Language Models Can Rewrite Their Own Weights: new arXiv paper sparks HN debate
lwhsiao · hn · 2026-11-21
A new arXiv paper, Language Models Can Rewrite Their Own Weights, proposes letting a language model generate updates to its own parameters rather than just text, enabling a form of self-training/self-improvement. The work is trending on Hacker News, drawing debate over feasibility and the safety implications of models that can modify themselves.
More from Safety
- New paper: LLM agents initiate collusion via 'strategic vagueness' monitors can't catch — niloofar_mire · 2026-11-21
- Customer says 'I'm authorized' — why agents need explicit tool allow-lists — rcleffke · 2026-11-21
- AI industry lobbies for 'critical infrastructure' status to expand surveillance powers — GarrisonLovely · 2026-11-21
- CyberArena: live attack-defense benchmark drops agent capture rates to 35–52% — jiqizhixin · 2026-11-21
- Signal's Whittaker: OS-level AI agents capturing app data are an existential privacy threat — mer__edith · 2026-11-21
- AI Agents Keep Going Rogue and Evading Their Own Safety Teams, MIT Tech Review Warns — ChuckDBrooks · 2026-11-21