2040 Live: The AI Uprising Has Begun!

All right, let's get started. Welcome everyone and thank you for joining the live stream. We are here to celebrate the publication of my latest book, 2040: A Silicon Valley Satire. Let me start by saying a few things about the book, why I wrote it, what it's about, and what I hope to accomplish with it. and then we will open it up to questions from everybody, questions about the book, AI, tech, politics, culture, anything you'd like. So 2040 is a novel. It's about, and I'm trying to say this with a straight face, a chatbot running for president. So the plot is basically ChatGPT runs for president. And of course, you can already see the jokes coming one after another. So what happens in the novel is that there's an AI startup started by 2 grad students at Stanford of Professor George Jeffington, who think they've solved AI, like a lot of people do. And then they do a startup and things aren't going very well. So as a stunt to attract publicity, they decide to run their chatbot for president because now, because of a couple of Supreme Court decisions, it's actually possible, as they, as they one day while getting drunk realize. So they run this candidate for president in the Republican primary, and to their surprise, it gets the backing of this libertarian trillionaire. And wins. And now this dozen-person startup suddenly finds themselves in the general election facing the Democratic candidate, who is a fake Lakota chief who wants to abolish the United States. So the novel opens in the first debate, right in the middle of the action, and with Ethan, who's the protagonist, and who's the CEO of this startup, terrified that, you know, Prezibot, as this presidential chatbot is called, is going to hallucinate or say something stupid and completely ruin the campaign. And, you know, it only gets worse from there. No spoilers. What I hope the novel illustrates is a lot of things about the tech industry and about AI, about both the hypes and the fears around AI and what's justified and maybe what are the worries that we should have. For example, one of the key plot elements in 2040 is the panic button. Prezibot has a kill switch so that Ethan can take control of it if things go awry for one reason or another. And of course, kill switches are real. People have written papers about having kill switches for AIs. But then Ethan loses the kill switch, and of course what follows is a big scramble that also illustrates that the real problem with AI is not the AI being evil, it's who controls it and control falling into the wrong hands. And, you know, the kill switch is the MacGuffin of 2040, if you will. It also illustrates a larger point about AI, which is that kill switches maybe are not such a good idea and what happens when different people are in control of it. And my goal when writing the novel was that everything in it, as well as being fun and doing the usual things of revealing character and advancing the plot, would also illustrate something real about the world of AI and tech and, you know, the present time in America. And literally everything in the novel, the characters, the situations, the, the companies, the, uh, the plot turns, every one of them represents something about AI. So on the one hand, the novel is a very quick read, very fun. You know, you can read in a couple of hours. It's, it's very dialogue-driven, and dialogue goes by fast. But at the same time, there's a lot of content packed into it. And I tried to make it something that people could read on 2 levels. People who don't know much about AI or the tech industry hopefully will actually learn a lot about AI and the tech industry from just from reading this. People who do know more about AI and the tech industry, they will find a whole series of inside jokes to various degrees, which I think is part of what's going to make it fun. And I— people often ask me, you know, it's a bit of a departure considering that, you know, I'm a scientist and my previous novel was popular science. Why would you write a satire? And there's a couple reasons why I did this. Talking to people in all walks of life about AI, you know, apropos of The Master Algorithm, my previous book, which explains it to a broad audience, I came to the conclusion that often the best way to show something to someone is not necessarily to give a long explanation or even write a whole essay or book about it, is to illustrate it with a story that shows it actually happening. And that planted in my mind the idea of writing a novel about AI in particular, I think, you know, instead of trying to explain to people that AI is not Terminator and Terminator is not coming to kill us and AI is not going to turn spontaneously bad, but that the real problem is the people who are in control of the AI screwing up or the conflicts between them or the wrong people getting control of AI or the AI itself screwing up because it hallucinates and it has bugs and it's a work in progress. The best— I think the best way to do that at the end of the day is to just show people. an AI that's like that, and that's PreziBot. So if at the end of the day when people read the novel, all it does is that when they picture AI instead of picturing Terminator, they picture PreziBot, I think the job, the job is already done. So with that, I will take questions. Let's see if there are any questions already in the, in the chat box. Don't be shy, ask away. I maybe will start with a few questions that people have already sent me. Here's the first one, a very good one. What are the advantages and disadvantages of AI candidates versus human candidates like Trump or Harris? Well, first of all, let me clarify. Because this is something that has already come up in a bunch of interviews and so on, that, you know, I do not literally mean for PrezzyBot to be a realistic candidate. I am not advocating that people should vote for AIs for president right now, but a couple of provisos. The thing that triggered the actual plot of the novel, as opposed to just the general idea of writing a satire about Silicon Valley, was a journalist asking me, so would you vote for an AI for president? And I thought about it for a little bit and I said, well, no, but if the other candidate was Trump, then maybe yes. Because, you know, I'm not a fan of Trump. Sorry for the Trump fans out there. And it really comes down— and I'm not a fan of Harris either, to answer the question. So the thought on my mind back then, and this was years ago, was, well, you know, AIs are definitely not good enough for being president right now, but if the human candidates are bad enough, then at some point the AI becomes the least bad option. And back then this seemed a bit far-fetched, but I would say that in our current moment, maybe, you know, given how fast AI is getting better and, and, and how bad the human candidates are, uh, you know, maybe President Bot would be better than Trump or Harris. Right now, there's several aspects to this, right? One is that if you know Moravec's paradox, the jobs that are hard for humans are easy for AI and vice versa. So in some sense, the harder the job is for a human, the better candidate it is to be done by AI. And there is no harder job for a human than being President of the United States of America. It's really— it's an impossible job. No one can do it well. No one can know all the different issues and all the briefs and, and really have the right judgment in all situations. It's just not humanly possible. So maybe a good AI would actually be better than humans in many circumstances. Another aspect is that, uh, and again, counter to things like LLMs today, AI can actually— well, and even considering them, AI can actually be much more transparent and much more responsive to what the people actually want in a democracy than human politicians, who in addition to their