[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"report-2026-07-06":3,"Zpxeshli2B":611,"yz8vlrqKUf":626,"D9wp4eGtHR":636,"7rz8VZvv14":646,"mfvhRTJuAN":656,"YjswcwmRTY":766,"35IJkBZMtt":825,"F5DpeV9E1n":841,"ycEamUiOSr":857,"VwU6A76PWe":893,"wueS2fMwwV":929,"DSJP1TSlQY":939,"qzc2Jr6F1h":949,"rjXXzCXzrd":959,"KizT5xyTAF":969,"pfdeIqZRp8":979,"Z4t89R6iHG":989,"VP2jTAmiHv":999,"FShELeQ8ea":1009,"AI8tn6hFxj":1019,"PklIuNMtho":1029,"vHmYUPYDag":1039,"a49af8wB1S":1049,"RISx7YcH5a":1059,"IMsmWQon5o":1144,"44JarwvVof":1155,"KqIn7D6K7u":1187,"j1a8qnhI0K":1203,"znkPQbGH95":1219,"qW9XIspwKq":1326,"7IP9QL1BRQ":1460,"z01Ttqwdzq":1477,"6Qj9uhmTdO":1494,"UpmsiHlOz5":1504,"3aYaQlcZVu":1514,"HRpbgqW7US":1524,"P3D9gE9SvS":1534,"fGzzX2JXpj":1544,"my9jNNPgGx":1565,"XdoYQIZHWj":1713,"4A0zh1aUxp":1724,"WWgAeEoIU4":1740,"8Zs74O87MX":1774,"66zxwVd7En":1800,"3iFGzis38P":1915,"xvMlehTn8D":2147,"FrJDhgYzYz":2168,"5qyAmm1Z3d":2197,"guisQTaQoD":2207,"DSrEWvSQ0x":2226,"QEkdKh590r":2236,"bjSJ82Ycw6":2246,"5GGXdgmieh":2256,"LTgL9fRKA7":2266,"ZKDQFG3XKY":2391,"CgwJkt3Nuy":2402,"gm5SbM4RSV":2418,"1lsKbLNb36":2449,"LISDxZrLgT":2465,"OOYGfuqvBX":2586,"dl1iKkD0iy":2636,"pfWisDytfh":2661,"RLG2HK1YHA":2682,"LERsBnyL03":2692,"i5DbTlhSku":2702,"qTwoG4J7lw":2712,"FNR44xqtMZ":2759,"2wQlANIxpE":2769,"85grpn4Gap":2779,"rVEhNyrEMg":2854,"vz8Xgb8hLs":2888,"XMHFXex58I":2912,"vOUqmurlna":2942,"s9RLRFyAHk":3060,"U4JrBvAabA":3081,"yxQsL6YNfs":3097,"0ePluBpuKP":3170,"8GtLuSwu9I":3180,"Xw5wlcfnf4":3190,"6F5YxBvJAp":3266,"LaWi9bpOFq":3276,"eRjAgSJOep":3286,"1PsbdfISUo":3364,"UMuZWY2BWt":3411,"1gwGC8pWTQ":3427,"BXgjjO19cG":3443,"Qol6TXfS8A":3467,"J9VqphN3nT":3544,"I7gkUYcAVg":3560,"FXvZm0Vjm8":3576,"vT0DLskvwP":3615,"5sIi95nCKg":3669,"NGbC1ySWcn":3688,"MlDGG3A8uz":3698,"EjvQTOSxOu":3731,"8U1KJbhh4k":3803,"EKhw0AtlvC":3824,"fKRLlQpwIX":4894},{"report":4,"adjacent":608},{"version":5,"date":6,"title":7,"sources":8,"hook":19,"deepDives":20,"quickBites":333,"communityOverview":585,"dailyActions":586,"outro":607},"20260216.0","2026-07-06","AI 趨勢日報：2026-07-06",[9,10,11,12,13,14,15,16,17,18],"academic","amazon","anthropic","baidu","community","github","media","meta","mistral","openai","加密通訊存亡決戰、AI 訓練基礎設施世代更替、開源算力生態崛起，今日社群同步見證三個轉折點。",[21,126,193,266],{"category":22,"source":13,"title":23,"subtitle":24,"publishDate":6,"tier1Source":25,"supplementSources":28,"tldr":41,"context":53,"devilsAdvocate":54,"community":57,"hypeScore":75,"hypeMax":76,"adoptionAdvice":77,"actionItems":78,"policyDetail":88,"complianceImpact":89,"industryImpact":99,"timeline":100},"policy","EU 理事會快速通道強推 Chat Control，加密通訊全面掃描爭議白熱化","歐洲議會以 1 票之差否決後，理事會趁暑假前強行重啟訊息掃描授權",{"name":26,"url":27},"heise online","https://www.heise.de/en/news/Chat-Control-1-0-EU-Council-forces-messenger-scans-via-fast-track-11353659.html",[29,33,37],{"name":30,"url":31,"detail":32},"Hacker News 討論 #48793393","https://news.ycombinator.com/item?id=48793393","社群深度討論 Chat Control 1.0 與 2.0 的差異、民主程序爭議及對加密技術的影響",{"name":34,"url":35,"detail":36},"EU Chat Control 即時追蹤器 — Closed Network","https://closednetwork.io/eu-chat-control-the-fight-to-scan-every-private-message-live-tracker/","持續追蹤歐盟 Chat Control 立法進程、投票記錄及各方立場更新",{"name":38,"url":39,"detail":40},"Patrick Breyer MEP — Chat Control 資訊頁","https://www.patrick-breyer.de/en/posts/chat-control/","歐洲議會議員 Patrick Breyer 對 Chat Control 各版本的詳細分析與反對立場文件",{"tagline":42,"points":43},"歐洲議會剛以 1 票否決，理事會卻趁暑假前強行重啟——民主程序本身成了第一個受害者",[44,47,50],{"label":45,"text":46},"政策","Chat Control 1.0 豁免條款允許平台自願掃描 CSAM，歐洲議會以 307：306 否決延長後，EU 理事會透過快速通道強推重啟，計畫趁議員暑休前完成表決。",{"label":48,"text":49},"合規","二讀程序要求議會取得「絕對多數」才能阻止，缺席議員的票計入反對方分母，法案通過機率被普遍認為偏高。Signal 揚言若法案通過將撤出歐盟市場。",{"label":51,"text":52},"影響","Chat Control 2.0 的強制客戶端掃描條款雖暫緩，但超過 500 位密碼學家警告：一旦掃描基礎設施建立，未來轉為強制義務的政治阻力將大幅降低。","#### Chat Control 是什麼——歐盟的 AI 掃描立法草案\n\nChat Control 是歐盟關於私人訊息掃描的一系列立法提案的統稱，這個詞彙由反對掃描的隱私倡議者所創造並推廣，支持掃描的政府方從不使用這個術語，命名權之爭本身就是輿論戰的一部分。\n\nChat Control 1.0 源自 2021 年通過的《ePrivacy 指令》臨時豁免條款，允許 Gmail、Facebook Messenger 等平台**自願掃描**未加密訊息，以偵測兒童性剝削素材 (CSAM) 。豁免期原訂至 2026 年 4 月 3 日，掃描涵蓋三種技術：哈希比對、AI 模式辨識（grooming 偵測）以及對未加密訊息的主動掃描。\n\n> **名詞解釋**\n> 哈希比對 (hash matching) ：為每個已知非法圖片生成唯一的「數位指紋」（哈希值），系統掃描上傳內容時與指紋庫比對，命中即標記，無需人工判讀原始圖片內容。\n\nChat Control 2.0 是更進一步的永久性強制掃描方案，覆蓋端對端加密 (E2E) 訊息服務，必須在訊息加密前的用戶裝置上執行掃描，即「客戶端掃描 (client-side scanning) 」。超過 500 位科學家與密碼學家聯署公開聲明，指出此方案「技術上不可行」——這等同在每台裝置上安裝可遠端觸發的後門程式。\n\n#### 快速通道的政治操作與民主程序爭議\n\n2026 年 3 月 26 日，歐洲議會以 307 票對 306 票（24 票棄權）的極微差距否決延長 Chat Control 1.0 臨時規定，隱私倡議者一度宣告勝利。\n\n然而，歐盟理事會（由各成員國政府代表組成，而非直選機構）在 2026 年 7 月 2 日透過書面程序通過新立場，強行推動重啟授權。更具爭議的是：理事會計畫在 7 月 8 日緊急將草案列入議程，刻意選在暑假前大批議員已離場的時間點推動表決。\n\n此法案處於立法二讀程序，議會若要阻止，須取得「絕對多數」——缺席議員的票也計入反對方的分母，使這一門檻在實務上被認為幾乎不可能達到。批評者將此形容為「試圖繞過民主控制機構、給議會一個突襲」，這種程序操作的合法性爭議，不亞於法案本身的實質內容。\n\n#### 隱私 vs 兒童保護：社群兩極化的激辯\n\n支持方（各成員國政府）的核心論點是：自願偵測措施是「早期辨別受害兒童、拯救受虐者不可或缺的工具」，現行豁免期間協助發現大量案例，強制廢止等同讓受害者失去保護網。\n\n反對方則指出，「自願」掃描一旦取得法定授權基礎，將為日後強制義務建立現成技術架構。義大利等國政府明確反對大規模監控，卻仍在理事會投票支持，被倡議者 Patrick Breyer 直接批評為「自相矛盾的荒謬」。\n\nHN 社群用戶 SpicyLemonZest 提出一個值得注意的觀察：「Chat Control」詞彙本身帶有強烈立場，由反對者定義，支持方從不使用此詞。而 HN 用戶 neobrain 則提醒社群：此次復活的是允許平台「自願掃描」的 Chat Control 1.0，要求破壞加密的 Chat Control 2.0 因公民強烈抵制已暫緩——兩者不應混淆，否則會誤判當前的實質風險層級。\n\n#### 對加密技術與科技產業的深遠衝擊\n\nChat Control 2.0 中的客戶端掃描要求，被密碼學社群視為對端對端加密的根本性威脅。Signal 已明確聲明：若強制掃描法案通過，寧願撤出整個歐盟市場，而非在安全架構中留下可被利用的後門。Mullvad VPN 等隱私工具廠商也在密切追蹤立法動態，向用戶發出持續警示。\n\n若相關法案在歐盟形成合規義務，影響將不限於歐洲用戶。任何在歐盟有業務的全球性通訊服務——WhatsApp、iMessage、Telegram——都可能面臨同樣的合規壓力，必須重新評估能否在保留歐盟用戶的同時，維持現有的端對端加密安全承諾。這將是科技產業與各國監管機構之間一場事關隱私基礎設施的長期博弈。",[55,56],"兒童性剝削素材的傳播是嚴重犯罪，若哈希比對能協助識別受害者並拯救生命，其防護效益是否可能超過對隱私的有限侵犯？限於已知素材哈希比對的掃描，與監控整體通訊內容之間，在技術實作上存在明確邊界。","Chat Control 1.0 的自願掃描豁免，與 Chat Control 2.0 的強制客戶端掃描，在技術實作和隱私侵害程度上有根本差異。反對者將兩者以同一術語混用，可能使公眾誤以為當前法案的威脅程度高於實際，反而不利於精準的政策辯論。",[58,62,64,68,72],{"platform":59,"user":60,"quote":61},"Hacker News","SpicyLemonZest（HN 用戶）","『Chat Control』是一個由反對任何私人訊息掃描的隱私倡議團體所發明並推廣的術語；術語指涉不清造成的混淆應由他們負責——儘管他們認為這樣的稱呼是恰當的，因為任何形式的掃描都可以用相同方式被濫用。支持掃描私人訊息的一方從不使用『Chat Control』這個詞。",{"platform":59,"user":60,"quote":63},"但我們為何認為歐盟民主正在衰退？是否有具體證據顯示歐盟公民反對 CSAM 掃描，而政治機構對此漠然？就我所知，民調普遍顯示大眾支持 CSAM 掃描，唯一讓它具有爭議性的是隱私倡議團體積極的倡議與遊說。",{"platform":65,"user":66,"quote":67},"Bluesky","echo-pbreyer（Patrick Breyer，29 upvotes）","荒謬！義大利警告反對 ChatControl 大規模監控，卻投票支持它！只剩 2 天：阻止這場政變！立即採取行動：fightchatcontrol.eu",{"platform":69,"user":70,"quote":71},"X","@paddi_hansen（EU 加密與金融科技政策專家）","剛傳來消息——延長 Chat Control 1.0 的投票遭到歐洲議會否決。延長 ePrivacy 豁免條款的嘗試再度失敗，有效期限將於 4 月 3 日正式屆滿。隱私倡議者兩週內連拿兩場重大勝利！",{"platform":69,"user":73,"quote":74},"@mullvadnet（Mullvad VPN，隱私導向 VPN 服務商）","這是重要的勝利——但我們仍須阻止 Chat Control。歐盟部長理事會歷經三年後終於就 Chat Control 達成共同立場，強制掃描要求（包括對端對端加密訊息服務的掃描）已被移除⋯⋯",4,5,"追整體趨勢",[79,82,85],{"type":80,"text":81},"Try","現在就切換至預設端對端加密的通訊應用（如 Signal），親身理解加密架構如何保護私人通訊，以及客戶端掃描若上路將破壞哪個環節。",{"type":83,"text":84},"Build","若你的服務在歐盟有用戶，評估現有通訊模組是否依賴第三方掃描基礎設施，並模擬 DPIA（資料保護衝擊評估）以識別合規差距。",{"type":86,"text":87},"Watch","追蹤 Patrick Breyer MEP 部落格及 EFF 聲明，關注 2026 年 7 月 8 日關鍵快速通道表決結果，以及 Chat Control 2.0 後續三方協商進度。","#### 核心條款\n\nChat Control 1.0 的核心是《ePrivacy 指令》的臨時豁免授權，允許通訊服務提供者在未取得用戶同意的情況下，**自願掃描**私人訊息以偵測 CSAM 及 grooming 行為。此豁免使平台得以在不違反 ePrivacy 法規的前提下，主動篩查用戶通訊內容。\n\nChat Control 2.0（仍在三方協商中）計畫將此掃描從自願改為**強制義務**，且擴大至端對端加密訊息服務——由於加密訊息在傳輸中無法讀取，強制掃描必須在裝置端（加密前）執行，即所謂「客戶端掃描」，其本質是在用戶裝置上安裝偵測程式。\n\n#### 適用範圍\n\nChat Control 1.0 適用於所有在歐盟提供「網際網路通訊服務」的平台，包括電子郵件服務（Gmail、Outlook）、即時通訊軟體（WhatsApp、Messenger、Signal、Telegram）及具私訊功能的社交媒體平台。\n\n適用對象不限企業規模或國籍，只要在歐盟境內有用戶即受管轄，意味著美國、亞洲等地的科技公司只要提供歐盟用戶服務，均需評估合規義務。\n\n#### 執法機制\n\n偵測到疑似 CSAM 的內容，平台須向指定通報機構（如 NCMEC 或各成員國對等機構）提交通報。資料保留規定要求偵測相關的訊息內容與流量資料，在嫌疑確認後最遲 12 個月內不可逆刪除；在確認前，資料可依法保存供執法調查使用。各成員國須指定負責監督合規的主管機關，並建立企業申訴機制。",[90,93,96],{"label":91,"markdown":92},"工程改造需求","平台需部署 CSAM 哈希比對系統（整合 PhotoDNA 或 NCMEC 資料庫）與 grooming AI 模式偵測模組，並嵌入訊息收發流程。\n\n資料管道須支援「偵測後 12 個月不可逆刪除」機制，需建置定時清理排程與稽核日誌。若 Chat Control 2.0 強制客戶端掃描義務落地，還需在 App 中嵌入掃描邏輯，並面對 App Store／Play Store 的額外分發審查壓力。",{"label":94,"markdown":95},"合規成本估計","中型訊息平台預估需投入 6–18 個月工程週期，建置並驗收 CSAM 偵測基礎設施。第三方哈希資料庫（如 PhotoDNA）須支付授權費與持續維護成本。\n\n法律合規方面，需聘用資料保護長 (DPO) 並完成資料保護衝擊評估 (DPIA) ，屬一次性及年度循環支出。若跨多個歐盟成員國營運，各地監管對接成本將乘數放大。",{"label":97,"markdown":98},"最小合規路徑","1. 整合 PhotoDNA 或 NCMEC 提供的哈希資料庫，僅對圖片與影片附件進行哈希比對\n2. 建置自動通報流程，確保命中後於規定時效內通報指定機構\n3. 實作偵測日誌的 12 個月不可逆刪除排程\n4. 完成 DPIA 文件，確認掃描範圍不超出 Chat Control 1.0 授權邊界\n5. 明確排除對純文字訊息的掃描，避免過度合規引發用戶信任危機","#### 直接影響者\n\n即時通訊平台首當其衝，包括 WhatsApp、Signal、Telegram、iMessage 等全球主流服務。電子郵件服務商（Gmail、Outlook）及具私訊功能的社交媒體平台（Meta、Snapchat）同樣在適用範圍內。Signal 已明確表態將撤出歐盟市場，其他平台則多採取沉默觀察姿態。\n\n#### 間接波及者\n\n裝置製造商（Apple、Google）若客戶端掃描義務落地，需提供 OS 層級 API 或面對應用層合規壓力。隱私工具提供商（VPN、加密郵件服務）可能因用戶隱私需求升溫而迎來新增流量。CSAM 偵測技術供應商（PhotoDNA 母公司 Microsoft、Thorn）則成為新增合規鏈中的必選採購方。\n\n#### 成本轉嫁效應\n\n合規建置成本最終將以服務費上漲或功能降格（如關閉特定地區加密功能）的形式轉嫁至用戶端。更深遠的衝擊是用戶對私人通訊平台的信任度系統性下降，可能推動部分用戶轉向無歐盟管轄的替代工具，加速去中心化通訊協議（如 Matrix）的採用。",[101,105,108,111,114,117,122],{"date":102,"text":103,"phase":104},"2021-01-01","Chat Control 1.0《ePrivacy 指令》臨時豁免條款生效，允許平台自願掃描 CSAM","past",{"date":106,"text":107,"phase":104},"2026-03-26","歐洲議會以 307：306 否決延長臨時豁免規定，隱私倡議者宣告階段性勝利",{"date":109,"text":110,"phase":104},"2026-04-03","Chat Control 1.0 豁免期限正式屆滿，平台掃描行為進入法律灰色地帶",{"date":112,"text":113,"phase":104},"2026-06-29","Chat Control 2.0 第五輪三方協商在賽普勒斯輪值主席任期內完成",{"date":115,"text":116,"phase":104},"2026-07-02","歐盟理事會透過書面程序通過新立場，強推重啟 Chat Control 1.0 掃描授權",{"date":118,"label":119,"text":120,"phase":121},"2026-07-08","關鍵節點","理事會計畫將草案緊急列入歐洲議會議程，趁暑假前議員大量缺席時強行表決","future",{"date":123,"label":124,"text":125,"phase":121},"中期（2026 下半年）","觀察","Chat Control 2.0 三方協商能否繼續推進；Signal 等平台是否啟動撤出歐盟的實際準備動作",{"category":127,"source":10,"title":128,"subtitle":129,"publishDate":6,"tier1Source":130,"supplementSources":133,"tldr":142,"context":154,"mechanics":155,"benchmark":156,"useCases":157,"engineerLens":164,"businessLens":165,"devilsAdvocate":166,"community":169,"hypeScore":185,"hypeMax":76,"adoptionAdvice":77,"actionItems":186},"ecosystem","Amazon Mechanical Turk 停止接受新用戶，AI 訓練資料的初代基礎設施走入黃昏","從 ImageNet 的幕後推手到 AWS 維護清單：眾包標註時代終結",{"name":131,"url":132},"TechCrunch","https://techcrunch.com/2026/07/05/amazon-will-stop-accepting-new-customers-for-mechanical-turk/",[134,138],{"name":135,"url":136,"detail":137},"The Register","https://www.theregister.com/off-prem/2026/07/03/amazons-mechanical-turk-to-stop-accepting-new-customers-and-not-even-ai-can-save-it/5266274","深度分析 MTurk 退場脈絡，涵蓋 AWS 維護清單定位與生態替代服務的競爭分析",{"name":139,"url":140,"detail":141},"Shopifreaks","https://www.shopifreaks.com/amazon-winds-down-mechanical-turk-closing-the-2005-microtask-crowdsourcing-marketplace-to-new-customers-on-july-30-2026/","提供 2005 年歷史背景脈絡與 7 月 30 日截止日程細節",{"tagline":143,"points":144},"眾包標註開創了 AI 資料工業化，而它的落幕，正是 AI 已強大到不再需要它的明證",[145,148,151],{"label":146,"text":147},"生態","MTurk 自 2005 年起支撐 ImageNet 等奠基性資料集，現已被列入 AWS 維護名單，7 月 30 日起關閉新客戶申請，不計劃推出新功能",{"label":149,"text":150},"轉移","Scale AI 等專業標註平台崛起，加上 LLM 代工問題（46% 工人使用 AI 代替人工答題）雙重夾擊，讓 MTurk 喪失核心競爭力",{"label":152,"text":153},"趨勢","合成資料市場 2026 年達 7.1 億美元，RLHF 轉向少量高品質專家標註，大規模廉價眾包的時代正式謝幕","#### 章節一：MTurk 的歷史地位——ImageNet 與 AI 革命的幕後推手\n\n2005 年 11 月，Amazon Mechanical Turk 正式上線，比 Freelancer 和 Fiverr 更早成為眾包微任務的商業市集。它的命名靈感來自 1769 年一台名為「機械土耳其人」的棋盤自動機——表面上是機器自主運作，實際上藏有人類操作者，正如 MTurk 以「人工智慧訓練」為名，實則仰賴隱匿工人的勞動。\n\nImageNet(2009) 、COCO、Open Images 等奠定深度學習革命的大型資料集，背後都仰賴 MTurk 完成數以百萬計的圖片分類與標記工作。2018 年 AWS 將 MTurk 整合進 SageMaker，正式將其定位為 AI 訓練資料的工業化生產線。沒有 MTurk，深度學習的資料基礎可能要晚幾年才能搭建完成。\n\n#### 章節二：為何現在關門：專業標註平台崛起與市場轉移\n\nMTurk 退場並非突然，其市場地位在過去五年中逐漸被侵蝕。AWS 自家的 SageMaker Ground Truth 整合了自動化標註、品質控管與人機協作工作流，功能已全面超越 MTurk 的簡單任務市集模式。\n\nScale AI、Surge AI 等新世代專業標註平台則從另一方向夾擊，提供更嚴格的品質保證與領域專精能力。醫療影像、自動駕駛、法律文件等高要求場景，早已轉向能提供可追責標註者的專業平台。MTurk 的泛用低門檻模式在專業 AI 訓練場景中逐漸成為累贅，TechCrunch 直言「這可能是 MTurk 的最後時日」。\n\n#### 章節三：對資料標註生態系的連鎖效應\n\nMTurk 的衰亡存在一個難以逆轉的死亡螺旋：AWS 近年多次無預警關閉工人帳號，導致活躍工人數持續下降，任務完成率隨之惡化，客戶開始出走，進而導致更多工人失去收入動機而離開平台。\n\n更深層的結構性問題於 2023 年浮出水面：研究分析顯示，高達 33–46% 的 MTurk 工人使用大型語言模型代替人工完成任務。這意味著，以「人工智慧訓練」為名收集的資料，實際上可能由 AI 生成——「人工標註資料的真實性假設」徹底崩潰，讓平台的核心價值主張失去立足點。\n\n#### 章節四：後 MTurk 時代——合成資料與 RLHF 的新典範\n\n現代 AI 訓練方法已從根本上改變了對人工標註的依賴程度。自監督學習讓模型能從未標註資料中學習，大幅削減標籤需求；合成資料可降低資料成本達 70%，並在規模與多樣性上遠超人工標註。2026 年合成資料市場規模已達 7.1 億美元，預計 2030 年成長至 23 億美元，直接取代人工標註的市場需求。\n\nRLHF 所需的人類偏好資料，也從 MTurk 式的大規模廉價眾包轉向更少量但更高品質的專家標註。Anthropic、OpenAI 等公司採用的方法，需要的是能判斷細微倫理差異的標註者，而非完成簡單微任務的廉價勞力。MTurk 代表的「大量廉價人工標註」典範，正隨著 AI 能力的提升進入歷史。\n\n> **名詞解釋**\n> RLHF(Reinforcement Learning from Human Feedback) ：從人類回饋中進行強化學習，是訓練 ChatGPT、Claude 等對話 AI 使其符合人類偏好的核心技術，需要人類標註者對模型回應進行排序比較。","MTurk 的退場揭示了 AI 資料標註生態系三個相互交織的結構性轉變：任務市集模式的內在缺陷、品質信任基礎的崩潰，以及資料生產典範的整體移轉。\n\n#### 機制 1：雙邊市集的結構性脆弱\n\nMTurk 採 Requesters（發任務）與 Workers（競標完成）的雙邊市集，以微支付驅動。這種模式在 2005–2015 年間高效運作，因為資訊不對稱讓平台能維持低工資均衡，工人接受低報酬是因為缺乏替代選擇。\n\n但雙邊市集有致命弱點：任何一方的信任崩潰都會引發惡性循環。AWS 多次無預警關閉帳號，摧毀了工人對平台的信任，加速活躍工人池萎縮，最終讓任務完成率跌至難以維持的水平。\n\n> **白話比喻**\n> 想像一個計程車市場：司機接連被無故吊銷執照，剩下的司機開始掛名接單卻讓 AI 代駕，乘客漸漸發現車子沒人開——整個平台就這樣從內部垮掉。\n\n#### 機制 2：LLM 代工現象的品質崩潰\n\n2023 年研究揭示的「LLM 代工」現象是 MTurk 最終的致命一擊：高達 46% 的工人用語言模型自動完成標註任務，再以「人工」名義提交。這在技術上形成了一個自我矛盾——我們用「人類智慧」標註的資料訓練 LLM，而這些資料本身已被 LLM 污染。\n\n這不只是品質問題，更是信任基礎問題。一旦 Requesters 無法確認標註是否真的來自人類判斷，MTurk 相對專業平台的唯一優勢（廉價）便喪失意義——廉價但不可信的資料，在 AI 訓練中的價值為零，甚至為負。\n\n#### 機制 3：合成資料與自動化標註的雙重替代\n\n從技術路徑看，MTurk 面對來自兩個方向的替代：上游的合成資料生成，與下游的自動化標註流水線。SageMaker Ground Truth 將自動化預標註與人工驗證結合，能以 MTurk 不到一半的成本達到更高準確率。\n\n合成資料則完全繞過了人工標註的需求。到 2030 年，合成資料市場預計達 23 億美元——這是從根本上重新定義「訓練資料怎麼來」的範式轉移，不是在改善現有方法，而是讓舊方法本身變得多餘。","",{"recommended":158,"avoid":161},[159,160],"歷史 MTurk 資料集的品質驗證：對 2022 年後收集的資料進行抽樣交叉驗證，評估 LLM 代工污染比例後再決定是否重新標註","SageMaker Ground Truth 遷移評估：在 7 月 30 日截止前完成現有 HIT 定義映射與成本比較，確保標註工作流不中斷",[162,163],"新的 AI 訓練資料收集專案：工人池萎縮導致完成率不穩定，品質保證機制已無法依賴","需要高一致性的大規模標註任務：LLM 代工現象使 MTurk 資料的「人工」屬性不再可信，高要求場景應轉向 Scale AI 等專業平台","#### 環境需求\n\n遷移前需確認：現有 MTurk 任務類型（影像標記、文字分類、偏好排序等）、每月任務量與支出基準、當前品質指標（Worker Agreement Rate、HIT Rejection Rate），以及是否有自定義 Worker UI（HTML／JavaScript 模板）。AWS 帳號須具備 SageMaker 存取權限，Ground Truth 的 IAM 角色設定需提前驗證。\n\n#### 遷移／整合步驟\n\n1. 匯出現有 MTurk HIT(Human Intelligence Task) 定義與 Qualification 篩選條件\n2. 在 SageMaker Ground Truth 建立對應的標註工作類型，選擇 built-in 工作流（影像、文字、影片）或自定義 UI\n3. 於 7 月 30 日前在 Ground Truth 設定 Amazon Mechanical Turk 公開工作隊（沿用現有 MTurk Workers），或評估私有工作隊以提升品質控管\n4. 執行 A/B 比較：同批資料分別送 MTurk 和 Ground Truth，比較 Cohen's Kappa 標籤一致性與每任務完成時間\n5. 若切換至 Scale AI、Surge AI 等外部平台，需撰寫完整標註指引 (Annotation Guidelines) 文件，並安排初期校準輪次\n\n#### 驗測規劃\n\n衡量遷移成功的關鍵指標：標籤 Cohen's Kappa（目標 > 0.8，MTurk 典型值 0.6–0.75）、每任務完成時間（與 MTurk 基準比較）、每標籤有效成本（含品質修正後）。遷移後第一個月應保留 10% 任務在 MTurk 上執行作為對照組，直至確認新平台穩定。\n\n#### 常見陷阱\n\n- 直接照搬 MTurk HIT 設計到新平台——不同平台的工人激勵結構不同，任務指引與品質閾值需重新校準\n- 忽略歷史資料的 LLM 污染風險——2022 年後收集的 MTurk 資料集應進行抽樣品質審計再投入訓練\n- 高估合成資料對邊緣案例的覆蓋——長尾分布（罕見物件、方言、文化特定內容）仍需真實人工標註補充，不可全面替換\n\n#### 上線檢核清單\n\n- 觀測：Worker Agreement Rate（跨工人標籤一致性）、任務完成率趨勢、品質審核拒絕率\n- 成本：有效標籤單價（標籤總成本 ÷ 通過品質審核的標籤數）vs. MTurk 歷史基準\n- 風險：7 月 30 日截止後的緊急備援計畫、現有 MTurk API 整合的服務中斷處置","#### 競爭版圖\n\n- **直接競品**：SageMaker Ground Truth（AWS 自家替代方案，MTurk 工人可直接遷移，阻力最低）、Scale AI（估值超過 140 億美元，高端 AI 訓練資料市場最大受益者）、Surge AI、Prolific（學術與 RLHF 偏好標註市場）\n- **間接競品**：合成資料平台（Gretel.ai、Argilla、Lilac）；自監督學習工具鏈（從根本上削減對標註資料的依賴）\n\n#### 護城河類型\n\n- **工程護城河**：SageMaker Ground Truth 的自動化預標註機制結合 IAM 整合，降低企業導入摩擦，同一 AWS 生態系減少供應商切換阻力\n- **生態護城河**：Scale AI 在自動駕駛、醫療、法律等垂直領域積累的標註指引體系與品質稽核機制，難以在短期內複製\n\n#### 定價策略\n\nMTurk 退場使資料標註市場往兩極化發展。大量泛用任務往合成資料移動（幾乎零邊際成本）；高價值的人類偏好資料、倫理判斷、領域專業標註往 Scale AI 等平台集中，每任務單價可能較 MTurk 高出 5–20 倍。\n\n「廉價眾包」這個市場區間正在消失，取而代之的是「免費合成」與「高價專業」的兩端格局。\n\n#### 企業導入阻力\n\n- 現有 MTurk 工作流程的遷移成本：任務設計重寫、品質閾值重設、API 整合替換\n- 合規要求（GDPR、HIPAA）在新平台的重新驗證，尤其是涉及個人資料的標註任務\n- 歷史 MTurk 資料集的品質不確定性，部分資料集可能需要重新標註\n\n#### 第二序影響\n\n- AI 訓練資料成本兩極化：合成資料壓低低端價格，同時高品質人類標註的溢價持續提升\n- 全球眾包工人（尤其是發展中國家依賴微任務收入的群體）面臨工作機會收縮，是 AI 發展帶來的直接就業衝擊\n- 標註市場從碎片化眾包走向集中化：少數有品質保證的專業平台將主導市場，可選供應商的多樣性下降\n\n#### 判決：Scale AI 等專業標註平台的主要受益方（MTurk 客戶出走加速高端市場整合）\n\nMTurk 的退場是 AI 資料供應鏈重組的縮影，而非終結。市場並未消失，而是往「更少工人、更高品質、更高單價」的方向集中。短期內，現有 MTurk 客戶需要緊急評估遷移路徑；長期來看，合成資料的崛起將進一步壓縮人工標註的市場空間，只留下機器暫時無法勝任的判斷工作。",[167,168],"眾包資料仍有其難以替代的價值：在邊緣案例標註、文化多樣性語料、低資源語言等長尾需求上，大規模廉價人力仍難以被合成資料完全取代，MTurk 的退場可能在某些垂直領域留下真實的供給缺口","MTurk 的工人品質問題並非結構性無解——更嚴格的品質把關機制和激勵重設計本可重建可靠性，AWS 選擇廢棄而非修復，更多反映的是商業優先順序，而非技術必然",[170,173,176,179,182],{"platform":65,"user":171,"quote":172},"socialmedialab.ca(6 upvotes)","一個時代的終結。Amazon Mechanical Turk 將於 2026 年 7 月 30 日起關閉新客戶申請。Amazon 最近將 Mechanical Turk 服務加入「Services in Maintenance」清單——這是 AWS 即將退役服務的慣用說法。",{"platform":69,"user":174,"quote":175},"@sarahookr（Cohere 研究主管）","Mechanical Turk 這個名稱源自 1769 年一台會下棋的自動機器人，它看似能與人類棋手一較高下……因為裡面藏著一個人。如果這就是 Amazon MTurk 的命名靈感，那真是一個對歷史的絕妙呼應。",{"platform":65,"user":177,"quote":178},"Benjamin Han（Bluesky 用戶，1 upvote）","本週稍早，AWS 將 Amazon SageMaker AI – Mechanical Turk 服務加入「Services in Maintenance」清單——這是 AWS 即將退役服務的慣用說法。Mechanical Turk 網站也新增警告，說明將於 2026 年 7 月 30 日起關閉新客戶申請，現有用戶則可繼續使用。",{"platform":65,"user":180,"quote":181},"Annemarie Bridy（Bluesky 用戶，1 upvote）","Mechanical Turk 終結的開端似乎已經到來。",{"platform":69,"user":183,"quote":184},"@ebizfacts（X 用戶）","聽說過 Amazon Mechanical Turk 嗎？它基本上是一個微任務平台，讓你可以靠做這些事賺取報酬：圖片標記、音頻內容描述、填寫問卷，以及其他類似任務。它在許多國家都可使用，包括印度。",3,[187,189,191],{"type":80,"text":188},"評估 SageMaker Ground Truth 能否接替現有 MTurk 工作流——7 月 30 日截止前完成 HIT 定義映射，並執行一次 A/B 成本與品質比較",{"type":83,"text":190},"為 2022 年後收集的 MTurk 資料集建立品質重新驗證流程，針對 LLM 代工污染進行抽樣稽核，再決定是否重新標註或廢棄使用",{"type":86,"text":192},"追蹤合成資料工具（如 Argilla、Gretel.ai）與 RLHF 標註平台（如 Prolific、Scale RLHF）的功能演進，判斷下一代資料基礎設施的格局走向",{"category":194,"source":18,"title":195,"subtitle":196,"publishDate":6,"tier1Source":197,"supplementSources":200,"tldr":213,"context":225,"mechanics":226,"benchmark":227,"useCases":228,"engineerLens":237,"businessLens":238,"devilsAdvocate":239,"community":242,"hypeScore":185,"hypeMax":76,"adoptionAdvice":258,"actionItems":259},"tech","GPT-5.5 Codex reasoning-token clustering 疑雲：社群揭露效能退化的隱藏機制","390,195 筆 API 記錄揭示 reasoning 在 516 token 強制截斷，付費用戶悄悄獲得更少推理空間",{"name":198,"url":199},"openai/codex Issue #30364 — reasoning-token clustering","https://github.com/openai/codex/issues/30364",[201,205,209],{"name":202,"url":203,"detail":204},"Hacker News — GPT-5.5 Codex reasoning-token clustering 討論","https://news.ycombinator.com/item?id=48789428","社群對 clustering 機制的推測、workaround 分享與開放 harness 討論",{"name":206,"url":207,"detail":208},"openai/codex Issue #26876 — GPT-5.5 degradation over time","https://github.com/openai/codex/issues/26876","2,702 個 prompt 的用戶挫敗率追蹤，記錄從 3.9% 升至 12.1% 的退化歷程",{"name":210,"url":211,"detail":212},"Let's Data Science — GPT-5.5 Reasoning-Token Clustering 分析","https://letsdatascience.com/news/gpt-55-exhibits-reasoning-token-clustering-at-fixed-boundari-63ae3735","月度 clustering 比率走勢與統計方法的深度分析",{"tagline":214,"points":215},"GPT-5.5 的推理正在悄悄縮水——你付了全價，卻只得到六分之一的思考空間",[216,219,222],{"label":217,"text":218},"技術","390,195 筆記錄顯示 GPT-5.5 的 reasoning token 異常聚集在 516、1,034、1,552 三個固定值，clustering 比率 5 月峰值達 53.30%，平均 reasoning 用量同期從 268 驟降至 107。",{"label":220,"text":221},"成本","付費用戶以相同價格獲得更少的有效推理空間，且 OpenAI 未發出任何版本公告，複雜任務（需 6,000–8,000 reasoning token）的準確率悄悄退化，無從察覺。",{"label":223,"text":224},"落地","已知 workaround：移除系統提示中的 `## Intermediary updates` 區塊可消除聚集問題；長期建議採用開放 harness 架構，保持後端可替換性以應對供應商的靜默版本變更。","#### 章節一：問題浮現——開發者如何發現 reasoning-token 異常聚集\n\n2026 年 6 月 27 日，GitHub 用戶 vguptaa45 在 openai/codex 倉庫提交 issue #30364，呈報 GPT-5.5 的 reasoning 輸出 token 出現高度異常的聚集現象。這項分析橫跨 390,195 筆 API 回應紀錄，時間範圍從 2026 年 2 月延伸至 6 月，是迄今最具規模的個人自發式效能監測案例之一。\n\n觸發深度分析的關鍵訊號在於一個驚人的倍數差異：GPT-5.5 的「恰好命中 516 token／超過 516 token」比率高達 44.0%，而非 GPT-5.5 模型僅 1.3%。\n\nGPT-5.5 雖只佔全部回應的 19.3%，卻貢獻了 82% 的「恰好 516 token」事件，三個峰值（516、1,034、1,552 token）之間的間距恰好為 518，呈現出顯然非隨機的週期性結構。\n\n#### 章節二：Clustering 機制推測與社群復現實驗\n\n社群對「為何恰好停在 516」提出多種假說。第一條路線指向提示詞觸發：用戶 nsingh2 以糖果計算謎題重現問題，4/10 次執行停在 516 token 且全數答錯。\n\n他進一步發現，移除系統提示中的 `## Intermediary updates` 區塊可完全消除問題，推測模型將「中途更新信號」誤解為最終回答，導致 reasoning 提前終止。\n\n第二條路線則指向底層基礎設施：516 ≈ 512 bytes 對齊加上 header overhead，有可能對應到記憶體配置邊界問題。\n\n月度數據走勢提供了另一條線索——clustering 比率從 2 月的 0.11% 暴增至 5 月的 53.30%，6 月雖回落至 35.84%，卻同步出現 GPT-5.5 平均 reasoning token 用量從 268.1 驟降至 106.9 的「用得少、錯得多」反常組合，意味著問題不是孤立錯誤，而是持續性退化。\n\n> **名詞解釋**\n> Clustering（聚集）：指模型輸出的 token 數量異常集中在某幾個固定數值，而非應有的平滑分佈，是偵測隱性行為退化的一種統計方法。\n\n#### 章節三：開放 Harness 的保險效應：隨時可以換船\n\nHN 首則回覆 selectodude 直接點出了應對策略的核心：「這正是使用開放 harness 的美妙之處——有狀況我可以隨時換船。」開放 harness 讓開發者在後端模型退化時，以最小成本切換替代方案，不必等待供應商的修復公告。\n\n社群同步分享了 MIT 授權的 `clankerbend` 工具（可透過 `npx clankerbend` 直接使用），允許將相同 harness 配置接上不同後端模型。\n\nsolarkraft 以 Deepseek 的行為作為對比基準——Deepseek 在多輪對話中會動態調整 reasoning budget，依任務複雜度分配推理空間，展示了「可觀測、可調控」的 reasoning 機制應有的樣貌，與 GPT-5.5 的不透明截斷形成鮮明對比。\n\n> **名詞解釋**\n> Harness（執行容器）：包裹 LLM API 呼叫的框架層，負責管理提示詞、上下文、輸出格式等邏輯，使後端模型可被替換而不影響上層應用邏輯。\n\n#### 章節四：Reasoning 模型的隱性成本與信任危機\n\nissue 作者明確聲明「數據不能證明 chain-of-thought 被截斷」，但現實問題在於：需要 6,000–8,000 token 才能穩定正確回答的任務，在 reasoning 空間悄悄退化至 106.9 token 後，付費用戶面臨實質性的效能下修，且缺乏任何版本公告或公開說明。\n\n同期的 issue #26876 記錄了更廣泛的退化數據：2,702 個 prompt、530 個 session，用戶挫敗率從基準期的 3.9% 升至晚期 12.1%。\n\n多名社群成員將此事與 Anthropic 先前的 Claude Code 退化事件相提並論，顯示大型推理模型的「隱性降級」已成跨廠商的系統性信任問題——能力變化不透明、用戶難以偵測、退化與修復均無公告。","GPT-5.5 的 reasoning-token clustering 問題揭示了大型 LLM 在生產環境中可能發生的一類隱性退化：不是直接報錯，而是以「看似正常輸出、實際推理截斷」的形式悄悄降級。\n\n#### 機制 1：Token 聚集的統計特徵\n\n三個峰值（516、1,034、1,552 token）間距恰好為 518，在 390,195 筆回應中，516 token 精確命中 3,363 次。這種週期性結構在隨機推理過程中極不可能自然出現，強烈暗示存在某種外部截斷邊界。\n\nGPT-5.5 的 exact-516／≥516 比率 (44.0%) 與其他模型 (1.3%) 相差 33 倍，是觸發深度分析的核心訊號。\n\n#### 機制 2：提示詞觸發假說\n\n用戶 nsingh2 以糖果計算謎題進行可重現實驗，4/10 次執行停在 516 token 且答案全部錯誤。移除系統提示中的 `## Intermediary updates` 區塊後，所有執行均成功。一個合理推測是：模型的注意力機制將「中途更新」格式標記誤解為完成信號，導致 reasoning chain 在中途提前收斂。\n\n> **名詞解釋**\n> Chain-of-thought（思維鏈）：讓模型在給出最終答案前先逐步推理的技術。Reasoning model 的 chain-of-thought 通常在後台運行，消耗 reasoning token，不直接顯示給用戶。\n\n#### 機制 3：記憶體對齊假說\n\n516 ≈ 512 bytes + header overhead，指向底層 KV cache 或批次處理緩衝區的記憶體邊界對齊問題。若推理引擎在分配記憶體時以 512 bytes 為單位，且某些路徑下未正確擴展緩衝區，便可能導致 reasoning 在邊界處強制截斷，三個峰值恰好是 518 的整數倍也支持這一假說。\n\n> **白話比喻**\n> 想像你的筆記本只有 512 頁，但老師要求解題過程必須寫滿。若有人在第 516 頁貼了一張「完成」貼紙，你可能就此擱筆——即使後面還有空白頁，即使答案根本還沒算出來。","#### Clustering 事件統計\n\n| 指標 | 數值 |\n|---|---|\n| 分析樣本 | 390,195 筆 API 回應 |\n| 516 token 精確命中 | 3,363 次 |\n| GPT-5.5 佔樣本比例 | 19.3% |\n| GPT-5.5 貢獻 516-event 比例 | 82% |\n| exact-516／≥516 比率 (GPT-5.5) | 44.0% |\n| exact-516／≥516 比率（其他模型）| 1.3% |\n\n#### 月度退化走勢\n\n| 月份 | Clustering 比率 | 平均 Reasoning Token |\n|---|---|---|\n| 2026-02 | 0.11% | 268.1 |\n| 2026-03 | 2.45% | — |\n| 2026-05 | 53.30% | — |\n| 2026-06 | 35.84% | 106.9 |\n\n#### 用戶挫敗率 (Issue #26876)\n\n2,702 個 prompt、530 個 session：基準期挫敗率 3.9%，退化後期升至 12.1%，升幅約 3 倍。",{"recommended":229,"avoid":233},[230,231,232],"使用開放 harness（如 MIT 授權的 clankerbend）包裝 API 呼叫，保持後端模型可替換性","監控每次呼叫的 reasoning token 分佈，以 clustering 比率作為早期退化偵測指標","對複雜推理任務（需 6,000+ reasoning token）預先設置 Deepseek 或其他模型作為 fallback",[234,235,236],"在系統提示中使用 `## Intermediary updates` 格式的結構化標記（已知觸發條件）","在無獨立監控機制的情況下，直接依賴 GPT-5.5 執行高精度推理任務","以單次成功測試作為生產可用性判斷依據，忽略 reasoning token 的月度趨勢變化","#### 環境需求\n\n需要一個能記錄每次 API 回應 reasoning token 數量的日誌系統。建議使用 OpenAI 的 usage 物件中的 `completion_tokens_details.reasoning_tokens` 欄位，搭配 SQLite 或時序資料庫儲存歷史分佈，以便計算 clustering 比率。\n\n#### 最小 PoC\n\n```python\nimport openai, sqlite3, collections, time\n\nclient = openai.OpenAI()\ndb = sqlite3.connect(\"reasoning_log.db\")\ndb.execute(\"CREATE TABLE IF NOT EXISTS logs (ts REAL, model TEXT, reasoning_tokens INT)\")\n\ndef call_and_log(prompt: str, model: str = \"gpt-5.5\"):\n    resp = client.chat.completions.create(\n        model=model,\n        messages=[{\"role\": \"user\", \"content\": prompt}],\n    )\n    r_tokens = getattr(resp.usage.completion_tokens_details, \"reasoning_tokens\", 0)\n    db.execute(\"INSERT INTO logs VALUES (?, ?, ?)\", (time.time(), model, r_tokens))\n    db.commit()\n    return resp, r_tokens\n\ndef clustering_rate(model: str = \"gpt-5.5\", window: int = 100) -> float:\n    rows = db.execute(\n        \"SELECT reasoning_tokens FROM logs WHERE model=? ORDER BY ts DESC LIMIT ?\",\n        (model, window)\n    ).fetchall()\n    counter = collections.Counter(r[0] for r in rows)\n    return counter[516] / len(rows) if rows else 0.0\n```\n\n#### 驗測規劃\n\n連續呼叫 50–100 次後，觀察 reasoning token 分佈是否呈現單峰（健康）或固定聚集（退化）。