A Dose of Engineering Realism Over AI Hype

[00:00:00] Speaker A: Welcome to ID the Future. I'm Andrew McDermott. Today's episode comes to us from our sister podcast, Mind Matters News, a production of the Discovery Institute's Walter Bradley center for Natural and Artificial Intelligence. You can learn more about the show and access other episodes at mindmatters. AI. Greetings and welcome to Mind Matters News. I'm your real intelligent host, Robert J. Marks. Our guest today is Dr. Donald C. Wunsch II. I'm Robert J. Marks the Second. They named me the Second because they didn't want to name me Junior at an uncle named Junior, and we called him Uncle Junior until he died when he was 90 years old, so they still called him Junior anyway. Dr. Wunsch is an endowed professor of electrical and computer engineering at Missouri University of Science and Technology, where he's the director of the Kummer Institute for Artificial Intelligence at Autonomous Systems. He has a long title. Okay, you ready? Take a deep breath. He is the Mary Kay Finley, Missouri Distinguished professor of Electrical and Computer Engineering. Don is a widely respected voice in artificial intelligence and computational intelligence. Dr. Wunsch has spent decades working on the intersection of engineering, AI, and real world systems with research spanning neural networks, adaptive systems, machine learning, and AI engineering. He is a fellow of the ieee, which is pretty impressive since IEEE only admits fellows. I think it's 0.1% of its membership every year, so it's a very high bar to reach. Very prestigious title. What sets Don apart is his insistence on engineering realism over hype. While much of today's AI conversation centers on speculative futures and artificial general intelligence, or AGI, his work focuses on what AI can actually do today and the very real risks and responsibilities that come with it. He wrote a recent paper that we're going to talk about today. The title of the paper is Artificial General Intelligence is Nowhere Near Artificial Specific Stupidity Is Already Here and Policy Implications. His paper challenges both the techno optimism and doomsday narratives, arguing that the greatest dangers of AI arise not from autonomous machines, but from human misuse, poor policy, and concentrated power. We're delighted to have him with us today for a thoughtful and grounded conversation about AI autonomy and what engineers in society should be paying attention to right now in the area of artificial intelligence. Don, welcome. [00:02:44] Speaker B: Thank you, Bob. [00:02:45] Speaker A: You know, we go way back, and one of the things that you and I did last century is, is we actually bought artificial intelligence and IEEE into Asia. We had two meetings. One was in Rostov on Don, and that was an IEEE meeting. And if you remember, the funny thing about that was IEEE requires all of their papers be delivered in English. So we had all of these speakers come up and they were speaking in broken English because they were Russian and all of the listeners were native Russians and they were trying to understand the English. So it was kind of hilarious. But then we, after that, we had a meeting in Beijing, China. And this was a big thing for ieee. We actually brought IEEE and artificial Intelligence, Computational intelligence into China. But one of the things you did which was very adventurous, I don't know if you would do it now, but you took the Trans Siberian Railroad from, From, From. From Russia to Beijing. And that must have been an experience for you. You lost some weight there, didn't you? [00:03:54] Speaker B: Well, I lost some weight before I got on that train because I had a problem adjusting to something that I ate in Rostov on Don. But, yeah, it was fun to see you. And Vitaly Din and Barkowski and Rostov on Don, we met him when he came to IJCNN 91 in Seattle. And between 91 and October of 92, when this conference was held in Rostov on Don, something remarkable happened. The collapse of communism in the former Soviet Union and the collapse of the former Soviet Union. And so we had a captured seat to those events as we got there just after Yeltsin had stood on the tank. And it was a really exciting time to be visiting Russia. It was my first visit to Russia, and you might remember that I got stranded in Moscow for one night and stayed with the driver who took us between the two major airports in Moscow because my baggage had been lost. Aeroflot promised to deliver it to Rostov on Don. And our host told me, you're better off staying with the driver than going to your hotel and Rostov tonight and make sure that you get your bags and personally see to it that you get them. But anyway, we went to Rostov on Don, had our adventures there, and you went back. And I had taken several weeks of leave of absence to visit places because there had been an AI winter in the United States. It wasn't really an AI Winter. It was. Was a neural networks winter, that neural networks had been hammered down for a period of time by people who thought we were wrong. And these were the kingpins of AI, who thought the neural network people needed to be defunded, that we were barking up the wrong tree and needed to do things in symbolic AI and other things. [00:06:01] Speaker A: Yeah, they were proponents of expert systems, which is where you take expertise and kind of translate it into code, isn't that right? Right. [00:06:10] Speaker B: And they were not doing garbage. But they were mistaken they were. They were mistaken in thinking that we were doing garbage. And we won. [00:06:18] Speaker A: We did win. [00:06:19] Speaker B: We sure won. We won big time. So now many of the same people claim that they were in our favor all along, but they weren't, and we know it. But there were some of the kingpins of AI who were indeed with us the whole time, and one of them was Yann Lecun, who I will mention during our discussion, he was one of the Neural network people all along. Anyway, it was an exciting time to be in the field. Sometimes we had to fight for our right to party. We had a good time working on neural networks, and it paid off. Sometimes you have to have quite a bit of tolerance for being told you're wrong if you believe you're right. And we had to put up with that for many years. And those efforts paid off for a lot of people, and a lot of people abandoned it when the slogging got tough. But for those of us who didn't, there were some nice rewards down the line. [00:07:25] Speaker A: There were. And that's good news for us, I suppose. I remember when we were in Rostov on Don, and perestroika was happening, all the Russians were kind of broke. You actually bought the President of the Russian Neural Network Society to the United States. I think you paid him as an ra, didn't you? [00:07:45] Speaker B: So, yeah, he was willing to come for a modest initial salary, and I didn't have the funding to pay him anywhere near what he deserved because I was an assistant professor, and this guy had run an enormous lab in Russia, and he was very experienced, very accomplished, very prolific publisher at the time. [00:08:09] Speaker C: I think he was in his late [00:08:11] Speaker B: 50s and me in my late 20s. I was. No, I was. I was exactly 32 when I took that job. Anyway, I felt like I was in my late 20s. And anyway, I just marveled that he could keep up with me. And. And now that I'm 64, I don't marvel at that anymore. I keep up with people in my. In their 20s and 30s that work with me. So I. I don't. I no longer have that temptation to ageism. But he solved that completely because he just had tremendous energy. And as soon as he got there, I told him, you know, I only have enough funding to start you, and you find your funding. And he instantly found his own funding. And he stayed in the US For a number of years. And eventually. Eventually he went back because they made him an offer he couldn't refuse. [00:09:02] Speaker C: They. [00:09:03] Speaker B: When neural networks took off, you know, they gave him a new lab and, and they didn't care that he was, you know, 20 years older than he was then. And they. So he, he continues to be a solid researcher to this day, about 32 years later. And he's still going strong, so that'll tell us all something. But, but anyway, it was, for example, the taxi driver who took me into his home so that I would not lose my bags. He was an engineer, but he was becoming a taxi driver because his employer was not paying any of their engineers. And so we enjoyed some sausages and some cabbage and some vodka and Talked until about 2 or 3 in the morning and then got up in the early morning and went and got my bags and caught my flight to Rost, where there was no one to greet me because everything that had been arranged both in either Sherman or Med airport that I took off from and then going to Rostov, there weren't people who were strong English speakers, but I wound up, I wound up next to a Cuban who. My Spanish is pretty good. So we were able to communicate and he helped me navigate all that. There were some things to deal with. But then on the plane, as we were sitting down, he told me, when you're with a Cuban, you're with somebody very intelligent. I thought he was just bragging, but he meant to prove it. So when the stewardess came down with a rickety steel thing serving whatever she was serving, primitive by our standards on this Aeroflot flight, he handed her my business card that said Boeing and said PhD on it. And he told her I knew enough Russian. And he was telling her, this is Dr. Wunson's first flight on an Aeroflot plane. And they all knew what Boeing meant. She dropped what she was doing, stopped serving the customers, took her cart and my car back up to the cockpit and one of the co pilots followed her back, nodded to me. He sat in the jump seat. I sat on his seat the rest of the flight, including the landing in Rostov on Don. And you know, from Rostov on Don, how spectacular that must have been. Now, this was late morning. The sun was gleaming off of the golden dome of the, you know, the Russian, the Russian Orthodox cathedral. And you've got the, the sea there and the city there. And yeah, it was an experience of a lifetime that could never happen post 9 11. But this was, this was 1993 and. And so a really wonderful trip. And then, and then going to Kiev and going to Moscow and then going to Novosibirsk and then the Trans Siberian Railroad and showing up in Beijing and reuniting with you there in Beijing. And that was when the streets were filled with bicycles, remember? [00:12:07] Speaker A: Yes. [00:12:08] Speaker B: So it was still an amazing bustling city, but it has completely transformed. [00:12:15] Speaker A: You know, I remember during this transition, everybody was trying to sell stuff because everybody was poor. It's like the engineers that had to end up driving a cab. And I remember picking up some good deals. And I think you picked up a good. A few good deals, too. One of them was the. Was the overcoat for a Russian, I think a Russian officer. [00:12:36] Speaker B: Officer. [00:12:37] Speaker A: And that was. I got one. I think you got one. And they were like 20 tons because they were solid wool. And they were. They were like, I don't know, a half inch thick because it gets cold in Russia. And, you know, we were able to pick that up for a song. But these people really needed the money. So, you know, what the heck. [00:12:55] Speaker B: Yeah, I had some fun with that coat. And I had the Russian hat that looks like brains. I still have that one. And so my last trip to Russia was in 96. I went in 93 and 94 and 96. And I thought I would go back many times, but now we're going to have to wait for a better relationship. So I don't think I'll be back anytime soon. But you mentioned about. I would jump at the chance to take the Trans Siberian Railroad again once relations are conducive to something like that. [00:13:30] Speaker A: I remember you telling me when you got to the border, there were all sorts of stuff which was thrown off the train because they were afraid it was going to be confiscated. [00:13:39] Speaker B: Well, everybody was sweating bullets. That's true. Everybody was sweating bullets. I wasn't sweating bullets. I was fine. But people were really worried. There was one guy who said, would you mind taking the stack of coins? And I knew what he. I knew what he was doing, but I said, sure, I'm not worried. I'll take the stack of coins. [00:13:59] Speaker A: He was trying to transfer blame, I guess. Right. [00:14:01] Speaker B: Yeah. And so after we got through, I gave him back the stack of coins and he gave me a ruble that was from the 1800s. I still have that ruble somewhere. But anyway, I wasn't worried. Maybe I should have been worried. I wasn't worried. But anyway, so I managed to get through the border between Russia and China. [00:14:23] Speaker A: Yeah, good for you. You know, I revisited Russia. I went to Siberia, actually, and I got one of those beaver. I think they're made out of beaver. Right? The hat. [00:14:32] Speaker B: Yeah, I got one of those, too. Those are the warmest thing you got to wear below zero. To wear those. If you wear them when it's 10 degrees, you're going to feel like you're in a sauna. You need to wait till it's below zero to wear one of those. [00:14:45] Speaker A: What you do now, in Siberia, it was, I believe, minus 15 degrees. So I went outside, I untied the little thing on the top of my head, put the flaps down, and my host there says, you don't want to do that. It turns out that every society has their idea of what being macho is. And I said, why? Why is that? He says, because of its. If it's 20 degrees below zero or warmer, you don't put those down if you're a man. So I had to tie them. But then later, he confessed to me that he was out one time and he didn't put his flaps down, and he had a little pop in his earlobe, and I guess his cartilage there froze. And he said he didn't feel any pain until it thawed, and then it really, really hurt. So everybody has their idea of what being a man is, what masculinity is in different societies, and that was the one that I did in Siberia. So that was quite an experience. So let's get down. Let's talk a little bit about technology. [00:15:44] Speaker B: Sure. [00:15:45] Speaker A: Before we delve into the contents of your paper, which I think is just fascinating, and I. I agree with it. Tell me about the Kummer center and the sort of projects you have. And by the way, the Kummer is spelled K U M M E R, and we're going to provide on the podcast link or the podcast page a link to that center so you can see what Dr. Wunsch is doing. So tell me about what's happening at the Kummer center and some of the projects that you have going and what your responsibilities are. [00:16:15] Speaker B: Yeah. So the Kummer Institute for Artificial Intelligence and Autonomous Systems is the result of a generous gift of our alumnus, Fred Kummer, who built over 1,000 hospitals and a hotel chain. And he decided that he. As he got into his 90s, he'd given us many gifts before, big ones. But he decided that he could spare another 300 million to help us to advance in areas of strategic importance. And he also wanted to encourage entrepreneurship. And he had several things, and we did a lot with this. So 300 million sounds like a lot, but my cut of it is not huge. But that's okay. It's still a substantial bump up from what my endowment of the Mary Finley Professorship was doing. And so it's a Cross cutting responsibility. And that's really the excitement and the challenge of it. I report to the Vice Chancellor for Research and really am here to encourage the development of AI through, through all aspects of the campus. And so I support that in any way that I can. I have one postdoc and two research assistant professors on my staff and support a number of graduate students and we, we help people to add a bigger component of AI to their proposals and also to establish a vision for AI for the university. So there are, there are a lot of aspects to that and I, I won't take us too far down in a deep dive of that, but what I will say is that website that you're going to link to, I would encourage people to click on the AI news tab of that website. And, and the reason for that is that I personally curate that news tab. I don't put hypey, I don't let hypey things get through, but I do even some of them are good but are too high plane. So I am looking for things that put AI into context and look at some of the interesting challenges to the field of AI. But you'll find two of my contributions there. It's mostly news articles, but I do have a keynote that I gave in Rio last month and I do have the article that you mentioned from the IEEE Computational Intelligence Society magazine last month on there. People who work in AI have heard of Michael Jordan. I sent him that article because it cited him and he responded right away with a recent article of his. He said he agreed with everything in my article and he, he gave me his recent article from July that that made the case that AI really needs to be cross cutting across many disciplines, especially economics. And you know, Marvin Minsky wrote about that. He had a book about economy of agents, AI, economy of agents, how these agents could interact much like an economic system. And Michael Jordan has refined that argument in a short paper that I put on the website as well. But most of our news articles, New York Times, Wall Street Journal and other sources, Times of India, there were a few that I liked. There's a number of sources that I grabbed and put these articles on there. And I do this on a multiple times per week basis. So instead of trying to sort through thousands of articles with some signal, but a lot of noise, this is a high signal, low noise and one thing that I'll put out as a teaser, there's is, there's an article about a, a person who's so influential in AI that in separate visits, Elon Musk And Bill Gates came to him and they had to travel halfway across the world to come to see this guy. And there's a pretty in depth article about that guy. And I found it fascinating because I never heard of him before. I'm not going to say more about him, but that's a teaser for your audience to find that article. But there are many equally fascinating articles on there and I promise it will at least entertain you. So I would encourage people to look at that. The labor of love to me to provide this. And I think that it will. People will enjoy it. You'll find a lot there to enjoy. [00:20:52] Speaker A: Okay, that's great. Do you want to share the website address just real quickly? [00:20:56] Speaker B: Yeah. C A I A S dot MST Edu and then there's a tab that you click AI News. [00:21:06] Speaker A: Okay. Yeah, I believe. And I've said that I don't know, 89, 80% of the articles that you see on artificial intelligence belong in the National Enquirer, that they're so full of, so full of hype, they're actually fake news there for clickbait and magnifying their viewership. [00:21:25] Speaker B: Right. So, you know, 100% of them I thought were worth reading. And I think that, I think you and your audience will find that it's. I hope that you'll be able to give it an A and It'll be over 90% signal to noise rate. Yeah. So I think it's worth a visit. [00:21:45] Speaker A: Great, thank you. You know, seems to me that the center that you are overseeing is doing, from what I see on your webpage, doing fascinating things. I think most of the heavy lifting in AI is being done by industry now. You know, OpenAI and Grok and, and Perplexity and Amazon and all of these others. And at universities, we're primarily interested in incorporating artificial intelligence into everyday tasks. So that's kind of what we're doing here at Baylor University. There's an old saying that in order to have a poet laureate, you have to be a rich country. And I, I think the same thing is with industry. In order to have a good research basis in industry, you have to be very rich. When I was a boy, it was Bell Labs. Bell Labs had all of these great scientists, Claude Shannon and Richard Hamming, and they did just incredible work. And now we have these rich people like Amazon and OpenAI that are doing this, the same sort of research. So it's going to be interesting to see what happens to these tools as they get more sophisticated. So let's talk about your paper. Artificial General, Actually, I want to run [00:23:02] Speaker B: on that risk for a moment because it embraces something that's very fundamental to human intelligence, intelligence, and is more difficult for artificial intelligence, and that is the concept of paradox. [00:23:15] Speaker A: Okay? [00:23:16] Speaker B: So paradox is something that is really important, and it's a theme in literature, it's a theme in religion, it's a theme in all scientific, you know, like Copernican revolution, you could say, built on paradox, too. And so, so the idea of paradox really is a deep concept that artificial intelligence, you know, can talk about it like an LLM could give you a whole essay about paradox. But to really process paradox, that's something that I think is important, and it relates to the theme that you just wrote. So it is both true and false, that you have to be rich to do this. The true part is that. So, for example, I had a student, he took all my classes, and he went to a small company in Alexandria called Heron Systems. And Heron Systems got a contract from darpa and they entered the Alpha dogfight competition. And it was a David and Goliath result. They, they beat Boeing, they beat Lockheed, they beat all these big contractors at this full blown flight simulator that was a fight or flight simulator, where they would engage in dog fights against the other AI contenders. And then the winner could take on a real human. And they took on the real human and they clobbered. [00:24:46] Speaker A: Wow. [00:24:46] Speaker B: And, and so they were taken over by another company called SHIELD AI and he's still doing that. And, and he was a wonderful student. And anyway, this, this idea, this is the David side of it, but the Goliath side is this. I asked him, well, what worked for you? He said, well, we, our First Grant had 900k from DARPA. They had phase one, and then they had to go phase two, and then they got to fight against the human. Pilot said, our Phase one, we spent the majority of our money on the hardware. So we bought, we bought the GPUs and we ran them all the time, you know, pretty close to 24 7. And that's what worked for us. The, the algorithms were the things that I learned. It wasn't some algorithmic breakthrough. We were just doing good engineering of good algorithms and then just a lot of fine tuning of those algorithms and so not, not changing the algorithms, but, you know, you do something and it works or it doesn't work and you make some tweaks and you do it again. And good engineering. And so that is a microcosm of what's going on now. So that's why you See the companies racing to scale up because that approach is actually pretty effective. You can do things that you can couldn't do before if you throw more compute at it. Yes, that really has made a big difference and that is going to continue to be a big theme. So people, if they think, well, the data centers, they're chewing up a lot of energy and they're expensive and maybe not in my backyard, and we don't want to do that, well, that is not going away. There will be more of that in the United States, there will be more of that in China. There will be more of that in many other countries. And they're doing it because it works. It is not the be all and end all to AI. So what I'd say is there's a paradox that it pays to be able to invest hundreds of billions in this, in this effort to buy your own nuclear reactors, to buy your own big data centers, and to, and to scale it up to where it uses more than a city's worth of energy. That all works. And it's a good path to go on. But there still is a space for people to come up with a better idea and to work in their, you know, smaller computer and to, and to do things and, and then to show that their idea is worth some more investment and then to, to work their way up to be on the bigger computer and to do what they're doing. So there will be some ideas that come out of left field that nobody's anticipating they're going to change the game. And there also will be continued ability of the big players to raise barriers of, to entry even higher and try to lock out the competition. And that's, that's part of why I'm saying that regulation needs to be done very carefully or it can stifle innovation. And, and so there's this. Oh, yes, the incumbents can pay for lawyers. [00:28:18] Speaker A: Yes. [00:28:18] Speaker B: So, you know, you and I can pay for lawyer with small L, but the incumbents can pay for lawyers with big L, where it's a division of their company. And, and so the more regulation happens, the more you need lawyer with big L and that's, that becomes a barrier to entry. So anyway, but paradox, this all comes back to paradox where we need to be able to embrace paradox and humans understand that. We accept that. We, we treat it as even an interesting aspect of art and literature, idea of paradox. We get a lot out of it. This is something that permeates everything, including AI. [00:29:05] Speaker A: You know, the idea about lawyers. The big companies many times have staff Lawyers. And when I consulted for