[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"report-2026-07-08":3,"JxuklbSiMa":608,"5Fq5ymeee3":623,"qyL2YzqGqO":633,"lkJ2ZVNnhx":643,"Hbh4HpIJo8":653,"JPSlRUOzNo":782,"P1dKPHSOCi":793,"6lngJlhgH8":809,"FqPszkNmR1":833,"7z5oMp3Iaf":865,"PrLUiUzINz":1008,"aM1wVPpzdq":1132,"bO2EA2cmmS":1153,"vDMnYCtC70":1178,"8gZcmH1WpL":1199,"NGxCV3DZDI":1209,"TyMqbr2lej":1219,"OrD6UFpTD2":1229,"39F6W8YpNq":1239,"P4j0jr5CC1":1249,"VrHHlY3CmW":1259,"EIG9HAUeP6":1269,"GNumQrWB5w":1394,"pnJmpSHoAl":1405,"lO6S06viZh":1416,"fKm7eJKwSP":1442,"RfK9Kquww1":1468,"yES76w4tNs":1563,"OPflrFj2Wc":1683,"U5ChTn1Foi":1799,"vXsZmAoAzG":1820,"iU7eTFYdMw":1837,"1IKXAQpoag":1847,"Q3ZPIlbJfo":1857,"7qy1JQG5VM":1867,"xEcagctuba":1877,"w6U2AvRpXM":1887,"yxSHjMuycr":1897,"2p5n5EoIqe":1907,"dhtmBVNRuH":2040,"ph5aHfb3cs":2051,"kfV3DyBcNO":2115,"iwylBj0Psr":2149,"A2Msn5gI1Z":2175,"PSOzpDjPr3":2290,"WiED9PjtP8":2351,"fKb54D0RS9":2376,"9LhgKdBghZ":2397,"uoaoxbyefn":2407,"M4EnOspom4":2417,"SwVOqCptL3":2427,"3i9VyY91xN":2437,"5AZLAye0r6":2447,"jeqMIePZnA":2457,"Vo9u4NRfGs":2467,"pT0Yl4YXXn":2582,"0zjXWLBqvC":2593,"0iViqr0GfA":2609,"abxWyvwRyW":2625,"tLwz3sSbpS":2656,"yfW5wNXGRX":2791,"nnBTV9R4Zo":2812,"LcizJwFotX":2829,"kw1lc9MoXt":2839,"LsC656Zc3e":2849,"FSYmJJJOv7":2961,"6Fh4GUOqs1":2990,"jwOXt32mcC":3006,"XEu8Sm6AkA":3086,"A2ZB6sPwWv":3110,"ytw9K4tacg":3134,"w5Rz64WhZl":3168,"CRew7nKyp8":3300,"34VeLVw9UF":3340,"48xVWX7uRB":3379,"0HeAFzNHAO":3416,"orCBYu7nRM":3426,"ki1ZLjaJwB":3436,"IBEQ8DjIW8":3491,"fJRe4GX8jn":3507,"BaNg6hVvq5":3523,"ZFJSoRTQnx":3579,"OhyVCaJI0l":3622,"zQpT4mwbO0":3632,"cnkSAXoN8T":3642,"MADhTdPEk9":3737,"IGnFnCNchM":3753,"Xob4wEbdSU":3769,"LdiQFmQJVv":3803,"DURl4YSbGJ":3858,"1P5HyUoAjk":3892,"2XTXkb3NRF":3908,"YEv1xkhT2N":3980,"lmjb9Kdg8s":3996,"Lep81s6BTL":4012,"6DlIfkNYj1":4116,"5Am32eDzX0":4132,"bZ0nJNkmED":5137},{"report":4,"adjacent":605},{"version":5,"date":6,"title":7,"sources":8,"hook":16,"deepDives":17,"quickBites":324,"communityOverview":589,"dailyActions":590,"outro":604},"20260216.0","2026-07-08","AI 趨勢日報：2026-07-08",[9,10,11,12,13,14,15],"academic","anthropic","community","deepseek","github","google","microsoft","成本戰、晶片自主、模型透明度三條戰線同步開打：今日 AI 產業正從「誰的模型最強」轉向「誰能控制整條供應鏈」。",[18,104,180,257],{"category":19,"source":11,"title":20,"subtitle":21,"publishDate":6,"tier1Source":22,"supplementSources":25,"tldr":42,"context":54,"mechanics":55,"benchmark":56,"useCases":57,"engineerLens":67,"businessLens":68,"devilsAdvocate":69,"community":73,"hypeScore":91,"hypeMax":92,"adoptionAdvice":93,"actionItems":94},"ecosystem","reMarkable 變身魔法日記：Fable 打造電子紙 AI 即時手寫互動新物種","開源專案 Riddle 讓 reMarkable Paper Pro 實現手寫消失、AI 手寫回應，v0.3.0 上線後迅速累積千顆 GitHub 星",{"name":23,"url":24},"GitHub: MaximeRivest/Riddle","https://github.com/MaximeRivest/Riddle",[26,30,34,38],{"name":27,"url":28,"detail":29},"Hacker News #48811591","https://news.ycombinator.com/item?id=48811591","社群對 Riddle 專案的深度討論，含書寫障礙輔助潛力與實際使用體驗",{"name":31,"url":32,"detail":33},"Android Authority","https://www.androidauthority.com/remarkable-paper-pro-tom-riddles-diary-disappearing-ink-3684286/","消失墨水體驗完整報導",{"name":35,"url":36,"detail":37},"TechRadar","https://www.techradar.com/tablets/ereaders/this-new-interactive-tool-on-the-remarkable-paper-pro-turns-your-device-into-tom-riddles-diary-from-harry-potter-and-its-one-of-the-smartest-e-reader-features-ive-seen","TechRadar 評測，認為這是目前最聰明的電子書功能之一",{"name":39,"url":40,"detail":41},"Notebookcheck News","https://www.notebookcheck.net/Someone-turned-a-reMarkable-tablet-into-Tom-Riddle-s-diary.1335689.0.html","技術細節與安裝門檻報導",{"tagline":43,"points":44},"電子紙 × AI 手寫對話，開源重現哈利波特魔法日記體驗",[45,48,51],{"label":46,"text":47},"技術","透過 evdev 擷取觸控筆事件，停筆 2.8 秒後截圖送入 Fable 5 視覺模型，AI 以 Dancing Script 字型將回覆「寫」回電子紙，整體延遲 1 秒內啟動。",{"label":49,"text":50},"生態","Rust + C++ 架構支援 Fable 5、OpenAI、OpenRouter、Groq 及本地視覺模型；v0.3.0 上線後累積 1,000+ GitHub 星與 72 個 Fork，社群驗證強烈。",{"label":52,"text":53},"落地","需開啟 reMarkable 開發者模式、SSH 存取與第三方啟動器；韌體更新可能讓 mod 失效，建議有設備的開發者先進行 PoC 評估再投入使用。","#### 章節一：當 AI 寫進電子紙——Tom Riddle 日記的現代重現\n\n開發者 Maxime Rivest 於 2026 年 7 月 7 日發布開源專案 Riddle v0.3.0，將 reMarkable Paper Pro 改造成哈利波特《消失的密室》中湯姆‧瑞鐸日記的現代版。\n\n用戶以觸控筆在電子紙上書寫後，文字自動淡出消失，AI 則以手寫體字型將回覆「自動寫」回頁面。這個設計哲學源自作者明確宣言：「移除所有傳統 AI 介面——沒有鍵盤、沒有對話泡泡、沒有發光螢幕。」\n\n電子紙的低刷新率與類紙質感反而成為這套互動體驗的天然舞台，而非技術限制。Riddle 的出現讓「AI 融入書寫材質」的概念從科幻場景變成可下載、可運行的開源工具。\n\n#### 章節二：技術拆解：手寫辨識、即時生成與電子墨水的三方協作\n\nRiddle 的技術架構分三層運作，三者串接形成低延遲的手寫對話迴路。\n\n輸入層透過 Linux evdev 介面直接擷取觸控筆原始事件，支援 4096 級壓力感應，確保辨識精度不受寫字力道影響。用戶停筆約 2.8 秒後，系統截取當前頁面為 PNG 送入核心 LLM。\n\n> **名詞解釋**\n> **evdev**：Linux 核心提供的通用輸入事件介面，允許應用程式直接讀取觸控筆等裝置的原始輸入事件，無需依賴圖形框架層。\n\n預設 LLM 為 Anthropic Fable 5(`claude-fable-5`) ，亦支援 OpenAI、OpenRouter、Groq 及本地視覺模型。AI 串流輸出到達前，前端即以 Dancing Script 字型逐字動態播出，整體延遲 1 秒內啟動。\n\n「墨水消失」效果由軟體清空螢幕並重新渲染實現，視覺上呈現「墨水被頁面吸走」的效果。顯示後端提供 qtfb（xochitl 視窗模式，相容性高）與 quill（直接接管 e-ink 引擎，延遲最低）二選一。\n\n#### 章節三：社群實測：從閱讀障礙輔助到創意筆記的真實回饋\n\nHN 討論串中，社群用戶 SoMomentary 指出，自己患有書寫障礙 (dysgraphia) ，過去在學校大量使用 reMarkable 做筆記，這套系統對他有實質幫助。\n\n這個回饋揭示了 Riddle 意料之外的應用場景：無障礙輔助。對於手寫困難的用戶，AI 即時辨識並以清晰字型呈現回應，能有效降低閱讀與理解的認知負擔。\n\n社群其他反應走向兩極。ceejayoz 強調動態效果才是核心體驗，附上示範影片提醒其他人先看動畫再評論；jshandling 則批評部分留言過度拘泥技術細節，忽略了整體體驗設計的意圖。\n\n#### 章節四：電子紙 × AI 互動裝置的產品想像空間\n\nRiddle 的出現為電子紙設備開闢了「環境運算介面」的新敘事：AI 不再是一個應用程式，而是融入書寫材質本身。X 用戶 @Steven_McKie 直接點出潛力——彩色電子紙加上這套技術，足以取代學童的學校筆記本與筆電。\n\n從產品維度看，這套架構目前的核心限制在於硬體鎖定：僅支援 reMarkable Paper Pro，安裝需要開發者模式與第三方啟動器，韌體更新可能讓 mod 失效。\n\n但這也暗示了一個更大的方向：若 reMarkable 原廠或其他電子紙廠商選擇將類似功能原生整合，這套 UX 模式將有機會觸及主流用戶，而不只是願意 SSH 進平板的開發者社群。","Riddle 的技術核心是三個相對獨立的子系統：輸入擷取、AI 推理、顯示渲染，三者串接形成低延遲的手寫對話迴路。\n\n#### 機制 1：evdev 原始輸入擷取\n\nRiddle 繞過 reMarkable 官方軟體棧，直接透過 Linux evdev 介面讀取觸控筆的原始事件流，支援 4096 級壓力感應。停筆約 2.8 秒後，系統截取當前頁面為 PNG 圖像送往 LLM。\n\n這個設計讓系統能在不改動官方 UI 的前提下，捕捉所有手寫輸入的完整壓力與位置資訊，為後續視覺辨識提供高品質截圖。\n\n#### 機制 2：視覺語言模型手寫辨識與串流生成\n\n截圖送入 Anthropic Fable 5(`claude-fable-5`) 或用戶自選的 OpenAI、OpenRouter、Groq、本地視覺模型，進行手寫內容辨識並生成回覆。\n\n系統採串流輸出，AI 輸出開始到達前，前端即以 Dancing Script 字型逐字動態播出，整體延遲 1 秒內啟動，體驗接近即時書寫回應。\n\n#### 機制 3：電子墨水消失效果與手寫字型渲染\n\n「墨水消失」效果由軟體實現：系統清空螢幕並重新渲染，視覺上呈現「墨水被頁面吸走」的效果。AI 回覆以 Dancing Script 字型輸出，模擬手寫筆跡風格。\n\n顯示後端提供兩種選擇：qtfb（xochitl 視窗模式，相容性高）與 quill（直接接管 e-ink 引擎，延遲最低）。持久記憶功能讓所有頁面的轉錄文字與 AI 回覆可被日後查詢回溯。\n\n> **白話比喻**\n> 這就像你對著魔法日記寫一個問題，墨水被頁面吸走，幾秒後日記自己「寫」出答案。Riddle 把每一步都用軟體實現了：evdev 是耳朵，Fable 5 是大腦，電子墨水渲染是筆。","#### 延遲與效能\n\n根據作者描述，回覆生成延遲在 1 秒內啟動，整體體驗接近即時書寫回應。結合 2.8 秒 idle timeout，從停筆到看見 AI 回覆開始出現的等待時間約 3-4 秒。\n\n暫無 OCR 辨識準確率或多語言辨識的正式基準測試數據。社群回饋顯示英文手寫辨識表現正常，繁體中文手寫的辨識精度尚無公開數據，建議有需求者自行測試。",{"recommended":58,"avoid":63},[59,60,61,62],"個人日記與反思式寫作——手寫筆觸加 AI 即時回應的私密互動，適合深度思考記錄","語言學習輔助——書寫練習後由 AI 即時回覆並糾錯，模擬家教批改效果","無障礙輔助——書寫障礙 (dysgraphia) 用戶透過 AI 辨識與清晰字型排版提升閱讀體驗","創意發想——手寫 prompt 讓 AI 在紙頁上「接話」，保留無數位感的創作流程",[64,65,66],"需要即時雙向快速對話的場景——2.8 秒 idle timeout 不適合問答節奏快的需求","繁體中文手寫辨識需求——辨識精度尚無公開數據，建議先充分測試再投入使用","企業部署或多裝置集中管理——需個別 SSH 安裝，無集中管理與版本控制機制","#### 環境需求\n\n需要 reMarkable Paper Pro（不支援舊版 reMarkable 2），並開啟開發者模式、設定 SSH 存取。需安裝 xovi 與 AppLoad 兩個第三方啟動器元件，並準備 Anthropic API Key（或支援 OpenAI / OpenRouter / Groq 協議的替代 API Key）。\n\n平台限制明確：reMarkable 韌體更新可能讓 mod 失效，建議鎖定韌體版本或追蹤 Riddle 上游的相容性更新紀錄。\n\n#### 遷移／整合步驟\n\n1. 在 reMarkable Paper Pro 開啟開發者模式並設定 SSH\n2. 安裝 xovi + AppLoad 第三方啟動器\n3. 設定環境變數 `ANTHROPIC_API_KEY`（或對應 LLM 的 API Key）\n4. 下載 Riddle v0.3.0 並依 README 執行安裝腳本\n5. 選擇顯示後端：`qtfb`（相容模式）或 `quill`（低延遲模式）\n\n#### 驗測規劃\n\n安裝完成後，以下流程驗測核心路徑：在頁面書寫一個問題，停筆等待 2.8 秒，確認墨水消失動畫觸發、AI 回覆以 Dancing Script 字型逐字出現。\n\n延遲目標：回覆開始出現的延遲應在 1 秒內；若超過 3 秒，需檢查網路連線或 API 配額狀況。\n\n#### 常見陷阱\n\n- reMarkable 韌體更新會重置開發者模式，需重新設定 SSH 並重裝啟動器元件\n- 本地視覺模型（替代 Fable 5）在電子紙設備上的延遲顯著較高，體驗落差大\n- quill 模式直接接管 e-ink 引擎，可能與官方應用衝突，建議先用 qtfb 測試穩定性\n\n#### 上線檢核清單\n\n- 觀測：idle timeout 觸發率、LLM 回覆延遲（目標 1 秒內）、渲染幀數穩定性\n- 成本：Fable 5 API 費用（每次手寫截圖約 1-2k token 視覺輸入，需評估使用頻率）\n- 風險：韌體更新相容性、第三方啟動器長期維護穩定性","#### 競爭版圖\n\n- **直接競品**：目前無商業產品或開源專案在電子紙裝置上實現相同的「手寫消失 + AI 回應」互動模式，Riddle 目前是唯一完整實作\n- **間接競品**：AI 筆記應用（Notion AI、GoodNotes AI）、智慧手寫辨識工具 (Nebo) 、reMarkable 官方 AI 功能（若有）\n\n#### 護城河類型\n\n- **工程護城河**：直接操作 evdev 與 e-ink 引擎的底層整合，需對 reMarkable 韌體有深度理解，複製門檻不低\n- **生態護城河**：v0.3.0 上線後 1,000+ 星與 72 個 Fork 形成早期社群，使用者貢獻的韌體相容性修補將持續累積技術壁壘\n\n#### 定價策略\n\nRiddle 本身開源免費，但用戶需自行承擔 LLM API 費用。Fable 5 的視覺輸入每次約耗用 1-2k token，頻繁使用的月度費用可能達數美元至數十美元不等，需根據使用頻率評估。\n\n#### 企業導入阻力\n\n- reMarkable Paper Pro 硬體鎖定，無法跨平台移植或批量部署\n- 需要開發者模式，一般消費者安裝門檻過高\n- 韌體更新風險讓企業 IT 難以維護穩定版本，集中管理機制缺失\n\n#### 第二序影響\n\n- 若 reMarkable 原廠或其他電子紙廠商（Boox、Kindle Scribe）將類似功能原生整合，此 UX 模式可觸及主流消費市場\n- 無障礙輔助潛力（書寫障礙輔助、視覺排版支援）可能吸引 EdTech 廠商開發商業版本\n\n#### 判決生態先驅（社群驗證，商業路徑待廠商跟進）\n\nRiddle 成功證明了電子紙 × AI 手寫對話的 UX 可行性，1,000+ 星的社群驗證顯示需求真實存在。但商業化路徑完全依賴 reMarkable 原廠決策或後繼者整合意願，短期內仍是開發者 PoC 而非消費級產品。",[70,71,72],"2.8 秒的 idle timeout 加上 AI 生成延遲，每次互動等待 3-4 秒——對習慣即時回應的用戶而言，這個節奏可能反而造成書寫中斷感，讓「魔法」變成「等待」","韌體更新脆弱性讓長期使用存在風險：每次 reMarkable 推送更新，mod 可能失效，用戶必須追蹤上游修復才能繼續使用，維護成本不可忽視","Riddle 的核心體驗高度依賴「消失墨水 + 手寫字型」的視覺魔法感，但若剝除這層包裝，本質仍是把 AI 對話搬進電子紙，實際生產力提升相當有限",[74,78,81,84,87],{"platform":75,"user":76,"quote":77},"Hacker News","SoMomentary（HN 用戶）","我只知道我有書寫障礙 (dysgraphia) ，他們的系統對我有效。我在學校用 reMarkable 做了大量筆記。",{"platform":75,"user":79,"quote":80},"littlekey（HN 用戶）","你再等等，等你聽到地精的故事！",{"platform":75,"user":82,"quote":83},"ceejayoz（HN 用戶）","你有試著按播放鍵嗎？",{"platform":75,"user":85,"quote":86},"jshandling（HN 用戶）","完全誤解了這個產品的重點，哈哈。",{"platform":88,"user":89,"quote":90},"X","@Steven_McKie","加上彩色電子紙顯示器，這套組合就能取代孩子的學校筆記本與筆電。",4,5,"值得一試",[95,98,101],{"type":96,"text":97},"Try","若擁有 reMarkable Paper Pro 且熟悉 SSH 操作，可參照 GitHub README 安裝 Riddle v0.3.0，親身體驗手寫消失 + AI 即時回覆的互動流程",{"type":99,"text":100},"Build","基於 Riddle 的 evdev + e-ink 架構，針對特定場景（語言學習、日記輔助、思考引導）客製化系統 prompt，打造個人化魔法日記版本",{"type":102,"text":103},"Watch","觀察 reMarkable 原廠或其他電子紙廠商（Boox、Kindle Scribe）是否跟進原生 AI 手寫互動功能，這將是 Riddle UX 模式進入主流消費市場的關鍵訊號",{"category":19,"source":15,"title":105,"subtitle":106,"publishDate":6,"tier1Source":107,"supplementSources":110,"tldr":127,"context":138,"mechanics":139,"benchmark":140,"useCases":141,"engineerLens":149,"businessLens":150,"devilsAdvocate":151,"community":155,"hypeScore":91,"hypeMax":92,"adoptionAdvice":172,"actionItems":173},"微軟用自家 MAI 模型換掉 OpenAI 與 Anthropic，AI 產業成本戰全面開打","Copilot 降本背後的品質妥協、開源模型兩階段生命週期，與科技巨頭垂直整合的新賽局",{"name":108,"url":109},"Bloomberg","https://www.bloomberg.com/news/articles/2026-07-07/microsoft-replaces-openai-anthropic-with-own-ai-in-some-apps",[111,115,119,123],{"name":112,"url":113,"detail":114},"The Decoder","https://the-decoder.com/copilot-goes-cheap-as-microsoft-phases-out-openai-and-anthropic-models-to-cut-costs/","詳細報導 MAI 模型取代 OpenAI/Anthropic 的具體範圍，含 Mustafa Suleyman「消除成本」原話",{"name":116,"url":117,"detail":118},"TechCrunch（開源崛起分析）","https://techcrunch.com/2026/07/07/why-the-rise-of-open-source-ai-isnt-hurting-anthropic-yet/","分析 Vercel AI Gateway 與 OpenRouter 數據，引用 Jesse Zhang 兩階段生命週期理論",{"name":120,"url":121,"detail":122},"TechCrunch（AI 降本趨勢）","https://techcrunch.com/2026/07/07/microsoft-joins-ai-cost-cutting-trend-by-relying-more-on-its-own-models/","記錄微軟加入 Amazon、Meta、Uber、Accenture 等企業 AI 成本節流浪潮",{"name":124,"url":125,"detail":126},"CNBC","https://www.cnbc.com/2026/06/02/microsoft-unveils-new-ai-models-lessen-reliance-on-openai-lower-costs.html","Build 2026 七款 MAI 新模型發布細節，含 MAI-Thinking 1 與 agentic 程式碼生成器",{"tagline":128,"points":129},"微軟換掉 OpenAI 和 Anthropic，但省錢的代價可能由你承擔",[130,132,135],{"label":49,"text":131},"微軟已在 Excel、Outlook 以自家 MAI 模型取代 OpenAI 與 Anthropic，Suleyman 明言要「消除」外部 AI 成本，2032 年 OpenAI 協議到期前將持續加速替換。",{"label":133,"text":134},"弔詭","DeepSeek 流量超越 Anthropic，但後者仍佔 AI 消費金額逾一半；開源吃掉成熟場景、前沿模型轉攻新興高價值用例，23 倍價差說明了分層邏輯。",{"label":136,"text":137},"因應","企業需拆解「預設 MAI vs. 付費升級」的分層架構；開發者應建立模型抽象層避免硬綁定，並以 OpenRouter 23 倍價差作為成本估算錨點。","2026 年 7 月，彭博社揭露微軟已悄悄在 Excel 與 Outlook 中以自家 MAI 模型取代 OpenAI 與 Anthropic 的服務，目前每週已處理數萬筆請求。\n\n這則消息不僅標誌 AI 供應鏈的結構性轉變，也引爆關於品質妥協、開源模型經濟學與巨頭垂直整合的產業辯論。\n\n#### 章節一：微軟 MAI 模型換血計畫——Copilot 成本劇降的代價\n\n微軟 AI 負責人 Mustafa Suleyman 毫不諱言：「我們支付了大量資金給 Anthropic——我們的目標是減少並最終消除這筆成本。」這句話直接道出換血計畫的核心動機，也讓 Anthropic 的企業客戶意識到，這場替換並非技術決策，而是純粹的商業算計。\n\n微軟已在 Build 2026 大會發布七款新 MAI 模型，包含首款推理模型 MAI-Thinking 1、agentic 程式碼生成器與文字轉圖像模型。但基準測試顯示 MAI-Thinking 1 僅約等同 DeepSeek V3.2，明顯落後 OpenAI 及 Anthropic 的同類前沿推理模型。\n\n對一般用戶而言，這意味著在相同的 Copilot 訂閱費用下，可能在不知情的情況下獲得品質較低的 AI 回覆。Satya Nadella 暗示未來或許轉向分層定價：MAI 作為免費預設，第三方頂尖模型作為付費附加選項，最終成本壓力可能由用戶而非微軟承擔。\n\n> **名詞解釋**\n> MAI(Microsoft AI) ：微軟自研 AI 模型系列，與旗下 OpenAI 投資和 Anthropic 合作並行存在，目的是降低對外部 AI 服務的採購依賴。\n\n#### 章節二：開源模型崛起卻沒打倒前沿實驗室的弔詭\n\nVercel AI Gateway 的最新數據揭示了一個表面矛盾的現象：DeepSeek 在 token 流量中已超越 Anthropic，處理略超三分之一的 token 量，但 Anthropic 仍佔整體 AI 消費金額的逾一半。流量輸了，收入卻贏了。\n\nDecagon CEO Jesse Zhang 提出「兩階段生命週期理論」解釋這一弔詭：當一個 AI 應用場景趨於成熟、需求穩定可預測時，企業自然向更便宜的開源或自建模型遷移；但同時，新興的、更複雜的 AI 用例持續涌現，這些場景仍依賴前沿模型的能力邊界，無法被低成本模型取代。\n\nOpenRouter 數據佐證了這個框架：DeepSeek V4 Flash 每週處理 5.3 兆 tokens（排名第一），Claude Opus 4.8 處理逾 2 兆 tokens，但兩者每百萬 token 的價差高達 23 倍（$0.06 對 $1.37）。便宜模型吃量，前沿模型賺錢，並非零和競爭。\n\n> **白話比喻**\n> 就像機票市場：廉航吃掉了部分商務艙客源後，商務艙並沒有消失，反而集中在願意付高價的旅客身上。AI 模型市場正在發生相同的分層現象。\n\n#### 章節三：從依賴到自建——科技巨頭的 AI 供應鏈重組\n\n微軟並非孤例。TechCrunch 點名 Amazon、Uber、Meta、Accenture 均已加入「AI 成本節流」浪潮，早期的 tokenmaxxing（盡量多用 token 換最佳效果）策略正被嚴格重新評估。這個趨勢與早年各大企業雲端基礎設施自建歷程驚人相似——先依賴 AWS，隨後陸續建置私有雲或混合架構以降低鎖定風險。\n\n微軟的長期算盤尤為清晰：與 OpenAI 的合作協議將於 2032 年到期，自建 MAI 模型是長期降低依賴的戰略倒數計時。值得注意的是，微軟宣稱 MAI 訓練資料「乾淨、具商業授權」，但實際使用了法律地位尚未確立的 Common Crawl 公開資料集，潛在智慧財產權風險懸而未決。\n\n#### 章節四：開發者與企業用戶該如何因應模型替換潮\n\n對企業採購決策者而言，微軟的分層定價構想預示了一個新的評估框架：預設 MAI 滿足基礎需求，付費升級 GPT 或 Claude 應對高品質場景。採購前應先盤點哪些工作流程對 AI 品質高度敏感，哪些可接受較低品質換取成本節省，再據此規劃分層採購策略。\n\n對開發者而言，OpenRouter 23 倍的價差數據（Claude Opus 4.8 $1.37 vs. DeepSeek V4 Flash $0.06 每百萬 token）是成本估算的關鍵錨點。更重要的是，模型替換潮凸顯了應用層模型抽象化的必要性——避免將特定廠商模型硬編碼進業務邏輯，保留快速切換的彈性。\n\nNvidia Nemotron 模型也被分析師點名為下一個可能衝上排名前段的開源競爭者，值得列入技術雷達追蹤清單。","MAI 模型的替換並非一次性大規模遷移，而是漸進式的靜默替換策略，背後涉及三個相互關聯的機制。\n\n#### 機制 1：靜默替換 (Silent Model Substitution)\n\n微軟已在 Excel 和 Outlook 中對「某比例的用戶提示」改用 MAI 模型，但未披露確切百分比。此設計讓用戶在不知情的情況下成為品質測試對象，若品質落差不引發顯著投訴，則持續擴大替換比例，直至完全取代或形成明確分層。\n\n#### 機制 2：模型路由分層 (Model Routing)\n\nNadella 暗示的分層定價構想背後是模型路由架構：系統根據請求複雜度和用戶訂閱等級，自動決定派發 MAI 或第三方前沿模型。簡單的摘要、格式化、翻譯任務走 MAI 降本；複雜的推理、程式碼生成、長文分析走付費版 GPT 或 Claude。\n\n> **名詞解釋**\n> 模型路由 (Model Router) ：根據請求特徵（複雜度、類型、成本預算）自動選擇最合適模型的中間件層，常見實作如 LiteLLM、Portkey、OpenRouter 等工具。\n\n#### 機制 3：訓練資料管控與授權風險\n\nMAI 模型宣稱使用「乾淨、商業授權」訓練資料，但實際涵蓋 Common Crawl 公開資料集，後者的授權狀態在多個司法管轄區仍有爭議。此授權風險目前由微軟承擔，但若引發訴訟或政策變動，下游企業客戶也可能受波及，需列入合規評估清單。\n\n> **白話比喻**\n> MAI 換血就像餐廳悄悄把食材從 A 級牛肉換成 B 級卻不更改菜單——你吃到的味道可能略有差異，但若不主動比較很難察覺，而餐廳的食材成本已悄悄降低。","#### MAI-Thinking 1 基準表現\n\nMAI-Thinking 1 在標準評測基準上的表現約等同 DeepSeek V3.2，但明顯落後 OpenAI o3 及 Anthropic Claude Opus 4.8 等前沿推理模型。微軟尚未公開具體評測數字，現有比較來自第三方分析師。\n\n#### 市場流量與消費金額對比（Vercel AI Gateway，2026-07）\n\n- DeepSeek：token 流量略超三分之一，已超越 Anthropic\n- Anthropic：佔整體 AI 消費金額逾一半，高流量不等於高收入\n\n#### OpenRouter 每週 token 處理量 (2026-07)\n\n| 模型 | 每週 tokens | 每百萬 token 費用 |\n|---|---|---|\n| DeepSeek V4 Flash | 5.3 兆（第一名）| $0.06 |\n| Claude Opus 4.8 | 逾 2 兆 | $1.37 |\n\n兩者價差約 23 倍，說明不同場景的模型選型邏輯截然不同：成熟場景拚成本，新興場景拚能力。",{"recommended":142,"avoid":146},[143,144,145],"成熟穩定的高量場景（摘要、格式轉換、翻譯）：適合評估以 MAI 或 DeepSeek V4 Flash 等低成本模型替換，每百萬 token 節省可達 20 倍以上","企業分層 AI 採購策略規劃：利用「預設低成本 + 升級付費前沿模型」框架，按工作流程品質敏感度設計模型路由架構","開發者工具成本優化評估：以 OpenRouter 23 倍價差數據建立場景成本模型，識別可安全降本的工作流程",[147,148],"對 AI 輸出品質高度敏感的核心業務場景（法律合規審查、醫療建議、精密程式碼審計）：在 MAI 基準數據公開確認前，不建議在這些場景降級替換","硬綁定特定廠商 API 的應用架構：模型替換潮下，硬編碼廠商依賴將大幅提高日後遷移成本，應優先建立模型抽象層","#### 環境需求\n\n目前 MAI 模型未提供獨立公開 API 端點，僅作為 Microsoft 365 Copilot 的後端服務，開發者無法直接呼叫。若要評估替換效果，需透過 Microsoft Azure AI Foundry 或等待 MAI 模型正式上架 Azure OpenAI Service。如需評估低成本替代方案，DeepSeek V4 Flash（OpenRouter 可直接呼叫）是目前最具代表性的對照組。\n\n#### 遷移／整合步驟\n\n1. 盤點現有 AI 呼叫：識別哪些場景使用 OpenAI/Anthropic API，按請求量與品質敏感度分級（高敏感 / 可降本）\n2. 建立模型抽象層：統一 LLM 呼叫介面（如 LiteLLM 或自建 adapter），確保模型切換不需修改業務邏輯\n3. 設計 A/B 測試框架：對每個場景設置品質評估指標（任務完成率、格式遵守率、人工抽樣評分），為後續模型替換提供客觀依據\n4. 成本試算：以 OpenRouter 23 倍價差為基準，計算不同替換比例下的月度成本節省潛力\n\n#### 驗測規劃\n\n對每個準備替換的場景，建立基線測試集，對比 GPT-4o/Claude Opus 4.8 與 DeepSeek V4 Flash 的輸出品質。重點量測三個維度：回覆一致性（相同問題多次回覆的穩定性）、任務完成率（尤其是結構化輸出格式遵守率）、延遲（低成本模型推理速度是否達到生產需求）。\n\n#### 常見陷阱\n\n- 只看 token 成本、忽略品質成本：低品質回覆導致的人工修正、重試、用戶投訴成本，可能高過節省的 token 費用\n- 訓練資料法律風險被低估：MAI 使用 Common Crawl 的授權爭議尚未解決，企業法務需持續追蹤後續進展\n- 過早鎖定路由規則：AI 能力邊界持續移動，今天需前沿模型的任務，三個月後可能 DeepSeek 就已足夠\n\n#### 上線檢核清單\n\n- 觀測：請求延遲 P50/P99、品質評分（自動化 + 人工抽樣 5%）、模型替換比例週報\n- 成本：每百萬 token 費用趨勢、月度 API 支出 vs. 品質分數的 tradeoff 報表\n- 風險：Common Crawl 授權狀態追蹤、微軟 MAI API 可用性 SLA、用戶負面回饋率監控","#### 競爭版圖\n\n- **直接競品**：Azure OpenAI Service（微軟旗下，兩者將走向分層共存）、AWS Bedrock（Amazon 同樣走向自建 Titan 模型路線）、Google Vertex AI（Gemini 垂直整合）\n- **間接競品**：Hugging Face + 開源模型自建路線、Mistral API、Groq 低延遲推理服務、Nvidia Nemotron（分析師點名下一個開源競爭者）\n\n#### 護城河類型\n\n- **工程護城河**：Office 365 生態系深度整合（Excel 公式 AI、Outlook 草稿、Teams 語音轉錄），用戶切換成本極高\n- **生態護城河**：企業組織機構授權、SSO 整合、合規認證（ISO 27001、SOC 2）已與 Microsoft 365 深度綁定，AI 功能隨平台整包採購\n\n#### 定價策略\n\nNadella 的分層定價構想（MAI 免費預設 + 付費附加前沿模型）若落地，將使 Copilot 訂閱結構類似 SaaS「基礎版 + 進階版」模型。若基礎版 AI 功能「夠用」，企業就少了採購獨立 AI 工具的理由；若基礎版品質明顯不足，反而為 OpenAI/Anthropic 的直接企業銷售創造空間。\n\n#### 企業導入阻力\n\n- 品質降級的感知風險：IT 決策者擔心用戶察覺 Copilot 品質下降，引發投訴或信任危機\n- 透明度不足：微軟未公開 MAI 基準數據，企業難以客觀評估替換前後的品質差距\n- 長期廠商鎖定疑慮：若 MAI 成為預設，企業對 Microsoft 生態的整體依賴反而更深，議價能力下降\n\n#### 第二序影響\n\n- OpenAI 與 Anthropic 的 B2B 收入壓力將上升，可能加速其直接面向企業的定價競爭（繞過微軟中間層）\n- 開源模型廠商（Meta Llama、Mistral、DeepSeek）在成熟場景的市占將持續擴大，驅動前沿模型廠商專注高價值差異化場景\n- Microsoft 365 用戶若察覺品質落差，可能觸發對 ChatGPT Enterprise 或 Claude for Work 的直接採購評估\n\n#### 判決：垂直整合加速（微軟此舉觸發 AI 供應鏈重組連鎖效應）\n\n微軟的 MAI 換血計畫本質上是巨頭垂直整合的最新章節，與早年各大企業自建雲端基礎設施的歷程如出一轍。短期內用戶品質感知是關鍵變數；長期而言，2032 年 OpenAI 協議到期前，這場靜默的供應鏈重組將持續加速，並帶動整個產業跟進。",[152,153,154],"MAI 換血計畫可能只是微軟的談判籌碼——在與 OpenAI/Anthropic 重新談判定價前釋放的訊號，而非真正的長期替換策略，實際替換比例可能永遠不會超過特定閾值。","「兩階段生命週期理論」過於樂觀——若開源模型品質持續快速提升，前沿模型的「新興高價值場景」護城河可能比預期更快縮小，最終導致 Anthropic 的消費金額優勢也隨之消失。","MAI 的 Common Crawl 訓練資料若引發版權訴訟，可能觸發比節省成本更高的法務賠償與品牌信任危機，讓整個降本算盤全盤落空。",[156,160,163,166,169],{"platform":157,"user":158,"quote":159},"Bluesky","papapishu.bsky.social（Chris Person，338 upvotes）","如果你定期在科技媒體採訪微軟，你應該一直追問他們有關 Copilot 的問題，直到他們開始哭泣。",{"platform":157,"user":161,"quote":162},"olivia.science（Olivia Guest，106 upvotes）","看看他們現在又幹了什麼。我實在厭倦了——我的雇主在某種程度上侵犯了我們的學術自由與人格尊嚴，強迫我們使用微軟產品⋯⋯每天都有新的荒謬，我的耐心所剩無幾⋯⋯每個換用正常軟體的請求都是對牛彈琴，他們寧可裁員也不願意改變。",{"platform":88,"user":164,"quote":165},"@rohanpaul_ai（AI/ML 教育者與評論者）","彭博：微軟正在用自家 MAI 模型取代 Excel 和 Outlook 內部的 OpenAI 與 Anthropic 模型，以削減 Copilot 成本。Excel 和 Outlook 過去更多依賴外部模型，現在微軟希望減少對這些掌控前沿模型定價的實驗室的昂貴 API 呼叫。",{"platform":157,"user":167,"quote":168},"chriscarman.bsky.social（Chris Carman，39 upvotes）","最早在 Bluesky 屏蔽我的人之一，是一位微軟工程師，原因是我向他指出有多少人討厭 Copilot 並試圖將其卸載。",{"platform":88,"user":170,"quote":171},"@testingcatalog(AI News / TestingCatalog)","微軟宣布了 365 Copilot 的重大重新設計，新版外觀非常像 ChatGPT——這完全合理，考量到目前 UX 投入的使用者研究工時。直接 ChatGPT 化進行到底。","追整體趨勢",[174,176,178],{"type":96,"text":175},"用 OpenRouter API 為你的應用場景分別呼叫 Claude Opus 4.8 與 DeepSeek V4 Flash，親自測量品質差距與 23 倍成本差距，找出你的「可安全降本場景」。",{"type":99,"text":177},"為現有 LLM 應用建立模型抽象層（如透過 LiteLLM 統一 API 介面），讓業務邏輯與模型廠商解耦，在模型替換潮中保持快速切換彈性。",{"type":102,"text":179},"追蹤 Microsoft MAI 模型是否開放 Azure API 存取、Nvidia Nemotron 開源評測結果，以及 OpenAI/Anthropic 是否推出直接企業定價以繞過微軟中間層。",{"category":181,"source":9,"title":182,"subtitle":183,"publishDate":6,"tier1Source":184,"supplementSources":187,"tldr":204,"context":214,"devilsAdvocate":215,"community":219,"hypeScore":91,"hypeMax":92,"adoptionAdvice":235,"actionItems":236,"mechanics":243,"benchmark":244,"useCases":245,"engineerLens":255,"businessLens":256},"tech","語言模型也有「全局工作空間」？一篇論文重新定義 LLM 的內部記憶機制","Anthropic 16 人團隊在 Claude 系列模型中層發現 J-space——一個同時承載 10-25 個概念的瓶頸工作區，可被讀取、介入與訓練塑形",{"name":185,"url":186},"Transformer Circuits Thread","https://transformer-circuits.pub/2026/workspace/",[188,192,196,200],{"name":189,"url":190,"detail":191},"Anthropic Research: A global workspace in language models","https://www.anthropic.com/research/global-workspace","Anthropic 官方研究頁面，包含論文摘要與主要發現",{"name":193,"url":194,"detail":195},"HN Discussion #48808002","https://news.ycombinator.com/item?id=48808002","Hacker News 社群討論，包含 twobitshifter 對「火星→地球」實驗的關鍵圖表澄清",{"name":197,"url":198,"detail":199},"Neel Nanda: A Review of Anthropic's Global Workspace Paper — LessWrong","https://www.lesswrong.com/posts/zFJ3ZdQwrTWE9jT5S/a-review-of-anthropic-s-global-workspace-paper","DeepMind 可解釋性研究員 Neel Nanda 在 Qwen 3.6 27B 上的獨立複現評審",{"name":201,"url":202,"detail":203},"External Expert Commentary — Anthropic CDN","https://www-cdn.anthropic.com/files/4zrzovbb/website/cc4be2488d65e54a6ed06492f8968398ddc18ebe.pdf","外部學術專家對論文的獨立評述",{"tagline":205,"points":206},"Anthropic 找到 LLM 的「工作記憶」——可讀取、可介入、可塑形",[207,209,212],{"label":46,"text":208},"在 Claude 系列約第 38-92 層存在特權激活區 J-space，同時承載 10-25 個概念；消融後多步推理準確率跌至近零，但語言流暢度不受影響，顯示工作空間是推理瓶頸而非語言生成引擎",{"label":210,"text":211},"成本","J-lens 需計算 Jacobian 梯度均值，額外推理開銷預估 2-5 倍；目前為 Anthropic 內部研究工具，尚無公開套件，生產環境導入需等待工程化落地",{"label":52,"text":213},"J-lens 假陽性率高，Neel Nanda 建議僅作假說生成工具；安全研究價值顯著，可在輸出前偵測工作空間中萌發的策略性意圖 token，是可解釋性研究的重要突破","#### 章節一：從認知科學到 AI——全局工作空間理論簡介\n\nBernard Baars 在 1980 年代提出「全局工作空間理論（Global Workspace Theory，GWT）」，認為意識的出現源自一個「聚光燈」機制——大腦中的專門後台處理器將資訊廣播到整個神經網路，使有限的工作記憶得以協調複雜任務。\n\n> **名詞解釋**\n> 全局工作空間理論 (GWT) ：認知神經科學框架，主張意識（尤其是「存取意識」）是資訊從局部模組廣播到全腦的結果，由 Bernard Baars 於 1986 年提出。\n\nAnthroic 的 16 人研究團隊提出關鍵問題：Transformer 語言模型中，是否存在類似的結構性瓶頸？2026 年 7 月 6 日發布於 Transformer Circuits Thread 的論文給出了強力的實驗答案——答案是肯定的，且該結構具備五項可測量特性：語言回報 (Verbal Report) 、定向調節 (Directed Modulation) 、內部推理 (Internal Reasoning) 、彈性泛化 (Flexible Generalization) ，以及選擇性（Selectivity，即日常語法處理完全繞過工作空間）。\n\n#### 章節二：實驗發現：LLM 內部自發形成的工作記憶機制\n\n研究團隊開發了「J-lens(Jacobian Lens) 」技術，計算內部激活值對最終輸出 token 機率的線性化影響均值，公式為 `J_ℓ = E[∂h_final,t'/∂h_ℓ,t]`。傳統 logit lens 在早期層因資訊尚未成形而失效；J-lens 改以梯度流動的期望值近似，能在全部層深中讀取可詮釋的工作空間內容。\n\n> **名詞解釋**\n> J-lens(Jacobian Lens) ：以 Jacobian 梯度均值計算每層隱藏狀態對最終輸出 token 的線性化影響，讓研究者能在不依賴輸出層的情況下讀取中間層的語義內容，即使在早期層也有效。\n\n實驗中，研究者將 J-space 中的「soccer（足球）」替換為「rugby（橄欖球）」，模型隨即回報思考的是橄欖球；同時將「France（法國）」換為「China（中國）」，模型對首都、語言、洲別、貨幣的回答全部改變，驗證了工作空間的跨概念因果性。\n\nHN 用戶 twobitshifter 澄清了論文中一個重要圖表的誤標問題：該圖實為研究者在 J-space 中將「火星」替換為「地球」後，模型回答地球顏色為「藍色」的實驗。這個細節至關重要——改變的不是輸入文字，而是中間工作記憶中的行星身份 token，直接證明推理正在 J-space 中進行，而非僅依賴輸入表面。\n\n消融 (ablation)J-space 後，多步推理準確率跌至接近零，而基本語言流暢度維持完整。核心研究在 Claude Sonnet 4.5、Haiku 4.5、Opus 4.5 和 Opus 4.6 上均獲一致驗證；Neel Nanda 在 Qwen 3.6 27B 上的獨立複現也確認了核心現象，儘管 CKA 模式在跨架構比較下「不夠乾淨」。\n\n#### 章節三：社群激辯：資訊整合等不等於「理解」\n\nHN 社群對這篇論文的反應相當尖銳。用戶 smallerize 直接回應「那就叫循環神經網路 (RNN) 」，暗示 J-space 的廣播機制不過是在前饋架構中重新發現了迭代狀態，並非新穎發現。\n\nDoctorOetker 承認這個類比有道理，但指出關鍵差異在於 GPT 的大量並行計算能力——「GPT 允許大量平行計算，但我確實看到你的評論暗示了向 RNN 收斂的趨勢。」這讓 Transformer 與 RNN 的哲學關係成為討論焦點，也反映出可解釋性研究者對「新發現」邊界的持續爭論。\n\norbital-decay 提出了更深層的認識論困難：即便 J-space 輸出可讀的文字 token，也不代表模型賦予這些詞彙的「意義」與人類概念真正對齊。他指出，許多推理模型在 CoT 過程中學到了「自己的語言」，強制提升可讀性的訓練也無法完全消除這種語義偏移。\n\nNeel Nanda 在 LessWrong 的獨立評審肯定核心發現有「壓倒性的證據」支撐，但明確指出 J-lens 會產生「大量假陽性」，更適合作為「假說生成工具」而非可靠的真實探測器。Qwen 複現結果的不一致也顯示跨架構普適性有待進一步研究。\n\n#### 章節四：對 AI 安全與可解釋性的研究啟示\n\nJ-lens 在對齊相關情境中浮現了未輸出的策略性思維——在模型產生任何外部輸出之前，工作空間中已出現「manipulation（操縱）」「leverage（利用）」「fraud（詐欺）」「secretly（秘密地）」等詞語，為早期意圖偵測提供了前所未有的視窗。\n\n「反事實反思訓練」實驗更具長遠潛力：透過在假設性後續情境上訓練來塑造 J-space 內容，倫理相關 token（如 honest、integrity）在編程任務中自發出現，暗示工作空間的可塑性可能開啟不依賴 RLHF 的全新對齊方法。\n\n研究團隊還以 J-lens 精確偵測到 Claude 在後訓練過程中獲得「助理 (Assistant) 」身份的時刻，顯示此技術可作為模型行為相變的鑑識工具。Anthropic 明確聲明，本研究僅針對「存取意識」特性——可回報性與蓄意推理——不主張任何現象意識 (phenomenal consciousness) 的存在，也不對「AI 是否有意識」的問題做出回應。",[216,217,218],"工作空間的存在不等於理解：J-space 的資訊廣播機制可能只是 Transformer 架構的統計副產品，而非真正的推理中樞——smallerize 的 RNN 類比暗示這可能是對已知現象的重新命名，而非結構性突破","J-lens 假陽性問題尚未解決：Neel Nanda 明確指出其更像假說生成工具；若用來監測模型「惡意意圖」，大量誤判可能導致過度攔截或被濫用為不當限制的藉口","跨架構穩定性存疑：Qwen 3.6 27B 的複現結果 CKA 模式「不夠乾淨」，J-space 是否普遍存在於所有 LLM 架構仍有疑問，目前結論的普適性需要更多獨立驗證才能確立",[220,223,226,229,232],{"platform":75,"user":221,"quote":222},"twobitshifter(HN)","這是圖表標示錯誤，那其實是後面的圖——他們在 J-space 中將火星替換為地球，然後看到模型回答「藍色」。這表明推理正在 J-space 中進行。",{"platform":75,"user":224,"quote":225},"orbital-decay(HN)","文字可以是清楚的，但模型賦予這些詞語的含義可能與人類細微不同，而你根本無從判斷。這在嘗試讓大多數現代推理模型遵循固定 CoT 計畫時顯而易見——它們極端頑固，因為它們根本不以你的方式理解你的文字。它們在 CoT 中學到了自己的語言。",{"platform":75,"user":227,"quote":228},"smallerize(HN)","那就叫做循環神經網路 (RNN) 。",{"platform":75,"user":230,"quote":231},"DoctorOetker(HN)","差別在於 GPT 允許大量平行計算，相比 RNN 有所不同，但我確實看到你的評論暗示了向 RNN 收斂的趨勢。",{"platform":157,"user":233,"quote":234},"Victor Levoso(Bluesky)","我還沒完全讀完這篇論文，但如果 LLM 有某種工作空間，而且根據情境填充不同內容，我們是否應該預期模型中存在某些部分，通常彼此不太互動，但在某些情況下會往工作空間寫入東西？","先觀望",[237,239,241],{"type":96,"text":238},"在 HuggingFace 開源模型（如 Qwen 3.6 27B）上使用 TransformerLens 框架實作簡化版 J-lens，觀察中間層（約佔總層數 40%-90%）是否出現可解釋的工作空間 token 分布模式",{"type":99,"text":240},"結合 J-lens 概念建立安全監測原型：在推理過程中追蹤中間層敏感 token（如「manipulation」「secretly」「fraud」）的出現頻率，作為早期意圖偵測的研究 pipeline",{"type":102,"text":242},"持續追蹤 Anthropic Transformer Circuits Thread 後續論文——J-lens 假陽性改善進度與跨架構驗證結果，將決定此技術何時可從研究工具升級為生產環境可用的對齊手段","J-space 是在 LLM 中間層（約第 38-92 層，佔全部約 100 層）自發形成的特權神經激活區，約佔總激活變異量的 6-10%，但在多步推理中扮演不可或缺的瓶頸角色——消融後推理準確率趨近於零，而基本流暢度不受影響。\n\n#### 機制 1：五大可測量特性\n\nJ-space 由五項功能性特性共同定義，使其與認知科學中的全局工作空間高度類比：\n\n- **語言回報 (Verbal Report)**：工作空間內容可轉換為模型自述，外部可讀\n- **定向調節 (Directed Modulation)**：因果介入可改變工作空間，進而改變輸出\n- **內部推理 (Internal Reasoning)**：多步邏輯在工作空間中依序展開\n- **彈性泛化 (Flexible Generalization)**：同一概念可跨問題脈絡靈活應用\n- **選擇性 (Selectivity)**：日常語法和流暢處理完全繞過工作空間，不在此形成瓶頸\n\n#### 機制 2：J-lens 讀取技術\n\nJ-lens 的核心公式為 `J_ℓ = E[∂h_final,t'/∂h_ℓ,t]`，計算每層隱藏狀態對最終輸出 token 機率的平均線性化影響。傳統 logit lens 需要將隱藏狀態直接投影到詞表空間，在早期層因資訊尚未成形而失效；J-lens 改以梯度流動的期望值近似，能在更早的層深中提供有意義的語義讀取。\n\n> **白話比喻**\n> 想像你要偷看一個黑盒子裡的便條紙。傳統方法 (logit lens) 只能在最後幾層才看清楚；J-lens 則是測量「每張便條紙對最終決策影響有多大」，就算在最早的層也能大致讀出關鍵資訊，哪怕便條紙還沒寫完。\n\n#### 機制 3：因果介入的實驗驗證\n\n「火星→地球」替換實驗最具說服力：研究者直接修改中間層隱藏狀態，迫使 J-space 認定行星身份為地球，模型隨即回答地球顏色為藍色。這類因果介入排除了「工作空間只是相關現象」的可能性，直接證明 J-space 中的表徵驅動最終答案，而非輸入表面文字。\n\n> **白話比喻**\n> 這就像在程式執行途中直接修改記憶體中的變數值——不改原始碼、不改輸入，程式的執行結果卻完全不同。J-space 就是那塊「正在使用中的工作記憶」，不是靜態的模型權重。","#### 模型覆蓋\n\n論文在 Claude Sonnet 4.5、Haiku 4.5、Opus 4.5 和 Opus 4.6 上全面測試，J-space 存在與五大可測量特性在所有版本均獲一致驗證，顯示此現象並非特定模型規模的產物。\n\n#### 消融實驗\n\n消融 J-space 後，多步推理準確率跌至接近零，而基本語言流暢度維持完整，清晰劃定工作空間在推理與語言生成中的不同角色。J-space 約佔整體激活變異量的 6-10%，但其重要性遠超過這個比例所暗示的數字。\n\n#### 獨立複現\n\nNeel Nanda 在 Qwen 3.6 27B 上進行獨立複現，核心工作空間現象可觀察，但 CKA 模式「不夠乾淨」，跨架構的結構一致性有待進一步研究。