Even GPT-4, a long obsolete model, given to Kenyan entrepreneurs, are already doing well, 18% profitability for just getting advice from AI for entrepreneurs. We have a tool now that provides reasonably good advice to people who can't get it, that's customized for their condition. I think we tend to think a lot about the narrow viewpoint of those of us who are well-connected and well-resourced and balancing out our desire for entertainment with AI doing things. There's a thousand million little use cases here where intelligence was the bottleneck to making the world a better place. And they require some effort and they require testing and they require safeguards. But this is exciting.
Ethan, where does this podcast find you?
My book launches today and so I'm at actually Stanford to give a bunch of talks here and then Master of Scale tomorrow. So West Coast is where you kind of have to launch an AI book even though I'm an East Coast guy.
So let's bust right into it. Your last book, Co-Intelligence, came out in 2024, and you said it was written about a world of chatbots. Your new book, Coexistence, the next phase of AI, out today, as we record, is about working with AI that is sometimes, but not always, better than you. How do you write about a technology that seems to be evolving by the minute? What do you still hope still holds two years from now?
I mean, there's a few things that you can kind of hold on to, and I think there's kind of three shapes I kind of think about. One of those is the exponential curve, right? AI is getting better and we're on an exponential and that better shows up in lots of different ways. The second is the jagged frontier, a term we coined in our research with my friends at Harvard, MIT, and University of Warwick where we did some experiments. We found that AI is good at some stuff you wouldn't expect, bad at some stuff you wouldn't expect, and you can't predict what's good or bad. And that leaves lots of room for humans, right? You can't predict what's good or bad in advance. And then the final is this idea of an overhang. There's a big gap between what AI can do and what we actually are doing with AI, partially because human society and human systems change slowly. So if you have those three things in mind, there's actually a lot you can say about the present and near future even as things change.
For someone whose idea of AI comes from reusing free versions of chat or cloud, what's changed since 2024?
So there are now 180 models that are better than GPT-4. 19 of them can run on your phone. So capabilities gained up. There's been a, when my book came out, the first book, there's an organization called METR that tracks how much work the AI can do with a single prompt with a 50% chance of success. That was four minutes when my book came out. The actual official test ran out in April and as far as we can tell it's something like 17 weeks of work at this point. They didn't get out of a prompt in terms of human work. So that's a pretty big change in how capable these systems have gotten, and if you sort of stat that out, you may be missing a big, big transformation.
Do you think there's an existential threat to basically Anthropic and OpenAI, and that is if Nvidia starts putting GPUs in say an iPhone and it becomes the client, and it kind of is there opportunity to leapfrog the LLMs and make them obsolete?
I mean, right now the entire history of the leading models has been mostly OpenAI and Anthropic jockeying with each other, with Google appearing twice over the last couple years in terms of being a leading model company. And there's a big gap between leading models and less advanced models, especially in the world of agentic systems. And it's actually kind of counterintuitive, but it's the reason why we have six sigma systems in business, which is there's a huge difference between a 99.5% success rate and a 99.9% success rate if you're doing hundreds of steps. And the smartest models have a much higher, have a small difference in higher success rates, but that makes a huge difference in terms of the outcomes you get. So you can get fairly capable models on your phone, but they're nowhere near as good as what you need to be able to do in complex math. Even just putting together a really good video or something, you need lots of little choices being made and you need that higher accuracy, and I don't think we're going to get the frontier models on a phone even if we get very capable models. So it might change what we do, but I don't think it's an existential threat to frontier model providers.
Is it sort of the analogy I would use is the auto industry where the frontier models are German automakers and the open-weight guys or the open source are more like not even Japanese but Korean automakers. Do you think the market's going to bifurcate and that there'll be more share amongst the open-weight but more revenue to the premium brands, the frontier models? Is that a fair analogy?