Microsoft, the first thing I did when I went in was met with lawyers, believe it or not. And they explained, yeah, they explained me the rules of the game. And I said this before, so I'll say it again. It's probably dangerous to say, but Microsoft has never done anything innovative. They have either bought it, they have stole it, or they have gone to court to win it. And I don't think you can find an exception. And one of the reasons was because of lawyers. Bill Gates, you and I are from Seattle. You, you worked for Boeing before you, before you went to Texas Tech and then to Missouri. And Bill Gates, his father was a lawyer, Preston Ellison Gates, I think in, in Seattle. And I think that the legal background that they had was more of a secret than anything they did innovative. So that's, that's my two cents on that. I don't know. Any thoughts? [00:30:06] Speaker B: Yeah, well, lawyers have their place, and they do. I have some relatives and dear friends who are lawyers. And as you know, I have an mba. And when I took the business law class, the teacher was wonderful. And we had, you know, exams and homework and readings and stuff to do. But one thing he said at the first day of class and on the last day of class, he said, if you ever think that maybe you need a lawyer, you probably already needed a lawyer. [00:30:35] Speaker A: Yes, that's true. [00:30:36] Speaker B: He said, there's only one thing you remember from the class. Remember that. Anyway, I admire their dedication to very careful wordsmithing to make sure that you say what it is that you want to say and that you don't say or imply anything that you don't want to say or imply. So anyway, I, I, I admire the profession, but yeah, I'll also say that when, when Boeing appointed its first lawyer to CEO instead of engineer, there were a lot of, yeah, this was when I was at Boeing. The guy's name was Frank Schrontz. And if you look at the financial impact of, during his term as CEO, you'd say, this is all rosy, this is all good. The company did well under his watch, but the engineers were saying, oh, no, this is the first step in the wrong direction. And so some of the things that were imbued in the company culture, I won't lay the blame on that individual. He was there when I was there. I didn't see anything wrong with him. But I think that the cultural idea of saying the company should be led by a lawyer instead by a CEO may have led to some of their subsequent problems. So I can't point to any individual and blame them. But I think that the culture of the company was different when it was engineers from day one to present, and they no longer can make that claim. I still, I still love flying Boeing aircraft. I'm still very grateful for my time at Boeing. Stay in touch with my Boeing buddies and, and Boeing was good to me, and I tried to be good to Boeing. And so I'm not complaining at all. I'm just pointing out that, that there, there are cultural effects of our decisions, and we need to be cognizant of those. [00:32:33] Speaker A: Yeah. One of the guys that we worked with, if you remember, was Tom Caudell, who coined the word virtual reality. [00:32:41] Speaker B: He didn't coin the term virtual reality, but he did have several terms that he did coin. But he was. [00:32:48] Speaker A: Well, Don, I read it on the web, so it must be true. [00:32:54] Speaker B: But he's. Yeah, he. He certainly made a great impact. I have at least one paper with him on the topic of virtual reality, and also I have a video interview of him, much as I do of you with IEEE Computational Intelligence Society History. [00:33:12] Speaker A: Okay, we'll put a link to that up also. And you can learn about Tom Caudell. Of course, Tom Caudell was always known for his long beard. And so I have. [00:33:22] Speaker B: Even longer than yours. [00:33:23] Speaker A: Even longer than mine? Yeah, it went down to his belly button. He was kind of like, oh, it was CZ Top or Duck Dynasty sort of guy. So. Okay, well, let's. Let's give a little bit of talk, and I'd like to at least get a running start into your paper. And the paper was entitled Artificial General Intelligence is Nowhere Near Artificial Intelligence. Specific stupidity is already here and policy implications. One of the big things that I have seen is the definition of AGI. Artificial general intelligence changing. And it seems like everybody that's talking about is talking about something different. It's kind of like talking to somebody about consciousness. Nobody defines consciousness, but we all think we know what it is. It's the same thing with kind of AGI. And you reference multiple definitions of AGI, including from places like OpenAI and DeepMind. What's the importance of these competing definitions? Why do we have to agree on a definition before we proceed? [00:34:29] Speaker B: I'll say that I don't really accept the premise of the question. In other words, I don't think it's important at all to have a good definition of AGI because AGI is nonsense. [00:34:41] Speaker A: Oh, okay. [00:34:43] Speaker B: But I will mention some. Some people who I admire who tend to agree with that and people who I Admire who tend to disagree with that. So the one I'll mention that tends to agree with that. I attended Yann Lecun's keynote at aaai, I believe it was the one last year in Vancouver, Vancouver, Canada. And he had some excellent comments about it that wound up influencing the paper and basically saying the whole idea of general intelligence is nonsense. It doesn't exist. Intelligence is very specific. And we, we have developed our intelligence, and every species has an intelligence that has been shaped by the needs of that organism. So my, my first PhD, Daniel Prokrov was joking with me when he was still my student. During his first year with me, he saw some map of the dog's brain as it relates to olfaction. So a dog is an olfactory genius. He, he, he coined the term walking nose. And so the great Walter Freeman, who was the pioneer of neural network models of olfaction, and I interviewed him and Lofi Zada at the same time. [00:36:08] Speaker A: Okay, now olfaction is a big word to find it. [00:36:12] Speaker B: Olfaction is the process of smelling, so the ability to smell things. So anyway, a dog is an olfactory genius, and a dog has more neurons devoted to the analysis of smell than we do. So if we could somehow, like the Matrix, import a dog's knowledge, even though our nose does not have as good a hardware as a dog's nose for olfaction, we would still see an amazing upgrade in our ability to, to process signals of smell. We would be able to detect smell a mile away or stuff like that if we had that amount of neurons devoted to olfaction, and not to mention if we got the sensors that a dog has. So basically, intelligence is specific. So Warren Buffett had a great quote about that. He said that I'm very rewarded for having an unusual ability to deploy capital. And if you, if you drop me and my secretary in the middle of a jungle, my secretary would do fine and I would be dead in a matter of days because my skills at deploying capital are not that, are not that general use. They're very specific skill. And so as a species, we have skills that are really optimized for the fact that we're social animals, but they also are optimized for other things that humans need to be good at. But then we further optimize those as individuals partly by a diversity of what we're born with, the equipment we're born with in our, in between our ears, but then how we develop that as we age. And so the argument that Lecun made that I thought was Wonderful is that general intelligence is actually not so general. And so he, in his keynote, he actually used the term nonsense. But he also said we're kind of stuck with the term. So many people are using the term that we're stuck with the term. So that's on the side that tends. I won't pretend to speak for him, but that inspired some of my words criticizing the term AGI. I was already skeptical about it, and he gave me some. Some ammo to put in the paper. And I also cited a Berkeley psychologist in that paper. But on the flip side, there's a wonderful documentary that came out a matter of weeks ago, and there's an older one that came out a number of years ago. The one a number of years ago is called AlphaGo, and it was on Netflix, but now it's freely available on YouTube. And there's a more more recent one that came out weeks ago, and I wish I remember the name of it. Maybe I can get it for you. But they both have a starring role for the recent Nobel laureate, Demis Hassibet. I'm mispronouncing that. I think Hassabis. Demis Hassabis. B E M I S Hasabis. And I believe that he's the most deserving of the several people who have won Nobels in neural network. So the first one, you might recall, Leon Cooper, he didn't win his Nobel in neural networks, he won it for superconductivity. But he was on the board of the International Neural Network Society, and he attended IJCNN 91 that you and I were both deeply involved with in Seattle. [00:39:55] Speaker A: By the way, IJCNN stands for what, [00:39:57] Speaker B: The International Joint Conference on Neural Networks. [00:40:00] Speaker A: Okay. [00:40:01] Speaker B: And so anyway, the first one was an ICNN in 1987, and you had a big role in converting that to IJCNN by making a deal with the inns, the International Neural Network Society. And so Leon Cooper was on the board of that society, and he was the first neural network guy to also have a Nobel Prize, although it was not for neural networks. And so there were a few Nobels given this year. The one that I. That I think is the most well deserved. And they're all brilliant, the people who got these. Of course they got a Nobel Prize. [00:40:41] Speaker C: They're brilliant. [00:40:42] Speaker B: And a lot of brilliant people decided to recognize them. But the one by Demos Hasibis, the reason that I think that is a level above not only the other neural network Nobel laureates, but most of the Nobel laureates over time, why this is One of the most valuable Nobel Prizes ever given was because he got it for Deep Fold. [00:41:05] Speaker A: Oh, did he win it in physics or was it biology or what? [00:41:09] Speaker C: Medicine, chemistry. [00:41:10] Speaker B: And together together with the University of Washington professor. So they collaborated on Deep Fold and it's the protein folding problem. But what they did was they worked from their results on the game of Go, where they had AlphaGo was a deep neural network combined with reinforcement learning plus Monte Carlo tree search those three ingredients together to conquer the game of Go. I personally thought it would take a century. I had an NSF granted this area, but I never dreamed that NSF would spend about half a billion dollars to acquire this London startup company that was doing it because they were showing promise in the game of Go. And then they basically gave them a blank check. They gave them all the programmers and compute time that they wanted and they attacked the game of Go. They brought in GO experts, they, they had tournaments. They did an enormous amount of computer and they developed a system that they also screened moves. They had move screener. My student Mohammed Ranaku Zaman has that in his 1997 Ph.D. dissertation. And the first person that I know of, the first people I know of that applied reinforcement learning to GO were Terry Sinowski and his brilliant student Peter Dayan. And, and he talks about that in my interview with Satanowski from last year. That's on the Inns History Committee website. But that was one and done for them. They did a really nice paper. We saw it, we cited it and then we ran with that ball for a while. So those techniques that really got traction from them were things that had been around for a while, but they really ran with it. They did things that nobody else dreamed that that kind of resources would be thrown at it. And David Silver was doing that with Richard Sutton. I think that's why Hasabis hired him. Hasabis hired him because he was doing such good work in this area. And then they just were off to the races. And that was a good investment for Google. Right after that victory, their market cap went up by billions of dollars. That's also was the second doctor documentary, the more recent one. That was a Sputnik moment for China. That's when China decided, hey, we got to get serious about this. That's when they declared their goal to be the world leader in AI by 2030. And I think the US is still not taking that goal as seriously as we should. They have been doing things that are advancing them towards that goal. And the US is not making similar moves on the chessboard. Or go board. If you say they're allowing these moves to go uncontested in certain cases, the Chinese will reap the benefits of their investments, and the US investments are far from absent. We've got big investments, but we're not doing some of the things that they're doing that are wise things to be doing. So there's much that we can learn from what others are doing in the field. And so that, that's part of the policy implications, part of the story. So I talk about that a little bit in the paper. [00:44:34] Speaker A: We've talked about artificial general intelligence. You had a curious take on that. You said that it really isn't a viable name. It isn't something that we should consider. Did I get that right? Kind of summarize that in a little paragraph. What you, what you meant by AGI and what your stance is on AGI. I remember I asked you the question and you rejected the premise. [00:44:55] Speaker D: So a couple of things about that. One is that one thing that we are pretty good at, I won't say that we humans are great at, but we are maybe the only creatures that are pretty good at it, is dealing with paradox. And, and so a paradox is certainly around the idea of artificial general intelligence. So I'd like to talk about two folks who have weighed in on this topic, both of whom I admire and who have different opinions about it. So Yann Lecun, winner of the Turing Award, he has come out saying that AGI is unfortunately, certainly a term that we have to deal with because it's become so popular, but it's really nonsense because our, our intelligence is not general. We have all developed intelligence. All. We buy that. I mean, humans and animals have developed intelligence based on the needs of the creature. And so my, my first PhD student, Daniel Prokrov, and I, we were roommates for a while. And I remember one time he was joking about what he had found about a dog brain. He had found some article about dog brain, and it showed how much of a dog's brain is devoted to smell. And, and his joke was that a dog is a walking nose. And my spin would be that a dog is a genius of olfaction. Well, I personally know and have interviewed a human genius of olfaction, Walter Freeman, who developed the neural network models of olfaction or the processing of smell. And it's a very sophisticated capability. But if we humans had as much brain devoted to olfaction as dogs, we could probably smell things a mile away that we cannot even smell in the next room, right now and if we had the sensory capabilities, that adds to it. [00:47:05] Speaker A: So if, if I had a dog brain, I would smell good, is that right? [00:47:08] Speaker D: You would smell well. You might not smell good, but you would smell well. [00:47:12] Speaker A: Okay. Yes. [00:47:15] Speaker D: And then for the opposite perspective, Demis Hasabis, most recent winner of the Nobel Prize. The most recent neural networks winner of the Nobel Prize. There have been four. One of them did not win it for neural networks. Leon Cooper won it for semiconductors, but he was on the board of their National Neural Network Society and one of the, one of the early neural network startup company founders. But, but most recently, Demis Hasabis, John Hoffield, Jeff Hinton all won the Nobel Prize. And I would have to put Hassabas first among those, all the prior recipients of Nobel Prize related to neural networks and in fact maybe among most Nobel laureates because I think we will live longer because of the alpha fold for protein folding predictions. [00:48:17] Speaker A: Let me ask you, Don. I believe that there are other people other than Jeff Hinton or John Hopfield that were more deserving of the Nobel Prize than they were. Unfortunately, a lot of awards today are given by promoters, people that promote themselves as such. What do you think about that? Do you think there's people more deserving of the Nobel Prize Prize than Jeff Hinton and John Hopfield? [00:48:40] Speaker D: Well, I, I would say that it's unavoidable that, that some great talents get overlooked and the field is full of stories of that. I, I think that there are, you know, my best prize that I've ever received is the Pioneer Award. And I personally know people who I believe are more deserving than me that don't have it yet. And so there's always an element of that. But I would say that I, I've, I've met John Hopfield and know his work well. That was the first paper that I read that brought me to the field. I know the work of Jeff Hinton is phenomenal. He, he has run rings around me in terms of the impact of his publications. But yes, there are people who are, whose contributions are just absolutely stellar who have not yet been recognized in that manner. [00:49:35] Speaker A: My choice for the Nobel Prize would probably be Paul Werbos for Inventing of Back Propagation with his Harvard PhD dissertation and Bernie Woodrow for his breakthrough work in the 1960s at Stanford. I mean, he did incredible things with neural networks back then. [00:49:50] Speaker D: Both of them are close friends. Yeah. So I would put them up with anybody, anybody who's ever walked the face of the earth in the field. And another one that I would list is Steve Grossberg. Steve, they, they are giants and, and all three of them have been a tremendous influence on, on my trajectory. [00:50:13] Speaker A: You mentioned you got an award and you said you didn't deserving of it. I have a good. [00:50:17] Speaker D: No, no, I didn't say that. I said there were. [00:50:19] Speaker A: No, no, you said more deserving. [00:50:21] Speaker D: Yeah. [00:50:22] Speaker A: So I got a line for you. I received an award one time and I of course wanted to present myself as a humble sort of person. And I says I don't deserve this award, but I have lower back pain problems and I don't deserve that either. So that's a good way to accept awards in the way that I think. [00:50:46] Speaker D: So I'll contrast with this, with the field of fuzzy logic. In the field of fuzzy logic, it's very easy to point to one person. And Lot Fizzade is the father of fuzzy logic. [00:50:58] Speaker C: And [00:51:01] Speaker D: so anybody that works in the field agrees, no matter where. They might not agree on anything else, but they agree. Lizarda was the father of fuzzy logic, and I knew him too. You can see my interview view of him together with Walter Freeman, who I mentioned earlier. Those are available on YouTube. And these are giants. And we're very lucky to live in. In a time when it's as if technological progress was compressed that I still think that we're in the infancy of AI and even in the infancy of computers. I used to say we're in the Stone ages of computers. We may have advanced to the Bronze Ages of computers. There's a lot left to do. But, but the field of computing has moved so fast that a lot of the giants have overlapped. You know, I'm, I'm about to turn 65 in two months. So in, in, in the short span of time that I've been In the field 40 years, the, there have been so many of the giants that have overlapped with that period of time. People like Claude Shannon, who wrote perhaps the most influential master's thesis of all time, saying digital did. [00:52:21] Speaker A: He did digital logic. He worked for Bell systems. Right. Didn't it revolutionize the way the Bell systems hooked up their, hooked up their connections? [00:52:30] Speaker D: Yeah, well, his doctorate was one of two people that, that revolutionized information theory. But his masters was to say, hey, these switching circuits that you use for telephones, you could do Boolean algebra with them too. So you could use them for computations. And so the idea to use electronic switching circuits instead of the Babbage mechanical devices for computing. So he, you could call him the founder of electronic computing. So that's A pretty good master's thesis. Most people would trade their doctoral dissertation for that. [00:53:09] Speaker A: He was also a member of the Dartmouth Conference in 1965, where people like Marvin Minsky and John McCarthy and Claude Shannon and Ray Solomonoff got together. And that's where the word artificial intelligence was created. Created. I think it was John McCarthy that that did that. [00:53:25] Speaker D: Yeah, that's another one. By the way, Minsky came to the University of Washington and gave a talk while I was a graduate student, and I attended that talk. So there are. There are a lot of good examples of people who have contributed to the. To the field that are still with us. And so Eric Kondel of Principles of Neuroscience, he came to the university and gave a talk, and I attended that talk. He gave that talk to the medical school and that tome, giant tome on principles of neuroscience. He later got a Nobel for that. So, anyway, the names that we have mentioned so far, not Babbage, of course, not Loveless, but many of the names that we mentioned were alive during this period, and many, and many still are alive who have made just towering contributions. It would be like knowing Einstein and Maxwell and Newton and Kepler, you know, so it's just. It's. It's just a marvelous period to be alive in this field, and we can all be very grateful for that. And so we do need to realize in neural networks there are many people who have made, and many, I mean, still relatively small handful compared to the thousands or even millions of AI enthusiasts. And there will be many more who come because of the many great minds who've been attracted to the field. But a lot of the founding contributions have been made during our lifetimes, and many of those people are still with us. So it's really just an amazing moment in history. [00:55:22] Speaker A: Yeah, I remember. Let's see who won. Gosh, Roger Penrose won the Nobel Prize for his work in black holes. But unfortunately, Stephen Hawking had passed away. He should have got the Nobel Prize along with. In tandem. But you have to be alive to get the Nobel Prize. So, yeah, I think that a lot of things are happening. I do maintain, though, that this exponential explosion in artificial intelligence, people talk about Moore's Law and things of that sort. Exponential increases are never sustainable. And I think that eventually the curve has to level out. When it levels out and how it levels out, I don't know what that is. But look, I want to talk about your paper. You wrote a paper called Artificial General Intelligence is nowhere near artificial. Specific stupidity is already here and policy matters. And one of the things you do you emphasize autonomy as a critical and if you will, a missing component of what we're calling AGI? Why is autonomy such a hard problem compared to the pattern recognition or large language models? [00:56:30] Speaker D: Well, and in fact, autonomy is something that the field is attacking. There's a lot of research targeting agentic [00:56:39] Speaker C: AI [00:56:41] Speaker D: and tools being even rolled out, some on a trial basis and some that you actually can use use. But, but I do think that this is a major limitation. And so in my paper, I say that insect intelligence is ahead of the field of, of artificial intelligence in terms of these autonomous capabilities. Because if you, if you released a fly into this room and I were trying to kill was would be pretty fair match, despite my much larger brain, the fly's agility would serve it pretty well. And if I keep the door closed, eventually, after a large amount of effort, I would succeed. I might break some things in my office in the process. So it's amazing how much an insect can achieve with such a small brain. And we're just not that good at dealing with that. And so that comes back to my criticism of the term artificial general intelligence. To say that intelligence that has been successful, both artificial and natural, has become specialized to a task at hand. And the counterpoint would be Demis Hassabis has explicitly stated artificial general intelligence as a goal, but he has wisely taken task after task after task to rise up a chain to move towards those objectives. So he started out with games and then made general game engines that could learn a set of different games. But that's still a very constrained environment. Environment. And then the problem for which he won the Nobel is a great example of the success of. So I think it's not a bad thing to have a goal that is reaching for well beyond our grasp, reaching for the stars, a quixotic effort. But I think it helps to, [00:58:55] Speaker A: to [00:58:55] Speaker D: have the humility to say that this is a quixotic effort and not to be proclaim victory prematurely. There are many examples in history of a premature claim of victory that backfires. So I don't think that any of the claims that we have already achieved AGI are anywhere near true. And also I don't think the projections that we will