Unverified50% confidenceTradeoffExact time
让模型更安全与让模型更好用之间存在真实的权衡，而非纯粹正和游戏
1
Sources
50%
Confidence
Long-term
Relevance
7/25/2026
First Seen
Sources
Gemini 3.0 Flash比3.5更受欢迎？揭秘AI「对齐税」真相
redditr/Bard7/12/2026
Related Claims
UnverifiedFor sufficiently powerful models, clearly conveying intent and context is more effective than providing detailed operational procedures.72% similarUnverifiedAI对齐研究中，模型倾向回避高度不确定性的开放问题的现象被称为奖励模型过拟合或规范游戏69% similarUnverified提示注入针对系统架构，越狱针对模型价值观对齐本身，这是两类攻击的根本区别68% similarUnverified模型竞赛的胜负取决于能力、安全与生态之间的可持续平衡，而非发布速度68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/611258API
curl https://kongchang.com/api/v1/knowledge/claims/611258MCP
get_claim(id=611258)