2026年5月19日，Google I/O大會的主舞台燈光亮起，工程師們宣布Gemini 3.5 Flash正式推出，同步發布的還有一個名叫Gemini Spark的裝置端代理人——延遲低於50毫秒，可離線運行，24小時待命。那個週末，「萬物皆代理人」（Everything becomes an Agent）成為科技媒體的標題關鍵字。
同一週，Meta在執行層面上也完成了屬於自己的宣言：裁員8,000人，約佔全球員工總數的10%，同時強制將逾7,000名員工調任至新成立的AI部門。Meta CTO Andrew Bosworth對此的解釋是：「AI代理人類工作，人類角色將專注於方向、審查與改進。」
兩件事不是矛盾，是同一個帳單的正反兩面。
宣言的成本，由誰來付
Meta這一輪重組的新單位包括負責AI雲端基礎設施的Massive Scale、用於軟體開發的內部AI代理人Hatch、Customer Experience AI，以及Meta AI。被替換的，是創意、行銷、內容創作等白領職位——這些職位的共同點是：可以被描述成工作流程、可以被拆解成步驟、可以被代理。Meta同步引進了名為「Model Capability Initiative」的監控工具，記錄員工電腦活動，用以訓練AI模型。人類在離開之前，還要先幫AI上一堂課。
2026年Meta的AI基礎設施投資預計達1,450億美元。裁員省下的人力成本，直接流向算力採購。這不是轉型，是資本重新分配的會計項目。
Google那邊則更像一場設計展。ADK 1.0多語言代理人框架讓開發者用單一API call就能啟動代理人；Jules負責非同步程式碼任務；Antigravity 2.0是代理人開發平台；Managed Agents API對外開放。Google宣示進入「代理人工作流時代」，把整套基礎設施攤開來給全世界的開發者用——門檻越低，替代發生的速度越快。
台灣造了刀，但不決定刀砍向哪裡
讓這一切成為可能的晶片，大多數在新竹和台南的無塵室裡製造出來。台積電控制全球72%的晶片市場，製造全球90%以上的先進AI訓練晶片，NVIDIA、Google、Apple都在這條供應鏈上。Gemini Spark的50毫秒推論、Meta Massive Scale的算力密度，物理上都有一部分來自台灣的光罩和蝕刻製程。
台灣也不是沒有試著往上走。AI Academy在2018年成立，四年內培訓超過7,000名AI人才。台灣AI行動計畫第一期投入NT$100.7億，第二期跟進NT$200億。行政院確認以半導體優勢為基礎打造AI強國。這些數字夠大，但它們描述的主要是供應鏈端的強化，而不是「誰決定AI代理什麼」這個層次的介入。
台灣是基礎設施的建造者。代理人的發號施令者在山景城和門洛帕克。這個位置不是一夜之間形成的，但「萬物皆代理人」這個宣言讓這個位置變得無法繼續模糊。
插畫家的問題，不是哲學問題
身為一個用手畫圖的人，我對「代理」這個詞沒有任何浪漫化的空間。Google I/O同週發布的Veo 3是視覺生成工具；Gemini Omni處理任意輸入輸出的多模態任務。這些工具不是未來的威脅，是現在競稿時已經存在的競爭對手。
Meta裁掉的那8,000人裡，有多少是設計師、有多少是內容創作者，財報沒有逐條列出。但「創意、行銷、內容創作等白領工作」這個描述，幾乎可以精確框住所有我認識的同行。Bosworth說人類未來的角色是「方向、審查與改進」。這句話說得很體面。實際上，那是一個人監管十個代理人的工作描述，不是一份創作者的職位說明。
台灣那7,000名AI人才在學什麼？他們訓練的模型將被用來代理誰？這兩個問題的答案，決定了台灣在這張帳單上是收款方還是付款方。
Google I/O的宣言和Meta的裁員公告壓縮在同一個新聞週期裡——這不是巧合，是同步。技術宣言需要財務空間，財務空間從人力成本裡擠出來。帳單已經寄出，地址欄寫得很清楚。
— 蔡子涵
延伸閱讀
The Week Google Said Everything Is an Agent
On May 19, 2026, Google I/O’s main stage announced Gemini 3.5 Flash and a device-side agent called Gemini Spark — under 50ms latency, offline-capable, always on. By that weekend, “Everything becomes an Agent” was the headline keyword across tech media.
The same week, Meta completed its own announcement at the operational level: 8,000 employees laid off — approximately 10% of its global workforce — while over 7,000 others were forcibly transferred into newly formed AI divisions. Meta CTO Andrew Bosworth offered the official framing: “AI agents human work, and humans will focus on direction, review, and improvement.”
These two events are not contradictions. They are the front and back of the same invoice.
Who Pays for the Manifesto
Meta’s restructured units include Massive Scale for AI cloud infrastructure, an internal AI coding agent called Hatch, Customer Experience AI, and Meta AI. The positions displaced — creative, marketing, content production — share one quality: they can be described as workflows, broken into steps, and delegated. Meta also introduced a monitoring tool called “Model Capability Initiative,” logging employee computer activity to train AI models. Employees were asked to teach the system that would replace them before they left.
Meta’s projected AI infrastructure investment for 2026 is $145 billion. The headcount savings flow directly toward compute procurement. This is not transformation. It is a capital reallocation line item.
Google’s side looked more like a product showcase. ADK 1.0 lets developers spin up agents with a single API call. Jules handles asynchronous coding tasks. Antigravity 2.0 is the agent development platform. Managed Agents API is open to the public. The lower the barrier, the faster displacement happens.
Taiwan Built the Knife
The chips that make all of this physically possible are largely fabricated in cleanrooms in Hsinchu and Tainan. TSMC controls 72% of the global chip market and manufactures over 90% of advanced AI training chips — for NVIDIA, Google, and Apple. Gemini Spark’s sub-50ms inference and Meta’s compute density both trace, in part, to Taiwanese photolithography and etching processes.
Taiwan has not been passive. The AI Academy was founded in 2018 and trained over 7,000 AI professionals within four years. The first phase of Taiwan’s AI Action Plan committed NT$100.7 billion; the second phase followed with NT$200 billion. The Executive Yuan has confirmed the semiconductor advantage as the foundation for an AI power strategy.
These are real investments. But they describe supply-chain reinforcement, not participation in decisions about what AI agents do or who they replace. The architects of “Everything becomes an Agent” are in Mountain View and Menlo Park. Taiwan built the infrastructure. It is not sitting at the table where the delegations are decided.
An Illustrator’s Question Is Not a Philosophy Question
As someone who draws for a living, I have no room to romanticize the word “agent.” Veo 3, Google’s visual generation tool released the same week as I/O, is not a future threat — it is already a competitor in current project pitches. Gemini Omni handles arbitrary multimodal inputs and outputs.
Of the 8,000 Meta employees cut, the breakdown by role was not itemized in public disclosures. But the phrase “creative, marketing, and content production white-collar roles” maps almost exactly onto every peer I know. Bosworth’s formulation — humans focusing on “direction, review, and improvement” — sounds measured. In practice, it describes one person supervising ten agents, not a creative professional’s job description.
What are Taiwan’s 7,000 AI-trained professionals learning to build? Whose work will the models they train be used to replace? The answers to those two questions determine whether Taiwan is on the receiving or paying end of this invoice.
Google’s manifesto and Meta’s layoff announcement landed in the same news cycle. That is not coincidence — it is synchronization. Technical declarations require financial headroom. Financial headroom is extracted from labor costs. The bill has been sent. The address line is filled in.
— 蔡子涵
Related Posts
https://justfly.idv.tw/s/LJCcEJD