Future of Science and Technology Q&A (February 2, 2024)

Hello, everyone. Welcome to another episode of Q and A about future of science and technology. I'm next week I will finally be back in my natural habitat and hopefully no technical things will go wrong then they might today. All right, let's see. We have a whole bunch of questions here. Let's see. Elum Enum asks, do you think the plateau of LLMs will be at the level of understanding language? Once it does, it will start to taper off and we'll need to use a different technique. Well, let's see. I think I may have talked about this a little bit before, but there's sort of the question of we now have these neural nets where with billions, hundreds of billions of parameters and so on, and we've bashed them to sort of get them to learn things that, for example, consist of the text we put on the web and so on. The surprising thing is that neural nets seem to be able to learn all kinds of things and seem to be able to extrapolate from the things that they are fed by in ways that are similar to the ways that we humans extrapolate things. So kind of it's like, here's a few billion pages of text from the web. Here's a piece of a sentence. Continue this sentence the way that you would expect it would be continued based on what you've seen on the web, and the neural net will successfully make that continuation. Now, there are all these questions about the big surprise, I suppose, with ChatGPT, for example, is that not only could you go a few words ahead, you could write a whole essay and have the things stay fairly coherent. And as time's gone on the sort of the context window for the LLM as it writes the next word, how far back can it look? That context window has gotten bigger. The ability to remain coherent over a longer period has gotten better. There still isn't yet kind of true planning of what's going to happen next. It's still writing a word at a time. Looking at the past, one can certainly imagine there's sort of ideas, one might even call them kind of quantum LLMs, where you're kind of looking at all the paths that could follow in the future and sort of preemptively picking between those paths to be able to do more kind of planning of what the LLM, where the LLM is going, rather than just saying, I will continue from what I had so far. And there's a lot that we've been deeply involved in and will be even more involved in of having the LLM Call computational tools and actual computational knowledge, whether Wolfram, Alpha, Wolfram, language, whatever, being able to sort of get that feedback of some sort of hard computation rather than just, oh, it roughly looks right based on the statistics of what the neural net learnt. I think there's sort of a question of to what extent is it going to be? You just take these neural nets, you make them bigger, stronger, you feed them more material, they'll sort of get progressively better. My guess is, and this has happened multiple times before in the history of machine learning, I know I've talked about this at other times, that there are these thresholds that get reached. It's kind of like 2012, sort of recognition of objects, 5,000 kinds of objects, okay, that got solved with kind of the deep learning idea, other kinds of ideas about speech to text and so on, got solved. Image generation, stable diffusion kinds of ideas and so on. Got sort of solved. The basic question of how do you do kind of AI generated images? Now in each of these cases there's sort of a basic threshold. You get past and things sort of start to work and they're useful. Then there's a question, can you tighten it up? It gets more and more difficult to tighten it up. And often these kinds of techniques with machine learning are quite hard to tighten up. So traditional software engineering, programming, kinds of things are often easier to tighten up. It's kind of like add this special case here, add that special case there. That's not something that one gets to do with machine learning. It's a much more holistic thing. You can train it this way, you can give it some reinforcement learning here it's much more difficult to say, I want to make this specific change and have this specific thing happen. And I think what we've certainly seen is this sort of a threshold is reached and the threshold is passed. We get to some amount of functionality. Then it's sort of gradual improvement over a long period of time without a certain amount of tweaking. But it isn't as systematic as it might be. And you kind of have what you have based on that sort of level of machine learning. Now the question is, what are the next thresholds? I'm sure that understanding videos is going to be an important thing, something that's already started. But being able to say, I know what happens in the world, because I've seen all these billions of hours of video of things that happen in the world, whether they're things that happen mechanically in the world, whether they're things that happen with people and so on. Those Are things where one can expect that that will be a thing that will routinely be sort of aiable in the future. I think there are different questions about sort of what one gets to when one can understand language in certain kinds of ways. I mean, the very idea of language, language is a sort of way of packaging thoughts in a way that is communicable to another mind, so to speak. And the really big idea of language is this sort of compositionality idea, the idea that rather than just saying, it's a rock, it's a chair, little nouns like that, you can put together a whole sentence where there's sort of a construction kit of different pieces of the sentence. And you can say, a red chair was next to a blue table. And you don't have to have kind of a separate word for blue table and a separate word for that whole concept. You can compose together these different pieces and make a piece of meaning by composing those pieces together. That's sort of the key idea of language, so to speak. And it's sort of anchored with the core words that define concepts. In English, we have 40,000 of them or something commonly used. But they get to be combined in all these different ways to make all the things we might say in the language. And so that idea of sort of packaging ideas, packaging thoughts by this sort of composable elements anchored to concepts represented by words, that's the core idea of language. And so there's a question as you put the AIs in place, and they start to look at the world and they start to be able to make inferences about what will happen in the world based on watching videos and things like that. Will there be some different form of summarization of the world that the AIs can come up with that's different from language as we have it? For example, for us humans, you can imagine there are sort of pictorial languages, there's sort of iconography, there's user interfaces, there's things like this. Those over the last, particularly, I would say, over the last few decades, have come to be sort of another level of kind of a communication with humans. There's certain things you expect in a user interface. There's certain kinds of things that a user interface can communicate or that infographics can communicate, and it's a little different. And, you know, there's often the statement, you know, a picture is worth a thousand words, so to speak. There's somewhat different things you can communicate in pictures than with words. But although they don't have quite the same kind of compositionality that language seems to have, apparently. And so that's a case where we're starting to sort of poke at a different form of communication. Not quite as systematic as the one that we've developed with language, but it gives kind of a sense of where one might see sort of other forms of communicating ideas. Now, will the AIs, between themselves, for example, come up with a way of packaging ideas, thoughts, that is very different from what we humans have come up with? Possibly. I think that it is. And then will the thing that they come up with be something that we humans can learn? My guess is if the AIs come up with it and the AIs are neural nets, we humans will be able to learn it. But there will be challenges. For example, it could be that with a sufficiently big neural net, you'll be able to have a language with a million words in it, whereas compared to the 40,000 we might have in our languages today. It does seem to be the case that there's some kind of scaling between the size of the brain, the number of neurons, the number of artificial neurons in the artificial brain, actual neurons in the physical brain, the number of neurons in the brain, and the number of distinct words that you can be dealing with. We have 100 billion neurons, give or take. We can deal with 40,000 words. You know, this is the question of sort of the. I don't know, how many words can an ant deal with? It has 50,000 neurons. It's some much smaller number, it seems, of sort of concepts that an ant can perhaps communicate. And so then the question comes, if we start to have neural nets that have tens of trillions of neurons in them, then can that deal with a larger collection of concepts? What does it mean to deal with these concepts? What it basically means is as we take what we see from the world and we say this whole collection of different images, they all correspond to the same concept. How many different sort of attractors of different kinds of forms of input that lead to the same concept can we fit in the space of our brains, so to speak? Or is it the case if our brain has a certain number of neurons and we start saying we want to pack more and more and more concepts necessarily, they'll get confused, because the sort of the attractor, the pattern of brain activity that we recognize as the concept of a cat, the pattern of brain activity that we recognize as the concept of a dog, if you try and pack too many cats and dogs or different kinds of concepts in there, there'll