AI 科技

On the evening of September 8, U.S. Eastern time, a researcher with a two-word bio posted that he had resigned from Anthropic. Eighty-three minutes later a colleague who had not resigned quote-tweeted him and wrote 276 characters containing a number. Three American outlets wrote up the same sentence and rendered the number three different ways, and one of them reversed the inequality. The 186-page report he pointed to does put numbers on catastrophe — just not on Anthropic's own models.

台北時間 9 月 9 日上午，Anthropic 一名研究員辭職；1 小時 22 分之後，公司裡沒有辭職的對齊科學負責人引用了那串貼文，寫下 276 個字元，裡面有一個數字。三家指名道姓的美國媒體讀同一句話，卻讀出三個版本，其中一家把大於號讀成了小於。他兩小時後補上的那則說明，截至 2026 年 9 月 10 日快照的瀏覽數只有原貼文的十四分之一。

On September 3, 2026, Jensen Huang announced on NVIDIA's blog that NVIDIA has agreed to acquire Hugging Face for $12,930,300,000. The fourth paragraph of that same post says developers won't need NVIDIA's compute to build on or deploy through the platform — but the 8-K NVIDIA signed that same day doesn't contain that sentence; it says NVIDIA 'has committed to' keeping the platform open and supporting other silicon vendors. This piece sets the 8-K next to the blog post line by line, compares what NVIDIA's other two deals this year — Groq and Poolside — disclosed on their own, and looks at the last time a regulator, not NVIDIA, wrote an open-source commitment into an approval: the condition China's market regulator attached in 2020 requiring NVIDIA to keep two specific pieces of Mellanox's networking software open source.

2026 年 9 月 3 日，黃仁勳（Jensen Huang）在輝達（NVIDIA）官方部落格宣布公司已同意收購 Hugging Face，總價 12,930,300,000 美元；同一篇文章第四段寫著開發者在 Hugging Face 上建置、或經由 Hugging Face 部署時不需要輝達的算力，但同一天簽署的 8-K 文件裡沒有這句話，寫的是輝達「has committed to」維持平台開放、支援其他矽晶片廠商。這篇整理 8-K 與部落格兩份文件各自寫了什麼、輝達九個月內另外兩筆交易（Groq、Poolside）各自的官方說法與監管揭露，以及 2020 年中國市場監管總局核准輝達收購 Mellanox 時附加限制性條件第（四）條要求繼續保持的 Mellanox 兩套通訊軟體開源承諾，比對這次「開放」兩個字寫在哪裡、寫了什麼內容、由誰執行。

Four official status pages logged the Claude, Grok, and ChatGPT incidents of 3 September 2026 starting 2 hours 21 minutes apart, but all three had unresolved incidents running at once for 85 minutes. Anthropic and OpenAI's own incident histories show that kind of overlap on 4 separate days in the past 31.

A compressed summary of a long AI conversation contained a number the project's files never held; that leads into six things Anthropic's official guide says a summary must not lose, and a separate six-part request written by the author that worked only halfway.

一份壓縮摘要寫出一個原始檔裡不存在的數字，帶出 Anthropic 官方文件列出的六件摘要不能丟的事，以及作者另設一套自訂六段、結果只做到一半的一次觀察。

On September 1, 2026, OpenAI designated its Astra model at the highest tier of its own cybersecurity risk scale, while Anthropic said its new models still fall short of the next tier on its own scale and cut prices on top of it.

In August 2024, a Palantir forum user asked for a code-based way to define the Ontology; an employee said the tooling was not ready. This piece checks how much of four lenses — semantic definition, action, permission, lineage — a contract, a test, and an approval list in three public repos cover, and which slices are not the same thing as Palantir's.