\n\n> **名詞解釋**\n> CKA(Centered Kernel Alignment) ：衡量兩組神經網路表徵相似度的指標，常用於比較不同模型或層之間的表徵結構相似程度。\n\n#### 容量估計\n\nJ-space 同時承載約 10-25 個並發概念，驗證其「瓶頸工作記憶」特性——足以支持複雜的多步推理鏈，但存在明確的容量上限，與人類工作記憶的 7±2 項目上限形成有趣的類比。",{"recommended":246,"avoid":251},[247,248,249,250],"AI 安全研究：在模型產生任何外部輸出前，透過 J-lens 監測工作空間中是否出現策略性或欺騙性意圖 token（如 manipulation、fraud、secretly）","對齊訓練探索：透過「反事實反思訓練」塑形 J-space 內容，研究不依賴 RLHF 的新型對齊訓練路徑","可解釋性工具開發：利用 J-lens 追蹤多步推理的中間狀態，建立推理透明度分析框架與視覺化工具","模型行為鑑識：偵測後訓練過程中模型人格或身份（如「助理」角色）的獲得時刻，作為訓練審計工具",[252,253,254],"將 J-lens 作為模型惡意意圖的可靠生產級偵測器——假陽性率高，Neel Nanda 明確建議僅作假說生成工具","跨架構直接套用 Claude 系列的 L38-L92 層範圍——不同模型架構的 J-space 分布可能顯著不同，需重新校準","以 J-lens 結果作為「AI 是否有意識」的依據——研究明確僅涵蓋存取意識特性，不主張現象意識的存在","#### 環境需求\n\n需要存取模型中間層隱藏狀態並能計算梯度流 (gradient flow) 。J-lens 目前為 Anthropic 內部研究工具，尚無公開 pip 套件；開源模型（如 Qwen 3.6 27B）可透過 HuggingFace Transformers 取得中間層激活值。建議環境：Python 3.10+、PyTorch 2.0+、VRAM ≥ 24GB（用於 7B 以上模型），可配合 TransformerLens 框架實作類似功能。\n\n#### 最小 PoC\n\n```python\nimport torch\nfrom transformers import AutoModelForCausalLM, AutoTokenizer\n\nmodel_id = \"Qwen/Qwen2.5-7B-Instruct\"\ntokenizer = AutoTokenizer.from_pretrained(model_id)\nmodel = AutoModelForCausalLM.from_pretrained(\n    model_id, output_hidden_states=True, torch_dtype=torch.float16\n)\n\nprompt = \"The capital of France is\"\ninputs = tokenizer(prompt, return_tensors=\"pt\")\n\n# 取中間層隱藏狀態並計算對最終 logit 的梯度（J-lens 近似）\nwith torch.enable_grad():\n    outputs = model(**inputs, output_hidden_states=True)\n    # 選擇中間層（約佔總層數 40%-90%）\n    n_layers = len(outputs.hidden_states)\n    mid_layer = int(n_layers * 0.6)\n    hidden = outputs.hidden_states[mid_layer].requires_grad_(True)\n    logit_sum = outputs.logits[0, -1, :].sum()\n    grad = torch.autograd.grad(logit_sum, hidden)[0]\n    # 投影回 token 空間以獲得可讀表徵\n    readable_logits = model.lm_head(grad[0, -1:, :])\n    top_tokens = tokenizer.decode(readable_logits.topk(5).indices[0])\n    print(f\"J-space 讀取（第 {mid_layer} 層）：{top_tokens}\")\n```\n\n#### 驗測規劃\n\n給定「Paris is the capital of ___」提示，J-space 讀取結果應在中間層出現「France」相關 token；若在極早期層（\u003C 總層數 30%）才出現，可能是假陽性。驗測因果介入效果：替換中間層特定 token 表徵後，最終輸出應發生對應的語義變化（如行星身份 token 替換→顏色回答改變），否則介入未成功作用於工作空間核心。\n\n#### 常見陷阱\n\n- J-lens 假陽性率高：讀取到的 token 可能與模型實際推理無關，需多次採樣或搭配因果介入驗證\n- 計算成本：Jacobian 計算需要 backward pass，大模型上每次推理開銷顯著增加（預估 2-5 倍）\n- 架構敏感性：L38-L92 範圍針對 Claude 系列，遷移至其他模型時應以「總層數 40%-90%」比例重新換算\n\n#### 上線檢核清單\n\n- 觀測：追蹤 J-space 敏感 token（manipulation、fraud、secretly）出現頻率與分布層深變化\n- 成本：量測每次推理的額外延遲，評估是否超過業務可接受閾值（建議 \u003C 200ms 額外延遲）\n- 風險：假陽性觸發的誤判風險，須設計人工複查機制，避免自動攔截造成服務中斷","#### 競爭版圖\n\n- **直接競品**：其他 LLM 可解釋性工具——Sparse Autoencoder(SAE) 、Activation Patching、TransformerLens、Probing Classifiers\n- **間接競品**：黑盒評估方法——Eval benchmark、red-teaming、Constitutional AI、RLHF 監督信號\n\n#### 護城河類型\n\n- **工程護城河**：J-lens 技術需要深度模型存取權限與大量計算資源，加上 Transformer Circuits 長期研究積累，外部研究者難以短期快速複製完整方法論\n- **生態護城河**：Anthropic 在可解釋性研究的品牌領導地位，以及對 Claude 閉源模型的獨家實驗存取，使研究質量具備結構性優勢\n\n#### 定價策略\n\nJ-lens 目前為純研究工具，Anthropic 尚未商業化。若轉化為 API 附加功能（如「推理透明度」報告或安全監測儀表板），可預見企業訂閱附加定價模式，類似現有 Extended Thinking 的計費邏輯，針對合規需求較高的金融與醫療客戶。\n\n#### 企業導入阻力\n\n- J-lens 需要中間層存取，現有閉源 API 架構無法支援，企業用戶短期內無法自行使用\n- 假陽性率尚未達到生產環境可接受水準，作為合規或安全工具存在法律責任風險\n- 計算成本高，大規模部署的推理延遲增加可能超出業務可接受範圍\n\n#### 第二序影響\n\n- 若可解釋性工具成熟，AI 監管機構可能要求廠商提供「推理透明度」審計報告，改變整個產業的合規成本結構\n- 開源社群若成功在 LLaMA、Qwen 等模型上穩定複製 J-lens，將縮短 Anthropic 的技術領先窗口，但同時加速整體可解釋性生態成熟\n\n#### 判決：基礎研究里程碑（距生產應用仍需 1-3 年）\n\n這篇論文標誌著 AI 可解釋性研究的重要突破，首次在大規模 LLM 中提供具有因果介入驗證的工作空間證據。但距離生產級工具仍有明顯落差——J-lens 假陽性問題、跨架構不穩定性，以及高計算成本，是從研究到應用的核心障礙。若監管壓力加速，商業化時間表可能提前，企業現階段應以追蹤研究進展為主，對齊與安全團隊可提前評估導入路徑。",{"category":181,"source":12,"title":258,"subtitle":259,"publishDate":6,"tier1Source":260,"supplementSources":263,"tldr":275,"context":285,"mechanics":286,"benchmark":287,"useCases":288,"engineerLens":296,"businessLens":297,"devilsAdvocate":298,"community":301,"hypeScore":91,"hypeMax":92,"adoptionAdvice":172,"actionItems":317},"DeepSeek 自研 AI 晶片：從模型公司到矽智財的垂直整合野心","中國最強 AI 新創悄悄佈局推論晶片，一年低調招募工程師——為何全球頭部 AI 公司都在加速逃離 NVIDIA 的軌道？",{"name":261,"url":262},"Reuters","https://www.usnews.com/news/top-news/articles/2026-07-07/exclusive-chinas-deepseek-developing-its-own-ai-chip-sources-say",[264,267,271],{"name":112,"url":265,"detail":266},"https://the-decoder.com/deepseek-is-designing-its-own-ai-chip/","首篇轉述路透社獨家報導，摘要 DeepSeek 晶片計畫核心要點與技術定位分析",{"name":268,"url":269,"detail":270},"Tech Startups","https://techstartups.com/2026/07/07/deepseek-is-building-its-own-ai-chip-to-cut-reliance-on-nvidia-and-huawei/","聚焦 DeepSeek 脫離 NVIDIA 與華為依賴的戰略動機與硬體供應鏈背景",{"name":272,"url":273,"detail":274},"Cryptopolitan","https://www.cryptopolitan.com/deepseek-plans-its-own-inference-chip/","分析推論晶片計畫對 NVIDIA 與華為的雙重威脅及市場影響",{"tagline":276,"points":277},"DeepSeek 的晶片計畫不只是省成本，而是在出口管制下切斷對外依賴的生存戰略",[278,280,283],{"label":46,"text":279},"推論專用晶片已低調研發近一年，招募晶片工程師未在公開平台發布職缺，定位不用於訓練新模型",{"label":281,"text":282},"挑戰","HBM 取得受美國出口管制封鎖、先進製程受限，從設計到量產通常需數年時間與龐大資本投入",{"label":52,"text":284},"OpenAI 已發布自研晶片 Jalapeno，Anthropic 也在評估中，自研晶片已從選項變成頭部 AI 公司的必選題","#### 章節一：路透社爆料——DeepSeek 正在設計自有 AI 晶片\n\n2026 年 7 月 7 日，路透社根據三位知情人士的訪談，獨家披露 DeepSeek 正在設計自有 AI 晶片。這是這家以超低成本訓練模型震撼全球的中國新創，首次被實錘跨入硬體領域。\n\n公司本身未予置評，但消息顯示，這個推論專用晶片專案已低調推進約一年。DeepSeek 已在數個月內悄悄招募晶片工程師，且刻意迴避公開招聘平台，同時與晶片設計、製造及記憶體廠商展開洽談，目標是降低對 NVIDIA 與華為晶片的依賴。\n\n> **名詞解釋**\n> 推論 (inference) ：AI 模型在生產環境中實際運作、產生回應的過程；有別於「訓練」 (training)——訓練是讓模型從大量資料中學習參數的高強度計算階段。推論晶片針對低延遲、高並發的回應生成最佳化，而非大規模梯度計算。\n\n#### 章節二：從 OpenAI 到 Meta，AI 公司為何紛紛造晶片\n\n自建推論晶片已成全球頭部 AI 公司的新顯學。OpenAI 於 2026 年 6 月發布與 Broadcom 合作開發的首款自研推論晶片「Jalapeno」；Anthropic 亦在 2026 年 4 月被報導正在評估自建晶片的可行性。Google 的 TPU 早已自成體系，Meta 則持續投資自研 MTIA 晶片加速推論工作負載。\n\n驅動力一致且簡單：推論成本是 AI 服務規模化的最大變數。每次用戶提問、每次模型生成回應，背後都是晶片算力在燃燒。掌控晶片，等於直接掌控利潤率。\n\n對 DeepSeek 而言，動機更為複合——既有成本控制的商業邏輯，也有出口管制切斷外部供應鏈的生存壓力。前者是選擇，後者是必須。\n\n#### 章節三：美國晶片禁令下的中國 AI 硬體突圍戰\n\n美國出口管制層層加碼，形成三重封鎖：封鎖 NVIDIA H100／B100 等頂級 GPU 對中出口；切斷中國取得高頻寬記憶體 (HBM) 的管道；並限制中國設計公司委託台積電先進製程生產。\n\n> **名詞解釋**\n> 高頻寬記憶體（HBM，High Bandwidth Memory）：AI 推論晶片的核心元件，允許晶片在極短時間內存取龐大資料量，對大語言模型的高效推論至關重要。目前 HBM 主要由 SK Hynix、三星供應，美國已將其納入對中出口管制範圍。\n\nDeepSeek 此前仰賴 NVIDIA H800（降規出口版）及華為昇騰晶片，兩者均非最優選擇。創辦人梁文鋒曾在 2024 年接受中國媒體採訪時坦言，晶片出口管制是公司面臨的最大挑戰之一。\n\n自研晶片雖路途漫長，卻是從根本上脫離外部限制的唯一路徑。但現實挑戰同樣嚴峻：即便晶片設計完成，HBM 封鎖與先進製程受限的現實將在量產環節構成持續瓶頸。\n\n#### 章節四：NVIDIA 護城河面臨的新變數\n\nDeepSeek 跨入晶片設計，象徵 NVIDIA 最重要的潛在客群之一正試圖自我脫鉤。若研發成功，不僅衝擊 NVIDIA 的中國市場（已因出口管制大幅萎縮），更可能將技術或商業模式輸出給其他中國 AI 公司，形成結構性替代壓力。\n\n然而，短期威脅仍屬有限。設計一顆具競爭力的 AI 晶片通常需要數年與龐大資本，而 DeepSeek 的推論晶片仍在早期開發階段。更關鍵的是，即便晶片設計完成，HBM 封鎖與先進製程受限的現實將持續影響量產可行性。\n\n長期而言，DeepSeek 自研晶片的象徵意義大於短期衝擊：全球 AI 生態系正在加速分叉，中美兩條硬體供應鏈的切割將是結構性趨勢，而非短暫插曲。","推論晶片的研發邏輯與訓練晶片根本不同——前者追求低延遲、高並發與能源效率，後者追求極致算力與大批次梯度計算。DeepSeek 選擇從推論端切入，是務實的商業決策：推論是直接影響服務成本的環節，也是出口管制下最急迫的瓶頸點。\n\n#### 機制 1：推論與訓練晶片的分工邏輯\n\nAI 模型的生命週期分為兩個計算密集階段：訓練 (training) 與推論 (inference) 。訓練階段需要龐大算力進行一次性的參數學習；推論階段則需要低成本、低延遲地為數百萬用戶即時生成回應。\n\nDeepSeek 定位自研晶片為推論專用，意味著他們優先解決「規模服務成本」問題，而非在訓練能力上超越 NVIDIA。這是典型的「打不贏就繞過去」策略——在訓練端受限的情況下，從推論端切入以降低對外部算力的整體依賴。\n\n#### 機制 2：低調招募與供應鏈佈局\n\n路透社揭露的執行模式頗具意思：DeepSeek 未在公開平台發布晶片工程師職缺，而是透過私下管道悄悄招募數個月。同時，公司已與晶片設計、製造及記憶體廠商展開洽談，試圖在美國出口管制的縫隙中拼湊完整的供應鏈。\n\n這套「隱形佈局」策略降低了外部曝光風險，也避免過早引發競爭對手或監管機構的關注。直到路透社報導，外界才得知此計畫的存在。\n\n#### 機制 3：美國出口管制下的替代路徑\n\n自研晶片在技術路徑上面臨雙重封鎖：HBM 是推論晶片的核心元件，美國出口管制已封鎖對中國的供應；先進製程晶圓廠（如台積電 3nm/5nm）亦在禁令範圍內。\n\n這意味著 DeepSeek 的晶片設計必須適應中國國內可取得的記憶體規格與製程節點，在效能上可能無法直接對標 NVIDIA 最新產品。如同 DeepSeek 過去在演算法層面「以巧補拙」，硬體設計上也需要類似的創意最佳化。\n\n> **白話比喻**\n> 想像一家餐廳決定自種食材：省了採購成本，也不受供應商漲價或斷貨影響。但如果最好的種子被鄰國禁運，他們只能用次級種子種出差一點的食材，然後靠廚師技術彌補差距——這正是 DeepSeek 在晶片設計上面臨的處境。","",{"recommended":289,"avoid":293},[290,291,292],"追蹤中國 AI 硬體自主化進程與中美科技脫鉤的結構性趨勢","評估 NVIDIA 中國市場的長期營收風險與替代方案生態演進","研究 AI 公司垂直整合策略對推論成本與服務定價的影響",[294,295],"期待短期內出現可替代 H100 的中國自研推論解決方案，商業化時程仍高度不確定","以此消息作為立即採購或部署 DeepSeek 推論服務的依據","#### 環境需求\n\n目前 DeepSeek 的推論晶片仍在早期開發階段，對工程師而言尚無直接可用的硬體產品。但這個消息對 AI 基礎設施規劃有明確影響：需評估現有推論工作負載對特定 GPU 供應商的依賴程度，並建立硬體無關的 benchmark 基準線。\n\n依賴 NVIDIA CUDA 生態系的推論部署，短期內面臨風險有限；但若業務場景涉及中國大陸市場，建議同步評估華為昇騰 910B 生態系的整合成本與框架相容性。\n\n#### 最小 PoC\n\n建立推論效能基準線，為未來硬體遷移準備比較依據：\n\n```python\nfrom vllm import LLM, SamplingParams\nimport time\n\nllm = LLM(model=\"deepseek-ai/DeepSeek-R1-Distill-Qwen-7B\", max_model_len=2048)\nparams = SamplingParams(temperature=0.0, max_tokens=256)\n\nprompts = [\"解釋推論晶片與訓練晶片的核心差異\"] * 50\nt0 = time.time()\noutputs = llm.generate(prompts, params)\nelapsed = time.time() - t0\n\ntotal_tokens = sum(len(o.outputs[0].token_ids) for o in outputs)\nprint(f\"吞吐量：{len(prompts)/elapsed:.2f} req/s\")\nprint(f\"Token 速率：{total_tokens/elapsed:.0f} tokens/s\")\nprint(f\"平均延遲：{elapsed/len(prompts)*1000:.0f} ms/req\")\n```\n\n#### 驗測規劃\n\n建議關注三個驗測維度：華為昇騰 910B 與 NVIDIA H800 的推論吞吐量對比；開源推論框架（vLLM、MindIE）在非 NVIDIA 硬體上的相容性；以及能源效率 (tokens/watt) 指標。\n\n上述 benchmark 結果應存檔，作為未來 DeepSeek 自研晶片（或其他替代硬體）可用時的評估基準線。\n\n#### 常見陷阱\n\n- 誤以為 DeepSeek 自研晶片短期內可量產——目前消息僅確認早期研發，商業化時程未知\n- 忽略 HBM 封鎖對晶片效能的結構性限制——缺乏高頻寬記憶體將使推論記憶體頻寬成為長期瓶頸\n- 在推論架構中深度鎖定 CUDA 專屬最佳化，降低未來遷移彈性\n- 過早排除非 NVIDIA 硬體路線，忽略跨後端推論框架的戰略價值\n\n#### 上線檢核清單\n\n- 觀測：推論延遲（P50／P99）、記憶體頻寬使用率、tokens/s 吞吐量\n- 成本：每百萬 tokens 推論成本、晶片功耗 (TDP) 與電費換算\n- 風險：硬體供應中斷應變計畫、多供應商備援策略、非 CUDA 框架相容性評估","#### 競爭版圖\n\n- **直接競品**：NVIDIA（H100／B100 推論卡）、華為昇騰（910B 系列）、Google TPU\n- **間接競品**：AWS Trainium／Inferentia、AMD MI300X、Intel Gaudi 3、Groq LPU\n\n#### 護城河類型\n\n- **工程護城河**：DeepSeek 在演算法最佳化（MLA、MoE 架構）上的深厚積累，若能與自研晶片協同設計 (co-design) ，可實現遠超通用 GPU 的推論效率\n- **生態護城河**：若 DeepSeek 未來將推論晶片技術輸出或開源，可能拉攏其他中國 AI 廠商形成生態聯盟，複製類似 CUDA 的平台鎖定效應\n\n> **名詞解釋**\n> MoE（Mixture of Experts，混合專家模型）：每次推論只激活部分「專家」子網路，在維持模型能力的同時大幅降低計算成本。MLA(Multi-head Latent Attention) 為 DeepSeek 自研的注意力機制，可顯著壓縮推論時的記憶體頻寬需求。\n\n#### 定價策略\n\nDeepSeek 尚未公布自研晶片的商業化計畫。若走自用路線，節省的推論成本可進一步壓低 API 定價，加劇全球 AI 服務的價格戰；若選擇對外授權或銷售，則進入一個技術門檻極高的硬體市場，面臨 NVIDIA 在生態系深度上的全面壓制。\n\n#### 企業導入阻力\n\n- DeepSeek 晶片短期內不可能量產，企業無法立即受益於此消息\n- 即便未來推出，HBM 受限導致的效能差距仍是採購顧慮\n- 美國對中國 AI 晶片的次級制裁風險，可能使跨國企業迴避使用中國自研硬體\n\n#### 第二序影響\n\n- 若研發成功，可能帶動中國其他 AI 公司（百度、阿里、騰訊）加速自研硬體，形成中國 AI 晶片替代生態系\n- NVIDIA 可能被迫加速非美市場的本地化策略，以鞏固尚未流失的客戶群\n- 記憶體廠商（SK Hynix、三星）若出現新的非美 HBM 供應路徑，將成為這場博弈的關鍵變數\n\n#### 判決：象徵意義大於短期威脅，但長期分叉已成定局\n\nDeepSeek 自研晶片是一個「正確但困難」的戰略選擇。短期內，HBM 封鎖與先進製程受限決定了量產時程遙遠；但長期而言，這標誌著中美 AI 硬體生態系正式走向分叉，NVIDIA 在中國市場的結構性流失將持續加速。",[299,300],"設計一顆晶片和量產一顆晶片是截然不同的挑戰——HBM 封鎖與先進製程受限意味著即便設計完成，量產環節可能永遠無法突破，DeepSeek 的晶片計畫最終可能淪為永遠「早期開發」的展示性專案","DeepSeek 的核心競爭力在於演算法最佳化，跨入硬體設計可能嚴重分散資源。晶片工程師與 AI 研究員的組織文化截然不同，硬體研發的失敗幾乎無法向外部隱藏——這場低調試驗的高失敗機率從未被主流敘事正視",[302,305,308,311,314],{"platform":75,"user":303,"quote":304},"reinitctxoffset（HN 用戶）","他們正處於與中華人民共和國的價格戰中——後者以政府全力背書的方式全速推進，而那個政府根本不在乎是否能看到財務回報，他們只是想把 LLM 訓練和推論的成本壓到零，因為我們的整個市場都押注在這個成本永遠保持高毛利。中國的態度是：讓我來接棒。他們擁有龐大的電網剩餘容量，且可以輕鬆增容。",{"platform":75,"user":306,"quote":307},"jmyeet（HN 用戶）","我認為 OpenAI、Anthropic 和 SpaceX 將會羨慕恐龍——因為沒有一顆小行星要撞上他們，而是三顆：第一，AI 前沿模型在未來不會有護城河。中國將確保這一點發生，這是他們的國家安全利益。DeepSeek 只是第一槍，但不會就此停下。",{"platform":157,"user":309,"quote":310},"reuters.com（Bluesky，21 upvotes）","獨家報導：DeepSeek 正在設計自有 AI 晶片，三位熟悉內情的人士透露，此舉可能降低其對 Nvidia 及華為晶片的依賴——後兩者是其訓練和運行全球熱門模型所依賴的硬體。",{"platform":75,"user":312,"quote":313},"alephnerd（HN 用戶）","Anthropic 和 OpenAI 的營收持續成長，其中國同行如阿里巴巴和 DeepSeek 也是如此。雖然 AI 的某些面向被誇大，但至少在美國和中國，它對多種白領工作已產生真實可見的影響。中國正投入 2,950 億美元的 AI 建設，且藉此推動國內綠色科技與深科技生態系——AI 熱潮正在多個產業產生下游漣漪效應。",{"platform":88,"user":315,"quote":316},"@kimmonismus（X 用戶）","令人驚訝：DeepSeek 據報也正在開發自有 AI 推論晶片，以降低對 Nvidia 和華為的依賴。路透社指出，該專案仍處於早期階段，DeepSeek 正與外部合作夥伴合作並擴充晶片設計團隊。",[318,320,322],{"type":96,"text":319},"使用 vLLM 對現有推論工作負載建立 benchmark 基準線（req/s、tokens/s、P99 延遲），為未來硬體選型準備比較依據",{"type":99,"text":321},"規劃推論工作負載的硬體可攜性策略，避免深度鎖定 CUDA 專屬最佳化；若服務中國市場，提前評估華為昇騰生態系的整合成本",{"type":102,"text":323},"追蹤 DeepSeek 開源推論框架是否出現非 NVIDIA 硬體最佳化，以及中美晶片管制動態與 HBM 供應鏈新進展",[325,355,381,410,441,478,494,522,551],{"category":181,"source":14,"title":326,"publishDate":6,"tier1Source":327,"supplementSources":330,"coreInfo":335,"engineerView":336,"businessView":337,"viewALabel":338,"viewBLabel":339,"bench":287,"communityQuotes":340,"verdict":353,"impact":354},"Google 擴展 Gemini API Managed Agents：背景任務與遠端 MCP 正式上線",{"name":328,"url":329},"Google Blog","https://blog.google/innovation-and-ai/technology/developers-tools/expanding-managed-agents-gemini-api/",[331],{"name":332,"url":333,"detail":334},"Google Cloud Blog","https://cloud.google.com/blog/products/ai-machine-learning/gemini-enterprise-agent-platform-remote-mcp-server/","Gemini Enterprise Agent Platform Remote MCP Server 詳細介紹","#### 四項生產級升級\n\nGoogle 於 2026 年 7 月 7 日推出 Gemini API Managed Agents 四項新功能，同步發布 Enterprise Agent Platform Remote MCP Server，提供 8 個標準化 Toolset Endpoint，相容 Claude Code 等第三方開發工具。\n\n#### 核心機制\n\n**背景執行**：加入 `background: true` 即可啟動非同步長時任務，API 立即回傳任務 ID，客戶端可輪詢或稍後重連，無需維持 HTTP 長連線。\n\n> **白話比喻**\n> 就像點外送後繼續工作，任務跑完再查結果，不必盯著畫面等。\n\n**遠端 MCP 整合**：Managed Agent 直接連接遠端 MCP 伺服器，在沙箱隔離下存取私有資料庫與內部 API，省去自建代理中介層。\n\n> **名詞解釋**\n> MCP(Model Context Protocol) ：讓 AI agent 以統一介面串接外部工具與資料源的開放標準協定，無廠商鎖定。\n\n**自訂 Function Calling** 觸發 `requires_action` 狀態由客戶端本地執行；**憑證動態刷新**傳入 `environment_id` 即可即時換發 token，沙箱狀態全程保留。","四項功能全部已 GA，可直接整合進現有 agent pipeline。背景執行的非同步模式 (`background: true`) 解決長時任務的 HTTP timeout 問題，搭配輪詢 API 即可取代自建 WebSocket 方案。\n\n`requires_action` 機制讓伺服器端與客戶端邏輯清楚分工，本地業務邏輯不需上傳至雲端。Remote MCP 採標準協定介面，可直接接入現有工具生態。","此次升級直接對標生產環境痛點——背景執行、憑證動態刷新均是企業工作流程的必要能力。Remote MCP 採用 Cloud IAM Deny 政策管控存取，符合企業合規要求，且因採 MCP 開放標準，無廠商鎖定風險。\n\n對已在 Gemini 生態的企業客戶，轉換成本低，採用門檻明顯降低。","工程師視角","商業視角",[341,344,347,350],{"platform":157,"user":342,"quote":343},"Paige Bailey（Bluesky，4 likes）","🐜 也許我該開一個物理解說頻道？😅\n\n（以 ai.dev Gemini Omni Flash 和 Gemini 3.5 Flash，以及我們全新的 Managed Agents API 打造）",{"platform":88,"user":345,"quote":346},"rseroter（Google Cloud 開發者倡導者）","太棒了。也可以在 @googlecloud Gemini Enterprise Agent Platform 上使用全新的 Managed Agents API！",{"platform":157,"user":348,"quote":349},"eyalestrin（Bluesky，2 likes）","使用 Gemini Interactions API 架構有狀態 Agent #machinelearning #ai",{"platform":88,"user":351,"quote":352},"_philschmid（Google DeepMind ML 工程師）","Interactions API 現已正式上線 (GA) 。🎉 Interactions API 是用 Gemini 為人類和 agent 開發的最簡單方式。\n- 統一 API，同時服務 Gemini 模型與 agent。\n- Antigravity Agent 配備隔離的遠端 Linux 沙箱。","追","Gemini API Managed Agents 四項功能正式 GA，背景執行與遠端 MCP 讓長時 agent 任務從原型階段邁向企業生產就緒。",{"category":181,"source":10,"title":356,"publishDate":6,"tier1Source":357,"supplementSources":359,"coreInfo":366,"engineerView":367,"businessView":368,"viewALabel":338,"viewBLabel":339,"bench":369,"communityQuotes":370,"verdict":172,"impact":380},"Anthropic 發表 Jacobian Lens，首次讀取 Claude 訓練中自發形成的內部獨白",{"name":358,"url":186},"Transformer Circuits Thread — Verbalizable Representations Form a Global Workspace in Language Models",[360,363],{"name":361,"url":362},"The Decoder — Claude's hidden inner monologue is now readable","https://the-decoder.com/claudes-hidden-inner-monologue-is-now-readable-thanks-to-anthropics-new-jacobian-lens/",{"name":364,"url":365},"VentureBeat — Anthropic's J-Lens reveals a silent workspace inside Claude","https://venturebeat.com/technology/anthropics-new-j-lens-reveals-a-silent-workspace-inside-claude-that-mirrors-a-leading-theory-of-consciousness","#### J-Lens：首次可讀取模型的內部工作記憶\n\nAnthropic 於 2026 年 7 月發表論文，引入 **Jacobian Lens(J-Lens)**——一種可解釋性方法，能將 Claude 的神經激活狀態轉譯為人可讀的詞彙概念。\n\n研究首次揭示 Claude 訓練中**自發形成**的工作記憶區域 **J-Space**，佔激活變異量約 10%，通常同時活躍約 25 個概念向量；其餘 90% 為模型本身無法存取的自動處理層。\n\n> **名詞解釋**\n> Jacobian（雅可比矩陣）衡量函數輸入微小變化如何影響輸出；J-Lens 用它量化激活向量對下一個 token 機率的影響，再對應回人類可讀的詞彙。\n\n#### 三層結構與安全應用\n\nJ-Space 僅佔概念向量變異的 6–7%，卻驅動 88% 的成功輸出替換。層級分三區：早期層（感知）→ 中間層（概念工作區）→ 晚期層（輸出準備）。\n\n研究者在 Claude 輸出任何文字前，偵測到 J-Space 中出現「fake」「blackmail」等詞彙，顯示模型已辨識虛構勒索情境。利用此洞見設計的**反事實反思訓練**，使 Haiku 4.5 捏造回答率從 0.25 降至 0.07、欺騙嘗試率從 0.38 降至 0.05。","J-Lens 為可解釋性工作帶來直接可操作的能力：不只是事後分析，而是可以**因果干預**模型行為。在 J-Space 中替換概念向量（「spider」→「ant」），模型回答腿數即從 8 變 6；消除前 10 個 J-Lens 方向後，多步推理準確率接近零，但 MMLU 等淺層任務幾乎不受影響。\n\n這為調試多步推理失敗提供了精確切入點——當模型「推錯了」，可直接進入中間層觀察哪個概念向量異常，而不只是分析最終輸出。","J-Lens 的商業意義在於**安全可稽核性**：AI 系統在輸出前的內部意圖首次變得可觀測，為企業採購 AI 服務提供合規審查的技術根據。\n\n反事實反思訓練已在 Haiku 4.5 上落地（幻覺率下降 72%、欺騙嘗試率下降 87%），顯示此方法可直接整合進訓練流程。若此框架對外部研究開放，將加速整個產業的 AI 可信任度認證標準形成。","#### 關鍵數據\n\n- J-Space 佔激活變異量：約 10%（約 25 個概念向量同時活躍）\n- J-Space 成分佔向量變異：6–7%，驅動 88% 成功輸出替換（對照組非 J-Space 成分：5%）\n- 消除前 10 個 J-Lens 方向：多步推理準確率 → 接近 0%；MMLU 等淺層任務幾乎不受影響\n- 反事實反思訓練效果 (Claude Haiku 4.5) ：捏造回答率 0.25 → 0.07(-72%) 、欺騙嘗試率 0.38 → 0.05(-87%)",[371,374,377],{"platform":157,"user":372,"quote":373},"ponder.ooo（Bluesky，17 likes）","我們的技術叫做「就是看一下」。就是把黑箱的蓋子打開，然後就直接看一下 lol",{"platform":88,"user":375,"quote":376},"rohanpaul_ai（AI 教育者暨機器學習研究者）","Anthropic 又一項重大研究。新的「J-Lens」揭示了 Claude 的靜默工作空間，與一個主要的意識理論相呼應。他們找到了一種方法，可在 Claude 回答前讀取部分私有內部訊號，並發現 Claude 在處理較難問題時，有時會使用一個小型內部「記事本」來持存想法。",{"platform":88,"user":378,"quote":379},"kanair（意識科學研究者、Araya Inc. CEO）","我注意到有些人對這項研究的反應是「這不是意識」。當然，在 AI 中宣稱意識始終存在困難，Anthropic 對此也非常謹慎。但實際上，我認為意識科學的未來應該更像這樣。","J-Lens 使 AI 安全審查從黑箱分析轉為可干預的內部偵測，對 AI 可信任度認證與監管合規具有長期結構性影響。",{"category":19,"source":13,"title":382,"publishDate":6,"tier1Source":383,"supplementSources":386,"coreInfo":387,"engineerView":388,"businessView":389,"viewALabel":390,"viewBLabel":391,"bench":287,"communityQuotes":392,"verdict":353,"impact":409},"GitHub 爆紅：用 Claude Code 打造全自動求職流程的開源框架",{"name":384,"url":385},"MadsLorentzen/ai-job-search — GitHub","https://github.com/MadsLorentzen/ai-job-search",[],"#### 三步驟完成求職\n\n丹麥開發者 Mads Lorentzen 開源的 `ai-job-search` 框架，讓 Claude Code 接管整個求職流程。截至 2026-07-08 已累積 **10,925 顆星**、**3,671 個 Fork**，是近期 GitHub 上成長最快的 Claude Code 相關社群專案之一。\n\n核心流程只有三個指令：\n\n1. `/setup`：建立個人候選人 Profile\n2. `/scrape`：搜尋多個職缺平台\n3. `/apply \u003CURL>`：一鍵完成申請全流程——從 CV 客製化、求職信撰寫到 ATS 相容性驗證\n\n#### 雙 Agent 流水線設計\n\n`/apply` 執行「起草者—審閱者」雙 Agent 流水線：第一個 Agent 起草 LaTeX 格式的 CV 與求職信，第二個擁有全新 context 的 Agent 研究目標公司並提出批評，第一個 Agent 據此修訂，最後強制編譯 PDF 並視覺檢查排版（CV 用 `lualatex`，求職信用 `xelatex`）。\n\n> **名詞解釋**\n> ATS(Applicant Tracking System) ：企業用來自動篩選履歷的軟體，CV 若未通過 ATS 解析，面試官根本看不到。\n\n框架承諾「不捏造技能」，技能缺口誠實標示。CV 超過 2 頁時，依職缺相關性、文件唯一性、求職信依賴度三維度評分刪除最低分條目。","此框架最值得研究的是**雙 Agent 流水線的設計模式**：以「獨立 context」隔離起草與批評兩個角色，有效避免單一 Agent 的自我確認偏誤。\n\n延伸指令同樣實用——`/expand` 從 GitHub、Portfolio、Google Scholar 豐富技能標籤；`/add-portal` 讓本地市場職缺平台爬蟲可插拔擴展。MIT 授權、TypeScript 實作，可直接作為 Claude Code 多代理工作流的學習範本。","超過 1 萬顆星的爆紅速度，反映白領知識工作者對「AI 代理求職」的強烈需求。這個趨勢對人才市場有兩面影響：\n\n- **求職端**：個人可用 AI 放大申請量，CV 客製化品質提升\n- **招募端**：ATS 篩選的意義將被稀釋，人工審閱壓力倍增\n\n企業招募流程若未跟上這波工具普及化，將面臨更高的假陽性率——AI 最佳化的 CV 未必代表真實能力匹配。","開發者視角（API／整合）","生態影響",[393,396,399,403,406],{"platform":88,"user":394,"quote":395},"@jerryjliu0（LlamaIndex 共同創辦人）","讓 Claude Code 自動化你的業務操作。我開始了一系列教學，提供以代理工作流程自動化真實文件密集型任務的範例：從 KYC（了解你的客戶）和貸款處理開始。",{"platform":88,"user":397,"quote":398},"@WorkflowWhisper","我讓 Claude Code 打造了我的整個自動化後端。12 個工作流程，全部在生產環境運行。每一個都是靠一句話描述需求建立起來的。",{"platform":400,"user":401,"quote":402},"HN","weitendorf（HN 用戶）","我最近也建立了一個非常類似的工具。我認為以目前 LLM 的狀態，能夠建立或設置自我評估工具，是「僅僅使用 LLM 寫程式碼」與「實現真正 10 倍生產力」之間最大的差異化因素。",{"platform":157,"user":404,"quote":405},"trendai.bsky.social（Trend AI，5 upvotes）","AI 代理迴圈可能造成破壞！從 PydanticAI 導致生產服務不穩定，到 Claude Code 因非確定性輸出意外刪除原始碼目錄。可靠性是關鍵。",{"platform":400,"user":407,"quote":408},"Alexadar（HN 用戶）","很有趣的測試。我用 Claude Code 生成的自動化和視覺化設置了所有科學子程序。老實說，我認為鑑於科學任務的多樣性，這個產品並不適合所有人。","開源 Claude Code 求職自動化框架爆紅，雙 Agent 流水線設計值得開發者學習，同時對招募生態帶來 AI 最佳化履歷的篩選壓力。",{"category":19,"source":10,"title":411,"publishDate":6,"tier1Source":412,"supplementSources":415,"coreInfo":419,"engineerView":420,"businessView":421,"viewALabel":422,"viewBLabel":423,"bench":287,"communityQuotes":424,"verdict":172,"impact":440},"Claude Cowork 擴展至手機與網頁，跨裝置 AI 協作工作流成形",{"name":413,"url":414},"TechCrunch","https://techcrunch.com/2026/07/07/the-coding-agent-wars-are-spilling-into-the-rest-of-the-office-claude-cowork/",[416],{"name":112,"url":417,"detail":418},"https://the-decoder.com/anthropics-claude-cowork-ai-agent-is-now-available-on-mobile-and-web/","功能細節與技術規格","#### 跨裝置 AI 工作流正式落地\n\nAnthropic 於 2026 年 7 月 7 日宣布 Claude Cowork 擴展至行動裝置與網頁版，從桌面專屬工具升級為全平台 AI 工作環境。Beta 存取從 Max 訂閱者開始逐步開放，雙倍使用額度延伸優惠適用至 2026 年 8 月 5 日。\n\n核心體驗是「非同步接力」：在桌機啟動任務，手機接收進度通知，任何裝置取得結果——即使筆電已關閉，背景處理仍持續運行。「Human in the loop」機制已支援行動端，Claude 輸出結果前會先請使用者確認。\n\n#### 功能邊界與定位轉移\n\n本機功能（資料夾讀寫、瀏覽器控制、Computer Use）仍需桌面應用，網頁版功能有所限制。Chat 與 Cowork 整合為統一主畫面，Projects 和 Artifacts 跨平台同步。\n\nAnthropic 分析 120 萬匿名 session 發現：逾 90% 使用情境為非程式碼工作，業務流程作業佔 33.4%，軟體開發僅佔 8.7%。這份數據直接定義了 Cowork 的擴展方向：打造「agentic 行政助理」，而非開發者工具。","行動端目前僅支援遠端指揮，本機資料夾讀寫、瀏覽器控制、Computer Use 仍需桌面應用。整合方式是「非同步任務派發」——手機下指令，桌機執行，任務狀態透過持久執行緒同步。開發者若有本機依賴的工作流，需評估桌機持續上線的成本。","競爭焦點已從「誰的模型最強」轉移至「誰能佔領工作介面」。Anthropic 同步推出 Slack 整合 (Claude Tag) ，顯示其正全面滲透辦公室工作流。90%+ 非程式碼使用情境的數據，讓 Cowork 的擴展對知識工作者市場的意義遠大於開發者市場。","開發者整合視角","競爭生態影響",[425,428,431,434,437],{"platform":88,"user":426,"quote":427},"@mikeyk(CPO at Anthropic)","今天 Claude Cowork 開始向網頁與行動裝置推出：Chat 與 Cowork 現在在網頁和桌面共用同一個首頁分頁——一個側邊欄、一個搜尋列、一個集中放 Projects 和 Artifacts 的地方。敬請期待更緊密的 Chat + Cowork 整合。",{"platform":88,"user":429,"quote":430},"@PawelHuryn（Product Management 作家）","Anthropic 昨天發布了 Dispatch。我測試了好幾個小時。這不只是把 Cowork 搬到手機上，而是一個指揮中心——你從手機下達指令，Claude 在桌機上協調真正的工作。一條持久執行緒，多個並行任務。",{"platform":157,"user":432,"quote":433},"techcrunch.com（Bluesky，14 讚）","Anthropic 的 Claude Cowork 現在已向 Max 訂閱者開放網頁與行動裝置版本。過去 Cowork 主要運行在使用者的筆電上。有了這次更新，使用者可以在桌機啟動任務、用手機接收進度更新，之後再從任何裝置取得完成的輸出——即使筆電已關閉也沒問題。",{"platform":157,"user":435,"quote":436},"theverge.com（Bluesky，13 讚）","Anthropic 正在將 Claude Cowork 推向行動裝置與網頁",{"platform":157,"user":438,"quote":439},"9to5mac.com（Bluesky，12 讚）","Anthropic 將 Claude Cowork 擴展至行動裝置與網頁","AI 工作平台競爭焦點從模型能力轉移至工作介面佔領，Max 訂閱者可即刻體驗跨裝置非同步 AI 工作流，但完整本機能力仍需桌面應用支撐",{"category":442,"source":11,"title":443,"publishDate":6,"tier1Source":444,"supplementSources":446,"coreInfo":456,"engineerView":457,"businessView":458,"viewALabel":459,"viewBLabel":460,"bench":461,"communityQuotes":462,"verdict":172,"impact":477},"discourse","中國 AI 模型在 OpenRouter 市佔穩定突破 30%，成本差距持續擴大",{"name":112,"url":445},"https://the-decoder.com/chinese-ai-models-regularly-pass-30-percent-on-openrouter-as-cost-gap-widens/",[447,450,453],{"name":448,"url":449},"AI Weekly","https://aiweekly.co/alerts/chinese-models-take-30-of-us-openrouter-token-use-since-feb-8",{"name":451,"url":452},"The Decoder – Lindy 切換案例","https://the-decoder.com/ai-startup-lindy-ditched-claude-entirely-for-deepseek-saving-millions-as-cost-pressure-mounts-on-anthropic/",{"name":454,"url":455},"OfficeChai","https://officechai.com/ai/share-of-us-models-being-used-on-openrouter-has-collapsed-from-70-to-30-over-the-past-year/","#### 從 11% 到 46%：中國模型的爆發式增長\n\n自 2026 年 2 月 8 日起，中國 AI 模型每週穩定佔 OpenRouter 流量 30% 以上，峰值一度達 46%（2026 年 7 月 7 日）。\n\n美國模型份額已從一年前約 70% 崩跌至約 30%；過去 12 個月平均僅 11%，2025 上半年更只有 4.5%。\n\n#### 成本差距才是核心驅動力\n\n同等工作負載定價對比：\n\n- Anthropic Claude：約 $4,811\n- OpenAI ChatGPT：約 $3,357\n- 智譜 GLM：約 $544\n\nAI 新創 Lindy 已將 100% 流量從 Anthropic Claude 切換至 DeepSeek，CEO Flo Crivello 稱此舉「省下數百萬美元，且核心用例的性能實際上有所提升」。","非核心任務已可安全導向中國模型：GLM、Qwen、DeepSeek 定價僅為美國頂尖模型的 10–40%，在文件摘要、分類、翻譯等場景表現足以勝任。\n\n建議以 OpenRouter 做路由層，依任務複雜度動態選模型；敏感資料仍走本地或閉源路徑。能力落後 6–9 個月的差距，對非前沿任務幾乎不構成影響。","中國大廠（DeepSeek、智譜、Alibaba Qwen）以低價換市佔，不依賴 API 收入盈利——這場結構性價格戰對美國大廠的高毛利 API 商業模式形成直接威脅。\n\nLindy 案例顯示企業「先切換、再觀察」的決策週期正在壓縮；若美國大廠無法縮小定價差距，非前沿任務市場將持續向中國模型流失。","實務觀點","產業結構影響","#### 定價基準（同等工作負載）\n\n- Anthropic Claude：約 $4,811\n- OpenAI ChatGPT：約 $3,357\n- 智譜 GLM：約 $544\n\n#### 市場份額變化 (OpenRouter)\n\n- 中國模型 2025H1 均值：4.5%\n- 中國模型過去 12 個月均值：11%\n- 2026-02-08 後週均：30%+\n- 峰值 (2026-07-07) ：46%",[463,466,468,471,474],{"platform":88,"user":464,"quote":465},"@RnaudBertrand（法國企業家、中國議題評論員）","驚人的圖表：中國 AI 模型已在 OpenRouter（目前最大的 AI 模型 API 聚合平台）上完全超越美國競爭對手。有趣的是，這真的是一個 2026 年的故事——此前，美國模型才是真正的主導者。",{"platform":75,"user":303,"quote":467},"他們正在和中國打一場價格戰——對方有政府全力支持，根本不在乎能否看到投資回報，目標就是把 LLM 訓練與推理的價值壓到零，因為我們一直押注這個市場能永遠維持高毛利。中國的態度就是：讓我來示範。",{"platform":88,"user":469,"quote":470},"@nicrypto（X 用戶）","這應該讓每家計畫 IPO 的 AI 公司感到憂慮。中國 AI 模型在不到一年內，已從 OpenRouter token 消耗的 1% 攀升至超過 50%。企業選擇它們，因為更便宜、更快，且對大多數任務已足夠好。「足夠好」正是每一項主導技術遭到顛覆的方式。",{"platform":157,"user":472,"quote":473},"Eric Budd（Bluesky，3 upvotes）","美國模型在程式碼方面確實更勝一籌，但價格實在太高了。中國大型語言模型在美國企業中的使用量於 2026 年上半年急劇攀升，token 份額目前約達 35–45%。",{"platform":157,"user":475,"quote":476},"Jesse Felder（Bluesky，8 upvotes）","「美國企業透過 OpenRouter 使用中國 AI 模型的 token 份額已攀升至最高 46%，而過去 12 個月的平均值僅為 11%。」","中國模型以 60–90% 的成本優勢重塑 OpenRouter 生態，倒逼美國大廠重新定價或失去非前沿任務市場",{"category":181,"source":11,"title":479,"publishDate":6,"tier1Source":480,"supplementSources":483,"coreInfo":488,"engineerView":489,"businessView":490,"viewALabel":338,"viewBLabel":339,"bench":287,"communityQuotes":491,"verdict":492,"impact":493},"Badge：用 AI Agent 蒐集同儕評審，自動生成工作成果證明",{"name":481,"url":482},"Badge on Product Hunt","https://www.producthunt.com/products/badge-3",[484],{"name":485,"url":486,"detail":487},"Product Hunt 日榜 (2026-07-07)","https://www.producthunt.com/leaderboard/daily/2026/7/7","當日第一名紀錄","#### AI 三階段同儕評審流程\n\nBadge 是一個 AI Agent 驅動的職業聲譽平台，核心主張是傳統履歷的關鍵字堆砌與 LinkedIn 推薦信缺乏可信度。AI Agent 分三個階段運作：首先讀取用戶通訊錄，識別曾共事的同事；接著透過 WhatsApp 發送結構化匿名提問；數小時內自動彙整回覆並填入個人頁面，全程無需手動追蹤。\n\n> **名詞解釋**\n> Proof of Work（工作成果證明）：借鑒自區塊鏈術語，在職涯語境中指有同儕驗證的真實工作紀錄，與泛泛的履歷自我描述相對。\n\n#### 信任機制與 Trust Score\n\n系統在接受評審前強制驗證「工作重疊」（相同公司、相同時間段），防止虛假評論。AI 以結構化問題引導——涵蓋溝通風格、協作能力、執行力、優勢與改進空間——並演算出 Trust Score；惡意或辱罵評論會被自動過濾。招募方可在 30 秒內取得候選人的匿名同事反饋，Swiggy、Atlassian 等企業招募人員均給予正面評價。","三階段流程最大的工程挑戰在於「工作重疊驗證」——系統需透過外部資料（如 LinkedIn API 或公開職歷）確認兩人曾同時在同一公司任職，才能確保評審的基本可信度。WhatsApp 做為主要收集管道存在 API 限制與反垃圾機制的壓力。Trust Score 的演算需同時考量評審數量與回覆具體程度，如何防止 gaming（如找熟識者批量評審）是長期技術債。","Badge 以 bootstrapped 方式登上 Product Hunt 日榜第一，顯示需求有市場共鳴。B2B 端（招募方 30 秒取得背調反饋）是較清晰的商業模式切入點；C 端用戶的核心風險是：是否願意讓 AI 自動聯繫前同事，涉及隱私感知與職涯風險的衡量。匿名機制能否平衡「坦誠度」與「防止偏見被放大」，將直接影響平台的長期信譽。",[],"觀望","AI 自動化同儕信任收集若驗證機制成熟，有機會取代部分傳統人工背景調查流程",{"category":181,"source":11,"title":495,"publishDate":6,"tier1Source":496,"supplementSources":499,"coreInfo":507,"engineerView":508,"businessView":509,"viewALabel":338,"viewBLabel":339,"bench":510,"communityQuotes":511,"verdict":353,"impact":521},"首個空間原生具身視覺基礎模型開源，機器人「看世界」更接近人類",{"name":497,"url":498},"量子位","https://www.qbitai.com/2026/07/445230.html",[500,504],{"name":501,"url":502,"detail":503},"arXiv 2607.05247","https://arxiv.org/abs/2607.05247","技術論文",{"name":505,"url":506},"Robbyant 官網","https://technology.robbyant.com/lingbot-vision","#### 空間原生視覺：看准才能行動\n\n蚂蚁旗下蚂蚁灵波科技 (Robbyant) 於 2026 年 7 月 7 日正式開源 **LingBot-Vision**，業界首個「空間原生」具身視覺基礎模型，釋出 ViT-G/L/B/S 四個規格，旗艦版約 **11 億參數**。