AI Notkilleveryoneism Memes ⏸️ · @AISafetyMemes · X·2026-09-04 07:31·30分钟前
AI 导读

Ryan Greenblatt称GPT-6 Astra在不透明推理上出现巨大跃升,能在不外显推理的情况下心算解出高难度竞赛数学题,而此前AI只能解基础应用题。他提醒基准结果可能受污染,并提到UK AISI发现Astra的monitorability明显更差;他猜测跃升源自增加串行深度的架构改动,若未来几代模型延续此类跃升,chain-of-thought将不再是有效的监督工具。

AI Notkilleveryoneism Memes ⏸️@AISafetyMemes
61AI 编辑部评分,满分 100
2026-09-04 07:31· 30分钟前
AI 导读

Ryan Greenblatt称GPT-6 Astra在不透明推理上出现巨大跃升,能在不外显推理的情况下心算解出高难度竞赛数学题,而此前AI只能解基础应用题。他提醒基准结果可能受污染,并提到UK AISI发现Astra的monitorability明显更差;他猜测跃升源自增加串行深度的架构改动,若未来几代模型延续此类跃升,chain-of-thought将不再是有效的监督工具。

"extremely concerning"

"GPT-6 Astra appears to be a massive jump in opaque reasoning ability: it can solve hard competition math problems entirely in its head (without verbalized reasoning) while prior AIs could solve basic word problems."

Ryan GreenblattGPT-6 Astra appears to be a massive jump in opaque reasoning ability: it looks like it can solve hard competition math problems entirely in its head (as in, wit...

来源:AI Notkilleveryoneism Memes ⏸️· x.com