100,000× effective compute by 2027
Three compounding vectors of progress, each contributing ~0.5 OOMs/year. Compounded: ~5 orders of magnitude from 2023 to 2027.
Compute Scaling
GPT-2 (2019) used 4e21 FLOP; GPT-4 used 8e24-4e25 FLOP — 3,000-10,000× in 4 years.
Algorithmic Efficiency
GPT-4o costs 6×/4× less than GPT-4. Gemini 1.5 Flash costs 85×/57× less.
Unhobbling
RLHF'd small model = non-RLHF'd model 100× larger. Context 2K → 1M+ tokens.
The Trillion-Dollar Cluster
From GPT-4 at 10MW to 100GW clusters by 2030 — greater than 20% of all US electricity consumption.
“Power is the binding constraint, not chips or investment.”
Path to superintelligence
Proto-engineer by 2026 · AGI 2027 · intelligence explosion 2028-29 · superintelligence by 2030.
“Multiple doublings a year” possible — a regime shift comparable to hunting → farming → industrialization.
Each working at 100× human speed. A decade of algorithmic progress compressed to one year.
Greater than 20% of all US electricity. The binding constraint is gas — not chips.