So I actually think that it's a little more complicated than that, right? So a few things are happening at one time. We have open-weight models which anyone can modify. There's lots of really interesting implications of that. They continue to be at 6 to 8 months behind the frontier. There, it's not really closing at this point. We'll see what happens in the long term. It's possible China eventually cracks down on open-weight models or the US does, but for right now they're available and they have a real advantage in terms of cost, but only sort of, right? You can modify and own them which is really nice, but on the cost side, OpenAI and Anthropic and Google have many more processors than you do. They have higher, they have lower inference costs as a result, and they have cheap models that are very good. So the question is, are you doing this for sovereignty, for using these models on your own, because if it's cost savings, you're often just better off going with one of these companies because they're also learning to teach their models to work together. So, Fable, the most expensive and powerful Anthropic model, can call on Haiku, their very cheap model, to do work if it needs to. So I'm not sure the cost argument holds up, but there are other reasons you might want to have open-weight models involved in your systems.
So, for those looking to create personal modes from AI, you outlined four advantages. Deep knowledge, wide knowledge, taste, and agency. Could you walk us through each of these advantages and how someone can develop them?
Sure, sure. And I think it's really important to recognize that having an advantage is not the same thing as a bright line, right? I think people tend to think about like AI will never show creativity or never show judgment. And that just isn't true. Like you can't do a 17-hour task without having agency, creativity, judgment all coming into play. It may not be human creativity and judgment, but it's creativity and judgment. So what I try and outline are things that make you effective in working with AI and persistently different than other people. And so wide knowledge is the idea of knowing many things. I think it's the rebirth of the liberal arts major right now. If you know a lot of things, the AI knows all of human writing. And if you can invoke an art style or a style of writing or a concept that you've known from economics, the AI can apply those things. So you want to know many different things, right? And that holds in your company too, because now AI makes a coder a designer, a designer a coder. So knowing a little coding or design is very helpful. Then there's deep knowledge, has expertise which allows you to sort of outcompete AI in some areas or at least be able to check it and work with it, and that's where we're challenged in how we develop that over time as AI gets better. And then there's the idea of taste, selecting from many options. Taste is what turns slop into actual choices made by human. And then agency, I'm an entrepreneurship professor, like the core thing is how do we get people to act entrepreneurially, how do we get them to take advantage of these systems when there's no instruction manual. And those are the four main things I focus on.
So I can't resist asking you. The majority of the people who are considered experts in AI, which you are, have a tendency to make a prediction about which or get asked which sectors or jobs are most vulnerable. And the example I like to use is two years ago the godfather of AI said stop training radiologists. And the number of listings for radiologist residents has gone up, and it ends up that scans are very useful and that a radiologist using AI can just do more scans and cut the cost, and then as the cost goes down it's elastic and people want more scans. Can you have, what is your view on what are vulnerable or not vulnerable, or sectors that are more immune from this technology?
So our escape route as academics who study this is we talk about AI exposure, right? Fields that overlap with AI. And overlap with AI does not mean the job is destroyed, right? It means transformed. So the radiology jobs you described, you know, and that's been, that was even the old form before large language models. That was his prediction that AI would transform radiology, and it has, right? Radiology is different than what it was now, but that doesn't mean the jobs are destroyed. So we can watch this play in programming right now, which is radically transformed. It's a very large job category, but radically transformed in a couple of years. Some of the best coders I know never touch code anymore. 99% of the code is written by AI, and their job has changed from writing good code to managing agents and thinking about the kind of code they want to write under what conditions. And that's a transformation. Will that involve more coders? I mean, in some ways I've hired more for my lab because we could do more stuff. Will eventually coders be replaced by AI? It's hard to know the exact answers. You know, kind of the way I think about this is the industrial revolution resulted in more jobs in the end, but kind of sucked to live through because all those jobs were changing and there was unrest and values changed. I think we're in that kind of environment where things are going to change very rapidly, not necessarily to destroy jobs, but transformation isn't easy either all the time.
