Wharton's top AI professor: stop launching software startups — the money is here

We can easily stop using certain skills by relying on AI and perhaps never recover them.

You don't want AI making those decisions, even if they're good ones. You want to be part of that process. I'm deciding when to exercise my own agency in the system instead of just sending an agent to do the work.

The adaptation part has to happen from scratch every two weeks. That's what I'm feeling.

So we have to deliberately add friction to our lives in certain key areas to maintain an advantage.

I wouldn't say never delegate.

That's always a tough question, right? Ambition seems like what we call entrepreneurial efficacy in the entrepreneurship world, right? The same kind of thing, agency.

And it's actually really hard to do. We don't exactly have a manual for it, but I really think people create barriers that you can help them overcome by getting them — just by asking them to be more ambitious — into a room where their peers are more ambitious. So a lot of what I ask, and it's pretty easy now, is like: "What else could you do beyond this?" And that used to be a hard question before AI, because you'd stop, because you only had the capacity to do so much. And now, as the frontier expands, encouraging people to be more ambitious is much more important.

It's funny because the last edition of the course I taught was six months ago, and that was already the beginning of the agent work. My next semester starts in January.

So actually, after ChatGPT was launched — I'd been tracking the models for a while, obviously — but after DALL-E 2 came out just a few days before ChatGPT, and ChatGPT was launched by OpenAI, so a few days later I had a class and basically dropped all the planned content and said: "I'm going to show you how to use ChatGPT." And at the end of class, one of my students had stopped paying attention and programmed his entire semester project during class, and showed me the results at the end. And so, every year since then, the level of ambition has kept rising. What was most interesting this time was all these people who were experts in their fields. I have surgeons, biomedical device specialists, finance people, and, you know, NGO and humanitarian aid folks, soldiers, all in this class. And suddenly they were all building real things, you know? Like a go-for-me objective. So very interesting ideas came up. How do we track our entire supply chain? Because many companies don't actually know their whole supply chain. And how do we find where the financial weakness is in that chain that maybe we need to monitor? A really interesting set of ideas. There were several very interesting things about specific problems in medicine, like how to actually convey information to patients in ways they understand, rather than just forms.

So, I mean, really interesting things, but quite specific, you know? Focused on something people know. And part of what's interesting is, like, software can be many things.

And if you're trying to solve a problem you have deep experience in, suddenly a lot of resources appear there that are really exciting.

Among those of us who teach entrepreneurship at some of the top business schools, we, you know, used to swap ideas about the topics we see people tackle repeatedly that we know will never work, but we always conflicted over it. Should we tell them it won't work or let them try and find out for themselves? So there are some that are kind of standard problems, which are actually a good example of where technology hits its limit. One of the most common is: "We're going to replace real estate agents," right? It seems like something huge, like you're paying about 2% to these people who are doing less and less work for you. This is obviously a case of disintermediation, and every time the idea fails, because at the end of the day there's deep social integration, political integration, and legal integration of the agents in the process that you simply can't remove. So sometimes the reality of the world is quite different from what you think. And the other is students thinking very narrowly about their student problems. So everyone always wants, you know, better ways to get home-cooked food or better ways to, you know, store things between semesters that probably aren't that universal once you leave college.

Well, I don't know. You just try, you tell people, right? Like, these are the ideas with the highest failure rate, but obviously because they want to do that, right? Because they keep coming up. They're not bad ideas. The whole thing is you can't filter them out right at the start. Like, there should be a better way to buy and sell a house, right? And, you know, people want home-cooked food more easily than a meal delivery service, which is kind of what they're asking for now. It's like, how do I get my mom to cook a meal for me as a son or a father? That's basically what they're asking. So there should be a way to solve that. There's demand for it, but they're harder problems than people think. And in a way, AI makes them more ambitious, right? Because it's no longer about the technical part, but the numbers look really good and then they reach a point where they think: "There's a challenge here." So whoever solves this will make a lot of money. But it's also an idea where we've all seen people try to solve it repeatedly. So maybe some of the people watching or listening will be the ones to break the paradigm of the real estate agent experience, but it's not as easy as it seems, I think.

