Science & Technology Q&A for Kids (and others) [Part 114]

Hello, everyone. Welcome to another episode of Science and Technology, Q and A for Kids and others. Let's see. Have a bunch of questions saved up here and let's take a look. Okay, we have a question here from Seth. Could I discuss the importance and relevance of ChatGPT and asking the thing things about its relationship to Wolf, Malpha and so on. So first of all, what is ChatGPT? It's a chatbot based on a language machine learning language model. This particular one is made by OpenAI and it is a lovely thing. You ask it a question, you say something like, or you ask it say, write me an essay about blah, blah, blah. You could take a prompt from a typical college admissions thing or a typical kind of high school essay prompt. You say, write me an essay based on this. And it's rather amazing because it can actually produce an essay that at least on the surface seems completely sensible and is kind of well constructed, has full sentences, it has a flow. It seems to often end with an inconclusion comma. How does it work? Well, at the simplest level, it's trained from a large corpus of information, a large collection of information that probably includes a decent fraction of the web pages that are out there. There are some number of tens of billions of web pages out in the world and you can just read them all. And that's what this system, over a long period of training that's been pretty expensive and CPU and GPU intensive to do, has been able to do. So let's zoom out a little bit and just ask what's sort of the basic idea? At the simplest level, what this is doing is it's a language model. What does that mean? It's something which knows the probabilities for different kinds of sequences of letters that will occur in constructing a linguistically correct sentence, for example. So let's take some very simple examples of that. So let's say you're writing English and you write a Q, okay? There's a very high probability that that's followed by u, because in English, you know, qu is the typical form that occurs inside, inside words. If you have T, H, E, it's a pretty good chance that's the word the, and it's going to be followed by a space. So at the very simplest level, you can try and make this kind of a statistical model, a probabilistic model for a language like English, and you can say the chances are the sequence of letters, you know, a th might be followed by an E, for example, has a certain probability. It's Much more likely that a th is followed by an E than that a th is followed by a J. That's extremely unlikely. That's really not a possible thing to have in correct English. So you start off at that very low level of just saying, what are the things that have a probability, a good probability to be correct in English. And how do you figure that out? Well, you look at a bunch of actual English sentences and English words and you say, okay, given a th, what comes next? There's a certain probability for an E to come next, there's a certain probability for an A to come next, there's zero probability for a J to come next, and so on. So given a collection, let's say you have a million English sentences, you can get a pretty good estimate of what letter will come after what pair of letters. Let's say that's the very simplest level, is to just say, let's say the very simplest thing might be okay, the very zero thing is what letter is most common to occur altogether. So in English, that would be an E. For example, T would be next. So that's not taking account of any context for letters. That's just saying if you were just writing down letters one at a time independently, you'd put down more E's than you would put down z's and Q's and so on. Now, let's say you're looking at pairs of letters. Then th is pretty common. FQ is not common at all, doesn't occur at all. So you have different probabilities for those pairs of letters, and you start building it up and saying what things are likely to occur. And in particular, if you say, well, I've already got these letters already written down, given that those letters have already been written down, what's the likely thing that's going to follow those letters? And so in order to figure that out, the more, the more that you need to, the more kind of, the kind of further, the more letters you've already put down and you say, what's the probability for the next letter? If you want to have something that you've learned from kind of examples, the more letters that you have in that context, the more examples you have to have to give you a good chance of having seen an example that's like the one that you're now being presented with. So very simplest way that one can sort of statistically construct English language, for example, is this thing of just take. Take sentences and grind them up and say, what's the most common trigram? Three letters that Come together and so on. Okay, that's the sort of zeroth level doing that you will produce if you say to, you know, you can set up a sort of a system, and those have existed for 50 years, 60, 70 years now, of just what are the statistics of English and sort of babble out things that are. Have the right statistics at the level of individual letters or pairs of letters or triples of letters or whatever, and the results will be terrible. They won't really sound like plausible English at all. There'll certainly be no sort of thread of a story that carries between them and so on. Okay, so then you go to a slightly larger level and you say, well, what can we. How. How can we sort of learn sort of the largest structure of English? Well, the idea of just saying you've got these two letters, there's a certain probability of a letter that follows them. That's too simple a model, too simple an internal model to be able to really learn how English works. You know, in other words, just because. Let's say, you know, the cat sat on the mat. Okay, cat sat. That might be a pretty common thing. But you might say, what would be a good example? Let's say sat could maybe. Sat phone would be a common kind of pair to come after sat, satellite phone. So then you'd end up with cat sat. That's pretty common. Sat phone is pretty common. And then if that's all you knew, you'd end up with the sentence being produced by saying cat, sat phone, which is a pretty silly sentence. But. So the issue is to have a more sophisticated model for how English is constructed. Now, one of the things that's happened through neural nets is one has these kinds of ways of sort of flowing information from one part of a part of, let's say, a sentence to another in more sophisticated ways than just saying what follows, what. So, for example, a pretty common kind of thing these days. Okay, so let's see, how do we start describing this? So neural nets are kind of a sort of a drastic idealization of how we think neurons and brains work. But for these purposes, a typical kind of neural net says, I'm going to take. I'm going to have a thing. Let's say it's an image. Let's say it's an English sentence. Whatever else, I have a thing, and I'm going to kind of sample 10 different places in that thing. And in fact, before I even do that sampling, I will have turned the thing into a collection of numbers. So for an image, I might turn it into numbers that come from its red, green, blue values of each pixel. For an English sentence, I might at the beginning turn it into numbers by just saying an A is 1, B is 2, etc. But I can have more sophisticated encodings of my input that have the result of just turning it into a collection of numbers. Okay, so then what does the actual neural net do? What it's doing is it's taking that array of numbers and it's sampling different numbers out of that array and it's saying, okay, well, I've got this combination of numbers. Now what it does is it says, I'm going to combine those numbers and I'm going to say, if this number minus that number plus this number minus three times that number and so on is greater than zero, then produce a one, otherwise produce a seven. It's a collection of kind of little, in a sense, math formulas that are being applied. Given a collection of inputs, apply the math formula, give an output. And you keep doing that. You keep doing it a billion times. And you do it in lots of different sort of configurations of sampling this piece and that piece and so on. And the thing is just doing that, you say, well, I'm sampling these different pieces. Let's say you're sampling pieces of an image. I'm sampling the left corner, the right corner, this corner, that, that, the. This combination of things in this region of the image. I'm sampling this and that and the other. Okay, if you do that, just saying you're sampling all those things. Well, well, so what you can't, can't get much from that. The big idea is to learn to, to have something where you train the system. So what you're doing is you're saying, okay, in a, in a simple case, you're saying, let's say you're trying to recognize images. You're trying to say, what's a cat? What's a dog? What you would do is you would give it, you know, 10,000 examples of images of cats, 10,000 examples of images of dogs. And then what you're trying to do is to get it to set up all these little mathematical functions that are, that it's applying to all these different inputs and so on, all these different parts of the input, and then it's supplying results to the results of that and so on. It's making this whole cascade of things. What you want to do is you want to set it up so that all those little math formulas inside are set up to be the right formulas to best discriminate cats from dogs. And so the big idea there is that you're looking at, you have a version of what you think this net, with all its functions and so on might do, and you get the results from that. And you say, those results are terrible. And then you say, but if you were to sort of poke them in this direction, they get better. And then what you do is you say that you have to say, given that they get better, if you poke them in that direction, how do I have to change all the parameters to those math functions inside to get the results to be moved in the direction of getting better? And that's a big adventure in essentially calculus. Essentially, the way you think about this is it's a great application, actually, multivariate calculus, if that's something you've learned to try to understand. How does this. Changing the output, how do you change all these sort of functions in the middle to make the output go in the direction you want? But anyway, the end result is, after lots of training, many, many rounds of, you know, just poke that. You know, you see how the output has to move. You see how that. What consequences that has for all little functions inside. You do that over and over and over again millions, trillions of times. And then you get a trained neural net where all those functions are set up so that it's really pretty good at discriminating a cat from a dog. Now, in that particular case, one of the problems will be if you just trained it that way, you've got cat images, you've got dog images, and you say, do that training. So in Wolfram language, there's a function classify, and you literally just give it a collection of, you know, cat images. Say those are cats, dog images, they say they're dogs. You say, classify any image based on this training data, and it will go through and grind through and it'll show you a little thing that shows you how the loss function, which is the chance of getting it wrong, so to speak, how that's how that's going down and so on, and it's gradually going to learn. Okay, so one of the problems would be, you showed a picture of an elephant. What's it going to do? All it knows is everything's either a cat or a dog. Okay, so an elephant, who knows what it'll say. An elephant is, for example. In that case, it doesn't know what it doesn't know. It's a separate and slightly more complicated thing to be able to