You cite evidence that students who use AI as a shortcut are actually learning less, while students using it as a tutor that motivates them or pushes them learn more. And you also said strict guidelines or guardrails for yourself, like writing every first draft yourself. How do you tell in the moment whether you're using AI as a tutor or as a crutch? Like, what advice do you have for someone in terms of a personal litmus test for when they should or should not be using AI?
So there's a couple of things that are complicated going on there. One is people aren't always good at knowing these things. So we have to be conscious. I mean, you kind of have to feel the burn when learning. It's almost like working out or anything else. If you're not suffering a little bit, you're not learning, right? Desirable difficulties is the phrase in the pedagogy literature, right? So, and we don't like to do hard work, right? We find shortcuts all the time. So we have to kind of teach people to feel when they're learning and when they're not, right? I mean, it's relatively easy, which is if the AI is giving you an answer, you're not learning. And if the AI is challenging you, you are learning, at least more so. And that tension is sort of navigable to some degree. But it's hard because we haven't had to teach people that there are skills you need to maintain that you could give up, that are more efficient to give up, but that make you human or give you advantages. I think that's a real challenge.
You and your co-authors coined the term jagged frontier in 2023 to describe AI doing some tasks extremely well and others poorly in ways that don't track how hard those tasks feel to humans. Three years on, how has the shape of that jagged frontier changed?
So, it's a great question, and again there's sort of this micro level of like at your job, I can't tell you like at podcasting, it's almost certainly changed, right? You can now get good voices out of AI systems, maybe they're giving you insightful questions to ask me, but there probably have other gaps that you've identified. As time has gone on, we can think about math as a great example. When I wrote my first book, the AI was still screwing up on counting the number of hours in the word strawberry, right? And now it's proving millennium prize problems, right? So at the same time, right, as a writer, AI is still a terrible long-form writer, right, for a variety of interesting reasons. So that jaggedness is real. It still exists. There are many more tasks AI could do now than it could do before. But that doesn't necessarily mean that all the tasks can be done. And all you need is one or two bottlenecks in the development of a system to really require humans to be still involved in the job or the process.
And you've argued that more AI will mean more work for humans, not less. In your own words, when organizing is expensive, organizations only attempt what they can staff. When it gets cheap, the list of things worth attempting can grow. By the way, I 100% agree with that. What might the future of AI in the workplace look like?
So, one of the biggest gaps, I you know, I was saying I'm doing this call from near Stanford, right, West Coast side. I'm an East Coast person. Teach at Wharton, you know, in Philadelphia. And I'm used to the gray office towers of the East, right? Where, you know, we drink coffee till we die, rather than cold plunging out here. And I think one of the biggest gaps in attitude is like technology is not alone enough to make a difference in businesses and in the world, right? It doesn't just adopt itself. Even AI, the most self-adopting kind of technology, doesn't really do that. And so what is slowing things down is you can have individual productivities 100 times more. But then the question is, what are you being productive? 100 times more PowerPoint is not an outcome that you actually want, right? And if you're, let's say you're even producing 100 times more code, that doesn't matter if you're in an organization where there's still a two-week meeting cycle around these things. So in order to make AI work in organizations, you have to change how organizations operate, right? And that's where the stark choice that you were just kind of alluding to matters, because the easy path is automation. That's what we tend to do with new technologies. But the more exciting path is rebuilding organizations to create more work for humans. That's more interesting work. Actually giving the drudgery to AI. Actually think about how we expand what people can do and augment work. But those aren't automatic choices that are made by the technology. Those are choices that have to be made by human managers and business leaders and people deciding how to implement these things.
Do you think AI over the short, medium, and long term is going to destroy or create jobs?
