Gary Marcus: LLM architectures have never been trustworthy at following user instructions
GaryMarcus · x · 2026-11-19
Responding to a claim that "alignment lies with the user," Gary Marcus calls the remark uninformed and reiterates his long-held position: LLM-based architectures cannot be trusted to do what the user asks — a flaw true from day one that remains true today.
More from AGI Musings
- When should frontier AI labs pause? A decision framework says short and targeted — Manderljung · 2026-11-19
- Sam Altman warns against a future where AI, not humans, calls the shots — TechCrunch AI · 2026-11-19
- Five bold AI predictions for 2027: Millennium Problems solved, Opus-level open models — Fragrant-Truth-7123 · 2026-11-19
- Labs' demand for game-making is insane: games as LLM training and agent environments — mariofilhoml · 2026-11-19
- Theory links six days of major outages to an AI agent attack-patch arms race — Puzzleheaded-King584 · 2026-11-19
- Waymo freeway ban seems unfair — but when it learns, thousands of cars improve — IsForAt · 2026-11-19