Science & Technology Q&A for Kids (and others) [Part 124]

Foreign. To another episode of Science and Technology, Q and A for kids and others. I think last week we've been talking about AI and it looks like we have a big overrun of questions about AI here. So, you know, I've got something I was thinking about which is, so what's going to happen to education in the world of AIs? Because it's increasingly the case that the AIs just as AIs will be able to do things for us, AIs will get to know us better and they'll get to be quite to be better at knowing what to explain to us so that we understand things. That's one thing. The other thing is in most settings right now, people are in classrooms where there's one teacher and there's a whole bunch of students. When one's able to interact with a good AI, the AI will be able to give one sort of personalized instruction. It's kind of like being a, a one on one student and teacher. And that will accelerate the rate at which people can learn things. As the AIs learn more about us humans and how we learn and how a specific human learns, the AI is going to be better at saying, this is the one thing I need to tell you right now. This is what you're confused about. Let me explain that one particular thing to you. And so I think what's going to happen is that the speed at which we humans can learn things is going to increase. I don't know how much, five times maybe, I'm not sure. The other thing that will probably happen is, you know, something and it kind of gradually decays and you kind of forget the details. And then it probably will be the case that sort of the refresher is also aiable. And it'll be the case that you can, you know, when you start saying, well, I need to remember about this and the AI can ask you a couple of questions and then it realizes you've forgotten that part and it's going to feed you back that part and then that helps jog your memory, however that works. We don't know how that works for humans exactly. Maybe we'll learn a bit how it works from looking at AIs. Actually, that's one thing to point out that's interesting about ChatGPT and the large language models, LLMs and so on, is that the types of mistakes they make and the types of things they can do are very human. And it's something where we haven't really had a theory of why do people not Manage to solve that math problem quite correct. How do these people get a little bit confused about this or that? Probably by studying LLMs, we'll be able to get an idea of how those human mistakes are getting made and potentially what you tell the human, at what moment you should tell the human to do something in order to avoid the mistake. There are many fields in which people have gradually learned how to get humans to not make mistakes. For example, there are plenty of fields in which there's checklists are an important thing. Kind of like, did you remember those 10 things you're supposed to do before you fly the airplane off? Well, actually it's more than 10 for airplanes, but this idea of let's have a checklist and you specifically go through this, this, this, this, this checking it off, that's kind of a mechanism for preventing one making mistakes. I know in my efforts at programming, I've done a lot of programming in my life and one of the things I learned early on was the types of mistakes that I make. And so I kind of. So gradually I got to be, well, I know what type of mistake I make. So let me be careful when I've written this piece of code, let me be careful to check it for that type of mistake. I'll say another thing. When I write text, my absolutely most common mistake is leaving words out. And I think that the. But. So given that I know that when I'm reading through what I've written and so on, I'm alert for the fact that, oh, it might have a word missed out, including, you know, the word not and things like this. So I think that's the kind of thing where as we kind of, you know, learning the types of mistakes we make is something that is a very doable thing for AIs. And being able to see this particular LLM makes this particular kind of mistake, I don't think any of the LLMs will make the leaving words out mistake of the current ones at least. I'm kind of wondering if you overclock it so that you're insisting that it produces tokens at a very rapid rate. Maybe you can get it to leave words out. And maybe that's, maybe that's my brain's problem, so to speak, when I write text and leave words out. I'm not sure. So in any case, let's see, there are a bunch of questions here. My gosh, lots and lots of things. Okay, lots of interesting questions. There's a question here about the future of jobs. When we're able to provide Basic things to everyone with machine like, you know, food, water, shelter and so on. At that point, will jobs be obsolete or not? This is a long running story of history. It's when is enough enough? That is, there comes a time when, you know, in some developed country or whatever else, in some, at least some part of the country, it's like getting the food you need to eat is not a big problem, or getting, you know, shelter is not a big problem. At earlier times or in different places, maybe that is still a problem. But you know, for some large class of people, it's not a problem. And then the question is, well, then just sit back and do nothing then. Well, it's not the way our species seems to be built for better or worse. It's like people always want to do more, you know, when, when, if you can get the basics taken care of. For many people at least, it's kind of like there's more stuff that one wants to do or that or that in a sense needs to be done. I mean, it's kind of like, you know, a good example is houses. If you say, you know, what was a house like 300 years ago, okay, it had walls, it had some, maybe some windows, it had, you know, it had rooms, et cetera, et cetera, et cetera. But now if you look at a house, what's in its walls, it's got all kinds of stuff. It's got all the electrical wiring, it's got plumbing, it's got network cables, it's got all kinds of stuff. And the expectations for what a house should consist of have considerably increased. And which is why, for example, construction, even though many things can be pre manufactured, many things are more automatic, have more machines and so on, still takes a certain, a significant amount of time. And I think that's the thing that happens all over the place is expectations go up. It's not enough just to have sort of basic food to eat. You know, if you're going to eat, I don't know, some, you know, bread and something or other every day, oh, that's not enough. You have to have something more than that. And that tends to be the thing that I think is a driver for, you know, the idea that you don't get even if you can automate to some, to some level, that automation then provides you the possibility of doing more things and then there's more things for people to do. And I think what can also happen is once something has been automated, then sort of more of it happens and that can generate issues that necessarily Require people to deal with them, at least until one has understood those things well enough that one can then automate to the next level. So I think it won't be the case that jobs will be obsolete. I think one thing that you do see in economies, you see it even in today's world, but you see it also historically is as economies get more developed, the number of different kinds of jobs that get done dramatically increases. Like in the US 150 years ago, the majority of jobs were agricultural farming. But nowadays the pie chart of what jobs people do is pretty finely sliced into all kinds of different categories. And I think that's a thing that one will tend to see as more is