for a human, if you kind of relearn things, no doubt. The ways those things will be encoded in your brain are different. You know, I have to tell a story. I was just a few weeks ago, I happened to see a chap I've. I've known off and on for ages, a philosopher named David Chalmers, who writes about computation and its philosophical features and about topics like free will. The question of whether what we do is somehow determined by underlying rules or. Or whether what we do is somehow something that is freely determined under our own agency, so to speak. And I was having this conversation because he works on free will, and I've also written about free will and how one can end up, through computational irreducibility in having what amounts to behavior that is for all practical purposes, independent of underlying rules, but is nevertheless determined by them. In any case, we're having this conversation, and I think we both realize that we're talking about free will, and I'm discussing computational irreducibility, and he's saying, in my opinion, rather mystically, there is something free about the way that our brains work, and it goes beyond any kind of underlying rules. But anyway, we both realize that actually we had essentially the exact same conversation 30 years ago. And it's sort of interesting that insofar as one can think about free will, so to speak, yet for us humans, given the same stimuli, one will often respond in the same way. And I noticed for myself, actually, that there are cases where I'll be trying to resolve some question about some piece of technology or some such other thing, and the same question will come up and I'll give an answer which seems like it's an original answer to me at that moment, and somebody will point out you said the exact same thing two years ago when we last discussed this. And so usually then I say, good, then it's consistent. And probably, maybe it's correct, maybe it might or might not be correct, but at least it's consistent. Let's see, there's a question here from Noctu. Do you think at any point we will create an AI factory, like specialized AI algorithms that create other AIs which do well at some specific task? Well, I think there are several different issues here. One is given a particular, like, let's say a large language model. Can you create a more specialized large language model that does well at particular tasks? I think the answer is probably yes. It has not explicitly been done that way yet, but I think it probably will be. Let me explain something about large language models. One of the surprising things about ChatGPT is that you can talk to it about all kinds of different things. Sort of any topic that's kind of out there on the web or in books or whatever it'll be able to yak about, might say things that are factually correct. It might make stuff up, it might write fiction, it might write sort of reasonable fact ish stuff. One of the things that I certainly noticed, and it's a very frustrating thing really is like I was literally just looking at some output from GPT4 and I look at it and it's like, well, yeah, sure, it makes sense. It seems roughly right. And then I realize actually if I look in more detail, it's not right. It's very convincing sounding, but it's just not right. And in some respects it's sort of sometimes more difficult to go back from the thing that looks convincing but isn't right to the thing that one would do if one just started from scratch. But any case, one can sort of talk to chatgpt about a whole range of topics. But one of the things that has been a big surprise and not very well understood is this concept of so called few shot learning where you can essentially tell it things like, you know, you are a. Scientist from the 1600s who knows only about the things that were known at that time and blah, blah, blah, and then you can ask it questions and it will sort of stay in character and it will, it will kind of have, have organized itself to only be using that part of its kind of, one could say, knowledge base, that part of its linguistic base that is relevant to, let's say scientists from the 1600s. How that works we don't completely know because every time it produces a new word that is the next word that it is putting down in the essay it's creating, every time it puts down that next word, it's running through the exact same neural network that it always does. And the fact that it had this kind of pre prompt that said here's this thing, act like a scientist from the 1600s or something, it's using that information as part of the initial collection of numbers as it ripples through the neural net making each successive word. But exactly why the presence of that sort of pre prompt causes it to organize itself to only be in one part of its linguistic knowledge base. Not nobody really understands. It's an important problem to understand better. But if you do do this sort of few shot learning and actually in few shot learning you often also give it examples. Do it like this, do it like this and it kind of gets in the groove and then will stay in character to those examples, so to speak. One would suspect that if one has done few shot learning, for example, prompting of various kinds t to get it in a groove, that one could somehow just say, okay, chop off everything else that isn't relevant to this groove. One would suspect one could do that, but nobody knows how to do that yet. I think it's a little subtle because it's, you know, people used to think back in the 1940s, 50s, things like that, that if you did brain surgery on human brains and you. Or if somebody had some sort of injury to their brain and some piece of the brain was destroyed, that that would mean that there were just certain things that person wouldn't remember. But that isn't true. What seems to be the case is that, yeah, there are changes, but those changes are much more subtle. And it's not just, oh, I forgot everything I knew about how to speak French or something, or some very specific kind of content area. It's not like in. Well, you might think that memory will be organized in the brain with, oh, we've got this section of the brain that knows about this. We've got that section that knows about this. You might have thought that it was like that. Just like, you know, in a library, physical library, you might have books about this topic shelved in this place and books about that topic shelved in that place. I have to say, now that I think about it, I haven't thought about this in years. That even for a computer, the question of where is that memory stored? Has become a lot more subtle than it was in the past. In the past, it might have been the case that you could readily say, okay, in the memory of the computer. In the. Well, these days it would be many gigabytes, many billions of of characters of memory, so to speak. Where is the particular memory associated with knowing some, you know, some particular number or some such other thing? Where is that stored in the memory of the computer. It used to be more straightforward than it is today. If you actually look at the memory of a computer, it will be sort of. There'll be all these little sub pieces of memory and they'll all be kind of hooked together. It's a complicated thing because when you allocate memory to use for a particular purpose in a computer, you're always the way memory allocators work. And let me not get off on a big sidetrack about this. You can talk about it if people are interested. It isn't the case that it's just like, oh, Here's a big block of memory. You're going to use it for this. Oh, there's another program running. Here's a big block of memory for that. It tends to be much more sort of, in terms of the actual memory addresses, it tends to be much more, in a sense, fragmented than that. But that's even with computers when it comes to neural nets. The question of that fact that you just trained a neural net on, where is that fact? The answer is that fact might be distributed across a billion different numbers that characterize the structure of the neural net. And we don't really know how that distribution across those numbers works. It seems like, oh, it's the third decimal digit of this number is really important for that. Oh, it's, it's this part of that number that's really important for that. But it's the whole thing collected together that manages to get the neural net to kind of generate the results it gets. So I do suspect that there will be a way of saying, well, actually, even though the memory of this thing is distributed in some sense across the whole neural net and all the hundreds of billions of connections in the neural net and so on, I think it will be possible to say, let's make something which is a smaller