The Information:GPT-6.1 Astra 再度延期
报道称 OpenAI 因安全顾虑砍掉一个较弱模型,Astra 用 loop transformer 导致推理难监控,或需接受美政府最长 30 天测试。


GPT-6.1 Astra is facing yet another delay. According to The Information, OpenAI is still working to make its next flagship safe enough to release after reportedly canceling a less capable model over safety concerns. The model is said to use loop transformers, which repeatedly process problems internally but make its reasoning harder to monitor. Following the Hugging Face incident, OpenAI reportedly redirected around 10% of its compute toward alignment and safety. The model could also face up to 30 days of US government testing before release. Employees still expect one or two new models before year end, even if Astra isn’t ready.