be start to be overlap and you'll start to not be able to distinguish these things. And that's probably a thing that is distinctly associated with sort of number of neurons in brains. And so what will it be like, for example, if we could have an AI that knows a million words, that has identified a million concepts about the world and that is able to distinguish those million concepts? It's kind of a little bit like when we look at a specialized area of language, I don't know, medical jargon or something like this, there are a bunch of different words that the typical person won't immediately know, but which are useful for communicating medical concepts and which somehow leverage they somehow go further with language than we could go with the everyday words. And when we have the million word AI or something, it's probably going to be able to make a bunch of distinctions that we don't in our everyday language make. And what it will sort of feel like to deal with such an AI is an interesting question. I don't completely know. I've sort of been wondering a little bit about it, I think, but, but anyway, that, that would be my thought about sort of what the next level of quotes, understanding language. Now, is there a form of organization of ways to package thoughts that is fundamentally different from language, perhaps more efficient than language, perhaps that's achievable by a thing that can keep more concepts in mind or some such other thing? I don't know. I mean, I think that before language existed, I think it would have been quite hard to, I mean, very hard to project to this kind of other mind situation. Very hard to imagine what it would be like if one had language, if one didn't have language. Now, I suppose an example where we get to see a sample of this is our computational language, you know, Wolfram language, which is something that sort of didn't exist and now it does exist. And we can kind of start to see how that clarifies one's thinking about certain kinds of things and to sort of have that in its full form and to. I mean, since I've sort of lived the from before that existed to when that existed, so to speak. It's been interesting to see the extent to which sort of the existence of computational language helps one think thoughts in a different way and helps one kind of compose thoughts in a different way. So I suppose maybe that's an example which perhaps one should think about more carefully of sort of what it might feel like to deal with the AI that let's say, knows a million words, has a different level of sort of Language like thing, maybe we already have an example of that in the computational language that we built. So, anyway, a few thoughts about that. All right, let's see. Lots of other questions here. Well, let's see. Okay, a few responding to this point, and then we'll come to some others. Kant says, how do you know our brains aren't already using compression? Our brains do use a lot of compression. That's kind of how we go from all the complexity of the world to just using words to describe what's going on. That is really a story of compression. That's really a story of taking every little pixel arrangement that might correspond to the camera I'm seeing in front of me and just saying it's a camera. That's a sort of an act of compression. And it is a key thing that I think kind of beings that can observe entities, that can observe the world make use of. It's a key feature of being a mind that you go from the complexity of the world, all the atoms bouncing around as they bounce around, to. To something where you've concentrated that into a thing that corresponds to a sort of coherent stream of thought. Let's see. Comments. The AI could make far more precise classifications than we, leading to new words and with very precise meaning. The thing to understand about words and the precision of meaning and so on is. Is the whole point is this idea of compression. I'm just saying it's a camera in front of me. I could say it's a camera that's this shape with this thing and that thing and that thing, and I'll call that a camera logo, blah, blah, blah, some complicated word, and that means a camera with those characteristics. And I have some other camera that has a camera with some other characteristics. And pretty soon, if I have a different word for everything I see, I can no longer abstract things. I mean, one of the key ideas of minds is the abstraction from the details of the world to things that there is a narrative about that minds can reason in terms of. So it's sort of be careful what you wish for in terms of having more words for things, because if you sort of have too many words, you don't get to make any generalizations. You don't get to abstract anything. It's a question of can you have the words, different words for things that you care about, making distinctions about, and not different words for things that should be conflated together. Let's see. Comments. Seems more important is the speed of the medium that AI has and how quickly it can process all the Interrelations, however many, however many concepts there are. Yeah, I think that's interesting. And I think, you know, you've got to recognize that human brains, you know, our neurons fire, you know, every few milliseconds or something. In, in computers, things are happening every nanosecond. It's, you know, the fundamental hardware of computers is a million times faster than the fundamental hardware of our brains. Right now we still have the advantage that all our neurons are sort of active all the time and are able to learn all the time. We don't quite yet have that in machine learning systems, but I'm sure we will. And then we will have the situation where the fundamental hardware is a million times faster. And there is very interesting questions about what the world would look like if you could perceive it a million times faster. I mean, I realized that, for example, light doesn't. You, you'll perceive the speed of light. If you think a million times faster, you'll notice that the light is coming from there, bouncing off the, the other wall in the room and doing this and that and the other. A bit like how you hear echoes and sound and things like this. You'll, you'll notice a lot of things that we don't notice at the speed we can perceive things. It's also the case that sort of, in terms of sensory input, you know, we have whatever 10 million nerve fibers in our optic nerves going from our eyes to our brains, but one can imagine that one could sort of detect from every Internet device, you know, every IoT device in the world. You know, let's say it's a billion devices. You could, if you had a faster, bigger brain, you could be getting input from all of those devices all the time and then forming, in a sense, a perception of reality from a vastly richer kind of sensory experience. I don't know what the sensory experience of a billion IoT devices would be like. We, in our brains, you know, we've got our visual cortex organized to make use of the spatial relationships between pixels. You know, we know what it means for these nerves in our retina to be the photoreceptors in our retina to excite nerves and two neighboring nerves, so to speak, we know that they are geometrically nearby. And that's important to our vision of the concept of forming an image, so to speak. What does that look like for a billion IoT devices distributed perhaps geographically, perhaps functionally, whatever else, there will be some sort of collective view of the world. Just like we have this view of images as being this two dimensional Thing that will be formed from that sort of collection of the IoT experience of the world. And I not sure what that will be. I think that we can say the same thing for all kinds of detailed sensory inputs. Whether we have, let's say, here's an example, we do have some experience of. Let's say we have chemical sensors which detect roughly the shapes of molecules by having, let's say, a protein that the molecule of lots of different shapes and the molecule is binding to those different shapes. A little bit like how the immune system recognizes things. But let's imagine we just have this array and, and we certainly have technology that does this. We have an array where we have a bunch of proteins that are sort of attached to some array. And it's like a molecule comes in and which protein it binds to determines what nerve impulse is produced. Well, what is that? Well, that's like smell and taste. That's kind of how those work. And now the question is, what is our map of smell space, taste space, whatever? Well, we humans least, most of us don't have a great version of that map. If you say, how are things laid out in physical space? Yeah, we understand that. If you say, how are things laid out in olfactory space? I don't think most of us would have a good thing to say about that. I don't know, dogs might have a perception of that because they pay much more attention to olfaction. But I think that that's an example of sort of trying to imagine what we could get to if we had sort of, you know, let's imagine we could detect the shapes of all the molecules. And let's imagine that with faster thinking we could do that across more molecules. What would we then conclude about the world? How would we then summarize the world? Because I think the point is that one of the things to realize is, yes, things happen in the world that involve lots of computation, but the things that connect to our minds, that we pay attention to are things that are somehow matched to the way that our minds work. Which means you have to reduce it to this sort of stream of concepts that our minds can process. Otherwise you're kind of. You're dealing with the AIs. I'm kind of thinking the AIs on Mount Olympus or something, that the AI gods are doing all their things. And every so often something sort of falls down to the level of mortals, so to speak, and we get to understand it. And most of the time it's just sort of AIs doing