Tag
A new METR study finds AI's ability to complete complex, hours-long tasks is doubling every 3.6 months, with faster progress on easier-to-verify tasks and potential human-equivalent 200-hour task completion by July 2027.
METR updated its widely cited time-horizon plot, finding frontier AI capabilities still doubling roughly every four months while showing that models remain weaker than humans at judgment-heavy, open-ended tasks such as designing AI evaluations and scoring philosophical arguments.
Toby Ord 的文章探讨 AI 模型通过更多 token 扩展的效果是否与人类通过更多时间扩展的效果相当,分析了 METR 的对比数据,并讨论了这一问题对 AI 自我改进与智能爆炸前景的影响。