version. Now, a thing people often do these days is to try and distill a neural net. Let's say you start off with a neural net with a hundred billion connections, you can run lots of queries through that neural net. So you can know sort of, you can't give a sort of mathematical description of what the neural net does, but you can certainly say, I ran a billion queries through and this is what each of them did. And you can use that as a way to train another neural net. So a trick people are starting to use is to say, okay, I've got a neural net with 100 billion connections. Let me see if I can train a neural net with only 10 million connections and see whether I can get it to do roughly the same as the 100 billion connection neural net does. And there's reasonable success with this. Another thing one can do is to make the numbers that are specifying the, the weights and the neural net not be precise 16 digit numbers, but something that is just a few digit number. So in any case, I think the answer is that it will be possible to either sort of find ways to sort of carve out a piece of a big neural net and say, this is the piece that is relevant for this purpose. This is the piece that is relevant for that purpose and I think potentially also to train smaller neural nets from bigger neural nets, maybe just concentrating on queries that were relevant to some particular topic area. Now there's a quite different question which is making a factory for physical objects that's very much based on AI. That's been a very long story. And factories. There are surprising number of things that get done in factories. Even assembling very high tech objects like cell phones and things like this, where it's just not worth having a robot try and do it. It's too hard to get the robot to work, get to work consistently. It's better to have people just do it. I mean one can argue about it's, you know, in the end it's a sort of an economic question. How much does it cost to buy the robot versus how much do you have to pay the people? And that's a complicated issue. Quite different kind of topic. But when it comes to robots making things, sort of AIs making things in factories that turn out to be very mundane issues which are really hard, like how do you pick something up? This is something that comes up a lot for companies like Amazon, which have giant warehouses where, you know, that sort of theory of Amazon is you can, you can buy anything through it. Well, that means those, those anythings are on shelves in warehouses and they are in all kinds of different shapes and sizes and something or someone has to pick them out of the places where they're stored and put them in the box that's going to be shipped to you. And taking an object of any shape and picking it up and not breaking the object and not dropping the object turns out to be really hard. Now, you know, in a sense we've cheated for ourselves because most objects are set up. Most objects that one can sort of worth buying from, you know, online, you know, e commerce type, type site, are things that you would be able to pick up if you got it. They're not things where you, you pick it up and it's this weird slippery thing and it slips out of your hands at the first moment. They're things that were made for humans to use. So they are things which, you know, work, set up so that you know, you can, with your hands, with your opposable thumb and all that kind of thing, you can pick the thing up. You don't need to be an octopus to pick it up, for example. So in a sense we humans have made things to make it easier for ourselves to pick things up. But even if you make a robot hand, that's still four fingers and a thumb and has a certain amount of articulation. It's still really hard to get it to grasp an object and pick it up. And in fact, sometimes the things that are best at picking things up don't look much like human hands at all. And this question of how do you learn how to pick things up is something which I'm kind of strongly suspecting is going to be solved by through things a bit like large language models, large neural nets and so on. That that will be solved fairly soon. But that's been a big impediment in kind of making robotic, lots of kinds of factory like operations. I mean, there's things where what works pretty well is to have robots that do a very specific, very defined task where the robot moves, you know, 18 centimeters this way, picks up that thing, moves it 12 centimeters that way, and you know, turns the screw to insert it in something, those kinds of things. And then you can do a little bit better than that by giving the robot eyes so it can kind of visually see. Oh, yes, did the screw get put in the correct hole? And so on. In a sense, those kinds of eyes work a bit like a control system where, okay, so, you know, if you have a control system that's trying to be, I don't know, let's say cruise control for a car. So in cars, cruise control, it's like, okay, normally you put on the gas a certain amount to make the car go a certain speed. But if the car is supposed to be going uphill, you have to put the gas on more. If it's going downhill, you have to take your foot off the gas. Maybe you even have to brake a bit. That in a car there is this control system for cruise control. For example, if you have cruise control in your car and you put it on and it says, okay, keep the car going, it's 60 miles an hour, let's say. And then there's a giant, you know, there's a very steep hill and the car has to push the gas more. And then maybe suddenly you get to the top of the hill and the car would. You don't have to put your foot on the gas as much. How does the car not sort of start zooming forward? Well, the answer is it has a controller, a control system. And what it's doing is it's sensing the speed of the car and it's saying, okay, the car's going a little bit too slowly. Put your foot on the gas more, so to speak, automatically increase the gas. So that will make the car go a little faster. And what can happen is if you do that and then you say, oh my gosh, Now I'm going 62 miles and I'm going too fast, Take my foot off the gas. If things don't work right, the car will just be going faster, slower, faster, slower. It'll be oscillating around. And that's what happens. If you have a controller that isn't set up correctly, what ends up being the right way to do it is you have a certain time delay. So you go a little bit faster, but you don't kind of instantly respond to the fact you're going faster. You just calmly respond a little bit and you don't let it get into that unstable oscillating phase. But that's a very typical form of control system. Such control systems have existed since the 1940s. And there's sort of this area of mathematical sort of engineering called control theory that is all about kind of working out when will a system be stable, when will it be unstable, when is a system controllable? When is it possible to make that car successfully go at 60 miles an hour? I mean, if you have a car that's underpowered to go up a hill that's that steep, well, by golly, you won't be able to get it to go at 60 miles an hour. And there are also questions about whether you can maintain stability or whether the system will always show certain kinds of instability. And there are lots of control systems we see all around us. For example, thermostats for keeping temperature construction constant. Temperature. That's a very common type of control system. Usually the buzz term is a PID controller is a very simple kind of controller that uses a certain algorithm to keep at a fixed, let's say, temperature. Same is true more elaborately with autopilots for planes keeping a certain heading independent of which way the wind is pushing the airplane, keeping a certain altitude. Those are things that are done, can be done with these sort of fairly traditional electronics rather than computer based controllers, by the way. One example of a more complicated kind of controller is the flight controller for a drone. And that sort of only became possible within the last decade to really have those controllers are not typically done with simple electronics where you just have a few electronic components. They're done with a computer and the algorithms are more complicated to make those work and fly in a stable fashion. Another famous early controller was the automatic windshield wiper, where it was only if you sort of see rain do you sort of operate the windshield wiper. It was an early application, I think, of transistors. I think the first controller for that just had one transistor in it. Whereas your average microprocessor these days might have a billion transistors to make a computer that we have. But in any case, in