Unverified50% confidenceOpinionExact time
Pachocki believes that current alignment techniques such as RLHF may have fundamental limitations when faced with superintelligent systems.
1
Sources
50%
Confidence
Long-term
Relevance
9/7/2026
First Seen
Sources
Related Entities
Related Claims
Unverified2023年多项研究证明，即使经过RLHF对齐的模型也难以完全抵抗精心设计的间接提示词注入70% similarUnverified仅靠微调阶段的RLHF或许永远无法产生真正鲁棒的安全对齐66% similarUnverifiedDFT对强关联体系（如很多超导材料）的描述存在系统性误差，这是AI在超导领域面临挑战的技术根源之一65% similarVerifiedTraditional RLHF methods have a fundamental limitation of 'reward hacking,' where models may learn to perform well in evaluation scenarios but fail in out-of-distribution scenarios64% similarUnverified大模型存在能力边界、幻觉问题、可靠性瓶颈等尚未彻底解决的挑战62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/873233API
curl https://kongchang.com/api/v1/knowledge/claims/873233MCP
get_claim(id=873233)