On September 9, 2026, Jacob Coxon, who spent three years working on pretraining research at OpenAI and Anthropic, announced his departure from Anthropic on X. In his statement, he claims that "neither company is acting responsibly." The Wall Street Journal headlined the news as a departure due to fears of "uncontrollable AI" (out-of-control AI).

What happened

Coxon's post on X begins with the phrase "I resigned from Anthropic today," after which he describes three years of pretraining work at both labs and claims that both OpenAI and Anthropic are racing toward a self-improving superintelligent system, risking lives. The statement sparked widespread discussion: a thread on Hacker News gathered 126 points and 137 comments. Coxon's Google Scholar profile confirms his affiliation with Anthropic: his list of works includes co-authorship on the GPT-4o system card (2024) and a publication on weight-sparse transformers (2025).

Context

The author's scientific profile is limited: according to Google Scholar, Coxon's h-index is 3, and almost all of his roughly 6,900 citations come from co-authorship on the GPT-4o system card, while his individual contribution to model risk assessment is not visible in open data. The thesis of a "race to self-improving superintelligence" is an interpretation, not an established fact: neither the quote from the statement nor The Wall Street Journal headline contains references to specific capability metrics, internal test results, or evaluation data. However, the news was published in a mainstream outlet, which moved the debate over AI existential risks from professional discussions into the general agenda.

Why this matters for the industry

A public departure by a pretraining researcher from Anthropic with accusations of a "race to superintelligence" is a rare primary signal that debates over risks within the lab are going beyond official statements. For the industry and investors, this is a reputational signal of internal conflict, and for ML engineers, it is an indicator of organizational risks in frontier labs on which production systems rely. The direct technical impact is zero: the news contains no new model, API, pricing, or latency or benchmark data, and no production system is changing its model, inference cost, or SLA. If 2-3 similar public cases follow the departure, it will become a systemic signal of a split within the labs, which will affect hiring, employee retention, and safety program priorities.

Why this matters for users

The event is a specific example of how the discussion of AI existential risks translates into personnel decisions in labs. Readers have access to both sides of the argument: the original statement on X and the Hacker News thread provide both the position that "the race is real and dangerous" and the position that "this is PR, not a proven threat." The direct practical effect for users is minimal: there is nothing to integrate into projects today, so the main result is the ability to separate the verified fact of the departure from unconfirmed claims about risks.

What is still unknown / limitations

The central technical claim about a "race to a self-improving superintelligent system" is not supported in the provided materials by evaluation results, capability metrics, or internal test data. It is impossible to reliably distinguish a departure due to beliefs from a departure due to conflict or PR from public data: part of the community on Hacker News interprets the statement as attention marketing rather than a moral act. The sources lack customer statements, data on vendor changes, and information about changes in lab policies, so any conclusions about the impact on the market and vendor strategy would go beyond confirmed facts.

Sources

Author

Look at AI, editorial team