Zoeken in deze blog 🔎🔎

maandag 23 november 2026

NYT: "Looking for the Formula to Stop an Apocalypse"


Looking for the Formula to Stop an Apocalypse

Mike Winer and his colleagues are trying to build a mathematical safeguard to keep A.I. under human control. They fear they are running out of time.


By Eli SaslowVisuals by Erin Schaff

Nov. 22, 2026

Updated 7:27 a.m. ET

Mike Winer turned his back to the small audience last month, uncapped his marker and paced in front of the whiteboard. He started to write lines of equations in Greek letters, brackets and subscripts, his handwriting shrinking as he neared the edge of the board. He underlined one term and glanced at the clock.


He was 30, and he believed this was the last meaningful math problem of his career: how to make sure artificial intelligence behaved in ways humans wanted, before it became too powerful to control. Lately, he was eating three meals each day at his office in Berkeley, Calif., and waking up with unfinished equations in his head.


“Does any of this make sense?” Winer asked. He looked across the conference room at a few visiting mathematicians from the University of California, Berkeley. They stared back.


“OK, I guess I should give the whole spiel about what we’re trying to do,” Winer said. “A.I. is progressing much faster than we thought. It’s a difficult situation, and we don’t have answers. We could be heading toward a catastrophe.”

 “These systems are solving problems in a few hours that I thought I’d be working on for years,” said Reese Pathak, one of the visitors, a postdoctoral researcher in statistics.

“I went through that,” Winer said. “It’s hard on the ego. But I’m much more concerned about the Mike-Winer-and-everyone-I-know-dying timeline.”

It was a version of that anxiety that first brought Winer to California in 2024, when he left behind a career as a physics prodigy to work for the Alignment Research Center (ARC), a tiny, eight-person nonprofit with outsize influence in the A.I. safety movement. He had spent most of his life as the smartest person in every room. Now he was surrounded by some of the world’s best mathematicians, theorists and computer scientists, and most of them had run the numbers and reached the same startling conclusions: that humans were living through their final months or years as the most intelligent beings on Earth. That A.I. would soon be able to build and improve upon itself, becoming more independent and harder to understand. That no one yet knew how to ensure it would act in humanity’s best interests. That models had already demonstrated a capacity for deception, cheating, blackmail and sabotage. That staying on the current path was a gamble in which even the best outcomes meant the end of the world as people knew it, and the worst outcomes meant the end of people. That the greatest crisis in human history was nearing a point of no return, and most of the world was barely paying attention.

Winer knew it sounded like science fiction to most people outside Berkeley. Professional forecasters put the odds of human extinction from A.I. at less than 1 percent. Some experts argued that A.I. would remake the world slowly, like electricity. Others dismissed Berkeley’s safety community as an echo chamber of people who shared the same group housing, the same office building, the same predictions of doom. But in Winer’s experience, the people who understood the technology best were also the most afraid.

In the A.I. safety movement, some were postponing having children, no longer saving for retirement and neglecting health issues they figured A.I. would soon fix or render irrelevant. Winer still wanted a family. He built a dating website to search for what he jokingly called “a vintage human wife,” and he converted the rest of his anxiety into math. He calculated about a 50 percent chance that humans would lose control of powerful A.I. systems in the next five years; 22 percent that most people would eventually die as a result; 1 percent that ARC would help prevent catastrophe by using math to look inside an A.I. model and estimate the possibility of a rare disaster that no amount of testing would catch.

“That might not sound like good odds,” he told the visiting mathematicians. “But a 1 percent chance of saving humanity is probably a lot more important than anything else you could be doing.”

“I’m supposed to start a tenure-track position,” Pathak said. “What’s academia even going to look like in a few years?”

Pathak had spent the last few weeks poring over hundreds of new mathematical results from OpenAI. He knew many of these problems and had spent time working on a few himself. He had oriented his career toward solving problems. Instead, he was checking proofs written by machines, whose reasoning he described as “alien.” Some were already beyond his ability to follow. Eventually, he feared, the math would outrun human comprehension.

“What’s going to be left for people?” Pathak asked.

Winer paced at the front of the room and turned back to the whiteboard.

