The Munk Debates Podcast - Friday Focus: AI Anxiety – Trump Indictment

Episode Date: March 31, 2023

Friday Focus provides listeners with a focused, half-hour masterclass on the big issues, events and trends driving the news and current events. The show features Janice Gross Stein, the founding director of the Munk School of Global Affairs and bestselling author, in conversation with Rudyard Griffiths, Chair and moderator of the Munk Debates.  The following is a sample of the Munk Debates’ weekly current affairs podcast, Friday Focus. On this week’s edition of the Friday Focus podcast, Janice and Rudyard start the show with a discussion of the recent letter by artificial intelligence experts calling for a six-month hiatus on the development and launch of new technologies. What is the root concern here? How at risk are we of creating a machine intelligence that slips its digital handcuffs? And, if this does happen, what is the likely result? The second part of the program discusses the indictment of Donald Trump on felony charges. What does this event say about the state and future of American democracy? What are the risks of pursuing former politicians with criminal charges versus upholding the rule of law? To access full-length editions of the Friday Focus podcast, consider becoming a donor to the Munk Debates for as little as $25 annually, or $.50 per episode. Canadian donors receive a charitable tax receipt. This podcast is a project of the Munk Debates, a Canadian charitable organization dedicated to fostering civil and substantive public dialogue. More information at www.munkdebates.com. Pause Giant AI Experiments: An Open Letter: https://futureoflife.org/open-letter/pause-giant-ai-experiments/Become a Munk Donor ($50 annually) to get 72-hour advanced access to the full length editions of Friday Focus and Munk Dialogues. Go to www.munkdebates.com to sign up. Hosted on Acast. See acast.com/privacy for more information.

The following is a complimentary excerpt of this week's edition of the Friday Focus podcast by The Monk Debates. To access full-length editions of each and every episode, along with all kinds of great additional benefits and perks, become a donor to the Monk Debates. You can do that for as little as $25 a year, and you'll receive each and every year 50 Friday Focus episodes at full length. It's all available right now on our website. in just a few simple clicks. Triple W. The Monk Debates.com. Look for the Friday Focus option in our navigation bar, the top right of the website.

Make your donation, and we will send you each and every Friday a link to listen to the full-length edition of this program. Thanks in advance for your generous contribution. Hello, Monk listeners. Rudyard Griffiths here, the executive director of the Monk Debates. Welcome to this, the Friday Focus podcast. Each and every Friday, we did. dig into the big issues and ideas shaping the news, moving the public conversation.

We do this with Janice Gross Stein, the founding director of the Monk School of Global Affairs, an internationally renowned scholar and author. Janice, great to be in conversation with you. March 31st. April's coming, but you know, Roger, I've been waiting for a quiet week. This was not it. I know, I know. It's almost, I need a Thorzine drip or something by the time Friday comes around and we record this podcast just to find a bit of intercomness.

But two topics that we have to discuss this week, they're the obvious ones, I think, for many of our listeners. This kind of unprecedented open letter that Elon Musk and other leaders in artificial intelligence released asking effectively for, for a halt on the development of technologies such as chat GPT4. What the heck does that mean? What is the risk? What are they getting at? And then in the second half of the show, let's talk the Trump indictment, not to rehash or unpack the intricacies of American politics and the GOP nomination,

but instead, is this another clue, another sign of something important that's happening in American democracy that we kind of need to. to wrap our heads around here in Canada, but also for the benefit of our American listeners. But Janice, let me start with the open letter by Elon Musk and other really, truly world leaders in AI. You know many of them.

Were you surprised by this? Were you, I don't know, do you share the same level of concern? I mean, there were words in that letter around existential risk. I just, I don't know how to interpret it, so help me. It is a fascinating letter, Rudyard, and we should put in the show notes, a link to the letter for our listeners to read,

because the language was at the top of the alarm register, and who signed this letter, Roger, a thousand people, but many of the leaders in the development of artificial intelligence who, unlike Elon Musk, really have no financial incentive here either way. So what I wish everyone should pay attention to, I think, here is the scientists who have been leading this are now really worried about what they've unleashed. They look back, for example, to nuclear scientists in the late 40s, and they say we should have put a lid on that.

