Business, Innovation, and Managing Life (July 17, 2024)

Hello everyone. Welcome to another episode of Q and A about business, innovation and managing life. I am back in my usual habitat after my three weeks in my spending time on my annual extreme professoring adventures involved with our Wolfram summer school and our Wolfram summer research program for high school students. Each about 60 something students. A lot of fun. I learned a lot. Hopefully students learned something too. It seems like they had fun. And the projects, by the way, from the summer school and summer research program, we do the very gutsy thing of posting everybody's project online as soon as it's finished at the end of our programs. And you can find those online. And they're pretty interesting collection, I would say almost all projects that I kind of made up, some that I'd hoped somebody would do and I'd been kind of saving up as things to do through the course of the year, some of which I made up sort of on the spot after meeting students and finding out what they were interested in. Any case, that's the. Oh, and the other thing I should say is that I met a couple of people at our summer school who have been long time participants in these live streams. That was really interesting to see who's behind the handles. Great. Let's see. All right. Another thing, last week one of the mentors at our summer school had brought the humanoid robot that they were working on. And so I did a live stream discussion with the humanoid robot. That was an interesting experience. Only after did I learn that I seem to be the human who has spent the longest time talking to a humanoid robot than anybody else. It was like an hour and a half. And the humanoid robot has fairly realistic eye motion and head motion and so on. And I was like sitting and talking to the robot for an hour and a half. And when I finished, I was very disoriented. It felt really weird. And then Vlad, the person who had built the robot, said, oh, by the way, I think I should fix the fact that it doesn't ever look away when you're talking to it. Humans, when you're talking to them, they'll mostly look straight at you, but occasionally they'll look away. But the robot didn't do that because of kind of a bug. And the result, I can tell you, at least for me, was a quite disorienting experience where it took me about 15 minutes before I felt able to kind of think straight and talk to actual humans and so on. Any case, an interesting experience and quite a particularly interesting, I would say, for the fact that the robot was using an LLM to have the discussion with me. And that LLM was running on a computer that was right there with the robot. It was a local LLM. And that's. So it's kind of impressive that a local LLM can have an hour and a half discussion with me. Maybe that's just me being prepared to do that. But it seemed like a reasonable hour and a half discussion just running off a local LLM. It's kind of a sign of the future, I think, of the power of local LLMs and also a feature of kind of the way that to think about LLMs, which is they are good as linguistic interfaces. They shouldn't be the place that you put kind of your reliance for all computation, so to speak. Maybe we'll talk about that a bit more. But let's see. Johnny asks here, based on my discussion with the robot, how do you imagine a future of robots in business? Robot coworkers, bosses, assistants, et cetera? Will robots be able to effectively communicate with their human companions and vice versa? I think there are definite use cases. I think that the kind of, there are various kinds of companion use cases that I can imagine, particularly if that robot had access to my complete archives, email and all the other kinds of things that I keep. And, and there are plenty of things that people keep on social media and so on. If it had had access to all those things, I would probably be able to have a pretty interesting discussion with it about my life and times. Particularly relevant, I would suspect, for people with senior folk and so on, where it's kind of really nice to have somebody to talk to. But it's not clear that one needs kind of the incisive think about the future type of type discussion. I don't know with respect to, for example, kids, I, I'm a little bit less, less optimistic there. I think there's been, it's, I think it's a little strange and it's kind of like, you know, just how much does one want to sort of dehumanize the gen, you know, successive generations? I, I kind of worry that it's just like the phenomenon that people say, well, my essay is going to be graded by an AI, so let me figure out how to write the essay so that I'll please the AI, so to speak. That's a path that, it's a curious path to go down. I suppose it's a generalization of things like let me write my website so that I'll improve my SEO. Forget any humans reading it. I just care about the SEO. So to speak. I think that's for the future of our species, so to speak. I think this kind of like I'm really doing this to please the AI or to interact better with the AI is a curious and dubious future, so to speak. So I suppose I'm a little less optimistic about, you know, robots for the kids type thing than I am for senior folk and so on. I think it's an interesting question whether like autism spectrum folk who have a hard time doing the kind of the eye contact with the robot, kind of the human, very human kind of interactions. Is the robot human enough that it triggers the I don't want to interact with the human, or is the robot exactly in the right place to be? Kind of like, I'm not quite human, so there's no kind of inhibition to interaction? I don't know. I think that experiment has not been tried yet with that particular robot. As far as other things are concerned, I think this question of sort of the humanoid form factor is an interesting one because the real reason why one might think a humanoid form factor for a robot was useful, one reason is that the environment that we have built for ourselves is built for us with legs and arms and things like this. And so if you want a robot to be able to kind of walk through your environment, pick things up that you would pick up and so on, then a humanoid robot is a good form factor for that. Now, you know, I have to say I've run into a bunch of people recently who are doing humanoid robot companies and things who say things like, well, we're going to have the robots, like tidy up rooms and so on. That strikes me as a really hard problem and not obviously one for which robots, until they are super cheap, will make a lot of sense. I think there are plenty of places if we look at the use of robots in the world at large. Well, there's manufacturing, but most of the time you don't really need or want a humanoid robot for that. You want something that has the, the right arm and eyes and sensors and so on to do the task you want. A big issue in that domain has been do you have a robot that is sort of built just for this particular place on this particular production line, or do you try and have a more general robot that is quickly and easily reprogrammable for different tasks? That industry has tended to be quite conservative and have robots that are like there for 30 years or more. But I suspect the future is more like it is with computers, where it's more general purpose and it's more like, well, we've got this robot. Well, we can reprogram it and reconfigure it to change what we're running on this production line as soon as different orders come in or we change our products or whatever. My guess is that's the future, although I think it's been slow to come. And I know quite a few companies and people who've tried hard to make that work and been disappointed with the extent to which folks who are actually running manufacturing lines say, look, we just want to have something that's reliable and will run this line for a long time. We don't really want this, reconfigure it in a day type type thing. But I think the question of, well, what other kinds of uses? I mean, there's all kinds of military uses for things, but I'm not sure whether there's. The configuration of a humanoid robot is really the thing one wants there. I mean, whether it depends on what one's trying to do, and it depends on whether it seems like a soldier from a distance type thing is what's relevant or whether it's more like, can you carry this heavy object kind of thing. I think in a