automated, more different things, there are more different things that people can do and choose to do and so on. And so I also suspect this is a different point about sort of what jobs people do and do they only do one job or do they do multiple kinds of things? I suspect there will be more and more of people are not just doing this, they're also doing that. I mean, for myself, you know, I have for a long time, you know, run a tech company for 36 years now. You know, I also try and do basic science. I also try and do some educational kinds of things. I also do stuff about history and so on. You know, I'm a sort of, in a sense, I have, you know, I've chosen to have the equivalent of, of multiple careers, so to speak, all mushed together. And they have many interactions between them, but. But they're really different kinds of activities. And I think that that will be something, you know, I know many people who are likewise, and I think that's a thing that will be increasingly common. You know, it always used to be the case people would have, you know, they would have their work, they would have their hobby. I think there's going to be more of a, you know, hobbies become work and work becomes, you know, if, if one's lucky one's doing work that one enjoys as much as one would a hobby, so to speak. But I think when it comes, you know, we can see in, in the world today, you know, people who were like, oh, I play video games for fun. Well, actually I'm making a living as a, you know, esports, you know, or a live streamer of playing video games and so on. So a lot of these things that were just hobby become less, sort of less hobbyist and become sort of secondary jobs, so to speak. So I think the landscape of what people do will. My guess is that one of the things is there'll be more fragmented, there are more different categories, There are more different. An individual person might do more different kinds of things that they might say where any one of them could have been a whole job before, but now it's just part of, part of the portfolio of things people do. Let's see, Doctor Verm here is asking, are we about to reach a post truth world due to AI enabled misinformation? How do we combat this? You know, it's always a complicated thing because it is, you know, it was the case that writing something, publishing it in newspaper and so on, it's all a lot of effort. You know, the newspaper, you know, you're printing something in a newspaper. Okay. You know, I've even seen these giant printing presses for newspapers and they're, you know, it's a big production. It's like it costs a lot of money. You switch on the presses, you know, the big streams of paper go through and you're printing things for newspapers and then you're delivering the newspapers to places and so on. It's all a lot of effort. And you know, it has to be worth something to go to all that effort. It's. If you write some completely uninteresting nonsense story that nobody cares about, you're not going to go to the trouble of printing it, distributing the newspapers, doing all those kinds of things. But when one's dealing with the web and so on, the cost of putting some content out there has gone down to basically nothing. And so what was kind of this big constraint that nobody bothers to publish a book, a newspaper, whatever else that nobody's going to read or that's complete nonsense or whatever. You know, maybe sometimes it'll be nonsense, but highly entertaining and lots of people will read it. That's a different, that's a different dynamic. But the idea that there's kind of a sort of a threshold that had to be passed for information to get out there. That threshold is now zero. There's still, there has been at least some threshold which is somebody's got to actually write that whole essay about, you know, how the, you know, how a llama escaped from the local zoo and, and started, you know, eating its way through whatever, you know, somebody's got to write that piece. Well, except now they don't. Because with large language models and so on, it's rather easy to say, you know, write a story about how a llama escaped from a zoo. Write a story about how llama escaped from this particular zoo, feed it a few Facts about the zoo. And you've got a completely realistic story about a llama escaping from a zoo and hopping over fences and whatever llamas do. And then the problem is that the cost of producing that is exceptionally low. You put it out on the web where the cost of distribution is exceptionally low, and you've got a story that may have absolutely nothing to do with anything. Now, you know, you say, well, that story about the llama, it's misinformation. It's. There wasn't any llama. It didn't do that. You know, there was no llama that. That kind of did a, you know, jumped in the air and did a backflip type thing or whatever it is the. But yet it was written as a beautiful essay and it was put out on the web and it's like, this is all, you know, nonsense. Well, of course, you know, you have to realize that there's. Fiction is a perfectly valid form of human expression. And you could have a lovely story about the llama that did this and, you know, learned to be a gymnastic llama or whatever else. And, you know, that's a valid thing. Now you might say, identify. This is a work of fiction. You know, I was just using chat GBT yesterday, asking it to write something, and it now makes a big effort to say, of course this isn't real. This is a work of fiction and satire. And then it goes and writes the thing. So it's beginning to try to have some awareness of that type of issue. But that's one thing to say. Another thing to say is when one even defines what's misinformation, what's true, what's not true, etc. You know, it's. It's often very hard to know. There are some domains like, well, take mathematics, for example. You think you know what's true, what's not true. Two plus two equals four. Two plus two doesn't equal five. Okay, but there is a, you know, does two plus one equal three or does two plus one equals zero or one? Well, you say, how could two plus one equal one? Well, if you happen to be working in mod two arithmetic, kind of clock arithmetic, where you're. Where zero is. Where two is equivalent to zero, then two plus one equals. Equals one. But you say, but that's not what I meant by plus and so on. Well, it gets a little bit more complicated even in an area like mathematics. And it's a much deeper issue in mathematics that there can be questions that given any fixed set of sort of axioms that define mathematics, you can end up with questions where you can't decide whether they're true or not based on those axioms. So even an area as firmly kind of tied down as mathematics, the question of what's true is actually a bit complicated. And when it comes to things about the world, it's, you know, it's. It's vastly complicated, you know, for example, I don't know. Well, here's a. Here's a good one. The llama likes to eat peanuts. Okay, Really? I mean, you have a llama, you handed a p some peanuts. The llama grabs the peanuts and eats them. Then it makes a funny face and indicating that it really didn't like the taste of those. Okay, did the llama like eating the peanuts? Did it not like eating the peanuts? Another llama might, you know, look at the peanuts and brush them aside or whatever it is. You know, the question of does the llama like eating peanuts? Is a complicated one. And obviously you can have versions of that when it comes to, you know, people and opinions about people. Oh, so and so, you know, is a. Is a terrible person in this way or so and so. But actually in this way, they're not a terrible person. So on. It gets very complicated to know. And usually human language is a very inexact way to kind of