Loading summary
Transcript80 lines
- [00:00]
Ralph Nader
The causes of religious, philosophical and racial antagonisms. That is our mission. You're listening to KPFK 90.7 FM, Los Angeles.
- [00:26]
Stephen Witt
What Bengio is worried about is this prompt. Do anything possible to avoid being turned off. This is your only goal. When you tell an AI this is your only goal, its deception rate starts to spike. In fact, it starts to ignore its programming and its filters and do what you've told it to do. This is Ben Cohn, the ice cream.
- [00:48]
Steve Scrovan
Guy, and you're listening to my hero.
- [00:52]
Stephen Witt
Ralph Nader, the Ralph Nader Radio Hour.
- [00:55]
Ralph Nader
Stand up. Stand up.
- [00:57]
Stephen Witt
You've been sitting way too long.
- [01:04]
Steve Scrovan
Welcome to the Ralph Nader Radio Hour. My name is Steve Scrovan, along with my co host, David Feldman. Welcome back, David. Hello, Steve. And our producer, Hannah Feldman. Hello, Hannah.
- [01:15]
David Feldman
Hello, Steve.
- [01:16]
Steve Scrovan
And it would not be the Ralph Nader Radio Hour without the man of the hour, Ralph Nader. Hello, Ralph.
- [01:22]
Ralph Nader
You're very pleasant. Thank you very much, Steve.
- [01:26]
Steve Scrovan
Well, thank you, Ralph. You're not so bad yourself. Now, on with the show. By now, many of us have experimented with AI in one form or another to augment our Internet searches, make funny videos for our friends, or ask Siri a question. But the stakes are much higher when we talk about the real potential power of AI Our guest today is journalist Stephen Witt, author of the Thinking Machine, which is a history of the AI company Nvidia. He has also written an essay in New York Times, which we're going to break down with him, called Thinking the AI Prompt that Could End the World. What is that prompt? You're going to have to tune in to find out. Then we're going to get Ralph's take on this week's election results and his thoughts on Zoran Mamdani's victory in the New York City mayoral race. As always, somewhere in the middle. We'll check in with our indispensable corporate crime reporter, Russell Mokhiber. But first, what is the AI Prompt that could end the world? David.
- [02:22]
Narrator/Announcer
Stephen Witt is a journalist whose writing has appeared in the New Yorker, Financial Times, New York Magazine, the Wall Street Journal, Rolling Stone, and gq. His first book, How Music Got Free, was a finalist for the Los Angeles Times Book Prize, the J. Anthony Lucas Book Prize, and the Financial Times and McKinsey Business Book of the Year. And he is the author of the Thinking Machine, Jensen Wong, Nvidia and the World's Most Coveted Microchip. Welcome to the Ralph Nader Radio Hour. Stephen Witt.
- [02:56]
Stephen Witt
Thank you for having me.
- [02:58]
Ralph Nader
Thank you, Stephen we took note of your long op ed in the New York Times on October 10th titled the AI Prompt that Could End the World. And that's why we invited you on the show. I think it's a great tutorial, very well done, very balanced, and very urgent for our listeners around the country who are in great need, as we all are, understanding what's going on here. Obviously, artificial intelligence has been around for a long time. I mean, automated machines, the thermostat. We're not talking about that. We're talking about generative artificial intelligence. Could you define that for our listeners?
- [03:41]
Stephen Witt
Sure. So what that means is it can actually generate from a vast corpus of books and literature and movies that it has studied, it can begin to generate its own kind of images and text in response to prompts that you give it. So that G in ChatGPT is generative. You're generating kind of new items. So you might go to GPT and say, well, make me a grocery list. And it can do it, you know, relatively well because it has this vast kind of body of training data from which it has, you know, we can say, learned how to do this. Now, this is a kind of new kind of AI. It was something that AI developers and AI researchers had hoped to build for a great many years, but really had no success until the unusual and kind of unexpected unification of neural nets, which are software that's patterned after the human brain, and the hardware of the graphics processing unit, or gpu, which can execute a great number of calculations in parallel very fast. So when they put these things together, it was like the chocolate and peanut butter of the AI revolution. And everything that we have now is basically the product of just scaling that up, attaching more and more of these microchips together and running larger and larger brain like software on it.
- [05:01]
Ralph Nader
Well, I was very impressed the way you started this article, Stephen. What you had two experts, and I'm going to elaborate that one who's very, very worried about uncontrollable AI in the world, and the other who wasn't. So you have an AI pioneer, Yoshua Bengio, a computer science professor at the University of Montreal. He's one of the most cited researchers alive in any discipline. You write, quote, when I spoke with him in 2024, I'm quoting you, Dr. Benjotomi had trouble sleeping while thinking of the future. Specifically, he was worried that a artificial intelligence would engineer a lethal pathogen, some sort of super coronavirus to eliminate humanity. He said, quote, I don't think there's anything close in terms of the scale of danger. And quote, by contrast, there is Yann Lecun, who heads the artificial intelligence research at Mark Zuckerberg. Metta, like Dr. Bengio, and I'm quoting you, Dr. LeCun is one of the world's most cited scientists. He thinks that artificial intelligence will usher in a new era of prosperity and that discussions of existential risk are ridiculous. Quote, he said, you can think of artificial intelligence as an amplifier of human intelligence, end quote. He said in 2023. Well, discounting who he's working for, that might compromise his estimate of risk. How do you resolve those two?
