Tag
Sean Goedecke argues that a recent paper showing frontier LLMs reading and editing neural network weights is not as superhuman as claimed, since a smart ML researcher could replicate the tasks with basic analysis.