Ethan Mollick is probably the single best explainer of what AI models can do for people—at work, at leisure, or in the messy space in between. More than an everyman’s Virgil into the weird underworld of AI capabilities, he’s also excellent at explaining how not to use AI tools. As the media theorist Marshall McLuhan once said, every technology is both an “extension” and an “amputation.” A machine that extends human intelligence is also a machine that cuts off the need for humans to be intelligent. When in possession of a tool that can reduce struggle and friction, we have to ask ourselves: What is the value of friction? When is the struggle a problem, and when is the struggle the point?
In today’s conversation, we talk about:
why so many experts keep underrating the rate of progress for model capabilities
why so many experts keep overrating how rapidly powerful AI will change the physical world and the macroeconomy
garbage cans
how AI is coming for the hard parts of being human
the sociology of AI swarms
what happens when personal AI agents get so good at prompting us—e.g., “Hey, it’s time to order food for the week. Same as last week, or do you want to try a new protein?”—that the age of prompting AI becomes obsolete?
2027 predictions, the next phase of AI, and his excellent new book Co-Existence
Derek Thompson: Your book opens with forecasters trying to predict the future of artificial intelligence and math, and they gave AI a 2.3 percent chance of winning Math Olympiad gold by 2025. Last year’s superforecasters gave a Millennium Prize problem a 1.7 percent chance of being solved. Now OpenAI has said it has already solved one. Starting with this forecasting failure to anticipate how powerful AI would be even in just a few months, what does that teach us about the capabilities of AI?
Ethan Mollick: I think it teaches us a lesson that feels like people have heard since COVID, which is humans don’t think in exponentials very well, no matter how hard we try. Even the world’s best forecasters, and they are really good, it’s a straight line on a logarithmic graph. The ability level has not actually varied very much in the last four years. It’s ticked up a little bit. And I think it’s really hard for us to grasp what that means. It’s better on a bunch of dimensions we’ve never had to measure before. It’s not horsepower, it’s not speed, it’s intelligence, whatever that is. I think that’s a lot of the failure I see constantly, an assumption that the future looks like the past, and that’s not looking true anymore.
Thompson: It seems to me that forecasters have been wrong in two directions. One group has underrated AI’s growing capabilities, for example in solving complicated math problems. But another group has overrated the real-world consequences. The unemployment rate looks basically normal. It’s hard to see AI in the productivity numbers. And outside the chip and memory companies, a lot of the economy is saying we’re not necessarily witnessing enormous ROI from all this AI. So if AI’s capabilities are growing so fast, why are the real-world consequences moving so slowly?
Mollick: This is where I’ll argue. I’m surprised how fast it is. I am a professor at a business school, an economic sociologist, and the lesson you learn is entrepreneurs should never win because all of the resources, all of the money, all of the customers are with large companies. Why do they win? Because big companies can’t change. So the fact that we’re actually seeing in surveys high levels of ROI already, even if that hasn’t been translated into enough to move the needle for a lot of these large companies, but they are seeing positive gains over a very short period of time and some of them are starting to reconfigure how they work. To me, that’s fast, not slow.
The current phase of AI we’re in is single-player. It is about individuals getting performance gains, and you can see that. I can vibe code an entire game overnight. I can write academic papers faster than ever before. But when I start writing to systems, multiplayer, many people working together, there are still huge gaps. How does that even work? You’ll see in coding, a coder will be a hundred times more productive, but they’re still in a two-week-long sprint that requires input from people every two weeks. You’re going to have huge amounts of token spend and you’re not going to see an impact. I don’t think that is a long-term persistent problem, but it is a very obvious short-term problem.
Thompson: You’ve argued that jobs are like O-rings and companies are like garbage cans. Why is the economy made up of garbage cans and O-rings?
Mollick: We have two tragedies here in different ways. The O-ring is named after the space shuttle Challenger, where 16,000 parts or whatever worked great that day in 1986, but the O-ring failed and caused a catastrophe. I talk about the jagged frontier a lot in this book and the last book, a term we coined in our papers. It’s the idea that AI is good at some stuff you wouldn’t expect, bad at some stuff you wouldn’t expect. It only takes one or two places for AI to be bad at a job for you to still need a human to sub in because you can’t have everything work great and then one thing fail at most jobs. It’s not additive, where little bits add and subtract from each other. So the O-ring means that jobs are not easily replaced one for another with AI because if AI can’t do human tasks exactly, you still have value accruing to human work.
You see that happening even in coding, where AI now writes code, but you still need human engineers to manage it. Maybe that will change in the next couple of months, and then we’ll see another part of the O-ring get stuck. Jobs are evolving very rapidly, but there are still these sticking points.