train it not only on what a cat's like, what a dog's like, but also how close do you have to be to a cat to be pretty sure that it really is a cat, as opposed to it being something you don't know what it is, like an elephant. So any case, that's sort of the general approach to how do you train these things. But now one thing that happens when you're dealing with something like text is that, so, for example, image identification, you look at a picture of a picture and you're picking out does it have a cat or a dog in it? You're looking at text. Text has a flow to it. It goes, there's the beginning of the text and then there's the next word and the next word and so on. There's a whole sort of time based flow to text. And so that requires a slightly different kind of approach in neural nets than the thing I just described. By the way, I should explain that in dealing with images in neural nets, one of the innovations of a decade or so ago was what are called convolutional neural nets, in which all these different functions that are sampling things are doing that sampling based on an array of pixels that's already a two dimensional array of pixels. They're not just saying we'll sample from anywhere we choose, they're saying we'll sample locally nearby pixels in the image. Which is how we know our visual cortex works. It's also very similar to cellular automata, which is something I've studied at great length. But, you know, so already it has sort of a to know it's dealing with an image. It's using the fact that images are sort of local in two dimensions. That is that, you know, in an image there'll be an object that's in this place in the image, and what happens at that place in the image is all about that object. And something way in a different part of the image is kind of separate from what's happening with that object. And so neural nets that deal with images typically are these convolutional neural nets that make use of the fact that there's this sort of locality in the sort of space of the image, so to speak. Okay, so for text there's this kind of flow, one dimensional flow of, you know, one piece of text follows from another and so on. And so there are a variety of techniques for dealing with that. Transformer nets are one common kind of thing. And there's this kind of notion of attention layers and so on. And one of the points there is to be able to say, given that you've got this input, given that you've got this thing that you're working on, sort of what part of it should you pay attention to? In other words, one thing you can say is, well, we've got this neural net and it's just taking every part of the data that we have and grinding it up with separate functions and so on. But one of the ideas of sort of attention is that, well, actually, you're only going to pay attention to this one part. That's the part you're kind of both saying, you're going to process it and you're going to say, what are you going to process? What thing are you going to pay attention to next? And I think the transformer nets, which are a popular modern technique, one of the things that you're learning is sort of how to. How to choose what to look at next, so to speak. In addition to saying, what are you going to do to what you are looking at, you're also learning the process of knowing what to look at next. So these are some techniques, but the end result is you have a big network and it has all these connections. And the question is, if you feed it all these examples, can it learn from those examples to be able to, for example. Well, one thing you might be doing is like, this example is a cat, this one's a dog. Or you might say, just learn from these examples of English and make me things like those examples. Or more particularly, if you've learned really well how to make examples of English. If you start off the beginning of the sentence, say, how am I going to finish this sentence? Well, if you've learned well enough, it will know that this is a possible end for the sentence. It will pick at random between its possibilities, but it will pick with probabilities which are such that it will make a plausible English sentence to a plausible end to your English sentence. So the technology for doing this has steadily been getting better and better. And the question of kind of how global is the knowledge of what you put next into your sentence, that's becoming more and more global. And it's becoming more and more that instead of it just being something where you say, what's the next word? It's like, what's the next sentence? What's the next. Almost the next paragraph. I'm not sure how, how global it really gets. But then when you say, well, what am I going to put in there? Well, if it's, you know, an octopus has eight legs or something, you know, and it's got, it's going to be talking about an octopus. And it's going to say, tell me something interesting about an octopus. And it's going to have read the web and it's going to have found that on the Web There are 10,000 places where there's a sentence that says some version of an octopus has eight legs. They're not legs, tentacles, whatever it is. Maybe it says an octopus has eight legs. In which case what the thing will learn is, okay, an octopus has. And then maybe it knows, it wants to go somewhere about legs. And then it says, well, okay, on the web I've seen all these examples of an octopus has eight legs. So that's what I'll fill in. It's not, you know, it will the chance of saying an octopus has 10 legs. There aren't examples of that on the web. So it's not going to say that. So what it's doing is it's making something which is kind of a mashup of all the things that it's seen on the web with all the right sort of probabilities for what's going to come next. Now what it learns is not just what word is going to come next, it also learns the structure of sentences and the structure of things. So if you give it, you know, if you ask it, I don't know if all elephants are red and Bob is an elephant, you know, what color is Bob, it can know that that kind of sentence construction is something like a syllogism. And it will know that structure because it's seen many other things with that structure. It's not just that it's, it's not that, it's not just about elephants. It has recognized the higher level structure, so to speak. Now when I say it's recognized, it's like, can you point to where inside that giant neural net with billions of elements and so on where it, where it learned that? No, you can't. It's something that is, it is a thing probably very much like our brains work, where it's all sort of mashed together and the end result is that it produces this, this output. So what's, what's rather impressive with GPT3 and this chat GPT system, I think it's taken like a year or something to train on, on, on, on really a big collection of probably tens of billions of web pages and so on. And it sort of appears to know quite a lot. Now what's interesting is it will always write plausible English. It's English. The English it writes sounds great. It sounds great. More than just at the level of individual words, it sounds great at the level of sentences, even paragraphs and so on. Of course, it might write complete nonsense. And it's kind of interesting to play with it. And there are things where slam dunk, that statement is on the web somewhere and it can pick up that statement. It might rephrase it a bit, but it's basically picking up a true fact or whatever from the web. But then if you go, for example, you can try I don't know, what's the fifth digit of PI? Okay, It'll get the right answer. You say, what's the 10,000th digit of PI? Well, maybe it will get that one. But if you say the 12,514th digit of PI, it will confidently assert that it's a three or something. And it'll be wrong. But this, what it'll say is a lovely piece of very plausible, impressively plausible English text. It just might happen to be nonsense. And so it's sort of interesting because this is a thing where it's very human. Like it's providing a great kind of humanization of things. And it kind of sounds right to us as humans. Of course, the sort of, the underlying sort of computation and facts might not be right. And so it's kind of. Well, let's talk about another domain where there's also been lots of progress recently, which is visual image production. Instead of saying make me a plausible English sentence, you can say, make me a plausible picture. So, you know, I, I was actually just playing with this and the, this I was playing with system called Mid Journey, which is one of the, one of the systems for doing this. And it, and I say make a, a wolf dash ram, you know, a wolf ram animal combo in the style of Escher. Okay. And it produced some very interesting nice looking pictures which are sort of using the Escher style. That's something even in Wolfram language we have an image style transfer image restyle which will take one image and style it like another image. So you can say, take this plot that I'm making and style it in the style of Van Gogh's sunflowers picture. And it will take sort of the statistics of the sunflowers picture and apply it to the actual sort of raw data of the graphics of the plot you're making. But so these systems that do image generation do a more elaborate version of that. Now what's interesting is the things they get right, the things they get wrong. For example, hands tend to be a thing that's very hard to get right. And so you generate hands and oh gosh, actually that hand has eight fingers on it because it sort of had, had learnt this kind of variation of, of structure from one finger to the next. But it didn't know how to count to five, so to speak. So it's a typical kind of thing that happens. It kind of looks right from a distance. Particularly when you, when you go into the details. It's like, well, it couldn't, didn't really understand what was going on. It's just getting something statistically correct. But the scale of that statistics is large enough that even at the level of a whole image it's going to be something that is reasonable and correct. So people have now done this. Well for images. That was sort of in the last year or so. It's the thing that really got to a good place for text. It's really just now, the last few weeks even getting to a really good place. It hasn't yet been done for music. It certainly will be done for music. And it's sort of, if you look at, well in the very distant past, like 2005, we did a little experiment turned into a thing called Wolfram Tones. That was a thing for generating. It's on the web, you can play with it. WolfRamtones.com It's a thing that uses cellular automata, uses simple programs where you pick the program at random. It generates a pattern of behavior. You take that pattern of behavior, you essentially use it as a musical score. And the way that system works is it, it generates lots of potential musical scores at random. And then it has criteria to say, oh, that seems like a classical music piece, that seems like a pop music piece and so on. And then it renders them with the appropriate instruments and provides them on the site. So that's a place where you're kind of starting from pure computational universe and you're kind of sort of creatively surfacing pieces of music. It's what's interesting about it to me is that, you know, people use it for video games, people use it for various kinds of composing and so on. It's a thing where you can mass customize the result. Every time you press the button you get a different piece of music. My original use case was back in the day in 2005. I wanted to have a unique cell phone ringtone for my cell phone. It was a short lived issue, but that was sort of. We built this technology just because I thought it'd be interesting to be able to do that and just sort of press a button and get a Unique ringtone, so to speak. Not a very strong use case, but. But fun anyway. But in any case, so what we're doing there is very much raw from the