Anders Sandberg: by the time an AI research-speed doubling is confirmed, takeoff may be near
anderssandberg · x · 2026-11-08
Oxford philosopher Anders Sandberg unpacks the METR report on recursive self-improvement: he hasn't strongly updated toward believing AI lab research speed is really doubling, but argues that's not reassuring.
- Using Bostrom's 2014 I'=(c+I)I takeoff model, human + AI effort yields returns to scale and a finite-time singularity; once capacity gains one extra doubling, only about half an initial doubling time (a year) remains
- The core problem: we're estimating a second derivative — Gaussian noise in benchmarks makes the change rate look flat until very late, delaying any confident update
- Real-world benchmarking of change rates is rare, compartmentalized, and quality-dependent, making updates even rarer
- His takeaway: act to mitigate faster-than-expected scenarios. He remains a long-ASI-timeline person, but "long" has shrunk from many decades to 10+ years
More from AGI Musings
- Nassim Taleb: People afraid of AI secretly feel like impostors with no edge — GhaffariMaani · 2026-11-08
- Google Brain saying: you get more compute only after proving you used it well — karinanguyen · 2026-11-08
- Embodiment vs physical instantiation: researchers debate whether LLMs already have a toehold — AnnaCiaunica · 2026-11-08
- Datacenter Bottlenecks Favor Hyperscalers: Compute Holders Win as Training Slows — adityaag · 2026-11-08
- AI safety researcher: METR 2x report matches my two-year-old prediction — JeffLadish · 2026-11-08
- If a lab achieves RSI, would it still need human researchers? — TheMrCurious · 2026-11-08