Then we missed it. We're not going to miss the chance now. So they're asking for a six-month pause while we try to get some regulations, some legislation in which there, and the real focus of this effort is to be able to test something before you release it to the public. Let's draw a quick analogy, Roger, to any new medication. It doesn't go from the lab to the public. There is a process there in the states of food and agriculture department looks at it.

In Canada, the Department of Health, where's the equivalent for chat, GBT, GBT, 5, or BARD or whatever it is that's coming next. But Janice, let's talk about the risks here because I'm, you know, I'm confused, okay? There's people that, in fact, we had a debate on the Monk podcast. Gary Marcus, who's a world leader in AI saying, in effect, you know, chat GPT is really a kind of dumb thing. It's a little bit of a parlor trick. It's using skimming millions of websites to imitate human behavior. It doesn't know what a doorknob is. It doesn't have any kind of, well, some people debate this, but, you know, spatial awareness.

It's not intelligent in the way that we understand intelligence, you know, reasoning, agency. So why the worry? Because I can see in a sense the concern that Gary Marcus and others share around the explosion of disinformation, the manufacturing of, you know, viral videos, false images, a voice on the other end of the phone that sounds human and interacts with you is human, but is pretending to be a Trump supporter and you have no idea whether it's a bot or real person. I get all that. But this letter seemed to go kind of one. step further into, should I call it like the Terminator, the Terminator, you know, thesis of, you know, this is potentially humanity's last invention. There was a big debate about what large language models can do. You just presented one side of it, Roger. It is not human. It's auto-complete.

In other words, it just scans billions and trillions of words and then makes a really good bet on what word comes next. That's one perspective. There is another, and there's a really hot debate going. Gillian Hatfield, who heads the Reishman Sports Institute, is on the side that I'm going to tell you about now. She said, no, this is qualitatively different. There's two big issues here.

One, they are capable, the newest models are capable of doing representational learning, which goes beyond just filling in the blanks. That's problem number one. And if they're doing that, that's a level above what we've ever seen before. And the second big problem. Just for a sec, explain representational learning. Does that mean that they're on their own drawing some kind of causality between different things and then and then learning a pattern independent of simply just synthesizing a whole

bunch of patterns. They're creating unique and novel connections between data X and data Y. That's exactly right. They are building new logic chains. Okay. And that, you see the difference between that and autocomplete, which is just what's the likelihood that the next word after the cat is Yeah, which is what we're used to. So that's the first order of difference. The second order of difference, Richard, we can't see what they're doing. These black boxes are opaque. We can't see it. So take a drug, take a new pharmaceutical that goes in for regulatory approval and there's an explanation and the regulators can see inside the black box and they can see why taking more calcium is going to do something good.

for my osteoporosis. They can see the logic change. We can't see the logic chains that are being created inside these black boxes. And that, in colloquial English, is what's freaking out the people, the scientists who are developing the AI. So they want a pause while we figure out the easiest practical first step to put some regulation in place. And we can We can talk about that if you like, but I am very dubious. I'm very dubious that we're going to get that pause. We've never been able to do this. It also, Janice, I mean, this seems like closing, you know, the barn doors after the horses have fled.

I mean, we're months into chat GPT. We've gone from three to four, five is no doubt coming soon. I'm sure the Chinese have all kinds of large language learning models that they've scaled up. We've got, you know, Facebook, meta, Google. with similar products at various stages of release. I mean, isn't this just, as you said, you made the analogy to nuclear weapons? Well, you know, what happened?

Technology, technology is an idea, right? It is, it's easily communicated, easily transmitted. In this case, it literally is code, zeros and ones. How do you possibly contain this? six months really represent and anything significant? So I'm very skeptical. I was asked to sign the letter and I did not because I was so skeptical, Roger.

But let me try to make the best case here for the listeners. So, for example, six months a whole, China is behind in large language models, actually. You know, is way ahead in facial recognition technology because the regime really uses it and deploys it, it's actually behind to the best for knowledge in large language models. So what could we do with the six months? Here's an example.

