Brian Greene has a friend, the physicist Michael Douglas, who gave him a piece of career advice that doubles as an obituary: choose the final problems you want to work on yourself. Within a few years, Greene recalls Douglas saying, physicists will be “out of business because the AI systems will just be doing it all.”
Greene, a professor of physics and mathematics at Columbia, calls the prospect “both exciting and terrifying.” On The Diary of a CEO, he describes a physicist’s lifeblood as looking out at the world, finding puzzles and hunting for explanations. If a machine can generate the answer, what becomes of the person whose life revolves around that hunt?
Yet the same possibility carries an extraordinary attraction: answers to questions Greene thought he might never see resolved in his lifetime. Douglas’s warning is a forecast Greene recounts, not a deadline he establishes. But it gives a personal edge to the question that runs through his discussion of creativity: could the next Newton be an artificial intelligence?
What made Newton different
Asked to choose between Isaac Newton and Albert Einstein, Greene first resists the contest. At that level, he says, it hardly matters who was smarter. Pressed, he gives Newton the edge because of the scale of the first step he took toward a mathematical account of nature.
Greene points to Newton’s laws of motion and gravity, and his development of calculus as a tool for describing physical change. Force equals mass times acceleration—the relationship still taught in school—connects a push or pull to how an object’s motion changes. His law of gravity describes how attraction depends on mass and distance.
Greene acknowledges Newton’s predecessors. What impresses him is the leap toward an architecture in which a few equations could explain so much of the physical world. That is a more demanding model of scientific achievement than solving a familiar problem faster.
Could a machine make a comparable leap? “Yes, it could absolutely be,” Greene says. He separates creativity into three broad categories to explain why.
Seeing more of the possibilities
The first kind is the ability to see a larger landscape of options. A strong chess player considers moves an ordinary player misses. A scientist may recognize an equation or approach that another researcher never thinks to try.
Greene expects AI to excel here because of the breadth of material systems can draw on. His language about machines having read every textbook and knowing everything humans have developed is sweeping; the useful distinction in his argument is between a person’s limited repertoire and a system able to explore many more possibilities.
His example is AlphaGo, the AI system built to play the board game Go. In Greene’s telling, its celebrated Move 37 initially looked like a blunder before people recognized its ingenuity. He sees it as a case of a machine finding an effective option outside human expectations.
DeepMind’s account places the move in the second game of AlphaGo’s March 2016 match against Lee Sedol, which the system won four games to one. AlphaGo combined learned judgments about promising moves and likely winners with a search through possible continuations. It trained on expert games, then improved by playing versions of itself.
That was not an exhaustive view of every possible game. It was a way to explore possibilities selectively and effectively. Nor did it eliminate human ingenuity: DeepMind also highlights Sedol’s unexpected Move 78 in game four, which helped him secure his one victory in the match.
Joining ideas from distant fields
Greene’s second category is bringing together ideas that nobody has previously connected in that way.
His example is Einstein’s use of the mathematics of curved spaces to rethink gravity in general relativity. Rather than treating gravity only as an attraction between masses, the theory describes it through the geometry of space and time. In Greene’s broad classification, the creative act is recognizing that a mathematical framework can transform a physical question.
Here, too, he expects AI’s breadth to matter. Drawing on work across different fields could allow a system to suggest combinations that a specialist, steeped in one discipline, would not consider. The proposed advantage is not merely remembering more facts, but finding productive relationships among them.
Inventing something nobody imagined
The third category is harder: a genuinely new framework, rather than a wider search or an unfamiliar combination. It is the kind of idea that leaves other scientists asking where it could possibly have come from.
“That’s a tough one,” Greene says. “We’re still able to win in that category.”
But he does not see it as permanently reserved for humans. He asks what might happen if an artificial system were allowed to grow up in an environment, acquiring experiences of the sort Newton had, rather than relying only on the material used to train it.
In this proposal, experience would give an artificial scientist encounters with the world from which to develop new ideas. Could it then produce something not already represented in its training data? “I don’t think there’s a barrier,” Greene says. He offers that as a possibility, not as a demonstrated route to a machine-made theory of nature.
A collaborator, not just a replacement
Earlier in the conversation, Greene offers a less binary future than humans doing science or machines taking it away. People are already working with artificial systems, he says, and he invokes a reported mathematical result as an illustration: an AI-assisted resolution of the Jacobian conjecture.
In a July 20, 2026 post, mathematician Oliver Knill reports that Levent Alpöge used Claude Fable to find a counterexample. Broadly, the conjecture asks whether a particular condition that lets a polynomial transformation be reversed locally also guarantees that it can be reversed globally. A counterexample would show that the condition is not enough.
Knill supplies a proposed three-variable transformation and symbolic checks intended to show both that it meets the condition and that two distinct inputs produce the same output. This is Knill’s report of an AI-assisted disproof, not an independently established mathematical consensus. Greene invokes the case as an example of a mathematician working with AI, rather than a machine pursuing research alone.
Further out, Greene imagines the partnership becoming harder to separate into human and artificial contributions. Perhaps there will be implants; perhaps integration will take another form. He suggests that the boundary between biological and artificial intelligence could eventually blur enough that the result would no longer be recognizably human. Evolution brought us here, he reasons; this need not be its endpoint.
That is his alternative to a simple replacement story, but it does not remove the loss Douglas’s advice brings into focus. A physicist might gain answers to lifelong questions while losing the familiar role of working them out personally. Greene’s hope is that collaboration opens questions he once thought beyond his lifetime—even as it changes what it means to spend a lifetime doing physics.