Quantilizer
Sample from a high quantile of a base distribution instead of maximizing.
Role. Target. Graph node quant.
A quantilizer is a proposed AI design that aims to reduce the harms from Goodhart’s law and specification gaming by selecting reasonably effective actions from a distribution of human-like actions, rather than maximizing over actions.
— LessWrong tag: Quantilization
Taylor (2016): the guarantee is relative to the base, the quantile, and the use pattern. One-shot bounds do not imply safety under repeated deployment. The orange edge to state utility is a limit: quantile → maximum recovers ordinary maximization. Lean (quantilizer_limit_is_argmax): once the quantile is no larger than a maximizer’s base mass, the support is exactly the argmax set; if the base puts no mass on any maximizer, it never is.