Frontier models already rival the human brain on energy per thought
ricklamers · x · 2026-10-13
- DavidSHolz's back-of-envelope math: a frontier model as a 10T MoE on GB300 NVL72 costs 5 Joules/token, or 50W at human thinking speed, vs. the brain's 20W—surprisingly similar on energy-per-thought.
- joefioti adds that high-batch inference is already nearly as efficient as the human brain, with another 10-100x efficiency gain on track.
- The thread contrasts grid power ($0.08/kWh) with food energy ($0.50/kWh), highlighting how fast inference economics are approaching biological baselines.
More from AGI Musings
- What if a pandemic virus were as smart as human experts across most topics — AndyMasley · 2026-10-13
- Bindu Reddy pushes back on Anthropic's small-model self-replication warning — bindureddy · 2026-10-13
- Researchers push back: release open weights if you want scientists to recalibrate on AGI — mjdramstead · 2026-10-13
- 'Claude is my CTO' is wrong: senior engineers matter more than ever — MartinGTobias · 2026-10-13
- Kurzgesagt quietly scrubbed its William MacAskill book promo after FTX/EA backlash — beffjezos · 2026-10-13
- Peter Diamandis on OpenAI's math progress and the case for universal high income — PeterDiamandis · 2026-10-13