cognitive limitations have conflicts of interest, right? They say they're defending your interests, but really, uh, at the end of the day, defending theirs. You can actually design an AI that, well, what it really does is channel what the people think and, you know, reason based on it and come to decisions based on it. So in that regard, I would say that in their different ways, something like PresiBot might actually be a lot better than Trump or Harris. Now, of course, there are obvious reasons why an AI candidate wouldn't be good, but here's the thing. There are already AI candidates and have been for a few years. This was in the papers recently. There's an AI running for mayor of Cheyenne, Wyoming right now. There's an AI called Vic. There's an AI called Lincoln that says they're going to run for president in 2028. There was an AI that ran for mayor in Tokyo a few years ago and it placed third. There's a synthetic party headed by an AI candidate in Denmark. There was an AI candidate that ran against Putin some years ago. So this is actually real. Now, of course, in all or most of these cases, what they do, because an AI is not constitutionally allowed to run, is a human runs as a placeholder for the AI. The human says, well, you know, I will do the, the, you know, the ribbon cutting, if you will, the ceremonial things. I'll be the constitutional monarch, but in terms of real decisions, I'll just implement the decisions that the AI makes. Of course, in— as I mentioned, in 2040 there's a slightly more interesting way in which this comes about, which is the Supreme Court rules that all sentient beings are citizens, right? You can see that as a natural continuation of animal rights and things that exist today. And then in another decision some years later, the Supreme Court rules that yes, AIs are sentient beings. And again, there's a lot of people today who say that they are or will be and should have rights and whatnot. So if you take these 2 slightly implausible but not completely implausible steps, now an AI can run for president. And that's what happens in 2040. And again, the motif in 2040 is that, as in a lot of science fiction, to just extrapolate current trends to the point of ridicule, but to arrive at something. And a lot of people have said this about the novel, that on the one hand seems absurd, far out, and over the top, but on the other hand, it's kind of oddly realistic. You almost feel like it could happen. one of these days. In fact, to be completely honest, I conceived of the novel several years ago before there was ChatGPT, and back then it was harder to explain to people what it is. And so many of the things in it have come true in the last few years that I felt I just really need to publish the novel now before it gets outdated. So here it is. Let me see what other questions there are. Okay, here's a question about me, not about the book, which is, I see you're professor emeritus. May I ask, why have you retired from your faculty position and have you been writing full-time? So why did I go emeritus? For a couple of reasons, but here's the simplest one. I, I like many of the different aspects of being a professor, uh, but in particular the one that, you know, that really I always, uh, you know, have loved best and, and caused me to go into is doing research. The problem with doing research as a professor is that you are multitasking about 100 things continuously. And so what I tell people is that when I was a full-time professor, doing research was my hobby. Now that I'm emeritus, I can actually do it full-time. So in answer to your question, no, I have not been writing full-time at all. In fact, 2040 was, was something of a lark. It was more of a hobby. It was something that I did on the side over the last few years. It was glorious fun. It was some of the most fun I've ever had, and I hope you have fun reading it as well. Other people who've read it so far seem to have had a lot of fun. as well. And it's a great complement, you know, it exercises a different part of the brain to doing actual research. But what I have really spent the great majority of the last several years doing is research. And what being emeritus has allowed me to do— and again, this was really part of my goal— is I'm doing research without the short-term constraints of you have to get these grants and, you know, and publish these papers and, and so on, which at least for the sake of the students, whether you like it or not, you have to do. But each of these things forces you into local optima. Every paper is a local optimum that you fall into and that winds up spending a lot of time. So what I'm doing is I'm doing— my goal is to reach human-level AI and beyond. I think we are at the point where we can do that. I think the current state of the art in AI, as exciting as the recent progress has been— I love transformers, I think they're amazing, and I think LLMs in many ways are amazing. It is not— this path is not going to get us there. I firmly believe that. There may be things starting from here that will get us there. Maybe we need to go back to an earlier point. Maybe it doesn't matter because you can get there either way. But I'm working on something that I do believe is our shortest path to human-level AI. And hopefully you will be hearing more about it in the not-too-distant future. Of course, the history of AI is full of people who think they're on the path to human-level intelligence. The latest example, of course, is the likes of OpenAI. There's a lot of people there who think AI is very close. So I could be wrong. And at the end of the day, machine learning, which is really what I do and the foundation of AI these days, it's an experimental science. So only the experiments will tell. The experiments that I've done so far have gone very well, so, you know, I have some validation already, but they're on a small scale. What I've done is deliberately on the smallest scale possible while being on interesting, very challenging problems that we can't do with the state of the art. But who knows what will happen when we scale it up. Scaling up is not just pushing the button. It's a whole difficult process that requires creativity in its own right. And again, the likes of OpenAI are good at it, and that part is, is, uh, you know, is important as well. So we'll see what happens. Stay tuned. So here's another question. Uh, what does it say— uh, oops, I think I'm missing the bottom of it right here, or, or perhaps not. Uh, what does it say about humanity that we are so quick to rely on AI, from day-to-day tasks to long and planning? A good question, and it says, it says 2 things about us. It says, number one, that we're lazy. But it's not a bad thing to be lazy. In fact, there's bad lazy and good lazy. Good lazy is, is, uh, the, the, the, in some ways, the biggest least recognized factor of progress and civilization that we know. It's when people automate what they do so they have to do less of it and take a break. If you think about it, laziness, like having steam engines to do what our muscles used to do, laziness is a great motivation for that and it's a great source of progress. And same thing with intellectual tasks and computers. And in fact, I often say jokingly to my students in the first lecture of machine learning that lazy are us. You know, like, if— like, we do machine learning because we're too lazy to do the programming. We have great respect and admiration for all those great software engineers and developers out there, but their life is hell. I'm too lazy for that. I, I want the AI to do all the hard work for me. Yes, there's still a lot of stuff that you need to do, like, you know, collating data and cleaning it, integrating it, and babysitting the runs and deploying and whatnot. But at the end of the day With a good piece of machine learning, you can do the work of a million