Model generation gains could soon let a single agent solve NS in 88 hours

ChaseBrowe32432 · x · 2026-09-19

A discussion around a compute-scaling plot: the poster extrapolates that one more model generation would allow solving NS with 16 sub-agents over 88 hours, and two generations would need just a single agent. He argues this is consistent with Noam's "10%" comment, possibly even conservative. A replier pushes back: the plot only covers 30x compute, which can't justify claiming generational model gains beat a 10,000x parallel scale-up.

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

Original post →

More from coding & agent

coding & agent channel →