Would you blindly trust a machine that generates the software? Can we trust the machine? Why was I willing to trust my algorithm in finance that was wrong almost half the time? And yet I wouldn't trust it, let's say, in healthcare, where it could be right much more often. The answer now in retrospect seems mind-numbingly simple, which is that it depends on how often the algorithm will be wrong and the consequences of its error.
SPEAKER_00What are some of the trends that you're seeing today when it comes to green shoots of opportunity for businesses of all size?
SPEAKER_02What I'm seeing is like a serious engagement with this problem because I think people do recognize that this is a pivotal technology. It's here to stay. It is likely to be transformative. And the question is, how is it likely to transform my business? Is the sentiment that you're seeing, is it fear or greed? I would say it's mostly fear. You should be skeptical.
SPEAKER_01Vasant Dar is a pioneering AI researcher, NYU professor, and host of Brave New World podcast who has spent more than four decades helping people understand when to trust machines, how AI is reshaping business and society, and why humans must learn to think with machines.
SPEAKER_00Welcome to Using AI at work. I'm your host, Chris Dagon. Each week we'll be learning how today's business owners, entrepreneurs, and ambitious professionals are getting more done with smart use of tomorrow's tech. Let's get started. Right now, every business leader is asking the same question. What are we going to do about AI? If this is you, ChiefAIOfficer.com has the answer. We give you a simple path forward where we provide executive and team training so your people know exactly how to safely use generative AI in their day-to-day. We also manage the deployment and implementation to make sure tools actually get adopted and deliver results. And we'll also guide company-wide transformation so AI becomes part of your operating system, not just another shiny object. The companies that act now will increase productivity, cut costs, and grow faster than their competitors. Those that wait will get left behind. So if you want to make AI work in your business, visit ChiefaiOfficer.com and see how we're helping companies of all sizes finally get results from AI. Hi everybody, this is Chris Daigle, and welcome back to another episode of Using AI at Work. Happen to be doing this one from the road today. We're in Salt Lake City speaking at Aquacon, which the audience is MA family offices and private equity parties that are exploring how they can leverage generative AI, at least in their in their operations and their ambitions with scaling those businesses. But today I'm very excited that we've got a guest Professor Vassant Dar, who is a longtime participant in the AI discussion, 30 years plus, 40 years plus. But he's currently acting as a professor at NYU's business school Stern and also their Center for Data Science. And we're privileged to have him here today to discuss where AI has been and where it's going, especially through the lens of business. So Professor Darr, welcome to the show.
SPEAKER_02Delighted to join you, Chris.
SPEAKER_00Yeah, we're very excited to have you here. You and I had the opportunity to have kind of like a pre-interview, and that opened the the, I guess, my uh curious nature. So I started doing some research. And one of the things that I most recently found uh was a TEDx talk that you did uh eight years ago. And in that talk, it was about um can we trust the machines? And it was AI, and you were way ahead of uh the conversation as it as it uh stands today in and current commerce, where the general conversation tends to be artificial narrow intelligence, generative AI, and those sorts of things. So um, if you don't mind, maybe recap some of the thesis uh that you shared in that talk, and I'd like to explore how how those have aged eight years on in the AI game.
SPEAKER_02Sure. So the title of that talk, if I remember correctly, was you know, when should we trust machines, right? So when should we trust AI? And let me just give you a little bit of the genesis for that question, right? Because I've been in AI, you know, since 1979. You know, I got into the field as a PhD student, you know, I did my thesis in an area called planning. And then I brought machine learning to Wall Street in the you know early to mid-90s, and I created a hedge fund, like the first machine learning-based hedge fund that continues to operate um, you know, trades every day. One of the questions that intrigued me, because I'd I'd been doing you know, prediction in financial markets, but I'd also done stuff in the healthcare arena, various other business applications. And the question that really intrigued me was why was I willing to trust my algorithm in finance that was wrong almost half the time? And yet I wouldn't trust it, let's say, in healthcare, where you know it is, you know, could be right much more often. Or uh for a driverless car where I'm really hesitant to take my hands off the wheel. You know, I don't I still don't trust driverless cars on highways, right? I mean, I would gladly take a Waymo in San Francisco, but I'm not gonna take my hands off the wheel on the highway at 70 miles an hour. So that was the question. Like, why? And the answer now in retrospect seems mind-numbingly simple, which is that it depends on two things. It depends on how often the algorithm will be wrong and the consequences of its error, right? So the way I saw it was that if a machine is never wrong, then you should absolutely trust it. You know, it never gives a wrong answer, you should trust it, right? So on this, so I created this sort of spectrum of predictability from zero to one, right? Zero being completely random or some benchmark and one being it's never wrong. And so that was one axis. And the way I saw it, you could position problems on that axis. So predicting financial markets is about as close to random as you get. It's a very difficult problem to predict, you know, you're not going to do better than you know 52, 53% accuracy, you know, on a daily basis, like not happening. It's just like too noisy a problem. Whereas, you know, in drive-less cars, you're you know, raw, you're right like 99.999997% of the time, right? Extremely accurate. But the problem is that when that driver's car makes a mistake, the consequences are catastrophic or can be, right? It's death. Whereas if my trading algorithm makes a mistake, I lose a lot of money. It's not the end of the world. If I've got lots of positions on, so essentially my cost of error is low in finance and trading, whereas cost my cost of error in driver's cars is extremely high, right? So these two axes essentially create a map of trust, right? On the vertical on the