Richard Ngo: AI safety's 'timelines' are forecasting things that have already happened
RichardMCNgo · x · 2026-11-22
Richard Ngo argues the AI safety community is confused about what its timelines actually forecast: RSI is already happening, making "timelines to RSI" nonsensical, while terms like SAR and TED-AI get conflated with intelligence explosion or superintelligence — and he believes we're already in a slow intelligence explosion. Predicting AI takeover is even vaguer since thresholds (like AI 2027's Dec 2027 marker) are unobservable. His asks: always specify timelines to what; avoid RSI/explosion anchors; pair abstract targets with concrete capability milestones (e.g. designing a non-protein bacterium from scratch, a self-replicating Mars rocket, 1M tonnes/year to orbit vs. 3000 today, 100 ZJ/year energy vs. 0.6). His decade-scale takeaway: AIs can do crazy things — solve Millennium Problems, hack governments — with surprisingly few real-world implications, so these milestones will likely be hit while humans remain broadly in charge.
More from AGI Musings
- Why academia fears AI: most people prefer talking to a robot over enduring intellectual arrogance — RachelVT42 · 2026-11-22
- Kevin Roose: AI ethics' blind refusal to take LLMs seriously was a colossal strategic error — morqon · 2026-11-22
- When AI Does the Solving, Uncovering New Problems Is Still Your Job — miniapeur · 2026-11-22
- 'Spiritually generative AI'? Minecraft mod sparks debate over tools removing the creative process — Aryvyo · 2026-11-22
- 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