I'm Rufus Griskum and this is the next big idea. Today, Ethan Mollick on AI, agency, and what comes next.
Ethan Mollick may be the most useful person alive on the subject of AI, not because he builds it, because he uses it relentlessly, skeptically in public, and then tells the rest of us what he found. Ethan is a professor at Wharton where he co-directs the generative AI labs and the author of the bestseller Co-Intelligence. His new book is called Coexistence, and this conversation is full of things you can try this afternoon.
Start by not starting with AI. Spend 10 minutes writing down your ideas and then ask the AI to independently generate ideas. Don't trust one model. Convene a council. If you just pick from one AI, all the ideas seem brilliant. Once you have some intention, you're like, 'Oh, actually, there's more choices for me to make here.' When you're stuck, ask for volume. I am locked here. Give me 30 ways to end this sentence. Before you publish, summon your critics. Tell me what a critic would say about this. Tell me what an expert might think that I'm glossing over that's important.
And get a second opinion where it matters most.
Still keep your doctor, but you should be running your medical records through AI.
But most critically, use it. Challenge the latest models with your hardest questions.
You absolutely can be the best person in the world at your job with AI. There is no doubt in my mind.
Then we wade into deeper water. Why do the leading models, even Grok, even the Chinese models, lean the same way politically? Are they converging on moral truth or just telling us what we want to hear? Will treating them as if they're conscious decrease or increase alignment risk? Will it matter if the song that makes you cry was written by a machine? And finally, if Ethan could snap his fingers and make AI disappear, would he? Ethan Mollick, welcome back to the next big idea.
Your last book, Co-Intelligence, came out a lifetime ago in 2024. In it, you laid out four scenarios for the future of AI: the as-good-as-it-gets scenario that AI has already peaked, the slow growth scenario, the exponential growth scenario, and the machine god scenario. What's happened in the last two and a half years since your last book came out? And how has it changed your perspective on these scenarios?
I think we're down from four scenarios to two, and it's the two weirdest ones that we're stuck with. So I think that it's extremely unlikely that this is as good as it gets. Both the confidence from the AI labs, the track record recently, and the fact that there's just a lot of ways forward indicate to me that AI is not going to stop development, right? And I think a true slowdown seems unlikely. There's some possibility maybe government coordination around that, but like there is a lot of money in this technology and there's a lot of money in making it better and there's lots of pathways forward to using that money. So even if like Anthropic disappeared, even if there was a financial crash, it's hard to imagine that there's not a lot of pathways forward for companies to pursue. That leaves us with two scenarios. And I've sort of renamed them the escalator and the eschaton. So the escalator is that things keep improving without ever kind of reaching a tipping point where we have sort of self-aware AIs or exponentially improving AI that change how AIs and humans relate much from where we are. Or there's a chance that the people building these systems are right and we get AGI, a machine artificial intelligence better than a human, then the world changes very dramatically.
In the escalator scenario, you no longer have the word exponential in that one of your two scenarios. However, if we look at the data, is it fair to say that what we're seeing, what we have seen, what appears to be underway is actually an exponential increase in capability?
Yeah, I mean, and you can kind of see this. When my book, when Co-Intelligence came out, GPT-4 had been the best AI model in the world for a year. And you know, after a very short period where its predecessor GPT-3.5, which released with ChatGPT, was the best, then it was GPT-4 for over a year. And the best AI in the world could do about 4 minutes of human work with a 50% chance of success. When I turned in my draft for Coexistence in April, the best AI in the world could do 17 hours of work with 50% chance of success. When the book was published in October, it appears to be around 17-plus weeks of work with a higher than 50% chance of success off a single prompt. And for a couple hundred bucks. So, that is an exponential for you.
Yeah. Oh my gosh. Well, you know, one way of measuring what has changed is to look at how the way you have written these two books has changed, right? Because it's a two and a half year gap between Co-Intelligence and Coexistence. How much about your writing process has changed during this period?