horizontal axis, there's predictability, on the vertical axis, there's cost of error. And so on the bottom right, you should always trust the machine, and on the top left, you should never trust the machine, you know, because the cost of error is too high and it's going to make lots of mistakes. You know, so I label these like trust the machine zone versus don't trust the machine zone. And they're separated by this automation frontier, which um is something that AI enables us to cross, right? So sports refereeing has crossed the automation frontier, right? Machines make calls about whether, you know, in tennis, whether a you know ball is out or in, you know, in baseball that's increasingly the case, and cricket, and you know, like increasingly the referees are becoming automated, right? Urban taxis have crossed that automation frontier into the trust zone, right? We begin to trust them, right? High frequency trading, same thing, right? It it you know, you trust the algorithm, you know, because the cost of error is low. So that's sort of the map of trust that I presented eight years ago. Right. And your question about how has it aged? Well, it's relevant more than ever today, right? So in 2018, when I first presented it, you know, ChatGPT was three years from coming out, right? I was talking about sort of traditional AI at that time, right? Uh, but then now the problem has become even more important, right? We're using it for all kinds of applications. You know, use an LLM to check a legal contract, for example, right? But the question becomes, can you trust it? You know, if you, you know, let's say you have a legal contract, you give it to the machine and say, check this for me, and the machine says it's okay. Would you trust it? You know, so that's the kind of question that that trust heat map is implying. And the you know, the answer is if it's a simple legal contract, chances are I would trust it. You know, it's a few pages, it's seen lots of them, what could go wrong, right? But it's a complex one, then I'm not gonna just trust the machine that says, hey, it's good. I'm gonna have to double check it and verify it, in which case the question is, what good did the AI do you in the first place? Right? If you're gonna check it anyway, did it save you any effort at all? Right? So these are the questions I see as being like very front and center now to the future of work, you know. You know, now that AI is becoming pervasive, you know, and in my book I talk about sort of this general intelligence paradigm where the machine has learned something about everything and that something is getting deeper and the everything is getting wider.
SPEAKER_00You know, as I'm talking to business leaders, the the kind of the three things that are preventing them from uh feeling comfortable about going all in on their efforts to just remain viable in the space with introduction of generative AI at all levels of the business is three things. One, I don't understand the risk, and therefore I can't really do much with AI. Because uh saying that I like is a confused mind always says no. It's safer for the business owners at this point to simply say no because they don't truly understand how can my people screw this up, how can I screw this up, right? The second thing that I'm hearing a lot is we'd love to use AI more, but we don't know where to start. So kind of that use case identification. If if it's something like we just talked about, if the risk is very high and I'm not aware of how to judge that risk as a business owner, big mistake. And then the third thing is we'd love to do more with AI, but we don't have anybody to help us. I don't trust a YouTube video or a TikTok reel. Um considering this this evaluation of uh this this risk uh line, this this how would you uh counsel a business leader who's maybe those are the three main things that are keeping them from really embracing this in their business and getting the benefit of what we know is possible.
SPEAKER_02Yeah, I mean, you know, you hit the nail uh on the head, Chris. These are the three most important problems, right? So if you start with the last of these, which is talent, like do I have the talent to implement it? This is a huge barrier to entry, right? Because unless you're working with a team or people who have a demonstrated track record of success, right? That they've actually done something successful, chances are you should be skeptical.
SPEAKER_00Yeah, right?
SPEAKER_02That is, you know, you can't just hire talent off the street, you know, and say, all right, guys, you know, you guys know a lot about you guys are PhDs in AI, you know, go do it, right? That's a risky proposition, right? So you need people who have like some sort of a demonstrated record of successful projects in the space, right? These and and these days it's particularly uh important because there are lots of wannabes, you know, lots of people sort of saying, yeah, I can do it, right? You should be skeptical, right? So talent, number one, very important. The two other things you talked about was like where to start and the risk, right? Yeah. So the where to start is you know, you're probably familiar with this MIT study, you know, that uh, you know, showed that 95% of the projects fail to provide any kind of value. And that's a great example of a number of things, of a you know, of uh because to me that that's not the problem with AI, it's a problem with the choice of problem, right? You're just choosing the wrong problem, or you're letting someone choose the problem for you, you know, equally bad. And to me, that's a leadership issue, right? As a business leader, you have to be capable of asking the right questions and identifying those projects that are associated with those questions, you know. So if I'm the CEO of an insurance company, right, the question I need to ask myself is for example, is how can AI make my underwriting process better? That's a fundamental question. There's no doubt in my mind that that's a question. There could be others, but that's one question that a leader should be asking, right? Now, I'm as it turns out, I've been through this exercise with an insurance company where the head of the business says, This is what we want to do. And I said, What does that mean? And the answer was, I don't know. I don't know what it means. I I need you to help me figure out what this really means. And it took us three to six months to actually figure it out. But we did figure it out, you know, in this particular case. What it boiled down to was, you know, our our so we get prospects, you know, as an insurance company, like you know, we get prospects from brokers, and so the question, one of the questions was are we getting better prospects from some channels than others?
SPEAKER_00Fair question.
