Low-stakes control (ControlConf talk)
A talk by Vivek Hebbar (Redwood Research) at FAR.AI's ControlConf. 30 minutes.
This talk introduces the low-stakes regime of AI control — settings where no single action can cause a catastrophe, so the danger accumulates across many weakly-incriminating actions. It extends the diffuse-threats framing from the threat-modeling readings in the first module and previews the methods the rest of this module examines in depth.