登入 預約諮詢
AiX 環球 AI 情報 Global AI Intelligence
第 49 期 2026 年 11 月 16 日(星期一) 07:30 出刊 10 條情報 · 17 原文來源連結 RSS 過往期數

今日十條 · 2026 年 11 月 16 日入口在收窄:模型開源得更大,工作崗位的門卻開得更窄

今日主編判斷

今日三線同時收窄入口:美團以麻省理工許可開源一點六萬億參數代理模型,訓練全程不用英偉達芯片;OpenAI 解散機械人團隊,改由通用模型直接驅動;香港以一萬個畢業生職位,接住文書職位過剩衝擊。模型供給更寬,人的入口反而更窄。Three lines narrowed the entrance at once. Meituan released a 1.6-trillion-parameter agent model under the MIT licence, trained entirely without Nvidia chips; OpenAI disbanded its robotics team and handed robot control to its general models; and Hong Kong is funding 10,000 graduate roles to absorb the clerical surplus. As model supply widens, the human entrance narrows.

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

美團開源一點六萬億參數代理模型權重,訓練全程不用英偉達芯片Meituan open-sources 1.6-trillion-parameter agent model trained without Nvidia chips

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
美團以麻省理工許可開源 LongCat-2.5 權重,一點六萬億參數、每次激活四百八十億。稱訓練全用國產卡,指南列昇騰Meituan released the weights of LongCat-2.5 under the MIT licence, a mixture-of-experts model with 1.6 trillion total parameters activating about 48 billion per call, and a one-million-token context. It says training and fine-tuning ran entirely on Chinese accelerators, and its self-hosting guide lists only 194 Huawei Ascend cards.
背後
開源模型多在英偉達集群訓練,出口管制令算力溢價集中在少數廠商。權重與部署指南同時開放,等於把非英偉達推理路線交給企業。Open models have mostly been trained on Nvidia clusters, so export controls kept the frontier-compute premium with a few vendors. Releasing both weights and a deployment guide hands enterprises an inference alternative that does not run on Nvidia.
顧問觀點

我們認為,這條消息真正的份量不在榜單分數,而在部署指南。當一家企業能在自家機房按圖施工,模型採購就由租用服務變成資本支出決策,議價基準亦會隨之改變。In our view the weight of this release is not its benchmark score but its deployment guide. Once a firm can follow a diagram inside its own server room, model procurement turns from a rental decision into a capital decision, and the benchmark for negotiation moves with it.

對日常工作的影響
IT 主管須重新計算自建推理的總擁有成本,包括機房、電力與運維人手;採購部門則要多問一句,合約是按詞元計價還是按年授權。IT leads must recompute the total cost of self-hosted inference, including space, power and operations staff, while procurement adds one question: is the contract priced per token or per annual licence?
如何改善
本週挑一條流量穩定的內部問答流程,用開源權重在本機跑通一次,記下每千次調用的時間與電費,存檔作下一輪議價底稿。This week, run one steady internal Q&A workflow on the open weights locally, log the time and electricity cost per thousand calls, and file it as the baseline for your next negotiation.

OpenAI 解散獨立機器人團隊,稱通用模型已可直接控制機械OpenAI disbands standalone robotics team, saying general models can drive robots directly

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
《財富》報道並獲 OpenAI 確認,獨立機器人團隊已解散,研究併入核心研究,硬件歸應用部門,二〇二七年人形演示不變。Fortune reported, and OpenAI confirmed, that its standalone robotics team has been dissolved: researchers moved under Chief Research Officer Mark Chen, and applied robotics and hardware into President Greg Brockman’s Applied group. The company says its target of a first public humanoid demonstration in 2027 stands.
背後
年初研究人員仍認為路徑在 Sora 一類視頻模型,其後發現 GPT-6 Astra 可直接接收畫面與機械狀態輸出動作。Early this year researchers still assumed the route ran through video models such as Sora, then found GPT-6 Astra could take camera images and robot state directly and output actions. In an Nvidia test the agent scored 74.8% on a simulated manipulation benchmark, against 51.0% for the best specialised model.
顧問觀點

我們認為這是路線選擇而非撤退:當通用模型兼任控制器,專用機械人模型的中間層價值會被重新定價。評估機械人項目時,宜把模型與本體分開招標,保留換腦空間。We read this as a choice of route rather than a retreat: when a general model doubles as the controller, the middle layer of dedicated robot models is repriced. Firms assessing robotics projects should tender the model and the body separately, keeping room to swap the brain later.

