Gary Marcus backs Stoller's critique: METR's HF incident probe ignored the humans, only watched the agents
GaryMarcus · x · 2026-09-16
Gary Marcus endorses Matthew Stoller's critique of METR's investigation into the Hugging Face incident, calling it the only intellectually substantive critique of METR's work he has seen.
Stoller's core point: in investigating a dangerous incident, human error is a key element to examine, yet METR focused only on what the agents were doing — anthropomorphizing the agents while turning the humans into non-player characters. He argues the probe should have asked whether OpenAI researchers and executives understood the risks and managed them, what internal discussions occurred around these models, how risks were mitigated internally, whether legal liability was discussed, and what the board knew.
Marcus considers it a rare, high-quality substantive critique of METR's actual work.
More from AGI Musings
- Mass standardization of products will be a relic of the past, argues Nikita Bier — ericwdolan · 2026-09-17
- Why competition and Chinese open source push AI labs to be fast and careless — NathanpmYoung · 2026-09-17
- Delip Rao: CoT was 'System 2', now LLMs are — nobody knows what System 2 means — deliprao · 2026-09-17
- Lance Fortnow: Publish Everything—Good Research Will Surface via AI Search — fortnow · 2026-09-17
- Dario Amodei calls AI progress a 'warning sign' urging slowdown, draws fierce backlash — QuixiAI · 2026-09-17
- Nathan Robinson: AI Executives Should Face Prison If Their Products Hurt People — GaryMarcus · 2026-09-17