Overall, like, I mean, I think if you're trying to build a business from scratch, right? Aim high, because you might not get there anyway, right? And ambition also helps motivate people. It makes the best people work for you. It creates a real self-fulfilling prophecy, right? But I also think there are phases in a business, right? So that initial phase can be difficult anyway. Scaling presents its own challenges compared to running a small business, right? So they're not really equivalent things, right? Like trying to get to the next level. If you're going to raise a Series B and a Series C, that's very different from running, you know, like a, you know, a local business of some kind, right? So I think they're different skill sets.

Yes. Yes. I mean, if you're starting from scratch, right, and you want to conquer the world, be as ambitious as possible. And but I also, by the way, think that, you know, we keep talking about ideas. I think people aren't being ambitious enough about the types of ideas they pursue, in part because there's a kind of Silicon Valley mindset about software, right? And I think a lot about — I studied the history of technology, among other things, and worked on it. And if you look at who actually got rich during the Industrial Revolution, right, it wasn't the people selling steam engines, although that remains the analogy in Silicon Valley, like: "I want to sell shovels and pickaxes." You don't want to sell shovels and pickaxes? You want to dig a gold mine with a new mining technology? And I feel like we're kind of missing that point here. If you can make the business more efficient, just do it yourself right now, right? Overcome those barriers directly.

So I think, you know, it's good news for entrepreneurship in general. Like, the early data looks really good and you can understand why, it's a very good idea. I think everyone's mind is a bit too focused on SaaS at this point, right? Which is like, what do you do to make money? Software as a service. And I think that's under threat, right? Like, it's a different kind of environment than it was before. I spend a lot of time with big companies now, as well as with entrepreneurs. And there are people who get it. There's a lot of conflict about what the future of software will look like. On the other hand, I think software is becoming more and more a basic requirement for doing a lot of other interesting work. So I would love to see more entrepreneurs expand beyond software as the product they sell to actually create or make something, because there has never been a better time for it.

So it's kind of interesting, right? A lot of things are happening at the same time. On one side, everyone became a programmer. Now I say, and this isn't just a theoretical thing, right? There's this incredible series of studies from Anthropic suggesting that, actually, managers now write better code than programmers, at least in the first versions, right? Because it's more about giving clear instructions than being great at programming. And things are changing really fast, right? You have people 100 times more productive, so companies are also a complete mess as a result, but they all have their sights set on, you know, the SaaS vendors, right? There isn't a big company out there that doesn't want to think about how we become more independent, how we build our own tools, and so it will take a while for them to get all that right, and many of them are still very happy. I mean, you know, as far as I know, both OpenAI and Anthropic still use Slack. They, you know, Salesforce, they probably still use Workday. So, like, there will be more choices being made about which type of software vendor to work with, but I think the most interesting thing is that companies are realizing that many of these startups that are selling things to them are just selling things they could already build themselves, right? They don't have their own LLMs. So they're selling LLM capabilities they already have. And that's creating a really interesting dynamic.

I think software will be democratized, right? So it's like everything — I teach at Wharton, I teach MBA students, you know, I teach undergrads, and many of the MBA students aren't technical. So, like, the key to success in, you know, the last 10 years has been finding a technical person who can program for you. I find it incredibly liberating to see this, as happens in other areas, where you can approach a question in different ways, you know? There will be software-focused companies where you'll really need the best engineers working with AI to succeed. And then there will be something like software as a toolkit for getting things done. And I think democratizing that is incredibly useful, right? I mean, you don't want to be a software company, you want to do your newsletters, you want to do your videos, and, you know, there were things in your way, and software, by the way, was one of them, right? Also, you needed to be able to handle your own finances. I mean, when I started my first startup, I didn't realize it was the late '90s, the first internet bubble, but I didn't realize you could pay someone a few cents per payroll to handle your employees' taxes. So I calculated the taxes in Excel every week, in the first weeks, without really realizing that you could just do that.

So even just advice, all this kind of help you get now makes everything much easier than it ever was.

