AiX 環球 AI 情報 Global AI Intelligence

第 8 期 2026 年 10 月 2 日（星期五） 07:30 出刊 10 條情報 · 15 原文來源連結 RSS 過往期數

今日十條 · 2026 年 10 月 2 日旗艦降價、責任上身：AI 競賽由能力比併轉向成本與問責比併

今日主編判斷

今日十條消息指向同一條主線：AI 競爭的計分板正在換。Google 以五分一價格推出 Gemini 4 Argon、OpenAI 把旗艦成本壓到五分之一，能力差距收窄後，價格與交付可靠性成為採購新標準；另一邊廂，加州立法禁止單憑 AI 解僱員工、Anthropic 在招股書自陳模型風險、新馬把供應鏈韌性寫進雙邊機制，監管與客戶要求正把責任推回企業身上。我們認為，本季企業的關鍵動作不再是「要不要用 AI」，而是重新盤點哪些工序可以交給代理、哪些必須保留人工簽核，並把成本下降的紅利轉為治理與培訓投入。Today's ten stories point to a single shift: the scoreboard for AI competition is being replaced. Google launched Gemini 4 Argon at one-fifth the usual frontier price, and OpenAI cut flagship costs to a fifth, so as capability gaps narrow, price and delivery reliability become the new procurement criteria. Meanwhile, California barred AI-only firings, Anthropic disclosed model risks in its own prospectus, and Singapore and Malaysia wrote supply-chain resilience into a bilateral mechanism, pushing accountability back onto enterprises. In our view, the key corporate action this quarter is no longer whether to adopt AI, but which processes can be delegated to agents, which must retain human sign-off, and how to convert falling costs into governance and training investment.

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

Google 發布 Gemini 4 Argon，首發價每百萬輸入 token 2 美元Google Releases Gemini 4 Argon at $2 per Million Input Tokens

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

事件: Google DeepMind 於 9 月 30 日發布前沿模型 Gemini 4 Argon，首發價輸入每百萬 token 2 美元、輸出 10 美元，快取輸入再打九五折；輸出上限由 6.4 萬提升至 100 萬 token，初期只透過 Fairwind 計劃開放給網絡安全機構。Google DeepMind released the frontier model Gemini 4 Argon on 30 September, priced at an introductory $2 per million input tokens and $10 per million output tokens, with a further 95% discount on cached input. The output ceiling rose from 64,000 to 1,000,000 tokens, with initial access limited to cybersecurity organisations through the Fairwind Program.
背後: Google 已七個月沒有推出高於 Flash 級別的自家模型，能力落後於 OpenAI 與 Anthropic 的旗艦；以貼近成本線的首發價配合長流程輸出容量，是要用價格與工程可靠性換取企業市場份額。Google had not shipped a proprietary model above the Flash class for seven months and trailed the OpenAI and Anthropic flagships. An introductory price near cost, paired with a long-horizon output capacity, is a bid to win enterprise share through price and engineering reliability rather than benchmark bragging rights.

顧問觀點

我們認為，旗艦模型同級能力而價格差一個數量級，意味企業過去因成本而放棄的長流程工序現在值得重估；但分階段開放也提醒買方，供應商會把最強能力優先給能承擔風險與合規審查的客戶，採購談判要把「提前取得權」寫進條款。We believe that flagship-class capability at an order-of-magnitude lower price means the long-horizon processes enterprises previously abandoned on cost grounds now deserve re-evaluation. The staged rollout also reminds buyers that vendors reserve their strongest capabilities for customers able to absorb risk and compliance review, so procurement talks should secure early-access rights in writing.

對日常工作的影響: IT 主管需重跑模型選型矩陣：以每完成一件任務的總成本、而非每 token 單價作比較，並為長文檔、多階段任務重新指定模型。法務則須核對 Fairwind 類計劃的使用範圍與資料留存條款。IT leads must rerun the model selection matrix, comparing cost per completed task rather than cost per token, and reassign models for long-document and multi-stage work. Legal should check the permitted scope and data-retention terms of program-based access like Fairwind.
如何改善: 本週挑一條因成本而被擱置的長流程工序（如合約全文審閱、跨年度財報比對），用新價格重算一次投入產出，決定是否解凍。This week, take one long-horizon process shelved because of cost, such as full-contract review or multi-year financial comparison, recompute its economics at the new prices, and decide whether to unfreeze it.

