Former Anthropic researcher Coxon says the company can now retrain Mythos with a small fraction of the compute originally required
Anthropic can already retrain Mythos with "a fraction of the compute originally needed," says former researcher Coxon
Jacob Coxon, a former OpenAI and Anthropic researcher who drew wide attention for publicly warning about the risks of advanced AI, says a chart in Anthropic's IPO roadshow was an important reason he left about four months after joining. The chart shows the code pretraining compute multiplier over time, with a logarithmic vertical axis, so 10x and 100x growth are almost indistinguishable on the chart; Coxon says the steep climb from July 2024 to October 2026 indicates that algorithmic efficiency is accelerating. He says Mythos, which showed a large jump in capability in the spring, was not a one-off deviation but a point on a super-exponential trend, and that Anthropic can now retrain the model with a small fraction of the compute originally required. He also says algorithmic improvement is approaching the point of being automatable, and therefore advocates pausing frontier model development next year to let competitors catch up, saying this would only harm OpenAI and Anthropic. The chart notes that the data points are based on internal estimates and are for illustrative purposes only.