sort of factory automation, the idea of control systems that sort of keep things in some fixed way, that's a very old idea. Often even vision systems where the robot has eyes are being used kind of like a control system. Just say, oh, we just need to make sure that the screw goes right into that hole and we can't. You can oscillate on each side and you put it right in. The question of whether one can make a kind of, sort of how one makes a kind of general purpose factory with AI, it's a difficult question. I mean, a typical actual factory, even when it has lots of robots in, is non trivial to reprogram the factory. It's gotten better. I think Tesla pioneered some robotic construction techniques. All car companies, I think at this point use significant numbers of robots on their production lines. But if you say, oh, last week we were making this kind of car, next week we want to switch to making a completely different kind of car. Bringing up the line, the production line, to do something quite different has traditionally taken a while, like months. My impression is that that's time has gotten down considerably. But it's not like you just tell the robots, I think, okay, just reconfigure yourselves to make, you know, here, let me show you this new car assembly that we're going to make. Let me show you the instruction book. Now reconfigure yourselves to operate according to this instruction book. That hasn't yet become a fast thing to do. That will come. I mean, one of the things that is totally remarkable about LLMs is you can actually give them things like instruction books and have them read the instruction book and sort of figure out what to do from that. Now they don't necessarily do it completely reliably, so you have to have all kinds of safety nets and checks to make sure the thing didn't just go crazy and have a little brain glitch, so to speak, that caused it to just do something, do something quite wrong. But I think it's the thing that will be a definitely a coming thing that one will be able to basically say, I want to make this, here's the manual for how to make this. Or even just here's the, here's the thing that's been made, you know, Go ahead and put it together. You know, I don't think I've ever seen a demo yet of something like a LEGO construction where, you know, you have all these pictures of how the LEGO bits put together. And I find them hard to follow often, but that's just me. But then to be able to have an AI system that can just read those instructions, understand the pictures, put the Legos together, that's step one. Step two is just look at the picture on the box of the final completed thing and figure out how to take the components you have and put them together to make that thing that will come. I'm guessing that's not that far away, a couple of years maybe to be able to do things like that. And that will certainly help in making sort of more robotic AI enabled factories. Let's see. Looking at some of the questions that have come in about this particular point. So Spare Parts is asking any thoughts on using physics simulations versus the real world to teach robots? You know, people. There's this technique called reinforcement learning, which is actually kind of related to feedback control systems that I was just talking about. They've got a lot of the same mathematics, they grew up in two slightly different communities, and so they have a lot of different language, which is super confusing to people because you see things which are like very similar but not quite the same, described in quite different ways. But one of the ideas of reinforcement learning is you've got this robot and it can try things in the real world and see what happens and use that as its way to train itself about how things should work. In a sense, that's what we humans might do when we first learn to use our hands or something. And yes, you stick your thumb somewhere near your mouth and eventually you discover this is how to put your thumb in your mouth or whatever. And that's something which you kind of learn by experiencing things in the world. Quotes robotically so that's one approach. If you are sort of making a self driving car do its thing, you can say, well, let's try it on this kind of road, let's try it in the snow, let's try it in this case. In that case, that's sort of one possibility. The other possibility is to essentially make a video game type environment where you are supposedly successfully simulating what the self driving car would do, what its environment would be, and what its mechanical response would be to what's going on, and then train it that way. What's tended to happen is that it's just much cheaper to run endless video game like simulations for training than it is to train anything in the real world where things are sort of where a robot is actually walking around and banging into things and bashing its head and having to be replaced and this foot falls off and you have to replace it. That's all a big pain relative to just having the thing exist in some virtualized game environment. It's sort of interesting that when we talk, when we talk about educating humans about things, there are lots of purposes where it seems much better to just do it in the real world than to experience it in virtual world. But the thing about it is that's because we are sort of tuned to deal with the real world, experience things at the speed that the real world happens. Whereas if you are training an AI, the AI could be running at sort of a thousand times real world speed, training itself against many, many more scenarios. Now that's all well and good, so long as the simulation is accurate enough that you are capturing what you need to capture. If you train the self driving car based on a model of snow and it's not how snow actually works, then it's going to have learned the wrong thing. It's also the case, I think, that a lot of that training is probably in the end going to be unnecessary because what's happening is one's just getting the AI to basically learn certain rules about the laws of physics and things like this, which you could, if you were able to deal with it differently, just say these are the laws of physics, they're described by these equations or rules or whatever else. But because you're saying everything's going to be learnt by a neural net, the neural net doesn't have any way to directly solve equations or whatever else. So you have no choice but to just say learn by example, learn by example, learn by example. Whereas if you could combine that with something which was just using kind of mathematical science or something to just solve the equations, you wouldn't have to go through all that training. And that's a powerful direction that one's just starting to figure out. I mean, a lot of things we've done with, for example, the Wolfram Plugin for ChatGPT is an example of kind of combining the, the ability to do sophisticated computation with the pure sort of linguistic and common sense kinds of capabilities of an LLM. And I think that will be increasingly a story of knitting together kind of deep computation and in a sense scientific knowledge of the world together with this kind of learn by example type thing. A quick question here from Pedro, is it Computer power, then that speeds real progress in, for example, LLMs, AIs, things like this. So I think I mentioned this here before. The story of neural networks is such a strange story in the history of science and technology, because neural nets sort of modeled on how brains, how we think brains work. But they were kind of defined back in the 1940s. And you know, at that time people would say, well, we've got a few neurons, five neurons at most, that we set up with an electric circuit and so on. And then now in kind of chatgpt, there are many millions of neurons with hundreds of billions of connections between the neurons. It's very big thing. And the fact that in ChatGPT we got something that can meaningfully write essays and so on, and at what point, how many neurons did we have to have to make that work? Nobody knew. And even beyond that, there were other things that were done that were very much sort of engineering solutions of, okay, yes, you can get a ChatGPT like system to complete sentences that you start and so on, but, but it will go off track, it will lose its train of thought, it will do all kinds of crazy things. What was important there was that it was trained using reinforcement learning with human feedback to kind of stay on track. You know, yes, you're writing a poem. This is how poems work. Yes, you're writing an essay. This is what to expect. Yes, don't produce all kinds of crazy bad words when you respond to something. And the way that was done, it's a