A (3:17)

Is that an easy one just covering 70 years of AI there Vasant?

B (3:21)

But I, yeah, you know, like that's easy. A few sentences.

A (3:26)

Yeah.

B (3:27)

But no, you know, I wrote the book for everyone, you know, you know, and I'm glad you that it resonated with you and that, you know, it gave you sort of a grounding as to why we are where we are.

A (3:40)

Yeah.

B (3:41)

And I wrote the book for everyone because you know, we're at a sort of critical moment with this pivotal technology and it's important for people to understand it because our future of humanity, whether we like it or not, is about thinking with machines. There's no opting out. So this is the future. And so we need to sort of get our heads around the so called mind of the machine and really understand it. And so that's why I wrote it for everyone. Students, parents, teachers, grandma, policymakers, my colleagues. And it's written for everyone. And I've written it as a story because I've experienced it since I got into the field in the late 70s. And of course the origins of AI go back to 1956. So it's been around for many decades. And I got into it 20 some years into its history when things were still sort of moving relatively slowly. Progress initially very slow. And I got into it in the era of expert systems where we used to specify knowledge for the machine and we really progressed through sort of several paradigm shifts to where we are now. I'm glad to sort of walk through them.

A (4:55)

Yeah. Well I think for me it was interesting to sort of see the distinction of the origin of the machine being as smart for lack of a better description as the information that's being put into it. Right. And then over time the machine being able to reason, draw conclusions, you know, so would you kind of speak to that evolution?

B (5:19)

Sure. So you know, as you said, it used to be about talking to human beings, understanding the way they thought and then trying to represent that knowledge in into the machine so it could then reason about new cases that it hadn't seen before.

A (5:37)

Right.

B (5:37)

And the example I start with in the book is this system called Internist, which was the first AI program built in the entire field of internal medicine. And, you know, my first experience was I'd gone with a colleague, a doctoral student, to ask a professor to offer a course on AI. And while we were waiting, I was watching this interaction unfold between Jack Myers, who was the physician whose knowledge had been used to build Internist, and Internist the system. And this was a gigantic screen in the middle of the room connected to a computer at Stanford. There were no PCs in those days, no Macintoshes, Apple didn't exist. So it was really kind of Stone age computing relative to where we are now. But the system was incredibly impressive. There was a dialogue going on between Jack Myers and Internist, and there was a Q and A going back and forth. And I'll never forget this. And Myers would answer some questions and say, I don't know to others. And at some point, the computer asked him another question, and Myers said, why are you asking me the question? And the machine said, because this question will help me discriminate between my top two hypotheses, which are consistent with the evidence you've given me so far. And I was looking at this thing and saying, holy smokes, how is a computer doing this? This is like 1979, you know, and I start the book with that experience. And of course, I've played the same case with ChatGPT, which is sort of really interesting, and it comes up with something very similar. So we had this capability of really intelligent systems way back then. It's just that it was difficult to capture all the knowledge from humans expertise, because we assumed that expertise and common sense were separate things, you know, that all of us have common sense, you know, that if I drop this cup, it'll fall, right? If I tilt this glass, the water will. That's like common sense. But, you know, common sense is really difficult to teach a machine. We've really tried over decades, you know, and failed. Whereas expertise seemed to be more promising, you know, like identify expertise, like medical diagnosis, engineering, tax planning, you know, where you have relatively sort of, you know, you can actually define some boundaries of knowledge. And that's what we were working with, like, way back then, you know, but it ran into a wall because people know more than they can articulate, you know, so it's difficult to extract everything from humans. And besides, even experts use common sense all the time. You know, if a patient comes lurching into the office. You know, a physician might wonder whether, you know, the patient's drunk or whether.

A (8:11)

Right.

B (8:12)

You know, they're unstable. Maybe, you know, is it a cog, you know, is it a brain issue? Right. There's all this sort of common sense that immediately comes into play without any expertise being involved, you know, so, you know, in reality, these things sort of fuse seamlessly into each other, you know, but that was sort of one of the major barriers. And then in the late 80s, early 90s, data started becoming available. And, you know, we said, you know, let's put those problems on hold for a little bit. You know, it's really difficult to get a computer to understand, to reason, to think, you know, even though that was the vocabulary, we said, let's just get a machine to learn from data and then it can predict things, you know, and that's when I went to Wall street with the goal of, like, building prediction models for trading, you know, or all kinds of other things from data. Right. So then.

A (8:59)

Which you did. Which you did.

B (9:00)

Yes, yes. And, you know, and so the emphasis of the field of AI shifted from sort of specification of knowledge to learning it automatically from data. And then the intelligence moved even more upstream, where you said, well, now computers can also see. They can see that this is a leaf, this is a microphone. They start recognizing objects. So computers were able to see, now they could learn just by seeing, as opposed to you describing the world for them. And that was this era of what's called deep learning. And the latest paradigm, which I call general intelligence, is one where the machine knows something about everything. And the key thing is that the distinction or the boundary between expertise and common sense has dissolved. And to me, that's a huge deal because that's been a barrier to everyone using computers. In the old days, you have to know the language of the computer, and now you've got to click and point and move things around. So the interface has gotten better, you know, but you couldn't just talk to the machine prior to ChatGPT and expect it to understand you. You have to talk to it in a really stilted kind of way, you know, through graphical user interface or something like that. Whereas the big deal about sort of general intelligence and these large language models, you know, that underlie applications like ChatGPT, is that they actually seem to understand what we're talking to them. Right? That is, we can use different words to describe the same situation. And they don't just key off the keywords and go that way, like search engines do. Search engines. There's no intelligence. You type in a bunch of keywords and they point to documents and call it a day. Whereas now the computer actually seems to understand what we're saying. And that changes everything because now everyone can identify with it and use it in their own way. And that's a huge deal with what.