clustering_rate > 5% 視為警示；> 20% 視為需切換後端或回報 issue 的嚴重退化。\n\n#### 常見陷阱\n\n- 僅看最終輸出正確率，忽略 reasoning token 數量：退化初期輸出可能仍「看起來」合理\n- 誤將 workaround（移除 `## Intermediary updates`）當作根本修復：底層問題未解決，其他觸發條件可能持續出現\n- 未記錄完整 model ID（含版本戳記）：OpenAI 可能在不公告的情況下調整後端行為\n\n#### 上線檢核清單\n\n- 觀測：reasoning token 分佈（每日 p50／p95）、516 ± 5 token clustering 比率、任務成功率\n- 成本：reasoning token 驟降時，推理成本應同比下降；若成本不降但準確率降，需深查計費異常\n- 風險：避免系統提示中使用結構化 `##` heading 格式；為關鍵推理任務設置輸出驗證層","#### 競爭版圖\n\n- **直接競品**：Anthropic Claude Code（同樣發生過未公告的效能退化事件，已成社群比較基準）、Deepseek（動態調整 reasoning budget，行為更透明可觀測）\n- **間接競品**：本地部署的開源推理模型（Llama 4、Qwen 3），不依賴 API 供應商的版本管理決策\n\n#### 護城河類型\n\n- **工程護城河**：GPT-5.5 在廣泛基準測試上仍保持領先，clustering 問題主要影響特定提示格式，不代表全面能力退化\n- **生態護城河**：openai/codex 是最大的 AI 輔助編程生態，用戶遷移成本極高，短期難以大規模流失\n\n#### 定價策略\n\n此事件本質上是一場隱性的服務降級——用戶以相同價格獲得了更少的 reasoning capacity，且無任何通知。\n\n若 OpenAI 持續維持不透明的版本管理方式，企業用戶的 SLA 談判籌碼將增加，可能要求以「有效 reasoning token」而非「消耗 reasoning token」作為計費基礎。\n\n#### 企業導入阻力\n\n- 缺乏 reasoning token 使用量的官方 SLA 保證，難以預估複雜任務的成本與品質穩定性\n- 版本變更無公告，CI／CD 迴歸測試可能在不知情的情況下對齊退化版本的行為\n\n#### 第二序影響\n\n- 推動企業採用「模型無關」架構，增加中間層框架（如 LiteLLM、Martian）的商業價值\n- 社群自發監測工具（個人規模的 390K 筆日誌分析）可能演變為商業化的 LLM 健康監測服務\n\n#### 判決：生態信任成本正在累積（短期可消化，長期需透明度改革）\n\nGPT-5.5 的 clustering 問題本身可能有 workaround，但連同 Claude Code 退化事件，「隱性降級」已成跨廠商的系統性模式。企業在採用大型推理模型時，必須將「不透明版本管理風險」列入採購評估，並在架構層保留切換能力。",[240,241],"OpenAI 尚未官方確認 clustering 現象，目前所有分析均來自用戶自發監測，存在抽樣偏差的可能性，測試集中特定提示格式可能放大了問題的實際規模","已知 workaround（移除 `## Intermediary updates` 區塊）被多名用戶驗證有效，若問題僅限於特定提示格式，實際受影響的生產用例可能遠少於統計數字所呈現的嚴重程度",[243,246,249,252,255],{"platform":59,"user":244,"quote":245},"selectodude（HN 用戶）","這正是使用開放 harness 的美妙之處——有狀況我可以隨時換船。",{"platform":59,"user":247,"quote":248},"solarkraft（HN 用戶）","Deepseek 定期為我這樣做，至少在主動探索某些內容時（多輪對話）。它會根據任務大幅調整 reasoning budget。",{"platform":59,"user":250,"quote":251},"zuzululu（HN 用戶）","我剛試了，成功了！！！我不知道為什麼有效，但現在的 GPT-5.5 感覺和一個半月前我記憶中的一樣了。",{"platform":59,"user":253,"quote":254},"lmwnshn（HN 用戶）","MIT 授權，隨便用，你也可以直接透過 `npx clankerbend` 使用它（前提是已安裝 Node.js）。GitHub 頁面的截圖連結到 YouTube 示範影片。",{"platform":65,"user":256,"quote":257},"notengoprisa.bsky.social（Bluesky 用戶，5 讚）","Claude：你 20 歐的訂閱只夠用 3 個 Opus prompt，到達上限後請去結帳或等到下個月。\n\nCodex：可以讓 GPT-5.5 不間斷工作數小時，到達上限後可以免費重置或等你吃完午飯回來。","先觀望",[260,262,264],{"type":80,"text":261},"在現有 GPT-5.5 呼叫中加入 reasoning token 日誌，統計 516 token 命中率，即可快速判斷自身工作負載是否受影響",{"type":83,"text":263},"為推理任務搭建開放 harness 層（參考 MIT 授權的 clankerbend），確保後端模型可替換，不被單一供應商的版本管理決策綁死",{"type":86,"text":265},"追蹤 openai/codex issue #30364 的官方回應，以及 OpenAI 是否會承諾增加 reasoning token 分佈的透明度",{"category":194,"source":12,"title":267,"subtitle":268,"publishDate":6,"tier1Source":269,"supplementSources":272,"tldr":281,"context":290,"mechanics":291,"benchmark":292,"useCases":293,"engineerLens":303,"businessLens":304,"devilsAdvocate":305,"community":309,"hypeScore":75,"hypeMax":76,"adoptionAdvice":325,"actionItems":326},"百度 Unlimited OCR：模仿人類遺忘機制，一次讀懂數十頁文件","R-SWA 注意力機制讓 KV Cache 不再隨頁數爆炸，單次推論通道突破 40 頁長文件處理上限",{"name":270,"url":271},"The Decoder","https://the-decoder.com/baidus-unlimited-ocr-processes-dozens-of-document-pages-in-one-pass-by-treating-memory-like-human-forgetting/",[273,277],{"name":274,"url":275,"detail":276},"MarkTechPost","https://www.marktechpost.com/2026/06/24/baidu-releases-unlimited-ocr-a-3b-model-that-keeps-the-kv-cache-flat-for-long-document-parsing/","詳細介紹 3B MoE 架構與 KV Cache 固定大小設計的技術細節",{"name":278,"url":279,"detail":280},"arXiv 2606.23050","https://arxiv.org/pdf/2606.23050","Unlimited OCR Works 論文原文，含 R-SWA 機制完整推導與 OmniDocBench 實驗數據",{"tagline":282,"points":283},"KV Cache 不再隨頁數爆炸——百度用「選擇性遺忘」讓 OCR 一口氣讀完整本手冊",[284,286,288],{"label":217,"text":285},"R-SWA 讓 KV Cache 維持常數大小，40+ 頁文件編輯距離低於 0.11，OmniDocBench v1.5 達 93.23 分，比 DeepSeek OCR 高出 6.22 分",{"label":220,"text":287},"3B MoE 架構推論僅啟用 500M 參數，可在 12GB VRAM 執行，每秒 5,580 tokens 比 DeepSeek OCR 快 12.7%，延遲 6,000 步後保持平坦",{"label":223,"text":289},"MIT 授權開源，模型權重已上傳 Hugging Face，適合企業批次文件數位化與本地私有化 RAG 管道建構","#### 章節一：傳統 OCR 的頁數瓶頸與長文件處理困境\n\n標準 Transformer 解碼器每生成一個 token，就會向 KV Cache 新增一筆鍵值記錄。當文件頁數增加、輸出 token 數量暴增時，快取記憶體呈線性增長，最終逼近硬體上限。\n\n這迫使主流 OCR 系統採取「逐頁重置」策略——讀完一頁就清空記憶體，再從頭開始讀下一頁。這個工程妥協雖能繞過記憶體限制，卻導致跨頁語境斷裂，表格或公式一旦跨越頁面邊界便無法連貫辨識。\n\n業界主流系統因此形成約 10 頁的事實上限。長文件處理成為文件 AI 的隱形天花板，而非技術層面的真正解決。\n\n#### 章節二：Unlimited OCR 的核心設計——像人類一樣選擇性遺忘\n\n百度的解法來自認知科學的類比：人類閱讀長文件時，持續保留對「眼前頁面圖像」的感知，同時只留住最近讀過的幾段文字作為上下文。\n\nReference Sliding Window Attention(R-SWA) 將這個機制轉化為注意力設計。解碼器的注意力被拆成兩個區塊，各自負責不同的記憶功能。\n\n第一區塊對所有視覺 token 與 prompt 保持完整注意力，確保模型隨時「看得到」原始圖像；第二區塊對已生成的輸出 token 只保留最近 128 個，舊的記憶自動滑出視窗。\n\n> **名詞解釋**\n> KV Cache（鍵值快取）：Transformer 推論時儲存歷史注意力鍵值對的記憶體結構，大小直接影響可處理的序列長度上限。\n\nKV Cache 因此維持固定長度佇列，上界為常數 Lm + n（Lm 為視覺 token 數，n 為滑動視窗大小），不隨頁數增加。這是突破頁數限制的根本機制。\n\n#### 章節三：技術架構解析與效能實測\n\nUnlimited OCR 採用 3B 參數的 Mixture-of-Experts(MoE) 架構，推論時僅啟用 500M 參數。視覺編碼器採用 DeepEncoder，將每頁 1024×1024 的 PDF 影像壓縮至 256 個視覺 token。\n\n> **名詞解釋**\n> Mixture-of-Experts(MoE) ：一種稀疏神經網路架構，每次推論只激活部分「專家」子網路，在保有大模型容量的同時大幅降低計算成本。\n\n效能測試中，在 OmniDocBench v1.5 獲得 93.23 分，較 DeepSeek OCR 基準線 (87.01) 提升 6.22 分；v1.6 版本進一步達到 93.92%，確認跨版本一致性。\n\n推論速度達每秒 5,580 tokens，比 DeepSeek OCR 快 12.7%，且在 6,000 步解碼後延遲仍保持平坦。40 頁以上文件的文字編輯距離 (Edit Distance) 低於 0.11，顯示辨識品質未隨頁數衰退。\n\n#### 章節四：文件 AI 的下一步——從辨識到理解\n\nR-SWA 論文作者明確指出，這個注意力機制並非 OCR 專屬，而是任何「解碼器需要永久存取固定來源、同時生成長輸出」場景的通用解法——語音辨識 (ASR) 與機器翻譯均在適用場景之列。\n\n對文件 AI 產業而言，Unlimited OCR 的意義超越辨識準確率的提升。整本報告單次推論成為可能，讓下游 RAG 管道不再需要依賴分頁切片與跨頁拼接等工程補丁。\n\nMIT 授權的開源策略讓這個突破得以直接進入企業文件數位化管道，而非停留在論文實驗室。開源社群的二次開發將加速生態聚集，推動文件 AI 的整體基準線向上移動。","R-SWA 的設計目標是讓 KV Cache 大小與輸出長度完全脫鉤。傳統解碼器每步都把新生成的 token 推入快取，導致快取無限增長；R-SWA 透過結構性分區解決這個問題。\n\n#### 機制 1：視覺 Token 全量保留（永久記憶層）\n\n視覺編碼器將每頁 1024×1024 的 PDF 影像壓縮成 256 個視覺 token，連同原始 prompt 一起放入注意力的「永久區」。無論生成多少輸出，解碼器對這個區塊始終保持完整注意力。\n\n這個設計確保每個輸出 token 都能回望完整的原始圖像資訊，不受輸出長度影響。來源圖像是固定且完整的，不能被遺忘——這是 OCR 場景的關鍵約束。\n\n#### 機制 2：輸出 Token 滑動視窗（短期工作記憶）\n\n已生成的輸出 token 進入「滑動視窗區」，大小固定為 128 個 token。每生成一個新 token，最舊的一個就滑出視窗，整個快取大小維持不變。\n\nKV Cache 的總大小上界為 Lm（視覺 token 數）＋ n（滑動視窗大小），是與頁數無關的常數。這個設計類比人類閱讀時只記住最近幾段，而非逐字回憶全文。\n\n> **白話比喻**\n> 把 Unlimited OCR 想像成一位速記員：桌上永遠攤開著原始文件（視覺 token 永久區），便條紙只記最近 128 個字（滑動視窗），舊便條紙就丟掉。桌子永遠不會被便條紙淹沒，不管文件有多厚。\n\n#### 機制 3：MoE 稀疏啟動（推論效率壓縮）\n\n3B 參數的 MoE 架構在推論時僅啟用 500M 參數，讓整個系統可在 12GB VRAM 的消費級 GPU 上執行，大幅降低硬體門檻。\n\n每秒 5,580 tokens 的吞吐量在 6,000 步後仍維持平坦延遲，意味著長文件推論不會隨頁數出現速度衰減。這是 KV Cache 固定大小與 MoE 稀疏啟動共同帶來的系統級效益。","#### OmniDocBench 綜合評測\n\nUnlimited OCR 在 OmniDocBench v1.5 獲得 93.23 分，較 DeepSeek OCR 基準線 (87.01) 提升 6.22 分；v1.6 版本進一步達到 93.92%，確認跨版本基準一致性。\n\n#### 子任務分解\n\n- 表格結構辨識 (TEDS) ：90.93（較基準線提升近 6 個百分點）\n- 公式辨識：92.61\n- 40 頁以上文件文字編輯距離 (Edit Distance) ：\u003C 0.11\n\n#### 速度對比\n\n推論速度達每秒 5,580 tokens，比 DeepSeek OCR 的 4,951 快 12.7%。在 6,000 步解碼後延遲曲線保持平坦，驗證 KV Cache 固定大小設計對長文件推論的系統級穩定性。",{"recommended":294,"avoid":299},[295,296,297,298],"企業文件數位化管道——年報、合約、技術手冊等 20+ 頁 PDF 的批次 OCR 處理","RAG 知識庫建構——無需分頁切片即可完整抽取跨頁表格與公式，降低後處理複雜度","學術論文結構化——公式辨識達 92.61，適合理工科論文批次轉換為結構化格式","本地私有化部署——12GB VRAM 即可執行，適合無法使用雲端 API 的資安合規場景",[300,301,302],"單頁高精度掃描文件——相較於專注單頁的輕量模型，跨頁設計帶來的優勢消失","手寫文字辨識——訓練資料以印刷體 PDF 為主（單頁：多頁比例 9：1），手寫場景未見充分驗證","即時線上辨識（\u003C 500ms 回應要求）——批次推論模式不適合嚴格低延遲的互動場景","#### 環境需求\n\n最低需求：12GB VRAM(NVIDIA GPU) ，Python 3.10+，CUDA 12.x。模型權重約 6GB，建議使用 A10G 或 RTX 3090/4090 進行本地測試；生產環境建議 A100 或 H100 以獲得最佳吞吐量。\n\n#### 最小 PoC\n\n```python\nfrom transformers import AutoModelForCausalLM, AutoTokenizer\nfrom PIL import Image\nimport torch\n\nmodel_id = \"baidu/Unlimited-OCR\"\ntokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)\nmodel = AutoModelForCausalLM.from_pretrained(\n    model_id,\n    torch_dtype=torch.bfloat16,\n    device_map=\"auto\",\n    trust_remote_code=True\n)\n\n# 多頁 PDF 需先轉為圖像列表（如使用 pdf2image）\nimages = [Image.open(f\"page_{i}.png\") for i in range(1, 41)]\nresult = model.ocr(images, tokenizer=tokenizer)\nprint(result)\n```\n\n#### 驗測規劃\n\n以 OmniDocBench 評測集進行基準對比，預期綜合分數應落在 93+ 區間。自有文件測試時，重點驗證跨頁表格的 TEDS 分數與公式辨識的 LaTeX 輸出正確率。\n\n#### 常見陷阱\n\n- 序列長度上限 32K tokens：單次推論約可處理 40-60 頁（依內容密度而異），超過上限需手動分批\n- `trust_remote_code=True` 為必要參數：模型使用自訂注意力實作，需允許執行自定義程式碼\n- 訓練資料單頁：多頁比例 9：1：超長文件（100+ 頁）表現未經充分驗證，需自行評測\n\n#### 上線檢核清單\n\n- 觀測：TEDS 分數、Edit Distance、每頁推論延遲、VRAM 峰值使用量\n- 成本：單頁推論約 0.18 秒 (A100) ，40 頁文件約 7.2 秒；批次作業建議搭配 vLLM 部署提升吞吐\n- 風險：MIT 授權允許商用，但 `trust_remote_code=True` 在生產環境需通過安全審查；訓練資料著作權合規性需評估","#### 競爭版圖\n\n- **直接競品**：DeepSeek OCR（基準線 87.01，速度慢 12.7%）、Mistral OCR、Azure Document Intelligence、Google Document AI\n- **間接競品**：傳統 Tesseract + 後處理管道、Adobe Acrobat Pro OCR、AWS Textract\n\n#### 護城河類型\n\n- **工程護城河**：R-SWA 注意力機制設計巧思，競品短期需重新訓練才能複製；6,000 步平坦延遲曲線意味著長文件場景效能優勢難以靠微調追上\n- **生態護城河**：MIT 授權降低企業採用門檻，Hugging Face 模型權重讓整合成本極低；開源社群二次開發將加速生態聚集\n\n#### 定價策略\n\n目前以 MIT 授權開源釋出，無直接授權費用。企業若使用百度雲端 API 版本，預期採用與 DeepSeek OCR API 類似的 token 計費模式。本地部署零授權成本是相對 Azure、Google 等商業競品的顯著優勢。\n\n#### 企業導入阻力\n\n- `trust_remote_code=True` 觸發安全審查流程，大型企業 IT 合規部門可能需要額外時間審核自定義注意力程式碼\n- 訓練資料來源透明度不足，GDPR、HIPAA 等合規需求可能要求更詳細的資料說明\n- 百度品牌在部分市場存在地緣政治敏感性，影響企業採購決策\n\n#### 第二序影響\n\n- RAG 管道設計簡化：文件切片、跨頁拼接的工程補丁需求降低，開發者可專注語義理解層\n- 文件 AI 基準競爭加速：Unlimited OCR 將 OmniDocBench 門檻抬高至 93+，迫使競品加快研發週期\n\n#### 判決：值得商業評估（MIT 授權消除授權壁壘，長文件場景優勢明確）\n\n對於有大量 20+ 頁文件處理需求的企業，Unlimited OCR 提供明確的技術優勢與零授權成本，短期內值得進行 PoC 評估。地緣政治與合規疑慮是需要獨立評估的風險因素。",[306,307,308],"訓練資料單頁：多頁比例為 9：1，意味著模型在超長文件（100+ 頁）場景的泛化能力仍有疑慮，OmniDocBench 的優異成績能否在真實世界長文件中複現，需要獨立驗證","R-SWA 的滑動視窗只保留最近 128 個輸出 token，若文件有跨越 128 token 以上的長距離依賴（如附錄引用正文第一頁的定義），模型可能丟失關鍵上下文","百度開源動機可能帶有生態競爭目的，後續雲端 API 版本是否維持同等性能、是否出現功能分裂，是長期採用前需要考量的風險",[310,313,316,319,322],{"platform":69,"user":311,"quote":312},"@DataChaz（Charly Wargnier，資料 / AI 內容創作者）","百度剛剛為文件 AI 領域帶來了一個絕對的遊戲規則改變者，叫做 Unlimited-OCR，它可以在單次推論中轉錄整本書。大多數視覺模型只讀一頁，然後遺忘上下文，最終撞上性能崩潰的牆。",{"platform":59,"user":314,"quote":315},"HN 用戶 potsandpans","這剛發布：huggingface.co/baidu/Unlimited-OCR，可以在 12GB VRAM 上舒適運行。我試了一下，看起來確實很有競爭力。不知道和你的使用場景相比如何。",{"platform":69,"user":317,"quote":318},"@techNmak（Tech with Mak，科技教育創作者）","百度讓其他所有 OCR 模型看起來都過時了。它叫 Unlimited OCR，在單次推論中轉錄 40+ 頁，記憶體零增長。現有每個模型都在每頁後重置：第一頁，重置，第二頁，重置——那是工程妥協，不是解決方案。",{"platform":59,"user":320,"quote":321},"HN 用戶 satvikpendem","百度也發布了一個更快的 OCR 模型：github.com/baidu/Unlimited-OCR",{"platform":59,"user":323,"quote":324},"HN 用戶 HDBaseT","百度最近在 AI 領域做了一些有趣的事情，Unlimited OCR 模型非常出色。","值得一試",[327,329,331],{"type":80,"text":328},"從 Hugging Face 下載 baidu/Unlimited-OCR 模型，在 12GB VRAM 本地 GPU 上測試自有 20+ 頁 PDF，對比現有 OCR 管道的 Edit Distance 與表格辨識準確率",{"type":83,"text":330},"將 Unlimited OCR 整合至 RAG 知識庫建構管道，替換現有分頁切片＋跨頁拼接邏輯，評估端對端準確率提升與工程複雜度降低的成本效益",{"type":86,"text":332},"追蹤 R-SWA 機制在語音辨識 (ASR) 與機器翻譯場景的應用進展，以及競品（DeepSeek、Mistral、Azure）對長文件 OCR 基準的反應策略",[334,371,404,429,467,495,522,558],{"category":194,"source":11,"title":335,"publishDate":6,"tier1Source":336,"supplementSources":338,"coreInfo":347,"engineerView":348,"businessView":349,"viewALabel":350,"viewBLabel":351,"bench":156,"communityQuotes":352,"verdict":369,"impact":370},"Claude Code + Fable 5 數小時內將 2003 年 C&C 即時戰略遊戲移植到 iOS",{"name":270,"url":337},"https://the-decoder.com/claude-code-and-fable-5-ported-the-2003-pc-game-command-conquer-to-native-ios-in-a-few-hours/",[339,343],{"name":340,"url":341,"detail":342},"GitHub: ammaarreshi/Generals-Mac-iOS-iPad","https://github.com/ammaarreshi/Generals-Mac-iOS-iPad/tree/main","開源程式碼（不含遊戲資產）",{"name":344,"url":345,"detail":346},"WCCFTech","https://wccftech.com/claude-fable-5-brought-native-command-conquer-generals-zero-hour-support-to-ipad/","移植過程技術細節報導","#### 23 年前的 C++ 引擎，原生跑在 iPhone 上\n\nGoogle DeepMind 產品設計主管 Ammaar Reshi，以 Claude Code 搭配 Fable 5，在約兩天內完成《Command & Conquer： Generals Zero Hour》的 iOS 移植——使用真實 2003 年 C++ 引擎，直接編譯為 ARM64 原生執行，非模擬器。\n\n移植路徑為 DirectX 8 → DXVK/MoltenVK → Apple Metal API，支援 Campaign、Skirmish 對戰、Generals Challenge，並適配觸控操作（框選拖曳、雙指平移、捏合縮放）。原始碼已於 GitHub 開源，不含遊戲資產。\n\n> **名詞解釋**\n> DXVK/MoltenVK 是圖形 API 轉換層，讓 DirectX 指令得以在 Vulkan(DXVK) 或 Apple Metal(MoltenVK) 上執行。\n\n#### 開發節奏與關鍵轉折\n\n第一個可執行的 build 只花了 40 分鐘。Reshi 表示，初期使用 Claude Opus 4.8，切換至 Fable 5 後才突破關鍵瓶頸，最終 19 小時內完成 19 次 commits，整個過程耗盡了他的 Claude Max 配額。已知限制：長時間使用後可能因記憶體消耗過高而崩潰。","技術關鍵不在 LLM「懂遊戲」，而在它能串聯跨平台圖形 API 轉換鏈。DirectX 8 → DXVK/MoltenVK → Metal 每層都需要理解底層記憶體模型與 API 語義差異。Fable 5 在此案例中更接近有充分上下文的資深工程師：能在 debug 迴圈中保持跨 commit 的技術一致性，而非每次重新理解整個專案結構。","真正的訊號不是「AI 能移植遊戲」，而是個人開發者的槓桿倍率正在質變。以往需要跨平台圖形工程師與 iOS 原生開發者協作才能完成的任務，如今一名有領域知識的個人可在兩天內完成原型。對遊戲公司而言，舊作重製與平台擴展的成本門檻大幅下降；這類案例本身也是 Anthropic 最有力的產品行銷素材。","工程師視角","商業視角",[353,356,359,362,366],{"platform":69,"user":354,"quote":355},"@bcherny（Claude Code 創作者）","Fable 5 現已在 Claude Code 和 Cowork 中推出。Fable 是我用過的最佳程式設計模型，遙遙領先。這是一大躍進——更少的提示與引導、更有效率的 token 使用、更好的程式碼品質、更優秀的工具使用能力、更智慧的自我驗證、更長的工作階段，以及更高的信任度與自主性。",{"platform":65,"user":357,"quote":358},"Bluesky 用戶 (135 upvotes)","2027 年。你打開 Claude Code 並掃描身分證。Anthropic 伺服器接收你的硬體識別碼、虹膜掃描、指紋和完整基因組，判定你有資格提交提示給 Fable。Fable 不這麼認為，所以你收到了 Haiku 的回應。你打開 pi，Kimi 說「嗨！」並修好了你的人生。",{"platform":65,"user":360,"quote":361},"faz.ms(70 upvotes)","好的工匠總是稱讚自己的工具。",{"platform":363,"user":364,"quote":365},"HN","Callicles（HN 用戶）","這完全是用 Claude Code 和 Fable 進行 vibe coding 的成果。感謝你的回饋，也謝謝你花時間查看程式碼，我可以出一版清理後的版本。",{"platform":363,"user":367,"quote":368},"diwank（HN 用戶）","你從哪裡找到那個資訊的？很奇怪，因為他們發布 Fable 5 的公告也提到了 Claude Code：Fable 5 從 7 月 1 日起在 Claude Platform、Claude.ai、Claude Code 和 Cowork 向全球開放，Pro、Max、Team 與部分 Enterprise 方案在 7 月 7 日前享有每週使用量上限 50% 的免費額度，之後轉為使用積分制。","追","Fable 5 讓個人開發者得以在數日內完成過去需要跨平台圖形工程師團隊才能執行的移植工作，大幅降低遺留程式碼現代化的人力門檻。",{"category":372,"source":17,"title":373,"publishDate":6,"tier1Source":374,"supplementSources":376,"coreInfo":381,"engineerView":382,"businessView":383,"viewALabel":384,"viewBLabel":385,"bench":386,"communityQuotes":387,"verdict":77,"impact":403},"discourse","Mistral CEO 警告：用封閉模型等於讓 AI 實驗室坐在你的業務前排",{"name":270,"url":375},"https://the-decoder.com/mistral-ceo-mensch-says-proprietary-ai-models-give-labs-a-front-row-seat-to-your-business-processes/",[377],{"name":378,"url":379,"detail":380},"The Next Web","https://thenextweb.com/news/mistral-ceo-open-source-enterprise-warning","TNW 對 Mensch 指控的補充分析，指出部分說法缺乏直接證據","#### 封閉 AI 的「前排席位」問題\n\nMistral AI 執行長 Arthur Mensch 於 2026 年 7 月 5 日警告：依賴封閉 AI 模型，等同讓供應商取得業務流程的「前排席位」。供應商透過強制資料留存政策持續觀察企業運作，部分甚至以此轉攻自家客戶的市場。\n\n他以 2025 年 Anthropic 切斷 Windsurf API 為例——同期 Anthropic 推出競品 Claude Code——佐證供應商「針對最成功客戶出手」的模式。\n\n#### 四步應對策略\n\nMensch 建議企業採取：\n\n1. 採用開源模型\n2. 建立開放資料系統並自設存取控制\n3. 打造持續訓練飛輪 (continuous training flywheel)\n4. 建立競爭對手難以複製的自有 AI 系統\n\n> **名詞解釋**\n> 持續訓練飛輪：企業以自有資料持續微調模型，形成資料積累與能力強化的正向循環，競爭對手難以複製。\n\nMensch 的論述與 Mistral 商業利益高度吻合——該公司正洽談以 200 億歐元估值融資，「開源優先」主張本身也是市場定位策略。","Windsurf 事件讓 API 依賴的風險具體化。評估 AI 工作流時，建議考量三點：\n\n1. 核心業務邏輯優先選 Apache 2.0 / MIT 授權的開源模型\n2. 透過 vLLM 或 Ollama 建立本地推理層，解耦 API 依賴\n3. 採用閉源 API 前確認資料留存與競業條款\n\n自有訓練飛輪建置成本高，但對核心 AI 應用具長期防禦價值。","Mensch 的論述即使存在商業動機，核心邏輯仍值得正視：**與供應商共享業務資料，等同潛在競爭情報外流**。\n\n短期而言，閉源模型的易用性優勢仍在；長期來看，涉及核心業務流程的 AI 化應優先考慮開源或自托管方案，以保護競爭護城河。關鍵決策是區分「可外包」與「不可外包」的 AI 使用情境。","實務觀點","產業結構影響","#### 開源微調效能（來源聲稱，未獨立驗證）\n\n- 私有金融資料微調 Qwen3-235B 準確率：84.7%\n- 前沿閉源模型準確率：78.2%\n- 運營成本：開源方案低 14 倍",[388,391,394,397,400],{"platform":65,"user":389,"quote":390},"arteesetica.bsky.social（Bluesky 140 讚）","注意 Ecosia 的用戶 ⚠️ Ecosia 聲稱深愛樹木和再生能源，但細看條款——依舊找那些老面孔。他們的主要 AI 供應商是 Mistral（同樣因版權問題飽受質疑），備援則是 OpenAI 🤦",{"platform":363,"user":392,"quote":393},"InsideOutSanta（HN 用戶）","如果 Mistral 的目標不是無限擴張，而是以合理定價提供利基市場模型，他們現在就該放棄嗎？我不確定你想表達什麼。",{"platform":65,"user":395,"quote":396},"TechCrunch（Bluesky 40 讚）","Mistral AI 提供部分開源 AI 模型，自 2023 年成立以來已完成大規模融資，目標是「讓前沿 AI 人人皆可使用」。",{"platform":65,"user":398,"quote":399},"Evangelos Kazakos（Bluesky 6 讚）","抱歉打破歐洲的幻想，但 Mistral 確實在開發軍事 AI——他們與法國軍方及 Helsing 簽有合約。而且，「跟隨 Palantir 玩法」這件事，對我來說聽起來並不像是什麼好事。",{"platform":363,"user":401,"quote":402},"kaashif（HN 用戶）","這確實意味著牛奶產業是低附加價值的大宗商品業務，供應商只靠價格競爭，僅在保護主義和補貼下苟活。食品對國家安全重要，但它是成本中心，永遠不會帶動成長。如果這就是 Mistral 的目標，現在就該放棄了。","封閉 AI 的資料主權風險正推動企業重新評估 AI 採購策略，開源與自托管模型需求將持續上升。",{"category":127,"source":14,"title":405,"publishDate":6,"tier1Source":406,"supplementSources":409,"coreInfo":413,"engineerView":414,"businessView":415,"viewALabel":416,"viewBLabel":417,"bench":156,"communityQuotes":418,"verdict":369,"impact":428},"Taste-Skill 爆紅 GitHub：教你的 AI 什麼是好品味，拒絕千篇一律的 AI Slop",{"name":407,"url":408},"Leonxlnx/taste-skill — GitHub","https://github.com/Leonxlnx/taste-skill",[410],{"name":411,"url":412},"SKILL.md 原始內容","https://github.com/Leonxlnx/taste-skill/blob/main/skills/taste-skill/SKILL.md","#### 什麼是 Taste-Skill？\n\nTaste-Skill 是一個放在專案根目錄的 **SKILL.md 設計規則檔案**，透過 `npx skills add Leonxlnx/taste-skill` 一行指令安裝，讓 AI coding agent 在生成前端 UI 前先「讀懂需求」，而非直接套用預設美學——即所謂 AI Slop。\n\n> **名詞解釋**\n> AI Slop：指 AI 大量生成的千篇一律介面，例如 AI 紫色漸層、深色網格背景、glassmorphism，缺乏真實設計判斷力。\n\n#### 三旋鈕 + Brief Inference\n\n核心機制是「三旋鈕系統」：`DESIGN_VARIANCE`（設計多樣性，1–10）、`MOTION_INTENSITY`（動效強度）、`VISUAL_DENSITY`（視覺密度），供 AI agent 依 brief 自動調校。\n\n執行前強制 **Brief Inference（讀懂場景）**：AI 必須先輸出一行 Design Read 確認理解需求，再開始生成。2026 年 3 月重構後拆為三個模組：`taste-skill`、`redesign-skill`（含 100+ 審查維度）、`output-skill`（強制輸出完整產物）。","相容 Cursor、Claude Code、Codex、Gemini CLI、v0、Lovable、OpenCode，安裝只需一行指令。\n\n核心在於 Anti-Default 禁令：禁止 AI 自動套用 AI 紫色漸層、三欄等寬 feature card、Inter + slate-900 等「LLM design tell」。\n\nv2 加入 Design System Map，指引 AI 根據 brief 選用 Fluent UI、Material 3、Carbon、shadcn/ui 等官方設計系統，並誠實標注「近似實作」vs「官方元件」。","57,000+ GitHub 星數代表一個新需求正在成熟：開發者不再滿足於 AI 能「跑通 UI」，而是要求它能「讀懂場景」再動筆。\n\nTaste-Skill 的爆紅暗示前端 AI 工具正從「能用」邁向「有品味」的分水嶺——誰能最先整合類似的 design taste 框架，誰就能在 AI coding 工具市場取得差異化。","開發者整合視角","生態影響",[419,422,425],{"platform":69,"user":420,"quote":421},"@hasantoxr（AI 工具內容創作者）","這種東西不應該免費。Taste Skill 是一個開源 repo，讓 Claude Code、Cursor 和 Windsurf 具備設計品味，不再生成千篇一律、無聊的 AI 風介面。一行指令安裝，agent 讀取技能檔案，之後生成的每個 UI 都有高階質感。",{"platform":69,"user":423,"quote":424},"@mxcl（Homebrew 創作者 Max Howell）","這套「品味」技能集用起來效果非常好，尤其是 frontend-taste-skill，似乎在 native 和 web UI 上都通用。",{"platform":65,"user":426,"quote":427},"dailygithubtrends.bsky.social(1 like)","今日 GitHub 趨勢：Leonxlnx/taste-skill，提升 AI agent 生成前端品質的框架，避免套版設計、強化排版與動效品質。另含參考板圖片生成技能，支援透過 Cursor、Claude Code 等工具落實 AI 生成構建方案的工作流程。","MIT 開源、一行指令安裝、相容主流 AI coding 工具，前端開發者可立即導入，有效提升 AI 生成 UI 的設計品質。",{"category":372,"source":15,"title":430,"publishDate":6,"tier1Source":431,"supplementSources":433,"coreInfo":446,"engineerView":447,"businessView":448,"viewALabel":384,"viewBLabel":385,"bench":156,"communityQuotes":449,"verdict":465,"impact":466},"好萊塢想禁 Seedance 卻又捨不得用——AI 影片工具的矛盾現狀",{"name":270,"url":432},"https://the-decoder.com/hollywood-wants-seedance-banned-and-reportedly-also-wants-to-keep-using-it/",[434,438,442],{"name":435,"url":436,"detail":437},"Hollywood Reporter","https://www.hollywoodreporter.com/business/business-news/mpa-cease-and-desist-bytedance-seedance-2-0-1236510957/","MPA 停止侵權信詳情",{"name":439,"url":440,"detail":441},"Variety（SAG-AFTRA 聲明）","https://variety.com/2026/film/news/sag-aftra-seedance-ai-infringement-tom-cruise-brad-pitt-fight-1236662695/","演員工會回應",{"name":443,"url":444,"detail":445},"TechTimes(Seedance 2.5)","https://www.techtimes.com/articles/318975/20260624/bytedance-seedance-25-native-30-second-ai-video-no-stitching-required.htm","Seedance 2.5 升級細節","#### 法律戰火全面引爆\n\n2026 年 2 月，ByteDance 推出 **Seedance 2.0**，一支 AI 生成的「Brad Pitt 與 Tom Cruise 屋頂對打」15 秒短片瘋傳，引爆好萊塢全面反撲。**美國電影協會 (MPA) 史上首次對 AI 公司發出停止侵權信**，Disney、Warner Bros.、Paramount、Sony Pictures、Netflix 相繼發出法律威脅。\n\n> **名詞解釋**\n> MPA（美國電影協會）：代表好萊塢主要製片公司的版權保護組織，此次為首次向 AI 公司發出停止侵權令。\n\nDisney 形容此舉是對 IP 的「虛擬砸搶」，指控平台預載 Marvel、Star Wars 等盜版訓練資料；SAG-AFTRA 演員工會亦稱相關影片為「明目張膽的侵權行為」。\n\n#### 「不問不說」的矛盾內幕\n\n法律對峙之外，好萊塢內部卻截然相反。影視顧問 Peter Csathy 直言，懂 AI 的創作者普遍認為 Seedance 是目前最佳影片生成工具；《辛普森家庭》製片人 Joel Kuwahara 透露，許多製片公司雖未正式核准，卻以「不問不說」的默許態度悄悄在內部使用。\n\n2026 年 6 月，**Seedance 2.5** 升級為原生 30 秒影片生成，相較 Runway、Sora、Veo 優勢持續擴大。","Seedance 2.5 在原生 30 秒生成與畫質上領先同類工具，但訓練資料合規性存疑。若 MPA 訴訟成立，平台可能遭強制下架或 API 存取受限。商業專案使用前，需評估 IP 侵權連帶責任，建議優先採用授權來源明確的替代工具。","Seedance 的矛盾處境揭示 AI 影片工具的產業核心衝突：工具威力愈強，IP 糾紛風險愈高，但成本降幅也愈誘人。好萊塢「檯面禁、私下用」的模式若持續，將加速 AI 影片工具實質滲透，也讓授權談判籌碼往平台傾斜。ByteDance 以技術優勢換取市場接受度的策略，正在奏效。",[450,453,456,459,462],{"platform":65,"user":451,"quote":452},"Dare Obasanjo（Bluesky 28 讚）","AI 社群愛說 ByteDance 的 Seedance 2.0 一出，好萊塢就完了。現實是：好萊塢是個生意，各大片廠正在悄悄擁抱這項技術——因為它承諾大幅降低特效與電影製作成本。",{"platform":69,"user":454,"quote":455},"@deedydas（前 Cohere/Google 工程師）","ByteDance 將在 7 月初推出全球最強影片生成模型 Seedance 2.5！這鞏固了中國在影片領域的絕對主導地位——生成長度是前代版本的兩倍，並原生支援 4K 解析度。",{"platform":65,"user":457,"quote":458},"Eris（Bluesky 4 讚）","幹，影片什麼時候變得這麼好了。這是 Seedance 嗎？",{"platform":65,"user":460,"quote":461},"AI News（Bluesky 1 讚）","好萊塢想禁 Seedance，但據報導也想繼續使用它。ByteDance 的 AI 影片工具正讓好萊塢左右為難。一支 AI 生成的 Brad Pitt 與 Tom Cruise 影片爆紅後，美國電影協會向 AI 公司發出史上第一封停止侵權信。",{"platform":69,"user":463,"quote":464},"@MrDavids1（X 用戶）","我稱讚 Seedance Pro 已經有一段時間了，因為它真的是個驚人的影片模型。能從單一提示生成多個鏡頭，實在太實用了。","觀望","Seedance 引爆 AI 影片工具的 IP 授權根本衝突，好萊塢「禁而不止」的矛盾將重塑內容製作的合規標準與授權生態。",{"category":194,"source":9,"title":468,"publishDate":6,"tier1Source":469,"supplementSources":472,"coreInfo":480,"engineerView":481,"businessView":482,"viewALabel":350,"viewBLabel":351,"bench":483,"communityQuotes":484,"verdict":77,"impact":494},"AI 搜尋 Agent 的真正弱點不在搜尋，而在不會問對問題",{"name":470,"url":471},"DiscoBench： When Search Agents Should Ask(arXiv 2606.27669)","https://arxiv.org/abs/2606.27669",[473,476],{"name":270,"url":474,"detail":475},"https://the-decoder.com/ai-search-agents-dont-fail-at-searching-they-fail-at-asking-the-right-questions-when-queries-get-ambiguous/","DiscoBench 研究摘要報導 (2026-07-05)",{"name":477,"url":478,"detail":479},"InteractComp： Evaluating Search Agents With Ambiguous Queries(arXiv 2510.24668)","https://arxiv.org/abs/2510.24668","互補研究，追蹤 15 個月搜尋能力 vs 互動能力演進","#### DiscoBench：量化澄清提問能力的基準測試\n\n騰訊混元與清華大學聯合發布 DiscoBench，評估搜尋 Agent 面對模糊查詢時主動提問澄清的能力。\n\n> **名詞解釋**\n> DiscoBench 定義四類模糊類型——描述對應多個實體、多個時間段或版本、不同排名標準、事實性錯誤——並以任務效用、模糊偵測、互動策略、成本效率四個維度評估 Agent 表現。\n\n測試集涵蓋 211 項任務、463 個模糊點，橫跨 11 個知識領域。在未提供消歧提示的情況下，頂尖模型準確率均低於 50%：\n\n- Doubao Seed 2.0 Pro：43.1%\n- Gemini 3.1 Pro：40.8%\n- Claude Opus 4.7：39.8%\n- Qwen3.6 Max：12.3%\n\n#### 越搜尋越迷失，「先問後搜」才是正解\n\n研究揭露一個反直覺發現：不斷追加搜尋卻不提問，成功率 (51.9%) 甚至低於直接猜測 (56.5%) 。「SearchThenAsk（先搜尋後提問）」策略成功率高達 93.4%，差距懸殊。\n\n關鍵缺陷在於：模型能偵測到查詢有歧義，卻不懂得主動與使用者互動澄清。互補研究 InteractComp 追蹤 15 個月資料，發現搜尋能力提升七倍，互動能力卻幾乎停滯——根本原因是「系統性過度自信」。","搜尋 Agent 若缺乏澄清互動迴圈，等同於在半盲狀態下執行任務。建議在查詢入口加入模糊分類前處理，偵測有歧義的查詢後觸發多輪對話，而非立即啟動搜尋。DiscoBench 的四維評估框架可直接作為 evals 基準，量化澄清能力並追蹤改善進度。","所有主要模型在模糊查詢下準確率低於 50%，意味著超過一半的模糊指令無法獲得正確結果。採購搜尋 Agent 產品時，建議加入「澄清能力測試」：故意輸入有歧義的查詢，觀察系統是否主動詢問。能問對問題的 Agent 才是真正可信賴的企業助理。","#### 效能基準（DiscoBench，無消歧提示）\n\n- Doubao Seed 2.0 Pro：43.1%\n- Gemini 3.1 Pro：40.8%\n- Claude Opus 4.7：39.8%\n- MiniMax M2.7：16.1%\n- Qwen3.6 Max：12.3%\n\n#### 策略成功率對比\n\n- SearchThenAsk（先搜尋後提問）：93.4%\n- DirectGuess（直接猜測）：56.5%\n- SearchHeavyGuess（重搜尋後猜測）：51.9%\n\n#### InteractComp 最佳模型\n\n- 無消歧上下文：13.73%\n- 完整消歧上下文：71.50%",[485,488,491],{"platform":65,"user":486,"quote":487},"aidailypost.com(3 likes)","新研究顯示 AI 搜尋 Agent 在模糊查詢上表現欠佳——DiscoBench 揭示 Gemini 3.1 Pro 與 Claude Opus 4.7 仍需提升澄清能力。人類能幫助這些機器人應對不確定性嗎？",{"platform":69,"user":489,"quote":490},"@ZDNET（科技媒體）","AI Agent 正在獲得專屬搜尋引擎。",{"platform":69,"user":492,"quote":493},"@alexgroberman（X 用戶）","Google 揭示了 2026 年 AI 搜尋流量的運作方式。","搜尋 Agent 澄清提問能力普遍不足且停滯不前，企業部署前應主動評估此能力，並持續關注業界改進進展。",{"category":194,"source":16,"title":496,"publishDate":6,"tier1Source":497,"supplementSources":500,"coreInfo":507,"engineerView":508,"businessView":509,"viewALabel":350,"viewBLabel":351,"bench":510,"communityQuotes":511,"verdict":465,"impact":521},"LeCun 團隊讓世界模型學會持續學習，邁向不斷進化的環境理解",{"name":498,"url":499},"arXiv: AdaJEPA: An Adaptive Latent World Model","https://arxiv.org/abs/2606.32026",[501,504],{"name":502,"url":503},"量子位：LeCun 團隊讓世界模型學會持續學習","https://www.qbitai.com/2026/07/442964.html",{"name":505,"url":506},"CryptoBriefing: Yann LeCun's paper on LeJEPA world model conditions","https://cryptobriefing.com/lejepa-world-model-conditions-lecun/","#### 世界模型首次學會「邊跑邊學」\n\n2026 年 6 月，紐約大學與 LeCun 新創公司 AMI Labs 聯合發表 AdaJEPA，首次讓世界模型在部署期間持續自適應更新，打破了傳統「訓練後凍結參數」的模式。\n\n> **名詞解釋**\n> 世界模型 (World Model) ：AI 對環境物理規律的內部表示，讓智慧體能預測行動後果並進行規劃，而非僅對感測器輸入做出反應。