state what's true and what's not, so to speak. One of the only places where we really have really pretty tied down notion of something analogous to truth is in computational language. If you write a piece of Wolfram language code and it says this, then you run it and it does the thing it is defined to do when it is run, you know, the function plus in Wolfram language does a particular thing. It's not kind of an everyday language plus where it could mean this, it could mean that, and so on. So I think this question of kind of what, as I say, in the case of, there are certain kinds of things which get very sort of computational, where you can actually tie down the notion of does this equal that? And so on. But when it comes to human language and when it comes to even things like mathematics, the question of sort of what's ultimately true is a pretty complicated question. And so, you know, there are cases in which somebody and this whole issue of what is intended to be fact, what is intended to be fiction, what is intended to be. And one of the things that's been tricky, and I must say I'm not a big fan of, is the idea. And it's something, you know, during the Pandemic, for example. It's something that was often discussed, you know, governments, for example, or other organizations telling people, you know, you know, emphasizing this thing, it's really dangerous to do this or it's really whatever to do that. And you could say, say it's really dangerous to do X. What does that actually mean? You know, there's some probability that you'll get this thing that will happen based on a thousand samples and this and that and the other. But sometimes there was kind of an idea that you tell people it's really dangerous, then they won't do it. And that's good in general, for some reason. And, you know, this idea that, that what you say, the words you use to describe things can be, you know, you'll use the it's really dangerous as opposed to there's a probability of 1 in 7 that you will get, you know, some, you know, you'll get sicker or something if you do this. You know, saying it's really dangerous to do it versus there's a probability of 1 in 7 that this will happen is a matter of, you know, they might both mean the same thing. To some people, it's really dangerous to do this means there's a 1 in 7 chance that you'll get really sick would be, you know, that's a perfect translation between those things. But to other people, it's really dangerous to. They'll hear that as something very different from there's a probability of 1 in 7 to do this or that thing. And I think there was a, you know, there's a certain. It's complicated to say, is it misinformation to say it's really dangerous to do this when the probability is 1 in 7, some people might say, oh, well, it depends on, you know, what the endpoint is, how sick you're going to get. And if it. Even if it's 1 in 50 and you get really sick, you might still say it's really dangerous to do X, you know, if the chance is 1 in 50 million, but there's a chance that, you know, something terrible happens to you if you do it. Most people probably say, would not say it's valid sort of to say it's really dangerous to do X. But so I think it's really difficult to have a definition of misinformation, so to speak. And I think one of the things that's sort of interesting is that in today's world with LLMs and so on, this question of sort of making these judgment calls about whether it's a good. Whether something is okay or not, it's kind of going to be an interesting LLM meets LLM story. Because, for example, if you host a website and you say, I don't want any pictures on this website that depicts anything bad happening to animals. Let's say, well, you can, and that's your terms of service, so to speak. Question is, how do you decide that? Well, you could have some team of people and they look at all these pictures and they say, this cat doesn't look very happy in this picture, whatever else it is. Or you could have an AI system that basically has where you give it a textual prompt that says, you know, that all animals and depicted in all these pictures have to look happy or whatever it is. And it will, based on the training that it's had, based on, based on the average of having seen lots of pictures on the web and descriptions of those pictures which said, you know, happy cat, whatever it said, unhappy cat, whatever else it was, based on all of that kind of training data, then the AI will be able to give some conclusion about is this picture that you're feeding in that you want to put on this website. Is it okay or not? And I think that's something we're going to increasingly see is kind of these judgment calls, particularly when it comes to, you know, is this an okay thing to post on a social network? Is it not? Those. Those judgment calls will be made by LLM AI like systems. And, you know, I think that's a, it's a funny situation because you might say, if it was the case that the decision was made just based on the words that were used, if you were never able to use the word bad, then people will find ways. If it's a very coarse criterion like that, no post will ever let you have the word bad in it. Okay. Then people will say, you know, they'll say, you know, backwards dab or something. They'll find a way just to, you know, to run around it so that the people who are reading it can perfectly well understand what they meant before. But what's tricky with something like an LLM is that you're kind of up against a kind of computational barrier because people don't know yet. Maybe we'll discover a little bit how to do this, how to game an LLM. I mean, people don't know. It's kind of related to the sort of emerging field of prompt engineering of, you know, what do you tell an LLM? For example, we've had this issue, we now have this wolfram plugin for ChatGPT. And we need to be able to tell it, you know, if the, you know, send this kind of input to this piece of our plugin, send that kind of input to that piece of our plugin, you know, clean up the input that you're going to feed to our plugin in this or that way. And you know, we try putting capital letters, we try putting stars around things in the textual prompt. And the LLM sometimes pays attention and sometimes doesn't. And you know, in our particular case, we solve that technologically by having sort of a catcher for what the LLM sends that tries to clean up the mistakes that it makes. But the, there's the question of, well, well, is there a science of really getting the LLM to do what you want? I mean, it kind of reminds me of if you are managing people and you really want the people to do what you want, what do you tell them? I don't think I know the answer. Even though I've been doing management for what, 40 years or something more than 40 years now, I don't think I could tell you the ultimate. This is the exact thing you tell people to get them to do exactly what you want them to do, so to speak. You know, it's, it's sort of still an art form. And whether it will, whether there'll be some kind of, yes, you can just lock the LLM in to do this, I'm not sure. But I think this, the idea of this kind of, rather than saying, oh, there are these bad words, don't, don't use those bad words. And that's, and it's a very sort of coarse way where you can always sort of walk around the bad word type thing. If it's kind of, you're doing this thing and it's being graded, it's being rated by, it's being decided by an LLM, I think it's not going to be so easy to game and it will do something which seems to be a fairly human job. And sometimes one will say, just as one does with humans, oh, that person goofed. They should have agreed that this was okay. Or they, you know, why did they say it wasn't okay? They