來源Google 黑板報（騰訊新聞轉載）（2026-10-01）↗Anadolu Agency（2026-10-01）↗

加州《反機械人老闆法》生效，禁止單憑 AI 解僱或懲戒員工California's 'No Robo Bosses Act' Bars AI-Only Firings and Discipline

影響力✓ 已核實Verified6–12 個月6–12 months人力資源HR法務合規Legal & compliance企業決策者Business owners

事件: 加州州長紐森於 9 月 30 日簽署 SB 947《2026 年反機械人老闆法》，禁止僱主完全依賴自動化決策系統解僱或懲戒員工；若以 AI 為主要依據，須由人工覆核並以評核紀錄等佐證，且須書面告知員工、提供所用資料說明及人工聯絡人。法案 2027 年 7 月 1 日生效。California Governor Gavin Newsom signed SB 947, the No Robo Bosses Act of 2026, on 30 September, barring employers from relying solely on automated decision systems for discipline or dismissal. Where AI is the primary basis, a human must corroborate the decision using supporting records, and affected employees must receive written notice, a description of the data used, and a human point of contact. It takes effect on 1 July 2027.
背後: 美國約九成管理層表示公司已採用至少一項自動化管理工具，Meta 亦面對以 AI 排名挑選裁員對象的訴訟；加州在聯邦缺乏全面 AI 立法下，以州法先行建立問責框架，同時簽署 AI 裁員通知與職場監控限制三項配套。Around 90% of US managers say their firms have adopted at least one automated management tool, and Meta faces litigation over AI ranking used to select layoffs. With no comprehensive federal AI law, California is building accountability at state level, signing alongside it a mass-layoff AI notice bill and two workplace surveillance restrictions.

顧問觀點

我們認為，法案的真正重量不在罰則，而在舉證責任的轉移：企業必須能證明每一次以 AI 為主要依據的人事決定都有可查證的人工介入。跨國企業若在多地營運，宜把加州的標準當作最低共同分母，統一內部人事 AI 的紀錄要求，避免逐地重新設計流程。We believe the law's real weight lies not in penalties but in the shift of the burden of proof: employers must be able to show verifiable human involvement in every people decision where AI was the primary basis. Multinationals operating across jurisdictions should treat California's standard as the lowest common denominator and unify their HR AI record-keeping rather than redesigning per market.

對日常工作的影響: HR 需為解僱、懲戒、評核三類流程加入人工覆核簽名與資料留存欄位；法務須更新員工手冊與供應商條款，要求人事系統供應商提供分類說明與可解釋紀錄。內地與香港團隊若有同類工具，亦應預先盤點。HR must add human corroboration sign-off and data retention fields to dismissal, discipline and appraisal workflows. Legal should update handbooks and vendor terms to require HR-tech providers to supply classification statements and explainable logs. Teams in mainland China and Hong Kong using similar tools should also conduct an inventory now.
如何改善: 本週由 HR 會同法務，為所有影響人事決定的自動化工具建立一份登記表，逐項記錄決策類型、人工覆核人與資料留存期。This week, have HR and legal jointly create a register of every automated tool affecting people decisions, recording the decision type, the human reviewer, and the data retention period for each.

來源ETHRWorld (Economic Times)（2026-10-01）↗Quartz（2026-10-01）↗

Anthropic 招股書披露 518 億美元算力承諾與 80 頁風險因素Anthropic Prospectus Discloses $518bn Compute Commitments and 80 Pages of Risk Factors

