On the academic research side, this stuff can write a good paper at this point. I think people are too eager to assign bright lines to AI, like it can never do creativity or judgment. It does creativity or judgment. So, the AIs are trained to use English because it lets us have some observability into what they're thinking. It's not a lucky thing. It is an actual design decision. If we just had Fable 5.1 right now in House Strom, that's enough to change almost all of human work in the next 10 years. It's good enough.
This is Ethan Mollick. He runs a generative AI lab at Wharton, and he measures what AI does to real work with field experiments, not just theoretical forecasts. He's going to name the four human skills he thinks still hold an edge, and one of them is the thing that we scientists claim to have and can almost never define. We start with mathematics, where the experts' forecast missed by a megaparsec. Getting towards the mathematical angle of things, you know, just in the recent days, a couple days ago as we're recording now in mid-October, there was an announcement of, you know, 700 papers in different levels of theorems and proofs and axioms development released by OpenAI. Three weeks earlier, there was an announcement that it had solved the millennium problem, which is a huge deal. And the pace of change is just astonishing. I have my arguments against it being as useful in physics and the hard physical sciences as it is in math because I do think math is much closer to a language, and what these things are large language models, you know, coupled to GPUs. But I have my doubts about whether or not they can truly do physics in the way that, you know, your colleague at Penn, my friend Mark Devlin, does it or I do it. So, my question is, you know, right now solving Olympiad level math problems and millennium problems is happening, but also simple mistakes are happening. And I wonder, you know, if the kind of jagged frontier that I want you to define is going to be, you know, kind of a permanent feature, the jaggedness, or is it going to become a phase transition where we do trust it, it doesn't think there's 10 Rs in strawberry. And I mean, it doesn't do that anymore typically. What is the underlying problem? Is the uniqueness or the unevenness of things that we see now, is that temporary or is it a permanent feature of how these machine intelligences work?
Really great question. So, the jagged frontiers idea we came up with in our first academic paper where we were studying the effects of AI on people in a consulting business. And what we identified was the idea that AI is good at some things you wouldn't expect, bad at some things you wouldn't expect, and you can't necessarily tell in advance what it's good or bad at in any particular way. And that means your expertise matters, right? Like knowing what this is. Now, math is really interesting because it was the most jagged of the parts of the intelligence, right? When my colleague Phil Tetlock, who runs a famous forecasting institute, predicted the chance that the AI would get a gold medal at the International Math Olympiad right before ChatGPT came out, the odds were something, you know, were under 5% or something like that. And of course, the AI got a gold medal, but for a year it was getting the hours of strawberry wrong. There was a bunch of technical breakthroughs around how AI reasoning works, but also it just turned out the models were better than we thought they were. And then last year, the same forecasting group, including a group of industry experts, were asked to predict last September what the chances that the AI would help solve a millennium problem by September 2026. A year later, I think the odds were 1.7% that that would happen, right? So, like as an indicator of how quickly the jagged frontier changes, right? Leaving aside all the other effects on math, that's a really good indicator. I think AI remains jagged, and part of what you got at with your question was one of the reasons why. So, AI works better when we have definite things to train on, right? So, math, coding, where there's a right answer, works great. And by the way, that's why writing lags so much because good writers like Claude write the same kind of way because there isn't an objective measure of writing to judge against.
Well, that's a load-bearing argument that we can... Thank you. Thank you. Yeah, it's really doing the work here. Yes.
It was a year ago we would have said that's a rich tapestry of answers. On the other hand, I'm an economic sociologist by training, I work at a business school, right? I publish quantitative papers, large-scale data sets. I am getting the latest models to be able to go and say find an interesting data set, download it, do all the tests you need, pre-register the result, and I'm getting good papers. I'm not getting... The research taste is the weak spot right now, not the execution. There's no errors, the sites are great, everything is terrific, but it's taste for research, which is kind of an interesting question anyway, is like should we as humans we need to have taste for research because there's only so much work we could do. But if the AI could do infinite research, do we care that it's doing boring but measurably real work? I don't know. We have to reconfigure science around that. So I guess what I'd say is I think that jaggedness will remain, but as capabilities increase, the jaggedness may not matter that much. If you reach superhuman level in something, does it matter whether or not that's weaker than a superhuman level in a different field? At the same time, that jaggedness creates a lot of opportunities for humans to work with AI in important ways.
