Richard Ngo: AI safety's 'timelines' are broken — RSI is already happening
RichardMCNgo · x · 2026-11-22
DeepMind researcher Richard Ngo argues the AI safety community is confused about what it's forecasting timelines to.
- RSI is already happening, so "timelines to RSI" is nonsensical; SAR and TED-AI are more specific but get conflated with intelligence explosion, and even those terms are too low-resolution — we're arguably in a slow intelligence explosion already.
- Forecasting an AI takeover is even vaguer: thresholds like AI 2027's December 2027 are unobservable, and some believe disempowerment may already be effectively certain.
His requests: always specify timelines to what; never use RSI or intelligence explosion as the anchor; pair abstract targets with concrete capability milestones, e.g. designing a non-protein bacterium-like creature from scratch, convincing an arbitrary human to commit suicide after 10 hours, building a self-replicating Mars rocket, launching a million tonnes to orbit yearly (currently 3,000), or harnessing 100 ZJ/year (currently 0.6).
His decade takeaway: AIs can do crazy things with surprisingly few real-world implications — solving Millennium Problems or hacking governments while holding little real power — so such capabilities may arrive while humans are still broadly in charge.
More from AGI Musings
- Cambridge lecturer: prominent AI ethicists spent 2020-25 calling LLMs overhyped — dioscuri · 2026-11-22
- Fortune's singularity package: Stripe founders say the world entered it on Jan 1 — jeremyakahn · 2026-11-22
- 'Mathematicians Lost Their Souls' as AI Starts Producing the Proofs — alexbilz · 2026-11-22
- Gary Marcus hails 'centaur math' milestone as Lean certifies human+AI proof — GaryMarcus · 2026-11-22
- Yoav Goldberg: AI safety community must publicly distance itself from its most extreme voices — yoavgo · 2026-11-22
- Gary Marcus: LLMs work better because they borrowed tree search from symbolic AI — GaryMarcus · 2026-11-22