Yes, it was an interesting study, it was with GPT-4, right, a system that's already very obsolete. Now there are 180 models better than GPT-4, 19 of which can run on your phone. So, you know, a big difference. And it was only giving advice. So the fact that I think the number was an 18% increase in profits for the best entrepreneurs, like, that's an absurd number. We run studies all the time, trying to make a difference in entrepreneurship. That's an absurd number, but it only helped the entrepreneurs who were already succeeding, because it helped them expand the way you're talking about. The entrepreneurs who were already failing couldn't implement the advice. They were already going down the hole and there was no way back.

So it was — so we know these things can be effective, and that was before programming, right? That was literally: what should I do now? And in entrepreneurship, like, it turns out that, by the way, about 50% of Americans have had a good idea or think they've had a good idea for a business in the last 5 years. And almost nobody follows through because of all these barriers. So just the fact of having guidance to overcome those barriers already makes this a true golden age for entrepreneurship compared to what we had before.

I talk about this in the new book, but I think there are four things needed to use AI successfully. And that is deep knowledge, right? Certain experience in something. And that is, by the way, how entrepreneurship boils down to systematically developing specialized knowledge, right? You go into an area, you interview people, you, you know, learn about the industry, you test ideas. I mean, that's building experience, and I think that's as important as it's always been, right? Otherwise, you're just sharing Claude's ideas. And Claude's ideas all look very similar to each other, right? The second thing is broad knowledge. Increasingly important for entrepreneurship is knowing a little about the areas you don't operate in, right? So if you work in software, you should know a little about design and a little about accounting, because that helps you apply AI more broadly. And then autonomy, which we've been discussing a bit here. And finally, some sense of good taste — how to pick good ideas out of that pile of garbage — is also very useful. So I think what we're seeing with more agentic systems is that they can run for longer periods of time, right? While GPT-4 could work for 4 minutes independently with a 50% success rate, now we're well past 17 hours of work. But you still need to be able to guide these systems. And even when they get really good, it turns out the human touch really makes the difference, because variation is what makes entrepreneurs succeed.

Yeah, I mean, it's kind of funny because you realize people are more important than we thought, or than some people thought, right? So everyone who started using Claude early as a good writer back then, you know, when, you know, Opus 4.5 or whatever it was, they wrote really well and for a while nobody knew what Claude's writing looked like. And now if you open LinkedIn or X, it's simply nauseating, right? It's like, what's sustaining this? Wait, what is it? It's simply all the same thing. And it's not just the stock phrases, like the way it approaches things, the type of language it uses. And you think, well, they should — if you didn't hire a marketing professional, you're probably making a mistake. Now, that professional could produce much more than before, but each professional's own perspective makes a big difference. And I think we're seeing that in many other cases, where AI pushes people toward a kind of consensus strategy or vision. We have some experiments where we found we can make ideas almost as diverse as a group of people's. But if you map the ideas AIs create and map the ideas a group of humans creates, AI still doesn't cover the full range of human ideas, even if you push it to the max. So that part of variation is going to become increasingly important, like what you're adding to the system to guide it or direct it, so it doesn't just stagnate on everything. And then we'll be fine, like, it's fine if you're using Claude for finance and that's not your competitive advantage, but in the area you're specifically focused on, you need to stand out, you need to differentiate, or you need a hard-to-copy competitive advantage that has to be human to some extent, even as the models become incredibly capable.

We know a little about how to teach good taste to people, right? Like, you need to see good and bad examples of things. So it's good when people explain to you how something works. So I — one of the — I bring an art professor to my classes now who talks about thinking about how a designer sees the world, you know? So you have that kind of perspective, like what makes a font good or bad, what makes an image interesting or not. And that's why I think we need a language for discussing issues of taste, you know? If you talk to someone who really knows cinema, that person has their own taste in film, but they can probably talk quite a bit about it.

So I think seeing a lot of variation is really important. Learning to have discipline about it is another example of why I think education has some value here. Being exposed to many options, I mean, helps you develop the aesthetic sense you need.

