Future of Science and Technology Q&A: Live from the Wolfram Summer School (July 3, 2024)

Hi, everyone. Welcome to another episode of Q and A about future of science and technology. Just for a change, I'm actually here with a live audience. This is part of our annual Wolfram Summer school. It's actually also the same time as a Wolfram summer research program for high school students. And so actually, one thing that's really cool is that there are several people here in person who have been asking questions on, on several or many of our live streams in the past. So we get to actually find out who some of those handles who've been there in handle form only are in real life. So let's see, let's start off here. There's a question here, very down to earth question from Peter. How do you see electricity being transmitted or provided to households in the future? These power poles and power lines are over 100 years old. It's kind of an interesting thing because lots of technologies sort of alternate between being local and being global. I mean, there was a time when I remember back in the early 80s, people discussing, would computers be personal computers that were local? Would they be mainframes that you communicated with? I even remember having this whole brainstorming session about computing power stations. That was what we called them back in the early 80s. And at the time, it seemed like the kind of computers on your desktop were definitively going to win. And then later on, you know, that those won for a while. And then we had cloud computing and kind of flopped back the other way. For electric power, it seems like the central generation, you know, like the computing power station idea, the central generation idea has been the winning idea. Is it going to go on being the winning idea? If you believe in photovoltaics, those kinds of things, that's local. If you believe in, well, geothermal, you need very deep geothermal stuff. And I don't think everybody's going to have that in their backyard. But there's this question of is electricity going to be generated centrally or in a distributed fashion? And, you know, there are lots of different issues that come up. Like, you know, electric power lines dissipate lots of lots of energy. You know, you. You ramp them up to high voltage, they dissipate less energy, but they still dissipate lots of energy. If one day somebody decided that, discovered a robust room temperature superconductor, all those electric distribution lines would immediately dissipate no energy. And it would make even more sense to have power centrally generated. I think it also depends on how you're generating the power. You know, nuclear power stations. Right now, nobody is imagining that There will be a nuclear reactor in your basement. It's not totally obvious. That's impossible. But it's not something that, well, I guess it has to be somewhat large because you've got to be up to the critical size of critical mass of whatever material you, whatever radioactive material you're using. But it's not, you know, it's a, it's a physics question. Whether we can have, you know, nuclear reactors in our basements doesn't seem promising right now, but not obvious. And you know, if it turns out that fusion reactors actually work again, it's a question of at what point those can be scaled down, whether the economies of scale are such that those are, those are huge or not. So I think it's a, you know, we're at a point where it's not obvious what's going to happen in terms of should things be local, should things be global, should things be local with computers, should things be global with computers and so on. And I think it's also somewhat amusing to me that sort of the location of computers, where do the server farms go? The fact that it's beginning to seem feasible that it makes sense to have servers that are either on satellites or on high altitude drones or something like that, making a, a reality to the idea that, you know, there is a cloud. The cloud. Where is the cloud? The cloud is in the clouds or above the clouds. I just think that's a sort of amusing thing that that can happen. Let's see, I'm going to move on here. Let's see a question here. How often will AI be revisited in future science and technology? Or do you think AI has firmly cemented its place? I mean, I suppose what you're asking is if you look at the history of what's now called AI, it's had many ups and downs. I mean, the term AI was invented in 1956, but people had had automata, mostly toy like things that were kind of, you know, here is a toy duck that seems to do things that are like what real ducks do and so on, those that existed for hundreds of years. I think that the question of sort of what's happened since 1956 is kind of a roller coaster of, oh, it's working great things are happening. You know, we've got very early neural nets from the 1950s, early 1960s. You know, we're going to be able to recognize images and then, oh, no, that doesn't work. And then by the 1980s, it's like, oh, actually we've got stuff working again. And then that kind of died off. And then 2011, kind of the image identification story, and now things can be done with deep neural nets and things start working again. And then in 2022, we had kind of the surprise of ChatGPT. So I guess one. One question is, you know, will there be a series of further kind of moments of surprise with what happens? It's worth remembering that the current kind of achievements have all been on the basis of kind of a single idea. And that single idea is this idea of neural networks and neural networks. You know, it's this sort of connections of arrays of numbers, sort of layers of arrays of numbers and so on, connected with what amounts to kind of the informational equivalent of wires, each with weights and so on. And it's been quite surprising that with that sort of same architecture, it's been possible to have a series of levels of achievement. Why has that worked? Probably one reason it's worked is because neural nets, as their name might suggest, are. Are based on things that neurons do in our brains. And probably the reason that the things that we have wanted to do are things that we have been successful at doing with neural nets are things that correspond to things that we do in our brains. Because the neural nets are kind of at least an idealization of what's happening in our brains. And so what we're doing is kind of being able to do things that are like what we're able to do in our brains. And so we're impressed with those things. There are plenty of things that computers can do that neural nets find it really difficult to do, but those have not been the things that have been the recent focus, at least of AI. So it's sort of a question of what will happen in the future, of what will be the future. If you're saying kind of the neural net path of AI, what will be the future of that? I mean, right now, neural nets we have are a bit smaller than brains. They're gaining on brains. And brains still have the feature that they kind of simultaneously learn and do things, whereas most, at least presumably brains, do that. I mean, there's some unclearness about the extent to which brains kind of learn a bit during the day and then solidify those memories in the night and things like this. So. So maybe those are a little bit like the difference between kind of training in a neural net and inference actually running the neural net. But we seem to have some features of, you know, we do both training and using our neural nets at the same time, which is something not yet Supported in the kind of hardware of current neural nets. But that will presumably come. But overall, we have kind of. Our brains still have not only that hardware advantage, but also just being bigger, 100 billion neurons, give or take, in our brains, which is bigger than the few million neurons typically that exist in artificial neural nets. Typical artificial neural net will maybe have a few million neurons, and then it might have many, you know, close to a trillion weights connections between those neurons. But so a question is, well, what happens if you exceed human kind of neuron counts, your artificial neural net? What does something remarkable then happen? And certainly we're. It's sort of. There's this interesting fact that sort of critters with smaller brains than ours seem to do less than we do. So that kind of is a. Is a clue that the bigger brain might be able to do a bunch more. And I think the, you know, the big thing that sort of what do we do that's qualitatively different from what other critters do? The main thing we seem to obviously qualitatively differently do is use human language and have this kind of modular compositional structure of language for packaging up ideas and communicating them, presumably. And that language makes us believe that we have the sort of processes of thought that we can then express in language. What happens, whether there's thought that can happen sort of without language is an interesting question. Whether we. To know that thought happened, we. We typically have to have something like language. To know that abstract thought happened, we typically have to have in the end, something like language. But whether there's a sort of a thought distinct from language is not clear. But it's sort of obvious that we seem to be able to do a bunch of things that. I mean, language is sort of a unique thing to us. And also, I think language is part of the story of formalizing thought. I mean, you know, as you. As you start sort of trying to piece together, should I do this, should I do that? You know, the tiger is trying to figure out should it jump this way, that way one has the impression might not be right because we don't kind of know what the inner experience of the tiger is. But we have the impression that it doesn't get to kind of build up levels of thinking about that. It just does what it does. Whereas for us, at least, my internal impression would be that I can sort of build up levels of reasoning and thinking about things that because each of those levels, I can sort of concretify my brain by playing to myself a certain kind of language, a Certain representation in language. I can kind of say my thoughts, so to speak, to myself. And that seems to be something that's necessary to kind of build up kind of higher levels of thinking. And that seems to be something that we're able to do. And, you know, what's the critical number of neurons you need to have language of the kind that we have? We don't know. We do know that we have had the experience with neural nets, artificial neural nets that have been built up, that the number of different kinds of things that can be, for example, recognized by neural nets seems to increase with the size of the neural net and also with the depth of the neural nets. But people were recognizing handwritten digits, for example, in the 1980s with neural nets with maybe half a million connections or something in them. Half a million. A million weights. To get to the point where we can do sort of fluent human language, type things seems to need maybe tens of billions of weights. So it's kind of this. There seem to be these sort of thresholds of. In the case of. Well, what happened at first was handwritten digits. There are only 10 of those. Then there's kind of the thing that, you know, can you do the whole Alphabet? Can you. Can you do alphabets that aren't just the, you know, the 26 letters in the English Alphabet? Can you then, you know, can you identify. I mean, in English, there are about 5,000 picturable nouns. Can you identify all of those kinds of things that became possible around 2011 with networks with, I don't know how many. It must be a few tens of millions of connections. I think so. You know, there seem to be these thresholds. If you have brains of a certain size, you can do a certain level of thing, you can deal with a certain number of kinds of objects. So for brains of our size, 100 billion neurons, we seem to