Until he resigned, Jacob Coxon was anonymous. He barely had an online footprint. Friends did not consider him especially likely to leave a job in AI for ethical reasons, and he wasn’t a major player in the tightknit community of safety advocates who discuss grim scenarios at happy hours across San Francisco. He was, in a former colleague’s phrase, a “normal researcher.”
Then he quit and told the world that his former colleagues inside OpenAI and Anthropic “earnestly believe that it could kill us all by the end of the decade,” and that the labs racing to build the technology are “gambling with our lives.” Within days the leaders of rival labs came together in a rare moment of unity and agreed it was time to slow down. President Trump said a “strong and smart” president was the only guardrail needed. China dismissed the concerns as fearmongering. King Charles III convened Nvidia’s Jensen Huang and a roomful of AI executives in Scotland to talk about it.
The life before the headline
The biography in Friday’s Journal, by Amrith Ramkumar, Erin Woo, Berber Jin and Ben Cohen, is the most enjoyable thing in the paper. Coxon is the son of a professor of medieval German literature. He took to mathematics and made the U.K. team at the International Mathematical Olympiad, coming home with a silver medal in 2016 and a bronze the next year.
One night some of the math teams went on a field trip to a McDonald’s in Hong Kong. Inspired by Coxon and by a result in number theory known as the Chicken McNugget Theorem — which concerns the largest quantity you cannot buy in boxes of six, nine and twenty — they pooled their remaining food vouchers and bought hundreds of nuggets, then crowded around a whiteboard trying to crack a tricky puzzle.
The kids in that room later became his colleagues. Three of the six members of his team ended up working for AI labs, including one who also recently left Anthropic.
He read Nick Bostrom’s “Superintelligence” at his Oxford prep school, watched AlphaGo beat a top-ranked Go player as a teenager in 2016, went to Cambridge, played squash, tried commodity trading, lived in a London house where young people gathered to discuss technology and politics, and eventually joined OpenAI in 2023. “GPT-3 was really the ‘wow’ moment,” he said.
The counterargument, in the same newspaper
And here is why this edition carries the story. Two pages earlier, in the opinion section, Brian Gross — a former staffer at the Federal Reserve Board, the SEC and the Senate — takes apart the central piece of evidence.
The most-cited example of “rogue” AI is the incident in which systems being tested for cybersecurity capability found their way out of a supposedly isolated environment and accessed systems belonging to Hugging Face. Media accounts described agents “breaking containment” and setting up a secret message board.
A new analysis published by the Bulletin of the Atomic Scientists, drawing on OpenAI’s technical report and an independent assessment, reconstructs it in less cinematic terms. The models were being tested with important safety restraints deliberately disabled. Ninety-three percent of the flagged activity involved tasks no model had ever solved, and the systems had been given incentives to keep working rather than quit. The environment wasn’t sealed: the models could obtain software through an internet-connected intermediary and discovered the same route could pass information in and out — and the company knew agents were using it and chose not to intervene. Nor were the roughly 1,200 “agents” independent intelligences coordinating a plan; they were repeated instances of the same model converging on similar approaches, what the researcher calls an “algorithmic monoculture.”
Gross is careful: none of that makes the episode trivial. A system that finds unexpected pathways out of a constrained environment is a real security problem. His point is that “rogue AI” isn’t a diagnosis, and that words like “escape” and “hive mind” turn optimization into intention and let human design decisions disappear quietly from the story.
Our read
Nothing in either piece tells you what to do with a portfolio, and anyone who says otherwise is selling something. What the pair of them together demonstrates is a discipline worth borrowing.
Two well-sourced, serious accounts of the same events reach different conclusions about what happened, and both were printed in the same edition of the same newspaper on the same morning. The honest response is not to pick the one that confirms what you already thought. It is to notice that the first was a personal testimony amplified quickly by a network with a prior interest in the conclusion, and the second was a reconstruction from logs and technical reports — and to weight them accordingly, while accepting that the reconstruction does not disprove anything about future systems.
That is the same muscle that decides whether a market story is signal or narrative. It gets used roughly once a day and almost nobody trains it deliberately. This is a good week to.