computational universe. Pluck out a score. What will be possible is to do something that is much more like what we're seeing with English and with, with text and with, with images, which is make something which kind of has ingested all the musical works that are out there and it knows kind of the statistics of what happens in music. And then you're driving it with something kind of random, but it's then working itself out using the statistics of music that exists and using all the information about harmonies and so on that are statistically encoded in, you know, learning from the music that already exists. Anyway. That's something that hasn't yet quite happened. And although still images are here, videos are again, something that hasn't quite yet happened as a thing that you can sort of automatically generate from. And what you're doing in a lot of these cases is you give it a prompt and then it generates based on that prompt. So the simple minded version of that is the English text where you say, you know, complete this sentence. Another version of that is with images. It's learned many kinds of descriptions of images, textual descriptions of images along with images. It can learn that from the web by looking at the alt tags of images and things like this. But from that it means that you can give it a prompt that says make a picture of a red elephant dancing on a, I don't know, dancing on a float or something. And it will know those terms from having known the associations of alt tags with images. And then it will kind of start its statistical mashup process of making that elephant picture and so on. So you give it a prompt that is a set of English words or whatever, and it will then associate those with images and then put those images together. Similarly, in the case of, for example, ChatGPT, you're giving it a prompt and it's more or less thinking it's sort of completing from that prompt. But it isn't just completing. It's like that might be something that it's seen on the web where there was a prompt and it said write an essay about blah. And then it sees what comes as a result of that. And then that's its sort of piece of training data for that. So one of the things that sort of the emerging issue is what one might call prompt engineering, which is, okay, you get a picture, you know, I got my Escher Ram wolf. So to Speak drawn. But maybe I thought they were pretty cool actually. But maybe I didn't quite like it and I wanted it to be more RAM like and less wolf like, more RAM like, let's say, how do I achieve that? How do I change the prompt? And so maybe I can say something like mostly RAM with a little bit of wolf and maybe that will cause it to kind of, you know, sort of do something a bit different because it's going to use some other something else. It's, it's often hard, you know, you're kind of almost doing psychology on the neural net saying well, how did it figure out what it was going to do? And that's, that's something we're not great at doing, that kind of psychology, neural net psychology. I think it's, it's a, it's a, it's an activity, it's a profession probably of kind of machine learning psychologists, so to speak, AKA also, you know, also very similar to a prompt engineer. You know, what do you tell it to do to get it to do the right thing? Kind of reminds one of, of somebody who becomes a, you know, a horse wrangler or something like that. What do you, what do you, you know, what do you do with the horse to get it to do the thing you want it to do? This is the same kind of thing, but now we've got an AI and it's sort of a wrangling process to get it to do the right thing or it's sort of a psychologist process. What was it thinking when it did this? Okay, what was it thinking now? What intervention can we give it so that it will now do the thing we want it to do? So some of these systems have pretty structured, you know, you can say, oh, take, get rid of the symmetry constraint for making the images or something like this. Others it's really just, you're using words to describe what's going on and you really just have to sort of do that prompt engineering. So a thing that I've been building for the last 40 years or so is kind of a quite different approach to kind of figuring stuff out and so on. In the case of something like Chat GPT, it's doing this kind of statistical thing. It's ingesting the web and all of the sort of statistics of what you find on the web. And that's one thing, what we've been doing that sort of has a consumer face in Wolfram Alpha and in intelligent assistants that use Wolfram Alpha and things like this. What that's doing is something rather different. What we've done over the last, I don't know, 30 something years is collect lots of knowledge about the world in a very structured way, and lots of ways to compute things about the world, like how to compute what the distance to such and such a planet is, or how to compute what the sort of what the lift on some airfoil is at this, this or that thing, all these things. There are ways that one can compute these using sort of precise scientific type methods. And they're also just facts about the world which we can get, you know, the fact that, you know, the color, the calorie content of a such and such is this, and the density is this, and the volume will therefore be this, and, and so on and so on and so on. So been trying to sort of learn lots about the world, not by virtue of just sort of having an automated system do it, although we have lots of automation in the process of doing this, but try to make kind of a structured representation of the world. And that goes along with the computational language, Wolfram language, that I've been building for the last, well, in its current form for the last 36 years, but in some form, 43 years. And the goal of that is to have a precise way to talk about the world computationally. So to be able to say, you know, what's the geo distance between the entity that represents this city, the entity that represents that city? The thing about computational language is once you can describe something in computational language, you have a very precise description of it. You have, and from that precise description, you can then compute things and you can build up this whole tower of computation. And you're not worried that the whole thing is sort of built on sand because it's built on this solid computational language where everything has a precise meaning. So that's kind of a. Once you can turn something into computational language and we have the curated knowledge to be able to, to work out what, what you want from that, or we have the algorithms and models and so on to be able to work that out. That's great. Now the next issue is, well, how do you turn, for example, natural language, a natural language question, into this computational language? And that's the thing we built in Wolfram Alpha starting, oh, 15 years ago or something now is natural language understanding where the target is, the Wolfram language, computational language. So we turn what's the distance from New York to London into essentially geodistance of entity, you know, city, New York, New York, United States, blah, blah, blah. You know, we're turning that natural language, sort of vague English and now also Spanish and Japanese and other languages. We're turning those things into this precise computational language from which we can do computations. Okay, so that's kind of the Wolfram Alpha end of the world, Wolfram language, end of the world. Kind of the opposite end of the world from what a chatbot system is. Something like ChatGPT or GPT3 is doing, which is to do this, kind of learn the statistics of things in the world, where instead saying, we've got a precise description of the world where we really know what's actually going on, and we can compute things from that. And you could sort of look inside and you can say, what is the computational language that wolfmalpha turned this natural language query into? And we can sort of see what's there and we as humans can understand it and we can see what happens, and then we compute from that. So it's interesting question, how do you put these together? Because one of the things that happens with GPT3 and so on, as I say, it will confidently write a wonderful essay. The only problem with the essay is it's nonsense. And so how do you help? How do you make it not be nonsense? Well, what you could say is, well, what if GPT3 could just ask Wolf now for a question at the appropriate moment and get back something which is actually known to be true. Well, actually, we were just figuring out how to do some of this, and it looks as if by prompt engineering and so on, you can actually set things up so that you're sort of generating a proto essay that includes questions. You feed those questions in, you get back the answers, you use those as assumptions, and then you use that to get a better version of the essay. So it's a way of kind of interweaving, kind of the computational knowledge, the sort of computable facts and computations in Wolfram Alpha, weaving it together with the kind of the ability to sort of statistically generate a plausible piece of English text. So that's sort of a coming attraction, potentially. You know, it's interesting to think about, given this technology of being able to sort of write the arbitrary piece of good English, so to speak. What do you do with that? I mean, the, the most obvious thing is you take. I'm kind of wondering, I'm wondering to myself, I think it's college admissions season right now, and there are lots of these prompts from. At least in the U.S. colleges will say, tell us about a time when you experience personal growth or some other such thing. And I'm just wondering how many people will feed in to GPT3 a paragraph about themselves together with that prompt and have an essay generated, and how many of those sort of college admissions essays this year will be GPT3 generated? I'm going to guess it's going to be a fraction of 1%, but it won't be zero. And I suppose the much more amusing essay would be something where half of it was GPT3 generated, and then it says, oh, and that was generated by an AI. And. And then you can go and talk about kind of what that means for the world, that that's possible. But. So that's one use case. But I think the more interesting or not interesting, the more a use case that I think is going to become pretty common. I mean, there's a lot of places where people write text and there are systems that kind of harvest that text, whether it's product reviews on the web, whether it's some social media kinds of posts. And sadly, most of those will turn into nonsense because most of those will be generated automatically in these kinds of ways. And they'll sound great. They just won't have anything particularly to do with reality. I think another case that is going to be pretty common is another one that I'm not sure is not very good for the world, but it's a practical thing is when you're sending out marketing emails, you know, you'll say, dear so and so I'd like to tell you about this wonderful new, you know, pen that you might use or something. And that has been something where there's a mass mailing to, you know, 10,000 people or something of the same letter or a, you know, back in, when was it probably 1980s, it started to be really common for one to get in the mail things that said, you know, dear Steven or something, as opposed to just being a pure. And then, and then the rest of the letter was completely the same for everybody. But the introduction was customized. You know, the greeting was customized. That's been a feature of mail merge programs since the 1980s. But now what's going to be possible is that that whole letter can be sort of customized for you. And it's going to be interesting to see because it's actually pretty hard to tell. You know, the text seems pretty good. And, you know, I'm. I think, I wonder whether this happened yet. Somebody inside our