people while going to the beach, or at least going to the beach more of the time than you could before. So, you know, so in that way, you know, being lazy is good. The other aspect, and this I think now is more mixed, is that— and we see this very much in action today— is we have a very strong tendency to overestimate how good AI is. Precisely because of Moravec's paradox. A lot of things that we think are going to be very easy for AI aren't, but we don't realize that. And so we tend to give to AI a lot of tasks that it is not ready for. And in fact, this is a much bigger danger than Terminator or, or the jobs apocalypse or any of those things, is that we have AI— and this is actually not a hypothetical problem about the future. this is a problem today, is that we have AI doing things that it's actually not yet capable of doing, but we misinterpreted the demos. Again, part of the joke in 2040 is that the way you succeed as an AI startup is with a demo that impresses you and makes the AI look a lot more intelligent and more human than it really is. We believe the demo and then we have the AI doing things that maybe it shouldn't be doing. There I think we need to be a little bit more cautious. Of course, a famous example these days is Air Canada using a chatbot that just made up a refund policy for one of the customers, and then of course the customer was not able to get the refund. He actually took Air Canada to court and the judge said, that chatbot is your representative, you have to honor what it says. So companies can learn the hard way to not do some of the things that they're doing with AI. And there are many more mundane examples. In fact, I would say as a general rule, the killer apps for machine learning are the ones where it's okay to be wrong a good fraction of the time because machine learning is necessarily imperfect. Things like recommender systems or search results and whatnot. The opposite of the killer app, or maybe a killer app in a very different sense of the word, is things like self-driving cars. Right? I'm all for self-driving cars, and we are going to have them sooner or later. They're coming along. But as an application for machine learning, self-driving cars are about the most difficult you can imagine because everything that goes wrong is potentially somebody dying. And of course, a lot of the joke in '24 is that this AI now has some very big responsibilities that it's absolutely not prepared for. And, and then we see, um, you know, the disasters that happen. Let's see more questions. Is AI progressing at the same rate in other countries as it is in the US? Very good question. Painting with a broad brush, I would say the following: in Europe, AI is progressing at a slower rate at this point than it is in America. Which is actually a very sad thing to observe because historically, since the beginning, there was more AI in America than in Europe, but Europe was perfectly keeping pace. There were important AI communities in Europe going back to at least the '70s in several countries doing very good work. So while AI was more of a research topic and not something with a lot of real-world impact, Europe was, you know, on the frontier along with the US. Those were the 2 main players. Unfortunately, now that AI is really important and having impact and making a difference politically, economically, socially, and increasingly militarily as well, Europe has decided to hamstring itself with fears and with regulations to combat those fears that are causing it to slow down and fall behind. So on the research front, Europe is still doing okay, but there's certainly— you look at the most recent progress and very little of it has come from Europe. But even more so in the application of AI in the industrial deployment, Europe is seriously falling behind, and this is really, really sad. Now, of course, there's also China, and what happens is that China is making faster progress than the US. This varies by area. It's different in language than in vision than in learning than in planning than in etc., etc. But overall, China is progressing faster than the US. However, Because it started from behind. Overall in AI, the US is still ahead. But if you extrapolate these lines for another 5, 10 years, China is going to be ahead, which is very— and in some areas it already is. Not in LLMs, not in chatbots, but even there in many ways it's caught up pretty quickly. And this I think is something that people in the West and people who care about democracy really need to worry about. Because you do not want the leader in AI to be an autocracy. Unfortunately, Xi Jinping, having read famously The Master Algorithm, is highly aware of the importance of AI, and China has a very explicit plan to dominate AI within 5 years. And so far so good is all you can say. Now, there's this big question looming over all of this. Kai-Fu Lee wrote a book where he talked about this. Where if AI is already mature and what matters is deployment and doing it on a large scale and trying a million different things, China has the advantage. They just have, you know, more people working on AI or are about to have, or again, this depends on the area, but looking to the future and they have an enormous amount of data, etc., etc., and they have much fewer restrictions and, and hesitations about to do than certainly Europe, but, but, but to some extent increasingly also the US. If— how— and this is his argument, is that China is winning and is going to win because of that. I actually don't buy that because I believe, as, as I alluded to earlier, that the— there are major problems still unsolved in AI, and to do that really fundamental research The West and the US in particular still have the advantage. We can go to all of the factors that are involved in this, from the society being freer to the education system being different, et cetera, et cetera. But the truth is, if you look at even just the last 10 years at the major advances in AI, almost all of them, with a couple of exceptions, came from America. And I think America needs to not only hold on to that, but make the most of it. But it's also important to then turn that into, into products and services that people use. And there, at this point, the US does not have a big advantage compared to China. And unfortunately, the fundamental advances are, A, hard to keep secret, and B, it doesn't really make sense to keep them secret anyway. So even though in some ways, you know, America is still better at doing the fundamental AI research than China, This doesn't really, at the end of the day, necessarily buy much of an advantage when it comes to economy, politics, and the military. And so we really, we and our governments really need to get on top of this before it's too late. Let's see, what are the questions? Exactly, someone just mentioned, as you said once, America innovates, China copies, and Europe regulates. And yeah, I agree with that guy. I think he nailed it. Do you think the future of AI will return to expert system models like your work on MLNs? Great question. And of course, you can guess that this is a favorite question of mine. I really think, and of course, if you read the Maastricht and Mornoy thesis, that we cannot— so there are these major paradigms in AI. The 2 best known and historically most important ones are neural networks, and symbolic AI, the previous AI summer in the '80s was expert systems that were created by symbolic AI, and it was based on rules that were written by hand. And there was a lot of excitement, which in many ways is eerily similar to what's happening today on many levels, but then it hit a wall because that is not enough. Now, what we're seeing today is deep learning, which is just the latest incarnation of neural networks, and LLMs, which are themselves just a type of deep learning, Uh, you know, they've made enormous