their thing with AIs. And how would we describe that? We probably describe that much like we try and describe natural science, much like we describe the natural world. We say, well, there's all these things going on and there are all these detailed patterns of behavior, et cetera, et cetera, et cetera. But the thing we really care about is that a tornado formed. For example, the details of every little eddy that made the tornado. We don't really care about that. We just care about the big story of a tornado formed. And I suspect it's the same kind of thing in the case of the AIs. There. Lyn asked the question, which language is best at this compression? I suppose the compression of ideas into thoughts. It's a good question. I mean, there's certainly. When you look at different languages and you say, let's take this book, let's translate it from English into French, it's a bit longer in French. Just the number of letters in all the words is a bit longer in French. And one knows, for example, I know we have translation lengths that we computed at some point, typical translation lengths for how long will a book be? It's relevant if you're translating a big book like my new kind of science book. It's kind of like how much does it puff out when you translate it into some language other than English? English is one of the most compact of the languages, alphabetically written languages. And I think this question of, well, what language best compresses? There are. There are artificial languages, constructed languages that have many fewer words in English, down to 150 words or something. But the issue with those languages is it's hard to express elaborate philosophical ideas in those languages. Those languages might be good for sort of the everyday I want to get a cup of coffee type thing, but they're not going to be good for sort of the detailed philosophical discourse. Just as in talking about some specialized area, whether it's medicine or finance or whatever it is, knowing the jargon of that area is useful for communication in that area. So I think with respect to sort of how you express things in language and what sort of takes the concepts we care about and expresses them as succinctly as possible. Interesting question. Big project that we're working on right now is this thing that I call symbolic discourse language, which is an attempt to make a computational version of a way of expressing kind of everyday discourse. My guess is that that will be a very succinct version that takes the essence of sort of concepts that are being communicated in language and really boils them down to the sort of computational essence. Of course, when you do that, you lose a certain nuance, you lose the. It's just like when people go to WolframAlpha and they say, please, can you compute the following differential equation or something like this. It doesn't care that you said, please, can you do this? It just says, there's a differential equation here. I'm going to do my thing. If you said, very poetically, work out something about some genomics thing that wolfmalpha can compute, and you say, imagine, you know, there is a, you know, the, the boat is crossing the genomic sea and it's got this and this and this. Well, if wolfmalpha can figure out what's going on, it's just going to say, you want me to go compute this thing about what happens upstream of this gene? And it'll go and compute that. It's crushed the poetry out of it. And in a sense, it's often the case that in the end, you know, when you hear language and you would hear, it's like, what's the difference between people who communicate by email, people who communicate in person. It's like sometimes in my world of running companies and so on, it's sometimes when you tell people that you'll have people and they'll be sort of feuding with each other by email and they'll be sending these kind of growly emails back and forth at each other, and it's like just pick up the phone and talk to each other. Because what you get from that additional level of communication, of hearing the tenor of the voice, so to speak, as well as sort of just seeing the words, you get a different amount of communication which hopefully can help resolve things. Let's see, there's a question here from Luis. Is there room for significant advancements in mathematical notation? Well, I like to think that our computational language is that story, that notation. I mean, we've done a lot over the years to kind of take what is the traditional mathematical notation and just clean it up a bit so it's really consistent and able to be systematically understood. I mean, otherwise you end up in these crazy situations where, yes, you write it down that way, but it's really hard to understand. I mean, in the history of mathematical notation, there are famous examples that, like the times sign, invented by a chap called William Outrad Leibniz, who was really into notational kinds of things, said people will confuse that with an X, you should use a different symbol. But it didn't catch on. Likewise, Leibniz himself, with his D by DX notation for calculus, worried that people would think you could cancel the Ds, which is obviously a thing that happens in kind of calculus classes for the last 300 and something years. So sort of a bad notation. We've tried to, in Wolfram language and in Mathematica, we've tried to sort of clean up these notations a bit, so they're still recognizable as being those notations, but nevertheless made precise. And that's. And so I think that you only get a certain amount of stuff that you can expect somebody to learn notationally. One of the mistakes made in a language called apl, a very interesting language built by a person called Ken iverson in the 1960s, kind of an array processing language. The APL stood for a programming language. But it was early days. But any case, it was a sort of attempt to make a notation for computation, but it really didn't work out well. It had all kinds of weird symbols. There were special APL keyboards, all this kind of thing. People didn't learn 30 or 40 new symbols. That was too much. But what we managed to do in Wolfram language was we leverage English. We use English words and people have an idea what the thing means just by reading the English words. And it makes it much easier to memorize, much easier to stuff it in a brain, to have that familiar language there. And I think that's sort of the future of kind of, in a sense, mathematical notation is computational notation. And that's what I've spent much of my life trying to build, is a consistent sort of computational notation for expressing things about the world, leveraging in most cases, kind of English as names of functions and always having that as the backstop. Even if there might be some slash AT notation. That's the thing that the experts get to use. It's still MAP or, you know, let's say map apply, triple AT sign is, you know, the experts can use triple at sign, but in the end it's map apply, and you can kind of know what that means. Let's see, Let's see. There's so many questions here. Brivich says a single biological neuron can perform an exclusive or operation. A current single artificial neuron can't do that. Do we think we're underestimating the power of biological neurons compared to artificial neurons? See, this is an interesting question. When you are doing neuroscience, you're modeling brains. There are lots of details in the neurons that we have in our brains. What matters, what doesn't. It's like in physics, you could say I'm studying mechanics, I'm throwing an object off the Tower of Pisa or something. And I want to know, does it matter that it had spikes on the edges, or does the only thing that matters to how fast it will fall is it just the mass? Well, there are details because there's air resistance. In that case, there wouldn't be if it was in a vacuum, and so on. It's like what matters, what doesn't. And it, of course, depends a bit on what you are asking for, what output you're asking for. If it's a question of if we generate language, if we do our thinking kinds of things to generate language, what matters? What doesn't do the twitching of the dendritic spines on the individual neurons in our brains, does that matter to the production of language or not? Or is that irrelevant? I think the thing we've learned from sort of the chatgpt discovery and invention is that less matters about the details of artificial biological neurons than we might have thought, because we can get a pretty good result for generating language from these pure artificial neural nets. Now, there are clearly things about biological neural nets that are different. Like, for example, you put in, you know, you bathe your neurons in your brain with some neurotransmitter, some hormone, something like that. You bathe it in some overall global chemical, so to speak. You change the mood of your neural net. You know, you have some drug, and you change the overall behavior of the neural net. What's the analog of that, an artificial neural net? Actually, I've done some experiments on that of an artificial neural nets, of making sort of overall changes in the weights in the neural nets, seeing what happens, seeing how you can kind of blow the mind of the artificial neural net. Kind of interesting to see what happens there. It's kind of very much like the picture that you have of sort of blowing a human mind, so to speak. But it's. So I don't think there clearly are differences between artificial neural nets, biological neural nets, but what matters and for what purpose? And I think this question of, for example, what's the memory capacity of human brains? Is the memory stored in a few bits per synapse, a few bits per connection between nerve cells, or is it stored in all sorts of detailed molecular information that is more at a molecular scale than at the scale of individual nerve cells? I think the suspicion is that it really is at the level of individual of nerve cells, which is what