have languages that have about 50,000 words in them. And there's a question. If we had brains with, you know, with a trillion neurons in them, you know, what would then happen? Would we be able to have much finer distinctions about the world? How would we deal with that situation? In a sense, one of the things that is notable about our brains and the way that we think about things is we always seem to remain coherent. We have a definite thread of experience. We don't have. At least we don't have the impression that we have many different sort of threads of thinking going on at the same time. We seem to concentrate things down to sort of one thread of thinking, and that's probably something for which there's some architecture in brains to make that possible. But it's something where we. The bigger the brain, perhaps the more difficult it is to synchronize things. I mean, it could be the case that. I think people say that smaller brains and other critters can have faster reaction times because there's sort of less going on. But perhaps part of the point is that, you know, the brain, if the brain is smaller, the synchronization time is correspondingly smaller. And before you have the next thought, so to speak, you have less, that you have to kind of keep in synchrony to sort of go on to the next thought. And so it could be that you have a bigger brain and you still have the same transmission rates for neural impulses and so on. It could be that with a bigger brain, you'll actually think progressively slower. Maybe you'll be able to think about more things, but you'll be able to think slower about those more things. What would it mean to think about more things? What would it mean to have kind of the. In typical languages, we might have, I don't know, 30,000 fairly common words, and then a few others are slightly more specialized. If we go into some specialized area, we might have a whole bunch of other words about medicine or law or physics or whatever else. And if you combine, I don't actually know how many sort of. Actually, I would guess it's on the order of 1,200,000 words that are sort of the complete inventory of what you would find in sort of the full dictionary that includes not just our everyday words that any given person uses, but just the full inventory of sort of what's out there altogether? And I suppose one can ask the question, does that full inventory fit in human brains or not? Is it actually even possible in our brains to kind of know the full. Every field with all of its different words, One has the impression that we're not that far away from that. But maybe there are limitations that are just. You have a brain of a certain size, you can only learn a certain number of distinctions, so to speak. What would it be like if you were able to have a much larger number of distinctions? And would that be sort of useful to the AIs to be able to deal with that? I think one thing I've been thinking about recently, it's my little, little obsession of recent times, but I don't think I've really figured it all out, is what if not only could you deal with kind of thoughts of the kind that we have in human language where you have this stream of, of modular words and so on, but you could also kind of imagine an image and have that image of a whole array of pixels and so on. And you could kind of take the mental images that we normally just sort of have somewhere inside our minds and concretify those mental images in the same way that we kind of concretify thoughts as sequences of words and so on. And of course with generative AI, we perfectly well can take, you know, take an image and take, take, have the AI kind of generate a rich image. And I suppose my, one of my little, you know, what, what would it be like if, well, us humans were reading out sort of our internal neural signals as images all the time. So I'm not just standing here yakking at you all, but instead I have some kind of display which is kind of displaying my current thought pattern, so to speak, in the same way that the words I'm saying are a very sort of narrow thin channel for expressing the thoughts that I might be having. And so I think the question, so that's kind of what is it like if you could have image based communication of thought? And I suppose the question is when we think about sort of the future of AI and what will the AIs, what new levels of capability, what new kinds of things could AIs do? One of the tricky things is there's things AIs could potentially do that are sort of computational kinds of things. The question is, do we have an interface to those things that's useful? Will we humans, the AIs can go off and compute all sorts of things, but do we, are they doing things that we humans will care about, so to speak, or will be able to connect to? And I think that's, that's sort of less obvious to, to me at least what the, you know, I've been really trying to imagine what does the kind of the, the, the trillion neuron quadrillion neuron alien mind, what is it potentially capable of doing that is as far away from what your average lizard does from our language as, as, as, as it is from, from going from, from our language on, so to speak. That, that's. Anyway, I think that's kind of a, a sort of an attempt to kind of talk about sort of where, what, what is the big step that goes beyond kind of the, the, the AI that we, that we see today. So let's see, I see a bunch of questions here from AC. Do you think LLMs have already passed the Turing Test, which is being asserted by many experts. If so, what does that mean for the future direction of AI research? What's missing? You know, I. The Turing Test is, you know, at the time of Alan Turing, when he wrote about kind of what it would mean for machines to think in 1950, it was a different time, a different set of expectations. And, you know, in. In that same paper, he's talking about things like telepathy and so on, which was a, you know, something people sort of generally believed was right around the corner. That was, you know, something that we haven't observed as of now. But it was kind of a different scientific and intellectual time. And I think at that time, it seemed inconceivable that you could just sort of type to a computer or talk to a computer and that it would come back with something that sort of behaved in a human like way. And indeed, that's something that for a long time seemed inconceivable until it turned out that in 2022, one got to the point where large language models could. Could produce pretty plausible kind of natural language text. You know, I think this question of, is it human? Like, how human, like, is it. Which was, you know, Alan Turing's test is just, you talk to the thing. Does it seem kind of like it's a human in there or not? Well, that's a weird test for many reasons. I mean, there are humans you can talk to where it's kind of a. You know, it's like they're very different from oneself and they're, you know, it's like talking to. Well, not to a machine, one thinks, but it's certainly not something which one can immediately identify as being like oneself. And I think that the question of whether, you know, when you. It's like people in past years sort of connect WolframAlpha to a Turing Test Test kind of system. And it's trivial to tell which is the system that has WolframAlpha connected to it and which doesn't. You just ask it a whole sequence of totally obscure different questions. If it can answer them, it's a machine. If it can't answer them, well, it might be a human then. So I mean that. Which is kind of the opposite of what you would think and what Alan Turing imagined, which is that the Turing Test is sort of the pinnacle of machine achievement. So I think in, you know, can you have a small talk type discussion with a machine? I don't know. Actually, I'm planning in a few days to try and do a discussion with an AI, actually a robot, even A humanoid AI. So we'll get a chance to sample how well that actually works in current times. I think that in, you know, I see kind of the modern round of large language models as being more than anything else, kind of a linguistic user interface kind of layer where you can, in the past you had that was, you know, we've defined ways to make kind of, to let one formalize things in computational language. I think that's, I was talking before about the extent to which having ordinary human language allows one to formalize thoughts and build kind of towers of consequences on those thoughts. I think that sort of tightening up that formalization, extending that formalization, that's kind of the big effort that we've had with Wolfram language over the last four decades now to sort of provide this level of formalization that allows one to kind of build towers of thinking which both one can do oneself and one can sort of delegate to one's machine. But I think in this question. Well, so I see kind of the current round of LLMs as providing this kind of layer of going from the language that we currently have, not the computational language that we could imagine having, but the language that we, we routinely use to talk about things and you know, providing kind of an interface between, from that to things like computational language. And I think that's kind of the, you know, that's the best use case for, for LLMs as they are today. When people say, oh, I'm going to, you know, AI has gone so far in recent years, it's obvious we just have to feed it enough training data and it's going to be able to figure out everything about math and physics and everything else. I think that is nonsense, actually. I think that's just not, it's not going to happen that way. I think feeding lots of kind of mathematical structured kind of information to an LLM is useful for helping it learn more about how things fit together. I think that LLMs learnt logic by having seen lots of kind of correct logical arguments fed to them and they were able just by their sort of nature of forming patterns and so on to recognize these kind of structures of logic. I think that's kind of that type of thing one can expect to continue when LLMs see lots of kind of step by step math or whatever else. But the idea that the consequence of that, and that can be very useful to LLMs. But the idea that the consequence of that is, oh, we'll be able to sort of figure out everything directly in the LLM, I just don't think that can happen. I think that's sort of the whole idea of computational irreducibility kind of, sort of tells one that that can't happen. But I think the use cases for LLMs that are really good are the ones that involve. They are the sort of linguistic interface that go between our natural language way of expressing ourselves and some underlying computational process or something else. Let's see. Jake asks, over time, AI training data will increasingly be AI generated. Will this feedback loop amplify errors and cause AIs to self destruct? You know, I think the more likely thing is that it's a kind of a reversion to the mean. In other words, if the, you know, the AI has learned from things, it's kind of interpolated all those things, it's generating something which is sort of some average of what it's been trained, and then you feed it more of that average and you're kind of averaging, averaging, averaging to get to the mean. Now you could change that. You could make an AI with attitude. You could inject some kind of totally crazy thing in there. What would happen then? I'm not sure. I think the most obvious thing and the thing that's sort of commercially happening right now is people sort of want the AI to be uncontroversial. So they're kind of making it kind of do the average things and so on and not be too wild. If you say, let me insist that the AI is wild, let me insist that the AI is always saying things that are extremely surprising, and then you put all those things on the web. I don't know what happens in that case. That's not what's happening right now. Most of what AIs are generating is sort of fairly banal kinds of