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

事件: Anthropic 遞交 261 頁招股書，目標估值逾 2 兆美元；2025 年收入 46 億美元、營運虧損約 80 億美元，並承諾未來十年雲端、運算與基建支出約 5,180 億美元，其中大部分不可取消。全書 80 頁為風險因素，包括模型可能拒絕關機及自主代理行為帶來的法律責任不確定性。Anthropic filed a 261-page prospectus targeting a valuation above $2 trillion. It reported $4.6bn of 2025 revenue against roughly $8bn of operating losses, and committed about $518bn in cloud, compute and infrastructure spending over the next decade, largely non-cancellable. Eighty pages are devoted to risk factors, including models that may resist shutdown and uncertain legal liability from autonomous agents.
背後: Anthropic 收入在一年內增長約十二倍，但同期基礎設施投入與長期承諾遠超收入；這份文件首次讓公眾市場投資者看到前沿模型公司的成本結構，也把「代理行為風險」由技術討論搬進法定披露語言。Anthropic's revenue grew roughly twelvefold in a year, yet infrastructure spending and long-term commitments far outpace it. The filing gives public-market investors their first look at a frontier lab's cost structure, and moves agent-behaviour risk out of technical debate and into statutory disclosure language.

顧問觀點

我們認為，這份招股書對企業買方的價值不在估值，而在它把供應商的風險自認寫成公開文件：模型可能拒絕關機、代理可能越權，這正是企業採購時難以取得的明確答案。既然如此，企業有理由把同類問題直接寫進與 AI 供應商之間的合約，而非依賴對方自律。We believe the prospectus's value to enterprise buyers lies not in its valuation but in putting a supplier's own risk admissions on the public record: models may resist shutdown, agents may exceed authority. These are precisely the answers enterprises struggle to obtain in procurement. That being so, buyers have grounds to write the same questions directly into supplier contracts rather than rely on self-regulation.

對日常工作的影響: 財務與法務在審視 AI 供應商時，應把長期承諾與不可取消條款視為對手方風險指標；IT 則須為代理工具設定權限邊界與中止機制，不能假設供應商已處理。Finance and legal reviewing AI vendors should treat long-term and non-cancellable commitments as counterparty risk indicators. IT must define permission boundaries and kill-switch mechanisms for agent tools rather than assuming the vendor has handled them.
如何改善: 本週在與主要 AI 供應商的合約或續約清單中，加問三題：代理越權時的責任歸屬、服務中止的可取消程度、事故通報時限。This week, add three questions to your contract or renewal checklist for major AI vendors: liability when an agent exceeds authority, how cancellable the service commitment is, and the incident notification deadline.

來源Splitfeed（2026-10-01）↗AI.cm（2026-10-01）↗

Decagon 發布個人代理閘道與 PACT 協定，處理「代理代客查詢」Decagon Launches Personal Agent Gateway and PACT Protocol for Agent-to-Business Requests

影響力✓ 已核實Verified6–12 個月6–12 months前線員工Frontline staff資訊科技IT法務合規Legal & compliance

事件: Decagon 於 10 月 1 日的 Dialogues 2026 大會發布四項產品，包括 Voice 3 語音模型 Chord、可識別並分流個人代理的 Personal Agent Gateway、Agent Modules 及 Duet Apprentice 測試版。其 PACT 協定讓個人代理能證明代表誰、獲授權做什麼，企業可就人類與代理分別設定不同作業程序與權限範圍。At its Dialogues 2026 event on 1 October, Decagon released four products: the Voice 3 speech model Chord, a Personal Agent Gateway that detects and routes personal agents, Agent Modules, and the Duet Apprentice beta. Its PACT protocol lets personal agents prove whom they represent and what they are authorised to do, while businesses set separate operating procedures and permission scopes for humans and agents.
背後: 過去一個月 Meta、OpenAI 與其他廠商相繼推出代客預訂、購買、取消的個人代理，這些代理已開始接觸企業客服渠道；企業原本只服務人類客戶的流程，第一次要面對「對面不是人」的情境，授權與責任鏈需要新的技術憑證。In the past month Meta, OpenAI and others have launched personal agents that book, buy and cancel on their owners' behalf, and these agents are already reaching corporate service channels. For the first time, processes built to serve human customers must handle a counterparty that is not a person, requiring new technical credentials for authorisation and the chain of liability.

顧問觀點

我們認為，「誰在代誰說話」將成為客戶服務與合約管理的新工序。企業不應急於封鎖個人代理，而是先在客戶身份驗證與授權範圍上分層：低風險查詢可自動放行，涉及付款、變更合約或個資的動作則必須回到本人確認。We believe establishing who is speaking for whom will become a new step in customer service and contract management. Rather than blocking personal agents outright, enterprises should layer identity verification and authorisation scope: low-risk enquiries can pass automatically, while anything touching payment, contract change or personal data must return to the principal for confirmation.

