Former Anthropic researcher cites IPO slide as reason for resignation
Jacob Coxon left the AI lab after seeing a chart showing accelerating efficiency gains that he believes could lead to dangerous self-improving systems.
The short version
- Jacob Coxon resigned from Anthropic in September 2026 after only four months on the job.
- The resignation was triggered by a slide showing superexponential gains in pretraining compute efficiency.
- Coxon argues that accelerating gains could lead to self-improving AI systems, posing an existential threat.
- Anthropic is currently seeking to raise approximately $64 billion in an IPO with a $2 trillion valuation target.
Why it matters
This story highlights a growing tension between the financial goals of AI labs and the safety concerns of their researchers. It suggests that the same data used to attract investors may be viewed as a warning sign by those tasked with managing AI risks.
A departure over data
Jacob Coxon, a former pretraining researcher at Anthropic, recently revealed that a specific slide from the company's IPO roadshow was a primary factor in his decision to resign. Coxon joined the company in May 2026 but left only four months later in September 2026. While many employees leave for better pay, Coxon stated that his departure was driven by safety concerns regarding the company's trajectory.
The slide in question showcased superexponential gains in pretraining compute efficiency. This data was presented as a major selling point for potential investors during Anthropic's current roadshow. However, for Coxon, the chart served as a warning sign rather than a success metric. He expressed that the data suggested a pace of progress that could lead to significant risks.
Coxon's departure is notable because it occurred while Anthropic is preparing for a massive public listing. The company is aiming to raise approximately $64 billion with a targeted share price between $70 and $78. By leaving early, Coxon forfeited unvested equity, a significant move given that the company is targeting a valuation near $2 trillion.
Understanding the math
To understand the concern, one must look at how the company plotted its efficiency gains. Pretraining is the phase where an AI model learns from massive datasets, and compute efficiency measures how much capability a lab achieves per unit of computing power. Anthropic displayed these gains on a log scale to illustrate the trend.
On a standard chart, exponential growth appears as a hockey stick shape. On a log scale, that same growth appears as a straight line. When a line on a log chart bends upward, it indicates that the growth is outpacing even exponential rates. This is what the company refers to as superexponential growth, meaning the gains are accelerating as they compound.
Coxon interpreted this upward bend as evidence of a dangerous trend. He argued that if efficiency gains continue to stack at this rate, the development of self-improving AI systems could happen much faster than the industry currently anticipates. He has publicly warned that such systems could pose existential threats to humanity.
Internal and external echoes
Coxon is not the only person within the AI industry expressing these specific anxieties. His concerns about rapid progress toward self-improving systems have been voiced by other figures in the field. Specifically, his assertions align with the sentiments expressed by Evan Hubinger, who leads alignment science at Anthropic.
The researcher's warnings have gained significant attention outside of the company as well. A post he shared on X regarding AI safety garnered more than 100 million views. This high level of engagement suggests that the tension between rapid scaling and safety is a major point of public interest.
Despite his brief tenure, Coxon's departure carries weight because of his background. He holds a mathematics degree from Cambridge University and previously worked at OpenAI from 2023 until his move to Anthropic in May 2026. His ability to interpret the technical nuances of the IPO slide adds credibility to his public warnings.
The stakes of the IPO
Anthropic is currently navigating a high-stakes transition to become a public company. The company plans to list on the Nasdaq under the ticker "ANT." The roadshow pitch relies heavily on the company's computational advantages, specifically the ability to get better results for every dollar spent on compute.
The efficiency slide is a cornerstone of this narrative because it supports the claim of accelerating progress. For investors, this represents a path to market dominance and high returns. For researchers like Coxon, however, that same data point represents a potential loss of control over the technology being developed.
The situation creates a difficult dynamic for Anthropic as it seeks to go public. The company must balance the need to show investors that their technology is improving at a rapid pace while also addressing the safety concerns of its own staff. Coxon's public comments highlight the friction between these two competing priorities.
What to watch
- How Anthropic responds to Coxon's public comments regarding their IPO materials.
- Whether other researchers at Anthropic express similar concerns about superexponential gains.
- The final valuation and share price of Anthropic's Nasdaq listing.
Try it
Look up the difference between linear, exponential, and log scales to see how data visualization can change the perception of growth rates.
Sources and links
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by Crypto Briefing · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
