{
 "generated": "2026-11-20",
 "sha": "",
 "node": {
  "date": "2026-07-15",
  "updated": "2026-07-15",
  "kind": "topic",
  "category": "模型前沿",
  "group": "agents",
  "page": "/deepdive/training-data-crisis/",
  "summary": "Karpathy：Llama 3 信息压缩率仅 0.07 bits/token，前沿模型庞大是数据质量低下的代价而非算力堆叠必然 · 私有企业数据收购 vs 公开数据供给萎缩 · 推理模型作为绕过数据瓶颈的新路径",
  "version": 1,
  "tags": [
   "training-data",
   "data-quality",
   "karpathy",
   "reasoning-models",
   "synthetic-data",
   "scaling-laws"
  ],
  "langs": [
   "zh",
   "en"
  ],
  "words": 3847,
  "n_sections": 5,
  "dd_id": "DD·0051",
  "related": [
   "loop-transformer",
   "six-frontier-ai-technologies",
   "karpathy-joins-anthropic",
   "ai-education-resources-map"
  ],
  "id": "dd:training-data-crisis",
  "type": "Article",
  "article": "training-data-crisis",
  "name": "垃圾数据与下一次能力跃升",
  "title": "垃圾数据与下一次能力跃升",
  "slug": "training-data-crisis",
  "kind_zh": "热点专题",
  "entry": "agent",
  "n_ents": 6,
  "n_entities": 6
 },
 "page": "https://xiaopingfeng.com/graph/e/dd/training-data-crisis/",
 "out": [
  {
   "type": "covers",
   "sentence": "垃圾数据与下一次能力跃升讲到Meta",
   "direction": "out",
   "id": "org:meta",
   "name": "Meta",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "垃圾数据与下一次能力跃升讲到Google",
   "direction": "out",
   "id": "org:google",
   "name": "Google",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "垃圾数据与下一次能力跃升讲到Andrej Karpathy",
   "direction": "out",
   "id": "person:andrej-karpathy",
   "name": "Andrej Karpathy",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "垃圾数据与下一次能力跃升讲到Epoch AI",
   "direction": "out",
   "id": "org:epoch-ai",
   "name": "Epoch AI",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "垃圾数据与下一次能力跃升讲到Sakana AI",
   "direction": "out",
   "id": "org:sakana-ai",
   "name": "Sakana AI",
   "role": "mention",
   "status": "agent",
   "evidence": []
  },
  {
   "type": "covers",
   "sentence": "垃圾数据与下一次能力跃升讲到Sam Altman",
   "direction": "out",
   "id": "person:sam-altman",
   "name": "Sam Altman",
   "role": "mention",
   "status": "agent",
   "evidence": []
  }
 ],
 "in": [
  {
   "type": "part_of",
   "sentence": "0.07 bits/token 揭示的真相属于垃圾数据与下一次能力跃升",
   "direction": "in",
   "id": "ddsec:training-data-crisis-sec-diagnosis",
   "name": "0.07 bits/token 揭示的真相",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "从互联网 到企业私有数据属于垃圾数据与下一次能力跃升",
   "direction": "in",
   "id": "ddsec:training-data-crisis-sec-frontier",
   "name": "从互联网 到企业私有数据",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "推理模型：绕过数据瓶颈的新路径属于垃圾数据与下一次能力跃升",
   "direction": "in",
   "id": "ddsec:training-data-crisis-sec-reasoning",
   "name": "推理模型：绕过数据瓶颈的新路径",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "对从业者 和企业管理者的含义属于垃圾数据与下一次能力跃升",
   "direction": "in",
   "id": "ddsec:training-data-crisis-sec-implications",
   "name": "对从业者 和企业管理者的含义",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "算力军备竞赛 只是上半场属于垃圾数据与下一次能力跃升",
   "direction": "in",
   "id": "ddsec:training-data-crisis-sec-synthesis",
   "name": "算力军备竞赛 只是上半场",
   "status": "derived",
   "evidence": []
  }
 ]
}