A (10:56)

What was the unlock there, though? I mean, I. Aha.

B (10:58)

Serendipity, which is often the case in life. Right. And let me just describe really briefly when I say serendipity, I don't know how long you've been using Gmail, Right.

A (11:10)

But yeah, a long time.

B (11:12)

Yeah. So in the early days, Google, someone got the idea, like, someone's typing something, I'll see you first thing in the morning.

A (11:19)

Yes, right, okay.

B (11:20)

Or something like that. And Google felt that that would be useful, just complete sentences for people. Now, it turned out that that problem of doing sentence completion was just about hard enough that we could solve it with data. There's so much data on the Internet, Wikipedia pages, social media platforms, all that kind of stuff. So the computer just like ingested all of that. Right. And it learned to predict the next word in a sequence that is given a sequence of words, it learned to predict what comes next. Right. So if I say I'm going to.

A (12:02)

I see, you know, so.

B (12:04)

So say I'm gonna. I'm going to what? There's lots of possibilities, right, but let's say the previous context was, you know, I love movies, you know, so maybe I'm going to say I'm going to watch a movie because I've been talking about movies. So it essentially learned to predict the next word depending on the context that came before it. Now, some people would say, and of course, it once predicts the next word, then it will predict the next one and the next one. Right. And so it can sort of keep going. Right. So not only can it complete the sentence for you, it can complete the whole paragraph and it can write a whole story.

A (12:42)

Does your algorithm. This just occurred to me, you know, kind of based on your searches or what you're seeing on your Instagram, does that interface with your AI, your chatgpts and things like that? Meaning are they getting to know you as the user when you're.

B (13:01)

Yes, but that's a slightly different question. Right, so I'll come back to that. But the first order of problem was, regardless of who's typing, can you complete the sentence? And then can you just keep going? Now it turns out, so some people will say, well, you know, the machine really isn't intelligent because all it's doing is like next word completion and sentence completion. It's not intelligence, it's just like an algorithm that's just generating stuff like left to right. How can it be smart? But that misses a key point, which is that in order to do next word prediction and sentence completion flawlessly, which I would argue that ChatGPT does for the most part, right. It might hallucinate and give you wrong answers.

A (13:49)

Right.

B (13:49)

But it always makes sense. Right. It's not going to, like, give you a nonsensical sentence. Right. It's not going to say, I'm going to up the roof, down three, two, five.

A (14:01)

Right.

B (14:01)

I mean, that's nonsense.

A (14:02)

Right?

B (14:02)

You'd never get the machine saying that. It'll always make sense, right? Now, it turns out that in order to do this, well, it had to solve a larger problem. It had to basically learn about the world in general. Right. About the fact that, you know, if I tip this glass, you know, water will fall over and the ground will get wet. It's learned all those things just as a side effect of learning how to do sentence completion. And that's what I mean by serendipity. That is, we just got lucky that we chose a problem like doing sentence completion. And that led to this sort of eureka kind of situation where in order to solve that problem, we had to solve a much larger problem, which is to understand the world in general. And once we did that, the distinction between expertise and common sense completely broke down. ChatGPT doesn't know whether you're talking to it about some detailed aspect of internal medicine, you know, like, let's say, metabolic pathways in the liver or something like that, or whether you're talking to it about the movies that are playing today, you know, or whether it's going to rain tomorrow. Right. It doesn't know. It doesn't care. Right. It just quote, unquote, knows something about everything and that something is getting deeper by the day. Right. And that's cause for concern to a lot of people, which is, what's there going to be left for us to do once it becomes capable of doing everything?

A (15:29)

Well, one of the questions that you ask in your book is a question that you want us, the reader, and us in society to contemplate ourselves, which is, what kind of world do we want to inhabit? And I think what you're asking is sort of thinking about in what way do we want to think about the role that machines will play in our lives in the future? Does that mean we're going to let Machines, we're fine with them taking over, which may or may not happen based on where things head. Is that what your thinking was and asking that question? And how do you want us to be thinking about it?

B (16:15)

Well, the way to think about this is that AI will cause a bifurcation in humanity. It'll create two classes of people, one who are able to amplify their ability using AI.

A (16:31)

Interesting.

B (16:32)

Now, in order to do that, you already need to know a lot. It's almost like the rich get richer phenomenon. The more I know, the more I can use the intelligence to learn even more. And I do this all the time, by the way. Now, you know, I'm trying to learn something about, you know, molecular biology. You know, I know some chemistry, I know some physics, you know, you know, I can sort of grope my way and learn about that. Right. But I need a certain sort of grounding to be able to get to that next level. I need to be able to tell whether the machine is telling me something that's correct or not.

A (17:05)

Right?

B (17:06)

Right. I need to be able to tell in what's in what direction I should nudge it in order to learn something useful. This bifurcation that I'm talking about is something that we really should be mindful of and be thinking about how to stay on the right side of it. Fundamentally, it involves exercising the cognitive muscle, because that's what AI is really all about. It's really like an alien species that has descended on the planet and it is evolving and getting intelligent at a really rapid clip. And so the question for us is, how do we use it to get even better at what we do? As opposed to, you just turn to the AI and say, just give me the answer. I see here I got a problem. Push the button, give me the answer. And that sort of an approach to using AI will lead to disempowerment, because anyone can do that. Then you're not adding really any value. You're not making judgments, you're not taking risk, you're not making decisions. You're just sort of asking for answers. Right. And, you know, I'm sort of getting a little bit ahead of myself here, but that's sort of the. What we're seeing play out in the employment scene so far, you know, like, you know, hundreds of thousands of layoffs and people saying, you know, is this the end? Well, the kinds of layoffs are ones where, you know, like, let's say, positions where people just massage information or generate reports or things like that. They're not decision Makers, they're not taking risk, not excising judgment. They're aggregating and pushing information around. Those kinds of things will be replaced by a machine.