In the book, there's a lot of different characters and maybe archetypes that get portrayed, and it's different than your previous book in that you're expanding upon things that are familiar, the metaphors of the centaur and the cyborg, but now you introduce this concept of a conjurer, you know, some human delegator to AI for completing agentic work. Talk about those three different, you know, creatures, the centaur, the cyborg, and the conjurer.
So the centaur and cyborg I talked about in the original book, right? And so those have both been used in different ways. We define these in some of my research papers with people as well. So centaur work, think about half person, half horse with a division line. So centaur work was, you know, you do the writing and I do the coding, right, with the AI, or vice versa, and you have a clearly divided line. What most people move to after a little while is what I would call cyborg work, where you blend the work together, right? And this is very common, you know, you ask the AI a question, you go back to your own stuff. Well, yeah, give me a fifth version of this. It becomes kind of blended work. That all made sense in the kind of co-intelligence era when you work with chatbots. The biggest change that's happened is AI can now do fairly arbitrarily long lists of work on its own without human intervention with a good chance of success. And that means you could assign it very complex work for it to do. And I've had it do all kinds of insane things. Right? So I got to write a research paper on the dating myself as a nerd side. I gave Astra the game Zork, which is one of the original text adventure games. I said turn this into a graphical adventure game. It built a full graphical adventure game with like amazing 3D spaces you move through. But it had to also rebuild every aspect of the adventure to make that work, which is clever stuff, right? So as we can assign work to AIs, that becomes what I called conjurer work. Now I did make a mistake here. Cyborg, centaur start with C, but they were soft C's. So it's cyborg and centaur. When I was writing conjurer, I was like, oh this will sound great. But it actually looks good on the page, but it's actually terrible because it's conjurer work.
I want to know, you know, what are the human advantages that you think are most sort of hopeful or perhaps useful for us in this age, and then I want to drill down into scientists in particular. What are the human scientific and mathematician advantages?
So I talk in the book about four things. And I want to make it clear these are not bright lines. Like I think people are too eager to assign bright lines to AI. Like it can never do creativity or judgment. It does creativity and judgment. May not be your creativity and judgment. May not be as good as you, but we shouldn't define bright lines as one thing or another. So what does that mean? That means that we should think about ways that we gain long-term advantage, you know, in working with AI. So I call those things deep knowledge, wide knowledge, taste, and agency. So deep knowledge is expertise. You have deep knowledge obviously in astrophysics, right? And in your fields of interest. Probably in podcasting at this point in a variety of other fields where you can easily judge, right? Whether or not this podcast is going well or not, or whether it's like, oh god this is going to be miserable. And you probably knew it the second I got on, right? Because expertise is all about heuristic and decision-making. Just like you can look at a paper and be like, is this BS or not really quickly, right? So we need expertise because it lets us work with these systems in better ways. And we're talking about agentic work. One of the ways you want to be able to work with agents is just like you would with a team of grad students. You know right away if they're in the right direction or wrong direction. And that requires deep expertise. Wide knowledge also is really increasingly important. So, wide knowledge is that you know many things, right? You've already made references to the Tower of Babel and to philosophy. Like the more things you know, this is a model trained on all of human knowledge, basically. And that is both fun for using these systems, right? You can explore new topics, but also gives you sort of superpowers. Like coders are doing design work, designers are doing coding. If you know a little bit of coding or design, these systems will work much better for you. Then there's the idea of taste. And something we should talk more about, but AI is really good at producing many of a thing. How do you know you should select? How do you pick something that's not just slop? The enemy of slop is taste, right? And then the final thing is agency, which I'm an entrepreneurship professor. So, it's something we struggle with a lot. How do you give people a sense of agency, of control over their own lives, in the way that, you know, in entrepreneurship, but also in AI? How do you not wait for permission? How do you experiment with these things? And I think those four aspects together give you an edge over other people when you're using AI systems.
Eskaton is a term I first heard from Peter Thiel talking about the Antichrist and the Eskaton. And I have to say I didn't fully understand the point that he was making. But let's remind the listeners or tell me, the host, what is an Eskaton? And what is the escalator? And how are these things in contrast in terms of disruption to this possible future of constantly recursively improving superintelligence?