place where people have certainly talked about humanoid robots is things like education. People always say whenever there's a new technology, people always say, oh, and it'll be useful for education. Only sometimes does that turn out to be true. I would say the vast minority of the time does that turn out to be true. And I think the concept that you've got some educational setting where it's like, talk to the humanoid robot. I'm pretty skeptical about that. I think that the value of having humans in the loop as teachers and so on is mostly about kind of the emotional content of students actually wanting to do things that a human will respond to. And I doubt that that will be achieved with a robot any more so than it will be achieved with kind of an AI tutor just running on a computer and so on. And that will be a different thing that's achieved there. So I would say that in terms of there's kind of the mechanical aspects of robots, a big area relevant for that, I suppose, is the construction industry, one of the larger industries that still has a lot of people involved and people doing diverse things and so on. I suspect that one day that will become much more heavily robotic. I mean, it's already, over the course of time, more and more stuff is kind of manufactured to be installed, so to speak, rather than having to be kind of custom built on Site and so on. But I suspect that the custom install the thing on site and get into those awkward places and so on. My guess is that's probably a good use case for a sufficiently cheap, sufficiently effective humanoid like robot. Partly because the things you're building are in the end going to be things that typically are things that are lived in or worked in by humans. And perhaps that doesn't drive the full story of whether a humanoid robot should be able to sort of hammer those nails in and so on. But I think it certainly supports the idea that that's more likely to be a good form factor for it. I think in terms of the. Well, I'm. There are cases where I could imagine, you know, it's always hard to predict what will make sense to kind of be sort of human interface. And by human interface in this case I mean actually sort of, you know, size and shape of a human type thing and moving like a human versus better to just be the functionality and so on. I mean, imagine that you're doing something and you need tools handed to you. Is it better to have a humanoid robot handing those tools to you with a hand like thing grabbing them, or is it better to have some quite different thing that's sort of optimized just for the purpose of handing you the right tool, so to speak? I'm not sure. I mean, I remember years ago, sort of an interesting dynamic of seeing automation happen was buying movie tickets at a movie theater. You know, it used to be the case you would walk up to a person, you would buy the ticket, et cetera, et cetera, et cetera. And then these screens started appearing which were very different in their functionality from walk up to a person, ask for the ticket, hand them a credit card, things like this, they worked in a different way. And for a while, at least in suburban movie theaters, the automated screens, nobody will be. There would still be people talking to the person. And then there was this transition that happened over a period of about less than a year I think, where sort of the flight from the talk to the person way of getting a movie ticket to the just press something on the screen way of getting a movie ticket happened. But again, you wouldn't have really what happened on the screen was a bit different from what happened when you walked up to the person and asked for the movie ticket. And the fact that that became the fact that that one sort of refactored the task based on the available interface was something I think was hard to expect. And I think the same is true with a lot of These kind of robotic tasks that it's kind of like, will the task get refactored to need something that is a different shape than the humanoid top robot, or will it be more convenient to just have a humanoid in the loop? A thing that is an interesting analogy to this is the linguistic user interface that's provided by LLMs and so on. I mean, there's a question of if you're trying to do some particular operation, what is the right way of describing that operation? And in many cases, perhaps the right interface is just natural language, just something very human. Now, that only works up to a point. I mean, this is the whole story of our computational language and so on, that when you're just describing something where it's like it's a sentence or two to say what to do. Okay, natural language is a very convenient way to do that. When it's a kind of a big complicated description, it's a big tower that all has to fit together in a nice consistent way. Well, natural language isn't a very good way to do that. At best you end up with legalese or the kind of language that you'll read in patents and things like this that's very formalized and really very hard to understand. And you'd be better off with something that's much more genuinely formalized, like computational language, for example. So it's sort of a question of are there tasks which kind of like tasks where it's convenient to just use natural language to describe the task? Are there tasks where it's just convenient to use a kind of human form factor type thing to do the task versus to have something which is more custom for that particular task? I mean, for example, robotic vacuum cleaners, let's say that's something where one might have thought, oh, it would be best to have kind of a humanoid robot that could use a vacuum cleaner that's a sort of upright thing and so on. But actually no, it's much more convenient to have this thing that's just very close to the ground and not tall at all doing it. But maybe there are other tasks. Maybe it's a situation where if there's enough diversity of tasks that that operate within the kind of environment that humans have been used to operating in, then it makes sense to have sort of the full human like robot rather than something that is a custom kind of robot for that particular purpose. So I'm not sure there, I think in terms of the, I mean, this question of the embodiment of kind of intelligence or intelligent like behavior, the Question of whether there's sort of a different issue of whether one wants something with hands that can pick things up and so on, or whether there's just importance to the very fact that the thing is in a human shaped box and the thing that's talking to you is a human shaped box rather than a computer shaped box and so on. Is there importance to that? As I say, I think in some of these cases where it's desirable to have kind of where people are more comfortable with people, at least with some, but are prepared to accept some kind of attenuated version of people, then that's a good thing. I mean, I do think that in the immediate future there'll be an interesting kind of novelty item of the robot greeter and the robot kind of thing you talk to at a trade show and so on. And that's clearly going to be a fine business for at least a while now. What will happen when, what will the robot greeter at a store or something, what will that be like? Will that turn out to be something where it's just a novelty item, or will it turn out that there's high utility that can be generated there that we don't yet know about because we just haven't had the experience of having robot greeters that are sort of cheap enough to just deploy widely exist? Let's see. Well, Lori is asking what business ideas can you think of that useful AI applications? How can we make building your own AI for your purpose easy and affordable, such as having a bot that helps me find weekly coupons and savings for my grocery trips or mapping ideal travel times? You know, I think the, the thing to understand is this thing that's called AI. The term AI was invented in 1956. It has gone through various ups and downs, various assumptions about what it would mean. And then there was the shock of ChatGPT that it was something nobody really expected. And suddenly one got to this level of capability where an AI could produce kind of reasonable human like language. Well, you know what, how can you use quotes AI? Well, it depends a bit what you mean by AI. A lot of what people will, you know, at some level, AI sort of defined by the thing that we all got excited