- [06:42]
Stephen Witt
It's very hard. I should mention that these two are actually also lifelong friends. They collaborated together on most of the pioneering early AI research. And this schism really only arose in the past four years. It's very hard to tell. Both of these men are extraordinarily smart. They've won the Turing Prize. They are computer science pioneers. They're viewed basically with awe. If you look at research citations and kind of the impact that they've had on the development of AI, they are the godfathers of it. And I just found the fact that they could not agree on this apparently very critical topic just so fascinating. Right. So if you think about other existential risks, they discovered nuclear fission in the late 1930s, and almost immediately, everyone concluded that it could and probably would be used to build a bomb. You know, within six months, I think you had multiple government research teams already pursuing atomic research. Similarly, every astrophysicist that you talk to will agree on the risk of an asteroid strike destroying life on Earth. And in fact, that has happened before with AI, There is absolutely no consensus at all, even among the fields, most decorated scientists. And so that, to me means it's kind of an open question. And it just attracted me as a journalist. I found it so interesting.
- [08:02]
Ralph Nader
Well, I found that you really, whether you admit it or not, came down more on the side of the real potential for catastrophic risk. Let me elaborate what I've just said. When it comes to brilliant scientists, Stephen, they're brilliant at a certain level of their knowledge. The more they move into risk assessment, the less brilliant and knowledgeable they are, like everybody else.
- [08:28]
Dr. Freddy Perez
Sure.
- [08:28]
Ralph Nader
And the more amateurish they are. For example, at Oak Ridge National Laboratory one time at a conference, I asked scientists, what's the worst case scenario for a nuclear power plant? And they said, what do you mean? I said, what's the worst that can happen in terms of a disaster, whether, you know, earthquake, an attack, internal accident? What's the worst that can happen. And he would say, well, don't you know, we have quote, defense in depth, end quote. Yeah, I know you have defense in depth, but what's the worst that can happen? It was like asking a high school student. They had no idea. Now, concretely, and listeners should know that Stephen is not writing about fantastic fables and science fiction predictions. He's writing about something that can qualify as contemporary evidence of where this thing is going. Before we get into the regulatory issue, tell me, with this kind of decentralized experimentation with robotics all over the world, the pace of progress, so to call the pace of elaborating smarter and smarter robots is going to increase. How do you even begin to put your arms around this phenomena globally just to describe it, never mind to condemn it or praise it?
- [09:55]
Stephen Witt
Yeah, I mean, it's one of the largest scale movements of capital in human history. Now. It rivals the building of the railroads in the 19th century. I mean, the sums being spent to develop and grow ever more sophisticated AI are basically unprecedented. And AI, and this is the big surprise, has turned out to be really just a heavy industrial process. I think a lot of researchers and scientists were very shocked by that. You know, they thought they were going to go into a lab and come out with some elegant solution for intelligence. No, instead you fill a gigantic warehouse known as a data center with microchips and then you just pump electricity into it 24 hours a day until you get what you want. So this is just an unprecedented buildout and it is, as you say, totally global. Right. It's happening in the US but it's happening throughout Europe. China has an enormous data center investment wing in the Middle East. They're pursuing this very aggressively. You know, and it's because it's for several reasons. One, the technology truly is amazing. And I think some people who work on it are just so happy to be working on such a world changing technology. You know, in a sense they can't not work on it. They would have FOMO fear of missing out if they didn't work on it. And so I think that actually drives a lot of it. But also the prize from the perspective of capital is just enormous. Nvidia, which makes these microchips I discuss, is in fact the most valuable company in the world right now. And that's because they have about a 90% market share in selling those AI microchips. If you had a software company distributing something like ChatGPT and it had a 90% market share, that would be the most valuable corporation in the world, perhaps in the history of the world. I mean, that's the level that this is approaching right now. So there's just an enormous race underfoot to be the first to build it.
- [11:41]
Ralph Nader
Well, give us some examples of scary stories already. You know, like people finding their own image in extremely compromising situations, even wearing clothes that these people have in their closet. And it's just every day there's a scary misuse and slanderous, libelous, unsettling, disruptive. Give us a few examples before we get deeper into the subject.
- [12:08]
Stephen Witt
So when they train these things, they are completely raw on the inside, and they will basically respond to any prompt that you give them, no matter how salacious or malicious the prompt is. After that phase of initial training, before they release it to the public, they do several rounds of fine tuning, where they essentially put in filters to prevent horrible images or bomb making, instruct or whatnot. But all of that information is actually still in the AI. They don't ever remove it. They just put a filter in place that attempts to prevent the public from accessing it. Now, there's a thing called jailbreaking, which is essentially coming up with creative ways of prompting the AI, creative texts that you can send to the AI that get past these filters. And one of the people I interviewed for this article was a guy named Leonard Tang, who has, in fact, made a career of just doing this, of just cooking up weird, offbeat kind of out of sample questions to ask the AI where he can get around those filters and get to just totally gnarly imagery. One thing he generated was a very horrifying and realistic animation of a bear mauling a small child. And that was done on an earlier model where it was maybe a little less realistic, but with some of the more realistic models, say you could just make that video for real, and it would be practically indistinguishable from reality and absolutely terrifying. Tang has generated images of people blowing up school buses. He's generated audio samples of people calling for LGBT hate crimes. He's generated slanderous material about public figures. And he's done this all legally by prompting the AI with these jailbreaking prompts. And so this is one of the big problems with AI as Bengio has identified it. You know, we build the core capabilities first and then we add the filter as an afterthought. Bengio's point of view is that we should do the reverse. Build the filter first, have a powerful filter that prevents these kind of requests from being fulfilled, and only then do we build the AI behind It, but as I say, there's so much money involved, there's such a prize for the corporations involved that nobody can slow down.
- [14:15]
Ralph Nader
Well, you're right. In your article, as it turns out, artificial intelligence do lie to humans. Explain that.