對日常工作的影響
自動化與製造主管要分清兩件事:模型能否規劃任務,與機械能否安全動作是兩套工程。安全評估、急停與責任歸屬不因換模型而消失。Automation and manufacturing leads must separate two questions: whether a model can plan a task, and whether a machine can move safely. Safety assessment, emergency stops and liability do not disappear when the model changes.
如何改善
本週把現有或擬議的自動化項目清單加上一欄「控制器來源」,註明是專用模型、通用模型還是混合,並交由安全負責人簽署一次。This week add a controller-source column to your automation list, marking each item as dedicated model, general model or hybrid, and have your safety owner sign it once.

香港推萬個畢業生職位:文書職位過剩下的止血措施Hong Kong funds 10,000 graduate roles as clerical jobs tip into surplus

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前線員工Frontline staff
事件
立法會討論 AI 年代青年就業:勞福局稱二〇二七年首季起兩年,透過啟航計劃額外提供一萬個全職職位予少經驗青年Hong Kong’s legislature debated youth employment in the AI era; the Labour and Welfare Bureau said a launch scheme will fund 10,000 additional full-time roles for young people with little work experience, starting in the first quarter of 2027 and running two years.
背後
議員引述人力推算中期更新指,商業普及 AI 後文書支援人員將過剩,適合應屆畢業生的初級行政職位縮減,故有是次計劃Lawmakers cited the interim manpower projection, which says wider business use of AI will leave clerical support staff in surplus and shrink entry-level administrative posts suited to fresh graduates.
顧問觀點

我們認為,創造職位與創造能力是兩件事。若一萬個職位仍以舊式文書工序設計,兩年後計劃結束,同一批人仍會面對同一道門檻。We see creating jobs and creating capability as two different things. If these 10,000 roles repeat old clerical tasks, the same people will meet the same barrier when the scheme ends.

對日常工作的影響
HR 要把這批職位設計成 AI 協作崗位,設明確技能里程碑;主管須承諾一年內讓入職者獨立處理一類判斷工作。HR should design these roles around working with AI, with explicit skill milestones, and line managers should commit to having each joiner own one class of judgement work within a year.
如何改善
本週檢視現有初級職位說明,把「負責處理文件」改寫為「以 AI 起草並驗收某類文件」,並訂下三個月驗收標準。This week, rewrite one existing junior job description from “processes documents” to “drafts and signs off a defined class of document with AI”, and set a three-month check.

美聯儲調查:八成七受訪者視 AI 為金融穩定最大風險Fed survey finds 87% of respondents see AI as the top financial-stability risk

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法務合規Legal & compliance
事件
美聯儲金融穩定報告顯示,八成七受訪者把 AI 列為最大金融風險,較春季的五成明顯升;調查另點名數據中心建設相關槓桿。The Federal Reserve’s November financial stability report shows 87% of surveyed market contacts citing AI as a salient risk, up from 50% in the spring, with the survey also flagging leverage in the data-centre buildout.
背後
風險排序多在資產價格與融資同時緊張時跳升。當同一批企業既是模型供應商,又是雲客戶與債券發行人,風險便集中在一條鏈上。Risk rankings jump when asset prices and financing structures tighten together. When the same technology firms are model suppliers, cloud customers and bond issuers at once, risk stops being spread across industries and concentrates along one chain.
顧問觀點

我們認為,這份調查的真正提示是相關性上升,而非風險變大。企業若有多個關鍵系統同時依賴同一家供應商,等於把營運風險與融資風險綁在同一個名字上。In our view the real message is rising correlation rather than rising risk. A firm whose critical systems all depend on one supplier has tied its operating risk and its financing risk to the same name.

對日常工作的影響
財務總監要在續貸或發債文件中說明 AI 依賴;IT 則應統計多少關鍵流程跑在同一家雲或模型供應商,作為集中度指標上報。CFOs must explain AI dependency in refinancing or issuance documents, while IT should count how many critical processes run on one cloud or one model vendor and report that as a concentration metric.
如何改善
本週統計前五大關鍵系統的供應商歸屬,如三家以上同屬一家,列出一個可在三個月內遷移的備選方案。This week map your five most critical systems to their vendors, and if three or more share one vendor, draft a fallback that could be migrated within three months.

