AiX 環球 AI 情報 Global AI Intelligence

第 40 期 2026 年 11 月 6 日（星期五） 07:30 出刊 10 條情報 · 16 原文來源連結 RSS 過往期數

今日十條 · 2026 年 11 月 6 日監管的線先劃在產品上，代理的賬單先寄到企業裡

今日主編判斷

今天兩條主線同時收窄：歐盟人工智能法透過溫莎框架走進北愛爾蘭，但只管嵌進實體產品的系統；美國中期選舉後，OpenAI 把算力園區的電費、用水與保密條款寫成社區契約。企業這端的訊號更直接：Uber 一季只有四程車由代理預訂，企業的代理要買的是可以核對的執行，而非示範。Two lines narrowed today. The EU AI Act reached Northern Ireland through the Windsor Framework, but only for AI embedded in physical products, while OpenAI turned campus energy, water and confidentiality into community contracts after the US midterms. The enterprise signal was blunter: four Uber rides in a quarter were booked by agents. What enterprises need now is auditable execution, not demonstrations.

今日十條

本欄由 AIX Society 編輯部整理公開資料後撰稿，每條均附原始來源連結；事實與觀點分列，觀點屬本會判斷。

按受眾篩選

歐盟人工智能法十一月十日起局部適用於北愛爾蘭，只管嵌進實體產品的系統EU AI Act to apply in Northern Ireland from 10 November, limited to AI embedded in physical products

影響力✓ 已核實Verified已在發生Happening now法務合規Legal & compliance企業決策者Business owners資訊科技IT

事件: 英國商業及貿易部十一月四日發布指引：歐盟人工智能法同月十日起經溫莎框架局部適用於北愛爾蘭，只涵蓋緊密整合於實體產品的系統。The UK Department for Business and Innovation published guidance on 4 November: from 10 November the EU AI Act applies in Northern Ireland through the Windsor Framework, but only to AI closely integrated into physical products.
背後: 判斷標準有二：出廠已預裝，或產品專為該系統而設。可下載的軟件、瀏覽器與 API 一律不在範圍。Two tests decide scope: the AI is pre-installed at the factory, or the product is designed solely to run that system. Downloadable software, browser tools and APIs stay out of scope.

顧問觀點

我們認為，這是英國境內首次由歐盟規則直接約束企業，但被刻意限在最窄的產品安全邊界。它真正改寫的是製造商的產品線決策，而非辦公室的工具政策。We believe this is the first time EU rules bind businesses inside the UK directly, yet the scope is deliberately pinned to the narrowest product-safety boundary. What it really rewrites is manufacturers' product-line decisions, not office tooling policy.

對日常工作的影響: 硬件與消費電子公司的法務主管，本週要盤點銷往北愛的含 AI 產品，並判定屬「緊密整合」還是單純預載軟件。Legal leads at hardware and consumer-electronics firms must this week inventory AI-bearing products sold into Northern Ireland and rule each one closely integrated or merely software preloaded.
如何改善: 本週把產品清單逐項標記「預裝、依賴、無關」三類，只就前兩類準備技術文件與風險管理紀錄。This week tag each product pre-installed, dependent or unaffected, then prepare technical files and risk-management records only for the first two.

來源nibusinessinfo.co.uk（英國政府商業資訊網）（2026-11-05）↗Data & Marketing Association（2026-11-05）↗

OpenAI 發表社區契約，承諾全額承擔算力園區電網成本並公開用水用電OpenAI publishes community compact, pledging full grid cost recovery and site-level water and power disclosure

影響力✓ 已核實Verified6–12 個月6–12 months企業決策者Business owners法務合規Legal & compliance資訊科技IT

事件: OpenAI 十一月五日發表《全國社區契約》及《算力園區政策藍圖》，承諾全額承擔園區用電與電網擴容成本，並公開每站用水、用電與噪音研究。On 5 November OpenAI published a National Community Compact and a Compute Campus Policy Blueprint, pledging to cover the full electricity and grid-expansion cost of its campuses and to disclose site-level water, power and noise data.
背後: 中期選舉後地方反對升溫，四十一項地方公投有四十項通過設限或否決發展方案，企業自願承諾須換成可驗收條款。Opposition hardened after the midterms, with 40 of 41 local ballot measures approving restrictions or rejecting pro-development proposals, so voluntary promises now have to become verifiable terms.