about with ChatGPT and LLMs and so on. At some level there's kind of a back end to that that's more about what one can generally do computationally. So for example, when it comes to invent a new thing, there's a lot that can be done by just sort of enumerating things computationally and then studying them computationally. Want to invent some new molecule? It's more a story of enumerating molecules and looking at their properties than it is saying, let's use an LLM and talk to it and have it use something like, kind of use its sort of human knowledge and human language capabilities to do this. I mean, there's, there's potential value to saying, go look at the literature that humans have written about this thing and go pick up relevant information. But the actual kind of, you know, know how to produce language like a large language model does is not so directly relevant. I some footnotes to that in the sort of surprise relevance of at least the underlying transformer architecture of LLMs to things like protein folding. But let's put that to the side for a moment. The thing that I think is the clear use case for sort of current LLMs is linguistic user interfaces to things. Being able to have sort of a, a way of us not having to crispen up what we want to ask, but instead just being able to ramble a bit and having the LLM deal with our ramblings and be able to turn that into something that can be crisply executed, so to speak. I view in our own efforts, I view our efforts to formalize everything in terms of computational language, Wolfram language and so on. That's kind of the gold standard for how you actually define what you want and how you actually build things where you can actually build sort of a tower of capabilities in the world and really expect computers and computation to help you the best way it can. When we built WolframAlpha 15 years ago now we set it up for the first time to be able to do sort of small scale but highly reliable natural language understanding. And that was useful to people. It is useful to people. You don't have to sort of read the documentation. You just type in the kind of crisply understood thing, whether it's a chemistry problem or whether it's some general kind of query, you can kind of enter it crisply and get the answer. What LLMs now add is the ability to kind of ramble at them and have them ramble back and have you be able to have a much more kind of traditional human interaction, so to speak. And there are plenty of cases where that kind of linguistic interface is very useful. I mean, for example, if you can do something like where you say, okay, let's search the documentation to find out how we do this thing, or let's say, let's talk to the linguistic interface and let's kind of ramble at it about what we want and it will then go and make use of hopefully behind the scenes, kind of what is actually in the documentation and bring back kind of the humanized version of what you might want to know and perhaps its guess about, let's say, a piece of code you might want to write or whatever else. It's worth saying, by the way, that in the actual technology of how that works, it's not a pure LLM. It's an LLM with a so called RAG retrieval Augmented Generation system, which is doing more traditional, although still sort of AI informed kind of lookups against an actual corpus of text and so on, typically using a vector database where what one's done is to break the text in little snippets and then say of each snippet what is the kind of meaning vector, what's associated with the snippet? And then say, when you ask about a thing like that, the meaning vector of the thing that you're asking about is compared to the meaning vectors of the things that were in the original document. And when there's a match it's like, hey, this might be interesting, that might be interesting. And if you look at some of the things we're building, which will come out soon, I hope to do with code assistants for Wolfram language, a lot of that has to do with surfacing things that might be relevant, then having the LLM kind of humanize its way of presenting those and synthesize them and put together something which is the useful thing to tell you at that point. But I think the thing to understand about sort of AI uses people sometimes say, well, I'm just going to have the AI run things. Okay, so long as running things is kind of a mushy type of thing, that's probably going to be okay. If there's a need for kind of precise essentially mathematical precision, then it's probably not a great idea to have this traditional machine learning system be doing that, because that's not what those systems do well. What those systems do well are the kinds of things that we humans do well which have a certain degree of a kind of mushiness to them. A certain degree of, well, if it's roughly this, then it should be roughly that. What the thing that is sort of the special feature of computers and computation is the idea that you can formalize things and then you can build these big towers on top of this formalization. It isn't something where you say a few words and maybe those were slightly misunderstood, maybe they were understood well enough to go to the next step in the conversation and so on. But you don't expect that those words will provide solid enough foundations that you can build a giant tower on top of the them. That's the thing that in the last few hundred years, through mathematics, through now computation, we actually have a way to do, to formalize things with enough precision that we can actually kind of build up and up and up. And that's a thing that isn't really a thing that we do directly in our minds. It's not a thing that current LLM style AIs do either. But it's something where you can provide, you can use the LLM, the AI to, to provide a linguistic user interface to those kinds of solid computational tools. And that's sort of a big thing that certainly we have kind of the unique technology to provide kind of the backing for that type of from the linguistic interface. What are you actually computing with? So in terms of, I think the specific question here was asking about weekly coupons, saving for grocery store trips and mapping travel times and so on. Many of those tasks are very traditional computational tasks and one of the main issues with them is just get the relevant data ingested. And getting that data ingested is usually something where yes, it's kind of messy to set up the, the pipes to get that data ingested. Once the pipes are set up, millions of pieces of data, billions of pieces of data can flow through those pipes. And it's really just a computational story. The process of getting the pipes set up. Can modern AI and LLMs be useful for that? We've certainly been trying to use them for that purpose for things like data connections and so on, and being able to take some, some random function somebody wrote somewhere and sort of interface it to our technology, stack things like this. I would say that the jury is still a bit out on how useful that's going to be, but I think it's a very, that's an under appreciated use of kind of modern AI is what one might call the offline uses, I.e. you use it to build a thing that you will then use many, many, many times in a more computational way. But the original building is done with the help of some AI. But once you've built it, it's a solid pipe, so to speak, that you can run lots of things through and that no longer has really an AI component to it. I would say that the obvious use cases for the current generation of kind of LLM type AI mostly revolve around the linguistic interface area, whether that's kind of customer service type interactions, whether that's medical history taking, whether that's tutoring kinds of things. Still we're working on that. But still it's not clear how well that's going to work. The main thing that we've learned about that is it's hard and it's difficult to kind of keep the thing on track and so on. And I think those are, it's kind of when language can be used as an interface. That's a place where the current generation of LLMs and AIs is really useful. And I think we still don't know all the different use cases to which that can be put. Let's see. Oh gosh, many, many questions here. Okay, Thyme is asking, what do you think of robots remotely operated by humans as a precursor to autonomous robots? I'm not sure. I mean, I kind of wonder with some of these sort of self