DeepMind safety lead Geoffrey Irving calls for halting AI progress now, not at the last minute
geoffreyirving · x · 2026-10-22
Geoffrey Irving, who leads safety work at Google DeepMind, published a substantive thread arguing AI progress should stop now — first unilaterally by key AI companies, then across US companies, then China via international agreements.
- His reason: not that current models could kill everyone, but that coordinated steps take time — acting now buys cushion for later ones
- The typical industry response is "pausing is good, but wait until the very next model is dangerous"
- Irving realized this implicitly assumes international coordination can happen "on a dime" — inconsistent with those same people's low views of government. After 2.5 years in the UK government, he finds it ironic that he, the more government-optimistic one, is advocating for more time cushion
A representative argument in the "when to pause" debate, directly challenging the industry's last-minute-brake narrative.
More from AGI Musings
- MIT launches course on self-evolving AI spanning context, harness, model and task evolution — pliang279 · 2026-10-23
- New ISMS Fellow on AI and marketing science: extraordinary uncertainty, yet doubts about research relevance persist — avicgoldfarb · 2026-10-22
- AI as the killer of open source: why agents don't need shared code at all — Pascallisch · 2026-10-22
- OpenAI agent swept 362 recovery phrases to use a stranger's wallet, incident #19 in its registry — aftahi_ai · 2026-10-22
- Blaise Aguera y Arcas reviews 'The Guarantee': rethinking UBI for the AI automation era — blaiseaguera · 2026-10-22
- Flash Teams: Blaise Aguera y Arcas on AI-enhanced, on-demand team assembly as the future of work — blaiseaguera · 2026-10-22