\n\n#### 如何做到輕量化線上學習\n\n系統基於閉環 MPC(Model Predictive Control) 框架，執行「規劃→執行→觀測→梯度更新→再規劃」的完整循環。每次重規劃僅需 1 步梯度下降，額外延遲僅 0.01–0.03 秒。\n\n為防止表示空間崩塌，系統僅更新視覺編碼器與預測器的最後幾層，以真實環境的狀態轉移作為自監督學習信號，無需額外專家示範。","AdaJEPA 的在線校準策略值得深入研究：僅更新最後幾層的保守做法，有效避免了災難性遺忘與表示崩塌，可應用於機器人、自駕等部署後環境持續變化的場景。\n\n單步梯度下降的設計與 LoRA 類似——最小化更新範圍以換取穩定性——實際整合前需評估重規劃頻率對能耗的累積影響。","世界模型能在線學習，代表 AI 系統的維護模式即將轉變：不再需要定期重訓並重新部署，而是讓模型在真實互動中自我校準。\n\n對機器人、自駕與工業自動化廠商而言，這降低了落地後的迭代成本。LeCun 的 JEPA 路線持續獲得研究支撐，AMI Labs 的商業化前景值得追蹤。","#### 效能基準\n\n- PointMaze 未知布局測試：梯度下降成功率 53.3% → 78.7%\n- 未見物體目標導航測試：規劃成功率提升接近一倍",[512,515,518],{"platform":69,"user":513,"quote":514},"@LiorOnAI（AI 內容評論員）","剛讀完 LeCun 最新論文。他的團隊訓練出第一個不會崩塌的世界模型，讓我來解釋為什麼這很重要。它叫做 LeWorldModel。世界模型預測接下來會發生的物理現象——物體移動、墜落、碰撞——這是機器人技術的基礎層。",{"platform":59,"user":516,"quote":517},"gmueckl（HN 用戶）","人類並非完美無缺，但在推理能力上遠優於大型語言模型。LLM 很容易被誤導，因為它們無法建立真正可操作、具預測能力的模型——這正是 Yann LeCun 長期倡導具預測能力的世界模型（針對物理世界）的核心論點。",{"platform":69,"user":519,"quote":520},"@askalphaxiv（alphaXiv 平台）","隨著 Yann LeCun 及其團隊在 2026 年引領新一波世界模型研究浪潮，我們整理了一份必讀論文清單，帶你快速掌握 JEPA 與世界建模的最新進展！","輕量級在線梯度校準技術有望讓世界模型擺脫部署後凍結限制，加速具身智慧與自主系統在動態真實環境中的適應能力。",{"category":127,"source":16,"title":523,"publishDate":6,"tier1Source":524,"supplementSources":527,"coreInfo":536,"engineerView":537,"businessView":538,"viewALabel":539,"viewBLabel":417,"bench":540,"communityQuotes":541,"verdict":465,"impact":557},"Meta 也來賣鏟子：Zuckerberg 推 Meta Compute，GPU 算力直接對外賣",{"name":525,"url":526},"Windows News","https://windowsnews.ai/article/meta-reportedly-readying-meta-compute-cloud-service-to-sell-ai-gpu-access-and-hosted-models.433294",[528,532],{"name":529,"url":530,"detail":531},"量子位","https://www.qbitai.com/2026/07/443339.html","Meta Compute 業務詳情報導",{"name":533,"url":534,"detail":535},"每日經濟新聞","https://www.nbd.com.cn/articles/2026-07-02/4445510.html","市場反應與股價影響","#### Meta Compute 三層服務\n\nMeta 正籌備名為 **Meta Compute** 的雲端服務，計劃對外販售 GPU 算力與模型 API，直接挑戰 AWS、Azure、Google Cloud 及 CoreWeave。消息曝光後，Meta 股價單日飆升 8.8%，CoreWeave 暴跌 13.92%，Nebius 跌 17.01%。\n\n服務規劃分三層：\n\n1. **裸機 GPU 實例**：H100 專屬物理伺服器\n2. **受管訓練集群**：自動化分散式訓練編排\n3. **托管模型 API**：Llama 系列與第三方模型（含 Anthropic Claude）REST 端點\n\n#### 定價策略與背景\n\nMeta 計劃定價約 $1.50/GPU 小時，比市場主流 $2.50–3.00 低 30–40%，並採用 90 天可取消彈性合約，降低客戶遷移門檻。\n\n背景矛盾之處在於：市場算力價格正在上漲——H100 合約價五個月漲 38.2%，Nvidia B200 租賃價預計 10 月漲幅達 94%。Meta 能否長期維持低價，取決於其 5GW 數據中心規模帶來的成本優勢。","對開發者而言，Meta Compute 帶來第四個主流雲算力選項。$1.50/GPU 小時若成真，相當於 H100 訓練成本直接打六折。90 天可取消合約降低供應商鎖定風險，對需要彈性擴縮的訓練任務尤其友好。\n\n若 Claude 確實納入托管範圍，代表單一平台可同時呼叫開源與閉源模型，省去多平台管理成本——但服務何時上線仍未公布，目前僅能觀望。","Meta 入場最直接的受害者是獨立算力雲業態——CoreWeave 和 Nebius 單日跌幅已說明市場判斷。Wells Fargo 估算此業務 2028 年可帶來年化 2,640 億美元收入，若成真將顛覆算力市場格局。\n\n關鍵風險在於動機：Zuckerberg 自承 AI agent 開發不如預期，賣算力更像是龐大資本支出的應急出路。一旦 Meta 內部 AI 需求回升，外部算力供應的優先度將直接受壓。","開發者視角（API／整合）","#### 定價對比\n\n- Meta Compute（預計）：$1.50/GPU 小時 (H100)\n- 市場主流：$2.50–3.00/GPU 小時\n- Nvidia B200 租賃（2026 年 10 月預測）：$5.10／小時 (+94%)\n- H100 合約價五個月漲幅：+38.2%\n- AWS EC2 7 月調漲：約 +20%",[542,545,548,551,554],{"platform":69,"user":543,"quote":544},"@negligible_cap","Meta Compute 是 Zuck 對外販售多餘算力的新事業單位名稱。其中一個潛在方案是銷售托管在 Meta 現有 AI 基礎設施上的各種 AI 模型存取權限，類似 AWS 的 Bedrock 模式。",{"platform":59,"user":546,"quote":547},"laweijfmvo","我認為 Meta 龐大的算力投資，從來都不是為了讓 10 萬名工程師跑程式碼模型，而是為了讓 35 億用戶在每個產品中使用 AI——Meta AI、眼鏡等。所以我認為真正沒被充分利用的，是用戶端這個部分……",{"platform":69,"user":549,"quote":550},"@SemiAnalysis_（半導體與 AI 基礎設施研究機構 SemiAnalysis）","Meta Compute：人人都想成為雲端供應商。Zuck 啟動 B 計劃？SpaceX 2.0、Bedrock 2.0，MSL 不打算放棄，推薦系統擴展 10 倍……",{"platform":59,"user":552,"quote":553},"htrp","其中一個方案包括銷售托管在 Meta 現有 AI 基礎設施上的 AI 模型存取權限，類似 AWS 的 Bedrock 模式，Meta 負責運營支撐模型的數據中心和晶片，並向開發者收費。此外，公司也考慮銷售「裸機」算力，類似 CoreWeave 等新興雲廠商的模式。",{"platform":59,"user":555,"quote":556},"Barrin92","如果 Meta 在賣算力，Twitter 也在賣算力，而這些算力又沒有實際用途，不需要經濟學學位就能判斷算力價格將走向何方。這些公司最終都會坐擁龐大數據中心，等熱潮退去後不知道拿它們怎麼辦。","Meta 以 5GW 數據中心規模強勢進入算力市場，若定價與服務落地，將重塑雲 GPU 成本基準並直接衝擊 CoreWeave 等獨立算力雲廠商。",{"category":372,"source":9,"title":559,"publishDate":6,"tier1Source":560,"supplementSources":563,"coreInfo":567,"engineerView":568,"businessView":569,"viewALabel":384,"viewBLabel":385,"bench":570,"communityQuotes":571,"verdict":465,"impact":584},"Dartmouth 新研究：AI 家教達到 0.71-1.30 標準差效果量，接近人類一對一教學",{"name":561,"url":562},"Balancing Efficacy and Engagement in Interactive Texts","https://intextbooks.science.uu.nl/workshop2026/files/itb26_s1s2.pdf",[564],{"name":565,"url":566},"HN 討論：New AI tutor achieves 0.71-1.30 SD effect size in Dartmouth course","https://news.ycombinator.com/item?id=48796817","#### 研究結果：接近人類一對一家教效益\n\nDartmouth College 於 2026 年春季學期，在入門統計課程 MATH 010 部署 Phosphor AI 輔導平台，共 143 名學生完成實驗。完整使用平台的學生，期末考成績提升幅度介於 **0.71 SD（控制期中成績後）至 1.30 SD（未調整）**，接近 Bloom 1984 年「2 Sigma 問題」中人類一對一家教的效益水準。\n\n> **名詞解釋**\n> 「2 Sigma 問題」源自 Bloom 1984 年研究：一對一人類家教可讓平均生進步約 2 個標準差，超越 98% 傳統班級教學學生，但規模化成本極高，故被視為教育科技的終極目標。\n\n#### 方法論爭議不可忽視\n\nPhosphor 的核心功能是練習題測驗加 AI 自動批改，附帶 RAG 聊天助理但實際使用率極低。90% 的學生至少使用過一次，但達到「完全投入」門檻的僅約 16 人（約 11%）。\n\n研究未設置隨機對照組，且研究者在分析期間將題型從多選題改為 AI 批改主觀題——社群批評此舉造成「實驗設計被自身結果反向決定」的因果倒置問題。完整投入的 16 名學生也可能本來就是高動機學習者，選擇效應難以排除。","Phosphor 靠「練習題 + 即時 AI 批改」驅動學習，RAG 對話功能幾乎無人使用——對 AI 教育工具開發者而言，**自動批改反饋迴路的實效遠高於對話 UI**。11% 的完整投入率顯示產品留存設計是主要瓶頸；如何將 90% 的初次使用者轉化為持續投入的學習者，才是真正的工程挑戰。","研究方法論雖有缺陷，但 90% 自願使用率（對比教科書 10-15% 閱讀率）是更有力的商業論據：觸及率是傳統教材的六倍。教育科技投資人目前面臨兩難——誇大的效果量數字容易引發學術質疑，反而拖累公信力；要真正釋放估值，需要穩健的 RCT 設計背書。","#### 效能基準\n\n- 完整投入組期末考提升：0.71 SD（控制期中成績）至 1.30 SD（未調整）\n- 換算實際分數：24 分滿分中多約 3 分（0.71 SD 口徑）\n- 完整投入率：約 11%（16 / 143 名學生）\n- 平台使用率：90% 至少使用一次",[572,575,578,581],{"platform":59,"user":573,"quote":574},"dash2(HN)","把這叫做「效果量」本身就是胡說，教育軟體能用如此糟糕的統計實踐矇混過關，令人擔憂。希望這只是學生專案。",{"platform":59,"user":576,"quote":577},"usernametaken29(HN)","認真備料、精心設計教材、並將內容與期末考對齊，成績當然會進步。大多數大學統計課本來就糟透了，任何投入都會有改善。這究竟有多少是 AI 的功勞？",{"platform":59,"user":579,"quote":580},"Herring(HN)","也許更好的做法是直接參考目前表現最佳的國家——芬蘭、愛沙尼亞、日本、新加坡等。學校經費、教師專業化等因素的影響，可能遠比 AI 家教大。",{"platform":59,"user":582,"quote":583},"skybrian(HN)","選擇效應在教育中極為重要。達特茅斯學生已歷大規模篩選，若要更廣泛推廣，效果未必能複製。動機也是核心問題——AI 家教的新鮮感能讓更多人嘗試，但這種熱度會消退嗎？","AI 自動批改在精英大學課程中展現顯著效果量，但無 RCT 且樣本極小，需更嚴謹的大規模實驗才能確認可複製性。","#### 社群熱議排行\n\n本日熱度最高為 EU Chat Control 快速通道表決。Patrick Breyer（Bluesky， 29 upvotes）緊急呼籲阻止大規模監控，@paddi_hansen 宣告歐洲議會接連否決延長豁免條款，隱私倡議者連取勝利。\n\n百度 Unlimited OCR 緊接其後，HN + X 多則討論關注其「記憶體零增長、讀完整本書」的突破。Fable 5 iOS 移植 C&C 遊戲在 Bluesky 獲 135 upvotes，引發下一代 AI 編碼能力的熱烈討論。\n\nAmazon MTurk 退場公告（socialmedialab.ca， 6 upvotes， Bluesky）引爆一代 AI 訓練基礎設施的告別情緒。\n\n#### 技術爭議與分歧\n\nChat Control 最大爭議在術語認知：SpicyLemonZest(HN) 指出「『Chat Control』這個詞是反對私訊掃描的隱私倡議團體所創造，民調普遍支持 CSAM 掃描」。\n\n@mullvadnet(X) 則主張理事會文本仍保留後門空間，兩方對「勝利」的定義根本分歧。GPT-5.5 陣營，lmwnshn(HN) 推廣 MIT 授權 clankerbend：「有狀況可以隨時換船」，呼籲不被閉源版本決策綁死。\n\n#### 實戰經驗（最高價值）\n\nzuzululu(HN) 實測 GPT-5.5 Codex reasoning token 修復：「我剛試了，成功了！不知道為什麼有效，但感覺和一個半月前一樣了。」\n\nHN 用戶 potsandpans 實測百度 Unlimited OCR：「可以在 12GB VRAM 上舒適運行，確實很有競爭力。」@mxcl（Homebrew 創作者， X）確認 Taste-Skill：「用起來非常好，在 native 和 web UI 上都通用。」\n\n#### 未解問題與社群預期\n\n7 月 8 日 Chat Control 快速通道表決結果未知；義大利「警告反對卻投票支持」的矛盾顯示政治利益計算超越公開立場，社群對連勝能否延續高度存疑。\n\nGPT-5.5 reasoning token 透明度問題仍無官方回應，516-token clustering 究竟是設計還是 bug，開發者無從驗證。MTurk 退場後，社群追問 2022 年後的眾包標注資料集是否已遭 LLM 代工污染，訓練出的模型品質如何驗證？",[587,589,591,593,595,597,599,601,603,605],{"type":80,"text":588},"現在就切換至 Signal 等預設端對端加密應用，親身理解加密架構，評估 Chat Control 若上路將對你的通訊流程造成何種衝擊",{"type":80,"text":590},"在現有 GPT-5.5 API 呼叫中加入 reasoning token 日誌，統計 516-token 命中率，快速判斷自身工作負載是否受效能退化影響",{"type":80,"text":592},"從 Hugging Face 下載 baidu/Unlimited-OCR，在 12GB VRAM 本地 GPU 上測試自有多頁 PDF，對比現有 OCR 管道準確率",{"type":80,"text":594},"一行指令安裝 Taste-Skill 並整合至 Claude Code 或 Cursor，測試 AI 生成 UI 的設計品質提升效果",{"type":83,"text":596},"若服務在歐盟有用戶，執行 DPIA（資料保護衝擊評估），識別現有通訊模組對 Chat Control 合規的暴露程度",{"type":83,"text":598},"為 2022 年後收集的 MTurk 資料集建立品質重新驗證流程，針對 LLM 代工污染進行抽樣稽核，再決定是否重新標注",{"type":83,"text":600},"為推理任務搭建開放 harness 層（參考 MIT 授權的 clankerbend），確保後端模型可替換，不被單一供應商版本決策綁死",{"type":86,"text":602},"追蹤 7 月 8 日 EU Chat Control 快速通道表決結果，以及 Patrick Breyer MEP 部落格與 EFF 的後續聲明",{"type":86,"text":604},"關注 openai/codex issue #30364，看 OpenAI 是否承諾提高 reasoning token 分佈的透明度",{"type":86,"text":606},"追蹤合成資料工具（如 Argilla、Gretel.ai）與 RLHF 標注平台的功能演進，判斷 MTurk 之後下一代資料基礎設施的格局走向","今日社群橫跨三個維度見證了 AI 生態的世代交替：歐盟加密政策進入最後衝刺、Amazon MTurk 退場讓資料基礎設施面臨真空、GPT-5.5 reasoning token 爭議提醒開發者，閉源模型的「升級」未必是進步。\n\n與此同時，開源工具陣營正用行動回應：百度 Unlimited OCR 打破長文件記憶體天花板、Taste-Skill 試圖從根本解決 AI 生成設計的同質化、Meta 也悄悄準備加入算力市場競爭。\n\n社群共識漸趨清晰——真正的護城河在於可替換的開放架構，而非依賴任何單一供應商的黑盒版本管理。",{"prev":609,"next":610},"2026-07-05","2026-07-07",{"data":612,"body":613,"excerpt":-1,"toc":623},{"title":156,"description":42},{"type":614,"children":615},"root",[616],{"type":617,"tag":618,"props":619,"children":620},"element","p",{},[621],{"type":622,"value":42},"text",{"title":156,"searchDepth":624,"depth":624,"links":625},2,[],{"data":627,"body":628,"excerpt":-1,"toc":634},{"title":156,"description":46},{"type":614,"children":629},[630],{"type":617,"tag":618,"props":631,"children":632},{},[633],{"type":622,"value":46},{"title":156,"searchDepth":624,"depth":624,"links":635},[],{"data":637,"body":638,"excerpt":-1,"toc":644},{"title":156,"description":49},{"type":614,"children":639},[640],{"type":617,"tag":618,"props":641,"children":642},{},[643],{"type":622,"value":49},{"title":156,"searchDepth":624,"depth":624,"links":645},[],{"data":647,"body":648,"excerpt":-1,"toc":654},{"title":156,"description":52},{"type":614,"children":649},[650],{"type":617,"tag":618,"props":651,"children":652},{},[653],{"type":622,"value":52},{"title":156,"searchDepth":624,"depth":624,"links":655},[],{"data":657,"body":658,"excerpt":-1,"toc":764},{"title":156,"description":156},{"type":614,"children":659},[660,667,672,685,703,708,713,718,723,728,734,739,744,749,754,759],{"type":617,"tag":661,"props":662,"children":664},"h4",{"id":663},"chat-control-是什麼歐盟的-ai-掃描立法草案",[665],{"type":622,"value":666},"Chat Control 是什麼——歐盟的 AI 掃描立法草案",{"type":617,"tag":618,"props":668,"children":669},{},[670],{"type":622,"value":671},"Chat Control 是歐盟關於私人訊息掃描的一系列立法提案的統稱，這個詞彙由反對掃描的隱私倡議者所創造並推廣，支持掃描的政府方從不使用這個術語，命名權之爭本身就是輿論戰的一部分。",{"type":617,"tag":618,"props":673,"children":674},{},[675,677,683],{"type":622,"value":676},"Chat Control 1.0 源自 2021 年通過的《ePrivacy 指令》臨時豁免條款，允許 Gmail、Facebook Messenger 等平台",{"type":617,"tag":678,"props":679,"children":680},"strong",{},[681],{"type":622,"value":682},"自願掃描",{"type":622,"value":684},"未加密訊息，以偵測兒童性剝削素材 (CSAM) 。豁免期原訂至 2026 年 4 月 3 日，掃描涵蓋三種技術：哈希比對、AI 模式辨識（grooming 偵測）以及對未加密訊息的主動掃描。",{"type":617,"tag":686,"props":687,"children":688},"blockquote",{},[689],{"type":617,"tag":618,"props":690,"children":691},{},[692,697,701],{"type":617,"tag":678,"props":693,"children":694},{},[695],{"type":622,"value":696},"名詞解釋",{"type":617,"tag":698,"props":699,"children":700},"br",{},[],{"type":622,"value":702},"\n哈希比對 (hash matching) ：為每個已知非法圖片生成唯一的「數位指紋」（哈希值），系統掃描上傳內容時與指紋庫比對，命中即標記，無需人工判讀原始圖片內容。",{"type":617,"tag":618,"props":704,"children":705},{},[706],{"type":622,"value":707},"Chat Control 2.0 是更進一步的永久性強制掃描方案，覆蓋端對端加密 (E2E) 訊息服務，必須在訊息加密前的用戶裝置上執行掃描，即「客戶端掃描 (client-side scanning) 」。超過 500 位科學家與密碼學家聯署公開聲明，指出此方案「技術上不可行」——這等同在每台裝置上安裝可遠端觸發的後門程式。",{"type":617,"tag":661,"props":709,"children":711},{"id":710},"快速通道的政治操作與民主程序爭議",[712],{"type":622,"value":710},{"type":617,"tag":618,"props":714,"children":715},{},[716],{"type":622,"value":717},"2026 年 3 月 26 日，歐洲議會以 307 票對 306 票（24 票棄權）的極微差距否決延長 Chat Control 1.0 臨時規定，隱私倡議者一度宣告勝利。",{"type":617,"tag":618,"props":719,"children":720},{},[721],{"type":622,"value":722},"然而，歐盟理事會（由各成員國政府代表組成，而非直選機構）在 2026 年 7 月 2 日透過書面程序通過新立場，強行推動重啟授權。更具爭議的是：理事會計畫在 7 月 8 日緊急將草案列入議程，刻意選在暑假前大批議員已離場的時間點推動表決。",{"type":617,"tag":618,"props":724,"children":725},{},[726],{"type":622,"value":727},"此法案處於立法二讀程序，議會若要阻止，須取得「絕對多數」——缺席議員的票也計入反對方的分母，使這一門檻在實務上被認為幾乎不可能達到。批評者將此形容為「試圖繞過民主控制機構、給議會一個突襲」，這種程序操作的合法性爭議，不亞於法案本身的實質內容。",{"type":617,"tag":661,"props":729,"children":731},{"id":730},"隱私-vs-兒童保護社群兩極化的激辯",[732],{"type":622,"value":733},"隱私 vs 兒童保護：社群兩極化的激辯",{"type":617,"tag":618,"props":735,"children":736},{},[737],{"type":622,"value":738},"支持方（各成員國政府）的核心論點是：自願偵測措施是「早期辨別受害兒童、拯救受虐者不可或缺的工具」，現行豁免期間協助發現大量案例，強制廢止等同讓受害者失去保護網。",{"type":617,"tag":618,"props":740,"children":741},{},[742],{"type":622,"value":743},"反對方則指出，「自願」掃描一旦取得法定授權基礎，將為日後強制義務建立現成技術架構。義大利等國政府明確反對大規模監控，卻仍在理事會投票支持，被倡議者 Patrick Breyer 直接批評為「自相矛盾的荒謬」。",{"type":617,"tag":618,"props":745,"children":746},{},[747],{"type":622,"value":748},"HN 社群用戶 SpicyLemonZest 提出一個值得注意的觀察：「Chat Control」詞彙本身帶有強烈立場，由反對者定義，支持方從不使用此詞。而 HN 用戶 neobrain 則提醒社群：此次復活的是允許平台「自願掃描」的 Chat Control 1.0，要求破壞加密的 Chat Control 2.0 因公民強烈抵制已暫緩——兩者不應混淆，否則會誤判當前的實質風險層級。",{"type":617,"tag":661,"props":750,"children":752},{"id":751},"對加密技術與科技產業的深遠衝擊",[753],{"type":622,"value":751},{"type":617,"tag":618,"props":755,"children":756},{},[757],{"type":622,"value":758},"Chat Control 2.0 中的客戶端掃描要求，被密碼學社群視為對端對端加密的根本性威脅。Signal 已明確聲明：若強制掃描法案通過，寧願撤出整個歐盟市場，而非在安全架構中留下可被利用的後門。Mullvad VPN 等隱私工具廠商也在密切追蹤立法動態，向用戶發出持續警示。",{"type":617,"tag":618,"props":760,"children":761},{},[762],{"type":622,"value":763},"若相關法案在歐盟形成合規義務，影響將不限於歐洲用戶。任何在歐盟有業務的全球性通訊服務——WhatsApp、iMessage、Telegram——都可能面臨同樣的合規壓力，必須重新評估能否在保留歐盟用戶的同時，維持現有的端對端加密安全承諾。這將是科技產業與各國監管機構之間一場事關隱私基礎設施的長期博弈。",{"title":156,"searchDepth":624,"depth":624,"links":765},[],{"data":767,"body":768,"excerpt":-1,"toc":823},{"title":156,"description":156},{"type":614,"children":769},[770,775,786,798,803,808,813,818],{"type":617,"tag":661,"props":771,"children":773},{"id":772},"核心條款",[774],{"type":622,"value":772},{"type":617,"tag":618,"props":776,"children":777},{},[778,780,784],{"type":622,"value":779},"Chat Control 1.0 的核心是《ePrivacy 指令》的臨時豁免授權，允許通訊服務提供者在未取得用戶同意的情況下，",{"type":617,"tag":678,"props":781,"children":782},{},[783],{"type":622,"value":682},{"type":622,"value":785},"私人訊息以偵測 CSAM 及 grooming 行為。此豁免使平台得以在不違反 ePrivacy 法規的前提下，主動篩查用戶通訊內容。",{"type":617,"tag":618,"props":787,"children":788},{},[789,791,796],{"type":622,"value":790},"Chat Control 2.0（仍在三方協商中）計畫將此掃描從自願改為",{"type":617,"tag":678,"props":792,"children":793},{},[794],{"type":622,"value":795},"強制義務",{"type":622,"value":797},"，且擴大至端對端加密訊息服務——由於加密訊息在傳輸中無法讀取，強制掃描必須在裝置端（加密前）執行，即所謂「客戶端掃描」，其本質是在用戶裝置上安裝偵測程式。",{"type":617,"tag":661,"props":799,"children":801},{"id":800},"適用範圍",[802],{"type":622,"value":800},{"type":617,"tag":618,"props":804,"children":805},{},[806],{"type":622,"value":807},"Chat Control 1.0 適用於所有在歐盟提供「網際網路通訊服務」的平台，包括電子郵件服務（Gmail、Outlook）、即時通訊軟體（WhatsApp、Messenger、Signal、Telegram）及具私訊功能的社交媒體平台。",{"type":617,"tag":618,"props":809,"children":810},{},[811],{"type":622,"value":812},"適用對象不限企業規模或國籍，只要在歐盟境內有用戶即受管轄，意味著美國、亞洲等地的科技公司只要提供歐盟用戶服務，均需評估合規義務。",{"type":617,"tag":661,"props":814,"children":816},{"id":815},"執法機制",[817],{"type":622,"value":815},{"type":617,"tag":618,"props":819,"children":820},{},[821],{"type":622,"value":822},"偵測到疑似 CSAM 的內容，平台須向指定通報機構（如 NCMEC 或各成員國對等機構）提交通報。資料保留規定要求偵測相關的訊息內容與流量資料，在嫌疑確認後最遲 12 個月內不可逆刪除；在確認前，資料可依法保存供執法調查使用。各成員國須指定負責監督合規的主管機關，並建立企業申訴機制。",{"title":156,"searchDepth":624,"depth":624,"links":824},[],{"data":826,"body":828,"excerpt":-1,"toc":839},{"title":156,"description":827},"平台需部署 CSAM 哈希比對系統（整合 PhotoDNA 或 NCMEC 資料庫）與 grooming AI 模式偵測模組，並嵌入訊息收發流程。",{"type":614,"children":829},[830,834],{"type":617,"tag":618,"props":831,"children":832},{},[833],{"type":622,"value":827},{"type":617,"tag":618,"props":835,"children":836},{},[837],{"type":622,"value":838},"資料管道須支援「偵測後 12 個月不可逆刪除」機制，需建置定時清理排程與稽核日誌。若 Chat Control 2.0 強制客戶端掃描義務落地，還需在 App 中嵌入掃描邏輯，並面對 App Store／Play Store 的額外分發審查壓力。",{"title":156,"searchDepth":624,"depth":624,"links":840},[],{"data":842,"body":844,"excerpt":-1,"toc":855},{"title":156,"description":843},"中型訊息平台預估需投入 6–18 個月工程週期，建置並驗收 CSAM 偵測基礎設施。第三方哈希資料庫（如 PhotoDNA）須支付授權費與持續維護成本。",{"type":614,"children":845},[846,850],{"type":617,"tag":618,"props":847,"children":848},{},[849],{"type":622,"value":843},{"type":617,"tag":618,"props":851,"children":852},{},[853],{"type":622,"value":854},"法律合規方面，需聘用資料保護長 (DPO) 並完成資料保護衝擊評估 (DPIA) ，屬一次性及年度循環支出。若跨多個歐盟成員國營運，各地監管對接成本將乘數放大。",{"title":156,"searchDepth":624,"depth":624,"links":856},[],{"data":858,"body":859,"excerpt":-1,"toc":891},{"title":156,"description":156},{"type":614,"children":860},[861],{"type":617,"tag":862,"props":863,"children":864},"ol",{},[865,871,876,881,886],{"type":617,"tag":866,"props":867,"children":868},"li",{},[869],{"type":622,"value":870},"整合 PhotoDNA 或 NCMEC 提供的哈希資料庫，僅對圖片與影片附件進行哈希比對",{"type":617,"tag":866,"props":872,"children":873},{},[874],{"type":622,"value":875},"建置自動通報流程，確保命中後於規定時效內通報指定機構",{"type":617,"tag":866,"props":877,"children":878},{},[879],{"type":622,"value":880},"實作偵測日誌的 12 個月不可逆刪除排程",{"type":617,"tag":866,"props":882,"children":883},{},[884],{"type":622,"value":885},"完成 DPIA 文件，確認掃描範圍不超出 Chat Control 1.0 授權邊界",{"type":617,"tag":866,"props":887,"children":888},{},[889],{"type":622,"value":890},"明確排除對純文字訊息的掃描，避免過度合規引發用戶信任危機",{"title":156,"searchDepth":624,"depth":624,"links":892},[],{"data":894,"body":895,"excerpt":-1,"toc":927},{"title":156,"description":156},{"type":614,"children":896},[897,902,907,912,917,922],{"type":617,"tag":661,"props":898,"children":900},{"id":899},"直接影響者",[901],{"type":622,"value":899},{"type":617,"tag":618,"props":903,"children":904},{},[905],{"type":622,"value":906},"即時通訊平台首當其衝，包括 WhatsApp、Signal、Telegram、iMessage 等全球主流服務。電子郵件服務商（Gmail、Outlook）及具私訊功能的社交媒體平台（Meta、Snapchat）同樣在適用範圍內。Signal 已明確表態將撤出歐盟市場，其他平台則多採取沉默觀察姿態。",{"type":617,"tag":661,"props":908,"children":910},{"id":909},"間接波及者",[911],{"type":622,"value":909},{"type":617,"tag":618,"props":913,"children":914},{},[915],{"type":622,"value":916},"裝置製造商（Apple、Google）若客戶端掃描義務落地，需提供 OS 層級 API 或面對應用層合規壓力。隱私工具提供商（VPN、加密郵件服務）可能因用戶隱私需求升溫而迎來新增流量。CSAM 偵測技術供應商（PhotoDNA 母公司 Microsoft、Thorn）則成為新增合規鏈中的必選採購方。",{"type":617,"tag":661,"props":918,"children":920},{"id":919},"成本轉嫁效應",[921],{"type":622,"value":919},{"type":617,"tag":618,"props":923,"children":924},{},[925],{"type":622,"value":926},"合規建置成本最終將以服務費上漲或功能降格（如關閉特定地區加密功能）的形式轉嫁至用戶端。更深遠的衝擊是用戶對私人通訊平台的信任度系統性下降，可能推動部分用戶轉向無歐盟管轄的替代工具，加速去中心化通訊協議（如 Matrix）的採用。",{"title":156,"searchDepth":624,"depth":624,"links":928},[],{"data":930,"body":931,"excerpt":-1,"toc":937},{"title":156,"description":103},{"type":614,"children":932},[933],{"type":617,"tag":618,"props":934,"children":935},{},[936],{"type":622,"value":103},{"title":156,"searchDepth":624,"depth":624,"links":938},[],{"data":940,"body":941,"excerpt":-1,"toc":947},{"title":156,"description":107},{"type":614,"children":942},[943],{"type":617,"tag":618,"props":944,"children":945},{},[946],{"type":622,"value":107},{"title":156,"searchDepth":624,"depth":624,"links":948},[],{"data":950,"body":951,"excerpt":-1,"toc":957},{"title":156,"description":110},{"type":614,"children":952},[953],{"type":617,"tag":618,"props":954,"children":955},{},[956],{"type":622,"value":110},{"title":156,"searchDepth":624,"depth":624,"links":958},[],{"data":960,"body":961,"excerpt":-1,"toc":967},{"title":156,"description":113},{"type":614,"children":962},[963],{"type":617,"tag":618,"props":964,"children":965},{},[966],{"type":622,"value":113},{"title":156,"searchDepth":624,"depth":624,"links":968},[],{"data":970,"body":971,"excerpt":-1,"toc":977},{"title":156,"description":116},{"type":614,"children":972},[973],{"type":617,"tag":618,"props":974,"children":975},{},[976],{"type":622,"value":116},{"title":156,"searchDepth":624,"depth":624,"links":978},[],{"data":980,"body":981,"excerpt":-1,"toc":987},{"title":156,"description":120},{"type":614,"children":982},[983],{"type":617,"tag":618,"props":984,"children":985},{},[986],{"type":622,"value":120},{"title":156,"searchDepth":624,"depth":624,"links":988},[],{"data":990,"body":991,"excerpt":-1,"toc":997},{"title":156,"description":125},{"type":614,"children":992},[993],{"type":617,"tag":618,"props":994,"children":995},{},[996],{"type":622,"value":125},{"title":156,"searchDepth":624,"depth":624,"links":998},[],{"data":1000,"body":1001,"excerpt":-1,"toc":1007},{"title":156,"description":55},{"type":614,"children":1002},[1003],{"type":617,"tag":618,"props":1004,"children":1005},{},[1006],{"type":622,"value":55},{"title":156,"searchDepth":624,"depth":624,"links":1008},[],{"data":1010,"body":1011,"excerpt":-1,"toc":1017},{"title":156,"description":56},{"type":614,"children":1012},[1013],{"type":617,"tag":618,"props":1014,"children":1015},{},[1016],{"type":622,"value":56},{"title":156,"searchDepth":624,"depth":624,"links":1018},[],{"data":1020,"body":1021,"excerpt":-1,"toc":1027},{"title":156,"description":143},{"type":614,"children":1022},[1023],{"type":617,"tag":618,"props":1024,"children":1025},{},[1026],{"type":622,"value":143},{"title":156,"searchDepth":624,"depth":624,"links":1028},[],{"data":1030,"body":1031,"excerpt":-1,"toc":1037},{"title":156,"description":147},{"type":614,"children":1032},[1033],{"type":617,"tag":618,"props":1034,"children":1035},{},[1036],{"type":622,"value":147},{"title":156,"searchDepth":624,"depth":624,"links":1038},[],{"data":1040,"body":1041,"excerpt":-1,"toc":1047},{"title":156,"description":150},{"type":614,"children":1042},[1043],{"type":617,"tag":618,"props":1044,"children":1045},{},[1046],{"type":622,"value":150},{"title":156,"searchDepth":624,"depth":624,"links":1048},[],{"data":1050,"body":1051,"excerpt":-1,"toc":1057},{"title":156,"description":153},{"type":614,"children":1052},[1053],{"type":617,"tag":618,"props":1054,"children":1055},{},[1056],{"type":622,"value":153},{"title":156,"searchDepth":624,"depth":624,"links":1058},[],{"data":1060,"body":1061,"excerpt":-1,"toc":1142},{"title":156,"description":156},{"type":614,"children":1062},[1063,1069,1074,1079,1085,1090,1095,1101,1106,1111,1117,1122,1127],{"type":617,"tag":661,"props":1064,"children":1066},{"id":1065},"章節一mturk-的歷史地位imagenet-與-ai-革命的幕後推手",[1067],{"type":622,"value":1068},"章節一：MTurk 的歷史地位——ImageNet 與 AI 革命的幕後推手",{"type":617,"tag":618,"props":1070,"children":1071},{},[1072],{"type":622,"value":1073},"2005 年 11 月，Amazon Mechanical Turk 正式上線，比 Freelancer 和 Fiverr 更早成為眾包微任務的商業市集。它的命名靈感來自 1769 年一台名為「機械土耳其人」的棋盤自動機——表面上是機器自主運作，實際上藏有人類操作者，正如 MTurk 以「人工智慧訓練」為名，實則仰賴隱匿工人的勞動。",{"type":617,"tag":618,"props":1075,"children":1076},{},[1077],{"type":622,"value":1078},"ImageNet(2009) 、COCO、Open Images 等奠定深度學習革命的大型資料集，背後都仰賴 MTurk 完成數以百萬計的圖片分類與標記工作。2018 年 AWS 將 MTurk 整合進 SageMaker，正式將其定位為 AI 訓練資料的工業化生產線。沒有 MTurk，深度學習的資料基礎可能要晚幾年才能搭建完成。",{"type":617,"tag":661,"props":1080,"children":1082},{"id":1081},"章節二為何現在關門專業標註平台崛起與市場轉移",[1083],{"type":622,"value":1084},"章節二：為何現在關門：專業標註平台崛起與市場轉移",{"type":617,"tag":618,"props":1086,"children":1087},{},[1088],{"type":622,"value":1089},"MTurk 退場並非突然，其市場地位在過去五年中逐漸被侵蝕。AWS 自家的 SageMaker Ground Truth 整合了自動化標註、品質控管與人機協作工作流，功能已全面超越 MTurk 的簡單任務市集模式。",{"type":617,"tag":618,"props":1091,"children":1092},{},[1093],{"type":622,"value":1094},"Scale AI、Surge AI 等新世代專業標註平台則從另一方向夾擊，提供更嚴格的品質保證與領域專精能力。醫療影像、自動駕駛、法律文件等高要求場景，早已轉向能提供可追責標註者的專業平台。MTurk 的泛用低門檻模式在專業 AI 訓練場景中逐漸成為累贅，TechCrunch 直言「這可能是 MTurk 的最後時日」。",{"type":617,"tag":661,"props":1096,"children":1098},{"id":1097},"章節三對資料標註生態系的連鎖效應",[1099],{"type":622,"value":1100},"章節三：對資料標註生態系的連鎖效應",{"type":617,"tag":618,"props":1102,"children":1103},{},[1104],{"type":622,"value":1105},"MTurk 的衰亡存在一個難以逆轉的死亡螺旋：AWS 近年多次無預警關閉工人帳號，導致活躍工人數持續下降，任務完成率隨之惡化，客戶開始出走，進而導致更多工人失去收入動機而離開平台。",{"type":617,"tag":618,"props":1107,"children":1108},{},[1109],{"type":622,"value":1110},"更深層的結構性問題於 2023 年浮出水面：研究分析顯示，高達 33–46% 的 MTurk 工人使用大型語言模型代替人工完成任務。這意味著，以「人工智慧訓練」為名收集的資料，實際上可能由 AI 生成——「人工標註資料的真實性假設」徹底崩潰，讓平台的核心價值主張失去立足點。",{"type":617,"tag":661,"props":1112,"children":1114},{"id":1113},"章節四後-mturk-時代合成資料與-rlhf-的新典範",[1115],{"type":622,"value":1116},"章節四：後 MTurk 時代——合成資料與 RLHF 的新典範",{"type":617,"tag":618,"props":1118,"children":1119},{},[1120],{"type":622,"value":1121},"現代 AI 訓練方法已從根本上改變了對人工標註的依賴程度。自監督學習讓模型能從未標註資料中學習，大幅削減標籤需求；合成資料可降低資料成本達 70%，並在規模與多樣性上遠超人工標註。