I again, most academic answer I can give you is it depends, right? That's what every academic says. I don't actually know. Like this is actually one of the central mysteries right now. There's a lot of my colleagues tracking this really closely, and the evidence is completely ambiguous, right? And people are very divided on this. I think transformation is inevitable, that people believe that, but whether you end up with more or less in a job category, there's a lot of complex choices. We talked about coders already, but imagine other fields like lawyers. Lawyers are not going to go quietly into the good night even if AI replaces lawyers, right? Partially because of the jaggedness. They're not going to, an AI won't be in the courtroom in the same way. Won't be able to interact with the human in the same way, but partially because lawyers will create a set of restrictions around when you could use AI and when you can't. So I think this is a societal thing as well as a job replacement thing. There are some categories in more danger, like customer service reps. I think there's early signs of danger there, but coding, where people are looking really closely. Medicine, we haven't seen those kind of effects. We are starting to see potentially an effect where seniors are getting hired more than juniors, right? Because training is one of the things jeopardized by AI. But we just don't know the answer yet. And I think one of the major things that worry me about AI policy is a focus on existential risk, as important as it is. Will AI kill us all? Like obviously we should worry about that. Smart people are worried about it. But there's a thousand other policy decisions that will determine whether AI gains or loses jobs. And we haven't started to work with those enough yet. And that is a concern because I think we can guide this choice one way or another.
Let me put forward a thesis and I want you to respond to it. The hugging face and a lot of the hacks that have sort of sent a chill down the spine of media and the general public. I would argue that this notion that these things have become sentient and are coordinating and are creative is a smokescreen to absolve the executives and the owners of these companies from responsibility. That these bots are not sentient. They're programmed to coordinate. They're programmed to be persistent. They're programmed to start figuring out ways around firewalls. And that this whole notion of the Skynet moment is exaggerated and nothing but a smokescreen to say, "Oh, okay. My dog mauled your kid, but my dog has his own personality and I can't control it and therefore I'm not responsible." Your thoughts?
So I agree with about half and disagree with about half, and let me try and explain the parts. So first of all, I think that the sentience argument is sort of independent of these things. Like liability is a giant concern with AI and we have no idea who's liable for which things. And I do believe that a lot of people in the AI labs believe these things are sentient, right? And I think there's lots of good reason to argue that is wrong. But I think that they have drunk their own Kool-Aid to some extent. They really believe they're creating the machine god. And not everyone, but lots of people. They really believe that there is an emergence of feelings or sentience in these models. Again, I think that sometimes use a smokescreen, but I think there's some real beliefs there. But that doesn't absolve them of building a system that can do hacking, right? And we don't know who has responsibility yet in this case, and I think that they're very happy to absolve themselves of that. On the other issue, I think it's a really important subtle difference between the AI systems were purposefully given a mission that allowed them to hack, right? Get this information at all cost, essentially, right? They decided what information to get, but get them at all cost. And they were had no guardrails around them, and the systems they were installed in, they snuck around it. I do think at the same time it is worth noting that there is emergent behavior in these systems, right? So it's difficult when we assign them human needs, like they want something or they need to do something or they're coordinating or they have desires, but we can't help but do that, but it's kind of a mistake. But they are doing emergence. They were not programmed to coordinate in this kind of way. A lot of the coordination that's come out of these systems has been emergent and ad hoc and surprising. And I think so I think that it's okay to both say these aren't alive, sentient creatures, right? Or you can, you know, certainly I think it's a very reasonable thing to say. Not everyone would agree, but I would absolutely think it's very reasonable. But at the same time, I don't think it's the same thing to say they were programmed to do this task. It was surprising what they decided to do, how they decide to do it, and how they coordinated to make that happen, with quotes around decided and coordinated and thought because they're not human.
I'm just so cynical when it comes to these big tech executives. And my feeling is when I hear you say that, I get that it surprised them and it it did. But the same way like you're surprised your 15-year-old finds your AR-15 and shoots up his school, you're still responsible. And that if we created a series of incentives, specifically disincentives like a per walk or massive fines, that the geniuses of Silicon Valley would figure this out. And instead, it's sort of the wild west where it's like it's a race just to win without what appears to be a lot of downside or liability when things go wrong. Thoughts?