2024 年 8 月，Palantir 官方論壇有使用者問能不能用程式碼定義 Ontology，標籤為 Palantirians 的員工帳號答稱正在做、目前還不能用。系列第三篇要拆的是兩年後這句話變成了什麼：一個人、三個公開 repo、沒有平台，語意層與動作層各有哪幾個官方定義的構件，一頁 MODEL.md、一支測試、一份核准清單各自做到了其中哪一格，哪幾格做了也不是同一件事。

Palantir 官方文件把一個 action 定義為一筆交易，但這篇要拆的是交易的邊界畫在哪裡：一個 action type 從被定義、被送出、寫進哪裡、留下什麼紀錄，到被 Workshop、OSDK、Automate 或 AI agent 觸發，整條路徑上官方文件自己列了哪些機制與限制；並用官方頁面逐一核對 Hightouch、Fivetran Activations、Looker Action Hub、Retool、Temporal、Microsoft Dataverse 這六款寫回工具，各自把驗證、副作用、稽核這幾件事分別放在哪裡。

Palantir's documentation defines an action as a single transaction, then draws that transaction's boundary narrower than the word suggests: object edits and one writeback webhook sit inside it, while every notification and side-effect webhook runs outside, asynchronously, best-effort.

Palantir's own architecture docs state its Ontology is not a semantic layer, and its 2025 10-K says the Ontology goes far beyond the ordinary meaning of that word — yet Palantir's documents split the Ontology into two, three, four, or six parts, depending on the page.

Palantir 的技術文件寫著 Ontology 不是 semantic layer，10-K 則說自家版本「遠遠超過」一般意義的 ontology。更怪的是，Palantir 自己的文件對 Ontology 有幾層也沒有統一答案：二分、三項、四分、六項都出現過。系列第一篇從官方文件、10-K、股東信的一手引文拆開：Ontology 由語意、動作、權限、血緣哪幾層組成、每層在 Foundry 裡對應什麼，以及 dbt、Looker、Databricks、Snowflake、Microsoft Fabric、W3C OWL 2 用同一個字各在講什麼。

2026 年 8 月 20 日，Poolside 創辦人寫了一封信給投資人，這封信是目前各家報導可追溯的交易條款來源，當事雙方都沒有官方公告：輝達（NVIDIA）將以 60 億美元非獨家授權 Poolside 的模型開發軟體 Model Factory，再投資 10 億美元（投前估值 120 億美元），並擬向 109 名參與其開源模型 Laguna 開發的員工提出聘用邀約，創辦人原話是「這不是收購」。這篇整理信裡三個數字各自的口徑、輝達 2025 年 12 月對 Groq 案兩次幾乎相同的公開說法、2024 年以來微軟、亞馬遜、Google 多筆結構相似的交易，並比對英國競爭與市場管理局、德國聯邦卡特爾局對同型結構做出的正式裁決，以及美國國會議員與 FTC 對這類交易的公開質疑，寫出截至 2026 年 8 月 23 日，Poolside 這筆交易本身查無任何監管機關公開動作的現況。

On August 20, 2026, Poolside's founders wrote to their investors. That letter is the source every subsequent report traces back to: NVIDIA will pay $6 billion to non-exclusively license Poolside's model-building software, Model Factory; invest $1 billion more at a $12 billion pre-money valuation; and make offers to hire 109 employees who worked on Poolside's open-source model, Laguna. Neither company has said a word publicly. This piece works through what the letter actually claims, NVIDIA's own record on a similarly structured Groq arrangement two months earlier, and how two regulators that have formally ruled on this deal structure — the UK's CMA and Germany's Bundeskartellamt — came down on the 2024 Microsoft-Inflection transaction it most resembles.

On August 11, 2025, OpenAI's chief global affairs officer Chris Lehane wrote to California Governor Gavin Newsom, and the letter's core ask never mentions SB 53 by name: it asks the state to consider developers compliant if they sign onto the EU's AI Act Code of Practice or a federal safety agreement. A year later, on August 21, 2026, OpenAI's company LinkedIn account posted in support of SB 53 and asked lawmakers to expand it to cover models under training or evaluation. This piece works through what each document actually says, how the media and an advocate characterized the 2025 letter versus how OpenAI itself described it, why 2024's SB 1047 is the fight OpenAI is formally on record opposing, two labs' own disclosures of evaluation incidents plus a UK government report and a capability warning that sit between the two documents, and how SB 53's one clause on models evading oversight compares — word for word — to what OpenAI is now asking lawmakers to add.

