#210 Pedro Domingos: Exploring AI’s Impact on Politics and Society

0:00This is a natural mistake that people make by analogizing between AI and humans. The only intelligence that we know is human intelligence, animal intelligence. So once we see a computer behaving in ways that seem intelligent, we can't help but project onto the computer all sorts of human characteristics that it just doesn't have, like ambition and the will to power and emotions and consciousness, etc., etc. None of this is there. Hey, it's time to make online shopping personal. Bloomreach personalizes the e-commerce experience, unifying real-time customer and product data to understand what customers really want.

0:42By connecting all that understanding to every channel, the e-commerce experience becomes limitless. Amplified by Loomy, Bloomreach's AI for e-commerce. This creates endless new paths to purchase, greater profitability, and fast business growth. So give Bloomreach a try. Hi, I'm Pedro Domingos. I'm a professor of computer science at the University of Washington in Seattle. I am an expert on AI and in particular machine learning. I'm originally from Portugal, got my PhD at UC Irvine back when hardly anybody else was doing machine learning. This was in the 80s. I've done a whole bunch of things in this field since.

1:31I've written a couple of books, one that called The Master Algorithm, introduced machine learning to a broad audience, turned into a surprise bestseller, and most recently, 2040, A Silicon Valley Satire, which, as the name implies, is a satire of AI in the tech industry and the politics and culture around it. Yeah, I enjoyed the book. It reminded me a little bit, when I was in high school, there was a book by a writer named Tom Robbins, Even Cowgirls Get the Blues. It's kind of a crazy satire with a lot of wild characters, but with an underlying philosophical theme. Can you, I mean, I have a lot of questions about the book.

2:20For listeners, it's at a very high level about what happens when AI has advanced to the point that we have an AI candidate for president and all of the crazy things that devolve from that. But without going into the plot so much, I'm interested. The book is largely around the theme of control and lack of control and the danger of losing control of AI. Understanding this is a satire, but that's for two years now, that's been in the wind. And I frankly think it's exaggerated, those fears. And I think regulation is moving along quickly, maybe because of the noise made by the research community. But it is moving along quickly.

3:23So how much of the – it's satire, but is it poking fun at those concerns? Or are you raising those concerns? Well, like all satire, it has an underlying serious intent. So 2040 is not about 2040 any more than animal farm is about animals or farming. It's about today. It's about 2024. And like a lot of classic science fiction, what it does is it extrapolates current trends to the point of absurdity. the thing that's kind of funny and maybe alarming in some ways is that, you know, it's not, a lot of this stuff is not that far away. In fact, a lot of people's reaction to the book is, yeah, this seems so far out and over the top, and yet it's oddly realistic, like it could happen anytime.

4:12And in fact, I conceived of this book a few years ago, but, you know, then I decided I really need to publish this because before it becomes outdated, because so much of the stuff in it was coming true. So it's a short, quick read, but it actually covers a lot of things, right? So every element of the book is there for a reason. It's there to illustrate something about today's world of tech and AI and whatnot. And as you say, one of the key elements is what I think are the misguided fears about AI versus the real ones. So part of what the novel tries to illustrate with PreziBot, who's this AI candidate for president, you know, just think chat GPT runs for president.

4:53That's basically what it really is. And you can already see, you know, the jokes coming one after another is that the real danger is not Terminator or some evil AI taking over and exterminating us. The real danger is that the AI is imperfect. The AI screws up, it hallucinates, it has bugs, it needs fixed. The even bigger danger that the humans developing and trying to control the AI are flawed humans, like we all are, and they screw up, they get into conflicts, and disasters happen without giving any spoilers. Those disasters are the result of the interaction of the flaws of the AI with the flaws and the intentions of the various humans, which I think is the real problem.

5:36And of course, what is always the real problem with technology is just that on AI, it might be on a bigger scale than ever before. Yeah. But where do you lie on that safety debate, you know, between Jeff Hinton and I don't know who. Jan LeCun. Jan LeCun, yeah. They're the principals in much of the debate. Where do you lie on that spectrum? I'm very much in agreement with Jan on most of this and in disagreement with Jeff. I have a lot of respect for Jeff as a researcher, but as someone to take advice from about real world issues, he's not the first person I would turn to. And in fact, he's very honest about it.