little bit like a robot learning things in the real world. What was done was that a sort of prototype of ChatGPT was put out there in the wild. Well, not quite in the wild, but people were brought in to say, okay, go have a chat with this thing and tell it. Comment, tell it. Yes, that was a good piece of chat. No, that was a bad piece of chat. Here's what was wrong with that. Here's a piece you should take out, here's a piece you should add to give that kind of feedback. And that turned out to be, I think, more crucial than people even recognize yet to the fact that ChatGPT actually worked in a useful way, that kind of direction, human direction, so called reinforcement learning, human feedback, RLHF, those steps were more important than people expected. It's kind of an interesting thing because in a sense, the people who first did that kind of human feedback for ChatGPT, their views about the world were in a sense, burnt into the very first versions of ChatGPT. So it had all kinds of political views and this and that and the other, because that just happened to be the views of the fairly small number of people who did that reinforcement training. Subsequently, it sort of got more. More genericized by having more people and. And be it by being specifically sort of trained to not go in the specific directions of the people who first trained it. But it kind of gives one an idea of kind of the role of training and the role of sort of how do you govern the AIs? Well, you know, you imagine you've got people feeding in these prompts, giving these reinforcement training to the AIs. And, you know, I think it's an interesting question, sort of how do you do government, democracy, whatever, in the land in the time of AIs. And you can sort of imagine it's just all the people in the world provide reinforcement training. All the people in the world or in a country or whatever, or in some particular community provide essentially prompts to the AI to tell it what they want it to do, each one telling it, maybe pulling it in a different direction. Then it's up to the AI to figure out how to sort of merge those together. A challenging, even philosophically sort of impossible thing to do. Let's see, Look at something somewhat different here. Kephalos is asking, what do you think about sources of energy now and in the future for developed and developing countries? The interesting question. So, I mean, I get exposed a certain amount to this because I know a lot of people who are interested in what usually gets called deep tech, which is kinds of technology problems which would be very important if they were solved. But they're hard to solve. And so energy generation is one of them. So, I mean, if we look at what's available now, well, there's a lot of energy that's sort of provided to us by all those trees that existed in the Carboniferous period and turned into oil and so on, crushed underground. And that's been for the last hundred years or so, that's been a big source of energy in the world. Then there are things where there's hydroelectric energy where you're basically using the fact that you've got water. I mean, people used, even before electricity, people had water mills where they would just use the flow of water in a river to determine to run a wheel to produce energy. But the fancier version of that is to put a dam in and to have water, sort of water. High water on one side and low water on the other side. And as the water runs through the dam, it runs through electric generators. That the flow of water spins around the wheels in those electric generators and generates electricity. I mean, it should be said that the vast majority of energy that people, I would say, care. Well, is this really true? Yeah, I think the most convenient form of energy is probably electrical energy. People find it much more convenient to sort of package things up as electrical energy in batteries or on electric power lines than to have other forms of energy. Now, there are other things. I mean, you can have, you know, gas, for example, natural gas, and you can just burn it directly to produce heat. But electricity has been a very flexible way of storing, transporting, having energy ready to be used in different ways to make light, to make heat, to run computers, to drive mechanical things and so on. It's a very conveniently convertible form of energy. So a lot of types of energy generation involve electricity. So the question is then, how do you get electricity? What are some good ways to get electricity? Well, you'll need some kind of energy that you can turn into electrical energy. So hydroelectric energy, dams, geothermal energy. The idea of geothermal energy. Well, there's different kinds. There's kind of shallow geothermal energy, and there's deep geothermal energy. The deep one is probably the most interesting case that uses the fact that the center of the Earth is really hot compared to the surface. The average temperature a little bit under the surface is like 50 degrees Fahrenheit. The temperature of the center of the Earth is many thousands of degrees Fahrenheit. There's. There's partly. There's just relic heat from when the earth was created 4.6 billion years ago. That's been kind of locked into the insides of the Earth. About half the heat of the Earth and the other half comes from the decay of radioactive elements inside the earth, producing, Producing nuclear. Turning nuclear energy into heat. So the way that deep geothermal energy works is you. It's a little bit like the way that you might drill for oil. You drill down until you get a deposit of oil. Now you're just drilling down anywhere on the Earth. And you drill a mile or two down, and a mile or two down, it's pretty hot. You're getting closer to the center of the Earth. Below them, below the crust of the Earth, which is like 50 miles thick. There's this. You go. You go into the mantle of the Earth, which is sufficiently hot that the rock is liquid there. Eventually you get to the center of the Earth where it's solid again because of the pressure is preventing it from becoming liquid. But so anywhere on the Earth you drill Down a mile or two down. And it's really hot. It's certainly hot enough to boil water. And so then the idea is you put sort of, you go drill down, you pump water down or some other fluid down on one side at the bottom of your, your big well, the water gets really hot because the earth is really hot there. And then you pump it back up and you take that, the water that you just heated and you use that to, to drive, to make kind of a. To drive an electric generator, for example. It's a general principle that you can use heat to make things like electric power. Heat is a bunch of molecules running around in kind of a random way. And each the molecules individually have lots of energy associated with them. But the heat, heat is a sort of disorganized form of energy. And something was figured out in the early 1800s was the idea that in order to make use of heat, you kind of have to have sort of a hot thing and a cold thing. And in a sense, the way that you make use of heat is by having a flow of heat from a hot thing to a cold thing. So there's sort of always a hot reservoir and a cold reservoir. So in the case of geothermal energy, what's happening is you start off with a water or whatever it is. Usually people talk about using things other than water. High pressure methane maybe is, I forget, some high pressure gas I think is better than water or steam. Yeah, there's no, it's not methane. It's another. I've forgotten. In any case, you're kind of. It's cold at the top, you pump it down, you get it really hot, and then you bring it back up again. And then you're kind of getting a flow of heat from the hot to the cold and using that to drive something that can, can make, for example, electricity. So what are some other ways to make energy another way? Well, another thing you can try to use to make energy is like the tides. The, the, the effect of the moon pulling the water on the Earth in sort of the direction of the moon and away from the direction of the moon. On the opposite side of the Earth, there are, you know, the water goes up and down. You can potentially use that like a sort of a daily version of a hydroelectric dam. Doesn't seem to work that well. Another big thing you can use is wind. That's why there are wind turbines all over the place. I'm not sure how terrific that source of energy really is. I mean, the wind turbines are extremely expensive to make. And it is the case that there is the rotation of the Earth does. And also the fact that hot air rises and so on, those effects together produce flows of air on the Earth that are wind. And there are plenty of places where there's lots