content that is sort of uncontroversial. Because most of the use cases, most of the reasons people put it on the web is to be kind of, yes, that's kind of sort of what we would expect to see rather than that's crazy and wild and not what we expect to see. So I think it's an interesting question. What would happen if you chose to specifically put content on the Web and put, you know, There are what, 5 billion pages or something of somewhat meaningful content on the Web? If you put in another 10 billion pages of AI generated content that has specifically been given an attitude, so to speak? I'm not sure what that would do. It's an interesting question, and I think that that's not where the commercial forces are today. But I'm not Sure. What would happen? Certainly right now, as you're sort of averaging, averaging, averaging, you're getting to the more and more kind of reversion to the mean more and more banal output. But that's not necessary that that would be the way it works. Mindchell asks if we can sustain mini brains or large clusters of human neurons for years. This approach might achieve artificial general intelligence before synthetic methods do. What do you think? Well, you know, it's kind of a weird thing. I saw one of these little organoids made from a million human nerve cells. I'm kind of wondering whether that was like clone nerve cells from the CEO of that company. I don't think so, but that would be kind of interesting. It's and avoids some kind of ethical approval types of things. But you know, in that case it's a bundle of a bunch of nerve cells. You probably grow much bigger organoids. People say that those organoids are not yet organ way that cells are organized in actual human brains. There isn't probably this sort of overall developmental flow that exists in actually having a brain be created. It's kind of like when you have blood vessels, for example. The big blood vessels are driven by all kinds of large scale features of flow. But by the time you're filling in the capillaries, they just sort of fill in where there's space for them to fill in. And they fill in in a somewhat more random way. They're nice cellular automaton models actually for that. But the thing that you. So I don't think it's obvious at all that the organoid you just sort of grow in a petri dish, so to speak, or from stem cells or whatever, that it doesn't have overall architecture that's just a feature of the current technology of those things. But so then the question is, you've got this organoid and it's got lots of nerve cells in it, and then what do you do with it? Well, for example, if it had eyes and ears and you'd built out that whole biological structure, then, well, you can communicate it with it by talking to it, so to speak. If you don't have that, you're a little bit more stuck in terms of how you talk to it. I mean, in a sense that's kind of. It's the flip side of kind of a brain computer interface problem. We imagine that brain computer interfaces are. You have a whole fully developed human brain and then you're putting in something which will be a direct brain interface where it's dealing with the Fully developed human brain and its kind of neural signals and so on. In this case, it's like you've got this thing that's sort of a brain like object, and you're saying, can you start off communicating with it? Can you make contact with it? With an ordinary human brain, you can tell the person, think about this or whatever else, and the person might understand you and be able to start thinking about that. And you'd see what effect that has on neural signals for the brain. Computer interface. But in the case of an organoid, the first question is, how do you communicate with the alien, so to speak? What do you do? And you can certainly make electrical connections to this thing, but then how do you start giving it any indication about what's going on? It's a very brain in a jar kind of story, quite literally. And it's like, how do you communicate with the brain in the jaw, so to speak? And so I wonder, you know, what could you, how could you give the brain in a jar a worldview? And you could certainly imagine a situation where you're giving it electrical signals. And for example, you know, with some electrical signals, you feed in some electrical signals, it generates some electrical signals. If you like what you see, you give it dopamine or some, some other something like that to make it feel good in some sense, otherwise you give it something else. And maybe by that you can educate the brain and the jaw, so to speak, to do things that are sort of useful perhaps, or, you know, how do you, how do you give it, you know, again, if it has sensory organs, then, you know, for example, another thing you could imagine is that just like when we, you know, we have to learn to see our, you know, when we are, we see arrays of pixels that come into our retinas and in our visual cortex as we, you know, when we're very young, we're kind of recognizing the correlations that exist in the world and that's causing connections to be produced in much the same way that neural net training for convolutional neural nets work. And from that we develop edge detectors and all those good kinds of things. And there are those terrible experiments with cats from a long time ago where cats were reared in this environment of only vertical stripes. And those cats never developed the kind of detectors for horizontal motion. And so we kind of know that we have to learn to see. And so one could imagine that this organoid could kind of learn to see in a similar kind of way. But can it learn to talk? Can it learn to, you know, Package up its thoughts in the way that we have with human language. That seems more challenging. And it seems like, you know, maybe if you give it sort of a audio input, it can kind of listen to these phonemes. And maybe if you give it, you know, some output, maybe there's a way of having it just by virtue of getting this feedback loop, of having it progressively learn what's going on. I'm not sure. I think that's an interesting possibility. Now, it's a funny thing. I mean, let's say that the hardware of. I tend to think that it would be quite embarrassing for us, in a sense, if all of our technological development is actually worse than what biology provides in this regard. I mean, it's kind of like, you know, is it better to ride horses or to drive cars? For a long time, horses were the only option. But eventually, you know, we got cars which are presumably a sleeker. You know, they're more efficient in many kinds of ways. Maybe we think that, you know, it's. They're not as natural as horses, and maybe we don't like that, but they certainly are more efficient in many ways. So the question is, when we're doing computational kinds of things, should we piggyback, horseback, ride kind of the computers, or should we have something which is pure engineered technology? Now, I think I would say that the. If we think about actual neural hardware as it exists today in brains, the typical sort of characteristic times associated with the electrochemistry of neurons is of order thousands of seconds, milliseconds, whereas the typical times associated with digital electronics are more like billionths of seconds, nanoseconds. And so there's clearly a big advantage there. We don't have, as I mentioned before, the same kind of parallelism of many things operating in parallel in digital electronics yet that we do in brains. But that seems to be only a matter of more a matter of how do you organize the computation. Because actually, here's the tricky thing. In brains, we got all these neurons doing their own thing, but somehow we concentrate that sort of thread of attention and experience down to the single thread, even though all the neurons are doing their own thing. When we do computing, we are usually thinking about things in this sequential way because that's how we think, because we've concentrated ourselves down into this thread of experience. And then the question is, well, how do we get all these components in our computational system to then do things in a very parallel way? Because we have, for our understanding, we've concentrated things down into this sequential thing, which happened in some mechanism in our brains. But now we have to kind of reverse that and make everything run in parallel, In a sense, in neural nets, we have done that in the sense that when a neural net runs, what is happening, the flow of data through a neural net kind of distributes everything sort of across all the neurons and so on. So in a sense, we're doing that in that case. But in general, in computation, that's not yet something that we routinely do. I think that, let's see, we were talking. So there's. There's currently an advantage in brains that things sort of run in parallel, even though in the end they concentrate into a single thread of experience. There's a big disadvantage by a factor of a million in terms of the speed with which the components run. You know, will it be the case in the end that it's better to use kind of biological like hardware than it is electronic like hardware? I strongly doubt it for the kinds of computations we want to do. Now, there's a little bit of a caveat to that because, you know, biology does things at a molecular scale, but it seems like the only stuff that matters in brains is stuff that happens at the level of whole neurons rather than at the level of individual molecules. I mean, people wonder whether, for example, what, what does a neuron. What is important about a neuron? Is it just the kind of the density of little calcium channels or something on the, on the synapses that it has with other neurons? Or is it much more detailed stuff about the individual molecules that exist inside that neuron? Is memory encoded in RNA molecules or something like this? Or is memory just encoded in the overall connection strength of one neuron to another? One might have thought that human memory is so detailed that, oh, it must be something down at the molecular level. We must be encoding things, you know, in individual molecules and so on. But I think one of the things we've learned from LLMs is that, you know, a trillion connections does a lot and that, you know, it probably isn't the case that if we imagine all the stuff that any one of us remembers that, that, oh, you can't fit that in a billion neurons. It seems very plausible that it can fit where sort of a one bit roughly per neuron is enough to encode everything that any of us know, kind of that, that. So I, I tend to think that the, the, the future will not be, you know, the future is more likely to be electronic than biological. Now you can also ask, could we go and actually do things at a molecular scale? You know, Right now, in a typical computer, every bit is about 100,000 electrons. Could we do things where we go down to, you know, every bit is just one electronic where we don't have, you know, in a typical computer today, the individual transistors are separated by, you know, at least, well, what is it? Tens of nanometers at least. So many, many atomic diameters. And, you know, can we go down to something which is more operating at the level of individual, you know, sort of individual atoms and so on? First there's the question of can we reduce the scale to individual atoms? Can we increase the speed so that it's, you know, instead of gigahertz, it's terahertz, for example. We can't go too much further than that because if you start trying to increase the speed too much, you kind of run into kind of limitations of sizes of atoms and so on. So you can't, you know, there's a limit to how fast you can go that way. But then the other question is, the thing that happens in lots of at least chemical processes is that you've got a zillion molecules doing things. You know, you'll have 10 to the 23 molecules involved in some, you know, beaker of. Of stuff. Whereas in, in. In digital electronics and all those molecules, if they, if they are able