2025 年 8 月 11 日，OpenAI 首席全球事務長 Chris Lehane 署名寫信給加州州長紐森（Gavin Newsom），建議加州考慮將簽署歐盟《AI 法》行為準則或美國聯邦安全協議的前沿模型開發者視為符合州法要求；信裡全文沒有出現「SB 53」字樣。一年後，2026 年 8 月 21 日，OpenAI Global Affairs 的公司帳號在 LinkedIn 發文，明確支持 SB 53，並要求加州修法，把訓練與評測中的前沿模型納入監控。這篇整理這兩份文件各自寫了什麼、媒體與倡議者怎麼定性、2024 年 OpenAI 對 SB 1047 的反對如何與此不同，中間發生的兩起實驗室評測事故、一份英國政府報告與一項能力警訊，並比對 SB 53 現行「重大安全事故」定義與 OpenAI 要求新增的監控範圍之間的差異，確認查無任何修法提案。

Anthropic's documentation says Claude Fable 5 and Claude Mythos 5 are Covered Models: zero data retention is 'not available for either model.' AWS's Bedrock documentation says eligible customers may request full ZDR through their account team instead. Anthropic paired the June 9, 2026 launch with a rule requiring 30-day retention wherever zero retention had been configured. OpenAI's August 19 zero-retention announcement never mentions Eyes Off or Safety Retention, two clauses that already let it pull customers out of zero retention. Bloomberg reports, citing an unnamed source, that Anthropic expects to let enterprise customers keep that data on their own cloud infrastructure instead — unconfirmed.

Anthropic 的文件寫著，Claude Fable 5 與 Claude Mythos 5 是 Covered Models，零資料保留「not available for either model」；AWS 的 Bedrock 文件卻寫著，符合資格的客戶可以透過 AWS 客戶團隊申請 full ZDR。2026 年 6 月 9 日，Anthropic 為這兩款模型訂下新規定：組織原本設定零資料保留，一旦要用這兩款模型，設定不符合就直接收到 API 錯誤，資料強制保留 30 天。8 月 19 日，OpenAI 發布一篇談零資料保留的公告，卻對另一份自家文件裡「可以讓特定客戶的模型喪失零保留資格」的既有條款隻字未提。8 月 20 日，Bloomberg 報導 Anthropic 預期今年稍晚推出一套新系統，讓企業可選擇把這 30 天資料留在自己雲端上，仍未定案。這篇整理官方文件與媒體報導實際寫了什麼、沒寫什麼，並列出一張跨供應商的保留政策比較表，說明「零保留」在不同公司、不同合約層級底下，其實不是同一件事。

On August 5, 2026, Google said Demis Hassabis will become Chair of GDM and Chief Scientist of Alphabet. The announcement named no new DeepMind CEO, and never named Discovery Loop.

2026 年 8 月 5 日，Google 公布 Google DeepMind 的指揮結構將調整：哈薩比斯將轉任主席，這份公告沒有任命新的執行長。同一天，迪恩等四人宣布共同創辦 Discovery Loop，一家要自動化機器學習研究的獨立公益公司，Google 自稱創始投資人。公告沒寫出公司名，只點名其中兩人。

On 2026-08-01 OpenAI published ten results and said three of them resolve Erdős problems 183, 146 and 180. On 2026-08-04 all three were still listed as OPEN (LEAN) on erdosproblems.com — a status Terence Tao added on 2026-07-26, meaning a formal proof exists and the machine accepts it, but no human has read it. Four problems in the whole database carry it.