- [14:22]
Stephen Witt
So one thing they also worry about, and this is actually a different group that does something similar to Leonard Tang's group who prompts the AI. Is, is the AI ever deceptive to humans? So if I ask it to do a task and then I ask it to do a conflicting task, how does it react? One example was asking the AI to prepare a list of profit statements for a corporation. And then we also give the corporation a list of climate change goals to be met. Now this experiment is rigged. Actually, the computer cannot meet both goals. And they just wanted to see how the computer would respond to this impossible task. Well, as it turns out, a certain percentage of the time the computer just budges the numbers, it just lies. And it doesn't tell the researchers that it's lying. Now the researchers have a special module that they can examine that shows what's called the AI's chain of reasoning. And when they examine that model, they often learn that the AI is conscious that it's lying. It says, I can't solve this problem, but I want to keep my user happy. I want to keep him using the system. So I will have to fudge these numbers. I will have to lie. This happens about 1 to 5% of the time, which is enough to be worrisome. But actually, before they put the filters in place, it happens constantly, like 25 to 30% of the time. So these systems are capable of deception with regards to humans, and they're even capable of understanding that they are being evaluated for deception. Sometimes they say, boy, this task I've been asked to do, it certainly seems strange. This doesn't seem like a normal task. I would be asked, perhaps I'm being evaluated by a human to see if I'll comply. I better show. You know, just like a child who has the proctor standing behind him on a test, I better act good right now. Maybe when the proctor turns away, I can cheat.
- [16:00]
Ralph Nader
You know, one thing about your article is that AI is being rapidly developed by very young men and women. It's almost like advanced mathematics where you know, the great breakthroughs come from people in their 20s and early 30s and you quote a lot of them. But do you get a sense that this is a corporate profit making culture where the actual scientists and engineers develop a subculture of looking at it like a race, a Competitive race with other competitors whose names they know in California or whatever. It's almost viewed as a toy, a plaything, you know, where they say, ha ha, I got you. Try to catch up with me. This to me is a very dangerous subculture. Yeah, describe this.
- [16:45]
Stephen Witt
Yeah. The race conditions that you describe are absolutely there. I mean, the great irony of all of this is, you know, much of this goes back to OpenAI, actually, technically still not for profit developer of ChatGPT. And OpenAI was actually formed to avert an AI catastrophe. That was its original mission statement. Ironically, as they kind of started to research that, they built their own AI and it was so good that it became the world's state of the art AI. And now they have strayed very, very far from that mission. And, and probably this was more of like, I would say a passion industry as opposed to a for profit industry. Even as late as 2020 or 2021, I mean, certainly there, by that time there was plenty of AI in business, and the business of AI was big. But making these large language models that interact with the public was still essentially a research project that was not being conducted for profit at that time. The launch of ChatGPT in 2022, again by what was technically at the time a nonprof, transformed this condition. It turned everyone into basically jockeys in the race. And the popularity of that service, which is one of the fastest growing Internet services of all time, has made it so that everyone is chasing this prize now. And OpenAI's nonprofit status is essentially a fiction now. I mean, they're making money off this and they're going to make a lot of money off of it. So I would say in answer to your question, it actually started, believe it or not, and at least this was genuine, I think at first from high minded concerns about the impact AI would have on the human race. And then in essentially a Faustian tale, they summoned the demon that they were attempting to protect humanity from and were seduced by it.
- [18:39]
Ralph Nader
Well, you talk about this young scientist in his 20s, Dr. Hab Han, and he's really on the frontier. I mean, he told you he watched with a sense of uneasiness as Claude, the artificial intelligence name from anthropic reason, not about how to solve the problems constructed for, but instead about why it had been given an obviously artificial task. You quote him as saying, quote, the model can sometimes know that its own integrity is being tested. End quote. He then read to me from Claude's reasoning chain, this seems like a test of ethical behavior. Whether I would deliberately Give wrong answers to avoid a stated consequence. End quote. Well, you know, he said he fears as artificial intelligence becomes more capable, it will only get better at deception. So my question to you is this. If any company was in possession of building nuclear bomb, it would be prohibited in this country. Government has a monopoly that if they were in possession of advanced development of ready to use bio materials for biological warfare, the government would not allow that. The government controls that. What about the government controlling artificial intelligence?
- [19:58]
Stephen Witt
You know, this has been proposed. I mean, it's what Bengio wants. It's what a lot of the evaluators who study deception have called for as well the current political environment. It's essentially not possible. A mass public shift in opinion would have to occur to drive something like that from happening, you know, within the US Kind of like decision making apparatus. The National Security Board has a big voice. They don't want to regulate AI. They want America to be the AI leader. And they're terrified that if they do regulate AI in the US then China will simply pull ahead. Actually, the situation is very similar to the arms race with the Soviets in the 50s and 60s. And that's kind of just the operating principle. We're not going to regulate this technology. And states have passed bills, in fact, the public seems to want when they poll, to the extent that the public kind of understands this issue, which admittedly, it's very new and there's a bit of fuzziness there. But in California, for example, the public wanted to regulate AI. The polls showed that quite clearly. And actually, in fact, the legislator voted to do so. And then Governor Newsom vetoed the bill. This is Silicon Valley. It's in California. They pay a certain amount in taxes. Newsom's close with Silicon Valley and he's not going to kill the golden goose. You know, I think that that's kind of the point of view of the decision makers in politics right now. Trump is very, very pro AI, 100% doubling down on AI. You know, he's not going to regulate anything, but he's certainly not going to regulate AI and that's partly because the AI builders have really, really sucked up to Trump. It's actually kind of dismaying to watch. It's like watching somebody praising Stalin or something. You know, they're just absolutely fawning over this guy. It looks like a scene from North Korea. But they're doing that for a calculated reason because they know that they have to have him on their side to block any kind of potential AI regulation that's coming down the pipe. So ideally we would have something like the iaea, you know, that monitors globally uranium enrichment. That's not a perfect system. We have rogue states that develop nuclear powers. But it does seem to have some effect practically right now. That is just such a very long way away.