to be directed to do different things, you've got a huge sort of computational power there. So I think there's sort of a different direction which is going down to the level of. Of kind of using individual molecules, individual atoms to do computation. And it's, you know, there are. There are different efforts to do that. For example, many of the efforts to make quantum computers are sort of attempting to do things at the level of individual atoms. You know, you can have a quantum computer where you've got lasers moving around individual atoms. Right now it's a few hundred atoms. Maybe it'll be more than that. But right now it's a very. It's kind of the lasers are being manipulated in a pretty macroscopic way, although the actual sort of storage of bits is roughly individual atoms, although individual atoms with kind of a large cocoon, a large kind of. It's like if that's the brain, there's a big body supporting the brain. You know, I was mentioning organoids before, at least one of these that I've seen. It's kind of like you've got the organoid brain and you've got the rest of the system, and the rest of the system has kind of an imitation of of components of the human body because it's got to keep this brain alive with glucose and supplying oxygen and all these kinds of things, just like, you know, actual biological bodies sort of support brains. But I think, you know, this question of, of, okay, we've got, you know, in, in quantum computing. There are different schemes for quantum computing, but if you can, you know, some of them deal with just atoms just hanging out there, sort of trapped in, in arrays of lasers and where you're moving individual atoms around, although you're moving them around incredibly slowly right now. But, you know, if you could figure out how to control that, maybe you could do that much faster. But that's sort of the atoms on its own and not even a gas that's in a state which is so, you know, where the atoms are so separated that they're just sitting there in these kind of, you know, micron apart, microns apart kind of wells defined by lasers and so on. But, you know, then the question is, can we, if we really want to use kind of molecular computing, we would like something much higher density. We'd like something like what happens in a liquid where there are gazillions of collisions every second. And so then the question is, can we. What would it take to, to make, to do computation instead of liquid phase computation? Sort of an interesting question because it's like, what can we compute that in the seemingly random collisions of molecules in a liquid? We can successfully kind of define the computation and read out the results. It's been my suspicion for a while that much of the dynamics of what happens in something like a liquid is more determined than you might think. I mean, you might, as a model of a liquid, you might just say, oh, the molecule's bouncing around randomly, but in actuality, the molecules collide, the molecules bounce back the way they came, those molecules might hit other ones. You have something which ultimately has kind of a causal sequence that is quite deterministic. I mean, you can maybe you have some sort of noise from the outside in the form of, let's say, you know, photons, electromagnetic radiation, something like this. But otherwise it's kind of. Or you have things happening from the walls of the container, but. But otherwise the molecules on their own are bouncing around in very deterministic ways. Can you use that? Well, that's a big challenge. I mean, in a sense, the second law of thermodynamics is all about the difficulty of using that. The second law of thermodynamics says that what starts as sort of organized motion of things, mechanical work, Tends to sort of inexorably degrade to this sort of seemingly random collection of atoms just bouncing around in a way that looks to us disorganized and random. And in a sense, that's the. To us is the important part of that. When we are sort of observers who kind of can't trace all these individual molecules, can't define the sort of states we'd need to have these molecules operate in ways that are useful to us. It just looks random. And we just say, that's not doing anything useful. That's why we call it heat. We say that's not what we want. It's different from mechanical work. But in actuality, those molecules are doing a computation. It just doesn't happen to be a computation that we can readily connect with. So then the question is, if you have something like a liquid, is there any way to. Which is kind of a little bit like these molecules just bouncing around? Is there any way to set things up so that we can encode sort of the thing we want in the configuration of molecules, then let it sort of be a liquid, and then read out the results? It seems hard. It seems like something that, in a sense, is sort of breaking out of the second law of thermodynamics, Breaking out of the law of entropy increase. Breaking out of the fact that to us, entropy always seems to increase. To us, the things that we see seem to be sort of given the constraints we have, seem like we have to just treat them as random. But in fact, there's one big example where it doesn't work this way, and that's biology. And my increasing suspicion is that in biology, one of the key sort of discoveries of biology is don't use liquids. That is, if you have a liquid, don't use things where the molecules really have to be thought about as just sort of bouncing around randomly. Instead, and this is sort of the meta discovery of lots of modern molecular biology, is that there's more and more sort of orchestration of the motion of molecules, More and more, oh, this molecule carries this other thing and arranges this other molecule so that it's able to react with this other one and so on, that it's kind of a big. It's more like a machine than it is like sort of traditional chemistry and traditional sort of liquids with molecules randomly bouncing around. It's kind of interesting to imagine what makes that machine. Can you have something where, for example, you have a sort of biomatrix of kind of sum of materials where the sort of the behavior of the Molecules is somehow orchestrated. You can imagine kind of evolution sort of having some. There being some definite rule for sort of creating this biomatrix. Maybe it's even one of my favorite cellular automaton rules that sort of creates this biomatrix. And then within that biomatrix, it is something where the sort of orchestration of the interactions between molecules to be able to do something which is sort of useful to the biological organism. I mean, I think another worthwhile analogy is you have a microprocessor where electrons are kind of being organized in a way to do something sort of. That is a useful computation. So similarly in biology, molecules are organized to do some kind of useful computation. And perhaps the sort of biomatrix is something like the circuit diagram of the microprocessor that's kind of arranging exactly how these molecules will interact with each other and so on. And there's sort of a question of, okay, if that's what's going on for a microprocessor, how do we describe it? What's the point of a microprocessor? It's presumably doing. It's doing some set of computations that we humans want to do. And that's kind of the point of a microprocessor. You typically are not feeding a microprocessor random voltages on its input, input pins. You're doing something which is what we humans want to do. And so then what we're doing in the case of. In biology, it's similarly the case that we don't expect that the sort of. The input to the biological organism is something kind of completely random. We expect that the kind of the biological organism exists in some environment and with some objectives, in a sense, that define a certain set of inputs and constrain what kind of outputs it will generate. So I think that's the one's guess is my guess would be that while leveraging existing biology with amino acids and who knows what else is probably not the sort of the path forward for computation, leveraging the sort of core idea of biology of using molecular scale things and moving molecules around rather than just moving electrons around might be a good thing to do. Although it's not obvious what the trade off will be when you're interested in just in computation between having your working fluid, in a sense, be electrons versus your working fluid be atoms of different kinds. I mean, if your goal is to make something that humans can eat as food or something, it's no good to just have electrons. You have to have actual atoms. But if your goal is to do Something abstract and computational. You might as well do that with electrons. It's kind of an interesting analogy, I suppose, to people thinking about kind of in the early days of thermodynamics, thinking about kind of what we're trying to make as a heat engine. And we could have, in that heat engine, we could have steam, we could have air, we could, could have some other substance. But in the end, there's sort of an abstract notion of a heat engine, just like there's an abstract notion of a computer that doesn't really depend on what that working fluid is, whether in the case of a computer, whether it's electrons or atoms or whatever else. Anyway, a few thoughts about that. Let's see, Spare Parts is asking, are those neural cats behind you? I believe the answer is yes. I did not pick this, this, this backdrop. I, I showed up here and the backdrop was, was there. Let's see. Okay. Misha is asking, is it possible that human machine integration or radical genetic modification can. Can allow humans to make significant leaps in rulial space? So, I mean, I sort of was trying to address a little bit of that earlier that, you know, how do we, when we think about rulial space, this is sort of the story of a ruliad concept and the idea that sort of different points in the ruliad correspond to kind of different ways of viewing the world, in a sense, different paradigms for thinking about the world. And the question is, how do we explore the ruliad in physical space? We explore things by sending spacecraft out in the ruliad. We explore things by sort of having paradigms that cover more, by having kind of a broader way of thinking about the world. Now there's a question of how do we get to that point? And for example, we could have gone so far. The problem is, in a sense, if we have a paradigm that's too big, that covers everything, in a sense, we no longer have some sort of coherent view of what's going on in the world. By the time we're covering everything, we're sort of too dilute to have any kind of coherent way of thinking or in a sense, coherent existence. I mean, insofar as we know our existence by the internal thoughts that we have, to say that we're kind of distributed across everything is kind of to say that we, in a sense, no longer exist. So the question is, what are the best ways to kind of expand in rulial space? I happen to think computational language is one of the key ideas that's kind of the analog of A spacecraft. It's kind of something which provides an organized structure for letting you move around in paradigm space, so to speak. Just like a spacecraft is sort of an organized structure that you get in and you, you know, rather than just trying to walk to Alpha Centauri, which is not going to work, you know, you get in a spacecraft and it has, it has technology that is appropriate maybe one day for getting you to Alpha Centauri. So similarly, I view sort of computational language as the analog of that, the tooling, the technological tooling necessary to sort of build up the structures that you need to move in rulial space. In physical space, you move by sort of firing rocket engines or something. In rulial space, you move by kind of building up paradigms and sort of building this bridge of paradigms that goes ever further. And I think that's something that we have with computational language. We have kind of