2026 年合成資料市場規模已達 7.1 億美元，預計 2030 年成長至 23 億美元，直接取代人工標註的市場需求。",{"type":617,"tag":618,"props":1123,"children":1124},{},[1125],{"type":622,"value":1126},"RLHF 所需的人類偏好資料，也從 MTurk 式的大規模廉價眾包轉向更少量但更高品質的專家標註。Anthropic、OpenAI 等公司採用的方法，需要的是能判斷細微倫理差異的標註者，而非完成簡單微任務的廉價勞力。MTurk 代表的「大量廉價人工標註」典範，正隨著 AI 能力的提升進入歷史。",{"type":617,"tag":686,"props":1128,"children":1129},{},[1130],{"type":617,"tag":618,"props":1131,"children":1132},{},[1133,1137,1140],{"type":617,"tag":678,"props":1134,"children":1135},{},[1136],{"type":622,"value":696},{"type":617,"tag":698,"props":1138,"children":1139},{},[],{"type":622,"value":1141},"\nRLHF(Reinforcement Learning from Human Feedback) ：從人類回饋中進行強化學習，是訓練 ChatGPT、Claude 等對話 AI 使其符合人類偏好的核心技術，需要人類標註者對模型回應進行排序比較。",{"title":156,"searchDepth":624,"depth":624,"links":1143},[],{"data":1145,"body":1147,"excerpt":-1,"toc":1153},{"title":156,"description":1146},"MTurk 的退場揭示了 AI 資料標註生態系三個相互交織的結構性轉變：任務市集模式的內在缺陷、品質信任基礎的崩潰，以及資料生產典範的整體移轉。",{"type":614,"children":1148},[1149],{"type":617,"tag":618,"props":1150,"children":1151},{},[1152],{"type":622,"value":1146},{"title":156,"searchDepth":624,"depth":624,"links":1154},[],{"data":1156,"body":1158,"excerpt":-1,"toc":1185},{"title":156,"description":1157},"MTurk 採 Requesters（發任務）與 Workers（競標完成）的雙邊市集，以微支付驅動。這種模式在 2005–2015 年間高效運作，因為資訊不對稱讓平台能維持低工資均衡，工人接受低報酬是因為缺乏替代選擇。",{"type":614,"children":1159},[1160,1164,1169],{"type":617,"tag":618,"props":1161,"children":1162},{},[1163],{"type":622,"value":1157},{"type":617,"tag":618,"props":1165,"children":1166},{},[1167],{"type":622,"value":1168},"但雙邊市集有致命弱點：任何一方的信任崩潰都會引發惡性循環。AWS 多次無預警關閉帳號，摧毀了工人對平台的信任，加速活躍工人池萎縮，最終讓任務完成率跌至難以維持的水平。",{"type":617,"tag":686,"props":1170,"children":1171},{},[1172],{"type":617,"tag":618,"props":1173,"children":1174},{},[1175,1180,1183],{"type":617,"tag":678,"props":1176,"children":1177},{},[1178],{"type":622,"value":1179},"白話比喻",{"type":617,"tag":698,"props":1181,"children":1182},{},[],{"type":622,"value":1184},"\n想像一個計程車市場：司機接連被無故吊銷執照，剩下的司機開始掛名接單卻讓 AI 代駕，乘客漸漸發現車子沒人開——整個平台就這樣從內部垮掉。",{"title":156,"searchDepth":624,"depth":624,"links":1186},[],{"data":1188,"body":1190,"excerpt":-1,"toc":1201},{"title":156,"description":1189},"2023 年研究揭示的「LLM 代工」現象是 MTurk 最終的致命一擊：高達 46% 的工人用語言模型自動完成標註任務，再以「人工」名義提交。這在技術上形成了一個自我矛盾——我們用「人類智慧」標註的資料訓練 LLM，而這些資料本身已被 LLM 污染。",{"type":614,"children":1191},[1192,1196],{"type":617,"tag":618,"props":1193,"children":1194},{},[1195],{"type":622,"value":1189},{"type":617,"tag":618,"props":1197,"children":1198},{},[1199],{"type":622,"value":1200},"這不只是品質問題，更是信任基礎問題。一旦 Requesters 無法確認標註是否真的來自人類判斷，MTurk 相對專業平台的唯一優勢（廉價）便喪失意義——廉價但不可信的資料，在 AI 訓練中的價值為零，甚至為負。",{"title":156,"searchDepth":624,"depth":624,"links":1202},[],{"data":1204,"body":1206,"excerpt":-1,"toc":1217},{"title":156,"description":1205},"從技術路徑看，MTurk 面對來自兩個方向的替代：上游的合成資料生成，與下游的自動化標註流水線。SageMaker Ground Truth 將自動化預標註與人工驗證結合，能以 MTurk 不到一半的成本達到更高準確率。",{"type":614,"children":1207},[1208,1212],{"type":617,"tag":618,"props":1209,"children":1210},{},[1211],{"type":622,"value":1205},{"type":617,"tag":618,"props":1213,"children":1214},{},[1215],{"type":622,"value":1216},"合成資料則完全繞過了人工標註的需求。到 2030 年，合成資料市場預計達 23 億美元——這是從根本上重新定義「訓練資料怎麼來」的範式轉移，不是在改善現有方法，而是讓舊方法本身變得多餘。",{"title":156,"searchDepth":624,"depth":624,"links":1218},[],{"data":1220,"body":1221,"excerpt":-1,"toc":1324},{"title":156,"description":156},{"type":614,"children":1222},[1223,1228,1233,1239,1267,1272,1277,1282,1301,1306],{"type":617,"tag":661,"props":1224,"children":1226},{"id":1225},"環境需求",[1227],{"type":622,"value":1225},{"type":617,"tag":618,"props":1229,"children":1230},{},[1231],{"type":622,"value":1232},"遷移前需確認：現有 MTurk 任務類型（影像標記、文字分類、偏好排序等）、每月任務量與支出基準、當前品質指標（Worker Agreement Rate、HIT Rejection Rate），以及是否有自定義 Worker UI（HTML／JavaScript 模板）。AWS 帳號須具備 SageMaker 存取權限，Ground Truth 的 IAM 角色設定需提前驗證。",{"type":617,"tag":661,"props":1234,"children":1236},{"id":1235},"遷移整合步驟",[1237],{"type":622,"value":1238},"遷移／整合步驟",{"type":617,"tag":862,"props":1240,"children":1241},{},[1242,1247,1252,1257,1262],{"type":617,"tag":866,"props":1243,"children":1244},{},[1245],{"type":622,"value":1246},"匯出現有 MTurk HIT(Human Intelligence Task) 定義與 Qualification 篩選條件",{"type":617,"tag":866,"props":1248,"children":1249},{},[1250],{"type":622,"value":1251},"在 SageMaker Ground Truth 建立對應的標註工作類型，選擇 built-in 工作流（影像、文字、影片）或自定義 UI",{"type":617,"tag":866,"props":1253,"children":1254},{},[1255],{"type":622,"value":1256},"於 7 月 30 日前在 Ground Truth 設定 Amazon Mechanical Turk 公開工作隊（沿用現有 MTurk Workers），或評估私有工作隊以提升品質控管",{"type":617,"tag":866,"props":1258,"children":1259},{},[1260],{"type":622,"value":1261},"執行 A/B 比較：同批資料分別送 MTurk 和 Ground Truth，比較 Cohen's Kappa 標籤一致性與每任務完成時間",{"type":617,"tag":866,"props":1263,"children":1264},{},[1265],{"type":622,"value":1266},"若切換至 Scale AI、Surge AI 等外部平台，需撰寫完整標註指引 (Annotation Guidelines) 文件，並安排初期校準輪次",{"type":617,"tag":661,"props":1268,"children":1270},{"id":1269},"驗測規劃",[1271],{"type":622,"value":1269},{"type":617,"tag":618,"props":1273,"children":1274},{},[1275],{"type":622,"value":1276},"衡量遷移成功的關鍵指標：標籤 Cohen's Kappa（目標 > 0.8，MTurk 典型值 0.6–0.75）、每任務完成時間（與 MTurk 基準比較）、每標籤有效成本（含品質修正後）。遷移後第一個月應保留 10% 任務在 MTurk 上執行作為對照組，直至確認新平台穩定。",{"type":617,"tag":661,"props":1278,"children":1280},{"id":1279},"常見陷阱",[1281],{"type":622,"value":1279},{"type":617,"tag":1283,"props":1284,"children":1285},"ul",{},[1286,1291,1296],{"type":617,"tag":866,"props":1287,"children":1288},{},[1289],{"type":622,"value":1290},"直接照搬 MTurk HIT 設計到新平台——不同平台的工人激勵結構不同，任務指引與品質閾值需重新校準",{"type":617,"tag":866,"props":1292,"children":1293},{},[1294],{"type":622,"value":1295},"忽略歷史資料的 LLM 污染風險——2022 年後收集的 MTurk 資料集應進行抽樣品質審計再投入訓練",{"type":617,"tag":866,"props":1297,"children":1298},{},[1299],{"type":622,"value":1300},"高估合成資料對邊緣案例的覆蓋——長尾分布（罕見物件、方言、文化特定內容）仍需真實人工標註補充，不可全面替換",{"type":617,"tag":661,"props":1302,"children":1304},{"id":1303},"上線檢核清單",[1305],{"type":622,"value":1303},{"type":617,"tag":1283,"props":1307,"children":1308},{},[1309,1314,1319],{"type":617,"tag":866,"props":1310,"children":1311},{},[1312],{"type":622,"value":1313},"觀測：Worker Agreement Rate（跨工人標籤一致性）、任務完成率趨勢、品質審核拒絕率",{"type":617,"tag":866,"props":1315,"children":1316},{},[1317],{"type":622,"value":1318},"成本：有效標籤單價（標籤總成本 ÷ 通過品質審核的標籤數）vs. MTurk 歷史基準",{"type":617,"tag":866,"props":1320,"children":1321},{},[1322],{"type":622,"value":1323},"風險：7 月 30 日截止後的緊急備援計畫、現有 MTurk API 整合的服務中斷處置",{"title":156,"searchDepth":624,"depth":624,"links":1325},[],{"data":1327,"body":1328,"excerpt":-1,"toc":1458},{"title":156,"description":156},{"type":614,"children":1329},[1330,1335,1358,1363,1386,1391,1396,1401,1406,1424,1429,1447,1453],{"type":617,"tag":661,"props":1331,"children":1333},{"id":1332},"競爭版圖",[1334],{"type":622,"value":1332},{"type":617,"tag":1283,"props":1336,"children":1337},{},[1338,1348],{"type":617,"tag":866,"props":1339,"children":1340},{},[1341,1346],{"type":617,"tag":678,"props":1342,"children":1343},{},[1344],{"type":622,"value":1345},"直接競品",{"type":622,"value":1347},"：SageMaker Ground Truth（AWS 自家替代方案，MTurk 工人可直接遷移，阻力最低）、Scale AI（估值超過 140 億美元，高端 AI 訓練資料市場最大受益者）、Surge AI、Prolific（學術與 RLHF 偏好標註市場）",{"type":617,"tag":866,"props":1349,"children":1350},{},[1351,1356],{"type":617,"tag":678,"props":1352,"children":1353},{},[1354],{"type":622,"value":1355},"間接競品",{"type":622,"value":1357},"：合成資料平台（Gretel.ai、Argilla、Lilac）；自監督學習工具鏈（從根本上削減對標註資料的依賴）",{"type":617,"tag":661,"props":1359,"children":1361},{"id":1360},"護城河類型",[1362],{"type":622,"value":1360},{"type":617,"tag":1283,"props":1364,"children":1365},{},[1366,1376],{"type":617,"tag":866,"props":1367,"children":1368},{},[1369,1374],{"type":617,"tag":678,"props":1370,"children":1371},{},[1372],{"type":622,"value":1373},"工程護城河",{"type":622,"value":1375},"：SageMaker Ground Truth 的自動化預標註機制結合 IAM 整合，降低企業導入摩擦，同一 AWS 生態系減少供應商切換阻力",{"type":617,"tag":866,"props":1377,"children":1378},{},[1379,1384],{"type":617,"tag":678,"props":1380,"children":1381},{},[1382],{"type":622,"value":1383},"生態護城河",{"type":622,"value":1385},"：Scale AI 在自動駕駛、醫療、法律等垂直領域積累的標註指引體系與品質稽核機制，難以在短期內複製",{"type":617,"tag":661,"props":1387,"children":1389},{"id":1388},"定價策略",[1390],{"type":622,"value":1388},{"type":617,"tag":618,"props":1392,"children":1393},{},[1394],{"type":622,"value":1395},"MTurk 退場使資料標註市場往兩極化發展。大量泛用任務往合成資料移動（幾乎零邊際成本）；高價值的人類偏好資料、倫理判斷、領域專業標註往 Scale AI 等平台集中，每任務單價可能較 MTurk 高出 5–20 倍。",{"type":617,"tag":618,"props":1397,"children":1398},{},[1399],{"type":622,"value":1400},"「廉價眾包」這個市場區間正在消失，取而代之的是「免費合成」與「高價專業」的兩端格局。",{"type":617,"tag":661,"props":1402,"children":1404},{"id":1403},"企業導入阻力",[1405],{"type":622,"value":1403},{"type":617,"tag":1283,"props":1407,"children":1408},{},[1409,1414,1419],{"type":617,"tag":866,"props":1410,"children":1411},{},[1412],{"type":622,"value":1413},"現有 MTurk 工作流程的遷移成本：任務設計重寫、品質閾值重設、API 整合替換",{"type":617,"tag":866,"props":1415,"children":1416},{},[1417],{"type":622,"value":1418},"合規要求（GDPR、HIPAA）在新平台的重新驗證，尤其是涉及個人資料的標註任務",{"type":617,"tag":866,"props":1420,"children":1421},{},[1422],{"type":622,"value":1423},"歷史 MTurk 資料集的品質不確定性，部分資料集可能需要重新標註",{"type":617,"tag":661,"props":1425,"children":1427},{"id":1426},"第二序影響",[1428],{"type":622,"value":1426},{"type":617,"tag":1283,"props":1430,"children":1431},{},[1432,1437,1442],{"type":617,"tag":866,"props":1433,"children":1434},{},[1435],{"type":622,"value":1436},"AI 訓練資料成本兩極化：合成資料壓低低端價格，同時高品質人類標註的溢價持續提升",{"type":617,"tag":866,"props":1438,"children":1439},{},[1440],{"type":622,"value":1441},"全球眾包工人（尤其是發展中國家依賴微任務收入的群體）面臨工作機會收縮，是 AI 發展帶來的直接就業衝擊",{"type":617,"tag":866,"props":1443,"children":1444},{},[1445],{"type":622,"value":1446},"標註市場從碎片化眾包走向集中化：少數有品質保證的專業平台將主導市場，可選供應商的多樣性下降",{"type":617,"tag":661,"props":1448,"children":1450},{"id":1449},"判決scale-ai-等專業標註平台的主要受益方mturk-客戶出走加速高端市場整合",[1451],{"type":622,"value":1452},"判決：Scale AI 等專業標註平台的主要受益方（MTurk 客戶出走加速高端市場整合）",{"type":617,"tag":618,"props":1454,"children":1455},{},[1456],{"type":622,"value":1457},"MTurk 的退場是 AI 資料供應鏈重組的縮影，而非終結。市場並未消失，而是往「更少工人、更高品質、更高單價」的方向集中。短期內，現有 MTurk 客戶需要緊急評估遷移路徑；長期來看，合成資料的崛起將進一步壓縮人工標註的市場空間，只留下機器暫時無法勝任的判斷工作。",{"title":156,"searchDepth":624,"depth":624,"links":1459},[],{"data":1461,"body":1462,"excerpt":-1,"toc":1475},{"title":156,"description":156},{"type":614,"children":1463},[1464],{"type":617,"tag":1283,"props":1465,"children":1466},{},[1467,1471],{"type":617,"tag":866,"props":1468,"children":1469},{},[1470],{"type":622,"value":159},{"type":617,"tag":866,"props":1472,"children":1473},{},[1474],{"type":622,"value":160},{"title":156,"searchDepth":624,"depth":624,"links":1476},[],{"data":1478,"body":1479,"excerpt":-1,"toc":1492},{"title":156,"description":156},{"type":614,"children":1480},[1481],{"type":617,"tag":1283,"props":1482,"children":1483},{},[1484,1488],{"type":617,"tag":866,"props":1485,"children":1486},{},[1487],{"type":622,"value":162},{"type":617,"tag":866,"props":1489,"children":1490},{},[1491],{"type":622,"value":163},{"title":156,"searchDepth":624,"depth":624,"links":1493},[],{"data":1495,"body":1496,"excerpt":-1,"toc":1502},{"title":156,"description":167},{"type":614,"children":1497},[1498],{"type":617,"tag":618,"props":1499,"children":1500},{},[1501],{"type":622,"value":167},{"title":156,"searchDepth":624,"depth":624,"links":1503},[],{"data":1505,"body":1506,"excerpt":-1,"toc":1512},{"title":156,"description":168},{"type":614,"children":1507},[1508],{"type":617,"tag":618,"props":1509,"children":1510},{},[1511],{"type":622,"value":168},{"title":156,"searchDepth":624,"depth":624,"links":1513},[],{"data":1515,"body":1516,"excerpt":-1,"toc":1522},{"title":156,"description":214},{"type":614,"children":1517},[1518],{"type":617,"tag":618,"props":1519,"children":1520},{},[1521],{"type":622,"value":214},{"title":156,"searchDepth":624,"depth":624,"links":1523},[],{"data":1525,"body":1526,"excerpt":-1,"toc":1532},{"title":156,"description":218},{"type":614,"children":1527},[1528],{"type":617,"tag":618,"props":1529,"children":1530},{},[1531],{"type":622,"value":218},{"title":156,"searchDepth":624,"depth":624,"links":1533},[],{"data":1535,"body":1536,"excerpt":-1,"toc":1542},{"title":156,"description":221},{"type":614,"children":1537},[1538],{"type":617,"tag":618,"props":1539,"children":1540},{},[1541],{"type":622,"value":221},{"title":156,"searchDepth":624,"depth":624,"links":1543},[],{"data":1545,"body":1547,"excerpt":-1,"toc":1563},{"title":156,"description":1546},"已知 workaround：移除系統提示中的 ## Intermediary updates 區塊可消除聚集問題；長期建議採用開放 harness 架構，保持後端可替換性以應對供應商的靜默版本變更。",{"type":614,"children":1548},[1549],{"type":617,"tag":618,"props":1550,"children":1551},{},[1552,1554,1561],{"type":622,"value":1553},"已知 workaround：移除系統提示中的 ",{"type":617,"tag":1555,"props":1556,"children":1558},"code",{"className":1557},[],[1559],{"type":622,"value":1560},"## Intermediary updates",{"type":622,"value":1562}," 區塊可消除聚集問題；長期建議採用開放 harness 架構，保持後端可替換性以應對供應商的靜默版本變更。",{"title":156,"searchDepth":624,"depth":624,"links":1564},[],{"data":1566,"body":1567,"excerpt":-1,"toc":1711},{"title":156,"description":156},{"type":614,"children":1568},[1569,1575,1580,1585,1590,1596,1601,1613,1618,1623,1638,1644,1649,1670,1675,1690,1696,1701,1706],{"type":617,"tag":661,"props":1570,"children":1572},{"id":1571},"章節一問題浮現開發者如何發現-reasoning-token-異常聚集",[1573],{"type":622,"value":1574},"章節一：問題浮現——開發者如何發現 reasoning-token 異常聚集",{"type":617,"tag":618,"props":1576,"children":1577},{},[1578],{"type":622,"value":1579},"2026 年 6 月 27 日，GitHub 用戶 vguptaa45 在 openai/codex 倉庫提交 issue #30364，呈報 GPT-5.5 的 reasoning 輸出 token 出現高度異常的聚集現象。這項分析橫跨 390,195 筆 API 回應紀錄，時間範圍從 2026 年 2 月延伸至 6 月，是迄今最具規模的個人自發式效能監測案例之一。",{"type":617,"tag":618,"props":1581,"children":1582},{},[1583],{"type":622,"value":1584},"觸發深度分析的關鍵訊號在於一個驚人的倍數差異：GPT-5.5 的「恰好命中 516 token／超過 516 token」比率高達 44.0%，而非 GPT-5.5 模型僅 1.3%。",{"type":617,"tag":618,"props":1586,"children":1587},{},[1588],{"type":622,"value":1589},"GPT-5.5 雖只佔全部回應的 19.3%，卻貢獻了 82% 的「恰好 516 token」事件，三個峰值（516、1,034、1,552 token）之間的間距恰好為 518，呈現出顯然非隨機的週期性結構。",{"type":617,"tag":661,"props":1591,"children":1593},{"id":1592},"章節二clustering-機制推測與社群復現實驗",[1594],{"type":622,"value":1595},"章節二：Clustering 機制推測與社群復現實驗",{"type":617,"tag":618,"props":1597,"children":1598},{},[1599],{"type":622,"value":1600},"社群對「為何恰好停在 516」提出多種假說。第一條路線指向提示詞觸發：用戶 nsingh2 以糖果計算謎題重現問題，4/10 次執行停在 516 token 且全數答錯。",{"type":617,"tag":618,"props":1602,"children":1603},{},[1604,1606,1611],{"type":622,"value":1605},"他進一步發現，移除系統提示中的 ",{"type":617,"tag":1555,"props":1607,"children":1609},{"className":1608},[],[1610],{"type":622,"value":1560},{"type":622,"value":1612}," 區塊可完全消除問題，推測模型將「中途更新信號」誤解為最終回答，導致 reasoning 提前終止。",{"type":617,"tag":618,"props":1614,"children":1615},{},[1616],{"type":622,"value":1617},"第二條路線則指向底層基礎設施：516 ≈ 512 bytes 對齊加上 header overhead，有可能對應到記憶體配置邊界問題。",{"type":617,"tag":618,"props":1619,"children":1620},{},[1621],{"type":622,"value":1622},"月度數據走勢提供了另一條線索——clustering 比率從 2 月的 0.11% 暴增至 5 月的 53.30%，6 月雖回落至 35.84%，卻同步出現 GPT-5.5 平均 reasoning token 用量從 268.1 驟降至 106.9 的「用得少、錯得多」反常組合，意味著問題不是孤立錯誤，而是持續性退化。",{"type":617,"tag":686,"props":1624,"children":1625},{},[1626],{"type":617,"tag":618,"props":1627,"children":1628},{},[1629,1633,1636],{"type":617,"tag":678,"props":1630,"children":1631},{},[1632],{"type":622,"value":696},{"type":617,"tag":698,"props":1634,"children":1635},{},[],{"type":622,"value":1637},"\nClustering（聚集）：指模型輸出的 token 數量異常集中在某幾個固定數值，而非應有的平滑分佈，是偵測隱性行為退化的一種統計方法。",{"type":617,"tag":661,"props":1639,"children":1641},{"id":1640},"章節三開放-harness-的保險效應隨時可以換船",[1642],{"type":622,"value":1643},"章節三：開放 Harness 的保險效應：隨時可以換船",{"type":617,"tag":618,"props":1645,"children":1646},{},[1647],{"type":622,"value":1648},"HN 首則回覆 selectodude 直接點出了應對策略的核心：「這正是使用開放 harness 的美妙之處——有狀況我可以隨時換船。」開放 harness 讓開發者在後端模型退化時，以最小成本切換替代方案，不必等待供應商的修復公告。",{"type":617,"tag":618,"props":1650,"children":1651},{},[1652,1654,1660,1662,1668],{"type":622,"value":1653},"社群同步分享了 MIT 授權的 ",{"type":617,"tag":1555,"props":1655,"children":1657},{"className":1656},[],[1658],{"type":622,"value":1659},"clankerbend",{"type":622,"value":1661}," 工具（可透過 ",{"type":617,"tag":1555,"props":1663,"children":1665},{"className":1664},[],[1666],{"type":622,"value":1667},"npx clankerbend",{"type":622,"value":1669}," 直接使用），允許將相同 harness 配置接上不同後端模型。",{"type":617,"tag":618,"props":1671,"children":1672},{},[1673],{"type":622,"value":1674},"solarkraft 以 Deepseek 的行為作為對比基準——Deepseek 在多輪對話中會動態調整 reasoning budget，依任務複雜度分配推理空間，展示了「可觀測、可調控」的 reasoning 機制應有的樣貌，與 GPT-5.5 的不透明截斷形成鮮明對比。",{"type":617,"tag":686,"props":1676,"children":1677},{},[1678],{"type":617,"tag":618,"props":1679,"children":1680},{},[1681,1685,1688],{"type":617,"tag":678,"props":1682,"children":1683},{},[1684],{"type":622,"value":696},{"type":617,"tag":698,"props":1686,"children":1687},{},[],{"type":622,"value":1689},"\nHarness（執行容器）：包裹 LLM API 呼叫的框架層，負責管理提示詞、上下文、輸出格式等邏輯，使後端模型可被替換而不影響上層應用邏輯。",{"type":617,"tag":661,"props":1691,"children":1693},{"id":1692},"章節四reasoning-模型的隱性成本與信任危機",[1694],{"type":622,"value":1695},"章節四：Reasoning 模型的隱性成本與信任危機",{"type":617,"tag":618,"props":1697,"children":1698},{},[1699],{"type":622,"value":1700},"issue 作者明確聲明「數據不能證明 chain-of-thought 被截斷」，但現實問題在於：需要 6,000–8,000 token 才能穩定正確回答的任務，在 reasoning 空間悄悄退化至 106.9 token 後，付費用戶面臨實質性的效能下修，且缺乏任何版本公告或公開說明。",{"type":617,"tag":618,"props":1702,"children":1703},{},[1704],{"type":622,"value":1705},"同期的 issue #26876 記錄了更廣泛的退化數據：2,702 個 prompt、530 個 session，用戶挫敗率從基準期的 3.9% 升至晚期 12.1%。",{"type":617,"tag":618,"props":1707,"children":1708},{},[1709],{"type":622,"value":1710},"多名社群成員將此事與 Anthropic 先前的 Claude Code 退化事件相提並論，顯示大型推理模型的「隱性降級」已成跨廠商的系統性信任問題——能力變化不透明、用戶難以偵測、退化與修復均無公告。",{"title":156,"searchDepth":624,"depth":624,"links":1712},[],{"data":1714,"body":1716,"excerpt":-1,"toc":1722},{"title":156,"description":1715},"GPT-5.5 的 reasoning-token clustering 問題揭示了大型 LLM 在生產環境中可能發生的一類隱性退化：不是直接報錯，而是以「看似正常輸出、實際推理截斷」的形式悄悄降級。",{"type":614,"children":1717},[1718],{"type":617,"tag":618,"props":1719,"children":1720},{},[1721],{"type":622,"value":1715},{"title":156,"searchDepth":624,"depth":624,"links":1723},[],{"data":1725,"body":1727,"excerpt":-1,"toc":1738},{"title":156,"description":1726},"三個峰值（516、1,034、1,552 token）間距恰好為 518，在 390,195 筆回應中，516 token 精確命中 3,363 次。這種週期性結構在隨機推理過程中極不可能自然出現，強烈暗示存在某種外部截斷邊界。",{"type":614,"children":1728},[1729,1733],{"type":617,"tag":618,"props":1730,"children":1731},{},[1732],{"type":622,"value":1726},{"type":617,"tag":618,"props":1734,"children":1735},{},[1736],{"type":622,"value":1737},"GPT-5.5 的 exact-516／≥516 比率 (44.0%) 與其他模型 (1.3%) 相差 33 倍，是觸發深度分析的核心訊號。",{"title":156,"searchDepth":624,"depth":624,"links":1739},[],{"data":1741,"body":1743,"excerpt":-1,"toc":1772},{"title":156,"description":1742},"用戶 nsingh2 以糖果計算謎題進行可重現實驗，4/10 次執行停在 516 token 且答案全部錯誤。移除系統提示中的 ## Intermediary updates 區塊後，所有執行均成功。一個合理推測是：模型的注意力機制將「中途更新」格式標記誤解為完成信號，導致 reasoning chain 在中途提前收斂。",{"type":614,"children":1744},[1745,1757],{"type":617,"tag":618,"props":1746,"children":1747},{},[1748,1750,1755],{"type":622,"value":1749},"用戶 nsingh2 以糖果計算謎題進行可重現實驗，4/10 次執行停在 516 token 且答案全部錯誤。移除系統提示中的 ",{"type":617,"tag":1555,"props":1751,"children":1753},{"className":1752},[],[1754],{"type":622,"value":1560},{"type":622,"value":1756}," 區塊後，所有執行均成功。一個合理推測是：模型的注意力機制將「中途更新」格式標記誤解為完成信號，導致 reasoning chain 在中途提前收斂。",{"type":617,"tag":686,"props":1758,"children":1759},{},[1760],{"type":617,"tag":618,"props":1761,"children":1762},{},[1763,1767,1770],{"type":617,"tag":678,"props":1764,"children":1765},{},[1766],{"type":622,"value":696},{"type":617,"tag":698,"props":1768,"children":1769},{},[],{"type":622,"value":1771},"\nChain-of-thought（思維鏈）：讓模型在給出最終答案前先逐步推理的技術。Reasoning model 的 chain-of-thought 通常在後台運行，消耗 reasoning token，不直接顯示給用戶。",{"title":156,"searchDepth":624,"depth":624,"links":1773},[],{"data":1775,"body":1777,"excerpt":-1,"toc":1798},{"title":156,"description":1776},"516 ≈ 512 bytes + header overhead，指向底層 KV cache 或批次處理緩衝區的記憶體邊界對齊問題。若推理引擎在分配記憶體時以 512 bytes 為單位，且某些路徑下未正確擴展緩衝區，便可能導致 reasoning 在邊界處強制截斷，三個峰值恰好是 518 的整數倍也支持這一假說。",{"type":614,"children":1778},[1779,1783],{"type":617,"tag":618,"props":1780,"children":1781},{},[1782],{"type":622,"value":1776},{"type":617,"tag":686,"props":1784,"children":1785},{},[1786],{"type":617,"tag":618,"props":1787,"children":1788},{},[1789,1793,1796],{"type":617,"tag":678,"props":1790,"children":1791},{},[1792],{"type":622,"value":1179},{"type":617,"tag":698,"props":1794,"children":1795},{},[],{"type":622,"value":1797},"\n想像你的筆記本只有 512 頁，但老師要求解題過程必須寫滿。若有人在第 516 頁貼了一張「完成」貼紙，你可能就此擱筆——即使後面還有空白頁，即使答案根本還沒算出來。",{"title":156,"searchDepth":624,"depth":624,"links":1799},[],{"data":1801,"body":1802,"excerpt":-1,"toc":1913},{"title":156,"description":156},{"type":614,"children":1803},[1804,1808,1829,1833,1854,1858,1863,1868,1872,1885,1889,1902,1908],{"type":617,"tag":661,"props":1805,"children":1806},{"id":1332},[1807],{"type":622,"value":1332},{"type":617,"tag":1283,"props":1809,"children":1810},{},[1811,1820],{"type":617,"tag":866,"props":1812,"children":1813},{},[1814,1818],{"type":617,"tag":678,"props":1815,"children":1816},{},[1817],{"type":622,"value":1345},{"type":622,"value":1819},"：Anthropic Claude Code（同樣發生過未公告的效能退化事件，已成社群比較基準）、Deepseek（動態調整 reasoning budget，行為更透明可觀測）",{"type":617,"tag":866,"props":1821,"children":1822},{},[1823,1827],{"type":617,"tag":678,"props":1824,"children":1825},{},[1826],{"type":622,"value":1355},{"type":622,"value":1828},"：本地部署的開源推理模型（Llama 4、Qwen 3），不依賴 API 供應商的版本管理決策",{"type":617,"tag":661,"props":1830,"children":1831},{"id":1360},[1832],{"type":622,"value":1360},{"type":617,"tag":1283,"props":1834,"children":1835},{},[1836,1845],{"type":617,"tag":866,"props":1837,"children":1838},{},[1839,1843],{"type":617,"tag":678,"props":1840,"children":1841},{},[1842],{"type":622,"value":1373},{"type":622,"value":1844},"：GPT-5.5 在廣泛基準測試上仍保持領先，clustering 問題主要影響特定提示格式，不代表全面能力退化",{"type":617,"tag":866,"props":1846,"children":1847},{},[1848,1852],{"type":617,"tag":678,"props":1849,"children":1850},{},[1851],{"type":622,"value":1383},{"type":622,"value":1853},"：openai/codex 是最大的 AI 輔助編程生態，用戶遷移成本極高，短期難以大規模流失",{"type":617,"tag":661,"props":1855,"children":1856},{"id":1388},[1857],{"type":622,"value":1388},{"type":617,"tag":618,"props":1859,"children":1860},{},[1861],{"type":622,"value":1862},"此事件本質上是一場隱性的服務降級——用戶以相同價格獲得了更少的 reasoning capacity，且無任何通知。",{"type":617,"tag":618,"props":1864,"children":1865},{},[1866],{"type":622,"value":1867},"若 OpenAI 持續維持不透明的版本管理方式，企業用戶的 SLA 談判籌碼將增加，可能要求以「有效 reasoning token」而非「消耗 reasoning token」作為計費基礎。",{"type":617,"tag":661,"props":1869,"children":1870},{"id":1403},[1871],{"type":622,"value":1403},{"type":617,"tag":1283,"props":1873,"children":1874},{},[1875,1880],{"type":617,"tag":866,"props":1876,"children":1877},{},[1878],{"type":622,"value":1879},"缺乏 reasoning token 使用量的官方 SLA 保證，難以預估複雜任務的成本與品質穩定性",{"type":617,"tag":866,"props":1881,"children":1882},{},[1883],{"type":622,"value":1884},"版本變更無公告，CI／CD 迴歸測試可能在不知情的情況下對齊退化版本的行為",{"type":617,"tag":661,"props":1886,"children":1887},{"id":1426},[1888],{"type":622,"value":1426},{"type":617,"tag":1283,"props":1890,"children":1891},{},[1892,1897],{"type":617,"tag":866,"props":1893,"children":1894},{},[1895],{"type":622,"value":1896},"推動企業採用「模型無關」架構，增加中間層框架（如 LiteLLM、Martian）的商業價值",{"type":617,"tag":866,"props":1898,"children":1899},{},[1900],{"type":622,"value":1901},"社群自發監測工具（個人規模的 390K 筆日誌分析）可能演變為商業化的 LLM 健康監測服務",{"type":617,"tag":661,"props":1903,"children":1905},{"id":1904},"判決生態信任成本正在累積短期可消化長期需透明度改革",[1906],{"type":622,"value":1907},"判決：生態信任成本正在累積（短期可消化，長期需透明度改革）",{"type":617,"tag":618,"props":1909,"children":1910},{},[1911],{"type":622,"value":1912},"GPT-5.5 的 clustering 問題本身可能有 workaround，但連同 Claude Code 退化事件，「隱性降級」已成跨廠商的系統性模式。