顧問觀點

我們認為，這一輪把「AI 基建」的社會成本從政治爭議搬到合約文本：電費誰付、資料誰公開、違約誰承接，日後都會寫進採購與選址條件。We believe this round moves the social cost of AI infrastructure from political argument into contract text: who pays for power, who publishes the data and who absorbs default will now sit inside procurement and siting terms.

對日常工作的影響: 企業設施與採購主管日後審視雲端或園區合約時，須把電力回收、用水上限與公開報告義務列為條款，而非只看每度電價。Facilities and procurement leads will have to read power pass-through, water caps and disclosure duties as contract clauses in cloud or campus deals, not just look at the tariff.
如何改善: 本週在現有數據中心或雲端合約中找出「用電成本轉嫁」條款，標明誰承擔電網擴容費，並記下一個續約日。This week locate the power pass-through clause in any data-centre or cloud contract, mark who bears grid-expansion cost, and note the next renewal date.

來源The Hindu（路透社）（2026-11-05）↗Silicon Report（2026-11-05）↗

Anthropic 分析二十萬次科研對話：數據分析佔三分之一，構思只佔百分之四點二Anthropic analyses 207,000 research sessions: data analysis takes a third, ideation only 4.2%

影響力✓ 已核實Verified已在發生Happening now企業決策者Business owners人力資源HR資訊科技IT

事件: Anthropic 十一月五日發表研究，以隱私保護方法分析八月二日至九月三十日共二十萬七千零四十九次科研對話，數據分析佔 Claude 會話三成四。Anthropic published a study on 5 November analysing 207,049 research sessions from 2 August to 30 September with privacy-preserving methods; data analysis accounted for 33.7% of Claude sessions.
背後: 研究人員在問卷中把構思與採集數據列為最重要，實際使用卻最少。能力前兩成的任務，使用量是工時推算的兩至四倍。Researchers ranked ideation and data gathering as most important, yet used AI least for them. Tasks in the top fifth by AI suitability drew two to four times the usage their share of work would predict.

顧問觀點

我們認為，這份數據說明人機分工已按能力自動排列，而不是按重要性排列。企業若只按「員工說最重要」來投放工具，會錯配資源。We believe the data shows division of labour is already sorting itself by capability rather than by importance. Firms that allocate tooling by what staff say matters most will misplace resources.

對日常工作的影響: 研發與知識部門主管，本月應比對團隊自述痛點與實際調用紀錄，找出「最重要但最少用」的工序，優先補工具或補人。R&D and knowledge-team leads should this month compare stated pain points with actual usage logs, isolate the most important but least assisted tasks, and fund tooling or people there.
如何改善: 本週抽十個工作日的團隊 AI 使用紀錄，按工序分類統計次數，與團隊自評的重要度排序並列比較。This week pull ten working days of team AI usage, count it by task type, and lay it beside the team's own ranking of importance.

來源Anthropic（2026-11-05）↗ExplainX（2026-11-05）↗

谷歌 Gemini Omni 1.2 Flash 全面開放，十五秒影片每秒收費七美仙Google makes Gemini Omni 1.2 Flash generally available at 7 US cents per second of 720p video

影響力✓ 已核實Verified已在發生Happening now市場推廣Marketing企業決策者Business owners

事件: 谷歌十一月五日把 Gemini Omni 1.2 Flash 全面開放，單段影片由十秒延至十五秒，720p 每秒 0.07 美元，較上一代低三成。Google made Gemini Omni 1.2 Flash generally available on 5 November; clips extend from 10 to 15 seconds and 720p output costs US$0.07 per second, 30% below the previous version.
背後: 同一片可多輪對話式修改，不必整段重生成，並可先出 360p 草稿再升 4K，工序由「重拍」變成「續修」。One clip can be edited across turns instead of re-rendered, and drafts can start at 360p before upscaling to 4K, turning re-shooting into continued editing.