the ability to fashion these kind of bricks that allow us to kind of move around in rulial space. So I think that's my best answer to that. I think the question of which way we go in rulial space is a whole different question. It's like, well, where do we explore in physical space? We go this direction, we go that direction. We'll end up with a different point of view about the universe. If we decide to succeed in doing a big expedition to the center of the galaxy, we'll have a certain view of what the universe is like when we're at the center of the galaxy. If we do another expedition and we manage to escape the galaxy, a challenging thing to do, its escape velocity is like a, a million miles an hour. If we succeeded in escaping the galaxy, we would have a different point of view about sort of what things are like in the universe. And so I think in sort of in rulial space, we can't go in all possible directions. We sort of don't have the computational resources and we don't have in our sort of coherent existence. We have to make choices about which direction to go. And then the question is, well, when we make those choices, we're sort of building these highways, this network in rulial space of where we have chosen to go. And that's something that's the ultimate reflection of kind of human choice in some sense. And I think that's the question of sort of how we drive down those roads. That's, I think more than anything, sort of a story of computational language where we build the roads. That's a matter of human choice that we can select that, we can sort of figure out what are the possible roads. Well, that's something where we're just enumerating programs and seeing what happens. But the typical thing is we enumerate programs, you see what happens. Many programs do things where we say that's really surprising, but we don't really have a sense of, we don't internalize what that program is doing. We don't see why we should care about that program and so on. Let's see, Rawr is asking what role do emotions play in language and information processing? Do emotions speed up communication? What other elements are important for AI development in communication beyond language? You know, I've always assumed that emotions are a very coarse sort of overlay on the fineness of human language. I mean, we know, you know, Charles Darwin wrote one of his later books about the, the expression of emotions in animals and man because he was trying to make the point that there's a lot of stuff that's in common between animals and people and I think other animals and people. And I think that in a sense that's a low bandwidth communication channel, that it's a small amount of information transmitted in kind of emotions and probably their sort of origin, they're literally coded in the concentrations of a limited number of neurotransmitters in the brain and so on. And I think that's, to my mind, that's sort of an overlay, a piece of metadata on top of the much more fine grained information that's available in sort of human language. And I think this notion of, oh, you know, can we get the AIs to have emotion? Sure we can. I don't think that's a challenging thing. Can we get AIs to have? I mean, I think that's, it's sort of an overlay that you can put there. And does it really have emotion? Like, does its heart rate increase? Obviously not. It doesn't have sort of all those features of human biology. And to say, does it really have all those sort of features that go along with emotions? Well, no, it wouldn't have that unless it was a full biological organism, just like a person. But that's inevitable. You're making something which is an idealization. It's kind of like in science, people are often making models of things. And one person will say, I've got this great model of how, I don't know, snowflakes grow or something. And another person will say, no, your model is really crummy. It doesn't reproduce this particular feature of Snowflakes. Well, in a sense, at some level, the only way you're going to reproduce every feature of snowflakes is to sort of be a whole snowflake, so to speak. So I think it's the same thing with this. You know, can you reproduce everything about what humans do? Let's see. Sophie asks, will AI make interdisciplinary learning and collaboration easier by facilitating that process, or will it create more misunderstanding between fields? I kind of think one of the things that's sort of an interesting stretch opportunity for LLMs is that they will be able to notice analogies between fields in a way that people like me are kind of proud of being able to know enough about different fields that one can kind of see this analogy from here to there. One kind of imagines that LLMs and so on, AIs Neural Net AIs will have the possibility of kind of having seen the kind of flow of how people talk about one field, the flow of how people talk about another field, and being able to make sort of a correspondence between those things. And that will seem like a really brilliant kind of human connection in terms of humans doing different kinds of fields. I don't know. It's a good question. I mean, whether, just as an LLM can readily translate pretty well from English to French or something, can an alarm translate pretty well between a category theorist and a person doing, I don't know, quantum computing or something like this? Possibly. Haven't tried it. It's a good idea. I mean, it's something where you could say, you need something where there's sort of a common theme. I don't know. The thing is that many different sort of descriptive languages describe different kinds of things. I mean, when it comes to, you know, sometimes there's sort of the physics way to describe something and the math way to describe the same thing. It's an interesting idea to have an LLM be the translator between those things. I mean, it might lead to all sorts of horror and people misunderstanding each other. You know, it's worth remembering the piece of history that machine translation was really launched as a serious endeavor in the early 1960s. And the original way it was launched was. Was people were worrying about sort of communication between east and west, you know, during the Cold War. And kind of the. The usual scheme would be, oh, you'd have an interpreter, you know, translating between, you know, the. The premier and the president and the whatever else. And. But, you know, it was kind of like, but what if the interpreter is kind of. Is corrupt? What if the interpreter kind of, you Know, throws the whole thing, let's use a machine to do this instead was the idea. And then the notion was AI looked promising at the time. And it's like, let's just make a machine that does this translation from one language to another. And, of course, that turned out to be more difficult than expected. And 50 years went by. And then finally we have something that does a pretty good job of translation between common languages. At least. Now. What does it mean to do translation? That's another interesting question. Is what you get when you do the translation sort of the same as what you started with? Are things translatable between human languages? People often say when they know some particular culture. Associated with some human language. They say that's just not translatable into some other language. And I'm sure that's true. Because in the end, what you're trying to do in human language. Is take sort of thoughts that are going on. Package them up in a transportable form. And move them to another mind where they're unpacked. And if the mind that they're being packaged up in. Has very different experiences. Than the mind that you're trying to communicate them to. It has sort of a very different cultural environment. Then it could very well be the case that there's just sort of no way to have that thought be unpacked. And sort of a similar thought in the other mind. I mean, it's kind of like one of the things that's sort of interesting about the idea of motion. You know, motion. We take for granted that you can take a thing and move it from one place to another. But that's something that, for example, in our models of fundamental physics. Is something you have to derive that motion is possible. You can take something and have it be somehow the same thing. And you ought to realize that it will typically be made of different atoms of space. But yet we identify it as being the same thing. And it's the same type of thing. When we're talking about language translation, we're saying that thought that you had over here. That you packaged up as kind of a concept particle or something. We moved it over here, we unpacked it. To what extent is it the same thing when we unpacked it? And I think it's something where you have to kind of. It's a question of the observer whether it seems like the same thing when we unpack it. It's a question. Did the environment in which it was packaged up and unpacked. Is it the same enough? So I think this question about whether if you translate from math to physics or something, some concept about gauge theory in physics, translating it to something about fiber bundles and mathematics or something, is it really the same thing or is it something where there is a mismatch in the translation, so to speak? It's an interesting question to what extent there's. It's possible to make those kinds of translations. Can LLMs do it? It would be an interesting experiment, but again, you have to find these places where things overlap. I mean, there are certainly commonalities in what happens in economics and what happens in evolutionary biology, let's say. But is there enough kind of overlap, cultural overlap, in a sense, enough of, oh, the evolutionary biologists sort of live in the same world as what was the other thing I just said? But anyway, the economists, to be able to sort of have a common discussion about them. And I think that's the same thing with human language. If you've got two people who kind of live in the same environment and you're trying to translate between languages, that's much more promising than translating between kind of the languages of some indigenous tribe in the Amazon jungle that has a completely different experience of life than the people living in the big city and the expression of kind of the forest spirits or something versus the expression of the subway tokens. That translation will not necessarily be easy. Let's see. Yao asks, when people discuss whether an LLM is sentient or not, the question that always comes up is whether it understands the prompts in its replies. And there's the Chinese room thought experiment. No, my favorite. That often gets brought up. So it says, I see two ways to look at this. One is that an LLM is just an advanced predictive text generator, and that sentience is something more than that. Another is that we sentient beings are actually just advanced predictive text action generators. What do you feel sentience really is? You know, this is always such a slippery thing because you come up with a word and everybody says, we're going to hang so much on this word. We're going to, you know, like, like for instance, people say, you know, oh, should AIs have rights? Should this and that and the other? Oh, they should have rights if they're sentient. You know, can we. Do they. Do they feel pain? Well, you know, if they, if they're sentient, then they feel pain. We've got this word that's kind of this word. Are we talking about words being a packaging of thoughts? This word sentient is kind of a packaging of some whole cluster of kind of ideas that we have, which we're now concentrating down onto this one word. And you know, what we can say about how AI's work is probably a whole cluster of things. Can we anchor it to this one word, you know, sentient or not? That's much less clear. It's not obvious. I mean, you know, we can ask, is the world good or bad? You know, say we can define that. You know, it's a. But we're not going to be able to answer that question. It's not something