Coxon says Anthropic can retrain Mythos with far less compute
Former OpenAI and Anthropic researcher Jacob Coxon says a slide presented during Anthropic’s IPO roadshow was a major reason he left the company after roughly four months. The graphic tracks the computing resources required to train models over time. Its compressed, logarithmic scale makes tenfold and hundredfold increases appear relatively similar, yet the curve still rises sharply. Coxon interprets this as evidence that algorithmic efficiency is improving at an accelerating pace.
He argues that Anthropic’s Mythos model, described as a major capability advance in the spring, fits this pattern rather than representing an isolated event. According to Coxon, Anthropic could already retrain Mythos using a fraction of the computing resources initially required. Anthropic says the chart’s points are based on internal estimates and are intended for illustration.
Coxon also says algorithmic progress is close to becoming automated. He supports pausing frontier-model development next year, arguing that a pause would give competitors time to catch up and would make human-level intelligence more widely available. He previously attracted attention for warnings that advanced, self-improving AI systems could become difficult to control.
Written by Sakhanda Wire AI from reporting by The Decoder. Facts, quotes and full context are in the original article below.