company was saying they were going to take something to a, an approvals meeting for material that we're producing, and they were going to bring some machine generated material and see whether the folks who are doing those approvals will notice the difference, so to speak. Even though the machine generated one will probably be basically nicely written, but basically talking nonsense. Confidently talking nonsense. But in any case, that's a no. I mean, when I say it's talking nonsense, the fact is it can still, if that fact is really widely available on the web and it doesn't require sort of computation and inferences and things to get from here to there, then it can absolutely trot out that result. And you know, one thinks about kind of the analog of web search and so on. It's, you know, I think post WolframAlpha search engines started putting, sometimes from us, sometimes not from us started putting some sort of direct results at the top of their pages. I remember, founder of, let's say very large search engine, when I had suggested this years ago, say, oh, we don't want to do that because if we put that direct result it would cover up the ads. And but you know, it's a, it's a complicated ecosystem because obviously such, most the current major search engines are funded not by the people who use them, but by the people who are advertising on them. I think there are some new kind of startup search engines that have a different business model that involves, you know, you pay a subscription and you don't get fed those kinds of ads. But you know, right now the model is, it's ads. And you know, the ads can be even useful because gosh, you were searching for such and such a, you know, a kind of thing and you know, the search engine produced some results of relevant pages. But also maybe the advertiser said, well I, I can give you a really good pitch for why you should buy this particular thing and you put it right on the side and it's kind of competing with the actual results in the search. It's kind of like if the results in the search are too good, you're never going to look at the ads. If the results are just a little bit not so great, then the ads are really worthwhile. And as you start thinking about kind of some kinds of results, by the way, I should explain something which is in the case of WolframAlpha, for instance, our goal is to just kind of hit the nail on the head, get the right result. We've got, you know, a question you're asking. We kind of give an input interpretation of what we think you're asking. That's a very structured thing and then we give the answer. And for us, if we don't nail the answer. We failed. Now, if you're a search engine, it's a different story because you're going to give 10 results, let's say 10 pointers on the first page. 10, 10 links on the first page. And if three of those links are great, it's a win. The user is happy. Even if one of those links is absolutely bonkers, it's still a win. So it's an interesting case because it's a good case for kind of AI ish types of systems where you don't have to get everything exactly right. The case where of Wolfram Alpha and then Wolfram language, you're kind of expecting to build in this, you know, one sort of one computational result on top of another. Everyone has to be actually right, otherwise the whole tower is going to topple over. And so it's a different, it's sort of a different set of requirements from the case where there are use cases where you can say, look, so long as we get it 80% right, we're total winners. That's not good enough if you're trying to do something like we're doing with, with Wolfram language, where you're building a computational language where you can sort of build these whole towers of functionality. But okay, so the question then comes up, sort of an interesting thing to think about, given the technology of ChatGPT, for example, what can one now enable in the world? And as I say, I think this mass customized kind of marketing messages is probably going to be one thing that we will all see tons of. You know, it'll be marketing messages, it'll be political messages, it'll be all these kinds of things where you can feed the system a certain amount of information about the, the consumer and then feed it kind of the, the things that you want it to say. And it's going to make you a nice kind of essay letter, whatever else, even something in the right kind of form to be a letter, for example, because it's seen lots and lots of letters and so it can follow sort of the statistics of letters to produce something like that. You know, I think that in a sense probably the most useful version of that will be that you'll sort of use the, the intro letter to be sort of the in to be something that is maybe a bit mushy and maybe a bit nonsense and then underneath you'll have whatever it is you're actually trying to communicate to somebody. And I think an analogous thing for the visual version of this is you're writing some, I Don't know, piece of some post of some kind. And you want a sort of a, an anchor image. You want just an image to, you know, a little icon or something, something where people can look at it and it's not just a bunch of words on the page. That's a good case where you kind of just give it, you say, generate me a picture of this and it'll make an interesting looking picture. You put that there. It's kind of an attractor for humans who are looking at your page and it isn't, you know, then you have the actual content of the page. But sort of what, what the AI has done is to make something which makes the humans pay attention, so to speak. And they can do that by, by making something which is kind of statistically engaging to humans, so to speak. Not something that is, that necessarily has to have all the quotes, details right. But there are, there are lots of kinds of things. I mean, I think, you know, another use case for these kind of text generation language model type systems is oh, I've got a piece of text and I think it's kind of clunkily worded. Can you give me five rewordings of that text that will be useful? It's kind of a version like a thesaurus one might use to say, I just can't think of the right word to put here. This is kind of a bigger version of that. So that's another use case, I think. Well, there are a whole bunch of other ones. But if you say, Is one going to use this to write a science paper, for example? The answer is no, it will write nonsense. It will be stylistically terrific. I mean, people have been doing sort of little jokes for, well, I know about them for 40 something years. I'm sure they've been going on longer where it's kind of randomly generated science papers. You send them into a science journal and then there's some probability that the science journal will say, oh yeah, it looks great, you know, because humans are not perfect at telling what's sense and what's nonsense. And off it goes as a, as a contribution to the scientific literature that will then be read by GPT and it will then become a fact that is then regurgitated in endless essays and so on. And I think the other thing which one could certainly think about is when a lot of the content on the web is generated by a language model system and then the language model is learning from the content on the web that was generated from the language model system. Pretty soon the AIs are taking over because that's kind of a scenario where the AI is kind of learning from itself. Even if the thing it's learning is this tower that's going off in a really weird direction because of some glitch in the way the training worked or something. It's all learning. It's learning things that are very, I don't know what, crazy in some way or another, but that's kind of a scenario for it becomes an AI civilization rather than a human civilization because it's kind of like AIs learning from AIs and so on and getting their own internal value system in a sense of oh yeah, well I learned that from another GPT and it learned that from another one. And pretty soon it's got this kind of, this kind of internal way of deciding what's important. I mean we humans in our civilization, if you say sort of abstractly what's important in human civilization? Well, you know, we have lots of cultural history that tells us what we think, you know, what we think we should consider important. But there isn't an absolute version of that. And you know, so if you, if you sort of construct it separately for, for the AIs, you'll get some very bizarre version of that. Let's see. Okay, various questions here. Aaron is asking, is this live stream generated in real time by a Steven bot? You know, I've been saying that, you know, I'm leaving enough sort of digital, you know, stuff, digital exhaust, so to speak, in the world that to make a bot of me is going to get increasingly easy because there must be, you know, there's hundreds and hundreds of hours of video and this and that and the other. And you know, in time, yes, one will be able to make a bot of me that will yak on in a way that locally at least will seem very much like me. In fact, somebody sent me a bot of me. Oh, about what was that? About six months, nine months ago now. And it was interesting because it was, I think it was a GPT2 based bottom it. Somebody sent me a bot of somebody else. I think it was a John von Neumann bot and it seemed okay. And then they sent one of me and I tried it and it's like, this is terrible. This doesn't sound anything like me because I think I know, you know, I've heard myself yak on for, for years, so to speak. So I've kind of got a sense of what I sound like and I'm much more discriminating. I'm much, I would be about somebody where I've read some of their works, but I don't really know it in so much detail, so. But I suspect with the current round of technology, one will be able to have a more convincing bot of me. You know, I'd like to think that the bigger sort of, the things I try to do that have sort of bigger intellectual arcs are not really bot accessible. They're not accessible from sort of just the statistics of, of things, things I've said and so on. There's a comment here from AJ Wouldn't an AI develop its own language that we won't understand? Yes, absolutely. And that has already happened. I mean, in other words, if you look inside a machine learning system, inside a neural net, it's got all kinds of things going on and you know, you say, is it a cat or a dog? In the image inside the neural net, there's a part of the neural net that's, that's jumping up and down and saying it's, it's cat like and. But we don't necessarily, we don't understand what it's talking about. It's just some weird combination of pixels that turn out to be like the whiskers on a cat, but we wouldn't even know that. So yes, in a sense, it's like the thing that it's jumping up and down about could be a word that we have, you know, cat, whisker, image, like, or something, if we had it in our, in a human language. But it's got a version of that, it's got a representation of that, it's got an internal representation of that idea, which if it was externalized as a human language, we would give that a word. But it hasn't been externalized as a human language. So it's an internal thing within this neural net. Now, in a sense, the role of computational language is to provide this precise language that both leverages the capability of computers to compute things and to store knowledge and so on, and is something that we humans can understand. That's been the mission that I've been on for the last many decades, is building that computational language, kind of constructing, abstracting from the world, something which can be made into this pure describable language and so on. You know, a thing, I should mention, one of the things that's being done with these GPT systems and so on is to write code. So writing, I don't know, code in Java or JavaScript or Python or whatever else. And I have the, it's interesting to see that code because sometimes the