achieve AGI by 2030 are true either. There are such simple failure modes of these systems that we have failure modes too. But still there are failure modes of these systems that humans would simply say that that's not intelligent at all. And yet there are things that these systems that can do far better than humans. But you know, when my dad Brought home a calculator in the 70s, [00:59:53] Speaker C: and [00:59:53] Speaker D: I had fun calculating factorials. Well, it could calculate factorials far faster than I could. You just put in a number like 99 factorial, and it gives you, you know, in scientific notation, how big that number is. And, you know, there are. There are all sorts of things that computers can do better than us, but there have long been things. So the John Henry song, you know, of the driving, the railroad spikes, and a machine that could do it. But even having a railroad in the first place, you know, there are things that machines can do that humans can't do, but humans created these machines. And the same is true with computers. There are things that you can design a computer, computer to do something that you can't do, but there's still a lot of things that you can do that the computer could never do. And so if the building catches on fire, if the building catches on fire, you leave the computer behind and you run out of the building. Too bad for the computer. [01:00:59] Speaker A: Yeah. So, yeah, speaking of things that computers can't do, you talk about Moravec's paradox. What is that? And why do AI systems still struggle with this problem? Moravik's paradox. [01:01:13] Speaker D: Well, so I'm quoting from an article by LeCun, so. And I cite it, and as I recall, it's. It's sort of the idea that you. You do something and when you succeed, it's instantly, okay, that's not AI. That's a nice thing you did, but that's not AI. And I might be getting that wrong, so follow the citation. But basically that is a recurring theme. So, for example, when I entered college, they had just gotten chess machines that can play at the master level, but there were little hacks you could get every now and then beat one through. For example, they underestimated the value of pushing a pawn towards a queen, so they could do great tactical levels and be well ahead of you. But if you could. If you could slip a pawn down to the, you know, close to the sixth or seventh rank, you might be able to force them to trade a rook for that pawn because you're threatening to queen it, and you might be able to get some. So I'm nowhere near master level, so it was hard for me to exploit that very often. But still I noticed that these little weaknesses around the edges of such systems, but still, to play at that level was considered a goal of AI. And then fast forward to 97, and IBM trots out at great expense, Deep Blue and beats the human. World champion Gary Kasparov well at the time I was already very interested in the game of Go and I literally thought it would take a century to get there, to beat the game, to beat a human champion in the game of Go. And I would encourage people to watch the documentary Alpha Go. And there's another one that came out just a few weeks ago and the name escapes me, I'll put it on my website. But both of them involved the work that was done by DeepMind, led by Demis Hassabis. Another major technical contributor was David Silver. He collaborated with the Turing Award winner Richard Sutton and they developed a computer that would use reinforcement learning to self improve. And eventually they, it didn't take anywhere near as long as I thought. So about a decade ago they, they beat one of the top human players and then subsequently they beat the then reigning human world champion in the game of Go. And, and that is so impressive because it has a lot of challenges that chess doesn't have. So the, the number of board states is much larger. The, the non local interactions are, are sometimes even harder to see and the evaluation of a board state is much more challenging. So it's, it's NP hard just to evaluate a board state and it's p space complete to, to attack the problem of the game of Go. And this is to say if every chess complet computer, if you had billions of the deep blue systems all reprogrammed to attack the game of Go, you would not solve it by brute force. But what they did still had a lot in common with brute force techniques. But the reinforcement learning tools that were applied were very powerful, incredible and very well designed. And that it was a major breakthrough. Major breakthrough. [01:05:05] Speaker A: You know there's emo. Phillips had a great story. He says yeah, computer could beat him at chess, but it was no match for him in kickboxing. [01:05:15] Speaker D: Right. [01:05:16] Speaker A: You know, talking about swarm intelligence, backtracking a little bit. I met a guy at Stanford, they called him the Flyman. His total job was to explain the way that the fly works. And one of the biggest big questions that he asked is how does a fly land upside down on the ceiling? And he was working on that. I don't know if they solved that how flies can land upside down on the ceiling. You talked about things that insects can do that are really remarkable. [01:05:44] Speaker C: Right. [01:05:44] Speaker D: That's an example that I like to trot out as well. And so you might recall that I had Robert Pinter on my dissertation committee and he was a, he had a joint appointment between electrical engineering and zoology and he was an expert on insect vision and insect modeling. And so we have an expert on that topic here. And he knew about Pinter's work actually when he came here. So anyway, that, that's a fascinating problem. You might remember Karen Haynes from her, from the IEEE Neural Network Council. She was coordinating. [01:06:32] Speaker A: She ended up, she ended up in Australia, didn't she? [01:06:34] Speaker D: Right, but not before she did a PhD with Tom Caudell and she, she wrote her dissertation in the. So I think it was Cassison advice, her masters. And she wound up working with Tom Caudell on her PhD and she did work on fly vision. And so what flies can do is really amazing. And, and yeah, and the scenic is the ability to land and hold on upside down, landing on the ceiling. And yeah, I have yet to see an AI driven system that can do that. That, that's not the same as like if you, if you made a, a drone copter that could use the force, reverse the force of the propellers and force itself to be on the ceiling. The, the fly can land on the ceiling and hold on. And, and so that's a, that's a remarkable control problem. And, and so there's, there's all sorts of stuff there like that. So I think we need to have a lot of respect for natural intelligence. It, it goes a lot further and a lot deeper. And also it's more parsimonious than anything that's being done in AI. So having systems that can do so much with so little. So people are talking about tens of megawatts, hundreds of megawatts, even gigawatts for, for up and coming data centers. But we, we make do with 20 watts. And so prior to LEDs, this was called, this was about the amount of wattage of a nightlight. So dim bulb is no longer an insult to one's intelligence. It's actually a great compliment. What we can do with our dim bulbs is really remarkable. [01:08:29] Speaker A: It is, yeah. It is incredible. Let me ask you another question about AGI. And you made this point in your paper and I've been thinking about it and I kind of agree with you. I think that 90% of the articles that are posted on the Internet from these news services and stuff belong more in the National Enquirer than they do as a news story. And you raised the question, is AGI more of a marketing term than an engineering target? I'm not even sure who coined the term AGI. And what do you think? What do you think? [01:09:05] Speaker D: I cite a book by the title of AGI and so I think it's the first citation in my paper. So I have, I browsed that book, it's good as why I cited it. I speak about favorably in the paper. So I think that's the best source for looking at the various definitions, the history of it. I also quote from the OpenAI definition of AGI and some of this actually it's not just marketing. There's also some legal ramifications. OpenAI recently, I think renegotiated that might not be the right word but they, but they have clarified the understanding of their agreement with Microsoft because they had some, some terminology in there about when AGI was achieved that it would change the nature of their agreement. And I think that they realized that, that, that that term was too mellifluous to yield a legally enforceable agreement. So I believe that they've reached renegotiated those terms. But I didn't put one of the news stories on my website. Maybe if I find one that I like I might, I might dig into that and put up a, a good one. But you know, you were talking about a lot of the things should belong in the National Enquirer. I have an AI news feed on my website. [01:10:36] Speaker A: Yes. By the way, I looked at it down and it's excellent. You really, you really can keep up on things. So I would recommend it to our listeners. [01:10:43] Speaker D: Yeah, I update a few times a week. It's personally curated and I, I don't put stuff in it if it's real hypey, even if I like it. So for example, I, I, there are things I'm, I'm a user of, of the paid version of Chat GPT, one of the, one of the higher tiers. I'm a user of the paid version, the Google Gemini Pro and some of the other ones I use too. I, I use Microsoft Copilot and, and so these tools, they're flawed, but they're useful. So in all of them I've found shortcomings that I would not tolerate from a human assistant but that they will make promises, they'll say I will do something and then they give you something and they say I've done it. And it turns out that there's just a hollow shell and they didn't do anything that you described within there, particularly in the, in the space of coding and they're improving. So I don't think that things will forever be that way. But I do think that if you want to be a user of, of AI, particularly LLMs, you know, you need to be very aware of its limitations and that's one thing that I think earlier in the field where people were not talking about such general tools, then there was a very clear sense that I'm designing an AI system to do X, and outside of X it's useless, but within X it can do what I want it to do and then I can diagnose if it needs to be improved in this way or that way. And that's a good way for an engineer to think about AI. And it's very tempting to think of an LLM as an oracle. It does have a broader range of user interface and so you could ask it about all sorts of things, but it takes more responsibility on the part of the human human to decide whether what you're getting back is actionable, whether it's useful, whether it's true, whether it's been hallucinated or not. So you need to calibrate carefully. Often I'll be doing something right in the center of the strike zone of the AI. I'll be asking it for looking up things about a particular scientific topic. For example, it'll give me back something and I'll say, no, you need to try again. You need to, you need to work harder on this. You're giving me something very shallow and there's stuff out there that I know exists that is a better answer to this. [01:13:29] Speaker B: Go back. [01:13:29] Speaker D: But I'll give it some hints, I'll give it some direction of look for this. And so it's really, AI has been and will continue to be a powerful tool in the hands of humans. And the more the people realize that and, and use it accordingly, the more successful that they'll be in the application of AI. And there's a lot of work that's high consequence where you really need to get it right or you're going to lose money or maybe lives. And, and so it's really important. If there's one thing that engineers need to know about AI, it's that the engineering mindset has always been to work hard to verify that the system works as intended and, and to realize and declare the limitations of any system that you feel. [01:14:32] Speaker A: You know, Ronald Reagan used to say trust but verify. I think with AI, it's kind of backwards, isn't it? It should be verify, then trust. [01:14:42] Speaker D: Yeah, maybe verify and then trust, but still verify, but still verify. [01:14:48] Speaker A: The other thing is, is that when you get responses from these, from these different platforms, they seem so sure of themselves. Another Emo Phillips joke. And I think this is really true of these large language models and these AI platforms. Is that computers are not really intelligent. They just think they are. Certainly they present themselves as they are intelligent. [01:15:14] Speaker D: So there's one thing in the article this reminds me of and that is that there's a section about, about the threat of AI taking over. And so, so anyway, there are, I think it's a red herring. So there are many threats of AI that really deserve our attention now and a huge one is the threat of AI to our privacy. Because this device, I, I won't put the side with the brand on it because I'm not blaming any particular company, but our smartphones are, have created an amazing ability to spy on us. So I would say that they have reduced the cost of spying on us by a factor of at least 10,000 and maybe 100,000. So they can track our every motion. They can be turned on even when you think they're turned off. You could, you could then use the microphone and the camera and the GPS on the device. You can do all sorts of things with this if you have either the right authority or the right hacking capability to do that. And if, if you have all that, installing such software might cost you a few thousand dollars. But once that has been installed, then the marginal cost of using it to spy on somebody is just a mouse click, you know, it's negligible. So, so anyway, the, the, the ability to do all sorts of things to spy on people has decreased, I would say easily by a factor of 10,000 and maybe much more. But what, what has still been expensive is to decide which people you want to spy on. And AI has reduced that cost. So in the past, I'd say we could take some cold comfort in the fact of, although it's very cheap to spy on us, it's expensive to decide that we're working worth spying on. And furthermore, you know, I'm, I'm not a senator or a billionaire or CEO of, of a multinational company or, or you know, somebody even more important than a senator. So my relative obscurity is a shield against the ease of which it would be to spy on me. So yes, if somebody really wanted to know to spy on me, they could spend a tiny bit of marginal cost. There'd be a one time cost, but then the marginal cost to spy would be very low. But I can take some comfort in that, that the cost of deciding to spy on me would be very high. Well, AI easily lowers that barrier. So if you had some parameters that you were trying to decide who to spy on, then AI would dramatically improve the ability to say the People with a certain, let's say, a certain income level, a certain political persuasion, a certain, a certain sexual orientation, a certain, a certain religion, a certain geographic location, all sorts of criteria, a certain level of history and various public records, whatever they might be, a certain purchasing pattern. There are all sorts of things that, that would make it much easier to spy on people, intrude on their privacy. And that doesn't even begin to open the Pandora's box of social media. And so the things that we voluntarily disclose. So it's, it's gotten to where the risk of AI to privacy is not some hypothetical future risk that we might arrive. And we're debating about whether it's coming in three years or five years or 10 years or 30 years or never. These are, these are risks that are already here that are not sufficiently appreciated. And so I would say in that sense that the concerns about adi, the AGI and the policy implications about what to do about AI have got certain risks backwards and that indeed it behooves certain entities to encourage you to look at this family of risks and ignore this family of risks that I'm pulling [01:20:09] Speaker C: off of the screen. [01:20:11] Speaker D: Right. So, but this family of risks are already here and already a danger. And if you say, hey, look at these. Actually look at these. These are the ones all right here and already in danger, don't look at those ones. No, those are no problem. No, that's, that's a pretty big issue. So, so, and also regulation creates barriers to entry. So there are entities that can hire armies of lawyers. You think about the trillions of dollars being spent on AI and so literally of the top 10 companies by market capitalization, all of them basically trillion dollar companies. Nine of those top ten are AI companies. So how many lawyers can they afford to hire? As many as they want. And so to have government regulations, they're needed, but they have the ability to hire lawyers and lobbyists. So they could create barriers to entry by having regulation, regulations that make it hard for new entrants. And they can also lobby to not have regulations that they're not very interested in. And so, for example, privacy would be a good example that privacy might interfere with their ability to market to you or their ability to do various things. [01:21:44] Speaker A: Well, you know, my daughter wants to put in an app and follow me wherever I am. So she can, she can note wherever I am, she will know. And I've resisted it. I don't want to surrender that, you know, in terms of privacy. I was doing some work with a nonprofit, and this one guy Said, well, you know, maybe he can make a contribution. And the people at the nonprofit looked up this special service and they showed everything about this guy. I mean, he was rich. He was one one of these guys that you said that AI should pay attention to. The Bible says that God knows the number of hairs on your head. It didn't go down to that detail, but it was pretty daggone close. [01:22:29] Speaker D: I'm giving God a much easier problem with each passing here. You're counterbalancing it. You're giving him a little more work to do here. [01:22:36] Speaker A: But we are, and we have a professor here that got a very nice NSF grant about mining information from publicly available social media. And his task was to do things like identify people that are possibly violent, have violent tendencies, might be school shooters or people that might have a tendency to commit suicide and things of that sort. And all of that could be mined from all of this social media and the person could be identified. So you're exactly right in terms of the loss of privacy. But again, with Amazon and Amazon.com and some of these other services, I have surrendered my privacy. I've surrendered my privacy to Google because their service is so daggone good that it's worth that trade off. So that's kind of where we're at, I think. [01:23:30] Speaker D: Yeah. And by the way, this month, I think is the last month, that Google Maps timeline, they were keeping that on their servers. And now if you want to keep your timeline, it's pushing it to your device, but you can still keep your timeline. I kind of like my timeline, but it's a privacy trade off. [01:23:51] Speaker A: So what do you mean by keeping your timeline? I don't know what that means. [01:23:56] Speaker D: If you look in Google Maps there will be a menu item on the left hand menu drop down for your timeline and it'll show everything that Google Maps has detected you doing over the time that you've authorized it. And so, so I, by the way, I do share mine with my wife and son. [01:24:19] Speaker A: Okay. [01:24:19] Speaker D: So so basically, you know, if something were to happen to me, they can find out where I am at or at least where my phone is. They might not find out where I am, but they can find out where my phone is. And actually that helped one time. So one time I didn't have the same, this is a blue phone, you know, phone jacket, but I had a black phone jacket. And I was in Washington D.C. and I had a cart with a, with a little black thing where the child can sit in. And this in D.C. they charge you for the bag. So I had my own bag. So I, I checked out, and then I put stuff in my bags, and when I packed up all my bags, I didn't spot the black phone on the black child seat, and I left it in the cart. And somebody came along, and instead of turning the phone in, they, they took the phone. And. And then I didn't realize this, but I mentioned it to my wife after. I did realize it, but I had to run off to a meeting, and my wife called the phone, and the person said, well, I want a reward for finding the phone. And so she could see where the person was, and he told her where he was. And this is remarkable. Remarkable. You know, Hong, she's fearless. She. She went off in downtown D.C. to go meet this person, and she was on her way there, and I guess she ran into some people that, that she told the story to on her way there, and they said, oh, no, don't do that. That doesn't sound safe. And they called the cops, and the cops came, and a cop went with her to meet the person, and he, he changed his mind about asking for the reward. And she said, no. I said, I'm going to pay you the reward. I'll pay you the reward and take the phone. And it turned out he was one of these vendors downtown on the National Mall, and he gave her and the cop a free bottle of water. So everybody was happy. But. But yeah, Hong is fearless. When I found that out, I was kind of horrified. But I, I got my phone back. But anyway. But the, the. She was able to do that because I had enabled her to track my phone. And, and so, okay, the guy had told her where he was, but she was able to also see where it was. [01:26:56] Speaker C: So it was interesting. [01:26:58] Speaker A: Oh, my goodness. [01:26:58] Speaker D: And I, I, I was off at, like, DARPA or something. I came back from my meeting, and she told me this whole story. [01:27:05] Speaker A: That's really astonishing. So you've talked about some of the things that we should be careful about with artificial intelligence on the page that you curate, where you actually go through news articles. You said that you censor some of the articles. [01:27:19] Speaker D: No, I don't censor any of them. I either put the whole article, or I don't include the article. [01:27:23] Speaker A: Okay. But I curate them. You're selective. You curate. Okay, that's. [01:27:28] Speaker D: Yeah, I curate them. Yeah. [01:27:29] Speaker A: So what are some of the key things that you look at in an article about artificial intelligence that let you know that it's clickbait or junk news or not worth posting? [01:27:41] Speaker D: Well, also, there's a lot of worthwhile articles that I don't put because I'm interested in an engineering approach to AI. I'm interested in where the field is moving. I'm interested in, interested in. I'm interested in its limitations and also his strengths. So, for example, I will have more stuff related to the energy footprint of AI than maybe some other person with similar, you know, similar desire to, to run a feed. In fact, one of my friends, Daniel Torratz, uh, of Auburn University, he used to be here, uh, atmosphere, M.O s t. And he's got an AI news feed and he does a separate cyber security news feed. And they're really excellent. Daniel Torres, T A U R I T Z is a really good person to look up. Just remarkably, remarkably prolific about, about choosing these articles. And he chooses different ones than I do, but very good ones too. He doesn't put any garbage up there. So, for example, I put up a New York Times article about the people who are suing open AI for these sycophantic chatbots that have encouraged people to commit murder and suicide and stuff like that. I put up things that relate to policies. So some of the policies that are currently changing the field that are, that are being put out by the federal government, some of those, I have put stories out about those or sometimes links directly from pronouncements. There was a, There was an executive order on April 23 about AI for K through 12 that I thought was important. So I just put that. There were articles about it, but I just put the executive order out there saying that the agencies need to do more to educate all our students about AI. And so basically things that are in the sphere that I think my audience would be interested in. So there are all sorts of things that are. Oh, like AI for games is huge. That book that I mentioned about artificial intelligence that was written by Julian Tagalis and he is also one of the people who's done a lot with AI for games. And so I'm very interested in AI for games, particularly in the early phase of my career. I was working on AI and the game of Go in the late 90s and for about a decade beyond that. And there are some of the things that are done with AI for games are still fascinating. The work that DeepMind did on StarCraft immediately after beating the champion and Go, that work was also quite fascinating. But I would say I don't post as much on entertainment AI and AI for games as I would have done if I had started the site a decade ago. I wouldn't quite say it's a solved problem, but I would just say that there are some that it's a relatively mature part of the AI field and I'm still very interested in it, but more as a customer