„As a Language Model“: Chat Template Switches LLM Self-Referential Voice Von Hackernews 27. September 2026 security ORIGINAL QUELLE:arxiv.org Quelle: Hackernews Comments Tags: cant, comments, ki-modell, language, Security