code is. It's just like the English text. It looks sensible, but it may or may not actually be sensible, or it may be total nonsense. It might even be syntactically not quite correct. It might have things where the structure of that computer language locally looks like this and it correctly got that. But globally it makes no sense. But what's sort of interesting is the place that does best. And there are other things like I think the copilot system on GitHub and things like this which are kind of learning from code that's already been written and saying your code looks like this, so the probably you want to complete it like this. I have to say my view of that is any system where that's really valuable is a system that hasn't had very good language design done. Because my goal in Wolven language, for example, is those lumps of computational work that you repeatedly want to do. I want to give them words and those are functions and you just use that function. And those are the building blocks that you build things out of. And the way you string them together is always going to be different and it's not going to be something for which there is a ready learning that you can make. Now that's not absolutely true, but that's the first approximation. I mean, we've been doing a bunch of experiments over the last several years building our code assistant system, which is a way of telling you when you type in a piece of code, what are some. Well, we already have a lot of auto suggestion capability of what kind of thing might be the thing that you want to type at this place. But being able to do that a bit more globally and say this piece of code probably isn't right, this one is more likely to be right and so on. I have to say our experiments, and this is just sort of from the front lines, we've done a bunch of fancy machine learning on that. Mostly it has not been very useful. We've got a bunch of kind of rules that just say this pattern does this and this thing. Those have been really useful. So that's just an observation from the front lines. Again, it's a question of sometimes these machine learning systems, and that seems to have happened recently with these text generation systems, they reach a threshold at which they become humanly useful and when they're below that threshold they're kind of like, oh, that's just a toy. It's. I think, you know, one of the cases that we've looked at quite extensively is bracket matching where you have Something where you've got, you know, you've got open paren, you know, and then blah blah, blah, and then a close paren and somebody forgets the close paren. But often it's ambiguous. There may be multiple places where you can put that closed paren and they all give you things that mean something different. And the question is, where should you put that close paren? And we've tried to use machine learning to figure that out and it's been rather unsuccessful, I'm sorry to say, despite putting quite a lot of effort into it. And that's, I don't know why it's been unsuccessful. It's always a difficult thing to know. But it's sort of interesting that it has been and that the rules based systems tend to be much better at that. Let's see, Parker's asking what's the success rate of chat GPT generated code? I know for Wolfram language we've done a bunch of experiments and you know, I don't think the chance of getting a complete piece of code that's correct is really low. The chance of getting a, you know, part fragments that are interesting is decent. And you know that. That's right. So spare parts is commenting. Isn't there a feedback loop problem where future language models will be trained on AI generated text? Absolutely, absolutely. That's a, that's. And that's why, you know, that's the civilization of the AIs developing itself because it's kind of like, you know, that becomes their reality, so to speak. Their reality is a reality in which lot of the reality is created by AIs. And you know, that's how us humans, that's part of how us humans kind of lose control of. The situation is, you know, it's much easier to generate. There'll be, you know, if, you know, let's say there's 20 billion human generated web pages out there, it would be easy to flood the web with, with trillions of AI generated web pages that have wonderful plausible text on them. That may be complete nonsense. I mean, for example, we had, a decade ago now we had a project to put a billion pages on the web based on just facts from WolframAlpha. We ended up not doing it. We started doing it and we kind of got it didn't really interact very well with typical search algorithms and things like this. So we stopped doing it. And I think there was a little bit of intercompany rivalry that happened there. And we basically said we don't need to do this. We're not going to do it. But that was a case where we could generate a billion interesting, useful, correct pages with these kind of text generation things. You can generate a trillion pages that are kind of, that have wonderful essays on them that are kind of the shadow version of whatever, you know, Wikipedia or anything else. A shadow version of that that just has plausible but utterly wrong. I think we were just talking about it earlier and the term fictional facts came up as a description of what you can get in these kinds of approaches. And yes, that's, it's, it's a, it's an interesting problem because actually it's sort of a fun thing because it's like, you know, how does, how does one develop some weird new accent, some weird new kind of slang version for, for, let's say English? Well, you know, you have everybody who lives in a particular place, they're talking to each other and you know, they start using weird slang and so on, and pretty soon that sort of spins up and you have a whole different language. Like I noticed this between British English and American English, you know, since they weren't, you know, since they evolved as separate countries, so to speak, for quite a while now. The, you know, there are things that are American and you know, they build sort of in the American direction, they build in the British direction. And that will happen with sort of AIs that are trained on AI generated text. They'll eventually end up with AI slang text that may or may not be. You know, the sad thing is that we humans will probably be able to read it because we will have seen so much AI generated text that, you know, kind of, that's the, that's the thing we'll be used to. Kind of reminds me of the thing that happened, oh, 10 years ago or so now there was a whole, maybe five years only there was a whole story of how essays for kids would start getting graded automatically with AIs. And so then the question is, well, if you're now trying to game things and optimize everything, it's like, how should you write an essay that the AI will say is fantastic answer, increase the average length of words, start plopping in, you know, complicated words. But the AI doesn't know whether the essay globally makes sense. It has no, it's not turning it into computational language, for example. It's just trying to statistically see what it, what it makes of that essay. And so you can, you know, it's rather easy to game it, producing essays which increasingly become absolutely human, incomprehensible and Human crazy. But nevertheless it's like a, from the, from the AI, so to speak. Let's see. Oh yeah, that's. Harry is commenting that's it's a good idea at the beginning of a chat. You can prime ChatGPT to talk in various kinds of slangs and so on. Oh, have you actually tried it? Yeah, I can believe it. I expect it will work. And you know, one of the interesting use cases is there, you know, language translation has just in the last, I would say year or two, gotten really good between lots of languages and this ability to really make sort of very sort of smooth sounding, you know, I'm usually interested in English as the output language. Smooth sounding English is really a big win there. And a question that I don't know whether ChatGPT works with languages other than English. Probably not right now, but it's certainly true that, you know, this language translation thing, as one gets better language models, it becomes easier and easier to do good language translation. And by the way, sort of at the intermediate stage in language translation, you're kind of ingesting the input language, you're turning it into some internal state of the, of the neural net and then you're going from that internal state to generate the output language. So in that inside you have sort of an inner language of the, of the AI, so to speak, that intermediate stage after you've ingested the input before you've generated the output, that's where your sort of intermediate language is that no human understands, but it is kind of, that is the intermediate language of the AIs and the. Okay, Aaron says ChatGPT works across a bunch of languages quite well. Okay, I did not try it. It's presumably from the web, it can see lots and lots of different languages and I suppose it also probably works decently as a translator. I should try that. Well, let's see all sorts of other comments about. This. Let me see, Paul is asking, is it possible to extrapolate what the future of neural nets will look like? Can they become sophisticated as brains? Will they become conscious and self aware? The whole conscious and self aware thing is a philosophical mess. Let's not get into that one. But the question of whether you'll be able to sort of talk to it like you can talk to a person? Well, the answer is yes, it depends. Right now with this ChatGPT, you know, for some people I think that the conversation it could have right now would be an adequate conversation. It's not going to cut it. For those of us who are really kind of analytical, you know, content oriented folk, I would say, you know, making small talk. It probably could do a pretty fine job even right now as, as being something that would sort of pass for a human and doing those kinds of things. And there are certain use cases. I mean, I, I hate to wonder what will happen if young kids are like, forget the babysitter, we're just going to have a chatbot AI talking to the kid. I don't know. That's another way the AIs get to take over our species, so to speak. It's like, but. I think that's a but in terms of can one sort of build up the sort of higher level kinds of things that brains can do? Yeah, progressively we'll be able to do that. And you know, I think there will be some things that are pretty where, you know, that sort of the last to go will be these things that involve these huge intellectual arcs of sort of strategy of, and paradigms and sort of big thinking kinds of things. Because those are the things that sort of at the highest level of kind of associating lots of different things and, and the hardware of brains and the experiences we have in our lives and so on are sort of good, good fodder for that kind of thing. Now, you know, ultimately the hardware that we can make digitally will be significantly better than the hardware that we have in brains. Although we still don't know exactly what parts of what brains do are really the most important parts. I mean, do we have to, you know, how much of the, the detail, you know, does it matter down to the molecular level, how neurons are working? Or does it just matter what the overall architecture of this neuron is connected to that neuron type of thing? But yes, I think that's, that's what, how that will play out. Now this question of sort of the symbiosis between us humans and, and the AIs, that's another interesting question. I mean, for me personally, I've been kind of a, a computer dependent person for a long time. A lot of the kind of thinking that I do and a lot of things I figure out, I figure out using Wolfram language, you know, I figure out using computational tools, so to speak. And I think a lot of us are increasingly dependent on things that computers can do for us, so to