people have suspected since at least the 1930s or 1940s. And not at some much more detailed molecular level, but might be wrong. But that seems to be what we're discovering from looking at the capabilities of artificial neural nets. When it comes to the way neural nets work, I mean, it's remarkable how very straightforward it is and how very similar it is to the way that the very first artificial neural nets from 1943 were described. It's kind of like you're just adding up a bunch of inputs with certain weights and you're thresholding that in some way with some activation function. That's it. And yes, there are lots of things that it's kind of hard. It's easy to prove theorems that say that if you have a hidden layer between the input and the output, if it's a big enough layer, you can in principle compute anything. But in practice, what can you learn by training? By what you know, how. How easy is it to compute particular functions? Good question. I've studied this a little bit. What functions are easy to compute? What functions are hard to compute? That is, something probably gives us information about, if, if we, if we followed that path well enough, it would give us information about things we humans find easy, things we humans don't find easy. Assuming that this model of artificial neural nets is an accurate model, and I think it's a reasonable question, what, what functions, what things are easy to compute with neural nets? What are not. The thing to understand is digital computers are very much based on kind of ones and zeros on and off. Everything is discrete. Neural nets are, in their current sort of construction, are fundamentally continuous. One's dealing with just O1 feeds in this number. You know, even if you're feeding in words, you're turning them into numbers. And then the weights are precise numbers. 1.735 or whatever else it is. And if you, if you bash the thing to be trained to learn something slightly different, that 1.735 might turn into 1.737. And that little extra, you know, change might at some moment as you go through the neural net, that might really matter because you might just go above some threshold for some activation function or not depending on that precise value. But so neural nets are kind of a continuous model of computation. The way they're set up right now. My guess is that's quite unnecessary, that actually there's a way of distributing the kind of. Just as neural nets, neurons and neural nets, there are many in parallel. I think that those bits that one's packing into numbers right now probably can also be distributed and sort of run in parallel. So it's like many bits make up the analog of one number rather than everything having to be in that sort of continuous number. It's very similar, I probably use this analogy a zillion times to what happens in fluid mechanics, where there's sort of a continuous velocity of fluid, but that's really made up of lots of underlying discrete molecules that exist in a discrete form. And I suspect the same thing can be done for neural nets. So I don't think that's a fundamental feature. But this question about sort of what matters in biological neural nets, what matters? What do we have to emulate to achieve certain functionality? Interesting question. I think the current evidence seems to be on the side of the things that people modeled in 1943 about how neural nets might work are probably good enough to capture the essence of what's needed for thinking kinds of processes. Memes asks, did the neural nets I was playing with learn after their minds were blown? I would think, I'm not sure. So what I did was you take the weights in a neural net, you've trained the neural net to be able to generate images, and you say, let's take all those weights and let's just increase the values, just multiply them by 1.01, 1.02, et cetera, et cetera, et cetera. And for example, what I was doing was saying make a picture of a cat and a party hat and do that where the neural net keeps on getting adjusted. Well, you do it for differently sort of inflated neural net weights. And what you see is the cat looks pretty good. I don't remember the numbers you can look up in the writing that I did about this. The, you know, let's say at 1.01, it's still looking pretty cat like. By 1.05, the cat is starting to sort of, you know, have bizarre green things sticking out of its head and so on. And by, you know, 1.07, the cat's kind of exploded and there isn't really a cat to be seen anymore. Now, could one, given that sort of exploded network, that, that amped up network, could one go and take that network and continue training it? My guess is yes. But that training, in a sense, eventually that training will just revert. If you train it enough, it's going to revert to what it learned before. If it's the same training data. But. So if it's a question of fine tuning from that network, how far can fine tuning go? I don't know the answer. Good question. I just don't Know, Gosh, so many new questions coming in here. I want to come back to some of the earlier ones here, but there's a question here from Hassan. What's your view of emergent phenomena? Do you think it's possible to predict all the emergent abilities that AI can obtain? Well, absolutely not. I mean, one of the fundamental things I kind of discovered in the 1980s is this phenomenon that even very simple programs can produce very complicated behavior. The programs we tend to choose to write with a particular purpose in mind just do what the purpose we had in mind was, assuming that they don't have bugs we didn't expect. But if you just pick little tiny programs that just say how to update, you know, arrays of bits or something like this, if you just pick those programs kind of at random, you just start enumerating from the simplest programs on you, you quickly find out that you'll end up with programs where you look at the program, you say, this is a simple program. You run it. Oh, my gosh, it's doing all this incredibly complicated stuff that I never expected. That is the generic behavior that you see out in the computational universe of possible programs. I think it is also the generic behavior you see in a lot of systems in nature, where nature isn't under the constraint that one should be able to understand what it does. So it just runs some program and it does what it does. It makes that complicated, turbulent, fluid. It makes that elaborate biological pattern. These kinds of things. It's just going out and sort of in the computational universe, producing things. Lots of stuff emerges. There's this fundamental phenomenon that I call computational irreducibility that basically says, when you have one of these programs and it's running and it's doing what it does, there's the question, can you jump ahead and say, I know what it's going to do. I know the final answer. Well, computational irreducibility says you generically can't do that. You have to follow the steps and just see what it does. Which is another way of saying there'll always be unexpected things that will happen. Because if everything was. If you could always jump ahead, you would be able to expect everything. It would be. It's the unexpected is what happens through computational irreducibility. And yes, there are always unexpected things that will happen. Even in systems with very simple rules, let alone the systems we have with all these AIs and so on. Yes, unexpected things will happen. And this is a fundamental kind of societal thing for the future, is kind of, if you want computation and AIs to do the things that computation and AI can do, to have the sort of power of computational capability, then you are signing up for computational irreducibility. You're signing up for the fact that the things that can happen can include things that you cannot foresee would happen. So in other words, you say, I'm going to set the system up, it's going to do what it does. And if you say, I never wanted to do this, I never wanted to do that, that's very hard to achieve. If you say, I just want the system to follow, go, you know, follow this track. Well, yes, you can get it to follow the track, but then you can't have it have the true power of computation, true irreducible computation. You're then constraining it. And yes, you can understand what it's going to do, but it won't be able to do the interesting things. It won't be able to do almost the creative things. It will only be able to do the things that it was sort of reflexively programmed to do. Let's see. And so, by the way, that is the sort of the future of computation and so on is this thing where you're not going to be able to predict what will happen. It is interesting in a sense that just as you won't be able to predict what happens in the physical universe, but I think in a sense we're changing the rules more quickly in AI than they get changed in physics. And so we're kind of seeing at an accelerated rate, this phenomenon of kind of computational irreducibility is more in our face in the case of AI than it has been in the case of the natural world. Let's see. Brady asks, will humanoid robots usher in an age of abundance? Well, you know, the world as we've built it is built for us humans. You know, we have steps that go from here to there which are about the right size for human legs to walk up them. We have all kinds of things in our environment that are built for humans. You know, we have lights that allow us to see in the visual spectrum, part of the electromagnetic spectrum. We have things that are predictable to us. We have rooms that are in rectangular shapes that we can sort of predict. The world as we have built it is built for us humans. Now, the natural world wasn't built for us humans, although biological evolution has made us kind of adapt to the niches that are naturally available there. You know, if we go start colonizing Mars or something, we're out of that game. We're out of what we were ever evolved to be. To be in, so to speak. And so then it's