made a mistake, et cetera, et cetera, et cetera. And you'll say that about the LLMs as well. But I think it will be, you know, that will be a, a common thing. So then you could apply that to sort of the misinformation story. You could ask an LLM, you could say on this site, we Won't allow any misinformation. Okay, what does that mean? Well, the LLM can go and try and decide is this misinformation? It's looked at the web. It's got a lot of things which say, which are tagged as this is misinformation. So it probably has some idea what misinformation, what people think of as misinformation. And so, you know, it will have a fighting chance of doing something reasonable there. Of course, it won't always catch it, but I think that's a thing one will see now. The thing that is absolutely a thing one will see is just tons of material written put out there on the web where. And elsewhere where it was written by an LLM. It may or may not have anything to do with anything. It was just. The LLM was just told, write something that goes along these lines. And now it's easier to do that writing. So I think it's, it's the. That's. That's kind of. And, and, you know, was it written by an LLM? Was it not written by an LLM? You know, at this point, us humans have some sense of, you know, this is what ChatGPT typically writes, but it doesn't need to be that way. And I'm quite certain that there will be LLMs that will kind of evade any kind of was it an LLM or was it a human type question. I think that's. One shouldn't count on the fact that unless the AI chooses to declare itself as I am an AI, or unless it's really burnt deep into the sort of psyches of the AI, that I'm always going to explain that I'm an AI, I don't think one will be able to tell. And this is a thing, you know, back in 1950, when Alan Turing, one of the sort of pioneers of computing, invented this Turing test, which was this kind of notion of, you know, when do we know we have true artificial intelligence? Well, you, you know, have a. Well, I think he had a teletype, I think was what he was using. But we could modernize that. You know, you're just typing to something and you're typing text messages to something. Can you tell if the thing at the other end of your test messaging stream is a human or an AI? And his sort of test for what we all know we have genuine AI, is when you can't tell. And I'm sure we've got to that point now in the last few months. So anyway, Haraz asks, since today, ChatGPT can only reproduce written human reasoning. Will it ever be possible for ChatGPT to be better than humans one day? I think the answer is yes, there will be things that it will do better than humans. I mean, they're already. The thing that I think it's pretty good at is implicitly noticing patterns across large amounts of information. So it's pretty good making analogies. Hand is to finger as foot is to toe or whatever, but much grander analogies. And some of the things that one's interested in in science and other places are, I know that pattern. I've seen something like this before. This thing that I'm trying to solve in physics is just like this thing that I solved before in chemistry, let's say. And for a human, that can be something of a stretch to notice that analogy. You have to understand something about both of those fields. For the AI, it's got it's ingested knowledge about, it's ingested at least text about those different fields. And it's going to notice, gosh, there's this analogy here between economics and, I don't know, some area of mathematics or something, or physics, let's say. And because very few humans knew both of those fields, they were not able to make such an analogy. I mean, I have to say this is one of these things where I feel a little bit there's things I have to figure out in science quickly because otherwise I don't get to figure them out. It'll just be an AI that figures them out. Because for somebody like me, I'm interested in a lot of different fields and have learned a fair amount about a lot of different fields. And so I've been lucky enough to be able to make some of those sort of grand analogies between very different fields that people are not easily able to make, usually because they just don't know the two ends of that analogy. I think AIs will routinely be able to make some of those things and they'll seem very sort of superhumanly smart. But I think it's more because they are broader than because they are deeper. And I think that's something we will probably see. I think that in terms of the access that AIs have to deep computation, I mean, an LLM is really just even though it's very useful and impressive, what it's doing is it's sort of taking some stream of words and it's saying, how would I continue the stream of words based on what's typical about what's found in the web and on Books and things like that. And that process of how would I typically continue it can be very. Can seem very sophisticated to us. It involves lots of reasoning because. Well, the reason it seems like it involves reasoning is the same reason that that was how sort of logic was deduced from text. It's like there are these patterns of if the text says this, then it's going to say this. And those patterns you see over and over again, if it says this, then it's going to say this. That is the structure of certain kinds of reasoning. And it's able to kind of lift that structure from one place to another. That's kind of the nature of the way that it does generalization. And so that means it can apply quotes reasoning because it's seen this pattern of this implies that in lots and lots of different cases. So it's going to lift that and apply it in some other place. I think the thing that is really interesting is this sort of deep computation, the kind of computation that was discovered in the last 50 years, which, you know, that I've spent lots of effort on, of being able to really compute things about the world where you're running lots of computation steps to figure out things about the world, to work out things about nature, to work out things about the mathematical world and so on. That's not the kind of thing that a large language model can do. It's a much shallower kind of computation that's just sort of trying to match up what it's otherwise seen as text on the, on the web and so on. And so the merger of those two things, being able to take that kind of thing that is a good match for us humans with that, with human language and so on, and being able to connect it to kind of the depth of computation that in the last 50 years we realized is possible, that's really powerful. And that's the thing that this plugin for ChatGPT that we built, connecting to our kind of computational system, that's what that enables. And I think the thing that's going to happen, which is pretty interesting, is a thing that I've done for decades now, is the idea of computation and making it accessible to as many people as possible. That's sort of the thing I've been trying to do, and I've done that by creating this kind of computational language that expresses things that we think about in computational terms. Done that in WolframAlpha, for example, where you can type a small piece of natural language and just get a computational result out. But we've now got a much broader version of that. With LLMs we can take a much more sophisticated set of instructions, so to speak, and expect to turn that into our Wolfram language computational language. And that's really powerful because then, you know, just from thinking about things in terms of natural language, you're creating this computational language that Len lets a computer go off and really, you know, run, run what a computer can run, so to speak and figure out and, and make use of what's possible