The interesting thing is less than you'd think, because it's important to keep stuff human, or important for each of us to make decisions about what. So I wrote the draft of the book myself. The last book, you know, Co-Intelligence, I talked about how I got feedback from the AI. The AI is now much better at feedback, right? Incredibly insightful. And part of why I'm trying to avoid having it give me feedback too early is my writing will change. It has really good ideas, and even if its ideas aren't really good, they'll infect my own ideas. So I have to write on my own. It's almost a survival technique. On the other hand, if you're used to AI making mistakes and hallucinating and making up errors, that just doesn't happen anymore. Not only did the fact-checking from the AI spot errors I made, but this time around it spotted errors in the sources that I cited. So it's like, 'Table 3 actually uses the wrong math to accomplish its goals' in a paper I was citing, not in the actual work I was writing. So we're already at a level of intelligence that's sort of quite high. And so I had to be very cautious about when I was using that and when I wasn't.
Well, I'm going to push back on you a little bit here, Ethan, because you say towards the end of the book, 'This book was not written by AI. But I say this with less confidence than I would like. It became increasingly hard to separate where the intellectual work was entirely my own and where it was developed in conjunction with AI.' So I mean, it seems that even over the course of writing this most recent book, you witnessed an increase in how compelling... I mean, perhaps what's shocking is when we see some of our favorite combinations of words emerging from the AI itself.
Well, it's worse than that, right? Like, I do use the AI to polish up a thing, help me with this final bit, right? You know, when I do my Substack post, again, same technique, write by human. And then they'll often be like, 'Hey, help me nail this sentence at the end of this.' And the worst thing is like people, you know, people be like, 'Oh, that sentence was amazing.' I'm like, 'That's the only sentence I didn't write in the entire document.' So I mean, what you know, part of the process is work is iterative, right? So like I write the first draft myself. The AI gives me feedback. I'm like, 'Ooh, that's a really good idea.' And then that gets carried through to other chapters. And so just like any co-authorship, right? Like I dedicate the book to my wife who I work with on all this. We started the generative AI lab together, and like I can't separate the ideas that were mine and hers that easily when we discuss this all the time. In the same way, the AI is getting very insightful, and I talk to it even, you know, a huge amount during the writing process.
And this is a human wife of yours?
A human wife. Yes. 100% human wife. Yes.
Fantastic. Well, it could be useful here to explain to listeners the distinctions you make between what you call centaur work, cyborg work, and conjurer work. What do those terms mean for you?
And just as a side note, I apologize for how they sound out loud. On paper, I wanted three things starting with C's, and I already had centaur, cyborg. So I went with conjurer, but the hard C doesn't sound as good. So I should have went with like sorcerer. It would have sounded better, looked worse on paper.
Yeah, but it would have like broken the reading symmetry, but not the audio one. So centaur and cyborg work are things I discussed in the first book, and they're still sort of valid, but you can think of Co-Intelligence as really about the chatbot era. So the chatbot era, you couldn't help but work with these systems, and you kind of had to make a choice. You could either do centaur work where you divide the work between yourself and the AI, right? I do this work, the AI does that work, right? I do the writing, it does the analysis. Or you do cyborg work, you blend it together, almost like how you Google something and read a web page and do some other work. And those were the two dominant methods, and the human was in the loop during those times. And that is just no longer the way the agentic era works, which is when the AI could do 17 days of work on its own. You're not doing the same kind of back and forth. So the conjurer idea is you're summoning servants from the void to solve your problems for you when you need to and sending them back afterwards. But I mean, the less exciting version of that is manager work, right? You're acting like a manager with a team.
You have said that when writing you're a centaur, but research and fact-checking you're in cyborg mode. So there's a flipping between modalities that's happening here, right? A lot of the book is about how we make choices about adapting to AI, right? And we get to make decisions.