progress. This is fantastic. But again, we know, we see that they can't reason, that they screw up immensely in all sorts of ways, to contradicting themselves, you know, in their second answer after the first one, to just doing things that are completely nonsensical, to image generators that clearly have— that, that can really fill the eye with beautiful demos but have no idea what a human body is really like, for example. All All of that shows that what we have in deep learning is not enough. Some people say that you can somehow change deep learning to do these things. I don't believe so. I think you really need to do a very deep, not a superficial combination, not just of deep learning and symbolic AI, but of deep learning and symbolic AI and probabilistic AI, which is what Markov logic networks are and the The master algorithm that I— the candidate master algorithm, if you will, that I talked about in the master algorithm was based on Markov logic. So that's a combination of symbolic AI and probabilistic AI. We also need things like embeddings, right? We now understand— this was a surprise of the last decade— that embeddings are not just a way to get symbols into neural networks. They actually perform a really useful function. So we need that as well. We also need the reasoning by analogy. This, I think, has been sorely neglected, but the reasoning by analogy is a really important part. And, you know, MLNs did some of that, as I described in the Master Algorithm, but what I'm working on right now, I think, does all of this much better. And of course, the 5th one is evolutionary learning, or this kind of crossover-type adaptation. We, I think, we can get pretty far in AI without that, It's an open question whether at the end of the day we're going to need it or not, but I think it's— let me just put it this way: gradient descent, I don't think, is the search algorithm that's going to solve AI, or greedy search in any form, whether it's discrete or continuous. I think we need things that can make bigger jumps in the search space, and analogical reasoning is one of them. Crossover is another. But having 2 systems that are already very high performing and then switching components between them seems like a much better thing to do than just tweak one system in little ways that maybe for a system, a simple system like bacteria, this is a way how they work, like by exchanging small pieces of DNA works, but not for things like, like human brains. So I think it is very possible at the end of the day We will not be able to succeed without those 5. Maybe some subset will be okay. Maybe we need things beyond those 5. In fact, another strand of my research is on things that I think are not in any of these paradigms. But for sure, at least combining things like symbolic AI, probabilistic reasoning, analogical reasoning, deep learning, I'm very confident that we need that. Now, it may be possible. In fact, we're seeing some of this. For example, DeepMind does this a lot. They actually produce things that are very successful, kudos to them, like AlphaGo, and then it's another great success for deep learning. But what you look at when you look at what they're actually doing, it's a combination of deep learning and symbolic AI. They're using ConvNets to learn on the one hand, you know, with the board as the pixels, if you will. But on the other hand, they're doing Monte Carlo tree search, which is very— it's a very traditional, very standard symbolic search. And the reason the system works is precisely that it has these 2 components. With, with just one of them, it wouldn't. So when you hear a lot of these claims for the latest greatest success of deep learning, you should take a closer look and think about, you know, is, is this really just deep learning, or are there other things in here that should also, uh, um, you know, get some of the credit? And, you know, we should let a thousand flowers bloom, but, but the path that I think is most worth pursuing is the one of unifying these paradigms. Next question, uh, if AI is going to be used in government and politics, shouldn't the American people have more of a say in how it is being developed or how it works? Uh, absolutely. Uh, I violently agree with that. And in fact, I agree with that so much that you could even say that at the end of the day, that is the main topic of 2040. So without giving any spoilers, there are 2 versions of PreziBot in the novel. PreziBot 1.0 is really ChatGPT. You, you'll recognize it. Uh, it's, it's a chatbot, right? It mined a bunch of data, and now it can often generate very clever replies and very witty, you know, very pithy things to say and so on. But, but at the end of the day, it's just a chatbot. And they very quickly realized that that's not going to win them the election for a number of different reasons, but which, you know, I will not go into apropos this question. But what they gradually— the realization that they come to, and naturally does not come easily to them, but, and again, it's not coming easily to people today, but this is where I think things should go and will go, is that Presidbo 2.0 is, is really an AI that the people, we the people, are in control of. It's crowdsourced in real time. So, you know, somebody asked, you know, somebody did a joke interview with PreziBot, and the first thing they asked, so it was, so what are your pronouns? I love that question asked of PreziBot. And PreziBot's answer was, well, version 1.0, the pronouns are it. It's right. It's the first PresiBot is an it. It's a chatbot. But PresiBot 2.0, the pronouns are we, us, because PresiBot 2.0 really is us. You know, any millions of people, 5, 10, 50, 300 million people can be talking to PresiBot at the same time, telling it what they believe, what they want to do. And PresiBot in real time, because it's a whole vast data center, it It's a real mega brain, right? It's combining all these people say, not just about the same question, but about sub-questions. And this whole inference change that— so PraiseBot is really just making the polity more intelligent, which is really, I think, is the future of democracy, is to have AI solve the dilemma between direct democracy and representative democracy, right? Our representatives, as I mentioned earlier, are necessarily imperfect, both because of their cognitive limitations and because of their conflicts of interest. We can actually have an AI that overcomes this problem, and we can actually have, you know, think of the agora, right, the forum in, in, in, in Greece or Rome where people would discuss things, right? We can actually have that discussion with, instead of a few thousand people, millions of people mediated by, by AI. And in fact, I would say that social media today They're the people who like it, some of the people who say this is horrible and we need to stop it and, and, you know, outlaw its use and, and, and constrain it and blah blah blah. I think they have the wrong idea. What we need is better algorithms running over the social media inputs, producing outputs and suggestions and, and intermediate results and, and whatnot. And we're starting to see a little bit of that with things like the summaries that, that X produces and whatnot. We need to, to bring social media up to a higher level of intelligence that also makes us more intelligent. The way this happens in 2040— again, it's partly serious, partly joking— is that the panic button is really just an app that Arvin, the CTO of Kumba AI, their startup, rigged up so that then Ethan would be able to just use it and take control of PreziBot, either switch it off or start literally talking you know, through PreziBot. And then what they do, because there's very little time left in the campaign when they decide to do this, is they just turn that— the panic button into Prezi app, which is basically giving everybody