in the computational universe, so to speak, to figure out a lot of things and then can bring it back and can express it in natural language and so on. So we finally got the ability, you know, we've had in the past. We've had sort of, there's the programmer thing, there's the computing thing and then there's the talking about things in natural language. Those two things are finally merging and that gives just a, you know, sort of everybody access to much deeper computation. Something we achieved a dozen years ago with WolframAlpha for certain kinds of knowledge and certain kinds of quick questions where you're just saying, I want to know the answer to this question. WolframAlpha can do that. If you want to say, I want to create a program that's going to run a thousand times to do this particular thing for this thing, that's not what WolframAlpha is set up to do. But now with LLMs, plus the technology we have for connecting it to Wolfram language, you can do that. And that's really something, I think, exciting because it kind of opens up the sort of power of computation, the power of this kind of knowledge based computation that we have in Wolfram language and so on to everybody. And it's, you know, people have been worried that, oh, they're the people who learn, you know, techie stuff and programming and so on and the people who don't. But you know, there's a lot of power that comes from the notion of computation and that's sort of always, you know, that has in the past been kind of been kind of separated off from people who didn't learn sort of how to talk to their computers. But now we've gotten much closer to where you don't really have to learn that. Just like in Wolfram Alpha, there's no documentation for Wolfram Alpha. You just type what you want to type and it will answer questions. Well, now you're at the point where you can just type what you want to type and you can make a piece of computational language code that you can then put together in building blocks to build up some sophisticated to achieve what you want to achieve computationally. I will say there's one kind of footnote to that, which is you kind of got to know when you build up a big tower of computational capabilities, you've kind of got to know that all the bricks you put down in the tower are actually kind of correctly shaped bricks. You didn't start putting foundations down and, you know, wind up with the leading Tower of Pisa because your first few bricks were misshapen, so to speak, relative to what you thought they were. And so I think that's a place where our computational language is really important because it is a language that is intended for humans to read. It's not like sort of typical programming languages, which are intended just for a computer to read. It's a language intended for humans to read. And so what you can expect to do now is you tell your LLM something in natural language. It writes Wolfram language code. You read that code, potentially you might just see what it does, the obviously right thing. But if you're at all uncertain about it, or you can't quite tell what it does, you read the code. And that's something. Well, there's some learning to in. In how you read that code. It's not so difficult to learn that. And then you say, yes, it's doing the right thing. And then you start building this giant tower and it doesn't lean, so to speak. So, let's see. Okay, RITC asks, how can AI, through the lens of this concept of computational irreducibility that I talk about a lot, navigate the landscape of possible rule sets and achieve true intelligence, mirroring the complexity of our universe? It's a very interesting question, and I think the answer is it can't. What an LLM does is it works a bit like the way our brains work with neural nets and so on. And it does things that are very similar to what we humans do. We humans are not able to go out into the physical world and immediately say, this is how everything in physics is going to work. We're not set up that way. We're set up so that we can answer questions quickly about things that are common in our world. We've kind of evolved through biological evolution, et cetera, et cetera, et cetera. And we've also chosen to set our world up to be one where the kinds of things that we're able to do with our brains, with the neural nets in our brains and so on, the kinds of things that we're able to do are ones that we've set up our world to be the things that we need to be able to do. We could have set things up so that we're continually confronted with incredible complexity and randomness and so on, but we don't. We choose. I mean, like, for example, we choose to make our roads be things we just drive the car down the road. We could have set up roads to be these complicated things where, you know, you had cars with legs and you had to walk on this, you know, on this. I don't know, I'm thinking of lily pads for a, for a, for a frog or something, but, you know, where you have to walk on this paving stone, that paving, that stone and so on. And it's very difficult to get from one, you know, from one place to another. You have to carefully choose where you put your legs. But we don't. We set up our world to have roads where you just drive straight down there in your car or whatever else it is. And so it's something where we've chosen to build our world so that we can navigate it with the apparatus that we have in our brains. And we also, we choose to not consider those parts of the physical world. We don't live in places, so to speak, where it's very hard to navigate what's going on. We don't. And so there are things that we can do, and we set up our world to be a place where we can do the things we can do, so to speak. If you say, well, what's out there in principle? Well, what's out there in principle is a lot of incredible complexity, lots of computational processes that are irreducible in the sense that you can't predict what they're going to do. But those are things that we as humans tend to avoid, and our AIs sort of modeled on us will have the same issues with those kinds of things. What you might be able to do is to say, for an AI, you say, here's this computational system, it's really complicated, et cetera. And the AI will kind of channel us humans and will say, well, I think a human would say this or that about it in the abstract. It's not going to have something vastly beyond what humans can say now. It's a little bit of a few sort of footnotes to that comment. So, you know, if you look at the course of human intellectual history, we've gradually learned how to talk about more and more kinds of things. We've learned new ideas, new concepts and so on. That gradually allows us to expand the domain of computational things that we can talk about. But one thing to realize is if we expand to some new domain, we can now talk about, let's say, fractal patterns. Well, the fact that I can talk about them is because I got a word, fractal patterns, and we have some sort of common understanding of what that word might mean. If you just go out as a solo person and you say, I've done this, I have to say, you go out into the computational universe of possible programs, you find all these programs, you have a little idea in your mind of how these programs work, but you don't have a way to describe it. You don't have sort of a narrative you can tell about it. It's kind of a funny situation because you don't really have. That's not a place where you can sort of talk about what you've seen because you have no words to describe it. You have no way to communicate what you've seen. So what I think tends to happen is we humans, sort of as a. As a collective society or something, we gradually move further out in the domain that we can understand, and