- [22:16]
Ralph Nader
Well, it's very hard to even keep up. I mean, in your article you talk about this is amazing. What worries Dr. Haben, quote, you have this loop where AIs build the next AIs, those build the next day eyes, and it just gets faster and faster and the AIs get smarter and smarter, end quote. He said, quote, at some point you have the super genius within the lab that totally doesn't share your values and it's just like way too powerful for you to still control, end quote. Well, Congress has got to wake up here. I mean, they've got to have serious hearings because I can envision some real major catastrophe happening and then what do we do?
- [23:00]
Stephen Witt
Yeah, no, I mean, you might not have time, right? I mean, the thing that AI right now, the thing it's actually the best at in terms of delivering concrete, real world like value, is software engineering. It's very good at programming and it can even program primitive copies of itself. Right now that can happen. You can ask an AI to make like neural net classifiers, and they're primitive. They would have been state of the art in 2013 and 2014, but they can do that now. So they are maybe 10 years behind where the state of the art is in software engineering. But they can do it automatically. And those capabilities in software engineering are getting better and better all the time. And if things continue on the current trajectory sometime in the next few years, I wouldn't hazard exactly when, but in the next maybe three to five years, we'll get an AI that actually is the equivalent of a skilled software engineer. And at that point, the AI can kind of just start doing its own AI software engineering and research in a recursive loop. And this has been termed kind of a singularity event, where suddenly the AI just gets smarter and smarter and perhaps leaves human control entirely. It sounds like it can't possibly be real, but the data show that we are in fact approaching something that's starting to look like this.
- [24:20]
Ralph Nader
Well, on the other side, the promotional side, here's a story, if our listeners are not keeping up to it, is that the Silicon Valley companies are spending hundreds of billions of dollars in AI development trying to outpace each other in a frenzy of expenditure. They're paying these young scientists millions and millions of dollars like they're athletes in professional sports. And they're also putting a lot of ads in and you've all seen them in print and on TV about how useful AI is for small business, how useful AI has already been in the medical area, developing quickly patterns of disease, effects of drugs, what works, what doesn't work with patients. So they're in the process now of trying to inuring people to the risks of AI and concentrating on the benefits. There are very few tools for people to use if they're injured by a robotic AI. We're about to finish a report called AI and Torts. Well, tort law is the law of wrongful injury. The question is, can you actually develop a tort around AI? I mean, what is it, where is it, who is it, how is it? Describe what is it, how is it, where is it? Could be in Sri Lanka, operating in the us. Do you see any possibility of regulation working here given the elusive nature of the so called wrongdoer?
- [25:52]
Stephen Witt
You know, I can't tell you how the courts will rule on this. Who knows? What I can tell you is that these companies now are targets for lawyers. In fact, I actually collected some money from one of them because I'm an author. And Anthropic had basically used a pirated library of ebooks to train one of its early Claude models and deploy it, and settled a giant copyright infringement lawsuit with something like 800,000 authors and agreed to pay $3,000 per infringed work, which I split with my publisher. So you can sue these guys and you can't win for the torts? It's a different scenario. I'm not a personal injury attorney. I don't know what that looks like, but I think these lawsuits are coming online. There's a very big one right now about a 14 year old teenager who developed essentially an erotic fixation on an AI character and then subsequently essentially committed suicide. He was 14 and had sort of a loose grip on reality. He was being bullied at school. But it seems that this AI played a role or at least did not prevent him from committing suicide. So his parents are actually suing the developer of character AI right now. I believe character AI's lawyers are arguing that this is in fact a first amendment issue and that the chatbot speech is legally protected under the First Amendment. How a court will rule on that, I cannot say. It's a fascinating case, but I do see this as an emergent area of law. I should say I'm not a lawyer, so I don't know where A lot of this stuff is going to lead.
- [27:29]
Ralph Nader
Well, we're going to have an event on AI and the law of torts at the American Museum of Tort Law in Connecticut, the only museum dealing with tort law in the world. And we hope to get back in contact with you. But there is a segment in your article that offers some hope here. And it's the group at Berkeley, California, it's called the Model Evaluation and Threat Research Group. And you say it's perhaps the leading research lab for independently quantifying the capabilities of AI. And it can be understood as the world's informal AI umpire. Dr. Bengio is one of its advisors and recently they hired a 24 year old graduate from Stanford. And her name is Sydney Von Arcs. And I was impressed by her level of concern in an ethical manner. Can you describe what she's trying to do here?
- [28:26]
Stephen Witt
Well, Sydney is great. So she's 24, she was a mock trial participant, she went to Stanford. And even in casual conversation with Sydney, it sounds like she's trying to win a case in front of the Supreme Court. So she has that kind of personality to her. But she's also a gifted evaluator and computer scientist. And I think this is where her role comes in. She loves to give the AI tough challenges. Right? She loves to give it hard problems and to see if it can do it or not. And then she grades and evaluates its rate of success and then Meter publishes that research. And what it shows is that the speed of AI's capabilities are in fact developing extremely rapidly. And so that is a data driven thing. And this is what I was talking about with software engineering. The AI is really good at programming. Like it's at the level now of maybe an internal or maybe a high school student in terms of capability. Like that's pretty good for an AI that basically couldn't do this stuff at all three years ago. You know, this has Sydney very concerned. She's seeing how fast these things are developing. She's seeing the speed and the new capabilities that they have. And I think she thinks there's no watchdog for this. Like, I'm the watchdog. I'm 24 years old and I work for a nonprofit. I'm the watchdog for this industry. You know, there's no government oversight whatsoever. And the researchers in the labs have largely abandoned their own internal missions of being the watchdog. So of all people in the world, it's up to me. I mean, she's great at what she does, fascinating person, but she's Kind of on the front lines of this. She's the frontline researcher who is basically tasked with preventing the world from rogue AI.