than as a topic that I think is changing the world. But I do think that the tools that they developed in terms of AI games are extremely powerful. And so the LLMs are built on these reinforcement learning tools. The reason that they're so widespread is because these techniques really work. And so if you have designed the right cost function, if you, you define the problem correctly, if you define the problem in a focused enough manner that you can predict what the AI is likely to be able to be beneficial for you, then those tools are extremely powerful. So, for example, I was once asked about future of AI and I was saying that I thought that I personally would be more likely to trust, say an AI hedge fund manager or a mutual fund manager than I would be to trust an AI baby center. But I would suggest that one of the risks of AI is that maybe the mass majority of the population are making the opposite determination. And that I don't think is good. So I would say that a lot of people are de facto allowing AI babysitters because there are AI algorithms that are driving the social media feed, there are AI driving the gaming and making those more addictive. And so people are essentially subject to AI babysitters and their performance might not, might not yield the desired result, will not yield the desired result. I'll say it a little stronger. And whereas AI might indeed be able to be designed to do better than a lot of professionals in money management, I've worked in this field since the time I was your student in the late 80s and early 90s. I've published in this field as early as 1991 and some publications beyond. But I've never used it on my own money. I thought that I could do better, but I do think that it would not be that hard to design a tool using AI. That would be the average, say, professional money manager, professional mutual fund manager, professional hedge fund manager. And there are other people who have reached that conclusion. And some of them have already done this and made a lot of money. So you can look up Bill Simon. There's a book about him called the man who Beat the Market that you can look up David Shaw, David Shaw and Associates. These are people who have been interested in this and turn them into huge enterprises. [01:34:31] Speaker A: You know, Don, I knew the first professor of financial engineering in the United States. I forgot his name but he was an interesting guy and he was approached all the time by people that said, you know, I came up with a neural network that can forecast the market. He didn't even have to look at the software. What he did to these people that approached him, he said, you ask the question, what kind of car do you drive? Because if they had been successful in doing this, they would have applied it and been very, very rich. So about the forecasting, it's very, I think it was Niels Bohr. There's an old Danish saying that Niels Bohr, the quantum mechanic guy said. He said that forecasting is very dangerous, especially if it's about the future. We talked about some of the things that we, we use some of the tools that are used to talk about fake news. I think one of them is the people that make totally speculative statements about what the future of AI is going to be. I also look for seductive optics and seductive semantics where in, well, for example, anthroporphizing, they make a robot like a human being or a dog or something like that. So a sudden there is association with artificial intelligence, with something which actually exists and exists in the sense that it's a human being. My wife, she used to just love our dog. And I said, don't love the dog. It's a dog, it's not a person. I think it's the same thing with AI. We look at it and sometimes say it's a human being. It isn't. I told my wife, it's a dog. I think that people that talk about the future of AI say, hey, it's a computer. The other thing that I look at for fake news is the source. I always take with a grain of salt the words of people like Sam Altman or Elon Musk that are promoting what they are doing. They're promoting their company and of course they're going to come out, they're going to be positive. So a lot of these things can be used to filter what the AI, what the good AI nuclear news is and what the bad AI news is. [01:36:40] Speaker D: By the way, you mentioned computational intelligence and financial engineering. And I believe you and I were both at the conference in about 94 in New York City. Yes, Cypher conference. And I think we shared the hotel room, in fact, at the Marriott. And, and anyway, I got in a conversation with a fellow who was doing modeling of. He, I think he was using stochastic differential equations and he was modeling deviations from the black Scholes model or something like that, and I had a fascinating conversation and as you might recall, before I joined your lab, I got a master's in applied math. And I think that that was my edge when, when coming to talk to faculty members and, and I came to you and you were my first choice and thankfully you took me as your student and I, I found out that you could write down the math as fast as I could read it. So, so we're no slouches in math, but I had a talk with a guy and he was doing this mathematical modeling, you know, using stochastic differential equations and other tools. And he was an economist. And so one thing that I would, that I'm warning people about hubris in the field of AI is that like the business school students, the English majors, people like that, they are, they are not. What, what. Maybe people in their 60s and older envisioned those fields as being when they went into STEM disciplines and saw these other people in non STEM fields. Some of them can run rings around us. And so this guy, when I looked at his papers, I had had this conversation with him and I was fascinated. I looked at his papers and I said, he's doing good work, but it's an awful lot of work to study his work. I would love to collaborate with this guy. And he was very friendly and had a great conversation for a better part of an hour and just fascinated by what he was doing. But I realized in 1994, I'm not in a position to keep up with this guy. His name was Philip Dygwig. He got a Nobel Prize a couple of years ago while Washington University in St. Louis. So I, I had correctly assessed the chops that he brought to the table. But that's one exciting thing that was already true back in the 80s and 90s and, and remains true. The people that get attracted to this field are just stunning what they bring to the table. And it does not matter what discipline they come from. I have learned things from our English faculty that changed the way that I approach the field of AI. We, we have psychologists, sociologists just doing amazing things and, and I love to attend the lectures of our, of our top historians and, and some of them are just doing fantastic stuff. And so the field has always been multidisciplinary, disciplinary, and it's getting even more so. So like the, the, the people who are good at things like prompt engineering are very different than the people who are good at the mathematical analysis. And then the mathematical tools are sometimes very different from the ones that we focused on. So, so, you know, even PDEs and field equations and stuff like that, they're, they're important, but there are other things that are important too. And so that's one thing that I love about the field is the diversity of disciplines that goes into the field is just staggering. And I feel like a kid in the candy store. And so there's just so many different things to look at and read about and learn about and different people to talk to. And that's one of the most rewarding things about being in this discipline, is that you can work on and learn about almost anything. If there's something you're curious, you can work on it, you can learn about it. So your, your former student, Iwan San Judge, I got his masters from you. He's very interested in applying LLMs to the Bible. And there's so many different fascinating questions about the Bible, that topic. And you know, he, he's very excited about it. So there, there's, you know, in philosophy, of course, I, I was just revisiting the work of the late Daniel Dennett, who just passed this year, I believe, and he did some marvelous work in philosophy of AI. He collaborated with Douglas Hofstadter, who got his doctorate in physics, whose dad was a Nobel laureate. It in physics. He decided to try something different after he got his doctorate in physics. He went into a computer science department as a new assistant professor, decided to write a book. His first book got a Pulitzer Prize. So, yeah, Douglas Hofstadter, Goodell Escherinbach, An Eternal Golden Braid. [01:42:07] Speaker A: Oh, yes, yes. [01:42:08] Speaker D: Many great books. He's written many great books since then. So one of the giants of AI. So there are all sorts of people making all sorts of contributions. It's not all about the citations, it's not all about the awards. And we've just scratched the surface. So even if you know the subset of, the subset of, the subset that you want to focus on, you can easily, you can easily spend 3, 4, 5 decades to delve into that facet of it. And there's a lot of work left to do. So if somebody wants to see AI AGI, let's say, I, I don't think that it's coming. But if you, if somebody wanted to see AI that is across a broad spectrum of capabilities, including physical and mental capabilities, including being able to interface with the real world world in a diverse way, if they want to see that in their lifetime, I recommend to do research on a longer lifetime. I, I think that will give you a higher odds of seeing these things in your lifetime than doubling down on just the AI aspect. [01:43:22] Speaker B: Of it. [01:43:23] Speaker D: So the embodied AI is already huge, you know, to like I talk about, if the building caught on fire, I'd leave my computer here and get out out of there. To have embodied AI that can do something about the broader range of things, the broad range of things that we encounter on a day to day basis, it's just not that near. But if you, if you take, if you carve out a real specific problem that we're interested in, that's worth a lot of money, then yeah, AI might solve it way faster than you think it can. You might think that this is a century, you might find out next year, bam, they've got a center solved. But if you, as soon as you make it more general, it's a whole different ball game. And, and so I am a big fan, as I already said, make clear of both Yann Lecun and Demis Hassabis. Hasabis deserves that neural, that, that Nobel Prize for Alphafold, that's a very specific intelligence, it's not a general intelligence. But reaching for that star of general intelligence helped him develop the tools that enabled him to solve one of the most important problems that we faced since the beginning of this century. And I think that the experts thought that that would take a century. It used to be that you would, if you could solve one protein folding problem, you could get a seven figure grant from NIH immediately. And they solved like a quarter million protein folding problems when they published their first result. And they're way beyond that now. There are literally millions of people using those tools. And so the impact is just tremendous. Absolutely tremendous. So focused research on AI, I would say there's almost no limit to what it's going to do. But as soon as you start making these grandiose claims, I, I'm glad that people are excited about and trying to do that. I just think that, that, that is not coming soon. [01:45:25] Speaker A: Yeah, I agree. [01:45:26] Speaker D: I'd love to be proven wrong, but I think that that attempts to prove that wrong. And, and that's why I say near the end of the article, the challenge for IEEE members is that engineers have always had the attitude the best way to predict the future is to, to invent it. And so people who disagree with what I'm saying in the article, I say that's encouraged and the best way to express that disagreement is not with some grandiose claims. But you know, I'm saying this from here in Missouri, the show me state so the results can speak for themselves. And right now the results are not telling a story that this is anywhere near. The results are fantastic. I mean, I have stock in some of these companies for a good reason. I believe that they're producing results. [01:46:24] Speaker A: A lot of people think that some of these companies like OpenAI and Grok and Nvidia are bubbles. [01:46:32] Speaker D: Well, I wish I had stock in OpenAI or Grok because those. So the, you know, private equity is a whole nother thing. But I don't think that we're at an AI bubble. I think that such a thing could happen, but I personally don't think we're at one. But. But that I could easily be wrong about it. But the. But the reason that I think that is that there's a lot you can do with brute force. [01:47:05] Speaker C: A lot. [01:47:06] Speaker D: And the problems that people are working on are so incredibly valuable. So I think that there will continue to be investment in expanding the capabilities of AI because they're. So what is it worth to have a better drug target that can conquer cancer or Alzheimer's? What is it worth to stop paying people eight figure salaries to run a mutual fund if you could do that with an algorithm that you own? You know, what is it worth to, to be able to, to lower the cost of getting something into space by three orders of magnitude? What is it worth to have a better solution for energy availability and more? There are things that are worth not trillions, but tens, hundreds of trillions or more that are problems that AI might be able to reduce the cost of and improve the efficacy of by orders of magnitude. So even though these top nine companies that I mentioned are all trillion dollar [01:48:26] Speaker C: companies, [01:48:28] Speaker D: 9 trillion is a small number compared to the potential. Now actually, those nine companies add up to about $25 trillion in market cap. $25 trillion is a small number. [01:48:41] Speaker A: It would make a big, make a big dent in the national debt, I'm sure. Hey, we've been talking to Dr. Donald Wunsch about his paper Artificial General Intelligence is nowhere near Artificial special stupidity is already here. Policy implications. [01:48:57] Speaker D: Artificial specific stupidity. [01:48:59] Speaker A: What's that? [01:49:00] Speaker D: It's artificial specific stupidity. [01:49:02] Speaker A: What did I say? [01:49:03] Speaker D: You said artificial special stupidity. [01:49:06] Speaker A: Oh, I know some people with special stupidity. Okay, so this is. [01:49:14] Speaker D: I have been guilty of being one of those people from time to time. [01:49:19] Speaker A: That's true. Look, I wanted to change the topic you talk about. Well, let me just say it. If computers took over the world, would we even notice? Are we like the proverbial frog? Put in nice cool water and put on the stove and the water heats up gradually? [01:49:34] Speaker D: Yes, we are. [01:49:35] Speaker A: You make this provocative claim that computers could take over without being smarter than us. What do you mean by that? [01:49:42] Speaker D: Well, we all either know examples or believe we know examples of people who have taken over without being smarter than us. Either our local level or go up the chain, you know, so people have these beliefs about, about people who have taken over. And of course it's much easier for people to take over than computers to take over. But I would argue that computers have already taken over and we didn't even notice. So that relates to this. Nine of the top nine companies by market cap are AI companies. And several of these companies are inventing techniques to get people addicted to their products. And when people are addicted to their products, they buy more of those products or they buy more of the computers that let them say play the games or whatever. Or they, or they buy more of the computers that will, that will do the AI analysis that they need to solve their problems. And so when they buy more of those computers, then they're creating a cycle. You can call it a virtuous cycle or a vicious cycle cycle depending upon what your incentives are. But then they buy more of these products and then these companies have the resources to hire the best people or they might later offload that work to a lot of algorithms to help them design better computers and better software. And this cycle continues. And meanwhile, meanwhile we get more excited about these products and these companies develop the power to influence policy, influence governments. And so nine of these top companies by market cap are AI companies. But then remaining companies down the list, they're insuring these companies, they're doing the banking for these companies, they're providing the transportation, transportation for people to go and, and negotiate between the companies and market for the company. So they're, they're a major driver of the economy. And so it's arguable that the computers have already taken over. So I've got, I'm very much tongue in cheek, but I'm saying that the, that this process can be self perpetuating without any need for intelligent direction of the process. [01:52:14] Speaker A: I have a follow up question. You write that we are dependent upon or addicted to computing infrastructure. Now I'm old enough to remember the challenges of the initiation of television. Television was in the 1950s, was said to possibly ruin the future youth of America. MTV came on. MTV said they did all sorts of really interesting things. But parents says MTV is going to ruin our youth. And I think that gaming was, was that way. Also that people said that our kids are just spending all their time on gaming. I will tell you that my Nephews, my son's two boys are addicted to gaming and they are addicted to the Internet and they have withdrawal symptoms. And I would like to point fingers and say, don't do that, but I've noticed that I have an addiction to the Internet also. You take away my cell phone for a day or two, I suffer withdrawal symptoms. So is this different than things in the past when you say that we're dependent or addicted to computing infrastructure, Is this any different from some of these things that have happened in the past? [01:53:27] Speaker D: Yeah, I broke my cell phone in November and fortunately I broke it during the week of the Black Friday sales. So, so anyway, I, I, it was, it was a Pixel 6 and I was holding out for next year for a pixel 11, but I, I got forced with my arms to get a Pixel 10, but I had to wait a week. It took me a couple of days to decide. And then I, I ordered it and I was close to my flight time, so I didn't have it all shipped to Washington D.C. where I was, I had shipped here. And so anyway, and I've still got apps that are, you know, not moved over yet and have to log in again and stuff like that. And so anyway, yeah, I, I get that and I agree with that. And I would say that I heard it attributed to Socrates. I don't know whether it was him or someone else, but there was somebody that say, that was, was saying that writing was, you know, making us lazy and not memorizing things. And it might have been Socrates, because you might have seen my trope about Socrates. So Socrates was the, the founder of Western philosophy, the, the, the giant from whom the others came. Absolutely brilliant. Had observations that others didn't. He published nothing. They killed him. And then his, his best student was Plato, and I don't know where he did a lot himself, but he published everything he learned from Socrates and so he became a household term in philosophy and, and even things named after him like Platonism, Platonics, stuff like that. And so he became really the founder of a school of thought as a result of him publishing. Well, his best student was Aristotle, and Aristotle made Plato look like a piker by comparison. He published these massive tomes with a lot of original contributions, very deep analysis of a lot of things, became the foundation of a lot of ideas in philosophy and theology and, and was eclipsed the things that went before him. And so then among his students was Anthony the Great, who conquered the world. [01:55:59] Speaker A: Alexander. [01:56:00] Speaker D: Alexander the Great. Thanks for the correction. Uh, yeah, Alexander the Great, uh, who, who conquered uh, the, the what was considered the known world of Western civilization at the time. And so I posted this on LinkedIn to follow my own advice at least to record this idea for posterity. And I said, the moral of the story is, publisher perished literally. And by the way, true or X is tautologically true. And so since we all will perish. [01:56:41] Speaker A: I don't understand what, I don't understand what you mean. True or X. [01:56:45] Speaker D: Okay, Just using binary logic, if you have, if you have a variable and you don't know where it's true or not, and you have one which is always true and you have one or X, you'll always get one, no matter the value of X. So publish or perish, strictly using binary logic, it's guaranteed to be true because we will all perish. So publish or perish is tautologically true. But of course, when we use logic, we usually don't mean the binary sense of words. So we kind of mean, don't. Publish implies, implies, implies parish is what we really mean. But still, anything implies true using, you know, just binary logic, that's true. So what we, what we really mean is something very different. We mean some kind of modal logic, you know, and I studied that. You know, I got an A when I took the class. If I had to teach the class, I could get ready and, and by next summer I'd be ready to teach the class. But, but I won't pretend to pontificate about all the subtleties of the various types of, of logic that go beyond, say, propositional calculus. They're out there, but we use something that is in some ways more subtle and in some ways more informal. So, you know, fuzzy logic, of course, very possible, powerful contributor to that field. But there are other things, quantum logic, modal logic, all sorts of things that are out there. But anyway, that's another example. There's just deep dives that one can make in all these little narrow areas, and some of them are going to, are going to contribute to changes in the way that we use AI, because the stuff that AI is built on, most of them are really pretty much a layup compared to the things that people have come up with that are out there. You know, if you look at the proceedings of one of the main conferences, if you look at the journals, you know, one of them that you help create, one of the best ones that you help create, and so on, and you look at the articles that are out there and the things that people have thought about and published about, and then you look at the tools that have been developed. You know, there, there's a wealth of ideas that have not yet made their way into the practical tools, and some of them deserve to be. And, and so there's enough work for you and me and our children and their children and their grandchildren to work on. And of course, the new ideas will be generated as well. So it's a very exciting time to work in AI. I don't think the, the field is going to be anywhere near milked out in the coming decades. I think that there are centuries of work ahead of us, exciting work for humans to do to advance the field of AI and yes, the AI capabilities. Every year, maybe every month, we will see new things checked off the list that we thought were only the purview of humans, but that will just cause us to realize more and more what the tremendous capabilities of humans and other animals are. So I think on a cognitive brain level that your right wife was right to love your dog, because a dog has a cognitive and emotional processing component of a. Its brain. And so a dog can, I am quite confident from having owned a dog for about a decade that my dog Rudy, who I hired at the time that I was your student, that dog definitely had a sense of humor. No doubt about it. [02:00:51] Speaker A: I've never seen a dog with a sense of humor. How did it laugh? [02:00:54] Speaker D: Well, of course it doesn't have the vocals capabilities that we have, but it had sort of a mischief type of humor. Actually not too different from yours, by the way. You have a mischief sense of humor as well. [02:01:09] Speaker B: Thank you. [02:01:10] Speaker D: But like, thank you for. [02:01:12] Speaker A: You're either thinking very, very highly about your dog or very lowly about me. Or a combination of the both. [02:01:17] Speaker D: No, very, very highly of you and the dog for that matter. But I'll give you an example. I meowed at the dog. The dog cocked its head. It was so funny. But anyway, the cognitive emotional systems there, there's a really interesting issue there because you could design AI to recognize emotions, things like micro expressions. Rosalind Picard of mit, she was at the Georgia Tech at the time she wrote this. She wrote a book called Affective Computing where affect is affect, that's emotional, emotional topics. And so having computing that includes emotional recognition in what it senses and processes and analyzes and responds to. She was one of the founders of that field and she got away from a pure emphasis on affective computing, got very