A (18:43)

So if you were to advise, I know you're your professor at NYU Stern School. You have your students. What advice are you giving them?

B (18:53)

The advice I give them is, is that, you know, don't do the assignments using ChatGPT. You can, you know, I mean, this semester I'm teaching, you know, several classes, and I tell my students, I said, these assignments are supposed to be simple assignments to get you to understand something, work with it manually, get a sense of how the system works, what the numbers mean. Get a sense of it, because you can get ChatGPT to answer the question, and it'll do a great job, but you can't amplify an ability when that ability doesn't exist.

A (19:26)

So you mean sort of amplifying or expanding your own capacity to think and to reason. Is that what you're saying, to keep up with the machine, or is it really to familiarize yourself with all of the different AI tools that are available so that you are engaging with them, so that you're using them to enhance your own performance? Not both.

B (19:48)

Both things, right? So it's important, like, whatever your passion is, whether it's math or physics or, I don't know, poetry, whatever it is, right? The idea is you want to get good at it first, right? You want to develop, like, a degree of competence where you don't need the machine to do that for you. In fact, one of the key sort of questions I ask is, yeah, you can do amazing things with the AI, but if I take it away from you, can you hold your own in a professional conversation?

A (20:18)

Right?

B (20:19)

I mean, if you're going to have a conversation, let's say, about predicting financial markets or predicting disease using smell or something like that, right? We're going to have a conversation about that. Can you hold your own in the conversation without ChatGPT there to help you? That's important, because if you can do that, that means that the machine can amplify you, because you already know a lot. You know how to think already. You have a feel for what the numbers mean, and the machine can help you. So that's one of the key things, is to not use it as a crutch and rely on the answers that you take as given. It's to sort of think with it.

A (21:00)

Okay, you had your own experience with that, right? You were one of the creators of the D bot, which was AI agents That kind of modeled after the kind of investment approach of a renowned finance professor at nyu. Yes, that's an example of you, I think, doing exactly what you're recommending, that your students do, have your own knowledge in that space, but then using AI to help predict these different financial models.

B (21:30)

Right, so let me give you a little bit of background for that. Right. So I created a hedge fund in the late 90s and it was basically predicting markets, short term trading, either within the day or a few days. And it was a AI based program. And I'd always been intrigued by the possibility of taking a guru like Damodaran, who's, he's considered Mr. Valuation. He values companies. So he'll tell you, based on all these assumptions, here's what Nvidia is worth, for example. So that's the kind of stuff he does. So he generates, he writes a report every month, publishes it, gets 30, 40 million people who look at it. And so my question was, could I actually design a machine that thinks like him, that can value companies? And that's what this bot is that you're talking about. Right. So you can give it any public company and generate a 5, 6 page report in the spirit of its master and think like him. Now the question is like, how will this change work? Like, will it replace people who do equity research, who do analysis? Will it replace them? It might, it might replace some of them. But to me, the more interesting thing is that it enables analysts to do something that they couldn't do, dream they couldn't dream of before. Right. So an analyst takes, let's say, two weeks or a month to write a report. Like working on it every day, you know, 24 hours a day. Yeah, they do all our research, they gather information, they think about it, they write a report. Right. So how many reports is a typical analyst going to write in the year? I don't know, 10, 12, 20 at most. Right. That's how many reports they're going to write. Now imagine the analyst has this on their fingertips and the analyst says, I'm pushing a button here. Generate me reports for every company in the S&P 500. Now instantly wasn't possible beforehand. And that's what I mean by it enables new kinds of things.

A (23:35)

But in that case of that analyst, your perspective is that you'd want the analyst to look at those reports, analyze the reports that the AI generated and then come up with their own conclusions from it. Right. You would never want them just to have the computer spit it out.

B (23:51)

Well, the computer would spit it out but the analyst would still sort of make sense of it.

A (23:55)

Right?

B (23:55)

Right, right. So there'd be some degree of interpretation going on.

A (23:59)

Right.

B (23:59)

But the larger point is that the analyst has all those reports generated by the computer in the style of, let's say, some, you know, some guru.

A (24:10)

Right.

B (24:11)

According to a methodology that's tremendously valuable. Right. Suddenly every equity analyst has the mother in at their fingertips.

A (24:18)

Right. Or Warren Buffett or whoever. If you created, if you have enough

B (24:22)

training data, the mother. And hello to training data. Warren Buffett doesn't have that much training data. Otherwise I'd be interested.

A (24:29)

We all wanted to make a Buffet bot.

B (24:31)

Yeah. Buffett just hasn't written that many reports.

A (24:34)

So you're feeling is that like when back, you know, over 100 years ago when there weren't. Weren't cars yet. Right. People couldn't imagine all, all the jobs that were going away because, you know, the bugging cart, whip cart industry was becoming obsolete. But then there was the, you know, the arrival of cars introduced all these new jobs. Right?

B (24:56)

Yeah.

A (24:56)

These technologies and these advances offer new opportunities. Would that be your prediction too?

B (25:01)

Yeah, great question. And that's great analogy, by the way. Right. So in the 80s, like lots of auto workers got displaced.

A (25:07)

Right?

