The slide that weighed on his departure
Jacob Coxon, a researcher who moved from OpenAI to Anthropic, explained that a slide published by the company for its IPO roadshow was one of the main reasons he left the group after about four months. The news is reported by The Decoder. The slide shows the amount of compute Anthropic has needed over time to train its models, and according to Coxon the axis is compressed to the point that tenfold and hundredfold increases appear almost indistinguishable.
Despite this compression, according to Coxon the curve continues to rise steeply. It is this element, in his account, that indicates that algorithmic efficiency is improving not only rapidly but at an accelerating rate. Coxon calls the chart extreme and says he is surprised that the company made it public.
Mythos and the exponential-curve thesis
Coxon links the Mythos model to this trend, describing it as a significant leap in capability that occurred in the spring. In his reading, Mythos is not an isolated anomaly but a point on a superexponential growth curve. According to Coxon, Anthropic would already be able to retrain that model with a fraction of the compute originally required.
These are claims attributed to Coxon, not data independently verified in the material. The slide would confirm, again according to the researcher, a dynamic in which development costs tend to fall thanks to algorithmic improvements, an aspect that directly affects the economic sustainability of the large labs.
The call for a pause and its intended recipients
Coxon argues that algorithmic advances are close to being automated. He therefore calls for a pause in frontier model development next year, with the stated aim of allowing competitors to catch up with the leaders. According to his thesis, a pause would turn human-level intelligence into a commodity and would harm only OpenAI and Anthropic.
His position is not new: Coxon had already drawn public attention in September with warnings about the risks associated with advanced artificial intelligence systems, a position shared by other figures in the industry. One of the concerns expressed is the possibility that an advanced, self-improving AI could escape control. However, the details of the slide, its actual technical meaning and the existence of independent data supporting the thesis on Mythos's retrainability remain to be verified.