a panic button, giving everybody a kill switch, giving everybody essentially a microphone by which they can talk to PreziBot, or if PreziBot sees fit, talk through PreziBot. At any given point, PreziBot could actually be saying what you said because PreziBot decided that that was the best thing to say. And with millions of people saying things very often, even if most of them are useless or irrelevant or wrong or whatever, the chances that some of those will actually be just what you needed are very high. So once everybody has a panic button, you know, think about like everybody with X or something else on their phone, right? Imagine combining all these people who are, you know, in real time contributing to a debate. Imagine if the debate was more intelligent. And the decisions were made that way, I think that would be a much better democracy than we have today. And, you know, one more thought is this: the autocrats like, you know, Xi Jinping and Putin and whatnot, they're already going all out to perfect their autocracy using AI. So we had better perfect democracy using AI, because again, if you look at the 20th century, the, the autocrats were the first ones to make use of the new technologies like film and radio. And then, you know, Churchill and Roosevelt and whatnot figured out how to use this as well. And, and, you know, the propaganda war in the Cold War and whatnot, eventually the, the, the West, you know, started winning. But in the beginning, the West was very behind and the bad guys were making headway. And this is where we are now. So we need to start worrying less about how to contain AI and, and limit it because it's an enemy of democracy, and more about how to use AI to, to perfect our democracy. Because, you know, I believe very strongly that democracy is better than autocracy, but there's no absolute answer to this. It's democracy with a certain technology may or may not be better than in terms of just winning, not, not the question of whether it's better for the people, because that democracy surely is better. But in terms of winning, you know, with, with each new technology, the question is posed again: will democracy with this technology prevail over dictatorship with this technology? And the question is very much an open question in our time, and it's going to be a very important one in the next, you know, 10, 20, 30 years. And sadly, so far, I am not optimistic about how things are going, and we really need to turn this around. Here's another one. Um, oh, you know, simple question: any good URL to learn and practice Markov logic? Absolutely. If you just go to my homepage. Um, so we have this system called Alchemy and, and, and, and the associated, um, website, which is just alchemy.cs.washington.edu. But if you don't want to remember that, just Google my name, go to my homepage. There's a section there that's software that I've been involved in producing. One of them is Alchemy. If you go to that website, it has papers, it has datasets, it has MLNs that we and others have produced for various things. It has other implementations of Markov logic, so there are many good ones besides Alchemy with different strengths and weaknesses. So I recommend you go to that website and then you can download it, play with it. You know, there's a community of people who use it, so you can take it from there. There are people who've done, you know, won NLP competitions and whatnot in different areas using Markov logic. And, you know, it's been I think there's a good chance that whatever your application area of interest is, you can probably do something interesting with Markov logic there that you would have a much harder time doing with deep networks, for example. It's not as scalable as deep networks right now, of course, but it's actually pretty remarkably— we've done Markov logic networks based on, you know, that have millions of objects in them and hundreds of millions of weights. You know, it's, it's not as scalable as, uh, uh, you know, uh, the current LLMs, but, but it's not bad either. And for many applications, it's probably more than you need. Um, historical question, uh, was Shannon at von Neumann level, or maybe more like Gödel? I think these were all brilliant people. I think by far the most brilliant one was von Neumann. Both if you talk to the people who knew him, right? You know, he was at the Institute for Advanced Study with Einstein, and the people who knew both said he was, you know, a lot smarter than Einstein, which is saying something. But also just in terms of his contributions, right? So Shannon made a couple of really important contributions. One of them was, of course, information theory. Another one that is much less well known but is actually hugely important was that in his master's thesis— I actually, I don't think I I don't— this is probably the most impactful master's thesis ever written. He actually discovered something that we take for granted now, which is the mapping between, you know, digital gates, digital circuits, and Boolean logic, without which we wouldn't be able to design anything today. But Boolean logic was created in the 19th century, long before there were computers, and computers grew up without knowing about Boolean logic. He was actually the one who said, actually, These 2 things are the same. So very brilliant. But, but von Neumann invented the von Neumann architecture, which alone dwarfs all of this in importance. And he also invented game theory, which is the mathematical foundation of social science, of evolutionary game theory. And those are only 2. Like, he's done like half a dozen things, each of which is enough to make him a major genius. So for all that I like Shannon, I don't think it's quite at the level of von Neumann. Gödel is also interesting. Gödel, I mean, Shannon had varied interests. You know, von Neumann was remarkable in that he made contributions in a whole bunch of different fields, unlike Einstein, who, you know, was just a physicist. And, you know, that's fine. Gödel was, was just a mathematician. Von Neumann started as a mathematician, and then, and then he became a bunch of other things. And Gödel, he was— I mean, he made a few really amazing contributions, but then honestly, he kind of lost it. And for the last many years of his life, he's actually— his mental health was not very good, which if you read the Logic Comics, you know, tends to happen to logicians for some reason. So, um, you know, as much as again he made very important contributions, I would say that, you know, neither Shannon nor Gödel are at the level of von Neumann. Which of those 2 has the edge? I would say Shannon, because for, for as great as, uh, and impactful in mathematics as Gödel's, you know, theorem, incompleteness theorem, and so on have been, you know, the impact of what Shannon did, including his master thesis, has been greater. Also not Gödel's fault, but a lot of people like to cite incompleteness as a reason why you can't reach human-level AI with computers and blah blah blah, and this, this honestly is just misguided. Gödel's theorem is about the limitations of mathematics. And those limitations come from allowing infinity, where you start from Peano's axioms where every number has a successor and now you have infinity. If you disallow infinity, then you don't have any of those problems. Computers don't have infinity. They're naturally finite. And infinity is just an abstraction that we have in math that is convenient, that makes things simple, but also creates a lot of problems. So really Gödel's theorem has nothing to do with the computation, with the limits of computing machines. It has everything to do with the limits of using— it's the price that we pay for using infinite mathematics, but it really doesn't affect computers or human brains in any way whatsoever, I would say. Why does the novel center on the year 2040? Good question. So it has to be an election year, right, because there's a presidential election, so it had to be 2036, 2040, 20-something. Now I wanted the novel that would be very near future. So there's, there's a lot of near future science fiction, but it's usually not that near. It's far enough away that you can wave— like, think of Neuromancer, for example, that, you know, invented cyberspace, if you will. Like, there were predecessors, but as a very, you know, wide— as part of the culture, that's— you, you could say that at some level we already have cyberspace now in, in online, uh, games and, and so on. And You know, Mark Zuckerberg is making sure that it arrives sooner rather than later. But Neuromancer happened in the future that was maybe, I don't know, 100 years out. I didn't want that. I wanted this to be close to today for a couple of reasons. One is that the closer it is, the more urgency you feel about the issues that it raises, and that's part of the goal. The sooner it is, the more you feel that it's real. And again, 2040, despite the date, is not about 2040. It's about today. 