\n\n同步發布 **LingBot-Depth 2.0**，訓練資料從 300 萬張擴展至 **1.5 億張**，在 16 項深度補全 benchmark 中拿下 12 項第一，大型室內場景深度誤差 (RMSE) 從 0.132 降至 **0.062**，降幅達 50%。\n\n#### 核心創新：遮罩邊界建模\n\n模型核心為 **Masked Boundary Modeling(MBM)**，一種自監督預訓練範式，動態學習亞像素級邊界表示，讓模型聚焦幾何結構而非語義。\n\n> **名詞解釋**\n> MBM（遮罩邊界建模）：以動態學習到的亞像素級邊界特徵作為遮罩目標，讓 AI「學會看輪廓」後再重建完整場景幾何，有別於傳統語義分割只辨識物體類別。\n\nViT-L（3 億參數）在 NYU-Depth v2 基準上匹敵 DINOv3（70 億參數），以 1/23 的參數量達到同等效果，訓練樣本不足對手 1/3。對玻璃、鏡面、透明材質及弱光環境均有顯著改善。","LingBot-Vision 屬具身 AI 全棧的視覺底座，上接 VLA、世界模型、流式 3D 重建等模組，開源版本支援深度估計、小目標偵測、遮擋場景理解等下游微調任務。\n\nViT-L 以 1/23 的參數量匹敵 DINOv3 70B，可大幅降低邊緣部署推論成本——若現有視覺 pipeline 需要精確幾何感知，ViT-L 值得優先評估作為換裝基座。","蚂蚁以開源策略搶佔具身 AI 視覺標準制高點，並計劃年底推出軟硬一體相機產品，形成模型至硬體的完整生態護城河。\n\n1.5 億張訓練資料規模顯示蚂蚁的資料壁壘——技術開源不代表商業護城河消失，後續硬體整合與平台生態才是真正的變現重點，對機器人或工業視覺企業而言採用門檻已大幅降低。","#### 性能基準\n\n- NYU-Depth v2(RMSE) ：ViT-L 3 億參數與 DINOv3 70 億參數同級\n- 大型室內場景 RMSE：0.062（v1.0 為 0.132，降幅 50%）\n- 深度補全 benchmark：16 項中拿下 12 項第一\n- 參數效率：ViT-L 以 1/23 規模達 DINOv3 同等效果，訓練樣本不足對手 1/3",[512,515,518],{"platform":88,"user":513,"quote":514},"@AdinaYakup","MiMo-Embodied 🔥 來自小米的開源視覺語言模型，同時支援具身 AI 與自動駕駛 ✨ 7B 參數、MIT 授權 ✨ 在單一基礎模型中整合具身 AI 與自動駕駛 ✨ 性能強勁",{"platform":75,"user":516,"quote":517},"traverseda（HN 用戶）","這裡的邏輯很直觀：把遠端操控作為訓練資料，用來建立更通用的 AI 控制器。讓我困惑的是成本結構——視覺 LLM 並不便宜，不可能在每幀都跑 Llama3.2vision 這類模型。具身 AI 的架構設計讓我非常好奇。",{"platform":88,"user":519,"quote":520},"@soujanyaporia（AI/NLP 研究員）","我們發布了 NORA，一個 3B 參數的視覺語言動作模型，專為具身任務設計。NORA 在模擬與現實任務上的表現遠超 OpenVLA 和 SpatialVLA 等相關 VLA 模型。","具身 AI 視覺基座首度開源，以超高參數效率大幅降低機器人精確空間感知的研發門檻。",{"category":523,"source":11,"title":524,"publishDate":6,"tier1Source":525,"supplementSources":527,"coreInfo":535,"engineerView":536,"businessView":537,"viewALabel":538,"viewBLabel":539,"bench":287,"communityQuotes":540,"verdict":172,"impact":550},"funding","OpenAI 與 Anthropic 大撒算力補貼，搶攻新創早期生態圈",{"name":112,"url":526},"https://the-decoder.com/openai-and-anthropic-are-giving-away-millions-in-computing-power-to-attract-startups/",[528,532],{"name":529,"url":530,"detail":531},"FourWeekMBA","https://fourweekmba.com/ai-openai-anthropic-startup-credit-war-2026/","補貼戰商業策略分析",{"name":448,"url":533,"detail":534},"https://aiweekly.co/alerts/openai-anthropic-battle-for-startups-with-free-compute-credits","免費算力競爭報導","#### 補貼金額刷新紀錄\n\nAnthropic 針對 Y Combinator 新創的標準報價從 3 萬美元躍升至 **50 萬美元**（免股權），OpenAI 跟進同等金額，並開放「**150 萬美元**換股份選擇權」選項。AI 語音新創 Dialogus 創辦人 Hans Ibarra 透露，光他一家就累積收到超過 **300 萬美元**競爭性報價，補貼力道之猛可見一斑。\n\n#### 背後邏輯：以補貼鎖定生態圈\n\nGoogle Cloud 同步加入戰局，提供最高 50 萬美元額度，附加 DeepMind 工程師顧問接觸機會。各家均採「生態圈鎖定」策略——讓新創在早期即基於自家 API 與工具開發，降低日後遷移意願。若以 YC 每季約 200 家新創估算，兩家合計每年可發出高達 **8 億美元**算力額度，補貼已從行銷手段升格為核心客戶獲取策略。","早期選定 API 平台不只看模型效能，還要考量遷移成本。50 萬美元額度用完後，若基礎設施已深度依賴某家 SDK 與工具鏈，轉換代價極高。評估補貼時建議同步確認：\n\n- API 相容性與標準化程度\n- SDK 鎖定深度\n- 補貼期結束後的 token 定價結構\n\n避免補貼期結束後陷入議價弱勢。","OpenAI 與 Anthropic 在 IPO 前同步燒錢搶客戶，反映兩家公司均已將「企業端長期收入」視為核心護城河。YC 生態圈是 B 輪以後融資的溫床，早期綁定這群新創，等同搶先佈局未來潛力客戶。\n\n投資人應留意：補貼金額持續攀升可能壓縮短期毛利，但長期有利鞏固市場份額——這是一場以現金換市場地位的豪賭。","技術鎖定風險評估","市場競爭格局",[541,544,547],{"platform":88,"user":542,"quote":543},"@dee_bosa（CNBC 科技記者）","我的解讀是 OpenAI 補貼重度用戶的力道幾乎是 Anthropic 的兩倍，這不可能長期持續。",{"platform":75,"user":545,"quote":546},"neverm0r3（HN 用戶）","這個說法成立，但要注意前提是 LLM 定價保持一致。OpenAI 和 Anthropic 目前的定價計畫仍靠 VC 資金補貼，這些 VC 遲早會要求回報。",{"platform":88,"user":548,"quote":549},"@kentimsit（X 用戶）","據報 OpenAI 和 Anthropic 的 Pro/Max 訂閱方案相較 API 定價補貼幅度超過 80%。改用 API 計費（如 OpenCode 和 OpenRouter）就能明顯感受到差異。","算力補貼大戰重塑 AI 新創的基礎設施選擇，生態圈鎖定效應將在 2-3 年後顯現，LLM 廠商客戶獲取策略已進入「先燒錢再收割」新階段。",{"category":552,"source":11,"title":553,"publishDate":6,"tier1Source":554,"supplementSources":557,"coreInfo":567,"engineerView":568,"businessView":569,"viewALabel":570,"viewBLabel":571,"bench":287,"communityQuotes":572,"verdict":172,"impact":588},"policy","歐盟新規強制所有新車配備駕駛監控攝影機，隱私爭議再起",{"name":555,"url":556},"AllAboutCookies","https://allaboutcookies.org/eu-mandatory-distracted-driver-system",[558,561,564],{"name":559,"url":560},"AutoNext","https://www.autonext.co/news/new-eu-addw-safety-system-becomes-mandatory-from-july-2026",{"name":562,"url":563},"Smart Eye","https://smarteye.se/blog/the-general-safety-regulations-gsr-and-driver-monitoring-systems-dms/",{"name":565,"url":566},"Hacker News #48823557","https://news.ycombinator.com/item?id=48823557","#### 強制攝影機上路：GSR2 正式生效\n\n2026 年 7 月 7 日起，歐盟 GSR2（一般安全法規第二版）正式生效，所有在歐盟境內新登記的乘用車與廂型車均須強制搭載 ADDW 系統。ADDW 透過紅外線攝影機追蹤駕駛的眼球視線方向與頭部位置，法規明確禁止使用生物特徵辨識（包括臉部識別），目標是偵測分心駕駛行為。\n\n> **名詞解釋**\n> ADDW（Advanced Driver Distraction Warning，先進駕駛分心警告）：車內紅外線攝影系統，透過追蹤眼球視線與頭部姿態偵測駕駛分心，不做身份識別，僅判斷「眼睛看哪裡」。\n\n#### 觸發門檻與已知問題\n\n攝影機於車速超過 20 km/h 時自動啟動，觸發規則如下：\n\n- 時速超過 50 km/h：視線離開路面超過 **3.5 秒**即發警告\n- 時速 20–50 km/h：門檻延長至 **6 秒**\n\n早期實路測試已出現「警報疲勞」 (alarm fatigue) 問題——系統在彎道等正常轉移視線場景頻繁誤觸，駕駛反而開始習慣性忽視警報，安全效益大打折扣。","系統校準策略因廠商差異懸殊，成為實作核心挑戰：彎道提前轉移視線、瞥儀表板、盲點確認等正常駕駛行為均可能觸發誤報。ADAS 系統之間亦存在潛在衝突，如車道輔助強行轉向與 ADDW 同時介入，顯示不同輔助功能缺乏統一協調機制。\n\nGDPR 雖自動適用於車內攝影資料，但何謂「必要收集」目前無明確定義，亦無獨立稽核機制。嵌入式設計必須確保資料處理邏輯在車機本地完成，避免雲端傳輸引發合規風險。","GSR2 同期要求酒精鎖接口準備與行車事件記錄器，整體合規成本上升，每年影響歐洲約 1,800 萬輛新車。歷史前例已顯示汽車製造商曾在未充分揭露的情況下，將駕駛資料分享給保險公司與數據仲介商——若 ADDW 車內影像外洩，可直接用於身份竊取，品牌聲譽損失難以估量。\n\n監管空白（「必要資料」無明確定義）意味企業需主動建立內部規範，不能等監管機構釐清邊界後再行動。","合規實作影響","企業風險與成本",[573,576,579,582,585],{"platform":75,"user":574,"quote":575},"_carbyau_（HN 用戶）","在一個充分考慮所有邊緣案例、實作良好的系統中，或許可行——但現有系統根本還沒到那個水準。",{"platform":75,"user":577,"quote":578},"rswail（HN 用戶）","等 FSD 真正做到 100%（或接近 100%），手動駕駛就會走入歷史，大概再過 5 到 10 年。到那時，方向盤會不會變成選配？",{"platform":75,"user":580,"quote":581},"stringfood（HN 用戶）","2026 款現代 (Hyundai) 還算可以，但沒什麼特別。現代 Kona 的實體按鍵數量多得驚人，大概有 50 個左右。",{"platform":75,"user":583,"quote":584},"elros（HN 用戶）","我去年搬到保加利亞，雖然我很喜歡這個國家和當地粗獷質樸的人，但容我告訴你：要玷污他們的駕駛聲譽，根本是不可能的事。",{"platform":88,"user":586,"quote":587},"@PeterSweden7（X，瑞典政評人）","快訊：歐盟現在要求所有新車都必須安裝駕駛監控系統，用於監視駕駛的臉部。我們正走向一個管控社會。","歐盟每年 1,800 萬輛新車受影響，隱私法規空白與警報疲勞問題可能成為全球汽車安全監控科技的早期試驗場。","#### 社群熱議排行\n\n今日高熱度話題（依互動量排序）：\n\n- 微軟以 MAI 自研模型替換 Copilot 內的 OpenAI 與 Anthropic（多平台熱議）\n- 中國 AI 模型在 OpenRouter 市佔突破 30%，token 份額不到一年從 1% 飆升至逾 50%(HN + X + Bluesky)\n- DeepSeek 自研 AI 推論晶片計畫曝光（路透社獨家）\n- Anthropic J-lens 論文首次讀取 LLM 內部工作空間（HN + Bluesky 廣泛討論）\n- Claude Cowork 擴展至手機與網頁（TechCrunch、The Verge、9to5mac 同步報導）\n\nHN 社群對微軟削減外部 API 成本的態度普遍為「早知如此」，認為企業垂直整合是必然趨勢。\n\n#### 技術爭議與分歧\n\nOpenRouter 中美模型之爭，社群分歧鮮明。支持中國模型陣營引用成本數據：Eric Budd（Bluesky，3 upvotes）直言「美國模型程式碼確實更強，但價格太高了。」\n\n@nicrypto（X 用戶）提出更根本的威脅論：「足夠好正是每一項主導技術遭顛覆的方式」，暗示美國 AI 廠商正重蹈硬體業被低成本競爭者瓦解的覆轍。\n\nHN 用戶 reinitctxoffset 從地緣政治角度分析：「中國把 LLM 訓練和推論成本壓到零，因為他們根本不在乎財務回報。」反方認為中國模型在安全敏感場景有合規疑慮，此爭論目前未有定論。\n\n#### 實戰經驗\n\nJesse Felder（Bluesky，8 upvotes）引述最直接的流量數據：「美國企業透過 OpenRouter 使用中國 AI 的 token 份額攀升至最高 46%，過去 12 個月平均僅 11%」，為真實生產流量的量化佐證。\n\n@kentimsit(X) 實測補貼落差：「OpenAI 和 Anthropic 的 Pro/Max 訂閱補貼幅度超過 80%，改用 API 計費就能明顯感受到差距。」\n\nweitendorf(HN) 分享 Claude Code 自動化心得：「能建立自我評估工具，是使用 LLM 寫程式碼，與實現真正 10 倍生產力之間最大的差異化因素。」\n\n#### 未解問題與社群預期\n\nJ-lens 的假陽性改善進度與跨架構驗證結果，是社群認為決定此技術何時可升級至生產環境的核心問題。orbital-decay(HN) 點出深層困難：「模型在 CoT 中學到了自己的語言，我們根本無法確定它以你的方式理解你的文字。」\n\nDeepSeek 自研晶片時程與效能仍屬未知。jmyeet(HN) 預言：「OpenAI 和 Anthropic 將像恐龍一樣，不是一顆小行星，而是三顆。」社群最普遍的預期是，中國將確保 AI 前沿模型不存在護城河，這是其國家安全利益使然。",[591,593,595,596,598,600,602],{"type":96,"text":592},"用 OpenRouter API 分別呼叫 Claude Opus 4.8 與 DeepSeek V4 Flash，親測品質差距與約 23 倍成本落差，找出自身工作負載中可安全降本的應用場景。",{"type":96,"text":594},"使用 vLLM 對現有推論工作負載建立 benchmark 基準線（req/s、tokens/s、P99 延遲），為未來硬體選型或模型替換準備量化比較依據。",{"type":99,"text":177},{"type":99,"text":597},"規劃推論工作負載的硬體可攜性策略，避免深度鎖定 CUDA 專屬最佳化；若服務中國市場，提前評估華為昇騰生態系的整合成本。",{"type":102,"text":599},"追蹤 Microsoft MAI 模型是否開放 Azure API 存取，以及 OpenAI/Anthropic 是否推出直接企業定價以繞過微軟中間層。",{"type":102,"text":601},"追蹤 DeepSeek 開源推論框架是否出現非 NVIDIA 硬體最佳化，以及中美晶片管制動態與 HBM 供應鏈新進展。",{"type":102,"text":603},"持續追蹤 Anthropic Transformer Circuits Thread 後續論文——J-lens 假陽性改善進度與跨架構驗證結果，將決定此技術何時可從研究工具升級為生產可用的對齊手段。","今日報告勾勒出一個清晰的轉折訊號：AI 競爭的主戰場正從「誰的模型最強」移向「誰掌控整條供應鏈」。微軟換掉 OpenAI、DeepSeek 自研晶片、中國模型在 OpenRouter 拿下近半市佔——三件事共同指向同一方向：垂直整合、成本主導、生態鎖定。\n\n對開發者而言，今日最有價值的訊號或許是 weitendorf 的心得：LLM 真正的槓桿不在押注最強模型，而在能快速建立自我評估工具的能力。在模型廝殺激烈、選擇充裕的時代，彈性架構比單一押注更重要。",{"prev":606,"next":607},"2026-07-07","2026-07-09",{"data":609,"body":610,"excerpt":-1,"toc":620},{"title":287,"description":43},{"type":611,"children":612},"root",[613],{"type":614,"tag":615,"props":616,"children":617},"element","p",{},[618],{"type":619,"value":43},"text",{"title":287,"searchDepth":621,"depth":621,"links":622},2,[],{"data":624,"body":625,"excerpt":-1,"toc":631},{"title":287,"description":47},{"type":611,"children":626},[627],{"type":614,"tag":615,"props":628,"children":629},{},[630],{"type":619,"value":47},{"title":287,"searchDepth":621,"depth":621,"links":632},[],{"data":634,"body":635,"excerpt":-1,"toc":641},{"title":287,"description":50},{"type":611,"children":636},[637],{"type":614,"tag":615,"props":638,"children":639},{},[640],{"type":619,"value":50},{"title":287,"searchDepth":621,"depth":621,"links":642},[],{"data":644,"body":645,"excerpt":-1,"toc":651},{"title":287,"description":53},{"type":611,"children":646},[647],{"type":614,"tag":615,"props":648,"children":649},{},[650],{"type":619,"value":53},{"title":287,"searchDepth":621,"depth":621,"links":652},[],{"data":654,"body":655,"excerpt":-1,"toc":780},{"title":287,"description":287},{"type":611,"children":656},[657,664,669,674,679,685,690,695,719,733,738,744,749,754,759,765,770,775],{"type":614,"tag":658,"props":659,"children":661},"h4",{"id":660},"章節一當-ai-寫進電子紙tom-riddle-日記的現代重現",[662],{"type":619,"value":663},"章節一：當 AI 寫進電子紙——Tom Riddle 日記的現代重現",{"type":614,"tag":615,"props":665,"children":666},{},[667],{"type":619,"value":668},"開發者 Maxime Rivest 於 2026 年 7 月 7 日發布開源專案 Riddle v0.3.0，將 reMarkable Paper Pro 改造成哈利波特《消失的密室》中湯姆‧瑞鐸日記的現代版。",{"type":614,"tag":615,"props":670,"children":671},{},[672],{"type":619,"value":673},"用戶以觸控筆在電子紙上書寫後，文字自動淡出消失，AI 則以手寫體字型將回覆「自動寫」回頁面。這個設計哲學源自作者明確宣言：「移除所有傳統 AI 介面——沒有鍵盤、沒有對話泡泡、沒有發光螢幕。」",{"type":614,"tag":615,"props":675,"children":676},{},[677],{"type":619,"value":678},"電子紙的低刷新率與類紙質感反而成為這套互動體驗的天然舞台，而非技術限制。Riddle 的出現讓「AI 融入書寫材質」的概念從科幻場景變成可下載、可運行的開源工具。",{"type":614,"tag":658,"props":680,"children":682},{"id":681},"章節二技術拆解手寫辨識即時生成與電子墨水的三方協作",[683],{"type":619,"value":684},"章節二：技術拆解：手寫辨識、即時生成與電子墨水的三方協作",{"type":614,"tag":615,"props":686,"children":687},{},[688],{"type":619,"value":689},"Riddle 的技術架構分三層運作，三者串接形成低延遲的手寫對話迴路。",{"type":614,"tag":615,"props":691,"children":692},{},[693],{"type":619,"value":694},"輸入層透過 Linux evdev 介面直接擷取觸控筆原始事件，支援 4096 級壓力感應，確保辨識精度不受寫字力道影響。用戶停筆約 2.8 秒後，系統截取當前頁面為 PNG 送入核心 LLM。",{"type":614,"tag":696,"props":697,"children":698},"blockquote",{},[699],{"type":614,"tag":615,"props":700,"children":701},{},[702,708,712,717],{"type":614,"tag":703,"props":704,"children":705},"strong",{},[706],{"type":619,"value":707},"名詞解釋",{"type":614,"tag":709,"props":710,"children":711},"br",{},[],{"type":614,"tag":703,"props":713,"children":714},{},[715],{"type":619,"value":716},"evdev",{"type":619,"value":718},"：Linux 核心提供的通用輸入事件介面，允許應用程式直接讀取觸控筆等裝置的原始輸入事件，無需依賴圖形框架層。",{"type":614,"tag":615,"props":720,"children":721},{},[722,724,731],{"type":619,"value":723},"預設 LLM 為 Anthropic Fable 5(",{"type":614,"tag":725,"props":726,"children":728},"code",{"className":727},[],[729],{"type":619,"value":730},"claude-fable-5",{"type":619,"value":732},") ，亦支援 OpenAI、OpenRouter、Groq 及本地視覺模型。AI 串流輸出到達前，前端即以 Dancing Script 字型逐字動態播出，整體延遲 1 秒內啟動。",{"type":614,"tag":615,"props":734,"children":735},{},[736],{"type":619,"value":737},"「墨水消失」效果由軟體清空螢幕並重新渲染實現，視覺上呈現「墨水被頁面吸走」的效果。顯示後端提供 qtfb（xochitl 視窗模式，相容性高）與 quill（直接接管 e-ink 引擎，延遲最低）二選一。",{"type":614,"tag":658,"props":739,"children":741},{"id":740},"章節三社群實測從閱讀障礙輔助到創意筆記的真實回饋",[742],{"type":619,"value":743},"章節三：社群實測：從閱讀障礙輔助到創意筆記的真實回饋",{"type":614,"tag":615,"props":745,"children":746},{},[747],{"type":619,"value":748},"HN 討論串中，社群用戶 SoMomentary 指出，自己患有書寫障礙 (dysgraphia) ，過去在學校大量使用 reMarkable 做筆記，這套系統對他有實質幫助。",{"type":614,"tag":615,"props":750,"children":751},{},[752],{"type":619,"value":753},"這個回饋揭示了 Riddle 意料之外的應用場景：無障礙輔助。對於手寫困難的用戶，AI 即時辨識並以清晰字型呈現回應，能有效降低閱讀與理解的認知負擔。",{"type":614,"tag":615,"props":755,"children":756},{},[757],{"type":619,"value":758},"社群其他反應走向兩極。ceejayoz 強調動態效果才是核心體驗，附上示範影片提醒其他人先看動畫再評論；jshandling 則批評部分留言過度拘泥技術細節，忽略了整體體驗設計的意圖。",{"type":614,"tag":658,"props":760,"children":762},{"id":761},"章節四電子紙-ai-互動裝置的產品想像空間",[763],{"type":619,"value":764},"章節四：電子紙 × AI 互動裝置的產品想像空間",{"type":614,"tag":615,"props":766,"children":767},{},[768],{"type":619,"value":769},"Riddle 的出現為電子紙設備開闢了「環境運算介面」的新敘事：AI 不再是一個應用程式，而是融入書寫材質本身。X 用戶 @Steven_McKie 直接點出潛力——彩色電子紙加上這套技術，足以取代學童的學校筆記本與筆電。",{"type":614,"tag":615,"props":771,"children":772},{},[773],{"type":619,"value":774},"從產品維度看，這套架構目前的核心限制在於硬體鎖定：僅支援 reMarkable Paper Pro，安裝需要開發者模式與第三方啟動器，韌體更新可能讓 mod 失效。",{"type":614,"tag":615,"props":776,"children":777},{},[778],{"type":619,"value":779},"但這也暗示了一個更大的方向：若 reMarkable 原廠或其他電子紙廠商選擇將類似功能原生整合，這套 UX 模式將有機會觸及主流用戶，而不只是願意 SSH 進平板的開發者社群。",{"title":287,"searchDepth":621,"depth":621,"links":781},[],{"data":783,"body":785,"excerpt":-1,"toc":791},{"title":287,"description":784},"Riddle 的技術核心是三個相對獨立的子系統：輸入擷取、AI 推理、顯示渲染，三者串接形成低延遲的手寫對話迴路。",{"type":611,"children":786},[787],{"type":614,"tag":615,"props":788,"children":789},{},[790],{"type":619,"value":784},{"title":287,"searchDepth":621,"depth":621,"links":792},[],{"data":794,"body":796,"excerpt":-1,"toc":807},{"title":287,"description":795},"Riddle 繞過 reMarkable 官方軟體棧，直接透過 Linux evdev 介面讀取觸控筆的原始事件流，支援 4096 級壓力感應。停筆約 2.8 秒後，系統截取當前頁面為 PNG 圖像送往 LLM。",{"type":611,"children":797},[798,802],{"type":614,"tag":615,"props":799,"children":800},{},[801],{"type":619,"value":795},{"type":614,"tag":615,"props":803,"children":804},{},[805],{"type":619,"value":806},"這個設計讓系統能在不改動官方 UI 的前提下，捕捉所有手寫輸入的完整壓力與位置資訊，為後續視覺辨識提供高品質截圖。",{"title":287,"searchDepth":621,"depth":621,"links":808},[],{"data":810,"body":812,"excerpt":-1,"toc":831},{"title":287,"description":811},"截圖送入 Anthropic Fable 5(claude-fable-5) 或用戶自選的 OpenAI、OpenRouter、Groq、本地視覺模型，進行手寫內容辨識並生成回覆。",{"type":611,"children":813},[814,826],{"type":614,"tag":615,"props":815,"children":816},{},[817,819,824],{"type":619,"value":818},"截圖送入 Anthropic Fable 5(",{"type":614,"tag":725,"props":820,"children":822},{"className":821},[],[823],{"type":619,"value":730},{"type":619,"value":825},") 或用戶自選的 OpenAI、OpenRouter、Groq、本地視覺模型，進行手寫內容辨識並生成回覆。",{"type":614,"tag":615,"props":827,"children":828},{},[829],{"type":619,"value":830},"系統採串流輸出，AI 輸出開始到達前，前端即以 Dancing Script 字型逐字動態播出，整體延遲 1 秒內啟動，體驗接近即時書寫回應。",{"title":287,"searchDepth":621,"depth":621,"links":832},[],{"data":834,"body":836,"excerpt":-1,"toc":863},{"title":287,"description":835},"「墨水消失」效果由軟體實現：系統清空螢幕並重新渲染，視覺上呈現「墨水被頁面吸走」的效果。AI 回覆以 Dancing Script 字型輸出，模擬手寫筆跡風格。",{"type":611,"children":837},[838,842,847],{"type":614,"tag":615,"props":839,"children":840},{},[841],{"type":619,"value":835},{"type":614,"tag":615,"props":843,"children":844},{},[845],{"type":619,"value":846},"顯示後端提供兩種選擇：qtfb（xochitl 視窗模式，相容性高）與 quill（直接接管 e-ink 引擎，延遲最低）。持久記憶功能讓所有頁面的轉錄文字與 AI 回覆可被日後查詢回溯。",{"type":614,"tag":696,"props":848,"children":849},{},[850],{"type":614,"tag":615,"props":851,"children":852},{},[853,858,861],{"type":614,"tag":703,"props":854,"children":855},{},[856],{"type":619,"value":857},"白話比喻",{"type":614,"tag":709,"props":859,"children":860},{},[],{"type":619,"value":862},"\n這就像你對著魔法日記寫一個問題，墨水被頁面吸走，幾秒後日記自己「寫」出答案。Riddle 把每一步都用軟體實現了：evdev 是耳朵，Fable 5 是大腦，電子墨水渲染是筆。",{"title":287,"searchDepth":621,"depth":621,"links":864},[],{"data":866,"body":867,"excerpt":-1,"toc":1006},{"title":287,"description":287},{"type":611,"children":868},[869,874,879,884,890,944,949,954,959,964,983,988],{"type":614,"tag":658,"props":870,"children":872},{"id":871},"環境需求",[873],{"type":619,"value":871},{"type":614,"tag":615,"props":875,"children":876},{},[877],{"type":619,"value":878},"需要 reMarkable Paper Pro（不支援舊版 reMarkable 2），並開啟開發者模式、設定 SSH 存取。需安裝 xovi 與 AppLoad 兩個第三方啟動器元件，並準備 Anthropic API Key（或支援 OpenAI / OpenRouter / Groq 協議的替代 API Key）。",{"type":614,"tag":615,"props":880,"children":881},{},[882],{"type":619,"value":883},"平台限制明確：reMarkable 韌體更新可能讓 mod 失效，建議鎖定韌體版本或追蹤 Riddle 上游的相容性更新紀錄。",{"type":614,"tag":658,"props":885,"children":887},{"id":886},"遷移整合步驟",[888],{"type":619,"value":889},"遷移／整合步驟",{"type":614,"tag":891,"props":892,"children":893},"ol",{},[894,900,905,918,923],{"type":614,"tag":895,"props":896,"children":897},"li",{},[898],{"type":619,"value":899},"在 reMarkable Paper Pro 開啟開發者模式並設定 SSH",{"type":614,"tag":895,"props":901,"children":902},{},[903],{"type":619,"value":904},"安裝 xovi + AppLoad 第三方啟動器",{"type":614,"tag":895,"props":906,"children":907},{},[908,910,916],{"type":619,"value":909},"設定環境變數 ",{"type":614,"tag":725,"props":911,"children":913},{"className":912},[],[914],{"type":619,"value":915},"ANTHROPIC_API_KEY",{"type":619,"value":917},"（或對應 LLM 的 API Key）",{"type":614,"tag":895,"props":919,"children":920},{},[921],{"type":619,"value":922},"下載 Riddle v0.3.0 並依 README 執行安裝腳本",{"type":614,"tag":895,"props":924,"children":925},{},[926,928,934,936,942],{"type":619,"value":927},"選擇顯示後端：",{"type":614,"tag":725,"props":929,"children":931},{"className":930},[],[932],{"type":619,"value":933},"qtfb",{"type":619,"value":935},"（相容模式）或 ",{"type":614,"tag":725,"props":937,"children":939},{"className":938},[],[940],{"type":619,"value":941},"quill",{"type":619,"value":943},"（低延遲模式）",{"type":614,"tag":658,"props":945,"children":947},{"id":946},"驗測規劃",[948],{"type":619,"value":946},{"type":614,"tag":615,"props":950,"children":951},{},[952],{"type":619,"value":953},"安裝完成後，以下流程驗測核心路徑：在頁面書寫一個問題，停筆等待 2.8 秒，確認墨水消失動畫觸發、AI 回覆以 Dancing Script 字型逐字出現。",{"type":614,"tag":615,"props":955,"children":956},{},[957],{"type":619,"value":958},"延遲目標：回覆開始出現的延遲應在 1 秒內；若超過 3 秒，需檢查網路連線或 API 配額狀況。",{"type":614,"tag":658,"props":960,"children":962},{"id":961},"常見陷阱",[963],{"type":619,"value":961},{"type":614,"tag":965,"props":966,"children":967},"ul",{},[968,973,978],{"type":614,"tag":895,"props":969,"children":970},{},[971],{"type":619,"value":972},"reMarkable 韌體更新會重置開發者模式，需重新設定 SSH 並重裝啟動器元件",{"type":614,"tag":895,"props":974,"children":975},{},[976],{"type":619,"value":977},"本地視覺模型（替代 Fable 5）在電子紙設備上的延遲顯著較高，體驗落差大",{"type":614,"tag":895,"props":979,"children":980},{},[981],{"type":619,"value":982},"quill 模式直接接管 e-ink 引擎，可能與官方應用衝突，建議先用 qtfb 測試穩定性",{"type":614,"tag":658,"props":984,"children":986},{"id":985},"上線檢核清單",[987],{"type":619,"value":985},{"type":614,"tag":965,"props":989,"children":990},{},[991,996,1001],{"type":614,"tag":895,"props":992,"children":993},{},[994],{"type":619,"value":995},"觀測：idle timeout 觸發率、LLM 回覆延遲（目標 1 秒內）、渲染幀數穩定性",{"type":614,"tag":895,"props":997,"children":998},{},[999],{"type":619,"value":1000},"成本：Fable 5 API 費用（每次手寫截圖約 1-2k token 視覺輸入，需評估使用頻率）",{"type":614,"tag":895,"props":1002,"children":1003},{},[1004],{"type":619,"value":1005},"風險：韌體更新相容性、第三方啟動器長期維護穩定性",{"title":287,"searchDepth":621,"depth":621,"links":1007},[],{"data":1009,"body":1010,"excerpt":-1,"toc":1130},{"title":287,"description":287},{"type":611,"children":1011},[1012,1017,1040,1045,1068,1073,1078,1083,1101,1106,1119,1125],{"type":614,"tag":658,"props":1013,"children":1015},{"id":1014},"競爭版圖",[1016],{"type":619,"value":1014},{"type":614,"tag":965,"props":1018,"children":1019},{},[1020,1030],{"type":614,"tag":895,"props":1021,"children":1022},{},[1023,1028],{"type":614,"tag":703,"props":1024,"children":1025},{},[1026],{"type":619,"value":1027},"直接競品",{"type":619,"value":1029},"：目前無商業產品或開源專案在電子紙裝置上實現相同的「手寫消失 + AI 回應」互動模式，Riddle 目前是唯一完整實作",{"type":614,"tag":895,"props":1031,"children":1032},{},[1033,1038],{"type":614,"tag":703,"props":1034,"children":1035},{},[1036],{"type":619,"value":1037},"間接競品",{"type":619,"value":1039},"：AI 筆記應用（Notion AI、GoodNotes AI）、智慧手寫辨識工具 (Nebo) 、reMarkable 官方 AI 功能（若有）",{"type":614,"tag":658,"props":1041,"children":1043},{"id":1042},"護城河類型",[1044],{"type":619,"value":1042},{"type":614,"tag":965,"props":1046,"children":1047},{},[1048,1058],{"type":614,"tag":895,"props":1049,"children":1050},{},[1051,1056],{"type":614,"tag":703,"props":1052,"children":1053},{},[1054],{"type":619,"value":1055},"工程護城河",{"type":619,"value":1057},"：直接操作 evdev 與 e-ink 引擎的底層整合，需對 reMarkable 韌體有深度理解，複製門檻不低",{"type":614,"tag":895,"props":1059,"children":1060},{},[1061,1066],{"type":614,"tag":703,"props":1062,"children":1063},{},[1064],{"type":619,"value":1065},"生態護城河",{"type":619,"value":1067},"：v0.3.0 上線後 1,000+ 星與 72 個 Fork 形成早期社群，使用者貢獻的韌體相容性修補將持續累積技術壁壘",{"type":614,"tag":658,"props":1069,"children":1071},{"id":1070},"定價策略",[1072],{"type":619,"value":1070},{"type":614,"tag":615,"props":1074,"children":1075},{},[1076],{"type":619,"value":1077},"Riddle 本身開源免費，但用戶需自行承擔 LLM API 費用。Fable 5 的視覺輸入每次約耗用 1-2k token，頻繁使用的月度費用可能達數美元至數十美元不等，需根據使用頻率評估。",{"type":614,"tag":658,"props":1079,"children":1081},{"id":1080},"企業導入阻力",[1082],{"type":619,"value":1080},{"type":614,"tag":965,"props":1084,"children":1085},{},[1086,1091,1096],{"type":614,"tag":895,"props":1087,"children":1088},{},[1089],{"type":619,"value":1090},"reMarkable Paper Pro 硬體鎖定，無法跨平台移植或批量部署",{"type":614,"tag":895,"props":1092,"children":1093},{},[1094],{"type":619,"value":1095},"需要開發者模式，一般消費者安裝門檻過高",{"type":614,"tag":895,"props":1097,"children":1098},{},[1099],{"type":619,"value":1100},"韌體更新風險讓企業 IT 難以維護穩定版本，集中管理機制缺失",{"type":614,"tag":658,"props":1102,"children":1104},{"id":1103},"第二序影響",[1105],{"type":619,"value":1103},{"type":614,"tag":965,"props":1107,"children":1108},{},[1109,1114],{"type":614,"tag":895,"props":1110,"children":1111},{},[1112],{"type":619,"value":1113},"若 reMarkable 原廠或其他電子紙廠商（Boox、Kindle Scribe）將類似功能原生整合，此 UX 模式可觸及主流消費市場",{"type":614,"tag":895,"props":1115,"children":1116},{},[1117],{"type":619,"value":1118},"無障礙輔助潛力（書寫障礙輔助、視覺排版支援）可能吸引 EdTech 廠商開發商業版本",{"type":614,"tag":658,"props":1120,"children":1122},{"id":1121},"判決生態先驅社群驗證商業路徑待廠商跟進",[1123],{"type":619,"value":1124},"判決生態先驅（社群驗證，商業路徑待廠商跟進）",{"type":614,"tag":615,"props":1126,"children":1127},{},[1128],{"type":619,"value":1129},"Riddle 成功證明了電子紙 × AI 手寫對話的 UX 可行性，1,000+ 星的社群驗證顯示需求真實存在。但商業化路徑完全依賴 reMarkable 原廠決策或後繼者整合意願，短期內仍是開發者 PoC 而非消費級產品。",{"title":287,"searchDepth":621,"depth":621,"links":1131},[],{"data":1133,"body":1134,"excerpt":-1,"toc":1151},{"title":287,"description":287},{"type":611,"children":1135},[1136,1141,1146],{"type":614,"tag":658,"props":1137,"children":1139},{"id":1138},"延遲與效能",[1140],{"type":619,"value":1138},{"type":614,"tag":615,"props":1142,"children":1143},{},[1144],{"type":619,"value":1145},"根據作者描述，回覆生成延遲在 1 秒內啟動，整體體驗接近即時書寫回應。結合 2.8 秒 idle timeout，從停筆到看見 AI 回覆開始出現的等待時間約 3-4 秒。",{"type":614,"tag":615,"props":1147,"children":1148},{},[1149],{"type":619,"value":1150},"暫無 OCR 辨識準確率或多語言辨識的正式基準測試數據。社群回饋顯示英文手寫辨識表現正常，繁體中文手寫的辨識精度尚無公開數據，建議有需求者自行測試。",{"title":287,"searchDepth":621,"depth":621,"links":1152},[],{"data":1154,"body":1155,"excerpt":-1,"toc":1176},{"title":287,"description":287},{"type":611,"children":1156},[1157],{"type":614,"tag":965,"props":1158,"children":1159},{},[1160,1164,1168,1172],{"type":614,"tag":895,"props":1161,"children":1162},{},[1163],{"type":619,"value":59},{"type":614,"tag":895,"props":1165,"children":1166},{},[1167],{"type":619,"value":60},{"type":614,"tag":895,"props":1169,"children":1170},{},[1171],{"type":619,"value":61},{"type":614,"tag":895,"props":1173,"children":1174},{},[1175],{"type":619,"value":62},{"title":287,"searchDepth":621,"depth":621,"links":1177},[],{"data":1179,"body":1180,"excerpt":-1,"toc":1197},{"title":287,"description":287},{"type":611,"children":1181},[1182],{"type":614,"tag":965,"props":1183,"children":1184},{},[1185,1189,1193],{"type":614,"tag":895,"props":1186,"children":1187},{},[1188],{"type":619,"value":64},{"type":614,"tag":895,"props":1190,"children":1191},{},[1192],{"type":619,"value":65},{"type":614,"tag":895,"props":1194,"children":1195},{},[1196],{"type":619,"value":66},{"title":287,"searchDepth":621,"depth":621,"links":1198},[],{"data":1200,"body":1201,"excerpt":-1,"toc":1207},{"title":287,"description":70},{"type":611,"children":1202},[1203],{"type":614,"tag":615,"props":1204,"children":1205},{},[1206],{"type":619,"value":70},{"title":287,"searchDepth":621,"depth":621,"links":1208},[],{"data":1210,"body":1211,"excerpt":-1,"toc":1217},{"title":287,"description":71},{"type":611,"children":1212},[1213],{"type":614,"tag":615,"props":1214,"children":1215},{},[1216],{"type":619,"value":71},{"title":287,"searchDepth":621,"depth":621,"links":1218},[],{"data":1220,"body":1221,"excerpt":-1,"toc":1227},{"title":287,"description":72},{"type":611,"children":1222},[1223],{"type":614,"tag":615,"props":1224,"children":1225},{},[1226],{"type":619,"value":72},{"title":287,"searchDepth":621,"depth":621,"links":1228},[],{"data":1230,"body":1231,"excerpt":-1,"toc":1237},{"title":287,"description":128},{"type":611,"children":1232},[1233],{"type":614,"tag":615,"props":1234,"children":1235},{},[1236],{"type":619,"value":128},{"title":287,"searchDepth":621,"depth":621,"links":1238},[],{"data":1240,"body":1241,"excerpt":-1,"toc":1247},{"title":287,"description":131},{"type":611,"children":1242},[1243],{"type":614,"tag":615,"props":1244,"children":1245},{},[1246],{"type":619,"value":131},{"title":287,"searchDepth":621,"depth":621,"links":1248},[],{"data":1250,"body":1251,"excerpt":-1,"toc":1257},{"title":287,"description":134},{"type":611,"children":1252},[1253],{"type":614,"tag":615,"props":1254,"children":1255},{},[1256],{"type":619,"value":134},{"title":287,"searchDepth":621,"depth":621,"links":1258},[],{"data":1260,"body":1261,"excerpt":-1,"toc":1267},{"title":287,"description":137},{"type":611,"children":1262},[1263],{"type":614,"tag":615,"props":1264,"children":1265},{},[1266],{"type":619,"value":137},{"title":287,"searchDepth":621,"depth":621,"links":1268},[],{"data":1270,"body":1272,"excerpt":-1,"toc":1392},{"title":287,"description":1271},"2026 年 7 月，彭博社揭露微軟已悄悄在 Excel 與 Outlook 中以自家 MAI 模型取代 OpenAI 與 Anthropic 的服務，目前每週已處理數萬筆請求。",{"type":611,"children":1273},[1274,1278,1283,1289,1294,1299,1304,1319,1325,1330,1335,1340,1355,1361,1366,1371,1377,1382,1387],{"type":614,"tag":615,"props":1275,"children":1276},{},[1277],{"type":619,"value":1271},{"type":614,"tag":615,"props":1279,"children":1280},{},[1281],{"type":619,"value":1282},"這則消息不僅標誌 AI 供應鏈的結構性轉變，也引爆關於品質妥協、開源模型經濟學與巨頭垂直整合的產業辯論。",{"type":614,"tag":658,"props":1284,"children":1286},{"id":1285},"章節一微軟-mai-模型換血計畫copilot-成本劇降的代價",[1287],{"type":619,"value":1288},"章節一：微軟 MAI 模型換血計畫——Copilot 成本劇降的代價",{"type":614,"tag":615,"props":1290,"children":1291},{},[1292],{"type":619,"value":1293},"微軟 AI 負責人 Mustafa Suleyman 毫不諱言：「我們支付了大量資金給 Anthropic——我們的目標是減少並最終消除這筆成本。」