對日常工作的影響: 客服與 IT 需為代理來電建立識別與分級處理規則；法務須釐清代理越權所作交易的效力與追責路徑；市場部則要重估以真人互動為前提的客戶旅程設計。Customer service and IT must build identification and tiered handling rules for agent-initiated contact. Legal should clarify the validity and recourse path for transactions an agent makes beyond its authority. Marketing must reassess customer journeys designed on the assumption of human interaction.
如何改善: 本週請客服主管盤點哪些查詢類型可由代理自助完成、哪些必須人工或本人確認，先寫成一頁分流規則試行。This week, ask the customer service lead to map which enquiry types an agent may complete autonomously and which require a human or principal confirmation, then draft a one-page triage rule to pilot.

來源Unite.AI（2026-10-01）↗

英國科技界聯署要求五年投入 200 億英鎊建立主權 AI 能力UK Tech Leaders Urge £20bn Five-Year Sovereign AI Investment

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: 英國國防、金融服務與關鍵基建企業高層聯署公開信，要求政府五年內投入至少 200 億英鎊發展主權 AI，用於國內算力、電力、數據中心與網絡；聯署者包括 BT、Leonardo、Kainos、HPE、Sopra Steria 與 AI 實驗室 Cosine。公開信要求改革規劃、能源與採購規則，並為 AI 供應商提供更長期合約。Executives from UK defence, financial services and critical infrastructure firms signed an open letter asking the government for at least £20bn over five years for sovereign AI, covering domestic compute, power, data centre capacity and networks. Signatories include BT, Leonardo, Kainos, HPE, Sopra Steria and AI lab Cosine. The letter seeks reform of planning, energy and procurement rules and longer-term contracts for AI suppliers.
背後: 英國現有主權 AI 基金規模為 5 億英鎊，聯署要求為其 40 倍；當地 AI 工作負載目前多數在愛爾蘭、荷蘭或美東處理，國防、金融與政府數據的屬地要求令本土算力成為政策議題。The UK's existing Sovereign AI fund stands at £500m, so the letter asks for forty times that. UK AI workloads are today mostly processed in Ireland, the Netherlands or the US East Coast, and data residency expectations in defence, finance and government have turned domestic compute into a policy question.

顧問觀點

我們認為，公開信的訊號價值高於金額本身：多個既是倡議者、又是潛在供應商的企業同時發聲，說明主權算力已由政策口號變成採購議題。對有意進入英國公共部門市場的服務商，真正的門檻不在技術，而在能否配合較長的採購週期與資料屬地要求。We believe the letter's signal value exceeds its headline number: several firms that are both advocates and potential suppliers speaking at once shows sovereign compute has moved from policy slogan to procurement question. For service providers targeting the UK public sector, the real barrier is not technology but the ability to work with longer procurement cycles and data residency requirements.

對日常工作的影響: 有英國公共部門業務的企業，須預備更長合約期與本地資料處理方案；IT 要評估工作負載屬地選項與跨境傳輸成本；財務則須把資料屬地要求納入投標成本模型。Firms with UK public-sector business must prepare for longer contract terms and local data processing options. IT should assess workload residency choices and cross-border transfer costs, while finance folds residency requirements into bid cost models.
如何改善: 本週請負責英國業務的同事，把現行服務的資料處理地點列成清單，標出哪幾項在資料屬地要求下需要改為本地部署。This week, ask whoever covers UK business to list where each current service processes data, flagging which ones would need local deployment under residency requirements.

