Former Anthropic researcher quits over an IPO roadshow chart

Jacob Coxson, a researcher who has worked at OpenAI and Anthropic, says a slide shown during Anthropic's IPO roadshow was a key reason he left. He only stayed at Anthropic for about four months.

The chart showed how much computing power AI models need for training. It used compressed axes, so even a 10x or 100x jump in compute looks small on screen, yet the curve still climbs sharply. He sees this as proof that algorithm efficiency is improving fast, and speeding up.

In his view, the big performance leap of Anthropic's Mythos model this spring was no accident, but a step in this super-fast growth. Retraining Mythos now only takes a small part of the original compute.

He wants to pause frontier AI model work next year to give rivals a chance to catch up, making human-level intelligence widely available. He had already publicly warned that advanced AI could slip out of human control.