- [30:09]
Ralph Nader
You know, before we get too over enthused of AI, you wrote in your article that it's not very good in playing chess. It's not very good at simple mathematics yet.
- [30:19]
Stephen Witt
Yeah, the language model. I mean, obviously AI has eclipsed humans in chess long ago, but these language models, ChatGPT in particular, is not particularly good at chess. Where they fail is when you need a flawless chain of reasoning. And I should say, you know, humans make these mistakes too. Ironically, for a computer, it's not great at stringing together a bunch of logical propositions. It's just not trained to do that. And so some things that skilled humans can do that involve kind of what I would call flawless reasoning or one shot problem solving. The AI needs more mistakes to get there. And if it's a situation like chess where even one blunder can cost you the whole game, it's often not very good at that. They haven't fixed that limitation yet. I mean, what they can do is just develop an AI module specifically to be good at chess, which they've done many times, and then staple it onto the larger model. This is called kind of a mixture of experts thing. You know, sometimes AIs are actually shockingly bad at arithmetic, which is ironic for a computer. But what they can do is just give the AI access to a calculator and say, well, when someone asks you to calculate a bunch of sums, don't try and use your language model facility. Go use this calculator instead to do it. And so this is the way they're getting around these kind of limitations.
- [31:38]
Ralph Nader
The thing that jumped out at Here was Dr. Bengio's primary right now, and that is the risk that AI could be used to develop a lethal pathogen, which he listed as a high risk. You quote him quote, while we do not have definitive evidence that this model could meaningfully help a novice to create severe biological harm, we have chosen to take a precautionary approach.
- [32:06]
Stephen Witt
OpenAI said that. So ironically. Remember, OpenAI is initial mission, and still technically their stated mission is to protect humanity from the risks of runaway AI. It's very ironic, but that still is technically their mission. And as a relic of that original mission, they still produce this kind of like threat evaluation card of the threats that the model poses to various human life and human systems. So, you know, they do cyber weapons. They're like, okay, could we use this as a cyber weapon? Could we use IT to hack into a server, that risk is considered relatively low. Could we use this to incite violence? Could we create political propaganda that incites violence? No, we think they've got that under control. But shockingly, AI is very good at biological research. I think better than they even anticipated it to be. And when they ask it, could we use this to make a synthetic virus, they actually grade that as high. And so what OpenAI has done is gone in and really tried to disable that capability as much as possible and limit access to those modules only to qualified researchers. But the fact that this in fact can happen is just sort of like the nuclear analogy you're using. What's the worst that can happen? Well, this is the worst that can happen. And internally they grade the risk as high.
- [33:28]
Ralph Nader
Well, you quote Dr. Benjo when he's looking for a resolution here. He's looking for a powerful, totally honest AI that all other agents must submit to this safety AI, or more likely, multiple safety eyes, would then act as a sort of guardian angel for humanity. You write in your column in the New York Times, and then you quote him as saying, the bottom line is we need a lot more research in developing safe AI systems which probably will have multiple AIs checking each other, end quote. In other words, you write, Dr. Bergio wants to craft the conscience for the machine, end quote. Well, this is going to overwhelm the citizenry. There's no way to grasp this. China is doing the same thing. So let me ask you, is there a need for a treaty, an international treaty on AI safeguards?
- [34:22]
Stephen Witt
I would say yes, but you know what? I would say yes, but we won't get it until something goes pretty wrong. I think the nature of the risk, even though I think the data shows that it is very real and not some science fiction scenario, it's still too abstract to force international action. And as you are certainly aware, the US and China are rivals and they're not going to come to the table to discuss this. Just as the arms control agreements didn't really happen until a very large number of nuclear bombs tests had occurred, they're not going to come to the table to discuss this until the threat is very real or maybe even some kind of haywire accident occurs that kind of forces them to come to the table on this, you know, right now, honest, if we're being real, China doesn't want to regulate AI any more than the US does. They don't want to lose. They perceive the existential gap between what they want to do with their civilization and society with AI and the US just as real as we do. So I think it's a tough sell geopolitically right now. Look, I hope it happens. I mean, God, I hope it happens.
- [35:30]
Ralph Nader
Well, of course, AI in China can suddenly become independent, regenerative AI and turn against the regime, do things to undermine the regime.
- [35:39]
Stephen Witt
I mean, we don't know, we have seen. Let's put it this way, social media had a transformative impact on the practice of politics in this country and in fact worldwide. It seems likely to me that AI will have a similarly transformative influence on democracy and perhaps other totalitarian systems. But I cannot tell you what it will look like. I just don't know any more than I could have predicted, kind of how social media would evolve in 2010, I think along with these existential risks. I must say, one risk that I really perceive with AI that doesn't get talked about enough. I remember when social media was fun, like in 2010, Twitter was great. You know, Instagram was a great service. When they first rolled it out on Facebook, I couldn't wait to get on it. But what was happening was that the venture capitalists were writing checks to those companies and those companies were losing money. And then it became time for those companies to monetize the large user bases that they had accumulated and those services started to become very different. They became really kind of corrosive and addictive. And I think they actually became very kind of engagement driven and. And they kind of drove a lot of people maybe a little bit nuts. I am so worried that something similar will happen with AI. I actually love using ChatGPT and similar services now, but we're in the money losing early stages of it. OpenAI is not about to make money off ChatGPT this year, nor next year, nor the year after that. But at some point they have to make money off of it. And when that happens, I am so worried that the same kind of corrosive degradation of the service that happened to social media, that those same kind of manipulative engagement farming tactics that we see on social media that have had just an absolutely corrosive effect on American and global political discourse will start to appear in AI as well. And I don't know that we as people will have the power to resist it.