顧問觀點

我們認為，短片單價跌到以秒計且可對話修改，改變的是創意工序的排程，而不是創意本身。審片、授權與版權核對會成為新的瓶頸。We believe second-priced, conversationally editable video shifts the scheduling of creative work rather than creativity itself. Review, likeness clearance and rights checks become the new bottleneck.

對日常工作的影響: 市場部影片主管，本月應以同一段素材實測「生成加兩輪修改」的總耗時，與現行外發剪輯的交付週期比較。Marketing video leads should this month time a generate-plus-two-edits cycle on one asset and compare it with the current outsourced editing turnaround.
如何改善: 本週用一條現成產品片試跑 Omni 1.2，記錄每輪修改耗時與每秒成本，據此訂出內部審片的最短流程。This week run one existing product clip through Omni 1.2, log time per edit round and cost per second, then set the minimum internal review path.

來源Google Cloud Blog（2026-11-05）↗Silicon Report（2026-11-05）↗

騰訊混元 Hy4 正式版發布，首次支援電腦操作，幻覺率降至百分之五點四五Tencent releases Hunyuan Hy4 with computer use for the first time, cutting hallucination rate to 5.45%

影響力✓ 已核實Verified已在發生Happening now資訊科技IT企業決策者Business owners

事件: 騰訊混元十一月四日發布 Hy4 正式版，770B 總參數、49B 激活、100 萬詞元上下文，首次支援視覺理解與電腦操作。Tencent released Hunyuan Hy4 on 4 November: 770B total parameters, 49B active, a 1M-token context, and computer use and visual understanding for the first time.
背後: 幻覺率由預覽版 15.76% 降至 5.45%，並針對 CAD、Blender 等專業軟件與跨應用長程任務專門訓練。Hallucination rate fell from 15.76% in the preview to 5.45%, with dedicated training for CAD and Blender and for long cross-application tasks.

顧問觀點

我們認為，真正值得記的不是參數量，而是幻覺率與錯誤率的下降幅度。企業敢把代理放進舊系統，看的正是「不出錯」而不是「很會講」。We believe the number worth recording is not the parameter count but the fall in hallucination and guess-error rates. What lets enterprises put agents into legacy systems is not eloquence but reliability.

對日常工作的影響: IT 與營運主管，本月可挑一張仍靠人手抄錄的報表，用電腦操作模式試做一次，記錄成功次數與需人工介入的環節。IT and operations leads could this month take one report still copied by hand, run it in computer-use mode once, and log successes and every point needing human intervention.
如何改善: 本週選一個純內部、無客戶資料的流程做三天試跑，只計成功完成次數，不作能力評價。This week pilot one purely internal, customer-data-free workflow for three days and count only completed runs, without scoring capability.

來源騰訊雲開發者社區（2026-11-05）↗騰訊雲技術百科（2026-11-05）↗

螞蟻靈光 2.0 上線一週託管逾一百一十六萬項事務，用戶突破二千萬Ant's Lingguang 2.0 manages 1.16m tasks in its first week, users pass 20 million

影響力✓ 已核實Verified已在發生Happening now企業決策者Business owners前線員工Frontline staff市場推廣Marketing

事件: 螞蟻集團旗下 AI 助理靈光十一月五日公布，十月二十七日上線的 2.0 版一週內管理逾 116 萬項事務、執行逾 360 萬次，用戶破 2,000 萬。Ant Group's AI assistant Lingguang said on 5 November that its 2.0 release, launched on 27 October, managed over 1.16 million tasks and ran more than 3.6 million executions in a week, with users passing 20 million.
背後: 新功能讓助理在用戶授權後脫離對話持續跟進事務；截至十一月三日已接入四十項外部連接器。A new function lets the assistant keep tracking a task outside the chat once authorised; by 3 November it connected 40 external connectors.