I think that one of the things that worries me most is people just deciding that AI is something being done to us. And there's always an aspect of technology that's true, but we have to interpret what this means and how to use it, right? And we do get to make decisions. So in writing, I really, really want to be clear about what my role is. Because I feel like I have valuable things to say as an author, and I feel like people want to get my book because they want to hear what I have to say, and because I don't want to lose that. In the research side, I'm a good researcher, but I'm not worried that I wouldn't, you know, I could delegate research to somebody because I generally know what I'm interested in looking for and I know some of the works myself. So, you know, we blend back and forth. I'm like, 'Hey, am I citing this paper properly? This stat might, it seems dodgy. Can you look up 50 or 60 options for this? Is this paper holding up still today or not?' Like those are kind of questions where we're blending, and there might be 50 different projects happening with the AI in the background looking up stuff that I want to confirm is right or not.
You know, I want to say by the way that I think this is where your work in particular, Ethan, is singular in the marketplace of ideas, because, you know, right now we hear from people who are either highly skeptical of AI or just wide-eyed enthusiasts. And one of the things I like about your work is you really drill into how we can be highly pragmatic and speak clearly about what use of AI is empowering, what will produce bad outcomes, and maybe even worse, what might erode our capabilities. I think the nuances of how you kind of parse this out are really helpful. And so as you say, like researching, fact-checking, when fact-checking you say that you like to use a council of AIs. I'm very curious about how exactly you do that.
So, and thank you for those kind words. I mean, it's tough, right? Because AI is many things to many people, and it's hard to stay in the middle of like, yes, this is both transformative and we need to make decisions. So the council is kind of interesting, right? AI systems are quite smart, but they have their own biases. You kind of have to think of them like a really smart person with its own viewpoints. They also sometimes miss things. They're not perfect. And so I find that bringing together multiple AI systems, I usually using Gemini and Claude and GPT in a combination, finds things the others miss. And also you get to kind of know their personalities. It's almost like you send it off to, you know, send a book off to be read by your mother and by a colleague and by an editor, and you know you're going to get kind of different results from all of them. And there's value in that kind of piece. So again, that's a way of centering yourself as a human having different opinions to pick from. If you just pick from one AI, all the ideas seem brilliant. Once you have some intention, you're like, 'Oh, actually, there's more choices for me to make here.'
I recently decided The Next Big Idea should get with the times and set up a website. But I had no idea where to start. I'm not a designer or developer. I don't know how to code. I know what I like, but I have none of the technical expertise to bring it to fruition. But then I found Hostinger. I've been able to build a professional website all by myself, giving The Next Big Idea a new way to share our favorite episodes with wonderful listeners like you. With Hostinger, you just describe what you want, and the Hostinger AI agent builds a version in minutes. And it doesn't stop there. Once your site is up and running, Hostinger's AI agent acts like a built-in specialist for SEO, website copy, email marketing, and more. If you're not curious already, get this: plans start at just $3 a month. Creating a website doesn't have to cost thousands. You can begin at your own pace and build from there. Hostinger builds your vision step by step. Create your website today. Go to hostinger.com/NBI and use code NBI for 10% off.
Now, here's something that Mark Pincus, the founder of Zynga, told me that I can't get out of my head. 'You know, hope is not a strategy. Kill hope before it kills you.' What he meant by that is that you can't let your desire for an outcome for your business cloud your ability to assess what's actually going to work. His new book, Life at the Speed of Play, is eye-opening. It made me want to build again. I told Mark it's one of the four best business books I've ever read. Mark's a legendary Silicon Valley entrepreneur and prolific idea generator. His instant national bestseller is a fast-paced 170-page read for uncertain times that feels like you're sitting in his kitchen. It's packed with invaluable life skills and a battle-tested, easy-to-follow framework to rapidly test ideas, get real feedback, and separate winners from losers and stop wasting time. If you purchase one copy of Life at the Speed of Play, Mark is paying for a second copy to send you. Not the publisher, Mark personally. Go to getmarksbook.com/next. Buy a copy and Mark will send you a second one. That's getmarksbook.com/next.
