Tagged AI alignment · 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.
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.
Constitutional AI keeps RLHF's reward-model-and-reinforcement-learning core intact and swaps the source of one training dataset, a document standing in for a person's judgment. That swap is narrower, and its limits more specific, than the name suggests.