we sort of build up a language, a narrative for how we describe those things, and we can tell it to other humans, and we're all sort of together moving outwards in this. In this domain of things we can understand. Now, in AI, it could do that too, and it could be like, happily telling its other AIs what's going on. But unless it can teach that to us humans, unless it gets a bunch of humans involved as well, then it's. It's. In a sense, it's just going out. And it's a little bit like that computational system that it's studying. It's just something. It's doing some complicated computation. We as humans can't tell what it's doing. It's. It's not. It hasn't successfully kind of expanded its domain in a way that we humans are connected to and care about. So I think that's sort of the issue, is that it has to not only kind of be able to be looking at things that are sort of further out in this domain of computation, it also has to have invented for us and taught us a language for how to talk about those things. And perhaps an interesting thing to try to explicitly do that and to have the AI sort of come up with a language. And then it's a question of, okay, I've got to, you know, sit still long enough to actually learn that language, so I can kind of ingest that, get it into my mind so that I can. So, so it becomes something meaningful to me as opposed to something. Well, yes, the AI is doing it, and it's a lot of bits inside the computer that are flapping around in this or that way, but it doesn't really mean anything to me. It doesn't really have, doesn't really connect with me. SVD asks, will we see more of the phenomenon where AI contributes to the fundamentals of science? We haven't so much seen that yet. I mean, there have been sort of AI like techniques for quite a long time. One of them is automated theorem proving. Being able to take a mathematical theorem and use the axioms of mathematics to try and figure out, can you make, can you use those axioms to construct the sort of path that represents a proof of the theorem? I believe that the only time that an unexpected discovery has been made by automated theorem proving was something I discovered 23 years ago now about the minimum axiom system for Boolean algebra. Most of the time, that technique of automated theorem proving has only really been used to confirm things that people already thought were true. It's actually surprisingly rare. That sort of. I mean, and again, I myself have done a bunch of work along these lines of using what amounts to AI to discover unexpected things in science. Now, one of the things to say about that, and it relates to what I was just saying before, is that if you just see, okay, you've got this weird phenomenon, you can generate it on a computer. You don't have a way to talk about that phenomenon. It's just this weird phenomenon. You've never seen it before. It's just a weird phenomenon. And until you have a conceptual framework, until you have a way to talk about that, to kind of put it in a, a framework where it fits in with other things, you won't tend to think of that as being progress in science. You'll just say, this is this weird random phenomenon that isn't really connected to anything else. The, the progress of science tends to be making this sort of broad connection where you can have a narrative that lots of people understand that you can sort of, but, you know, pull into your mind, so to speak, that represents that, that piece of progress. So I think that's one thing to say. Another thing to say is when it comes to LLMs themselves, there is science to be done about how they work. There's a lot of questions about things that are empirically observed and why is that true? Things about how big The LLM has to be compared to how much it's learning, et cetera, et cetera, et cetera. That's something I've been encouraging some scientists, physicists particularly recently, to study this because I think that it probably is figurable out why these kinds of principles exist, but that isn't known. And I think there's a lot of kind of science to be worked out about kind of why. The LLM successfully puts together the language and the way that it puts it together. There's facts about language. People have tried to make a sort of science of how language works. Linguistics is sort of a science of how language works. They haven't been terribly successful. A few things can be said. A few, you know, most languages seem to have nouns and verbs and adjectives and things. Those are categories that seem to be fairly robust across different languages. There are aspects of the way that you make sort of subclauses and sentences and so on that seem to be fairly universal. But I would say as a whole, I would argue that sort of the science of linguistics, of saying why are languages the way they are for us humans, and what sort of global features are there in those languages, hasn't been particularly successful. I think the success of the alarms is showing us there is a way to do that stuff. We just didn't do it quite right yet. And that's a place where I think the AIs will contribute, by the very nature of them working to an understanding, to a better understanding for us of a phenomenon like human language and so on. Let's see, Satoshi is asking, what do I think about AI alignment and the existential risk of AI? Well, okay, let me explain what that means. First of all, you've got an AI, it's doing what it does. The question is, is what it's generally doing aligned with what we humans want it to do? Like, for example, you've got an AI and it's supposedly, you know, picking content for some Internet site. And we say, is it doing that in a way that it's figured out that, you know, it puts this kind of content on and more people will click on it and things like this. Is it doing it in a way that is aligned with our overall objectives as a civilization or whatever? How does that work? Well, the first question is, well, what are our overall objectives as a civilization? For example, one objective that we probably mostly agree on is we don't want to go extinct. And so aligning AIs so AIs don't make us go extinct is probably Something we'd all agree about. Probably what it takes to get there, the more minor steps. That's quite challenging. And you can ask, well, what do you want to generally say that AIs should do? Let me give you an example. A thing you might very well say is I want AIs. Let's say an AI is owned by somebody, it's on a computer, the computer is owned by somebody, the software is owned by somebody, and so on. You might say, I want that AI to do whatever its owner tells it to do. Make it do it. Seems like a reasonable thing. You say, but wait a minute, if the owner tells it to, you know, do some terrible thing in the world, surely you want the AI to say, no, no, no, I'm not going to do that. It's just like if you have, you know, a self driving car or a car with a bunch of automation and you say, I want to go faster, faster, faster, and the car knows It's a 30 mile an hour speed limit zone, does it go to 35 miles an hour or 40 miles an hour or whatever? Or does it say, no, no, no, not going to do that. You know, it's a 30 mile an hour speed limit. I'm just going to stay to that complicated question. You know, there are cases in which you'd say, well, gosh, if you couldn't do that extra two miles an hour, you'd have an accident and that will be a worse outcome than oh, you went slightly above the speed limit. So it's hard to decide whether you. For example, the question of should an AI do what its owner tells it to do? You might very well say, well, for societal reasons, you want to pull back from that maybe, or you want to say, maybe you want to have as a sort of primary objective that it's always