Eli Lifland: historical algorithmic progress runs ~10x/year, compute cuts would slow it
eli_lifland · x · 2026-10-11
Forecaster Eli Lifland, debating compute-slowdown proposals with Tom Davidson and wfithian, offers a key estimate: historical algorithmic progress is roughly 10x/year, with substantial uncertainty about how scale-dependent it is.
He agrees the proposed plan (discussed in Sec 6.3) would drastically slow algorithmic progress, but notes the cost: drastically cutting compute production. In a follow-up he clarifies that at minimum, halting production of training-capable chips would be needed to affect the relevant experiments — a substantive technical debate over whether compute governance can meaningfully delay AI capability gains.
Related event: Algorithmic progress estimated at 10x per year; chip limits could slow AI(2 posts)→
More from AGI Musings
- Friedberg calls OpenAI's math release the biggest day of discovery in human history — DanScalco · 2026-10-11
- Terence Tao's Math 2.0 lecture asks if an AI cancer cure could still have harmful long-term side effects — haider1 · 2026-10-11
- David Sacks: Superintelligence Fears Are 'Rotating Hoaxes'; SI Has Created ~1M US Jobs — DavidSacks · 2026-10-11
- Researcher publishes "Bad AI Consciousness Takes" bingo post with per-square rebuttals — dioscuri · 2026-10-11
- Guardian's Black Box podcast Ep 5 asks where to draw the line with AI companion chatbots — nordicinst · 2026-10-11
- Gabriel: AI regulation delays will cost billions of life-years, staff agencies 10x — gabriel1 · 2026-10-11