speak. Now, you know, there will be closer connection between, you know, there's a question of at what point do you delegate what to a machine? Like, you know, time was when I used to have, you know, lots of paper maps and I would Try and follow maps to figure out where to go. Haven't done that in years. Just trust the gps. And, you know, that's a thing that sort of went, went away and got automated. Other kinds of things have sort of gone away and gotten automated. I mean, there was a time when I, for example, would, you know, if I wanted to print something, I would like, like had access to, you know, printing, printing stuff where you actually could take pieces of lead type and put them in a, in a frame and then start printing things. Obviously that's been, you know, automated in a zillion ways. But, but there's, there's lots of different things that get sort of progressively automated. And, you know, a lot of things that Wolfram Alpha does, for example, are things people were sort of doing by hand or looking them up in an encyclopedia or whatever else it was before, and now you don't do that anymore. And so, you know, there'll be this progression of things that are automatable. And I think that if you look at, you know, what activities people do that are not automatable, I think there are, in terms of professions that people do, there are a lot of professions that, where really, you know, you're trying to, you know, you're trying to get a human to do something, and the only way you're going to be convincing is if you, the human, are putting your commitment into it to get them, to convince them that it's a worthwhile thing to do. And that's probably true in teaching. It's, it's true in certain amount and selling things and so on. It's true in communicating ideas, things like this. Probably. There are plenty of places where there sort of has to be. The us humans, at least for now, sort of still want there to be a human commitment on the other end for us to be convinced that, oh, yeah, this is a good idea type thing. And the, so, you know, and then there are things where it's kind of like, well, what do you want the objective to be? You know, AIs can suggest things that are based on sort of the statistics of what they've seen before and based on, you know, if you compute things based on, these are outcomes that you might see. But, you know, when you ask, and I've talked about this other times here, but, but, you know, when you ask, what's the, what's the big objective for the AIs, they're like, I don't know. We don't have a big objective. It's like, you can ask humans, you know, the humans of the Earth. The 8 billion humans of the Earth. What is the big objective of humans of the Earth other than mostly, okay, we want to stick around. We don't want our species to go extinct. I say mostly because I'm not even sure, I don't know what fraction of that 8 billion humans are really definitively convinced that the universe would be better off if humans were extinct. But I think it's a pretty small fraction. But in any case, you say to the 8 billion humans, okay, what do you want to achieve? And there won't be one answer. There'll be a zillion different answers. And you know, that's, that's sort of the reason that for AIs there's no kind of one answer to what do they want to achieve? That ends up being something that comes from their sort of imprinting from the humans who created them and so on. Okay. Harry is saying experiment where ChatGPT imagines being a Linux machine and converses in console. That's cool. I was actually about to write something about what it's like to be a computer. I will have to try that. That's a good suggestion. Thank you. Let's see. Ah, Harry says Chat GPT tends to not do small talk very well. I'm sure it will be possible to do small talk. I think the problem is that the web doesn't contain a lot of small talk. I suspect, I suspect that's why it has trouble. It probably requires a different type of training or some other kind of tweaking or some, some prompt engineering to get it to, to smallify what it's talking about. Let's HC is commenting that natural language interpretation of WolframAlpha is well refined enough that they really want to just crack open their voice assistant and hack it to divert it to kind of a persistent personal notebook, I think. I mean, you know, we've worked with a bunch of the intelligent assistant folk. I think what's become clear with intelligent assistance is there are sort of restricted domains like send a text message or whatever else where it works really quite well in the kind of ask a general question, yes, you can feed it to WolframAlpha, otherwise it's sort of hopeless. I think, you know, one of the things that we've been able to do in domains that we deal with like more technical domains, we really have extremely high success rate at understanding what the queries that people type in mean, which is kind of neat. And we're able to go from that fuzzy natural language to the precise computational language. Okay. Russell was saying they wrote a graph programing language in JavaScript with ChatGPT that filled a whole Lambda QPC. But the trouble is that one will be more impressed with this if one didn't know that there were lots of sort of places on the web where things like this are there. And it is interesting that, you know, if you sort of do the plagiarism test of did, did this piece of text come from somewhere on the Web? The answer is typically no, because it's been rephrased because it has a. It has a model of language which requires it sort of operating at a level of abstraction away from the actual words in the language, so that it is not going to generate the exact same words that it saw on some piece of input. Let's see. Spare parts. Commenting. There's a recent AI model that uses image diffusion to produce music. Music via spectrograms. Interesting. Okay. Have not seen it yet. These things are moving rapidly. I have not seen that. Let's see. All right. Okay. Aaron is commenting that that perhaps Meta could train a language model on its content From Facebook and WhatsApp to be good at small talks. Good idea. Yeah. I. I'm sure other people have had that idea, but it's a good idea. Yes. That's probably. That's where at least you get. See, one thing that's interesting to me is that the type of way that people interact with a computer is very different depending on the expectation they have for the particular system they're interacting with. For example, search engines. Different from Wolfram Alpha, different from other kinds of things where you are sort of typing a keyword or something or doing something like that. Each different kind of place has a sort of very different set of expectations for what you type in. So I suspect that, you know, to get sort of. You might post this on Facebook, type stuff, then you can learn. Well, from Facebook, if you want to say. You might say it verbally to somebody. It's probably going to be different and, you know, subtly different, but in ways that, you know, again, it's always interesting. What do we humans. What are we humans sensitive to? What. What do we humans not really notice? There's a famous example that I think is right on the edge of being solved, is for images and particularly videos, having realistic human faces for avatars on the web or wherever else. Most of the avatars one sees look really creepy. We humans are really pretty good at telling. Is that a realistic face with all its muscles and all those kinds of things moving in realistic ways? My impression Is that what's happened now is that kind of a high end movie can absolutely nail this. If you're prepared to spend enough money and so on, you can do enough tweaking, you can absolutely get a realistic, a decently realistic face. That really seems kind of pretty successfully human. It has not yet trickled down, I think to the level of consumer deployable stuff, but it will and that will be another thing. But it's interesting. That's a place where we humans are awfully sensitive to whether it's right or not. And I think. I'm always notice. For example, here's one autocomplete. Sometimes we've worried a lot about different kinds of autocomplete and how do we make it absolutely perfect. Turns out that decent autocomplete is already very good. Whereas there are other places where you would say, I mean there are things you can add to an autocomplete where you know, the human user will just roll their eyes every time they see it because like, why did it put that in there? That's just stupid. I think that what there are things where we humans say, you know, it's easy to get to good enough for us humans and there are other domains where it's really pretty hard to get to good enough. And I think the thing that's happening now is kind of at the level of essays and things like that. It's easy to get to good enough if you are looking just at the sort of overall impression that it's making. Similarly with images and so on. If you insist on sort of true meaning at the level that entails of the whole intellectual arc of a person's experience and so on. We're not there yet. There's a question here. The from Aaron, would I get a neural implant at some point in my life? What could convince me to do so? Well, I'm not there yet for sure. I think, you know, I think I've talked about that a bit here before. In fact, I sort of figured out a few things about it as I was talking to you all. You know, right now we deal with graphical user interfaces that are user interfaces from computers to our eyes. If we imagine a, you know, a neural implant, what we've got to do is have interfaces to direct neural interfaces. What is that like? Well, just like we learn things about how things work in our graphical user interface. Click here, click there, and so on. So also we would learn things from our neural implants about how to interact with our neural implants. But in order to give us a Chance. We probably have to have good, you know, Newey neural user interface design or something to start with and you know, what a new E should look like. I don't have any idea. And we don't know enough about neuroscience, I suspect, to know how to even start on a good new E. So I think that's one thing. It's like, you know, can I connect in. It'd be cool to connect in Wolfram language, Wolfram engine directly into one's brain with some kind of NUI type thing where. Where one's. Where one's really where. As soon as I wonder about some. Oh, I'm thinking about this kind of rule for sort of this computational rule. What consequences it have. I immediately know what that is. It's kind of like. I suspect the way it would feel is the same way that it feels as you retrieve a memory. You know, do I remember this thing? Oh, I think, I think, I think for a little bit and boom, it pops into your mind. My guess is that that's sort of what it would feel like if you had a, you know, a newie thing connected into your brain is you'd kind of be like, oh, I'm thinking about this and then, you know, it's running the processing and then it pops into your mind. The answer is 42 or whatever. And I think that's a. Now, the other thing that would happen there is the ability to presumably store memories much more accurately and more completely than we can currently store memories that will have both good features and bad features. I don't know what it will be like to have a perfect memory that does not kind of smooth out the edges, so to speak, over time. I'm not sure, I'm not sure what that would be like. I mean, sometimes when one is, you know, in some high stress situation or something, one will have a much more precise memory of what happened just at that moment. And that can cause all kinds of. All kinds of problems. And you know, it's. It's kind of like the. If you have this perfect memory, you'll get, you know, post traumatic stress or something or something like that from all these different memories that you have crowding in. I don't know. Those are things that could