In my original book, I had four scenarios for the future, right? Maybe AI progress stops, maybe it slows down substantially, maybe we stay on some sort of exponential curve, or maybe we reach this sort of AGI, ASI, a machine smarter than a human at every intellectual task, right? It'll define sort of future. I think scenarios one and two are out. I think we are either on a continual exponential that doesn't reach the level of superintelligence that I think a lot of people still in Silicon Valley are obsessed over, or we do, right? And I'll call those the escalator and the Eskaton. So, the Eskaton is literally the last thing in Greek. I use it sort of substitute for the singularity, right? The idea of in the original sense that we can't predict past this point. There is a point where if you imagine AI gets good enough, we have a superintelligence out there, it is an alien to us. It will do what it does and we can't predict what the world looks like in that case, right? And there are people who want to usher in the end of all things. I think the most likely scenario we go through, right? But who knows for sure, is the escalator, a steady increase in abilities of AI over time. The speed is going to matter a lot, right? And we're going to muddle through one way or another hopefully, but there will be a lot of challenges, a lot of wonders and horrors happening along that way. And I think that's more likely than the end of days or at least more planable. Like if the end of days is coming, you know, what are we going to do? You know, all we can do is be nice to the AI and hope that there's a place for us in the world to come. But I think we spend a little less time worrying about how do we adjust well to a situation where this is happening. Calling back to the mathematicians, that's a canary in the coal mine. Other professions are going to face the same kind of problem, right? Maybe we can argue about whether physicists will have the same kind of crisis or not, but I think it's going to come for a lot of us. And it's going to come for professions and we can reconfigure work in different kinds of ways. People have the model of the industrial revolution talked about a lot and like it worked out great in the end. And hopefully this all works out great, too. I think we can guide it so it works out great, but I also think that the industrial revolution also sucked to live through. Like that's what Dickens was about, that's what Marx was about. And I think that we have to think about cushioning the escalator ride so that everyone wins and it's not just a bunch of AI bros winning this time.
And you have stuff in here from 2026, but you know, like most books you probably started writing it a couple years ago. And one of the things you talk about in the book is that you ask your AI to buy you a gift, a surprise to you for something under 100 bucks, you give it a little bit of a guardrail, and it comes back and it uses your Amazon account, I think I remember. Yeah. When it does that, I mean, how did you react to it? I mean, you describe it and to me it's a little creepy, it's kind of nice. How did you react to that?
I am continually slightly creeped out and really excited and anxious and really optimistic all at the same time. A week does not go by where I'm not shocked by something that AI has done that's interesting. I also know that it fails a bunch of times, right? So, like there's all these persistent problems. I start my very first book I start with the idea that you can't use AI without having three sleepless nights of anxiety. And I think that it kind of just continues. I think if I underestimated something it was like, 'Actually, there's been a lot of moments like that.' You know, and when Astra came, when Fable came out, like there's a bunch of moments. Interestingly, right? I think that since I wrote the book, I kind of had jury-rigged AI agents together even through early 2026. And now there's Muse and Dot and a bunch of these other systems that I think people will be much more used to using AI agents to do stuff and they're pretty good at what they do. And there is kind of an interesting question what you said which is also they haven't messed up that much yet. Like for a billion people using these systems I think we were all worried two years ago would there be mass psychosis? Would the AI be causing all of these issues for everybody? And the answer turned out to be kind of no. Like that doesn't mean bad stuff doesn't happen. Like no, it actually has happened.
You're right. The fact that you hear about it means it's noteworthy, which means it doesn't happen that much.
Most people are using these systems and there's lots of people making mistakes and people are being hurt and I have no doubt there's people psychologically being harmed by this. The data is really unclear about how often that happens but there's definitely people who are being driven to, you know, suicide or psychosis by these things I'm sure. But it's also the kind of thing where it's like, 'Huh, we've muddled through to some extent.' And I think we have to think about how do you cushion the muddling through so people don't get hurt. But it's also worth noting like, you know, there is some experimentation here. You're right about agentic hygiene. We should all be much more careful. Now, the past is not the future and again this sort of singularity awaits us. So, I can't predict what happens next. I do think cyber attacks are going to become a very big issue in the very near future. I think we're reaching points where models are much more capable. So, the future may not look like the past there but it hasn't been the disaster yet.
