The Age of Co-Existence

Few people have navigated the AI revolution with as much curiosity, pragmatism, and intellectual honesty as Ethan Mollick. He is an associate professor at Wharton, an AI researcher and the author of the New York Times bestseller Co-Intelligence. He was also named one of TIME Magazine’s Most Influential People in Artificial Intelligence.

I was really excited to sit down with him ahead of his new book, Co-Existence, in which he writes about working with an AI that is “sometimes, but not always”, better than us. Our conversation covered a lot of ground, from the art of math to the jagged frontier, the future of education and the value of apprenticeships. This interview is my attempt to capture its highlights, but you can listen to the full podcast interview here:

The Jagged Frontier

I’ve always been fascinated by Ethan’s concept of the “jagged technology frontier”, which describes the uneven impact of artificial intelligence capabilities: how AI assistance brilliantly improves performance for some tasks but worsens it for others, even within the same knowledge workflow and a seemingly similar level of difficulty. Writing is a perfect example: "If you were doing your ‘Never Normal’ newsletter entirely with AI, people would stop reading it”, Ethan explained. “It can't pull the argument together. It's still too fluffy."

But the frontier keeps moving. Math was the obvious weak spot just two years ago and has now been partially conquered, as OpenAI’s recent Navier–Stokes breakthrough illustrated. This evolution towards increased co-existence between humans and AI is just as exciting as it can be scary to some. Just imagine how overwhelming OpenAI’s recent math announcements must feel to someone who’s doing a PhD in mathematics right now.

Ethan made clear that every field, from coding to math and medicine, will be undergoing some form of AI-driven transformation at one point, but that doesn’t mean that these industries will be devoid of human employees. It will mean, however, that in field after field, we're going to need to re-examine and re-define what the end-goal is. “For instance, in math”, he explained, “proving theorems is only part of the game. In the process of proving, you learn things. It is a craft, and you are a guardian of a great intellectual tradition. There is even a debate among mathematicians about whether a solution is worth much if humans didn't produce it, because we didn't learn anything.”

He sees something similar happening within organisations. "They are beginning to ask more fundamental questions. Like “was producing the PowerPoint or the report ever the real goal”? Now that AI allows us to generate thousands of them, it becomes clear that the presentation in itself was never the objective. It was merely a means to achieve something else. That search for meaning is going to be a wave across everything. And there are losses and gains coming for all of us."

It’s an interesting idea: when the output becomes cheap, you are forced to rediscover the purpose. And that’s a question many of us haven’t asked ourselves in years.

Serious About Learning

Ethan is contrarian about education. Contrary to what some techno-positivists, like Marc Andreessen believe, he feels that it will only become more valuable. “Universities are uniquely positioned to push us to develop both breadth and depth of knowledge and to certify that we’ve actually acquired it”. “That mission”, he says, "feels more real than ever." The disruption to teaching styles will be more temporary than in other fields, he believes. "Education is always a mess. We still pretend case studies are a good way to teach, while there's no research that says they are. What we do know is that active learning and tutoring work. So, imagine what will happen if we have AI tutors outside the classroom and active learning inside it.”

What worries him is the next phase, though. What used to happen was that a generalist left an institution like London Business School and became a specialist at a bank that taught them the trade. That was a great deal, and it rested on apprenticeship. Juniors were desperate to prove themselves and they learned by doing. "But that all broke, because everybody is using AI for everything."

His new book proposes two answers. Companies need to get serious about learning, even though this will sometimes be inefficient. That might mean making people work without AI for a while. Or they will need to partner much more deeply with universities, with for instance a sponsored specialist year paid for by the employer. In both cases, learning has to be treated as a deliberate investment instead of a by-product of getting work done.

