Tagged scalable oversight · show all articles
Five problems in AI alignment remain open by the researchers' own account — not by outside speculation — from judging outputs no human can fully check to deciding whose values a system should encode when people disagree.
Aligned is not one property but two — whether the specified objective is right, and whether the trained system actually pursues it. A documented case shows exactly where the two split apart, and why neither one alone is what people mean by safe.
Two questions decide how AI alignment and safety practice matures by 2035: whether verification moves beyond behaviour, and whether the field becomes an externally certified discipline or stays self-governed by labs. Four scenarios, each with a falsifier.
Long before deep learning existed, a mathematician warned that a machine pursuing a fixed purpose could not be trusted to want what its operator meant. The distance from that warning to a shipped alignment pipeline is the field's actual, dated history.