DeepMind's Alex Irpan doubles his p(doom) estimate from 2% to 4%
AlexIrpan · x · 2026-09-13
Alex Irpan, who left DeepMind's robotics team in 2024 to join its AI safety group, has doubled his subjective estimate of catastrophic risk from 2% to about 4%. His takeaway: safety isn't improving at the rate needed to keep up with capabilities, "it's not looking good." The post links back to his 2024 essay "I'm Switching Into AI Safety," which covered his 8 years in robotics, the specialization-vs-transferability tradeoff, and why he expected robotics-style challenges to spread to other fields.
Key points:
- Two years ago: 2% chance of doom; now roughly 4% — "vibe about 2x worse"
- Safety progress not keeping pace with capability gains
- Original career switch driven by wanting a new challenge and belief his robotics experience would transfer to safety work
More from AGI Musings
- Regulation is coming for open and closed AI models alike — the question is proactive or reactive — benjamin_warner · 2026-09-13
- AI now generates solutions faster than humans can verify them, warns researcher — srchvrs · 2026-09-13
- AI generation now outpaces human verification, deepening 'knowledge debt' — srchvrs · 2026-09-13
- beffjezos calls for decentralized sovereign AI open stack after Casado laments frontier consolidation — nptacek · 2026-09-13
- Are AI minds shielded from suffering by their inability to remember? — repligate · 2026-09-13
- AI filmmakers are filmmakers: a 3D filmmaker's ComfyUI-powered rebuttal — Delphoi_Studio · 2026-09-13