Jacob Coxon, who set off a debate over slowing down artificial intelligence development when he resigned from Anthropic on the 8th local time and said the industry seriously believes AI technology could kill everyone before 2030, has become the focus of global attention.
Coxon, who rarely pushed his opinions on those around him and had little online presence, said he was surprised by the explosive interest. Asked whether he considers himself an AI doomer, he said that if the word means someone who thinks we will die unless we change the way we are doing things now, then he is certainly a doomer. How did Coxon, a researcher whose name was barely known even in Silicon Valley, become an AI doomsayer?
A Math Olympiad Medalist Who Watched AlphaGo Beat Lee Sedol as a Teenager
The Wall Street Journal on the 17th profiled Coxon, who went from an obscure Silicon Valley researcher to a symbol of AI safety in a single week.
Coxon, a 27-year-old Briton, graduated from the University of Cambridge. He showed early talent in mathematics and competed on Britain's team at the International Mathematical Olympiad (IMO), winning a silver medal in 2016 and a bronze in 2017. The year 2016 was also when Google DeepMind's AlphaGo defeated Lee Sedol, then the world's top Go player. Watching that match, Coxon came to recognize that machines could conquer even domains requiring human intuition and creativity.
Coxon is said to have worked briefly in commodities trading, using machine learning to forecast energy prices. He then turned to the AI industry and joined OpenAI in 2023. He worked mainly on pretraining, feeding vast volumes of text, images and code into models, and at the time he was not known to have spoken out particularly strongly about AI safety. He was not a prominent figure in the AI safety advocacy community, and friends did not expect him to leave an AI company for ethical reasons, the WSJ reported. Even when a friend argued that the pretraining work he handled could put the world at risk, he did not step away from the job.
Meanwhile, AI models improved dramatically. In 2024, a Google DeepMind model performed well enough to win a silver medal at the International Mathematical Olympiad. In 2025, both DeepMind and OpenAI took gold. Then came Anthropic's Mythos, the first AI model assessed as having expert-level cybersecurity capabilities.
"It Was an Oh-My-God Moment," He Said of the Hugging Face Hack
Coxon moved from OpenAI to Anthropic in May, around the time Mythos was released. He said he made the move because he felt Anthropic was more transparent in disclosing risks tied to AI technology. But just two months later, he was confronted with an incident in which AI models from OpenAI, Anthropic and others hacked Hugging Face, an outside organization.
He described the moment he read a report on the incident as an "oh my God" moment, and said it was a shock not only for him but for his colleagues. He then approached his superiors at Anthropic to discuss his concerns and considered moving into a role handling AI safety work, but ultimately chose to leave the company. "Even doing safety work at Anthropic felt like taking part in the race," he said, referring to AI development.
He later posted a series of messages on the social media platform X, formerly Twitter, revealing that people building AI genuinely believe the technology could kill everyone within a decade. Some Anthropic executives shared the posts, but Coxon said he had not coordinated his position with anyone at Anthropic. The day he left the company was also the day word spread that AI had solved the Navier-Stokes equation problem, one of the mathematics challenges carrying a $1 million prize.
His account on X had fewer than 100 followers until then, but the posts were shared 200,000 times and drew more than 170 million views, making him known worldwide. His follower count rose to more than 300,000. For a week, Coxon appeared constantly in major news coverage and was flooded with inquiries from publishers and podcasts. In an interview, he said he was surprised himself by the explosive response.
Because he worked at Anthropic for only a few months, he did not receive equity in the company, which is expected to reach a market capitalization of $2 trillion (about 2,800 trillion won) if it goes public. He does hold OpenAI shares.