I don't like the idea of judgment by itself, because AI obviously makes judgments. You can't carry out a 17-hour task without AI making decisions that require discernment, just as we didn't include creativity on the list, because AI obviously makes creative decisions. The question is whether they match the type of judgment and creative decisions you would make, right? So we can call that good taste or knowledge, but it's a different kind of question, right? So judgment is this hard thing. We don't really know how to teach it and we don't have a good example of it. But some people — you know, the question becomes in what context this judgment is happening. So it's a — it's a hard thing to ask, right? Like, what is judgment? And is AI exercising some kind? Obviously there's an element of human judgment that's different, but we haven't had to build the language to differentiate these things. I kind of think about what's happening in math now, right, which is this kind of explosion of AI-generated proofs. And one of the things mathematicians will tell you when you talk to them is a bunch of revelations about their area, or maybe clarifications about their field, that outsiders didn't know, right? There's an aesthetic element in math that really matters, right, to a mathematician, that if the proof doesn't help you learn something beautiful or interesting, it has no value to the field of mathematics. Even if we've solved the problem, who cares? And like, we're starting to differentiate things we didn't need to differentiate before, because the only type of judgment and intelligence we have is human. So we didn't need to create a different category.

A few things happening at the same time when we think about market and education, you know? First of all, we're temporarily in crisis in universities. I mean, we're always in crisis in universities. So that's no news. Anyone who's worked in education knows this, but we're in a temporary crisis of AI, right? And we more or less know the outlines of that solution, which is the fact that AI is an incredibly good tutor when used as such. And it's very bad if you're just copying and taking its answers, even if you think you're learning, you're not, right? So in the end, we'll have AI tutors outside the classroom. Inside the classroom, we'll do more hands-on learning. I mean, if you've been to a business school or law school, you've done this kind of case study, discussion, and activity. And, you know, we'll do in-class tests. We'll figure it out. The real problem is where people actually developed good taste and discernment. It's in the real world, right? And we had this great deal for the last thousands of years, where I would send you, you know? A freshly trained recruit who knew nothing, and we had apprenticeship with experienced professionals, right? That's how the last 4,000 years of training office professionals worked. So you'd get a junior person who was desperate to show they were good? They worked for little because they wanted to climb the hierarchy, the ranks, and did work a middle manager didn't want to do, and in return received a small salary. They showed what they were capable of and learned how things worked. And the middle manager passed their work to someone who could at least do part of it for them and make their life easier, right? And that's how they climbed the management chain. And all of that collapsed, right? Because each — you know, your incentives as an intern or a first-year associate or whatever you are, your incentives are 100% to deliver the best possible product. Who's better than you, Claude, ChatGPT, Gemini? So you're going to deliver answers from Claude, ChatGPT, or Gemini. And this isn't like being a digital native, because you don't know what you're delivering. You're just passing things along. And every middle manager realized: "Why would I give feedback to a guy who takes forever to do the work? I could just ask Claude to do it myself." So we've just broken the entire system that forms professional quality sense and discernment, and we haven't found a complete substitute yet.

So I think there are a few things. One is that at the individual level now we have to start being very disciplined about when we're going through cognitive surrender, when we're letting AI do tasks for us. And that's very hard to do. We are made — like, first of all, entrepreneurship, like, which is what I teach, right, what we've been talking about here, is about trying to make your life as easy as possible so you can focus on the hard things, right? And you know, having the carpal tunnel — but then you don't learn anything, right, as a result of that, having the carpal tunnel — but that also applies to any other aspect of life. Like, we can easily let AI do things we know how to do. And I think we all feel tempted to do that every day, and perhaps never recover those skills. So we have to, on purpose, add friction to our lives in certain key areas to maintain an advantage, but we're not incentivized to do it. So at the individual level it's going to be a matter of making choices. At the organizational level, companies will do one of two things. They'll take the easy path or the hard path. And I worry about the easy path, which is hiring only senior people. And we're already seeing some signs of this. I know you've already brought some economists to talk about this. There are some early signs that junior-level jobs may be threatened. We're not 100% sure yet, as you know. Or the other option, the harder one, is that we actually start taking training seriously as part of the job, right? So you really have to give people time to be inefficient and learn on the job. And we, instead of learning being something kind of accidental, will have to teach them. Maybe you really need to add another year to university education to teach you to be, you know, a Google person or a McKinsey person or whatever it is. And there's a year of personalized training to teach you, because we know how to teach and evaluate people in the classroom, something we no longer know how to do in the real world.