driving taxes and things, you know, just what, how much is there a human behind the scenes and how well does that kind of work when it comes to, you know, drones being used, particularly military drones where there's, you know, a human pilot that's obviously a, you know, that serves that objective. Well, I think there's the case where it's send in the robot, it's too dangerous for a person type thing. That's an obvious use case for remotely operated things. I think there's a big bank of screens and every so often a robot gets confused and some human has to step in. Possibly that's a good backup mechanism. I'm not sure. I have to say that I've been surprised at certain kinds of things that have not caught on. Like for example, I had thought that remote tourism where somebody, you say, I'd really like to see the Amazon jungle, let's say, but I don't want to get all those vaccinations and I don't want to be, It's too rugged a situation for me personally, but I'd love to get somebody who, who hangs out there routinely to go and you know, take a camera and I'm kind of just wearing my VR headset and so on and I'm doing remote tourism. That hasn't caught on for some reason. And I, I think these, these cases where it's, you know, I, I don't really know the extent to which kind of the idea of have the robots do it most of the time and just ask the humans when they get in trouble. I don't know to what extent that, when that's a useful thing certainly for self driving cars. That's been a thing. It seems to be okay. I think that there are cases where when it's kind of like, well, almost all of this can be done automatically, but there's this piece that can't be done automatically. It's often more difficult for one to jump in in that situation than one might expect. It's like, well, let me understand what you did elsewhere. I mean like for example, I've not so much recently, but when LLMs were first a thing, I quite often would try and create LLM based text and say, well, let me now tweak that text. It was extremely difficult and confusing to tweak the text. It's a thing where it feels like the LLM has created something which is in and of itself kind of this integrated thing. And it's pretty hard to tell. It's like, okay, the LLM has written this document, check the document. That's often something, at least for me that's pretty difficult to do because it kind of seems like it's right. And then you have to really go in and basically check every word to know whether it isn't right. And by the time you're checking every word, you might as well be writing the document. So I'm not sure to what extent this gluing in sort of humans to the corners that the robots can't deal with, I don't know to what extent that will work. And certainly if you have your robot that's off doing something in some, I don't know, construction site or something and then the thing freezes and says I don't know what to do, come help me. And then it takes an hour to put up the ladder and this and that and the other to be able to get to the robot. That's probably not a great situation. Let's see. Right. So Cost says, believe in another crucial thing. Not sure what that quite refers to is that not only technologies adapt to people's demands, but people quickly adapt to technologies. Recalling how weird the computer mouse was for us maybe 30 years ago. Well, you know, I have to say when I first used a mouse, which must have been 1980, something like that, I found it immediately pretty intuitive, which is interesting. It's not obvious. One would. I mean, I have to say when I first saw a window system, which must have been around 1979, Xerox, Xerox PARC, an Alto, I think the first thing I said is how is this possibly going to work? You've got these things where they're like windows, they're like pieces of paper, and you can cover up one with another. And if you want to copy something from one to another, it's like, well, it might not be visible on the screen and then you can't do the copy. And that just seems completely crazy. And so my intuition was this isn't going to work because there are all these corner cases where this sort of physicalized metaphor doesn't work. I was wrong about that. I mean, I remember that every time I'm in a situation where I'm trying to move a file from here to there. And I can't do it with a GUI because I can't do it easily with the GUI because the destination I'm trying to move it to is covered up by something else or it goes out of focus or something like this. It's. It's like I'm kind of remembering back to when I first saw that such a system. And it's like, this just can't work. Well, it turns out that the 99% it can work is good enough and that it becomes a very usable thing. Now, in terms of things which are and are not intuitive, there are things which are and are not learnable. I mean, there are things where there is a manual task. You know, I think flying a helicopter, at least without autopilot stuff and so on, it's kind of on the edge of, oh, you can do that as a human. You can pull the levers and make the controls and so on. But for example, a quadcopter, people were not able to fly quadcopters. They didn't have the flight controllers, the automatic flight controllers that now exist that made quadcopters possible. That's something that can work. But if you say as a human, you've got two levers for these different or three levers or five levers or whatever it is for all the different rotors and move them all around and so on. Humans just couldn't do that. Maybe with enough training and practice, people could have done it, but as a practical matter, it wasn't something that was readily trainable for. I mean, there's always a question of how much training you prepared to accept before you use something. I mean, you might have thought bicycles were would never catch on because you need to learn to ride a bicycle. It's not something you can immediately just get on a bicycle and drive off type thing. And I think it's sort of an interesting issue. How much can you expect people to learn before they're sort of proficient at something. I think when it comes to, as I say, well, for example, touchscreens are probably, you know, maybe those are more intuitively obvious than mice. But I have to say, when I first saw a mouse, used a mouse, it seemed very obvious to me. I remember when my kids were really young, got them touchscreen computers, they weren't so much of a thing back then, but that was easier for them than even this big ball mouse type thing that didn't require fine motor control. So maybe some evidence that the direct, oh yeah, you point at the animal on the screen type thing was easier, at least at that stage. But as I say, for me at least the mouse was a pretty immediate yes, we can see what's happening. There are other things where it's like that interface is just not really, you know, humans don't get it. Now similarly, when you see user interfaces where there are just like, oh, there are weird glitches and humans just get it wrong all the time. I remember long, long ago a Windows system being shown a Windows system where. The scroll bar had in modern scroll bars pretty common to have arrows at the top and bottom of the scroll bar. You can, you can press on to move up and down. But somebody had the bright idea of let's put the arrows on the thumb on the thing that moves in the scroll bar. Okay, is that a good idea or not? Well, it's actually a very bad idea because what you want to do when you're clicking on an arrow is repeatedly click to move up a page, up a page, up a page and so on. But if you put the arrow on this thumb, then when you move up a page, the thumb has moved and so you can no longer just click, you can no longer just repeatedly click because you have to move the mouse, move the cursor to the place where the arrow moved to to be able to click again. But things like that are quite unobvious until you actually try them and see what happens. And I think so. I mean now in terms of, of what humans adapt to and so on, I mean, it's a well known thing that humans will often say my computer is learning from me. And really all that's happening is they're learning from the computer. A classic example of this was early kind of Palm Pilot handwriting recognition, kinds of things where people would draw these, where sort of the idea was, it was even kind of stated as that you would kind of learn a different form of handwriting to be able to communicate with the