[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-ai-weekly-2026-w33-zh":3,"article-related-ai-weekly-2026-w33-zh":27,"series-industry-742c1e45-13c7-43a9-84aa-33b10e58d9af":70},{"id":4,"slug":5,"title":6,"content":7,"summary":8,"source":9,"source_url":10,"author":11,"image_url":12,"cover_image":12,"category":13,"language":14,"translated_content":11,"related_article_id":15,"keywords":16,"key_takeaways":11,"views":24,"created_at":25,"published_at":26,"topic_cluster_id":11},"742c1e45-13c7-43a9-84aa-33b10e58d9af","ai-weekly-2026-w33-zh","AI 週報：2026-08-03 ~ 2026-08-10","\u003Cp>本週主旋律是 AI 從「能用」轉向「可管、可算、可部署」。OpenAI、DeepSeek 與 GPT-5.6 Sol 的更新都指向同一件事：企業開始把成本、速率與治理放進同一張表裡看。\u003C\u002Fp>\u003Cp>另一條線則是端側與上下文能力同步前進，面壁智能的 SALA 把本地上下文做大，代表模型競爭正在從單純參數與榜單，轉到記憶、延遲與落地場景的整體表現。\u003C\u002Fp>\u003Ch2>趨勢雷達\u003C\u002Fh2>\u003Ctable>\u003Ctr>\u003Cth>維度\u003C\u002Fth>\u003Cth>信號\u003C\u002Fth>\u003Cth>本週變化\u003C\u002Fth>\u003Cth>影響\u003C\u002Fth>\u003C\u002Ftr>\u003Ctr>\u003Ctd>模型\u003C\u002Ftd>\u003Ctd>中\u003C\u002Ftd>\u003Ctd>AURORA-LM 將擴散方法搬進文字潛空間，嘗試同時兼顧生成與還原品質。\u003C\u002Ftd>\u003Ctd>研究路線更重視表示空間設計，後續可能影響長文本生成與可逆編碼方案。\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>Agent\u003C\u002Ftd>\u003Ctd>中\u003C\u002Ftd>\u003Ctd>DeepSeek V4 Flash 以更低成本提供 API 相容與 Codex 適配，直接對準 Agent 接入。\u003C\u002Ftd>\u003Ctd>Agent 產品會更在意單次任務成本與工具鏈相容性，而不是只看模型分數。\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>開源\u003C\u002Ftd>\u003Ctd>弱\u003C\u002Ftd>\u003Ctd>本週沒有新的重量級開源模型發布，延續上週開放權重制度化的討論。\u003C\u002Ftd>\u003Ctd>開源陣營短期比拼的重點，會從「是否開放」轉向授權、治理與商用邊界。\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>算力與基建\u003C\u002Ftd>\u003Ctd>強\u003C\u002Ftd>\u003Ctd>OpenAI API 新增支出上限、Fast 模式、GPT Transcribe 與 Terraform provider；GPT-5.6 Sol 也展示用 Codex 改寫 GPU 核心、把成本降 20%。\u003C\u002Ftd>\u003Ctd>雲端 AI 基礎設施正在產品化為可控預算、可編排部署與可持續優化的系統。\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>應用\u003C\u002Ftd>\u003Ctd>中\u003C\u002Ftd>\u003Ctd>圖生圖接單流程、Tesla Model Y L 規格與端側上下文擴展，顯示 AI 應用正往內容生產與裝置內體驗分流。\u003C\u002Ftd>\u003Ctd>應用層會更強調交付流程、硬體整合與使用場景，而不是只談模型能力。\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>政策與監管\u003C\u002Ftd>\u003Ctd>平靜\u003C\u002Ftd>\u003Ctd>無重大進展\u003C\u002Ftd>\u003Ctd>監管面暫時沒有新變化，但支出上限、Terraform 這類治理工具已提前把合規需求寫進產品。\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftable>\u003Ch2>重點事件\u003C\u002Fh2>\u003Ch3>OpenAI API 把治理做進產品\u003C\u002Fh3>\u003Cp>\u003Cstrong>發生什麼。\u003C\u002Fstrong> \u003Ca href=\"\u002Fnews\u002Fopenai-api-changelog-spend-caps-transcribe-fast-mode-zh\">OpenAI API changelog\u003C\u002Fa> 新增 Fast mode、硬性支出上限、GPT Transcribe 與 Terraform provider，讓開發者能直接在 API 層設定成本與部署規則。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786335622166-pb5e.png\" alt=\"AI 週報：2026-08-03 ~ 2026-08-10\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786335623797-0a9h.png\" alt=\"AI 週報：2026-08-03 ~ 2026-08-10\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>\u003Cstrong>為什麼重要。\u003C\u002Fstrong> 這代表 AI 平台不再只賣能力，而是把預算控制、語音能力與基礎設施編排一起打包。對企業來說，採用門檻會下降，但治理責任也會更明確。\u003C\u002Fp>\u003Cp>\u003Cstrong>誰受影響、下一步觀察。\u003C\u002Fstrong> 受影響的是平台工程團隊、SaaS 產品與大量呼叫 API 的 Agent 服務；接下來要看 Fast mode 的實際延遲表現，以及支出上限是否能被大型組織直接納入採購流程。