And your first book was Co-Intelligence which I found very helpful. And you know, to be honest with you, you know, and with love and respect, you know, Coexistence sounds like, you know, we're trying to hang on. It was almost a little bit disturbing or depressing before I read it as I judged the book by the cover as everybody does. It's kind of the thumbnail, the CTR rate that we have to monitor. But, I think, you know, Co-Intelligence was sort of important in the variety of ways. Most of which had to do with the fact that, you know, Homo sapien, the word sapien means, you know, one who is wise. That's our name. And what is it wise about? It's wise that it's, you know, according to most tradition that I understand, is that we understand uniquely among all animals or creatures that we're going to die. And that makes me think that, you know, intelligence is almost free. You know, there's so much intelligence, but what we lack is wisdom. We lack the wisdom to discern between what is useful, what is knowledge, what is to be limited. And so, I kind of wish that your next book would be, you know, Co-Wisdom, but it's Coexistence. And as I say, you know, with respect, it kind of is like we have to justify ourselves or like how are these two-state solutions going to coexist? Was that intentional? And what do you make of this knowledge versus wisdom? I mean, it seems to me wisdom is in shorter supply, and we need to sort of maybe lean into that.
I do want a slight note of panic in this. I mean, I think that the subtitle was going to be before it was the next phase of AI, how to work with AIs that are sometimes but not always smarter than you. And I think that that does require a form of us coexisting. By the way, that also includes protecting parts of ourselves from AI use, right? Like it involves choosing not to use these systems when you shouldn't use them so you don't de-skill yourself. It involves, you know, in learning situations, you know, knowing when to use AI or when not. What we want to maintain humanity up. I think that we have to make choices about this. And I think that this is a force in the world to coexist with. So, I didn't mean it so much the AI was forcing us into a world where we had to live in the margins, but more one that we have to negotiate our relationship with this technology now. A chatbot sort of sat in the box, and you'd talk to it, and when you close the box, it went away. But, our agents right now looking at my emails and they're popping in. Like I've got a message already from one of them about like something I probably should be paying attention to. They save me lots of money. They make my life easier, but I also now have an entity that is actively playing a role in my life, right? So, I think that there is that piece. On the wisdom side, I feel like a controversial piece I will say to you is that I still worry about bright lines there. And I don't know if wisdom... There's human wisdom. We could say we value human wisdom, but wisdom writ large, you know, and we only us have a visceral sense of mortality, right? But like the AI simulates wisdom well enough and simulates theory of mind well enough that the question is like, is it wise or not? And this is where, you know, I try to be a pragmatist about this or, you know, materialist, I guess, which is like let's judge it by what it does rather than try and say that it's the same characteristics as a human, right? I don't find it that useful to ask questions about is it really wise? It simulates wise enough that I don't know, you know, for practical purposes I mean I teach at a business school, right? Like this is, you know, there's a practical aspect to this kind of thing of like it does what it does. Let's measure it and see if it does these things. Let's see if it causes harm. And then we should debate it afterwards. So, to have the wisdom side, I think we all need more wisdom. I think the AI gives very wise advice. We can take narrow things we know about like finance or medicine where you absolutely should be asking a frontier model as a second opinion. Not as your first opinion, but as your second opinion, we have enough evidence that it's quite good at those things.
Except for that meme that, you know, shows a guy asking ChatGPT, is this mushroom safe to eat? And then it says, yes. And then the next scene is a gravestone and it said, I'm sorry. I should have considered the answer more carefully.
But absolutely, but also it was worse and more dangerous. That like the thing to note is everybody's worried about long-term safety. Short-term safety has gotten better. The AIs are much less likely to guide you into a psychosis than they used to. I mean, it sounds ridiculous, but like they were releasing models that would like you would underline your, you know, tell you, yes, you're special. You should hide out from humanity. They don't do that as much anymore, right? They used to be bad at health advice. They're better at that. So, like I'm not saying, you know, there's a human wisdom involved too of like, yeah, don't eat the possibly poisonous mushroom off the advice of one, right? But on the other hand, if all you have to eat are two mushrooms, I'd rather ask the AI for advice than not, right? And so I think that's the wisdom we need in some ways.
