Scaling debate: 30x on the plot can't prove model gains beat 10,000x parallel compute

scaling01 · x · 2026-09-19

A technical spat over test-time compute scaling. One side claims the plot shows a generational model step-up beats a 10,000x parallel compute scale-up; scaling01 counters that the plot only spans 30x compute, so it can't support that claim. The other side extracts data points: the internal model gains 0.17 logits per compute doubling while Astra's serial test-time compute yields 0.18 per doubling, arguing serial scaling is at least as efficient as the model gap.

Related event: Debate Erupts Over Agent Swarm's Contribution to Navier-Stokes Proof(11 posts)→

Original post →

More from Models

Models channel →