\u003C\u002Fp>\u003Ch3>DeepSeek V4 Flash 讓 Agent 更在意總成本\u003C\u002Fh3>\u003Cp>\u003Cstrong>發生什麼。\u003C\u002Fstrong> DeepSeek V4 Flash 以 API 相容、Codex 適配與更低定價切入 Agent 工作流，重點不是單點能力，而是直接降低任務執行成本。\u003C\u002Fp>\u003Cp>\u003Cstrong>為什麼重要。\u003C\u002Fstrong> Agent 產品的競爭焦點正在變成「每次任務花多少錢、能不能穩定接工具、失敗後怎麼重試」。當模型價格下來，產品設計會更敢做多步驟流程。\u003C\u002Fp>\u003Cp>\u003Cstrong>誰受影響、下一步觀察。\u003C\u002Fstrong> 受影響的是客服自動化、程式助理與內部知識工作流；接下來看是否有團隊把 DeepSeek V4 Flash 直接放進生產 Agent，並公開任務成功率與平均成本。\u003C\u002Fp>\u003Ch3>GPT-5.6 Sol 把推理成本往下壓\u003C\u002Fh3>\u003Cp>\u003Cstrong>發生什麼。\u003C\u002Fstrong> GPT-5.6 Sol 的做法是用 Codex 改寫生產 GPU 核心，據稱把推理成本砍 20%，把優化點從模型本身移到執行層。\u003C\u002Fp>\u003Cp>\u003Cstrong>為什麼重要。\u003C\u002Fstrong> 這類案例說明，AI 成本優化不只靠換更小模型，還能透過 kernel、編譯與推理路徑重寫來擠出空間。對大規模推理服務而言，這比單次模型升級更實際。\u003C\u002Fp>\u003Cp>\u003Cstrong>誰受影響、下一步觀察。\u003C\u002Fstrong> 受影響的是雲端推理平台、模型服務商與做高頻請求的應用；接下來要看這種 kernel 級優化是否能在其他模型與硬體組合上複製。\u003C\u002Fp>\u003Ch3>SALA 把端側上下文做大\u003C\u002Fh3>\u003Cp>\u003Cstrong>發生什麼。\u003C\u002Fstrong> 面壁智能的 SALA 以混合注意力擴展端側大模型的上下文容量，目標是在本地裝置上處理更長的輸入與更連續的對話。\u003C\u002Fp>\u003Cp>\u003Cstrong>為什麼重要。\u003C\u002Fstrong> 端側 AI 的瓶頸常常不是模型會不會答，而是記不記得前文、延遲能不能接受、裝置熱不熱。上下文擴展一旦穩定，手機、筆電與車載場景的體驗會更完整。\u003C\u002Fp>\u003Cp>\u003Cstrong>誰受影響、下一步觀察。\u003C\u002Fstrong> 受影響的是端側模型開發者、手機廠與車機供應鏈；接下來看 SALA 是否進入實機測試，以及在長對話、文件摘要與離線助理上的延遲數據。\u003C\u002Fp>\u003Ch2>下週觀察\u003C\u002Fh2>\u003Cul>\u003Cli>觀察 OpenAI API 的 Fast mode 是否被大型 SaaS 團隊用來重排請求路由，特別是 8 月中旬的企業測試回饋。\u003C\u002Fli>\u003Cli>留意 DeepSeek V4 Flash 是否出現更多 Agent 案例，尤其是支援多工具調度的開發框架與 2026-08-中旬的整合公告。\u003C\u002Fli>\u003Cli>追蹤面壁智能 SALA 是否在 2026-08-下旬釋出端側實測，重點看上下文長度與功耗數據。\u003C\u002Fli>\u003Cli>關注 AURORA-LM 與 ExtractBench 這類研究是否帶動企業文件抽取、長文本生成的後續論文或開源實作。\u003C\u002Fli>\u003Cli>看 Tesla Model Y L 的交付與市場反應，判斷 AI 車載與車機體驗是否會成為下一波硬體差異化入口。\u003C\u002Fli>\u003C\u002Ful>","本週主旋律是 AI 走向可治理、可控成本與端側擴展：OpenAI 強化支出上限與 Fast 模式，DeepSeek 與 GPT-5.6 Sol 持續壓低推理成本，面壁智能則把上下文能力往端側推進。","oracore.dev","https:\u002F\u002Foracore.dev\u002Fnews\u002Fai-weekly-2026-w33-zh",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786335622166-pb5e.png","industry","zh","a5fe53a8-a51c-4005-b927-0bcd383c234c",[17,18,19,20,21,22,23],"AI 週報","趨勢雷達","人工智慧","科技新聞","推理成本","Agent","端側 AI",0,"2026-08-10T04:00:28.661553+00:00","2026-08-10T04:00:28.63+00:00",{"tags":28,"relatedLang":29,"relatedPosts":33},[],{"id":15,"slug":30,"title":31,"language":32},"ai-weekly-2026-w33-en","AI Weekly: 2026-08-03 ~ 2026-08-10","en",[34,40,46,52,58,64],{"id":35,"slug":36,"title":37,"cover_image":38,"image_url":38,"created_at":39,"category":13},"2154ae8a-ee87-4b04-bf30-62619fbed031","crypto-infrastructure-era-ai-agents-zh","AI Agent 讓加密貨幣回到基礎設施","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786347179913-0eep.png","2026-08-10T07:32:31.882495+00:00",{"id":41,"slug":42,"title":43,"cover_image":44,"image_url":44,"created_at":45,"category":13},"fa5bffe8-c724-42ba-bbe5-94591fb9e766","kimi-k3-gpu-api-cost-comparison-zh","Kimi K3 自托管成本很高","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786323790394-kbfm.png","2026-08-10T01:02:51.081557+00:00",{"id":47,"slug":48,"title":49,"cover_image":50,"image_url":50,"created_at":51,"category":13},"f5581f94-ba2b-4164-b3b0-5d634239e90a","cursor-should-not-copy-vs-code-design-zh","Cursor 