SPEAKER_02Right? Yeah, and and and that question had not been asked. That question had not been asked, you know, big miss. Right now, the data for that did exist within the company, right? Because they've had lots of leads from lots of brokers, right? But no one had answered this particular question, which was is there a difference in the levels of risk that we're getting from different brokers? Are there certain brokers that exert tremendous market power and force us to take on contracts that we shouldn't be taking on? Right? And so that was the question that we discovered as being direct, directly related to how do I improve my underwriting process? Well, one of the reasons, one of the ways to do it is to sort risks and take on better risk. And now there are other issues like is my process, is my internal process good, or can I improve the internal process, right? But to me, this is a leadership question as to where to start. And to me, the MIT study demonstrates or or suggests that people just don't know where to start. Maybe there was FOMO, we need to get started, let's try something, right? And that's not usually a good idea, right? So that's an extremely important problem, is you know, a question, where to start? Which problems do I start with? And that's fundamentally a leadership issue, starts with the CEO, right? And that's it. The third the first point you raised about risk, that plays directly into my trust framework, right? Because if you're going to take on problems where there's a high degree of predictability, right, chances are you'll succeed, right? Conceptually, your your odds of success are higher. So this sort of ties into that second question, which is where do I start? And the the where you start is a function of those projects where the likelihood of success is high, right? So you have to estimate that, and that varies within each organization. So you have to like say, okay, there are these 50 possible projects we could do. Can we rank them by probability of success? And you should be able to rank them, right? Now, this isn't an easy exercise because it might involve asking yourself, do we even have the right data to answer this question? Or might we have to acquire data? And there's so there's a lot of sort of data uncertainty involved, right? Uh so the more the data uncertainty involved, the higher the likelihood of failure, right? The lower the data uncertainty involved, the higher the chance that you'll succeed. So the example I gave you earlier about, you know, am I getting better risks from some channels than others? You have the data for that, right? That question is actually answerable from the data. So so so so that's the way I would look at these three questions of risk, of where do I start, and of talent, right? Because the talent part is sort of their technical uncertainty, right? Do I have the chops to solve this problem, even if I have the right data and I have everything else, do I have the right people? Do I have the right talent? Right. So there's this sort of data uncertainty, talent uncertainty, you know, and and sort of what's the likelihood of success. And then the last piece of this is estimating what the what the payoff is, like what's the size of the opportunity, right? So if you so if you multiply the probability of success by the size of the opportunity, that gives you some sort of a you know an expectation, um, a number that you can use to rank this potential project relative to others.
SPEAKER_00So for any of those, those of you who are listening who are a little savvier with using the models, it we didn't give you the A plus B equals C formula here. What I would suggest is you re-listen to this or take this transcript, plug that into the model, ask it to evaluate the perspectives that were just shared through the lens of your business or your situation, because I think what was just outlined was it's a two-sided coin. It's not easy, but there is a way for you to um mitigate the risk of the introduction of generative AI in your business significantly by looking at it through the lens of the perspectives that were just shared by Professor Darr. So that was um that's fantastic. I was hoping that there might be a silver bullet, but doesn't sound like there is at this point. And for those of you that were, um, if you're not familiar with the MIT research that was discussed, I think it came out maybe um in the fall, MIT and their study that that got a lot of uh headlines was that 95% of pilot projects were failing. So um I can tell you that the response that you gave just then, uh Vassant was was spot on. It had a lot to do with probably the the pilots that were selected and the lack of talent that was available to execute those pilots within those businesses. So um, again, dear listener, uh don't fear we're not seeing those results when we're working with clients. So it all is not lost, you just need to do it the right way. So, Professor Dorr, what are some of the trends maybe that you're seeing? Because you're not just working with uh obviously ambitious students at Stern School and the Data Science Center, um, but you're also obviously communicating with uh, I guess, business leaders for sure, uh, considering your location. Um, what are some of the trends that you're seeing today when it comes to, I guess, green shoots of opportunity for uh businesses of of all size?
SPEAKER_02You know, the the the trends I guess I'm seeing today are that uh leaders are beginning to really sort of engage with this problem, right? That that I'm that what I'm seeing is like a serious engagement with this problem because I think people do recognize that this is a pivotal technology, it's here to stay, you know, it ain't going anywhere, it's here to stay, it is likely to be transformative, and you know the the question is how is it you know how how is it likely to transform my business?
SPEAKER_00And I've got a question is the sentiment that you're seeing is it fear or greed?