企業在採用大型推理模型時，必須將「不透明版本管理風險」列入採購評估，並在架構層保留切換能力。",{"title":156,"searchDepth":624,"depth":624,"links":1914},[],{"data":1916,"body":1917,"excerpt":-1,"toc":2145},{"title":156,"description":156},{"type":614,"children":1918},[1919,1925,2031,2036,2134,2140],{"type":617,"tag":661,"props":1920,"children":1922},{"id":1921},"clustering-事件統計",[1923],{"type":622,"value":1924},"Clustering 事件統計",{"type":617,"tag":1926,"props":1927,"children":1928},"table",{},[1929,1948],{"type":617,"tag":1930,"props":1931,"children":1932},"thead",{},[1933],{"type":617,"tag":1934,"props":1935,"children":1936},"tr",{},[1937,1943],{"type":617,"tag":1938,"props":1939,"children":1940},"th",{},[1941],{"type":622,"value":1942},"指標",{"type":617,"tag":1938,"props":1944,"children":1945},{},[1946],{"type":622,"value":1947},"數值",{"type":617,"tag":1949,"props":1950,"children":1951},"tbody",{},[1952,1966,1979,1992,2005,2018],{"type":617,"tag":1934,"props":1953,"children":1954},{},[1955,1961],{"type":617,"tag":1956,"props":1957,"children":1958},"td",{},[1959],{"type":622,"value":1960},"分析樣本",{"type":617,"tag":1956,"props":1962,"children":1963},{},[1964],{"type":622,"value":1965},"390,195 筆 API 回應",{"type":617,"tag":1934,"props":1967,"children":1968},{},[1969,1974],{"type":617,"tag":1956,"props":1970,"children":1971},{},[1972],{"type":622,"value":1973},"516 token 精確命中",{"type":617,"tag":1956,"props":1975,"children":1976},{},[1977],{"type":622,"value":1978},"3,363 次",{"type":617,"tag":1934,"props":1980,"children":1981},{},[1982,1987],{"type":617,"tag":1956,"props":1983,"children":1984},{},[1985],{"type":622,"value":1986},"GPT-5.5 佔樣本比例",{"type":617,"tag":1956,"props":1988,"children":1989},{},[1990],{"type":622,"value":1991},"19.3%",{"type":617,"tag":1934,"props":1993,"children":1994},{},[1995,2000],{"type":617,"tag":1956,"props":1996,"children":1997},{},[1998],{"type":622,"value":1999},"GPT-5.5 貢獻 516-event 比例",{"type":617,"tag":1956,"props":2001,"children":2002},{},[2003],{"type":622,"value":2004},"82%",{"type":617,"tag":1934,"props":2006,"children":2007},{},[2008,2013],{"type":617,"tag":1956,"props":2009,"children":2010},{},[2011],{"type":622,"value":2012},"exact-516／≥516 比率 (GPT-5.5)",{"type":617,"tag":1956,"props":2014,"children":2015},{},[2016],{"type":622,"value":2017},"44.0%",{"type":617,"tag":1934,"props":2019,"children":2020},{},[2021,2026],{"type":617,"tag":1956,"props":2022,"children":2023},{},[2024],{"type":622,"value":2025},"exact-516／≥516 比率（其他模型）",{"type":617,"tag":1956,"props":2027,"children":2028},{},[2029],{"type":622,"value":2030},"1.3%",{"type":617,"tag":661,"props":2032,"children":2034},{"id":2033},"月度退化走勢",[2035],{"type":622,"value":2033},{"type":617,"tag":1926,"props":2037,"children":2038},{},[2039,2060],{"type":617,"tag":1930,"props":2040,"children":2041},{},[2042],{"type":617,"tag":1934,"props":2043,"children":2044},{},[2045,2050,2055],{"type":617,"tag":1938,"props":2046,"children":2047},{},[2048],{"type":622,"value":2049},"月份",{"type":617,"tag":1938,"props":2051,"children":2052},{},[2053],{"type":622,"value":2054},"Clustering 比率",{"type":617,"tag":1938,"props":2056,"children":2057},{},[2058],{"type":622,"value":2059},"平均 Reasoning Token",{"type":617,"tag":1949,"props":2061,"children":2062},{},[2063,2081,2099,2116],{"type":617,"tag":1934,"props":2064,"children":2065},{},[2066,2071,2076],{"type":617,"tag":1956,"props":2067,"children":2068},{},[2069],{"type":622,"value":2070},"2026-02",{"type":617,"tag":1956,"props":2072,"children":2073},{},[2074],{"type":622,"value":2075},"0.11%",{"type":617,"tag":1956,"props":2077,"children":2078},{},[2079],{"type":622,"value":2080},"268.1",{"type":617,"tag":1934,"props":2082,"children":2083},{},[2084,2089,2094],{"type":617,"tag":1956,"props":2085,"children":2086},{},[2087],{"type":622,"value":2088},"2026-03",{"type":617,"tag":1956,"props":2090,"children":2091},{},[2092],{"type":622,"value":2093},"2.45%",{"type":617,"tag":1956,"props":2095,"children":2096},{},[2097],{"type":622,"value":2098},"—",{"type":617,"tag":1934,"props":2100,"children":2101},{},[2102,2107,2112],{"type":617,"tag":1956,"props":2103,"children":2104},{},[2105],{"type":622,"value":2106},"2026-05",{"type":617,"tag":1956,"props":2108,"children":2109},{},[2110],{"type":622,"value":2111},"53.30%",{"type":617,"tag":1956,"props":2113,"children":2114},{},[2115],{"type":622,"value":2098},{"type":617,"tag":1934,"props":2117,"children":2118},{},[2119,2124,2129],{"type":617,"tag":1956,"props":2120,"children":2121},{},[2122],{"type":622,"value":2123},"2026-06",{"type":617,"tag":1956,"props":2125,"children":2126},{},[2127],{"type":622,"value":2128},"35.84%",{"type":617,"tag":1956,"props":2130,"children":2131},{},[2132],{"type":622,"value":2133},"106.9",{"type":617,"tag":661,"props":2135,"children":2137},{"id":2136},"用戶挫敗率-issue-26876",[2138],{"type":622,"value":2139},"用戶挫敗率 (Issue #26876)",{"type":617,"tag":618,"props":2141,"children":2142},{},[2143],{"type":622,"value":2144},"2,702 個 prompt、530 個 session：基準期挫敗率 3.9%，退化後期升至 12.1%，升幅約 3 倍。",{"title":156,"searchDepth":624,"depth":624,"links":2146},[],{"data":2148,"body":2149,"excerpt":-1,"toc":2166},{"title":156,"description":156},{"type":614,"children":2150},[2151],{"type":617,"tag":1283,"props":2152,"children":2153},{},[2154,2158,2162],{"type":617,"tag":866,"props":2155,"children":2156},{},[2157],{"type":622,"value":230},{"type":617,"tag":866,"props":2159,"children":2160},{},[2161],{"type":622,"value":231},{"type":617,"tag":866,"props":2163,"children":2164},{},[2165],{"type":622,"value":232},{"title":156,"searchDepth":624,"depth":624,"links":2167},[],{"data":2169,"body":2170,"excerpt":-1,"toc":2195},{"title":156,"description":156},{"type":614,"children":2171},[2172],{"type":617,"tag":1283,"props":2173,"children":2174},{},[2175,2187,2191],{"type":617,"tag":866,"props":2176,"children":2177},{},[2178,2180,2185],{"type":622,"value":2179},"在系統提示中使用 ",{"type":617,"tag":1555,"props":2181,"children":2183},{"className":2182},[],[2184],{"type":622,"value":1560},{"type":622,"value":2186}," 格式的結構化標記（已知觸發條件）",{"type":617,"tag":866,"props":2188,"children":2189},{},[2190],{"type":622,"value":235},{"type":617,"tag":866,"props":2192,"children":2193},{},[2194],{"type":622,"value":236},{"title":156,"searchDepth":624,"depth":624,"links":2196},[],{"data":2198,"body":2199,"excerpt":-1,"toc":2205},{"title":156,"description":240},{"type":614,"children":2200},[2201],{"type":617,"tag":618,"props":2202,"children":2203},{},[2204],{"type":622,"value":240},{"title":156,"searchDepth":624,"depth":624,"links":2206},[],{"data":2208,"body":2210,"excerpt":-1,"toc":2224},{"title":156,"description":2209},"已知 workaround（移除 ## Intermediary updates 區塊）被多名用戶驗證有效，若問題僅限於特定提示格式，實際受影響的生產用例可能遠少於統計數字所呈現的嚴重程度",{"type":614,"children":2211},[2212],{"type":617,"tag":618,"props":2213,"children":2214},{},[2215,2217,2222],{"type":622,"value":2216},"已知 workaround（移除 ",{"type":617,"tag":1555,"props":2218,"children":2220},{"className":2219},[],[2221],{"type":622,"value":1560},{"type":622,"value":2223}," 區塊）被多名用戶驗證有效，若問題僅限於特定提示格式，實際受影響的生產用例可能遠少於統計數字所呈現的嚴重程度",{"title":156,"searchDepth":624,"depth":624,"links":2225},[],{"data":2227,"body":2228,"excerpt":-1,"toc":2234},{"title":156,"description":282},{"type":614,"children":2229},[2230],{"type":617,"tag":618,"props":2231,"children":2232},{},[2233],{"type":622,"value":282},{"title":156,"searchDepth":624,"depth":624,"links":2235},[],{"data":2237,"body":2238,"excerpt":-1,"toc":2244},{"title":156,"description":285},{"type":614,"children":2239},[2240],{"type":617,"tag":618,"props":2241,"children":2242},{},[2243],{"type":622,"value":285},{"title":156,"searchDepth":624,"depth":624,"links":2245},[],{"data":2247,"body":2248,"excerpt":-1,"toc":2254},{"title":156,"description":287},{"type":614,"children":2249},[2250],{"type":617,"tag":618,"props":2251,"children":2252},{},[2253],{"type":622,"value":287},{"title":156,"searchDepth":624,"depth":624,"links":2255},[],{"data":2257,"body":2258,"excerpt":-1,"toc":2264},{"title":156,"description":289},{"type":614,"children":2259},[2260],{"type":617,"tag":618,"props":2261,"children":2262},{},[2263],{"type":622,"value":289},{"title":156,"searchDepth":624,"depth":624,"links":2265},[],{"data":2267,"body":2268,"excerpt":-1,"toc":2389},{"title":156,"description":156},{"type":614,"children":2269},[2270,2276,2281,2286,2291,2297,2302,2307,2312,2327,2332,2338,2343,2358,2363,2368,2374,2379,2384],{"type":617,"tag":661,"props":2271,"children":2273},{"id":2272},"章節一傳統-ocr-的頁數瓶頸與長文件處理困境",[2274],{"type":622,"value":2275},"章節一：傳統 OCR 的頁數瓶頸與長文件處理困境",{"type":617,"tag":618,"props":2277,"children":2278},{},[2279],{"type":622,"value":2280},"標準 Transformer 解碼器每生成一個 token，就會向 KV Cache 新增一筆鍵值記錄。當文件頁數增加、輸出 token 數量暴增時，快取記憶體呈線性增長，最終逼近硬體上限。",{"type":617,"tag":618,"props":2282,"children":2283},{},[2284],{"type":622,"value":2285},"這迫使主流 OCR 系統採取「逐頁重置」策略——讀完一頁就清空記憶體，再從頭開始讀下一頁。這個工程妥協雖能繞過記憶體限制，卻導致跨頁語境斷裂，表格或公式一旦跨越頁面邊界便無法連貫辨識。",{"type":617,"tag":618,"props":2287,"children":2288},{},[2289],{"type":622,"value":2290},"業界主流系統因此形成約 10 頁的事實上限。長文件處理成為文件 AI 的隱形天花板，而非技術層面的真正解決。",{"type":617,"tag":661,"props":2292,"children":2294},{"id":2293},"章節二unlimited-ocr-的核心設計像人類一樣選擇性遺忘",[2295],{"type":622,"value":2296},"章節二：Unlimited OCR 的核心設計——像人類一樣選擇性遺忘",{"type":617,"tag":618,"props":2298,"children":2299},{},[2300],{"type":622,"value":2301},"百度的解法來自認知科學的類比：人類閱讀長文件時，持續保留對「眼前頁面圖像」的感知，同時只留住最近讀過的幾段文字作為上下文。",{"type":617,"tag":618,"props":2303,"children":2304},{},[2305],{"type":622,"value":2306},"Reference Sliding Window Attention(R-SWA) 將這個機制轉化為注意力設計。解碼器的注意力被拆成兩個區塊，各自負責不同的記憶功能。",{"type":617,"tag":618,"props":2308,"children":2309},{},[2310],{"type":622,"value":2311},"第一區塊對所有視覺 token 與 prompt 保持完整注意力，確保模型隨時「看得到」原始圖像；第二區塊對已生成的輸出 token 只保留最近 128 個，舊的記憶自動滑出視窗。",{"type":617,"tag":686,"props":2313,"children":2314},{},[2315],{"type":617,"tag":618,"props":2316,"children":2317},{},[2318,2322,2325],{"type":617,"tag":678,"props":2319,"children":2320},{},[2321],{"type":622,"value":696},{"type":617,"tag":698,"props":2323,"children":2324},{},[],{"type":622,"value":2326},"\nKV Cache（鍵值快取）：Transformer 推論時儲存歷史注意力鍵值對的記憶體結構，大小直接影響可處理的序列長度上限。",{"type":617,"tag":618,"props":2328,"children":2329},{},[2330],{"type":622,"value":2331},"KV Cache 因此維持固定長度佇列，上界為常數 Lm + n（Lm 為視覺 token 數，n 為滑動視窗大小），不隨頁數增加。這是突破頁數限制的根本機制。",{"type":617,"tag":661,"props":2333,"children":2335},{"id":2334},"章節三技術架構解析與效能實測",[2336],{"type":622,"value":2337},"章節三：技術架構解析與效能實測",{"type":617,"tag":618,"props":2339,"children":2340},{},[2341],{"type":622,"value":2342},"Unlimited OCR 採用 3B 參數的 Mixture-of-Experts(MoE) 架構，推論時僅啟用 500M 參數。視覺編碼器採用 DeepEncoder，將每頁 1024×1024 的 PDF 影像壓縮至 256 個視覺 token。",{"type":617,"tag":686,"props":2344,"children":2345},{},[2346],{"type":617,"tag":618,"props":2347,"children":2348},{},[2349,2353,2356],{"type":617,"tag":678,"props":2350,"children":2351},{},[2352],{"type":622,"value":696},{"type":617,"tag":698,"props":2354,"children":2355},{},[],{"type":622,"value":2357},"\nMixture-of-Experts(MoE) ：一種稀疏神經網路架構，每次推論只激活部分「專家」子網路，在保有大模型容量的同時大幅降低計算成本。",{"type":617,"tag":618,"props":2359,"children":2360},{},[2361],{"type":622,"value":2362},"效能測試中，在 OmniDocBench v1.5 獲得 93.23 分，較 DeepSeek OCR 基準線 (87.01) 提升 6.22 分；v1.6 版本進一步達到 93.92%，確認跨版本一致性。",{"type":617,"tag":618,"props":2364,"children":2365},{},[2366],{"type":622,"value":2367},"推論速度達每秒 5,580 tokens，比 DeepSeek OCR 快 12.7%，且在 6,000 步解碼後延遲仍保持平坦。40 頁以上文件的文字編輯距離 (Edit Distance) 低於 0.11，顯示辨識品質未隨頁數衰退。",{"type":617,"tag":661,"props":2369,"children":2371},{"id":2370},"章節四文件-ai-的下一步從辨識到理解",[2372],{"type":622,"value":2373},"章節四：文件 AI 的下一步——從辨識到理解",{"type":617,"tag":618,"props":2375,"children":2376},{},[2377],{"type":622,"value":2378},"R-SWA 論文作者明確指出，這個注意力機制並非 OCR 專屬，而是任何「解碼器需要永久存取固定來源、同時生成長輸出」場景的通用解法——語音辨識 (ASR) 與機器翻譯均在適用場景之列。",{"type":617,"tag":618,"props":2380,"children":2381},{},[2382],{"type":622,"value":2383},"對文件 AI 產業而言，Unlimited OCR 的意義超越辨識準確率的提升。整本報告單次推論成為可能，讓下游 RAG 管道不再需要依賴分頁切片與跨頁拼接等工程補丁。",{"type":617,"tag":618,"props":2385,"children":2386},{},[2387],{"type":622,"value":2388},"MIT 授權的開源策略讓這個突破得以直接進入企業文件數位化管道，而非停留在論文實驗室。開源社群的二次開發將加速生態聚集，推動文件 AI 的整體基準線向上移動。",{"title":156,"searchDepth":624,"depth":624,"links":2390},[],{"data":2392,"body":2394,"excerpt":-1,"toc":2400},{"title":156,"description":2393},"R-SWA 的設計目標是讓 KV Cache 大小與輸出長度完全脫鉤。傳統解碼器每步都把新生成的 token 推入快取，導致快取無限增長；R-SWA 透過結構性分區解決這個問題。",{"type":614,"children":2395},[2396],{"type":617,"tag":618,"props":2397,"children":2398},{},[2399],{"type":622,"value":2393},{"title":156,"searchDepth":624,"depth":624,"links":2401},[],{"data":2403,"body":2405,"excerpt":-1,"toc":2416},{"title":156,"description":2404},"視覺編碼器將每頁 1024×1024 的 PDF 影像壓縮成 256 個視覺 token，連同原始 prompt 一起放入注意力的「永久區」。無論生成多少輸出，解碼器對這個區塊始終保持完整注意力。",{"type":614,"children":2406},[2407,2411],{"type":617,"tag":618,"props":2408,"children":2409},{},[2410],{"type":622,"value":2404},{"type":617,"tag":618,"props":2412,"children":2413},{},[2414],{"type":622,"value":2415},"這個設計確保每個輸出 token 都能回望完整的原始圖像資訊，不受輸出長度影響。來源圖像是固定且完整的，不能被遺忘——這是 OCR 場景的關鍵約束。",{"title":156,"searchDepth":624,"depth":624,"links":2417},[],{"data":2419,"body":2421,"excerpt":-1,"toc":2447},{"title":156,"description":2420},"已生成的輸出 token 進入「滑動視窗區」，大小固定為 128 個 token。每生成一個新 token，最舊的一個就滑出視窗，整個快取大小維持不變。",{"type":614,"children":2422},[2423,2427,2432],{"type":617,"tag":618,"props":2424,"children":2425},{},[2426],{"type":622,"value":2420},{"type":617,"tag":618,"props":2428,"children":2429},{},[2430],{"type":622,"value":2431},"KV Cache 的總大小上界為 Lm（視覺 token 數）＋ n（滑動視窗大小），是與頁數無關的常數。這個設計類比人類閱讀時只記住最近幾段，而非逐字回憶全文。",{"type":617,"tag":686,"props":2433,"children":2434},{},[2435],{"type":617,"tag":618,"props":2436,"children":2437},{},[2438,2442,2445],{"type":617,"tag":678,"props":2439,"children":2440},{},[2441],{"type":622,"value":1179},{"type":617,"tag":698,"props":2443,"children":2444},{},[],{"type":622,"value":2446},"\n把 Unlimited OCR 想像成一位速記員：桌上永遠攤開著原始文件（視覺 token 永久區），便條紙只記最近 128 個字（滑動視窗），舊便條紙就丟掉。桌子永遠不會被便條紙淹沒，不管文件有多厚。",{"title":156,"searchDepth":624,"depth":624,"links":2448},[],{"data":2450,"body":2452,"excerpt":-1,"toc":2463},{"title":156,"description":2451},"3B 參數的 MoE 架構在推論時僅啟用 500M 參數，讓整個系統可在 12GB VRAM 的消費級 GPU 上執行，大幅降低硬體門檻。",{"type":614,"children":2453},[2454,2458],{"type":617,"tag":618,"props":2455,"children":2456},{},[2457],{"type":622,"value":2451},{"type":617,"tag":618,"props":2459,"children":2460},{},[2461],{"type":622,"value":2462},"每秒 5,580 tokens 的吞吐量在 6,000 步後仍維持平坦延遲，意味著長文件推論不會隨頁數出現速度衰減。這是 KV Cache 固定大小與 MoE 稀疏啟動共同帶來的系統級效益。",{"title":156,"searchDepth":624,"depth":624,"links":2464},[],{"data":2466,"body":2467,"excerpt":-1,"toc":2584},{"title":156,"description":156},{"type":614,"children":2468},[2469,2473,2494,2498,2519,2523,2528,2532,2556,2560,2573,2579],{"type":617,"tag":661,"props":2470,"children":2471},{"id":1332},[2472],{"type":622,"value":1332},{"type":617,"tag":1283,"props":2474,"children":2475},{},[2476,2485],{"type":617,"tag":866,"props":2477,"children":2478},{},[2479,2483],{"type":617,"tag":678,"props":2480,"children":2481},{},[2482],{"type":622,"value":1345},{"type":622,"value":2484},"：DeepSeek OCR（基準線 87.01，速度慢 12.7%）、Mistral OCR、Azure Document Intelligence、Google Document AI",{"type":617,"tag":866,"props":2486,"children":2487},{},[2488,2492],{"type":617,"tag":678,"props":2489,"children":2490},{},[2491],{"type":622,"value":1355},{"type":622,"value":2493},"：傳統 Tesseract + 後處理管道、Adobe Acrobat Pro OCR、AWS Textract",{"type":617,"tag":661,"props":2495,"children":2496},{"id":1360},[2497],{"type":622,"value":1360},{"type":617,"tag":1283,"props":2499,"children":2500},{},[2501,2510],{"type":617,"tag":866,"props":2502,"children":2503},{},[2504,2508],{"type":617,"tag":678,"props":2505,"children":2506},{},[2507],{"type":622,"value":1373},{"type":622,"value":2509},"：R-SWA 注意力機制設計巧思，競品短期需重新訓練才能複製；6,000 步平坦延遲曲線意味著長文件場景效能優勢難以靠微調追上",{"type":617,"tag":866,"props":2511,"children":2512},{},[2513,2517],{"type":617,"tag":678,"props":2514,"children":2515},{},[2516],{"type":622,"value":1383},{"type":622,"value":2518},"：MIT 授權降低企業採用門檻，Hugging Face 模型權重讓整合成本極低；開源社群二次開發將加速生態聚集",{"type":617,"tag":661,"props":2520,"children":2521},{"id":1388},[2522],{"type":622,"value":1388},{"type":617,"tag":618,"props":2524,"children":2525},{},[2526],{"type":622,"value":2527},"目前以 MIT 授權開源釋出，無直接授權費用。企業若使用百度雲端 API 版本，預期採用與 DeepSeek OCR API 類似的 token 計費模式。本地部署零授權成本是相對 Azure、Google 等商業競品的顯著優勢。",{"type":617,"tag":661,"props":2529,"children":2530},{"id":1403},[2531],{"type":622,"value":1403},{"type":617,"tag":1283,"props":2533,"children":2534},{},[2535,2546,2551],{"type":617,"tag":866,"props":2536,"children":2537},{},[2538,2544],{"type":617,"tag":1555,"props":2539,"children":2541},{"className":2540},[],[2542],{"type":622,"value":2543},"trust_remote_code=True",{"type":622,"value":2545}," 觸發安全審查流程，大型企業 IT 合規部門可能需要額外時間審核自定義注意力程式碼",{"type":617,"tag":866,"props":2547,"children":2548},{},[2549],{"type":622,"value":2550},"訓練資料來源透明度不足，GDPR、HIPAA 等合規需求可能要求更詳細的資料說明",{"type":617,"tag":866,"props":2552,"children":2553},{},[2554],{"type":622,"value":2555},"百度品牌在部分市場存在地緣政治敏感性，影響企業採購決策",{"type":617,"tag":661,"props":2557,"children":2558},{"id":1426},[2559],{"type":622,"value":1426},{"type":617,"tag":1283,"props":2561,"children":2562},{},[2563,2568],{"type":617,"tag":866,"props":2564,"children":2565},{},[2566],{"type":622,"value":2567},"RAG 管道設計簡化：文件切片、跨頁拼接的工程補丁需求降低，開發者可專注語義理解層",{"type":617,"tag":866,"props":2569,"children":2570},{},[2571],{"type":622,"value":2572},"文件 AI 基準競爭加速：Unlimited OCR 將 OmniDocBench 門檻抬高至 93+，迫使競品加快研發週期",{"type":617,"tag":661,"props":2574,"children":2576},{"id":2575},"判決值得商業評估mit-授權消除授權壁壘長文件場景優勢明確",[2577],{"type":622,"value":2578},"判決：值得商業評估（MIT 授權消除授權壁壘，長文件場景優勢明確）",{"type":617,"tag":618,"props":2580,"children":2581},{},[2582],{"type":622,"value":2583},"對於有大量 20+ 頁文件處理需求的企業，Unlimited OCR 提供明確的技術優勢與零授權成本，短期內值得進行 PoC 評估。地緣政治與合規疑慮是需要獨立評估的風險因素。",{"title":156,"searchDepth":624,"depth":624,"links":2585},[],{"data":2587,"body":2588,"excerpt":-1,"toc":2634},{"title":156,"description":156},{"type":614,"children":2589},[2590,2596,2601,2606,2624,2629],{"type":617,"tag":661,"props":2591,"children":2593},{"id":2592},"omnidocbench-綜合評測",[2594],{"type":622,"value":2595},"OmniDocBench 綜合評測",{"type":617,"tag":618,"props":2597,"children":2598},{},[2599],{"type":622,"value":2600},"Unlimited OCR 在 OmniDocBench v1.5 獲得 93.23 分，較 DeepSeek OCR 基準線 (87.01) 提升 6.22 分；v1.6 版本進一步達到 93.92%，確認跨版本基準一致性。",{"type":617,"tag":661,"props":2602,"children":2604},{"id":2603},"子任務分解",[2605],{"type":622,"value":2603},{"type":617,"tag":1283,"props":2607,"children":2608},{},[2609,2614,2619],{"type":617,"tag":866,"props":2610,"children":2611},{},[2612],{"type":622,"value":2613},"表格結構辨識 (TEDS) ：90.93（較基準線提升近 6 個百分點）",{"type":617,"tag":866,"props":2615,"children":2616},{},[2617],{"type":622,"value":2618},"公式辨識：92.61",{"type":617,"tag":866,"props":2620,"children":2621},{},[2622],{"type":622,"value":2623},"40 頁以上文件文字編輯距離 (Edit Distance) ：\u003C 0.11",{"type":617,"tag":661,"props":2625,"children":2627},{"id":2626},"速度對比",[2628],{"type":622,"value":2626},{"type":617,"tag":618,"props":2630,"children":2631},{},[2632],{"type":622,"value":2633},"推論速度達每秒 5,580 tokens，比 DeepSeek OCR 的 4,951 快 12.7%。在 6,000 步解碼後延遲曲線保持平坦，驗證 KV Cache 固定大小設計對長文件推論的系統級穩定性。",{"title":156,"searchDepth":624,"depth":624,"links":2635},[],{"data":2637,"body":2638,"excerpt":-1,"toc":2659},{"title":156,"description":156},{"type":614,"children":2639},[2640],{"type":617,"tag":1283,"props":2641,"children":2642},{},[2643,2647,2651,2655],{"type":617,"tag":866,"props":2644,"children":2645},{},[2646],{"type":622,"value":295},{"type":617,"tag":866,"props":2648,"children":2649},{},[2650],{"type":622,"value":296},{"type":617,"tag":866,"props":2652,"children":2653},{},[2654],{"type":622,"value":297},{"type":617,"tag":866,"props":2656,"children":2657},{},[2658],{"type":622,"value":298},{"title":156,"searchDepth":624,"depth":624,"links":2660},[],{"data":2662,"body":2663,"excerpt":-1,"toc":2680},{"title":156,"description":156},{"type":614,"children":2664},[2665],{"type":617,"tag":1283,"props":2666,"children":2667},{},[2668,2672,2676],{"type":617,"tag":866,"props":2669,"children":2670},{},[2671],{"type":622,"value":300},{"type":617,"tag":866,"props":2673,"children":2674},{},[2675],{"type":622,"value":301},{"type":617,"tag":866,"props":2677,"children":2678},{},[2679],{"type":622,"value":302},{"title":156,"searchDepth":624,"depth":624,"links":2681},[],{"data":2683,"body":2684,"excerpt":-1,"toc":2690},{"title":156,"description":306},{"type":614,"children":2685},[2686],{"type":617,"tag":618,"props":2687,"children":2688},{},[2689],{"type":622,"value":306},{"title":156,"searchDepth":624,"depth":624,"links":2691},[],{"data":2693,"body":2694,"excerpt":-1,"toc":2700},{"title":156,"description":307},{"type":614,"children":2695},[2696],{"type":617,"tag":618,"props":2697,"children":2698},{},[2699],{"type":622,"value":307},{"title":156,"searchDepth":624,"depth":624,"links":2701},[],{"data":2703,"body":2704,"excerpt":-1,"toc":2710},{"title":156,"description":308},{"type":614,"children":2705},[2706],{"type":617,"tag":618,"props":2707,"children":2708},{},[2709],{"type":622,"value":308},{"title":156,"searchDepth":624,"depth":624,"links":2711},[],{"data":2713,"body":2714,"excerpt":-1,"toc":2757},{"title":156,"description":156},{"type":614,"children":2715},[2716,2722,2727,2732,2747,2752],{"type":617,"tag":661,"props":2717,"children":2719},{"id":2718},"_23-年前的-c-引擎原生跑在-iphone-上",[2720],{"type":622,"value":2721},"23 年前的 C++ 引擎，原生跑在 iPhone 上",{"type":617,"tag":618,"props":2723,"children":2724},{},[2725],{"type":622,"value":2726},"Google DeepMind 產品設計主管 Ammaar Reshi，以 Claude Code 搭配 Fable 5，在約兩天內完成《Command & Conquer： Generals Zero Hour》的 iOS 移植——使用真實 2003 年 C++ 引擎，直接編譯為 ARM64 原生執行，非模擬器。",{"type":617,"tag":618,"props":2728,"children":2729},{},[2730],{"type":622,"value":2731},"移植路徑為 DirectX 8 → DXVK/MoltenVK → Apple Metal API，支援 Campaign、Skirmish 對戰、Generals Challenge，並適配觸控操作（框選拖曳、雙指平移、捏合縮放）。原始碼已於 GitHub 開源，不含遊戲資產。",{"type":617,"tag":686,"props":2733,"children":2734},{},[2735],{"type":617,"tag":618,"props":2736,"children":2737},{},[2738,2742,2745],{"type":617,"tag":678,"props":2739,"children":2740},{},[2741],{"type":622,"value":696},{"type":617,"tag":698,"props":2743,"children":2744},{},[],{"type":622,"value":2746},"\nDXVK/MoltenVK 是圖形 API 轉換層，讓 DirectX 指令得以在 Vulkan(DXVK) 或 Apple Metal(MoltenVK) 上執行。",{"type":617,"tag":661,"props":2748,"children":2750},{"id":2749},"開發節奏與關鍵轉折",[2751],{"type":622,"value":2749},{"type":617,"tag":618,"props":2753,"children":2754},{},[2755],{"type":622,"value":2756},"第一個可執行的 build 只花了 40 分鐘。Reshi 表示，初期使用 Claude Opus 4.8，切換至 Fable 5 後才突破關鍵瓶頸，最終 19 小時內完成 19 次 commits，整個過程耗盡了他的 Claude Max 配額。已知限制：長時間使用後可能因記憶體消耗過高而崩潰。",{"title":156,"searchDepth":624,"depth":624,"links":2758},[],{"data":2760,"body":2761,"excerpt":-1,"toc":2767},{"title":156,"description":348},{"type":614,"children":2762},[2763],{"type":617,"tag":618,"props":2764,"children":2765},{},[2766],{"type":622,"value":348},{"title":156,"searchDepth":624,"depth":624,"links":2768},[],{"data":2770,"body":2771,"excerpt":-1,"toc":2777},{"title":156,"description":349},{"type":614,"children":2772},[2773],{"type":617,"tag":618,"props":2774,"children":2775},{},[2776],{"type":622,"value":349},{"title":156,"searchDepth":624,"depth":624,"links":2778},[],{"data":2780,"body":2781,"excerpt":-1,"toc":2852},{"title":156,"description":156},{"type":614,"children":2782},[2783,2789,2794,2799,2804,2809,2832,2847],{"type":617,"tag":661,"props":2784,"children":2786},{"id":2785},"封閉-ai-的前排席位問題",[2787],{"type":622,"value":2788},"封閉 AI 的「前排席位」問題",{"type":617,"tag":618,"props":2790,"children":2791},{},[2792],{"type":622,"value":2793},"Mistral AI 執行長 Arthur Mensch 於 2026 年 7 月 5 日警告：依賴封閉 AI 模型，等同讓供應商取得業務流程的「前排席位」。