Jacob Coxon: What Did the Former Anthropic Researcher Say About the Superintelligence Race?

Jacob Coxon spent the past few years working close to the center of the artificial intelligence boom. The 27-year-old British researcher worked on advanced AI systems at OpenAI and later Anthropic, two companies competing at the frontier of increasingly capable models. Then, in September 2026, he walked away.

Coxon resigned from Anthropic on September 8 and said he was leaving the AI industry, warning that leading laboratories were moving too quickly toward what he describes as self-improving superintelligence. His departure quickly became much bigger than an ordinary Silicon Valley resignation. Coxon argued that researchers understand some of the potentially catastrophic risks of advanced AI, yet competitive pressure keeps pushing companies to build more powerful systems.

His claims remain predictions rather than established outcomes. There is no certainty that self-improving superintelligence is close, or even that AI development will follow the path Coxon fears. Still, his background inside OpenAI and Anthropic has made his warning difficult to dismiss as criticism coming entirely from outside the industry.

Who Is Jacob Coxon?

Jacob Coxon is a British mathematician and artificial intelligence researcher who has worked at both OpenAI and Anthropic.

Before entering the AI industry, Coxon studied mathematics at the University of Cambridge. His mathematical background eventually took him toward machine learning and the increasingly competitive field of frontier AI research.

He spent roughly three years working in AI, first at OpenAI and later at Anthropic. Much of his research involved pretraining, one of the most important stages in building large AI models.

Pretraining involves exposing a model to enormous quantities of information so it can learn statistical patterns and develop capabilities that can later be refined through additional training.

Coxon was therefore not simply observing the AI boom from a distance. His work placed him directly within teams involved in developing increasingly capable models.

He also contributed to GPT-4o. OpenAI’s official contributor list includes Coxon among the model’s core contributors, placing his name alongside researchers and engineers who worked across different parts of the project.

That experience has become particularly relevant following his resignation because Coxon’s criticism concerns an industry in which he personally participated.

Why Did Jacob Coxon Quit Anthropic?

Coxon announced his resignation in a series of posts on September 8, 2026.

His central concern was straightforward but dramatic. He believes frontier AI companies are competing to develop systems that could eventually become capable of helping improve AI itself.

Coxon described the direction as a race toward “self-improving superintelligence.”

In his account, competitive pressure creates a difficult situation. A company might recognize serious risks associated with increasingly autonomous and powerful AI, but executives and researchers could still feel pressure to continue because another company might reach the same capabilities first.

Coxon said neither OpenAI nor Anthropic was behaving responsibly enough under those conditions.

He also made one of his strongest claims about attitudes inside the industry, saying people developing advanced AI genuinely believe there is a possibility that it could cause catastrophic harm before the decade ends.

That statement attracted enormous attention online.

It is important to distinguish Coxon’s account from a proven industry-wide consensus. Different AI researchers hold very different views about the likelihood, timing and severity of risks from advanced artificial intelligence.

What Is Self-Improving Superintelligence?

The idea at the center of Coxon’s warning is sometimes described as recursive self-improvement.

Today’s AI models are largely created and improved through work performed by human researchers, engineers and large computing systems. But AI is increasingly being used to assist with coding, scientific research and AI development itself.

The concern is what could happen if future models became substantially better at these tasks.

Imagine an AI system capable of helping researchers design a more powerful AI model. That improved model could theoretically contribute to designing an even more capable successor.

If that cycle became fast enough, proponents of the risk scenario argue that AI capabilities could advance faster than humans could reliably understand, evaluate or control them.

Whether such a rapid feedback loop is technically possible remains debated.

Coxon’s argument is that society should not wait until the question has already been answered through a real-world experiment.

Why Coxon Talks About an AI “Endgame”

Coxon has also drawn attention to language reportedly used among people thinking about advanced AI, including terms such as “crunchtime” and “endgame.”

His objection is partly about who gets to make decisions carrying potentially global consequences.