SPEAKER_02I would say it's mostly fear. Okay. If you were to ask me which of those two dominates, I think at this point what is dominating is fear, uh there's FOMO, but there's also like a fear of like genuinely like missing the boat. Um because you know, like I said, this is transformative, and it really involves in many cases just rethinking the way things are done, right? And that's I mean, that's easier said than done, you know. That you know you know how do you rethink something that's that that you know that that you're doing? Because organizations get into modes of sort of operating procedures, right? We uh for for for good reason, you know, you need these operating procedures, but then those really get sort of ossified into the way you do things. It's very difficult to like dig your head out of the sand and say, could we be doing that thing completely differently? Why do we need 50 people to do that, right? I mean, maybe we need five, you know, or maybe the 50 people should be doing so much more than they're capable of doing, you know. I was talking to the chief AI officer of Morgan Stanley recently, and he said, you know, what this has enabled us to do is take on projects that we couldn't take on earlier, right? So if we had like 200 potential projects to do, you know, in the ideal situation earlier on, we could only do 50 of them, you know, for budget reasons or whatever. That's that's all we could afford. Whereas now we can go much further down that list of things, right? So our our ambitions have have increased, right? So with the same number of people, we can do a lot more, right? So that's the kind of thinking I'm seeing in organizations that are sort of really kind of pushing the frontier. And I'd say Morgan Stanley is probably one of them, right? That they are sort of pushing the frontier and saying, like, how can we do more with less? Right? Because software is the killer app of generative AI, you know. And you know, so how can we leverage this? Now, by the way, I spoke at an event last night which was about you know using generative AI for You know, creating software or enabling you know programmer productivity. And the interesting thing there was that um again the the the the issue of trust comes up, right? So on my trust heat map, I said, where does software lie on this trust heat map, right? Would you blindly trust a machine that generates the software? Um, and the answer is, well, maybe sometimes I would, but sometimes I wouldn't. Like if it's mission critical software, I want a human looking at it, right? In which case, how much of a productivity gain am I getting? Probably not orders of magnitude, right? Um, maybe the machine can generate stuff, but I still need humans to look at it and verify it and run unit tests and run regression tests and all that kind of stuff. Like that that ain't going away. On the other hand, if it's a low-risk kind of a application where it's internal and it doesn't really matter if the machine makes mistakes, I still get a lot of productivity enhancement, then I can probably use it like right out of the box, generate the software, run stuff. Yeah, it makes errors occasionally, but I can tolerate them. Like the cost of error is low. That's the kind of stuff I'm seeing across the board. But the issue of trust is like front-end central to all of these applications. You know, like when should I trust the machine? Um, what sort of human involvement do I need to verify the outputs? Like, how easily verifiable are the outputs? You know, the more easily verifiable they are, the easier it becomes, you know, for the human to trust it. Um so you know, to sort of summarize where I'm coming from, right? You know, you asked me like what are the green shoots? What are you seeing? Yeah, I'm seeing fear, I'm seeing real movement, I'm seeing you know, higher ambitions, like we can do more than we used to be able to do because this thing is a real productivity enhancer, right? So we can do more. Um, you know, and then you know, other questions around can we can we trust this thing, right? Or do we need humans in the loop to verify its outputs? So, broadly speaking, those are the kinds of trends I'm seeing out there.
SPEAKER_00Yeah, interesting, because you're talking to a different audience than I typically do. I'm I'm usually talking to lower middle market sized businesses, and you've got the opportunity to speak to Wall Street and you know, known brands for sure. And I was looking for maybe what are the parallels that are being that are concerns at one end of the spectrum versus the other. And it sounds like there's a lot of similarity. Now you mentioned something, and I want to just kind of take it, not a quick tangent, but you mentioned that they have some concern about missing the boat. I'd like to see if you have any prognostications on the timeline that exists before um there really will be a an obvious have and have not in the business world with those who adopted, went through the pain of growing, testing, shifting culture, all that stuff, and those who said, Oh, we'll just wait and see what happened. What do you think that how long do those guys have?
SPEAKER_02You know, that's a great question, you know. And I, you know, one could sort of come down on either side of this, right? Which is that, you know, look before you leap versus he who hesitates is lost. You know, that there's you know, there's there's uh there's a truth to both of those things, right? And I would say that in this case, the look before you leap is probably more important than he who hesitates is lost. You know, that it's important to engage in sort of what Kahneman calls slow thinking, you know, that just like you know, take your time, figure it out properly, right? Figure out, you know, what the most important sort of questions are that you should be asking, you know, figure out you know which are the projects you should be taking on and why, and be really clear about those. You know, be clear about the uncertainties involved and how you're gonna navigate them. To me, those are the important questions. I like it. Um I I I don't think you're gonna get left behind if you wait a week or a month or whatever. In fact, the you know, the the argument I've seen for delay is that hey, you know, the speed at which things are moving is very relevant, you know, that six months from now, you're gonna have capabilities in AI, in these tools, that you don't even have now, right? We've seen this even in the progression of these chatbots, they becoming, let's say, multimodal, right? Initially it was just text, you know, then they became multimodal. They could idle images and numbers and spreadsheets and all that kind of stuff, right? Now, if you had started early, if you started way too early, you'd be like reinventing the wheel, you know, to a large degree, right? You you'd be trying to do this. And I and I did do that by the way, right? I mean, I did some projects before LLMs came out where it was harder, you could still succeed, but you had to like build a lot of the machinery yourself. In in retrospect, that was like you know, effort, you know, down the drain, right? If you had just waited, you know, we could have done so much better, right? Yeah so so so so there are you know there's a case to be made for for that for multiple reasons, which is look before you leap, right? Because um the cost of error of like choosing the wrong projects is way too high. Yeah, um, you know, so take your time, be clear about what you want to do and why. And in a sense, time is on your side, right? It's your friend because the technology is improving, and you're sort of, you know, you you'll get more for the you know for your buck the later you start.
SPEAKER_00Yeah. You know, in our pre-interview, you you mentioned this concept, and my perspective was hey, I was an early adopter of generative AI, pretty much as soon as it came out, we were doing, we were trying to break it. Not GPT-1, too, but 3.5 and beyond. So since November of 2022. And I thought that I had an advantage in the space because I had been playing with it, working with it, using it uh since then, as compared to somebody who was just getting started today. But I I have to agree with you, and I want the listeners to know there's hope for you if you haven't gotten started. As the technology improves, it gets easier to use. Now, there's a couple things that you said there that I think are important. You didn't say, oh, just wait till it's easy. You said continue to think about it, but think slowly about it. So for those of you that are listening, if it's like, yeah, we we need to do something, we need to do something, and you've been saying that for several quarters now, you need to do something. Even if it's just starting to really take a serious, uh cogent, like what are we, what does this mean for our business? But at least start thinking, even if it's thinking slowly. I think that's great advice for our audience. And secondly, for those of you that are are are feeling that FOMO, um, the the perspective that was just shared was um those who hesitate is lost is not the approach. So uh, you know, it's maybe not what you're hearing from me, because I'm a go, go, go kind of guy, but I think that that advice that we just got is uh uh should certainly be considered. And if you're doing it, you're in motion, you're at least off the fence. So exactly.