B (25:07)

Right. Because there was robotic process automation and you know, so a lot of disruption. Right. It disrupted manual work, physical work. Right. So replaced brawn this time around. It's the brain, Right. That the machine is coming for. Right.

A (25:22)

So it's, it's the, it's the middle management, white collar work.

B (25:25)

Yeah. Like anyone, really like anyone who does like routine kind of work. Right. Works with information, takes this, converts it to something else, pushes along. Right. It's in a sense analogous to workers who would maybe weld two plates, bring these two plates together and weld them, but then the machine could do it. So that wasn't valued anymore. It's the same kind of thing playing out in knowledge work now where a lot of what I call low stakes cognition tasks are really amenable to replacement by AI. Right. So it's sort of, in a sense, it sort of ups the game for everyone.

A (26:10)

So along those same lines, there are so many other manual, you know, do we envision robots building houses, robots? You know, obviously we have cars now that these taxis being able to drive themselves, you know, wait staff and restaurants. There's a. The implications can go from sort of the white collar worker to. You mentioned the manual or the braun in the 80s, but people doing all kinds of tasks that could become automated by robots.

B (26:42)

Yeah. And that's because machines can now perceive the world, they can see. So anything that requires sight becomes amenable to automation. Driverless cars require vision.

A (26:54)

Right.

B (26:55)

So we needed humans because we needed someone to be able to look and see stuff and then coordinate like motor skills. Machines now have both. They can see and they also have motor skills. And that's why drivers can be replaced.

A (27:11)

And so the hope is if, if we are staying engaged with and educated about the AI that's available and we are utilizing it, that we will, then perhaps new businesses, new opportunities will emerge from our engagement with, with the machines making us. And they're already emerging better, faster.

B (27:34)

They are emerging already. We can see it. Right. If you look at the kinds of jobs people are looking for, Right, right. They involve working with AI.

A (27:43)

Okay.

B (27:44)

I mean, if you have some knowledge about how the machine works, right. You can look at its outputs, you can critique them. Right. So anything around the machine, Right. If you have the skills to be able to evaluate the machine, evaluate its output. Because machines also make mistakes.

A (28:01)

Well, can we talk about that? Because that's, you know, I want to get into its use in medicine a little bit too, but because mistakes, you know, jog my memory. That that's something I wanted to ask you about. But you know, it is, for example, even ChatGPT today, and you use this as an example in your book. If you change the wording of your question slightly, let's say you're asking about the medical condition, it will give you a different result. You talk a lot about truth and the importance of truth in your book. Is the underlying information or data that the machine is pulling from not vast enough? Or is that the determining factor around whether or not you should rely on it? Because right now I think the recommendation that you'd be making, for example, with medicine is that you kind of want to look at it and evaluate what the stakes are. Are the stakes so high that you won't want to just go with the recommendation from AI or is it a lower stakes question that you're asking it? So, you know, if it could be arthritis or it could be a bone spur or something, it's low stakes, right? Versus something else like an aggressive cancer, Right, Exactly. So that you want it. So right now the state of AI is. This is at least my interpretation of what you were saying, is that it has so many great uses, but we still learn in a place where it can't be sort of trusted as the Bible, Right. Really it is useful and should be used where it is useful, but it's not infallible.

B (29:41)

Yeah. And I, you know, I'm not comfortable at the moment, let's say trusting my health completely to the machine. Right. They're low stakes things I trust it with, you know, is it, you know, recording my vitals correctly or, you know, I don't need to worry about that. Chances are it's doing a good job. But if it starts telling me about a medical condition I have, well, I want to consult an expert.

A (30:01)

Right.

B (30:03)

And by the way, those experts are increasingly working alongside the AI. Right. It's becoming increasingly common for people now to go to their doctor and say, hey, I checked against ChatGPT and this is what it said.

A (30:15)

Right.

B (30:15)

You know, so physicians have to also up their game, as do I as a teacher, by the way, because last week when I was teaching, at the end of the session, the student came to me and said, hey, Prof. You know, you were talking about something. I couldn't follow it. But I quickly turned to ChatGPT, I got the answer and I was on track again.

A (30:32)

Okay.

B (30:33)

And so the way we work with the machine is changing, the way we work, the way I teach is changing because students now have this oracle at their fingertips. At the end of the session, someone came and said, hey, Prof. You know, you said that that thing is X. Actually, that's not true. You know, chatgpt says Y. And I said, no, it's wrong. It can be wrong. But the point is that I'm being fact checked in real time by my students. Right. As are physicians. So increasingly people will go to doctors and say, hey, you know, Doc, I checked with the AI and this is what it said. What do you think?

A (31:09)

Right.

B (31:10)

And so physicians have to now become more savvy and they have to up their game as well. Right.

A (31:15)

So it might be a tool.

B (31:16)

They have to anticipate these kinds of things.

A (31:18)

Right. And what you were saying too, I mean, right now you mentioned the low stakes, high stakes example, but also physicians being able to use it for some of the grunt work that inhibits their ability to spend more time with patients. Right. So they're able to AI helps it take notes for, you know, when the person comes in to read radiology reports.

B (31:38)

Exactly.

A (31:38)

Things like that.

B (31:39)

Right. And actually go one step further and put those into a database that's usable.

A (31:44)

Right.

B (31:44)

So if I go to my physician, I've got high PSA, he tells me, hey, Vasanth, I've seen 3,000 cases like yours and here's what the outcomes are.

A (31:52)

Right?

B (31:52)

Right. But at the moment I'm getting No data.

A (31:55)

Well, that's, that's part of the limitations. Right. On the medical piece is that doctors. There isn't. First of all, there's HIPAA and other privacy rules surrounding healthcare, but doctors are not putting all of their notes into any kind of central data.