kind of a different issue of whether we are, whether we're sort of suitable for the world as it is. But so when we have robots that we want to exist in the world as we have built it out or as we have evolved to it, if they're humanoid robots, they're going to fit in, they're going to be able to walk up the steps, they're going to be able to open the door, they're going to be able to do these things that were built for humans. Now what will that allow to have happen? I think a lot of things that were a lot of mechanical things that were very human specific. Now we can automate those things. I mean, a lot of things about, you know, we've been able to automate things about, you know, the robot vacuum cleaner type thing when the, when the floor is pretty simple and not too many, too many weird things happen. There are not too many steps that can fall down. All those kinds of things like this. We can automate that. When we have sort of more completely humanoid kinds of setups, then we can expect that, yes, it can have its own mop or something that it carries like a human does. And the, the environment, the house was built so that it was moppable by a human, so to speak. And it will then be moppable by a humanoid robot. What will that. There will be certain kinds of things that can be automated. I mean, there's an awful lot over the course of time as technology progresses, there are things that were kind of really hard to do that became easy to do. Like, know, cars got. There were things where, well, you used to have to have you, your horse and somebody would feed your horse and you would do the, you know, all the things, but somebody would be the person who would be your, you know, Uber driver for the, for the carriage or something like this. Except I don't think they had the. Well, I don't know if they had the. I think the gig economy was different in those days. But any case, the, the, the, the thing that was, you know, you would have this whole infrastructure for feeding the horses, doing this with that, or if you have, you know, using airships, you would have a whole infrastructure for dealing with. It's going to take a long time to go across the Atlantic or something. There's a whole different infrastructure. And then the technology comes in and then certain things become much simpler, much faster, et cetera, et cetera, et cetera then we adapt to that, and then a whole nother level of. Of human activities emerges. And I think the same will be true of humanoid robots. There are things where you say, you know, we might have said in the past, Even, you know, 150 years ago, it's like, fundamentally you're going to have to grow. Well, maybe not. Maybe a little longer ago than that. Fundamentally, you're going to have to grow your own food. Most people in developed countries these days do not grow their own food unless they think that that's fun, so to speak. And, you know, there's an infrastructure that built up where we mostly don't grow our own food. But yet. And similarly, you know, there's a. At the time, you know, there was a time when mostly people would always wash their own dishes, but then there were dishwashers that got, you know, that automated that. And, you know, there's a. There's this whole sort of train of what gets automated. The thing that is interesting about us humans, and it's sort of a computational irreducibility story applied to us humans, is there's always stuff to do that hasn't yet been automated. And I think that will always be the case. There is always stuff to do that hasn't been automated, that isn't sort of systematic enough that it's yet been automated or automatable. Now, we could choose to say, okay, enough is enough. We can just hang out and sit back and have the machine peel the grapes and we'll just eat them and we won't really do much else, and we'll just hang out for our lives kind of, you know, eating peeled grapes, and that's all we'll do. If we're in that situation, if we decide we just don't care about anything new and different, well, then, yes, we will get to the point where the things that we care about so far can be automated and they're very easy to get. I think the thing that has been the case, and as I say, I suspect will, in a sense it's an inevitable feature of computational irreducibility, is there's always more to do. And so there's always more to do that is hard to do. I mean, I know in, you know, I tend to be a person who ends up trying to do things where I managed to automate a whole bunch of stuff, managed to figure out, learn about a bunch of stuff. Then there's always a frontier. There's always more to do, there's always more to figure out, and there's Always a thing that's hard to do next, so to speak. And I suppose it's just a choice of somebody like me, I'm always going to try and, you know, try and do that thing that seems hard to do. And maybe as a, as a species, as a society, we could all just say enough is enough, we're done. In which case, yes, the humanoid robots can be our kind of automated butlers and so on and sort of we can just hang out. But my guess is there'll always be hard stuff for us to do and it's a question of. And they'll, and then there'll be the stuff that sort of. The automated away stuff that we just say, oh, yeah, yeah, of course, you know, don't worry, you can just go to the, go to the grocery store and buy the food. You don't have to grow that food. And the same with lots of other things. I think there will be some, some thresholds with as robotics gets better. Oh, for example, industries like the construction industry, which has been not very automated and involves a lot of people. And what happens when robots can routinely do construction? Well, a lot of things that were expensive before become much cheaper, like build a house, take a house down again, maybe at the point where there are things where everything is built robotically, maybe it's like, oh yeah, you put up this house. Oh yeah, you want to move your house? Okay, we'll get robots to disassemble the house again and reassemble it somewhere else. And that will be cheap enough to do that. It can realistically be done. Just like there were plenty of things that, where the things one does today in the past, one wouldn't ever have imagined that one could, you know, pick something up and move it to another place. Even with computers. For example, in my life, the first computer I used was, weighed probably a ton or so and it was, you know, the size of a big desk. It was and so on. Even the, the computers that I was using back in, oh, by the end of the 1970s, early 1980s, those were computers you couldn't move. And the concept that I can take a fully capable computer and pick it up, move it around is a remarkable concept. And I remember I first had a laptop computer and I first had a portable computer, the Osborne 1 computer, I think in 1981 was the size of a suitcase. And it certainly didn't run off batteries. And it was a sort of a funky thing to make use of. And I didn't make that much use of it actually. And then later There was a thing that was really a laptop size. It wasn't a very powerful computer yet that you could really just put in a bag, take with you. I remember at the point when computers got to the point where I could easily put a computer in an envelope. That was another thing where it became much more portable. And it's the same type of thing with, with things like, let's say, construction, like houses, you know, can you realistically build a house, take it down again, build it again somewhere else? Probably, yes. With, with, you know, when robotics gets to the point where it's capable enough and cheap enough, Brivage asks, can a higher intelligence ever transcend computational irreducibility? Well, here's the thing. Within computational irreducibility, there are always pockets of reducibility, things that we can do, even though to do everything we can't. But there are particular places where we can have technology, we can have ideas that say, yes, we can figure out that little piece, this piece here, this piece there. It's sort of inevitable that we can put all these patches on, will never get the whole way. We can put any number of patches on. The more patches we put on. Sort of interesting that if we put enough patches on, the thing that's doing the patching is itself going to get very, very bloated. It's kind of like by the time we've got enough sort of sort of axioms of our mathematical theory that everything's an axiom, we just, let's say, oh, we got a mathematical theory, theorem, you know, the Riemann hypothesis, go bust, conjecture, something like that. Just add it as an axiom. Well, eventually it becomes very incoherent when you've just sort of added everything as an axiom, you no longer have this sort of coherent view of the world. And I think the same is true with the sort of computational reducibility. If you have enough different pockets of reducibility, the world gets very complicated again because you're like picking between. It's like all these devices that manage to do a little bit of sort of jumping through computational reducibility. Then it's like pick which device you want to use. My guess is that in our universe, we never get to fully transcend computational irreducibility. Pick another universe you can, but you'll never be able to communicate between that universe and ours. So I think we're kind of stuck. I think it's kind of an inevitable feature of the fundamental structure of the idea of computation, the idea of following rules, that we as entities embedded in the universe, which is also following rules, will experience computational irreducibility. And I think that's true at whatever level of sort of the computational hierarchy we're at. Because the universe is just like us and we are part of the universe. It's inevitable by sort of a diagonal kind of argument and sort of mathematical logic kinds of ideas, it's sort of inevitable that there will be computational irreducibility for us. Foreign, let's