we can concentrate down onto one word. So I don't think we can in any sense concentrate down this question onto the sentient or not type question. But there's a question of what do we mean by this? What could we possibly mean by this? We all have this feeling that I think I know. I have the feeling that I'm sentient. I have the feeling that I have some kind of inner experience and so on. Now, do you all have inner experience? Well, I don't really know. All I know is my own inner experience. But I kind of figure you guys look enough like me and you do things that are similar enough to me. And I kind of know your history enough to know you're kind of like me. And so I can reasonably extrapolate from my own inner experience to your inner experience and say, yeah, if I'm sentient in the sense that I feel that I'm experiencing things, then probably so are you. Well, now I look at this AI that's in the cloud, in a box, whatever else, and I say, is that like me? Well, how do I tell, you know, it's not. It doesn't have arms and legs and talk. Talk in the same way and so on. You know, is it. Is it in fact kind of, you know, should I. Should I transport my notion of sentience to it? And I think that's a, you know, that that's a thing where you kind of edge into it as. As it has more attributes like us. It feels more as if you can transport your internal feeling of sentience to saying it's sentient too. But you can always say, well, look, it doesn't have arms and legs and so on, and so it doesn't have the same attributes as me. As I mentioned, next week I plan to have a little discussion like this with a humanoid robot. That it will be interesting to see the extent to which one sort of feels that the humanoid robot might. Makes it. The humanoid aspect of it makes it feel more sentient than if it's Just an AI in a box, so to speak. But I kind of feel that this idea of, does an AI have an inner experience? Do any of you have inner experiences? Does a cat have an inner experience? I think the notion of inner experience isn't as sacrosanct as you might imagine. I think that in many senses it's perfectly reasonable to say that the AI has an inner experience. I mean, just imagine you're thinking from inside the computer. What does the computer. The computer has certain experiences, the computer has certain memories. The computer has a lifetime. The computer, it's booted up, it does its thing, it remembers its life, then it crashes or it shut down or whatever, and that's the end of its life. And then it restarts again with the same sort of genetic material and has another lifespan. And I kind of think that from inside the computer, we can easily extrapolate to what it feels like to be inside the computer. And I don't think it's that different from what it feels like to be us, I think. And I think, you know, what would it. There's this question of what is sentient. I mean, in many traditions, one imagines that much of nature is sentient. But you know, too often in sort of typical Western scientific traditions, that's not what one does. One says there's sentience in humans, but we can't transport that as far as, you know, the forest spirits or whatever. And I think from a sort of scientific point of view, it's not obvious where that distinction is. I mean, to me, in terms of thinking about the ruliad, there's this kind of distance in rulial space, and the distance from us to the cats and dogs, us to the AIs, us to the forest spirits and so on, these are different distances. And I think it's not so much a question of sort of is the forest sentient, so to speak. It's more a question of sort of how far away from our internal feeling of sentience is that. So that's a few thoughts on that. Ra asks, is it possible for AI to achieve true randomness? Well, depends what you mean by true randomness. I mean, if you mean that randomness is typically, we can't predict, we don't know what's going to happen. Well, that's often the story of computational irreducibility. We know the rules, but to work out what will happen, we just have to follow all the steps and see what happens. So we can't predict it. So we just say it seems random to us. That Kind of randomness, very easy for an AI to achieve. Just, you know, it does computation, it achieves randomness in that sense. Now, sometimes by true randomness, people want sort of a higher level. They want it to be the case that they. This is not a good sign that my computer thinks it is about to run out of power. That is a very bad sign. Maybe it's a sign that we should wrap up soon, but I think it is. Here's now me following wires to see what's going on that it thinks it isn't plugged in. So. Somebody very helpfully here pressed the power button to actually make it be plugged in. Good stuff. Okay, we're not, we're not out of juice yet. Well, the, so this question of, of true randomness and what, you know, what's kind of the. What do we mean by that sometimes? Well, it's not clear that our universe has true randomness in it, in a sense. If our universe has a definite rule running from a definite initial state, which in our models of physics it does, and in our kind of idea of the ruliad, it does, then there isn't any true randomness in the world. People imagine that, oh, there's a program that's been defined from the outside, and that program that's been defined from the outside is fundamentally unknowable to us. If there's something fundamentally unknowable to us, then we can describe it as true randomness. And that's a bar far beyond what one will achieve with an AI. As a practical matter, if you ask a typical LLM, pick a number between 1 and 10, it's going to say 7. It's going to follow the cultural traditions of humans that it's read about on the web and so on. It's a very poor generator of randomness as a practical matter. But if you connect it to our computational language or something, then it's not going to have any problem sort of calling random integer and generating a random integer. Let's see. Multicell asks, why is there no latency when we look around and construct a scene on the fly? Or is it our perception which makes it seem like there's no latency? It's a good question, actually. I mean, when we look around, first thing is that, you know, if, if our brains work much faster than they do, we'd notice all the individual photons arriving. But we don't. Photons arrive from, you know, if we're doing astronomy, then the photons might take, you know, millions of years to arrive. But if we're doing everyday life, the photons typically take, you know, microseconds to arrive. It's not like. And so that whereas our brains take milliseconds to kind of see what we've seen, now there's the question of why, when we kind of turn our head around, why, how long it takes for us to have the impression of a scene form? It's an interesting question. I think that the thing to realize is, in our eyes, there are, you know, we get to low light levels, there are individual photons hitting individual photoreceptors on our retinas. And yet we. We always have the perception that we're seeing a continuous image, even though in actuality, the. Its individual photoreceptors that are being excited. Now, I think that's unsurprising because I think it's kind of like in an artificial neural net. We are. It is by its nature, kind of interpolating its inputs. And so I think similarly, we are kind of, you know, interpolating those individual points of light, so to speak, to synthesize kind of an overall structure that is the image that we perceive. So I think that's. But now the question is, when we turn our head and we are now sort of reconstituting the image that we've seen, does that happen in a time? You know, could we turn our heads so fast that the image has not been reconstituted in our brains? You know, I kind of suspect that in VR, some of those kinds of things are happening. I mean, in other words, we. We turn our head. The accelerometers, notice that we turned our head is the image being repainted in the VR system. And that's kind of a simulation of what it feels like to not have the image be properly formed as we turn our head. And by the way, the main consequence of that is you get terrible motion sickness. It's sickening to not have the image be formed quickly. Now, I mean, as a question of how much time it takes to. To sort of form the thought of what you are seeing. I suspect that, you know, it's like when people do these experiments and you've got a video, and somewhere in the middle of the video of the football game or something, somebody walks out in the middle in a monkey suit or something that. And, you know, people don't notice if that happens quickly and they're paying attention to other things. My guess is that there is a thought formation time, a kind of scene formation time. I would guess it's on the order of hundreds of milliseconds, I don't know. And it could be that our brains have a lot more in them than the brains of lower animals, and that those things form the image much more, form the scene much more quickly, and are much more quickly able to say, there's a thing when you try and swat the mosquito and it jumps out of the way. And it has a much simpler neural pathway to figure out, oh, you're coming after me, and we don't manage to get it. Okay, a few more questions maybe. Let's see. Derwin asks, what new types of auxiliary jobs do you think will be necessary for the ubiquitous integration of AI into society to properly balance it with human interests, such as the alignment problem? What role, if any, do you see Wolfram Research playing in that AI economy? Well, to answer the second question first, I think our Wolfram Institute is probably likely to play a more direct role in kind of some of these questions about AI alignment. Our company, Wolfram Research, we make tools, computational language and so on, which makes it possible to do things like define computational contracts for AIs and so on. But it's not the decision of which computational contract to have. And how that sort of integrates with society is kind of a different type of question. But now, how does that work? I mean, what will be the new jobs? I think we had the prompt engineer job. That's kind of a little bit gone. Gone by the wayside, I think, as the kind of, as the prompt engineers. You know, for a while. It's kind of a funny thing that people said back five, 10 years ago, people were saying, you know, the machine learning is coming. Machine learning is going to take all jobs. But the thing that always amused me at that time was I was thinking to myself, the very first job to be automated is going to be the machine learning engineers. Because all this sort of tweaking of machine learning systems, this is something you can machine learn how to do. And so I think when everybody said, well, everybody's going to become a prompt engineer, in some sense, the reason that hasn't happened is because one could prompt engineer the prompt engineering, so to speak. One can train things to not need those levels of human input. But the question of what the AIs will do is how do we direct the AIs to do the things we want them to do? That's kind of like AI management or something. And yes, that's a fundamentally human role, because left to their own devices, the AIs will just. In my vision of things, the AIs just go off and explore the computational universe for the rest of eternity and have a great time doing it in some sense of us being able to project what having a great time is. But it's exploring the computational universe of all possible programs and so on, exploring the ruliad and most of what's happening there has not been colonized by humans. Most of it is just they're out in sort of outer rullial space, so to speak, dealing with things which we humans have never really had a connection to. So I think there's this question of how do we manage the AIs to go in directions that we humans want to go in. And that's something that necessarily involves humans and involves the interaction with other humans and