device. It didn't really ever Catch on that. Well, I mean, some people used it, but it wasn't a thing. I have to say I did the let me learn this and it was like, let me not learn this. This is a waste of time, you know, I might as well just be typing. But yeah, I mean this, this question of how do you adapt to the machine versus how does the machine adapt to you is an interesting one. I think that in, I mean it relates to a lot of things like does the language you speak affect the way you think about things? To what extent are you adapting to the kind of tooling that you have for communication? And we know particularly for computer languages that there's a substantial effect there, that the way the language is constructed has a big effect on the way you think about things. So in that sense you're adapting to the machine rather than the machine adapting to you. My feeling is that that's often a better thing than something where there is a sort of a, where, you know, if it's a well designed thing that you can go a distance with, so to speak, then better that you learn to use it and that you adapt to it than that your bad habits are then fed to the computer, so to speak, and it just amplifies your bad habits. But that's a, you know, it's a complicated collection of issues there. Let's see. RBS is commenting. It's very useful for us humans to understand what the AI knows when it outputs its LLM computations. Yeah, well, do you know what you know? Do you know how your thinking process works? We can introspect to some extent, but I think there's a lot about how brains work that is not what we will be thinking about when we're thinking about thinking, so to speak. Actually, I've just been studying quite a bit some things about the foundations of machine learning. And I would say one of the striking conclusions is that when you learn something with machine learning, the thing you've learned is not really a mechanically describable thing. It's not something where, yes, you figured it out by machine learning, but now you've built a machine that you can readily see the cogs and levers inside to see how it works. It doesn't work like that. What seems to happen is that you're dealing with these kind of irreducible computations that look really quite complicated and random, but just happen to be kind of entrained in the right way to do the particular thing you want. So it's not really the case that you can expect that there's kind of a mechanistic explanation, a narrative explanation for what's going on. It's more like, well, we piggybacked on this complicated thing that's happening anyway, and we managed to figure out how to take that complicated thing and use it for this particular purpose. It's kind of more like saying, rather than if you have some system for pulleys and so on, for getting a weight up something that's up a hill or something like this, you can go and look at the pulleys and see how they work. The alternative is to use a donkey to get the thing up the hill. And yes, you can get the donkey to do it, but you don't necessarily know. You don't know much about how the donkey works inside. And I think that machine learning is much more like the donkey case than it is like the pulleys case. And so then you can say, well, let's now become anatomists and study what donkeys are like inside. The problem is, it looks to me like in a lot of these machine learning situations, every different thing you learn is like a different type of donkey or a different type of animal. One of them is a giraffe, and another one's a donkey, and another one's an armadillo. And they probably have even more diversity than these biological organisms. So it's one of these things where you're kind of capturing something out in the computational universe. The computational universe of possible programs is incredibly diverse. And as you do each particular piece of machine learning, you're kind of catching one of those pieces. But there isn't a kind of an overall story about what's going on there. So when you say, what is the LLM thinking about? That's going to be a thing, I think that's pretty hard to have a narrative description of. And I think what will happen is, and this is a very global point, that you're sort of a choice between saying, yes, I need to understand what it's doing inside, and then saying, so that means that there's limitations on what it can do. You can only have this set of cogs and levers. Whatever you can do with those cogs and levers, you can do. But if you want to have a whole donkey or something, then you're out of luck because you are saying, I need to understand how it works. And that prevents you from having things which are computationally much more powerful. I mean, I tend to think this is one of the coming kind of major decisions that our species has to make. Do we allow computational irreducibility and in the systems we deal with, which means they have high capability but can generate surprises, or do we insist on kind of the mechanical situation of saying we know exactly the rules by which the thing works and we can make those rules, we can set those rules up so that we can see sort of all the way through how things work? Let's see. Alistair comments. Maybe some layered hybrid architecture may work with LLMs providing the base and other modules, connecting and correcting what's there. Yeah, well, this is what we've been building since January of last year of using our computational knowledge stack and computational language as kind of a backend for LLMs. And, and I think that's, that's really the, the obvious and inexorable future. And it's kind of, you know, I think the sort of. There's been this, this great enthusiasm for like AI neural nets. They're going to do everything. Well, the fact is, if you look at kind of the, the sequence of what's happened, it's like, yeah, we were all really surprised at the end of 2022 that a neural net did something that we never imagined a neural net could do, at least anytime soon. And so the conclusion from that is it's kind of like, well, we've seen this surprise, so we are convinced that we'll see every other possible surprise. In other words, what that showed us is that in the end the sort of The LLM like AIs will be able to do everything. I don't think that's the correct conclusion. I think what we saw is something which, as much as anything, tells us about the fact that there is more regularity and simplicity in the task of generating language and doing this kind of common sense discussion than we imagined there was. There's more regularity in that. That was easier than we thought. We just didn't have a way to have any machinery to do it. Now we do. But that doesn't mean that every other task that we thought is hard will also kind of fall to the same methodology. I don't think it will. I think that the thing that might have been the case is back 300 years or so before there were lots of things that were done in sort of formalized ways using mathematics and now computation and so on. It might have been the case that the things that people mostly kind of figure out off the top of their head, so to speak, would be much more amenable to sort of the neural net AI approaches that we have, which are in a Sense imitating the kinds of things that the layers of neurons in our brains manage to do. So I think it's sort of an interesting dynamic. There's a lot of kind of enthusiasm about AI, although I have to say it seems like in the last few months it has been starting to kind of level off or die down. But there was sort of a belief AI is going to do everything. It's going to be different this time. We finally have a technology that's going to be the technology that wins against all technology, so to speak. I've never thought that. I thought that this is another very useful tool that particularly emphasizes the linguistic user interface capability. But it isn't something where it's kind of like this will take all, so to speak. I think it's also kind of interesting to realize that there are moments when one is surprised by the arrival of technology. As with ChatGPT, I think that there are many cases where we might be very surprised, but it sort of takes a little longer for things to soak in. Like the fact that we can produce stem cells, for example. For that was a thing that was sort of discovered and I suppose it was discovered sort of sudden in its discovery, but it took a while for it to be clear what was really going on. It's