Former Anthropic researcher Jacob Coxon says Mythos could now be retrained for a fraction of its original compute - eGamers.io
Former Anthropic researcher Jacob Coxon claims the company's Mythos AI model could now be retrained using only a fraction of its original computational power, based on a slide from Anthropic's IPO roadshow.
Coxon, who worked at Anthropic for about four months, said a slide from the company's IPO roadshow was a key factor in his departure. The slide charts Anthropic's compute needs for training AI models over time, with a compressed axis that makes exponential increases appear similar.
He interprets this as evidence that algorithmic efficiency is improving rapidly, with the pace of those gains accelerating. "It's an insane plot and I'm surprised they made it public," Coxon commented.
Mythos Efficiency
In Coxon's view, Anthropic's Mythos model sits on this same efficiency curve. When launched in spring, Mythos represented a significant capability improvement, but he believes it follows a pattern rather than being an anomaly.
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 didn't specify how small that fraction might be, basing his assessment on his interpretation of the chart and his experience at the company.
Call for Pause
The researcher believes algorithmic improvements are "on the verge of being automated," which supports his call for a pause on frontier model development next year. He argues this would give competitors time to catch up with leading AI labs.
According to Coxon, a temporary pause would "commoditize human-level intelligence" and would primarily harm only OpenAI and Anthropic, while most other companies would experience minimal negative impact.
Related