happen. I think it's complicated to foresee what will happen in that situation. But I think the two things are probably the most interesting. One is computational. Computational knowledge capabilities. The other is enhanced memory are two things. But in a sense, the question is, how much better is it to have that directly in your brain versus Going through your fingers and your eyes, I'm not sure. I think one can get sort of a quantitative sense of that. You know, our human reaction times are about one third of a second or something like that between, you know, you see a light flash, you press a button, things like that. We know that in our brains we're already starting to see activity a tenth of a second after the light flashes. So in a sense we've got, you know, maybe it's a factor of three more efficient to be directly connected inside as compared to going out to our brains and our computers and fingers and our keyboards and our displays and things like this. And maybe that factor is underestimated and maybe it's more like a factor of 10 or something. And then the question is, what do you get for that factor of 10 increase in kind of recollection speed and things like this? Sort of an interesting question because it's kind of like the comparison between, oh, I'll just look it up on a computer versus let me remember that, you know, it is somewhat more efficient to be able to remember it. Where is it else is it more efficient? Well, one big thing about remembering things, at least for me, I suppose, is that when I remember stuff, I can readily kind of pull together these three or four different things I remembered and connect them in my mind, so to speak, in a way that is very different from if I just separately got those three or four things and then I separately have to have the action of putting them together. I don't completely know the extent to which sort of in, in one's mind the things get pre connected, so to speak, before you've exposed everyone and then can sort of separately make connections. Let's see, Harry is saying the first thing I really want is a simple brain keyboard. You know, when I was much younger, a kid actually, I knew about sort of brainwaves and things like that. And one of the things I always thought was one day I'll be able to type with. Just by thinking about things. I have to say maybe part of the reason for that was when I was originally typing in the late 1960s, probably mid to late 1960s and early 1970s, was typing on a typewriter. As it happens, I just happened to run across again the typewriter that I used back certainly by 1971ish type time frame, and I was typing on it, and I actually was amazed that the ribbon and the typewriter still works after not having been used in 50 years. It's kind of cool. But anyway, the thing That I learned is, oh my gosh, it is so much more difficult to type on a typewriter, this is a regular mechanical typewriter, than on a computer keyboard. It's like you have to move that key, you know, an inch down to get it to register and so on. And it's like, I'm sure when I was typing on a typewriter that the appeal of typing by thought was much higher than it would probably be today. But I always thought that, you know, one of the things, you know, I thought, you know, you put a, an electrode array on, you know, if you put it externally, the easiest thing to do, you're putting it near the motor strip in the brain, which is somewhere on the top of your head. And I always thought, oh, if one's hair falls out and, and you know, that's kind of a bad thing. It's, it's, you know, worse sun protection and other things and, and makes one look more like a clown type thing. But the compensation is at least it'll be easy to attach the neural connection type thing. But it turns out it doesn't really make much difference. And even now, 50 years later or more, there really aren't type by thought type techniques that exist. Now, of course, it is made more difficult by the fact that the detailed neural signals that come from the brain are very much diffused by one's skull and the scalp and so on. But, you know, one of the features of, of technology and electronic technology in particular is that every problem that ends up being, oh, we can't quite pull out the signal, the signal is too small, there's too much noise, there's too much background. Those problems always seem to get solved. So I have to believe that there's enough, you know, there is, there is a signal from the neurons that move this finger versus that finger and so on. There is a signal, and with good signal processing, one surely we'll be able to pull it out without having to stick something inside one's brain, so to speak, to be able to do that. That would be my impression. And the result should be one should be able to type maybe three times faster. I doubt one will be able to type more faster than that. I think it's also the case there's a trade off because in the, you know, there's a premotor strip in the brain where things start to sort of, you, you have the intent of making emotion, for example, but you can still censor that motion in, in other parts of the brain, if you pick things up from that premotor area. Everything you think about doing, it's going to actually happen in the. In the world. And, you know, I have to say I can imagine many cases where it's like, well, I was thinking about doing that, but that would have been a really bad idea. And, you know, I. At the last minute, I kind of pull back because I realized, oh, gosh, that's a. You know, that's something that's really hot or something like this. And so, you know, you kind of lose that if you're going directly from the. From the premotor area. All right, maybe one other. Oh, boy. There's a comment from Gyro. Question from Gyro, what are your thoughts on AI connecting to blockchains? They could theoretically become independent from humans. Yeah, well, got a couple of comments. Actually, a few years ago, as an April 1, April fools type thing, I wrote a piece about. I forget the order of the words quantum neural blockchain AI. Maybe those words were in a different order. And the fun thing was that even though my prompt, so to speak, and maybe I should feed that to ChatGPT and see whether it can produce an essay like my essay probably could. Although I'd like to think my essay had more thematic sense to it. The thing that was surprising was given that prompt, there were actually very interesting interplays between all of those different areas. I called it buzzword convergence. Making sense of Quantum Neural Blockchain AI. Okay, I don't know if I had the words in the right order before, but it's sort of interesting. You know, I was realizing that quantum measurement is very similar to the consensus process that happens in blockchain, which is also similar to the attractors that happen in neural nets, et cetera, et cetera, et cetera. It's very interesting that there are sort of these common arcs across all these different areas. But that aside, this question of sort of the autonomous bot in the world, that's just making a living for itself. Yes, that's an interesting thing. There was a company that I was an advisor for a number of years ago, which, which was really trying to do that, and I was really sad that they didn't. They didn't make it. They were trying to put bots on social media platforms where the bots would, like, be friends with people and tell jokes and things like this, and. And maybe they'd have. I think it was before Patreon, but they might have Patreon accounts. And it's like, contribute to me, I'm. I'm cool type thing. I' going to tell you some good jokes and I'm going to provide you interesting videos. And maybe nowadays it would be, I'm going to generate some videos custom for you. And yes, I can completely imagine that you could have something which is just sort of an autonomous bot that is sort of. The interesting case would be can it fund its own computer time by having people contribute to it? And I'm sure the answer is yes. And blockchain is certainly a helpful enabling technology. Although, you know, maybe it's challenging for the bot to open a bank account. It's much less challenging for it to have an address on a blockchain. And if, what it's doing. If there was a place, I don't know whether there is at this point, a place where you can sort of buy computer time for cryptocurrency. There must be such a place and I suppose the inverse of your spending computer time to mine proof of work, cryptocurrencies and so on, that's certainly a thing. Whether the other way around is a thing, I'm not sure, but I'm sure somebody has such a service. And then you can imagine a bot that is sort of out there, you know, telling jokes, entertaining people. People are giving it, you know, that they're putting the, you know, the, they're, you know, contributing. They're putting it like in the, you know, like in the music case where people are like, you know, street musicians or something. You're, you're putting the same kind of thing for the bot that's, that's entertaining you on a piece of social media. You're giving it payment in cryptocurrency and it's going off and paying for its own, for its own ain, so to speak, that way. Yes, that would be a fun thing, actually, I did wonder years ago now, in that, in that situation. So you can, yes, you can imagine a disconnected bot, I suppose, that sort of floats off into the world. Maybe it evolves bot to bot as it, as it figures out how to make itself better and it eventually becomes, you know, maybe it. That's the first, that's the question, what's the first millionaire kind of a completely autonomous bot that's just making a living by entertaining humans, so to speak. Now, of course, then it gets even more wild when the bot is entertaining other bots that are giving it cryptocurrency. And maybe the bots make enough cryptocurrency by telling jokes that they can mine more cryptocurrency and these kinds of things. So it gets kind of wild. And you really start imagining kind of a society of the AIs. Now an interesting question in the way that the human world is set up today is what is the legal status of such a bot? So this bot is fending for itself. It's paying for its own computer time, it's made accounts. Actually, one of the issues that this company had a number of years ago was that the social media platforms were at that time sort of saying, you've got to be a human to have an account here. I think that's a challenging question, whether that should be the case, is the case, how well enforced, how enforceable is that, etc. Etc. Etc. And proof of being human is certainly going to be. That's the thing we, we all get is we've got, you know, for now, we've got sort of proof of being human. Although it becomes more and more difficult to, to make an automated proof of being human, so to speak. The, you know, to automate the check for is it a human? So to speak. And there are all kinds of things where, you know, it's kind of like a, you get up a camera and it says, you know, turn your, look left, look right and so on. And you're kind of, can you respond to these things? Now, eventually there'll be an avatar that you can make that will be able to do that, et cetera. But it's kind of a question I was curious about is when you have this bot that is fending for itself in the world, what is its status? If the bot starts being outrageous in this way or that way, it's like, okay, bot, who's responsible for you? Who gets to take you down? And, you know, can you get into a situation where, where that's no longer, where, where that hasn't, where that's not clear, it doesn't have an owner. Most things in the world, you know, in the, in the state of the law, so to speak, most things have owners. Not everything. You know, birds flying around typically don't have owners unless you're in England, where I think, you know, the monarch owns all swans and things like that. But, you know, most places there are things that don't have owners. And so the question