← Back to feed

Reliably aligning LLMs is impossible in principle

AI · Security · 2026-11-02 06:27

作者认为 LLM 对齐不仅做不到,连“成功”的定义都无法表述;并对比此前把 AI 视为盲优化器的安全研究范式。

Source: www.seangoedecke.com


Auto-generated summary for reference only. RSS