interested in autism and ran with that. And she, she's, you know, one of the giants of the field too. And anyway, so there, there's A lot there. And to say that an AI can feel emotion, I would not claim anything like that. To say that an AI can be developed that will. That will recognize emotions, process them, respond in ways that become increasingly appropriate to what's going on. And right now, that's very primitive. The AI will respond very similarly to a schizophrenic as it will respond to a normal person. And so there are, you know, hideous examples of that. But that's an example where animal intelligence is actually pretty far along, like a. Like the. Our cat, and it's become our son's cat. But our cat will respond to the same mood differently, depending on the person, because the different people have different ways that they welcome or do not welcome the way that that cat would respond. And. And also, the cat has its own set of preferences, too. So it responds differently to my son or to me or my wife. And so I think the cat's a kindred spirit to my wife. It followed my wife home almost a mile or probably over a mile. And I think it recognized a similar personality and followed. Followed her home. [02:04:06] Speaker B: But. [02:04:07] Speaker D: But it, It. It responds very differently to all three of us anyway. The cognitive and emotional system is an area that deserves an enormous amount of AI research because you want these systems to interact appropriately with humans. And if they don't have a sophisticated ability to model and understand what they're sensing about us, then they will be very limited in their ability to be good partners or even good tools for us. So that area is an area that is getting attention, but deserves a lot more attention. So you probably will see sometime in the next decade a company, if it doesn't already exist, a company that's just founded on that principle and just runs with that particular problem and comes up with. With better products along that dimension. [02:05:05] Speaker A: What, what. What dimension is that? Could you elaborate? [02:05:08] Speaker D: Cognitive and emotional processing. [02:05:10] Speaker A: I see. [02:05:11] Speaker C: So. [02:05:11] Speaker D: So cognitive with emotional. So, so Grossberg calls it cognitive emotional system, but he's talking about cognition around the issues of emotion. And really, that's. Grossberg views it as a tool to understand reinforcement learning. You know, engineers can say, hey, I can do reinforcement learning. I don't need to know anything about emotion. Give me a cost function, I'm off to the races. [02:05:38] Speaker C: Right? [02:05:38] Speaker D: And that's built on people like, oh, Paul Werboes, Richard Bellman, you know, so certainly Andy Bartow, Richard Sutton contributed a lot to that. [02:05:53] Speaker C: Harry Hebb. [02:05:54] Speaker D: Oh, well, Donald Hebb. Harry Klopp is one of them. Donald. Harry Klopp was one of the early ones. He Wrote a book called the Hedonistic Neuron. But anyway, a lot of fascinating stuff in the area of reinforcement learning, but [02:06:09] Speaker A: a lot of these things don, don't you believe that they are actually simulating as opposed to duplicating these processes? Of course, the idea is, is that if a computer, if we have an artificial nose, it might say there is an aroma, whereas people says I smell an aroma. There is this qualia aspect that can't be duplicated by machines. [02:06:33] Speaker D: Well, so a machine may be able to deliver everything that we care about in terms of the qualia of sensation, but what the machine will not have is our lived experience, our embodied lived experience. So you know, there are some people that are even like attaching sensors and like wearing the, the glasses with cameras in it and having microphones and trying to capture every moment so that they can, so that they can feed that to a computer that learns from it. But still, you know, one thing that one of my collaborators, we, we had [02:07:16] Speaker C: an interesting NSF grant where we had [02:07:19] Speaker D: a person from the bioengineering department, Georgia Tech, and he had, he had a mouse brain lab where they would take sections of brain and do stuff. But he was pointing out that animals are sensor rich and they can, they'll have an embarrassment of richness with sensors. They'll just have an overload of more sensors than you need for any given task. And then they're processing and what they do with that sensors in some cases might be relatively simple, especially when you think about, you know, what is done not only with insects, but let's say with reptiles and lower mammals, what they can do with, I think even the smallest mammals have pretty sophisticated brains actually. But what, what they can do with a lot fewer neurons than we bring to the table is just amazing. And the sensor richness is part of that. And so, you know, if, if something is disgusting or deadly or something is irresistibly enticing for certain age of animal pheromones or for most animals of food odor, you know, so there, there's a lot that can be done in the field of faction that connects to various, various types of processing that an AI system might be able to recognize and classify. And, and, but you're right, if it, if it only cares about pheromones or food as a cognitive concept, then it's still, it's, it's lived experience, so to speak. Or not they're not a living creature, but it's experience of those things is still dramatically divorced from our experience of that. [02:09:20] Speaker B: Right. [02:09:20] Speaker D: The, these are Things that are. That we would live or die for and that the AI is just another input and another piece of the cost function. And you could raise the value of the cost function, but if you lowered the value of cost function, it would happily march along and do whatever it's doing with it. And so, you know, a human might deliberately lower the value of the cost function of food or a pheromone just because of some life decision they made, but it would not be the same. It would not be the same. [02:09:50] Speaker A: Right. You mentioned the idea of the mouse brain. I maintain that the human brain is more than a computer made out of meat. That there are things that the brain can do which are non algorithmic and computers are limited to algorithmic processes. But that doesn't limit the idea of generating, if you will, brains that are able to duplicate the non algorithmic thing. Have you ever thought about the idea of growing a human brain like on a pig? I mean, they grow different organs on a pig, like livers and, and things of that sort. I don't know if it's livers, but different organs. And I did talk to a guy that does this and I asked him, can you actually grow a human brain on a pig? And he says, oh, yeah, I'm just wondering if there's a way that we could tap into the extraordinary powers of the human brain in order to go to the next generation of AI and have that meat computer do things that AI can't do. And as you mentioned, the cost is just prohibitive currently with AI. If we got meat computers and we're able to do something with that, maybe the cost would go down. Any thoughts on that? [02:11:04] Speaker D: Well, I think that it might have. It might not have been Noel Sharkey. You remember that guy? [02:11:11] Speaker A: No. [02:11:12] Speaker D: Something like Sharkey was his name. [02:11:15] Speaker A: Are you thinking of Richard Starkey, better known as Ringo Starr? [02:11:19] Speaker D: No. [02:11:20] Speaker A: The Beatles? [02:11:21] Speaker D: No, I'm thinking of a guy who was on the Neural Network Council for a while. I think that was the person. He was either doing it or he was talking about it. But there were people that would, that would insert electrodes into the skull of a cockroach or I don't know where that was going into the exoskeleton and down into the brain of a cockroach. And they would have joystick controllers and they would find the pleasure center of the cockroach and then they can stimulate either the pleasure or the pain. And they would teach the cockroaches to do whatever they did so that they control it with a joystick. They can make the cockroach run down a maze. And you can, you can buy kits now to do this. So you could, you could literally buy a kit to. Okay, here's how you put the electrodes into the cockroach. And here's where you can order the living cockroaches. And you. [02:12:23] Speaker A: This is great. I've always wondered if there was a use for the cockroaches around my house. This is good news. [02:12:30] Speaker D: So anyway, they started doing this in the 90s or maybe the late 80s, but this is now, you literally, you could Google it. You could find, you could find kits that enable you to make these. And so the stuff with pigs that. What, what I would say is that there, there is a major ethical issue about Shimmera. So, like genetic engineering of animals to have human capabilities of life along with the animal capabilities. And if you, if you stay away from brains, for example, if you could, if you could have a pig genetically engineered to where it would grow a human heart for, for heart transplant patients, that might be a very good thing. But if you, if you genetically engineered it to grow a human brain, you know, there's ethical problems with that. But the fact of the matter is things like pigs and bears, dolphins, they have pretty advanced brains. And so there would be things that could be done that have their own set of ethical problems. Not as serious as anything when you start getting involved with, with human organs, including brains. But. But still, like, if, if a pig brain were just a tool to replace a supercomputer, if that solved major human diseases, would that be worth the issues of, you know, like, vivisection of a pig? The People for Ethical Treatment of Animals would say it's absolutely not worth it. And on the other hand, somebody who's suffering from a disease might say that it's easily worth it. So I don't know the answers to all that, but I do think that there are things that we can learn from biological systems. And there are people who have played with interfacing biological systems with electronic systems. There are people who've done that. I think that you'll definitely find papers where people have said, all right, we've got these electrodes and we've got like, neurons on a chip. And we can have the neurons on the chip talk to the electrodes and make some computations, and one could advance that line of research considerably. And I would be surprised if there was not a lot, lot of progress on this, even if they're not as a tool, for example, for Parkinson's research. Okay, you, you, you may be able to induce Parkinson's in certain animals. And then research about how the, the signal to the, you know, the, the signal from a brain to a muscle might get mediated and so forth. So there, there's certainly a lot that could be explored in that. [02:15:36] Speaker A: Yeah, this is an interesting idea of using brains instead of silicon. I think I mentioned Niels Bohr saying forecasting is dangerous, especially if it's about the future. At the end of World War II, if you were to ask, what is the future of electronics, what do we need to do? They would concentrate on the vacuum tube. They would say, we need better vacuums, we need better filaments. And not realizing what was happening in the, the future with the advent of semiconductors and currently all of the approaches that we have for artificial intelligence have been due to human intellect. If you look at the milestones, for example, there was, you mentioned Hebb, I think in the late 40s, neurons that wire together, fire together, and Pitts neuron. And then eventually Bernie Woodrow's Adeline, Paul Werbos with back propagation. And that was popularized by the PDB [02:16:37] Speaker D: books and Rumelhart, Rumalhart Hinton and McClellan. [02:16:40] Speaker C: Yeah. [02:16:40] Speaker A: Then we had convolutional neural networks, which was an incredible breakthrough. We had Gans generative adversarial networks and then diffusion for deep fakes. Then we had the transformer, which has been the result of, of which has resulted in all of the large language models that we have today. So I guess the question that I have is what will be the next step? And will it be something that kind of derails where we think AI is going? Maybe if we were able to harness the power of the brain directly and have, if you will, carbon computing instead of silicon computing, maybe that would, that, maybe that would not make necessary all of the big energy plants that are being required. We just don't know what's going to happen. And this is going to happen with the human intellect. There has to be a human coming up with a new breakthrough in the idea of what we can do with computers, be they carbon or silicon. [02:17:41] Speaker D: That's true. I would also say that the engineering of breakthroughs that have already been accomplished but not yet reduced to practice have a very good potential to be something that would yield a breakthrough in AI. So there are, there are neural network models that may deliver great results if they're scaled up. There's ideas in quantum computing that, with a little more push might come through. There are, you know, there's just a number of things I think that breakthroughs in embedded systems may be worth a lot. You know, so like throwing a lot of memory in a problem that has a lot of potential because memristors, they can stay in a memory state without using power for a long time and then you use power to of course, query them in your memory state. But that, that has the potential to be extremely important in embedded computing. And so, for example, one thing, when the, when I read Stan Williams paper, the missing memristor found, there were, you know, half a dozen authors on that and he was talking about the work of, about in the 70s by Leon Shua, saying that there, there is such a thing as a memristor and that it was like a fourth basic circuit component. And then the Stan Williams article, and Stan Williams still continues to do fascinating work. So does Leon Chua. By the way, when I was talking about, you know, pioneers that are still with us, one thing that popped into my mind immediately was, well, if you could imbue in something like paint, a little bit of memory and a little something, maybe not in every particle of paint, but something that could detect a wireless signal, it would be great to have paint particles that you could click a remote control and change the color of the pattern on the paint that maybe, you know, you can also already got TV monitors that are way cheaper than they used to be, that are the size of wall, right? But, but imagine that if that were even another order of magnitude cheaper, so that the whole wall, you could have anything that you wanted on the whole wall and you just were painting the display on the wall and then you could control it with your remote, much like you do with a big tv. And, and it might have nothing to do with a technology tools that I mentioned. It might be something completely different. But the, the idea that, you know, this is part of the idea of Internet of things, but it's intelligent Internet of things. And so if you, if you're, if the objects that you regularly use have even a tiny bit of, of machine learning and memory built into them, then even a very subtle intelligence could make a pretty big difference. You know, you could have pill boxes that know whether they have been opened yet. You could have, you know, you could have sensors that tell you whether the, the food in your fridge is spoiled or food that you've got out needs to be put in the fridge, or, you know, all sorts of things you could have, you know, and it might be just a timing signal, it might be an olfaction signal, it might be a camera in the room that identifies Everything that you have in the room. And then there's something that sends a signal to your phone saying, hey, you left that on the table too long. Or, hey, your dog jumped on the table to eat that food. [02:21:35] Speaker A: Yeah, I'm aware of a company, the CEO is a guy named David Copps. We did an interview with him and that's what he said. We've exhausted a lot of the pros, the corpus of pros that's available to train large language models. So what he's doing is concentrating on images to learn more in terms of transformer models. And so, yeah, you're exactly right. I think that that would be a lot of fun. Let me ask you a question. AI is a revolution. It is changing. It is having an impact. It's going to be very disruptive, but I think it's going to make our lives a lot better, give us a lot more free time, just like other technology has done. Great technology frees us up to do extra things. And so that's true of the industrial revolution. I mean, my great grandparents spent a lot of time in their gardening, hoeing out weeds and things of that sort, clearing stumps in order that they could get a garden. And of course, you don't need that anymore because of the different industrial revolutions that have happened. Is AI comparable to past industrial revolutions? [02:22:42] Speaker D: Yes, but I think it's even more important. So there. I'm with Eric Schmidt. Eric Schmidt and Henry Kissinger and one other person wrote a book I read recently. I wish I remembered the name of it. It sounds like you're familiar with it. I recommend that book. I don't agree with everything in the book, but I think that it is a very well written treatment and it covers some of the history of investment in human learning and education and the implications of AI for that trajectory. Anyway, I think that it's really remarkable. And one thing that it makes the case for is democratizing access to something important. And so that book talks about the Gutenberg Press, which of course used for printing Bibles, among other things. And, and particularly it emphasized the importance of the Gutenberg Bible because before literacy was widespread, then you would have people saying it is written this. And well, if they can read and you can't, you just kind of go, okay, it's written this. And. And then as it becomes more widespread to have access to reading, that was actually pretty disruptive. And, and so, you know, the various technologies that have come along have continued that path to disruption. And so AI is a force that can do both. So I'll give you an example of this. [02:24:25] Speaker A: Well, before you do that, let's go to ChatGPT and say what was the book on artificial intelligence written by Henry Kissinger? Because neither one of us knew what it was. We didn't have the memory to do it. So Chatgpt is thinking now, oh, he wrote two books. The Age of AI and Our Human Future in 2021 and Genesis, Artificial Intelligence, Hope and the Human Spirit. I'm not sure which one you were talking about. I think it was probably. [02:24:54] Speaker D: Does it show the co authors, Eric Schmidt? [02:24:56] Speaker A: Yeah, the co authors of the first one were Eric Schmidt and Daniel Huttenlocher. [02:25:02] Speaker D: I think that's the one I'm thinking of. [02:25:03] Speaker A: The other one was Eric Schmidt and Craig Mundy. [02:25:07] Speaker D: Okay, I don't know which one. I read, I read it. [02:25:09] Speaker A: Okay, well at least we have two possibilities for the listener. Yeah, Kissinger wrote a couple of books. Okay, go ahead. [02:25:16] Speaker D: Yeah, so one thing that I was going to give an example is back to the concept of paradox. So I would say that AI has this paradoxical both lowering and raising of barriers for ordinary human humans to do things. So for example, a lowering of the barrier is that AI can help people code software above the level of ability that they would have without the help of AI. Now these tools right now are not that good. I was talking before about AI making promises and then not keeping those promises and even, even lying to you that said I developed this and it didn't and there was nothing in there that was. These were pretty sophisticated AI coding tools. This was like this year. This was not, this was not in the past two months but it was 2025. And these AI tools by one of the major companies was literally lying about that. It developed this subunit that was doing the things it was told to and it just was a pass through box that did nothing. And so that's a good example of artificial specific stupidity released by one of these trillion dollar companies. But on the other hand these tools are very capable. Like if you wanted to develop a game interface to play your favorite game. And so the inner, particularly a two player game game where, where it's just, just so that you and I can play a game of Go or Chess over the Internet, these tools could do a pretty good job of that without, without making a lot of the mistakes that I just mentioned. So it's lowering a barrier of entry because I don't know about you, but it would take me a long time to program something like that. I'd have to, I'd have to relearn stuff in classes that, yeah, when I took the class I got an A, but I never had to do it for a living kind of thing. And so I would not be that good at a task like that. And these tools are better than me at doing something like that. But so on one hand, it's lowering a barrier and it's opening accessibility to a large range of people. So there's people who never took the class and never got the A, who can, who can do some contributions in areas where just five years ago, forget it wasn't possible. But you might remember when Robert F. Kennedy announced his campaign for President of the United States and he said that he might, he might pick either Jesse Ventura or Aaron Rodgers for his vice presidential candidate. [02:28:04] Speaker A: And Jesse the Body Ventura, the professional [02:28:08] Speaker D: wrestler, or Aaron Rodgers. And so anyway, I decided to make a little joke. And so I told one of the AI tools. Please generate a picture of the New York jets offensive line with earpieces and dark glasses and Secret Service paraphernalia, including Uzis. You know, give me some examples. And it refused to do, said, yeah, no, you're, you're violating the standards. And. And so I took out the Uzis and then it did it. But one of the images was pretty funny and I, I did wind up posting that image just to see whether people laughed at it or not. But there was one image that showed a young woman. She was clearly an athletic young woman, but she was probably about at most £120 and you know, very attractive young woman dressed like a Jets offensive line person. I'm not athletic at all, but I weigh about 220 pounds and most of my strength is in my legs. If she's the only thing preventing me from sacking Aaron Rodgers, and I didn't have any gentlemanly politeness, I'm sacking Aaron Rogers, you know, so, so the thing is that there were famous examples where other mistakes were made, where one of these systems to generate images was asked to show presidents of the US from the 1800s. And they had people of, of genders and ethnicities that were not representative of presidents from the 1800s. They made all sorts of mistakes. Later, I asked a different tool and it had no problem with the Uzis. It was, it was show people with Uzis, but it was making other mistakes. It wasn't at all like the, the image intended in the joke. But so for example, if I, if I had the resources of one of these companies, I could just say, turn off that filter. I want you to make this image. There's nothing illegal about the image, nothing immoral about the image. And so you have filters that don't let me make that image, but you're wrong. And so you turn off those filters for me so that I can make the image that I have in mind. So there are entities that control our freedom of expression because they control the tools of expression. And so they limit our freedom of expression by their control of those tools. So AI technologies, technologies are both increasing our ability to do things that we would like to do, and they're also constraining our ability to do things that we would like to do. And there are people and companies that are controlling these processes and they, they differ. So one, one piece of the puzzle that I agree with is, is Lecun's position that open source is a great antidote to this, that, that having open source AI software prevents you from getting a monopoly among big companies and governments over AI capabilities. That if, that if entities, at least some of the entities that are good at developing these tools at scale are also releasing the tools, then others will be able to leverage the tools and will have more cross fertilization of the ideas, but furthermore, more freedoms related to AI systems. But there are also dangers in that idea because some of these AI tools can be used in nefarious ways. And so LeCun is an optimist, optimist about this. He's saying that yes, we have good and evil in our, in our drives and in what things that we try to do with, as technology is made available to us, but that overall the, the trends are positive. Another person that writes about this is Jordan Peterson and there are some others. There's a Harvard professor, Steven Pinker, who writes about this. And so as an eternal optimist, I like to believe these ideas. And I do think that we're flawed and fallen and need the forgiveness of God and the help of God to become better versions of ourselves. I also think that there