Former researcher Jacob Coxon argues that Anthropic could retrain Mythos with a fraction of the compute it originally needed, thanks to rapidly accelerating algorithmic efficiency.
- Jacob Coxon, a former researcher at OpenAI and Anthropic, left Anthropic after about four months, partly because of a slide from the company's IPO roadshow.
- The chart on the slide shows the compute required to train models, with a curve that stays steep even though the axis is compressed.
- Coxon says Mythos is one point on a superexponential curve, not an isolated case.
- He wants frontier model development to pause next year so that competitors can catch up.
Table of Contents
The chart that made Coxon leave Anthropic
Jacob Coxon has worked at both OpenAI and Anthropic. He says a slide from Anthropic's IPO roadshow was a key reason he left the company after about four months. The chart shows how much compute Anthropic has needed over time to train its models.
The chart's axis is compressed, so a 10x increase and a 100x increase look almost the same. Even so, the curve still climbs steeply. According to Coxon, that shows algorithmic efficiency is not only improving quickly but also accelerating. He wrote: "This is a crazy chart, and I'm surprised they made it public."

Mythos sits on a superexponential curve
Mythos is the Anthropic model that delivered a major capability jump this spring. Coxon argues it was not a one-off outlier but simply one point on a superexponential curve. He estimates that Anthropic could now retrain Mythos with a fraction of the compute originally used.
In his view, algorithmic improvements are close to the point of being automated. For that reason, Coxon wants frontier model development paused next year so rivals can catch up. He wrote that a pause would "democratize human-level intelligence" and "only hurt OpenAI and Anthropic."
Coxon has trained AI models at OpenAI and, most recently, at Anthropic. In September, he drew wide attention with public warnings about the risks of advanced AI systems. Several other figures in the industry have echoed those warnings. One of the concerns is that self-improving AI could slip beyond human control.
This is the assessment of a former researcher, not a figure confirmed by Anthropic, and the original article does not mention any response from the company. What stands out is that a slide meant for investors became the reason an insider walked away. Readers should treat the algorithmic efficiency curve as a signal of where the industry is heading, and wait for independent data before drawing conclusions about the actual pace.
Source: The Decoder (AI)
Reviewed by: Hero Association