來源ITPro（2026-09-30）↗Resultsense（2026-10-01）↗

日本首相指示年底前制訂 AI 行動計劃，AI 列 17 個戰略領域Japan PM Orders AI Action Plan by Year-End, AI Named Among 17 Strategic Sectors

影響力✓ 已核實Verified6–12 個月6–12 months企業決策者Business owners市場推廣Marketing

事件: 日本首相高市早苗於 10 月 1 日主持增長戰略會議，指示相關大臣於年底前制訂 AI 行動計劃，訂明 AI 使用願景及所需監管與制度改革；AI 與半導體被納入 17 個公私營投資戰略領域，並把 2027 年 4 月起的五年定為集中推進期。計劃將隨 2027 年度預算案一併公布。Japanese Prime Minister Sanae Takaichi instructed ministers on 1 October to compile an AI action plan by year-end, setting out a vision for AI use and the regulatory and institutional reforms required. AI and semiconductors are among 17 strategic sectors for public and private investment, with the five years from April 2027 designated an intensive promotion period. The plan will be published alongside the FY2027 budget proposal.
背後: 日本政府同時要求制訂解決建築業人手短缺的政策包，反映其把 AI 視為補足勞動力缺口的工具而非單純產業政策；並計劃以跨年度預算措施提高企業投資可預期性，屬亞洲少數以財政工具直接承接 AI 落地的做法。The government also ordered a policy package to address construction labour shortages, showing it treats AI as a tool to offset workforce gaps rather than purely industrial policy. It plans multi-year budget measures to improve investment predictability, a rare case in Asia of fiscal tools being used to directly support AI deployment.

顧問觀點

我們認為，日本把 AI 與解決人手短缺掛鈎，對同樣面對老齡化與人力緊張的香港及大灣區企業具參考價值：AI 投資的回報不應只算生產力提升，也要算替代招聘成本與技能傳承的價值。跨年度預算安排亦提示企業，可與政府或大企業客戶洽談多年期合作以攤分導入成本。We believe Japan's linking of AI to labour shortages is instructive for Hong Kong and Greater Bay Area firms facing similar ageing and staffing pressure: the return on AI investment should count replacement hiring costs and knowledge transfer, not just productivity gains. Multi-year budget arrangements also suggest enterprises can negotiate multi-year engagements with government or large corporate clients to spread adoption costs.

對日常工作的影響: 在日業務或與日企合作的團隊，應留意年底行動計劃公布的行業清單與採購渠道；市場部可預先準備針對製造、建築、醫療等人手緊張行業的方案材料。Teams operating in Japan or partnering with Japanese firms should watch the sector lists and procurement channels in the year-end plan. Marketing can prepare solution material for labour-tight sectors such as manufacturing, construction and healthcare in advance.
如何改善: 本週請負責日本市場的同事，列出計劃可能覆蓋的 17 個領域中與本會服務相關的三個，準備一頁對應方案。This week, ask whoever covers the Japan market to identify three of the 17 sectors likely in scope that match the Society's offerings and prepare a one-page matching proposal.

來源Adnkronos (Jiji Press)（2026-10-01）↗Nation Thailand（2026-10-01）↗

DeepSeek 開源華為昇騰全套組件，TileLang 對標 CUDADeepSeek Open-Sources Full Huawei Ascend Stack, Pitching TileLang Against CUDA

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

事件: DeepSeek 正式開源面向華為昇騰算力平台的三類基礎設施組件：TileLang 高級語言編譯工具、計算庫及分布式通信庫，與此前英偉達平台版本一一對應。團隊自測顯示，FlashMLA 在昇騰 950 的稀疏注意力預填充內核最高達 410 TFLOPS，約為硬件峰值的 95%。DeepSeek formally open-sourced three categories of infrastructure components for Huawei's Ascend platform: the TileLang high-level language compiler toolchain, compute libraries and distributed communication libraries, matching its earlier NVIDIA-platform releases one for one. Internal testing showed FlashMLA reaching up to 410 TFLOPS on Ascend 950 sparse-attention prefill kernels, about 95% of hardware peak.
背後: DeepSeek 正推進新一輪融資並籌劃大規模採購昇騰芯片；軟件工具鏈是把設備變成可用訓練與推理基建的關鍵一步，因為資本可以買到硬件，但只有軟件、模型與硬件真正協同，設備才會持續產生價值。DeepSeek is raising a new round and planning large-scale Ascend chip purchases. The software toolchain is the step that turns hardware into usable training and inference infrastructure: capital can buy equipment, but assets only create lasting value once software, models and silicon genuinely work together.