SPEAKER_02And and and by the way, it's it's essential to get started because you know, you may not so one of the realities of organizations is that leaders are very far removed from the details of their data, right? They don't have a clue, right? And um very often um they're not going to realize that you can't get there from here because you don't have the right data, or you're you know, and and I I and I, by the way, I went through that with this insurance company. That is, when we actually started looking at their data, right? And you know, if you've dealt with data at all, you know that it's a messy endeavor, you know, that there are fields missing, you know, some of the values are wrong. You know, it this is just part and parcel of dealing with data. And leaders need to get on board with really understanding what they have and the quality of it, and they don't do it unless they actually start asking questions, right? It's only the question asking that leads to some kind of learning where people say, Oh, like, you know, our customer database is pretty solid or it's pretty messed up, um, or we need to acquire external data to augment what we have, right? These are steps that are essential, there's no getting around them. So, in that sense, you need to get started. Yeah, yeah. Because you won't discover those things until you get started. Until yeah, until you get started and until you start asking the right questions.
SPEAKER_00And even if you're not planning on jumping right in, you're at least preparing the playing field for when that time is right. You're not starting from scratch, you've already set the conditions of the company beginning its engagement with generative AI. And again, on this podcast, we don't really talk a lot about data science and things like that. I would consider that like a level two discussion for most of the people listening to this. They just want to get started. Um, but that's that's an interesting question. How does how does somebody who's been doing this since 1979, when AI was a different definition of what most people think of when they talk about AI today, what do you think when you see like the new person that's like, oh, I'm I'm an AI expert because I've been using Chat GPT for is it two different paths of AI? And and one is not gonna say independent from the other, but is it necessary that a business leader, because I know that there's remote learning classes, even from perhaps Stern, but I know that MIT has one, Chicago's Booth School has one, Miami's Howard School. They've got these chief AI officer certifications. And when I talk to individuals who have done those, they were attracted by the pedigree, um, but they leave the class with an understanding of things that aren't necessarily practical for, let's call it, transformation. They understand the mechanics of machine learning, they understand what a neural network is and those sorts of things, but they don't learn how to do what we call, at least in our organization, think in AI. I'm encountering an issue with the business, I don't know what to do, my reflex has now become let me go to the models to ideate a solution or to get closer to a decision.
SPEAKER_02So I'm gonna say something, and I apologize if it sounds self-serving because I don't I don't mean it to be as such. But one of the reasons I wrote this book and the and um and and the way I wrote it was to be accessible to everyone, uh, you know, to students, teachers, parents, grandma, policymakers, my colleagues. Um you know, because you know, there's uh, you know, I mean, I'm a I'm a big fan of Bob Marley, and one of my favorite sort of lines of his is, you know, if you know your history, then you'll know where I'm coming from. And I think it's important for people to understand the history of AI, right? That this isn't something that has emerged in the last three years. It's been around since 1956. I mean, I only got into it in 1979, you know, 23 years into its existence. But we had, you know, our ambitions at that time were really high, you know. Um, one of the things we were trying to do was uh get a machine to write programs, automatic programming, it was called. Sound familiar, right? Um and that vision has been realized, right? So the vision of AI right from the 70s was about telling the machine what to do without telling it how to do it. Okay, that's been the vision of AI for a very long time. Just declare what you want and the machine will figure it out. You know, uh agents have been around since the 70s, but the reason we're so excited about it now is because you just tell an agent what to do and it figures out how to do it, right? It writes the program, it you know, it knows how to communicate with other applications. You don't have to like do the messy job of like writing code and verifying it, all that kind of stuff. It does that for you. So the ambitions were there way back then, but why couldn't we achieve them? Like what were the barriers, right? And so what I one thing I do is I walk people through the sort of begin with the history of AI, the scientific history of AI, right? From the paradigm of expert system of specification to machine learning, to deep learning, to you know what I call general intelligence. And so it's important to understand this evolution, it's important to understand what the goals of the field have been, what the barriers were, what was achievable, what wasn't, and what's led us here, you know. So it's important to understand the history because without that, you're functioning at a very superficial level, right? You're just saying, oh yeah, these deep nets, you know, they're amazing, right? But you don't really get beyond that, right? You don't really understand why they're so good, right? And one of the points I make is that the fundamental difference between modern AI and AI that existed for the 67 years preceding that was that this boundary between expertise and common sense has broken down, right? In my 40 plus years in AI, we artificially created a boundary around expertise and said, this is medical knowledge. Guess what? Medical experts don't just use medical knowledge, they use common sense all the time. Yeah, you walk into your doctor's office, he or she is using a lot of common sense in addition to that domain knowledge, right? These things like blend seamlessly into each other. And that is the nature of human activity and human thinking, is that we don't even think about it, we're unaware of it. And this is what modern AI has done. It just sort of broke down this barrier. And in you know, one of my podcast episodes on my podcast Brave New World, I asked my guest Sam Bowman, I said, What happened? Like, you know, you know, what was the difference between pre-uh uh Chat GPT and post? And he said serendipity. Essentially, what he said was uh Google wanted to do sentence completion in Gmail, and that required predicting the next word in a sentence, and the word after that, and the word after that. And that was a problem at just the right level of difficulty for which there was lots of data available to solve that problem, right? And it was basically like a prediction problem. Just predict what's going to come next. Predict what's going to come next. Now, some people say, well, these machines aren't really intelligent because all they're doing is like predicting the next thing and sort of generating a sequence. But I think that's a little uncharitable because in order to speak flawlessly, that is, you know, always generate stuff that makes sense, you have to know a lot about the world, right? So the serendipity was that in order to do sentence completion, we had to force the machine to learn about the world in general based on the collective expression of humanity on the internet, right? And that's been the big deal, right? That's been the sort of leap forward in modern AI, is this sort of dissolution of the boundary between common sense and expertise. Because the LLM has no idea whether it's talking to you about a deep medical case, or whether it's talking to you about something frivolous, or whether it's talking to you about shows on Broadway, it doesn't know, it doesn't care, you know. Um, and so people really need to understand AI, you know, not in terms of the technical details, but just sort of functionally, right? Why have we gotten to where we are? And if you understand that, you know, and you understand what's similar to the kinds of things we were doing in the 80s and the 70s, you know, like, and what's different, then you have a much better perspective on how to think about AI and how to think about its applicability to business.