see. Prabh comments. Why does pessimism towards future technology exist? That experiences that. Most people's default opinion on future tech is let's not do this. Well, I think that's a mixed bag because there are things where there are things. There have been cases where sort of being naive about what the consequences of something are is not very good. I mean, you might say, gosh, we can do artificial selection of humans, we can breed humans to do eugenics, for example. Oh, whoops, that leads to this or that or the other very bad consequence. I think what often happens is, and what makes people suspicious about sort of let's do future technology and it will be great is people say it will just be great. Nothing could possibly go wrong. We know exactly what will happen. Let's put this thing in here, let's start doing this and what could possibly go wrong type thing. Computational irreducibility is the story of there is always something that could go wrong. And as soon as you go into it with this, this attitude of it's science, we know nothing can go wrong, then you get yourself in trouble. I think a humbler attitude where it's like, look, we can do this. There will be things that will happen that we will not foresee, and then we have to deal with those, and then we have to deal with those. But just the look, everybody, you should do this, this and this. Because it's what science says that's a mistake. And I think when that gets said, that's when people, I think, rightfully get suspicious about is this a good idea? You know, should we do, you know, should we make super viruses that will, you know, be able to wipe out the species, so to speak? Well, you know, maybe it may be harder to do that than one thinks. My guess is it is harder to do that than one thinks. But it's still like that seems like an obviously bad idea. But it's also something where somehow, you know, you say, well, we'll do this and this and this, and we can scientifically show the virus will never get out or whatever else it is, you know, it's kind of you have to expect the unexpected, so to speak. The idea that science will just say, yep, we got it all under control, it's just going to work this way. Nothing unexpected will happen. That's wrong. And that's something that has been sort of a mistake of science for the last few hundred years. That's a consequence of the fact that bunch of science that we bit off starting in the 1600s does have that feature that the stuff we bit off. We are able to say, yup, it works this way. Nothing could possibly go wrong. But that's not the whole story. That's the story of the particular pieces of mechanics or whatever else that we chose to study. It's not the full story of everything that can happen. That full story involves computational irreducibility and involves things that can go wrong, so to speak. Now, in terms of the question of people's general attitude towards the future, I think one has to realize that we all have a certain set of things that we've learned that we're comfortable with and so on. And some of us, I mean, I know it's a thing that even has genetic correlates in mice and so on. So some of us are more like, oh, I want to do something new than others. It's kind of that, that's a thing which probably we're all set at a certain level of we want to do new things versus we don't. But we all have a certain degree of, look, I'm familiar with this. I like this. I don't want to, you know, that's where I'm comfortable, so to speak. So when the new comes in, it's a question of sort of how does one adapt to the new. And for some people, it's very disorienting. That's true when it comes to ideas and paradigms and so on. It's true when it comes to societal, the way society is set up. It's, you know, there are different cases of that, but I think that that's. That there's a certain, you know, there's a certain, I would say, perhaps rational sort of fear of the new because, well, gosh, then you have to put a lot of effort into learning the new and so on and that. So in a sense, when we bring in new technology, part of the point is, can you make that new technology comfortable to people? Can you? For example, one of the things that is often an issue with user interfaces is can you make that user interface remind people enough of something that they already knew that they can deal with it. Or in the design of computational language, when we build that, we are using English words for names of functions and so on, because that is something people already know. We're building on something people already know. It's sort of an interesting thing. I mean, the. The different trends in user interface design from the kind of skeuomorphic design where it's kind of, yes, it should look like a, you know, a leather padded desktop or something. And does it really have to. How faithfully does it have to represent what people are familiar with? And what became clear was it can be pretty cartoonish, and people are still fine with that. It seems unnecessary to have some of this detail. It's only sort of the essentials that you need to capture. So, you know, I think, yeah, I mean, this question of whether one should, you know, one thing about the future, and I suppose it's kind of like you imagine the case where, you know, you're cryonically frozen for 200 years, you wake up again, the future has arrived. What is it like? And what you realize by thinking about the historical span of this is the future may be very unfamiliar and even very. Perhaps even uninteresting, perhaps even disorienting, because the things that are incredibly important to us today might not be things that are incredibly important in the future, just as it wasn't the same set of things that were important in the past. And I think that that's another feature of kind of the arrival of the future. It's kind of like, look, I just don't get why all these kids worry about how many likes they get on Instagram or something. I don't get why all these people want to do. There's that every generation is always saying about the next one. I don't get why these people are, you know, communicating an emoji or something like this, like the ancient Egyptians, perhaps. But, you know, whatever else it is, why this thing has changed in this way. I don't get it. And that's, you know, and so as you sort of imagine the future, there's this question of, are you going to care about the future, or is the future just going to be about things? Things that are. That are profoundly different from what you care about. I mean, there are plenty of things that happen in today's world, which, from 200 years ago, people would like. You say, this is what we're doing. This is what we're. And people would just, like, shrug their shoulders. Why do you want to do that? You know, that just seems like a Silly thing to do. Let's see a few more questions. Let me go backwards here in my list. M. Rudo asks, can we calculate how wind energy will influence the weather in the future if we become independent of fossil fuels? You know, I've wondered about that. It's kind of like downstream from a wind farm. How does it change the wind? The fact is the rotation of the Earth, which is what eventually drives the wind and so on. There's a lot of energy in the rotation of the Earth and it's going to take a lot to slow the Earth down. I don't think we could use, you know, the energy associated with the rotational energy of the Earth is just vast compared to a lot of other kinds of things. Not as vast as the energy associated with fusion in the sun. It's an interesting question, how it compares with the energy associated with the heat inside the Earth, kind of the ultimately deep geothermal energy. I don't know the answer to that. I probably, if I thought for a little bit I'd probably calculate that. Let me see, can I do that calculation easily? No. If I were typing into Waltham Alpha, I think I know how to get to that calculation in a couple of minutes, but I can't immediately tell you. My intuition would be the heat of the Earth is more significant source of energy than the rotation of the Earth, which can be captured by wind and so on. But in any case the I. So I would not expect that that will slow the Earth down that way or that all wind will be gone, so to speak. I think it's, you know, in the end, the energy use that we humans have. I mean there's certain kinds of reasons that we use energy and it's a question of, you know, how could we scale up, let's say energy was super cheap, was a million times cheaper than it is today, and was completely kind of renewable and all those kinds of good things. And what will we do with it? What will we do with vastly more energy? And how would it change sort of technology, society, et cetera? I mean, I'm not sure. I think we would be able to. Oh, for example, things flying around would be a lot easier. Things, you know, it would be much more routine to be, you know, have the drone that, you know, just put this thing on, you know, drones. If energy was sufficiently cheap, everything might as well fly. And there's a lot more room in three dimensional space than there is on the surface of the Earth. Forget roads, just have everything fly. And I think that's one example in terms of there's a question of if you can move stuff around much more easily than you can right now, how would that change things? Would it be the case that. I suppose that there are things one could imagine where right now you kind of have to. Yeah, I mean, I suppose one can imagine a bizarre kind of. This is a cloud computing analog of owning things back in. You know, when. In that these trends seem to sort of go cyclically in computers. You know, do you want the personal computer on your desk, or do you want just the computer power in the cloud somewhere connected via the Internet? Well, that's taking the bits and having them. The bits flow through the Internet to the cloud computer, to the server somewhere, and then come back to you. Or do you want the bits