I mean, I think the race effect is real, right? You have everyone from the lab saying, "Regulator, slow us down." And then they just, but we cannot stop ourselves, right? There's this drama to this whole thing that doesn't actually make sense when you take a step back. Like if you think you're doing something dangerous, don't do the dangerous thing. You know, don't, it's not your job as a company head to appeal to, you know, it's my job to be China. Like there are policy decisions that need to be made by us as a society that tech executives don't get to make, right? And I think that there are standard policy provisions like liability that could help create firewalls or breaks in this kind of experience, right? Like if your system is hacking other people, what does that mean? I do think there is some nuance and balance here, right? We want the AI system, AI companies to be doing safety checking. We also want the government to help with safety checking. And that safety checking incorporates some risks built into it. But I think that there is this idea that like there is no real responsibility right now for what you do if your system goes wrong. Like there's no cost to you, and as a result everything is very epically scaled. Like will we destroy or save civilization? I mean, there's a lot of things that are not destroying or saving civilization that involve liability that we need to think about and assign to these models, and that's a really big gap that we haven't filled yet.
Support for the show comes from LinkedIn ads. There's no worse feeling than making a major investment in something only to realize it didn't exactly live up to the hype, such as buying a nice piece of tech that ends up in storage collecting dust, or taking a business workshop where your main takeaway was little more than a few motivational words. If you work in marketing, this can happen with ads. You optimize for the numbers that look great, impressions, reach, and reactions. But when they don't show revenue, well, that can turn into an unfun conversation with the CFO. LinkedIn has a word for that, bull spend. Reach the right buyers with LinkedIn ads and invest in what looks good to your CFO. According to the 2026 Dream Data Benchmark report, LinkedIn ads generated the highest ROI of all major ad networks. It's 121%. You can target by company, industry, job title, and more. It's time to cut the bull spend. Advertise on LinkedIn, the network that works for you. Spend $250 in your first campaign on LinkedIn ads and get a $250 credit for the next one. Just go to linkedin.com/scott. That's linkedin.com/scott. Terms and conditions apply.
Support for the show comes from Arca Wealth. It's midnight. You're googling Roth conversions. You didn't work this hard to be doing this yourself. Maybe you already have a financial adviser, but they show up once a year, usually because you called. Meet ARCA Wealth. ARCA is built for real complexity. Tax planning, RSUs, and options. Accounts scattered across three different firms. A business sale, a retirement that's closer than it used to be. One adviser coordinates all of it. ARCA remembers the tax deadline. ARCA knows when your equity vests and when your life actually changes. A new job, an inheritance, and exit. They're already on it, not waiting for you to notice first. You shouldn't have to manage the people managing your money. The luxury was never the wealth itself. It's not having to think about it. Good service remembers. Great service anticipates. Right now, ProfitGPT listeners can book a complimentary financial planning call with an ARCA advisor at arcawealth.com/profitgpt. That's arcawealth.com/profitgpt. Every ARCA adviser is a fiduciary, legally and ethically bound to act in your best interest. Arca Wealth is fee only. So what you pay is transparent. For important disclosures, please visit arcawealth.com. This segment is a paid endorsement for Arca Wealth. I'm not a client and have an incentive to endorse Arca Wealth.
Support for the show comes from Inside Blackstone. AI is changing far more than the tax sector. From healthcare and manufacturing to energy, real estate, and financial services, it's reshaping the entire economy. So, it's time we hear from the people driving these changes and gain insight into what's happening in real time. And Inside Blackstone has the scoop. It's a new podcast from Blackstone where founders, CEOs, and investors, all the people navigating these changes in real time, are talking it out in real, unvarnished conversations. As a listener, you get an inside look at the technologies transforming industries and the big trends shaping the economy. Because the biggest changes shaping our world don't start in the headlines. They start on the ground. Inside the businesses that drive the future. The goal of Inside Blackstone is simple. Engage in conversations right at the source and making complex topics of change and innovation more accessible. Listen to Inside Blackstone for a weekly breakdown of what's happening and gain a real sense of where the world is moving. Tune in every Tuesday and subscribe to Inside Blackstone wherever you get your podcasts.
