Why do LLMs make stuff up? New research peers under the hood. - Ars Technica

Claude’s faulty “known entity” neurons sometime override its “don’t answer” circuitry.