Parallel scaling debate: the bottleneck is wall-clock time, not token efficiency

scaling01 · x · 2026-09-19

Continuing the test-time compute debate, scaling01 adds that token efficiency is a moot point — nobody argues parallel is more token-efficient. The real constraint is time: even if a parallel approach were 100x more token-efficient, sequential execution would still take a year.

Original post →

More from Models

Models channel →