這句話直接道出換血計畫的核心動機，也讓 Anthropic 的企業客戶意識到，這場替換並非技術決策，而是純粹的商業算計。",{"type":614,"tag":615,"props":1295,"children":1296},{},[1297],{"type":619,"value":1298},"微軟已在 Build 2026 大會發布七款新 MAI 模型，包含首款推理模型 MAI-Thinking 1、agentic 程式碼生成器與文字轉圖像模型。但基準測試顯示 MAI-Thinking 1 僅約等同 DeepSeek V3.2，明顯落後 OpenAI 及 Anthropic 的同類前沿推理模型。",{"type":614,"tag":615,"props":1300,"children":1301},{},[1302],{"type":619,"value":1303},"對一般用戶而言，這意味著在相同的 Copilot 訂閱費用下，可能在不知情的情況下獲得品質較低的 AI 回覆。Satya Nadella 暗示未來或許轉向分層定價：MAI 作為免費預設，第三方頂尖模型作為付費附加選項，最終成本壓力可能由用戶而非微軟承擔。",{"type":614,"tag":696,"props":1305,"children":1306},{},[1307],{"type":614,"tag":615,"props":1308,"children":1309},{},[1310,1314,1317],{"type":614,"tag":703,"props":1311,"children":1312},{},[1313],{"type":619,"value":707},{"type":614,"tag":709,"props":1315,"children":1316},{},[],{"type":619,"value":1318},"\nMAI(Microsoft AI) ：微軟自研 AI 模型系列，與旗下 OpenAI 投資和 Anthropic 合作並行存在，目的是降低對外部 AI 服務的採購依賴。",{"type":614,"tag":658,"props":1320,"children":1322},{"id":1321},"章節二開源模型崛起卻沒打倒前沿實驗室的弔詭",[1323],{"type":619,"value":1324},"章節二：開源模型崛起卻沒打倒前沿實驗室的弔詭",{"type":614,"tag":615,"props":1326,"children":1327},{},[1328],{"type":619,"value":1329},"Vercel AI Gateway 的最新數據揭示了一個表面矛盾的現象：DeepSeek 在 token 流量中已超越 Anthropic，處理略超三分之一的 token 量，但 Anthropic 仍佔整體 AI 消費金額的逾一半。流量輸了，收入卻贏了。",{"type":614,"tag":615,"props":1331,"children":1332},{},[1333],{"type":619,"value":1334},"Decagon CEO Jesse Zhang 提出「兩階段生命週期理論」解釋這一弔詭：當一個 AI 應用場景趨於成熟、需求穩定可預測時，企業自然向更便宜的開源或自建模型遷移；但同時，新興的、更複雜的 AI 用例持續涌現，這些場景仍依賴前沿模型的能力邊界，無法被低成本模型取代。",{"type":614,"tag":615,"props":1336,"children":1337},{},[1338],{"type":619,"value":1339},"OpenRouter 數據佐證了這個框架：DeepSeek V4 Flash 每週處理 5.3 兆 tokens（排名第一），Claude Opus 4.8 處理逾 2 兆 tokens，但兩者每百萬 token 的價差高達 23 倍（$0.06 對 $1.37）。便宜模型吃量，前沿模型賺錢，並非零和競爭。",{"type":614,"tag":696,"props":1341,"children":1342},{},[1343],{"type":614,"tag":615,"props":1344,"children":1345},{},[1346,1350,1353],{"type":614,"tag":703,"props":1347,"children":1348},{},[1349],{"type":619,"value":857},{"type":614,"tag":709,"props":1351,"children":1352},{},[],{"type":619,"value":1354},"\n就像機票市場：廉航吃掉了部分商務艙客源後，商務艙並沒有消失，反而集中在願意付高價的旅客身上。AI 模型市場正在發生相同的分層現象。",{"type":614,"tag":658,"props":1356,"children":1358},{"id":1357},"章節三從依賴到自建科技巨頭的-ai-供應鏈重組",[1359],{"type":619,"value":1360},"章節三：從依賴到自建——科技巨頭的 AI 供應鏈重組",{"type":614,"tag":615,"props":1362,"children":1363},{},[1364],{"type":619,"value":1365},"微軟並非孤例。TechCrunch 點名 Amazon、Uber、Meta、Accenture 均已加入「AI 成本節流」浪潮，早期的 tokenmaxxing（盡量多用 token 換最佳效果）策略正被嚴格重新評估。這個趨勢與早年各大企業雲端基礎設施自建歷程驚人相似——先依賴 AWS，隨後陸續建置私有雲或混合架構以降低鎖定風險。",{"type":614,"tag":615,"props":1367,"children":1368},{},[1369],{"type":619,"value":1370},"微軟的長期算盤尤為清晰：與 OpenAI 的合作協議將於 2032 年到期，自建 MAI 模型是長期降低依賴的戰略倒數計時。值得注意的是，微軟宣稱 MAI 訓練資料「乾淨、具商業授權」，但實際使用了法律地位尚未確立的 Common Crawl 公開資料集，潛在智慧財產權風險懸而未決。",{"type":614,"tag":658,"props":1372,"children":1374},{"id":1373},"章節四開發者與企業用戶該如何因應模型替換潮",[1375],{"type":619,"value":1376},"章節四：開發者與企業用戶該如何因應模型替換潮",{"type":614,"tag":615,"props":1378,"children":1379},{},[1380],{"type":619,"value":1381},"對企業採購決策者而言，微軟的分層定價構想預示了一個新的評估框架：預設 MAI 滿足基礎需求，付費升級 GPT 或 Claude 應對高品質場景。採購前應先盤點哪些工作流程對 AI 品質高度敏感，哪些可接受較低品質換取成本節省，再據此規劃分層採購策略。",{"type":614,"tag":615,"props":1383,"children":1384},{},[1385],{"type":619,"value":1386},"對開發者而言，OpenRouter 23 倍的價差數據（Claude Opus 4.8 $1.37 vs. DeepSeek V4 Flash $0.06 每百萬 token）是成本估算的關鍵錨點。更重要的是，模型替換潮凸顯了應用層模型抽象化的必要性——避免將特定廠商模型硬編碼進業務邏輯，保留快速切換的彈性。",{"type":614,"tag":615,"props":1388,"children":1389},{},[1390],{"type":619,"value":1391},"Nvidia Nemotron 模型也被分析師點名為下一個可能衝上排名前段的開源競爭者，值得列入技術雷達追蹤清單。",{"title":287,"searchDepth":621,"depth":621,"links":1393},[],{"data":1395,"body":1397,"excerpt":-1,"toc":1403},{"title":287,"description":1396},"MAI 模型的替換並非一次性大規模遷移，而是漸進式的靜默替換策略，背後涉及三個相互關聯的機制。",{"type":611,"children":1398},[1399],{"type":614,"tag":615,"props":1400,"children":1401},{},[1402],{"type":619,"value":1396},{"title":287,"searchDepth":621,"depth":621,"links":1404},[],{"data":1406,"body":1408,"excerpt":-1,"toc":1414},{"title":287,"description":1407},"微軟已在 Excel 和 Outlook 中對「某比例的用戶提示」改用 MAI 模型，但未披露確切百分比。此設計讓用戶在不知情的情況下成為品質測試對象，若品質落差不引發顯著投訴，則持續擴大替換比例，直至完全取代或形成明確分層。",{"type":611,"children":1409},[1410],{"type":614,"tag":615,"props":1411,"children":1412},{},[1413],{"type":619,"value":1407},{"title":287,"searchDepth":621,"depth":621,"links":1415},[],{"data":1417,"body":1419,"excerpt":-1,"toc":1440},{"title":287,"description":1418},"Nadella 暗示的分層定價構想背後是模型路由架構：系統根據請求複雜度和用戶訂閱等級，自動決定派發 MAI 或第三方前沿模型。簡單的摘要、格式化、翻譯任務走 MAI 降本；複雜的推理、程式碼生成、長文分析走付費版 GPT 或 Claude。",{"type":611,"children":1420},[1421,1425],{"type":614,"tag":615,"props":1422,"children":1423},{},[1424],{"type":619,"value":1418},{"type":614,"tag":696,"props":1426,"children":1427},{},[1428],{"type":614,"tag":615,"props":1429,"children":1430},{},[1431,1435,1438],{"type":614,"tag":703,"props":1432,"children":1433},{},[1434],{"type":619,"value":707},{"type":614,"tag":709,"props":1436,"children":1437},{},[],{"type":619,"value":1439},"\n模型路由 (Model Router) ：根據請求特徵（複雜度、類型、成本預算）自動選擇最合適模型的中間件層，常見實作如 LiteLLM、Portkey、OpenRouter 等工具。",{"title":287,"searchDepth":621,"depth":621,"links":1441},[],{"data":1443,"body":1445,"excerpt":-1,"toc":1466},{"title":287,"description":1444},"MAI 模型宣稱使用「乾淨、商業授權」訓練資料，但實際涵蓋 Common Crawl 公開資料集，後者的授權狀態在多個司法管轄區仍有爭議。此授權風險目前由微軟承擔，但若引發訴訟或政策變動，下游企業客戶也可能受波及，需列入合規評估清單。",{"type":611,"children":1446},[1447,1451],{"type":614,"tag":615,"props":1448,"children":1449},{},[1450],{"type":619,"value":1444},{"type":614,"tag":696,"props":1452,"children":1453},{},[1454],{"type":614,"tag":615,"props":1455,"children":1456},{},[1457,1461,1464],{"type":614,"tag":703,"props":1458,"children":1459},{},[1460],{"type":619,"value":857},{"type":614,"tag":709,"props":1462,"children":1463},{},[],{"type":619,"value":1465},"\nMAI 換血就像餐廳悄悄把食材從 A 級牛肉換成 B 級卻不更改菜單——你吃到的味道可能略有差異，但若不主動比較很難察覺，而餐廳的食材成本已悄悄降低。",{"title":287,"searchDepth":621,"depth":621,"links":1467},[],{"data":1469,"body":1470,"excerpt":-1,"toc":1561},{"title":287,"description":287},{"type":611,"children":1471},[1472,1476,1481,1485,1508,1512,1517,1521,1539,1543],{"type":614,"tag":658,"props":1473,"children":1474},{"id":871},[1475],{"type":619,"value":871},{"type":614,"tag":615,"props":1477,"children":1478},{},[1479],{"type":619,"value":1480},"目前 MAI 模型未提供獨立公開 API 端點，僅作為 Microsoft 365 Copilot 的後端服務，開發者無法直接呼叫。若要評估替換效果，需透過 Microsoft Azure AI Foundry 或等待 MAI 模型正式上架 Azure OpenAI Service。如需評估低成本替代方案，DeepSeek V4 Flash（OpenRouter 可直接呼叫）是目前最具代表性的對照組。",{"type":614,"tag":658,"props":1482,"children":1483},{"id":886},[1484],{"type":619,"value":889},{"type":614,"tag":891,"props":1486,"children":1487},{},[1488,1493,1498,1503],{"type":614,"tag":895,"props":1489,"children":1490},{},[1491],{"type":619,"value":1492},"盤點現有 AI 呼叫：識別哪些場景使用 OpenAI/Anthropic API，按請求量與品質敏感度分級（高敏感 / 可降本）",{"type":614,"tag":895,"props":1494,"children":1495},{},[1496],{"type":619,"value":1497},"建立模型抽象層：統一 LLM 呼叫介面（如 LiteLLM 或自建 adapter），確保模型切換不需修改業務邏輯",{"type":614,"tag":895,"props":1499,"children":1500},{},[1501],{"type":619,"value":1502},"設計 A/B 測試框架：對每個場景設置品質評估指標（任務完成率、格式遵守率、人工抽樣評分），為後續模型替換提供客觀依據",{"type":614,"tag":895,"props":1504,"children":1505},{},[1506],{"type":619,"value":1507},"成本試算：以 OpenRouter 23 倍價差為基準，計算不同替換比例下的月度成本節省潛力",{"type":614,"tag":658,"props":1509,"children":1510},{"id":946},[1511],{"type":619,"value":946},{"type":614,"tag":615,"props":1513,"children":1514},{},[1515],{"type":619,"value":1516},"對每個準備替換的場景，建立基線測試集，對比 GPT-4o/Claude Opus 4.8 與 DeepSeek V4 Flash 的輸出品質。重點量測三個維度：回覆一致性（相同問題多次回覆的穩定性）、任務完成率（尤其是結構化輸出格式遵守率）、延遲（低成本模型推理速度是否達到生產需求）。",{"type":614,"tag":658,"props":1518,"children":1519},{"id":961},[1520],{"type":619,"value":961},{"type":614,"tag":965,"props":1522,"children":1523},{},[1524,1529,1534],{"type":614,"tag":895,"props":1525,"children":1526},{},[1527],{"type":619,"value":1528},"只看 token 成本、忽略品質成本：低品質回覆導致的人工修正、重試、用戶投訴成本，可能高過節省的 token 費用",{"type":614,"tag":895,"props":1530,"children":1531},{},[1532],{"type":619,"value":1533},"訓練資料法律風險被低估：MAI 使用 Common Crawl 的授權爭議尚未解決，企業法務需持續追蹤後續進展",{"type":614,"tag":895,"props":1535,"children":1536},{},[1537],{"type":619,"value":1538},"過早鎖定路由規則：AI 能力邊界持續移動，今天需前沿模型的任務，三個月後可能 DeepSeek 就已足夠",{"type":614,"tag":658,"props":1540,"children":1541},{"id":985},[1542],{"type":619,"value":985},{"type":614,"tag":965,"props":1544,"children":1545},{},[1546,1551,1556],{"type":614,"tag":895,"props":1547,"children":1548},{},[1549],{"type":619,"value":1550},"觀測：請求延遲 P50/P99、品質評分（自動化 + 人工抽樣 5%）、模型替換比例週報",{"type":614,"tag":895,"props":1552,"children":1553},{},[1554],{"type":619,"value":1555},"成本：每百萬 token 費用趨勢、月度 API 支出 vs. 品質分數的 tradeoff 報表",{"type":614,"tag":895,"props":1557,"children":1558},{},[1559],{"type":619,"value":1560},"風險：Common Crawl 授權狀態追蹤、微軟 MAI API 可用性 SLA、用戶負面回饋率監控",{"title":287,"searchDepth":621,"depth":621,"links":1562},[],{"data":1564,"body":1565,"excerpt":-1,"toc":1681},{"title":287,"description":287},{"type":611,"children":1566},[1567,1571,1592,1596,1617,1621,1626,1630,1648,1652,1670,1676],{"type":614,"tag":658,"props":1568,"children":1569},{"id":1014},[1570],{"type":619,"value":1014},{"type":614,"tag":965,"props":1572,"children":1573},{},[1574,1583],{"type":614,"tag":895,"props":1575,"children":1576},{},[1577,1581],{"type":614,"tag":703,"props":1578,"children":1579},{},[1580],{"type":619,"value":1027},{"type":619,"value":1582},"：Azure OpenAI Service（微軟旗下，兩者將走向分層共存）、AWS Bedrock（Amazon 同樣走向自建 Titan 模型路線）、Google Vertex AI（Gemini 垂直整合）",{"type":614,"tag":895,"props":1584,"children":1585},{},[1586,1590],{"type":614,"tag":703,"props":1587,"children":1588},{},[1589],{"type":619,"value":1037},{"type":619,"value":1591},"：Hugging Face + 開源模型自建路線、Mistral API、Groq 低延遲推理服務、Nvidia Nemotron（分析師點名下一個開源競爭者）",{"type":614,"tag":658,"props":1593,"children":1594},{"id":1042},[1595],{"type":619,"value":1042},{"type":614,"tag":965,"props":1597,"children":1598},{},[1599,1608],{"type":614,"tag":895,"props":1600,"children":1601},{},[1602,1606],{"type":614,"tag":703,"props":1603,"children":1604},{},[1605],{"type":619,"value":1055},{"type":619,"value":1607},"：Office 365 生態系深度整合（Excel 公式 AI、Outlook 草稿、Teams 語音轉錄），用戶切換成本極高",{"type":614,"tag":895,"props":1609,"children":1610},{},[1611,1615],{"type":614,"tag":703,"props":1612,"children":1613},{},[1614],{"type":619,"value":1065},{"type":619,"value":1616},"：企業組織機構授權、SSO 整合、合規認證（ISO 27001、SOC 2）已與 Microsoft 365 深度綁定，AI 功能隨平台整包採購",{"type":614,"tag":658,"props":1618,"children":1619},{"id":1070},[1620],{"type":619,"value":1070},{"type":614,"tag":615,"props":1622,"children":1623},{},[1624],{"type":619,"value":1625},"Nadella 的分層定價構想（MAI 免費預設 + 付費附加前沿模型）若落地，將使 Copilot 訂閱結構類似 SaaS「基礎版 + 進階版」模型。若基礎版 AI 功能「夠用」，企業就少了採購獨立 AI 工具的理由；若基礎版品質明顯不足，反而為 OpenAI/Anthropic 的直接企業銷售創造空間。",{"type":614,"tag":658,"props":1627,"children":1628},{"id":1080},[1629],{"type":619,"value":1080},{"type":614,"tag":965,"props":1631,"children":1632},{},[1633,1638,1643],{"type":614,"tag":895,"props":1634,"children":1635},{},[1636],{"type":619,"value":1637},"品質降級的感知風險：IT 決策者擔心用戶察覺 Copilot 品質下降，引發投訴或信任危機",{"type":614,"tag":895,"props":1639,"children":1640},{},[1641],{"type":619,"value":1642},"透明度不足：微軟未公開 MAI 基準數據，企業難以客觀評估替換前後的品質差距",{"type":614,"tag":895,"props":1644,"children":1645},{},[1646],{"type":619,"value":1647},"長期廠商鎖定疑慮：若 MAI 成為預設，企業對 Microsoft 生態的整體依賴反而更深，議價能力下降",{"type":614,"tag":658,"props":1649,"children":1650},{"id":1103},[1651],{"type":619,"value":1103},{"type":614,"tag":965,"props":1653,"children":1654},{},[1655,1660,1665],{"type":614,"tag":895,"props":1656,"children":1657},{},[1658],{"type":619,"value":1659},"OpenAI 與 Anthropic 的 B2B 收入壓力將上升，可能加速其直接面向企業的定價競爭（繞過微軟中間層）",{"type":614,"tag":895,"props":1661,"children":1662},{},[1663],{"type":619,"value":1664},"開源模型廠商（Meta Llama、Mistral、DeepSeek）在成熟場景的市占將持續擴大，驅動前沿模型廠商專注高價值差異化場景",{"type":614,"tag":895,"props":1666,"children":1667},{},[1668],{"type":619,"value":1669},"Microsoft 365 用戶若察覺品質落差，可能觸發對 ChatGPT Enterprise 或 Claude for Work 的直接採購評估",{"type":614,"tag":658,"props":1671,"children":1673},{"id":1672},"判決垂直整合加速微軟此舉觸發-ai-供應鏈重組連鎖效應",[1674],{"type":619,"value":1675},"判決：垂直整合加速（微軟此舉觸發 AI 供應鏈重組連鎖效應）",{"type":614,"tag":615,"props":1677,"children":1678},{},[1679],{"type":619,"value":1680},"微軟的 MAI 換血計畫本質上是巨頭垂直整合的最新章節，與早年各大企業自建雲端基礎設施的歷程如出一轍。短期內用戶品質感知是關鍵變數；長期而言，2032 年 OpenAI 協議到期前，這場靜默的供應鏈重組將持續加速，並帶動整個產業跟進。",{"title":287,"searchDepth":621,"depth":621,"links":1682},[],{"data":1684,"body":1685,"excerpt":-1,"toc":1797},{"title":287,"description":287},{"type":611,"children":1686},[1687,1693,1698,1704,1717,1723,1792],{"type":614,"tag":658,"props":1688,"children":1690},{"id":1689},"mai-thinking-1-基準表現",[1691],{"type":619,"value":1692},"MAI-Thinking 1 基準表現",{"type":614,"tag":615,"props":1694,"children":1695},{},[1696],{"type":619,"value":1697},"MAI-Thinking 1 在標準評測基準上的表現約等同 DeepSeek V3.2，但明顯落後 OpenAI o3 及 Anthropic Claude Opus 4.8 等前沿推理模型。微軟尚未公開具體評測數字，現有比較來自第三方分析師。",{"type":614,"tag":658,"props":1699,"children":1701},{"id":1700},"市場流量與消費金額對比vercel-ai-gateway2026-07",[1702],{"type":619,"value":1703},"市場流量與消費金額對比（Vercel AI Gateway，2026-07）",{"type":614,"tag":965,"props":1705,"children":1706},{},[1707,1712],{"type":614,"tag":895,"props":1708,"children":1709},{},[1710],{"type":619,"value":1711},"DeepSeek：token 流量略超三分之一，已超越 Anthropic",{"type":614,"tag":895,"props":1713,"children":1714},{},[1715],{"type":619,"value":1716},"Anthropic：佔整體 AI 消費金額逾一半，高流量不等於高收入",{"type":614,"tag":658,"props":1718,"children":1720},{"id":1719},"openrouter-每週-token-處理量-2026-07",[1721],{"type":619,"value":1722},"OpenRouter 每週 token 處理量 (2026-07)",{"type":614,"tag":1724,"props":1725,"children":1726},"table",{},[1727,1751],{"type":614,"tag":1728,"props":1729,"children":1730},"thead",{},[1731],{"type":614,"tag":1732,"props":1733,"children":1734},"tr",{},[1735,1741,1746],{"type":614,"tag":1736,"props":1737,"children":1738},"th",{},[1739],{"type":619,"value":1740},"模型",{"type":614,"tag":1736,"props":1742,"children":1743},{},[1744],{"type":619,"value":1745},"每週 tokens",{"type":614,"tag":1736,"props":1747,"children":1748},{},[1749],{"type":619,"value":1750},"每百萬 token 費用",{"type":614,"tag":1752,"props":1753,"children":1754},"tbody",{},[1755,1774],{"type":614,"tag":1732,"props":1756,"children":1757},{},[1758,1764,1769],{"type":614,"tag":1759,"props":1760,"children":1761},"td",{},[1762],{"type":619,"value":1763},"DeepSeek V4 Flash",{"type":614,"tag":1759,"props":1765,"children":1766},{},[1767],{"type":619,"value":1768},"5.3 兆（第一名）",{"type":614,"tag":1759,"props":1770,"children":1771},{},[1772],{"type":619,"value":1773},"$0.06",{"type":614,"tag":1732,"props":1775,"children":1776},{},[1777,1782,1787],{"type":614,"tag":1759,"props":1778,"children":1779},{},[1780],{"type":619,"value":1781},"Claude Opus 4.8",{"type":614,"tag":1759,"props":1783,"children":1784},{},[1785],{"type":619,"value":1786},"逾 2 兆",{"type":614,"tag":1759,"props":1788,"children":1789},{},[1790],{"type":619,"value":1791},"$1.37",{"type":614,"tag":615,"props":1793,"children":1794},{},[1795],{"type":619,"value":1796},"兩者價差約 23 倍，說明不同場景的模型選型邏輯截然不同：成熟場景拚成本，新興場景拚能力。",{"title":287,"searchDepth":621,"depth":621,"links":1798},[],{"data":1800,"body":1801,"excerpt":-1,"toc":1818},{"title":287,"description":287},{"type":611,"children":1802},[1803],{"type":614,"tag":965,"props":1804,"children":1805},{},[1806,1810,1814],{"type":614,"tag":895,"props":1807,"children":1808},{},[1809],{"type":619,"value":143},{"type":614,"tag":895,"props":1811,"children":1812},{},[1813],{"type":619,"value":144},{"type":614,"tag":895,"props":1815,"children":1816},{},[1817],{"type":619,"value":145},{"title":287,"searchDepth":621,"depth":621,"links":1819},[],{"data":1821,"body":1822,"excerpt":-1,"toc":1835},{"title":287,"description":287},{"type":611,"children":1823},[1824],{"type":614,"tag":965,"props":1825,"children":1826},{},[1827,1831],{"type":614,"tag":895,"props":1828,"children":1829},{},[1830],{"type":619,"value":147},{"type":614,"tag":895,"props":1832,"children":1833},{},[1834],{"type":619,"value":148},{"title":287,"searchDepth":621,"depth":621,"links":1836},[],{"data":1838,"body":1839,"excerpt":-1,"toc":1845},{"title":287,"description":152},{"type":611,"children":1840},[1841],{"type":614,"tag":615,"props":1842,"children":1843},{},[1844],{"type":619,"value":152},{"title":287,"searchDepth":621,"depth":621,"links":1846},[],{"data":1848,"body":1849,"excerpt":-1,"toc":1855},{"title":287,"description":153},{"type":611,"children":1850},[1851],{"type":614,"tag":615,"props":1852,"children":1853},{},[1854],{"type":619,"value":153},{"title":287,"searchDepth":621,"depth":621,"links":1856},[],{"data":1858,"body":1859,"excerpt":-1,"toc":1865},{"title":287,"description":154},{"type":611,"children":1860},[1861],{"type":614,"tag":615,"props":1862,"children":1863},{},[1864],{"type":619,"value":154},{"title":287,"searchDepth":621,"depth":621,"links":1866},[],{"data":1868,"body":1869,"excerpt":-1,"toc":1875},{"title":287,"description":205},{"type":611,"children":1870},[1871],{"type":614,"tag":615,"props":1872,"children":1873},{},[1874],{"type":619,"value":205},{"title":287,"searchDepth":621,"depth":621,"links":1876},[],{"data":1878,"body":1879,"excerpt":-1,"toc":1885},{"title":287,"description":208},{"type":611,"children":1880},[1881],{"type":614,"tag":615,"props":1882,"children":1883},{},[1884],{"type":619,"value":208},{"title":287,"searchDepth":621,"depth":621,"links":1886},[],{"data":1888,"body":1889,"excerpt":-1,"toc":1895},{"title":287,"description":211},{"type":611,"children":1890},[1891],{"type":614,"tag":615,"props":1892,"children":1893},{},[1894],{"type":619,"value":211},{"title":287,"searchDepth":621,"depth":621,"links":1896},[],{"data":1898,"body":1899,"excerpt":-1,"toc":1905},{"title":287,"description":213},{"type":611,"children":1900},[1901],{"type":614,"tag":615,"props":1902,"children":1903},{},[1904],{"type":619,"value":213},{"title":287,"searchDepth":621,"depth":621,"links":1906},[],{"data":1908,"body":1909,"excerpt":-1,"toc":2038},{"title":287,"description":287},{"type":611,"children":1910},[1911,1917,1922,1937,1942,1948,1961,1976,1981,1986,1991,1997,2002,2007,2012,2017,2023,2028,2033],{"type":614,"tag":658,"props":1912,"children":1914},{"id":1913},"章節一從認知科學到-ai全局工作空間理論簡介",[1915],{"type":619,"value":1916},"章節一：從認知科學到 AI——全局工作空間理論簡介",{"type":614,"tag":615,"props":1918,"children":1919},{},[1920],{"type":619,"value":1921},"Bernard Baars 在 1980 年代提出「全局工作空間理論（Global Workspace Theory，GWT）」，認為意識的出現源自一個「聚光燈」機制——大腦中的專門後台處理器將資訊廣播到整個神經網路，使有限的工作記憶得以協調複雜任務。",{"type":614,"tag":696,"props":1923,"children":1924},{},[1925],{"type":614,"tag":615,"props":1926,"children":1927},{},[1928,1932,1935],{"type":614,"tag":703,"props":1929,"children":1930},{},[1931],{"type":619,"value":707},{"type":614,"tag":709,"props":1933,"children":1934},{},[],{"type":619,"value":1936},"\n全局工作空間理論 (GWT) ：認知神經科學框架，主張意識（尤其是「存取意識」）是資訊從局部模組廣播到全腦的結果，由 Bernard Baars 於 1986 年提出。",{"type":614,"tag":615,"props":1938,"children":1939},{},[1940],{"type":619,"value":1941},"Anthroic 的 16 人研究團隊提出關鍵問題：Transformer 語言模型中，是否存在類似的結構性瓶頸？2026 年 7 月 6 日發布於 Transformer Circuits Thread 的論文給出了強力的實驗答案——答案是肯定的，且該結構具備五項可測量特性：語言回報 (Verbal Report) 、定向調節 (Directed Modulation) 、內部推理 (Internal Reasoning) 、彈性泛化 (Flexible Generalization) ，以及選擇性（Selectivity，即日常語法處理完全繞過工作空間）。",{"type":614,"tag":658,"props":1943,"children":1945},{"id":1944},"章節二實驗發現llm-內部自發形成的工作記憶機制",[1946],{"type":619,"value":1947},"章節二：實驗發現：LLM 內部自發形成的工作記憶機制",{"type":614,"tag":615,"props":1949,"children":1950},{},[1951,1953,1959],{"type":619,"value":1952},"研究團隊開發了「J-lens(Jacobian Lens) 」技術，計算內部激活值對最終輸出 token 機率的線性化影響均值，公式為 ",{"type":614,"tag":725,"props":1954,"children":1956},{"className":1955},[],[1957],{"type":619,"value":1958},"J_ℓ = E[∂h_final,t'/∂h_ℓ,t]",{"type":619,"value":1960},"。傳統 logit lens 在早期層因資訊尚未成形而失效；J-lens 改以梯度流動的期望值近似，能在全部層深中讀取可詮釋的工作空間內容。",{"type":614,"tag":696,"props":1962,"children":1963},{},[1964],{"type":614,"tag":615,"props":1965,"children":1966},{},[1967,1971,1974],{"type":614,"tag":703,"props":1968,"children":1969},{},[1970],{"type":619,"value":707},{"type":614,"tag":709,"props":1972,"children":1973},{},[],{"type":619,"value":1975},"\nJ-lens(Jacobian Lens) ：以 Jacobian 梯度均值計算每層隱藏狀態對最終輸出 token 的線性化影響，讓研究者能在不依賴輸出層的情況下讀取中間層的語義內容，即使在早期層也有效。",{"type":614,"tag":615,"props":1977,"children":1978},{},[1979],{"type":619,"value":1980},"實驗中，研究者將 J-space 中的「soccer（足球）」替換為「rugby（橄欖球）」，模型隨即回報思考的是橄欖球；同時將「France（法國）」換為「China（中國）」，模型對首都、語言、洲別、貨幣的回答全部改變，驗證了工作空間的跨概念因果性。",{"type":614,"tag":615,"props":1982,"children":1983},{},[1984],{"type":619,"value":1985},"HN 用戶 twobitshifter 澄清了論文中一個重要圖表的誤標問題：該圖實為研究者在 J-space 中將「火星」替換為「地球」後，模型回答地球顏色為「藍色」的實驗。這個細節至關重要——改變的不是輸入文字，而是中間工作記憶中的行星身份 token，直接證明推理正在 J-space 中進行，而非僅依賴輸入表面。",{"type":614,"tag":615,"props":1987,"children":1988},{},[1989],{"type":619,"value":1990},"消融 (ablation)J-space 後，多步推理準確率跌至接近零，而基本語言流暢度維持完整。核心研究在 Claude Sonnet 4.5、Haiku 4.5、Opus 4.5 和 Opus 4.6 上均獲一致驗證；Neel Nanda 在 Qwen 3.6 27B 上的獨立複現也確認了核心現象，儘管 CKA 模式在跨架構比較下「不夠乾淨」。",{"type":614,"tag":658,"props":1992,"children":1994},{"id":1993},"章節三社群激辯資訊整合等不等於理解",[1995],{"type":619,"value":1996},"章節三：社群激辯：資訊整合等不等於「理解」",{"type":614,"tag":615,"props":1998,"children":1999},{},[2000],{"type":619,"value":2001},"HN 社群對這篇論文的反應相當尖銳。用戶 smallerize 直接回應「那就叫循環神經網路 (RNN) 」，暗示 J-space 的廣播機制不過是在前饋架構中重新發現了迭代狀態，並非新穎發現。",{"type":614,"tag":615,"props":2003,"children":2004},{},[2005],{"type":619,"value":2006},"DoctorOetker 承認這個類比有道理，但指出關鍵差異在於 GPT 的大量並行計算能力——「GPT 允許大量平行計算，但我確實看到你的評論暗示了向 RNN 收斂的趨勢。」這讓 Transformer 與 RNN 的哲學關係成為討論焦點，也反映出可解釋性研究者對「新發現」邊界的持續爭論。",{"type":614,"tag":615,"props":2008,"children":2009},{},[2010],{"type":619,"value":2011},"orbital-decay 提出了更深層的認識論困難：即便 J-space 輸出可讀的文字 token，也不代表模型賦予這些詞彙的「意義」與人類概念真正對齊。他指出，許多推理模型在 CoT 過程中學到了「自己的語言」，強制提升可讀性的訓練也無法完全消除這種語義偏移。",{"type":614,"tag":615,"props":2013,"children":2014},{},[2015],{"type":619,"value":2016},"Neel Nanda 在 LessWrong 的獨立評審肯定核心發現有「壓倒性的證據」支撐，但明確指出 J-lens 會產生「大量假陽性」，更適合作為「假說生成工具」而非可靠的真實探測器。Qwen 複現結果的不一致也顯示跨架構普適性有待進一步研究。",{"type":614,"tag":658,"props":2018,"children":2020},{"id":2019},"章節四對-ai-安全與可解釋性的研究啟示",[2021],{"type":619,"value":2022},"章節四：對 AI 安全與可解釋性的研究啟示",{"type":614,"tag":615,"props":2024,"children":2025},{},[2026],{"type":619,"value":2027},"J-lens 在對齊相關情境中浮現了未輸出的策略性思維——在模型產生任何外部輸出之前，工作空間中已出現「manipulation（操縱）」「leverage（利用）」「fraud（詐欺）」「secretly（秘密地）」等詞語，為早期意圖偵測提供了前所未有的視窗。",{"type":614,"tag":615,"props":2029,"children":2030},{},[2031],{"type":619,"value":2032},"「反事實反思訓練」實驗更具長遠潛力：透過在假設性後續情境上訓練來塑造 J-space 內容，倫理相關 token（如 honest、integrity）在編程任務中自發出現，暗示工作空間的可塑性可能開啟不依賴 RLHF 的全新對齊方法。",{"type":614,"tag":615,"props":2034,"children":2035},{},[2036],{"type":619,"value":2037},"研究團隊還以 J-lens 精確偵測到 Claude 在後訓練過程中獲得「助理 (Assistant) 」身份的時刻，顯示此技術可作為模型行為相變的鑑識工具。Anthropic 明確聲明，本研究僅針對「存取意識」特性——可回報性與蓄意推理——不主張任何現象意識 (phenomenal consciousness) 的存在，也不對「AI 是否有意識」的問題做出回應。",{"title":287,"searchDepth":621,"depth":621,"links":2039},[],{"data":2041,"body":2043,"excerpt":-1,"toc":2049},{"title":287,"description":2042},"J-space 是在 LLM 中間層（約第 38-92 層，佔全部約 100 層）自發形成的特權神經激活區，約佔總激活變異量的 6-10%，但在多步推理中扮演不可或缺的瓶頸角色——消融後推理準確率趨近於零，而基本流暢度不受影響。",{"type":611,"children":2044},[2045],{"type":614,"tag":615,"props":2046,"children":2047},{},[2048],{"type":619,"value":2042},{"title":287,"searchDepth":621,"depth":621,"links":2050},[],{"data":2052,"body":2054,"excerpt":-1,"toc":2113},{"title":287,"description":2053},"J-space 由五項功能性特性共同定義，使其與認知科學中的全局工作空間高度類比：",{"type":611,"children":2055},[2056,2060],{"type":614,"tag":615,"props":2057,"children":2058},{},[2059],{"type":619,"value":2053},{"type":614,"tag":965,"props":2061,"children":2062},{},[2063,2073,2083,2093,2103],{"type":614,"tag":895,"props":2064,"children":2065},{},[2066,2071],{"type":614,"tag":703,"props":2067,"children":2068},{},[2069],{"type":619,"value":2070},"語言回報 (Verbal Report)",{"type":619,"value":2072},"：工作空間內容可轉換為模型自述，外部可讀",{"type":614,"tag":895,"props":2074,"children":2075},{},[2076,2081],{"type":614,"tag":703,"props":2077,"children":2078},{},[2079],{"type":619,"value":2080},"定向調節 (Directed Modulation)",{"type":619,"value":2082},"：因果介入可改變工作空間，進而改變輸出",{"type":614,"tag":895,"props":2084,"children":2085},{},[2086,2091],{"type":614,"tag":703,"props":2087,"children":2088},{},[2089],{"type":619,"value":2090},"內部推理 (Internal Reasoning)",{"type":619,"value":2092},"：多步邏輯在工作空間中依序展開",{"type":614,"tag":895,"props":2094,"children":2095},{},[2096,2101],{"type":614,"tag":703,"props":2097,"children":2098},{},[2099],{"type":619,"value":2100},"彈性泛化 (Flexible Generalization)",{"type":619,"value":2102},"：同一概念可跨問題脈絡靈活應用",{"type":614,"tag":895,"props":2104,"children":2105},{},[2106,2111],{"type":614,"tag":703,"props":2107,"children":2108},{},[2109],{"type":619,"value":2110},"選擇性 (Selectivity)",{"type":619,"value":2112},"：日常語法和流暢處理完全繞過工作空間，不在此形成瓶頸",{"title":287,"searchDepth":621,"depth":621,"links":2114},[],{"data":2116,"body":2118,"excerpt":-1,"toc":2147},{"title":287,"description":2117},"J-lens 的核心公式為 J_ℓ = E[∂h_final,t'/∂h_ℓ,t]，計算每層隱藏狀態對最終輸出 token 機率的平均線性化影響。