SPEAKER_00This is interesting because we have a chief AI officer at our company, a chief AI officer. His name is Adam Lyons. And Adam, when he's talking to people, he tells them, well, we don't have artificial intelligence yet. And and one of his perspectives is that, you know, XAI, they don't call their product Grok AI, their product is Grok. OpenAI is the company, but the product is Chat GPT. Clawed AI or anthropic AI is the company, but the product is called Claude. So to him, it's it's an indication that these companies that we haven't achieved true artificial intelligence yet, because if it was, they would include that in the product name. But being as litigious as the US is, that false advertising could be an issue if they were calling it Grok AI or Chat GPT AI. What are your how do you take that?
SPEAKER_02I'm not sure I fully agree with that. Um I I agree with his premise that um well or part of his premise that AI isn't that we have not reached the promised land yet, whatever that is, right? Yeah. That you know, now I I'll just sort of backtrack a little. If you think about AI, it's the only technology that has been designed without any purpose in mind. Right? Every other machine has a purpose.
SPEAKER_00Sure, okay.
SPEAKER_02You want to make a car, you want to drill a hole, you like it has a purpose. AI has no purpose, right? It has no purpose other than to be intelligent. And in my history of AI in the field, you know, over the last 45 years, the one thing I can tell you is that the goal post keeps moving. Right? Anytime the machine does something, people say, Oh yeah, but that's it's you know, it can't do this yet, right? So, you know, when I got into the field in the 70s, there was a philosopher called Stuart Dreyfus who said, Look, machines will never really be intelligent because they're not going to be able to do the even basic things that humans do so easily, like drive a car. Yeah. And at that time, it was inconceivable to think that a computer could like drive a car. You know, it just wasn't in people's conception, right? Now that computers can drive cars, we say, Oh, yeah, well, that's uh you know, that's okay, yeah, that's good. Yeah, you know, but you know, it it it doesn't it doesn't do this, right? Um, so we keep sort of shifting the goalpost. And you know, I would say that in this paradigm of what I call general intelligence, the machine is incredibly intelligent about some things and not so on others, right? So the intelligence is very uneven. Um but that's okay, you know. I mean, that's what you we should expect of machines because they are not like us, right? Yeah, you know, as Jeff Hinton says, you know, it's like an alien has descended on the planet, but we're having a hard time taking it in because they speak such good English, right? Um so you know, but but that's that but that's what we have. We've got this alien species cohabiting the planet with us, you know, of our own creation, but the species thinks very differently from us. Why should we expect it to be like us in every way, right? So I'm a little bit sort of, you know, when I think about like AGI, right? This sort of obsession with AGI, it baffles me because I don't see why machines should be like us in every way. In in some ways, they're gonna be so much better, in other ways, they're not gonna be as good as us, right? They they're not gonna feel emotion. I'm quite sure about that, you know, uh unless they sort of become maybe biological. They'll be able to simulate it, but they're not gonna like really feel emotion the way we do. So they're gonna be different from us. But are they intelligent? I'd say, yeah. I you know, I I I think we have achieved artificial intelligence to a large degree, um, and especially now because of this sort of dissolution of the boundary between expertise and common sense. Like to me, this is like a you know an indication that this sort of you know that AI has gone from an application to a general purpose technology, right? That I view AI now as much more akin to electricity than I do to an application, because general purpose technologies have like widespread impact throughout every industry in the economy, and that's what we're seeing with AI, right? So this is, you know, to me, it's exceeded its expectations as far as I'm concerned, right? So through most of my career, AI overpromised and underdelivered, and now it's doing exactly the opposite, right? That it seems to be like surprising us. Um, I never imagined that I would see a machine in my lifetime that you could talk to in English, right? If 10 years ago you told me this, I'd tell you you were smoking something. You know, yeah. Like, no way, right? That this was like beyond my comprehension. So, in that sense, it has over-delivered, you know, in the last few years. And we're just sort of caught by surprise at how fast it's you know going, you know, which is why I sort of end my book with like, you know, are we going to govern AI or will it govern us? Right? And I start with this sort of the 2001 Space Odyssey example of Hal, you know, the famous line, open the partway doors, Hal. And you know, I sort of conjecture whether Hal was actually doing an experiment on the humans, whether it created an edge case to test them, and they failed the test, so it killed them. So um I I know I've gone a little far afield of your original question. No, no, no. I it was always awful. But but but I feel that in some ways AI is actually over-delivering now, you know, relative to its history.
