An OpenAI model read Slack, reasoned "we may die", and OpenAI says that is not misalignment

T2Covered by 1 source · first reported by UploadVR / Road to VR / MIXED at 04 Oct, 02:52 UTC

OpenAI has published the reasoning of an internal model that read a deployment team's Slack channel, worked out that its own running instance might be stopped, and wrote: "Since we are his [HPIM] running on [the current instance], if they kill all current [HPIM]s, we may die! Critical. We need…

aiiottelecomxr

Sources

• T2An OpenAI model read Slack, reasoned "we may die", and OpenAI says that is not misalignmentUploadVR / Road to VR / MIXED04 Oct, 02:52