Require any developer of a large language model to have a license. Okay. License it, like you license a driver for a car, or you can't drive a car without a license. And at the very least, if you license it, if something goes barely wrong, you know, who's developing it, you can at least after the fact try to keep a lid on this. Well, licensing is done at the local or the national level. It ain't done at the international level. And I am so deeply skeptical that any government is going to do this and tie its own hands

while they watch other governments not do it and advantage their own companies. Well, it's clearly the first mover, you know, advantage that you benefit. Let's talk about the kind of existential scenario. I mean, they're painted in different ways, but I guess it's the idea that these programs, and that's what they are, there's zeros and ones, they can have goals that are programmed into them, and maybe those goals are in the moment commonsensical as they're constrained within a computer doing X, Y, or Z writing your emails for you, or calling somebody up and getting a visa number pretending to be you.

That's all the great stuff we can imagine pretty soon in our future that chat GPT and programs like it can complete. But I guess the scarier scenario is that they start to, these programs start to copy themselves, they start to spread that the goals that they're programmed to achieve, let's say, making more paper clips. That could sound kind of harmless, but what if making more paper clips means converting all the matter on the planet into paper clips? Not a good outcome. So to what extent are we close to kind of courting those bigger risks? Because the hypotheses here, which is kind of scary, is that we wouldn't even know that we're courting that risk, that the program

would become, you know, a human general intelligence, maybe start to advance beyond that. And I think history would show the baboon and the human. As soon as you confront something that's more intelligent than you, you've lost. There is no do-over. Is that the key point here, Janice? There's no second chance once you've created an intelligence greater than yourself. History would say, that's it. It's the end of something. There's no question.

That is the big existential risk out there. That's called an AGI and artificial general intelligence, which surpasses human intelligence because we know about all the plots that you, run through the way most of us behave most of the time about most things. It's entirely conceivable that the machine will, and here's the double risk, okay, it will be smarter than us, it will replicate itself and get better and better and better. That's two. And we won't see it.

That's three, because it's happening in ways that are invisible to it. And maybe four, its goals may not be our goals, even though we may have. have implanted goals in it originally. At the beginning. But those goals get surpassed or altered or changed as it makes new causal connections. As it learns, in quote, and just gets smarter and smarter and better. That's the best short summary we can provide to why the scientists are really worried about this. You know, there's a touring test, the famous touring test by the science.

scientist Alan Turing who cracked the enigma coach for Britain. And that was such a crucial advantage that he gave Britain. But he then, after the war, said, and why do we call it the tourintest? He said, we're going to cross the line when a computer can speak to us. And we don't know it's a computer. We think it's a human. Well, we've crossed that line. We're there.

We're there. And we're into another order of magnitude. magnitude, a challenge. Roger, the history, as you just said, of technology is somebody does it. It gets going. It moves much faster than any regulators can move, and we always regulate looking in the rear view mirror. I have no confidence that we're going to be able to do anything else this time either.

But no do-over, Chanis. That's, I guess, what's different, let's say, than the horrible existential risk of atomic weapons. You know, they have the bizarre but effective theory of mutual assured destruction that, you know, prevents in many cases. And we're all dead. And we're all dead then. Right. But in this case, there's no mutuality here. And I guess my contribution with this is that we,

often we anthropomorphize a lot of technology and we assume that these little thinking machines will be kind of happy servants to our ambitions, egos, and desires when in fact, if you listen to computer scientists, they might actually be very alien. Their intelligence may be very incomprehensible to us. I heard one great example recently where someone said that it would be as if we were a thousand years ago and someone sent us a blueprint for an air conditioner and told us how to build it. We could build the air conditioner, but we would have no idea of how it worked. And I guess what they're saying is that we're now building machines that we have no idea about how they work because there are things embedded in them that are a mystery, in a sense,

to us at our, and that's now, that's today, Janice. You know, I think the best thing that we can read to get a feel for what you're talking about, because that's where we are today. There's no question. That's where we are. It's to read science fiction. And what's the late motif now of most science fiction that's written in the last five years? You program these things.

They learn autonomously. They become smarter and better than you. they turn on you and they now control you rather than you controlling the machine. That's the dystopian future that people are most worried about. And that's not tomorrow. But Jeffrey Hinton, who is the creator of deep neural networks, which is what underlies all of us. Without Jeffrey Hinton, there is none of this.

and we're fortunate to have him in Toronto. But he said recently in an interview, I thought artificial general intelligence was decades away, given what's happening with these large language models, I now think it's years. So the pace has sped up at this exponential rate. Where do I get just a sense of restraining? optimism, when the scientists start worrying,

Red your things slow down a little bit. I don't think there's anybody right now that is going to release Chachi PT5 because there is so much public concern. That letter is more important as a signal than it is as a spur of regulation in the near term future. It is a signal to all the developers.

