Jacob Coxon, a 27-year-old machine learning researcher who worked at the forefront of foundation model development at both OpenAI and Anthropic, announced his resignation from Anthropic and his departure from the artificial intelligence industry altogether. His exit, first reported by The Wall Street Journal, was accompanied by a direct public indictment of the sector’s sprint toward artificial superintelligence, warning that unchecked self-improving systems could slip beyond human control in the near future.
Coxon spent the past three years focused on pretraining—the resource-heavy foundation phase where frontier systems learn broad representations from vast datasets before post-training refinement. During his tenure at OpenAI, he served as a core contributor to GPT-4o and co-authored its safety documentation, before moving to Anthropic eight months ago.
"I resigned from Anthropic today," wrote @hilbertspaess, Coxon’s public handle. "I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives."
The resignation marks a rare break in the ranks from an engineer embedded directly in the foundational pretraining pipelines that power top-tier commercial models. Observers point out that Coxon’s technical credentials distinguish his critique from external speculation. "Jacob Coxon was a core contributor to GPT‑4o and co-author of its System Card, cited 6,800+ times," noted @sachi_gkp. "His warning isn’t proof of catastrophe—but it is a serious insider signal: AI capability may be accelerating faster than our ability to control it."
A Shrinking Window for Containment
At the core of Coxon’s exit is a belief that current development trajectories are collapsing the safety margins labs previously operated under. According to reporting on his resignation, Coxon projects that recursive self-improvement—where models autonomously design, evaluate, and train successor iterations—could produce unmanageable dynamics within the next three years.
"We’re on track for a lot of the most aggressive of these scenarios where by the end of next year things could be out of control already," Coxon stated, as cited by @wallstengine.
The scenario of automated research loops has transitioned from a long-term theoretical risk to an active product roadmap among leading San Francisco labs. In recursive architectures, systems that write code and evaluate synthetic data could compress years of architectural refinement into weeks, potentially creating capabilities that developers cannot audit or reliably constrain before deployment.
Lab Cultures and the Regulatory Impasse
Coxon’s resignation highlights an acute structural tension within corporate AI laboratories, particularly Anthropic, which was founded in 2021 by former OpenAI researchers specifically pledging a safety-first charter. While researchers inside these labs frequently acknowledge catastrophic hazards, competitive market pressures and national security narratives continue to drive capability advances forward.
Colleagues within the research community acknowledged the gravity of the critique even while remaining inside corporate labs. "Jacob’s thread is very worth reading," wrote Anthropic researcher @saprmarks, noting they were speaking in a personal capacity rather than on behalf of the company. "AI developers believe their technology could cause catastrophe, yet the race dynamic persists."
That dynamic has renewed calls for binding oversight that supersedes voluntary corporate pledges. Without enforceable caps across competing firms and jurisdictions, individual researchers who refuse to advance frontier systems simply leave vacant seats for others to fill. As @Afinetheorem pointed out, individual departures underline the need for external state power: "I would bet the *vast* majority would prefer a rigorous regulatory regime that binds on American and Chinese firms, esp the 15 or so working at frontier, so that care about safety is not punished."
Whether Coxon’s exit prompts formal scrutiny from lawmakers or shifts internal protocols at Anthropic remains unclear. As frontier developers prepare their next generation of training runs, the question is whether technical dissent from core pretraining engineers will translate into slowed deployment schedules or merely register as an isolated casualty of an accelerating buildout.