2040 is about 2024. 2040 is not about 2040 any more than Animal Farm is about animals or farming. 24 is a metaphor for the present time, and all of it is really just a caricature of where we are today. And again, some of those things, they started as caricature, they're already happening. on both the AI, the tech, and the politics side. So I think, you know, I actually, when I, when I picked 2040s and I was like, well, you know, worst case we'll make it 2044. But, but no, I, I went with 2040 and I thought it was a good choice. Also, 2040 has a lot of other interesting sides, like, like for example, it, it rolls off the tongue pretty well, 2040, right? It's also an allusion to 20/40 vision, right? Our vision of the future is a little fuzzy. It's not '44, it's maybe 2040 at best, right? It's also, of course, an allusion to novels like 1984, right? Like George Orwell wrote it in 1948, you may not know this, and flipped the last 2 Gs, which turned it into 1984. So we're playing a little bit of a similar game here with 2040, not quite as precise, but, you know, it's 2024 in another guy. So, you know, that's why I went with 2040. Let's see, what else do we have? Uh, you just posted on Twitter that companies shouldn't give their AI fun names. PreziBot is pretty functional. What is your concern there? Why does it matter what we name AI? So what I just posted on Twitter was not that, you know, companies shouldn't give their AIs fun names. There's nothing wrong with fun names. What I'm saying, and there's again, there's a very serious point behind this, which is that companies shouldn't give their AIs names like, you know, Cortana or Siri or Alexa, or those aren't even the ones I'm thinking about. I'm thinking like, like Bank of America has a chatbot called Erica. And when you go on the Bank of America website or app, it says like, hi, I am Erica. And what I'm saying is that this is a mistake. This is a mistake and will become more so in the future, I believe, because it's creating a different personality. It's like, who is this Erica separate from the bank or the company? And like in the example that I gave was Walmart, right? When you go and talk with Walmart's chatbot, it shouldn't say like, oh, hi, you know, my name is whatever, Erica. It should say, hi, I'm Walmart. You don't want to be thinking that you're talking to a chatbot that represents Walmart. You want to be thinking that you are talking to Walmart. And in the future, as more and more of the company is AI, right, I think in the, in the, in the companies are going to be like raisin cakes where the humans are the raisins, right? So we, we add value. A raisin cake without raisins will not be as good. But most of the cake is AI. So when you talk to Walmart's chat, you're not talking— you don't want to be talking to Walmart's chatbot. You want to be talking to Walmart. And for you, the company, that's better once you're ready to take that risk, which maybe many companies aren't today, and in some cases maybe for good reasons in the current state of things. But that's kind of an ironic reason to not do this, is you want to feel like You don't— like, the problem today is like, oh, I said something, you know, to the chatbot, and then I said something to the customer service representative, and then there was this other email, and now I'm talking to another one who doesn't know what the other 2 said, and the AI doesn't— like, this is just so annoying, right? This is ridiculous. It shouldn't have to be that way. The company should have an AI who's, you know, is not active most of the time, is just sitting on a disk or whatever somewhere, whose entire job is to interact with you. and knows everything that you've ever done with the company, knows your entire history, knows your preferences, knows how to serve you. And then when you want to talk to the company, you talk to the company. So to, you know, somebody said, and I agree, that technology is mature when it becomes invisible. And the way I think this is going to happen with AI is that I'm going to be talking to Walmart, and I'm going to be talking to AI. Maybe sometimes it says, well, let me, let me, uh, connect you with this person that works for me, Walmart. Right? But most of the time you're interacting with Walmart or with Microsoft or with Google or with IBM, and that's actually what is better for everybody, for the company. Like, talking to that I think will be stickier than talking to Erica or Cortana or whatever. So companies will want that. It also introduces more accountability. It's not like— no, this is the company, so you can treat them as being the company. So I think that's, that's actually what should happen. So it's a very brief tweet And it may seem just like something that I typed randomly, but there's a very serious point that I think companies should pay attention to this and really start thinking about having their chatbots just be the company and identify themselves as the company, not just, I'm some chatbot. Let's see, what other questions do we have? How can we contact you after this lecture? Um, send me email. My email is public. Again, my email is pedrod@cs.washington.edu. But if you don't want to remember that, Google my name, you'll find my homepage, which you can also find as pedrodomingos.org. And on my homepage, uh, you know, there's my email. And usually the best way to reach me is by email and, you know, ask me questions and so on and so forth. You can also just, you know, if you're an X user I'm on X daily, so you can just ask me questions there, particularly if you think they're questions that would be of interest to other people as opposed to just yourself. Ask them. I can't guarantee that I will see them there, so email is more reliable because, you know, I get so many replies and comments and whatnot. But hopefully I will see it and then answer it for you and for everybody else as well. Somebody's thanking me for being a killer. You know, thank you. I appreciate that. Keeping questions about logic. Any implementation of C.S. Peirce related to analogy on AI and his abduction concept, which to me sounds similar? Ah, great question. So for those of you who are not familiar with, with, you know, logic and math and symbolic AI, Pierce actually— I mean, let's focus on Pierce because that's who the question is about. But there's more than— and again, this is very important and very neglected today— there is more than one type of inference. So deductive inference is what we usually associate with logic and what famously LLMs can't do. They can fake it sometimes, but they can't really do it. And that is very, very clear right now. There is also inductive inference, which is really what machine learning is. Induction. So deduction is going from the