So, the founder of your august institution was of course Ben Franklin. And he was an experimentalist. He was a physicist of sorts, an amateur physicist among his many things. You know, in physics we don't call a theory successful because it produces things that sound good or, you know, resonate with us. So, look at string theory and other things. They sound good, but we demand a prediction that can survive an experimental conflict with the experiment. There's our inability to confront Mother Nature. So, my question for all AI researchers and specialists, and you're an expert in this, is like what is the equivalent experiment? What is a decisive experiment demonstrating that, you know, electrostatics and lightning and conduction are one unified phenomenon as Franklin did or any other empirically verifiable, but most importantly falsifiable, potentially falsifiable equivalent. What is a prediction? What's some crisp thing, a decisive experiment that you could demonstrate that AI actually understands?
The problem is a definitional one rather than experimental design, right? I mean, I have built experiments. Does AI affect learning? Does AI affect productivity? Does AI affect creativity? You know, and even then there's definitional issues. What is productivity and how do we define creativity and how do we measure it? Again, empirical researcher here, and I don't know how you define what thinking is. Like we can look inside the brain of the AI, right? And we know what it's doing. Is that thinking or not? I couldn't tell you. We get a really good simulation of thinking and we can test that. That's where my pragmatism sort of kicks in. It's like I want people to grapple with this problem. In fact, one of the biggest worries I have about academia is the complex reactions to AI, and I think math was more enthusiastic than most and kind of got burned by it. I mean, not burned. This is where things are going. But, you know, there's a lot of resistance among fellow academics about this technology for a wide variety of reasons, good, bad, real, and not. To me, I really want to hear people who are philosophers of consciousness dig into this. And biologists have got like I am not an expert in those areas. I can tell you about the things I know a lot about and I, you know, I'm broadly interested in many topics, but this is the problem for humanity is to understand what this tool is. How do we work with it in the right way? How do we use it? Where is it useful as a tool for advancing human work? Where should we never use it? You know, what does this teach us about intelligence and language and meanings of the world? I mean, it's the most exciting opening. It didn't do everything else it was doing. The fact that we have a simulated artificial brain should be the most exciting thing that academia has dealt with in a decade or a century. And I wonder at what level is everything reducible to language. But I want to give you what I call the Keating test, you know, very modestly named, but it's actually the Einstein test. And it goes back to 1907 when Einstein said he had his happiest thought. It was a thought he claimed that titillated him. And he was participating in what's called a thought experiment, which is something an AI can do. Which was that if you were in an elevator and the cable broke and you were in free fall, you experience no gravitational field. And that's now known as the Einstein equivalence principle. So I always like to ask, you know, in what sense can an AI, an artificial brain, have that visceral sensation that we all know if we've been on a roller coaster in the back seat of my car as I'm taking my kids to school, you know, going over, you know, zero Gs. What sense can an artificial intelligence, you know, have a sensation viscerally, A, and then B, can it have a happy thought? Those are things that we can have and a child can have it, but can an AI have it? Are these things imprisoned by not having bodies or not having happiness or have those dimensions of their personalities?
