A lot of OpenAI's AGI safety research focus on specifying goals without reward functions — since it's clear that we can't hand-program objective functions for complex real-world tasks. One potential paradigm for avoiding reward functions:
Iterated Amplification: An AI safety technique that has the potential to let us specify behaviors and goals that are beyond human scale: blog.openai.com/amplifying-a…
Oct 22, 2018 · 5:07 PM UTC
36
1
169