Delegating scaling-law asymptote fitting to Codex while planning a birthday party

dlwh · x · 2026-10-27

dlwh says he got a better asymptote estimate by simply asking Codex to do it—freeing himself to plan his kid's birthday party and prompt Codex about other things. The simple form used: loss = E + A compute^(-alpha), where E is the loss floor being estimated while small models are still far above it. Also a real example of offloading grunt work to a coding agent.

Related event: Marin releases open scaling suite Delphi; author's revised prediction cuts error to 0.4%(8 posts)→

Original post →

More from coding & agent

coding & agent channel →