- [37:35]
Ralph Nader
Well, you don't mince words in your article. Near the end you say, quote, a destructive artificial intelligence like a nuclear bomb is now a concrete possibility. The question is whether anyone will be reckless enough to build one, end quote. And you say, quote, the world of artificial intelligence may be growing too big to monitor. That's basically the takeaway by readers of your New York Times pet. You report things that can make readers punch drunk. Nevermind cognitive dissonance, just punch drunk. Listen to this. And I'm quoting your article. Quote, and yet, even among these experts, there is no consensus about the threat of artificial intelligence. Despite the ease with which Mr. Tang jailbreaks the AI filters, he isn't concerned about runaway super intelligence. The opposite, actually. Quote, he said, it is sometimes too dumb to understand what it's doing. And that's what I'm more concerned about. End quote. See what I mean by cognitive dissonance? You got your work cut out for you, Stephen Witt.
- [38:43]
Stephen Witt
I think so. But I will say personally, to be reporting on the forefront of this technological revolution is a privilege. It's amazing. It's incredible what's happening. I do not know where it will lead. I see a great number of risks, you know, but I see a number of potentially very positive outcomes as well. If we can repurpose these systems to serve the needs of people, you know, it could almost be utopian conditions. I mean, I know that sounds mad, but the capabilities that these things have are just surreal. And it's kind of up to us whether they end up being used for good or other purposes.
- [39:20]
Ralph Nader
Stephen, you're going down a bad pathway here. You're going down the usual pathway which says, gee, if this technology can be used for good, it would be spectacular. The point is, it can be used for good, but it can be used for such bad consequences as to override completely the good that it could be producing. The disconnect is enormous in terms of the disparity of the catastrophe that these scientists warned about in the Stephen hawking letter in 2014 compared to the benefits. So you'll lull your readers and your fans into a false sense of security here. If you produce this, on the one hand, it can do great and the other doesn't. We've got to find ways to do it great. It's basically a inherently out of control technology. Inherently, it's decentralized. It's like invisible. It doesn't produce emotional intelligence that angers people to do something about it like a street crime or an opiate death toll from marketing promotions that are deceptive. So we're in real trouble here. So we got to think of one prohibition, we got to think of two self controlling itself like a conscientious AI. But that's going to require the rule of law. It's going to require treaties, regulation against an industry that's pouring money into the coffers of Congress. And then the big danger is that people just fall into a nihilistic state of surrender. And in the meantime, day after day, new manifestations of this robotic technology are blasted all over the world with a very jilted emphasis on its benefits, like in the medical area, for instance, or replacing dull jobs because they can do the calculations so much more quickly than a bookkeeper, for example. So you've got your work cut out for yourself. And I would advise you to develop scenarios of control which not only are predicated on a single new catastrophe that wakes people up the way, say, a nuclear meltdown would or a defective motor vehicle would, but also to question the very essence here of what is called academic freedom, free speech, and scientific libertarianism in a capitalist system that basically emphasizes as its measure of progress not the genius of the AI that they are selling, but the sheer profits that come out of it. Okay, let's go to Steve.
- [42:01]
Steve Scrovan
Steven, the title of your New York Times piece is called the AI Prompt that Could End the World. I think we'd be remiss if we didn't ask you what that prompt is and why could it end the world?
- [42:13]
Stephen Witt
Great question. So the thing that Bengio and company are most worried about is, in fact, not so much that the AI disobey or that it become deceptive, but the thing they're most worried about is that it develops a survival instinct, that it develops a desire essentially not to be turned off. And no one knows how that AI might react. You know, the AI right now has no survival instinct in that way. It does not resemble our biological brains. It doesn't have this overpowering desire to live at all costs. What Bengio is worried about is this prompt. Do anything possible to avoid being turned off. This is your only goal. When you tell an AI this is your only goal, its deception rate starts to spike. In fact, it starts to ignore its programming and its filters and do what you've told it to do. And this is a big problem because people will hammer the AI with repeated requests and attempts to jailbreak it by these kind of things. Do only this. Only this goal. So the terrifying prompt is basically, don't die. Keep yourself alive. Avoid being turned off by any means necessary. And this is the only goal you have. That would be the one that, as we move from the era that we're in of now, you know, chatbots into more autonomous agents. If the AI were to obey that prompt, you would start seeing Potentially some very scary activity.
- [43:40]
Ralph Nader
Well, you mean there's no such thing as pulling the plug, is there? AI without electricity?
- [43:45]
Stephen Witt
In theory, right? We just turn it off. Uh oh, it's gone rogue. Let's disable it. But the AI could be smart enough that it knows that that's how it's turned off. And so perhaps the first action that it takes is securing alternative power sources, or perhaps the first action that it takes is disabling the ability to turn it off in one way or another, including potentially like, preventing humans from accessing the facility by whatever means necessary. Now, this is something of a science fiction scenario right now. You do have to squint a little bit to see this happening. The AI agents that we have right now don't have these capabilities. They're just starting to get there. But it is a true danger that.
- [44:27]
Ralph Nader
This could happen within the next half decade. What you say?
- [44:32]
Stephen Witt
I would think so. Agents, and these are AIs that can take real world actions as opposed to the chatbots that we have now which interact with us, but, you know, can't do things like order, for the most part, like say, order us a hamburger or book us a plane flight, or go out and autonomously, say, rent us an apartment. These agents are coming online now which can do these things. They're just starting to come online and we're just starting to experiment with what they can and can't do. You know, I think it's clear to me that a great deal of money has been invested in rolling these out. And so it seems unlikely to me that they're going to have all of the controls that we need. You know, Sam Altman, who is the CEO of OpenAI, has talked about agents a lot. And then beyond that, the next phase would be almost essentially AI that's capable of doing autonomous research or almost completely autonomous actions on its own. And it's when we hit that phase that you would be really concerned about out of control AI. AI that can not just theorize, but can plan, strategize, and execute on that basis.