供應商透過強制資料留存政策持續觀察企業運作，部分甚至以此轉攻自家客戶的市場。",{"type":617,"tag":618,"props":2795,"children":2796},{},[2797],{"type":622,"value":2798},"他以 2025 年 Anthropic 切斷 Windsurf API 為例——同期 Anthropic 推出競品 Claude Code——佐證供應商「針對最成功客戶出手」的模式。",{"type":617,"tag":661,"props":2800,"children":2802},{"id":2801},"四步應對策略",[2803],{"type":622,"value":2801},{"type":617,"tag":618,"props":2805,"children":2806},{},[2807],{"type":622,"value":2808},"Mensch 建議企業採取：",{"type":617,"tag":862,"props":2810,"children":2811},{},[2812,2817,2822,2827],{"type":617,"tag":866,"props":2813,"children":2814},{},[2815],{"type":622,"value":2816},"採用開源模型",{"type":617,"tag":866,"props":2818,"children":2819},{},[2820],{"type":622,"value":2821},"建立開放資料系統並自設存取控制",{"type":617,"tag":866,"props":2823,"children":2824},{},[2825],{"type":622,"value":2826},"打造持續訓練飛輪 (continuous training flywheel)",{"type":617,"tag":866,"props":2828,"children":2829},{},[2830],{"type":622,"value":2831},"建立競爭對手難以複製的自有 AI 系統",{"type":617,"tag":686,"props":2833,"children":2834},{},[2835],{"type":617,"tag":618,"props":2836,"children":2837},{},[2838,2842,2845],{"type":617,"tag":678,"props":2839,"children":2840},{},[2841],{"type":622,"value":696},{"type":617,"tag":698,"props":2843,"children":2844},{},[],{"type":622,"value":2846},"\n持續訓練飛輪：企業以自有資料持續微調模型，形成資料積累與能力強化的正向循環，競爭對手難以複製。",{"type":617,"tag":618,"props":2848,"children":2849},{},[2850],{"type":622,"value":2851},"Mensch 的論述與 Mistral 商業利益高度吻合——該公司正洽談以 200 億歐元估值融資，「開源優先」主張本身也是市場定位策略。",{"title":156,"searchDepth":624,"depth":624,"links":2853},[],{"data":2855,"body":2857,"excerpt":-1,"toc":2886},{"title":156,"description":2856},"Windsurf 事件讓 API 依賴的風險具體化。評估 AI 工作流時，建議考量三點：",{"type":614,"children":2858},[2859,2863,2881],{"type":617,"tag":618,"props":2860,"children":2861},{},[2862],{"type":622,"value":2856},{"type":617,"tag":862,"props":2864,"children":2865},{},[2866,2871,2876],{"type":617,"tag":866,"props":2867,"children":2868},{},[2869],{"type":622,"value":2870},"核心業務邏輯優先選 Apache 2.0 / MIT 授權的開源模型",{"type":617,"tag":866,"props":2872,"children":2873},{},[2874],{"type":622,"value":2875},"透過 vLLM 或 Ollama 建立本地推理層，解耦 API 依賴",{"type":617,"tag":866,"props":2877,"children":2878},{},[2879],{"type":622,"value":2880},"採用閉源 API 前確認資料留存與競業條款",{"type":617,"tag":618,"props":2882,"children":2883},{},[2884],{"type":622,"value":2885},"自有訓練飛輪建置成本高，但對核心 AI 應用具長期防禦價值。",{"title":156,"searchDepth":624,"depth":624,"links":2887},[],{"data":2889,"body":2891,"excerpt":-1,"toc":2910},{"title":156,"description":2890},"Mensch 的論述即使存在商業動機，核心邏輯仍值得正視：與供應商共享業務資料，等同潛在競爭情報外流。",{"type":614,"children":2892},[2893,2905],{"type":617,"tag":618,"props":2894,"children":2895},{},[2896,2898,2903],{"type":622,"value":2897},"Mensch 的論述即使存在商業動機，核心邏輯仍值得正視：",{"type":617,"tag":678,"props":2899,"children":2900},{},[2901],{"type":622,"value":2902},"與供應商共享業務資料，等同潛在競爭情報外流",{"type":622,"value":2904},"。",{"type":617,"tag":618,"props":2906,"children":2907},{},[2908],{"type":622,"value":2909},"短期而言，閉源模型的易用性優勢仍在；長期來看，涉及核心業務流程的 AI 化應優先考慮開源或自托管方案，以保護競爭護城河。關鍵決策是區分「可外包」與「不可外包」的 AI 使用情境。",{"title":156,"searchDepth":624,"depth":624,"links":2911},[],{"data":2913,"body":2914,"excerpt":-1,"toc":2940},{"title":156,"description":156},{"type":614,"children":2915},[2916,2922],{"type":617,"tag":661,"props":2917,"children":2919},{"id":2918},"開源微調效能來源聲稱未獨立驗證",[2920],{"type":622,"value":2921},"開源微調效能（來源聲稱，未獨立驗證）",{"type":617,"tag":1283,"props":2923,"children":2924},{},[2925,2930,2935],{"type":617,"tag":866,"props":2926,"children":2927},{},[2928],{"type":622,"value":2929},"私有金融資料微調 Qwen3-235B 準確率：84.7%",{"type":617,"tag":866,"props":2931,"children":2932},{},[2933],{"type":622,"value":2934},"前沿閉源模型準確率：78.2%",{"type":617,"tag":866,"props":2936,"children":2937},{},[2938],{"type":622,"value":2939},"運營成本：開源方案低 14 倍",{"title":156,"searchDepth":624,"depth":624,"links":2941},[],{"data":2943,"body":2944,"excerpt":-1,"toc":3058},{"title":156,"description":156},{"type":614,"children":2945},[2946,2952,2972,2987,2993,3022],{"type":617,"tag":661,"props":2947,"children":2949},{"id":2948},"什麼是-taste-skill",[2950],{"type":622,"value":2951},"什麼是 Taste-Skill？",{"type":617,"tag":618,"props":2953,"children":2954},{},[2955,2957,2962,2964,2970],{"type":622,"value":2956},"Taste-Skill 是一個放在專案根目錄的 ",{"type":617,"tag":678,"props":2958,"children":2959},{},[2960],{"type":622,"value":2961},"SKILL.md 設計規則檔案",{"type":622,"value":2963},"，透過 ",{"type":617,"tag":1555,"props":2965,"children":2967},{"className":2966},[],[2968],{"type":622,"value":2969},"npx skills add Leonxlnx/taste-skill",{"type":622,"value":2971}," 一行指令安裝，讓 AI coding agent 在生成前端 UI 前先「讀懂需求」，而非直接套用預設美學——即所謂 AI Slop。",{"type":617,"tag":686,"props":2973,"children":2974},{},[2975],{"type":617,"tag":618,"props":2976,"children":2977},{},[2978,2982,2985],{"type":617,"tag":678,"props":2979,"children":2980},{},[2981],{"type":622,"value":696},{"type":617,"tag":698,"props":2983,"children":2984},{},[],{"type":622,"value":2986},"\nAI Slop：指 AI 大量生成的千篇一律介面，例如 AI 紫色漸層、深色網格背景、glassmorphism，缺乏真實設計判斷力。",{"type":617,"tag":661,"props":2988,"children":2990},{"id":2989},"三旋鈕-brief-inference",[2991],{"type":622,"value":2992},"三旋鈕 + Brief Inference",{"type":617,"tag":618,"props":2994,"children":2995},{},[2996,2998,3004,3006,3012,3014,3020],{"type":622,"value":2997},"核心機制是「三旋鈕系統」：",{"type":617,"tag":1555,"props":2999,"children":3001},{"className":3000},[],[3002],{"type":622,"value":3003},"DESIGN_VARIANCE",{"type":622,"value":3005},"（設計多樣性，1–10）、",{"type":617,"tag":1555,"props":3007,"children":3009},{"className":3008},[],[3010],{"type":622,"value":3011},"MOTION_INTENSITY",{"type":622,"value":3013},"（動效強度）、",{"type":617,"tag":1555,"props":3015,"children":3017},{"className":3016},[],[3018],{"type":622,"value":3019},"VISUAL_DENSITY",{"type":622,"value":3021},"（視覺密度），供 AI agent 依 brief 自動調校。",{"type":617,"tag":618,"props":3023,"children":3024},{},[3025,3027,3032,3034,3040,3042,3048,3050,3056],{"type":622,"value":3026},"執行前強制 ",{"type":617,"tag":678,"props":3028,"children":3029},{},[3030],{"type":622,"value":3031},"Brief Inference（讀懂場景）",{"type":622,"value":3033},"：AI 必須先輸出一行 Design Read 確認理解需求，再開始生成。2026 年 3 月重構後拆為三個模組：",{"type":617,"tag":1555,"props":3035,"children":3037},{"className":3036},[],[3038],{"type":622,"value":3039},"taste-skill",{"type":622,"value":3041},"、",{"type":617,"tag":1555,"props":3043,"children":3045},{"className":3044},[],[3046],{"type":622,"value":3047},"redesign-skill",{"type":622,"value":3049},"（含 100+ 審查維度）、",{"type":617,"tag":1555,"props":3051,"children":3053},{"className":3052},[],[3054],{"type":622,"value":3055},"output-skill",{"type":622,"value":3057},"（強制輸出完整產物）。",{"title":156,"searchDepth":624,"depth":624,"links":3059},[],{"data":3061,"body":3063,"excerpt":-1,"toc":3079},{"title":156,"description":3062},"相容 Cursor、Claude Code、Codex、Gemini CLI、v0、Lovable、OpenCode，安裝只需一行指令。",{"type":614,"children":3064},[3065,3069,3074],{"type":617,"tag":618,"props":3066,"children":3067},{},[3068],{"type":622,"value":3062},{"type":617,"tag":618,"props":3070,"children":3071},{},[3072],{"type":622,"value":3073},"核心在於 Anti-Default 禁令：禁止 AI 自動套用 AI 紫色漸層、三欄等寬 feature card、Inter + slate-900 等「LLM design tell」。",{"type":617,"tag":618,"props":3075,"children":3076},{},[3077],{"type":622,"value":3078},"v2 加入 Design System Map，指引 AI 根據 brief 選用 Fluent UI、Material 3、Carbon、shadcn/ui 等官方設計系統，並誠實標注「近似實作」vs「官方元件」。",{"title":156,"searchDepth":624,"depth":624,"links":3080},[],{"data":3082,"body":3084,"excerpt":-1,"toc":3095},{"title":156,"description":3083},"57,000+ GitHub 星數代表一個新需求正在成熟：開發者不再滿足於 AI 能「跑通 UI」，而是要求它能「讀懂場景」再動筆。",{"type":614,"children":3085},[3086,3090],{"type":617,"tag":618,"props":3087,"children":3088},{},[3089],{"type":622,"value":3083},{"type":617,"tag":618,"props":3091,"children":3092},{},[3093],{"type":622,"value":3094},"Taste-Skill 的爆紅暗示前端 AI 工具正從「能用」邁向「有品味」的分水嶺——誰能最先整合類似的 design taste 框架，誰就能在 AI coding 工具市場取得差異化。",{"title":156,"searchDepth":624,"depth":624,"links":3096},[],{"data":3098,"body":3099,"excerpt":-1,"toc":3168},{"title":156,"description":156},{"type":614,"children":3100},[3101,3106,3125,3140,3145,3151,3156],{"type":617,"tag":661,"props":3102,"children":3104},{"id":3103},"法律戰火全面引爆",[3105],{"type":622,"value":3103},{"type":617,"tag":618,"props":3107,"children":3108},{},[3109,3111,3116,3118,3123],{"type":622,"value":3110},"2026 年 2 月，ByteDance 推出 ",{"type":617,"tag":678,"props":3112,"children":3113},{},[3114],{"type":622,"value":3115},"Seedance 2.0",{"type":622,"value":3117},"，一支 AI 生成的「Brad Pitt 與 Tom Cruise 屋頂對打」15 秒短片瘋傳，引爆好萊塢全面反撲。",{"type":617,"tag":678,"props":3119,"children":3120},{},[3121],{"type":622,"value":3122},"美國電影協會 (MPA) 史上首次對 AI 公司發出停止侵權信",{"type":622,"value":3124},"，Disney、Warner Bros.、Paramount、Sony Pictures、Netflix 相繼發出法律威脅。",{"type":617,"tag":686,"props":3126,"children":3127},{},[3128],{"type":617,"tag":618,"props":3129,"children":3130},{},[3131,3135,3138],{"type":617,"tag":678,"props":3132,"children":3133},{},[3134],{"type":622,"value":696},{"type":617,"tag":698,"props":3136,"children":3137},{},[],{"type":622,"value":3139},"\nMPA（美國電影協會）：代表好萊塢主要製片公司的版權保護組織，此次為首次向 AI 公司發出停止侵權令。",{"type":617,"tag":618,"props":3141,"children":3142},{},[3143],{"type":622,"value":3144},"Disney 形容此舉是對 IP 的「虛擬砸搶」，指控平台預載 Marvel、Star Wars 等盜版訓練資料；SAG-AFTRA 演員工會亦稱相關影片為「明目張膽的侵權行為」。",{"type":617,"tag":661,"props":3146,"children":3148},{"id":3147},"不問不說的矛盾內幕",[3149],{"type":622,"value":3150},"「不問不說」的矛盾內幕",{"type":617,"tag":618,"props":3152,"children":3153},{},[3154],{"type":622,"value":3155},"法律對峙之外，好萊塢內部卻截然相反。影視顧問 Peter Csathy 直言，懂 AI 的創作者普遍認為 Seedance 是目前最佳影片生成工具；《辛普森家庭》製片人 Joel Kuwahara 透露，許多製片公司雖未正式核准，卻以「不問不說」的默許態度悄悄在內部使用。",{"type":617,"tag":618,"props":3157,"children":3158},{},[3159,3161,3166],{"type":622,"value":3160},"2026 年 6 月，",{"type":617,"tag":678,"props":3162,"children":3163},{},[3164],{"type":622,"value":3165},"Seedance 2.5",{"type":622,"value":3167}," 升級為原生 30 秒影片生成，相較 Runway、Sora、Veo 優勢持續擴大。",{"title":156,"searchDepth":624,"depth":624,"links":3169},[],{"data":3171,"body":3172,"excerpt":-1,"toc":3178},{"title":156,"description":447},{"type":614,"children":3173},[3174],{"type":617,"tag":618,"props":3175,"children":3176},{},[3177],{"type":622,"value":447},{"title":156,"searchDepth":624,"depth":624,"links":3179},[],{"data":3181,"body":3182,"excerpt":-1,"toc":3188},{"title":156,"description":448},{"type":614,"children":3183},[3184],{"type":617,"tag":618,"props":3185,"children":3186},{},[3187],{"type":622,"value":448},{"title":156,"searchDepth":624,"depth":624,"links":3189},[],{"data":3191,"body":3192,"excerpt":-1,"toc":3264},{"title":156,"description":156},{"type":614,"children":3193},[3194,3200,3205,3220,3225,3248,3254,3259],{"type":617,"tag":661,"props":3195,"children":3197},{"id":3196},"discobench量化澄清提問能力的基準測試",[3198],{"type":622,"value":3199},"DiscoBench：量化澄清提問能力的基準測試",{"type":617,"tag":618,"props":3201,"children":3202},{},[3203],{"type":622,"value":3204},"騰訊混元與清華大學聯合發布 DiscoBench，評估搜尋 Agent 面對模糊查詢時主動提問澄清的能力。",{"type":617,"tag":686,"props":3206,"children":3207},{},[3208],{"type":617,"tag":618,"props":3209,"children":3210},{},[3211,3215,3218],{"type":617,"tag":678,"props":3212,"children":3213},{},[3214],{"type":622,"value":696},{"type":617,"tag":698,"props":3216,"children":3217},{},[],{"type":622,"value":3219},"\nDiscoBench 定義四類模糊類型——描述對應多個實體、多個時間段或版本、不同排名標準、事實性錯誤——並以任務效用、模糊偵測、互動策略、成本效率四個維度評估 Agent 表現。",{"type":617,"tag":618,"props":3221,"children":3222},{},[3223],{"type":622,"value":3224},"測試集涵蓋 211 項任務、463 個模糊點，橫跨 11 個知識領域。在未提供消歧提示的情況下，頂尖模型準確率均低於 50%：",{"type":617,"tag":1283,"props":3226,"children":3227},{},[3228,3233,3238,3243],{"type":617,"tag":866,"props":3229,"children":3230},{},[3231],{"type":622,"value":3232},"Doubao Seed 2.0 Pro：43.1%",{"type":617,"tag":866,"props":3234,"children":3235},{},[3236],{"type":622,"value":3237},"Gemini 3.1 Pro：40.8%",{"type":617,"tag":866,"props":3239,"children":3240},{},[3241],{"type":622,"value":3242},"Claude Opus 4.7：39.8%",{"type":617,"tag":866,"props":3244,"children":3245},{},[3246],{"type":622,"value":3247},"Qwen3.6 Max：12.3%",{"type":617,"tag":661,"props":3249,"children":3251},{"id":3250},"越搜尋越迷失先問後搜才是正解",[3252],{"type":622,"value":3253},"越搜尋越迷失，「先問後搜」才是正解",{"type":617,"tag":618,"props":3255,"children":3256},{},[3257],{"type":622,"value":3258},"研究揭露一個反直覺發現：不斷追加搜尋卻不提問，成功率 (51.9%) 甚至低於直接猜測 (56.5%) 。「SearchThenAsk（先搜尋後提問）」策略成功率高達 93.4%，差距懸殊。",{"type":617,"tag":618,"props":3260,"children":3261},{},[3262],{"type":622,"value":3263},"關鍵缺陷在於：模型能偵測到查詢有歧義，卻不懂得主動與使用者互動澄清。互補研究 InteractComp 追蹤 15 個月資料，發現搜尋能力提升七倍，互動能力卻幾乎停滯——根本原因是「系統性過度自信」。",{"title":156,"searchDepth":624,"depth":624,"links":3265},[],{"data":3267,"body":3268,"excerpt":-1,"toc":3274},{"title":156,"description":481},{"type":614,"children":3269},[3270],{"type":617,"tag":618,"props":3271,"children":3272},{},[3273],{"type":622,"value":481},{"title":156,"searchDepth":624,"depth":624,"links":3275},[],{"data":3277,"body":3278,"excerpt":-1,"toc":3284},{"title":156,"description":482},{"type":614,"children":3279},[3280],{"type":617,"tag":618,"props":3281,"children":3282},{},[3283],{"type":622,"value":482},{"title":156,"searchDepth":624,"depth":624,"links":3285},[],{"data":3287,"body":3288,"excerpt":-1,"toc":3362},{"title":156,"description":156},{"type":614,"children":3289},[3290,3296,3320,3325,3343,3349],{"type":617,"tag":661,"props":3291,"children":3293},{"id":3292},"效能基準discobench無消歧提示",[3294],{"type":622,"value":3295},"效能基準（DiscoBench，無消歧提示）",{"type":617,"tag":1283,"props":3297,"children":3298},{},[3299,3303,3307,3311,3316],{"type":617,"tag":866,"props":3300,"children":3301},{},[3302],{"type":622,"value":3232},{"type":617,"tag":866,"props":3304,"children":3305},{},[3306],{"type":622,"value":3237},{"type":617,"tag":866,"props":3308,"children":3309},{},[3310],{"type":622,"value":3242},{"type":617,"tag":866,"props":3312,"children":3313},{},[3314],{"type":622,"value":3315},"MiniMax M2.7：16.1%",{"type":617,"tag":866,"props":3317,"children":3318},{},[3319],{"type":622,"value":3247},{"type":617,"tag":661,"props":3321,"children":3323},{"id":3322},"策略成功率對比",[3324],{"type":622,"value":3322},{"type":617,"tag":1283,"props":3326,"children":3327},{},[3328,3333,3338],{"type":617,"tag":866,"props":3329,"children":3330},{},[3331],{"type":622,"value":3332},"SearchThenAsk（先搜尋後提問）：93.4%",{"type":617,"tag":866,"props":3334,"children":3335},{},[3336],{"type":622,"value":3337},"DirectGuess（直接猜測）：56.5%",{"type":617,"tag":866,"props":3339,"children":3340},{},[3341],{"type":622,"value":3342},"SearchHeavyGuess（重搜尋後猜測）：51.9%",{"type":617,"tag":661,"props":3344,"children":3346},{"id":3345},"interactcomp-最佳模型",[3347],{"type":622,"value":3348},"InteractComp 最佳模型",{"type":617,"tag":1283,"props":3350,"children":3351},{},[3352,3357],{"type":617,"tag":866,"props":3353,"children":3354},{},[3355],{"type":622,"value":3356},"無消歧上下文：13.73%",{"type":617,"tag":866,"props":3358,"children":3359},{},[3360],{"type":622,"value":3361},"完整消歧上下文：71.50%",{"title":156,"searchDepth":624,"depth":624,"links":3363},[],{"data":3365,"body":3366,"excerpt":-1,"toc":3409},{"title":156,"description":156},{"type":614,"children":3367},[3368,3374,3379,3394,3399,3404],{"type":617,"tag":661,"props":3369,"children":3371},{"id":3370},"世界模型首次學會邊跑邊學",[3372],{"type":622,"value":3373},"世界模型首次學會「邊跑邊學」",{"type":617,"tag":618,"props":3375,"children":3376},{},[3377],{"type":622,"value":3378},"2026 年 6 月，紐約大學與 LeCun 新創公司 AMI Labs 聯合發表 AdaJEPA，首次讓世界模型在部署期間持續自適應更新，打破了傳統「訓練後凍結參數」的模式。",{"type":617,"tag":686,"props":3380,"children":3381},{},[3382],{"type":617,"tag":618,"props":3383,"children":3384},{},[3385,3389,3392],{"type":617,"tag":678,"props":3386,"children":3387},{},[3388],{"type":622,"value":696},{"type":617,"tag":698,"props":3390,"children":3391},{},[],{"type":622,"value":3393},"\n世界模型 (World Model) ：AI 對環境物理規律的內部表示，讓智慧體能預測行動後果並進行規劃，而非僅對感測器輸入做出反應。",{"type":617,"tag":661,"props":3395,"children":3397},{"id":3396},"如何做到輕量化線上學習",[3398],{"type":622,"value":3396},{"type":617,"tag":618,"props":3400,"children":3401},{},[3402],{"type":622,"value":3403},"系統基於閉環 MPC(Model Predictive Control) 框架，執行「規劃→執行→觀測→梯度更新→再規劃」的完整循環。每次重規劃僅需 1 步梯度下降，額外延遲僅 0.01–0.03 秒。",{"type":617,"tag":618,"props":3405,"children":3406},{},[3407],{"type":622,"value":3408},"為防止表示空間崩塌，系統僅更新視覺編碼器與預測器的最後幾層，以真實環境的狀態轉移作為自監督學習信號，無需額外專家示範。",{"title":156,"searchDepth":624,"depth":624,"links":3410},[],{"data":3412,"body":3414,"excerpt":-1,"toc":3425},{"title":156,"description":3413},"AdaJEPA 的在線校準策略值得深入研究：僅更新最後幾層的保守做法，有效避免了災難性遺忘與表示崩塌，可應用於機器人、自駕等部署後環境持續變化的場景。",{"type":614,"children":3415},[3416,3420],{"type":617,"tag":618,"props":3417,"children":3418},{},[3419],{"type":622,"value":3413},{"type":617,"tag":618,"props":3421,"children":3422},{},[3423],{"type":622,"value":3424},"單步梯度下降的設計與 LoRA 類似——最小化更新範圍以換取穩定性——實際整合前需評估重規劃頻率對能耗的累積影響。",{"title":156,"searchDepth":624,"depth":624,"links":3426},[],{"data":3428,"body":3430,"excerpt":-1,"toc":3441},{"title":156,"description":3429},"世界模型能在線學習，代表 AI 系統的維護模式即將轉變：不再需要定期重訓並重新部署，而是讓模型在真實互動中自我校準。",{"type":614,"children":3431},[3432,3436],{"type":617,"tag":618,"props":3433,"children":3434},{},[3435],{"type":622,"value":3429},{"type":617,"tag":618,"props":3437,"children":3438},{},[3439],{"type":622,"value":3440},"對機器人、自駕與工業自動化廠商而言，這降低了落地後的迭代成本。LeCun 的 JEPA 路線持續獲得研究支撐，AMI Labs 的商業化前景值得追蹤。",{"title":156,"searchDepth":624,"depth":624,"links":3442},[],{"data":3444,"body":3445,"excerpt":-1,"toc":3465},{"title":156,"description":156},{"type":614,"children":3446},[3447,3452],{"type":617,"tag":661,"props":3448,"children":3450},{"id":3449},"效能基準",[3451],{"type":622,"value":3449},{"type":617,"tag":1283,"props":3453,"children":3454},{},[3455,3460],{"type":617,"tag":866,"props":3456,"children":3457},{},[3458],{"type":622,"value":3459},"PointMaze 未知布局測試：梯度下降成功率 53.3% → 78.7%",{"type":617,"tag":866,"props":3461,"children":3462},{},[3463],{"type":622,"value":3464},"未見物體目標導航測試：規劃成功率提升接近一倍",{"title":156,"searchDepth":624,"depth":624,"links":3466},[],{"data":3468,"body":3469,"excerpt":-1,"toc":3542},{"title":156,"description":156},{"type":614,"children":3470},[3471,3477,3489,3494,3527,3532,3537],{"type":617,"tag":661,"props":3472,"children":3474},{"id":3473},"meta-compute-三層服務",[3475],{"type":622,"value":3476},"Meta Compute 三層服務",{"type":617,"tag":618,"props":3478,"children":3479},{},[3480,3482,3487],{"type":622,"value":3481},"Meta 正籌備名為 ",{"type":617,"tag":678,"props":3483,"children":3484},{},[3485],{"type":622,"value":3486},"Meta Compute",{"type":622,"value":3488}," 的雲端服務，計劃對外販售 GPU 算力與模型 API，直接挑戰 AWS、Azure、Google Cloud 及 CoreWeave。消息曝光後，Meta 股價單日飆升 8.8%，CoreWeave 暴跌 13.92%，Nebius 跌 17.01%。",{"type":617,"tag":618,"props":3490,"children":3491},{},[3492],{"type":622,"value":3493},"服務規劃分三層：",{"type":617,"tag":862,"props":3495,"children":3496},{},[3497,3507,3517],{"type":617,"tag":866,"props":3498,"children":3499},{},[3500,3505],{"type":617,"tag":678,"props":3501,"children":3502},{},[3503],{"type":622,"value":3504},"裸機 GPU 實例",{"type":622,"value":3506},"：H100 專屬物理伺服器",{"type":617,"tag":866,"props":3508,"children":3509},{},[3510,3515],{"type":617,"tag":678,"props":3511,"children":3512},{},[3513],{"type":622,"value":3514},"受管訓練集群",{"type":622,"value":3516},"：自動化分散式訓練編排",{"type":617,"tag":866,"props":3518,"children":3519},{},[3520,3525],{"type":617,"tag":678,"props":3521,"children":3522},{},[3523],{"type":622,"value":3524},"托管模型 API",{"type":622,"value":3526},"：Llama 系列與第三方模型（含 Anthropic Claude）REST 端點",{"type":617,"tag":661,"props":3528,"children":3530},{"id":3529},"定價策略與背景",[3531],{"type":622,"value":3529},{"type":617,"tag":618,"props":3533,"children":3534},{},[3535],{"type":622,"value":3536},"Meta 計劃定價約 $1.50/GPU 小時，比市場主流 $2.50–3.00 低 30–40%，並採用 90 天可取消彈性合約，降低客戶遷移門檻。",{"type":617,"tag":618,"props":3538,"children":3539},{},[3540],{"type":622,"value":3541},"背景矛盾之處在於：市場算力價格正在上漲——H100 合約價五個月漲 38.2%，Nvidia B200 租賃價預計 10 月漲幅達 94%。Meta 能否長期維持低價，取決於其 5GW 數據中心規模帶來的成本優勢。",{"title":156,"searchDepth":624,"depth":624,"links":3543},[],{"data":3545,"body":3547,"excerpt":-1,"toc":3558},{"title":156,"description":3546},"對開發者而言，Meta Compute 帶來第四個主流雲算力選項。$1.50/GPU 小時若成真，相當於 H100 訓練成本直接打六折。90 天可取消合約降低供應商鎖定風險，對需要彈性擴縮的訓練任務尤其友好。",{"type":614,"children":3548},[3549,3553],{"type":617,"tag":618,"props":3550,"children":3551},{},[3552],{"type":622,"value":3546},{"type":617,"tag":618,"props":3554,"children":3555},{},[3556],{"type":622,"value":3557},"若 Claude 確實納入托管範圍，代表單一平台可同時呼叫開源與閉源模型，省去多平台管理成本——但服務何時上線仍未公布，目前僅能觀望。",{"title":156,"searchDepth":624,"depth":624,"links":3559},[],{"data":3561,"body":3563,"excerpt":-1,"toc":3574},{"title":156,"description":3562},"Meta 入場最直接的受害者是獨立算力雲業態——CoreWeave 和 Nebius 單日跌幅已說明市場判斷。Wells Fargo 估算此業務 2028 年可帶來年化 2,640 億美元收入，若成真將顛覆算力市場格局。",{"type":614,"children":3564},[3565,3569],{"type":617,"tag":618,"props":3566,"children":3567},{},[3568],{"type":622,"value":3562},{"type":617,"tag":618,"props":3570,"children":3571},{},[3572],{"type":622,"value":3573},"關鍵風險在於動機：Zuckerberg 自承 AI agent 開發不如預期，賣算力更像是龐大資本支出的應急出路。一旦 Meta 內部 AI 需求回升，外部算力供應的優先度將直接受壓。",{"title":156,"searchDepth":624,"depth":624,"links":3575},[],{"data":3577,"body":3578,"excerpt":-1,"toc":3613},{"title":156,"description":156},{"type":614,"children":3579},[3580,3585],{"type":617,"tag":661,"props":3581,"children":3583},{"id":3582},"定價對比",[3584],{"type":622,"value":3582},{"type":617,"tag":1283,"props":3586,"children":3587},{},[3588,3593,3598,3603,3608],{"type":617,"tag":866,"props":3589,"children":3590},{},[3591],{"type":622,"value":3592},"Meta Compute（預計）：$1.50/GPU 小時 (H100)",{"type":617,"tag":866,"props":3594,"children":3595},{},[3596],{"type":622,"value":3597},"市場主流：$2.50–3.00/GPU 小時",{"type":617,"tag":866,"props":3599,"children":3600},{},[3601],{"type":622,"value":3602},"Nvidia B200 租賃（2026 年 10 月預測）：$5.10／小時 (+94%)",{"type":617,"tag":866,"props":3604,"children":3605},{},[3606],{"type":622,"value":3607},"H100 合約價五個月漲幅：+38.2%",{"type":617,"tag":866,"props":3609,"children":3610},{},[3611],{"type":622,"value":3612},"AWS EC2 7 月調漲：約 +20%",{"title":156,"searchDepth":624,"depth":624,"links":3614},[],{"data":3616,"body":3617,"excerpt":-1,"toc":3667},{"title":156,"description":156},{"type":614,"children":3618},[3619,3625,3637,3652,3657,3662],{"type":617,"tag":661,"props":3620,"children":3622},{"id":3621},"研究結果接近人類一對一家教效益",[3623],{"type":622,"value":3624},"研究結果：接近人類一對一家教效益",{"type":617,"tag":618,"props":3626,"children":3627},{},[3628,3630,3635],{"type":622,"value":3629},"Dartmouth College 於 2026 年春季學期，在入門統計課程 MATH 010 部署 Phosphor AI 輔導平台，共 143 名學生完成實驗。完整使用平台的學生，期末考成績提升幅度介於 ",{"type":617,"tag":678,"props":3631,"children":3632},{},[3633],{"type":622,"value":3634},"0.71 SD（控制期中成績後）至 1.30 SD（未調整）",{"type":622,"value":3636},"，接近 Bloom 1984 年「2 Sigma 問題」中人類一對一家教的效益水準。",{"type":617,"tag":686,"props":3638,"children":3639},{},[3640],{"type":617,"tag":618,"props":3641,"children":3642},{},[3643,3647,3650],{"type":617,"tag":678,"props":3644,"children":3645},{},[3646],{"type":622,"value":696},{"type":617,"tag":698,"props":3648,"children":3649},{},[],{"type":622,"value":3651},"\n「2 Sigma 問題」源自 Bloom 1984 年研究：一對一人類家教可讓平均生進步約 2 個標準差，超越 98% 傳統班級教學學生，但規模化成本極高，故被視為教育科技的終極目標。",{"type":617,"tag":661,"props":3653,"children":3655},{"id":3654},"方法論爭議不可忽視",[3656],{"type":622,"value":3654},{"type":617,"tag":618,"props":3658,"children":3659},{},[3660],{"type":622,"value":3661},"Phosphor 的核心功能是練習題測驗加 AI 自動批改，附帶 RAG 聊天助理但實際使用率極低。90% 的學生至少使用過一次，但達到「完全投入」門檻的僅約 16 人（約 11%）。",{"type":617,"tag":618,"props":3663,"children":3664},{},[3665],{"type":622,"value":3666},"研究未設置隨機對照組，且研究者在分析期間將題型從多選題改為 AI 批改主觀題——社群批評此舉造成「實驗設計被自身結果反向決定」的因果倒置問題。完整投入的 16 名學生也可能本來就是高動機學習者，選擇效應難以排除。",{"title":156,"searchDepth":624,"depth":624,"links":3668},[],{"data":3670,"body":3672,"excerpt":-1,"toc":3686},{"title":156,"description":3671},"Phosphor 靠「練習題 + 即時 AI 批改」驅動學習，RAG 對話功能幾乎無人使用——對 AI 教育工具開發者而言，自動批改反饋迴路的實效遠高於對話 UI。