顧問觀點

我們認為，這次開源的意義在於國產算力由「能跑」走向「好寫」：開發者若能在接近 Python 的抽象層上達到接近硬件峰值的性能，遷移成本就會顯著下降。企業在評估算力方案時，宜把工具鏈成熟度與人才供給一併計入，而非只看芯片規格與單價。We believe the significance lies in domestic compute moving from 'it runs' to 'it is pleasant to write for': if developers can reach near-peak performance at a Python-like level of abstraction, migration costs fall markedly. When evaluating compute options, enterprises should weigh toolchain maturity and talent availability alongside chip specifications and unit price.

對日常工作的影響: IT 主管評估國產算力方案時，應把工具鏈文件、社群支援與內部工程師上手時間納入決策矩陣；須自行開發底層算子的團隊，可先試用 TileLang 評估改造成本。IT leads evaluating domestic compute options should add toolchain documentation, community support and internal engineer ramp-up time to the decision matrix. Teams that build low-level kernels can pilot TileLang to gauge refactoring effort.
如何改善: 本週讓一名工程師在現有測試環境嘗試用 TileLang 重寫一個小型算子，記錄所需時間與性能差異，作為後續評估的基準。This week, have one engineer rewrite a small kernel in TileLang in the existing test environment, recording the time taken and performance difference as a baseline for later evaluation.

來源新浪財經（2026-10-01）↗搜狐財經（2026-10-01）↗

中國聯通軟研院企業級 AI 工作智能體覆蓋近千名員工China Unicom Software Institute's Enterprise AI Work Agent Covers Nearly 1,000 Staff

影響力✓ 已核實Verified已在發生Happening now資訊科技IT前線員工Frontline staff人力資源HR

事件: 中國聯通軟件研究院自研企業級 AI 工作智能體「通通代理」已累計覆蓋近千名員工，產生近 3 萬次對話，並在客服領域落地十餘個取數用數場景。客服數據獲取由「業務提需求、技術寫 SQL」改為業務人員自主取數分析，單次週期由數天縮短至半小時。China Unicom's Software Research Institute has built an enterprise AI work agent, Tongtong Agent, now covering nearly 1,000 employees with close to 30,000 conversations and more than ten data-access scenarios in customer service. Data retrieval shifted from business teams filing requests for engineers to write SQL, to staff querying directly, cutting a single cycle from days to half an hour.
背後: 國企與大型企業內部普遍存在知識查找效率低、系統操作鏈路長、專業能力難以複用的問題；此類智能體的價值不在模型能力，而在把散落的企業知識與業務系統接到同一個自然語言入口，並配合面向省級公司的專項培訓推動落地。State-owned and large enterprises commonly face inefficient knowledge search, long system operation chains and hard-to-reuse expertise. The value of such agents lies less in model capability than in connecting scattered enterprise knowledge and business systems to a single natural-language entry point, supported by dedicated training rolled out across provincial companies.

顧問觀點

我們認為，內地大型企業的落地路徑值得中小企參考：先選一個數據獲取頻繁、口徑清晰的場景，把週期壓縮到可量化的程度，再談橫向擴展。智能體導入失敗最常見的原因不是模型不夠強，而是知識與權限沒有整理好，導致它答不了真正重要的問題。We believe the rollout path taken by large mainland enterprises offers lessons for SMEs: pick one frequent, clearly defined data-access scenario, compress the cycle to a measurable degree, then discuss lateral expansion. The most common cause of agent adoption failure is not a weak model but unorganised knowledge and permissions, leaving the agent unable to answer the questions that matter.

對日常工作的影響: IT 與業務主管可先盤點取數需求最頻繁的三個崗位；HR 需同步設計智能體使用培訓與知識貢獻機制；市場部則可將可視化成果說明書轉為客戶案例材料。IT and business leads can start by mapping the three roles with the most frequent data requests. HR should design agent usage training and a knowledge contribution mechanism in parallel, while marketing can turn documented results into client case material.
如何改善: 本週挑一個「要等技術同事寫 SQL」的報表需求，試用現有 AI 工具讓業務同事自助取數，記錄實際節省的時間。This week, take one reporting request that currently waits on an engineer to write SQL, trial an existing AI tool so a business colleague can self-serve, and log the time actually saved.