not even clear now how stable those stem cells are, how much they can be used for medical purposes, et cetera, et cetera, et cetera. That tends to be a, you know, particularly things in the medical domain tend to be like the drama of discovery is definitely blunted by the fact that, oh yes, there's all these clinical trials and sometimes these things fail at a late stage. There's and turn out to have some weird side effect, et cetera, et cetera, et cetera. It's less like, here it is, boom, there's a website, you can just do something surprising with it. I mean, I think with ChatGPT that happened. I think that happened with WolframAlpha back 15 years ago now. And there are these occasional moments when it happens, but in a lot of industries the great surprise is not such an immediate surprise. I mean, I think the same, oh, I don't know, self driving cars, for example. It's kind of like maybe there'll be a moment when there's sort of the ultimate surprise of it just completely works. I kind of doubt that, but it's sort of a more of a slow progression type thing. Same is probably true with AI tutoring that there may be, okay, hopefully we'll have soon finally sort of an AI tutoring system that works but it's not a five minute demo because in five minutes you can't tell anything. It's like does some student learn this particular thing over the course of a course that lasts two months and do they come out of it with acing everything so to speak? That's the test. It's not a thing that has the moment of kind of wow, I never expected something like this could work. I mean there are so, I think so as a question of kind of what will happen with the current sort of enthusiasm around AI you know, the dynamics of things are such that there's kind of a tremendous, oh, you know, okay. Things that happen like in many kinds of settings people, the term AI becomes a, A, a kind of a brand term for lots of things. You know, I was, I was a little bit amused A, I was just communicating with a person I worked with many, many, many years ago. We, we invented something together which is sort of a, a mathematical method of analysis for things. And my friend in, in talking about this said and the machine learning thing that we did back in whenever it was 1978 or whatever and it's like that was not machine learning, that was a piece of, you know, mathematical anal. Pretty neat. But it was, you know, to call that machine learning is to sort of back project from today's cool brand onto something which was really not like that. And in fact often when you hear people say, and it's been true for quite a few years when you hear people say oh yes, we're going to use machine learning for this, what it means is that they're using, often it means they're using some simple regression analysis that has been something that has existed as something in statistics for well, the basic idea for 150 years and the slightly more refined idea for 100 years. So there's a certain sense in which that once the brand of AI is there lots of different things that one does what we call AI. It's a little bit confusing because right before the ChatGPT success there seemed to be a lot of feeling that you know, AI and machine learning and so on, like we have to kind of regulate against it because oh, we don't want AI to be used in this or that because it will be somehow it will, it will lead things astray. And then so there was a sort of a push to don't call the thing AI because then it will be subject to this, you know, regulation about, about oh, you can't use AI for this or that. And then Suddenly there was a. But AI really works, so let's call everything AI. And so that was a bit of a confusion there. But I think in the. And, you know, I mean, it's similar a little bit to what happened with Blockchain, where a company that was doing something quite mundane could sort of say, and it's like blockchain, or we're adding blockchain to this, and suddenly everybody would be more excited about it. Now, in a sense, when a company does something like that, it is a signal because some companies are frankly too sleepy and too sort of dyed in their wool to be able to even surface something like, and we'll use blockchain for this. It's kind of a signal, it's kind of a way of saying we're actually going to be paying attention to new things and so on. So it's not completely wrong to say, oh, the valuation of the company should go up. If it just mentions the trendy new thing, that's a sign that the company isn't completely asleep at the wheel. On the other hand, the idea that that's going to sort of in some sort of, I don't know, efficient market hypothesis or something, some kind of, you know, that that's going to mean that, oh, yes, we can know that this company is going to generate more profits at some time in the future, that, well, maybe it does mean something about that because it means the thing isn't completely asleep, but it certainly doesn't. You know, you can't necessarily draw a straight line from that to, oh, the products it's going to make are immediately going to be wonderful. But I think the, and this question of, of whether there are sort of these new categories of products that AI introduces, modern AI introduces, I think the answer is yes, there are a bunch of cases where that will work, but there are others where it's really just not a good idea. It's something where it's just not going to work. It's something where people are saying, oh, we'll put an AI interface on this thing. But it didn't need an AI interface. It's something where, if anybody had designed the website properly, you didn't need an agent to go read the website for you, and so on. So I think. Now, what is the exact dynamics of how all this works? I mean, a lot of money has been put into things that are called AI. And what tends to happen when lots of money is put into something based on kind of a, kind of a hope, a story, a brand is. Sometimes it just doesn't work out very well. And sometimes that kind of war chest of money gets sort of backfilled with real things. And what ends up happening is that some sort of steady, major, solid thing is built up that way. I'm not sure what's going to happen in this case. I think with blockchain, for example, where there were kind of large war chests generated, I think that's still very much unclear what's going to happen. I mean, there's still a fair amount of activity of trying to get communities around blockchains to actually build real stuff. How that's going to work is unclear to me. At least I will say that I think that the whole question of the economic value of blockchains, cryptocurrencies and so on is something I happen to have been thinking about a bit because I'd been for a while interested in the foundations of economics and understanding what makes there be value in the world. What really is this thing we call money, and so on. And I have to say my provisional conclusion is that there is something perfectly real to cryptocurrencies. It's not that by the time there's a big enough network of people using that and assuming value in it, that it is a thing that has value and that the fact that you can't necessarily buy a cup of coffee with it is not really the relevant thing. But, you know, more to understand there. But I think that's so. I wouldn't say that, but I think the kind of. The investment around kind of sort of blockchain as a thing that will have obvious utility in the world, there was a lot of money put into that. It's still in the process of attempting to be backfilled with kind of real use cases and so on. I would say. And I think in the AI case there is more immediate utility, but I think that the. And it will be able to do everything kind of story is. I don't think that story is right. And I think that the thing that's important is where can you put this AI layer on top of things that maybe are already there, already sort of computationally capable and so on, and really add lots of value to it and potentially unlock capabilities that were already there, but which nobody was really able to get access to. I mean, I'm kind of hoping with our technology stack that sort of the AI layers that we've built will make it. I think. I know it's the case that they'll make it a bit more accessible to people. It's something where It's a question of, well, do you just ramble at the AI or do you learn something yourself? In the end, it's going to be much more efficient to learn the thing yourself