Physix Frontier badgePhysix Frontier
PublishChinese | EN

Physix Frontier · News Briefing Card (Hacker News · Oct 14, 2026)

Study: AI Inference Scaling Efficiency Lags Human Time Investment

KEY FACTS

  • Toby Ord published a piece examining whether adding tokens to AI models and adding time for humans yield equivalent gains.
  • A METR chart shows that each doubling of resources yields roughly 4 times the score gain for humans compared to AI.
  • If this advantage persists, AI would need about 10,000 times more resources to match the improvement humans get from 10 times more time.
  • The UK AISI released a chart in July 2026 showing that AI time horizons grow sublinearly with token budgets.
  • All tested models have a scaling exponent gamma below 1, indicating that AI scaling efficiency is lower than that of humans.

KEY DATA

4xScore gain ratio per doubling of resources, humans vs AI
about 10,000xResource multiplier for AI to match humans' 10x time improvement
8 hoursMaximum single-session human work time limit
July 2026AISI chart release date

PHYSIX OBSERVATION

If AI scaling efficiency remains lower than that of humans, the current path of boosting capability by piling on compute will hit a cost bottleneck. This reminds the industry that simply adding tokens may not proportionally convert into intelligence, and improvements in model architecture and inference efficiency may be more critical than resource investment.

Source: Hacker News report

Study: AI Inference Scaling Efficiency Lags Human Time Investment | Physix Frontier Briefing