Not Quite The Industrial Revolution

I recently read and loved Cory Doctorow's ‘Reverse Centaur’ and Sarah O'Connor's ‘We Are Not Machines’. Both, like many contemporary books on AI, refer to the Industrial Revolution. GDP exploded back then while working conditions did not really improve and wages stayed flat for a very long time. So I asked Ethan whether our society is ready for something that may evolve even faster and more intense.

Ethan agrees the parallels are partly real. “Living through it was chaotic and painful for many, but the outcome was undeniably positive. Wages only caught up after a fight between capital and labour, fought through strikes and unions.” “The good news”, he continued, “is that we have far better tools today for dealing with boom-and-bust cycles and displaced workers.”

At the individual level, he says, we do get to choose between a centaur and a reverse-centaur path. For those of you not familiar with these concepts: a centaur is a human who uses AI to enhance their capabilities while remaining in control, whereas Cory Doctorow’s reverse centaur flips that relationship. Here, the AI is in charge, and the human is reduced to carrying out its instructions or checking its results. Ethan thinks Doctorow underestimates the technology, though, because he treats it as just another automation lever.

Ethan went on to refer to a study from Epoch showing that a current frontier model can deliver somewhere between two and eighteen weeks of engineering work. A world in which humans just verify the output would be a disaster, he said. But that isn't where we're going either, because AI will become superhuman at verification too.

"The people being attacked this time are the ones with the most money and the most levers of power.", he continued. “Mathematicians are pushing back now”, he said, "but wait until they come for the doctors and the lawyers. A third of Congress is lawyers." Do we really think the bar associations will quietly accept AI? He sketches two futures. In one, guild-like protections slow everything down. In the other, it all happens "everywhere, all at once", and we end up in a post-industrial or even post-need society. "It's science fictional."

Positive Policy

We also talked about political reactions to AI disruption, and his honest observation is that nobody is keeping up, including them. He thinks it is right to take existential risk seriously. But he also fears that debate sucks all the air out of the room. Europe too has built a lot of anticipatory regulation around hypothetical future dangers, while other urgent questions go unanswered. What do we do about the cybersecurity situation we face right now? How do we encourage augmentation instead of automation? "You're not going to do that by making it harder to fire people."

He shared a comment from someone at the Federal Reserve, who compared it to giving a baby an AK-47 and worrying about what happens when it becomes a teenager. We worry so much about the future, while the present needs policy.

And he wants policy to be positive, not just defensive. Too many European organisations are waiting for regulatory approval or explicit political encouragement before taking the initiative on AI. In his view, every human should also have access to advanced AI for, for instance, a second medical opinion. And we should launch a crash project to build certified AI tutors for every student in the world. "If you only think regulation and stopping, you’ll get into trouble."

I liked how direct Ethan was on sovereignty. "Sovereign is an illusion." If your strategy depends on open-weights models that may or may not remain available, you are not sovereign in any meaningful way. China doesn't seem particularly worried about safety and sees this as a race. It will likely have its own "Mythos moment" at some point, and then it may well decide that its frontier weights stay at home. (The Mythos moment refers to the turning point in April 2026 when Anthropic announced Claude Mythos, a frontier AI model capable of autonomously finding zero-day vulnerabilities, chaining complex exploits, and executing multi-step cyberattacks at a scale and speed beyond human capacity.) "Then what's the deal? You can't run a one-year-old open-weights model and expect it to behave like the most recent frontier models."

You Can't Delegate This

I ended by asking for his final word of advice, and he didn't hesitate. "A huge amount depends on you using AI yourself, whoever you are. I can't emphasize enough how much you need to use a frontier model. This can't be delegated down."

When he wrote Co-Intelligence, he said, you could still afford to wait. That is no longer true. His biggest fear is that people form their opinions from a model they tried a while ago, or from news articles and consultancy reports. There is a big difference between that and using a frontier model yourself. “So take out the $20 subscription (or the $100 or $200 one), download a coding agent like Claude Code or Codex, and put it on real, ambitious work. Until you do that, I don't think you can have an understanding of what's happening."