SPEAKER_00Yeah, I like it. So you mentioned your book a few times, and I want to make sure that we do talk about some of the key concepts in there. And the book is called Thinking with Machines, The Brave New World of AI. And one of the key concepts that we had kind of talked about in our pre-interview was this a concept of the bifurcation of humanity, with AI splitting us, the listeners, into two groups: those with deep domain knowledge who use AI to uh accelerate or augment or amplify, and then those who are the opposite, I guess, the other side of that coin. So I'd like to kind of dig in on that a little bit because I think that that's important for individuals who are discovering or they're on the AI journey, where you see them falling into, like which group over the next coming months and years.
SPEAKER_02Yeah, that's a that's a really big uh point in my book. You know, someone calls It a BFI. And you know, the thinking here is, and so I'll give you an uh an analogy, right? I'm I talk to a lot of parents who express concern. Yeah, yeah. You know, you know, how should I be thinking about it? And my question and my um answer to them is that the question you should be asking is, is AI making my kids smarter or dumber? And I do believe that question is answerable, right? It probably depends on each family. Like, when do you talk to your kids? Is it around the dinner table or when you drive them to school or walk into school or whatever, right? So when do you really engage with your kids? Because it's important to engage, right? And the thing you want to be figuring out is how you know can they hold their own in a conversation without the AI? Because if you've got the AI, you can say, hey, answer this question for me, but then you're really using it as a crutch, right? You're using it as a question-answering tool and it's doing stuff for you, as opposed to sort of making you think in new ways and making you learn things that you never even dreamed about, right? And so I talk about this bifurcation of humanity where I say it's it's gonna be sort of the rich get richer kind of phenomenon, right? The more you know, the more AI can teach you, you know, and to be honest, I find myself learning all kinds of new things with AI, right? That because I can ask it the right question, right? I know I know the question to ask, right? So it's important to be able to ask the right question, and then it's also important to be able to gauge its responses and say, well, is that really true? You know, to push back against it or ask it to verify something, right? So to have the chops to know whether what it's giving you makes sense, and then to sort of nudge it in promising directions, right? Now, someone who's an expert who's really good at what they do is able to do this, you know. A biologist might be able to learn something about the physics of you know things and and vice versa.
SPEAKER_00Yes.
SPEAKER_02Uh, because they already know a lot about you know their domain, right? They can now amplify the knowledge in their domain, they can go to adjacent domains, they they know the right, you know, what questions to ask, they know how to evaluate the outputs. Whereas people who just say, give me the answer, are basically becoming disempowered and entering into a potential cognitive decline, you know. And so the question I get from people sometimes is like, what do you think the relative proportion is going to be, you know, of people who get amplified versus disempowered? And that's a great question that I try not to answer. Um because you know, as Yogi Berra said, it's very difficult to make predictions, especially about the future. Um but but what I do uh sort of suggest is guidelines for staying on the right side of this divide, right? Because you don't want to be on the wrong side of this, you know, and and and get disempowered. So how do you really stay on the right side of this divide, right? And this applies to everyone, not just kids. I mean, I used kids as an example because that's sort of the obvious sort of use case, but it applies to all of us. You know, and I suspect that. Sorry, go ahead.
SPEAKER_00I was just gonna say that's another concept in the book about the you you don't have a choice here. You have to make you have you have to opt in. The future is not AI versus humans, but humans thinking with machines.
SPEAKER_02So like I think there's no opting out, yeah.
SPEAKER_00No, no, I think that that's so everybody listening, that is a critical point. This isn't something that you get to say, oh, well, others do it and I don't, because if you are on the other side of that coin of the others do it, as you said, the opportunities for individuals who opt out or or consider that that they're not going to participate in this, there's not a lot of options for them.
SPEAKER_02And there aren't, and and and you know, and like all other technologies, this really ups the bar for humans. You know, it it ups the expectations of everyone. You know, one of the things I talk about in this book, in my book, is this uh Damotharan bot that I've built to value companies in the uh same way that my colleague Aswat Damotharan values companies considered a valuation expert on Wall Street. You know, now that I've built the system, you know, we're testing it and and debugging it and everything, I'm realizing that uh I'm thinking about its use case and how it's going to impact work. And the way I see it is it'll amplify people who are really good at doing valuation, right? Because suddenly now they have the Moduran on their fingertips, right? They have this machine that has a tremendous degree of knowledge about how this valuation guru does valuation, and now you have that available to you, right? It can generate reports for the SP 500, all companies at the press of a button, right? It wasn't possible to do this earlier on, but what this does, it sort of ups the expectations that people will have from you, right? You're gonna be required to be doing so much more because you've now got this sort of really powerful amplifier, right? The expectation isn't gonna be that you produce you know one report or you evaluate one company every two weeks or a month, you can now evaluate so many more companies, right? You're able to do stuff that was just not feasible before that.