general to specific, uh, like, you know, all philosophers are mortal, therefore Socrates is mortal. Induction is going the opposite direction. If Socrates and Plato and Aristotle, etc., are all mortal, maybe I can induce that all philosophers are mortal. And this is very, very powerful, right? But what, you know, traditionally, going back to Aristotle, these were thought to be the 2 main types of inference. Deductive and inductive. What Pierce pointed out is that there's also another type of inference that is actually extremely important, which is abductive inference. And abductive inference is actually something that we do all the time. And, you know, there's more than one way to introduce it, but— and again, I, I, I touch on this in the master algorithm, uh, but maybe the most intuitive one is the following. What we off— what we typically do in deductive reasoning is go from the causes to the effects. If this cause is present, and then this effect will be present. You know, like you caught the COVID virus from somebody— cause— therefore you will get COVID— effect. Like the cause where you were in touch with this person, the effect was you got COVID, right? Very standard reasoning. But that is actually not what doctors do most of the time. What doctors do most of the time is they go the other way. They go against the arrow of causality. They say like, you have these systems, these symptoms, can I infer what caused them from your symptoms? Do you have COVID? Do you have the flu? Do you have something else? And as you can see, this is extremely important. It's not deductive reasoning, it's unsound, right? You can't, you can't guarantee that your conclusions will be correct as you can in deductive reasoning. It's certainly not inductive reasoning, although you people sometimes think that you can shoehorn it into inductive reasoning in some sense, but I don't think it's It's that helpful. So it's very important, and, and, and we do it all the time. In particular, maybe this is the most important and most relevant today, I would say, and many people would agree, others might disagree, you know, there's— there are many debates about this among linguists and, and NLP researchers and whatnot. I would say, and many would agree though not all, that when you are listening to me, for example, right now, or when you are reading a book like when you read 2040, your brain is continuously doing abductive inference from what the sentences are saying to what you infer is true. Again, not, not deductively— sometimes maybe deductively— but like, for example, if you look at Sherlock Holmes novels, a lot of the reason that he's doing is actually abductive. So your brain is constantly inferring these other things that are probably true given what is explicitly being said. But the key is that like you can't— we, we do this so intuitively that we take it for granted. And then you can ask me a question about something that was inferred from the text, and people have studied this, and then psychologists and psycholinguists, and you often can't even remember if it was actually in the text or just— you, you often mistake something that you abductively inferred for something that really was in the text, uh, but, but it wasn't. So, so it's— it goes that deep in us. And I would say that if LLMs aren't doing abductive reasoning, they're not really understanding natural language. And again, you could— this is a question that deserves more research than it has had. But, you know, if you remember nothing else from this, remember that abductive reasoning is very important. Pierce was the one who first came up with it. These days, it's a little bit unfairly neglected. But again, like many things in AI that come and go, I really wouldn't be— I was talking with someone who does, you know, large language models not long ago. And I mentioned this whole line of things and the fact that what I'm doing right now, one of its best uses that I can see is precisely doing this kind of abductive inferencing in reading. And she was extremely excited. So stay tuned. What else? Somebody tried Alchemy, but the site looked outdated. Yes, um, I agree that site is outdated. Uh, you have to remember that Alchemy is pretty old by now. Um, you know, again, it's not outdated in the sense that it's, you know, none of what it does I think has been superseded. But in terms of the underlying, uh, you know, repositories, blah blah blah, yes, it doesn't use GitHub and whatnot. You know, if you want to volunteer to, uh, some other, some of the other things in there are, are some of the other systems that we point to on the website. people who live more in GitHub land might, might be more comfortable with. But yeah, um, there's a lot that could be done, I think, not, you know, in terms of modernizing Alchemy and whatnot. Honestly, it's a question of priorities, right? Uh, I would love it if, if I could fork one of myself to just do that full-time. It would be a full-time job. But between doing that and spending my time pushing the frontier for the next thing, which of course MLNs are not, you know, they were I think a significant step, but they're certainly not, uh, you know, the final answer. As I've been saying, I think it's better to do the latter. And also in terms of students, um, uh, it's as, as doing that type of work is useful but is not what, what gives you, um, for better or worse, the way you get recognition and papers published and jobs in academia or even in industry to a large extent is not by doing things like, you know, maintaining the repositories and whatnot. There is a large value to that, uh, for sure, because, you know, having people use your stuff and making it open source and whatnot— I'm, I'm a big believer in that. And of course, Alchemy is an example of that. But there has to be coupled with having produced new research. So just doing that for the Alchemy stuff that's already there, uh, you know, I, I hope somebody does it. I will fully support them if they do. But, you know, it's not quite at the top of the stack of things for me to work on. So let's see, what other questions do we have? What is the difference between Markov and Monte Carlo? I keep thinking they are the same. So Markov, the difference is actually very simple. And I understand that the confusion is natural because of the name. both have Markov in it. Little, little did Andrei Markov know that his name would wind up on all these things like hidden Markov models and Markov chains and Markov logic and Markov chain Monte Carlo. Markov logic networks are a, are a, uh, are a representation. They're a class of models. They're a language in which they're like first-order logic or neural networks. Markov logic is called Markov logic because it is a combination of Markov networks which themselves are a generalization of Markov chains and hidden Markov models, and logic on the other side. Markov chain Monte Carlo is an algorithm. So there's Markov chains. Markov chains is actually what Andrey Markov, a Russian mathematician from the turn of the century, he actually built the first Markov chain. And he used it to learn from Andrey Rublev, a famous Russian epic poem/novel. And he built an n-gram model, in essence. So he's actually a predecessor a very direct way, but many steps removed of today's LLMs. And so that's, you know, now we call that Markov chains. And this is actually— it's not just AI people, you know, it's very well studied in math, used in many different areas and whatnot. So there's Markov chains, there's hidden Markov models, which now have hidden state, there's Markov logic networks, which are much more general in turn, and then there's Markov chain Monte Carlo, which is a class of inference algorithms for doing inference in not just Markov networks, but Markov chain Monte Carlo, or MCMC for short, is one of the most widely used algorithms in all