A chart about the computing power needed to train Anthropic’s models became a turning point for former researcher Jacob Coxon. He says it helped convince him to leave the company after about four months, and that its implications extend beyond one model: in his view, algorithmic efficiency is improving at an accelerating rate.
What Coxon saw in the chart
The chart appeared in material Anthropic released for its IPO roadshow. It tracked how much compute the company needed to train its models over time. Coxon says the chart’s axis was compressed so that a 10x increase and a 100x increase looked barely different.
Even with that compressed scale, he says, the curve still rose steeply. Coxon interprets that shape as evidence that algorithmic efficiency is not just improving quickly, but accelerating. The chart, in his reading, suggests that models may be trained with less compute than earlier efforts required.
He described the plot as “an insane plot” and said he was surprised Anthropic made it public. The source does not provide the chart’s underlying figures, so the case presented here is Coxon’s interpretation of what it shows.
Why Mythos matters to his argument
Coxon connects the chart’s trend to Anthropic’s Mythos model, which he describes as a major capability jump in the spring. He argues that Mythos should not be treated as a one-off departure from the pattern. Instead, he sees it as another point along what he calls a “superexponential” trajectory.
“They could already retrain Mythos with a fraction of the compute originally needed,” Coxon writes. The claim links two ideas: that algorithmic improvements are making training more efficient, and that those gains could let a company reproduce a model using much less compute than before.
That possibility matters to his broader concern because he sees efficiency gains as continuing rather than ending with a single breakthrough. If his reading is right, a capability jump may signal an ongoing trend. The source, however, gives Coxon’s assessment rather than independent measurements to establish how much compute a retraining run would require.
A proposed pause in frontier development
Coxon says algorithmic improvements are “on the verge of being automated.” Based on that concern, he wants frontier model development paused next year, giving competitors time to catch up.
He argues that a pause would “commoditize human-level intelligence” and “only harm OpenAI and Anthropic.” In this framing, slowing the leading companies could spread access to capabilities by allowing other competitors to close the gap. The source does not detail how such a pause would work or how long it would last.
His proposal follows from his view that progress in model training is accelerating and may become automated. If advances keep reducing the resources needed to reproduce powerful models, then the lead held by frontier developers could be less durable. That is the logic behind his call for competitors to catch up while development is paused.
His warnings about advanced AI
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, concerns echoed by other people in the industry.
One fear described in the source is that advanced, self-improving AI could get out of control. Coxon’s comments about compute and algorithmic efficiency sit alongside that concern: the faster models can improve, and the less compute needed to train them, the more pressing questions about the pace and control of development may become.
His departure, as he describes it, was tied in part to a chart that made the company’s efficiency trend visible. His interpretation of that chart, and the call for a pause that follows from it, offer a window into why he believes model development has reached a consequential point.