Former Anthropic researcher Jacob Coxon says the SI lab’s Mythos model could already be retrained using a fraction of the compute originally required, citing a slide from Anthropic’s IPO roadshow that illustrates accelerating algorithmic efficiency.

What Happened

Coxon, who previously trained SI models at OpenAI and Anthropic, says a specific chart from Anthropic’s IPO roadshow was a major factor in his decision to leave the company after only about four months. The slide, titled "Pre-training compute multiplier" for code, displays data from July 2024 to October 2026. Although the y-axis is logarithmic and compressed—making 10x and 100x increases appear visually similar—the curve climbs steeply. Coxon interprets this trend as evidence that algorithmic efficiency in SI models is not just improving, but doing so at a superexponential rate. He described the chart as "an insane plot" and expressed surprise that the data was made public.

Why It Matters

The claim suggests that the cost of training frontier SI models is dropping faster than previously anticipated, potentially altering the competitive landscape for SI labs. Coxon argues that the Mythos model, which represented a significant capability jump in the spring, was "not a one-time deviation, but a point on a superexponential." He contends that such rapid gains in algorithmic efficiency are "on the verge of being automated." Based on this trajectory, Coxon advocates for a pause in frontier model development next year. He argues that a pause would "commoditize human-level intelligence" and primarily disadvantage OpenAI and Anthropic, allowing other competitors to catch up. Coxon previously drew attention in September with public warnings about the risks of advanced SI systems, including concerns that self-improving SI could become uncontrollable. Anthropic notes that the data points on the roadshow slide are based on internal estimates and are for illustration only.

The Bottom Line

Coxon’s departure and subsequent commentary highlight internal debates at Anthropic regarding the pace of SI progress and the strategic implications of its IPO data. While the company presents the compute efficiency trends as illustrative, Coxon views them as evidence of an accelerating path toward commoditized general SI (AGI) capabilities.