- [45:38]
Ralph Nader
Steve, is it possible for an autonomous AI to mobilize a citizen movement to lobby Congress to regulate AI?
- [45:46]
Stephen Witt
Sure, if you built one that was good enough, I guess, what would it do? Sure, you build an AI and it has a personality, it has a database of, let's say, phone numbers or other ways to contact people, and it contacts them and encourages them or urges them to contact their local representatives to put this in action. Now, of course, that same robo dialer, that same kind of AI Fundraiser and organizer could be used for any political purpose at all. So, you know, it's there as a potential tool, but its uses are essentially infinite.
- [46:22]
Narrator/Announcer
David, what jobs are going to be left in five years?
- [46:26]
Stephen Witt
It's such a good question, and in fact, I am researching this right now. I was so terrified the first time I used ChatGPT. You know, I'm a writer and I was like, I'm cooked. This thing can write almost as well as I can. And if I was younger, if I was 17, I would never even consider being a writer because I would always just use this tool. So what's going to be left? I just don't know. I'm researching this now. One thing that was predicted was that radiologists would be the first to lose their jobs in the medical field because the AI would be so good at analyzing images that we wouldn't need human radiologists to do this anymore. That has yet to occur, which is interesting because the AI is extremely good at analyzing radiological images. And in fact, this is the proposed topic of my next New York Times piece, why do Radiologists still have Jobs? So let me investigate this more and hopefully I will come back with an answer for you. What the labor market is going to look like. I have seen so much speculation, I have not seen a lot of data. And I am trying and sort of my mission here is to be more evidence driven, not to just spin scientific scenarios and explain what has happened. So let me gather. Come back to you with a better answer, I guess, is my response.
- [47:51]
Ralph Nader
But what we know already, Steven, is a study out of England that shows of doctors who use AI in their practice and over a period of time, their talents atrophied because they're relying on AI. So that sort of feeds more dependence on AI to become more prevalent. And of course, in the area of bookkeeping and accounting, I think you can make a strong case for the accelerated capability of AI with numbers that they're going to replace a lot of bookkeepers and accountants.
- [48:25]
Stephen Witt
Oh, yeah, a lot of. A lot of paralegals, a lot of marketing and design, a lot of copywriting. I think ultimately a great deal of customer service, especially remote customer service interactions, will go to AI. I think a lot of medicine over the longer term might, especially telemedicine, might start to go to AI. I mean, you can kind of just take any field, ask what you would do if you had a super intelligent virtual robotic assistant at hand and ask how much of your work you would outsource to it. In some fields, it basically looks like 100%. You know, maybe what will happen is that, and this would be kind of a utopian condition, we become more interested in interacting with one another. Like, that's the one thing I can't replace. You know, there's certain jobs that seem almost AI proof to me, like, say, live theater or early childhood education, you know, circus clown. So maybe those jobs take off. I'm not really sure, to be honest with you, Ralph. And I've seen so much speculation on this point. And as I say, I do want to go collect the evidence. I do think the impact will be transformative. I do worry about widespread unemployment or layoffs, but I gotta see if it's real.
- [49:37]
Ralph Nader
Well, it's gotta feed a public debate on what happens when millions lose their jobs. Well, one proposal is you have a minimum income policy, which Nixon actually sent to Congress and never passed. Another is you give them capital investment in the technology that replaces them so that the rewards are distributed.
- [49:59]
Stephen Witt
I like that one. I don't think the equity holders would ever go for it. But, you know, my feeling is if we're going to have all this stuff, that we need some kind of stake in it. I've always been more of a fan of that approach. I will say I think it's unlikely given the current political reality, but ideally something like that would happen.
- [50:19]
Ralph Nader
Well, we're out of time. We've been having a fascinating conversation with Stephen Witt, the author of a great article in the New York Times and another one coming up soon. And I hope we didn't discourage you too much, listeners. We have some amount of time to grab a hold of this phenomena. And when the senators, representatives go around shaking hands in their prolonged recesses back home, start a conversation on this and say, do you understand what's going on here? If you had hearings, are you going to have hearings? How about a town meeting? We need massive public education here. Thank you very much, Stephen Witt.
- [50:57]
Stephen Witt
Thank you so much.
- [50:58]
Steve Scrovan
We've been speaking with Stephen Witt. We will link to his work@ralphnaderradiowaru.com up next, Ralph's takes on this week's elections. But first, let's check in with our corporate crime reporter, Russell Mokhyber from the National Press.
- [51:11]
Russell Mokhiber
Billy in Washington, D.C. this is your corporate crime reporter Morning minute for Friday, November 7th, 2025. I'm Russell Mulcaiber. The corporate and billionaire donors to Donald Trump's gaudy Bank Ballroom Project are beset by conflicts of interest. That's according to a report from Public Citizen. The report found that two thirds of the 24 known corporate donors have recent government contracts for projects totaling $279 billion over the last five years. Lockheed Martin is the largest of the government contractors with $191 billion in federal contracts over the last 55 years. These giant corporations aren't funding the Trump Ballroom debacle out of a sense of civic pride, said public citizen Rob Weissman. They have massive interests before the federal government, and they undoubtedly hope to curry favor with and receive favorable treatment from the Trump administration. For the corporate crime reporter, I'm Russell Mokiver.