6:19He says he never actually worried about any of these AI alarmist scenarios until recently. It was only when ChatGPT came out that he started thinking, oh, wow, you know, superintelligence is almost here. And then what's what it's going to do to us? And I think him and a lot of other people like him, they're making two mistakes, at least. One is that superintelligence is not almost here by any means. And I think in the years since then, it's becoming increasingly clear that no, this AI still has a long way to go, number one. But number two, and maybe less obvious, but in the longer term, more important, superintelligence is no extinction danger to humanity.

7:01This is a natural mistake that people make by analogizing between AI and humans. The only intelligence that we know is human intelligence, animal intelligence. So once we see a computer behaving in ways that seem intelligent, we can't help but project onto the computer all sorts of human characteristics that it just doesn't have, like ambition and the will to power and emotions and consciousness, et cetera, et cetera. None of this is there. And it won't be there unless we put it there. In particular, I think what we can and should and will do is we're going to develop AI. Remember, AI is under our control.

7:40We're going to develop it as an extension of our intelligence. It's not some other creature that is going to come and attack us. That is just a fantastical sci-fi scenario that, unfortunately, Hollywood has promoted heavily because it makes for successful movies. People think of Terminator when they think of AI. If after reading the book, what they think of instead is Prezi, but then it has already achieved its aim. Yeah, yeah. And the whole Terminator scenario, because people talk to me about this all the time, they're talking about embodied AI and they're really talking about advanced robotics, which have a really long way to go.

8:25I mean, there's some incredible things happening and things have moved very swiftly in the last year or two, but it's going to be a long time before you have a robot that can navigate the messiness of the world, let alone take action in the world in a way that would be threatening to humans. But that's my view. But the question of superintelligence, I mean, there's a lot of confusion, particularly in the public space about that, because, you know, there is a lot of superintelligence already in AI. it's not this monolithic thing that we're going to reach someday that when I think that that where suddenly you wake up one morning and there's this AI system that's infected the internet and it's smarter than humans I hear this from AI researchers in the alarmist camp that you know it'll be smarter than you you won't be able to stop it because it'll be smarter than you.

9:38But there is a lot of super intelligence. I mean, I would argue that, you know, that a lot of the, like, for example, ChatGPT within the bounds of its hallucinatory tendencies is super intelligent. I mean, it has access to more information than I do. it's really a question of how reliable it is. But let's move on from that. You talk, you're very critical of the Silicon Valley culture in the book. Yeah, can you talk about that a little bit? Because I'm on the East Coast. I'm not steeped in Silicon Valley culture, but it amazes me that you've got this concentration of billionaires in and around San Francisco that are creating superhuman technology, even if it's not consistent or across all domains.

10:57Yet they live in this, I'll probably get in trouble, but I lived in the Bay Area for a while, but this real cesspool of crime and homelessness and social dysfunction. and to me that you know i keep on thinking uh what you know i should email elon the musk and and jeff bezos i mean between them they could solve the homelessness problem like that uh with a fraction of their wealth uh so yeah can you talk a little about a little bit about that dysfunction so 2040 as the subtitle says is very much a satire of silicon valley and the culture of Silicon Valley around tech and around other things. And it has several elements.

11:45One of them is what you're referring to, which I'm always struck when I go there these days. There's such a contrast between the utopian high-tech side and the dystopian. In San Francisco, you see the self-driving cars go by on Market Street, or you're in one. And then behind them, you see these homeless people injecting. and lines sprawling on the ground and piles of trash and poop. And it's like it's you know, it makes a new romancer look like a children's story. It's like, how could this happen? And again, part of the goal of a satire and this one in particular is to say like, you know, if if we let this go on, like, where is this going to end and how do we stop it?

12:30And so one of the elements of the satire is also there's there's this giant company called HappyNet, which is the ultimate everything app which exists in China and every you know US tech billionaire wishes that's what they have it's an app that does everything for you and there's the headquarters which is the super high tower which is of course a caricature of the Salesforce tower you know soaring high above the streets in French Francisco and the CEO of this company looks down on the masses and you know he says they're just pixels and I can see the whole picture yeah underground underground, under San Francisco, there's a vast data center.