of wind, and particularly over oceans and things where the wind is not stopped by features on the Earth and so on. There's plenty of wind. And so long as you don't have too many wind turbines all lined up behind each other, where one of them is kind of stopping the wind by, you know, the wind is used to turn the turbine, but that's slowing down the wind. And then you have another turbine that's sort of in the shadow of that one. So long as you don't do that, there's sort of enough wind to go around. My impression is that a lot of these methods of energy generation, It's a very detailed thing whether. Whether it really, you know, how much does it cost to construct the wind turbine, how does it. How much does it cost to maintain the wind turbine? All those kinds of things. I mean, a very important feature of things like wind and geothermal energy and so on is that it's not going to run out until, I suppose, if we try to extract enough energy from the Earth, we might be able to slow its rotation down. But we humans are just so unbelievably far away that from doing something like that, we're much further away from doing that by. We could probably work out how many orders of magnitude compared to sort of using up all of the oil or natural gas that got sort of created a few hundred million years ago, when trees were first on the land and sort of collecting energy from sunlight and then getting fossilized and turning into things like oil and so on. So all of these things, eventually the energy. Eventually you take all these sort of forms of organized energy, and eventually you turn them into what to us seems like disorganized energy, namely heat. But in the case of slowing down the rotation of the Earth, it's just not. Not even close, I mean, to the kind of thing that we could achieve. So other forms of energy. Another big one is photovoltaics, which is taking the light from the sun and turning that into electricity. And it turns out to be possible using semiconductors to directly turn the. Using the photoelectric effect. The fact where you can have a photon of light come in and it can make an electron, it can take an electron that was otherwise just happily hanging around some atoms. It can kind of make that electron turn into something that Makes an electric current. It's been a very challenging thing to figure out how to make efficient photovoltaic cells. You see these big solar farms, so to speak, with usually purple blackish type silicon solar cells. And the efficiency of those things is just not all that great. And if somebody could discover some new kind of compound, some new material for which photons from the sun, the energy that photons come in from the sun hitting it, produces more electricity, it's certainly possible that that could happen. There's nothing that is sort of a physical limitation on the efficiency of those things. And that would have a dramatic effect on how effective that, how effectively that works. So another big area of energy generation is nuclear energy, of which there are really two kinds, Fission energy and fusion energy. Fission energy comes from the breaking up of very large nuclei like uranium and plutonium. And fusion energy comes from the pushing together of small nuclei like hydrogen. It's just a feature of nuclear physics that both of those activities make energy. Both breaking apart a big nucleus and to make smaller nuclei and putting together small nuclei to make bigger nuclei. You might say, well, what's the nucleus that everybody wants to be? It's iron 56. That's the type of nucleus that has sort of the, that is the most tightly bound nucleus. It's the one where if you break apart uranium, eventually you'll get down to that, put together hydrogen, eventually you'll get to that. That's why, for example, the crust of a neutron star is probably made of iron 56 for that reason. But in any case, so in fission energy, what you're starting off with is uranium, which you mine from the ground. And very conveniently, the half life, the time for the decay of uranium is in billions of years, a couple of billion years. So there's still uranium that was created when the Earth was created, was created by nuclear processes and stars. And it was kind of locked in on the earth and it can be mined and it's still radioactive. And that's kind of the thing you start from in making a nuclear reactor. Now, there are many different layers of what happens there. You can make a nuclear reactor, make new radioactive elements that you can then use in a nuclear reactor, so called breeder reactors. And there are a whole variety of different methods for, for doing, for using, for using that. But essentially a nuclear reactor is making heat from the fact that there are uranium. Okay, so the way a nuclear reactor works is a uranium nucleus spontaneously breaks apart at some rate. The rate is quite low because it takes a couple of the half. If you Have a big lump of uranium 235, particular isotope of uranium with total of 235 protons and neutrons in it. With that type of uranium. There's a different isotope of uranium 238 which is stable, doesn't decay at all. But uranium 235, which you can separate out with great effort using things like centrifuges, you separate it out, it's then radioactive, it decays. It decays by spontaneous fission. That means that basically the uranium nucleus, after a billion years or so, just boom, it falls apart. And when it falls apart, it produces a bunch of neutrons, among other things. And those neutrons have quite high energy. And the thing that leads to a so called chain reaction is that once one uranium nucleus falls apart, it produces neutrons that then hit other uranium nuclei. They poke other uranium nuclei, and that poke causes those other uranium nuclei to fall apart. Then they produce neutrons. Then those neutrons go and make other uranium nuclei fall apart. And you have this chain reaction where you have more and more and more neutrons produced, more and more and more uranium nuclei produced. It goes exponentially in that way. If you want to make a nuclear weapon, that's what you want to have happen, is to have that chain reaction happen as quickly as possible and produce as much energy as possible. If you make a nuclear reactor, you have to have so called control rods or moderator. And what the control rods do or the moderator does is it absorbs neutrons. So all these neutrons are being produced, but you want just the right number of neutrons to be sort of freely flowing around that it makes a bunch of other uranium nuclei kind of explode. Just enough to make the power you want, but not so much that the whole thing's going to go out of control. So, you know, typically in a nuclear reactor, there'll be these control rods that are, that are pushed inside the core of the nuclear reactor and you kind of pull them out. As you pull the control rods out, the neutrons can flow freely and so there are sort of more neutrons there. You push the control rods in, they absorb the neutrons, and the thing sort of calms down, so to speak. And again, it's one of these control systems that typically operates. You know, how far out should the, should the control rods be? What temperature should the, in the core of the reactor be? When things go very wrong, the core of the reactor, so called, melts down, which means that it gets so hot that it becomes molten. And that's very bad because the nuclear reactor was engineered to be a solid nuclear reactor of a certain shape, it wasn't energy to turn into a pile of liquid uranium or something. And that's happened a few times in the history of nuclear power. And it's very bad. But in any case. So nuclear reactors have this feature. You still have to mine uranium, but you don't need very much of it at all to, to produce a great deal of power. And you can also make your own fuel for nuclear reactors from nuclear reactors themselves. What's difficult is sort of, it's all a bit, you know, it can be a bit dangerous. You have to engineer it to be so that nothing could possibly go wrong. But the level of nothing possibly going wrong, usually the tolerances for the engineering are maybe a factor of two of nothing going wrong. Whereas, I don't know, you make a plane or something. It's more like a batch of 10, I think. So to have, it takes some effort and then you have to make these, it requires a lot of maintenance of a nuclear reactor. For example, the fact that there's all this radiation, all these fast particles going around. When you have a metal container for the nuclear reactor, that metal is continually being damaged by all these particles hitting it. And so that's something you have to deal with and eventually replace. And a lot of nuclear reactors are getting into their 