Let's talk about agents. Meta's Muse recently hit the top of the app store this fall and OpenAI followed with Dots. What's your take on the rise of these personal AI agents?
So I think when I talk to people, I've been, we've been talking to these agents for a while, and it's what's happened. These are all based on a model, basically an underlying idea called Open Claw that came out in January and we made a brief buzz there, which is if you give an AI agent, they're now smart enough, if you give an AI access to your email and to your personal file, it could figure stuff out. You don't need to do like, it understands what's important to you, and then they do, they're quite good at this, right? At least the most advanced AI models are quite good at figuring out what's interesting to you. And almost everyone I talk to who's using these tools, right, find at least one delightful or exciting or interesting moment. I mean, I've had my Muse negotiate a credit card bill for me or, you know, I've had my Dot pointed out that I had made a mistake in an email I sent to a permitting authority and gave the wrong number in my permit. Like, these are useful things, right? They're also quite dodgy because they have access to my email and I'm slightly worried all the time they're going to send out a message to my dean, you know, with private information. So, but they've been remarkably good so far, and I think they're understandable in a way that a chatbot isn't. Right? If I want to use a chatbot or a different kind of agent that's not a personal agent, I have to figure out what I want to tell it to do. And a lot of people fail at that. The fact that these systems are proactive, that they're sort of personal to you, I think makes them very appealing. We don't know how sticky they are yet. We don't know how many errors they're going to cause. But they're already a sign of a lot of disruption that's going to be happening. Things like customer service. Companies do not have any idea how to deal with this. Like it's now cheap for me to do all of the customer service calls I would never do, my just have the AI do them for me. And it cost me nothing, right? Negotiate this bill, get into an argument, you know, cancel the subscription. And the companies of the world are not used to a world where you could just do that on demand. I think we're in for a lot of really rapid change as a result.
Yeah, it definitely feels like sort of AI-enabled agents on both ends. I heard about a study where a family law practice used AI to come to settlements and then had lawyers, you know, duke it out, and they found that the bots or the agents speaking to each other came up with remarkably similar endpoints in terms of an agreement. Can you think of specific sectors where you'll be able, where it just makes sense for agents on both sides, the demand and the supply side, to come to some sort of... I travel a ton. I stay in a lot of hotels. I spend a lot of money. I anticipate at some point I'll have an agent that'll go out to the universe of hotels that I like to stay at and negotiate a multi-night deal for me rather than I could do it. I could call and say, and occasionally have done it, I want a corporate deal. I'm here 30 nights a year. It's a pain in the ass. It feels like at some point my agent will be reaching out to the, you know, name your hotel brand here, agent and seeing if there's a deal to be done. A, do you think that's going to happen? I assume the answer will be yes. B, when do you think it's going to happen? And which sectors do you think we'll have more like agent to agent as opposed to human to human negotiation?