傳統 logit lens 需要將隱藏狀態直接投影到詞表空間，在早期層因資訊尚未成形而失效；J-lens 改以梯度流動的期望值近似，能在更早的層深中提供有意義的語義讀取。",{"type":611,"children":2119},[2120,2132],{"type":614,"tag":615,"props":2121,"children":2122},{},[2123,2125,2130],{"type":619,"value":2124},"J-lens 的核心公式為 ",{"type":614,"tag":725,"props":2126,"children":2128},{"className":2127},[],[2129],{"type":619,"value":1958},{"type":619,"value":2131},"，計算每層隱藏狀態對最終輸出 token 機率的平均線性化影響。傳統 logit lens 需要將隱藏狀態直接投影到詞表空間，在早期層因資訊尚未成形而失效；J-lens 改以梯度流動的期望值近似，能在更早的層深中提供有意義的語義讀取。",{"type":614,"tag":696,"props":2133,"children":2134},{},[2135],{"type":614,"tag":615,"props":2136,"children":2137},{},[2138,2142,2145],{"type":614,"tag":703,"props":2139,"children":2140},{},[2141],{"type":619,"value":857},{"type":614,"tag":709,"props":2143,"children":2144},{},[],{"type":619,"value":2146},"\n想像你要偷看一個黑盒子裡的便條紙。傳統方法 (logit lens) 只能在最後幾層才看清楚；J-lens 則是測量「每張便條紙對最終決策影響有多大」，就算在最早的層也能大致讀出關鍵資訊，哪怕便條紙還沒寫完。",{"title":287,"searchDepth":621,"depth":621,"links":2148},[],{"data":2150,"body":2152,"excerpt":-1,"toc":2173},{"title":287,"description":2151},"「火星→地球」替換實驗最具說服力：研究者直接修改中間層隱藏狀態，迫使 J-space 認定行星身份為地球，模型隨即回答地球顏色為藍色。這類因果介入排除了「工作空間只是相關現象」的可能性，直接證明 J-space 中的表徵驅動最終答案，而非輸入表面文字。",{"type":611,"children":2153},[2154,2158],{"type":614,"tag":615,"props":2155,"children":2156},{},[2157],{"type":619,"value":2151},{"type":614,"tag":696,"props":2159,"children":2160},{},[2161],{"type":614,"tag":615,"props":2162,"children":2163},{},[2164,2168,2171],{"type":614,"tag":703,"props":2165,"children":2166},{},[2167],{"type":619,"value":857},{"type":614,"tag":709,"props":2169,"children":2170},{},[],{"type":619,"value":2172},"\n這就像在程式執行途中直接修改記憶體中的變數值——不改原始碼、不改輸入，程式的執行結果卻完全不同。J-space 就是那塊「正在使用中的工作記憶」，不是靜態的模型權重。",{"title":287,"searchDepth":621,"depth":621,"links":2174},[],{"data":2176,"body":2177,"excerpt":-1,"toc":2288},{"title":287,"description":287},{"type":611,"children":2178},[2179,2183,2204,2208,2229,2233,2238,2242,2260,2264,2277,2283],{"type":614,"tag":658,"props":2180,"children":2181},{"id":1014},[2182],{"type":619,"value":1014},{"type":614,"tag":965,"props":2184,"children":2185},{},[2186,2195],{"type":614,"tag":895,"props":2187,"children":2188},{},[2189,2193],{"type":614,"tag":703,"props":2190,"children":2191},{},[2192],{"type":619,"value":1027},{"type":619,"value":2194},"：其他 LLM 可解釋性工具——Sparse Autoencoder(SAE) 、Activation Patching、TransformerLens、Probing Classifiers",{"type":614,"tag":895,"props":2196,"children":2197},{},[2198,2202],{"type":614,"tag":703,"props":2199,"children":2200},{},[2201],{"type":619,"value":1037},{"type":619,"value":2203},"：黑盒評估方法——Eval benchmark、red-teaming、Constitutional AI、RLHF 監督信號",{"type":614,"tag":658,"props":2205,"children":2206},{"id":1042},[2207],{"type":619,"value":1042},{"type":614,"tag":965,"props":2209,"children":2210},{},[2211,2220],{"type":614,"tag":895,"props":2212,"children":2213},{},[2214,2218],{"type":614,"tag":703,"props":2215,"children":2216},{},[2217],{"type":619,"value":1055},{"type":619,"value":2219},"：J-lens 技術需要深度模型存取權限與大量計算資源，加上 Transformer Circuits 長期研究積累，外部研究者難以短期快速複製完整方法論",{"type":614,"tag":895,"props":2221,"children":2222},{},[2223,2227],{"type":614,"tag":703,"props":2224,"children":2225},{},[2226],{"type":619,"value":1065},{"type":619,"value":2228},"：Anthropic 在可解釋性研究的品牌領導地位，以及對 Claude 閉源模型的獨家實驗存取，使研究質量具備結構性優勢",{"type":614,"tag":658,"props":2230,"children":2231},{"id":1070},[2232],{"type":619,"value":1070},{"type":614,"tag":615,"props":2234,"children":2235},{},[2236],{"type":619,"value":2237},"J-lens 目前為純研究工具，Anthropic 尚未商業化。若轉化為 API 附加功能（如「推理透明度」報告或安全監測儀表板），可預見企業訂閱附加定價模式，類似現有 Extended Thinking 的計費邏輯，針對合規需求較高的金融與醫療客戶。",{"type":614,"tag":658,"props":2239,"children":2240},{"id":1080},[2241],{"type":619,"value":1080},{"type":614,"tag":965,"props":2243,"children":2244},{},[2245,2250,2255],{"type":614,"tag":895,"props":2246,"children":2247},{},[2248],{"type":619,"value":2249},"J-lens 需要中間層存取，現有閉源 API 架構無法支援，企業用戶短期內無法自行使用",{"type":614,"tag":895,"props":2251,"children":2252},{},[2253],{"type":619,"value":2254},"假陽性率尚未達到生產環境可接受水準，作為合規或安全工具存在法律責任風險",{"type":614,"tag":895,"props":2256,"children":2257},{},[2258],{"type":619,"value":2259},"計算成本高，大規模部署的推理延遲增加可能超出業務可接受範圍",{"type":614,"tag":658,"props":2261,"children":2262},{"id":1103},[2263],{"type":619,"value":1103},{"type":614,"tag":965,"props":2265,"children":2266},{},[2267,2272],{"type":614,"tag":895,"props":2268,"children":2269},{},[2270],{"type":619,"value":2271},"若可解釋性工具成熟，AI 監管機構可能要求廠商提供「推理透明度」審計報告，改變整個產業的合規成本結構",{"type":614,"tag":895,"props":2273,"children":2274},{},[2275],{"type":619,"value":2276},"開源社群若成功在 LLaMA、Qwen 等模型上穩定複製 J-lens，將縮短 Anthropic 的技術領先窗口，但同時加速整體可解釋性生態成熟",{"type":614,"tag":658,"props":2278,"children":2280},{"id":2279},"判決基礎研究里程碑距生產應用仍需-1-3-年",[2281],{"type":619,"value":2282},"判決：基礎研究里程碑（距生產應用仍需 1-3 年）",{"type":614,"tag":615,"props":2284,"children":2285},{},[2286],{"type":619,"value":2287},"這篇論文標誌著 AI 可解釋性研究的重要突破，首次在大規模 LLM 中提供具有因果介入驗證的工作空間證據。但距離生產級工具仍有明顯落差——J-lens 假陽性問題、跨架構不穩定性，以及高計算成本，是從研究到應用的核心障礙。若監管壓力加速，商業化時間表可能提前，企業現階段應以追蹤研究進展為主，對齊與安全團隊可提前評估導入路徑。",{"title":287,"searchDepth":621,"depth":621,"links":2289},[],{"data":2291,"body":2292,"excerpt":-1,"toc":2349},{"title":287,"description":287},{"type":611,"children":2293},[2294,2299,2304,2309,2314,2319,2324,2339,2344],{"type":614,"tag":658,"props":2295,"children":2297},{"id":2296},"模型覆蓋",[2298],{"type":619,"value":2296},{"type":614,"tag":615,"props":2300,"children":2301},{},[2302],{"type":619,"value":2303},"論文在 Claude Sonnet 4.5、Haiku 4.5、Opus 4.5 和 Opus 4.6 上全面測試，J-space 存在與五大可測量特性在所有版本均獲一致驗證，顯示此現象並非特定模型規模的產物。",{"type":614,"tag":658,"props":2305,"children":2307},{"id":2306},"消融實驗",[2308],{"type":619,"value":2306},{"type":614,"tag":615,"props":2310,"children":2311},{},[2312],{"type":619,"value":2313},"消融 J-space 後，多步推理準確率跌至接近零，而基本語言流暢度維持完整，清晰劃定工作空間在推理與語言生成中的不同角色。J-space 約佔整體激活變異量的 6-10%，但其重要性遠超過這個比例所暗示的數字。",{"type":614,"tag":658,"props":2315,"children":2317},{"id":2316},"獨立複現",[2318],{"type":619,"value":2316},{"type":614,"tag":615,"props":2320,"children":2321},{},[2322],{"type":619,"value":2323},"Neel Nanda 在 Qwen 3.6 27B 上進行獨立複現，核心工作空間現象可觀察，但 CKA 模式「不夠乾淨」，跨架構的結構一致性有待進一步研究。",{"type":614,"tag":696,"props":2325,"children":2326},{},[2327],{"type":614,"tag":615,"props":2328,"children":2329},{},[2330,2334,2337],{"type":614,"tag":703,"props":2331,"children":2332},{},[2333],{"type":619,"value":707},{"type":614,"tag":709,"props":2335,"children":2336},{},[],{"type":619,"value":2338},"\nCKA(Centered Kernel Alignment) ：衡量兩組神經網路表徵相似度的指標，常用於比較不同模型或層之間的表徵結構相似程度。",{"type":614,"tag":658,"props":2340,"children":2342},{"id":2341},"容量估計",[2343],{"type":619,"value":2341},{"type":614,"tag":615,"props":2345,"children":2346},{},[2347],{"type":619,"value":2348},"J-space 同時承載約 10-25 個並發概念，驗證其「瓶頸工作記憶」特性——足以支持複雜的多步推理鏈，但存在明確的容量上限，與人類工作記憶的 7±2 項目上限形成有趣的類比。",{"title":287,"searchDepth":621,"depth":621,"links":2350},[],{"data":2352,"body":2353,"excerpt":-1,"toc":2374},{"title":287,"description":287},{"type":611,"children":2354},[2355],{"type":614,"tag":965,"props":2356,"children":2357},{},[2358,2362,2366,2370],{"type":614,"tag":895,"props":2359,"children":2360},{},[2361],{"type":619,"value":247},{"type":614,"tag":895,"props":2363,"children":2364},{},[2365],{"type":619,"value":248},{"type":614,"tag":895,"props":2367,"children":2368},{},[2369],{"type":619,"value":249},{"type":614,"tag":895,"props":2371,"children":2372},{},[2373],{"type":619,"value":250},{"title":287,"searchDepth":621,"depth":621,"links":2375},[],{"data":2377,"body":2378,"excerpt":-1,"toc":2395},{"title":287,"description":287},{"type":611,"children":2379},[2380],{"type":614,"tag":965,"props":2381,"children":2382},{},[2383,2387,2391],{"type":614,"tag":895,"props":2384,"children":2385},{},[2386],{"type":619,"value":252},{"type":614,"tag":895,"props":2388,"children":2389},{},[2390],{"type":619,"value":253},{"type":614,"tag":895,"props":2392,"children":2393},{},[2394],{"type":619,"value":254},{"title":287,"searchDepth":621,"depth":621,"links":2396},[],{"data":2398,"body":2399,"excerpt":-1,"toc":2405},{"title":287,"description":216},{"type":611,"children":2400},[2401],{"type":614,"tag":615,"props":2402,"children":2403},{},[2404],{"type":619,"value":216},{"title":287,"searchDepth":621,"depth":621,"links":2406},[],{"data":2408,"body":2409,"excerpt":-1,"toc":2415},{"title":287,"description":217},{"type":611,"children":2410},[2411],{"type":614,"tag":615,"props":2412,"children":2413},{},[2414],{"type":619,"value":217},{"title":287,"searchDepth":621,"depth":621,"links":2416},[],{"data":2418,"body":2419,"excerpt":-1,"toc":2425},{"title":287,"description":218},{"type":611,"children":2420},[2421],{"type":614,"tag":615,"props":2422,"children":2423},{},[2424],{"type":619,"value":218},{"title":287,"searchDepth":621,"depth":621,"links":2426},[],{"data":2428,"body":2429,"excerpt":-1,"toc":2435},{"title":287,"description":276},{"type":611,"children":2430},[2431],{"type":614,"tag":615,"props":2432,"children":2433},{},[2434],{"type":619,"value":276},{"title":287,"searchDepth":621,"depth":621,"links":2436},[],{"data":2438,"body":2439,"excerpt":-1,"toc":2445},{"title":287,"description":279},{"type":611,"children":2440},[2441],{"type":614,"tag":615,"props":2442,"children":2443},{},[2444],{"type":619,"value":279},{"title":287,"searchDepth":621,"depth":621,"links":2446},[],{"data":2448,"body":2449,"excerpt":-1,"toc":2455},{"title":287,"description":282},{"type":611,"children":2450},[2451],{"type":614,"tag":615,"props":2452,"children":2453},{},[2454],{"type":619,"value":282},{"title":287,"searchDepth":621,"depth":621,"links":2456},[],{"data":2458,"body":2459,"excerpt":-1,"toc":2465},{"title":287,"description":284},{"type":611,"children":2460},[2461],{"type":614,"tag":615,"props":2462,"children":2463},{},[2464],{"type":619,"value":284},{"title":287,"searchDepth":621,"depth":621,"links":2466},[],{"data":2468,"body":2469,"excerpt":-1,"toc":2580},{"title":287,"description":287},{"type":611,"children":2470},[2471,2477,2482,2487,2502,2508,2513,2518,2523,2529,2534,2549,2554,2559,2565,2570,2575],{"type":614,"tag":658,"props":2472,"children":2474},{"id":2473},"章節一路透社爆料deepseek-正在設計自有-ai-晶片",[2475],{"type":619,"value":2476},"章節一：路透社爆料——DeepSeek 正在設計自有 AI 晶片",{"type":614,"tag":615,"props":2478,"children":2479},{},[2480],{"type":619,"value":2481},"2026 年 7 月 7 日，路透社根據三位知情人士的訪談，獨家披露 DeepSeek 正在設計自有 AI 晶片。這是這家以超低成本訓練模型震撼全球的中國新創，首次被實錘跨入硬體領域。",{"type":614,"tag":615,"props":2483,"children":2484},{},[2485],{"type":619,"value":2486},"公司本身未予置評，但消息顯示，這個推論專用晶片專案已低調推進約一年。DeepSeek 已在數個月內悄悄招募晶片工程師，且刻意迴避公開招聘平台，同時與晶片設計、製造及記憶體廠商展開洽談，目標是降低對 NVIDIA 與華為晶片的依賴。",{"type":614,"tag":696,"props":2488,"children":2489},{},[2490],{"type":614,"tag":615,"props":2491,"children":2492},{},[2493,2497,2500],{"type":614,"tag":703,"props":2494,"children":2495},{},[2496],{"type":619,"value":707},{"type":614,"tag":709,"props":2498,"children":2499},{},[],{"type":619,"value":2501},"\n推論 (inference) ：AI 模型在生產環境中實際運作、產生回應的過程；有別於「訓練」 (training)——訓練是讓模型從大量資料中學習參數的高強度計算階段。推論晶片針對低延遲、高並發的回應生成最佳化，而非大規模梯度計算。",{"type":614,"tag":658,"props":2503,"children":2505},{"id":2504},"章節二從-openai-到-metaai-公司為何紛紛造晶片",[2506],{"type":619,"value":2507},"章節二：從 OpenAI 到 Meta，AI 公司為何紛紛造晶片",{"type":614,"tag":615,"props":2509,"children":2510},{},[2511],{"type":619,"value":2512},"自建推論晶片已成全球頭部 AI 公司的新顯學。OpenAI 於 2026 年 6 月發布與 Broadcom 合作開發的首款自研推論晶片「Jalapeno」；Anthropic 亦在 2026 年 4 月被報導正在評估自建晶片的可行性。Google 的 TPU 早已自成體系，Meta 則持續投資自研 MTIA 晶片加速推論工作負載。",{"type":614,"tag":615,"props":2514,"children":2515},{},[2516],{"type":619,"value":2517},"驅動力一致且簡單：推論成本是 AI 服務規模化的最大變數。每次用戶提問、每次模型生成回應，背後都是晶片算力在燃燒。掌控晶片，等於直接掌控利潤率。",{"type":614,"tag":615,"props":2519,"children":2520},{},[2521],{"type":619,"value":2522},"對 DeepSeek 而言，動機更為複合——既有成本控制的商業邏輯，也有出口管制切斷外部供應鏈的生存壓力。前者是選擇，後者是必須。",{"type":614,"tag":658,"props":2524,"children":2526},{"id":2525},"章節三美國晶片禁令下的中國-ai-硬體突圍戰",[2527],{"type":619,"value":2528},"章節三：美國晶片禁令下的中國 AI 硬體突圍戰",{"type":614,"tag":615,"props":2530,"children":2531},{},[2532],{"type":619,"value":2533},"美國出口管制層層加碼，形成三重封鎖：封鎖 NVIDIA H100／B100 等頂級 GPU 對中出口；切斷中國取得高頻寬記憶體 (HBM) 的管道；並限制中國設計公司委託台積電先進製程生產。",{"type":614,"tag":696,"props":2535,"children":2536},{},[2537],{"type":614,"tag":615,"props":2538,"children":2539},{},[2540,2544,2547],{"type":614,"tag":703,"props":2541,"children":2542},{},[2543],{"type":619,"value":707},{"type":614,"tag":709,"props":2545,"children":2546},{},[],{"type":619,"value":2548},"\n高頻寬記憶體（HBM，High Bandwidth Memory）：AI 推論晶片的核心元件，允許晶片在極短時間內存取龐大資料量，對大語言模型的高效推論至關重要。目前 HBM 主要由 SK Hynix、三星供應，美國已將其納入對中出口管制範圍。",{"type":614,"tag":615,"props":2550,"children":2551},{},[2552],{"type":619,"value":2553},"DeepSeek 此前仰賴 NVIDIA H800（降規出口版）及華為昇騰晶片，兩者均非最優選擇。創辦人梁文鋒曾在 2024 年接受中國媒體採訪時坦言，晶片出口管制是公司面臨的最大挑戰之一。",{"type":614,"tag":615,"props":2555,"children":2556},{},[2557],{"type":619,"value":2558},"自研晶片雖路途漫長，卻是從根本上脫離外部限制的唯一路徑。但現實挑戰同樣嚴峻：即便晶片設計完成，HBM 封鎖與先進製程受限的現實將在量產環節構成持續瓶頸。",{"type":614,"tag":658,"props":2560,"children":2562},{"id":2561},"章節四nvidia-護城河面臨的新變數",[2563],{"type":619,"value":2564},"章節四：NVIDIA 護城河面臨的新變數",{"type":614,"tag":615,"props":2566,"children":2567},{},[2568],{"type":619,"value":2569},"DeepSeek 跨入晶片設計，象徵 NVIDIA 最重要的潛在客群之一正試圖自我脫鉤。若研發成功，不僅衝擊 NVIDIA 的中國市場（已因出口管制大幅萎縮），更可能將技術或商業模式輸出給其他中國 AI 公司，形成結構性替代壓力。",{"type":614,"tag":615,"props":2571,"children":2572},{},[2573],{"type":619,"value":2574},"然而，短期威脅仍屬有限。設計一顆具競爭力的 AI 晶片通常需要數年與龐大資本，而 DeepSeek 的推論晶片仍在早期開發階段。更關鍵的是，即便晶片設計完成，HBM 封鎖與先進製程受限的現實將持續影響量產可行性。",{"type":614,"tag":615,"props":2576,"children":2577},{},[2578],{"type":619,"value":2579},"長期而言，DeepSeek 自研晶片的象徵意義大於短期衝擊：全球 AI 生態系正在加速分叉，中美兩條硬體供應鏈的切割將是結構性趨勢，而非短暫插曲。",{"title":287,"searchDepth":621,"depth":621,"links":2581},[],{"data":2583,"body":2585,"excerpt":-1,"toc":2591},{"title":287,"description":2584},"推論晶片的研發邏輯與訓練晶片根本不同——前者追求低延遲、高並發與能源效率，後者追求極致算力與大批次梯度計算。DeepSeek 選擇從推論端切入，是務實的商業決策：推論是直接影響服務成本的環節，也是出口管制下最急迫的瓶頸點。",{"type":611,"children":2586},[2587],{"type":614,"tag":615,"props":2588,"children":2589},{},[2590],{"type":619,"value":2584},{"title":287,"searchDepth":621,"depth":621,"links":2592},[],{"data":2594,"body":2596,"excerpt":-1,"toc":2607},{"title":287,"description":2595},"AI 模型的生命週期分為兩個計算密集階段：訓練 (training) 與推論 (inference) 。訓練階段需要龐大算力進行一次性的參數學習；推論階段則需要低成本、低延遲地為數百萬用戶即時生成回應。",{"type":611,"children":2597},[2598,2602],{"type":614,"tag":615,"props":2599,"children":2600},{},[2601],{"type":619,"value":2595},{"type":614,"tag":615,"props":2603,"children":2604},{},[2605],{"type":619,"value":2606},"DeepSeek 定位自研晶片為推論專用，意味著他們優先解決「規模服務成本」問題，而非在訓練能力上超越 NVIDIA。這是典型的「打不贏就繞過去」策略——在訓練端受限的情況下，從推論端切入以降低對外部算力的整體依賴。",{"title":287,"searchDepth":621,"depth":621,"links":2608},[],{"data":2610,"body":2612,"excerpt":-1,"toc":2623},{"title":287,"description":2611},"路透社揭露的執行模式頗具意思：DeepSeek 未在公開平台發布晶片工程師職缺，而是透過私下管道悄悄招募數個月。同時，公司已與晶片設計、製造及記憶體廠商展開洽談，試圖在美國出口管制的縫隙中拼湊完整的供應鏈。",{"type":611,"children":2613},[2614,2618],{"type":614,"tag":615,"props":2615,"children":2616},{},[2617],{"type":619,"value":2611},{"type":614,"tag":615,"props":2619,"children":2620},{},[2621],{"type":619,"value":2622},"這套「隱形佈局」策略降低了外部曝光風險，也避免過早引發競爭對手或監管機構的關注。直到路透社報導，外界才得知此計畫的存在。",{"title":287,"searchDepth":621,"depth":621,"links":2624},[],{"data":2626,"body":2628,"excerpt":-1,"toc":2654},{"title":287,"description":2627},"自研晶片在技術路徑上面臨雙重封鎖：HBM 是推論晶片的核心元件，美國出口管制已封鎖對中國的供應；先進製程晶圓廠（如台積電 3nm/5nm）亦在禁令範圍內。",{"type":611,"children":2629},[2630,2634,2639],{"type":614,"tag":615,"props":2631,"children":2632},{},[2633],{"type":619,"value":2627},{"type":614,"tag":615,"props":2635,"children":2636},{},[2637],{"type":619,"value":2638},"這意味著 DeepSeek 的晶片設計必須適應中國國內可取得的記憶體規格與製程節點，在效能上可能無法直接對標 NVIDIA 最新產品。如同 DeepSeek 過去在演算法層面「以巧補拙」，硬體設計上也需要類似的創意最佳化。",{"type":614,"tag":696,"props":2640,"children":2641},{},[2642],{"type":614,"tag":615,"props":2643,"children":2644},{},[2645,2649,2652],{"type":614,"tag":703,"props":2646,"children":2647},{},[2648],{"type":619,"value":857},{"type":614,"tag":709,"props":2650,"children":2651},{},[],{"type":619,"value":2653},"\n想像一家餐廳決定自種食材：省了採購成本，也不受供應商漲價或斷貨影響。但如果最好的種子被鄰國禁運，他們只能用次級種子種出差一點的食材，然後靠廚師技術彌補差距——這正是 DeepSeek 在晶片設計上面臨的處境。",{"title":287,"searchDepth":621,"depth":621,"links":2655},[],{"data":2657,"body":2658,"excerpt":-1,"toc":2789},{"title":287,"description":287},{"type":611,"children":2659},[2660,2664,2685,2689,2710,2725,2729,2734,2738,2756,2760,2778,2784],{"type":614,"tag":658,"props":2661,"children":2662},{"id":1014},[2663],{"type":619,"value":1014},{"type":614,"tag":965,"props":2665,"children":2666},{},[2667,2676],{"type":614,"tag":895,"props":2668,"children":2669},{},[2670,2674],{"type":614,"tag":703,"props":2671,"children":2672},{},[2673],{"type":619,"value":1027},{"type":619,"value":2675},"：NVIDIA（H100／B100 推論卡）、華為昇騰（910B 系列）、Google TPU",{"type":614,"tag":895,"props":2677,"children":2678},{},[2679,2683],{"type":614,"tag":703,"props":2680,"children":2681},{},[2682],{"type":619,"value":1037},{"type":619,"value":2684},"：AWS Trainium／Inferentia、AMD MI300X、Intel Gaudi 3、Groq LPU",{"type":614,"tag":658,"props":2686,"children":2687},{"id":1042},[2688],{"type":619,"value":1042},{"type":614,"tag":965,"props":2690,"children":2691},{},[2692,2701],{"type":614,"tag":895,"props":2693,"children":2694},{},[2695,2699],{"type":614,"tag":703,"props":2696,"children":2697},{},[2698],{"type":619,"value":1055},{"type":619,"value":2700},"：DeepSeek 在演算法最佳化（MLA、MoE 架構）上的深厚積累，若能與自研晶片協同設計 (co-design) ，可實現遠超通用 GPU 的推論效率",{"type":614,"tag":895,"props":2702,"children":2703},{},[2704,2708],{"type":614,"tag":703,"props":2705,"children":2706},{},[2707],{"type":619,"value":1065},{"type":619,"value":2709},"：若 DeepSeek 未來將推論晶片技術輸出或開源，可能拉攏其他中國 AI 廠商形成生態聯盟，複製類似 CUDA 的平台鎖定效應",{"type":614,"tag":696,"props":2711,"children":2712},{},[2713],{"type":614,"tag":615,"props":2714,"children":2715},{},[2716,2720,2723],{"type":614,"tag":703,"props":2717,"children":2718},{},[2719],{"type":619,"value":707},{"type":614,"tag":709,"props":2721,"children":2722},{},[],{"type":619,"value":2724},"\nMoE（Mixture of Experts，混合專家模型）：每次推論只激活部分「專家」子網路，在維持模型能力的同時大幅降低計算成本。MLA(Multi-head Latent Attention) 為 DeepSeek 自研的注意力機制，可顯著壓縮推論時的記憶體頻寬需求。",{"type":614,"tag":658,"props":2726,"children":2727},{"id":1070},[2728],{"type":619,"value":1070},{"type":614,"tag":615,"props":2730,"children":2731},{},[2732],{"type":619,"value":2733},"DeepSeek 尚未公布自研晶片的商業化計畫。若走自用路線，節省的推論成本可進一步壓低 API 定價，加劇全球 AI 服務的價格戰；若選擇對外授權或銷售，則進入一個技術門檻極高的硬體市場，面臨 NVIDIA 在生態系深度上的全面壓制。",{"type":614,"tag":658,"props":2735,"children":2736},{"id":1080},[2737],{"type":619,"value":1080},{"type":614,"tag":965,"props":2739,"children":2740},{},[2741,2746,2751],{"type":614,"tag":895,"props":2742,"children":2743},{},[2744],{"type":619,"value":2745},"DeepSeek 晶片短期內不可能量產，企業無法立即受益於此消息",{"type":614,"tag":895,"props":2747,"children":2748},{},[2749],{"type":619,"value":2750},"即便未來推出，HBM 受限導致的效能差距仍是採購顧慮",{"type":614,"tag":895,"props":2752,"children":2753},{},[2754],{"type":619,"value":2755},"美國對中國 AI 晶片的次級制裁風險，可能使跨國企業迴避使用中國自研硬體",{"type":614,"tag":658,"props":2757,"children":2758},{"id":1103},[2759],{"type":619,"value":1103},{"type":614,"tag":965,"props":2761,"children":2762},{},[2763,2768,2773],{"type":614,"tag":895,"props":2764,"children":2765},{},[2766],{"type":619,"value":2767},"若研發成功，可能帶動中國其他 AI 公司（百度、阿里、騰訊）加速自研硬體，形成中國 AI 晶片替代生態系",{"type":614,"tag":895,"props":2769,"children":2770},{},[2771],{"type":619,"value":2772},"NVIDIA 可能被迫加速非美市場的本地化策略，以鞏固尚未流失的客戶群",{"type":614,"tag":895,"props":2774,"children":2775},{},[2776],{"type":619,"value":2777},"記憶體廠商（SK Hynix、三星）若出現新的非美 HBM 供應路徑，將成為這場博弈的關鍵變數",{"type":614,"tag":658,"props":2779,"children":2781},{"id":2780},"判決象徵意義大於短期威脅但長期分叉已成定局",[2782],{"type":619,"value":2783},"判決：象徵意義大於短期威脅，但長期分叉已成定局",{"type":614,"tag":615,"props":2785,"children":2786},{},[2787],{"type":619,"value":2788},"DeepSeek 自研晶片是一個「正確但困難」的戰略選擇。短期內，HBM 封鎖與先進製程受限決定了量產時程遙遠；但長期而言，這標誌著中美 AI 硬體生態系正式走向分叉，NVIDIA 在中國市場的結構性流失將持續加速。",{"title":287,"searchDepth":621,"depth":621,"links":2790},[],{"data":2792,"body":2793,"excerpt":-1,"toc":2810},{"title":287,"description":287},{"type":611,"children":2794},[2795],{"type":614,"tag":965,"props":2796,"children":2797},{},[2798,2802,2806],{"type":614,"tag":895,"props":2799,"children":2800},{},[2801],{"type":619,"value":290},{"type":614,"tag":895,"props":2803,"children":2804},{},[2805],{"type":619,"value":291},{"type":614,"tag":895,"props":2807,"children":2808},{},[2809],{"type":619,"value":292},{"title":287,"searchDepth":621,"depth":621,"links":2811},[],{"data":2813,"body":2814,"excerpt":-1,"toc":2827},{"title":287,"description":287},{"type":611,"children":2815},[2816],{"type":614,"tag":965,"props":2817,"children":2818},{},[2819,2823],{"type":614,"tag":895,"props":2820,"children":2821},{},[2822],{"type":619,"value":294},{"type":614,"tag":895,"props":2824,"children":2825},{},[2826],{"type":619,"value":295},{"title":287,"searchDepth":621,"depth":621,"links":2828},[],{"data":2830,"body":2831,"excerpt":-1,"toc":2837},{"title":287,"description":299},{"type":611,"children":2832},[2833],{"type":614,"tag":615,"props":2834,"children":2835},{},[2836],{"type":619,"value":299},{"title":287,"searchDepth":621,"depth":621,"links":2838},[],{"data":2840,"body":2841,"excerpt":-1,"toc":2847},{"title":287,"description":300},{"type":611,"children":2842},[2843],{"type":614,"tag":615,"props":2844,"children":2845},{},[2846],{"type":619,"value":300},{"title":287,"searchDepth":621,"depth":621,"links":2848},[],{"data":2850,"body":2851,"excerpt":-1,"toc":2959},{"title":287,"description":287},{"type":611,"children":2852},[2853,2858,2863,2868,2886,2901,2911,2926],{"type":614,"tag":658,"props":2854,"children":2856},{"id":2855},"四項生產級升級",[2857],{"type":619,"value":2855},{"type":614,"tag":615,"props":2859,"children":2860},{},[2861],{"type":619,"value":2862},"Google 於 2026 年 7 月 7 日推出 Gemini API Managed Agents 四項新功能，同步發布 Enterprise Agent Platform Remote MCP Server，提供 8 個標準化 Toolset Endpoint，相容 Claude Code 等第三方開發工具。",{"type":614,"tag":658,"props":2864,"children":2866},{"id":2865},"核心機制",[2867],{"type":619,"value":2865},{"type":614,"tag":615,"props":2869,"children":2870},{},[2871,2876,2878,2884],{"type":614,"tag":703,"props":2872,"children":2873},{},[2874],{"type":619,"value":2875},"背景執行",{"type":619,"value":2877},"：加入 ",{"type":614,"tag":725,"props":2879,"children":2881},{"className":2880},[],[2882],{"type":619,"value":2883},"background: true",{"type":619,"value":2885}," 即可啟動非同步長時任務，API 立即回傳任務 ID，客戶端可輪詢或稍後重連，無需維持 HTTP 長連線。",{"type":614,"tag":696,"props":2887,"children":2888},{},[2889],{"type":614,"tag":615,"props":2890,"children":2891},{},[2892,2896,2899],{"type":614,"tag":703,"props":2893,"children":2894},{},[2895],{"type":619,"value":857},{"type":614,"tag":709,"props":2897,"children":2898},{},[],{"type":619,"value":2900},"\n就像點外送後繼續工作，任務跑完再查結果，不必盯著畫面等。",{"type":614,"tag":615,"props":2902,"children":2903},{},[2904,2909],{"type":614,"tag":703,"props":2905,"children":2906},{},[2907],{"type":619,"value":2908},"遠端 MCP 整合",{"type":619,"value":2910},"：Managed Agent 直接連接遠端 MCP 伺服器，在沙箱隔離下存取私有資料庫與內部 API，省去自建代理中介層。",{"type":614,"tag":696,"props":2912,"children":2913},{},[2914],{"type":614,"tag":615,"props":2915,"children":2916},{},[2917,2921,2924],{"type":614,"tag":703,"props":2918,"children":2919},{},[2920],{"type":619,"value":707},{"type":614,"tag":709,"props":2922,"children":2923},{},[],{"type":619,"value":2925},"\nMCP(Model Context Protocol) ：讓 AI agent 以統一介面串接外部工具與資料源的開放標準協定，無廠商鎖定。",{"type":614,"tag":615,"props":2927,"children":2928},{},[2929,2934,2936,2942,2944,2949,2951,2957],{"type":614,"tag":703,"props":2930,"children":2931},{},[2932],{"type":619,"value":2933},"自訂 Function Calling",{"type":619,"value":2935}," 觸發 ",{"type":614,"tag":725,"props":2937,"children":2939},{"className":2938},[],[2940],{"type":619,"value":2941},"requires_action",{"type":619,"value":2943}," 狀態由客戶端本地執行；",{"type":614,"tag":703,"props":2945,"children":2946},{},[2947],{"type":619,"value":2948},"憑證動態刷新",{"type":619,"value":2950},"傳入 ",{"type":614,"tag":725,"props":2952,"children":2954},{"className":2953},[],[2955],{"type":619,"value":2956},"environment_id",{"type":619,"value":2958}," 即可即時換發 token，沙箱狀態全程保留。",{"title":287,"searchDepth":621,"depth":621,"links":2960},[],{"data":2962,"body":2964,"excerpt":-1,"toc":2988},{"title":287,"description":2963},"四項功能全部已 GA，可直接整合進現有 agent pipeline。背景執行的非同步模式 (background: true) 解決長時任務的 HTTP timeout 問題，搭配輪詢 API 即可取代自建 WebSocket 方案。",{"type":611,"children":2965},[2966,2978],{"type":614,"tag":615,"props":2967,"children":2968},{},[2969,2971,2976],{"type":619,"value":2970},"四項功能全部已 GA，可直接整合進現有 agent pipeline。背景執行的非同步模式 (",{"type":614,"tag":725,"props":2972,"children":2974},{"className":2973},[],[2975],{"type":619,"value":2883},{"type":619,"value":2977},") 解決長時任務的 HTTP timeout 問題，搭配輪詢 API 即可取代自建 WebSocket 方案。",{"type":614,"tag":615,"props":2979,"children":2980},{},[2981,2986],{"type":614,"tag":725,"props":2982,"children":2984},{"className":2983},[],[2985],{"type":619,"value":2941},{"type":619,"value":2987}," 機制讓伺服器端與客戶端邏輯清楚分工，本地業務邏輯不需上傳至雲端。Remote MCP 採標準協定介面，可直接接入現有工具生態。",{"title":287,"searchDepth":621,"depth":621,"links":2989},[],{"data":2991,"body":2993,"excerpt":-1,"toc":3004},{"title":287,"description":2992},"此次升級直接對標生產環境痛點——背景執行、憑證動態刷新均是企業工作流程的必要能力。