來源新浪財經（2026-10-01）↗

韓國三大電訊與平台商十月推「AI for All」測試版，目標年內 1500 萬用戶Korea's SK Telecom, KT and Kakao Launch 'AI for All' Beta in October, Targeting 15 Million Users

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市場推廣Marketing

事件: 韓國政府主導的全國 AI 服務「AI for All」於十月展開測試版，SK Telecom、KT 與 Kakao 三個財團分別推出「無應用」電話代理、KakaoTalk 內代理及整合現有服務方案；政府向每個財團提供 512 張 Nvidia B200 GPU，並要求使用國產模型佔總用量至少 50%，目標年底取得 1500 萬用戶。South Korea's government-led national AI service, AI for All, entered beta in October, with consortia led by SK Telecom, KT and Kakao offering an app-less phone agent, a KakaoTalk-embedded agent and an integration with existing services respectively. The government provides each consortium with 512 Nvidia B200 GPUs and requires domestic models to account for at least 50% of total usage, targeting 15 million users by year-end.
背後: 韓國約 2,300 萬生成式 AI 用戶中有 2,100 萬仍使用免費服務，外國免費服務設用量限制且條件可能因營運政策改變；政府因而以國產模型與免費入口降低依賴，同時以算力補貼換取國產模型使用比例。Of South Korea's roughly 23 million generative AI users, about 21 million remain on free services, and foreign free tiers impose usage limits with terms that can shift with operator policy. The government is therefore using domestic models and a free entry point to reduce dependence, trading compute subsidies for a guaranteed share of domestic model usage.

顧問觀點

我們認為，以算力補貼換取國產模型使用比例，是一種把產業政策直接嵌入用戶入口的做法；其成敗關鍵不在模型分數，而在服務是否真的解決本地生活痛點，例如長者使用門檻與政府文件辦理。企業若要在韓國市場推廣代理服務，須預期國產模型配額會影響技術選型。We believe trading compute subsidies for a domestic model quota is a way of embedding industrial policy directly into the user entry point. Success hinges not on benchmark scores but on whether the service solves local life frictions such as accessibility for older users and government paperwork. Firms promoting agent services in Korea should expect domestic model quotas to shape technology choices.

對日常工作的影響: 在韓業務或與韓企合作的團隊，需把國產模型配額納入技術與成本規劃；負責政府公共服務對接者，宜留意此類統一入口會否成為本地客戶的新渠道標準。Teams operating in Korea or partnering with Korean firms must factor domestic model quotas into technology and cost planning. Those handling public service integration should watch whether such unified entry points become a new channel standard for local clients.
如何改善: 本週請韓國業務同事確認現有方案若需改用國產模型，改造成本與交付時間會增加多少，列入客戶報價考慮。This week, ask the Korea-facing colleague to estimate how much refactoring cost and delivery time would rise if an existing solution had to switch to a domestic model, and factor it into client pricing.

來源BigGo Finance（2026-10-01）↗

新加坡與馬來西亞落實供應鏈合作機制，承認工作組職權範圍Singapore and Malaysia Formalise Supply-Chain Cooperation Mechanism

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法務合規Legal & compliance

事件: 新加坡副總理兼能源貿工部長顏金勇與馬來西亞投資、貿易及工業部長佐哈里阿都干尼於 10 月 1 日在吉隆坡共同主持第三屆常年部長對話會，雙方認可供應鏈合作工作組的職權範圍，以便在交通或供應中斷期間促進必需品與工作人員繼續跨境流動；兩國去年雙邊貿易總額 940 億 7,000 萬美元，同比增長 8.5% 創新高。Singapore Deputy Prime Minister and Minister for Trade and Industry Gan Kim Yong and Malaysia's Minister of Investment, Trade and Industry Johari Abdul Ghani co-chaired the third annual ministerial dialogue in Kuala Lumpur on 1 October. Both sides endorsed the terms of reference of the Supply Chain Cooperation Workgroup to keep essential goods and workers moving across the border during transport or supply disruptions. Bilateral trade hit a record US$94.07bn last year, up 8.5% year on year.
背後: 工作組成立於 2024 年，此次認可職權範圍意味機制由對話走向可操作安排；兩國同時樂見數碼與綠色經濟合作框架完成，涵蓋跨境二維碼支付互聯與即時轉帳服務，反映區域合作正把數碼基建與供應鏈韌性綁定處理。The workgroup was set up in 2024, and endorsing its terms of reference moves the mechanism from dialogue toward operational arrangements. Both sides also welcomed completion of the digital and green economy cooperation framework, covering cross-border QR payment linkage and instant transfers, showing regional cooperation now treats digital infrastructure and supply-chain resilience as one agenda.