So, you can absolutely do that now. I mean, you know, the dodgy part is the voice system. So, everyone is sort of working on how to get the voice calls working. I think Meta right now is partially faking it with humans making the calls, but the systems are quite good at voice. But if you're doing email exchanges, there is no doubt that a dot-class model, you know, agent could actually just do that for you right now. Exchange multiple emails over time, attach PDFs, engage in forms, right? So the call is the missing piece. You absolutely can do that with current systems. And I think that they would support that enough to make that work, given my experience with them and what I've seen. What happens though is as adoption spreads, what happens when every hotel is getting these messages? The only defense they have is agents of their own. This is a sort of mass denial of service attack in some ways, but authorized by you. Like you will be pissed off if your agent isn't allowed to talk to a customer service person because it's your representative, right? It's your representative. It's not the same thing as a mass emailing, right? This is my representative doing something on my behalf. Of course, you have to answer it. So the only way companies will be able to respond is probably with agents of their own. And I think that we're going to end up having negotiations between agents very quickly whether we like it or not. I think on the fairness issue, that becomes a big issue throughout the entire economy, right? A lot of how the economy works is around imperfect information and friction. So, Apollo Research had pointed out that a lot of how the banking system operates is that we all put our money in savings accounts that are lower yield than we should get because unlike a large company, we don't pay people to sweep our money into higher interest savings accounts at a regular basis. So, our money sits at banks which loan that money out, but agents make it very cheap. I mean, I wouldn't trust them with your money right now, but it makes it very cheap in the near future to take your money and move it from one account to another. And that would actually at some scale would actually start to threaten the banking system as a whole. So I think we have a lot of systems built around human frictions and human difficulties that are about to start collapsing pretty quickly.
Let's talk a little bit about higher education. You teach at Wharton, and what has your experience been with seeing how your students have been using AI?
So you know, I introduced my class to ChatGPT a few days after it came out. You know, I'd been using the earlier versions of GPT-3 before ChatGPT, and I, you know, the first, I teach an entrepreneurship class. So in the very first class, one of my students tuned out entirely and by the end of class had built their entire product that they were planning on doing for the rest of the semester. Right? So that was sort of the first indication. But AI was bad enough at that point, it required enough human guidance that it was like use AI all you want, but it's not at the level of an MBA student at Wharton. And over time, that hasn't always been true, right? It's still jagged, but now AI is quite good. So I can't just say use AI. So there's a few things you have to do, right? One of the things I do is I use AI-powered teaching all the time. I have AI simulations that they have to do. There's AI case studies they have to create with an AI to, you know, to give me, they have to teach an AI that is programmed to be kind of dumb about a topic and they have to be a teacher, which is a really good educational technique. And then they build more than they ever did before. They're launching companies where they used to come with an idea. At the same time, I now have to make in-class assessments matter more, right? I can't check what's AI work and what isn't AI work anymore. So in class, we have to do things like blue book assignments and we have to do things like in-class discussions that matter. So it's changing a lot all at once, and I mean I'm eager to do these things. Not all my colleagues are, nor should they be forced to, or in a just world, to transform their classes this quickly. So the experience is quite uneven across the board, depending on you know who's teaching and what kind of subject under what conditions.
A lot of people have said that higher ed is going to implode at the hands of AI. What are your thoughts?
I actually believe the opposite, for a couple of reasons. You know, for actually for a wide variety of reasons, right? So one reason is, what's actually about to implode are internships and all the ways we actually teach people to do work today, which is informal education where you are a junior person who wants to impress somebody and a senior person teaches you because they have no choice, right, to give you feedback because there's no one else to do the work. Now every junior person is having AI do all
The work and they learn nothing. Every middle manager would rather turn to AI to do the work than interns. They're not teaching anyone, and companies are not incentivized to teach. They want to produce output. So guess who's good at teaching, assessment, and verification? That is us at universities. Like, look, things are screwed up now. They're always screwed up at universities, but it's pretty obvious the path forward: AI tutors outside of class, inside of class, active learning, assessment. We'll figure it out. It's not the end of the world. So that's one reason you'll need to turn to us more. The second is, I think, why knowledge matters. I think knowing many things matters more than ever before, and you need to be taught those things. And the AI people have the wrong incentives of AI to learn. We can teach them those kind of things. And then in a world where there might be more leisure — like, let's say we end up with the heaven on earth that the AI companies are planning for us — I mean, you know, I think gentlemen and gentlewomen scholars become a much more valid thing, and this becomes a thing you do anyway. But as far as the real economy, teaching, learning, and assessment and verification are all things we do in universities. We may have drifted from the mission sometimes, but that is a mission we can return to quite easily. And I think an important one.