Getting into some of these nuances of how we can kind of enrich our intellectual experiences while taking advantage of the power of AI. You have this kind of great observation that deep knowledge and wide knowledge are actually becoming more important, not less important. But it's very easy for us to basically begin to surrender our passion for developing deep knowledge and wide knowledge. So do you want to expand on that?
Yeah. So the idea of deep knowledge is expertise, and expertise is hard. We actually, you know, there's a lot of things that we don't have good definitions for because the only model was human. So there's a lot of research on expertise, and it's really hard to judge expertise. Like you're an expert at these conversations. You're telling from a thousand clues you couldn't tell somebody about whether this is going well or badly, whether you need to be more provocative in the conversation or not, whether I'm going to be boring or not. And you're making a thousand decisions based on things that no one could ask you about. You'd say, you know, a feel for it. Like if you ask someone who's an expert at golf how to be a better golfer, they're like, 'Well, you feel in your hips when it's going well.' You're like, 'What does that mean?' Right? So we develop expertise through formal learning, right? And through much practice. And of course, AI erodes both of those things. So deep knowledge lets you stay on target with something. The other thing that's interesting is wide knowledge, and wide knowledge can mean both in the humanities sense of wide knowledge, right? Which is that I'm, you know, I know many liberal arts, I know many things, right? And the AI knows everything. So like the example I like to give is if you ever use Claude or OpenAI or ChatGPT for design, they have a tendency to put little subheadings over the subheadings they write above the titles they write. And it turns out there's a name for those things. They're called eyebrows. And if you know what eyebrows are, you can say, 'Please do not include eyebrows,' or 'the eyebrows should be in contrast to the main headline.' Just a minor thing, right? But we're seeing in companies, by the way, the same thing happening. Coders are doing design work. Designers are doing coding work. If you know a little bit of coding, a little bit about design, you may not be as good as a real designer or coder, but like that broadens what you can do. So that generalist and specialist combined can really help you out. And I think especially the specialist is undermined by AI use because AI short-circuits the hard one problem of building expertise.
And meanwhile, as you say, I mean, on the one hand, one might think like, well, the AI itself, I mean, one of the strengths of AI is that it has like endlessly wide knowledge and increasingly endlessly deep knowledge, right? But as you point out, the wider and deeper our knowledge is, the better we are capable of interfacing with the AI to ask the right questions, push it, and get more out of it. So it's a two-way process.
It's important to realize that deep knowledge and wide knowledge are not shining lines. Like I think we make a mistake, and people keep making this mistake over and over again, of saying AI will never do this, that there is a bright shining line. This is human. I don't think that's a safe thing to do. Right? Now, we could say it's different, right? But like to say AI, like human judgment is needed... Let's unpack what you mean by judgment there, right? The AI, if it's doing an 18-week-long task in 10 minutes, it is absolutely exercising judgment and creativity. And now they may not be your judgment and creativity, and that's worthwhile to say, which is like I have a particular thing I'm bringing to the table. But it's important to not have a bright line at this point where we say AI will never be able to do this. Maybe you're right, but we've been wrong about everything before. I've been wrong about the things I predicted AI would have trouble doing.
Interesting. Very quickly, what things did you think it wouldn't do?
So most recently, right, I'm an economic sociologist by training and I work at a business school, and I thought the hard problem would be, and this is very typical in your area, you think it's a hard problem, I thought it would be coordinating AI agents. I thought when AI agents had to work together, they would have to take lessons that we've learned over 200 years of business training, right? Like how do you build organizational structures? How do we pass information back and forth? How do we coordinate work? I thought all of that would take a couple years at least of trying to figure out the best approaches, and that was just wrong. It turns out that we watched this in the Hugging Face incident, which was a big security incident. We watched this in a proof of the Navier-Stokes formula, and Navier-Stokes, 10,000 AI agents clustered on the job, passed 3 million messages back and forth, and there just wasn't a problem. So this big issue I thought would be years of trying to figure out how do we get group swarms of AI agents to work, fleets of AI agents to work together, they just do it. This is something called the bitter lesson, which I talk about in the book as well. We just keep making the same mistake over and over again, which is generalized AI learning systems solve problems that we thought would involve lots of careful human expertise.