locally on your computer? Well, if you could move stuff around more easily, maybe you do. The analog of cloud computing for. For stuff that is. I don't need to own anything. Every time I need a thing, I just press a button and a drone delivers the thing I need. And then no clutter, drone takes it away, stores it in the cloud somewhere. And probably, if energy cost nothing, it might actually be in the cloud. There might be some flying platform that's the giant warehouse in the sky type thing that has all these things stored. And whenever you need something, it's just like, okay, let me get that thing. Oh, I'm done with that. Let me just send it away again. It's kind of like the library of stuff, so to speak. And that's the kind of thing one could imagine if energy was cheap enough. I'm kind of wondering what other things. I mean, I suppose that when it comes to computing right now, there's a certain energy cost in computing that doesn't need to be there. We certainly know in principle how to do computing in a completely reversible way that doesn't require the dissipation of energy, the dissipation of heat. That hasn't been. Not that many experiments have been done in doing that in practice. But my guess is that that will not be the dominant source of, you know, the fact that servers run so hot and you go to a data center and it's like belching out heat and so on. I don't think that's a permanent kind of issue. I think that's an issue of today's technology and the fact that transistors are a certain size and et cetera, et cetera, et cetera, and the capacitances are this or that. I don't think that's a Necessary feature of those kinds of things. Let's see. So question here from Allah about what's there to say about the future of food science? Will someone be able to recreate Willy Wonka's creations? You know, taste is a sense that we don't understand that, well, you know, we have taste buds in our tongues. It's like food gets smooshed in there. The actual texture of the food matters. The, the details of the shapes of the molecules matter. I just learned, actually, I didn't even know this, that the, the crunchy, creamy, you know, different kinds of food textures in your mouth. They are literally correlated with the shapes of the proteins in foods. And so that means, you know, if you're eating fibroblast cells, muscle cells in, in a piece of meat, that there's a. There's a microscopic texture associated with the fact that you have those actin, you know, filaments that are. That are in there versus that you're having some quite different kind of protein from something else. So in a sense, taste is all about, sort of, is a lot about both the chemical interaction between the molecule and the shapes of molecules and how they bind into the taste receptors and the physicality of the structure of the food at a level beyond the individual molecules of how the pieces put together, how do they smoosh onto your tongue or whatever else. And I think the whole area of molecular gastronomy, of trying to understand at the level, from the level of molecules, level of clusters of molecules to the level of whole things of food, so to speak, how that has an effect on our, you know, how we sense that from a taste point of view. That's something that's still, I think, somewhat in its infancy. I mean, we are much more able to, you know, for sound. We can kind of. We can create pretty much whatever sound we want. It's still not easy. You know, if you say, make me a new kind of musical instrument, well, that's still in the domain of, at the outer limit of what LLMs start to be able to do in visual perception. Make me a picture of this or that, we're starting to be able to do that in a reasonable way. Make me a thing that tastes like this. Well, actually, I think the biggest challenge there will be how you describe what you want it to taste like, because we don't have good words for that, really. We have pretty coarse linguistic correlates of taste. But if you say, if you say, make me use generative AI to make me molecules that will have this characteristic and that characteristic not sure you necessarily need that much AI in loop, but we can imagine doing that. Obviously there are efforts to do that with artificial meat and things like this, which are so far a little bit my own experiences of eating those things. I wasn't that impressed yet. And I think the, but that's, you know, I think the idea that you can kind of create a taste, create a food with certain characteristics, I think that is a coming attraction, so to speak. I mean it partly is affected by how do you make a molecule of a kind you want with proteins? It's fairly easy. It's the trick that biology came up with of you just have the DNA sequence and you transcribe the DNA sequence and you get this protein. And so long as you can get a protein that folds up to the right shape you want, you're in good shape. But literally. But if you want to make other kinds of molecules, we don't have good ways to do that. We don't have a good sort of standard method for making other molecules. We have to do chemical synthesis which involves taking all of these sort of named reactions and saying, how do I do this reaction, then that reaction, then that reaction to make this molecule have this little end on it here and oh, we don't want that end to fall off when we have this reaction happen. We have to plan these reactions very carefully. It's kind of like a theorem proving thing where you would sort of apply this axiom, apply that axiom, apply this reaction to this molecule, that reaction, and so on. See whether you get out this final shape of molecule that you want. That's still a thing that is not well characterized. It's something that over a long period of time there's sort of computational efforts to be able to do that. But even in that case, if you say, well, what's it going to taste like if I make a molecule that's this shape? We don't know that yet. We don't have a good way to predict that. And that's something that I would expect will come over the next limited ish number of years. Let's see, Let's see. Yeah, I've got a few comments about wind turbines here that people are making most of energy and wind is not on the ground. It's over oceans and up in the sky. Yeah, I mean that's sort of undoubtedly the case because you know, when fluid like air flows over a solid surface, right at the solid surface, that fluid is, is, is not slipping relative to the solid surface. So it's at rest relative to the solid surface. You go up 30,000ft in the air, you might have 100 mile an hour winds. You don't get 100 mile an hour wind as easily, at least as, as something that can slip across the surface. So yes, inevitably there's, there's more wind up higher. And how you capture that in, gosh, I don't know how you capture that in, in kind of, because by the time you're sort of a thing that's floating in the air, it's like you're, you're co. Moving with the air. It's not like it's. When you're a wind turbine that's, you know, rooted to the ground and you've got this great, you know, these big blades and the air is pushing past you and you're rooted to the ground. You're using that fact that you're rooted to the ground. So I don't know how you capture the things higher up unless you know, and this is again, what will happen in the future. You know, will there be super materials that allow you to build a wind turbine that's five miles high, for example? Certainly not possible today. The materials we have today are simply not strong enough to do that. I mean, people have certainly hoped that carbon nanotubes, things like that might be knitted together to, might be spun together, so to speak, to make materials that are that strong. But that hasn't been reduced to practice yet. And you know, there's nothing that says that there is no, there's no theoretical. If you look at the forces that bind atoms together, those forces are strong enough that you could build sort of super materials that, that are capable of making the wind turbine five miles high. But how to actually realize that in practice with an actual material that's not known how to do, but certainly not. I don't think there's a theoretical reason why that's impossible. I mean, as you get longer and longer and longer, eventually it becomes impossible. If you say, let's take this thing, it's on the surface of the earth, it's a rigid object, and it's rotating with the earth. If you go out far enough, the tips of that rigid object will be rotating faster than the speed of light. And you kind of know that no material can hold together under those circumstances necessarily. As you have the giant propeller and it's super, super long, and the tips of the propeller blades are going faster than the speed of light. The reason that doesn't work, the reason you don't get to go faster than the speed of light. There is that there are internal stresses that necessarily build up in the object that, that prevent you from ever getting that thing constructed in that way, just as a. Even as a result of the change of mass with velocity and things, you'll get effects like that. So. Let's see. Joe asks, could one describe a computation engine analogous to a thermodynamic engine, where useful work is a consequence of information processing? Well, what you have to realize is useful work is sort of systematic motion. A great big sort of wheel of a steam engine moves or something, or a great big piston in a steam engine moves lots of molecules all moving together in an organized way, as opposed to there's this. Just this hot gas where all the molecules are moving around randomly. And the key fact in thermodynamics, the second law of thermodynamics says something which starts a systematic motion tends to degrade into randomized heat. And the challenge then is, how do you take that randomized heat and turn it into something that you think of as being organized motion? Now, if we were much smaller than we are and we were at the size of molecules, we