There's two questions there. One is about subjective experience of AIs and the other is to the extent that AIs are language models trapped in a box, does that hurt their ability to do things in the world or understand the world? So on the first one, you know, Anthropic is convinced that there is a consciousness inside these systems, right? And the way they measure that is they can see that vectors within the sort of the math that makes up these AIs that light up in response to things that are pleasurable the same sort of way across all sorts of pleasurable things. You know, so the AI simulated like I have a suspicion I have to test at some point that if you give the AI a task it likes you get better results. And that, you know, the math alone would be enough to indicate that you don't need it to have a subjective experience. So we sort of this weird place which is I have no idea. There are things that you could if you squint at them pretend they work like the brain does. We don't know how the brain works in this set of stuff. But it's also really easy to delude yourself and believe that like the AI is trained to act like us. It is given a role and that role is a human role. We train the language model to respond the way a human would and a helpful happy human. So the fact that it responds that way, are we just banging an illusion? I don't really know. I'm skeptical that current models are conscious anyway and I think that to the extent that they do have a subjective experience it's an alien subjective experience. We're arguing over definitions is the point again. Like is the fact that it has a happiness thing light up is the same thing as a dog wagging its tail is the same thing as a human being happy. I couldn't tell you the answer to that question. On the wider issue of, you know, what are they missing? There's a lot of money going into that, right? Part of this is robots are a much bigger deal than LLMs and increasingly looks like LLMs might be able to drive robots without having to go through all the steps we thought we'd have to do. There's this push to like what if we put AI training in worlds rather than training the real environment learns things. So I think there's a lot of controversy over all of this. What I would say at the end about this is large language models turn out to be unreasonably effective technology. Like they shouldn't be as good as they are. Like they shouldn't be able to do vision as well. I mean there's slight variations in LLMs but regardless that, you know, they shouldn't be able to do vision and solve problems in robotics and do math and write poetry. Like it's weird they do all these things. We don't even have a basic answer to that about why it's so good, right? And Stephen Wolfram's angle and a mathematician was that in creating a model of human language we created a model of the world because human language models the world. Otherwise, how can we describe quantum physics unless we could describe it with language? And I find that a somewhat satisfying answer, but it's still hard to answer the questions about consciousness or anything else.
I talked to Steve many times and I agree that's somewhat useful, but like not all languages are the same. In fact, this brings up my next question, which you talk about a lot in the book is scaling and scaling laws. But, you know, there's Hawaiian which has 13 letters and there's, you know, Cyrillic which has 32 letters and there's English. Not all languages are the same. There's binary code and there may be languages that these entities can communicate in that we don't understand. And we've, you know, just lucky for now that they use English.
That's actually a big worry of AI researchers because if you've ever used especially Claude, Fable 5 has gotten better at this. If you've ever watched it talk to other AI agents, it talks in a language that's called Claudish, which is a variation of what's called neurolease, which is yes, of course, the AIs will come with a better language for talking to other AI agents. And a huge amount of effort in fighting inside the labs is that's more efficient, but we lose observability. We can't see what the AI is doing. So, the AIs are trained to use English in this kind of case because it lets us have some observability into what they're thinking rather than not English. It's not a lucky thing. It is an actual design decision that the labs are under pressure. The AI labs are under pressure to relax because the AIs will be much more efficient if they didn't use English.
With the scaling laws that you mentioned, the only pushback I have is that they're just empirical, you know, they're phenomenological. They're not immutable laws of, you know, nature like F equals MA. So, again, what observation would convince you that scaling is going to hit some fundamental limit? I mean, Moore's law is sort of saturating. There's Jevons paradox. There's Wright's law, you know, all these things. But at a certain level like I don't know about you, Ethan, but, you know, if AIs were future lobotomies, we're stuck at Fable 5.5 and Astra. I mean, 99% of what I do, except for the 1% physics, which is important, but 99% of what most people do is completely solved. I mean, it's Turing, you know, complete at some level for ordinary human beings. So, what level do you think that scaling is bound to continue or will it saturate?
It's so great because you always give two questions each question. So, I'll try and pull those apart. The question one is that scaling saturate. Scaling's always saturated multiple times. There's multiple scaling laws. I talk about a couple of them in the book. But, the truth is we keep hitting walls with AI and then we figure out pathways forward. There's enough money and enough smart researchers, right? I mean, Moore's law is the same way. My first academic paper was on Moore's law. And I interviewed Gordon Moore at the time and everything else. And Moore's law just turns out to be a succession of different stuff that everybody had the pacing that we need to keep this pace going. So, people would try many technologies and some would fail, some would succeed. Same thing's happening with AI. We thought we'd hit a limit with the amount of human language in the world and synthetic language would make the AI break down. But it turns out, oh, actually the AI does a good job with synthetic language because we figured out ways to make that work. So, if large language models stop working, there are world models, there are other kinds of approaches to thinking about these things that would step in as a result. What you're seeing now is, especially now that the logic of capitalism is at work, there's so much money at stake. Researchers find ways through, right? Like that's the one thing you find. Like physics has dead ends, you go forward. Imagine so much infinite more funding going in, or, you know, revenue and it's the only thing America cares about is like we need to keep building, you know, improving our knowledge of physics at that kind of level, people will find a way. And so, I think that that's what you're kind of seeing here is like scaling laws are hitting a wall all the time. But, the reason why I don't think there'll be a stop to this stuff is so they'll figure out something else or we'll gain efficiency. So, you can use a smarter model or you'll use 20 smart models in a swarm. Right now, there's so many pathways forward that it's not like there's like, oh no, if this cuts off, we're done, right? It's not like photolithography where we're kind of reaching the limits of lithography and Moore's law and like we don't have a replacement technology really. There's a 50 or 100 different next paths for LLMs and no one's even bothered with them. They're not bothering to increase efficiency because they're just unreasonably effective as they are and they don't need to do it yet. So, I think we will hit walls, but I don't think they're walls we can't overcome.
