A slide from Anthropic’s IPO roadshow was one of the main reasons Jacob Coxon left the company, he said. He had worked there for about four months.
The slide charts the compute Anthropic needs to train its models over time. The axis is squeezed so hard that a 10x jump and a 100x jump look almost the same, and the curve still rises sharply.
Coxon takes this as a sign that algorithmic efficiency is getting better fast, and that the pace of those gains is picking up. “It’s an insane plot and I’m surprised they made it public,” he said.
Coxon puts Mythos on a trend line
In Coxon’s view, Anthropic’s Mythos model sits on that same curve. Mythos was a big step up in capability when it launched in the spring, and he said it was part of a pattern.
“Mythos was not a one-time deviation, but a point on a superexponential. They could already retrain Mythos with a fraction of the compute originally needed,” Coxon said.
He did not say how small that fraction is. The slide, as he describes it, shows a trend line with no figure for any individual model. His point about Mythos comes from how he reads the chart and from what he learned while working at the company.
The case for pausing next year
According to Coxon, algorithmic improvements are “on the verge of being automated.” He uses that to support his call for a pause on frontier model development next year, which would give competitors time to close the gap with the leading labs.
He says few companies would lose anything from it. A pause would “commoditize human-level intelligence. It would only harm OpenAI and Anthropic,” Coxon said.
Coxon’s background
Before he joined Anthropic, his most recent employer, Coxon trained AI models at OpenAI. He has been called the researcher who brought AI fears into the mainstream.
His public warnings in September about the risks of advanced AI systems got wide attention, and others in the industry echoed them. Among those concerns is the worry that advanced AI that improves itself could slip out of control. His reading of the roadshow slide ties into that worry. If automated research starts producing the gains that lower training costs, the curve he called insane would steepen with no one needing to push it.