Private technology companies currently make many important choices about how quickly frontier systems are developed, tested and released. Coxon argues that a decision to pursue potentially transformative superintelligence should not effectively emerge from competition among private laboratories.

He believes governments and competing AI companies will eventually need some form of coordination.

The problem is that slowing development is difficult when every participant fears somebody else will continue.

If one American company pauses while another keeps developing, the company that stops could lose its technological advantage. The same argument becomes even more complicated internationally, particularly when competition between the United States and China enters the discussion.

Coxon sees this dynamic as one of the fundamental problems surrounding AI safety.

Why Not Stay at Anthropic and Work on Safety?

Coxon’s departure created another debate almost immediately.

If he believes advanced AI poses such a serious risk, some critics asked why he would leave Anthropic rather than remain inside the company and work to make its systems safer.

It is a reasonable question, particularly because Anthropic has publicly emphasized AI safety as an important part of its identity.

Coxon’s decision suggests that he no longer believed working inside a frontier laboratory was the best way for him to address the problem.

There was also a personal financial cost associated with leaving. Coxon told Axios that he departed around two months before his Anthropic equity was scheduled to begin vesting. He used that fact to push back against suggestions that his public warning was intended to increase the company’s value.

Reaction to his resignation has nevertheless been sharply divided.

Some researchers and AI safety advocates have treated it as evidence that concerns inside frontier laboratories deserve greater public attention. Others argue that predictions of human extinction or runaway superintelligence rely heavily on uncertain assumptions about capabilities that do not currently exist.

What Anthropic Researchers Have Said

Coxon’s concerns are not entirely isolated.

Following his resignation, other researchers connected with Anthropic publicly discussed serious risks associated with increasingly powerful AI systems.

That does not mean every Anthropic employee agrees with Coxon’s conclusions or timeline. AI safety itself covers a broad range of concerns, from current problems such as cybercrime and manipulation to hypothetical future scenarios involving systems exceeding human capabilities.

Anthropic has continued to publicly emphasize safeguards and research aimed at understanding dangerous model behavior.

The wider debate is therefore not simply between people who think AI is safe and people who think it is dangerous. Much of the disagreement concerns which risks deserve the most attention, how quickly capabilities might develop and what level of intervention is justified before those questions are settled.

His Warning Has Attracted Skepticism

Not everyone accepts Coxon’s interpretation of where artificial intelligence is heading.

Some researchers argue that today’s models remain fundamentally limited and that predictions about autonomous superintelligence extrapolate too aggressively from recent improvements.

Others worry that focusing heavily on extinction scenarios could distract policymakers from problems already happening today.

Those include AI-assisted cybercrime, misinformation, surveillance, fraud, job disruption and the use of increasingly autonomous systems in sensitive environments.

There is also disagreement about regulation. Some people inside the technology industry support stronger government oversight, while others worry that restrictive rules could protect established AI companies by making it harder for smaller competitors to operate.

Coxon’s resignation has therefore become part of a much larger argument about both technological risk and who should control AI development.

What Jacob Coxon’s Resignation Really Means

Jacob Coxon’s decision does not establish that superintelligence is coming soon. It also does not prove that artificial intelligence will escape human control or produce the catastrophic consequences he fears.

What makes his resignation notable is his position before making those warnings.

Coxon studied mathematics at Cambridge, worked on pretraining at OpenAI and Anthropic and contributed to GPT-4o. He had direct exposure to frontier AI development before deciding he no longer wanted to participate in the race.

At 27, he has now moved from a relatively unknown researcher to one of the most discussed voices in the current AI safety debate.

His resignation ultimately leaves the technology industry with an uncomfortable question. If AI systems continue becoming more capable and begin contributing significantly to the development of their successors, when should companies slow down?

Coxon decided that point had already arrived.

Many researchers disagree about whether he is right. But his departure has pushed a debate that once belonged mostly to specialized AI safety circles much further into public view.