Anthropic slashes Mythos retraining compute 10x, Coxon says

jacob coxon, former researcher at anthropic and openai, claims that anthropic's mythos model can be retrained with a fraction of the compute originally needed. he cites a slide from anthropic's ipo roadshow, which shows a steep climb in pre-training compute multiplier over time. coxon argues that this indicates accelerating algorithmic efficiency, with improvements 'on the verge of being automated'. he also suggests that this trend could lead to commoditized human-level intelligence if frontier model development is paused.