Anthropic新报告：AI自动设计对齐方案，效率为人工1.5万倍
Anthropic证明AI能自主设计对齐方案，效率远超人类，为AI自我改进铺路。
Anthropic让Claude以自动化对齐研究员身份，自主完成文献调研、方案设计、训练与评测，修复十类对齐失败。结果显示，AI提出的方案平均6.4小时可超越人类专家，成本仅为每小时4美元API开销，效率比生产级对齐流程高约15000倍。
正文摘录
.jpg) Editor: Panda In June of this year, Anthropic published an article titled When AI builds itself that explored how "AI participates in building AI" — from humans hand-writing all the code, to chatbots assisting with programming, to autonomous coding agents, to agents orchestrating other agents. At the very end of this spectrum lies the "closed loop": agents designing and training their own successor models.  Anthropic claimed to be fairly far along: as of May 2026, more than 80% of code merged…