顧問觀點

我們認為，代理競爭的焦點已由「答得準」轉向「管得住」：能否在無人盯着的情況下持續推進一件事，才決定用戶是否把事務交出去。We believe agent competition has shifted from answering accurately to managing reliably: whether a task can keep advancing unwatched is what decides if users hand it over.

對日常工作的影響: 客服與營運主管，本月可挑一類需多日跟進的客戶事項，試用持續跟進模式，記錄自動提醒與人工接手的分界。Customer-service and operations leads could this month take one multi-day customer matter, run it in persistent-tracking mode and log where auto-reminders end and people take over.
如何改善: 本週列出一份「跨三日以上」的事務清單，只挑其中一項交給代理，並在日誌記下每個需要人工決定的節點。This week list the matters that routinely run past three days, hand exactly one to an agent and log every point that needed a human decision.

來源人民政協網（2026-11-05）↗

韓國外資峰會錄得二十億美元申報，九成七投向首都圈以外，含龜尾人工智能數據中心Korea's investment summit draws US$2.03bn in pledges, 97% outside Seoul, including an AI data centre in Gumi

影響力✓ 已核實Verified6–12 個月6–12 months企業決策者Business owners資訊科技IT

事件: 韓國 Invest Korea 峰會十一月四日至六日舉行，七家外企申報投資二十億二千七百五十萬美元，九成七落在首都圈以外，涵蓋人工智能數據中心與半導體。At Korea's Invest Korea Summit, held 4 to 6 November, seven foreign firms pledged US$2.0275 billion, 97% of it outside the Seoul metropolitan area, spanning AI data centres and semiconductors.
背後: 英國 Nscale 將在龜尾建亞洲最大人工智能數據中心，美國 Amkor 擴建光州封裝廠，新加坡與挪威企業投資再生能源。Britain's Nscale will build Asia's largest AI data centre in Gumi, Amkor will expand its Gwangju packaging plant, and Singapore and Norwegian firms are funding renewables.

顧問觀點

我們認為，這說明算力正被推到土地與電力較寬裕的二線城市，政府以區域均衡換取投資。選址邏輯已從「靠近客戶」轉為「靠近電」。We believe this shows compute being pushed to second-tier cities with land and power headroom, with regional balance traded for investment. Site logic has shifted from proximity to customers to proximity to power.

對日常工作的影響: 供應鏈與選址主管，日後評估亞太算力或封裝基地時，應把電力接入、許可審批與地方補貼條款列為首要篩選條件。Supply-chain and site leads should rank grid connection, permitting and local incentives first when assessing Asia-Pacific compute or packaging bases.
如何改善: 本週在既有選址評分表中加一欄「電力與許可年期」，為每個候選地填上可查證的供電承諾與審批時程。This week add a power-and-permit-tenure column to your existing site scorecard and fill in a verifiable supply commitment and approval timeline for each candidate.

來源Seoul Economic Daily（2026-11-05）↗

OpenAI 推出 GPT-6 Cyber，僅開放經審核的國防級客戶，Firefox 挖出八十一項高危漏洞OpenAI launches GPT-6 Cyber for vetted defenders only; finds 81 high-severity Firefox flaws

影響力✓ 已核實Verified已在發生Happening now資訊科技IT法務合規Legal & compliance企業決策者Business owners

事件: OpenAI 十一月五日推出 GPT-6 Cyber，只向經審核的 Daybreak Red 客戶開放，在 ExploitGym 取得 48.5%，並在 Firefox 找出 81 項此前未知的高危漏洞。OpenAI launched GPT-6 Cyber on 5 November for vetted Daybreak Red customers only; it scored 48.5% on ExploitGym and found 81 previously unknown high-severity flaws in Firefox.
背後: 上一代只 32.1%，旗艦 Astra 為 42.4%。OpenAI 將其網絡安全能力評為準備框架下的最高級 Critical，並以 IP 白名單與身分核驗限制存取。The prior version scored 32.1% and the flagship Astra 42.4%. OpenAI rates it Critical, the top cyber tier in its Preparedness Framework, gated by IP allowlisting and identity checks.