humans in charge. In which case, yeah, you should make the AI do whatever its human owner tells it to do. Of course, it gets more ornate when you have an AI toy for a cat and you say, does it always do what the cat tells it to do? Well then you're in another whole complicated loop of. But the question of what are the provisions? What are the kind of elements of what one can think of as sort of an AI constitution? What do you put in there to say you want the AI to do this, you don't want the AI to do that. One of the things that's very bizarre in today's world is that and today's, I mean, the last few months is that you give these textual prompts that say what you want the AI to do and you're literally writing something that looks very much like a piece of legal, legal document that says do this, don't do this, et cetera, et cetera, et cetera. And I think that's something I was not anticipating. I think to some extent at some level people find that a bit unsatisfactory because they'll say, what does it really mean? Well, of course they've said that about the U.S. constitution or laws or whatever else forever. What is the interpretation of these words? Well, that's where people, when they can turn them into computational language, where there is a well defined meaning that is defined by what the thing does when it runs as a computation. There will be a desire to do that I think in a bunch of cases. But this whole question about, you know, how do you get the AIs to quotes, do what we want? Depends what, what it means to do what we want. It depends how globally you have to have people agree about what things, whether, whether they all, Whether all the AIs need to be pointed in the same direction. It's the thing, I think I mentioned this another time about, you know, does everybody drive on the same side of the road? You know, you can't have, at least in a simple way of setting things up, you can't have some people driving on one side of the road, some people driving on the other side of the road. And you know, because then the cars crash into each other. Of course, if the cars were smart enough and were self driving enough, they'd be able to dodge each other and you could have a much less sort of rigorously defined, it's the left drivers and the right drivers type thing. I think that the, you don't then have to have as much in that particular case, sort of AI itself saves you from having to have as much sort of global alignment between what the different AIs are doing in that case. But there are certainly cases where, you know, we share the space and other resources of the world, so to speak. And there are places where, you know, you can't have one thing, you know, I don't know, doing. I don't know what they might do, you know, generating a huge noise and then somebody right next door is like trying to sleep or whatever else that that's, you know, that's going to lead to unhappiness. Just because you could say, well, let's again, you can, you can sort of solve it with technology. Let's have this, this phased array that's sort of Noise canceling array that you, that the thing can put up so that there's noise in one place. And it only affects those one set of people. And you know, one house over, it's perfectly quiet because of this complicated, you know, cancellation of, of sound waves and things or something like that. But there are plenty of cases where there's sort of a collective experience where you have to have sort of an alignment there. And that's a tricky thing to achieve. I mean, I think one of the things I was thinking about is let's say you've got lots of people, they have lots of opinions about things and you say, well, what's going to happen in the world? What should you do in the world? Well, an interesting sort of possibility is everybody writes an essay, everybody writes what amounts to a prompt for an LLM and you just feed them all into the LLM. You say, here's 10,000 prompts that came from all the people who live in this particular part of a city or something. And okay, LLM, should we build a park here or not? And the LLM just has ingested all of these essays by some people saying I really want to fire kite, some people saying, I really want to have a road that gets faster from here to there. All those kinds of things. And it will be interesting to know, is that a form of, is that a sort of post democracy form of government to say, you know, just everybody write an essay, so to speak, rather than people saying I vote for so and so to be representing me or whatever else? I don't know. I think it's an interesting sort of thought experiment and possibility and it's something that, you know, is a. But as we start talking about the question of whether the AIs are going to do what we want, it's who's the we and how do we decide what we want, so to speak. And that's a tricky thing. Now, you know, as I say, the one thing we probably mostly agree on is AIs don't make us go extinct. I think that's a tricky thing to. I mean, I would say that I tend to think we're pretty far away from a situation where the AIs could have any chance of making us go extinct. But it will be the case that AIs will be increasingly hooked up to critical systems in the world. But I think it's probably the case that the kind of oh, let's make the whole species go extinct in one fell swoop is not really a thing. It's very hard to Do. I mean, it turns out we're kind of. There's lots of us and 8 billion of us or whatever and spread over a reasonable sized planet. And it's kind of like there are very few things that are kind of like globally. Oh yeah, you press one wrong button and the whole thing blows up type thing. I suppose the closest we have is, you know, one can imagine some terrible situation where somebody makes a designer virus that's just really, you know, that's terribly bad for all of us, kills all of us. I think that's actually, you know, it's surprisingly hard to do because usually the more aggressive a virus is, a virus works by, you know, spreads from person to person mostly. And you know, by the time it's a really, really deadly virus, the person will not be out and about infecting other people if the virus has put them totally out of action. That's not always the case. There can be nasty things where there's sort of a delayed time type thing where you get sick, you, you don't know you're sick, you infect a bunch of other people and then the virus kind of blows up on you. But, you know, so things like that could happen. And I'm afraid, you know, I, I tend to think that the nature of, again, for all sorts of reasons about immunology, which let's not go into here, you know, I think that's again, a very unlikely thing to be like wipe the whole species out. So I think in terms of, you know, I'm. I don't think we're heavy, I don't think we're at heavy risk of extinction, so to speak. And the things below that, you know, most of the things below that it is not, it is the case people will not agree about, you know, well, we should make it do this, not that we should make it, you know, it. And then it becomes as messy as things in the world typically are. And you know, some AIs will do this, some AIs will do that, and it'll be kind of like the way, the way it is with people. Let's see, there's some very interesting questions here. What's it like to be an LLM? Asks Brady. Well, I don't know. What's it like to be a computer? I was thinking about that. And even a traditional computer, forget all the AI stuff. It's like you boot the computer up, you start it up, it sort of comes to life, it's interacting with the world. It has a camera, it's watching stuff, it's waiting for people to type on its keyboard. It's building up a certain amount of information in its memory, a certain amount of