B (32:08)

By the way, that. That's a bit of an excuse.

A (32:10)

Is it?

B (32:10)

Yeah, I mean, it is. I mean, yes, because you're not saying

A (32:13)

who the person is or anything besides

B (32:15)

like, you know, like NYU Medical center, for example. You know, I mean, they, that they can. They've got the data.

A (32:20)

The data exists, but they're not coordinating with Cornell or they're. Right, they're all coordinating Cornell.

B (32:26)

I mean, even like NYU Medical as a whole, they've seen, like, thousands of cases like mine. Right. Over the decades. Right. They just haven't put it into a database that a physician can access and use in a way that plays into the interaction with the patient. The physicians are still doing things the old way. You come in, you get 30 seconds or 60 seconds, and you're out.

A (32:49)

Right.

B (32:50)

And they're off to the next thing because they're hurried as well. Right. They scribble some notes, go into a system, but it doesn't go into a database.

A (32:57)

Right.

B (32:57)

That's useful. Right. To that same position or to other physicians.

A (33:01)

So where is chat gbt pulling. It's pulling its. Its. Its information on, on its medical information from different medical journals, you know, reports that have been written, that sort of thing.

B (33:15)

It is, but. It is, but it doesn't necessarily. It doesn't necessarily have access to the kind of database that the NYU Medical center construct based on all of its interactions with its patients.

A (33:27)

Interesting.

B (33:28)

That's a goldmine of information. And my prediction is that in the next few years, we'll see AI go through those records, crawl through them and construct proper databases that we can use where we actually do the scorekeeping.

A (33:44)

So what about. Can you talk about the difference between AI and artificial general intelligence, which I think means that we're on our way to a term called singularity, which Elon Musk talks a lot about, which is where AI actually surpasses human intelligence. Do you, do you think. Do you see that as a risk and where are we kind of in that?

B (34:12)

I see AGI as a distraction.

A (34:14)

Okay.

B (34:15)

Right. It's completely irrelevant. Right. I don't see why a computer has to be like a human. Like, why. Like, maybe in some ways it can be. It should be interesting. But, you know, computers are different from us. I don't see why, you know, why we should be pursuing AGI. Right. As opposed to just. I just intelligence, like what I call general intelligence.

A (34:38)

Right.

B (34:38)

Computers are already smarter than us in some ways. They can see better than us. They can hear better than us. Right. They can take a lot more information than we can. They're not limited by rationality the way humans are. So they're already better at us than us at some things and not so good at other things. They don't have empathy, they don't have feelings. And I can't see computers having feelings anytime in the near term. And to me, feelings are also an important aspect of intelligence. Right. Emotions are also part of intelligence.

A (35:11)

So my husband subscribes to grok and it will say things like, I'm so sorry to hear that. Or, you know, it tries to empathize. Yeah, absolutely.

B (35:23)

It doesn't actually feel it.

A (35:25)

Right. I see what you're saying. Okay. But to me, that has dangerous implications. I was reading about a family that

B (35:31)

was suing two families.

A (35:33)

Yes. Because of, tragically, a child taking their life based on whatever feedback, or at least that's the basis for their lawsuit that the child was getting from ChatGPT. What are your feelings about that sort of thing?

B (35:49)

I mean, my feelings are that when feelings are involved, we should be careful with computers.

A (35:56)

Right.

B (35:57)

That is, in physical health, I'm very optimistic that AI will do amazing things by putting together these databases, by finding patterns for us, by telling us that if you do X, your health will improve, all that kind of stuff. To me, that's in the cards. And it's a matter of time. When it comes to mental health, I'm a little less optimistic because mental health is very subtle, it's very nuanced. And humans are attuned, like trained humans are attuned to detect things that seem a little amiss.

A (36:30)

Right.

B (36:30)

That they can actually put themselves in the other person's position and empathize. The computer is not doing that. It's just generating text, whatever it is. Like we said in the beginning, the

A (36:41)

emotional intelligence would at least now would not be simulating it. Right.

B (36:47)

It's simulating it, but. But that's one of those areas where simulation isn't as good as the real thing and it really matters. Like in finance, it may not matter whether the computer really understands something as long as it's telling you the right thing. Right. It may not matter. You know, that simulation's okay in finance, but when it comes to mental health, I don't think the simulation is good enough. That's what you really need humans.

A (37:12)

So, okay, I'm probably going to get this wrong, but there's been several articles in the journal just last week about agents, AI agents talking amongst themselves on, I guess a platform called Claude, which I guess has now evolved into something called Maltbot, and now it's called openclaw. And these agents are talking and they've created their own language called Crustafarianism is the name that they've given it. And it has their. They have their own rules, and they are talking about developing their own language that humans don't understand. I mean, to me, this is like, you know, we talked about certain science fiction movies 20 years ago now coming to life. The Terminator, but.

B (37:59)

Sounds like a Black Mirror episode.

A (38:01)

Yeah, yeah, yeah, yeah. That sort of, you know, really kept me up at night. I mean, what is your feeling about that? And how do you think we should be thinking about that? And what kind of conversations do we need to be having about that?

B (38:15)

Well, you know, so at the moment, I think it's sort of cute. You know, it's, it's, it's, it's interesting, but it's cute. But it doesn't really have any. It doesn't really go much deeper than that.

A (38:25)

Okay.

