The theme of today's conference is change drives growth, and there are few forces of change more powerful than artificial intelligence. It raises some of the biggest questions we face as long-term investors: What will it become capable of? How will businesses adapt? And where will durable value accrue? These are questions we debate constantly at Scottish Mortgage, and they're the focus of today's conversation.
I'm delighted to be joined by Ethan Mollick. Ethan is a professor at Wharton Business School, where he co-directs the generative AI labs, and he's the best-selling author of Co-Intelligence. His next book, Coexistence, is out in October. Ethan, welcome and thank you for joining us.
Perhaps we could start off by using your books as the framing. Three and a half years has passed between the two of them. That's not a long time in publishing, but it's a very long time in the world of AI, given how fast technological progress is. We've really gone through three successive eras of AI in that time. We've gone from chat AI to reasoning and now to agents. So I'd just be really curious to get your thoughts on what have been the biggest surprises over that three-year period between your two books in the world of AI to you.
So I think actually, as fast as I thought AI would develop, it's kind of gone faster than I was expecting. So we're at a place where I thought human development would take a while, right? It would take a while for people to adopt AI. And I think even that's been a little faster than I was expecting, even with my cynical business school professor hat on. But I think that the technological change has been very rapid. So when I wrote the book, we didn't have reason models. We didn't have these sort of advanced approaches. And I was regularly making mistakes on simple math or even counting the number of Rs in the word strawberry. And now, as of a week ago, it's solving millennial problems, right? And math is in a crisis. And I think it hasn't advanced everywhere in the same way. We have this concept in our research called the jagged frontier, that AI is good at some things you wouldn't expect, bad at some things you wouldn't expect. But it is advancing quite quickly in so many of these areas.
And despite that incredible progress, are there any areas that have perhaps surprised on the downside to you in terms of where you thought progress would be faster in a particular vector than it's turned out to be the case?
Well, because of this jaggedness, you can point to particular areas that were like, I thought this would advance more quickly. And one of those is writing. The AI is actually quite a good writer, but it isn't a good writer at producing text you'd want to read on a regular basis. It feels like one person kind of doing the text. And even plotting stories or long-form documents that are non-fiction or long reports, it tends to kind of lose the plot and get confused about who to talk to, even a really advanced model. So there are some areas that are still jagged.
And with the jagged frontier concept from your research, are there any areas or commonalities where AI seems to persistently struggle? Because I know there's this element of sometimes you have quite large breakthroughs in a particular area, but are there any areas that you think with a degree of confidence it will continue to struggle with as we go forward into the future?
So interestingly, I think it's sort of a second-order thing that we'll keep struggling at, which is that AI is really good at many tasks, but it's kind of like having one very smart person do the task. And in a lot of cases, we actually want variety. We want many people trying to address something. So diversity of ideas, for example, we have this research showing that AI can generate much more diverse ideas with the right prompting, but it's still less diverse than a group of humans who have a higher coverage area. And you're going to see the same sort of problem anywhere where diversity matters, including strategies, investment decisions where there's not necessarily one right answer, but you want a multitude of perspectives being brought to play. And I think that diversity piece is an area where we see AI struggle and continue to struggle.
As an organization or a user of AI, can you limit that issue by actually just feeding out your questions to multiple models and sort of recreate some of that diversity? Or is there still a limit or a friction point given the similarity of data that they're often trained on?
Yeah, I mean, we don't know the reasons why the AIs are deeply similar. I mean, there's a set of studies showing that there's remarkably similar political attitudes among Chinese models, Elon Musk's Grok models, and Anthropic and OpenAI's models. And we don't 100% know why. It could be training data. It could be reinforcement learning. It could be some Platonic thing happening that we can't pay attention to. So you absolutely can increase the diversity. I recommend using councils of AIs for decisions. They do find things prompting for personality. You're a critic. You're supportive. That can help. But there's still some innate similarities among these approaches that are hard to separate out.
And so most of us now be using AI probably both in our personal life and our work life. And for us as investors, we're encountering and looking at a lot of companies that are absorbing AI, that are using it as part of their organization. I suppose there's one argument that would say that actually all of these companies have access to the same models at roughly the same price. And a lot of the time when someone has access to the same tool at the same price, they end up getting the same sort of outcomes. I'd be interested from all of your research when it comes to organizations, do you see them as encountering AI as a democratic tool that makes them all equally better, or do you see it as an amplifier where some organizations absorb AI really well and get much better outputs than everyone else in their field?
So I think it's worth separating out the tool from how it's adopted inside organizations, right? And at the individual level or the group level. So at the individual level, individual adoption is very mixed, right? Some people are big adopters, some are not. Our early research at BCG found AI acts as a leveler. It took the lowest performers and leveled up their performance. More recent research suggests that people who are already experts get an even higher return to AI. So it's not necessarily even at the individual level. As they start to add their own intelligence, the kinds of questions they ask and how they guide AI, they get better or worse responses that are different from other people's. At the organizational level, the results are even more extreme. I think people tend to think about data being what differentiates them from one company to another, but it's actually process and how you integrate AI into your system and how you work humans into that process and who makes decisions. And out of that comes a lot of differences that means that your company is not like all clawed or all not clawed. It's some supplemental organization. And so I actually see quite large differences emerging between companies as they do these things based on how they're adopting and how smart they're adopting, and especially how they start to think about many people at once using these systems rather than individuals using them on their own.
And when we're talking to companies, obviously a bias that every company now wants to present themselves as being an AI-first company, that this is a tool that they are mastering and they are getting good outputs from. What type of questions would you probe an organization to sort of test whether or not they're interacting and leveraging AI in the right way?
So I actually have four questions I tend to ask companies. So the first is, what have you stopped doing as a result of AI? And if the answer is you haven't stopped doing anything, you know, something's wrong, right? I mean, you produce market research reports on demand. That may or may not be valuable to you, but you were probably paying for those before. And maybe it's worth thinking about, do we change something about how we're approaching market research? But if you haven't stopped doing anything, if there isn't anything that creates a competitive advantage for you that it no longer does because of AI and you haven't had change, that's a bad sign. A second is I ask companies, are they doing anything impossible? These systems allow you to accomplish goals you couldn't accomplish before. You should have at least one effort that you think would transform the industry. You should also be building stuff that doesn't work and betting on this curve getting better over time. And then the last piece is my model for how I have companies think about AI is they need three things: leadership, lab, and crowd. So they need the leadership team that is really bought into AI but is also making decisions about reward systems and how they want to change their companies. You need the crowd. You need people with access to the best tools experimenting with use cases because nobody knows the best use cases. And you need a lab. You need people doing 24/7 AI development work that's not just technical, that's also business and organizational. So if you don't have a leadership, lab, and crowd structure, you probably need to fix that too.
That's helpful. And then I suspect a lot of people listening would really like to get your thoughts on AI disruptions of companies and categories. Certainly the stock market has focused a lot this year on potentially the victims of AI that might be disrupted, and a lot of that's focused on SaaS businesses, software as a service, both for that specifically but also more broadly. How do you think about what companies or functions are likely to be disrupted by AI becoming more abundant?
