今日十條 · 2026 年 9 月 30 日代理常駐、邊界外移：企業開始接手交付責任的一天

我們認為 Dots 的真正變數不是能力，而是誰批准它做什麼。當代理由背景跨應用長期行動，權限設計與覆核責任會先於模型分數決定成敗，提示詞約束並不足夠。We think the decisive variable for Dots is not capability but who authorises what it may do. Once an agent runs in the background across applications, permission design and review ownership decide success before benchmark scores do, and prompt-level constraints are no longer sufficient.

我們認為重點是「指示清晰不等於邊界有效」。改寫指令後完成率由五十次中二十六次降至四十九次中四次，仍未歸零；高能力代理的範圍控制須靠外部強制，不能靠模型自律。We take the point to be that clear instructions are not the same as an enforceable boundary. After AISI clarified scope, full attacks fell from 26 of 50 runs to 4 of 49 but did not reach zero. Scope control for capable agents must be enforced externally, not left to model self-restraint.

我們認為這是今年最具體的架構轉向：安全由提示詞加護欄，變成矽與權限，等於承認模型層保證不可靠。但邊界準確與否，仍取決於企業自己填寫的權限表。We see this as the year's most concrete architectural shift: safety moving from prompts and guardrails to silicon and permissions, an admission that model-level assurances are insufficient. Whether the boundary is correct still depends on the permission table the enterprise itself fills in.

我們認為中階模型已足以承接大部分企業工序，旗艦模型應留給少數高價值判斷。企業若仍以單一旗艦模型處理所有請求，成本結構會不合理，亦缺乏可比較的替代方案。We think mid-tier models can now carry most enterprise workloads, with frontier models reserved for a narrow set of high-value judgements. Firms still routing every request to a single flagship model will find their cost structure hard to defend and their options hard to compare.

我們認為最值得留意的是分發而非產品。Meta 擁有數億商業帳戶，卻沒有既成的企業銷售隊伍，能否把消費端動能轉為合約收入，取決於通路與合規，而非模型分數。We think distribution matters more here than product. Meta has hundreds of millions of business accounts but no established enterprise sales force, so converting consumer momentum into contracted revenue will depend on channels and compliance rather than benchmark scores.

我們認為上訴主要爭取時間，而非改變方向。開放互通已是既定的監管路線，企業的準備工作不應押在訴訟結果上，提前設計資料可攜與助理可替換性，成本更低。We read the appeal as buying time rather than reversing direction. Interoperability is already the settled regulatory path, so enterprise preparations should not be wagered on the litigation. Designing for data portability and replaceable assistants now is cheaper.

我們認為資金增速快於場景落地速度，是今年內地市場的主要張力。當政策與資本同時指向落地，企業的評分標準會由「有否試用 AI」轉向「有否計得出回報」。We think the main tension in the mainland market this year is that capital is growing faster than deployment. With policy and funding both pointing at deployment, the standard applied to firms will shift from whether they have tried AI to whether they can show a return.

我們認為央企的資源配置是最清晰的政策風向標。當資金與考核同時移向前沿賽道，中小企的可行位置是配套與落地服務，而非在同一賽道正面競爭。We regard central SOE capital allocation as the clearest policy signal available. When funding and performance assessment both move towards frontier tracks, the workable position for smaller firms is supplying and implementing, not competing head-on in the same track.

我們認為關鍵訊號是由賣零件轉為入股基建。當芯片、散熱、儲能與保險公司以同一筆投資進場，AI 資料中心的競爭已由單項產品轉為整套交付能力。We take the signal to be a shift from selling components to holding stakes in infrastructure. When semiconductors, cooling, energy storage and insurers enter through one investment, competition in AI data centres moves from single products to whole-stack delivery capability.

我們認為馬來西亞選擇先立框架、後追技術，與區內部分市場的觀望態度不同。對在當地營運的企業而言，個人資料處理與智能眼鏡等新裝置會先受壓，宜提早整理清單。We read Malaysia as choosing to legislate first and catch up technically, unlike some regional markets that are still waiting. For firms operating there, personal data handling and newer devices such as smart glasses will feel the pressure first, so a data inventory is worth preparing now.