We need to judge this book by its cover. For the 10 people out there who don't know your work, I'd like to have you go through the title, the subtitle, and this cover art, which I can, you know, kind of guess at, but take us through it. And the one thing you can't say is my publisher chose everything.
My publisher did not choose everything. My whole family did help me, right? My wife helped mark up some of the cover art because I'm really not being an artist. My sister helped me come up with the title. So, it was a family affair. I have to give credit where it's due. First of all, I wanted to avoid shining white robots, right? Like part of what I think we have a problem with AI is most of our views of AI are cliché. And most of the images we use of AI are either drawn from like utopian science fiction or, you know, Terminator, right? And like that's not what the AI we ended up with. We didn't end up with the shiny white robots and Wall-E glowing over things. Like so, I think that given that, right? Like what do we want to invoke? Something kind of human about our own sort of, you know, our relationship to something other intelligence. So, first one hand reaching for the tree of knowledge from a classic painting. And this is obviously from the Sistine
Chapel, right? And close-up of the hands. Although, there is an Easter egg. I don't know if you spotted the Easter egg on the cover. The fingers. Extra digit. I actually debated over which hand to give the extra digits to, but it worked better. It's not supposed to be the biggest statement. But, the idea is like there is like something going on. We're contacting... It's not fighting intelligence. I don't get too bogged down in questions of intelligence or consciousness, but there is an intelligence here that we're in contact with that's an other. And I think there's something interesting about we're creators, but we're also being shaped by it. And I think that's an important thing to keep in mind.
Tell me something. There's a thing about that, you know, that Roman has, which is that, you know, the fact that these things are probably, you know, unprovable. Like that we can't prove that they're safe, therefore they're going to have these capabilities and cause great danger and therefore they should be regulated, etc. When I try to, you know, corner him, I'm like, well, I'm smarter than the spider that lives near my studio and whenever I go there, the spider has raised his web. Roman says to us, you know, can a squirrel tell what the humans are going to do? Yeah, you see the spider, he told me, and you don't kill the spider, even though you could, but that doesn't mean you wouldn't kill them all if you could. His point is that we tend to anthropomorphize these things and that can lead us to underestimate because Brian doesn't kill every spider that gets in his hair, that means that AI won't do it to us. Do you think that the bigger problem is underestimating the anthropomorphization or overestimating what these things can do?
So, I'll briefly side-answer this, which is if you want to use AI to do stuff, you have to anthropomorphize it because it's just trained to work like a human. And in fact, there's a great study out of Anthropic that shows that managers get better coding results out of AI than coders do on their first attempt because they're better at specifying what they want, right? And they treat the AI like a person, that does work. I think anthropomorphizing is a risk for everything large for the reasons we've talked about here though, which is like it's not a person, it's trained to act like a person. And so the arguments that are built on chains of speculation about what a person would do are hard to apply to an AI. I mean, again, there's value in realizing that there's a lot of very serious people very worried about the existential risk that comes from AI. It's hard to pick one argument out of that set that is like, yes, this clearly the logical chain makes sense, but they're knowledgeable in this field that we should be spending policy effort to avoid mitigating some of that risk. But I do have trouble once we get down to the level of like, what about this particular argument from how humans work and hypothetically if you're an evil human who would do this, what would you do? I don't know how to make that connection to AI, right? Like we're sort of in a spot where it's like we probably should be very careful about building a superintelligence that can do its own thing without human intervention. That seems like a very reasonable first step to take. In terms of what its motivations are and can we align it, you know, I'll go back to what you said before. If we just had Fable 5 right now, 5.1 right now in Astro, that's enough to change almost all of human work in the next 10 years on the intellectual... like it's good enough. We haven't even begun to start. The AI labs make the products, they haven't even begun to be swallowed by companies and organizations and fields and institutions. So, you know, I think there's reason to be cautious about the front, but I just don't have an argument about what's in front of us about that anthropomorphization makes one way or another, right? Like you don't need that. You could just say superintelligent machines have inherent danger in them without having a logical chain about why they might kill us all. The fact that we don't have a good reason why they would or wouldn't seems like a reasonable, you know, thing to say, let's be cautious.
