Coxon says Anthropic could retrain Mythos with a fraction of the compute

Coxon says Anthropic could retrain Mythos with a fraction of the compute

Jacob Coxon, a former OpenAI and Anthropic researcher described as someone who brought AI fears into the mainstream, says a slide Anthropic released for its IPO roadshow was a key reason he left the company after only about four months.

The chart shows how much compute Anthropic needs over time to train its models. Its axis is compressed, so a 10x and a 100x increase barely look different. Even so, the curve still climbs steeply. Coxon says that shows algorithmic efficiency is improving not just quickly but at an accelerating rate. "It's an insane plot and I'm surprised they made it public," he writes.

Coxon argues that Mythos, Anthropic's model that was a major capability jump in the spring, was part of this accelerating trend. In his words: "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."

He also says algorithmic improvements are "on the verge of being automated." That is why he wants frontier model development paused next year so competitors can catch up. A pause, he writes, would "commoditize human-level intelligence. It would only harm OpenAI and Anthropic."

Coxon previously trained AI models at OpenAI and, most recently, Anthropic. He drew wide attention in September with public warnings about the risks of advanced AI systems, echoed by others in the industry. One of their fears is that advanced, self-improving AI could get out of control.

Key facts

  • Jacob Coxon says a slide from Anthropic's IPO roadshow was a key reason he left the company after about four months.
  • The chart shows the compute Anthropic needs over time to train its models; its compressed axis makes a 10x and a 100x increase look barely different.
  • Coxon argues the steep curve shows algorithmic efficiency improving at an accelerating rate, and that Anthropic could already retrain Mythos with "a fraction of the compute originally needed."
  • He says algorithmic improvements are "on the verge of being automated" and wants frontier model development paused next year so competitors can catch up.

Why it matters

The claim is that the compute needed to reach a given capability level is falling faster and faster. If Coxon is right, a model like Mythos, a major capability jump in the spring, would be cheaper to rebuild than it was to build the first time. He draws a policy conclusion from this: pause frontier development next year so competitors can catch up. He also says a pause would commoditize human-level intelligence and would only harm OpenAI and Anthropic.

Who it affects

Anthropic and OpenAI are the companies Coxon names as the ones a pause would harm. Competitors are the ones he wants to give time to catch up. Anyone following Anthropic's IPO roadshow materials is affected too, since the slide is the trigger for his departure and his argument.

How to use it

There is nothing to install or adopt here. It is one person's reading of a public chart. The practical takeaway is to look at the chart's compressed axis yourself: on it, a 10x and a 100x increase in compute look nearly the same, and Coxon's argument rests on the curve still climbing steeply despite that.

How solid is it

This is Coxon's own argument, not a verified finding. The source gives no figures for the compute Anthropic needs or for how much less compute retraining Mythos would take, only "a fraction" per Coxon. The source does not say Anthropic has confirmed or commented on the claim, and it gives no response from OpenAI or Anthropic to the proposed pause. Coxon's standing is that he trained models at OpenAI and, most recently, Anthropic.

Risks and caveats

The source gives no date for the IPO or the roadshow and does not say when Coxon left, so the timeline is unclear. He spent only about four months at Anthropic. The headline claim about retraining Mythos is his inference from a chart with a compressed axis. His call for a pause is a position tied to his public warnings about advanced AI, including the fear that self-improving AI could get out of control.

“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”

— Jacob Coxon