13:08It's there to minimize latency and blah, blah. But then what Ethan, the protagonist, Ethan is the CEO of this Kumbaya startup that created Prezabot. And at least in the beginning, he's the guy who nominally controls it. Then other things happen. But at one point, he discovers that this underground data center is full of homeless people squatting there. Because life is better there. It's not as cold and it's more sheltered. And there are all these guys, there's robots. So that the surveillance cameras will not notice them. And again, there's multiple layers of commentary in this. But of course, part of the question is like, how did we wind up here?

13:54And I think Dave Newell, who's the CEO of this company, Of course, he's inspired by the Elon Musks and Mark Zuckerbergs and Jeff Bezos and whatnot of this world. And you say that they could solve the homeless problem with a stroke of the pen. Actually, it's much more complicated than that. And again, I try to not, as much as it's a short book, I try to not oversimplify. And the homelessness problem in San Francisco and other places, it does not exist because of lack of money to solve it at all. Yeah. Even without the billionaires, you know, we have put billions and billions of dollars in the problem when it gets worse.

14:31It exists because I mean, again, we could go into that, but it's more of a political problem than a technical problem. Part of what I spoof very much in this book is that the technical people's naive notion that all these problems have tech solutions. Right. The first chapter is called Optimize America with an exclamation mark, because that's the slogan of the PreziBot campaign. and then they very quickly discovered that things are a lot more gnarly than that if you will and and they are yeah yeah and and when i said uh they they could solve it i'm i'm i was i wrote a lot of when i was at the times about um the the housing first policy i don't know if you've read about it this guy sam sam barras who uh developed that during the coach administration in new york and and it's been implemented in some places very successfully and the the thing that really prevents it from being implemented in larger cities is coordination between and among institutions.

15:44You just can't get everyone on the same page, and as a result, you end up with a fragmented system, and each one's trying to address a different aspect of the problem. But anyway, that's for another conversation. I mean, my view is if they would build housing for the homeless, that would be a leap toward a solution. And certainly the tech billionaires have that money. You can't get it politically because the politics are so divisive that you can't come to a consensus. But yeah, and Ethan lives in this tiny little flat, which I thought was funny. I mean, that's more of a New York trope to me than San Francisco.

16:41I mean, San Francisco apartments are actually larger than New York apartments. But, yeah, so. By the way, that was actually inspired by reality. In particular, the apartment that he lives in was inspired by a piece in 1843, The Economist's Lifestyle magazine, where they talked about, and they showed various ways in which San Francisco is so expensive now that it's becoming like Manhattan. And of course, it's also a tech mecca. So people are coming up with all of these ways to make ever smaller apartments have all the things in them that you need, including beds that rise into the ceiling and so on and so forth.

17:20And of course, that's what I'm making fun of in Ethan's apartment. Again, it's completely real, that part. Yeah. Can you, I mean, the idea of of technology as as a decision maker is something that you know interests me and geez I gotta fly down here and is you know with with agentic which is a word I would never have used a year ago no one would have used it a year ago that's right models and and with these reasoning models with Q-Star or Strawberry or GPT-5 or whatever you want to call it, that next generation, the idea of models that can make decisions and take actions is becoming closer to reality.

18:23To me, that's a very positive thing. in the book it's it you know it's it's sort of the the door to all the the the craziness but how do you feel about that about and i have a question too i had a conversation the other day about truth and ai is a truth engine which uh was fascinating but but how do you feel about the potential for agentic reasoning AI models to make decision-making, to make decisions, maybe not autonomously, but certainly to aid in making decisions with humans to avoid some of the foibles that humans are prone to. AIs make decisions. That's what they've done since the beginning.

19:25And there's a whole theory, which you study in AI 101, about how to make decisions optimally and how to do search for the best decisions and et cetera, et cetera. This is what classic AI is about. If you went to an AI conference in 1980, it was all about automated reasoning and decision making and planning and whatnot. Some of that is not forgotten, unfortunately, but this is what AI does. So AI isn't really AI until it's making decisions about something. Now, the question, the interesting questions are, number one, decisions in service of what goals? And this is one of the crucial things is that the difference between AI and an audio computer program is that the computer program at some level doesn't make any decisions.