so on. I would say that there'll be that's AI interpreters, AI scientists. Just like we have the natural world operating according to the rules it operates according to, so we have AIs operating according to their rules. There's the question of can we make a bridge between the way the AIs operate and our understanding of things? We'll need AI scientists to do that. We'll need to understand, insofar as we can understand, something about the weather or something like that, and we need scientists to do that. So similarly, we need to understand at least something, even if we can't understand every detail of what the AIs are doing. So that's kind of a replication of that. I suppose that many of the things that we do to sort of interact with the natural world will kind of have to replicate for the AIs, because the AIs are another piece of the natural world, not natural in the sense of the nature as we've experienced it so far, but part of the sort of natural computational universe. And we're kind of chopping off pieces of that and figuring out how to kind of fashion those for human purposes. And it's just like we start with nature and we have to sort of build our civilization on top of that. And I think sort of all the layers that are involved in doing that are things that we can expect to kind of have to replicate in the AI case, as we had to build them in the case of sort of coexisting with nature. I mean, that would be my first thought about that kind of question. I think that as a practical matter, it's kind of. Well, I think that using kind of the computational language, computational thinking, that's a whole other set of sort of more structured occupations. Then there's the kind of, okay, how do we deal with the sort of neural net type AIs? And that's much more of this kind of wrangling and interfacing with them as we interface with things in nature. Let's see. Quimby asks, do you see there being more specialized computing hardware in the future where the computations are more directly embedded in physical processes, rather than needing to construct a given computation within a universal computer? That's kind of asking for analog computers, computers which can be the analog of some system you're trying to study. And, you know, that was the early history of computing was let's make a thing which is kind of a model and analog of what we want to study. I think it's a big question of what do you want to study? I could imagine that just visiting a quantum computing company and talking about their use cases and saying, well, we're going to use it. People want to use it as a simulator. What do they want to simulate? Maybe the number one thing they want to simulate is, is a quantum computer, just like this quantum computer. And so in a sense, that's an inevitable feature that you make a thing which is like the thing you're trying to simulate. Maybe it has certain simpler features. So, for example, let's say you're trying to simulate electric circuits. You can make an analog computer that is made with electric circuits. But the analog computer may be easier to tweak than the actual circuit itself. You may be able to have a variable resistor you can change, whereas in the actual circuit you're trying to simulate, it's a fixed resistor. But I think this question of sort of how close do you get to the thing that you're trying to simulate? Maybe if you can have a parameterized version, a tweakable version of the thing you're trying to simulate, maybe that's an interesting thing. What I haven't seen, and one could imagine it, is kind of, let's say you're trying to do protein folding. I haven't seen a case where somebody says, oh, okay, we've got this particular kind of instrumented protein that has special things on it that make NMR work better or something like this. And let's just look at the analog, you know, the folding of this kind of analog protein, which we can tweak in this way and that way, rather than looking at the folding of our actual protein. You know, I think there will be things that one can imagine. Things like that. I'm a bit skeptical about. Kind of the question is how much can you piggyback on existing physics and have it be general, Having physics be something where the physics of your machine, of your computer in quotes, is just the same as the physics of the thing you're trying to compute. Okay, that's easy to understand. That will work. The question is, the remarkable thing about universal computation is that it seems like we can cover everything with this sort of fixed set of underlying primitives. And that's not something that seems to be the case even if we do real number computation, computation that doesn't just involve discrete elements, but involves arbitrary continuous numbers. Even in that case, we don't know how to have kind of a universal set of primitives that can operate in that environment. So this question of just how far away from the detailed physics of your quotes computer can you get in what you simulate? That's really the question. And I guess I don't know of examples where you can get very far away from what you're simulating. I mean, sometimes in old analog computers, it was electric circuits, they have certain kinds of linear differential equations. You can get to things which have those kinds of linear differential equations. Or in Kelvin's old Tide computer, you have these disks that are going around, and that's kind of like the periodic, the periodicities of the Earth, moon, sun system with respect to the tides. But so I think that's the issue. The issue is we don't have another kind of pocket of simulatability, so to speak, a pocket of universality around anything other than universal computers. It's not saying that can't exist, but that's not something we know about right now. If we did know about that, then maybe there will be other things for which there are sort of, you know, sort of analog, universal within some domain kinds of computers. Let's see. Question, The question from Teo here, how do you envision hypergraph based models advancing our understanding of quantum mechanics, general relativity, their potential unifications? Specifically, how might these models address challenges like quantum gravity, nature of space time, the emergence of fundamental particles? You know, I've talked about this so much over the last few years. I think it's probably not. I shouldn't try to compress a recapitulation of that. But our physics project is all about kind of how do you go from what's underneath space time to kind of all the things that we observe? I have to say, in the last two days I made a little bit of progress in understanding that and understanding how things like particles can emerge and kind of how we can have this Kind of network of causality of different events and how we can sort of imagine fishing kind of the things like particles out of that. The main challenge I see right now is not knowing what the machine code is. It's writing programs that are the ones that are interesting for the kinds of things that we do experiments on. In other words, I think we know. I'm rather confident that we know what the truly underlying machine code. At the love of the RULIAD and Eames and all these kinds of computational operations, we know what the underlying machine code is. Now the question is, can we write a program, in effect, in that machine code, that represents the collision of an electron and a photon or something? And that's hard, but that's what we're trying to do. And it seems like we can get to some sort of general features of physics, like general relativity or the general features of quantum mechanics or features of statistical mechanics. Getting to the specifics of, oh, it's an electron. It's just like this. That's much more challenging. I think a lot of the things about, for example, quantum gravity and what happens in our models is that gravity, gravitational phenomena and phenomena about the structure of space time and phenomena about the structure of quantum mechanics, they are actually, in a sense, the same thing. They're different. They're just projections. They're reflections of the same underlying object. We kind of see how that works. And the interface between those two is kind of the land of quantum gravity. And I think at this point, I'm, you know, I no longer in any way see what has been the traditional issue of sort of, how do you bring general relativity and quantum mechanics together? I will say that there are lots of little challenges. I mean, for example, there's a notion of the speed of light, which is kind of the maximum rate. As you go a certain amount of time, how far away can you get in space? The maximum distance away you can get is the speed of light times the time you went. Well, there's also the notion of. In quantum mechanics, you have these many different paths of history. You have these many. This, what we call branchial space that covers many different sort of threads of history. And the question is, sort of is there that there's an analog of the speed of light in branchial space? It's the maximum speed. We've called it the maximum entanglement speed. I'm not sure that's the best name for it, but it's an okay name. It's the maximum rate at which things can be transmitted in branchial space. And there's a question of what is the significance of that. And I think just this morning I understood something about how that might work. And it's very weird. And it's all about. Actually it comes back to something which is kind of at the core of particle physics. These things called Feynman diagrams, which invented by my sometime friend Dick Feynman as a way of representing kind of the kind of the behavior of elementary particles, things like electrons and photons and quarks and gluons, all those kinds of things. And in a sense that what Feynman diagrams do is they say there are these events, there are these interactions between particles, and then the Feynman diagram has a bunch of lines that represent the sort of space time history of those particles in between those events. So in a sense, a Feynman diagram is like a thing that gives you the causal connections between events. The events are the interactions between particles and the particles drawn as lines. In a Feynman diagram, if it's a photon, it's a wiggly line and so on, drawn as lines, those represent the kind of the particles or the carriers of causality between events. And so what you have is something where you've got a particular sort of history corresponds to kind of these different, kind of these different sort of elementary pieces of causality, which are these particles. Now, in fact, in our models there's something much below that. These particles are really giant bundles of causality. In our models, there are sort of very elementary causal edges that are sort of far below the structure of whole particles and so on. And these kind of elementary causal edges are not only what give you the connection between sort of events that are like interactions between particles, they're also what kind of knit together the structure of space time. And so the question then is in that sort of setting where the speed of light is. If you are saying these different interaction events are sort of arranged so that we can think of them as being different places in space, what are the things that are arranged in being different places in branchial space, different possible threads of history. So I've been trying to understand what does the so called light cone, the maximum distance that you can go in a given time, defined by the speed of light, what does that in physical space correspond to? The entanglement cone, so to speak, in branchial space. And I think the beginning of that is a story of the following thing that in Feynman diagrams, these things that represent interactions between particles, any given interaction, any given possible. Oh, we've got an electron interacting With a photon. Well, there are many different things that can happen. There are all sorts of internal