不該跟著 VS Code 新版長相走","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786278763461-hkgn.png","2026-08-09T12:32:20.427633+00:00",{"id":53,"slug":54,"title":55,"cover_image":56,"image_url":56,"created_at":57,"category":13},"3d745dd7-d34a-427f-9403-94244d93e9ef","wall-street-tokenization-slowly-not-celebrate-pilot-zh","華爾街該慢慢代幣化資產，不該急著慶祝試點","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786237371565-bds4.png","2026-08-09T01:02:28.25862+00:00",{"id":59,"slug":60,"title":61,"cover_image":62,"image_url":62,"created_at":63,"category":13},"661c033e-8701-4b5d-ba69-1843545254a4","asset-tokenization-turns-ownership-into-software-zh","資產代幣化把權利寫成軟體","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786235588664-yp9g.png","2026-08-09T00:32:45.231134+00:00",{"id":65,"slug":66,"title":67,"cover_image":68,"image_url":68,"created_at":69,"category":13},"d9997b41-a0f6-49f5-8d2f-0fbf9733374e","institutional-crypto-tops-30b-tokenization-zh","機構加密資產突破 300 億美元","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786233783789-584a.png","2026-08-09T00:02:28.579957+00:00",[71,76,81,86,91,96,101,106,111,116],{"id":72,"slug":73,"title":74,"created_at":75},"ee073da7-28b3-4752-a319-5a501459fb87","ai-in-2026-what-actually-matters-now-zh","2026 AI 真正重要的事","2026-03-26T07:09:12.008134+00:00",{"id":77,"slug":78,"title":79,"created_at":80},"83bd1795-8548-44c9-9a7e-de50a0923f71","trump-ai-framework-power-speech-state-preemption-zh","川普 AI 框架瞄準電力、言論與州權","2026-03-26T07:12:18.695466+00:00",{"id":82,"slug":83,"title":84,"created_at":85},"ea6be18b-c903-4e54-97b7-5f7447a612e0","nvidia-gtc-2026-big-ai-announcements-zh","NVIDIA GTC 2026 重點拆解","2026-03-26T07:14:26.62638+00:00",{"id":87,"slug":88,"title":89,"created_at":90},"4bcec76f-4c36-4daa-909f-54cd702f7c93","claude-users-spreading-out-and-getting-better-zh","Claude 用戶更分散，也更會用","2026-03-26T07:22:52.325888+00:00",{"id":92,"slug":93,"title":94,"created_at":95},"bd903b15-2473-4178-9789-b7557816e535","openclaw-raises-hard-question-for-ai-models-zh","OpenClaw 逼問 AI 模型價值","2026-03-26T07:24:54.707486+00:00",{"id":97,"slug":98,"title":99,"created_at":100},"eeac6b9e-ad9d-4831-8eec-8bba3f9bca6a","gap-google-gemini-checkout-fashion-search-zh","Gap 把結帳搬進 Gemini","2026-03-26T07:28:23.937768+00:00",{"id":102,"slug":103,"title":104,"created_at":105},"0740e53f-605d-4d57-8601-c10beb126f3c","google-pushes-gemini-transition-to-march-2026-zh","Google 把 Gemini 轉換延到 2026 年 3…","2026-03-26T07:30:12.825269+00:00",{"id":107,"slug":108,"title":109,"created_at":110},"e660d801-2421-4529-8fa9-86b82b066990","metas-llama-4-benchmark-scandal-gets-worse-zh","Meta Llama 4 分數風波又擴大","2026-03-26T07:34:21.156421+00:00",{"id":112,"slug":113,"title":114,"created_at":115},"183f9e7c-e143-40bb-a6d5-67ba84a3a8bc","accenture-mistral-ai-sovereign-enterprise-deal-zh","Accenture 攜手 Mistral AI 賣主權 AI","2026-03-26T07:38:14.818906+00:00",{"id":117,"slug":118,"title":119,"created_at":120},"191d9b1b-768a-478c-978c-dd7431a38149","mistral-ai-faces-its-hardest-year-yet-zh","Mistral AI 迎來最硬的一年","2026-03-26T07:40:23.716374+00:00"]