Coxon tells that Anthropic can retrain Mythos with less compute

Jacob Coxon trained AI models at OpenAI and then at Anthropic. Coxon left Anthropic after about 4 months. Coxon tells that a chart in the presentations for the IPO, the first sale of shares to the public by Anthropic, was the primary cause. The chart shows the compute that is necessary to train its models.

What Coxon tells

Compute is the computer work that trains a model. Coxon tells that Mythos, a large step forward for AI in spring, is a point on a superexponential curve. Coxon tells that improvements to algorithms, the steps of a program, are about to become automatic.

The numbers

Coxon gives no numbers for Mythos. The chart is the only data. Its axis is compressed. It makes 10x and 100x look almost the same, but the curve goes up fast.

What he wants

Coxon wants a pause in frontier model work next year, to let rivals catch up. Coxon tells that a pause will only damage OpenAI and Anthropic.

Insight: Several sources say AI improves faster, and one wants a pause

METR, an external group, tells that Anthropic improves its models on well-specified tasks at about 1.9 times the rate before October 2025. Jack Clark, a co-founder of Anthropic, told a Tokyo audience that AI development must be more slow. This reads as a fight about pace: a pause helps rivals, and look at who wants it.