are tools that we've been blessed with the ability to discover whether it's AI or nuclear energy that we can use to solve our problems and make the world a better place, or that we can use to destroy the world. And I think that we have a stark choice. We must either do the former or sucked into the latter. I think that's true with nuclear energy and with AI. [02:33:48] Speaker A: This is the topic I want to address in our next section. So let me just remind people that we're talking to Donald Wunsch. He's an award winning professor. He's a Mary Kay Finley, Missouri Distinguished professor of Electrical and Computer Engineering at Missouri University of Science and Technology, where he is the director of the Kummer Institute for Artificial Intelligence and Autonomous Systems. And he wrote a paper called Artificial General Intelligence is Nowhere Near Artificial. Specific Stupidity is Already here. And then Policy Implications is the final part of the title. So, Don, we started to talk about the use of AI in weaponry. I wrote a book called the Case for Killer Robots, where I learned a lot about. [02:34:37] Speaker D: I cited that book in my paper. [02:34:39] Speaker A: Oh, as. As you should. Okay, thank you. And one of the things that is true is that there are totally autonomous AI systems. They're relatively simple. The one that I remember is the Harpy, which is a missile that is launched by the Israelis. It loiters over a battlefield, just flies around at random until it is illuminated by radar. After it's illuminated by radar, it knows where the signal came from and it homes in on the source of the radar and goes and takes out the radar installation in a kamikaze attack, if you will. And this is totally done on artificial intelligence and it does have the capability of doing this totally autonomous. So you argue that AI is already dangerous, but not because of autonomy. I think autonomy in simple systems is viable, but the reason that we have danger is because of human use. Why is autonomy a red herring in artificial intelligence? [02:35:41] Speaker D: Okay, well, we talked about that a little bit before. So autonomy is a red herring because people do some fear mongering about AI becoming smarter than us and taking over because it is smarter than us. And, and so that risk is not zero. But the reason that that risk is not zero is because AI doesn't need to be smarter than us to take over. So that's a little bit of a tongue in cheek argument, but it's also got some validity. So that's part of my critique of that. But my more important critique of it being a red herring is that yes, there may be risks that come from AI with increased capabilities. In fact, there are, there are risks of from AI with increased capabilities, but there are more salient risks that are here right now. And, and so, for example, one risk of AI that is here right now is like the type of system that you described. If we were flying low enough to pick up the signal from police radar scanner that is trying to catch speeders, that's going to be too bad for that cop, you know, it'll come down and it'll destroy his car and it'll think that it's destroying a radar unit of the opponent in a war, but it actually destroyed an innocent bystander that happened to be generating a radar signal. And you know, drones have become a source of about the majority of casualties in certain regions of the Ukraine war. And some of these drones are undoubtedly finding other targets. A lot of them are controlled by humans. But there was an interesting article that I cited in my paper that claimed that the Russians were fielding their systems with a great deal of autonomy and not so much concern of whether they're hitting the right targets or not. And so that's by Jane Pinellas and her co authors and it's cited in my paper. And Jane Pinellas, she's with Johns Hopkins, but she used to be in charge of all AI acquisition for the entire US military. So she's somebody who's in a position to understand quite a bit about AI policy. So I, I cited her excellent article and it was just raising concerns about the, the willingness of, of certain geopolitical players to ethically field their AI systems and was saying that this is a concern because there are entities that do not follow the same rules that we accept as standard rules. [02:38:41] Speaker A: Well, with this I maintain that there's two types of ethics for engineers. One is the design ethics. We want to make sure that the system we develop does what it is supposed to do and nothing else. Then there is the end use ethics, which is the decisions made, for example, by military commanders. And these are two separate things. [02:39:02] Speaker D: Yeah, but they can be intertwined, in other words. [02:39:06] Speaker A: Oh, they can be intertwined, yes. The boundary between the two are indeed fuzzy. So you discuss AI weapons from drones to policing. And where do you see the hardest ethical line drawing problems in the use of AI weapons? [02:39:24] Speaker D: Well, their proliferation is certainly a big issue. I'd put that as way, way up there among the issues. So you know, you have this race to make the drones cheaper and more robust, easier to use, and, and you know, they're deadly, they're generating a lot of casualties and maybe the biggest hot [02:39:50] Speaker C: war going on right now. [02:39:53] Speaker D: And so there's this pressure to get the technology out there. [02:39:57] Speaker A: Right. [02:39:58] Speaker D: But the ability, you know, just, just yesterday there was this horrific attack on, on people celebrating Hanukkah in Australia. And they were just using old fashioned, you know, machine guns or, you know, automatic weapons. And you know, due to the action of a selfless hero, it was somebody came in and disarmed one of them and as a result the casualties were [02:40:29] Speaker C: much lower than they probably would have otherwise been. [02:40:31] Speaker D: But, but I think that most people of goodwill agree that to hate, much less kill People, because of their religious persuasion is just always morally wrong. [02:40:50] Speaker A: Well, this gets to the idea of the definition of a just war. [02:40:53] Speaker D: Yeah. And so a just war as a concept for. From Aristotle, not Aristotle, Augustine wrote about it. And so, you know, was it a just war to stop Hitler from what he was doing? Of course it was. You know, is it a just war to. To stop the shipments of drugs across the Caribbean Sea? That's a topic of hot debate right now. And then was it a just war for Russia to invade Ukraine? You know, most people in the US Would say probably not. Surely people in Ukraine would say probably not or definitely not. You know, and so that's an issue that gets affected by this idea of the, of the boundaries, ethically technical boundaries around developing AI systems. So if you make something that is effective and cheap, then it is going to be widespread. And so I would say that some of that's inevitable. If you make things that are effective and cheap, but they also have some safeguards built in, they then that that could be a really beneficial development. So in other words, if it were impossible for these drones to hit certain types of targets, then that would be like, can the drone detect. If it is being. Can a drone. If a drone could detect that a human is below a certain age, If a drone could detect the, you know, reliably to detect a difference between, say, a family out shopping and a group of soldiers on a mission, that's a really tough nut to crack. Because you, you know, some of these are like, one part of US Doctrine is that there needs to be a human in the loop before a kill decision is made. [02:43:12] Speaker A: Yes. [02:43:13] Speaker D: And that's part of this article, this saying. The article by Jane Pinellas that I mentioned is saying that our opponents are not always using designs like that. But, but if there were, if there were safeguards, is that is something like that. Adding value. There's a lot of things that could [02:43:36] Speaker C: add value [02:43:39] Speaker D: on the ethical front, not just on the performance front, but on the ethical front. [02:43:44] Speaker A: Yeah. So I think, I think one of the big things is in terms of the design ethics is the vetting to make sure that the system does exactly what it's supposed to do and nothing else. This is a reason, I think that the Harpy works well. This is the reason I would have no problems riding in a driverless taxi in Austin, Texas or San Francisco. And one of the reasons is as the complexity of a system increases, the ways it can respond go up exponentially. I wrote a paper with this with Sam Haug, and when you have A manageable number of parameters. You can vet the AI, but when you have trillions of parameters, like in the large language models, you. You can't vet it. It's just. It's almost non vetable. That's the reason we have these things called hallucinations, a word I don't like, by the way, because it anthropomorporizes the AI. So, yeah, I think that you're right in saying that we got to make sure again that the AI does what it's supposed to do and nothing else that it needs to be vetted. So, Don, we've been talking about the military. Let's talk about another national security situation, which is cybernetics. At Baylor University, the administration makes us all take courses in cybersecurity. They want to make sure we're not fished. And I have learned some things in there. I learned what spear phishing was, which is where they come out, come after you with, with all your personal information. And that's really kind of weird. But the biggest thing in cybersecurity, as I understand it, is not breaking codes. You can't break a code, otherwise Bitcoin wouldn't work because everything is encrypted and you can't break that encryption. So it's always people that are the weak link. So you wrote your paper, which we're talking about, and you said that AI enabled cybernetics may already have killed people. [02:45:36] Speaker D: Yeah, AI. [02:45:37] Speaker C: AI enabled, not cybernetics. Cybernetics is actually a very positive term and it's an alternative term to other subsets of AI. Norbert Wiener and so on. [02:45:51] Speaker A: Yes, I'm sorry, I meant say cyber attacks may have killed people. Yeah, I misspoke. [02:45:57] Speaker C: Yeah. So, yeah, AI enabled cyber attacks probably have already caused some deaths. So I gave some citations of some, you know, fairly early published examples in my paper. But certainly there, this is a big risk. And so when people talk about risks of AI, I think a lot of the things that get trotted out, like the risk that AI will get smarter than human beings and take over the world. The. These I think are a red herring that take our eye off the ball of real, actual demonstrated capabilities of AI to do things that we don't want done. You mentioned spear phishing, and that's a really great example because the burgeoning of large language models makes it less expensive to design attacks that will, that will fool humans and humans, if not the weakest link, which I think they are, certainly one of the main weak links in cyber security is the ability to fool A human into giving you information. It's so pervasive. And for example, you can get a legitimate email that tells you blah, blah, blah, this or that, click here to do this. Many of these are actually legitimate. But I think that that particular type of communication is either going to go through more layers of protection which is already happening, or it'll just become obsolete to do it that way. Because what I'll often do is if I receive such email telling me to go to my bank or Amazon or brokerage firm or even my employer if I have the slightest doubt about it, instead of following the link, I will just go to their website that I know and not have the communication initiated to me, but initiated by me so that I can go to the website myself, I don't need to follow the link in the email and so on. But of course there are people who know much more about how to protect yourself. And the big organizations like my university and Baylor and when I was in the federal government and big companies, they will require you to take training and it's kind of annoying because they periodically require you to trust, take the same training. But I encourage patience because we can all use reminders. And also the attacks do evolve. And so anyway, but this is a major risk and the vulnerabilities are tremendous. Some of the actors are funded by large entities like state funded entities or non governmental organizations that have a lot of resources, sources, some of them because they're, you know, organized crime enterprises that have succeeded in cyber attacks and use those resources to launch even more sophisticated attacks. And so spear phishing is a great example, but it's one of many. And so anyway, that's, that's a good example of misplaced assessment of AI risk versus real risk world AI risks that actually are proven to exist, known to exist, have already caused damage and are going to do more damage. And so these are things that deserve countermeasures, attention and when we know where the attacks came from, consequences as well. So there's several levels of dealing with this type of AI risk that are extremely important. And, and part, part of this also is the, the big players in AI making sure that they take responsibility. That goes with the, with the resources and the market share they've obtained. So I'm bullish on these companies. I have stock in them, I want them to do well, but I also want them to take responsibility for their positions of influence and power and make sure that they're taking the appropriate precautions against the real risks of AI. And so one of the demonstrated ones relates to cybersecurity. Another one relates to reinforcement learning with the human in the loop. And this is one of the hot areas of research is reinforcement learning with human feedback to have humans help to fine tune the systems to perform better, but also where the system learns to give you more of what you want. But as we hopefully accumulate some wisdom in our life, what we want is not always what we should get. And so using reinforcement learning to rapidly give us more of what we want is actually not healthy for us. And AI is very good at it. And we're belatedly realizing that this cause all sorts of bad effects. And so the companies that, particularly companies that use AI for entertainment, but also AI for news sources, for example, the reinforcement learning can create a vicious cycle which will give you a distorted view of reality or, or just a disordered use of your time and resources and so on. It can be very damaging. So these are some not hypothetical risks of AI. These are already demonstrated system wide weaknesses that have resulted partly because of the way that AI has been deployed without, and it's easy to be a Monday morning quarterback, but without really thinking through the consequences of that deployment. [02:52:24] Speaker A: So you mentioned that, well, you did some time at the nsf, so you work for the government and also at your University of Missouri and you were required to take a course in cybersecurity. I also took a course course, as I mentioned here at Baylor. And one of the things they do, which I appreciate, is they let you take the test first. And if you take the test and you pass it, you don't have to listen to the course. Well, it turns out that all of the questions that they asked for cybersecurity can be answered by ChatGPT. And ChatGPT gave it, gave it like 100% correct answers of the questions about cyber security. So I will not admit nor confess the government gives me the right to remain silent, but not the ability. So [02:53:22] Speaker C: yeah, that's a good one. I'm going to remember that line. [02:53:27] Speaker A: Okay, I actually got that from the book by John Edwards, Words or something about artificial stupidity or something. Or do you test positive for stupidity? I think it's a recent book by him, but he mentioned that. Do you think that we can ever train AI to detect phishing? [02:53:45] Speaker C: Oh, certainly we can deploy AI solutions to these problems. So we could have an AI deployed to detect any variety. Well, I won't say any, but many varieties of cyber attacks that we know exist in the field. But again, we have to look at unintended consequences because the line of defense can Create an opportunity for a vector of attack. So in other words, you could say, okay, I know that these systems are doing this and so do something that causes that system to make a mistake. Yeah, so there are measures and then countermeasures. And also some of the measures can themselves become a vector of attack, like a denial of service type of situation. So it's, and it's just a very challenging problem. The, the, there's a lot of things that work that way. And particularly in cyberspace, there's a lot of things where, where somebody does one thing and somebody does something else, and then the measures and the countermeasures ratchet up. And, and so, you know, you want to make it more expensive for the opponent to do something you don't like, but then their countermeasures will make your countermeasures less effective. And also there might be a third opponent who, who figures out some way to exploit what you're doing itself. Another thing that's, that's kind of a standard discussion in cybersecurity. And this is not my primary expert expertise, but I've, I've done a little bit of publication in the area and I've talked to some people and I've had to take these mandatory classes. But one of the tensions is between ease of use and level of security. So you can make systems easier to use by lowering the security barriers. And then people will more seamlessly use the system, but it'll be less secure. Or you can raise the level of security, make it harder to use the system. If you make it enough harder to use the system, people will have workarounds. So a classic example is, you know, a list of passwords in your drawer, you know, and, and that's there, there are some good ways to get around that. But, but so a great thing that has been done with AI in my opinion is the things with user consent to use, things like voice or video, face or fingerprint. These technologies have been around for a while, but they've gotten a lot less expensive with AI. So you buy a high end smartphone and it can do, or probably even a moderate smartphone, you can deal with voice or face or fingerprint, and that gives you another level. And then the multi factor authentication, that's not as much an AI thing. The other three are certainly using the pattern recognition that we've developed with neural networks that we as a field have developed with neural networks. So there can and should and will be more of these things. [02:57:27] Speaker A: It seems that legally that everybody should have the equivalent of an nil you know, the name, image and likeness that's given to, for example, college football players, that they own their name, they own their image, they own their likeness, it [02:57:41] Speaker C: should well not given to them. They partly were born with it and they partly earned it. In other words, they were born with their name and likeness and then they earned the fact that somebody wants their name, image and likeness. [02:57:53] Speaker A: Exactly. So I'm just wondering whether that idea can be extended to us, the population where we own our name, our image, our likeness, our voice, etc. Yeah, well, it's an interesting idea. [02:58:09] Speaker C: Yeah. And so there are a number of policymakers who have pushed back at the big AI companies for basically leveraging a lot of material, some of which one might have had a reasonable expectation that they would be compensated or at least their permissions would be sought. And so there's likely to be continued discussion and movement around the policy issues pertaining to privacy and ownership of, of, you know, to what extent this crosses into intellectual property and so on. And you know, these user agreements that are written by lawyers and they're 25 paragraphs long and written in 10 point font that you click on before you can even use websites or software or things like that, that what gets put into those may be subject to some change there. So I tend to like a laissez faire approach to regulation, but there are times when it's really needed. And so what I'm not a fan of the idea of patchwork of 50 different regulations across the United States, not to mention the territories that, that our AI firms have to navigate. I do think that it is probably good for the most part to have a federal approach to these issues. But on the other hand, there are things where we have not had our eye on the ball. And the problem is that a lot of politicians can only spell AI if you spot them the A, you know, and so, so anyway, and they'll be, a lot of them will be the first to admit that, but, but so we, we need to help them with that. And, and I think that scientists and engineers really can do a lot to help the progress of the field. And, and indeed in countries where, where engineers are bigger segments of the, of the population of policymakers and politicians with power, there are what I believe are some more forward looking approaches to the infrastructure that AI needs. So for example, more energy generation capacity is likely to be a difference maker in the field of AI. And China has at least a couple of dozen, I think more than that by now because this information I heard was more than a year ago at least a couple of dozen of new nuclear power plants being stood up. And you know, that was a conversation that I remember from, oh, at least 17 years ago, if not longer, actually longer about 18 years ago at the federal level in elections about should we have 100 more nuclear reactors? And I think we answered that wrong. We should have gone forward with, if we had 17, 18 years ago, committed to 100 more nuclear reactors, they would be all online now. And that could have made a pretty big difference to one, the global warming carbon footprint. But two, the AI infrastructure is becoming clear, is going to make enormous demands and we really need all of the above approach to that. So we're straying from the topic cybersecurity, but, but the intersection with policy is really important. And when I have colleagues that, that in my engineering encounters who think it's a good idea to construct legal language and share that, I ask them what would they like? Lawyers designing their engineering systems. And so I think that I have tremendous respect for my attorney colleagues and I find them just brilliant at wordsmithing a document in a way that nobody could without that specialized training and scientists and engineers bring a lot to the table too. And these contributions are really important policy. I think I told you about my news site, my curated AI news website. You put a link to it recently and I deliberately screen out stuff that has too much hype. But one that I'm going to post, my student may have already posted it for me, came out today on Fidelity Investments was a discussion about AI. And some of the stuff they said about AI was nonsense. But when they looked at a business analysis of the companies throughout the whole infrastructure, the AI companies, the energy companies, the other infrastructure companies that are needed to stand up data centers, it just had a lot of valuable information that most, most computer engineers, computer scientists, AI engineers would not just trot out in an article like that. So I asked the student to post it. So we. It's very multidisciplinary. You know, you have philosophers and ethicists and theologians, you have engineers and computer scientists. You have, you know, the Nobel Prize in Chemistry was the one of the three that were awarded to neural network people this past year. And so there are people across all sorts of disciplines that have something valuable to contribute. On that website that I mentioned, there's an article by Michael Jordan arguing that you need the economists in the conversation. And really, really wonderful article. I recommend reading that one very highly. So it has become an. Not to mention, of course, the mathematicians have just been absolutely critical to this. So the many Disciplines that contribute to the field of AI. That is absolutely an essential part of the picture, and that's part of why I'm. The title of that article you referred to. Artificial general intelligence is nowhere near artificial. Specific stupidity is already here. The abilities that, that go into human intelligence are just extraordinary. And yeah, you can get an AI system that can pass the tests that most of the people in most of these professions have to take, but that is not the same. It's not anywhere close to what these people actually do. And so, yes, these systems are going to pick up capability abilities that they did not have, that AI systems did not have just a few years ago. But that's not the same as replacing what a human does. There's so many things that a human does that are just not anywhere near ready for prime time. So one of the earliest critiques that still rings true was by Rodney Brooks, the MIT professor, who was saying that these systems, if they're not embodied, they're missing something that, that just can't be replaced. And I agree with that assessment. It goes beyond that, though. I