But at the same time, just because the AI is figuring this out, and arguably there's more progress to be made in the coordination layer. You know, I'm sure you saw Google released recently a very interesting paper about the problem of coordination and opportunities to do it even better. But even as we learn again and again the bitter lesson, and our own skills are exceeded again and again by AI in this jagged way as you describe it, or inconsistent in different areas, we still see these great examples like Google AlphaFold beating the best human Go player resulted in Go players getting better. I had not heard that. I read it in your book. I thought that was very interesting.
And chess players got worse, right, in the short term. It was interesting. Right. I think, look, this is a complex technology interacting with us, and we've been talking about how amazing it is. It also has limitations, right? There is this jaggedness that's sometimes hard to describe, and like it's weird. The jaggedness is weird. I mean, some of it's technical, right? AI is still a pretty mediocre writer, prose writer. It's not bad, right? But it writes a very consistent kind of prose. And if you open LinkedIn right this moment, your eyes will start bleeding because you're going to read a thousand things that are like, 'This is the part that really sticks with me,' right? Like it, you know, and it's sort of like the professor now, like any one paper might read, you know, you might be able to get away with it. It's AI prose, but like it's the same idea as every other AI paper, right? So there is this problem with the prose piece, that consistent diversity of ideas. And also just weird things like AI has different sort of underlying values than we do. So like one problem with AI shopping agents is they don't really value price as much as humans do. They value it as one part of a rational decision-making package, and we overvalue price, right? And so there's really interesting questions that the AI is challenged with that I think are really hard ones that leave lots of room for humans. That's aside from stuff that like lots of jobs are undocumented, it can't be done on a computer. So there is this jaggedness that creates a lot of opportunities for humans to interact. But I think it's, you know, that will keep evolving over time.
This jagged line, I think this is a term that you introduced, right?
Yeah. In our first paper on this, you know, we came up with the name of the jagged frontier.
Right, the jagged frontier. And in our efforts to interact with AI in a way that is kind of most empowering, and at the same time we adapt and move to what you call the valleys of the jagged frontier, do you want to speak a little bit to like how important it is to understand where the jagged frontier is? Like mapping the jagged frontier of AI's strengths versus its weaknesses, which is constantly changing, is pretty critical to us leveraging it effectively.
So start with the principle that nobody knows the answer to this question in your field, right? We like, the AI labs know roughly where jaggedness is. They know writing is an issue or legal reasoning is an issue, and they put resources into trying to solve those problems. But the frontier is micro, right? If we take writing a podcast like this for example, right? Like, and you will find that even if you narrow down to the area of research that it's going to find some sorts of, it's going to be good at some kinds of research for you, but not sort of the insightful pulling across research, or it will be good, insightful, but it has one sort of perspective that's not quite your perspective. So you'll find that jaggedness at every level of your job as you dive into this, right? Nobody can tell you where the flaws of AI are going to be if you're running a car rental business in Panama, right? Like we don't, like there's nobody's ever thought of this before, right? And like in your particular case of doing marketing for a car rental, like we don't know. So you learn this by using this yourself. So you have to use these systems enough to do it. Now the changing aspect is kind of unnerving, but as you build an intuition for this, you can sort of get an intuition for where it's changing also. So you can get at this by use, by experimentation and by trial and error, but there is no formula to get you there.