顧問觀點

我們認為，企業要學的不是這個模型的戰績，而是分級開放的做法：能力愈強，存取門檻、授權邊界與事後紀錄就要愈硬。We believe the lesson for enterprises is not the model's scoreboard but the tiered-access design: the stronger the capability, the harder the gate, the scope and the audit trail.

對日常工作的影響: IT 與安全主管，本月須檢查自家代理中是否存在可被改作掃描用途的工具組合，並確認有無授權範圍清單。IT and security leads should this month check whether any agent toolchain can be repurposed for scanning, and whether an authorised-scope list exists.
如何改善: 本週為每一個可執行動作的代理寫一頁「可觸及範圍」清單，並在系統內以白名單方式落實。This week write a one-page allowed-scope list for every action-capable agent and enforce it as an allowlist inside the system.

來源Web Pulse（2026-11-05）↗Aiholics（2026-11-05）↗

Uber 上季由第三方代理完成的行程只有四程，同期總行程約三億二千萬Uber fulfilled just four rides through third-party agents last quarter against about 320 million trips

影響力✓ 已核實Verified已在發生Happening now企業決策者Business owners市場推廣Marketing資訊科技IT

事件: Uber 十一月四日業績會披露，上季由第三方人工智能代理完成的行程合計四程；萬聖節一週代理下的外賣訂單約五百張，同期總行程約三億二千萬。On its 4 November earnings call Uber disclosed that third-party AI agents fulfilled four rides in total last quarter, and about 500 delivery orders in Halloween week, against roughly 320 million trips in that week.
背後: 行政總裁稱應用程式平均五十秒完成叫車，七成半行程一鍵下單；代理要在既有習慣之外提供更快的路徑才有意義。The chief executive said the app averages 50 seconds to book a ride and 75% of trips are ordered in a single tap; an agent has to beat an existing habit, not merely exist.

顧問觀點

我們認為，這是企業代理投資最需要的一組數字：上線不等於使用，使用不等於完成交易。把「能否被托付」當成驗收標準，比看示範影片可靠。We believe this is the number enterprises most need: going live is not usage, and usage is not a completed transaction. Testing whether an agent can be entrusted beats watching a demo.

對日常工作的影響: 老闆與市場主管，本月應把代理試點的驗收指標由「接入數」改為「完成率」，並要求每週回報實際完成量。Owners and marketing leads should this month replace integration counts with completion rate as the pilot's acceptance test and require a weekly completed-volume report.
如何改善: 本週在試點報表加一欄「代理自行完成的交易數」，與人工介入數並列，一週後再決定是否擴大。This week add a column for transactions the agent completed unaided, beside those needing human intervention, and decide on expansion after one week.

來源AI Updates（2026-11-04）↗Another News（引業績會逐字稿）（2026-11-04）↗

美國十月裁員公告四萬三千九百九十四人，人工智能只佔一千二百六十五人US October layoffs announced at 43,994, of which only 1,265 cited AI

影響力✓ 已核實Verified已在發生Happening now人力資源HR企業決策者Business owners

事件: Challenger 十一月五日公布，美國十月宣布裁員 43,994 人，按年減七成一；歸因人工智能的只有 1,265 人，年內累計十二萬一千四百零一人。Challenger reported on 5 November that US employers announced 43,994 job cuts in October, down 71% year on year, with only 1,265 citing AI against 121,401 for the year to date.
背後: 十月招聘計劃 444,692 人，按年增五成七，多為季節工；公司稱僱主給出的理由已由科技轉向經濟環境。October hiring plans reached 444,692, up 57% year on year and largely seasonal; the firm said employers' stated reasons have shifted from technology to economic conditions.