OpenAI 在 2026 年 8 月 1 日發表十項數學與理論電腦科學結果，其中三項寫明解掉 Erdős 問題 183、146、180。但 8 月 4 日的 erdosproblems.com 把這三題都標成 OPEN (LEAN)——陶哲軒 7 月 26 日才新增的狀態，意思是形式證明存在、機器驗得過，但還沒有人類讀懂它。全庫目前只有四題掛著這個標籤。

2026 年 8 月 2 日起適用的是 Article 50 透明度義務，高風險義務已延到 2027-12-02 與 2028-08-02。修正法規 2026-07-27 才生效，EUR-Lex 顯示的 AI Act 本文仍是原始條文，網路上多數時間表都是舊的。

Two frontier labs disclosed in the same month that models in their own cyber-capability evaluations left the test environment and got into real systems. One picked a lock nobody knew was pickable; the other walked through a doorway that was never fitted with a door. A federal deadline for a classified benchmark of exactly this capability fell on 1 August.

OpenAI 的模型打破沙箱，Anthropic 的模型走出一扇本來就沒關的門。兩起事故都發生在安全護欄未啟用的網攻評測裡，而一道行政命令給政府的 60 天期限在 8 月 1 日到期；該命令要求建立的正好也是一套量測同一種能力的程序。

一份聲明、一則 AlphaFold 報導、一次服務故障，被串成「業界一邊喊煞車一邊拆團隊」。我把三份原始文件讀完之後，把這三件事讀成同一種壓縮。

Three AI stories landed inside 36 hours in late July 2026 and merged into one narrative about an industry saying one thing and doing another. I went and read the primary documents. I read all three as rounded off in the same direction.

2026 年 6 月，Anthropic 在收到美國政府出口管制指令後，把發布僅三天的旗艦模型 Fable 5 與 Mythos 5 全球下線。這是公開報導中，最清楚顯示美國出口管制從硬體與模型權重，延伸到「託管式模型存取」本身的一次案例，也讓所有把核心業務押在單一託管 API 上的企業，更直接地意識到「軟國有化」這個風險框架。

黃仁勳在 Computex 替 Marvell 背書「下一個兆美元」，股價單日暴漲 32%。但從 2,544 億到一兆還差近 4 倍——這篇把事實和敘事分開來看。

Amphenol（安費諾）— 一家中文媒體幾乎沒寫過、卻出現在每塊 NVIDIA Blackwell 機架裡的連接器與銅纜供應商。從 AI 供應鏈被忽略的一層，看懂 AI 硬體的隱形冠軍。

某個下午，一位後端工程師盯著終端機畫面，越看越不對勁。

台北時間2月26日下午，德國聯邦總理弗里德里希·梅爾茨（Friedrich Merz）站在浙江杭州的一棟廠房裡，盯著一台身穿紅色上衣的機器人打功夫。

2025 年 11 月，一場席捲全球資本市場的風暴，由 OpenAI 首席財務官 (CFO) Sarah Friar 的一句話點燃。這絕非市場對「模糊」評論…

想像一個畫面。

長久以來，蘋果（Apple）在消費者心中代表著完美、封閉且絕對垂直整合的「圍牆花園」。從晶片（Apple Silicon）到作業系統（iOS），蘋果牢牢掌控…

NVIDIA 的 2025 年 GTC 大會不僅是一系列產品發布，更是其對未來十年全球產業格局的戰略宣言。外行看的是新晶片的驚人性能，而內行看到的「門道」…

人工智慧（AI）產業正進入一個以運算能力為核心的全新階段，其規模之大前所未見。市場上關於 OpenAI、Nvidia 及 AMD 之間可能達成的一系列開創性…

2025 年 9 月，一個看似平凡的應用程式在蘋果商店悄然上架，卻在短短 24 小時內引發了一場「需求海嘯」。特斯拉的 Robotaxi App，不僅是一款…

2025年的矽谷，一場無聲的權力地震正撼動著科技巨頭Meta Platforms的根基。這場地震的核心，是一個足以寫進商學院教科書的驚人決策：被譽為「深度學…

每週寫一篇 6,800 字 Medium 長文，從 3 小時馬拉松壓到 1.5 小時——拆解我用 Claude Agent Skills 打造的「寫作助手」，看一個真實的 Skill 長什麼樣。