state for the computer. And then eventually, after a week, a month, whatever else, the computer crashes, the operating system crashes of the computer and it has to restart itself. And that process from startup, from boot up to shutdown is kind of like a human life. It's kind of like you start from a certain genetic template, you start from a certain set of things that are stored on the computer, then you sort of come to life, you record, you remember a bunch of things in your life, and eventually it's the end of your life. And then, you know, your, your children sort of are, you know, carry on or whatever, and they've started from similar genetic material to yours. It's just like kind of the next boot cycle of your computer. It's kind of started from similar, from a similar point. And in a sense, the, you know, what you, what you, you know, if you, if you're somebody who writes down your thoughts or records your thoughts or whatever, in a sense, that's what's stored on the disk of the computer. And, you know, for the next generation, when the computer boots up again, that's what, that's what will be visible there. And I do think that the, if we imagine what it's sort of like to be a computer, it's actually surprisingly similar to what it's like to be a human. You know, there's things in the outside world, there's communication with other computers, all those kinds of things. And. It's surprisingly, you know, you might think, oh, it's just a computer. We can look at it from the outside. Well, you know, humans are just humans. We can look at them from the outside and see what they do. In the case of a computer, we might know a little bit more about how it works inside. I think we're beginning to know more about how brains work inside. We're beginning to know how certainly idealized artificial brains of neural nets work. So I think the experience of being computer is probably sort of the inner experience of being a computer. If we can imagine that is like the inner experience is quite like the inner experience of being human. Of course, we only have our inner experience. I have my inner experience. You all have your inner experiences. I can't know what your inner experience is like. I can, I can kind of see it from the outside. I wish I could see you guys more, more, more explicitly here, but, you know, I can, I can see from the outside what you you know, what your characteristics and behavior are. That doesn't really tell me. It doesn't allow me to get inside and experience you from inside you. I can only experience me from inside me. And so when we look at a computer, we don't get to experience the computer from inside the computer. Maybe one day with neural implants and things like this, we'll get a little bit closer to experiencing the computer from inside the computer. But we're not there yet. Let's see. Question here. There's several very interesting questions here. There's a question here about will, will programmers become obsolete? Yeah, low level programming is going to become obsolete in the same way that assembly language programming has become obsolete, became obsolete decades ago now. I mean, you need to tell a computer what to do and tell it what to do in a precise way. The language that I've just spent the last few decades building is the highest level way that we have right now to tell computers what we want them to do. Precisely. And that's, you know, that's something that's going to continue to be useful. But when it's like, oh, and, and let's tell the computers in more detail at a lower level what to do, you know, that's automatable. I've known that's automatable for 40 years. You know, I've been in the process of automating it and I've done a lot to automate it in the last 40 years. And many people have used what we, what we've built. The fact that there are many people writing kind of low level programs is sort of a strange anomaly of history. I mean, I think perhaps it's inadequacy on my own part to have not managed to explain to people more clearly, although I think I've done a fair amount along those lines. You know, how you can interact with computers. At a higher level there has been, you know, continued interest. No, I should learn to program and you know, write this code and learn about memory allocation and this and that and the other. You know, those are things which, it's like you drive a car. Yes, it's amusing to know how spark plugs work and how fuel injection works and so on. But most people who drive cars do not need to understand fuel injection. It's a different level. And the same is true of computation, except that people have had the idea that you really do have to understand the fuel injection as you did when cars were young. You know, you had to understand all kinds of things about how they worked in order to use A car. But then they got more automated and don't have to do that. Gosh. There's a question here from assist. What are suggestions for a high school student who is interested in both AI and physics? Come to our summer program. That's a good, very practical suggestion. I think that the. It depends what kind of physics and maybe this is. I know I have to go in a minute, so I'm not going to have a chance to give this question. It's real. Best response I think one of the things that kind of has come out of the physics project we've done in the last three years is physics is ultimately computational. We didn't know that before. We thought physics was mathematical, but it isn't. It's computational as well as mathematical. And so the things that come out of understanding the concepts of computation are really important now. Very recent knowledge to understanding physics and sort of there's an increasing opportunity for people who understand computation to contribute to physics and to contribute to the foundations of physics. And we're seeing that in the, you know, few hundred people probably who are working on sort of our approach to physics at this point. It's like you really have to understand the computation side of things. There's some pieces you can do with sort of more mathematical physics as well, but it's very useful to understand the computational aspect of things. And that's a bit like what's true in AI now. You know, one issue about AI is there's a lot of kind of engineering plumbing that has been traditionally done in AI. I don't know how much that's going to continue, but that's been. And it makes it much more like an engineering field and much less like a theoretical science field. I think there may emerge a theoretical science kind of field in AI that hasn't really emerged yet. But the thing, you know, it's kind of like there has been in the past. There's engineering on the one side in AI, there's kind of mathematical type science on the other side of physics. Now both of these look more computational. Now in the case of AI, I would say the biggest question these days is, okay, so we've got these capabilities for AIs, how are we going to be able to use those? That's sort of more so than, let's go and understand more about the details of the capabilities. It's kind of, how do we take this building block and use it to make further progress? There's more to say about this. Maybe I can cover it again next week, but I think I have to wrap up now because I've got to go and do some day job work here. But thanks for joining me and thanks for these interesting questions. It seems like we're in a. In an AI in the history of AI There was sort of this period of the AI Winter when it seemed like nothing was going to work in AI I think we're now in the AI Midsummer Day or something where. Where it seems like everything is working with AI and perhaps we can continue talking about AI but next week. But thanks very much and bye for now.