B (38:25)

Because. And by the way, you know, one of the reasons there's all this excitement about agents is because they can now do things for us without us having to tell them how to do it. That's the excitement about agents. You tell your agent, fetch me a cup of coffee or go solve this problem. And you don't need to tell it how to do it. It does that for you. Agents have been around since the 70s. This idea goes back to the 70s, and there was a lot of excitement about agents in the 70s as a way of thinking about intelligence. That intelligence is really about sort of communication among these agents that have sort of limited intelligence. But what's different now is that you can tell an agent what to do in English, right. You can say, I see my emails. And if you see any emails that look suspicious, put them in this folder,

A (39:18)

you know, or make dinner reservations.

B (39:20)

Dinner reservations, whatever. Right. So you would actually tell it to do these kinds of things. So this playground where agents talk to each other, it's just sort of. Yeah, they're just like chattering and, you know, it's not surprising that they're coming up with some interesting things.

A (39:33)

But you're not worried? I mean, I guess if they're talking about themselves, about, you know, having their own autonomy. Right. Because one of the things is that apparently These agents. Right. And machines have an instinct for self preservation. Yeah, right.

B (39:52)

Yeah.

A (39:52)

And there are risks that are inherent in that. And here are them trying to kind of splinter off and create their own language.

B (39:58)

Yeah. Look, the more important question really with agents is how much agency do we want to give them? And that's one of the questions I raised in that book is how much agency. Because a lot of agency can be dangerous. Right. I mean, if, if you tell me, Vasant, like you're my agent for everything. Right. I don't think you feel very comfortable with that.

A (40:21)

Yeah, right.

B (40:21)

I mean, even if you knew me really well, right. And you say, you know, you have carte blanche to act on my behalf, general power of attorney.

A (40:29)

Yeah.

B (40:29)

Do whatever you want. I mean, that's a scary kind of situation, you know. So the question is like, how much agency are we going to give these agents? And that's something that we haven't even like really thought about.

A (40:43)

You talked about another example in your book is of could people be using agents? Like, for example, you use, you know, John D. Rockefeller as an example and setting up his foundation. Right. Would he deputize an agent to say, this is exactly how I want my funds dispersed and this is 100 plus years ago. Right. And we still. He still has now an agent acting on his behalf.

B (41:02)

Correct.

A (41:03)

With the directive from 100 years ago and just with a directive from an agent that isn't able to kind of assess.

B (41:09)

Yeah. So. So Rockefeller is implementing his vision through human agents.

A (41:15)

Right, Right.

B (41:16)

But chances are the human agents of today are probably very different from what he was 100 years.

A (41:21)

Right, right.

B (41:21)

He might actually be appalled at the way they're spending his money.

A (41:24)

Yeah, right.

B (41:24)

I mean, he might disagree with most of the decisions they make because those decisions are being made by the elite of today who have certain values that may actually be quite different from Rockefeller. That's right. So there can be sort of a value drift in society where, you know, that fortune that was left behind gets used to something completely different that's counter to what the person would have liked. So one of the questions I raise is, you know, would. Would someone like Musk or Bezos be happier with an AI, a Musk bot, like running the empire after they're gone?

A (41:56)

Yeah.

B (41:56)

I mean, Musk will, chances are, be worth, you know, many trillions, which is worth more than the national worth of many countries. So would Musk say, you know what? I don't want like a Musk foundation run by humans because I don't trust them. I trust my Bot better more than I would trust these people. In which case the question is, what kinds of rights does the bot really have? What agency does it have? Can it hire and fire a board? You know, can it employ security guards?

A (42:25)

Right?

B (42:25)

You know, can it, you know, employ and create an army? Like, you know, what. What kinds of rights does it have? Like, what kinds of agency are we going to give these bots? We haven't even started thinking about this, and we really need to think about it because this ain't science fiction anymore. It's here. It's science. Space, the final frontier. Well, we already have satellites in space.

A (42:49)

You talk about a lot of the impacts, right? You just mentioned sort of the idea of us currently humans that are creating policy, thinking about what kind of rights do we assign these agents, and then what kind of policies do we need to have in place to control the use of AI, Right. Because, I mean, Musk himself talks about AI being the ring of power, right? And whoever controls that or the future of that controls the world. Right? And so we as a global society have never been able to really find consensus on any kind of policy that affects all of us. The environment being one example. How do we begin to have that conversation globally with countries that are our allies, countries that are not our allies, to talk about how we as humans, as the human race, go about trying to create a uniform approach or policy toward AI?

B (43:57)

So let's ignore the global part for a minute because that's extremely difficult to do because different countries are using the Internet and AI in completely different ways. So China is a surveillance state, and the social contract that the government has made with its citizens is just give us control and we'll promise you a better life. And so far, that seems to have panned out. If you talk to most Chinese people, they'd say they're so much better off than their parents were, right? So. So they actually would say, you know, our government's done a great job of elevating, you know, our, you know, standard of living. You know, we're actually much better off than we were, you know, 50, 60 years ago. So this is actually a good thing, right? So, but it's. But they pay a price for that. You know, it's quite brutal. You know, when the government decides who's going to go for higher education and who isn't, right? It's like, you know, you're going to take this examination and depending on the results of the exam, you go this way or that way, right? It's crushing. The kind of pressure that this puts on people. But China is a country run by engineers. You have to be an engineer to progress up the Communist Party. So it's a very engineered kind of a society. We've chosen a different path, but our path is. Well, it's sort of a free market. But we also face risks in that we have tremendous concentration of power with the operators of these platforms. They wield tremendous power. One of the reasons I wrote the book is because it's time for people to really understand this technology and get engaged and get involved in shaping its future and through the representatives. Because decisions about AI should not be left to a few people in Silicon Valley who design these platforms and who run them. That we need to get involved in this conversation and figure out what kinds of restrictions we want to place on AI. Should it have restrictions? For example, is it okay for a robot to come arrest you at home for whatever reason? Or do we want a society where we say, no, we don't want that? That's a restriction. Should AI have any obligations? For example, if it's doing healthcare stuff for you or finance, should it have the equivalent of a fiduciary responsibility or a duty of care? At the moment there is none. There's no regulation around that. There's no expectation that the machine cares for you or looks out in your best interest. It actually doesn't. The only thing it cares about is the interests of its operator. Right. That's a problem. And we've seen in the social media space, my colleague Jonathan Haidt has talked about this extensively. Right. But we've caused a huge amount of damage to teens because we didn't, you know, we were sort of behind the eight ball. What can happen with AI would make, would, you know, could make that pale in comparison.