20:05It just does what you told it to do in one situation or another. The difference with AI is that it does make its own decisions, but they are guided by the objectives that you put in. An AI system can never do anything that is not, by design, subserving that objective. And so this is unintuitive, and I understand people's fears, but you could actually have a super intelligence that is vastly more intelligent than us, but you don't have to fear it because it's at our service. The intelligence is just doing our bidding. Now, there's a second part, which is what happens when the AI starts actually taking actions in the real world, like, say, a robot versus chat GPT.

20:46Now, of course, there are many risks that weren't there before. But again, the bigger risk is that the AI will do damage, like a self-driving car might run somebody over, right? The self-driving car is deciding whether to go left or right or stop or accelerate, and it can make mistakes. So those are the things that I think we really need to fear, and I think we see some of that. And it's interesting seeing how this is shifting, even in just the last couple of years, is that we have a tendency, and the book very much illustrates that, to prematurely put faith in AI as being more capable than it really is.

21:20Because, again, the demos are very impressive. It looks very intelligent. We're like, oh, wow, this AI is going to do all these things. And then it blows up in your face. So people are not, or you can jailbreak it and manipulate it. So the problem is not the AI manipulating the humans. It's very much the humans manipulating the AI. And I think even from just a year ago to now, by interacting with things like ChetGPT, people are coming to see, which I think is very healthy, AI as it really is, and what the real warriors are versus the imaginary ones. Yeah, yeah. I had this conversation. There are a couple of things I wanted to ask you that were sort of straying from the book a little.

22:02So I've talked to people who believe that AI will be an arbiter of truth. And I don't believe that at all because truth is so subjective. I think the best that could happen is with a very carefully constrained model, you could produce evidence-based consensus on any issue based on the training data. But how do you feel about that, about using AI to settle questions where the truth is unclear or hidden from a human observer? I wouldn't use the expression AI as an arbiter of truth because arbiter is a very dangerous word. I don't think AI can or will or should have the final say on what is the truth.

23:09Having said that, AI can and should, and I hope will be a fantastic tool to help us get closer to the truth. Very much so. Because it can gather and process evidence on a larger scale than any human can, in particular, combining, again, in 2040, there are two versions of PreziBot. The first version is ChatGPT. The second version, and more interesting one, is a crowdsourced AI that in real time takes input from people and decides what is true and what to do and where to go and whatnot. This, I think, is the real role of AI, is to increase our collective intelligence and get us closer to the truth.

23:50Now, you can never get to the truth because this is one of the basic lessons in AI. It doesn't matter how much computing power or how much intelligence you have. Even just to play the game of chess optimally, you need to compute it bigger than the universe. So the idea that some AI, again, this is part of people's failure of imagination, like, oh, it's more intelligent than us. Therefore, it's infinitely intelligent. This is a complete mistake. Between our intelligence and God, right, there's a vast, vast, vast gap. And yes, AI will go beyond us, but AI is not going to do miracles. Now, unfortunately, what is happening right now is that AI, instead of getting us closer to truth, is actually arguably getting us farther from the truth.

24:31Because what happens with all the generative AI is that it makes stuff up. That's what generative means. It means making stuff up. And so what it's creating is like these deep fakes. And even like the problem with ChatGPT, which you alluded to, that it's unreliable. Unreliable is that like ChatGPT, these LMs literally have no notion what is the truth. They just generate text that makes people happy. And that is similar to other texts that people have generated. And that, in fact, does not get as closer to the truth. It's actually completely ungrounded. And not completely, because that language is, again, at least some of it, right, comes from, say, good reporting and so on.

25:12But most of it doesn't, and ChatGPT can't tell the difference. Yeah, yeah. But, and right now, it's purely probabilistic. So if there isn't, you know, a probability curve that centers on sort of consensus about an issue, it'll come up with its own, you know, based on whatever distribution is in the training data. And if you had clean enough training data, presumably that covered enough of the world or intellectual thought, it would give the consensus on any issue. I mean, you were talking before about the second iteration of Prezibot, which is crowdsourced. I mean, the problem with the training data that is being used in a lot of these models, it's effectively crowdsourced.