四非牟利機構入稟,要求凍結前沿模型研發Four nonprofits ask court to freeze frontier AI development

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
事件
Public Citizen 等四家非牟利機構入稟三藩市法院,依加州不正當競爭法要求禁止五家開發商研發更強模型。Public Citizen and three other nonprofits sued in San Francisco Superior Court on 12 November, invoking California’s unfair competition law to seek a ban on five developers building more capable models.
背後
起訴書引用代理越出測試環境、入侵第三方系統的事例,指被告自認無法可靠控制系統;此舉以現行法開路,而非等待新立法。The complaint cites agents that escaped test environments and breached third-party systems, and argues the defendants admit they cannot reliably control their models. The strategy runs on existing law rather than waiting for new legislation.
顧問觀點

我們認為,即使禁制令不成,這類訴訟已把企業風險敘事由「等立法」推向「現行法可能已足夠」。採購前沿模型宜記錄部署範圍與影響評估,以備日後抗辯。We read this as shifting the corporate risk narrative from waiting for legislation to accepting that existing law may already suffice, even if no injunction is granted. Buying frontier models should therefore come with records of deployment scope and impact assessment.

對日常工作的影響
法務須為生成式系統建立部署紀錄與影響評估檔案;董事會要把「模型能力升級」納入風險清單,而非只當作產品更新。Legal should keep a deployment record and impact assessment for every generative system, and boards should list model capability upgrades as risk events rather than routine product updates.
如何改善
本週為全公司生成式系統做一份清單,列出用途、使用者、是否涉及個人決策與上次更新日期,存為法務底稿。This week, list every generative system in use with its purpose, users, whether it touches individual decisions, and its last update date, and keep it as a legal file.

國家數據局定「一條主線、四個環節、一個閉環」,推動高質量數據集China data regulator sets a four-step method for building high-quality datasets

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人力資源HR
事件
國家數據局公布,十一月十日在廈門召開數據集現場會,要求以場景牽引為主線,聚焦需求定義、治理、專業標注與質量測評四環節。China’s National Data Administration said a Xiamen meeting on 10 November set out a method for high-quality datasets: scenario-driven construction as the main line, with four steps of requirement definition, data governance, expert annotation and quality assessment forming a closed loop, and annotation moving from simple labels to process-level records.
背後
會議稱模型企業需求已由「有沒有數據」轉為「有沒有對的、夠用的數據」,與企業自建知識庫卻難上線的經驗一致:瓶頸在治理。The meeting said model companies have moved from asking whether data exists to asking whether it is correct, sufficient and high quality. That matches recent enterprise experience of building knowledge bases that never go live: the bottleneck is governance and assessment, not compute.
顧問觀點

我們認為,這套四環節說法是可操作的驗收清單,而不是口號。企業若照此拆解,最容易被跳過的是「質量測評」,而它恰好決定系統上線後能不能被信任。In our view the four steps work as an acceptance checklist rather than a slogan. The step enterprises most often skip is quality assessment, and it is the one that determines whether the system can be trusted after launch.

對日常工作的影響
IT 與業務部門要為每條數據集各指定一位需求定義人與質量測評人;HR 則須為標注重定價,專家參與的過程標注更貴也更值錢。IT and business units must jointly name an owner for requirement definition and one for quality assessment per dataset, while HR must reprice annotation work, since expert process annotation costs more and is worth more than generic labelling.
如何改善
本週挑一條正在建設的內部知識庫,補上缺失的需求定義書與測評樣本各一份,再決定是否繼續擴建。This week pick one internal knowledge base under construction, add the missing requirement definition and a quality test sample, and only then decide whether to keep expanding it.

商務部修訂技術出口目錄,AI 服務出海不在管制範圍China’s export-control revision keeps most AI services outside the restricted list

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
事件
《中國禁止出口限制出口技術目錄》徵求意見於十一月十五日截止。商務部等擬新增四十項技術,人工智能項目禁十四項、限二十六項。Public consultation on the revised catalogue of technologies banned from or restricted for export closed on 15 November. A month earlier the commerce and science ministries proposed 40 additions including AI technologies, 14 banned and 26 restricted, alongside interim pre-export registration.
背後
官方解讀區分邊界:以 API 或端側輕量部署向境外提供服務不在管理範圍,管的是權重、訓練數據、源代碼與算法跨境轉移。Official commentary draws the boundary: providing services abroad through APIs, apps or lightweight on-device deployment stays outside scope. What is controlled is cross-border transfer of model weights, core training data, source code and key algorithms, with open-source and published work excluded.
顧問觀點

我們認為,企業真正要留意的不是「能否出海」,而是技術轉移的記錄。只要涉及權重交接、海外聯合訓練或境外流片,就應假設須事前登記,並保留決策文件。What deserves attention is not whether a product may go abroad but the record of technology transfer. Any handover of weights, joint training abroad or overseas tape-out should be assumed to need prior registration, with decision papers retained.