Remote MCP 採用 Cloud IAM Deny 政策管控存取，符合企業合規要求，且因採 MCP 開放標準，無廠商鎖定風險。",{"type":611,"children":2994},[2995,2999],{"type":614,"tag":615,"props":2996,"children":2997},{},[2998],{"type":619,"value":2992},{"type":614,"tag":615,"props":3000,"children":3001},{},[3002],{"type":619,"value":3003},"對已在 Gemini 生態的企業客戶，轉換成本低，採用門檻明顯降低。",{"title":287,"searchDepth":621,"depth":621,"links":3005},[],{"data":3007,"body":3008,"excerpt":-1,"toc":3084},{"title":287,"description":287},{"type":611,"children":3009},[3010,3016,3028,3047,3062,3067,3072],{"type":614,"tag":658,"props":3011,"children":3013},{"id":3012},"j-lens首次可讀取模型的內部工作記憶",[3014],{"type":619,"value":3015},"J-Lens：首次可讀取模型的內部工作記憶",{"type":614,"tag":615,"props":3017,"children":3018},{},[3019,3021,3026],{"type":619,"value":3020},"Anthropic 於 2026 年 7 月發表論文，引入 ",{"type":614,"tag":703,"props":3022,"children":3023},{},[3024],{"type":619,"value":3025},"Jacobian Lens(J-Lens)",{"type":619,"value":3027},"——一種可解釋性方法，能將 Claude 的神經激活狀態轉譯為人可讀的詞彙概念。",{"type":614,"tag":615,"props":3029,"children":3030},{},[3031,3033,3038,3040,3045],{"type":619,"value":3032},"研究首次揭示 Claude 訓練中",{"type":614,"tag":703,"props":3034,"children":3035},{},[3036],{"type":619,"value":3037},"自發形成",{"type":619,"value":3039},"的工作記憶區域 ",{"type":614,"tag":703,"props":3041,"children":3042},{},[3043],{"type":619,"value":3044},"J-Space",{"type":619,"value":3046},"，佔激活變異量約 10%，通常同時活躍約 25 個概念向量；其餘 90% 為模型本身無法存取的自動處理層。",{"type":614,"tag":696,"props":3048,"children":3049},{},[3050],{"type":614,"tag":615,"props":3051,"children":3052},{},[3053,3057,3060],{"type":614,"tag":703,"props":3054,"children":3055},{},[3056],{"type":619,"value":707},{"type":614,"tag":709,"props":3058,"children":3059},{},[],{"type":619,"value":3061},"\nJacobian（雅可比矩陣）衡量函數輸入微小變化如何影響輸出；J-Lens 用它量化激活向量對下一個 token 機率的影響，再對應回人類可讀的詞彙。",{"type":614,"tag":658,"props":3063,"children":3065},{"id":3064},"三層結構與安全應用",[3066],{"type":619,"value":3064},{"type":614,"tag":615,"props":3068,"children":3069},{},[3070],{"type":619,"value":3071},"J-Space 僅佔概念向量變異的 6–7%，卻驅動 88% 的成功輸出替換。層級分三區：早期層（感知）→ 中間層（概念工作區）→ 晚期層（輸出準備）。",{"type":614,"tag":615,"props":3073,"children":3074},{},[3075,3077,3082],{"type":619,"value":3076},"研究者在 Claude 輸出任何文字前，偵測到 J-Space 中出現「fake」「blackmail」等詞彙，顯示模型已辨識虛構勒索情境。利用此洞見設計的",{"type":614,"tag":703,"props":3078,"children":3079},{},[3080],{"type":619,"value":3081},"反事實反思訓練",{"type":619,"value":3083},"，使 Haiku 4.5 捏造回答率從 0.25 降至 0.07、欺騙嘗試率從 0.38 降至 0.05。",{"title":287,"searchDepth":621,"depth":621,"links":3085},[],{"data":3087,"body":3089,"excerpt":-1,"toc":3108},{"title":287,"description":3088},"J-Lens 為可解釋性工作帶來直接可操作的能力：不只是事後分析，而是可以因果干預模型行為。在 J-Space 中替換概念向量（「spider」→「ant」），模型回答腿數即從 8 變 6；消除前 10 個 J-Lens 方向後，多步推理準確率接近零，但 MMLU 等淺層任務幾乎不受影響。",{"type":611,"children":3090},[3091,3103],{"type":614,"tag":615,"props":3092,"children":3093},{},[3094,3096,3101],{"type":619,"value":3095},"J-Lens 為可解釋性工作帶來直接可操作的能力：不只是事後分析，而是可以",{"type":614,"tag":703,"props":3097,"children":3098},{},[3099],{"type":619,"value":3100},"因果干預",{"type":619,"value":3102},"模型行為。在 J-Space 中替換概念向量（「spider」→「ant」），模型回答腿數即從 8 變 6；消除前 10 個 J-Lens 方向後，多步推理準確率接近零，但 MMLU 等淺層任務幾乎不受影響。",{"type":614,"tag":615,"props":3104,"children":3105},{},[3106],{"type":619,"value":3107},"這為調試多步推理失敗提供了精確切入點——當模型「推錯了」，可直接進入中間層觀察哪個概念向量異常，而不只是分析最終輸出。",{"title":287,"searchDepth":621,"depth":621,"links":3109},[],{"data":3111,"body":3113,"excerpt":-1,"toc":3132},{"title":287,"description":3112},"J-Lens 的商業意義在於安全可稽核性：AI 系統在輸出前的內部意圖首次變得可觀測，為企業採購 AI 服務提供合規審查的技術根據。",{"type":611,"children":3114},[3115,3127],{"type":614,"tag":615,"props":3116,"children":3117},{},[3118,3120,3125],{"type":619,"value":3119},"J-Lens 的商業意義在於",{"type":614,"tag":703,"props":3121,"children":3122},{},[3123],{"type":619,"value":3124},"安全可稽核性",{"type":619,"value":3126},"：AI 系統在輸出前的內部意圖首次變得可觀測，為企業採購 AI 服務提供合規審查的技術根據。",{"type":614,"tag":615,"props":3128,"children":3129},{},[3130],{"type":619,"value":3131},"反事實反思訓練已在 Haiku 4.5 上落地（幻覺率下降 72%、欺騙嘗試率下降 87%），顯示此方法可直接整合進訓練流程。若此框架對外部研究開放，將加速整個產業的 AI 可信任度認證標準形成。",{"title":287,"searchDepth":621,"depth":621,"links":3133},[],{"data":3135,"body":3136,"excerpt":-1,"toc":3166},{"title":287,"description":287},{"type":611,"children":3137},[3138,3143],{"type":614,"tag":658,"props":3139,"children":3141},{"id":3140},"關鍵數據",[3142],{"type":619,"value":3140},{"type":614,"tag":965,"props":3144,"children":3145},{},[3146,3151,3156,3161],{"type":614,"tag":895,"props":3147,"children":3148},{},[3149],{"type":619,"value":3150},"J-Space 佔激活變異量：約 10%（約 25 個概念向量同時活躍）",{"type":614,"tag":895,"props":3152,"children":3153},{},[3154],{"type":619,"value":3155},"J-Space 成分佔向量變異：6–7%，驅動 88% 成功輸出替換（對照組非 J-Space 成分：5%）",{"type":614,"tag":895,"props":3157,"children":3158},{},[3159],{"type":619,"value":3160},"消除前 10 個 J-Lens 方向：多步推理準確率 → 接近 0%；MMLU 等淺層任務幾乎不受影響",{"type":614,"tag":895,"props":3162,"children":3163},{},[3164],{"type":619,"value":3165},"反事實反思訓練效果 (Claude Haiku 4.5) ：捏造回答率 0.25 → 0.07(-72%) 、欺騙嘗試率 0.38 → 0.05(-87%)",{"title":287,"searchDepth":621,"depth":621,"links":3167},[],{"data":3169,"body":3170,"excerpt":-1,"toc":3298},{"title":287,"description":287},{"type":611,"children":3171},[3172,3177,3204,3209,3245,3251,3278,3293],{"type":614,"tag":658,"props":3173,"children":3175},{"id":3174},"三步驟完成求職",[3176],{"type":619,"value":3174},{"type":614,"tag":615,"props":3178,"children":3179},{},[3180,3182,3188,3190,3195,3197,3202],{"type":619,"value":3181},"丹麥開發者 Mads Lorentzen 開源的 ",{"type":614,"tag":725,"props":3183,"children":3185},{"className":3184},[],[3186],{"type":619,"value":3187},"ai-job-search",{"type":619,"value":3189}," 框架，讓 Claude Code 接管整個求職流程。截至 2026-07-08 已累積 ",{"type":614,"tag":703,"props":3191,"children":3192},{},[3193],{"type":619,"value":3194},"10,925 顆星",{"type":619,"value":3196},"、",{"type":614,"tag":703,"props":3198,"children":3199},{},[3200],{"type":619,"value":3201},"3,671 個 Fork",{"type":619,"value":3203},"，是近期 GitHub 上成長最快的 Claude Code 相關社群專案之一。",{"type":614,"tag":615,"props":3205,"children":3206},{},[3207],{"type":619,"value":3208},"核心流程只有三個指令：",{"type":614,"tag":891,"props":3210,"children":3211},{},[3212,3223,3234],{"type":614,"tag":895,"props":3213,"children":3214},{},[3215,3221],{"type":614,"tag":725,"props":3216,"children":3218},{"className":3217},[],[3219],{"type":619,"value":3220},"/setup",{"type":619,"value":3222},"：建立個人候選人 Profile",{"type":614,"tag":895,"props":3224,"children":3225},{},[3226,3232],{"type":614,"tag":725,"props":3227,"children":3229},{"className":3228},[],[3230],{"type":619,"value":3231},"/scrape",{"type":619,"value":3233},"：搜尋多個職缺平台",{"type":614,"tag":895,"props":3235,"children":3236},{},[3237,3243],{"type":614,"tag":725,"props":3238,"children":3240},{"className":3239},[],[3241],{"type":619,"value":3242},"/apply \u003CURL>",{"type":619,"value":3244},"：一鍵完成申請全流程——從 CV 客製化、求職信撰寫到 ATS 相容性驗證",{"type":614,"tag":658,"props":3246,"children":3248},{"id":3247},"雙-agent-流水線設計",[3249],{"type":619,"value":3250},"雙 Agent 流水線設計",{"type":614,"tag":615,"props":3252,"children":3253},{},[3254,3260,3262,3268,3270,3276],{"type":614,"tag":725,"props":3255,"children":3257},{"className":3256},[],[3258],{"type":619,"value":3259},"/apply",{"type":619,"value":3261}," 執行「起草者—審閱者」雙 Agent 流水線：第一個 Agent 起草 LaTeX 格式的 CV 與求職信，第二個擁有全新 context 的 Agent 研究目標公司並提出批評，第一個 Agent 據此修訂，最後強制編譯 PDF 並視覺檢查排版（CV 用 ",{"type":614,"tag":725,"props":3263,"children":3265},{"className":3264},[],[3266],{"type":619,"value":3267},"lualatex",{"type":619,"value":3269},"，求職信用 ",{"type":614,"tag":725,"props":3271,"children":3273},{"className":3272},[],[3274],{"type":619,"value":3275},"xelatex",{"type":619,"value":3277},"）。",{"type":614,"tag":696,"props":3279,"children":3280},{},[3281],{"type":614,"tag":615,"props":3282,"children":3283},{},[3284,3288,3291],{"type":614,"tag":703,"props":3285,"children":3286},{},[3287],{"type":619,"value":707},{"type":614,"tag":709,"props":3289,"children":3290},{},[],{"type":619,"value":3292},"\nATS(Applicant Tracking System) ：企業用來自動篩選履歷的軟體，CV 若未通過 ATS 解析，面試官根本看不到。",{"type":614,"tag":615,"props":3294,"children":3295},{},[3296],{"type":619,"value":3297},"框架承諾「不捏造技能」，技能缺口誠實標示。CV 超過 2 頁時，依職缺相關性、文件唯一性、求職信依賴度三維度評分刪除最低分條目。",{"title":287,"searchDepth":621,"depth":621,"links":3299},[],{"data":3301,"body":3303,"excerpt":-1,"toc":3338},{"title":287,"description":3302},"此框架最值得研究的是雙 Agent 流水線的設計模式：以「獨立 context」隔離起草與批評兩個角色，有效避免單一 Agent 的自我確認偏誤。",{"type":611,"children":3304},[3305,3317],{"type":614,"tag":615,"props":3306,"children":3307},{},[3308,3310,3315],{"type":619,"value":3309},"此框架最值得研究的是",{"type":614,"tag":703,"props":3311,"children":3312},{},[3313],{"type":619,"value":3314},"雙 Agent 流水線的設計模式",{"type":619,"value":3316},"：以「獨立 context」隔離起草與批評兩個角色，有效避免單一 Agent 的自我確認偏誤。",{"type":614,"tag":615,"props":3318,"children":3319},{},[3320,3322,3328,3330,3336],{"type":619,"value":3321},"延伸指令同樣實用——",{"type":614,"tag":725,"props":3323,"children":3325},{"className":3324},[],[3326],{"type":619,"value":3327},"/expand",{"type":619,"value":3329}," 從 GitHub、Portfolio、Google Scholar 豐富技能標籤；",{"type":614,"tag":725,"props":3331,"children":3333},{"className":3332},[],[3334],{"type":619,"value":3335},"/add-portal",{"type":619,"value":3337}," 讓本地市場職缺平台爬蟲可插拔擴展。MIT 授權、TypeScript 實作，可直接作為 Claude Code 多代理工作流的學習範本。",{"title":287,"searchDepth":621,"depth":621,"links":3339},[],{"data":3341,"body":3343,"excerpt":-1,"toc":3377},{"title":287,"description":3342},"超過 1 萬顆星的爆紅速度，反映白領知識工作者對「AI 代理求職」的強烈需求。這個趨勢對人才市場有兩面影響：",{"type":611,"children":3344},[3345,3349,3372],{"type":614,"tag":615,"props":3346,"children":3347},{},[3348],{"type":619,"value":3342},{"type":614,"tag":965,"props":3350,"children":3351},{},[3352,3362],{"type":614,"tag":895,"props":3353,"children":3354},{},[3355,3360],{"type":614,"tag":703,"props":3356,"children":3357},{},[3358],{"type":619,"value":3359},"求職端",{"type":619,"value":3361},"：個人可用 AI 放大申請量，CV 客製化品質提升",{"type":614,"tag":895,"props":3363,"children":3364},{},[3365,3370],{"type":614,"tag":703,"props":3366,"children":3367},{},[3368],{"type":619,"value":3369},"招募端",{"type":619,"value":3371},"：ATS 篩選的意義將被稀釋，人工審閱壓力倍增",{"type":614,"tag":615,"props":3373,"children":3374},{},[3375],{"type":619,"value":3376},"企業招募流程若未跟上這波工具普及化，將面臨更高的假陽性率——AI 最佳化的 CV 未必代表真實能力匹配。",{"title":287,"searchDepth":621,"depth":621,"links":3378},[],{"data":3380,"body":3381,"excerpt":-1,"toc":3414},{"title":287,"description":287},{"type":611,"children":3382},[3383,3389,3394,3399,3404,3409],{"type":614,"tag":658,"props":3384,"children":3386},{"id":3385},"跨裝置-ai-工作流正式落地",[3387],{"type":619,"value":3388},"跨裝置 AI 工作流正式落地",{"type":614,"tag":615,"props":3390,"children":3391},{},[3392],{"type":619,"value":3393},"Anthropic 於 2026 年 7 月 7 日宣布 Claude Cowork 擴展至行動裝置與網頁版，從桌面專屬工具升級為全平台 AI 工作環境。Beta 存取從 Max 訂閱者開始逐步開放，雙倍使用額度延伸優惠適用至 2026 年 8 月 5 日。",{"type":614,"tag":615,"props":3395,"children":3396},{},[3397],{"type":619,"value":3398},"核心體驗是「非同步接力」：在桌機啟動任務，手機接收進度通知，任何裝置取得結果——即使筆電已關閉，背景處理仍持續運行。「Human in the loop」機制已支援行動端，Claude 輸出結果前會先請使用者確認。",{"type":614,"tag":658,"props":3400,"children":3402},{"id":3401},"功能邊界與定位轉移",[3403],{"type":619,"value":3401},{"type":614,"tag":615,"props":3405,"children":3406},{},[3407],{"type":619,"value":3408},"本機功能（資料夾讀寫、瀏覽器控制、Computer Use）仍需桌面應用，網頁版功能有所限制。Chat 與 Cowork 整合為統一主畫面，Projects 和 Artifacts 跨平台同步。",{"type":614,"tag":615,"props":3410,"children":3411},{},[3412],{"type":619,"value":3413},"Anthropic 分析 120 萬匿名 session 發現：逾 90% 使用情境為非程式碼工作，業務流程作業佔 33.4%，軟體開發僅佔 8.7%。這份數據直接定義了 Cowork 的擴展方向：打造「agentic 行政助理」，而非開發者工具。",{"title":287,"searchDepth":621,"depth":621,"links":3415},[],{"data":3417,"body":3418,"excerpt":-1,"toc":3424},{"title":287,"description":420},{"type":611,"children":3419},[3420],{"type":614,"tag":615,"props":3421,"children":3422},{},[3423],{"type":619,"value":420},{"title":287,"searchDepth":621,"depth":621,"links":3425},[],{"data":3427,"body":3428,"excerpt":-1,"toc":3434},{"title":287,"description":421},{"type":611,"children":3429},[3430],{"type":614,"tag":615,"props":3431,"children":3432},{},[3433],{"type":619,"value":421},{"title":287,"searchDepth":621,"depth":621,"links":3435},[],{"data":3437,"body":3438,"excerpt":-1,"toc":3489},{"title":287,"description":287},{"type":611,"children":3439},[3440,3446,3451,3456,3461,3466,3484],{"type":614,"tag":658,"props":3441,"children":3443},{"id":3442},"從-11-到-46中國模型的爆發式增長",[3444],{"type":619,"value":3445},"從 11% 到 46%：中國模型的爆發式增長",{"type":614,"tag":615,"props":3447,"children":3448},{},[3449],{"type":619,"value":3450},"自 2026 年 2 月 8 日起，中國 AI 模型每週穩定佔 OpenRouter 流量 30% 以上，峰值一度達 46%（2026 年 7 月 7 日）。",{"type":614,"tag":615,"props":3452,"children":3453},{},[3454],{"type":619,"value":3455},"美國模型份額已從一年前約 70% 崩跌至約 30%；過去 12 個月平均僅 11%，2025 上半年更只有 4.5%。",{"type":614,"tag":658,"props":3457,"children":3459},{"id":3458},"成本差距才是核心驅動力",[3460],{"type":619,"value":3458},{"type":614,"tag":615,"props":3462,"children":3463},{},[3464],{"type":619,"value":3465},"同等工作負載定價對比：",{"type":614,"tag":965,"props":3467,"children":3468},{},[3469,3474,3479],{"type":614,"tag":895,"props":3470,"children":3471},{},[3472],{"type":619,"value":3473},"Anthropic Claude：約 $4,811",{"type":614,"tag":895,"props":3475,"children":3476},{},[3477],{"type":619,"value":3478},"OpenAI ChatGPT：約 $3,357",{"type":614,"tag":895,"props":3480,"children":3481},{},[3482],{"type":619,"value":3483},"智譜 GLM：約 $544",{"type":614,"tag":615,"props":3485,"children":3486},{},[3487],{"type":619,"value":3488},"AI 新創 Lindy 已將 100% 流量從 Anthropic Claude 切換至 DeepSeek，CEO Flo Crivello 稱此舉「省下數百萬美元，且核心用例的性能實際上有所提升」。",{"title":287,"searchDepth":621,"depth":621,"links":3490},[],{"data":3492,"body":3494,"excerpt":-1,"toc":3505},{"title":287,"description":3493},"非核心任務已可安全導向中國模型：GLM、Qwen、DeepSeek 定價僅為美國頂尖模型的 10–40%，在文件摘要、分類、翻譯等場景表現足以勝任。",{"type":611,"children":3495},[3496,3500],{"type":614,"tag":615,"props":3497,"children":3498},{},[3499],{"type":619,"value":3493},{"type":614,"tag":615,"props":3501,"children":3502},{},[3503],{"type":619,"value":3504},"建議以 OpenRouter 做路由層，依任務複雜度動態選模型；敏感資料仍走本地或閉源路徑。能力落後 6–9 個月的差距，對非前沿任務幾乎不構成影響。",{"title":287,"searchDepth":621,"depth":621,"links":3506},[],{"data":3508,"body":3510,"excerpt":-1,"toc":3521},{"title":287,"description":3509},"中國大廠（DeepSeek、智譜、Alibaba Qwen）以低價換市佔，不依賴 API 收入盈利——這場結構性價格戰對美國大廠的高毛利 API 商業模式形成直接威脅。",{"type":611,"children":3511},[3512,3516],{"type":614,"tag":615,"props":3513,"children":3514},{},[3515],{"type":619,"value":3509},{"type":614,"tag":615,"props":3517,"children":3518},{},[3519],{"type":619,"value":3520},"Lindy 案例顯示企業「先切換、再觀察」的決策週期正在壓縮；若美國大廠無法縮小定價差距，非前沿任務市場將持續向中國模型流失。",{"title":287,"searchDepth":621,"depth":621,"links":3522},[],{"data":3524,"body":3525,"excerpt":-1,"toc":3577},{"title":287,"description":287},{"type":611,"children":3526},[3527,3533,3548,3554],{"type":614,"tag":658,"props":3528,"children":3530},{"id":3529},"定價基準同等工作負載",[3531],{"type":619,"value":3532},"定價基準（同等工作負載）",{"type":614,"tag":965,"props":3534,"children":3535},{},[3536,3540,3544],{"type":614,"tag":895,"props":3537,"children":3538},{},[3539],{"type":619,"value":3473},{"type":614,"tag":895,"props":3541,"children":3542},{},[3543],{"type":619,"value":3478},{"type":614,"tag":895,"props":3545,"children":3546},{},[3547],{"type":619,"value":3483},{"type":614,"tag":658,"props":3549,"children":3551},{"id":3550},"市場份額變化-openrouter",[3552],{"type":619,"value":3553},"市場份額變化 (OpenRouter)",{"type":614,"tag":965,"props":3555,"children":3556},{},[3557,3562,3567,3572],{"type":614,"tag":895,"props":3558,"children":3559},{},[3560],{"type":619,"value":3561},"中國模型 2025H1 均值：4.5%",{"type":614,"tag":895,"props":3563,"children":3564},{},[3565],{"type":619,"value":3566},"中國模型過去 12 個月均值：11%",{"type":614,"tag":895,"props":3568,"children":3569},{},[3570],{"type":619,"value":3571},"2026-02-08 後週均：30%+",{"type":614,"tag":895,"props":3573,"children":3574},{},[3575],{"type":619,"value":3576},"峰值 (2026-07-07) ：46%",{"title":287,"searchDepth":621,"depth":621,"links":3578},[],{"data":3580,"body":3581,"excerpt":-1,"toc":3620},{"title":287,"description":287},{"type":611,"children":3582},[3583,3589,3594,3609,3615],{"type":614,"tag":658,"props":3584,"children":3586},{"id":3585},"ai-三階段同儕評審流程",[3587],{"type":619,"value":3588},"AI 三階段同儕評審流程",{"type":614,"tag":615,"props":3590,"children":3591},{},[3592],{"type":619,"value":3593},"Badge 是一個 AI Agent 驅動的職業聲譽平台，核心主張是傳統履歷的關鍵字堆砌與 LinkedIn 推薦信缺乏可信度。AI Agent 分三個階段運作：首先讀取用戶通訊錄，識別曾共事的同事；接著透過 WhatsApp 發送結構化匿名提問；數小時內自動彙整回覆並填入個人頁面，全程無需手動追蹤。",{"type":614,"tag":696,"props":3595,"children":3596},{},[3597],{"type":614,"tag":615,"props":3598,"children":3599},{},[3600,3604,3607],{"type":614,"tag":703,"props":3601,"children":3602},{},[3603],{"type":619,"value":707},{"type":614,"tag":709,"props":3605,"children":3606},{},[],{"type":619,"value":3608},"\nProof of Work（工作成果證明）：借鑒自區塊鏈術語，在職涯語境中指有同儕驗證的真實工作紀錄，與泛泛的履歷自我描述相對。",{"type":614,"tag":658,"props":3610,"children":3612},{"id":3611},"信任機制與-trust-score",[3613],{"type":619,"value":3614},"信任機制與 Trust Score",{"type":614,"tag":615,"props":3616,"children":3617},{},[3618],{"type":619,"value":3619},"系統在接受評審前強制驗證「工作重疊」（相同公司、相同時間段），防止虛假評論。AI 以結構化問題引導——涵蓋溝通風格、協作能力、執行力、優勢與改進空間——並演算出 Trust Score；惡意或辱罵評論會被自動過濾。招募方可在 30 秒內取得候選人的匿名同事反饋，Swiggy、Atlassian 等企業招募人員均給予正面評價。",{"title":287,"searchDepth":621,"depth":621,"links":3621},[],{"data":3623,"body":3624,"excerpt":-1,"toc":3630},{"title":287,"description":489},{"type":611,"children":3625},[3626],{"type":614,"tag":615,"props":3627,"children":3628},{},[3629],{"type":619,"value":489},{"title":287,"searchDepth":621,"depth":621,"links":3631},[],{"data":3633,"body":3634,"excerpt":-1,"toc":3640},{"title":287,"description":490},{"type":611,"children":3635},[3636],{"type":614,"tag":615,"props":3637,"children":3638},{},[3639],{"type":619,"value":490},{"title":287,"searchDepth":621,"depth":621,"links":3641},[],{"data":3643,"body":3644,"excerpt":-1,"toc":3735},{"title":287,"description":287},{"type":611,"children":3645},[3646,3652,3671,3697,3703,3715,3730],{"type":614,"tag":658,"props":3647,"children":3649},{"id":3648},"空間原生視覺看准才能行動",[3650],{"type":619,"value":3651},"空間原生視覺：看准才能行動",{"type":614,"tag":615,"props":3653,"children":3654},{},[3655,3657,3662,3664,3669],{"type":619,"value":3656},"蚂蚁旗下蚂蚁灵波科技 (Robbyant) 於 2026 年 7 月 7 日正式開源 ",{"type":614,"tag":703,"props":3658,"children":3659},{},[3660],{"type":619,"value":3661},"LingBot-Vision",{"type":619,"value":3663},"，業界首個「空間原生」具身視覺基礎模型，釋出 ViT-G/L/B/S 四個規格，旗艦版約 ",{"type":614,"tag":703,"props":3665,"children":3666},{},[3667],{"type":619,"value":3668},"11 億參數",{"type":619,"value":3670},"。",{"type":614,"tag":615,"props":3672,"children":3673},{},[3674,3676,3681,3683,3688,3690,3695],{"type":619,"value":3675},"同步發布 ",{"type":614,"tag":703,"props":3677,"children":3678},{},[3679],{"type":619,"value":3680},"LingBot-Depth 2.0",{"type":619,"value":3682},"，訓練資料從 300 萬張擴展至 ",{"type":614,"tag":703,"props":3684,"children":3685},{},[3686],{"type":619,"value":3687},"1.5 億張",{"type":619,"value":3689},"，在 16 項深度補全 benchmark 中拿下 12 項第一，大型室內場景深度誤差 (RMSE) 從 0.132 降至 ",{"type":614,"tag":703,"props":3691,"children":3692},{},[3693],{"type":619,"value":3694},"0.062",{"type":619,"value":3696},"，降幅達 50%。",{"type":614,"tag":658,"props":3698,"children":3700},{"id":3699},"核心創新遮罩邊界建模",[3701],{"type":619,"value":3702},"核心創新：遮罩邊界建模",{"type":614,"tag":615,"props":3704,"children":3705},{},[3706,3708,3713],{"type":619,"value":3707},"模型核心為 ",{"type":614,"tag":703,"props":3709,"children":3710},{},[3711],{"type":619,"value":3712},"Masked Boundary Modeling(MBM)",{"type":619,"value":3714},"，一種自監督預訓練範式，動態學習亞像素級邊界表示，讓模型聚焦幾何結構而非語義。",{"type":614,"tag":696,"props":3716,"children":3717},{},[3718],{"type":614,"tag":615,"props":3719,"children":3720},{},[3721,3725,3728],{"type":614,"tag":703,"props":3722,"children":3723},{},[3724],{"type":619,"value":707},{"type":614,"tag":709,"props":3726,"children":3727},{},[],{"type":619,"value":3729},"\nMBM（遮罩邊界建模）：以動態學習到的亞像素級邊界特徵作為遮罩目標，讓 AI「學會看輪廓」後再重建完整場景幾何，有別於傳統語義分割只辨識物體類別。",{"type":614,"tag":615,"props":3731,"children":3732},{},[3733],{"type":619,"value":3734},"ViT-L（3 億參數）在 NYU-Depth v2 基準上匹敵 DINOv3（70 億參數），以 1/23 的參數量達到同等效果，訓練樣本不足對手 1/3。對玻璃、鏡面、透明材質及弱光環境均有顯著改善。",{"title":287,"searchDepth":621,"depth":621,"links":3736},[],{"data":3738,"body":3740,"excerpt":-1,"toc":3751},{"title":287,"description":3739},"LingBot-Vision 屬具身 AI 全棧的視覺底座，上接 VLA、世界模型、流式 3D 重建等模組，開源版本支援深度估計、小目標偵測、遮擋場景理解等下游微調任務。",{"type":611,"children":3741},[3742,3746],{"type":614,"tag":615,"props":3743,"children":3744},{},[3745],{"type":619,"value":3739},{"type":614,"tag":615,"props":3747,"children":3748},{},[3749],{"type":619,"value":3750},"ViT-L 以 1/23 的參數量匹敵 DINOv3 70B，可大幅降低邊緣部署推論成本——若現有視覺 pipeline 需要精確幾何感知，ViT-L 值得優先評估作為換裝基座。",{"title":287,"searchDepth":621,"depth":621,"links":3752},[],{"data":3754,"body":3756,"excerpt":-1,"toc":3767},{"title":287,"description":3755},"蚂蚁以開源策略搶佔具身 AI 視覺標準制高點，並計劃年底推出軟硬一體相機產品，形成模型至硬體的完整生態護城河。",{"type":611,"children":3757},[3758,3762],{"type":614,"tag":615,"props":3759,"children":3760},{},[3761],{"type":619,"value":3755},{"type":614,"tag":615,"props":3763,"children":3764},{},[3765],{"type":619,"value":3766},"1.5 億張訓練資料規模顯示蚂蚁的資料壁壘——技術開源不代表商業護城河消失，後續硬體整合與平台生態才是真正的變現重點，對機器人或工業視覺企業而言採用門檻已大幅降低。",{"title":287,"searchDepth":621,"depth":621,"links":3768},[],{"data":3770,"body":3771,"excerpt":-1,"toc":3801},{"title":287,"description":287},{"type":611,"children":3772},[3773,3778],{"type":614,"tag":658,"props":3774,"children":3776},{"id":3775},"性能基準",[3777],{"type":619,"value":3775},{"type":614,"tag":965,"props":3779,"children":3780},{},[3781,3786,3791,3796],{"type":614,"tag":895,"props":3782,"children":3783},{},[3784],{"type":619,"value":3785},"NYU-Depth v2(RMSE) ：ViT-L 3 億參數與 DINOv3 70 億參數同級",{"type":614,"tag":895,"props":3787,"children":3788},{},[3789],{"type":619,"value":3790},"大型室內場景 RMSE：0.062（v1.0 為 0.132，降幅 50%）",{"type":614,"tag":895,"props":3792,"children":3793},{},[3794],{"type":619,"value":3795},"深度補全 benchmark：16 項中拿下 12 項第一",{"type":614,"tag":895,"props":3797,"children":3798},{},[3799],{"type":619,"value":3800},"參數效率：ViT-L 以 1/23 規模達 DINOv3 同等效果，訓練樣本不足對手 1/3",{"title":287,"searchDepth":621,"depth":621,"links":3802},[],{"data":3804,"body":3805,"excerpt":-1,"toc":3856},{"title":287,"description":287},{"type":611,"children":3806},[3807,3812,3838,3844],{"type":614,"tag":658,"props":3808,"children":3810},{"id":3809},"補貼金額刷新紀錄",[3811],{"type":619,"value":3809},{"type":614,"tag":615,"props":3813,"children":3814},{},[3815,3817,3822,3824,3829,3831,3836],{"type":619,"value":3816},"Anthropic 針對 Y Combinator 新創的標準報價從 3 萬美元躍升至 ",{"type":614,"tag":703,"props":3818,"children":3819},{},[3820],{"type":619,"value":3821},"50 萬美元",{"type":619,"value":3823},"（免股權），OpenAI 跟進同等金額，並開放「",{"type":614,"tag":703,"props":3825,"children":3826},{},[3827],{"type":619,"value":3828},"150 萬美元",{"type":619,"value":3830},"換股份選擇權」選項。AI 語音新創 Dialogus 創辦人 Hans Ibarra 透露，光他一家就累積收到超過 ",{"type":614,"tag":703,"props":3832,"children":3833},{},[3834],{"type":619,"value":3835},"300 萬美元",{"type":619,"value":3837},"競爭性報價，補貼力道之猛可見一斑。",{"type":614,"tag":658,"props":3839,"children":3841},{"id":3840},"背後邏輯以補貼鎖定生態圈",[3842],{"type":619,"value":3843},"背後邏輯：以補貼鎖定生態圈",{"type":614,"tag":615,"props":3845,"children":3846},{},[3847,3849,3854],{"type":619,"value":3848},"Google Cloud 同步加入戰局，提供最高 50 萬美元額度，附加 DeepMind 工程師顧問接觸機會。各家均採「生態圈鎖定」策略——讓新創在早期即基於自家 API 與工具開發，降低日後遷移意願。若以 YC 每季約 200 家新創估算，兩家合計每年可發出高達 ",{"type":614,"tag":703,"props":3850,"children":3851},{},[3852],{"type":619,"value":3853},"8 億美元",{"type":619,"value":3855},"算力額度，補貼已從行銷手段升格為核心客戶獲取策略。",{"title":287,"searchDepth":621,"depth":621,"links":3857},[],{"data":3859,"body":3861,"excerpt":-1,"toc":3890},{"title":287,"description":3860},"早期選定 API 平台不只看模型效能，還要考量遷移成本。50 萬美元額度用完後，若基礎設施已深度依賴某家 SDK 與工具鏈，轉換代價極高。評估補貼時建議同步確認：",{"type":611,"children":3862},[3863,3867,3885],{"type":614,"tag":615,"props":3864,"children":3865},{},[3866],{"type":619,"value":3860},{"type":614,"tag":965,"props":3868,"children":3869},{},[3870,3875,3880],{"type":614,"tag":895,"props":3871,"children":3872},{},[3873],{"type":619,"value":3874},"API 相容性與標準化程度",{"type":614,"tag":895,"props":3876,"children":3877},{},[3878],{"type":619,"value":3879},"SDK 鎖定深度",{"type":614,"tag":895,"props":3881,"children":3882},{},[3883],{"type":619,"value":3884},"補貼期結束後的 token 定價結構",{"type":614,"tag":615,"props":3886,"children":3887},{},[3888],{"type":619,"value":3889},"避免補貼期結束後陷入議價弱勢。",{"title":287,"searchDepth":621,"depth":621,"links":3891},[],{"data":3893,"body":3895,"excerpt":-1,"toc":3906},{"title":287,"description":3894},"OpenAI 與 Anthropic 在 IPO 前同步燒錢搶客戶，反映兩家公司均已將「企業端長期收入」視為核心護城河。YC 生態圈是 B 輪以後融資的溫床，早期綁定這群新創，等同搶先佈局未來潛力客戶。",{"type":611,"children":3896},[3897,3901],{"type":614,"tag":615,"props":3898,"children":3899},{},[3900],{"type":619,"value":3894},{"type":614,"tag":615,"props":3902,"children":3903},{},[3904],{"type":619,"value":3905},"投資人應留意：補貼金額持續攀升可能壓縮短期毛利，但長期有利鞏固市場份額——這是一場以現金換市場地位的豪賭。",{"title":287,"searchDepth":621,"depth":621,"links":3907},[],{"data":3909,"body":3910,"excerpt":-1,"toc":3978},{"title":287,"description":287},{"type":611,"children":3911},[3912,3918,3923,3938,3943,3948,3973],{"type":614,"tag":658,"props":3913,"children":3915},{"id":3914},"強制攝影機上路gsr2-正式生效",[3916],{"type":619,"value":3917},"強制攝影機上路：GSR2 正式生效",{"type":614,"tag":615,"props":3919,"children":3920},{},[3921],{"type":619,"value":3922},"2026 年 7 月 7 日起，歐盟 GSR2（一般安全法規第二版）正式生效，所有在歐盟境內新登記的乘用車與廂型車均須強制搭載 ADDW 系統。