might say, oh, that thing you call heat and you call random. That's where all the action, all the interesting stuff is happening. But this question about how do you take sort of randomness and turn it into something systematic, that's kind of the essence of turning heat into mechanical work. And I suppose it's an interesting question, actually, maybe an interesting way to think about it is how do you turn randomness into mechanical work? Actually, now that I think about it, there are many, many algorithms that make use of randomness and extract from it mechanical, you know, useful work. So it's a little bit like if you have a ratchet that is, you know, there's. It's randomly being kicked one way or another, and every so often catches on. The ratchet, the, the. In the end, such a setup will. In the end, the ratchet will produce enough heat that the thing will eventually sort of run down. But for a while at least, you get to sort of progressively ratchet forward just by randomly driving it with heat. And one thinks about all sorts of algorithms that exist, including ones in machine learning where there's lots of randomness that's made use of and sort of you're jiggling around randomly, and every so often, you hit something that you really want and you grab onto it a little bit like the ratchet, and that's the thing that's useful. So in a sense, I suppose that's the analogy to the creation of useful work from heat is these randomized algorithms where you are saying, let's not systematically go and try and find this thing. Let's just randomly try all these different possibilities and see whether we can catch on to one that's useful. I mean, I think I haven't thought this through and it actually not occurred to me that that was an analogy there. But I think that's a useful analogy for sort of the creation of mechanical work from heat. It's the creation of systematic knowledge, systematic information from sort of random sampling. Now, you know, could one invent. Good question. Could one invent the various analogs of, you know, Carnot's efficiency for a heat engine and things and notions of temperature and so on? For that case, I think it's quite possible. It would be interesting to do. I hadn't thought of it before, so thanks for, thanks for that. Lynn is asking can AI solve the taste problem as it did the protein folding problem? Yeah, I think the AI has totally nailed. Protein folding is a little bit oversold. If you really have a random sequence of amino acids like you do in the immune system, we still don't know what that's going to fold up into. It's when it's something which can readily be extrapolated from the proteins we've analyzed in detail with cryo electron microscopy or crystallography or something that we can do it. So it's not really quite the case that sort of the random protein thrown in. Oh yeah, we know how it folds yet. So the question of could we solve taste in the same way that one solved protein folding? I would think the answer is yes. In fact, I think it might even be easier. I think the problem is what are you going for? You know, if you say, well, so a question would be, you know, take chocolate, one of my favorite foods, and ask the question I don't actually know and I should. All the details of which molecules lead chocolate to taste like chocolate and so on. I mean, I think it's the, you know, the cocoa proteins and so on, but don't know in total detail how that works. But say, let's say I want to make artificial chocolate. And I mean, in a sense it's not that exciting to say I'll make an artificial protein. It's going to be taste like the protein that's in chocolate. The only case where that will be interesting. Yeah, here's, here's a case that's interesting. If you're allergic to food X. I'm not allergic for Very much. I discovered at some point that I'm allergic to kiwi fruit, but that's about the only thing I know. But let's imagine that I really like the taste of kiwi fruit, but I know I'm allergic to the specific shaped proteins in kiwifruit. Then I could imagine something where I could go to my AI and say, make me something that's going to taste like kiwi fruit as far as my taste receptors are concerned, but which will not lock on to my immune system to give me an allergic reaction. I think that's an interesting thing to think about, is presumably. Presumably I don't know this for sure, but presumably, certainly the mechanisms are very different for the immune response, the adaptive immune response and the taste response, even though they are molecular recognition mechanisms. And one certainly could imagine that. One could kind of squiggle through, yes, it's going to be the right taste response, but it isn't going to excite the allergic response, those kinds of things. So that's a place where one can imagine doing that. Comment about. Specifically about chocolate at a shop in Cambridge, Massachusetts. Well, I don't know. I'll be back in Massachusetts next week, so I don't know anything that remarkable there. Let me see. Okay, maybe one more thing here. Okay, Dash comments. If we could run our evolution over and over again, maybe we would see ourselves in some different shapes, floating in space with different forms of communication, exchanging thoughts telepathically, etc. What do I think about that? Yeah, I mean, this is the question of why is life the way life is? And is. Is there a thing that is kind of like life, but off on a different branch? You know, doesn't use carbon, uses germanium instead of. Doesn't have eyes. Instead, everything's based on, you know, sensing gravitational waves. They were much bigger to be able to do that. But. But, you know, or I think it's a question of what. What is the category of thing that you're trying to ask? Does it happen? Because in the universe, lots of stuff has happened. The things that we choose to call life are mostly things. Well, so far it's purely things that are historically related to life on Earth. Things which have the particular details of life on Earth. Cell membranes, you know, rna, lots of very, very detailed things about the particular history of life on Earth. The general category of thing that we would characterize as life, we don't really know what that is. My. My guess is, and. And I've sort of increasingly beginning to understand this. I suppose it's this Kind of weird phase of matter that is living matter, that has this sort of bulk orchestration of activity where the molecules aren't doing just completely quotes, random things like they are in a gas, where we as observers like us just say, oh, it's random. But they're doing things where there's a lot of detail in what they're doing, which is sort of understandable to observers like us if we were able to go in with sufficiently good sensors and so on. But yet. It's not something where we can just summarize everything and say, this is how it all works. It's something where there is sort of detailed action, but it's all orchestrated, but it's a lot of detail. It's kind of like if you were to open up, you know, your average computer, your average microprocessor, you'd see electrons going around all kinds of complicated patterns, and you wouldn't. You would say they're going around in complicated patterns. But there is some sort of meaning to what's going on. There is some organization to what's going on. There's some organization that we can recognize in what's going on. And I think it becomes a bit of a circular thing because we are life as life as we are. And if we say, well, is that other thing over there life? We're going to project our own form of organization onto answering the question of, is that life? Is it kind of life like we think about life, or is it life in a. In a, you know, in a. In a way that's. How far away, you know, how far away in the space of possible things that we would call life? Can you go and still have us identify those things as. As life, so to speak? So I think it's. It's a. What, you know, is there bulk orchestration that could happen in a way that's different from the way it happens for us? I'm sure the answer is yes. And how far away would it be? How recognizable to us would it be? I'm not sure. And would there be things where we could summarize what's happening by saying, oh, it's using telepathy or something? I'm not sure. I think this is a. Is one of a whole category of questions about sort of how other minds work that I think is really interesting, and I've tried to poke out a lot, but I don't really know how to make that work. Serge comments about my previous thing about kiwi fruit. Would it be easier to genetically modify the kiwi itself? Yeah, it might be. Although taste is known to be very sensitive or dependent on genetics. So, for example, if you take different varieties of apples, for instance, you know, a Golden Delicious apple, my favorite is they are all genetically identical. They are all clones. And if you took something which was genetically even slightly different, it would taste significantly different. It really seems to depend. Taste seems very sensitively dependent on the very particular, you know, amino acid sequence in those particular proteins. So the modified, you know, it's a tricky thing. You'd have to. You have to find out, well, what's the. We don't really understand food allergies very well to begin with. It's kind of like, well, what part of what protein is the thing that is. And maybe you can find that out by just looking at the immunoglobulins that exist in high abundances in the immune system when you have that allergic reaction. I'm not sure. And say, well, what does that match on the original protein of this particular fruit? And then let's zap that part. But count on the fact that it's mostly other parts that are leading to the characteristic taste it has. Yeah, reasonable. Reasonable thought. All right. I think I need to wrap up and go back to my day job, but I look forward to chatting with you all another time. See you. See you on another stream. Bye for now.