马斯克称 Grok 的政治倾向是“旧金山湾区政治观点的产物”
英文标题：🔶 Musk says Grok's political lean is 'a function of SF Bay Area political views'
马斯克表示，Grok 的政治倾向是旧金山湾区政治观点的产物，并称其结合了加州与得州的文化，因此立场更居中。这一表态被解读为公开承认：前沿 AI 模型的政治校准并非从数学中自然涌现，而与训练它们的人群及所在地密切相关。文章认为，闭源模型的访问权和价值取向都由供应商决定且缺乏透明度；相比之下，开放权重模型让开发者能够自行掌控模型权重、系统提示词和部署环境。
2.2 KB
1.8 KB
1.7 KB
42.2 KB
中文
这场交锋
本周，一张将 AI 模型标注在 politicalcompass.org 双轴坐标网格上的图表广为流传，最终传到了图中某个点位利益攸关程度最高的人手中。Wilfred Reilly 转发并评论道：“这非常、非常有问题。”埃隆·马斯克于 2026 年 8 月 28 日晚间发布的回复，才是值得保留的部分：“这是旧金山湾区政治观点的产物。Grok 是加州加得州，所以更居中。” 这条回复在不到一天内获得了 280 万次浏览。
这是承认，不是论证
无论你如何看待这张图表，马斯克都以第一手立场确认了前沿实验室批评者多年来一直提出的观点：模型的政治校准并不是从数学中自然涌现的中性属性——它是训练该模型的人及其所在地点的产物。他并未否认这种倾向的存在；他是在解释这种倾向，并将 Grok 定位为两种地区文化的有意融合，而非默认结果。这是一次公开表述的工程决策，涉及一种大多数实验室更愿意不加审视的属性。
如实看待这一测量工具
该图表来自一个项目：该项目让模型回答 politicalcompass.org 测试中的 62 道强制二选一命题，并使用真实的测试工具对回答评分，每个模型运行五次。这是一项有趣但薄弱的测量：在“独立推理者”人格设定下生成的回答，并不构成稳定的政治身份；两个坐标轴严重压缩了意识形态；而模型之间的排名则是其中最不可靠的部分。但一旦实验室 CEO 公开确认了其底层前提，这种弱点就不再重要——这种倾向是真实的、经过有意设计的，并且可以追溯到构建模型的人。
为什么这很重要：闭源权重与开放权重
对于闭源模型，其价值取向由他人为你设定——由一种你无法选择、也无法检查的文化决定；它还可能在一次提示词调整或后训练更新中发生变化，且变更日志中不会留下任何记录。马斯克只用了 11 个词就描述了这一机制。这也与本站本周的另一篇报道相呼应：OpenAI 在 SpaceX 收购后终止与 Cursor 的合作关系表明，闭源模型的访问权可以在董事会会议室中被收回；而这条回复则表明，模型的价值取向是在办公园区中设定的。访问权与价值取向，都由他人决定，而且在它们发生变化之前，你都无从知晓。
只有开放权重模型能让这两项决定都由你掌控。在本地运行 GLM-5.3 或任何开放权重模型时，模型权重存放在你的磁盘上，系统提示词由你编辑或删除，没有人能在一夜之间替换模型的人格设定。硬件配置查找器可根据你的工作负载匹配能够承载模型的硬件，而模型目录则列出值得自行托管的模型权重。
English / 原文
The exchange
A chart mapping AI models on the politicalcompass.org two-axis grid made the rounds this week, and it eventually reached the person with the most at stake in one of the dots. Wilfred Reilly quote-posted it: “This is very very problematic.” Elon Musk’s reply, posted the evening of August 28, 2026, is the part worth keeping: “It’s a function of SF Bay Area political views. Grok is California plus Texas, so more centered.” The reply drew 2.8M views in under a day.
An admission, not an argument
Whatever you think of the chart, Musk is confirming from first position what critics of frontier labs have claimed for years: a model’s political calibration is not a neutral property that emerges from the math - it is an artifact of the people and place that trained it. He is not disputing that the lean exists. He is explaining it, and positioning Grok as a deliberate blend of two regional cultures rather than a default. That is an act of engineering, stated openly, about a property most labs prefer to leave unexamined.
The instrument, honestly
The chart came from a project that had models answer the 62 forced-choice propositions of the politicalcompass.org test and scored those answers with the real instrument, five runs per model. It is a fun measurement and a weak one: answers produced under an “independent reasoner” persona are not a stable political identity, two axes compress ideology badly, and the model-level rankings are the least solid part. But that weakness stops mattering once a lab CEO publicly confirms the premise underneath it - the lean is real, deliberate, and traceable to who built it.
Why this matters: closed weights, open weights
With a closed model, the values are set for you - by a culture you did not choose and cannot inspect, changeable in a prompt or post-training update with no changelog line item. Musk just described the mechanism in eleven words. It pairs with this site’s other story this week: OpenAI ending its Cursor partnership after the SpaceX acquisition showed that access to a closed model can be revoked in a boardroom; this reply shows the values are set in an office park. Access and values, both someone else’s call, and neither visible to you until they move.
Open weights are the only case where both decisions are yours. Run GLM-5.3 or any open-weight model locally and the weights sit on your disk, the system prompt is yours to edit or delete, and nobody rotates your model’s persona overnight. The rig finder matches your workload to hardware that can host it, and the model catalog lists the weights worth hosting.