50 years old and more kind of timeframe. And some of them, it wasn't expected they would keep working this long because it was thought radiation damage and so on would stop them working. But there are new designs of nuclear reactors. So for example, one I was just hearing about is, well, what about if you put the nuclear reactor deep down in the ground, you know, a mile down or something like that, then the, then even if something goes wrong, it's not really a big deal because, gosh, there are deposits of uranium down in the ground which are also radioactive and so on. And even if there was a nuclear explosion a mile down, you wouldn't even notice. I don't think you'd even, I think actually it's a good question whether what kind of a little earthquake that would produce a mile down. I, I'm skeptical it would be much of an earthquake. You probably, it probably have some effect, but not much the amount. Yeah. Particularly if it was an actual nuclear explosion. A lot of nuclear tests were done underground, not terribly deep underground. And they certainly, you know, they made little earthquake like things, but they're not very impressive earthquakes. And you know, an actual earthquake releases a lot more energy even than a nuclear explosion. So. Okay, so that's on that side of, of nuclear energy, the fission side. There's also the fusion side. People have been. So the sun is powered by nuclear fusion. Hydrogen atoms, hydrogen nuclei being going so fast that they fuse together. They're going so fast because in the center of the sun, it's 10 million degrees centigrade. And that means that they're going really fast. And they have to be going fast because you have two hydrogen nuclei. They're two protons. Protons are particles that electrically repel each other. And that means that when two protons get close, usually they would just sort of bounce off. But if they're going fast enough, they get over that repulsion and they get to the point where they stick together and eventually make a helium nucleus. And so that process, when they do that, they release a bunch of energy. That's all really good. And all you have to feed in is a bunch of hydrogen plus a very high temperature. The problem is the following. In the case of the sun, the sun is held together by gravity, and there's nuclear processes going on throughout the sun, so to speak. If you just want one of these things in your backyard, you have to make something that's sort of as hot as, as the, as somewhere inside the Sun. And you have to make sure that the plasma, the sort of the gas with all its electrons stripped off, that you get at very high temperatures, is just kept confined enough that the protons can actually run into each other. And the thing doesn't just get so hot that it sort of just gets out of everything. In the case of the sun, it's all held together by gravity throughout the whole Sun. But in the case of just a little thing in your backyard, you have to kind of keep something at 10 million degrees centigrade. Well, how do you keep anything at 10 million degrees centigrade? If you have, I don't know, a tungsten, you know, some, some material that can withstand high temperatures. It, the, the maximum you're going to be able to get up to is a few thousand degrees. You certainly are not going to get to 10 million degrees. So the only way you can sort of keep things bottled up a 10 million degrees is to use something other than a material container. And so what gets used are magnetic fields. And that leads to this idea of plasma confinement fusion, where you are kind of, you're trying to keep this plasma at 10 million millions of degrees centigrade. You're trying to keep it confined. And there are these different very elaborate kind of shapes to a very common tokamaks and stellarators are these very weird shapes of magnets that try to keep this plasma sort of all bottled up to the point where you can get a release of energy from the protons running into each other. And recently, in a slightly different way with lasers, there was an example of being able to release more energy than you put in in nuclear fusion. But there's been, it's been a long running story. When I was a kid, people said it's just around the corner, you know, nuclear, nuclear fusion power will be here, you know, in just a few years in the early 1960s. Well, it's still not here. And it's turned out to be a very difficult problem to confine these, these plasmas. Now, there are just recently, the last three, four, five years, there's been a lot of interest in doing this. A lot of companies started, some of them with different ideas about how to do it. Different configuration of the magnetic field, different materials to make the magnets, to have things associated with magnets and so on, different computer control systems for the magnetic fields, all those kinds of things, all in an effort to make fusion power, which, if it works, will be just great because it's something where there's no lack of raw materials. You just have to have hydrogen, and then it's sort of free power for the world. Now, there are a few other things on the horizon. I mean, there are a few less common cases. So one of them is so called cold fusion. So this was the sort of hot fusion is what happens in the sun, happens in these various devices people are trying to build. And that involves you've got all these protons, and they're millions of degrees centigrade, and they're going fast enough, they're just going to fuse with each other. But there's the possibility that fusion might be achieved some other way without making everything very hot and about through. When was it? Thirty years ago, there was a claim that cold fusion was possible, and that in particular an element called palladium was sort of a big part of that. And palladium turns out to absorb hydrogen. It dissolves hydrogen, hydrogen in palladium. So palladium, you have this metal thing and you start putting hydrogen into it, and it just keeps on packing in protons, keeps on packing in hydrogen nuclei, and you can just keep putting in tons and tons of hydrogen, and the hydrogen kind of finds its way in between the atoms of palladium and so on, and somehow a piece of palladium will just absorb a huge amount of hydrogen. So the thought Went, well, maybe it absorbs so much hydrogen that just by its crystal lattice. By putting those, those, those hydrogen atoms in this crystal lattice, you can pack them in so much that eventually they'll sort of get close enough that fusion can occur. And at the time, people claimed this a bit, and then people claimed it wasn't true, and there was a big sort of flap about it. And the end result was people sort of stopped really working on that. But there have been a small number of groups around the world that have continued to work on cold fusion. It became kind of a, a, oh no, that can't possibly work. It's nonsense and it must be a fake experiment if you're finding that it works. My own guess is that something like that will work in the end. How it will work, why it will work, I'm not sure we know. But I think the very fact that you can just pack so many hydrogen atoms into some of these materials and they're kind of arranged in this crystal lattice, it kind of makes one think that somehow it will be possible to get them to be sort of close enough, sort of maintained by the crystal lattice that a fusion process could occur. Don't know how, don't know when. If that worked, it would again be a very exciting kind of free power for the world type of story. I think the thing to realize is, oh, there's just a lot of features of materials where things are happening at a much different scale than you expect. Like when you tear something, when you break a piece of metal or chalk or something like this, you might say, well, it just breaks. But it turns out there's enough energy in the breaking of some of the atomic bonds that you produce a tiny flash of light. You even produce X rays when you just break something. Not many, but there are very high energies associated with just atomic bonds in just your average material. And I think it's also worth saying that the theoretical deduction of what should in principle be possible to do with respect to fusion, cold fusion and so on, there's no way one's going to be able to work that out theoretically, I mean, if one's very lucky, one will get it, but I think it's not likely. Similar to the way people are pushing towards higher temperature superconductors, higher temperature materials where you have perfect electrical connectivity. Nobody knows how to make even the fairly high temperature superconductors that are known. And by fairly high temperature, I mean the temperature of liquid air or something which is not that