Why compute-extrapolation arguments for AGI timelines are flawed, and why Kurzweil's 1999 bet was luck
phl43 · x · 2026-09-06
phl43 lays out a systematic critique of a common AGI prediction genre: estimate the compute needed to emulate a human brain, extrapolate compute growth, and conclude AGI arrives shortly after crossing that threshold.
Even granting the safest assumption—that compute keeps growing at a constant or accelerating rate—he argues the reasoning collapses because our understanding of the brain is far too poor: the models used to estimate emulation requirements are almost certainly wrong, possibly by a lot, likely missing key biological mechanisms.
Context: the debate started when @akarvonen defended Kurzweil's 1999 prediction built on extrapolating compute growth against human-brain FLOPS ('in retrospect, that's all you need to predict AGI'). phl43 calls it a terrible argument and plans a full post; @stanislavfort adds that we don't even know how the brain implements its learning algorithm.
Related event: Debate Erupts Over Whether Compute Extrapolation Can Predict AGI(6 posts)→
More from AGI Musings
- Video editor says AI agent Astra makes him believe his job is a year from replacement — Street-Ad3815 · 2026-09-06
- Bryan Cantrill rails against LLM-authored writing: 78% of devs stop reading on sight — mgill25 · 2026-09-06
- Post-GPT-6, Reddit rethinks when AGI will actually change the world — NoSignificance152 · 2026-09-06
- Fredde r: GTA VI will be a museum to the last grand 100% human-made digital world — ZeroStateReflex · 2026-09-06
- Blogger argues the next trillion-dollar AI business will be an AI game studio — _jaydeepkarale · 2026-09-06
- a16z's Julie Yoo on why everyone will have a personalized AI doctor — a16z Podcast · 2026-09-06