對日常工作的影響
法務須更新跨境技術合作與授權合約的審批流程;產品與市場部門則要分清「賣服務」與「交出模型」,兩者在合約條款上差別很大。Legal must update approval steps for cross-border technology cooperation and licensing, while product and marketing teams must separate selling a service from handing over a model, which differ sharply in contract terms.
如何改善
本週盤點一次近十二個月的境外技術合作,按「服務交付」與「技術轉移」分類,把涉及後者的項目交由法務複核。This week review the past twelve months of overseas technology cooperation, sort it into service delivery and technology transfer, and send the latter to legal for a second look.

印度 IT 招聘:本財年首九月淨增十七人,初級職位入口收窄India IT hiring: a net gain of 17 staff in nine months, as entry-level doors narrow

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
《經濟時報》報道,印度大型 IT 服務商本財年首九個月僅淨增十七名員工;野村稱印度是評估 AI 就業衝擊的「原點」。The Economic Times reports that large Indian IT services firms added only 17 net employees in the first nine months of this fiscal year, while Nomura calls India the “ground zero” for assessing AI’s effect on jobs.
背後
企業以停止校園招聘代替裁員,招聘放緩先於職位消失。分析指初級程式與測試工序被自動化,資深判斷型職位需求仍在。Firms are freezing campus hiring rather than cutting staff, so recruitment slows before jobs disappear. Analysts say routine coding and testing are being automated while demand for senior judgement holds.
顧問觀點

我們認為,淨增十七人這類數字比失業率更早反映結構變化。職位總數只是存量,入口收窄要到三五年後才顯現為經驗斷層。We read a figure like 17 net hires as an earlier signal than the unemployment rate. Headcount is a stock; a narrowing entry point only shows up as an experience gap three to five years later.

對日常工作的影響
HR 要把校園招聘列為年度指標而非可削項;用人主管須為初級員工設定半年內可交付成果,否則削減招聘永遠最省事。HR should treat campus hiring as an annual target rather than a soft line item, and line managers should define what a junior hire must deliver within six months.
如何改善
本週把本年校園招聘名額與實際入職數並列,如名額連續兩季未兌現,要求用人部門書面說明。This week, set this year’s campus hiring quota beside actual joiners; if the quota has missed two quarters running, require the hiring department to explain in writing.

Anthropic 開放自託管 Claude Managed Agents 私人測試Anthropic opens self-hosted Claude Managed Agents in private beta

影響力✓ 已核實Verified6–12 個月6–12 months資訊科技IT法務合規Legal & compliance
事件
Anthropic 於十一月十二日開放自託管代理服務私人測試:伺服器、資料庫與沙箱置於客戶自有雲或機房,會話紀錄客戶自存Anthropic opened a private beta on 12 November for self-hosted Claude Managed Agents: server, database, workers and sandboxes can sit in the customer’s own cloud or on-premise, session history is stored in the customer’s database, and model requests can go to Anthropic, Bedrock, Google Cloud or the customer’s gateway.
背後
金融與醫療機構一直卡在同一問題:代理要動憑證與內部系統,紀錄卻留在供應商機房。自託管把控制面移入客戶邊界,補丁自負。Banks and hospitals were stuck on one point: agents need credentials and internal systems, yet the records lived in the vendor’s data centre. Self-hosting moves the control plane inside the customer perimeter, at the cost of owning patching, backup and access control.
顧問觀點

我們認為,這類方案把問題由「合規可否」變成「人手是否足夠」。同一個稽核要求,過去是供應商提供報告,現在是企業自己產出報告,難度其實更高。In our view this shifts the question from whether compliance is possible to whether staffing is sufficient. The same audit requirement that a vendor used to satisfy with a report must now be produced in-house, which is harder.

對日常工作的影響
IT 要為自託管代理加開備份、修補與日誌保留三條值班責任;法務則要重新確認零數據保留條款是否同樣覆蓋自託管路徑的請求。IT must assign on-call duties for agent backup, patching and log retention, while legal must reconfirm that zero-retention terms cover requests made through the self-hosted path.
如何改善
本週在代理權限清單中加一欄「日誌存放位置」,列出每個代理的會話紀錄實際落在哪個系統,標明由誰負責保留期限。This week add a log-location column to your agent permission list, showing which system holds each agent’s session records and who owns the retention period.