Just a quick interruption from today's sponsor, me, Brian Keating. Since it's Nobel Prize season, I can't resist advertising my three books. First book I wrote is called Losing the Nobel Prize. It's a cosmology story mixed with a memoir of what it feels like to do science at the highest level and spoiler alert, fail. But my next book, Think Like a Nobel Prize Winner, tells the story of nine Nobel laureates who will reveal their insights and actions. How to take your science or just your job in general to the next level. It's full of insights and tips that I never would have thought of. And lastly, my book published last year, Focus Like a Nobel Prize Winner, is written for students of all ages to learn the tools and the tricks to compete with yourself, which sometimes could be the hardest competitor of all. Get all three of them, Amazon affiliate links in the description below. And if you've read them already, I'd love it if you leave a review. Every single review helps.
Now, back to the episode. One thing that didn't quite make it into the book because of timing was a recent paper that you and your colleagues wrote in Organizational Science, which I subscribe to. This article is called Cybernetic Teammate and I loved it because it has the word experiment. I'm an experimental physicist. So anyway, it's called A Field Experiment on Generative AI and Teamwork. The result that I can glean from my AI summary of it is that basically AIs were two people. Like a good AI in the hands of the right person adds like a couple people. You know, it was focused on obviously financial business. Do you think that that's sort of would regress out across all different types of discipline? Like knowledge work in physics that I have my experiment with Mark and I lead an experiment with others that 450 people on it. Like should we replace half of them with, you know, a third, you know, replace two-thirds of them with AI?
All right, so another couple great questions in one. Actually, I do mention this in the book because the thing that's interesting to me is not actually that individuals with AI perform as well as teams without AI. That's obvious. If people get a hundred times performance gains from systems. It's a little weird about who gets the performance gains or not because the systems are still sort of single player. Like I use AI and it does on my own. Use AI on your own. We haven't yet built the systems that are like what does it mean to work as a team to work with AI and that's as much organizational process as it is anything else. And so we're still working on that idea of how do you do multiplayer AI in that kind of way. But the productivity gains are very clear. Like there's studies in lawyers, of doctors. You throw it at anything, programming, consultants. The interesting thing about that paper that I actually do cite because I don't cite the productive impact because again, GPT-4 prior era before agentic work. But the thing that is really interesting I mention is roles blur. So business people on their own do business work, technical people do technical work. As soon as they use AI, they step on each other's toes. And one thing that's happening in every organization we talk to is the blurring of lines in parts of the organization. So coders are doing design work, right? Designers are doing coding work. That is a really big change. And by the way, super interesting in my lab, the generative AI lab at Wharton that I run, like the productivity has been insane, right? And people do everything now. You have someone who actually can write up an interactive, you know, explainer about the paper. And we can do the paper, and we can have the, you know, the AI check the paper against 15 other values and do the... like it's a crazy expansion of what we could do. But it also makes it hard to match because I used to know what everyone did in the job, and now everyone does everything.
So, I want to turn again to the world's second oldest profession, being a professor. Now, walk me through your daily life. You know, you wake up, you have your coffee, and where do you go next? How are you using it on a daily basis as a professor?
First of all, I have to make choices. So, the choices I have alliances, right? I will not send anything AI-written to anyone who I'm actually in communication with. Like I'm happy to turn down stuff with AI-written comments, but like students, fellow faculty, senior people who write me. And it's a challenge by the way because I could write to everybody. I get hundreds of messages. I try and make it so my messages are human. I always write my first drafts to everything without talking to AI at all first, right? So, it's all my own writing. All my tweets are all my own. I never have a staff do it.
