Coxon: Anthropic can retrain Mythos with a fraction of the compute

Coxon's argument rests on a chart Anthropic itself published for its IPO roadshow, which he describes as "an insane plot" and says he is surprised the company made public. The chart plots pre-training compute needs over time with a compressed axis, so a 10x and a 100x increase barely look different, yet the curve still climbs steeply. That steepness is what Coxon reads as superexponential algorithmic progress, and he treats Mythos as a point on that trend rather than an outlier.

His proposed pause is framed as a leveling move rather than a brake on progress: he writes that it would "commoditize human-level intelligence" and "only harm OpenAI and Anthropic." The claim is notable because it comes from someone who previously trained models at both OpenAI and Anthropic and who drew wide attention in September with public warnings about advanced AI risks, including the fear that self-improving systems could get out of control.

Anthropic's own caveat on the slide is that the data points are based on internal estimates and are for illustration only, which leaves the strength of the trend open to interpretation.