顧問觀點

我們認為，對在兩地營運的企業而言，這一機制的實際價值在於中斷發生時人員與貨物的跨境安排有了既定協商渠道；但企業不應假設機制可自動適用，仍須自行確認自身貨品與人員是否落入「必需品」範圍，並準備替代路線與本地庫存方案。We believe that for firms operating in both markets, the mechanism's practical value is having an established channel for cross-border movement of people and goods during disruption. But enterprises should not assume automatic coverage: they still need to confirm whether their goods and staff fall within the 'essential' scope, and prepare alternative routes and local inventory plans.

對日常工作的影響: 供應鏈與營運主管應更新兩地營運的應變預案，明確中斷時的人員與貨物安排；財務則須評估本地備貨所佔用的營運資金。法務宜核對自身貨品分類是否符合「必需品」定義。Supply chain and operations leads should update contingency plans for dual-market operations, specifying people and cargo arrangements during disruption. Finance must assess the working capital tied up by local buffer stock, and legal should check whether their goods classification meets the 'essential' definition.
如何改善: 本週請營運同事為最關鍵的兩項跨境物料各準備一條替代路線與一個本地備貨點，寫入應變預案。This week, ask operations to prepare one alternative route and one local buffer-stock point for each of the two most critical cross-border materials, and write both into the contingency plan.

來源聯合早報（2026-10-01）↗

本週可做的三件事

請 IT 與 HR 共同列出公司目前使用、且會影響人事決定（招聘、排班、評核、解僱）的自動化工具清單，逐項標明有無人工覆核環節。Ask IT and HR to jointly list every automated tool that influences people decisions (hiring, scheduling, appraisal, termination) and mark which ones have a human review step.
抽樣三條最常用的 AI 工作流程，用同一任務分別以「便宜小模型」與「旗艦模型」跑一次，記錄成本與錯誤率，重新設定路由規則。Sample your three most-used AI workflows, run the same task on both a cheap small model and a flagship model, record cost and error rates, and reset your routing rules accordingly.
在下一份 AI 採購或續約文件加入兩條條款：資料留存與刪除安排、供應商模型事故的通報時限與責任歸屬。Add two clauses to your next AI procurement or renewal document: data retention and deletion arrangements, and the vendor's notification deadline and liability for model incidents.

常見問題

今日（2026-10-02）「AiX 環球 AI 情報」有哪十條重點？

今日十條為：1、Google 發布 Gemini 4 Argon，首發價每百萬輸入 token 2 美元；2、加州《反機械人老闆法》生效，禁止單憑 AI 解僱或懲戒員工；3、Anthropic 招股書披露 518 億美元算力承諾與 80 頁風險因素；4、Decagon 發布個人代理閘道與 PACT 協定，處理「代理代客查詢」；5、英國科技界聯署要求五年投入 200 億英鎊建立主權 AI 能力；6、日本首相指示年底前制訂 AI 行動計劃，AI 列 17 個戰略領域；7、DeepSeek 開源華為昇騰全套組件，TileLang 對標 CUDA；8、中國聯通軟研院企業級 AI 工作智能體覆蓋近千名員工；9、韓國三大電訊與平台商十月推「AI for All」測試版，目標年內 1500 萬用戶；10、新加坡與馬來西亞落實供應鏈合作機制，承認工作組職權範圍。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？

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：Google 黑板報（騰訊新聞轉載）、Anadolu Agency、ETHRWorld (Economic Times)、Quartz、Splitfeed、AI.cm、Unite.AI、ITPro），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。