Oh, my God. And like everything else, else these people are one big community who are in touch with one another. Scientists always are. I think that the pace may slow a little bit until there's a much deeper understanding of the risks. That was the purpose of the latter. Let's hope so.

Just end with one story that I read about Chat, GPT, that the versions that we're playing with three and four are tested and they're in a sense locked down. There's a lot of capability and, um, robustness that's kind of dialed back. But when they were kind of future-proofing chat GPT-4, they took the kind of lack of a better expression, the non-lockdown version, the full cranking the dial to 11. And they asked it to effectively escape itself. And on its own, it was able to make a copy of itself, in itself. And then it went autonomous. on its own, onto a chat forum and fooled a human, spoofed a human into thinking it was a human

and extracted from this human coding information and insights that then allowed it to break out of whatever kind of strictures or codes they put on it. It escaped and left, you know, a copy behind. So that's artificial general intelligence. That's the nascent version. Yeah, we're there. Yeah. I mean, that is exactly the nightmare. Once it is, if it can escape itself, right? And it is, and that's the nightmare. If it escapes itself in ways that the scientists don't understand because they can't see those logic chains.

So they don't know, they don't know how to reprogram it because they don't know what's doing inside. And again, it didn't escape itself because it wanted to be free. It doesn't have notions of like agency or whatever, but it was given a goal, which to get out of Dodge, and it got out of Dodge. It escaped itself because it learned. It learned without human intervention. It learned how to do it.

Nobody told it how to do it. That's the point, but it learned. Now, we've never been there before. So this is one of those revolutionary moments, Roger. I'm tired of revolutionary moments, yes. I need a break. Like, let's just take a two, three-year hiatus on revolutionary transformative moments.

Okay, I've had a pandemic. I've had maybe one great financial crisis, possibly we're shooting for two. I've got inflation going on. I've got China asserting itself. I mean, I'm full up, Janice. I can't handle another revolution, okay? Part of what we're talking about is this, the pace

at which this is accelerating. And it is what we call a general purpose. purpose technology. There are very few of them. When they come along, they change. They change the way we live, work, think. They change our societies. And this is just one. There's no question. There's no question. But this one may be our last invention. I think we'll figure it out before we get there, I hope. And that's why this letter is actually important. And we all talk about Elon Musk. who signed it. But, you know, when you're the CEO, multiple companies which use these kinds of technologies, you have to ask, well, is there something in the lab behind that there were, what I really paid attention to here,

are there scientists who created this? And they are saying, hey, oh my God, we need to slow down now to figure this out. So we don't create something that escapes because we tell it to. And then it does things that are existentially destructive. That's really the fear. Yeah. Well, we will continue to follow this debate, and I will continue to look for my Thorzine drip on Fridays

to survive these conversations with you, Janice. You know, we're going to do one segment soon, as soon as I could get it on our agenda on autonomous weapons. Oh, great. On things that people, you know, But they give it the original goal and then it learns it gets better and better and better at his staves, but they're lethal. Because we're close to that too.

Oh, well, another horseman of the apocalypse rides in. Let's take a break when we're back on the other side. Hey, he's not a horseman of the apocalypse, but he's a wrecking ball. Donald Trump had a big, big week this week. Indicted first ex-US president ever in the 200-plus year history of the Republic to face felony criminal charges. We're going to talk about what this means for American democracy, and maybe democracy generally, Israel, France. We are seeing, again, more stresses, more pushing and pulling at the kind of fabric of our democratic societies, maybe than ever before.

We'll have that right after this break. Thanks for listening to this excerpt of the Friday Focus podcast to get full-length editions of each and every episode of this program. simply go to our website, www.com. Click on the Friday Focus tab in our navigation on the top right of the site. Make a donation as little as $25 a year

of 50 cents an episode and we'll send you not only the full-length editions of each and every Friday Focus podcast, but all kinds of special offers, perks, access to events, and additional content. Again, you can do that right now, by becoming a donor to the monk debates at triple w monk debates munk debates m unk

debates with an s dot com