Ex-Anthropic Researcher Coxon Says an IPO Roadshow Slide Drove Him to Quit

The argument reaches investors just as they are pricing the company. Anthropic is expected to price its Nasdaq listing next week. In the roadshow, the company has pitched its ability to get more out of each unit of compute as a competitive advantage. Coxon draws a different conclusion: that the same gains make the field more dangerous and should prompt a pause.

According to Coxon, the chart plots how much compute Anthropic needs over time to train its models, on an axis so compressed that 10x and 100x increases look almost the same. The curve still climbs steeply. "It's an insane plot and I'm surprised they made it public," he wrote.

Coxon said Mythos, the Anthropic model that marked a large jump in capability in the spring, fits this trend and was not a one-off. "They could already retrain Mythos with a fraction of the compute originally needed," he wrote. He also said algorithmic improvements are "on the verge of being automated."

The roadshow video, which Anthropic posted on its US investor site, presents the company's view of efficiency. In it, co-founder Tom Brown describes "total effective compute" as physical capacity multiplied by the compute multipliers the company discovers. He gives the example of a 4x multiplier letting Anthropic train the same model with 25 units of compute instead of 100. According to developer Simon Willison's transcript of the video, the company says it expects such gains to speed up as it approaches recursive self-improvement, with Claude eventually finding multipliers itself.

Coxon wants frontier model development paused next year so competitors can catch up. He wrote that a pause would "commoditize human-level intelligence. It would only harm OpenAI and Anthropic." He drew wide attention in September with public warnings about the risks of advanced AI, including the fear that self-improving systems could slip out of human control.

The Decoder did not report any response from Anthropic to Coxon's claims.