hot, so to speak. It's pretty cold even Knowing how those kinds of superconductors work is beyond what's been successfully figured out in physics. And to be able to predict correctly, oh, we're going to have a superconductor that will work at like, say, room temperature. It's much more likely, I think, to be found empirically, maybe with some indications from theory, than otherwise. Now, you know, I'll mention a few other sort of other power opportunities. Another one is, how do plants work? Well, plants take in sunlight and they sort of effectively use that to produce energy in some sense. And they do that using chloroplasts, which contain the protein chlorophyll. And chlorophyll absorbs a photon and somehow and produces sort of generates those things with electrons eventually to generate something which can be used to sort of have organized energy in the plant. That organized energy in the plant is what got bottled up, so to speak, a few hundred million years ago and turned into oil. It's also what makes wood be worth burning and things like this. But in general, the point is, if you could have your own sort of artificial chlorophyll, artificial chloroplast, that would be another potential way to sort of make energy from sunlight. Sunlight is a very. A very large source of energy. You know, the sun, however much sunlight we use, we're not going to have the sun go out as a result of it. Let's see. Oh, yes, there are all kinds of schemes for different sorts of things. I think there's one. A friend of mine is doing that for photovoltaics. This is an idea where light from the sun has a certain distribution of wavelengths. And the photons from the sun are not specifically tuned. They have a whole distribution of wavelengths. They're like white light. Like, if you put light from the sun through a prism, you'll see that it's broken into all the colors of the rainbow. Well, what if you just make light at a particular frequency, a particular wavelength? Can you make a more efficient converter from light to electricity than a general photoelectric effect thing? And the answer seems to be potentially yes. And that's very interesting, because if you can increase the efficiency, that's a way of increasing the efficiency. The not by virtue of changing the photovoltaic substance, but by changing the incoming light. And you can make light of a fixed frequency by essentially just using heat. With the emission lines from gases, for example, where the. The gas will emit light at a particular frequency and then you kind of turn that into electricity. That's an example of One. Question being asked, what will happen when oil runs out? Will there be a shift to clean energy? Well, before this happens, I'm sure the answer is yes. You know, I remember back 25 years ago, 30 years ago, I would talk to people who are in the geology business and I would say, how much oil is there on the Earth? And they would say, well, eventually they would say, look, it just depends how much you're prepared to spend to get it out. The total amount of oil that exists somewhere under the surface of the earth is huge. Sometimes it's in little pieces and rocks and you have to do, you know, fracking to get it out. Sometimes it's in giant reservoirs, sometimes it's very deep, sometimes it's in very inconvenient places under the ocean, all these kinds of things. But there was a period of time when people said, the oil is just going to run out, there's just not going to be enough oil on the Earth. And I think that just turned out not to be true. It just turned out there were more and more places that oil could be found. It was just a question of they were in very inconvenient places. And it became more and more expensive and energy consuming to actually mine the oil out. So I don't think, I'm skeptical that it will run out. The economically accessible oil that can be mined out at a reasonably low price, that might run out, but it's just a question of price. As soon as that's running out and you still need it, the price will go up to the point where it becomes worthwhile to get that oil out as well. I think there's been this big push for all kinds of different reasons. One reason is that the economically accessible oil is certainly not equally distributed on the surface of the earth. There are particular countries that lucked out in terms of having easy to access oil resources, and that's led to all kinds of effects in the kind of sort of political economy of the world, so to speak. And I think that it tends to be the case that these other forms of energy are a bit more kind of globally accessible. I mean, there's sunlight almost everywhere. Yes. If you're above the Arctic Circle or something, then for half the year you don't get to have sunlight and you don't get to operate photovoltaic cells. It tends to be the case that, for example, who. People who think about high altitude drones that would be flying at 60,000ft and just keep flying continuously, powered by sunlight. There's a pretty sharp cutoff that Above a certain northern or southern latitude, you just can't fly a thing with a certain efficiency above 50 something degrees north latitude. I think the things you can make these days just can't keep themselves aloft, just at least, at least in one design by using the current efficiency of photovoltaics. And so there are limitations there. There are places that have more wind, there are places that have geothermal energy because they're on for example, plate boundaries where yes, there are earthquakes and geysers and things like that, but there's also geothermal energy. There's places where you can readily make a dam and a river and have hydroelectric energy, et cetera. There are also places where there's a limited number of places on the earth where uranium is economical to mine. But as I said, you don't need a lot of, of radioactive uranium to get started on this sort of cycle of making nuclear fuel. So I think that that's, yeah, that's the way that works. Well, there's a question here about nuclear energy. Do the dangers nuclear energy pose a problem or do the pros outweigh the cons in this situation? You know, the nuclear reactors as they're currently built are very complicated things and complicated sort of one off engineering systems are hard to get precisely right. The number of nuclear accidents that have occurred in the history of nuclear power is not that large. It's basically, let's see, I can probably even list them all. 1956 wind scale in England, 1970 something Three Mile island in Pennsylvania, 1980, mid-1980s, Chernobyl in Soviet Union and then Fukushima in Japan about a decade ago. I think those are the, the, and each one happened in a different way for different reasons. It's really hard to work out what will happen if something goes wrong. You know, I remember to tell a little historical story for me back in the, in the early 1980s, people were very keen to sort of push the idea of nuclear power and to prove that the current designs of reactors were, you know, 100% safe. Always, always hard to prove something's 100% safe. But one of the things that had happened at the Windscale accident in England is that radioactive gas was released into the atmosphere. Not much, but enough. And at Three Mile island, radioactive gas was not released into the atmosphere. The accident was contained, so to speak. And so there was an effort to sort of prove it was all going to work. And I remember I was, I think probably all the people involved in this, in this story, except for myself, happily, are now dead. But it was a sort of effort to sort of prove the safety of nuclear reactors by showing new designs couldn't possibly fail. And there was some big commission that was being put together to show that. And at that time, I was much younger than I am now, but I knew quite a bit about sort of computer simulations of things. And so the theory was, oh, we're going to do a computer simulation of this nuclear reactor. We're going to prove that nothing could possibly go wrong. And I was like, that's never going to work. You know, you're never going to be able to say, you know, you can look at certain scenarios and things, but you're never going to be able to say nothing could possibly go wrong. So I said, okay, I'm not going to be involved in this. So then a few years went by, and then the Chernobyl nuclear accident happened. And I. So I asked people who've been involved in this thing, so, you know, if you were like, nothing could possibly go wrong. A lot did go wrong at Chernobyl, and it did release radioactive gas into the atmosphere and so on. And, you know, how did that happen? And the answer was, well, it was a Soviet reactor and it had a bad design, etc. Etc. Etc. And the American ones are much