Definition
Plain language
A pattern where each time you multiply the effort, you get a steady proportional improvement.
As stated in the literature
Relationship of the form loss = a·C^(-b) + floor; fitted compute-scaling curves with an irreducible offset, giving consistent fractional error reduction per compute multiple.
Also called: power laws, power-law
Why it matters: It lets researchers forecast what a much bigger training run will buy before spending the money, and the floor term warns that no amount of compute drives error to zero.
For example, every tenfold increase in training compute might shave the same fraction off the error, so the gains keep coming but each one costs ten times more than the last.
Heard on the show
“A power-law on the weights in one run, a concave transform on the values in the other.”Episode 176 — An AI Designed Its Own Psychology Studies, Then Confirmed What It Found