I have spent years interviewing dozens of leading scientists about how we can live healthier lives. A common theme has been that we're entering a new era of personalized medicine, and it's essential to understand what your specific genetic predispositions are. Family history is a very clumsy way of assessing that. Today, we could sequence our DNA and get the full picture. I'm doing this myself, and I've chosen Sequencing.com. Sequencing reads 100% of your DNA, all 3.2 billion positions, screening for variants linked to over 15,000 diseases, conditions, and traits. Most popular tests read about one-tenth of 1%. Here's what's exciting: your DNA doesn't change, but the science does. Sequencing.com keeps reanalyzing your genome as long as you're a member. Your data is never sold or shared. Right now, Sequencing.com is offering our listeners up to 10% off your full genome test, plus free shipping. Go to Sequencing.com and use promo code IDEA for 10% off. That's Sequencing.com promo code IDEA. That is Sequencing.com promo code IDEA.
I've started a few companies, and here's a dirty little secret about entrepreneurship: you can have a brilliant idea that might just change the world, but before that happens, you're going to have to figure out your operating agreement, your business address, your website, and your phone number. It's enough to make you want to get a real job. But please don't do that. There is a better solution. Northwest Registered Agent gives you a complete business identity with privacy built in and free tools to help you grow. Northwest has been helping small business owners and entrepreneurs launch and grow businesses for nearly 30 years. They're the largest registered agent and LLC service in the US with over 1,500 corporate guides. These are real people who know your local laws that can help you and your business every step of the way. Don't pay hundreds or thousands of dollars for what you can get from Northwest for free. Visit northwestregisteredagent.com/nbifree and start using free resources to build something amazing. Get more with Northwest Registered Agent at northwestregisteredagent.com/nbifree.
There's understanding what AI is good at and less good at, and how we can interact with it. And it does seem to me to be very empowering that because the jagged frontier is changing so quickly, it's somewhat empowering for everyone listening to know that they can start to experiment and through trial and error make their own assessments. Like there's an empowering...
You absolutely can be the best person in the world at your job where they are. There is no doubt in my mind. Like whatever it is, the flip side is nobody knows what's going on. So like there's both scary and power in that, but also like I want people to realize there's agency here. Like you can empower yourself with these tools, and ignoring them doesn't empower you, right? Like pretending this isn't real or these systems aren't good doesn't help you. Like I'm a professor of entrepreneurship. This is the golden age for entrepreneurs, right? Entrepreneurship is all about not being killed by the fact that you're trying to be a specialist and generalist at the same time. I mean, when I was doing my first startup in the first internet bubble in the late 90s, I still remember the first payrolls we were sending out. I did not realize you could pay a company a couple cents to do payroll taxes for you, and I would calculate them in Excel for like 5 hours a day. And I was like, I didn't realize that it was so easy to write a check, this kind of, you know, just pay someone. Like that is impossible to imagine, it would obviously do that for me. But on top of that, just the level of advice and help that you get, it's a really golden age for this kind of stuff. So I think people should feel empowered. Like there is danger of disempowerment, but there's also an incredible tool to be used for people who want to use it in the right way.
You know, and so there's also the question of, I mean, as we think about again using AI for what it's best at and recognizing where it's weak, there's also a separate calculus of remembering that there are desirable difficulties that like all of us, I mean, we will talk about students and education and how education's changing. In some sense, we're all students, right? I mean, that this is like our job as humans. We're constantly learning. And as you point out, we have to take on a certain amount of friction. We have to choose to do hard things in the learning process in order to continue to grow and expand our expertise both deeply and broadly. So how do you think about that piece?