26:41But the crowd is unreliable because there's probably more unintelligence within the crowd than there is intelligence for the AI to draw on. But I'm fascinated by the idea that you could either through careful curation of training data come up with a system that could reliably answer questions that are facing humans or governments in real time. Is, you know, Prezibot and Prezibot 2 aside, is that something that you think is possible? It's possible. And here's the big difference, or at least one big difference between Prezibot 1.0 and 2.0, is that Prezibot 1.0, like ChatGPT, it is learned from past data.

27:48And that data, by definition, is stale, not only because time has passed, but because you are not like the chat GPT right is now answering something that it hasn't answered before it's in a new dialogue press what is in a new situation and then what it has to do and this is really the crucial point is that it has to generalize to the new situation from the ones that are in the text and the more different they are the worse it gets right this is also true of humans when humans are much better at generalizing from their experience to new situations than the best machine learning algorithms that we have today.

28:25And hopefully we'll get better algorithms. That's what I work on, you know, for a living, if you will. But that aside, the difference between the crowdsourcing, you're right, you know, like it's also crowdsourcing in the sense that it came from people, but this is crowdsourcing in real time, meaning like when there's a situation, what PraziVa 2.0 does, and actually technically today this is possible, it's like it listens to millions of people saying what they know and what they believe and what should happen and making suggestions. And, you know, there are several examples of this in the book.

28:56And so this can actually work much, much better because it's taking input from our human intelligences in the moment about the actual situation, about doing some very far-fetched thing that will be unreliable and prone to hallucination and whatnot. And I think there's a lot to do to perfect that, but this is very much a possibility and a very good one that is being very neglected right now. And it's not just at the level of, you know, a president of the United States, the level of a company. Right. Again, the company that created Prezibot created as a demo to then tell CEOs to companies. Right.

29:32And if you think about the job of a company or the CEO of a company, it's a lot of the decisions are very mundane. But it's screwing up all the time because it needs to do this for the customer. It needs this knowledge that actually somebody else has. But I don't even know that it has. And I'm handing it off from one customer service rep to another, to the chatbot, to like there's this L. that people live in, and AI can overcome that. Yeah. Do you think that the risks, that AI governance, and we've seen, as I said, a lot of activity, you know, the Europeans are working hard to protect the consumers, and the U.S.

30:16is relying more on existing institutions. But do you think that the governance is keeping up? Because one of the themes in the book is that this race for innovation in Silicon Valley is throwing us blindly into danger. But I kind of think that the regulation is a pretty fast follower. What do you think?

30:56Where to start? One of the central plot elements in 2040 is PreziBot's panic button. The panic button is the kill switch of the AI. And this is a real thing. People have proposed having kill switches. Their research papers aren't having kill switches. I would not be surprised at all if a bunch of regulations in various countries demand kill switches. And so President Biden has a kill switch, which is in the hands of Ethan, the CEO of the company. And then Ethan loses the kill switch. And then the fun begins. And of course, this is, again, there is a fun element to that, of course, of the chaos that ensues and the people fighting over it and so on.

31:35But this is making a real point. And the point is these measures that you put in place to make AI safe, if you're not careful, make it less safe. And kill switches are a great example of that. Right. Kill switch is like the ultimate guarantee against Terminator, but actually it's nothing of the kind. It's just another vector by which chaos can come in. And all of these regulations are, let me put it this way, the real problem in AI is who controls it. Yeah. and the kill switch and the kill switch in a way is the embodiment of that in in in 2040 and then the realization that he has is instead of having people fight over the kill switch i should give everybody one because this is a democracy so the kill switch originally is is an app that you know arvin the cto and his friend uh coded up quickly uh that literally has a red button that you push or that you can use it to talk you know press you can talk through prezibot what they do right because they're very short of time is they let everybody, they call it Prezi app, they create Prezi app and anybody can push the button and talk into Prezi app.