Researcher argues compute buys capability, but value hits zero once new demand dries up
DimitrisPapail · x · 2026-10-20
AI researcher Dimitris Papail pushes back on a discussion of log-log relationships between training compute and value, arguing:
- Compute buys capability at a constant increment per 10x, and steps can stack — which may be the whole point of continued scaling
- But each capability step brings fewer new customers, since most demand already sits below the frontier
- Worse than log-log is zero: once capability crosses the threshold where it stops bringing new customers, compute stops creating value — unless you expand into new domains like bio and robotics
More from AGI Musings
- Reddit essay: why demanding proof of AI consciousness before moral consideration is a double standard — Afflictionxx · 2026-10-20
- Economist Paul Novosad: AI-written text deserves a steep attention discount, often near 100% — paulnovosad · 2026-10-20
- When AI Takes 80% of Routine Tasks, Problem-Definition and Risk-Taking Decide Your Worth — AryHHAry · 2026-10-20
- The most valuable research asset: secret negative results on dead-end directions — fkasummer · 2026-10-20
- Chart goes viral: AuDHD traits poised for a generational comeback in the AI era — edgarpavlovsky · 2026-10-20
- Software engineer asks: why are we so eager to automate away each other's jobs? — Out-Live-In-Death · 2026-10-20