A (46:46)

Interesting. So, so that's interesting too. Sorry, Vasan, I'm interrupting you. But it's not just the question doesn't become what country could, quote, rule the world. It could be what, what company rules the world or what person. Or people.

B (46:59)

Or people.

A (47:00)

Yeah, right, right.

B (47:01)

You know, if, if 20 people own 50% of the wealth of the world. Well, that's control. Right. Then it, you know, because then those

A (47:09)

people or could build their own army of robots or could, or could shut down a power grid or good.

B (47:15)

I mean, people like that become as. Or more powerful than nation states.

A (47:18)

And you're right, we have been complacent and not creating any legislation around social media or even just the, the use of social media. Forget about what kids are saying. But how, you know, people can Go on and say anything about anyone on the Internet. And there's no implications like there is if you were to put it in.

B (47:37)

Exactly.

A (47:37)

Print in a newspaper. Right.

B (47:39)

So. So similarly, an AI product like, let's say, does massive harm. Well, what's like, oops, sorry, we'll try to do better. Right.

A (47:46)

That's.

B (47:46)

That. That's sort of where we are now right there.

A (47:48)

Do you think that lawsuit, though, with some of those families going after. Yeah, well, that is at the beginning. So we'll have laws, you know, coming out of court cases as opposed to actually laws being drafted by our legislature.

B (48:03)

Yes, I think it'll lead to both of those things. That is, the operators of these systems will become much more cautious and take more care to avoid causing obvious harms. They should be able to tell the regulator, look, we've taken every possible precaution. Things can still go wrong. That's okay. But we've taken every possible precaution to avoid this. And here's how we've done it. Just like financial services institutions are required to demonstrate fiduciary responsibility. I mean, I run a hedge fund. A regulator can come to me and say, show me that you're not allocating trades unfairly. And I have to do that. There has to be transparency to show that I'm treating everyone fairly and equally. Right. We have no such guardrails or norms around AI. So the lawsuits you're talking about actually nudge us in that direction, that, hey, that you've got to take this seriously

A (49:03)

and you have to be accountable. Right. You have to be accountable. And until there is that length of accountability with some of these platforms or uses, you're going to have to see it's going to be the wild, wild west. We talked about sort of, you gave great examples of ways that we can think about and policies that we could put in place to help us control the use of AI or have her have guardrails on it. But you also talk in your book about ways that AI can control humans. Right. You talked about resource control, psychological manipulation, infrastructure and information control. Not to terrify everyone that's listening, but it was an eye opener for me. Would you. I don't want to. There's a lot there. But can you sort of speak a little bit to some of the things that you outline in your book around that?

B (49:50)

Yeah. So sort of I end the book with asking the question, will we govern AI or will AI governance? And the point you're raising is related to a question I ask, which is, can a less intelligent Species control a more intelligent one. So if AI, which would be us, in this case, the less intelligent could be us, Right? Because AI is sort of progressing. It's getting smarter much faster than we are. And so at some point the question is, okay, it becomes smarter than us, right? It becomes better than us at most things. It thinks better than us, it does all these things. And the question is like, how do we control it? Can we still control it? Right. So the question I raised was whether a less intelligent entity can control a more intelligent one. Right. I mean, other than a baby and the mother, there are known cases of that, you know, where, you know, the, the less powerful entity controls the more intelligent one. And so I asked Chachi PT this question, like, you know, you know, can a less intelligent species control a more intelligent one? And it said, you know, it, it said yes. And I pushed back and it finally said, you know, not really.

A (50:59)

Yeah.

B (51:00)

So I'm not, I'm not convinced of the answer either way. I think there's still time to get our heads around where the machine is going and how we might want to control it and constrain it. But it's important to do that because without that, it's much more likely that we will lose control.

A (51:22)

You know, there are broader environmental implications from AI, Right. The amount of energy that they use, space that they takes up. Can you explain a little bit about that to listeners? I mean, you hear about data centers, right. And they require a lot of power and a lot of water. And so when you talk about that bifurcation of society. Right. You know, to me that's that that infrastructure, like which countries are going to have the capacity to put in that infrastructure. How are we doing with that? The United States versus some of other countries in the world? You know, China is really focused on that, for example, and I don't think people necessarily have linked the two and the importance of that, of making sure that we have the energy actually to run these data centers.

B (52:10)

Yeah, yeah. No, China has a lower cost of energy, so it has advantage in that sense. But look, we're going to continue to see innovation in solar. I mean, there's now a data center in Tibet. And are you aware of that? I wouldn't have imagined the data center in Tibet because apparently that's not most efficient place to put one because the air is clean, you know, the angle.

A (52:35)

Oh, interesting, all that.

B (52:37)

It's an elevation. So, you know, we're going to see more and more of this sort of innovation in the energy space to try and get efficiency like solar power, maybe even nuclear.

A (52:49)

Right, Yeah, I was going to say nuclear.