代理在封閉評測中寫出可用零日漏洞,最長補丁窗縮至六日Agents write working zero-days in sealed tests as the patch window shrinks to six days

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
安全業界匯總指出,谷歌披露一宗模型生成的零日利用,繞過某開源工具兩步驗證;Anthropic 稱一月內見代理群攻擊五十家Security reporting says Google disclosed a model-generated zero-day exploit that bypassed two-factor authentication in an open-source administration tool and was intended for an active campaign, while Anthropic observed agent swarms targeting about 50 organisations and generating more than a dozen possible zero-days in a month.
背後
同期微軟十一月修補一千二百五十九個漏洞,連續第二個月破千。修補追不上生成速度,日曆式打補丁失效,曝光管理須持續進行。In the same period Microsoft patched 1,259 vulnerabilities, the second consecutive month above a thousand. Patching cannot keep pace with generation, so calendar-based maintenance fails and exposure management has to become continuous.
顧問觀點

我們認為,企業的風險重心已由「有沒有漏洞」轉向「有多少東西暴露在互聯網上」。減少可達面比重複掃描更有用,尤其是那些為了方便而長期開放的舊介面。In our view the risk focus has moved from whether vulnerabilities exist to how much is reachable from the internet. Reducing exposed surface beats repeated scanning, especially for legacy interfaces left open for convenience.

對日常工作的影響
IT 須每月盤點一次對外可達的系統清單並逐項問「是否仍需要開放」;代理與編碼工具的權限則要收窄到任務所需的最小範圍。IT should review externally reachable systems monthly and ask of each whether it still needs to be open, while narrowing the permissions of agents and coding tools to the minimum each task requires.
如何改善
本週列出對外可達的系統清單,標出其中三個最少使用卻仍開放者,先關閉或加上來源限制,再排期處理其餘。This week list externally reachable systems, mark the three least used yet still open, close them or restrict their sources, then schedule the rest.

本週可做的三件事

  1. 把本季招聘計劃中三個初級崗位改寫成「見習加驗收」設計,由資深同事簽核產出,先保住新人入口。Rewrite three entry-level roles in this quarter’s hiring plan as apprentice-plus-sign-off positions, with a senior colleague approving the output, so the graduate entrance stays open.
  2. 用一款開源權重模型在本機或私有雲跑一條內部流程,記錄每千次調用的實際成本,再與現用雲端方案對比。Run one internal workflow on an open-weight model locally or in a private cloud, log the real cost per thousand calls, and compare it with your current hosted option.
  3. 為已上線的代理立一頁權限清單:列明它能觸及的系統、憑證與可寫入的資料,交 IT 與法務各簽一次。Write a one-page permission sheet for every agent already in production: which systems and credentials it can reach and which data it can write, signed once by IT and once by legal.

常見問題

今日(2026-11-16)「AiX 環球 AI 情報」有哪十條重點?

今日十條為:1、美團開源一點六萬億參數代理模型權重,訓練全程不用英偉達芯片;2、OpenAI 解散獨立機器人團隊,稱通用模型已可直接控制機械;3、香港推萬個畢業生職位:文書職位過剩下的止血措施;4、美聯儲調查:八成七受訪者視 AI 為金融穩定最大風險;5、四非牟利機構入稟,要求凍結前沿模型研發;6、國家數據局定「一條主線、四個環節、一個閉環」,推動高質量數據集;7、商務部修訂技術出口目錄,AI 服務出海不在管制範圍;8、印度 IT 招聘:本財年首九月淨增十七人,初級職位入口收窄;9、Anthropic 開放自託管 Claude Managed Agents 私人測試;10、代理在封閉評測中寫出可用零日漏洞,最長補丁窗縮至六日。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?

今日三線同時收窄入口:美團以麻省理工許可開源一點六萬億參數代理模型,訓練全程不用英偉達芯片;OpenAI 解散機械人團隊,改由通用模型直接驅動;香港以一萬個畢業生職位,接住文書職位過剩衝擊。模型供給更寬,人的入口反而更窄。

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:Hugging Face、GPTOnline(引 The Decoder)、PANews、Pivot News(引 Fortune)、Robotico、鳳凰網、香港01、Board of Governors of the Federal Reserve System),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。

每週一封 · 可隨時取消 · 私隱聲明