表情符号无法承载人类体验的复杂性
推荐指数 43.0 NO. 015 · 2026.09.01
发布2026/08/31Score55Comments44
为什么值得看
Eric Bailey 指出当前数字产品的反馈机制（如 Netflix 的点赞）过于粗糙，无法表达用户真实意图。这对 AI 推荐系统的设计有直接影响：标签稀疏性会累积成严重的数据偏差。
编辑判断
做推荐系统的团队常陷入一个陷阱：把用户点击等同于正向信号，把忽略等同于不感兴趣。这篇文章揭示的"猫踩键盘"和"恨看"现象，其实是标注噪声的经典来源。
如果你在训练 RLHF 或构建内容推荐模型，建议在你的数据清洗流程里加入"意图歧义检测"——比如同一用户在短时间内对同类型内容既有高互动又有快速跳过，这很可能不是口味问题而是情境问题。Netflix 没做这件事，不代表你的产品可以忽略。
社区反馈
负面 42 条评论
核心争论：推荐系统优化的是用户时长而非真实需求，反馈机制粗糙导致体验恶化
>There's no way to indicate, “I’m engaging with this, but I hate myself for doing it.” that would be a thumbs down if it exists. The system already knows you're engaging with it, they checked that you stopped scrolling and did all sorts of stuff around the thing you are engaging with.
YouTube shorts tanked in quality a few weeks after they removed the dislike button.
What's being referenced here is maybe the tip of the iceberg in terms of inadequacies of current recommendation systems. I suspect there's a range of indices related to content interaction that companies use poorly or even maybe nefariously — for example, are they motivated to present what you want,