{
 "generated": "2026-11-20",
 "sha": "",
 "node": {
  "date": "2026-06-21",
  "updated": "2026-06-21",
  "kind": "topic",
  "category": "评测与实验",
  "group": "probes",
  "page": "/deepdive/ai-taste-probe/",
  "summary": "🥣 粽子续集，加测「豆腐脑甜还是咸」并追问这到底是不是模型的「真实偏好」· 同一模型跨题口味常对不上（8/17 翻盘）、没有稳定的「口味人格」· 四层可解释性阶梯：行为采样 → logprobs → 可操控性 → 机理层 · 结论：想看得越深，模型就得越开放",
  "version": 3,
  "tags": [
   "AI",
   "Interpretability",
   "可解释性",
   "RLHF",
   "Preference",
   "logprobs",
   "豆腐脑",
   "SAE",
   "Neuronpedia",
   "可观测性"
  ],
  "langs": [
   "zh",
   "en"
  ],
  "words": 6818,
  "n_sections": 6,
  "dd_id": "DD·0020",
  "related": [
   "duanwu-ai",
   "loop-transformer",
   "llm-pollution-media-chain",
   "ai-jargon-index"
  ],
  "id": "dd:ai-taste-probe",
  "type": "Article",
  "article": "ai-taste-probe",
  "name": "一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性",
  "title": "一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性",
  "slug": "ai-taste-probe",
  "kind_zh": "热点专题",
  "entry": "agent",
  "n_ents": 21,
  "n_entities": 21
 },
 "page": "https://xiaopingfeng.com/graph/e/dd/ai-taste-probe/",
 "out": [
  {
   "type": "covers",
   "sentence": "一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性讲到Anthropic",
   "direction": "out",
   "id": "org:anthropic",
   "name": "Anthropic",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性讲到Claude",
   "direction": "out",
   "id": "app:claude",
   "name": "Claude",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性讲到Google",
   "direction": "out",
   "id": "org:google",
   "name": "Google",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性讲到DeepSeek",
   "direction": "out",
   "id": "app:deepseek",
   "name": "DeepSeek",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性讲到DeepSeek",
   "direction": "out",
   "id": "org:deepseek",
   "name": "DeepSeek（深度求索）",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性讲到Gemini",
   "direction": "out",
   "id": "app:gemini",
   "name": "Gemini",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性讲到Meta",
   "direction": "out",
   "id": "org:meta",
   "name": "Meta",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性讲到SpaceXAI",
   "direction": "out",
   "id": "org:spacexai",
   "name": "SpaceXAI（原 xAI）",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性讲到Grok",
   "direction": "out",
   "id": "app:grok",
   "name": "Grok",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性讲到OpenRouter",
   "direction": "out",
   "id": "app:openrouter",
   "name": "OpenRouter",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性讲到GLM-5.2",
   "direction": "out",
   "id": "model:glm-5-2",
   "name": "GLM-5.2",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性讲到Arize AI",
   "direction": "out",
   "id": "org:arize",
   "name": "Arize AI",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性讲到OpenRouter",
   "direction": "out",
   "id": "org:openrouter",
   "name": "OpenRouter",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性讲到Tencent",
   "direction": "out",
   "id": "org:tencent",
   "name": "Tencent（腾讯）",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性讲到Z.ai",
   "direction": "out",
   "id": "org:zai",
   "name": "Z.ai",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性讲到MiniMax",
   "direction": "out",
   "id": "app:minimax",
   "name": "MiniMax",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性讲到Qwen",
   "direction": "out",
   "id": "app:qwen",
   "name": "Qwen",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性讲到Alibaba",
   "direction": "out",
   "id": "org:alibaba",
   "name": "Alibaba（阿里巴巴）",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性讲到MiniMax",
   "direction": "out",
   "id": "org:minimax",
   "name": "MiniMax",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性讲到Moonshot AI",
   "direction": "out",
   "id": "org:moonshot",
   "name": "Moonshot AI（月之暗面）",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性讲到Kimi",
   "direction": "out",
   "id": "app:kimi",
   "name": "Kimi",
   "role": "mention",
   "status": "agent",
   "evidence": []
  }
 ],
 "in": [
  {
   "type": "part_of",
   "sentence": "承上：我们已经知道的三件事= 引自《AI 的甜咸之争》属于一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性",
   "direction": "in",
   "id": "ddsec:ai-taste-probe-recap",
   "name": "承上：我们已经知道的三件事= 引自《AI 的甜咸之争》",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "🥣 豆腐脑，甜的还是咸的？= 第二个探针属于一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性",
   "direction": "in",
   "id": "ddsec:ai-taste-probe-doufu",
   "name": "🥣 豆腐脑，甜的还是咸的？= 第二个探针",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "同一个模型，两道题口味一样吗？= 偏好的稳定性属于一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性",
   "direction": "in",
   "id": "ddsec:ai-taste-probe-consist",
   "name": "同一个模型，两道题口味一样吗？= 偏好的稳定性",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "能不能真的看进模型的脑子里？= 可解释性的阶梯属于一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性",
   "direction": "in",
   "id": "ddsec:ai-taste-probe-methods",
   "name": "能不能真的看进模型的脑子里？= 可解释性的阶梯",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "能不能定位「甜」在模型里的那根弦？= SAE 特征 + 因果干预属于一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性",
   "direction": "in",
   "id": "ddsec:ai-taste-probe-why",
   "name": "能不能定位「甜」在模型里的那根弦？= SAE 特征 + 因果干预",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "看 AI 的六种「仪器」= 从产品日志到神经元属于一碗豆腐脑，撬开 AI 的黑盒 · 从口味偏好到可解释性",
   "direction": "in",
   "id": "ddsec:ai-taste-probe-panorama",
   "name": "看 AI 的六种「仪器」= 从产品日志到神经元",
   "status": "derived",
   "evidence": []
  }
 ]
}