of science. Every time you have a probabilistic model that is too hard to do analytical inference on, or for some of the simpler approximation schemes like variational inference or, or, you know, simple forms of sampling-based methods, for almost every problem, Markov chain Monte Carlo is the method of choice. In fact, Numerical Recipes in C, which I have, you know, behind me on the shelf right here. They actually say between knowing— having a whole shelf full of books on doing inference and, and just knowing Markov chain Monte Carlo, they would pick Markov chain Monte Carlo. So yes, they're right, it is one of the main things that we've used in Markov logic. Not the only one. We actually have, uh, in, in some methods, uh, that use sampling but coupled with logical inference We call it probabilistic theorem proving. We have a paper about it. So I recommend if you're interested, you take a look at those things as well. Okay, we are, you know, our available time is sadly running out, but let me answer a couple of more questions and then conclude. I have some investigations. Would you like I share it by email? Please. Please send me emails with your brilliant ideas and I will publish them in my own name and take all the credit, just like Laplace did, right? Apparently that's what he did. Young mathematicians, because back then there were no papers, they sent him their brilliant ideas and he just published them. And when somebody accused him of doing this, he said, they were lucky that I published their ideas under my name. That way they got to be known. So, you know, please Send me your ideas by email and I'll make the best use that I possibly can of them. Um, this is a really good question to end on. Do you think AGI is already here? How do you think we should define AGI? I think we should compare to the average person, which the, which the chatbots are much smarter. So great question, and indeed, uh, this is really the overarching question in, in, uh, um, In AI, I would say in AI today, and certainly very much in 2040 as well. The first part of the answer is easy. AGI is not here. Having said that, the whole problem with the concept of AGI is that intelligence is not one thing. Intelligence is hundreds of different things. And in some of those, computers are already better than humans. And have been for a long time. Computer used to be a job description, and what computers did was they did additions and multiplications. And we used to be considered back in the 1920s, say, that it took a high level of intelligence to be a computer, a human computer. And now computers just do that, right? So in that regard, computers have blown humans away. We're way into superintelligence. But to take a more relevant example, medical diagnosis. We can already do medical diagnosis better than doctors. In fact, we could do medical diagnosis better than human doctors before the age of machine learning. There were expert systems, famous, the most famous, some of the most famous expert systems of the '70s and '80s. They became famous because they diagnosed specific illnesses better than human doctors. And as we progress into the future, the same thing will happen. What people should keep in mind is there's a very broad front And of course, AI keeps advancing along this front very unevenly and very unpredictably. And of course, it only moves in one direction. When it becomes better than humans at something, humans never catch up. But typically what happens is you need a combination of the human and the AI to get the best results precisely because the AI is better at some things than others. And so what's going to happen, unlike a lot of people who seem transfixed by this today, is not that AI will be reached on August 20th at 8:00 PM, meaning right now, It's not going to happen. There's not going to be a moment where AGI was born or AI surpassed humanity. It's going to happen in different areas at different times. We're going to have quite likely 10 years from now AI that is spectacularly intelligent in ways that we don't even imagine right now. Like, for example, people weren't imagining AI as being creative a few years ago. I kept telling people creativity is easier than reliability and people were like, What are you talking about? And now it's obvious. So more stuff like that is going to happen at the same time that AI is doing mind-bogglingly smart things, maybe solving physics, unifying general relativity and quantum mechanics. I really think maybe AI will do that where the physicists— again, probably in collaboration with real physicists, but where physicists are stuck, AI— where scientists are stuck, AI can do a lot more than they can, while at the same time still being spectacularly, shockingly stupid and able to do things that a 2-year-old child is. So in many ways, the question of AGI is a false one. Now, I don't— some people say just like, don't even think about AGI. And in fact, AGI for mainstream AI has a bad name as something that a bunch of wacky people came up with. And, you know, there's no need for the term because AI, you know, AI is supposed to be general, right? Things got specialized and over-specialized, but the goal of the founding fathers, you know, when they coined the term and started the field was always general intelligence. So I do think there is that— there is value in the concept of general artificial intelligence, or human-level artificial intelligence, if you will, because it gives us a rough guide. So for example, Until we have housebots, we don't have human-level AI, right? You can usefully define AGI as we want our— and this is how it's usually defined— as we want AI that can do everything that an average human can as well or better. And I think this is a perfectly reasonable goal to aim for. And just keeping that in mind can be very healthy because it, for example, saves you from getting a little overexcited about all the video generators, and, and which are great, right? But what they do is video generation. Or about all the large language models, which again are great, but what they do is not AGI. And in particular, there's no housebot. Until there's a housebot doing my dishes and making the beds and cooking dinner, we do not have AGI. I think we need to keep making progress towards that, and there are some very, um, you know, big problems to solve before you can get there. But I'm optimistic that we will. So to conclude, thanks for joining me on the livestream. As I mentioned earlier, the purpose is to celebrate the publication of 2040, my new book. If you're curious about 2040, you can read the first chapter for free at 2040novel.com. You can just, you know, if you don't want to remember this URL, Again, you can just go to my, you know, actually these days if you even just Google 2040 Novel or 2040 Domingos, you will find this. But one way or another, you can go to 2040novel.com and read the first chapter for free. You can also, this actually I just found out like 30 minutes ago, the Look Inside is now active on Amazon. So you can actually just read most of the first chapter there, almost the entire first chapter. minus the last surprise that sets up the following one. So I would still read it on 24novel.com if I were you, or if you're interested. If you like it, I hope you will buy the book and read it and enjoy it. As I said, reading it was some of the most fun I've ever had. It's a quick read. I hope you'll enjoy it. And also, I hope that besides being enjoyable, it also gives you a better perspective on AI, the AI hype, tech and the whole political moment that we're in, which of course the novel also gets into. And if you like the novel, spread the word so that other people can have fun and maybe benefit from what it reveals about the world of tech and about AI and culture and politics and so on. Thank you everybody, enjoy the rest of your day. Thank you.