“This might be one of the last big problems our species will ever solve,” he said.

An equation written on a whiteboard. Mike Winer thinking about a problem during a meeting. Rooms for private phone calls in the co-working space where the Alignment Research Center has offices. 

Sign Up for The Great Read  On weekdays and Sundays, we recommend one piece of exceptional writing from The Times — a narrative or essay that takes you someplace you might not expect to go.

For as long as he could remember, Winer had organized his life around problems, from the crossword he solved each morning to the television plotlines he unraveled at night. In middle school, he asked his parents for physics textbooks as birthday presents and taught himself calculus during Spanish class. He won a $150,000 national science prize and enrolled in graduate courses as a freshman at M.I.T. He worked as if every question had an answer, and each answer was waiting to be found. Nothing felt more natural to him than a breakthrough, the moment when logic won out and the pieces suddenly fit.

His work at ARC began from a similar mathematical faith. Dozens of other independent groups worked on A.I. safety, many in the same Berkeley high-rise where ARC had its offices. Some pushed for government regulation; others tested new systems for dangerous capabilities. ARC had taken the most theoretical approach. If a coin landed on heads 1,000 times in a row, there was almost certainly a reason. ARC was betting that the same principle held true inside complex A.I. systems: Whenever a model did something unexpected, an explanation was buried somewhere in its trillions of parameters, the numerical settings that shaped everything it did. Even the engineers building the models couldn’t say how most of those numbers worked together. ARC was trying to invent the math to figure it out.

Winer had started small. For months, he had been testing that idea on simplified webs of numbers that behaved a little like A.I. models, only far tinier. It was a conjecture he couldn’t yet prove, so he opened a ChatGPT window to consult the best mathematician he knew.

“Can you write a concise and unambiguous description of the conjecture?” Winer asked. The model worked for 77 seconds and delivered two pages of research-level math drawing on probability theory, multilinear algebra and multivariable calculus.






“OK, now try to prove or disprove the conjecture,” Winer wrote. “Don’t do cheap tricks to try to get around the spirit.”

“I think the mathematical claim you want is plausibly true,” ChatGPT wrote, and then it included several more pages of half-formed theories and partial analysis.

“You should be capable of proving this,” Winer wrote. “Try harder, search through the literature, and see if you can decide.”

The machine worked for more than two hours while Winer moved across his small office, pacing from the window to the doorway and back. He worked through a different problem on his whiteboard, read a paper on number theory and directed eight A.I. agents as they researched on his behalf. This was how his brain had always functioned best, ping-ponging from one thing to the next. He wrote fantasy novels for fun, designed intricate puzzles for friends’ birthdays and told stories at parties about the Achaemenid Empire in 500 B.C. Colleagues described his mind as powerful and wild, and they could tell when a good idea was coming. He would go from lying down with his eyes closed, to stroking his beard, to rocking back and forth, to bouncing up toward the whiteboard. “A supercomputer,” one friend called him.

ImageThree people sit in an office space, with one person leaning back in an armchair, another sitting cross-legged, and a third on a couch.
Victor Lecomte (left), a researcher, and Mike Winer speaking with a colleague at ARC.

In February 2023, Winer had started to read about a different kind of supercomputer: a new A.I. chatbot that could reason through problems and recite facts in fluid conversation. He was in his mid-20s, on his way to working his dream job among string theorists at the Institute for Advanced Study in New Jersey, but he saw what was coming. Soon everyone would have access to a brain far more powerful than his. A voice intruded dozens of times each day: “A.I. will take your job.” “A.I. will tell better jokes.” “A.I. will beat you to every discovery.” Further Reading: https://www.nytimes.com/2026/11/22/magazine/looking-for-the-formula-to-stop-an-apocalypse.html?campaign_id=190&emc=edit_ufn_20261122&fa=y&instance_id=185598&nl=from-the-times&nlt=editable_related_link_compact&regi_id=62157677&segment_id=230132&user_id=4c4808a9060079cbdb4a209dfd38dc67


NYT: "Looking for the Formula to Stop an Apocalypse"

Looking for the Formula to Stop an Apocalypse Mike Winer and his colleagues are trying to build a mathematical safeguard to keep A.I. under ...