Weight smuggling likely defeats attempts to cap FLOPs per training run.

A LessWrong post argues that verification mechanisms are unlikely to stop training runs from being chained together or aggregated in parallel, a practice the author calls weight smuggling, making it hard to enforce caps on FLOPs per training run.

Why it matters: It suggests that compute-based limits on individual training runs, a common proposal for controlling advanced AI development, may be difficult to verify or enforce in practice.

Written by AI from LessWrong's published text.