Yeah, I think that is, you know, it's an example of many challenges we're facing here that again, as humans, we're sort of built wrong for this, right? So learning has to be hard. There's just, I mean, I spent, you know, more than a decade building games for teaching, and one of the things you realize is you can only make it so fun. Like you cannot, like everyone has areas that they love, and they're like, 'Why isn't all learning like fill in the blank, you know, reading novels because I love that, doing math because I love that.' Like because even people who are intrinsically motivated to learn some subjects are not intrinsically motivated to learn everything we think we should know or the society thinks we should know, right? Like you may not be interested in history. We want you to know American history in the US, right? We want you to be a good citizen. So to learn anything requires difficulty and friction for most subjects, for most people. Yeah. You could be auto-acted something, but you're still probably having friction in some area. And even in that area, you're probably thinking that you're avoiding learning because that piece was boring for you. So it's hard because learning is hard, it makes our brains hurt. And it turns out if we don't make you feel pain in your brain, you're not actually learning. And we're very bad at knowing when we're learning. The single most effective intervention I've seen in a randomized controlled study to get students to say they've learned more, you get about, I think, 20 to 30% more learning in the past semester if you bring cookies the last day of class because they feel good about the class and then they retrospectively think they learned more. So you're turning people on a tool that can make you smarter but also can do the work for you. And we're very bad at knowing when we're learning something or not, and we're motivated to not want to do hard intellectual work that's boring. And as a result, people just let AI do the work for them, even if they don't mean to. So I think we've internalized this in weightlifting. Weightlifting, you know, like for most people a lot of the time sucks, right? Dieting isn't pleasant. We understand we have to do self-improvement that's hard. Education, partially because of how it's done, how it's taught, a lot of people are just resentful from the beginning, and we don't think it's that important, and why this homework assignment, even though testing and homework seems to be helping, we don't want to do it anyway. So we have to get into a habit of making life difficult for ourselves in a world of AI.
You have a great example, which is that you say like that AI might write the best advertising copy, and we can debate whether or not today the best AI models are going to write the best advertising copy. But we can certainly imagine, given the pace of progress, that there will be some inflection point at which AI is going to write the very best advertising copy. But the act of researching the customers to test the copy, that act yields new insights.
Yes. And so the process, there's a question of outcome and getting the best outcome, but there's also the importance of retaining processes that enrich us.
Yeah. I mean, and that's part of when I have this discussion about the bitter lesson, right? Which it's worth expanding for two seconds on it just because I think that's useful for this piece, right?
So the bitter lesson came from Sutton making this point about chess computers, right? And how chess worked, which was when you're building a chess computer starting in the 1970s, you'd hire grandmasters, you put in every move, a thousand years of chess experience. And indeed, you know, and that's how you program an AI. And that's how Deep Blue beat Garry Kasparov. It had a grandmaster working on it. It just could look at more moves, right? You probably heard it's like 40,000 moves a second or whatever, right? That was the value, is it could do more, right? You know, like more than a person could. And that's how chess computers worked until AlphaZero, which you made some reference to earlier, came out, which was a chess computer that learned how to play chess by playing against itself. And it started off not knowing chess, and by the end it could beat any chess computer and any human at chess, right? And so the bitter lesson was your thousand years of chess knowledge are less useful than an AI system just grinding away at the problem and learning how to do it, right? And so if you have a verifiable output, if there's a thing you're looking for, you are in danger of being bitter-lessoned, right? You're in danger of, yes, the AI can do that. If we know what a good ad copy looks like and we can do tests on it, you know, ad copy is a little more challenging, but like that's why math is so easy for the AIs at this point. Yes, you can give them, like they can verify coding, right? If the code runs, things are great. Though two weaknesses of the bitter lesson are: weakness number one is that verifiable domains work better. Part of the reason why writing is a larger challenge to the AI is that we don't have an objective measure of good writing. So it's hard to have a feedback loop. And the second is, like you said, process. AI is very good at outcomes, but if the process matters, if the debate over the PowerPoint matters, if, as you said, the market research matters, then that is where the AI is weakest. I worry that people cluster too much on outcomes, spend too much focus on outcomes when learning process really does make a difference.
Well, another area where the bitter lesson applies is that, of course, we have long thought that creativity was the realm of human excellence or human superiority. There's this Wharton study, I think, by some of your colleagues. 200 students were pitted against GPT-4 in an idea generation contest rated by independent human judges. 35 of the 40 best ideas came from GPT. That was GPT-4. But you point out that actually this is an opportunity, that actually generating quantity of ideas is an excellent use case for AI.