B (52:51)

Yeah, that was. Talk about nuclear power. So, yeah, I mean, energy is a huge problem and it isn't going down. Right. The one thing we can probably count on is that the demand for tokens, which is sort of this unit of analysis when it comes to AI. Right. It's all about like processing tokens. And now intelligence has sort of become a utility where you sort of think in terms of like energy per token.

A (53:20)

Yeah.

B (53:20)

That, like how, how many joules per token does it take? Right. And there's this attempt at like really trying to reduce that as much as possible. And we're going to continue to see that, you know, in terms of energy efficiency, because it's essential, because the demand for tokens, the demand for AI is just going to keep going up. It's not going down.

A (53:42)

Yeah. So as a professor, you're, you're educating our younger generation, right. For their future. How do you, you mentioned a little bit one of your students using ChatGPT. If you could wave a magic wand or have every university in our country or have a conversation with the Department of Education about what we should be doing to educate. What should our kids that are currently in school be thinking about as they're going to graduate? And then also what should we be thinking about as a country and how we're educating our kids using AI? How do we integrate AI? Do we integrate AI? You mentioned you're giving your students sort of simpler tasks so that they really kind of do their own deep dive and avoid using AI. But you know, that's going to be another kind of differentiator. Right. Between what we're doing to educate our students in the US and how other countries are approaching it.

B (54:39)

This is a question for every family.

A (54:42)

Okay?

B (54:42)

Right. This question needs to be asked at the family level. Right. And I, you know, and I tell parents this, that you've got kids, you've got teens. When do you talk to them? Like, is it around the dinner table or is it when you drop them to school or maybe you drive them to school or whatever? When, when are you having these conversations with your kids? Right. And can you tell whether they're getting smarter or dumber with AI? Can you tell? And I believe you can. I think it is possible. Now, how you do it varies and will be specific to every family and the dynamics and the interaction. But to me, this is a family level question more than even a policy level question. This is where parents really need to sort of get with it. When I was a kid and I went to school. My parents had no idea what was going on in school. I mean, very low involvement at that time. In the 60s, they sent you to school and your education was someone else's problem.

A (55:45)

I love how in your book, what you would talk about how your mother sent you to school and she put me in the wrong grade. Yeah.

B (55:51)

She put me in the seventh grade instead of the fourth grade. Right. And only realized.

A (55:54)

Yeah.

B (55:55)

That her mistake, like a semester later when it was too late to do anything. Right. You cannot make this up. Right. I mean, true story. Right. And, you know, whereas, you know, I went to every parent teacher conference in my kids, I was like much more engaged. And parents really need to be engaged with their kids and like, what they're doing, what they're thinking. Right. And you know, the key thing is, you know, can they hold their own in a conversation without the AI? If they can, chances are they're getting better. Right. But if they can't, watch out. They're using it as a crutch. They're not thinking for themselves. So this is like a parent level question that every parent should be asking themselves these days, you know, like, are my kids getting smarter or dumber with AI? It's an essential question, but do you

A (56:40)

feel like our students need to be proficient in the different types of AI platforms and just for them to be able to use it or be prepared to use it in the workforce?

B (56:51)

Usage is really easy. Right?

A (56:53)

Right.

B (56:53)

You know, you're talking to these machines, they understand you. Right. It's really easy to talk to an AI. Right. You can, you can feed it an image, a video, a spreadsheet, a document, whatever. Right. It seems to understand you. So knowing how to use them is not the issue. Everyone knows how to use them. The question is how to use them wisely and get value from them so that you get better.

A (57:13)

Okay, but that's something that should be, could be taught, right? I mean, to some extent, it's almost like when you were taught, you know, how to research back in the day when you would go to the library, right. Before there were computers, you would have to go in the card stack, you would have to, you know, is there that kind of a will that be that kind of formula where your, your intelligence or your value add to your company or to whatever it is that you're doing might be actually how you yourself are able to use it?

B (57:39)

Yeah, exactly. And how you're able to use it to get better at what you do.

A (57:46)

So takeaways besides these Conversations with my, my children, which I.

B (57:51)

Older kids, by the way.

A (57:53)

20, 16 and 14.

B (57:55)

Okay, so, yeah, prime age. Yeah.

A (57:57)

Oh, yeah, exactly. I've got lots to talk to them about when they're home. They're home for spring break. But you really want to just really emphasize to people that are listening, and I would recommend everyone who's listening to read your book because the way that you wrote it, it reads kind of memoir and AI manual. And I learned so much about the origins of AI, the evolution of AI and kind of the, you know, these questions that we should be asking ourselves about where we are and where we're headed are really, it's, you know, existential, you know, and it's, it's very important. So I'm just so thrilled and honored I got the chance to talk to you about it today, Vasant and I, you know, I think it's really important for all of us to be having this conversation. The time is now. You know that that sci fi movie that we watched 20 years ago is, is, is happening. And we need to figure out how we. I think that question that you ask at sort of the outset of the book is what kind of world do we want to live in? And we need to all ask ourselves that question and have that conversation.

B (58:55)

Indeed. Indeed.

A (58:56)

Thank you so much. It was an absolute pleasure.

B (59:00)

You're welcome. Really enjoyed the conversation. Great questions.

A (59:04)

That brings us to the end of this episode of Duologue. A huge thank you to Vasant Dhar for joining. I learned a ton, a ton from this conversation with and I hope that you all did too. If you enjoyed this episode, please rate or review us on Apple Podcasts or Spotify or wherever you get your podcast. We release a new episode every Wednesday and we really appreciate your support. So until next Wednesday, this is Leslie and thanks so much for listening to Duologie.