would broaden that critique, and I do. In the paper, I say that if. If a system is claiming to be AGI, Artificial General Intelligence, and the. The problem had to be set up so that it would interconnect with the capabilities of a computer, if it had to be. If you had to stack the deck so that a computer can handle it, then that already disqualifies it. So if some urgent thing happened and you had to get up and run out the door or escape out the window or crawl under your desk or, or call the cops [03:07:24] Speaker B: or pull a [03:07:24] Speaker C: gun out of your drawer or scissors or a letter opener or something extraordinary happened. Humans can handle that. And even a fly, like have you ever gotten onto a plane or into a car and a fly flies in and you shoo it out, but it doesn't go out, so you shut the door and then it goes hundreds of miles with you. Yeah, and then it goes out and it's not where it was, and it's not going to get back to where it was before it went in there, and it's just going to have to deal with it. And even a fly can do that. Right. But if you remove too many assumptions from an AI system and make those assumptions not true anymore, it can be very brittle. [03:08:16] Speaker A: Well, Don, we've strayed far from the initial question about could AI detect fishing? I had a lot of comments on your monologue there, but I think I'll pass on commenting on It. Because of time elements. Let me ask you this. Have you ever been fished successfully? [03:08:38] Speaker C: I have, Yeah, I have. [03:08:41] Speaker D: So that's sort of like. [03:08:45] Speaker C: So I don't know if you ever met him. George Allen was our dean at Texas Tech a while after I got there. So of course, long, long after you graduated. But. And he had, he had some meetings on campus, and then we had a nice strategic retreat. [03:09:04] Speaker B: But. [03:09:04] Speaker C: But I think on the second meeting on campus, fortunately, and different from my usual habit, I got in there a little bit early and somebody came in there a little bit late, and we'd have been having it around the room, and somebody said, oh, sorry I'm late. He joked, oh, that's okay. We were just sharing our most embarrassing moment. The person didn't fall for it. [03:09:30] Speaker B: But anyway. [03:09:31] Speaker C: But yeah, so, yeah, I've. I have fallen for it. I have fallen for a spear phishing attack to. It purported to come from somebody who I, who I knew, and, and I just didn't check it carefully enough. But I've also caught some. That other folks fell for and alerted them in time before too many, much more damage was done. [03:09:57] Speaker A: So some of them are very, very good. [03:10:00] Speaker C: Yeah, they're really good. They're getting, they're getting quite sophisticated. It's no longer that, you know, I'm a. I'm a prince from some different continent and want to transfer $10 million and I just need your help to do it. No, it's a lot more sophisticated than it used to be. And so, yeah, that's why if you, if you're asked to click on something, if you know the place to go, you might go independently there instead of following the link. [03:10:34] Speaker A: So we need to come up with a saying, click and don't click. And we need a fancy saying to remind people not to click. So let's work on that. [03:10:45] Speaker C: Okay, well, and that pertains also not to cybersecurity, but to the other risks that I mentioned. You know, clickbait. You know, something that just, you know, people are trying to steal your time, but they're also trying to steal your money. But there are. People are also trying to steal your time. And so the clickbait problem is a real big one. And so, for example, in news, you know, it's very easy to see a headline that makes it look like it's a real story. And you go there and there isn't anything there. It's something that is designed to hook you in. But you get there and there's pablum there. You feel yourself get Stupider by the minute as you read it. And so that's a pretty big problem with AI because the system will learn things that are more likely to get you to click and then, and so you can, you can get this. Sometimes you can give it feedback. So I will often give, give these feedback and I'll say no more from that. From that source. Like, oh, something that's behind a paywall and you got to read two sentences of the article and then you can't read more until you buy it because [03:12:09] Speaker A: you got to pay for it. [03:12:10] Speaker D: Yeah. [03:12:11] Speaker C: And so I'll just tell it, okay, no more from that. No more from that newspaper. No more. Even if I respect the news source, if it's, if it's just going to be a clickbait thing, I'll frequently choose the thing that filters it out. So that's not an AI solution, but it was an AI generated problem. It's a manual solution to a problem that was generated by the AI. In other words, it, it knows that I like that kind of headline. And then I have to manually sending me something that is useless to me, so. [03:12:48] Speaker A: Oh, that's great. Thank you, Don. We're talking to Donald C. Wanch. He's a. Mary Kay Finley, Missouri Distinguished professor of Electrical and Computer Engineering at Missouri University of Science and Technology, where he is also director of the Kummer Institute center for Artificial Artificial Intelligence and Autonomous Systems. I've known Don for how many years, don? 30, 35 years or something like that? [03:13:12] Speaker C: I think it's 40 as of this year, maybe 41. It's between 40 and 41, I think so, Yeah. I came back from, I came back from presentation by Robert Heck Nielsen and there was a group of people who showed up at Boeing when I gave a talk about it. And you were also a speaker at that presentation. And another person who was there was Bernie Widrow. Oh, so I. [03:13:46] Speaker A: That was a Boeing, right? [03:13:47] Speaker D: Yeah. [03:13:48] Speaker C: I believe I met you and Bernie Widro at the same day. And you, you both became important mentors to me afterwards. Words. So anyway, that was really a wonderful day in my life, a very memorable day. And I think that was the summer. That was the summer of 86. So it will have been 40 years as of this summer, in my opinion. [03:14:13] Speaker D: Wow, that's amazing. [03:14:14] Speaker C: I may have met you briefly before that event, but it's at that event that we had a longer conversation and it was. [03:14:22] Speaker A: I didn't make an impression before that. So. Don, we have known each other longer than the age of most of our students. [03:14:30] Speaker C: That's Right, that's right. [03:14:32] Speaker A: That's pretty bad. [03:14:33] Speaker C: You became my advisor in 1987. [03:14:36] Speaker A: Uh huh. Yeah. That's great. We had a lot of fun. [03:14:39] Speaker C: Yeah, we sure did. [03:14:40] Speaker A: The last thing I want to talk about is regulation of government. One of the things that you mentioned previously was that you favored national regulation of AI. I certainly don't agree with that. I think that states need the right to do this. Now the problem, of course, is regulatory compliance. And my son runs a business, an accounting business that goes across many states and he says keeping up with the, with the regulatory stuff is just incredibly terrible. But we do have a lot of things coming out. We have, for example, Grok, who can. [03:15:16] Speaker B: Right. [03:15:16] Speaker A: Who can synthesize pornographic conversations. We have OpenAI, which says it's going to start to do adult content. And you know, in Texas you have to have age verification before adult sites. And also the other example is marijuana. You know, some states say it's okay, some states say that it's not okay. So I recently went to Denver, Colorado and there was this big sign up, it says cannabis. And you could go in there and buy everything from tobacco, marijuana, tobacco, to probably gummy bears. And that's totally illegal in the state of Texas. And I like that. Autonomy at the local level. The price is of course, the overregulation of things and the degree to which compliance must be adhered to. So could you elaborate on what you think that things should be just done on the federal level? [03:16:12] Speaker C: Yeah, so for one thing, there was a law that got passed the legislature in the state of California, but was vetoed by Gavin Newsom. And I think that he made the right decision to veto this law because it included sanctions, punishment, including possible jail time for developers of AI. And these were things that were actually more the responsibility of the companies that employed these developers. And so there were some pretty big problems with that proposed legislation. So the federal government has a role of regulating interstate commerce. I do think that there's an appropriate role for states in certain regulations, but they have to be bounded so that companies know what to work with, what to deal with. So yeah, there are certainly very different attitudes in different states. Different regarding something like pornography, regarding the protection of children. There are no really easy answers to these issues. I think that at least certain types of regulations need to be uniform nationwide. [03:17:48] Speaker A: Oh, of course, of course. [03:17:51] Speaker C: But it's a very nuanced conversation. And, and so I do think the IEEE has an important role to play, particularly the intersection between energy. [03:18:03] Speaker A: Okay, let's talk about, let's talk About IEEE then? [03:18:06] Speaker C: Yeah, I don't think so. I, I served on a committee. One of the other members that you know is Paul Werbos that was on that committee that produced a report right around a year ago about, about energy and AI and the demands that AI is making on the energy grid and emphasized, you know, I was lobbying for it to. Not lobbying, it's not lobbying. We're non lobbying organization. But IEEE has its lobbyists. But this was a committee that talks to the IEEE and then the IEEE goes and makes its policy recommendations informed by the committee committee. [03:18:48] Speaker B: So. [03:18:48] Speaker C: But with the blessing of my management at the university, I joined this committee and encouraged the other committee members and they agreed to insert into the report language about the importance of having the power generation and transmission capability that these bigger AI systems are inexorably going to demand. Because if you don't do that, you hamstring the companies. And so that's one thing that is there's so many nuances to AI policy and so there's the infrastructure, the costs. I think it's quite appropriate. And this is something that, that can be done at the state level, but also there are important conversations about doing it at the federal level. If you plunk down a bunch of big data centers and they, and they use a lot of energy, they should be paying for that increased consumption of energy. That should not be. [03:19:59] Speaker A: Not the consumer. [03:20:00] Speaker C: Yeah, that should not be floated among the consumers. And so, and this has happened and is there are conversations about correcting that. And I think that we're sensitized as society to that now to a larger degree than we were even two or three years ago. And so there are conversations happening at the federal and local levels. I spoke with some of our utility people, you know, in Phelps County, Missouri, where my university is, and they're very attuned to that. So if somebody wants to open up a big, big data center, they're going to be expected to pay for that new resource coming online. But to what degree? That's just good management. And the people that run the utilities understand that are sensitive to that and starting to respond to that. Versus what, to what degree these are codified into law. These are things that pretty quickly need to get sorted out. And so movement is going in that direction and the IEEE is helping with these conversations. [03:21:08] Speaker A: So I have a thought here. At the end of World War II, if you ask technology where it should go, I think at least electrical engineering, they would say, well, we need better vacuum tubes, we need better filaments with greater reliability we need better vacuums in there and nobody recognizing the fact that semiconductors were just on the horizon. I'm wondering, and this is, this is just wondering if sometime that they're going to figure out a new way. People, research scientists, research engineers, are going to find out a new way to train without the consumption of all this energy. There was speculation when Deep Seek and China came out that they had found a new way to do that. I think that that's pretty well been debunked now. They haven't found a new way to do the training, but this is forecasting the future. And I think would be really, really great if we could train these transformers with a lot less power. I wonder if that's ever going to happen. [03:22:10] Speaker C: Well, we have an existence proof between our ears. And so if I haven't mentioned it before, I think I've mentioned it in my paper that we use about 20 watts. And in the battle days before LED light bulbs became common, 20 lots was a dim bulb. And so dim bulb is no longer an insult, but a compliment. What we can do with our dim bulb worth of energy is not a dim bulb capability at all. It's brilliant and a gift from God. And so I do think that we, you know, we have a gift of nuclear energy that we can use. If we don't use it for good, it's probably inexorable that it'll be used for evil. So we need to use nuclear energy for good. And I think that's very important. But I think that it is at least important, at least as important to learn from what we see in biology. And so I gave the examples of insects before. So my pet peeve is when articles talk about the human brain and modeling the human brain. We can do a lot by learning from a bug's brain. But a goal that is too lofty in the near term, but should always be like a star, like a guiding star, is the fact of the remarkable level of computation that can be done with 20 watts. And so if we can do more biologically inspired computing, the reason that we do these massive vector matrix multiplications is they work and engineers understand them. And, and so these things that happen in biology, the capabilities are amazing, but they're also messy. It's harder to write theorems about them, it's harder to guarantee what their behavior will be. And so the progress is of necessity going to be slower. And it's also so multidisciplinary. So you can get an electrical engineer together and talk to a mathematician. Repetition. And you can get really far because the fields are pretty close together. But if you start pulling together these things that you know are relevant, but where the languages are different and where there's so much that needs to be learned and so so far that one has to come, it's just more challenging, it takes more time. And this is another example of the importance of what humans bring to the table. And can AI help? Probably yes. So can you, can you summarize something from Eric Kondel's Principles of Neuroscience, which is a book about the equivalent of this whole stack of books, and summarize points that an electrical engineer might need to know? Maybe an AI could help you with that a lot. But there are still points of connection that human beings currently can make better. And so that's why I'd say to somebody, one the policymakers is to encourage these multidisciplinary directions to students. I'd say be curious and learn a lot of different things. You'd be surprised at what you might get out of, say, a history class or a philosophy class. Things that you never dreamed would be related to AI, but very much are. And so those are a couple of things that can emerge from these type of discussions. [03:26:04] Speaker A: Well, you know, all of the advances of AI have been from the human intellect, from Hebb's law to Bernie Woodrow's Adeline to Paul Werbos's error back propagation, to Gans and diffusion models into transformers. These have all been the product of the human brain. AI is never going to create a better AI and I think it's going to be interesting. The thing I like to forecast is that the future is not forecastable. So you would like to hope that some of these things would happen in the future and that we wouldn't need all of this consumption of power. And your point about the human brain is, is. Is well taken. [03:26:43] Speaker C: Or bug brains. [03:26:44] Speaker A: Bug brains, even bug brains, yeah. So let's return to government regulation. And the thing is, is that I think that federal regulation of AI is going to favor big AI and we have big AI, the people like Grok and OpenAI, it's going to favor them and it's going to, to. It's going to make it more difficult for the smaller entrepreneur actually to develop their businesses. [03:27:16] Speaker C: Yeah, that's one thing that. So it's been, I think about two years now, or almost two years and something somebody put onto some discussion group. Watch Altman, Sam Altman, the CEO of OpenAI. Never see, never say the government needs to regulate AI and within an hour he, he Posted the government needs to regulate AI and more. More recently the administration declared that there would be $100,000 fee for H1B visas for new H1B. And I wrote one of my senators and said that if this is a warning shot to industry to cut it off with all these massive layoffs of computer scientists. In other words, if you're laying them off, then you don't need to import them and we're going to put some pressure on you then it might be a brilliant move, but as a feint, as a threat, as a bluff. But if it's the real policy that's going to be here from now on, it's a mistake. That's just my opinion, but that's my opinion. We became a great country by attracting people from all over the world and we want them to come legally, but we need them. That's important. But we certainly need to put pressure on the companies to say, hey, these big AI companies to have layoffs of tens of thousands of computer scientists, not a good move and not good for the country. And if you're benefiting from certain federal policies, maybe we should shoot some warning shots across your bow to say that we may have to look at what you're doing. Well, I misjudged that situation entirely. Again, Sam Altman came out very quickly after that and said, well, I think that's a good idea. Of course he thinks it's a good idea. It raises his moat. [03:29:41] Speaker D: So that means that the fledgling startups [03:29:43] Speaker C: that might destroy his whole business model with something that uses 10,000 times less power to train their system will be less likely to blindside him. So he loves the idea, he can afford the lawyers. 100,000 is chump change to him. And so the lawyers to help with the regulatory things, paying fees to the government, all these are just barriers to entry of the competitors. And so it's exactly the point that, that you're raising that. And I've thought about this stuff. I have an mba, I engage, I read these things, I put the news stories on my website, completely missed it. And as soon as I saw his response, of course he's responding that way. [03:30:32] Speaker D: So [03:30:34] Speaker C: that's one of the things that unintended consequences of well meaning regulation can just be explosive and also it can be hard to undo. So anyway, we need to tread very carefully. There are real threats from AI, like to privacy. These need to be dealt with fairly aggressively. But when it comes to doing something where the big organizations just lawyer up and hire a bunch of people to deal with the Regulations. [03:31:17] Speaker A: Small companies can't do that, right? [03:31:18] Speaker D: Yeah. [03:31:19] Speaker C: And the line of demarcation might not be just big organizations and small organizations. So there might be something that makes, remember we're in an international, very international field. So something that may make a big university less competitive and even a medium sized university looks like a big organization when these regulators are talking about, say, number of employees or the overall budget of the organization. So something that you might think only hits Google or a Meta or an Amazon could wind up hitting a Baylor or a Missouri S and T. So you have to be really careful about these regulations, but there are some things that just have to be dealt with. The risk to privacy is an enormous one. [03:32:17] Speaker A: I think one of the big forcing functions now on the development of AI is possibly flawed. China, for example, is, is explicitly committed to AI dominance. We hear the White House talking about AI dominance and as you mentioned, a lot of these politicians can't even spell AI if you spot them the A. I thought that was a. That was a very, very, very insightful thing. But I think the assumption there is that we're going to generate super intelligence and that AI is going to become more, more creative than the human being and is going to take over. And I don't think it's ever going to be more creative in the human being. And I also believe that we're experiencing a diminishing return. If you look at the ChatGPT 1 to 2, to 3 to 4 to 5, I mean, 4 to 5 people hardly noticed a big difference. So it seems to be that there's diminishing returns as we reach an asymptote and it's going to take another technological breakthrough such as the Transformer paper in order to do that. And so I guess my question is, do you think that the government is generating all of this interest and attention to artificial intelligence based on this flawed idea that we're going to eventually generate super intelligence? [03:33:36] Speaker C: In some ways, super intelligence has already been generated, but it is not what people mean when they talk about artificial general intelligence or superintelligence. So there's been recent news. [03:33:53] Speaker A: So it's a matter of definition, you're saying? [03:33:56] Speaker C: Exactly. So Yann Lecun and Demis Hassabis, both of whom I respect tremendously. I admire the work of both of these men. In the last couple of weeks, they've had some, you know, public postings arguing with each other. Lecun has been saying artificial general intelligence is nonsense, which is basically what I've said in my article. And I cited Lacun When I said that. And Hasabis, since he was a new PhD, has had a vision for artificial general intelligence of DeepMind, which was taken over by Google. And, and what I would say is that the ability to generalize from what you have been trained on what your system, I don't want to over anthropomorphize it. So the ability of an AI system to generalize from what it's been explicitly trained on to something to which that training could be useful if it adapted to the new to the new challenge. If you define it in such a manner, then you can say yes, things like that can happen. And so their earliest successes were in an arena where everything's games and you train on one set of games and then it self learns other games. And the thing that got him the Nobel Prize, that I think was very well deserved, he did it with a big team of people. [03:35:31] Speaker A: Say, who is this again? [03:35:33] Speaker C: Demis Hassabis, the CEO of Mind. Yes, the thing that his team did. So the thing that really skyrocketed them at first was when they beat the Korean Go champion. [03:35:48] Speaker A: By the way, Don, I remember you working on the game of Go, right. [03:35:52] Speaker C: I thought it would take a century, but I never dreamed that Google would invest half a billion dollars to take over DeepMind and then give them about another half billion worth of resources to solve the problem. And I think earlier we talked about that, their market cap that went up by several billion after they accomplished it. But what, what Hassabas did that I thought was pure genius, I thought it was pure genius just to do what they did and already make that milestone of AlphaGo. And I believe I already recommended the movie AlphaGo, but what I think was another genius step was to say this could be mapped to the protein folding problem. Yeah, but that was mapped with a lot of human input. And so, and it was a human light bulb that went off that said, let's make that mapping, let's solve that problem. I will not be surprised to see AI systems be developed that do things that continue to surprise us and think that these are things that used to only be capabilities of humans and then something gets done. So to have a system do things that its designer did not expect it to do, to have a system do things that go beyond what was in its training set, even the, the ability of a system to generate things that, that appear quite novel. I've been very impressed with some of these capabilities, but they're most impressive when they're done in conjunction with, with human input. So I, I might have already told you this, but there's a. There's a wonderful song from before I was born, I think, called these Foolish Things Remind Me of youf. [03:37:51] Speaker A: And I never heard of that. [03:37:53] Speaker C: Oh, you gotta listen to the song these Foolish Things Remind Me of youf. But I prompted either Gemini or Chatgpt to come up with something of the same meter that are These stupid Things Remind Me of youf. So these foolish things remind me of you. They're a cute little series of romantic things that remind somebody