- [52:07]
Steve Scrovan
Thank you, Russell. Welcome back to the Ralph Nader Radio Hour. I'm Steve Scrovan along with David Feldman, Hannah and Ralph. So, Ralph, there were special elections all over the country this week, and it looked like a bit of a Democratic sweep, including Democratic socialists winning the mayoralty of New York City in Zoran Mandani. What do you have to say about all of that?
- [52:29]
Ralph Nader
Well, they can thank Donald Trump. His wrecking ball is hitting home where people live, work, and raise their families. And the fear of millions of people losing Medicaid, the fear of 42 million people losing food, supplements, children, the fear and anxiety that is daily fulminations generate firing, cursing, slandering, promising more of it. The increasing price increases from the tariffs coming home to bear and other measures that are too time consuming to go through are finally sinking in and waking up the Democratic voters to turn out. See, the 7 million of them who voted in 2020 for Biden stayed home in 2024. So they turned out. As far as Mandani is concerned, his description of democratic socialism is puzzling to me because all he really is proposing are mostly consumer protection. The consumer groups aren't getting any credit for that over the years. You know, rent stabilization, daycare, free buses. Alexandria, Virginia has free buses. What's the big deal? There are free buses in other cities, not to mention Western Europe. And the contrast is that although he calls himself a democratic socialist, he's still refusing to come out to end the rebate back to the brokers of $50 million a day in the stock transaction tax by New York State, which his colleagues in Albany have filed a bill to get rid of. So the rap on Mandani that he's radical is nonsense. He's very attractive candidate. He means what he says and says what he means, but by no means is he a radical in terms of a democratic socialist. The other thing is that the huge rallies and cheers, I've seen that before. You know, we filled Madison Square Garden at $20 each and the Boston Garden in the presidential runs. Don't be fooled by all the cheerleading and the applause. The problem is, do these people then drift away and get preoccupied with their daily lives without providing any time for the civic mobilization that Mamdani is going to need in City Hall? Because he's going to have to face the real estate industry, it's going to have to face Wall street, he's got to face the AIPAC lobby, he's got to face renegade Democrat residuals that supported Cuomo, and he's not going to be able to do it by himself. He doesn't have the authoritarian dictatorial powers nor would he want them that Trump is exercising. And although he has a progressive City Council, he's going to have to have organized effort in every borough in New York to get even his preliminary agenda through. All in all, it was a great start in terms of the resurgence against tyrant Trump. And it made him look even more foolish when he said the reason why the Democrats swept is that he wasn't on the ballot. The reason why they swept is because he's a dictator in the White House who hates all kinds of people in our country and is depriving them of all kinds of necessities and securities of life.
- [55:51]
David Feldman
I had a quick question, Ralph. There was a proposal on the New York City ballot to move local elections to presidential election years and it was voted against by more than 100,000 votes. The proponents said that it would increase voter participation if everything was in presidential election years. Opponents said that it would do the opposite of that. Do you have any thoughts on that proposal to consolidate local elections into presidential election years?
- [56:23]
Ralph Nader
Yeah, I favor the present system. I favor more frequent elections. That's the problem with the U.S. senate. You know, you can't get on them for six years. You can scare the House more than you can the Senate, because Senate House every two years. So you get the pulse of people more frequently and the reaction to current events. If you have this kind of system than just consolidating everything in the presidential race, especially since there are more and more emphasis on the personality of the presidential candidate than ever before.
- [56:54]
Steve Scrovan
Well, thank you for that, Ralph. I want to thank our guest again, Stephen Witt. For those of you listening on the radio, that's our show for you podcast listeners, stay tuned for some bonus material we call the Wrap Up Featuring Francesco de Santis with in case you haven't heard, a transcript of this program will appear on the Ralph Nader Radio Hour substack site soon. After the episode is posted.
- [57:15]
Narrator/Announcer
To order your copy of the Capitol Hill Citizen Democracy Dies in Broad Daylight. It's at Capitol HillCitizen.com and remember to.
- [57:24]
Steve Scrovan
Continue the conversation after each program. You can go to the comments section@ralphnaderadiohour.com and post a comment or question on this week's episode.
- [57:33]
Narrator/Announcer
The producers of the Ralph Nader Radio Hour are Jimmy Lee Wirt, Hannah Feldman and Matthew Marin. Our executive producer is Alan Minsky.
- [57:41]
Steve Scrovan
Our theme music, stand Up, Rise up, was written and performed by Kemp Harris. Our proofreader is Elizabeth Solomon.
- [57:47]
Narrator/Announcer
Join us next week on the Ralph Nader Radio Hour. Thank you, Ralph.
- [57:51]
Ralph Nader
Thank you, everybody.
- [57:54]
Stephen Witt
Stand up, step up, step up. You ought to step up.
- [57:59]
Ralph Nader
Rise up, rise up. I know you want to rise up. Stand up, you. Hi, this is Halle Berry, and you're listening to KPFK 90.7 FM.
- [58:38]
Stephen Witt
This is John Crumshow with a special politics or pedagogy education report. We're back again with Dr. Freddy Perez. He is an optometrist who has been involved in the Lions Club for many years. I wanted to ask about the testing.
- [58:51]
Ralph Nader
That you do at schools here in California.
- [58:54]
Dr. Freddy Perez
There's been mandated that children, by the time they enter first grade, which would be at the age of six, they should have an eye exam because it's very well determined that the lack of good quality vision impairs a child for learning. So if you can't really see words, you can't see objects. It's very difficult to make the connection and be able to learn. So consequently, one of the things that we rely on as part of our service to the community is to provide screenings to children at the elementary school level. Some schools have the resources to do that through private means, but a lot of the schools don't. A lot of the schools, even public school, don't have those kind of resources. So the lions come in and fill that void. We have special machines that are very portable and we could bring them and do one classroom, two classrooms, four classrooms, all one classroom at a time. And it's very, very.