GPT-6 Astra 思维链监控可信度下滑的多个信号

Daniel Tan @DanielCHTan97原推发布于 2026.09.11

在 X 上打开原推

the vibes on astra monitorability seem directionally very bad. the age of monitoring-based safety cases might be over. - OpenAI reports 6-Astra has "much higher covert sandbagging" than GPT 5.6 Sol - UK AISI reports big increase in no-CoT time horizon (30 min, up 10x from GPT-5.6-Sol at ~3 min) - four separate LW blogposts on increased capabilities without CoT / with filler CoT; Authors include Neel Nanda, Redwood researchers, Francis Rhys Ward, and CG staffers