顧問觀點

我們認為，年內仍有約兩成裁員以人工智能為由，但十月已明顯收縮，說明「AI 裁員」既是真實趨勢，也是方便的解釋。企業不宜以此作為人力規劃的唯一依據。We believe AI still accounts for about a fifth of this year's cuts, yet October's collapse shows the label is both a real trend and a convenient explanation. It should not be the sole basis for workforce planning.

對日常工作的影響: HR 主管本月覆核任何以人工智能為由的減員方案時，應要求附上崗位任務分析，而非部門整體比例。When HR leads review any AI-cited headcount reduction this month, they should demand a task-level role analysis rather than a departmental percentage.
如何改善: 本週為下一個擬議減員方案補做崗位任務清單，逐項標明可由系統完成的比重及驗證方式。This week add a task list to the next proposed reduction, marking for each task the share a system could do and how that would be verified.

來源The Edge Malaysia（引路透社）（2026-11-05）↗

本週可做的三件事

請法務把公司在北愛爾蘭銷售的每一件含 AI 實體產品列出，逐一核對是否屬「緊密整合」範圍。Ask legal to list every AI-enabled physical product you sell into Northern Ireland and test each against the "closely integrated" test.
要求 IT 提交一份代理執行紀錄樣本：誰授權、做了什麼、事後能否還原。Ask IT for one sample agent execution log: who authorised it, what it did, and whether it can be reconstructed afterwards.
在下次預算會上，把「代理試點轉化率」列為與採用率並列的指標。At the next budget meeting, put agent conversion rate alongside adoption rate as a headline metric.

常見問題

今日（2026-11-06）「AiX 環球 AI 情報」有哪十條重點？

今日十條為：1、歐盟人工智能法十一月十日起局部適用於北愛爾蘭，只管嵌進實體產品的系統；2、OpenAI 發表社區契約，承諾全額承擔算力園區電網成本並公開用水用電；3、Anthropic 分析二十萬次科研對話：數據分析佔三分之一，構思只佔百分之四點二；4、谷歌 Gemini Omni 1.2 Flash 全面開放，十五秒影片每秒收費七美仙；5、騰訊混元 Hy4 正式版發布，首次支援電腦操作，幻覺率降至百分之五點四五；6、螞蟻靈光 2.0 上線一週託管逾一百一十六萬項事務，用戶突破二千萬；7、韓國外資峰會錄得二十億美元申報，九成七投向首都圈以外，含龜尾人工智能數據中心；8、OpenAI 推出 GPT-6 Cyber，僅開放經審核的國防級客戶，Firefox 挖出八十一項高危漏洞；9、Uber 上季由第三方代理完成的行程只有四程，同期總行程約三億二千萬；10、美國十月裁員公告四萬三千九百九十四人，人工智能只佔一千二百六十五人。每條均附本會顧問觀點、對日常工作的影響與一項可即時執行的建議。

本會對今日 AI 局勢的整體判斷是甚麼？

今天兩條主線同時收窄：歐盟人工智能法透過溫莎框架走進北愛爾蘭，但只管嵌進實體產品的系統；美國中期選舉後，OpenAI 把算力園區的電費、用水與保密條款寫成社區契約。企業這端的訊號更直接：Uber 一季只有四程車由代理預訂，企業的代理要買的是可以核對的執行，而非示範。

這些情報的資料來源是甚麼？如何核實？

本欄每日由 AIX Society 編輯部檢索公開資料後撰稿，每條新聞均附原始來源連結（本期包括：nibusinessinfo.co.uk（英國政府商業資訊網）、Data & Marketing Association、The Hindu（路透社）、Silicon Report、Anthropic、ExplainX、Google Cloud Blog、騰訊雲開發者社區），並將事實與本會觀點分列。讀者可按來源連結自行核實。

訂閱每週電子報

每週一封：AI 轉型的實戰觀察、工具實測與案例拆解。