little diagrams. You have a whole collection of graphs that represent the different things that can happen in the interaction between an electron and a photon. And as you try and work it out very precisely and you get more and more different pieces to those graphs and so on, you have a whole collection of graphs and each of those graphs is in a sense a different thread of history. We can think of it that way in branchial space that it's a different thread of history. So in a sense there's a maximum rate at which you can sort of expand with respect to this sort of different threads of history. And that's defined by sort of the maximum rate at which you can generate new diagrams. So in some sense I think that this sort of limit has to do with, it's sort of a graph theoretical thing about how many different ways are there to make these graphs. And if you're kind of, if you're, if you're asking for sort of too many ways to make these graphs. Well, okay, in the end this is rather technical and maybe I shouldn't, shouldn't talk about it. And I also. It's something I only thought about, sort of think I somewhat figured out this morning. So it's not, well, annealed yet. But I think that the coupling constants, the interaction probabilities play a role that is similar to the speed of light role in physical space. That's my current guess anyway, that got very technical. Let's see, let's see. Well, I think maybe a last question or so and then I'm being told for, for folks who are in the room here that, that because many of them missed dinner, there's going to be pizza delivered. So but that's not relevant to people in the outside world. But let's see, Jake is commenting. Are we programmed by evolution to be sentient? If so, can't we program a machine to be sentient? Well, you know, I talked about the slipperiness of the concept of sentience, but I have to say I've sort of changed my view about evolution recently. Biological evolution. I kind of made this sort of very minimal model of biological evolution and the thing that. Well, first of all, I've been surprised by how far it can go just sort of tweaking the rules for something. It's just there's the question is what happens when you change the rules? What happens? Well, there's this whole computational irreducibility phenomenon that says just by knowing the rules, you can't immediately deduce what will happen. But then when you say, well, I want some definite thing to happen, I'm trying to evolve to something which has a certain degree of fitness. The thing that's interesting and surprising is that the way that that is achieved often looks completely coincidental. It's not that you evolve it, as an engineer might progressively incrementally develop something so that it has this lever and this wire and so on, so gradually get towards the thing you want. It's rather. It looks like it has the structure and it's kind of random, and then suddenly, boom, it does the right thing. It's this kind of weird phenomenon. That sort of evolution is something that, because it's a fundamentally computational. It's a story of computational irreducibility that it is possible that it sort of can achieve so much is a story of kind of computation, universality and so on, that you can get these systems to do all these very different kinds of things, but you don't do it in a way that looks engineered or that has kind of a narrative we can understand. You're kind of. It's always something that happens with computational irreducibility, and it just sort of happens to work, rather than it being something where we can give a narrative about why it works. And by the way, I think that machine learning is very much that story that you train a neural net and it happens to work. And if you say, show me the mechanism by which it works, it's not the same thing. That will be a different problem to try and train something or try and create something with mechanism. In a sense, that's what we're doing in building computational language. We're spending all this effort to sort of engineer something which is kind of for which there is an understandable human narrative and which is also able to tap into sort of the power of computation. So that's probably. Oh, boy. Okay, there's one question I'm sort of tempted to address here. Oh, a couple more, and then we'll. Until the. Until the pizza arrives for the locals here. Captain asks, do you think hydrogen has a future in computing? Will it play a major role in energy and possible propulsion to get us to Mars? Well, gosh, I don't know about hydrogen in computing. I mean, I think, you know, the fact that computers guzzle power right now is just a technological detail, in a sense. And the fact that even, you know, people imagine this. You know, the cloud I was mentioning earlier, the cloud might be in the clouds type. Thing. Thing you're using solar power, you're cooling things by virtue of them being, you know, in space or in the upper atmosphere or something like that. I don't think that the energy part of the computing story is really a continuing issue. I think in terms of how do we get to Mars. I think, you know, there are basically we're, we're, we're leveraging chemical reactions of different kinds. Whether hydrogen's the right one or not, or whether we should be using more, you know, electrical, you know, ion type things or whether we should be using, I, I'm not sure. Certainly, you know, chemical rockets. Chemical reactions have managed to get rockets that happily, you know, send unmanned probes to Mars. So one sort of assumes that that will, that will continue to work. Let's see just one more. Interesting. Okay, Sam asks, is the Ruliad a meta theory or does it actually exist? Boy, this was supposed to be about science and technology. That's pretty philosophical. But let me see if I can say anything useful about that. What does it mean to actually exist in the sense the does when you define a formal thing like a number? Does a number actually exist? The thing that the Ruliad does is it conflates things which, which are like numbers with things that are like the reality that we perceive. And I think the thing that is very, a little bit confusing about the story of the Ruliad is the Ruliad is a thing more like a number. It's a, it's an abstract thing that sort of just is. There isn't a, there isn't really a sense of do you have that number or do you not have that number? Two just is a thing that. But now the question of whether how we relate that abstract thing to the real thing that, you know, we can pick up and feel and experience and so on, that feels like reality to us. How those connections get made, that's the non trivial thing. And the thing that is confusing and kind of hard to wrap one's head around is the fact that in the story of the Ruliad we are part of the Ruliad. So it is a question of. It's not like we're saying, can we pick up the Ruliad from the outside and say it exists? We can pick it up. No, we are inside it and so we are part of it. And so the question is, does our perception of reality, how does our perception of reality arise from sort of our experience of the Ruliad? Because it's just our perception of reality. There isn't some real Sort of gold standard real reality. It is merely at any point, it's merely sort of our perception of reality that we get to experience. And so I think the thing that in a sense is the existence of the Ruliad, I think is as inevitable as the existence of the number two. What is not so inevitable is that we exist within the ruliad. We are more surprising. It's not obvious that there would be entities within the ruliad that would be observers like us. But it's something you can scientifically investigate. You can ask things like, you know, in principle, you can ask what is the density of entities? What is the density of structures in the RD that are kind of as. That behave in ways that are like us? That's a science question. That's a math question. It's not even a science question. It's really a math question. It's an abstract question. Given a definition of what observers like us are like, what is their density in the ruliad? In a sense, that's the ultimate, you know, if you're interested in, in SETI or something, and finding the extraterrestrial intelligences, all that kind of thing, that's the ultimate thing. It's like, what is the density of intelligences in the ruliad? What is the density of observers like us in the Ruliad? I don't have any idea. I don't know whether it is at all surprising that we are the only intelligences that we know about like us. I think it probably isn't surprising. I think there's so many. There's several sort of powers of exponentials of what, you know, of our particular pocket, just like us and the Ruliad, that it's kind of not surprising that in the universe as we have so far perceived it, and remember, we perceived only a tiny, tiny fraction of the ruliad. We perceived a fair fraction of what we imagine as physical space, but we've perceived only a tiny part of either branchial space or rulial space. And sort of out there in rulial space, you know, there are all kinds of, you know, one can anticipate that there are all kinds of civilizations having a great time out there in rulial space, but they are not, you know, they're far away from us in rulial space. We are, our paradigms have not reached that point where we can sort of perceive those kinds of things. The reason that the ruliad is not a pure meta theory, the reason that it's not just a theory about all possible theories is because it is anchored by the fact that we imagine that we are observers of the kind we are. So in other words, if it was just a matter of the abstractness of, you know, twos and pluses and so on, then yes, it would be a theory of all possible theories. But what anchors it is the fact that we, we have a notion of what observers like us are like. And that's kind of the anchoring point that allows us to make definite statements about the ruliad as we perceive it. We can also discuss the pure ruliad. There's a whole kind of science to be built about the whole ruliad. And there are many things that are sort of abstract pieces of computer science about, you know, P versus MP and so on that probably can be answered by thinking about the ruliad. But when it comes to physical reality, it's all sort of anchored in the fact that we imagine that we know kind of what we are like as observers. Okay. Nuno asks and I'm just if the ruliad is correct, not sure quite what correctness means given that it's an inevitable object. You know, it's, it's what kind of technology do you think it will bring us? It will bring us everything, every possible technology. The only question is what technology will we care about? In other words, it is if we just start thinking about all possible programs. They all do something. The question is do they do us something that is relevant to us? So it's much more a reflection of us, how broad minded in a sense are we, than it is a reflection of sort of what's out there in the ruliad or in sort of the computational universe. So I think in terms of our understanding, the fact that there is this infinite sort of rulial space to explore is already very invigorating in a sense in terms of thinking about technology, what is the right direction to go? It's kind of like if you're exploring the surface of the earth, where's worth going? Is it you discover the lutetium mine? Do you say that's the most exciting thing ever or do you say what the heck am I going to use lutetium for? Or do you have to wait a few hundred years until it turns out it has some electro optical property that's incredibly important. All right, I think we should wrap up here. I'm apparently folks on YouTube are saying that should let the people here eat pizza. So. So let's go do that. Thanks for joining me. See you another time. I think I may do another one of these on Friday about different topic. And then planning probably next Monday to do the interview with a humanoid robot. See you then, Sam.