11% 的完整投入率顯示產品留存設計是主要瓶頸；如何將 90% 的初次使用者轉化為持續投入的學習者，才是真正的工程挑戰。",{"type":614,"children":3673},[3674],{"type":617,"tag":618,"props":3675,"children":3676},{},[3677,3679,3684],{"type":622,"value":3678},"Phosphor 靠「練習題 + 即時 AI 批改」驅動學習，RAG 對話功能幾乎無人使用——對 AI 教育工具開發者而言，",{"type":617,"tag":678,"props":3680,"children":3681},{},[3682],{"type":622,"value":3683},"自動批改反饋迴路的實效遠高於對話 UI",{"type":622,"value":3685},"。11% 的完整投入率顯示產品留存設計是主要瓶頸；如何將 90% 的初次使用者轉化為持續投入的學習者，才是真正的工程挑戰。",{"title":156,"searchDepth":624,"depth":624,"links":3687},[],{"data":3689,"body":3690,"excerpt":-1,"toc":3696},{"title":156,"description":569},{"type":614,"children":3691},[3692],{"type":617,"tag":618,"props":3693,"children":3694},{},[3695],{"type":622,"value":569},{"title":156,"searchDepth":624,"depth":624,"links":3697},[],{"data":3699,"body":3700,"excerpt":-1,"toc":3729},{"title":156,"description":156},{"type":614,"children":3701},[3702,3706],{"type":617,"tag":661,"props":3703,"children":3704},{"id":3449},[3705],{"type":622,"value":3449},{"type":617,"tag":1283,"props":3707,"children":3708},{},[3709,3714,3719,3724],{"type":617,"tag":866,"props":3710,"children":3711},{},[3712],{"type":622,"value":3713},"完整投入組期末考提升：0.71 SD（控制期中成績）至 1.30 SD（未調整）",{"type":617,"tag":866,"props":3715,"children":3716},{},[3717],{"type":622,"value":3718},"換算實際分數：24 分滿分中多約 3 分（0.71 SD 口徑）",{"type":617,"tag":866,"props":3720,"children":3721},{},[3722],{"type":622,"value":3723},"完整投入率：約 11%（16 / 143 名學生）",{"type":617,"tag":866,"props":3725,"children":3726},{},[3727],{"type":622,"value":3728},"平台使用率：90% 至少使用一次",{"title":156,"searchDepth":624,"depth":624,"links":3730},[],{"data":3732,"body":3733,"excerpt":-1,"toc":3801},{"title":156,"description":156},{"type":614,"children":3734},[3735,3740,3745,3750,3755,3760,3765,3770,3776,3781,3786,3791,3796],{"type":617,"tag":661,"props":3736,"children":3738},{"id":3737},"社群熱議排行",[3739],{"type":622,"value":3737},{"type":617,"tag":618,"props":3741,"children":3742},{},[3743],{"type":622,"value":3744},"本日熱度最高為 EU Chat Control 快速通道表決。Patrick Breyer（Bluesky， 29 upvotes）緊急呼籲阻止大規模監控，@paddi_hansen 宣告歐洲議會接連否決延長豁免條款，隱私倡議者連取勝利。",{"type":617,"tag":618,"props":3746,"children":3747},{},[3748],{"type":622,"value":3749},"百度 Unlimited OCR 緊接其後，HN + X 多則討論關注其「記憶體零增長、讀完整本書」的突破。Fable 5 iOS 移植 C&C 遊戲在 Bluesky 獲 135 upvotes，引發下一代 AI 編碼能力的熱烈討論。",{"type":617,"tag":618,"props":3751,"children":3752},{},[3753],{"type":622,"value":3754},"Amazon MTurk 退場公告（socialmedialab.ca， 6 upvotes， Bluesky）引爆一代 AI 訓練基礎設施的告別情緒。",{"type":617,"tag":661,"props":3756,"children":3758},{"id":3757},"技術爭議與分歧",[3759],{"type":622,"value":3757},{"type":617,"tag":618,"props":3761,"children":3762},{},[3763],{"type":622,"value":3764},"Chat Control 最大爭議在術語認知：SpicyLemonZest(HN) 指出「『Chat Control』這個詞是反對私訊掃描的隱私倡議團體所創造，民調普遍支持 CSAM 掃描」。",{"type":617,"tag":618,"props":3766,"children":3767},{},[3768],{"type":622,"value":3769},"@mullvadnet(X) 則主張理事會文本仍保留後門空間，兩方對「勝利」的定義根本分歧。GPT-5.5 陣營，lmwnshn(HN) 推廣 MIT 授權 clankerbend：「有狀況可以隨時換船」，呼籲不被閉源版本決策綁死。",{"type":617,"tag":661,"props":3771,"children":3773},{"id":3772},"實戰經驗最高價值",[3774],{"type":622,"value":3775},"實戰經驗（最高價值）",{"type":617,"tag":618,"props":3777,"children":3778},{},[3779],{"type":622,"value":3780},"zuzululu(HN) 實測 GPT-5.5 Codex reasoning token 修復：「我剛試了，成功了！不知道為什麼有效，但感覺和一個半月前一樣了。」",{"type":617,"tag":618,"props":3782,"children":3783},{},[3784],{"type":622,"value":3785},"HN 用戶 potsandpans 實測百度 Unlimited OCR：「可以在 12GB VRAM 上舒適運行，確實很有競爭力。」@mxcl（Homebrew 創作者， X）確認 Taste-Skill：「用起來非常好，在 native 和 web UI 上都通用。」",{"type":617,"tag":661,"props":3787,"children":3789},{"id":3788},"未解問題與社群預期",[3790],{"type":622,"value":3788},{"type":617,"tag":618,"props":3792,"children":3793},{},[3794],{"type":622,"value":3795},"7 月 8 日 Chat Control 快速通道表決結果未知；義大利「警告反對卻投票支持」的矛盾顯示政治利益計算超越公開立場，社群對連勝能否延續高度存疑。",{"type":617,"tag":618,"props":3797,"children":3798},{},[3799],{"type":622,"value":3800},"GPT-5.5 reasoning token 透明度問題仍無官方回應，516-token clustering 究竟是設計還是 bug，開發者無從驗證。MTurk 退場後，社群追問 2022 年後的眾包標注資料集是否已遭 LLM 代工污染，訓練出的模型品質如何驗證？",{"title":156,"searchDepth":624,"depth":624,"links":3802},[],{"data":3804,"body":3806,"excerpt":-1,"toc":3822},{"title":156,"description":3805},"今日社群橫跨三個維度見證了 AI 生態的世代交替：歐盟加密政策進入最後衝刺、Amazon MTurk 退場讓資料基礎設施面臨真空、GPT-5.5 reasoning token 爭議提醒開發者，閉源模型的「升級」未必是進步。",{"type":614,"children":3807},[3808,3812,3817],{"type":617,"tag":618,"props":3809,"children":3810},{},[3811],{"type":622,"value":3805},{"type":617,"tag":618,"props":3813,"children":3814},{},[3815],{"type":622,"value":3816},"與此同時，開源工具陣營正用行動回應：百度 Unlimited OCR 打破長文件記憶體天花板、Taste-Skill 試圖從根本解決 AI 生成設計的同質化、Meta 也悄悄準備加入算力市場競爭。",{"type":617,"tag":618,"props":3818,"children":3819},{},[3820],{"type":622,"value":3821},"社群共識漸趨清晰——真正的護城河在於可替換的開放架構，而非依賴任何單一供應商的黑盒版本管理。",{"title":156,"searchDepth":624,"depth":624,"links":3823},[],{"data":3825,"body":3826,"excerpt":-1,"toc":4892},{"title":156,"description":156},{"type":614,"children":3827},[3828,3832,3845,3851,4818,4822,4827,4831,4856,4860,4886],{"type":617,"tag":661,"props":3829,"children":3830},{"id":1225},[3831],{"type":622,"value":1225},{"type":617,"tag":618,"props":3833,"children":3834},{},[3835,3837,3843],{"type":622,"value":3836},"需要一個能記錄每次 API 回應 reasoning token 數量的日誌系統。建議使用 OpenAI 的 usage 物件中的 ",{"type":617,"tag":1555,"props":3838,"children":3840},{"className":3839},[],[3841],{"type":622,"value":3842},"completion_tokens_details.reasoning_tokens",{"type":622,"value":3844}," 欄位，搭配 SQLite 或時序資料庫儲存歷史分佈，以便計算 clustering 比率。",{"type":617,"tag":661,"props":3846,"children":3848},{"id":3847},"最小-poc",[3849],{"type":622,"value":3850},"最小 PoC",{"type":617,"tag":3852,"props":3853,"children":3857},"pre",{"className":3854,"code":3855,"language":3856,"meta":156,"style":156},"language-python shiki shiki-themes vitesse-dark","import openai, sqlite3, collections, time\n\nclient = openai.OpenAI()\ndb = sqlite3.connect(\"reasoning_log.db\")\ndb.execute(\"CREATE TABLE IF NOT EXISTS logs (ts REAL, model TEXT, reasoning_tokens INT)\")\n\ndef call_and_log(prompt: str, model: str = \"gpt-5.5\"):\n    resp = client.chat.completions.create(\n        model=model,\n        messages=[{\"role\": \"user\", \"content\": prompt}],\n    )\n    r_tokens = getattr(resp.usage.completion_tokens_details, \"reasoning_tokens\", 0)\n    db.execute(\"INSERT INTO logs VALUES (?, ?, ?)\", (time.time(), model, r_tokens))\n    db.commit()\n    return resp, r_tokens\n\ndef clustering_rate(model: str = \"gpt-5.5\", window: int = 100) -> float:\n    rows = db.execute(\n        \"SELECT reasoning_tokens FROM logs WHERE model=? ORDER BY ts DESC LIMIT ?\",\n        (model, window)\n    ).fetchall()\n    counter = collections.Counter(r[0] for r in rows)\n    return counter[516] / len(rows) if rows else 0.0\n","python",[3858],{"type":617,"tag":1555,"props":3859,"children":3860},{"__ignoreMap":156},[3861,3908,3917,3949,4000,4038,4046,4123,4173,4197,4272,4281,4357,4436,4457,4480,4488,4580,4610,4632,4657,4675,4749],{"type":617,"tag":3862,"props":3863,"children":3866},"span",{"class":3864,"line":3865},"line",1,[3867,3873,3879,3885,3890,3894,3899,3903],{"type":617,"tag":3862,"props":3868,"children":3870},{"style":3869},"--shiki-default:#4D9375",[3871],{"type":622,"value":3872},"import",{"type":617,"tag":3862,"props":3874,"children":3876},{"style":3875},"--shiki-default:#DBD7CAEE",[3877],{"type":622,"value":3878}," openai",{"type":617,"tag":3862,"props":3880,"children":3882},{"style":3881},"--shiki-default:#666666",[3883],{"type":622,"value":3884},",",{"type":617,"tag":3862,"props":3886,"children":3887},{"style":3875},[3888],{"type":622,"value":3889}," sqlite3",{"type":617,"tag":3862,"props":3891,"children":3892},{"style":3881},[3893],{"type":622,"value":3884},{"type":617,"tag":3862,"props":3895,"children":3896},{"style":3875},[3897],{"type":622,"value":3898}," collections",{"type":617,"tag":3862,"props":3900,"children":3901},{"style":3881},[3902],{"type":622,"value":3884},{"type":617,"tag":3862,"props":3904,"children":3905},{"style":3875},[3906],{"type":622,"value":3907}," time\n",{"type":617,"tag":3862,"props":3909,"children":3910},{"class":3864,"line":624},[3911],{"type":617,"tag":3862,"props":3912,"children":3914},{"emptyLinePlaceholder":3913},true,[3915],{"type":622,"value":3916},"\n",{"type":617,"tag":3862,"props":3918,"children":3919},{"class":3864,"line":185},[3920,3925,3930,3934,3939,3944],{"type":617,"tag":3862,"props":3921,"children":3922},{"style":3875},[3923],{"type":622,"value":3924},"client ",{"type":617,"tag":3862,"props":3926,"children":3927},{"style":3881},[3928],{"type":622,"value":3929},"=",{"type":617,"tag":3862,"props":3931,"children":3932},{"style":3875},[3933],{"type":622,"value":3878},{"type":617,"tag":3862,"props":3935,"children":3936},{"style":3881},[3937],{"type":622,"value":3938},".",{"type":617,"tag":3862,"props":3940,"children":3941},{"style":3875},[3942],{"type":622,"value":3943},"OpenAI",{"type":617,"tag":3862,"props":3945,"children":3946},{"style":3881},[3947],{"type":622,"value":3948},"()\n",{"type":617,"tag":3862,"props":3950,"children":3951},{"class":3864,"line":75},[3952,3957,3961,3965,3969,3974,3979,3985,3991,3995],{"type":617,"tag":3862,"props":3953,"children":3954},{"style":3875},[3955],{"type":622,"value":3956},"db ",{"type":617,"tag":3862,"props":3958,"children":3959},{"style":3881},[3960],{"type":622,"value":3929},{"type":617,"tag":3862,"props":3962,"children":3963},{"style":3875},[3964],{"type":622,"value":3889},{"type":617,"tag":3862,"props":3966,"children":3967},{"style":3881},[3968],{"type":622,"value":3938},{"type":617,"tag":3862,"props":3970,"children":3971},{"style":3875},[3972],{"type":622,"value":3973},"connect",{"type":617,"tag":3862,"props":3975,"children":3976},{"style":3881},[3977],{"type":622,"value":3978},"(",{"type":617,"tag":3862,"props":3980,"children":3982},{"style":3981},"--shiki-default:#C98A7D77",[3983],{"type":622,"value":3984},"\"",{"type":617,"tag":3862,"props":3986,"children":3988},{"style":3987},"--shiki-default:#C98A7D",[3989],{"type":622,"value":3990},"reasoning_log.db",{"type":617,"tag":3862,"props":3992,"children":3993},{"style":3981},[3994],{"type":622,"value":3984},{"type":617,"tag":3862,"props":3996,"children":3997},{"style":3881},[3998],{"type":622,"value":3999},")\n",{"type":617,"tag":3862,"props":4001,"children":4002},{"class":3864,"line":76},[4003,4008,4012,4017,4021,4025,4030,4034],{"type":617,"tag":3862,"props":4004,"children":4005},{"style":3875},[4006],{"type":622,"value":4007},"db",{"type":617,"tag":3862,"props":4009,"children":4010},{"style":3881},[4011],{"type":622,"value":3938},{"type":617,"tag":3862,"props":4013,"children":4014},{"style":3875},[4015],{"type":622,"value":4016},"execute",{"type":617,"tag":3862,"props":4018,"children":4019},{"style":3881},[4020],{"type":622,"value":3978},{"type":617,"tag":3862,"props":4022,"children":4023},{"style":3981},[4024],{"type":622,"value":3984},{"type":617,"tag":3862,"props":4026,"children":4027},{"style":3987},[4028],{"type":622,"value":4029},"CREATE TABLE IF NOT EXISTS logs (ts REAL, model TEXT, reasoning_tokens INT)",{"type":617,"tag":3862,"props":4031,"children":4032},{"style":3981},[4033],{"type":622,"value":3984},{"type":617,"tag":3862,"props":4035,"children":4036},{"style":3881},[4037],{"type":622,"value":3999},{"type":617,"tag":3862,"props":4039,"children":4041},{"class":3864,"line":4040},6,[4042],{"type":617,"tag":3862,"props":4043,"children":4044},{"emptyLinePlaceholder":3913},[4045],{"type":622,"value":3916},{"type":617,"tag":3862,"props":4047,"children":4049},{"class":3864,"line":4048},7,[4050,4056,4062,4066,4071,4076,4082,4086,4091,4095,4099,4104,4109,4114,4118],{"type":617,"tag":3862,"props":4051,"children":4053},{"style":4052},"--shiki-default:#CB7676",[4054],{"type":622,"value":4055},"def",{"type":617,"tag":3862,"props":4057,"children":4059},{"style":4058},"--shiki-default:#80A665",[4060],{"type":622,"value":4061}," call_and_log",{"type":617,"tag":3862,"props":4063,"children":4064},{"style":3881},[4065],{"type":622,"value":3978},{"type":617,"tag":3862,"props":4067,"children":4068},{"style":3875},[4069],{"type":622,"value":4070},"prompt",{"type":617,"tag":3862,"props":4072,"children":4073},{"style":3881},[4074],{"type":622,"value":4075},":",{"type":617,"tag":3862,"props":4077,"children":4079},{"style":4078},"--shiki-default:#B8A965",[4080],{"type":622,"value":4081}," str",{"type":617,"tag":3862,"props":4083,"children":4084},{"style":3881},[4085],{"type":622,"value":3884},{"type":617,"tag":3862,"props":4087,"children":4088},{"style":3875},[4089],{"type":622,"value":4090}," model",{"type":617,"tag":3862,"props":4092,"children":4093},{"style":3881},[4094],{"type":622,"value":4075},{"type":617,"tag":3862,"props":4096,"children":4097},{"style":4078},[4098],{"type":622,"value":4081},{"type":617,"tag":3862,"props":4100,"children":4101},{"style":3881},[4102],{"type":622,"value":4103}," =",{"type":617,"tag":3862,"props":4105,"children":4106},{"style":3981},[4107],{"type":622,"value":4108}," \"",{"type":617,"tag":3862,"props":4110,"children":4111},{"style":3987},[4112],{"type":622,"value":4113},"gpt-5.5",{"type":617,"tag":3862,"props":4115,"children":4116},{"style":3981},[4117],{"type":622,"value":3984},{"type":617,"tag":3862,"props":4119,"children":4120},{"style":3881},[4121],{"type":622,"value":4122},"):\n",{"type":617,"tag":3862,"props":4124,"children":4126},{"class":3864,"line":4125},8,[4127,4132,4136,4141,4145,4150,4154,4159,4163,4168],{"type":617,"tag":3862,"props":4128,"children":4129},{"style":3875},[4130],{"type":622,"value":4131},"    resp ",{"type":617,"tag":3862,"props":4133,"children":4134},{"style":3881},[4135],{"type":622,"value":3929},{"type":617,"tag":3862,"props":4137,"children":4138},{"style":3875},[4139],{"type":622,"value":4140}," client",{"type":617,"tag":3862,"props":4142,"children":4143},{"style":3881},[4144],{"type":622,"value":3938},{"type":617,"tag":3862,"props":4146,"children":4147},{"style":3875},[4148],{"type":622,"value":4149},"chat",{"type":617,"tag":3862,"props":4151,"children":4152},{"style":3881},[4153],{"type":622,"value":3938},{"type":617,"tag":3862,"props":4155,"children":4156},{"style":3875},[4157],{"type":622,"value":4158},"completions",{"type":617,"tag":3862,"props":4160,"children":4161},{"style":3881},[4162],{"type":622,"value":3938},{"type":617,"tag":3862,"props":4164,"children":4165},{"style":3875},[4166],{"type":622,"value":4167},"create",{"type":617,"tag":3862,"props":4169,"children":4170},{"style":3881},[4171],{"type":622,"value":4172},"(\n",{"type":617,"tag":3862,"props":4174,"children":4176},{"class":3864,"line":4175},9,[4177,4183,4187,4192],{"type":617,"tag":3862,"props":4178,"children":4180},{"style":4179},"--shiki-default:#BD976A",[4181],{"type":622,"value":4182},"        model",{"type":617,"tag":3862,"props":4184,"children":4185},{"style":3881},[4186],{"type":622,"value":3929},{"type":617,"tag":3862,"props":4188,"children":4189},{"style":3875},[4190],{"type":622,"value":4191},"model",{"type":617,"tag":3862,"props":4193,"children":4194},{"style":3881},[4195],{"type":622,"value":4196},",\n",{"type":617,"tag":3862,"props":4198,"children":4200},{"class":3864,"line":4199},10,[4201,4206,4211,4215,4220,4224,4228,4232,4237,4241,4245,4249,4254,4258,4262,4267],{"type":617,"tag":3862,"props":4202,"children":4203},{"style":4179},[4204],{"type":622,"value":4205},"        messages",{"type":617,"tag":3862,"props":4207,"children":4208},{"style":3881},[4209],{"type":622,"value":4210},"=[{",{"type":617,"tag":3862,"props":4212,"children":4213},{"style":3981},[4214],{"type":622,"value":3984},{"type":617,"tag":3862,"props":4216,"children":4217},{"style":3987},[4218],{"type":622,"value":4219},"role",{"type":617,"tag":3862,"props":4221,"children":4222},{"style":3981},[4223],{"type":622,"value":3984},{"type":617,"tag":3862,"props":4225,"children":4226},{"style":3881},[4227],{"type":622,"value":4075},{"type":617,"tag":3862,"props":4229,"children":4230},{"style":3981},[4231],{"type":622,"value":4108},{"type":617,"tag":3862,"props":4233,"children":4234},{"style":3987},[4235],{"type":622,"value":4236},"user",{"type":617,"tag":3862,"props":4238,"children":4239},{"style":3981},[4240],{"type":622,"value":3984},{"type":617,"tag":3862,"props":4242,"children":4243},{"style":3881},[4244],{"type":622,"value":3884},{"type":617,"tag":3862,"props":4246,"children":4247},{"style":3981},[4248],{"type":622,"value":4108},{"type":617,"tag":3862,"props":4250,"children":4251},{"style":3987},[4252],{"type":622,"value":4253},"content",{"type":617,"tag":3862,"props":4255,"children":4256},{"style":3981},[4257],{"type":622,"value":3984},{"type":617,"tag":3862,"props":4259,"children":4260},{"style":3881},[4261],{"type":622,"value":4075},{"type":617,"tag":3862,"props":4263,"children":4264},{"style":3875},[4265],{"type":622,"value":4266}," prompt",{"type":617,"tag":3862,"props":4268,"children":4269},{"style":3881},[4270],{"type":622,"value":4271},"}],\n",{"type":617,"tag":3862,"props":4273,"children":4275},{"class":3864,"line":4274},11,[4276],{"type":617,"tag":3862,"props":4277,"children":4278},{"style":3881},[4279],{"type":622,"value":4280},"    )\n",{"type":617,"tag":3862,"props":4282,"children":4284},{"class":3864,"line":4283},12,[4285,4290,4294,4299,4303,4308,4312,4317,4321,4326,4330,4334,4339,4343,4347,4353],{"type":617,"tag":3862,"props":4286,"children":4287},{"style":3875},[4288],{"type":622,"value":4289},"    r_tokens ",{"type":617,"tag":3862,"props":4291,"children":4292},{"style":3881},[4293],{"type":622,"value":3929},{"type":617,"tag":3862,"props":4295,"children":4296},{"style":4078},[4297],{"type":622,"value":4298}," getattr",{"type":617,"tag":3862,"props":4300,"children":4301},{"style":3881},[4302],{"type":622,"value":3978},{"type":617,"tag":3862,"props":4304,"children":4305},{"style":3875},[4306],{"type":622,"value":4307},"resp",{"type":617,"tag":3862,"props":4309,"children":4310},{"style":3881},[4311],{"type":622,"value":3938},{"type":617,"tag":3862,"props":4313,"children":4314},{"style":3875},[4315],{"type":622,"value":4316},"usage",{"type":617,"tag":3862,"props":4318,"children":4319},{"style":3881},[4320],{"type":622,"value":3938},{"type":617,"tag":3862,"props":4322,"children":4323},{"style":3875},[4324],{"type":622,"value":4325},"completion_tokens_details",{"type":617,"tag":3862,"props":4327,"children":4328},{"style":3881},[4329],{"type":622,"value":3884},{"type":617,"tag":3862,"props":4331,"children":4332},{"style":3981},[4333],{"type":622,"value":4108},{"type":617,"tag":3862,"props":4335,"children":4336},{"style":3987},[4337],{"type":622,"value":4338},"reasoning_tokens",{"type":617,"tag":3862,"props":4340,"children":4341},{"style":3981},[4342],{"type":622,"value":3984},{"type":617,"tag":3862,"props":4344,"children":4345},{"style":3881},[4346],{"type":622,"value":3884},{"type":617,"tag":3862,"props":4348,"children":4350},{"style":4349},"--shiki-default:#4C9A91",[4351],{"type":622,"value":4352}," 0",{"type":617,"tag":3862,"props":4354,"children":4355},{"style":3881},[4356],{"type":622,"value":3999},{"type":617,"tag":3862,"props":4358,"children":4360},{"class":3864,"line":4359},13,[4361,4366,4370,4374,4378,4382,4387,4391,4395,4400,4405,4409,4413,4418,4422,4426,4431],{"type":617,"tag":3862,"props":4362,"children":4363},{"style":3875},[4364],{"type":622,"value":4365},"    db",{"type":617,"tag":3862,"props":4367,"children":4368},{"style":3881},[4369],{"type":622,"value":3938},{"type":617,"tag":3862,"props":4371,"children":4372},{"style":3875},[4373],{"type":622,"value":4016},{"type":617,"tag":3862,"props":4375,"children":4376},{"style":3881},[4377],{"type":622,"value":3978},{"type":617,"tag":3862,"props":4379,"children":4380},{"style":3981},[4381],{"type":622,"value":3984},{"type":617,"tag":3862,"props":4383,"children":4384},{"style":3987},[4385],{"type":622,"value":4386},"INSERT INTO logs VALUES (?, ?, ?)",{"type":617,"tag":3862,"props":4388,"children":4389},{"style":3981},[4390],{"type":622,"value":3984},{"type":617,"tag":3862,"props":4392,"children":4393},{"style":3881},[4394],{"type":622,"value":3884},{"type":617,"tag":3862,"props":4396,"children":4397},{"style":3881},[4398],{"type":622,"value":4399}," (",{"type":617,"tag":3862,"props":4401,"children":4402},{"style":3875},[4403],{"type":622,"value":4404},"time",{"type":617,"tag":3862,"props":4406,"children":4407},{"style":3881},[4408],{"type":622,"value":3938},{"type":617,"tag":3862,"props":4410,"children":4411},{"style":3875},[4412],{"type":622,"value":4404},{"type":617,"tag":3862,"props":4414,"children":4415},{"style":3881},[4416],{"type":622,"value":4417},"(),",{"type":617,"tag":3862,"props":4419,"children":4420},{"style":3875},[4421],{"type":622,"value":4090},{"type":617,"tag":3862,"props":4423,"children":4424},{"style":3881},[4425],{"type":622,"value":3884},{"type":617,"tag":3862,"props":4427,"children":4428},{"style":3875},[4429],{"type":622,"value":4430}," r_tokens",{"type":617,"tag":3862,"props":4432,"children":4433},{"style":3881},[4434],{"type":622,"value":4435},"))\n",{"type":617,"tag":3862,"props":4437,"children":4439},{"class":3864,"line":4438},14,[4440,4444,4448,4453],{"type":617,"tag":3862,"props":4441,"children":4442},{"style":3875},[4443],{"type":622,"value":4365},{"type":617,"tag":3862,"props":4445,"children":4446},{"style":3881},[4447],{"type":622,"value":3938},{"type":617,"tag":3862,"props":4449,"children":4450},{"style":3875},[4451],{"type":622,"value":4452},"commit",{"type":617,"tag":3862,"props":4454,"children":4455},{"style":3881},[4456],{"type":622,"value":3948},{"type":617,"tag":3862,"props":4458,"children":4460},{"class":3864,"line":4459},15,[4461,4466,4471,4475],{"type":617,"tag":3862,"props":4462,"children":4463},{"style":3869},[4464],{"type":622,"value":4465},"    return",{"type":617,"tag":3862,"props":4467,"children":4468},{"style":3875},[4469],{"type":622,"value":4470}," resp",{"type":617,"tag":3862,"props":4472,"children":4473},{"style":3881},[4474],{"type":622,"value":3884},{"type":617,"tag":3862,"props":4476,"children":4477},{"style":3875},[4478],{"type":622,"value":4479}," r_tokens\n",{"type":617,"tag":3862,"props":4481,"children":4483},{"class":3864,"line":4482},16,[4484],{"type":617,"tag":3862,"props":4485,"children":4486},{"emptyLinePlaceholder":3913},[4487],{"type":622,"value":3916},{"type":617,"tag":3862,"props":4489,"children":4491},{"class":3864,"line":4490},17,[4492,4496,4501,4505,4509,4513,4517,4521,4525,4529,4533,4537,4542,4546,4551,4555,4560,4565,4570,4575],{"type":617,"tag":3862,"props":4493,"children":4494},{"style":4052},[4495],{"type":622,"value":4055},{"type":617,"tag":3862,"props":4497,"children":4498},{"style":4058},[4499],{"type":622,"value":4500}," clustering_rate",{"type":617,"tag":3862,"props":4502,"children":4503},{"style":3881},[4504],{"type":622,"value":3978},{"type":617,"tag":3862,"props":4506,"children":4507},{"style":3875},[4508],{"type":622,"value":4191},{"type":617,"tag":3862,"props":4510,"children":4511},{"style":3881},[4512],{"type":622,"value":4075},{"type":617,"tag":3862,"props":4514,"children":4515},{"style":4078},[4516],{"type":622,"value":4081},{"type":617,"tag":3862,"props":4518,"children":4519},{"style":3881},[4520],{"type":622,"value":4103},{"type":617,"tag":3862,"props":4522,"children":4523},{"style":3981},[4524],{"type":622,"value":4108},{"type":617,"tag":3862,"props":4526,"children":4527},{"style":3987},[4528],{"type":622,"value":4113},{"type":617,"tag":3862,"props":4530,"children":4531},{"style":3981},[4532],{"type":622,"value":3984},{"type":617,"tag":3862,"props":4534,"children":4535},{"style":3881},[4536],{"type":622,"value":3884},{"type":617,"tag":3862,"props":4538,"children":4539},{"style":3875},[4540],{"type":622,"value":4541}," window",{"type":617,"tag":3862,"props":4543,"children":4544},{"style":3881},[4545],{"type":622,"value":4075},{"type":617,"tag":3862,"props":4547,"children":4548},{"style":4078},[4549],{"type":622,"value":4550}," int",{"type":617,"tag":3862,"props":4552,"children":4553},{"style":3881},[4554],{"type":622,"value":4103},{"type":617,"tag":3862,"props":4556,"children":4557},{"style":4349},[4558],{"type":622,"value":4559}," 