than to have to ramble at the AI. But it's easier to get started rambling at the AI, so to speak. I think the, I mean, there are other issues like, okay, is it worth spending a gazillion dollars to train an LLM? Is it worth kind of trying to ingest more and more training data from the world? I'm not sure. I think that the thing that is more likely to be the sort of emerging direction is more the kind of small LLM that provides that layer of wrapping that layer of kind of linguistic user interface to something which has kind of solid capability underneath. I mean, as I say, my surprise was that I could talk to a robot for an hour and a half. Okay, it crashed twice in the middle of that, but basically talked to it for an hour and a half talking to a thing that was running on a little tiny computer and saw underneath the robot that was just a local LLM. The fact that that worked tells me that there's really a lot of potential for things that will ultimately be completely local LLMs, and I think will ultimately this sort of common sense linguistic interface will be something that we take for granted, just like we take graphical user interfaces for granted and so on in operating systems. And I think it won't be the case that it's sort of, you know, just like it's not GUI all the way down. In the end, the GUI has to actually do something underneath. But the GUI is still useful in terms of our interaction with things. I mean, there are cases where it really is just a gui. You know, you're drawing something, it's kind of more or less just a gui, but in most cases there's something where there's a there there underneath the gui, so to speak. And my guess is the same will be the case for linguistic user interfaces. RBS is asking, what's the gold in AI, LLMs, etc? Is there some simpler algorithm that can learn instead of big neural networks, like trying to find gold in a gold mine? Well, I think the answer is yes. I think that neural nets, which are very old idea 1943, they are a way that one can make use of certain kinds of methods, particularly from calculus, to be able to incrementally learn things. They are also very hard to understand in terms of how they really work. And the amount of foundational knowledge about them is very Small, it's really as an engineering matter, yes, you can make them work as a. What's really going on here? One really doesn't know. I mean one can make certain statements, but the core sort of what's really going on, one doesn't really know. And in a situation like that, to say, oh, neural nets as they are now are the best thing we could ever have doesn't ring true. Because until we really understand what aspect of them is really important to their operation, we can't really say whether we need to have this or that piece. And as I say, the things that I happen to be working on right now are certainly suggesting that there are some rather different ways to think about this and that a lot of what's there in neural nets today can be kind of stripped away without removing the essence of the operation of neural nets and machine learning and so on. Let's see. LC is commenting. What do you make of the apparent disconnect between the heavy capex into AI infrastructure versus the lagging revenues from applications? Are we in for a 2000 style telecom vs fiber like setback? Well, I have to say that if this whole question, it really is in part a core technology and a science question. Do you need all of that sort of investment in infrastructure to be able to stand up sort of the AI capabilities we need? Or is it the case that, and if you believe that the kind of neural net AI is going to keep on going up and up and up, then it makes sense to invest in kind of more and more and more infrastructure because to even to reach that, to go further up on that curve, you're going to need more infrastructure. If on the other hand it's like, well, we discovered something and now we're going to tune it up, we're going to tweak it, we're going to get it so that in the end you can just run it inside your phone, then it's a completely different story in terms of infrastructure. And I have to say in the history of the computer industry, this whole progression from let's centralize, let's distribute, let's centralize, let's distribute. It goes back and forth. This was a place where it seemed like, well, with cloud providers and cloud services and so on, that was a centralization move. With more and more power in mobile devices, that was a decentralization move. Sort of the LLM epoch seemed like it was pulling things into the centralization again, by the way, other kinds of machine learning, the sort of the 2011 style deep learning kind of Advances, they all got into mobile devices in the end. They're all things that you can run on processors that will go in a mobile device. Well, that hasn't happened quite yet for the LLMs. My guess is that it will. And so my guess is that the big story will mostly be in use cases. My guess is that the infrastructure there will be some core infrastructure. It's just computers where you just can't get away from that. You've got to have computers, maybe even GPUs, maybe not GPUs. We'll see. I consider that's something that science will decide, so to speak. I think that. But, but I think it's. And then, then there's the question of the applications that are built on that. And will that be where the real value is? It's always complicated where value lies and what people just sort of expect to be generic and so on. I mean it's like in the, in the, for the Internet, it's like connecting to the Internet. Well, that could have been a, you know, a pay for everything type story. It wasn't something like email. In some countries, email was a big business. In other places, email was kind of like, oh, you can expect it for free. The tradition has been web search is kind of for free and paid for by advertising. I'm not sure that's going to continue. I think there's lots of arguments why it makes sense to have kind of pay for your web search and get web search optimized for you type thing. And this question of where people expect things to be free versus not. And then how does the ecosystem really work? I mean, could it be the case that AI LLMs will be free but they'll be stuffed full of ads all the time? It's a possibility. I think probably people won't like that and that won't be the thing that emerges. And rather people will prefer to just pay for it rather than. I mean it's. I don't know whether. I'm sure there are services where you get cell phone service for free, but you have to listen to ads in order to get that. But I think most people won't go for that. So I mean, I think this question of where the value is still a bit unclear. And it's also the case that sometimes there's sort of a core set of capabilities that provide most of the value and sort of the little apps on top aren't really that valuable. And other cases where the infrastructure eventually gets sort of taken for granted and gets sort of a low Margin kind of thing. And it's the kind of the application that people use that becomes the valuable thing. I think it's a bit unclear how that's going to work with kind of the AI setup. I have to say that in the particular case that we're in, where we've built our whole computational language and our whole sort of computational knowledge infrastructure, which is very unique, I mean, there's really nothing else like it. Nobody else has been as crazy as I've been over the last four decades of building this kind of computational representation of the world and so on. That's just a unique thing that we have. And it's sort of, to me it seems like that's a thing that should be a place where there's a lot of value in this kind of modern world of AI. But exactly how that value is being derived is not yet clear. Something that very relevant for kind of our business is figuring out how that value really gets derived. And certainly in the moments when people say it's just going to be a neural net, it's just going to be neural nets everywhere, that value is less obvious to see. Although I think that it's becoming more obvious as one realizes, yeah, neural nets do what they do and they provide these linguistic user interfaces, but they're not the whole story, so to speak. Let's see. Hyvie asks for full robot integration into human society. Will we see robot coffee shops where robots will be able to go and refuel and charge? What business opportunities