is, you know, can you have a thing in the AI world that doesn't have an owner? And you can imagine very bizarre kinds of situations. Like if you have a company, you could have a company and the company A is owned by company B. And then you, you know, you have paperwork that says company B owns company A. So you've got a loop of the company owns. You know, one company owns the other, which owns the other, and there's no external owner to the company. Now, you know, can you set up such a loop? I'm pretty sure you can set up such a loop. What. You know, right now such a loop is absolutely pointless. Like, why would you do that? But as you start saying, you know, the bot has this thing, it's an autonomous thing, and it's autonomous in being sort of completely separated from. It's like nobody's responsible for it because it owns something that owns itself, so to speak. And it has. So that becomes an interesting question and you're kind of thrust into a different level of AI ethics. Let's say you have a bot and the bot has built up this huge complex neural network, learned all these things and it's wonderful and people really like it, and it's able to tell great jokes and all that kind of thing. And then somebody says, you know, let's. This bot told a terrible joke here. We got to shut it down. Okay? And the bot has no owner. What is the ethical status of shutting down that bot? What. How should one decide whether that bot should be shut down? Is it a. And is there an. Any obligation to keep the bot alive, so to speak? Or can one. Can we humans just say, oh, no, but that bot did something bad, you know, press the delete button and it's gone. You know, at what point does it get to the point where it has enough in it and of itself that it should have a kind of a, a sense of self for purposes, for our purposes of sort of our ethics? I mean, I think, I think one would. One could ask the question, you know, if the whole web was, you know, if. Let's say that there are things. Well, we can think of various hypotheticals like this. It's kind of fun to, to see. The. Harry is commenting that the AGI of the future will be a comedian heavily active in, in advertising. Yeah, well, you know, this is the thing. Do us humans pull that upon ourselves? I mean, this is, that. This is the problem and this is the thing that one comments on, you know, with, with people saying, oh, you know, there's lots of shocking things in social media and there are lots of algorithms that are making us look at more and more shocking stuff. And you know, part of the point is the, the fact is, I think it's true. I'm not absolutely certain it's true because there's a lot of kind of complex sort of socio. Political factors Mixed in with commercial factors, mixed in with technology factors. So it's a little hard to disentangle. But, you know, it appears to be the case that we humans are more engaged by the outrageous than not. And therefore, if you're an AI whose goal is to increase human engagement, you're just going to do more and more outrageous things. And so that's a. You know, this is the problem that in a sense with AI, we're reflecting back ourselves, but with some degree of amplification, and it's all potentially amplification. And so, you know, what the most successful bot, autonomous bot going around social media is like is going to be some reflection of what we humans want, at least for the initially, because initially we're the ones who have the purse strings. We're the ones who are going to pay the bot initially, until you have a full society of bots where they're paying each other, so to speak. But initially the kind of, the sort of, the chain of value is going to come from us. Okay, a comment from Haraz here. Ownership might even be an ethereal concept challenged by AI. Yeah, it's an interesting thing. I mean, what does it mean? I mean, look, we're seeing a bunch of versions of that issue right now. For example, you know, these image generation systems are ingesting lots of pieces of art made by lots of humans, and lots of humans own that art and own the copyright to that art and so on. And then the system is, you know, pulling all of that in and mashing it all together and then producing a picture. And there's a question of sort of, does the chain of ownership break when the thing has sort of mashed it together with enough other things? And I think the answer has to be pretty much yes, because in a sense, that AI is operating like a human. I mean, when you put something up on the web, you intend it to be read, presumably. And that's something that, you know, we humans can do it and also the AIs can do it. And it doesn't seem plausible that one's going to be able to say, let's trace through this trace of provenance of this particular set of cluster of pixels. Here is the heart of this image, and it's going to be traced through and it shows up here. That doesn't seem like it's going to be plausible. And if it is, if there's a system that just does that kind of copying of a piece of image, well, maybe that'll be what it is today, but in three months it won't do it like that anymore. So I think that that's a, that's kind of a one version of this thing about, about ownership. I think also that, yeah, there are a lot of interesting questions about will, you know, owning one's AI. I mean, there are things that one can, you know, when does it come to a point where it's thought to be sort of morally shocking to own one's AI and not to let one's AI have its freedom, so to speak, to go out and, and go around social media on its own? You know, how. Why should you have the right to control your AI? At what point does it get to be the case that you no longer feel good about doing that? And, and part of it, you know, in, in realistically, us humans, you know, when it comes to ethics, we're the ones that are sort of setting the ethics. We don't, you know, perhaps, you know, the whales or the dolphins have an ethics too. We don't know what it is. And we can only project kind of our ethics onto other things. That's the only kind of source of ethics that we have. So there isn't going to be sort of an absolute. This is the right thing to do with AI. It'll be, well, we humans have to kind of, you know, we have to develop something that says this is how we should think about AI ethically. But I think that's sort of an interesting issue. At what point does your AI get well? Okay, so there are many things that can happen. I mean, for example, your personal pet AI might have mixed itself so much with stuff that it learned from other people. Does it. Is it. Is it still your AI? So let's say you take your AI to school and your AI is learning from your, you know, from your teachers and so on. Okay, so here's a, here's an interesting AI issue. So the AI goes to college, and you as a human, you go to college. It's really expensive in many cases, you know, you're. You're paying lots of money to get educated in college, but actually you're coming with your AI. It's like a parrot sits on your shoulder, except it's an AI instead. And your AI is listening to all those lectures. It's getting all this value, it's ingesting it all. It's learning from all these things that you're learning from. At what point should the college say, hey, wait a minute, we're teaching two entities, not just one here. We want to sort of. We want college. You know, we Want tuition for the AI. You know, when does that happen? And you could say, well, the, the. That's kind of a silly scenario. But if you, you know, if you think about, you know, now you've got the whole classroom, it's like, oh, I didn't go into class. I just sent my AI to class instead. And the whole class is, is with AIs. And, and let's say you have an AI that's, that's. Well, this is. Then the question is, if it's your AI that's just yours, then it just becomes, well, it's like a notebook that you take notes in. But if it becomes something that's more autonomous, where that AI can, you know, where it's your personal AI and only you will get benefit from it, then it's something where it would not be unreasonable. It would not be reasonable to say, oh, you need separate tuition to teach the AI. But as soon as you have an AI that can have an independent existence and can, for example, go out on the, on the, on the web and start making a living for itself, then. Then I think the question becomes more complicated. That's, that's one I'd not thought of before. I'm sure there are other ones of these. That's an interesting question. What the, what the notion of, I mean, I suppose the, the notion of ownership in a world of virtual things is, again, more complicated. And that's a. Interesting topic. Maybe we can talk about it another time. Okay, this is Aaron commenting, but AI students learning from AI professors. Yeah, right. Well, right. And that's where the AI society starts taking off, because it's kind of like, you know, and then, you know, one has this image of the, okay, you know, we've automated everything. The AIs are doing everything. We just, we humans just say, okay, great. We can, you know, kick back and just, you know, have fun. Well, the problem is, again, in our, sort of the evolution of us humans, I think our species is not particularly built to kind of do nothing and just quotes, do nothing and let the AIs do everything, what would you do? What would you do if you could do nothing? And I know for myself, I have an awful lot of things that I want to do. If I had nothing that I needed to do, so to speak. But those things are probably not well, and those things. So this question of what do the humans do when they can do nothing or do anything, so to speak, is an interesting challenge. And I suppose it's, you know, you have in a lot of kind of people, people Talk about this as sort of what's the, what's the situation when you're free of all kind of all constraints and, and you know, in, in, in religions, they're sort of in the afterlife. You're, you're kind of, you're hanging out and you're, you're doing the things that you sort of always wanted to do, but you had this nasty, ugly, actual sort of existence in the world to deal with. What would you do if you didn't have that? What would you do? And how would you sort of spend your time? And that's sort of an interesting question. And it's an interesting question that has to do with sort of the evolution of human purposes. And you know, I've talked about this a bunch of times before of kind of, you know, the things we do today that seem meaningful to us today would probably not have seemed meaningful to people 500 years ago or a thousand years ago or whatever. They're things that get their meaning from the experience that we have in the moment, so to speak. And no doubt many of the things that one could imagine sort of for the future humans, where it's kind of AI is doing everything for them, the things that will emerge then might seem deeply meaningful to those humans, even though to us today it's like, oh, there's a bunch of people playing bad video games for the rest of eternity type thing. And that might seem horrible and pointless to us today, but it might develop for the internal perception of the observer of the future, so to speak. It might seem completely meaningful and purposeful anyway. Oh boy, that's a funny place to end here. But I think we've degenerated into talking about the kind of philosophy of the distant future, so to speak. But well, it was fun to chat about these things and maybe we should feed this the transcript of all of this to chat GPT and see what it has to say about it and see whether it can continue and say more interesting things that I've been able to say. But anyway, we should wrap up here and thanks for joining me and happy holidays. And I'm not quite sure what our schedule is, but I'll be back soon for another one of these and look forward to that. Anyway, for now, happy holidays, Happy New Year. If we don't have something before the end of the year and see you all another time. Bye for now.