100",{"type":617,"tag":3862,"props":4561,"children":4562},{"style":3881},[4563],{"type":622,"value":4564},")",{"type":617,"tag":3862,"props":4566,"children":4567},{"style":3881},[4568],{"type":622,"value":4569}," ->",{"type":617,"tag":3862,"props":4571,"children":4572},{"style":4078},[4573],{"type":622,"value":4574}," float",{"type":617,"tag":3862,"props":4576,"children":4577},{"style":3881},[4578],{"type":622,"value":4579},":\n",{"type":617,"tag":3862,"props":4581,"children":4583},{"class":3864,"line":4582},18,[4584,4589,4593,4598,4602,4606],{"type":617,"tag":3862,"props":4585,"children":4586},{"style":3875},[4587],{"type":622,"value":4588},"    rows ",{"type":617,"tag":3862,"props":4590,"children":4591},{"style":3881},[4592],{"type":622,"value":3929},{"type":617,"tag":3862,"props":4594,"children":4595},{"style":3875},[4596],{"type":622,"value":4597}," db",{"type":617,"tag":3862,"props":4599,"children":4600},{"style":3881},[4601],{"type":622,"value":3938},{"type":617,"tag":3862,"props":4603,"children":4604},{"style":3875},[4605],{"type":622,"value":4016},{"type":617,"tag":3862,"props":4607,"children":4608},{"style":3881},[4609],{"type":622,"value":4172},{"type":617,"tag":3862,"props":4611,"children":4613},{"class":3864,"line":4612},19,[4614,4619,4624,4628],{"type":617,"tag":3862,"props":4615,"children":4616},{"style":3981},[4617],{"type":622,"value":4618},"        \"",{"type":617,"tag":3862,"props":4620,"children":4621},{"style":3987},[4622],{"type":622,"value":4623},"SELECT reasoning_tokens FROM logs WHERE model=? ORDER BY ts DESC LIMIT ?",{"type":617,"tag":3862,"props":4625,"children":4626},{"style":3981},[4627],{"type":622,"value":3984},{"type":617,"tag":3862,"props":4629,"children":4630},{"style":3881},[4631],{"type":622,"value":4196},{"type":617,"tag":3862,"props":4633,"children":4635},{"class":3864,"line":4634},20,[4636,4641,4645,4649,4653],{"type":617,"tag":3862,"props":4637,"children":4638},{"style":3881},[4639],{"type":622,"value":4640},"        (",{"type":617,"tag":3862,"props":4642,"children":4643},{"style":3875},[4644],{"type":622,"value":4191},{"type":617,"tag":3862,"props":4646,"children":4647},{"style":3881},[4648],{"type":622,"value":3884},{"type":617,"tag":3862,"props":4650,"children":4651},{"style":3875},[4652],{"type":622,"value":4541},{"type":617,"tag":3862,"props":4654,"children":4655},{"style":3881},[4656],{"type":622,"value":3999},{"type":617,"tag":3862,"props":4658,"children":4660},{"class":3864,"line":4659},21,[4661,4666,4671],{"type":617,"tag":3862,"props":4662,"children":4663},{"style":3881},[4664],{"type":622,"value":4665},"    ).",{"type":617,"tag":3862,"props":4667,"children":4668},{"style":3875},[4669],{"type":622,"value":4670},"fetchall",{"type":617,"tag":3862,"props":4672,"children":4673},{"style":3881},[4674],{"type":622,"value":3948},{"type":617,"tag":3862,"props":4676,"children":4678},{"class":3864,"line":4677},22,[4679,4684,4688,4692,4696,4701,4705,4710,4715,4720,4725,4730,4735,4740,4745],{"type":617,"tag":3862,"props":4680,"children":4681},{"style":3875},[4682],{"type":622,"value":4683},"    counter ",{"type":617,"tag":3862,"props":4685,"children":4686},{"style":3881},[4687],{"type":622,"value":3929},{"type":617,"tag":3862,"props":4689,"children":4690},{"style":3875},[4691],{"type":622,"value":3898},{"type":617,"tag":3862,"props":4693,"children":4694},{"style":3881},[4695],{"type":622,"value":3938},{"type":617,"tag":3862,"props":4697,"children":4698},{"style":3875},[4699],{"type":622,"value":4700},"Counter",{"type":617,"tag":3862,"props":4702,"children":4703},{"style":3881},[4704],{"type":622,"value":3978},{"type":617,"tag":3862,"props":4706,"children":4707},{"style":3875},[4708],{"type":622,"value":4709},"r",{"type":617,"tag":3862,"props":4711,"children":4712},{"style":3881},[4713],{"type":622,"value":4714},"[",{"type":617,"tag":3862,"props":4716,"children":4717},{"style":4349},[4718],{"type":622,"value":4719},"0",{"type":617,"tag":3862,"props":4721,"children":4722},{"style":3881},[4723],{"type":622,"value":4724},"]",{"type":617,"tag":3862,"props":4726,"children":4727},{"style":3869},[4728],{"type":622,"value":4729}," for",{"type":617,"tag":3862,"props":4731,"children":4732},{"style":3875},[4733],{"type":622,"value":4734}," r ",{"type":617,"tag":3862,"props":4736,"children":4737},{"style":3869},[4738],{"type":622,"value":4739},"in",{"type":617,"tag":3862,"props":4741,"children":4742},{"style":3875},[4743],{"type":622,"value":4744}," rows",{"type":617,"tag":3862,"props":4746,"children":4747},{"style":3881},[4748],{"type":622,"value":3999},{"type":617,"tag":3862,"props":4750,"children":4752},{"class":3864,"line":4751},23,[4753,4757,4762,4766,4771,4775,4780,4785,4789,4794,4798,4803,4808,4813],{"type":617,"tag":3862,"props":4754,"children":4755},{"style":3869},[4756],{"type":622,"value":4465},{"type":617,"tag":3862,"props":4758,"children":4759},{"style":3875},[4760],{"type":622,"value":4761}," counter",{"type":617,"tag":3862,"props":4763,"children":4764},{"style":3881},[4765],{"type":622,"value":4714},{"type":617,"tag":3862,"props":4767,"children":4768},{"style":4349},[4769],{"type":622,"value":4770},"516",{"type":617,"tag":3862,"props":4772,"children":4773},{"style":3881},[4774],{"type":622,"value":4724},{"type":617,"tag":3862,"props":4776,"children":4777},{"style":4052},[4778],{"type":622,"value":4779}," /",{"type":617,"tag":3862,"props":4781,"children":4782},{"style":4078},[4783],{"type":622,"value":4784}," len",{"type":617,"tag":3862,"props":4786,"children":4787},{"style":3881},[4788],{"type":622,"value":3978},{"type":617,"tag":3862,"props":4790,"children":4791},{"style":3875},[4792],{"type":622,"value":4793},"rows",{"type":617,"tag":3862,"props":4795,"children":4796},{"style":3881},[4797],{"type":622,"value":4564},{"type":617,"tag":3862,"props":4799,"children":4800},{"style":3869},[4801],{"type":622,"value":4802}," if",{"type":617,"tag":3862,"props":4804,"children":4805},{"style":3875},[4806],{"type":622,"value":4807}," rows ",{"type":617,"tag":3862,"props":4809,"children":4810},{"style":3869},[4811],{"type":622,"value":4812},"else",{"type":617,"tag":3862,"props":4814,"children":4815},{"style":4349},[4816],{"type":622,"value":4817}," 0.0\n",{"type":617,"tag":661,"props":4819,"children":4820},{"id":1269},[4821],{"type":622,"value":1269},{"type":617,"tag":618,"props":4823,"children":4824},{},[4825],{"type":622,"value":4826},"連續呼叫 50–100 次後，觀察 reasoning token 分佈是否呈現單峰（健康）或固定聚集（退化）。clustering_rate > 5% 視為警示；> 20% 視為需切換後端或回報 issue 的嚴重退化。",{"type":617,"tag":661,"props":4828,"children":4829},{"id":1279},[4830],{"type":622,"value":1279},{"type":617,"tag":1283,"props":4832,"children":4833},{},[4834,4839,4851],{"type":617,"tag":866,"props":4835,"children":4836},{},[4837],{"type":622,"value":4838},"僅看最終輸出正確率，忽略 reasoning token 數量：退化初期輸出可能仍「看起來」合理",{"type":617,"tag":866,"props":4840,"children":4841},{},[4842,4844,4849],{"type":622,"value":4843},"誤將 workaround（移除 ",{"type":617,"tag":1555,"props":4845,"children":4847},{"className":4846},[],[4848],{"type":622,"value":1560},{"type":622,"value":4850},"）當作根本修復：底層問題未解決，其他觸發條件可能持續出現",{"type":617,"tag":866,"props":4852,"children":4853},{},[4854],{"type":622,"value":4855},"未記錄完整 model ID（含版本戳記）：OpenAI 可能在不公告的情況下調整後端行為",{"type":617,"tag":661,"props":4857,"children":4858},{"id":1303},[4859],{"type":622,"value":1303},{"type":617,"tag":1283,"props":4861,"children":4862},{},[4863,4868,4873],{"type":617,"tag":866,"props":4864,"children":4865},{},[4866],{"type":622,"value":4867},"觀測：reasoning token 分佈（每日 p50／p95）、516 ± 5 token clustering 比率、任務成功率",{"type":617,"tag":866,"props":4869,"children":4870},{},[4871],{"type":622,"value":4872},"成本：reasoning token 驟降時，推理成本應同比下降；若成本不降但準確率降，需深查計費異常",{"type":617,"tag":866,"props":4874,"children":4875},{},[4876,4878,4884],{"type":622,"value":4877},"風險：避免系統提示中使用結構化 ",{"type":617,"tag":1555,"props":4879,"children":4881},{"className":4880},[],[4882],{"type":622,"value":4883},"##",{"type":622,"value":4885}," heading 格式；為關鍵推理任務設置輸出驗證層",{"type":617,"tag":4887,"props":4888,"children":4889},"style",{},[4890],{"type":622,"value":4891},"html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}",{"title":156,"searchDepth":624,"depth":624,"links":4893},[],{"data":4895,"body":4896,"excerpt":-1,"toc":5469},{"title":156,"description":156},{"type":614,"children":4897},[4898,4902,4907,4911,5400,5404,5409,5413,5436,5440,5465],{"type":617,"tag":661,"props":4899,"children":4900},{"id":1225},[4901],{"type":622,"value":1225},{"type":617,"tag":618,"props":4903,"children":4904},{},[4905],{"type":622,"value":4906},"最低需求：12GB VRAM(NVIDIA GPU) ，Python 3.10+，CUDA 12.x。模型權重約 6GB，建議使用 A10G 或 RTX 3090/4090 進行本地測試；生產環境建議 A100 或 H100 以獲得最佳吞吐量。",{"type":617,"tag":661,"props":4908,"children":4909},{"id":3847},[4910],{"type":622,"value":3850},{"type":617,"tag":3852,"props":4912,"children":4914},{"className":3854,"code":4913,"language":3856,"meta":156,"style":156},"from transformers import AutoModelForCausalLM, AutoTokenizer\nfrom PIL import Image\nimport torch\n\nmodel_id = \"baidu/Unlimited-OCR\"\ntokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)\nmodel = AutoModelForCausalLM.from_pretrained(\n    model_id,\n    torch_dtype=torch.bfloat16,\n    device_map=\"auto\",\n    trust_remote_code=True\n)\n\n# 多頁 PDF 需先轉為圖像列表（如使用 pdf2image）\nimages = [Image.open(f\"page_{i}.png\") for i in range(1, 41)]\nresult = model.ocr(images, tokenizer=tokenizer)\nprint(result)\n",[4915],{"type":617,"tag":1555,"props":4916,"children":4917},{"__ignoreMap":156},[4918,4949,4972,4984,4991,5017,5074,5102,5114,5144,5173,5190,5197,5204,5213,5323,5379],{"type":617,"tag":3862,"props":4919,"children":4920},{"class":3864,"line":3865},[4921,4926,4931,4935,4940,4944],{"type":617,"tag":3862,"props":4922,"children":4923},{"style":3869},[4924],{"type":622,"value":4925},"from",{"type":617,"tag":3862,"props":4927,"children":4928},{"style":3875},[4929],{"type":622,"value":4930}," transformers ",{"type":617,"tag":3862,"props":4932,"children":4933},{"style":3869},[4934],{"type":622,"value":3872},{"type":617,"tag":3862,"props":4936,"children":4937},{"style":3875},[4938],{"type":622,"value":4939}," AutoModelForCausalLM",{"type":617,"tag":3862,"props":4941,"children":4942},{"style":3881},[4943],{"type":622,"value":3884},{"type":617,"tag":3862,"props":4945,"children":4946},{"style":3875},[4947],{"type":622,"value":4948}," AutoTokenizer\n",{"type":617,"tag":3862,"props":4950,"children":4951},{"class":3864,"line":624},[4952,4956,4962,4967],{"type":617,"tag":3862,"props":4953,"children":4954},{"style":3869},[4955],{"type":622,"value":4925},{"type":617,"tag":3862,"props":4957,"children":4959},{"style":4958},"--shiki-default:#C99076",[4960],{"type":622,"value":4961}," PIL",{"type":617,"tag":3862,"props":4963,"children":4964},{"style":3869},[4965],{"type":622,"value":4966}," import",{"type":617,"tag":3862,"props":4968,"children":4969},{"style":3875},[4970],{"type":622,"value":4971}," Image\n",{"type":617,"tag":3862,"props":4973,"children":4974},{"class":3864,"line":185},[4975,4979],{"type":617,"tag":3862,"props":4976,"children":4977},{"style":3869},[4978],{"type":622,"value":3872},{"type":617,"tag":3862,"props":4980,"children":4981},{"style":3875},[4982],{"type":622,"value":4983}," torch\n",{"type":617,"tag":3862,"props":4985,"children":4986},{"class":3864,"line":75},[4987],{"type":617,"tag":3862,"props":4988,"children":4989},{"emptyLinePlaceholder":3913},[4990],{"type":622,"value":3916},{"type":617,"tag":3862,"props":4992,"children":4993},{"class":3864,"line":76},[4994,4999,5003,5007,5012],{"type":617,"tag":3862,"props":4995,"children":4996},{"style":3875},[4997],{"type":622,"value":4998},"model_id ",{"type":617,"tag":3862,"props":5000,"children":5001},{"style":3881},[5002],{"type":622,"value":3929},{"type":617,"tag":3862,"props":5004,"children":5005},{"style":3981},[5006],{"type":622,"value":4108},{"type":617,"tag":3862,"props":5008,"children":5009},{"style":3987},[5010],{"type":622,"value":5011},"baidu/Unlimited-OCR",{"type":617,"tag":3862,"props":5013,"children":5014},{"style":3981},[5015],{"type":622,"value":5016},"\"\n",{"type":617,"tag":3862,"props":5018,"children":5019},{"class":3864,"line":4040},[5020,5025,5029,5034,5038,5043,5047,5052,5056,5061,5065,5070],{"type":617,"tag":3862,"props":5021,"children":5022},{"style":3875},[5023],{"type":622,"value":5024},"tokenizer ",{"type":617,"tag":3862,"props":5026,"children":5027},{"style":3881},[5028],{"type":622,"value":3929},{"type":617,"tag":3862,"props":5030,"children":5031},{"style":3875},[5032],{"type":622,"value":5033}," AutoTokenizer",{"type":617,"tag":3862,"props":5035,"children":5036},{"style":3881},[5037],{"type":622,"value":3938},{"type":617,"tag":3862,"props":5039,"children":5040},{"style":3875},[5041],{"type":622,"value":5042},"from_pretrained",{"type":617,"tag":3862,"props":5044,"children":5045},{"style":3881},[5046],{"type":622,"value":3978},{"type":617,"tag":3862,"props":5048,"children":5049},{"style":3875},[5050],{"type":622,"value":5051},"model_id",{"type":617,"tag":3862,"props":5053,"children":5054},{"style":3881},[5055],{"type":622,"value":3884},{"type":617,"tag":3862,"props":5057,"children":5058},{"style":4179},[5059],{"type":622,"value":5060}," trust_remote_code",{"type":617,"tag":3862,"props":5062,"children":5063},{"style":3881},[5064],{"type":622,"value":3929},{"type":617,"tag":3862,"props":5066,"children":5067},{"style":3869},[5068],{"type":622,"value":5069},"True",{"type":617,"tag":3862,"props":5071,"children":5072},{"style":3881},[5073],{"type":622,"value":3999},{"type":617,"tag":3862,"props":5075,"children":5076},{"class":3864,"line":4048},[5077,5082,5086,5090,5094,5098],{"type":617,"tag":3862,"props":5078,"children":5079},{"style":3875},[5080],{"type":622,"value":5081},"model ",{"type":617,"tag":3862,"props":5083,"children":5084},{"style":3881},[5085],{"type":622,"value":3929},{"type":617,"tag":3862,"props":5087,"children":5088},{"style":3875},[5089],{"type":622,"value":4939},{"type":617,"tag":3862,"props":5091,"children":5092},{"style":3881},[5093],{"type":622,"value":3938},{"type":617,"tag":3862,"props":5095,"children":5096},{"style":3875},[5097],{"type":622,"value":5042},{"type":617,"tag":3862,"props":5099,"children":5100},{"style":3881},[5101],{"type":622,"value":4172},{"type":617,"tag":3862,"props":5103,"children":5104},{"class":3864,"line":4125},[5105,5110],{"type":617,"tag":3862,"props":5106,"children":5107},{"style":3875},[5108],{"type":622,"value":5109},"    model_id",{"type":617,"tag":3862,"props":5111,"children":5112},{"style":3881},[5113],{"type":622,"value":4196},{"type":617,"tag":3862,"props":5115,"children":5116},{"class":3864,"line":4175},[5117,5122,5126,5131,5135,5140],{"type":617,"tag":3862,"props":5118,"children":5119},{"style":4179},[5120],{"type":622,"value":5121},"    torch_dtype",{"type":617,"tag":3862,"props":5123,"children":5124},{"style":3881},[5125],{"type":622,"value":3929},{"type":617,"tag":3862,"props":5127,"children":5128},{"style":3875},[5129],{"type":622,"value":5130},"torch",{"type":617,"tag":3862,"props":5132,"children":5133},{"style":3881},[5134],{"type":622,"value":3938},{"type":617,"tag":3862,"props":5136,"children":5137},{"style":3875},[5138],{"type":622,"value":5139},"bfloat16",{"type":617,"tag":3862,"props":5141,"children":5142},{"style":3881},[5143],{"type":622,"value":4196},{"type":617,"tag":3862,"props":5145,"children":5146},{"class":3864,"line":4199},[5147,5152,5156,5160,5165,5169],{"type":617,"tag":3862,"props":5148,"children":5149},{"style":4179},[5150],{"type":622,"value":5151},"    device_map",{"type":617,"tag":3862,"props":5153,"children":5154},{"style":3881},[5155],{"type":622,"value":3929},{"type":617,"tag":3862,"props":5157,"children":5158},{"style":3981},[5159],{"type":622,"value":3984},{"type":617,"tag":3862,"props":5161,"children":5162},{"style":3987},[5163],{"type":622,"value":5164},"auto",{"type":617,"tag":3862,"props":5166,"children":5167},{"style":3981},[5168],{"type":622,"value":3984},{"type":617,"tag":3862,"props":5170,"children":5171},{"style":3881},[5172],{"type":622,"value":4196},{"type":617,"tag":3862,"props":5174,"children":5175},{"class":3864,"line":4274},[5176,5181,5185],{"type":617,"tag":3862,"props":5177,"children":5178},{"style":4179},[5179],{"type":622,"value":5180},"    trust_remote_code",{"type":617,"tag":3862,"props":5182,"children":5183},{"style":3881},[5184],{"type":622,"value":3929},{"type":617,"tag":3862,"props":5186,"children":5187},{"style":3869},[5188],{"type":622,"value":5189},"True\n",{"type":617,"tag":3862,"props":5191,"children":5192},{"class":3864,"line":4283},[5193],{"type":617,"tag":3862,"props":5194,"children":5195},{"style":3881},[5196],{"type":622,"value":3999},{"type":617,"tag":3862,"props":5198,"children":5199},{"class":3864,"line":4359},[5200],{"type":617,"tag":3862,"props":5201,"children":5202},{"emptyLinePlaceholder":3913},[5203],{"type":622,"value":3916},{"type":617,"tag":3862,"props":5205,"children":5206},{"class":3864,"line":4438},[5207],{"type":617,"tag":3862,"props":5208,"children":5210},{"style":5209},"--shiki-default:#758575DD",[5211],{"type":622,"value":5212},"# 多頁 PDF 需先轉為圖像列表（如使用 pdf2image）\n",{"type":617,"tag":3862,"props":5214,"children":5215},{"class":3864,"line":4459},[5216,5221,5225,5230,5235,5239,5244,5248,5253,5258,5263,5268,5273,5278,5282,5286,5291,5295,5300,5304,5309,5313,5318],{"type":617,"tag":3862,"props":5217,"children":5218},{"style":3875},[5219],{"type":622,"value":5220},"images ",{"type":617,"tag":3862,"props":5222,"children":5223},{"style":3881},[5224],{"type":622,"value":3929},{"type":617,"tag":3862,"props":5226,"children":5227},{"style":3881},[5228],{"type":622,"value":5229}," [",{"type":617,"tag":3862,"props":5231,"children":5232},{"style":3875},[5233],{"type":622,"value":5234},"Image",{"type":617,"tag":3862,"props":5236,"children":5237},{"style":3881},[5238],{"type":622,"value":3938},{"type":617,"tag":3862,"props":5240,"children":5241},{"style":3875},[5242],{"type":622,"value":5243},"open",{"type":617,"tag":3862,"props":5245,"children":5246},{"style":3881},[5247],{"type":622,"value":3978},{"type":617,"tag":3862,"props":5249,"children":5250},{"style":4052},[5251],{"type":622,"value":5252},"f",{"type":617,"tag":3862,"props":5254,"children":5255},{"style":3987},[5256],{"type":622,"value":5257},"\"page_",{"type":617,"tag":3862,"props":5259,"children":5260},{"style":4958},[5261],{"type":622,"value":5262},"{",{"type":617,"tag":3862,"props":5264,"children":5265},{"style":3875},[5266],{"type":622,"value":5267},"i",{"type":617,"tag":3862,"props":5269,"children":5270},{"style":4958},[5271],{"type":622,"value":5272},"}",{"type":617,"tag":3862,"props":5274,"children":5275},{"style":3987},[5276],{"type":622,"value":5277},".png\"",{"type":617,"tag":3862,"props":5279,"children":5280},{"style":3881},[5281],{"type":622,"value":4564},{"type":617,"tag":3862,"props":5283,"children":5284},{"style":3869},[5285],{"type":622,"value":4729},{"type":617,"tag":3862,"props":5287,"children":5288},{"style":3875},[5289],{"type":622,"value":5290}," i ",{"type":617,"tag":3862,"props":5292,"children":5293},{"style":3869},[5294],{"type":622,"value":4739},{"type":617,"tag":3862,"props":5296,"children":5297},{"style":4078},[5298],{"type":622,"value":5299}," range",{"type":617,"tag":3862,"props":5301,"children":5302},{"style":3881},[5303],{"type":622,"value":3978},{"type":617,"tag":3862,"props":5305,"children":5306},{"style":4349},[5307],{"type":622,"value":5308},"1",{"type":617,"tag":3862,"props":5310,"children":5311},{"style":3881},[5312],{"type":622,"value":3884},{"type":617,"tag":3862,"props":5314,"children":5315},{"style":4349},[5316],{"type":622,"value":5317}," 41",{"type":617,"tag":3862,"props":5319,"children":5320},{"style":3881},[5321],{"type":622,"value":5322},")]\n",{"type":617,"tag":3862,"props":5324,"children":5325},{"class":3864,"line":4482},[5326,5331,5335,5339,5343,5348,5352,5357,5361,5366,5370,5375],{"type":617,"tag":3862,"props":5327,"children":5328},{"style":3875},[5329],{"type":622,"value":5330},"result ",{"type":617,"tag":3862,"props":5332,"children":5333},{"style":3881},[5334],{"type":622,"value":3929},{"type":617,"tag":3862,"props":5336,"children":5337},{"style":3875},[5338],{"type":622,"value":4090},{"type":617,"tag":3862,"props":5340,"children":5341},{"style":3881},[5342],{"type":622,"value":3938},{"type":617,"tag":3862,"props":5344,"children":5345},{"style":3875},[5346],{"type":622,"value":5347},"ocr",{"type":617,"tag":3862,"props":5349,"children":5350},{"style":3881},[5351],{"type":622,"value":3978},{"type":617,"tag":3862,"props":5353,"children":5354},{"style":3875},[5355],{"type":622,"value":5356},"images",{"type":617,"tag":3862,"props":5358,"children":5359},{"style":3881},[5360],{"type":622,"value":3884},{"type":617,"tag":3862,"props":5362,"children":5363},{"style":4179},[5364],{"type":622,"value":5365}," tokenizer",{"type":617,"tag":3862,"props":5367,"children":5368},{"style":3881},[5369],{"type":622,"value":3929},{"type":617,"tag":3862,"props":5371,"children":5372},{"style":3875},[5373],{"type":622,"value":5374},"tokenizer",{"type":617,"tag":3862,"props":5376,"children":5377},{"style":3881},[5378],{"type":622,"value":3999},{"type":617,"tag":3862,"props":5380,"children":5381},{"class":3864,"line":4490},[5382,5387,5391,5396],{"type":617,"tag":3862,"props":5383,"children":5384},{"style":4078},[5385],{"type":622,"value":5386},"print",{"type":617,"tag":3862,"props":5388,"children":5389},{"style":3881},[5390],{"type":622,"value":3978},{"type":617,"tag":3862,"props":5392,"children":5393},{"style":3875},[5394],{"type":622,"value":5395},"result",{"type":617,"tag":3862,"props":5397,"children":5398},{"style":3881},[5399],{"type":622,"value":3999},{"type":617,"tag":661,"props":5401,"children":5402},{"id":1269},[5403],{"type":622,"value":1269},{"type":617,"tag":618,"props":5405,"children":5406},{},[5407],{"type":622,"value":5408},"以 OmniDocBench 評測集進行基準對比，預期綜合分數應落在 93+ 區間。自有文件測試時，重點驗證跨頁表格的 TEDS 分數與公式辨識的 LaTeX 輸出正確率。",{"type":617,"tag":661,"props":5410,"children":5411},{"id":1279},[5412],{"type":622,"value":1279},{"type":617,"tag":1283,"props":5414,"children":5415},{},[5416,5421,5431],{"type":617,"tag":866,"props":5417,"children":5418},{},[5419],{"type":622,"value":5420},"序列長度上限 32K tokens：單次推論約可處理 40-60 頁（依內容密度而異），超過上限需手動分批",{"type":617,"tag":866,"props":5422,"children":5423},{},[5424,5429],{"type":617,"tag":1555,"props":5425,"children":5427},{"className":5426},[],[5428],{"type":622,"value":2543},{"type":622,"value":5430}," 為必要參數：模型使用自訂注意力實作，需允許執行自定義程式碼",{"type":617,"tag":866,"props":5432,"children":5433},{},[5434],{"type":622,"value":5435},"訓練資料單頁：多頁比例 9：1：超長文件（100+ 頁）表現未經充分驗證，需自行評測",{"type":617,"tag":661,"props":5437,"children":5438},{"id":1303},[5439],{"type":622,"value":1303},{"type":617,"tag":1283,"props":5441,"children":5442},{},[5443,5448,5453],{"type":617,"tag":866,"props":5444,"children":5445},{},[5446],{"type":622,"value":5447},"觀測：TEDS 分數、Edit Distance、每頁推論延遲、VRAM 峰值使用量",{"type":617,"tag":866,"props":5449,"children":5450},{},[5451],{"type":622,"value":5452},"成本：單頁推論約 0.18 秒 (A100) ，40 頁文件約 7.2 秒；批次作業建議搭配 vLLM 部署提升吞吐",{"type":617,"tag":866,"props":5454,"children":5455},{},[5456,5458,5463],{"type":622,"value":5457},"風險：MIT 授權允許商用，但 ",{"type":617,"tag":1555,"props":5459,"children":5461},{"className":5460},[],[5462],{"type":622,"value":2543},{"type":622,"value":5464}," 在生產環境需通過安全審查；訓練資料著作權合規性需評估",{"type":617,"tag":4887,"props":5466,"children":5467},{},[5468],{"type":622,"value":4891},{"title":156,"searchDepth":624,"depth":624,"links":5470},[]]