would working robots open up? You know, it reminds me of the electric car recharging stations that have popped up all over the place. I would assume that in a situation of truly autonomous robots, and that's another interesting question is whether it's like, well, you can't fly a drone in most places unless you can see it. And there's sort of an idea, well, you can get a self driving car to drive around without you being there. What about a robot? A robot that is sort of far away from you and sort of just on its own hanging out, going to a coffee shop type thing. Is that going to work out? And sometimes these things are determined by specific incidents that happen where some disaster happens and people say, no, no, no, we don't want that. I think that the notion of kind of a, a, you know, the charging stations for robots, I don't think charging stations for, for electric cars have ended up. I mean, I'm sure there's some company somewhere that has done really well from them. But, but I think a lot of that is the actual car companies themselves installing things and so on. I'm not sure the how that ecosystem work, but I haven't sort of heard of it as being an immediate. Yes, this is a huge, huge business. I have to say when it comes to AI and thinking about AI and so on. I mean there clearly are sort of niches there like sort of the science of LLMs and AIs that's more for the science crowd, but that's a somewhat untapped market. I mean it's something where you see endless people just running an LLM and trying psychology tests on it and so on. Yeah, that's a thing. But how does the LLM work inside what science can you do about the insides? Well, I'm realizing that science is actually quite difficult, but I think that there's high value in it. And there was a brief time when sort of being a prompt engineer was a thing. I think that's less true now. You still have to write prompts but there's a bit less craziness in the way that the underlying models deal with prompts. And I think sort of the prompt engineering as the high value kind of activity has probably not made it, it wasn't obvious how that was going to go. I kind of think that at some moment the kind of the AI psychologist thing is going to be significant of you know, we've got an AI system. Why is it doing this? You know, with computer systems kind of debugging is bundled in with software development to some extent. I mean maybe DevOps is a little bit more, can be a little bit more on the debug this complicated system side. But I think the idea of sort of after the, after the fact, sort of send in the bug finders, you know, send in the AI psychologists may very well be be more of a thing. All right, maybe one or two more questions and. Let's see. Goldie asks how was your annual summer of professoring? And comments kudos to all the student projects. How was it? It was great. I, I, you know, I have to say that it's been my experience at least in the, in later in my life, so to speak. I didn't think this earlier on that teaching is a great way to learn and that I mean I think when I was younger I was partly because sort of teaching came bundled with research and research was what I really wanted to do. I was mostly like teaching is a necessary evil and rather than teaching has intrinsic value on its own. On its own. But that's certainly what I found now, and I have to say, I mean, we're very fortunate that we end up getting just wonderful students for the educational programs that we have and from whom I can learn a lot. And also my kind of role in these things is coming up with lots of projects for people. And that process is both interesting for me as kind of a creative act and also something where projects that I've kind of wondered, well, I wonder what happens with this. Okay, somebody works on it and they make progress. And watching them make progress and talking to them about that progress, I learned something about that project. Plus also these educational programs that we do and have been doing now for 22 years or so, they are. I just. One ends up meeting just all sorts of interesting people who, who end up being. Staying in one way or another, in the orbit or reappearing many years later. And it's just a sort of a great, A great way to meet a different cross section of people than one one meets in other ways. And that's, that's very, very nice. And I kind of feel like for me, there's a certain getting people to kind of get to the point where they have a new capability, whether it's sort of superpower of computational language or whether it's the kind of knowing that they can go from the concept of a project to an actual project and knowing that that's a thing that's within their capability. It's very satisfying to watch people kind of make that climb to in these kinds of areas. I mean, I'm not sure that I can rationally explain why I find that so satisfying, but I do. And it's something where it's kind of one feels like one's really contributing something, and it's a good feeling. Let's see. See, There's a question here from Anon. Will you let future robots enroll in the summer school? Interesting question. How do I feel about being a teacher of robots? I think I've done that all my life. I think that in a sense, a lot of what I've tried to do is to take things that I sort of understand and build them, formalize them so that they can be automated. So I guess my feeling is that's in a sense you get. In a sense, you get much higher leverage from teaching, from creating a system that is sort of broadly deployable as a piece of technology than you do to taking one person and educating them in some way. Now, you know, that can have lots of taking one person educating them. They may go off and do amazing things. So there's lots of, kind of, there's lots of kind of fan out leverage there. But it's, it's a less certain thing than if you just build a piece of technology and then can deploy it elsewhere. So I think insofar as it's like, okay, we can take an AI and teach it things, that seems like a good deal. The only problem is we know so little about how AI's work. It could be the case that, oh, you teach it a thing, but then you teach it another thing and it forgets the first thing it learned. I mean, there's a whole black art of when you try and fine tune an LLM and you try and add things to the training of an alarm that you need to kind of remind it every so often of things that it knew before, where it starts forgetting and just concentrating on the thing it just learned. It's kind of like it was cramming for an exam and the only thing it could remember is the details for that. And if you asked it a common sense thing, it's forgotten that. So it's something where I feel like building systematic technology. A clear win. Thinking through things and getting things to the point where you have enough clarity to build that technology, well, that's a clear win. Just sort of telling the AI something and it might forget it doesn't feel so good. I mean, in a sense, when you tell a human something, they might just forget it. But somehow for me, there's an additional satisfaction of even the moment when the human first gets it, even if they forget it in the future, is something quite fulfilling. I don't know whether that same thing will be equally fulfilling. I mean, this sort of relates to questions, things we were talking about earlier about the role of AIs in human interaction and whether if I teach an AI something and finally after the thing's been confused and confused and confused, it finally gets it. Is that exciting? I don't know. Actually, I haven't had the experience yet. I don't really know whether that's something where I would say, wow, I really feel good, or whether it feels like something, which for me is not very satisfying. I don't play video games, but I know certainly plenty of people who do. And when they reach oh yeah, we got to this level, it feels very satisfying to them. To me, given my particular psychology, I'm like, I don't care. And so I can't really say whether either for me or for other people, we just made a breakthrough. The AI just understood this one thing that we explained to it, even though it might forget it later, what that will feel like. All right, I should probably wrap up there, but thank you for some interesting questions. I see many more here that I didn't have a chance to answer today and look forward to answering them another time. Thanks very much and bye for now.