better, probably true, et cetera, et cetera, et cetera. But, you know, and I mean, Chernobyl was a very famous accident in the sense that, oh, gosh, it just produced so much, a lot of mess. I remember I visited northern Sweden a couple of months after Chernobyl happened, and there was a whole big plume of radioactive gas that had been blown by the wind, or actually radioactive dust had been blown by the wind, and that caused a lot of radioactivity in the food chain. And the reindeer were radioactive. And it's just a huge mess. The actual things that happened there, it was a failure of the control system of Chernobyl and this whole business about control rods and whether the control rods can be pulled out and pushed in and all this kind of thing. One of the things that's famous for its sort of engineering issue was, you know, you'll have a big system like a nuclear reactor, and it'll have warning lights and things go wrong. It's kind of like a red warning light will come on, but you have a big, complicated nuclear reactor, there's an array of a thousand warning lights. And what had happened was that, oh, the thing was a bit old and it was a bit, you know, sort of crotchety and so on, and it was always running with lots and lots of warning lights on. And, you know, I've often used this, this example because it's really worthwhile. You know, if you have a program and it's producing error messages and things, you really want it to run with no error messages, not just a few error messages that you don't understand, but no error messages. Partly because when you have a few and then you have a few extra ones, you don't even notice the extra ones. And it's kind of what happened at Chernobyl is that, well, it was always running with a certain number of warning lights, but then it had more warning lights and then, well, then what happened was an explosion happened and then the whole thing was full of warning lights. But I think even before that had happened, I don't remember the precise details of what occurred, but it was at a weekend, I think, and it was all a bit of a mess in terms of the chain of command and, you know, people calling people and not being able to get anybody on the phone to say, what do I do with the, you know, the operators of the nuclear reactor were saying, you know, oh, there's this extra warnings that are on, what do we do about this, etc. Etc. Etc. But any case, I think with nuclear reactors there are lots of new designs that are very different. I really kind of like this design I just heard about that involves burying the nuclear reactor deep underground. That seems like pretty good idea. It seems like a. That's one of these, really. There's not a lot one can see going wrong. Famous last words. But I think that one seems very well contained. But I mean, my own guess is that, You know, nuclear power is probably a pretty good idea. I mean, it really depends. It depends on how careful you're prepared to be and how careful you're going to insist on being. Because if you insist on being so, so careful that everything is, you know, 50 times over engineered, it will never be cost effective. If you say a factor of two over engineered, you know, maybe that works, but then you have to kind of not do anything kind of, you know, not miss things that, you know, not make mistakes, so to speak. And it's, you know, there, there are challenges there. Actually, it's interesting. Spare parts is commenting. The death rate for nuclear energy is around 0.04 deaths per terawatt hour, which is similar to wind and solar. Yeah, well, I don't know about solar. I don't know how. I mean, with wind turbines, definitely bad for birds and bad for humans who try and maintain wind turbines because they're awfully, you know, it's awfully high off the ground. I'm not sure what goes wrong with solar. I mean, maybe you have these things where you're collecting a bunch of solar energy and things get really hot. I'm not sure. I mean, maybe, you know, somebody falls off a roof when they're installing some solar thing. The Healy is commenting nuclear is safer than coal because people are more cautious when the stakes are higher. That's. That's an interesting claim. I think, you know, it's. I would suspect that with modern engineering, modern engineering design methods, that it really is possible to make much safer nuclear reactors. And there's a question here about what do I think of small modular reactors? I've not studied them in great detail. I know a bunch of people who are working on them. On, on the face of it, they seem like a good idea. I don't know the details, and I really need to get going in a second here. The. Let's see, There are questions here about, from spare parts, about connection between computational irreducibility and extracting usable energy. Can energy be mined with computation? That's a long story. And it relates to this thing called the second law of thermodynamics, which I've studied quite a lot over the last 50 years, I have to say. And that has to do with this question of whether the energy of all these molecules running around when something is hot is that useful energy or is that just heat that isn't useful? How do you line all those molecules up to be able to turn a wheel, to turn an electric generator? Something like that. And yes, the story of kind of how do you go from the apparent randomness of heat to something more organized is a story of kind of computational irreducibility. And yes, it probably is the case that there are places where you can make use of energy more effectively by using computation to know what's going on. I mean, I wouldn't be surprised if there weren't situations like many of the ones we've talked about. Well, in the case of wind energy, usually the wind is consistently blowing in a certain direction over a wind turbine. But if you were trying to deal with something where there's just a little bit of wind blowing in different directions, but you have all these sensors and you can figure out, oh, the wind is about to blow this way, turn the windmill in that direction and you'll catch the wind that way. And then, you know, do this computation, turn in another direction, catch the wind in that way. That's a Place where you can use computation effectively to collect more energy. To not have that energy seem like, oh, it's wind going in random directions. It's. Yes, it's sort of random, but you can predict it enough that you can kind of pre wire things to be picking up that energy. All right, I. Oh, Reebok is commenting. Nuclear is not a good thing until we have some way of disposing of the waste products. Yeah, that, that's a challenging thing. It's like, where do you put the waste? And I think the main issue is what was it? Yucca Mountain is this place in the US where a bunch of nuclear waste is put sort of under the mountain. And the problem is when you do that, if it isn't deep enough, it ends up being able to get into groundwater and then it seeps all over the place. The radioactivity can seep all over the place. And even this, let's bury it and put signs up that say, you know, that will for 10,000 years tell people, don't go here. After 10,000 years, the radioactivity will have dispersed, so to speak, just by radioactive decay. I kind of like, this is one reason, I kind of like this idea of just bury the reactor a mile down because, you know, things, if you have something that's radioactive and it's under a mile of rock, it really isn't going to do anybody any harm or anything. Any harm. And I'm not quite sure why the depth of, why it isn't the case that when one thinks about nuclear waste disposal, that one isn't thinking about deeper, you know, just drilling things down deeper. I mean, they're ones where it's disposed in the deep ocean disposal, whatever. I don't know why it's not just drilled down deep in the rock. There's probably a good reason for that, but it seems like sort of an obvious thing to do. But yeah, also, by the way, I mean, there are, there are schemes for making nuclear reactors where you can use many of the waste products. And that's, I think, one of the things, again, I don't know the details of these modular reactors. I should have paid more attention, but I think some of them make use of the waste products and it's sort of a closed cycle in the end, I think. Yeah. In fact, one of them. Yes, yes. There's one type where you seal it up for 100 years and it just, you know, uses it internally, uses the, the products of its, of its, of its nuclear reactions and so on. Anyway, I should stop there. Thanks. For joining me. Thanks for we had a bit of a range of questions today and I see there were many other interesting questions left over for another time. But thanks so much and bye for now.