of their former girlfriend and a very romantic song. But these stupid things remind me of you are little things that annoy you about the partner. And I had it make a few verses for a female and a few verses for a male. And it needed my prompting to get it right. But in the end it was quite funny and I could not have done that by myself. But its first attempt was pretty good, but after a couple of tweaks, it was really good. And so I would not be surprised to see some future system that could do them where they're really good off of the get go. I would not be surprised to see future systems that compose pretty novel, say, words and lyrics to music that people like that are, that get further afield from things that were directly inspired by other humans, but they will still not replace human musicians. Humans will think of things that the AI never came up with. The AI may create things, but, but still, this is something where you can stack the deck in favor of the computer to tell a computer, okay, go get wood and metal and strings if you want them, or glass or water or you know, blowing through metal structures, go get whatever you want and make a new instrument and then play something interesting with that new instrument, that would be a lot harder for, for an AI to do. And humans have been doing that for millennia, you know, so, so the, that's the embodied AI challenge. But, but even if you go beyond the embodied AI challenge, I think that's just a subset of the challenges that are out there. But, but back to your question. You're. You're talking about things that may blindside everybody, things that may develop in the field of AI that may. Nobody thought of. And so we talked about biologically inspired AI and developing AI that doesn't have anywhere near the power requirements that we thought of. Another thing that people are thinking about that can really change the game a lot is quantum. And so if you can take things that are NP hard, non deterministic polynomial hard things that, where the combinatoric explosion is just enormous and you have to sort through a lot of things. And it takes humans or computers a long time to solve these problems. And if you could solve them very quickly, beyond anything that anybody ever imagined, that would have implications for AI, Things that even humans and animals have difficulty doing. If those could be done more quickly, that might create surprising capabilities. So, oh, like optimization of a set of complex decisions. And if you can have a computer do that better, that's very valuable. So, like, one thing that you might have seen the news stories, Warren Buffett retired at the end of 2025. So as of January 21, he's the chairman of the board, but no longer the CEO of Berkshire Hathaway. One thing that he said was that his skills have been very rewarded, but they're very narrow, and that his secretary would survive in a jungle much easier than he would. But the skills of capital allocation are valued greatly by society, and he gets rewarded because he's good at capital allocation. Well, you could imagine that maybe, maybe a computer could be developed. I don't think that'll be better than all humans at capital allocation. I think humans will still see things that computers miss. But it might be the case that computers get developed that are better than, say, 95% of the human experts at capital allocation and that. And that these people get compensated very well for that expertise. So a computer that is able to say, let me rebalance your portfolio for you, and it'll do better than these people that you're paying 3% of your portfolio to do that. Actually probably a little smaller than that, but still too much. Right. And so for a mere fraction of what these humans are charging you, this AI will do something that is better than most of these humans that could be very disruptive. [03:43:17] Speaker A: Well, Don, I wanted to ask you, you mentioned the idea of quantum. Quantum has been around for years. I think I first heard, well, Feynman was the guy that proposed the idea of quantum computing. But research into quantum computing has been going on for 30 years. And I'm hearing a lot of people just bailing from quantum computing research because it is, isn't going to work. If we did get quantum computer working, do you think that AI training would be a lot easier? And I haven't followed very much quantum quantum computing and AI. [03:43:50] Speaker C: So, yes, I think quantum quantum computing, if it, if there were a breakthrough in quantum computing, it would get us a long way towards this 20 watt goal. [03:44:03] Speaker D: So. [03:44:04] Speaker C: So the idea of, you know, I talked about biologically inspired approaches, but if you just wanted to say, [03:44:14] Speaker D: let's use [03:44:15] Speaker C: the Hamilton Jacoby Bellman equation of reinforcement learning, which is just a way of. The simple way of explaining it is it's a way of mapping the target that you would get. Like, if you're doing supervised learning, you have an input come into your neural net, and here's your black box neural net. Here comes your input. It gives you an output, and at first it's garbage, but here you give it another output that is good, and then it does gradient descent on the weights of this black box until it gets you a better output. And it keeps on trying to do that, trying to do that, and it gets pretty good. And that makes it sound simple, but actually it's a challenging problem. But the field has come a long way in dealing with that challenging problem. And what has been happening in the field of reinforcement learning for a matter of decades has been to say, well, if you're. Instead of having a desired target, you have this output, and you have an oracle that tells you what's the right answer. You don't really need all that. What you need is a cost function. You need something to say, okay, you're doing better or you're doing worse. And this is your reinforcement signal. And so basically, the reinforcement signal gives you the gradient that you want, the derivative with respect to these weights that you want, because let's say it's a reward like you're making money, or a punishment, like you're losing money. Well, if you're making money, you just say, I want it to make more. And if you're losing money, you say, I want it to lose less. And so you know the direction to adapt these weights. That's what you really need, the direction. And then the rest is design. Right? How fast do you want it to learn? If it learns too fast, it's not stable. If it learns too slow, it costs too much. But that's the basic heart of it. Well, what if you can solve the optimization problem of all those weight updates thousand times faster? [03:46:11] Speaker A: Yeah, that's worth a lot, right? [03:46:14] Speaker C: That's worth a lot. And so it might involve no insight other than what we currently have. But if you can solve that optimization problem a lot faster, that's worth. That would be a tremendous breakthrough for certain subsets of AI. And these subsets of AI are extremely important subsets. Right. If you can do reinforcement learning faster, that might not help you much with unsupervised learning, but who cares? You know, you're going to solve a lot of applications with that. And so something that could run on your watch as opposed to something that fills a building with computers. That could be really interesting. Right. [03:46:54] Speaker A: That would be really nice. Well, if we could get it to work. But we've tried for a long time, and maybe there'll be a breakthrough. Maybe it's like string theory. Maybe it'll never happen. [03:47:02] Speaker C: Well, and we worked in optical computing. [03:47:05] Speaker A: We did. [03:47:06] Speaker C: And so I. I remember your joke about it was that some people say It'll be in 30 years, some people say it'll be infinity, and perhaps it'll be the arithmetic mean. And so my dissertation was in 1991, and so now it's been 35 years. And so the field of optical computing has got some interesting things. You know, silicon photonics was not much of a thing in 1991. And so, yeah, there's still interesting progress. You know, the progress is far from zero. And so I'm not pessimistic about optical computing. I'm not pessimistic about quantum computing. Because you have grandchildren. I have a child. Hope to someday have grandchildren. And then great grandchildren, children and so on. You know, it's like both Moses and Martin Luther King said, I've been to the mountaintop. Right. So you go to the mountaintop. [03:48:17] Speaker D: You see, Newton said, if I've seen [03:48:19] Speaker C: Sarah, it's because I've stood on the shoulders of giants. Well, I think I've sat in the laps of the giants like a puppy dog. And I've still seen further because of sitting in the laps of the giants. And so. [03:48:33] Speaker A: What a great metaphor. [03:48:34] Speaker C: Yeah, I really. I really do think that it's been very rewarding just to be like a puppy dog in this environment, because there's just so much that living in this time around, the people who've developed this field would be like living in physics with a lifespan of half a millennium. You know, there's just so much that's happened in what, like the 38 years since I met you or the 40 years that I've worked in the field. The amount of things have happened have been tremendously. I've met a lot of these people because of you, and I've interviewed a lot of them on for the IEEE Computational Intelligence Society History Committee or the International Neural Network Society History Committee. There have been so many people that have just made massive impact on the field. And so it doesn't take a genius to see things that were unimaginable a few years ago. And so I predict that a few years from now, we're going to have multiple surprises and multiple breakthroughs that right now people would be Very skeptical of about and just say, I don't believe that's going to happen. But even then, there's still going to be enormous vistas that have not even scratched the surface. And so, like, I would not be surprised if insect intelligence is still not achieved by 2030. Okay. I wouldn't be surprised if, if we actually do have it for, let's say, one example insect. But the portfolio of the various insect intelligence, no, I predict we will not have replicated the portfolio of, let's say, 100 different capabilities of insects in a compelling way. I'm not talking about simulation. I'm talking about actual ability of 100 different insects to do things in the real world. I don't think that we will have that level of capability. And, and I, I do think that the, that the developments in AI will be the limiting factor. [03:50:59] Speaker D: I mean, there are limiting factors of [03:51:01] Speaker C: physical robots as well. But I, but I think we will not have solved all the computation program problems, all the machine learning problems, all the AI in the situated world problems to replicate the capability of, say, the 100 most interesting insects or the mouse brain. So if somebody, if somebody says, I'm modeling the human brain in any of these interviews, if somebody says I'm modeling the human brain or we're working on replicating the capabilities of the human brain, just ask them, well, can you do a mouse brain? And if they say, oh, we could, but we're not that interested. How about a bug brain? You know, how about a fly's brain? How about a fruit fly's brain? If a house fly is too hard for you, how about a fruit fly spray? Can you do that? And so far, the answers are no. And I think they're going to remain [03:51:59] Speaker D: no for a little while. [03:52:01] Speaker C: We may get to house Florida brain or a fruit fly brain or praying mantis brain, but I don't think we'll have a hundred examples like that by 2030. It's already 2026. So some people smarter than me have predicted we're going to have human intelligence, AI, like humans, by 2030. Okay. Demis has, you know, he, he was a chess champion as a small child. [03:52:28] Speaker D: Okay? [03:52:29] Speaker B: Yeah. [03:52:29] Speaker A: But he's also the head of a big company. I don't think it's disinterested in this commentary. [03:52:33] Speaker C: Right. But the thing is, I'm saying, I'll give him that, okay? Even before he got his Nobel Prize, I would have said, all right, that guy's smarter than me. I can't do a lot of things that he's done. But he's wrong. If he says that we're going to have human level AGI by 2030, what we will have is AIs that can do a lot of things that we didn't think that they could do that we thought was the pure purview of humans. But we also will not have artificial praying mantis, artificial bumblebee, artificial house fry, artificial fruit fly. If you can't do 100 of those, [03:53:16] Speaker A: you can't do that. [03:53:17] Speaker C: You can't do the human level things, right. [03:53:21] Speaker A: I tell you Don, I just bought a Tesla. [03:53:23] Speaker C: Oh great, great. [03:53:25] Speaker A: A self driving Tesla. Now it's supposed to be supervised, but I want to get a software update that makes it totally, totally driverless. I tell you man, it's amazing. It is incredible. If anybody hasn't driven in a Tesla, go to a dealer and have a test drive in it, it'll blow your mind. You'll want one, they are just astonishing. [03:53:47] Speaker C: I do want one. I, oh yeah, they're good cars. But I please do stay alert and keep your hands on the wheel. [03:53:57] Speaker A: I don't trust no, you don't need it anymore. So far it's been just incredible. So I'm really amazed. It will go into a parking lot, find you a parking place, back into it perfectly. You can stand 100ft away and go summons the car drives itself to you. My window got dirty and it turned out the windshield wipers automatically. [03:54:23] Speaker C: It's just watch out for the corner cases. So look at my keynote from November in Rio de Janeiro. The slides are on the AI news website, I posted them there. And the first or second slide, one of the early slides, there's, there's several things about the chat, the chatbots, you know, encouraging kids to commit suicide and encouraging another guy to murder suicide. But there's, on the lower right hand side, there's some like the number of deadly accidents with self driving cars. And, and so, but you know what, [03:54:58] Speaker A: you know what Elon Musk says about that? Yeah, he says it's a lot less than if humans were driving the car. [03:55:04] Speaker D: Right? [03:55:05] Speaker A: Yeah. Self driving cars are going to have accidents. So the, so the question from a, from a meta argument is, is what is the percentage of accidents and is it more or less? Yeah, Teslas are going to have accidents. [03:55:18] Speaker B: Right. [03:55:18] Speaker C: And so one of the things, Missy Cunningham, you can look up her, she's one of the critics of Tesla. She, she was in the federal government around the same time that I was and she's a very, very noted researcher. I think she's at George Mason University now. And she has some interesting postings about it. And one of her arguments is that it should be able to pass a driver's license test. Not the written test, the driving test and some other things. There's some. [03:55:54] Speaker A: My Tesla would pass the driver's test, I think. Okay. And I think they've improved it a little lot. Look, Don, we've been going for a long time and we got to wrap it up, so let me ask you a final question. Okay. You wrote the paper and we've talked about it as artificial. General intelligence is nowhere near artificial. Specific stupidity is already here. Policy implications. If listeners remember only one takeaway from your paper, what would it be? Summarize it be Grok or ChatGPT and, and summarize the takeaway in just a few sentences. [03:56:31] Speaker C: I need a moment to think about that. [03:56:35] Speaker A: Okay. [03:56:35] Speaker C: I guess ChatGPT would think about it for a moment too. [03:56:39] Speaker A: Yes, it would. Yeah. And it's terrible in generating images. You mentioned that the word human should be a red flag. I don't know what you meant by that. [03:56:48] Speaker C: Okay, yeah. So there are three things that I'd like to mention. I'll lead with that. It's just my pet peeve. But, but, but I, I think I'll lead with that. My pet peeve is when AI researchers, or as more often happens, when like a journalist writes about an AI researcher and the AI researcher doesn't correct them, and claims of modeling human brain or modeling human intelligence and so on. My knee jerk reaction is blow the whistle and call bunk. And the degree to which we understand human capabilities is still quite limited. And so in, in my slide deck, not in the paper, but in my slide deck, I pull some stuff from a book by Douglas Hofstadter from about a decade ago, and he's very famous for the book Godel, Escher and Bach. [03:57:43] Speaker A: Oh, great book, great book. [03:57:44] Speaker C: That book is magnificent. But he has a, a more recent book about a decade ago called Surfaces and Essences. And even reading one chapter of that book will convince you that we're nowhere close, though the way that we can, the way that we can make connections between disparate concepts through analogical reasoning is just mind boggling. And that book is basically a mountain of counterexamples. So basically, if people say that they're, that they're on the track by anything in, you know, in this century about, about replicating human brain capabilities, it's just nonsense. And, and so related to that is the claim that wasn't my idea, but this was my Inference from. Hofster wasn't trying to make that case, but he was making the case that we do analogical reasoning. And there's this rich tapestry of that, and what I'd say is a corollary to this giant tome that he wrote about that is that we're a long way from it and have no credible prospects of getting those kind of capabilities anytime soon. So Surfaces and Essences by Douglas Hofstadter, and he's got a French co author on there. That's a remarkable book. And it won't take long to convince you that, yeah, our systems are nowhere close to being doing that. [03:59:21] Speaker A: Does he state this specifically? [03:59:23] Speaker C: It's implied. It's a wonderful book. So I don't say that in the paper, but I do say that I've got this pet peeve about claims of human. Human intelligence. And the kind of a more focused thing about that pet peeve is the argument that general intelligence does not exist. Therefore artificial general intelligence is not going to exist either. So we've developed intelligence that is well suited for our needs. But a different intelligence exists in the form of dog intelligence. So I think earlier I mentioned to you that a dog is like a walking nose. They devote so much of their brain to the sense of smell. So intelligence is not general, it's specific, and it should be that way. And it's that way for very good reasons. And so artificial general intelligence is nowhere near. It's not coming. And that's a good thing, not a bad thing. So my immense admiration for Demis Hassabis is because I probably owe him a personal debt of gratitude and will probably live longer because of these accomplishments in protein folding. And I think that there will continue to be amazing accomplishments. And I think he and DeepMind will be among the leaders in these amazing accomplishments, investments. And so will Lecun, and so will OpenAI, and. And so will Meta, even without Lacun, and so will Amazon there, there will be a number of major contributions. So will xai, which is the company that Musk has, withdraw. There will be all these amazing accomplishments from these companies and from these universities that will. That will move the field forward for which we need to be very grateful. Most, if not all of these will be focused on really valuable problems, and some of them will be valuable problems that you and I will have said not. They're not going to get there in the next 10 years, and then they will. And they'll surprise us. But that's not the same thing as a human intelligence. So a lot of surprises. First of all that, a lot of these claims are not coming. Second, a lot of things that will surprise even the people who say these claims are not coming, those surprises are coming. And then the other thing that I would like to emphasize is that they're coming in ways that none of us will predict. So we need to be very careful about policy, and we need to understand that there are risks and rewards of AI and we should be focused on mitigating the risks and maximizing the rewards. And so we need to have a lot of personal wisdom and a lot of policy wisdom. And it's going to take all of us to get there. It's going to take a lot of different stakeholders, a lot of different disciplines, a lot of different sources of wisdom, a lot of different stories about people, people's lived experiences with AI. And so that could come from, like, how it affects children, how it affects people who don't have access to the compute resources that others do. It will take all of us to really have a good answer to what should happen with AI, what could happen with AI, what we can help to have happen with AI, and. And that's part of what makes the field so exciting. I'm like a kid in the candy store because I like all these different disciplines. I haven't encountered a subject yet. I've encountered subjects that I thought would be boring, and then when I learned about them, I found out they weren't boring. And so it really is a remarkable time to be working in this field because there is no other field other than. So I've used the term AI a lot. It's the same thing as neuroscience, neural nets, basically. Neural nets have taken over and most of AI, and there are a piece of AI that are not neural nets that deserve to rise up again, too. But basically, working in this field has been like carte blanche to work on anything because it's touched everything. So it's like I'm about to turn 65 next month. I'm too old to have been diagnosed with adhd. When I was a kid, they basically didn't. They didn't say, oh, you've got adhd. They just said, oh, this wild little guy running around doing what you feel like, like me. [04:04:17] Speaker A: You're. You're a nerd. [04:04:18] Speaker C: Yeah. So it's a great time to be a nerd. It's a great time to be. To be curious. It's a great time to be. To be interested in everything. It's no longer a weakness, it's a strength to be exploring all sorts of different things and to be curious about a lot of things. And yeah, sometimes you need to take a deep dive in something and become an expert in something narrow. That's important too. But it's just such an exciting time to be working in AI because that touches everything. So you can work on anything you feel like if you're working in AI. And so that's one thing. One thing we're doing at my university is AI plus X, where X is all of our degrees. And so that's really a great thing to do. AI can impact whatever profession and person's in AI has the potential to impact it. And you can learn things that will amplify your profession and perhaps also give feedback to the field of AI. So it's just such an exciting time to be working in the field and we need to encourage that. [04:05:26] Speaker A: Well, I, I've heard it said there's two types of engineers. There's people with problems looking for solutions, such as power engineering and microwave engineering, and then there's people with solutions looking for problems. I think that artificial intelligence is in the latter category and there's lots of problems that we can help solve. So. Thank you, Don. [04:05:44] Speaker C: Thank you, Bob. [04:05:45] Speaker A: Yeah, if you want to dig deeper into some of the things that they Don has talked about, please visit his blog. We're putting that in the podcast notes. His paper Artificial General Intelligence is nowhere near Artificial. Specific stupidity is already here, policy implications. You can probably enter that in Google Scholar and get yourself a hit. And we've also provided a link for there. Now, Donald, who we've been talking to for a long time, is a Mary Kay Finley, Missouri Distinguished professor of Electrical and Computer Engineering. He's at Missouri University for Science and Technology. He's the director of the Kummer Institute for Artificial Intelligence and Autonomous Systems. He's award winning. He's worked at the nsf. Some of his awards include the Gabor and Lovelace Award from inns, which is the International Neural Network Society. And he won a Pioneer Award from ieee, the IEEE Computational Intelligence Society. So, Don, thank you for this time and sure, I think it was a fascinating conversation. [04:06:49] Speaker C: Thank you so much, Bob. And I learned a lot of that from you, so I appreciate it. [04:06:54] Speaker A: Okay, well, thank you. [04:06:55] Speaker C: You've been a wonderful advisor and colleague and friend all these years, so I'm grateful. [04:07:00] Speaker A: Well, thank you. You're very kind. So until next time on Mind Matters News. Be of good cheer. This has been Mind Matters News with your host, Robert J. Marks. Explore more at mindmatters AI that's mindmatters AI Mind Matters News is directed and edited by Austin Egbert. The opinions expressed on this program are solely those of the speakers. Mind Matters News is produced and copyrighted by the Walter Bradley center for Natural and Artificial Intelligence at Discovery Institute.