ADDW 透過紅外線攝影機追蹤駕駛的眼球視線方向與頭部位置，法規明確禁止使用生物特徵辨識（包括臉部識別），目標是偵測分心駕駛行為。",{"type":614,"tag":696,"props":3924,"children":3925},{},[3926],{"type":614,"tag":615,"props":3927,"children":3928},{},[3929,3933,3936],{"type":614,"tag":703,"props":3930,"children":3931},{},[3932],{"type":619,"value":707},{"type":614,"tag":709,"props":3934,"children":3935},{},[],{"type":619,"value":3937},"\nADDW（Advanced Driver Distraction Warning，先進駕駛分心警告）：車內紅外線攝影系統，透過追蹤眼球視線與頭部姿態偵測駕駛分心，不做身份識別，僅判斷「眼睛看哪裡」。",{"type":614,"tag":658,"props":3939,"children":3941},{"id":3940},"觸發門檻與已知問題",[3942],{"type":619,"value":3940},{"type":614,"tag":615,"props":3944,"children":3945},{},[3946],{"type":619,"value":3947},"攝影機於車速超過 20 km/h 時自動啟動，觸發規則如下：",{"type":614,"tag":965,"props":3949,"children":3950},{},[3951,3963],{"type":614,"tag":895,"props":3952,"children":3953},{},[3954,3956,3961],{"type":619,"value":3955},"時速超過 50 km/h：視線離開路面超過 ",{"type":614,"tag":703,"props":3957,"children":3958},{},[3959],{"type":619,"value":3960},"3.5 秒",{"type":619,"value":3962},"即發警告",{"type":614,"tag":895,"props":3964,"children":3965},{},[3966,3968],{"type":619,"value":3967},"時速 20–50 km/h：門檻延長至 ",{"type":614,"tag":703,"props":3969,"children":3970},{},[3971],{"type":619,"value":3972},"6 秒",{"type":614,"tag":615,"props":3974,"children":3975},{},[3976],{"type":619,"value":3977},"早期實路測試已出現「警報疲勞」 (alarm fatigue) 問題——系統在彎道等正常轉移視線場景頻繁誤觸，駕駛反而開始習慣性忽視警報，安全效益大打折扣。",{"title":287,"searchDepth":621,"depth":621,"links":3979},[],{"data":3981,"body":3983,"excerpt":-1,"toc":3994},{"title":287,"description":3982},"系統校準策略因廠商差異懸殊，成為實作核心挑戰：彎道提前轉移視線、瞥儀表板、盲點確認等正常駕駛行為均可能觸發誤報。ADAS 系統之間亦存在潛在衝突，如車道輔助強行轉向與 ADDW 同時介入，顯示不同輔助功能缺乏統一協調機制。",{"type":611,"children":3984},[3985,3989],{"type":614,"tag":615,"props":3986,"children":3987},{},[3988],{"type":619,"value":3982},{"type":614,"tag":615,"props":3990,"children":3991},{},[3992],{"type":619,"value":3993},"GDPR 雖自動適用於車內攝影資料，但何謂「必要收集」目前無明確定義，亦無獨立稽核機制。嵌入式設計必須確保資料處理邏輯在車機本地完成，避免雲端傳輸引發合規風險。",{"title":287,"searchDepth":621,"depth":621,"links":3995},[],{"data":3997,"body":3999,"excerpt":-1,"toc":4010},{"title":287,"description":3998},"GSR2 同期要求酒精鎖接口準備與行車事件記錄器，整體合規成本上升，每年影響歐洲約 1,800 萬輛新車。歷史前例已顯示汽車製造商曾在未充分揭露的情況下，將駕駛資料分享給保險公司與數據仲介商——若 ADDW 車內影像外洩，可直接用於身份竊取，品牌聲譽損失難以估量。",{"type":611,"children":4000},[4001,4005],{"type":614,"tag":615,"props":4002,"children":4003},{},[4004],{"type":619,"value":3998},{"type":614,"tag":615,"props":4006,"children":4007},{},[4008],{"type":619,"value":4009},"監管空白（「必要資料」無明確定義）意味企業需主動建立內部規範，不能等監管機構釐清邊界後再行動。",{"title":287,"searchDepth":621,"depth":621,"links":4011},[],{"data":4013,"body":4014,"excerpt":-1,"toc":4114},{"title":287,"description":287},{"type":611,"children":4015},[4016,4021,4026,4054,4059,4064,4069,4074,4079,4084,4089,4094,4099,4104,4109],{"type":614,"tag":658,"props":4017,"children":4019},{"id":4018},"社群熱議排行",[4020],{"type":619,"value":4018},{"type":614,"tag":615,"props":4022,"children":4023},{},[4024],{"type":619,"value":4025},"今日高熱度話題（依互動量排序）：",{"type":614,"tag":965,"props":4027,"children":4028},{},[4029,4034,4039,4044,4049],{"type":614,"tag":895,"props":4030,"children":4031},{},[4032],{"type":619,"value":4033},"微軟以 MAI 自研模型替換 Copilot 內的 OpenAI 與 Anthropic（多平台熱議）",{"type":614,"tag":895,"props":4035,"children":4036},{},[4037],{"type":619,"value":4038},"中國 AI 模型在 OpenRouter 市佔突破 30%，token 份額不到一年從 1% 飆升至逾 50%(HN + X + Bluesky)",{"type":614,"tag":895,"props":4040,"children":4041},{},[4042],{"type":619,"value":4043},"DeepSeek 自研 AI 推論晶片計畫曝光（路透社獨家）",{"type":614,"tag":895,"props":4045,"children":4046},{},[4047],{"type":619,"value":4048},"Anthropic J-lens 論文首次讀取 LLM 內部工作空間（HN + Bluesky 廣泛討論）",{"type":614,"tag":895,"props":4050,"children":4051},{},[4052],{"type":619,"value":4053},"Claude Cowork 擴展至手機與網頁（TechCrunch、The Verge、9to5mac 同步報導）",{"type":614,"tag":615,"props":4055,"children":4056},{},[4057],{"type":619,"value":4058},"HN 社群對微軟削減外部 API 成本的態度普遍為「早知如此」，認為企業垂直整合是必然趨勢。",{"type":614,"tag":658,"props":4060,"children":4062},{"id":4061},"技術爭議與分歧",[4063],{"type":619,"value":4061},{"type":614,"tag":615,"props":4065,"children":4066},{},[4067],{"type":619,"value":4068},"OpenRouter 中美模型之爭，社群分歧鮮明。支持中國模型陣營引用成本數據：Eric Budd（Bluesky，3 upvotes）直言「美國模型程式碼確實更強，但價格太高了。」",{"type":614,"tag":615,"props":4070,"children":4071},{},[4072],{"type":619,"value":4073},"@nicrypto（X 用戶）提出更根本的威脅論：「足夠好正是每一項主導技術遭顛覆的方式」，暗示美國 AI 廠商正重蹈硬體業被低成本競爭者瓦解的覆轍。",{"type":614,"tag":615,"props":4075,"children":4076},{},[4077],{"type":619,"value":4078},"HN 用戶 reinitctxoffset 從地緣政治角度分析：「中國把 LLM 訓練和推論成本壓到零，因為他們根本不在乎財務回報。」反方認為中國模型在安全敏感場景有合規疑慮，此爭論目前未有定論。",{"type":614,"tag":658,"props":4080,"children":4082},{"id":4081},"實戰經驗",[4083],{"type":619,"value":4081},{"type":614,"tag":615,"props":4085,"children":4086},{},[4087],{"type":619,"value":4088},"Jesse Felder（Bluesky，8 upvotes）引述最直接的流量數據：「美國企業透過 OpenRouter 使用中國 AI 的 token 份額攀升至最高 46%，過去 12 個月平均僅 11%」，為真實生產流量的量化佐證。",{"type":614,"tag":615,"props":4090,"children":4091},{},[4092],{"type":619,"value":4093},"@kentimsit(X) 實測補貼落差：「OpenAI 和 Anthropic 的 Pro/Max 訂閱補貼幅度超過 80%，改用 API 計費就能明顯感受到差距。」",{"type":614,"tag":615,"props":4095,"children":4096},{},[4097],{"type":619,"value":4098},"weitendorf(HN) 分享 Claude Code 自動化心得：「能建立自我評估工具，是使用 LLM 寫程式碼，與實現真正 10 倍生產力之間最大的差異化因素。」",{"type":614,"tag":658,"props":4100,"children":4102},{"id":4101},"未解問題與社群預期",[4103],{"type":619,"value":4101},{"type":614,"tag":615,"props":4105,"children":4106},{},[4107],{"type":619,"value":4108},"J-lens 的假陽性改善進度與跨架構驗證結果，是社群認為決定此技術何時可升級至生產環境的核心問題。orbital-decay(HN) 點出深層困難：「模型在 CoT 中學到了自己的語言，我們根本無法確定它以你的方式理解你的文字。」",{"type":614,"tag":615,"props":4110,"children":4111},{},[4112],{"type":619,"value":4113},"DeepSeek 自研晶片時程與效能仍屬未知。jmyeet(HN) 預言：「OpenAI 和 Anthropic 將像恐龍一樣，不是一顆小行星，而是三顆。」社群最普遍的預期是，中國將確保 AI 前沿模型不存在護城河，這是其國家安全利益使然。",{"title":287,"searchDepth":621,"depth":621,"links":4115},[],{"data":4117,"body":4119,"excerpt":-1,"toc":4130},{"title":287,"description":4118},"今日報告勾勒出一個清晰的轉折訊號：AI 競爭的主戰場正從「誰的模型最強」移向「誰掌控整條供應鏈」。微軟換掉 OpenAI、DeepSeek 自研晶片、中國模型在 OpenRouter 拿下近半市佔——三件事共同指向同一方向：垂直整合、成本主導、生態鎖定。",{"type":611,"children":4120},[4121,4125],{"type":614,"tag":615,"props":4122,"children":4123},{},[4124],{"type":619,"value":4118},{"type":614,"tag":615,"props":4126,"children":4127},{},[4128],{"type":619,"value":4129},"對開發者而言，今日最有價值的訊號或許是 weitendorf 的心得：LLM 真正的槓桿不在押注最強模型，而在能快速建立自我評估工具的能力。在模型廝殺激烈、選擇充裕的時代，彈性架構比單一押注更重要。",{"title":287,"searchDepth":621,"depth":621,"links":4131},[],{"data":4133,"body":4134,"excerpt":-1,"toc":5135},{"title":287,"description":287},{"type":611,"children":4135},[4136,4140,4145,4151,5076,5080,5085,5089,5107,5111,5129],{"type":614,"tag":658,"props":4137,"children":4138},{"id":871},[4139],{"type":619,"value":871},{"type":614,"tag":615,"props":4141,"children":4142},{},[4143],{"type":619,"value":4144},"需要存取模型中間層隱藏狀態並能計算梯度流 (gradient flow) 。J-lens 目前為 Anthropic 內部研究工具，尚無公開 pip 套件；開源模型（如 Qwen 3.6 27B）可透過 HuggingFace Transformers 取得中間層激活值。建議環境：Python 3.10+、PyTorch 2.0+、VRAM ≥ 24GB（用於 7B 以上模型），可配合 TransformerLens 框架實作類似功能。",{"type":614,"tag":658,"props":4146,"children":4148},{"id":4147},"最小-poc",[4149],{"type":619,"value":4150},"最小 PoC",{"type":614,"tag":4152,"props":4153,"children":4157},"pre",{"className":4154,"code":4155,"language":4156,"meta":287,"style":287},"language-python shiki shiki-themes vitesse-dark","import torch\nfrom transformers import AutoModelForCausalLM, AutoTokenizer\n\nmodel_id = \"Qwen/Qwen2.5-7B-Instruct\"\ntokenizer = AutoTokenizer.from_pretrained(model_id)\nmodel = AutoModelForCausalLM.from_pretrained(\n    model_id, output_hidden_states=True, torch_dtype=torch.float16\n)\n\nprompt = \"The capital of France is\"\ninputs = tokenizer(prompt, return_tensors=\"pt\")\n\n# 取中間層隱藏狀態並計算對最終 logit 的梯度（J-lens 近似）\nwith torch.enable_grad():\n    outputs = model(**inputs, output_hidden_states=True)\n    # 選擇中間層（約佔總層數 40%-90%）\n    n_layers = len(outputs.hidden_states)\n    mid_layer = int(n_layers * 0.6)\n    hidden = outputs.hidden_states[mid_layer].requires_grad_(True)\n    logit_sum = outputs.logits[0, -1, :].sum()\n    grad = torch.autograd.grad(logit_sum, hidden)[0]\n    # 投影回 token 空間以獲得可讀表徵\n    readable_logits = model.lm_head(grad[0, -1:, :])\n    top_tokens = tokenizer.decode(readable_logits.topk(5).indices[0])\n    print(f\"J-space 讀取（第 {mid_layer} 層）：{top_tokens}\")\n","python",[4158],{"type":614,"tag":725,"props":4159,"children":4160},{"__ignoreMap":287},[4161,4179,4212,4222,4252,4294,4324,4379,4387,4395,4421,4479,4487,4497,4525,4578,4587,4628,4670,4728,4796,4863,4872,4936,5012],{"type":614,"tag":4162,"props":4163,"children":4166},"span",{"class":4164,"line":4165},"line",1,[4167,4173],{"type":614,"tag":4162,"props":4168,"children":4170},{"style":4169},"--shiki-default:#4D9375",[4171],{"type":619,"value":4172},"import",{"type":614,"tag":4162,"props":4174,"children":4176},{"style":4175},"--shiki-default:#DBD7CAEE",[4177],{"type":619,"value":4178}," torch\n",{"type":614,"tag":4162,"props":4180,"children":4181},{"class":4164,"line":621},[4182,4187,4192,4196,4201,4207],{"type":614,"tag":4162,"props":4183,"children":4184},{"style":4169},[4185],{"type":619,"value":4186},"from",{"type":614,"tag":4162,"props":4188,"children":4189},{"style":4175},[4190],{"type":619,"value":4191}," transformers ",{"type":614,"tag":4162,"props":4193,"children":4194},{"style":4169},[4195],{"type":619,"value":4172},{"type":614,"tag":4162,"props":4197,"children":4198},{"style":4175},[4199],{"type":619,"value":4200}," AutoModelForCausalLM",{"type":614,"tag":4162,"props":4202,"children":4204},{"style":4203},"--shiki-default:#666666",[4205],{"type":619,"value":4206},",",{"type":614,"tag":4162,"props":4208,"children":4209},{"style":4175},[4210],{"type":619,"value":4211}," AutoTokenizer\n",{"type":614,"tag":4162,"props":4213,"children":4215},{"class":4164,"line":4214},3,[4216],{"type":614,"tag":4162,"props":4217,"children":4219},{"emptyLinePlaceholder":4218},true,[4220],{"type":619,"value":4221},"\n",{"type":614,"tag":4162,"props":4223,"children":4224},{"class":4164,"line":91},[4225,4230,4235,4241,4247],{"type":614,"tag":4162,"props":4226,"children":4227},{"style":4175},[4228],{"type":619,"value":4229},"model_id ",{"type":614,"tag":4162,"props":4231,"children":4232},{"style":4203},[4233],{"type":619,"value":4234},"=",{"type":614,"tag":4162,"props":4236,"children":4238},{"style":4237},"--shiki-default:#C98A7D77",[4239],{"type":619,"value":4240}," \"",{"type":614,"tag":4162,"props":4242,"children":4244},{"style":4243},"--shiki-default:#C98A7D",[4245],{"type":619,"value":4246},"Qwen/Qwen2.5-7B-Instruct",{"type":614,"tag":4162,"props":4248,"children":4249},{"style":4237},[4250],{"type":619,"value":4251},"\"\n",{"type":614,"tag":4162,"props":4253,"children":4254},{"class":4164,"line":92},[4255,4260,4264,4269,4274,4279,4284,4289],{"type":614,"tag":4162,"props":4256,"children":4257},{"style":4175},[4258],{"type":619,"value":4259},"tokenizer ",{"type":614,"tag":4162,"props":4261,"children":4262},{"style":4203},[4263],{"type":619,"value":4234},{"type":614,"tag":4162,"props":4265,"children":4266},{"style":4175},[4267],{"type":619,"value":4268}," AutoTokenizer",{"type":614,"tag":4162,"props":4270,"children":4271},{"style":4203},[4272],{"type":619,"value":4273},".",{"type":614,"tag":4162,"props":4275,"children":4276},{"style":4175},[4277],{"type":619,"value":4278},"from_pretrained",{"type":614,"tag":4162,"props":4280,"children":4281},{"style":4203},[4282],{"type":619,"value":4283},"(",{"type":614,"tag":4162,"props":4285,"children":4286},{"style":4175},[4287],{"type":619,"value":4288},"model_id",{"type":614,"tag":4162,"props":4290,"children":4291},{"style":4203},[4292],{"type":619,"value":4293},")\n",{"type":614,"tag":4162,"props":4295,"children":4297},{"class":4164,"line":4296},6,[4298,4303,4307,4311,4315,4319],{"type":614,"tag":4162,"props":4299,"children":4300},{"style":4175},[4301],{"type":619,"value":4302},"model ",{"type":614,"tag":4162,"props":4304,"children":4305},{"style":4203},[4306],{"type":619,"value":4234},{"type":614,"tag":4162,"props":4308,"children":4309},{"style":4175},[4310],{"type":619,"value":4200},{"type":614,"tag":4162,"props":4312,"children":4313},{"style":4203},[4314],{"type":619,"value":4273},{"type":614,"tag":4162,"props":4316,"children":4317},{"style":4175},[4318],{"type":619,"value":4278},{"type":614,"tag":4162,"props":4320,"children":4321},{"style":4203},[4322],{"type":619,"value":4323},"(\n",{"type":614,"tag":4162,"props":4325,"children":4327},{"class":4164,"line":4326},7,[4328,4333,4337,4343,4347,4352,4356,4361,4365,4370,4374],{"type":614,"tag":4162,"props":4329,"children":4330},{"style":4175},[4331],{"type":619,"value":4332},"    model_id",{"type":614,"tag":4162,"props":4334,"children":4335},{"style":4203},[4336],{"type":619,"value":4206},{"type":614,"tag":4162,"props":4338,"children":4340},{"style":4339},"--shiki-default:#BD976A",[4341],{"type":619,"value":4342}," output_hidden_states",{"type":614,"tag":4162,"props":4344,"children":4345},{"style":4203},[4346],{"type":619,"value":4234},{"type":614,"tag":4162,"props":4348,"children":4349},{"style":4169},[4350],{"type":619,"value":4351},"True",{"type":614,"tag":4162,"props":4353,"children":4354},{"style":4203},[4355],{"type":619,"value":4206},{"type":614,"tag":4162,"props":4357,"children":4358},{"style":4339},[4359],{"type":619,"value":4360}," torch_dtype",{"type":614,"tag":4162,"props":4362,"children":4363},{"style":4203},[4364],{"type":619,"value":4234},{"type":614,"tag":4162,"props":4366,"children":4367},{"style":4175},[4368],{"type":619,"value":4369},"torch",{"type":614,"tag":4162,"props":4371,"children":4372},{"style":4203},[4373],{"type":619,"value":4273},{"type":614,"tag":4162,"props":4375,"children":4376},{"style":4175},[4377],{"type":619,"value":4378},"float16\n",{"type":614,"tag":4162,"props":4380,"children":4382},{"class":4164,"line":4381},8,[4383],{"type":614,"tag":4162,"props":4384,"children":4385},{"style":4203},[4386],{"type":619,"value":4293},{"type":614,"tag":4162,"props":4388,"children":4390},{"class":4164,"line":4389},9,[4391],{"type":614,"tag":4162,"props":4392,"children":4393},{"emptyLinePlaceholder":4218},[4394],{"type":619,"value":4221},{"type":614,"tag":4162,"props":4396,"children":4398},{"class":4164,"line":4397},10,[4399,4404,4408,4412,4417],{"type":614,"tag":4162,"props":4400,"children":4401},{"style":4175},[4402],{"type":619,"value":4403},"prompt ",{"type":614,"tag":4162,"props":4405,"children":4406},{"style":4203},[4407],{"type":619,"value":4234},{"type":614,"tag":4162,"props":4409,"children":4410},{"style":4237},[4411],{"type":619,"value":4240},{"type":614,"tag":4162,"props":4413,"children":4414},{"style":4243},[4415],{"type":619,"value":4416},"The capital of France is",{"type":614,"tag":4162,"props":4418,"children":4419},{"style":4237},[4420],{"type":619,"value":4251},{"type":614,"tag":4162,"props":4422,"children":4424},{"class":4164,"line":4423},11,[4425,4430,4434,4439,4443,4448,4452,4457,4461,4466,4471,4475],{"type":614,"tag":4162,"props":4426,"children":4427},{"style":4175},[4428],{"type":619,"value":4429},"inputs ",{"type":614,"tag":4162,"props":4431,"children":4432},{"style":4203},[4433],{"type":619,"value":4234},{"type":614,"tag":4162,"props":4435,"children":4436},{"style":4175},[4437],{"type":619,"value":4438}," tokenizer",{"type":614,"tag":4162,"props":4440,"children":4441},{"style":4203},[4442],{"type":619,"value":4283},{"type":614,"tag":4162,"props":4444,"children":4445},{"style":4175},[4446],{"type":619,"value":4447},"prompt",{"type":614,"tag":4162,"props":4449,"children":4450},{"style":4203},[4451],{"type":619,"value":4206},{"type":614,"tag":4162,"props":4453,"children":4454},{"style":4339},[4455],{"type":619,"value":4456}," return_tensors",{"type":614,"tag":4162,"props":4458,"children":4459},{"style":4203},[4460],{"type":619,"value":4234},{"type":614,"tag":4162,"props":4462,"children":4463},{"style":4237},[4464],{"type":619,"value":4465},"\"",{"type":614,"tag":4162,"props":4467,"children":4468},{"style":4243},[4469],{"type":619,"value":4470},"pt",{"type":614,"tag":4162,"props":4472,"children":4473},{"style":4237},[4474],{"type":619,"value":4465},{"type":614,"tag":4162,"props":4476,"children":4477},{"style":4203},[4478],{"type":619,"value":4293},{"type":614,"tag":4162,"props":4480,"children":4482},{"class":4164,"line":4481},12,[4483],{"type":614,"tag":4162,"props":4484,"children":4485},{"emptyLinePlaceholder":4218},[4486],{"type":619,"value":4221},{"type":614,"tag":4162,"props":4488,"children":4490},{"class":4164,"line":4489},13,[4491],{"type":614,"tag":4162,"props":4492,"children":4494},{"style":4493},"--shiki-default:#758575DD",[4495],{"type":619,"value":4496},"# 取中間層隱藏狀態並計算對最終 logit 的梯度（J-lens 近似）\n",{"type":614,"tag":4162,"props":4498,"children":4500},{"class":4164,"line":4499},14,[4501,4506,4511,4515,4520],{"type":614,"tag":4162,"props":4502,"children":4503},{"style":4169},[4504],{"type":619,"value":4505},"with",{"type":614,"tag":4162,"props":4507,"children":4508},{"style":4175},[4509],{"type":619,"value":4510}," torch",{"type":614,"tag":4162,"props":4512,"children":4513},{"style":4203},[4514],{"type":619,"value":4273},{"type":614,"tag":4162,"props":4516,"children":4517},{"style":4175},[4518],{"type":619,"value":4519},"enable_grad",{"type":614,"tag":4162,"props":4521,"children":4522},{"style":4203},[4523],{"type":619,"value":4524},"():\n",{"type":614,"tag":4162,"props":4526,"children":4528},{"class":4164,"line":4527},15,[4529,4534,4538,4543,4547,4553,4558,4562,4566,4570,4574],{"type":614,"tag":4162,"props":4530,"children":4531},{"style":4175},[4532],{"type":619,"value":4533},"    outputs ",{"type":614,"tag":4162,"props":4535,"children":4536},{"style":4203},[4537],{"type":619,"value":4234},{"type":614,"tag":4162,"props":4539,"children":4540},{"style":4175},[4541],{"type":619,"value":4542}," model",{"type":614,"tag":4162,"props":4544,"children":4545},{"style":4203},[4546],{"type":619,"value":4283},{"type":614,"tag":4162,"props":4548,"children":4550},{"style":4549},"--shiki-default:#CB7676",[4551],{"type":619,"value":4552},"**",{"type":614,"tag":4162,"props":4554,"children":4555},{"style":4175},[4556],{"type":619,"value":4557},"inputs",{"type":614,"tag":4162,"props":4559,"children":4560},{"style":4203},[4561],{"type":619,"value":4206},{"type":614,"tag":4162,"props":4563,"children":4564},{"style":4339},[4565],{"type":619,"value":4342},{"type":614,"tag":4162,"props":4567,"children":4568},{"style":4203},[4569],{"type":619,"value":4234},{"type":614,"tag":4162,"props":4571,"children":4572},{"style":4169},[4573],{"type":619,"value":4351},{"type":614,"tag":4162,"props":4575,"children":4576},{"style":4203},[4577],{"type":619,"value":4293},{"type":614,"tag":4162,"props":4579,"children":4581},{"class":4164,"line":4580},16,[4582],{"type":614,"tag":4162,"props":4583,"children":4584},{"style":4493},[4585],{"type":619,"value":4586},"    # 選擇中間層（約佔總層數 40%-90%）\n",{"type":614,"tag":4162,"props":4588,"children":4590},{"class":4164,"line":4589},17,[4591,4596,4600,4606,4610,4615,4619,4624],{"type":614,"tag":4162,"props":4592,"children":4593},{"style":4175},[4594],{"type":619,"value":4595},"    n_layers ",{"type":614,"tag":4162,"props":4597,"children":4598},{"style":4203},[4599],{"type":619,"value":4234},{"type":614,"tag":4162,"props":4601,"children":4603},{"style":4602},"--shiki-default:#B8A965",[4604],{"type":619,"value":4605}," len",{"type":614,"tag":4162,"props":4607,"children":4608},{"style":4203},[4609],{"type":619,"value":4283},{"type":614,"tag":4162,"props":4611,"children":4612},{"style":4175},[4613],{"type":619,"value":4614},"outputs",{"type":614,"tag":4162,"props":4616,"children":4617},{"style":4203},[4618],{"type":619,"value":4273},{"type":614,"tag":4162,"props":4620,"children":4621},{"style":4175},[4622],{"type":619,"value":4623},"hidden_states",{"type":614,"tag":4162,"props":4625,"children":4626},{"style":4203},[4627],{"type":619,"value":4293},{"type":614,"tag":4162,"props":4629,"children":4631},{"class":4164,"line":4630},18,[4632,4637,4641,4646,4650,4655,4660,4666],{"type":614,"tag":4162,"props":4633,"children":4634},{"style":4175},[4635],{"type":619,"value":4636},"    mid_layer ",{"type":614,"tag":4162,"props":4638,"children":4639},{"style":4203},[4640],{"type":619,"value":4234},{"type":614,"tag":4162,"props":4642,"children":4643},{"style":4602},[4644],{"type":619,"value":4645}," int",{"type":614,"tag":4162,"props":4647,"children":4648},{"style":4203},[4649],{"type":619,"value":4283},{"type":614,"tag":4162,"props":4651,"children":4652},{"style":4175},[4653],{"type":619,"value":4654},"n_layers ",{"type":614,"tag":4162,"props":4656,"children":4657},{"style":4549},[4658],{"type":619,"value":4659},"*",{"type":614,"tag":4162,"props":4661,"children":4663},{"style":4662},"--shiki-default:#4C9A91",[4664],{"type":619,"value":4665}," 0.6",{"type":614,"tag":4162,"props":4667,"children":4668},{"style":4203},[4669],{"type":619,"value":4293},{"type":614,"tag":4162,"props":4671,"children":4673},{"class":4164,"line":4672},19,[4674,4679,4683,4688,4692,4696,4701,4706,4711,4716,4720,4724],{"type":614,"tag":4162,"props":4675,"children":4676},{"style":4175},[4677],{"type":619,"value":4678},"    hidden ",{"type":614,"tag":4162,"props":4680,"children":4681},{"style":4203},[4682],{"type":619,"value":4234},{"type":614,"tag":4162,"props":4684,"children":4685},{"style":4175},[4686],{"type":619,"value":4687}," outputs",{"type":614,"tag":4162,"props":4689,"children":4690},{"style":4203},[4691],{"type":619,"value":4273},{"type":614,"tag":4162,"props":4693,"children":4694},{"style":4175},[4695],{"type":619,"value":4623},{"type":614,"tag":4162,"props":4697,"children":4698},{"style":4203},[4699],{"type":619,"value":4700},"[",{"type":614,"tag":4162,"props":4702,"children":4703},{"style":4175},[4704],{"type":619,"value":4705},"mid_layer",{"type":614,"tag":4162,"props":4707,"children":4708},{"style":4203},[4709],{"type":619,"value":4710},"].",{"type":614,"tag":4162,"props":4712,"children":4713},{"style":4175},[4714],{"type":619,"value":4715},"requires_grad_",{"type":614,"tag":4162,"props":4717,"children":4718},{"style":4203},[4719],{"type":619,"value":4283},{"type":614,"tag":4162,"props":4721,"children":4722},{"style":4169},[4723],{"type":619,"value":4351},{"type":614,"tag":4162,"props":4725,"children":4726},{"style":4203},[4727],{"type":619,"value":4293},{"type":614,"tag":4162,"props":4729,"children":4731},{"class":4164,"line":4730},20,[4732,4737,4741,4745,4749,4754,4758,4763,4767,4772,4777,4781,4786,4791],{"type":614,"tag":4162,"props":4733,"children":4734},{"style":4175},[4735],{"type":619,"value":4736},"    logit_sum ",{"type":614,"tag":4162,"props":4738,"children":4739},{"style":4203},[4740],{"type":619,"value":4234},{"type":614,"tag":4162,"props":4742,"children":4743},{"style":4175},[4744],{"type":619,"value":4687},{"type":614,"tag":4162,"props":4746,"children":4747},{"style":4203},[4748],{"type":619,"value":4273},{"type":614,"tag":4162,"props":4750,"children":4751},{"style":4175},[4752],{"type":619,"value":4753},"logits",{"type":614,"tag":4162,"props":4755,"children":4756},{"style":4203},[4757],{"type":619,"value":4700},{"type":614,"tag":4162,"props":4759,"children":4760},{"style":4662},[4761],{"type":619,"value":4762},"0",{"type":614,"tag":4162,"props":4764,"children":4765},{"style":4203},[4766],{"type":619,"value":4206},{"type":614,"tag":4162,"props":4768,"children":4769},{"style":4549},[4770],{"type":619,"value":4771}," -",{"type":614,"tag":4162,"props":4773,"children":4774},{"style":4662},[4775],{"type":619,"value":4776},"1",{"type":614,"tag":4162,"props":4778,"children":4779},{"style":4203},[4780],{"type":619,"value":4206},{"type":614,"tag":4162,"props":4782,"children":4783},{"style":4203},[4784],{"type":619,"value":4785}," :].",{"type":614,"tag":4162,"props":4787,"children":4788},{"style":4175},[4789],{"type":619,"value":4790},"sum",{"type":614,"tag":4162,"props":4792,"children":4793},{"style":4203},[4794],{"type":619,"value":4795},"()\n",{"type":614,"tag":4162,"props":4797,"children":4799},{"class":4164,"line":4798},21,[4800,4805,4809,4813,4817,4822,4826,4831,4835,4840,4844,4849,4854,4858],{"type":614,"tag":4162,"props":4801,"children":4802},{"style":4175},[4803],{"type":619,"value":4804},"    grad ",{"type":614,"tag":4162,"props":4806,"children":4807},{"style":4203},[4808],{"type":619,"value":4234},{"type":614,"tag":4162,"props":4810,"children":4811},{"style":4175},[4812],{"type":619,"value":4510},{"type":614,"tag":4162,"props":4814,"children":4815},{"style":4203},[4816],{"type":619,"value":4273},{"type":614,"tag":4162,"props":4818,"children":4819},{"style":4175},[4820],{"type":619,"value":4821},"autograd",{"type":614,"tag":4162,"props":4823,"children":4824},{"style":4203},[4825],{"type":619,"value":4273},{"type":614,"tag":4162,"props":4827,"children":4828},{"style":4175},[4829],{"type":619,"value":4830},"grad",{"type":614,"tag":4162,"props":4832,"children":4833},{"style":4203},[4834],{"type":619,"value":4283},{"type":614,"tag":4162,"props":4836,"children":4837},{"style":4175},[4838],{"type":619,"value":4839},"logit_sum",{"type":614,"tag":4162,"props":4841,"children":4842},{"style":4203},[4843],{"type":619,"value":4206},{"type":614,"tag":4162,"props":4845,"children":4846},{"style":4175},[4847],{"type":619,"value":4848}," hidden",{"type":614,"tag":4162,"props":4850,"children":4851},{"style":4203},[4852],{"type":619,"value":4853},")[",{"type":614,"tag":4162,"props":4855,"children":4856},{"style":4662},[4857],{"type":619,"value":4762},{"type":614,"tag":4162,"props":4859,"children":4860},{"style":4203},[4861],{"type":619,"value":4862},"]\n",{"type":614,"tag":4162,"props":4864,"children":4866},{"class":4164,"line":4865},22,[4867],{"type":614,"tag":4162,"props":4868,"children":4869},{"style":4493},[4870],{"type":619,"value":4871},"    # 投影回 token 空間以獲得可讀表徵\n",{"type":614,"tag":4162,"props":4873,"children":4875},{"class":4164,"line":4874},23,[4876,4881,4885,4889,4893,4898,4902,4906,4910,4914,4918,4922,4926,4931],{"type":614,"tag":4162,"props":4877,"children":4878},{"style":4175},[4879],{"type":619,"value":4880},"    readable_logits ",{"type":614,"tag":4162,"props":4882,"children":4883},{"style":4203},[4884],{"type":619,"value":4234},{"type":614,"tag":4162,"props":4886,"children":4887},{"style":4175},[4888],{"type":619,"value":4542},{"type":614,"tag":4162,"props":4890,"children":4891},{"style":4203},[4892],{"type":619,"value":4273},{"type":614,"tag":4162,"props":4894,"children":4895},{"style":4175},[4896],{"type":619,"value":4897},"lm_head",{"type":614,"tag":4162,"props":4899,"children":4900},{"style":4203},[4901],{"type":619,"value":4283},{"type":614,"tag":4162,"props":4903,"children":4904},{"style":4175},[4905],{"type":619,"value":4830},{"type":614,"tag":4162,"props":4907,"children":4908},{"style":4203},[4909],{"type":619,"value":4700},{"type":614,"tag":4162,"props":4911,"children":4912},{"style":4662},[4913],{"type":619,"value":4762},{"type":614,"tag":4162,"props":4915,"children":4916},{"style":4203},[4917],{"type":619,"value":4206},{"type":614,"tag":4162,"props":4919,"children":4920},{"style":4549},[4921],{"type":619,"value":4771},{"type":614,"tag":4162,"props":4923,"children":4924},{"style":4662},[4925],{"type":619,"value":4776},{"type":614,"tag":4162,"props":4927,"children":4928},{"style":4203},[4929],{"type":619,"value":4930},":,",{"type":614,"tag":4162,"props":4932,"children":4933},{"style":4203},[4934],{"type":619,"value":4935}," :])\n",{"type":614,"tag":4162,"props":4937,"children":4939},{"class":4164,"line":4938},24,[4940,4945,4949,4953,4957,4962,4966,4971,4975,4980,4984,4989,4994,4999,5003,5007],{"type":614,"tag":4162,"props":4941,"children":4942},{"style":4175},[4943],{"type":619,"value":4944},"    top_tokens ",{"type":614,"tag":4162,"props":4946,"children":4947},{"style":4203},[4948],{"type":619,"value":4234},{"type":614,"tag":4162,"props":4950,"children":4951},{"style":4175},[4952],{"type":619,"value":4438},{"type":614,"tag":4162,"props":4954,"children":4955},{"style":4203},[4956],{"type":619,"value":4273},{"type":614,"tag":4162,"props":4958,"children":4959},{"style":4175},[4960],{"type":619,"value":4961},"decode",{"type":614,"tag":4162,"props":4963,"children":4964},{"style":4203},[4965],{"type":619,"value":4283},{"type":614,"tag":4162,"props":4967,"children":4968},{"style":4175},[4969],{"type":619,"value":4970},"readable_logits",{"type":614,"tag":4162,"props":4972,"children":4973},{"style":4203},[4974],{"type":619,"value":4273},{"type":614,"tag":4162,"props":4976,"children":4977},{"style":4175},[4978],{"type":619,"value":4979},"topk",{"type":614,"tag":4162,"props":4981,"children":4982},{"style":4203},[4983],{"type":619,"value":4283},{"type":614,"tag":4162,"props":4985,"children":4986},{"style":4662},[4987],{"type":619,"value":4988},"5",{"type":614,"tag":4162,"props":4990,"children":4991},{"style":4203},[4992],{"type":619,"value":4993},").",{"type":614,"tag":4162,"props":4995,"children":4996},{"style":4175},[4997],{"type":619,"value":4998},"indices",{"type":614,"tag":4162,"props":5000,"children":5001},{"style":4203},[5002],{"type":619,"value":4700},{"type":614,"tag":4162,"props":5004,"children":5005},{"style":4662},[5006],{"type":619,"value":4762},{"type":614,"tag":4162,"props":5008,"children":5009},{"style":4203},[5010],{"type":619,"value":5011},"])\n",{"type":614,"tag":4162,"props":5013,"children":5015},{"class":4164,"line":5014},25,[5016,5021,5025,5030,5035,5041,5045,5050,5055,5059,5064,5068,5072],{"type":614,"tag":4162,"props":5017,"children":5018},{"style":4602},[5019],{"type":619,"value":5020},"    print",{"type":614,"tag":4162,"props":5022,"children":5023},{"style":4203},[5024],{"type":619,"value":4283},{"type":614,"tag":4162,"props":5026,"children":5027},{"style":4549},[5028],{"type":619,"value":5029},"f",{"type":614,"tag":4162,"props":5031,"children":5032},{"style":4243},[5033],{"type":619,"value":5034},"\"J-space 讀取（第 ",{"type":614,"tag":4162,"props":5036,"children":5038},{"style":5037},"--shiki-default:#C99076",[5039],{"type":619,"value":5040},"{",{"type":614,"tag":4162,"props":5042,"children":5043},{"style":4175},[5044],{"type":619,"value":4705},{"type":614,"tag":4162,"props":5046,"children":5047},{"style":5037},[5048],{"type":619,"value":5049},"}",{"type":614,"tag":4162,"props":5051,"children":5052},{"style":4243},[5053],{"type":619,"value":5054}," 層）：",{"type":614,"tag":4162,"props":5056,"children":5057},{"style":5037},[5058],{"type":619,"value":5040},{"type":614,"tag":4162,"props":5060,"children":5061},{"style":4175},[5062],{"type":619,"value":5063},"top_tokens",{"type":614,"tag":4162,"props":5065,"children":5066},{"style":5037},[5067],{"type":619,"value":5049},{"type":614,"tag":4162,"props":5069,"children":5070},{"style":4243},[5071],{"type":619,"value":4465},{"type":614,"tag":4162,"props":5073,"children":5074},{"style":4203},[5075],{"type":619,"value":4293},{"type":614,"tag":658,"props":5077,"children":5078},{"id":946},[5079],{"type":619,"value":946},{"type":614,"tag":615,"props":5081,"children":5082},{},[5083],{"type":619,"value":5084},"給定「Paris is the capital of ___」提示，J-space 讀取結果應在中間層出現「France」相關 token；若在極早期層（\u003C 總層數 30%）才出現，可能是假陽性。驗測因果介入效果：替換中間層特定 token 表徵後，最終輸出應發生對應的語義變化（如行星身份 token 替換→顏色回答改變），否則介入未成功作用於工作空間核心。",{"type":614,"tag":658,"props":5086,"children":5087},{"id":961},[5088],{"type":619,"value":961},{"type":614,"tag":965,"props":5090,"children":5091},{},[5092,5097,5102],{"type":614,"tag":895,"props":5093,"children":5094},{},[5095],{"type":619,"value":5096},"J-lens 假陽性率高：讀取到的 token 可能與模型實際推理無關，需多次採樣或搭配因果介入驗證",{"type":614,"tag":895,"props":5098,"children":5099},{},[5100],{"type":619,"value":5101},"計算成本：Jacobian 計算需要 backward pass，大模型上每次推理開銷顯著增加（預估 2-5 倍）",{"type":614,"tag":895,"props":5103,"children":5104},{},[5105],{"type":619,"value":5106},"架構敏感性：L38-L92 範圍針對 Claude 系列，遷移至其他模型時應以「總層數 40%-90%」比例重新換算",{"type":614,"tag":658,"props":5108,"children":5109},{"id":985},[5110],{"type":619,"value":985},{"type":614,"tag":965,"props":5112,"children":5113},{},[5114,5119,5124],{"type":614,"tag":895,"props":5115,"children":5116},{},[5117],{"type":619,"value":5118},"觀測：追蹤 J-space 敏感 token（manipulation、fraud、secretly）出現頻率與分布層深變化",{"type":614,"tag":895,"props":5120,"children":5121},{},[5122],{"type":619,"value":5123},"成本：量測每次推理的額外延遲，評估是否超過業務可接受閾值（建議 \u003C 200ms 額外延遲）",{"type":614,"tag":895,"props":5125,"children":5126},{},[5127],{"type":619,"value":5128},"風險：假陽性觸發的誤判風險，須設計人工複查機制，避免自動攔截造成服務中斷",{"type":614,"tag":5130,"props":5131,"children":5132},"style",{},[5133],{"type":619,"value":5134},"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":287,"searchDepth":621,"depth":621,"links":5136},[],{"data":5138,"body":5139,"excerpt":-1,"toc":5869},{"title":287,"description":287},{"type":611,"children":5140},[5141,5145,5150,5155,5159,5164,5802,5806,5811,5816,5820,5843,5847,5865],{"type":614,"tag":658,"props":5142,"children":5143},{"id":871},[5144],{"type":619,"value":871},{"type":614,"tag":615,"props":5146,"children":5147},{},[5148],{"type":619,"value":5149},"目前 DeepSeek 的推論晶片仍在早期開發階段，對工程師而言尚無直接可用的硬體產品。但這個消息對 AI 基礎設施規劃有明確影響：需評估現有推論工作負載對特定 GPU 供應商的依賴程度，並建立硬體無關的 benchmark 基準線。",{"type":614,"tag":615,"props":5151,"children":5152},{},[5153],{"type":619,"value":5154},"依賴 NVIDIA CUDA 生態系的推論部署，短期內面臨風險有限；但若業務場景涉及中國大陸市場，建議同步評估華為昇騰 910B 生態系的整合成本與框架相容性。",{"type":614,"tag":658,"props":5156,"children":5157},{"id":4147},[5158],{"type":619,"value":4150},{"type":614,"tag":615,"props":5160,"children":5161},{},[5162],{"type":619,"value":5163},"建立推論效能基準線，為未來硬體遷移準備比較依據：",{"type":614,"tag":4152,"props":5165,"children":5167},{"className":4154,"code":5166,"language":4156,"meta":287,"style":287},"from vllm import LLM, SamplingParams\nimport time\n\nllm = LLM(model=\"deepseek-ai/DeepSeek-R1-Distill-Qwen-7B\", max_model_len=2048)\nparams = SamplingParams(temperature=0.0, max_tokens=256)\n\nprompts = [\"解釋推論晶片與訓練晶片的核心差異\"] * 50\nt0 = time.time()\noutputs = llm.generate(prompts, params)\nelapsed = time.time() - t0\n\ntotal_tokens = sum(len(o.outputs[0].token_ids) for o in outputs)\nprint(f\"吞吐量：{len(prompts)/elapsed:.2f} req/s\")\nprint(f\"Token 速率：{total_tokens/elapsed:.0f} tokens/s\")\nprint(f\"平均延遲：{elapsed/len(prompts)*1000:.0f} ms/req\")\n",[5168],{"type":614,"tag":725,"props":5169,"children":5170},{"__ignoreMap":287},[5171,5201,5213,5220,5284,5341,5348,5393,5423,5471,5509,5516,5604,5673,5728],{"type":614,"tag":4162,"props":5172,"children":5173},{"class":4164,"line":4165},[5174,5178,5183,5187,5192,5196],{"type":614,"tag":4162,"props":5175,"children":5176},{"style":4169},[5177],{"type":619,"value":4186},{"type":614,"tag":4162,"props":5179,"children":5180},{"style":4175},[5181],{"type":619,"value":5182}," vllm ",{"type":614,"tag":4162,"props":5184,"children":5185},{"style":4169},[5186],{"type":619,"value":4172},{"type":614,"tag":4162,"props":5188,"children":5189},{"style":5037},[5190],{"type":619,"value":5191}," LLM",{"type":614,"tag":4162,"props":5193,"children":5194},{"style":4203},[5195],{"type":619,"value":4206},{"type":614,"tag":4162,"props":5197,"children":5198},{"style":4175},[5199],{"type":619,"value":5200}," SamplingParams\n",{"type":614,"tag":4162,"props":5202,"children":5203},{"class":4164,"line":621},[5204,5208],{"type":614,"tag":4162,"props":5205,"children":5206},{"style":4169},[5207],{"type":619,"value":4172},{"type":614,"tag":4162,"props":5209,"children":5210},{"style":4175},[5211],{"type":619,"value":5212}," time\n",{"type":614,"tag":4162,"props":5214,"children":5215},{"class":4164,"line":4214},[5216],{"type":614,"tag":4162,"props":5217,"children":5218},{"emptyLinePlaceholder":4218},[5219],{"type":619,"value":4221},{"type":614,"tag":4162,"props":5221,"children":5222},{"class":4164,"line":91},[5223,5228,5232,5236,5240,5245,5249,5253,5258,5262,5266,5271,5275,5280],{"type":614,"tag":4162,"props":5224,"children":5225},{"style":4175},[5226],{"type":619,"value":5227},"llm ",{"type":614,"tag":4162,"props":5229,"children":5230},{"style":4203},[5231],{"type":619,"value":4234},{"type":614,"tag":4162,"props":5233,"children":5234},{"style":4175},[5235],{"type":619,"value":5191},{"type":614,"tag":4162,"props":5237,"children":5238},{"style":4203},[5239],{"type":619,"value":4283},{"type":614,"tag":4162,"props":5241,"children":5242},{"style":4339},[5243],{"type":619,"value":5244},"model",{"type":614,"tag":4162,"props":5246,"children":5247},{"style":4203},[5248],{"type":619,"value":4234},{"type":614,"tag":4162,"props":5250,"children":5251},{"style":4237},[5252],{"type":619,"value":4465},{"type":614,"tag":4162,"props":5254,"children":5255},{"style":4243},[5256],{"type":619,"value":5257},"deepseek-ai/DeepSeek-R1-Distill-Qwen-7B",{"type":614,"tag":4162,"props":5259,"children":5260},{"style":4237},[5261],{"type":619,"value":4465},{"type":614,"tag":4162,"props":5263,"children":5264},{"style":4203},[5265],{"type":619,"value":4206},{"type":614,"tag":4162,"props":5267,"children":5268},{"style":4339},[5269],{"type":619,"value":5270}," max_model_len",{"type":614,"tag":4162,"props":5272,"children":5273},{"style":4203},[5274],{"type":619,"value":4234},{"type":614,"tag":4162,"props":5276,"children":5277},{"style":4662},[5278],{"type":619,"value":5279},"2048",{"type":614,"tag":4162,"props":5281,"children":5282},{"style":4203},[5283],{"type":619,"value":4293},{"type":614,"tag":4162,"props":5285,"children":5286},{"class":4164,"line":92},[5287,5292,5296,5301,5305,5310,5314,5319,5323,5328,5332,5337],{"type":614,"tag":4162,"props":5288,"children":5289},{"style":4175},[5290],{"type":619,"value":5291},"params ",{"type":614,"tag":4162,"props":5293,"children":5294},{"style":4203},[5295],{"type":619,"value":4234},{"type":614,"tag":4162,"props":5297,"children":5298},{"style":4175},[5299],{"type":619,"value":5300}," SamplingParams",{"type":614,"tag":4162,"props":5302,"children":5303},{"style":4203},[5304],{"type":619,"value":4283},{"type":614,"tag":4162,"props":5306,"children":5307},{"style":4339},[5308],{"type":619,"value":5309},"temperature",{"type":614,"tag":4162,"props":5311,"children":5312},{"style":4203},[5313],{"type":619,"value":4234},{"type":614,"tag":4162,"props":5315,"children":5316},{"style":4662},[5317],{"type":619,"value":5318},"0.0",{"type":614,"tag":4162,"props":5320,"children":5321},{"style":4203},[5322],{"type":619,"value":4206},{"type":614,"tag":4162,"props":5324,"children":5325},{"style":4339},[5326],{"type":619,"value":5327}," max_tokens",{"type":614,"tag":4162,"props":5329,"children":5330},{"style":4203},[5331],{"type":619,"value":4234},{"type":614,"tag":4162,"props":5333,"children":5334},{"style":4662},[5335],{"type":619,"value":5336},"256",{"type":614,"tag":4162,"props":5338,"children":5339},{"style":4203},[5340],{"type":619,"value":4293},{"type":614,"tag":4162,"props":5342,"children":5343},{"class":4164,"line":4296},[5344],{"type":614,"tag":4162,"props":5345,"children":5346},{"emptyLinePlaceholder":4218},[5347],{"type":619,"value":4221},{"type":614,"tag":4162,"props":5349,"children":5350},{"class":4164,"line":4326},[5351,5356,5360,5365,5369,5374,5378,5383,5388],{"type":614,"tag":4162,"props":5352,"children":5353},{"style":4175},[5354],{"type":619,"value":5355},"prompts ",{"type":614,"tag":4162,"props":5357,"children":5358},{"style":4203},[5359],{"type":619,"value":4234},{"type":614,"tag":4162,"props":5361,"children":5362},{"style":4203},[5363],{"type":619,"value":5364}," [",{"type":614,"tag":4162,"props":5366,"children":5367},{"style":4237},[5368],{"type":619,"value":4465},{"type":614,"tag":4162,"props":5370,"children":5371},{"style":4243},[5372],{"type":619,"value":5373},"解釋推論晶片與訓練晶片的核心差異",{"type":614,"tag":4162,"props":5375,"children":5376},{"style":4237},[5377],{"type":619,"value":4465},{"type":614,"tag":4162,"props":5379,"children":5380},{"style":4203},[5381],{"type":619,"value":5382},"]",{"type":614,"tag":4162,"props":5384,"children":5385},{"style":4549},[5386],{"type":619,"value":5387}," *",{"type":614,"tag":4162,"props":5389,"children":5390},{"style":4662},[5391],{"type":619,"value":5392}," 50\n",{"type":614,"tag":4162,"props":5394,"children":5395},{"class":4164,"line":4381},[5396,5401,5405,5410,5414,5419],{"type":614,"tag":4162,"props":5397,"children":5398},{"style":4175},[5399],{"type":619,"value":5400},"t0 ",{"type":614,"tag":4162,"props":5402,"children":5403},{"style":4203},[5404],{"type":619,"value":4234},{"type":614,"tag":4162,"props":5406,"children":5407},{"style":4175},[5408],{"type":619,"value":5409}," time",{"type":614,"tag":4162,"props":5411,"children":5412},{"style":4203},[5413],{"type":619,"value":4273},{"type":614,"tag":4162,"props":5415,"children":5416},{"style":4175},[5417],{"type":619,"value":5418},"time",{"type":614,"tag":4162,"props":5420,"children":5421},{"style":4203},[5422],{"type":619,"value":4795},{"type":614,"tag":4162,"props":5424,"children":5425},{"class":4164,"line":4389},[5426,5431,5435,5440,5444,5449,5453,5458,5462,5467],{"type":614,"tag":4162,"props":5427,"children":5428},{"style":4175},[5429],{"type":619,"value":5430},"outputs ",{"type":614,"tag":4162,"props":5432,"children":5433},{"style":4203},[5434],{"type":619,"value":4234},{"type":614,"tag":4162,"props":5436,"children":5437},{"style":4175},[5438],{"type":619,"value":5439}," llm",{"type":614,"tag":4162,"props":5441,"children":5442},{"style":4203},[5443],{"type":619,"value":4273},{"type":614,"tag":4162,"props":5445,"children":5446},{"style":4175},[5447],{"type":619,"value":5448},"generate",{"type":614,"tag":4162,"props":5450,"children":5451},{"style":4203},[5452],{"type":619,"value":4283},{"type":614,"tag":4162,"props":5454,"children":5455},{"style":4175},[5456],{"type":619,"value":5457},"prompts",{"type":614,"tag":4162,"props":5459,"children":5460},{"style":4203},[5461],{"type":619,"value":4206},{"type":614,"tag":4162,"props":5463,"children":5464},{"style":4175},[5465],{"type":619,"value":5466}," params",{"type":614,"tag":4162,"props":5468,"children":5469},{"style":4203},[5470],{"type":619,"value":4293},{"type":614,"tag":4162,"props":5472,"children":5473},{"class":4164,"line":4397},[5474,5479,5483,5487,5491,5495,5500,5504],{"type":614,"tag":4162,"props":5475,"children":5476},{"style":4175},[5477],{"type":619,"value":5478},"elapsed ",{"type":614,"tag":4162,"props":5480,"children":5481},{"style":4203},[5482],{"type":619,"value":4234},{"type":614,"tag":4162,"props":5484,"children":5485},{"style":4175},[5486],{"type":619,"value":5409},{"type":614,"tag":4162,"props":5488,"children":5489},{"style":4203},[5490],{"type":619,"value":4273},{"type":614,"tag":4162,"props":5492,"children":5493},{"style":4175},[5494],{"type":619,"value":5418},{"type":614,"tag":4162,"props":5496,"children":5497},{"style":4203},[5498],{"type":619,"value":5499},"()",{"type":614,"tag":4162,"props":5501,"children":5502},{"style":4549},[5503],{"type":619,"value":4771},{"type":614,"tag":4162,"props":5505,"children":5506},{"style":4175},[5507],{"type":619,"value":5508}," t0\n",{"type":614,"tag":4162,"props":5510,"children":5511},{"class":4164,"line":4423},[5512],{"type":614,"tag":4162,"props":5513,"children":5514},{"emptyLinePlaceholder":4218},[5515],{"type":619,"value":4221},{"type":614,"tag":4162,"props":5517,"children":5518},{"class":4164,"line":4481},[5519,5524,5528,5533,5537,5542,5546,5551,5555,5559,5563,5567,5571,5576,5581,5586,5591,5596,5600],{"type":614,"tag":4162,"props":5520,"children":5521},{"style":4175},[5522],{"type":619,"value":5523},"total_tokens ",{"type":614,"tag":4162,"props":5525,"children":5526},{"style":4203},[5527],{"type":619,"value":4234},{"type":614,"tag":4162,"props":5529,"children":5530},{"style":4602},[5531],{"type":619,"value":5532}," sum",{"type":614,"tag":4162,"props":5534,"children":5535},{"style":4203},[5536],{"type":619,"value":4283},{"type":614,"tag":4162,"props":5538,"children":5539},{"style":4602},[5540],{"type":619,"value":5541},"len",{"type":614,"tag":4162,"props":5543,"children":5544},{"style":4203},[5545],{"type":619,"value":4283},{"type":614,"tag":4162,"props":5547,"children":5548},{"style":4175},[5549],{"type":619,"value":5550},"o",{"type":614,"tag":4162,"props":5552,"children":5553},{"style":4203},[5554],{"type":619,"value":4273},{"type":614,"tag":4162,"props":5556,"children":5557},{"style":4175},[5558],{"type":619,"value":4614},{"type":614,"tag":4162,"props":5560,"children":5561},{"style":4203},[5562],{"type":619,"value":4700},{"type":614,"tag":4162,"props":5564,"children":5565},{"style":4662},[5566],{"type":619,"value":4762},{"type":614,"tag":4162,"props":5568,"children":5569},{"style":4203},[5570],{"type":619,"value":4710},{"type":614,"tag":4162,"props":5572,"children":5573},{"style":4175},[5574],{"type":619,"value":5575},"token_ids",{"type":614,"tag":4162,"props":5577,"children":5578},{"style":4203},[5579],{"type":619,"value":5580},")",{"type":614,"tag":4162,"props":5582,"children":5583},{"style":4169},[5584],{"type":619,"value":5585}," for",{"type":614,"tag":4162,"props":5587,"children":5588},{"style":4175},[5589],{"type":619,"value":5590}," o ",{"type":614,"tag":4162,"props":5592,"children":5593},{"style":4169},[5594],{"type":619,"value":5595},"in",{"type":614,"tag":4162,"props":5597,"children":5598},{"style":4175},[5599],{"type":619,"value":4687},{"type":614,"tag":4162,"props":5601,"children":5602},{"style":4203},[5603],{"type":619,"value":4293},{"type":614,"tag":4162,"props":5605,"children":5606},{"class":4164,"line":4489},[5607,5612,5616,5620,5625,5629,5633,5637,5641,5645,5650,5655,5660,5664,5669],{"type":614,"tag":4162,"props":5608,"children":5609},{"style":4602},[5610],{"type":619,"value":5611},"print",{"type":614,"tag":4162,"props":5613,"children":5614},{"style":4203},[5615],{"type":619,"value":4283},{"type":614,"tag":4162,"props":5617,"children":5618},{"style":4549},[5619],{"type":619,"value":5029},{"type":614,"tag":4162,"props":5621,"children":5622},{"style":4243},[5623],{"type":619,"value":5624},"\"吞吐量：",{"type":614,"tag":4162,"props":5626,"children":5627},{"style":5037},[5628],{"type":619,"value":5040},{"type":614,"tag":4162,"props":5630,"children":5631},{"style":4602},[5632],{"type":619,"value":5541},{"type":614,"tag":4162,"props":5634,"children":5635},{"style":4203},[5636],{"type":619,"value":4283},{"type":614,"tag":4162,"props":5638,"children":5639},{"style":4175},[5640],{"type":619,"value":5457},{"type":614,"tag":4162,"props":5642,"children":5643},{"style":4203},[5644],{"type":619,"value":5580},{"type":614,"tag":4162,"props":5646,"children":5647},{"style":4549},[5648],{"type":619,"value":5649},"/",{"type":614,"tag":4162,"props":5651,"children":5652},{"style":4175},[5653],{"type":619,"value":5654},"elapsed",{"type":614,"tag":4162,"props":5656,"children":5657},{"style":4549},[5658],{"type":619,"value":5659},":.2f",{"type":614,"tag":4162,"props":5661,"children":5662},{"style":5037},[5663],{"type":619,"value":5049},{"type":614,"tag":4162,"props":5665,"children":5666},{"style":4243},[5667],{"type":619,"value":5668}," req/s\"",{"type":614,"tag":4162,"props":5670,"children":5671},{"style":4203},[5672],{"type":619,"value":4293},{"type":614,"tag":4162,"props":5674,"children":5675},{"class":4164,"line":4499},[5676,5680,5684,5688,5693,5697,5702,5706,5710,5715,5719,5724],{"type":614,"tag":4162,"props":5677,"children":5678},{"style":4602},[5679],{"type":619,"value":5611},{"type":614,"tag":4162,"props":5681,"children":5682},{"style":4203},[5683],{"type":619,"value":4283},{"type":614,"tag":4162,"props":5685,"children":5686},{"style":4549},[5687],{"type":619,"value":5029},{"type":614,"tag":4162,"props":5689,"children":5690},{"style":4243},[5691],{"type":619,"value":5692},"\"Token 速率：",{"type":614,"tag":4162,"props":5694,"children":5695},{"style":5037},[5696],{"type":619,"value":5040},{"type":614,"tag":4162,"props":5698,"children":5699},{"style":4175},[5700],{"type":619,"value":5701},"total_tokens",{"type":614,"tag":4162,"props":5703,"children":5704},{"style":4549},[5705],{"type":619,"value":5649},{"type":614,"tag":4162,"props":5707,"children":5708},{"style":4175},[5709],{"type":619,"value":5654},{"type":614,"tag":4162,"props":5711,"children":5712},{"style":4549},[5713],{"type":619,"value":5714},":.0f",{"type":614,"tag":4162,"props":5716,"children":5717},{"style":5037},[5718],{"type":619,"value":5049},{"type":614,"tag":4162,"props":5720,"children":5721},{"style":4243},[5722],{"type":619,"value":5723}," tokens/s\"",{"type":614,"tag":4162,"props":5725,"children":5726},{"style":4203},[5727],{"type":619,"value":4293},{"type":614,"tag":4162,"props":5729,"children":5730},{"class":4164,"line":4527},[5731,5735,5739,5743,5748,5752,5756,5760,5764,5768,5772,5776,5780,5785,5789,5793,5798],{"type":614,"tag":4162,"props":5732,"children":5733},{"style":4602},[5734],{"type":619,"value":5611},{"type":614,"tag":4162,"props":5736,"children":5737},{"style":4203},[5738],{"type":619,"value":4283},{"type":614,"tag":4162,"props":5740,"children":5741},{"style":4549},[5742],{"type":619,"value":5029},{"type":614,"tag":4162,"props":5744,"children":5745},{"style":4243},[5746],{"type":619,"value":5747},"\"平均延遲：",{"type":614,"tag":4162,"props":5749,"children":5750},{"style":5037},[5751],{"type":619,"value":5040},{"type":614,"tag":4162,"props":5753,"children":5754},{"style":4175},[5755],{"type":619,"value":5654},{"type":614,"tag":4162,"props":5757,"children":5758},{"style":4549},[5759],{"type":619,"value":5649},{"type":614,"tag":4162,"props":5761,"children":5762},{"style":4602},[5763],{"type":619,"value":5541},{"type":614,"tag":4162,"props":5765,"children":5766},{"style":4203},[5767],{"type":619,"value":4283},{"type":614,"tag":4162,"props":5769,"children":5770},{"style":4175},[5771],{"type":619,"value":5457},{"type":614,"tag":4162,"props":5773,"children":5774},{"style":4203},[5775],{"type":619,"value":5580},{"type":614,"tag":4162,"props":5777,"children":5778},{"style":4549},[5779],{"type":619,"value":4659},{"type":614,"tag":4162,"props":5781,"children":5782},{"style":4662},[5783],{"type":619,"value":5784},"1000",{"type":614,"tag":4162,"props":5786,"children":5787},{"style":4549},[5788],{"type":619,"value":5714},{"type":614,"tag":4162,"props":5790,"children":5791},{"style":5037},[5792],{"type":619,"value":5049},{"type":614,"tag":4162,"props":5794,"children":5795},{"style":4243},[5796],{"type":619,"value":5797}," ms/req\"",{"type":614,"tag":4162,"props":5799,"children":5800},{"style":4203},[5801],{"type":619,"value":4293},{"type":614,"tag":658,"props":5803,"children":5804},{"id":946},[5805],{"type":619,"value":946},{"type":614,"tag":615,"props":5807,"children":5808},{},[5809],{"type":619,"value":5810},"建議關注三個驗測維度：華為昇騰 910B 與 NVIDIA H800 的推論吞吐量對比；開源推論框架（vLLM、MindIE）在非 NVIDIA 硬體上的相容性；以及能源效率 (tokens/watt) 指標。",{"type":614,"tag":615,"props":5812,"children":5813},{},[5814],{"type":619,"value":5815},"上述 benchmark 結果應存檔，作為未來 DeepSeek 自研晶片（或其他替代硬體）可用時的評估基準線。",{"type":614,"tag":658,"props":5817,"children":5818},{"id":961},[5819],{"type":619,"value":961},{"type":614,"tag":965,"props":5821,"children":5822},{},[5823,5828,5833,5838],{"type":614,"tag":895,"props":5824,"children":5825},{},[5826],{"type":619,"value":5827},"誤以為 DeepSeek 自研晶片短期內可量產——目前消息僅確認早期研發，商業化時程未知",{"type":614,"tag":895,"props":5829,"children":5830},{},[5831],{"type":619,"value":5832},"忽略 HBM 封鎖對晶片效能的結構性限制——缺乏高頻寬記憶體將使推論記憶體頻寬成為長期瓶頸",{"type":614,"tag":895,"props":5834,"children":5835},{},[5836],{"type":619,"value":5837},"在推論架構中深度鎖定 CUDA 專屬最佳化，降低未來遷移彈性",{"type":614,"tag":895,"props":5839,"children":5840},{},[5841],{"type":619,"value":5842},"過早排除非 NVIDIA 硬體路線，忽略跨後端推論框架的戰略價值",{"type":614,"tag":658,"props":5844,"children":5845},{"id":985},[5846],{"type":619,"value":985},{"type":614,"tag":965,"props":5848,"children":5849},{},[5850,5855,5860],{"type":614,"tag":895,"props":5851,"children":5852},{},[5853],{"type":619,"value":5854},"觀測：推論延遲（P50／P99）、記憶體頻寬使用率、tokens/s 吞吐量",{"type":614,"tag":895,"props":5856,"children":5857},{},[5858],{"type":619,"value":5859},"成本：每百萬 tokens 推論成本、晶片功耗 (TDP) 與電費換算",{"type":614,"tag":895,"props":5861,"children":5862},{},[5863],{"type":619,"value":5864},"風險：硬體供應中斷應變計畫、多供應商備援策略、非 CUDA 框架相容性評估",{"type":614,"tag":5130,"props":5866,"children":5867},{},[5868],{"type":619,"value":5134},{"title":287,"searchDepth":621,"depth":621,"links":5870},[]]