Takeoff speeds, not AGI timelines, drive the real disagreement on AI risk
SamuelAlbanie · x · 2026-10-22
DeepMind researcher Samuel Albanie highlights Ajeya Cotra's essay "Takeoff speeds rule everything around me," noting that naive extrapolation works well for scaling laws and task horizons but breaks down for cyber vulnerabilities and high-severity agentic incidents — and most AI-progress disagreements stem from this.
Key points from the essay:
- A decade ago the x-risk debate centered on AGI timelines; today even skeptics accept 5-10 year timelines, yet disagreement over imminent risk remains huge.
- The author argues this disagreement is secretly still about timelines: skeptics combine short AGI timelines with very slow transformation of the real world, i.e. slow takeoff.
- Since the classic AGI-to-superintelligence definition is slippery, takeoff should be operationalized via real-world outcomes, with physical-tech progress (hardware, batteries, materials, chemistry, robotics, spacecraft) as the key yardstick.
- Thought experiment: if 2028 brings AI automating all intellectual work in these fields, how many "human-pace progress years" occur per calendar year? An accompanying chart shows fast takeoff (2029), slow takeoff (2032), and no-takeoff scenarios diverging dramatically.
Related event: METR data sparks debate: naive AI extrapolation works—sometimes(2 posts)→
More from AGI Musings
- Digital collage revives Mark Weiser's 1991 ubiquitous computing classic — GrantCuster · 2026-10-22
- Kelsey Tuoc: Living Under Annihilation's Shadow Harms People, Even If They're Right — EigenGender · 2026-10-22
- Researcher Warns AI Cyberattacks on Grid Possible Within a Year, Shares Prep List — cephaloform · 2026-10-22
- NYU professor: writing solo is more prone to the curse of knowledge than AI writing — ipeirotis · 2026-10-22
- Panel with Anthropic and H&F discusses AI deployment and the future of knowledge work — mathemagic1an · 2026-10-22
- Peter Yang asks Musk why game studios are shutting down while AI makes development easier — petergyang · 2026-10-22