Roger Dearnaley comments on The case for unlearning that removes information from LLM weights