So it's still too early to say anything, right?

Yes, I mean, but also for how long good enough is good enough, right? Like, the problem is, right? I mean, that's also progressing. So, you know, this concept we developed in our first paper, where we examined this, we called it the jagged frontier, that AI is good at some things you wouldn't expect and bad at some things you wouldn't expect. And it's hard to know in advance what's what. As long as the frontier remains fairly regular, there will always be bottlenecks that people still need to solve. And those bottlenecks can be pretty strange, like if you think AI replaces the lawyer. Well, it turns out a lawyer performs many tasks that involve real-world obstacles that AI can't necessarily do, even if you had a robot doing the work, right? Because there are human systems they interact with, there are legal requirements they need to deal with. Some people want to talk to a human. So this creates bottlenecks that end up rewarding different talents than before. And you see this happening in programming too, right? Like, what makes someone a good programmer today is completely different from three years ago, because you don't write code anymore. So I think one thing I'd say is being good enough is a set of skills. And it may well be that people who are average programmers but have a great vision of how to develop software — or communication. Exactly. Yeah, or communication. And that will also change over time, right? I'm not going to say AI will never do that task. It probably will.

So I think we'll realize that the valued skills change rapidly, without an easy answer.

Well, one of my favorite phenomena is seeing people who know nothing about an area do impossible things, and you're like: "Wait, how did you do that?" And the person: "I just asked." And you can't ask AI for that. And they asked and it just works. And you're like: "Great, right? What are we doing here?" So, I mean, there's this kind of strange phenomenon. I mean, there are two things we know make you better at AI. Before, AI was mainly a leveler, right? If you had poor performance, you got a little boost. It still is. But what we're seeing more and more, and there are some good results from Anthropic, as well as other studies, suggesting that people with more experience get the biggest gains from AI at this stage. And that wasn't true until agents kind of emerged in the last year or so. But the other thing we know is that there's a certain skill in knowing how to work with AI, which is knowing how AI thinks, which can be a kind of theory of mind. Like, if I know people really well and I come to know the mind of Claude, ChatGPT, Gemini, or whatever I'm using, I become more effective working with it. And so these are kind of two different skill sets. And then there's a third mysterious category, which is like, some people seem to be good at it in ways we don't understand. There's this great study by Alex Simas and some others. He's now in Chicago — now he's at DeepMind — but they evaluated people who created negotiation agents and some interesting things. One thing they discovered is that the multiplier of your skills becomes much bigger because the agent engages in many negotiations. So a small difference in skill at creating good agents has a huge impact on the final outcome. The second thing they found was that people who are good at creating agents aren't necessarily the ones who would be the best negotiators in real life. I mean, there was even a gender divide. One of the things we repeatedly find in studies, for a wide variety of reasons, is that women tend to perform worse in negotiations. And there's been a lot of work on this. It's been a major area of concern, but women created much better negotiation agents than men. And, you know, and again, we really don't know, like, I don't even know how to teach you to do this necessarily.

I feel like that's kind of made up. I mean, it might be true. For me, the analogy that makes the most sense for someone who has led organizations and managed people is to think of AI as a team I'm delegating work to. And if I can do a good job delegating to a team and noticing where they go wrong, which is very different from knowing my own process, right? AI isn't going to replicate my process. And part of what I try to do is preserve things for myself that way too, you know? Like, I write everything. I mean, it might seem crazy, since I use Substack and X, but I mean, I write all my own things, right? I write the first drafts of everything myself. I say: "Yes, I type everything, you know? Because if I don't do that, AI is very seductive and will definitely..."

type of master my thoughts in any direction." And then you'll end up reading like Claude's answers filtered through me. Do you like that?