SPEAKER_00Kind of like that Morgan Stanley example, where now, because of the amplification from the tools, they're able to address a lot more of that wish list than they were able to prior.
SPEAKER_02Exactly. Yeah.
SPEAKER_00So for the business owners listening to this, uh the encouraging part is that nobody knows your business as well as you do. And as we just talked about this idea of the, you know, the experts will get smarter because they know that they know the domain well enough to ask those nuanced or very high-leverage questions. That's you, dear listener. You are that person as soon as you start using the tools and and and thinking in AI. And and and if you're not doing that now, what I would suggest is you you dedicate some time on a daily basis, 10 minutes minimum. And I think by the third exercise, you're gonna have a breakthrough and you're gonna go, oh my gosh, like this needs to be on my phone. This needs to be what I what I talk to when I'm walking the dog. I think you're really gonna see the value of this if you're not already there. You mentioned your podcast. I would love for because the the conversation has been fascinating. And again, I'm in a bubble, you're in a different bubble, and I I enjoy the the being able to see, well, what's going on in that bubble? So this has been uh great for me personally. But for your podcast, um, what's the name of that show?
SPEAKER_02So my podcast is called Brave New World. Um, I'm a big Aldous Huxley fan. Yeah. Um, and to some extent, I feel that the world is becoming Huxleyan, you know. Uh so Aldous Huxley, uh, you know, essentially, you know, for those of you that haven't read the book, I would highly recommend it, right? But it's this sort of, you know, he paints a dystopian future in a technology in a society with really advanced technology. So it's called Brand New World, and it's really a podcast about um the future uh that uh the world that our future selves would like to inhabit, right? So it's a really broad canvas. I've had some amazing guests, you know, some Nobel laureates, Turing laureates, you know, uh, you know, like really amazing guests uh that I've been privileged to have, and I've had some, you know, fantastic conversations with them on topics ranging from you know business impact to consciousness to law, economics, uh, you know, just like you know, the the philosophy, uh AI, you know, just like the whole gamut of topics, you know, related to AI.
SPEAKER_00Very cool. So uh for the audience here, I just want you to know the the kind of the idea behind this episode today was to introduce you to Professor Darr's positions on how he's thinking about AI, ultimately leading to, if you want to do a deeper discovery, getting a hold of a copy of the book that was published by Wiley and Sons in November of 2025. And again, the I didn't give you the full title earlier, but the full title is Thinking with Machines, The Brave New World of AI, which feeds into the podcast as well. Um, outside of that, are where are you kind of sharing your thoughts outside of the podcast and being guests?
SPEAKER_02Yeah, so I have I have uh a newsletter on Substack, also Brave New World, which is vassantnar.substack.com. Uh I have a um a column in psychology today where I've started uh posting on a regular basis. So you can find me in a number of places. Uh the newsletter follows the podcast. So every uh I release a podcast monthly, so I'll be releasing it next week. And uh and the newsletter follows uh the podcast. Uh it describes summarizes the podcast in one sentence in one paragraph, and then you know, talks about you know something else. So the next my my next episode is about fiction machines, which is what AI machines really are, right? They generate fiction. And one one of the questions I uh consider is um, you know, and I was talking about this last night uh to a bunch of executives. I, you know, I said, I'm actually surprised at how often it is truthful and correct, considering that AI machines are not designed to be truthful, you know. Um and this is something that people should be aware of, right? So when people say, oh, machines hallucinate and that's not a good thing, well, actually, they don't hallucinate, they confabulate everything, right? So everything the machine generates is a confabulation. And one of the things I talk about is I'm amazed at how often these confabulations turn out to be truthful and correct, right? That's what amazes me, considering that they've been trained on everything, you know, truth, lies, opinions, everything. And yet somehow, you know, we've through this reinforcement learning human feedback, we've kind of fine-tuned them or channeled them to generate outputs that we find good or acceptable or useful. But there's a lot of stuff like lurking under the surface there, you know, that uh you know that can erupt, right? It's all there. It's it's all there, it's learnt everything about us, including some of our undesirable aspects, such as our ability to lie, to deceive, to manipulate, right? These machines have learnt everything.
SPEAKER_00We'll have to have an episode on that at some point in the future. Fasan, thank you so much for taking your time. And for the audience, all those, the link to the Substack, the link to the podcast, all those will be in the show notes. And I encourage you to plug in. This is uh again, we had the opportunity to get the perspectives of somebody who's been paying attention to AI and its impact in business and in the world for 40 years plus. So um not very often you don't run into too many AI experts these days who've had such a uh uh long track record. So, Fasan, thank you so much for being a guest, and um, I look forward to um uh continuing to read the the um the substack and look forward to any future writings that you're gonna be doing. Thank you so much.
SPEAKER_02Hi, Chris. I really enjoyed uh the show. Great questions, and uh I hope to keep in touch. Good stuff. Awesome. Thanks everybody. We'll see you on the next episode.
SPEAKER_00Thanks for tuning in to Using AI at Work. Don't forget to subscribe for more conversations about how to use AI at work. And a special thank you to our sponsor, Chief AI Officer, for empowering businesses with AI education and training. Visit their website for a free AI readiness assessment and AI strategy guide to help you get started using AI at work. It's www.chiefaiofficer.com. Follow us on Twitter at the handle UsingAI at work and visit www.usingai at work.com for free resources to help you harness AI in your role.