Then the garbage-can model is one of the more depressing models of organizational theory, which is that companies are bundles of problems and solutions all looking for each other, piled on top of each other. There is no plan, or at least the plan is only a small part of how organizations work. Organizations are full of contingencies and evolutionary dead ends, and there are loops of work that no one even knows exist anymore. Everything is tossed together, like waste in a garbage can, and sometimes it works out and sometimes it doesn’t. That makes the job of redesigning organizations to work with AI more complicated than just, “We’ll supplement this task,” because that task on paper doesn’t represent the actual garbage can of the company itself.
Thompson: I think the most important thing that’s happened in the last few weeks has been the release and success of Muse, the Meta agent. What is Muse teaching us?
Mollick: For those who don’t know, Muse is a personal AI agent. OpenAI has one called Dot. Lots of other ones are going to be coming out in the near future. They’re all based on this idea from early last year called OpenClaw, which was the same kind of proactive AI assistant.
They’re only capable now because the models have gotten good enough that you can throw undifferentiated piles of information at them and they are really good at figuring out how humans think and what the problems are. It turns out the current models are good enough that if you hook them up to your Gmail, they figure out what’s going on. You don’t need to do a lot of work behind the scenes.
They solved two big problems with AI. One is, when faced with a box that you could ask to do anything, nobody knows what to type in the box. Even if AI could do that stuff, people didn’t know what to type to make things happen. The second was: How do we get AI to be proactive and solve problems for us? Muse did both by basically saying, “People have problems. Look through your email and you’ll find problems to solve,” and then offering to solve those problems.
I’ve had Muse and Dot. I was locked in some subscription fee on American Airlines. I couldn’t figure out what it was. The AI figured it out, joined a chat with American Airlines, canceled the thing, then had to fill out a secondary form and got that back to me. My Dot, or a precursor version I had hacked together, messaged me that I had sent the wrong tax number on a form for a permit in my township. It checks whether I’m going to be on the right call.
There’s this proactive element to these systems that makes them useful to people who didn’t know they had problems to solve because there’s a lot of friction in everyday life that AI is really good at navigating. That friction is important for companies. It’s important for the entire U.S. financial system that it is not easy for you to transfer money from low-interest savings accounts to higher-interest savings accounts because low-interest savings accounts are where banks make their money. It’s important for customer service organizations that it’s hard to get customer service calls answered because otherwise everyone would ask for refunds. Muse breaks that friction down and just does stuff for you in a way that doesn’t require imagination. I think it makes the personal AI real.
Thompson: There are two ways to think about these agents. On the user side, how comfortable will people be giving private information to an agent that’s essentially spying, in a friendly way, on their email and calendar? But on the other side, what happens when millions of personal agents call customer service, fight charges, claim rebates, negotiate with companies, or send thousands of job applications? What’s going to happen on the corporate side of the internet economy in response to all these agents?
Mollick: Everybody’s worried about this right now. First of all, you can’t influence the agents through marketing or anything else. We tried to do that. One of the things I did for the book was put a page together for agents, and the agents hated the page, so I had to change it after interviewing a bunch of them. We have research showing you can’t really influence them in a consistent way. That’s bad because they’re doing shopping. When they just go for the lowest-cost product, how do we think about that scrambling everything?
The other big piece, like you said, is systems are built with friction that agents tear through, and we count on that friction. It’s the same thing as customer service, job applications, writing letters of recommendation. I think companies are not ready for what’s about to happen.
Amazon banned Muse. That feels like the kind of mistake Walmart made when they didn’t move quickly enough in e-commerce. Everything we’ve talked about today is AI reconfiguring parts of the world. We are in the renegotiation phase of a lot of work that is still jointly human and AI, but that breaks human-only systems. The co-intelligence era was great because you just had a chatbot that would give you advice. Now these systems are taking action in the world. Companies are going to be jammed with customer service calls. Things that used to be harder, like comparison shopping, are now going to be easy.
Companies are taking cold comfort from the earlier days of the internet with, “It took years for anyone to adopt.” Once someone uses a Muse system, most of them tell me amazing stories about this great thing that happened as a result. There’s already been a couple million people adopting and stickiness seems high. These are going to be products that matter, and you don’t need a billion people using these systems to screw over every way we do business today.
You should already be using AI proactively as a second opinion on taxes, legal issues, and medical issues. We have enough evidence that that stuff is true. Now it’ll just start doing it for you. I think we’re in for a very wild time even before we’re into the wild time of whatever AGI-ish worlds look like.
Thompson: Last year, you told me that 2026 would be the year of AI agents. I’d say you nailed that prediction. So, what’s your AI prediction for 2027?