Whether a system is an optimizer is an internal property

Created time
Sep 25, 2022 10:31 AM
Main Box
Tags
Machine Learning
Philosophy
Behaviourism
Interpretability
AI Alignment
Hubinger states that whether an AI is optimizing for something depends on the internal structure of the agent. Thus, we can not judge, based on the fulfillment of a certain metric, whether something is truly an optimizer or not.
He gives the example of a water bottle cap, that sufficiently holds in the water, but isn’t actually an optimizer, because it is not searching the solution space for optimizing towards a given metric.