2026
8 篇文章讓搜尋爬蟲進來、訓練爬蟲出去:Cloudflare 把混合用途爬蟲拆成三種行為
Cloudflare 推出 Disallow AI Training,讓站長在保留搜尋收錄的同時拒絕 AI 訓練,並把爬蟲控制拆成 Search、Training、Agent 三類。
閱讀文章 ↗Bot Preference Sync:Cloudflare 把 AI 機器人政策寫回 robots.txt
Cloudflare 推出 Bot Preference Sync,把 AI 搜尋、代理、訓練的政策自動同步到 robots.txt。文章拆解運作方式、出版者預設,以及透明度要求。
閱讀文章 ↗General Intuition 募 3.2 億美元:用遊戲數據教 AI 直覺
新創 General Intuition 以 23 億美元估值完成 3.2 億美元融資,Khosla Ventures 領投。它用數億小時帶動作標籤的遊戲影片訓練單一世界模型,同一個模型能打遊戲、跑模擬,也能控制實體機器人。
閱讀文章 ↗Anthropic 加碼 Google 與 Broadcom:數十億美元運算投資背後的訊號
Anthropic 宣布與 Google 和 Broadcom 簽署多年期、數 GW 等級的 TPU 運算合約,預計 2027 年起陸續上線。本文從產品建構者角度,解析這筆投資對 Claude 生態系、客戶成本與 AI 基礎設施趨勢的意義。
閱讀文章 ↗樂天開源 Rakuten AI 3.0:6,710 億參數、GENIAC 打造的日本最大模型
2026 年 3 月 17 日,樂天以 Apache 2.0 釋出約 6,710 億參數的 MoE 模型 Rakuten AI 3.0——日本最大高效能 AI 模型,訓練獲經產省與 NEDO 的 GENIAC 計畫支持。本文解析其日語基準成績、每 token 約 400 億活躍參數的架構,以及開源 671B 的現實限制。
閱讀文章 ↗Meta 向 Google 租 TPU:自建雲巨頭回頭當對手的客戶
2026 年 2 月 26 日,路透社報導 Meta 簽下數十億美元多年期協議,租用 Google Cloud TPU 訓練下一代模型,並洽談購買數百萬顆 TPU 自行部署。本文解析交易動機與對 Nvidia 的衝擊。
閱讀文章 ↗OIST 研究:讓 AI「自言自語」再配上工作記憶,學得更快也更會舉一反三
沖繩科學技術大學院大學(OIST)1 月 28 日發表研究:在主動推論框架上加入多槽工作記憶與「內在語言」機制,讓 AI 在稀疏資料下學會泛化與多工切換,表現明顯提升,為不依賴海量資料集的輕量學習路線提供新證據。
閱讀文章 ↗憶阻器訓練新法 EaPU:AI 訓練能耗比 GPU 低近百萬倍
中國研究團隊在 Nature Communications 提出 EaPU 訓練法,把憶阻器的隨機切換特性轉為機率式權重更新,寫入次數減少逾 99%,訓練能耗比 GPU 低近六個數量級,裝置壽命延長約千倍。本文解析原理、實測結果與距離 LLM 的路還有多遠。
閱讀文章 ↗
2025
4 篇文章Meta 入主 Scale AI 後:OpenAI 終止合作、Google 傳跟進
2025年6月中旬,Meta 以143億美元入股 Scale AI 並帶走執行長 Alexandr Wang 後,OpenAI 證實逐步終止與 Scale 的資料合作,Reuters 報導最大客戶 Google 也打算切斷關係,AI 資料供應鏈的中立性焦慮全面浮上檯面。
閱讀文章 ↗Meta 敲定 Scale AI 投資:143 億美元換 49% 股權
2025 年 6 月 12 日,Meta 敲定對 Scale AI 約 143 億美元的投資,換取 49% 股權,估值達 290 億美元。創辦人亞歷山大·王卸任執行長、加入 Meta 超級智慧團隊並留任董事;策略長 Jason Droege 接任,公司強調維持獨立。
閱讀文章 ↗Meta 傳洽談投資 Scale AI,規模可望超過百億美元
2025 年 6 月 8 日彭博報導,Meta 正洽談對資料標註公司 Scale AI 進行數十億美元投資,金額可能超過 100 億美元,有望擠進史上最大私人公司募資事件之列;條件尚未敲定,雙方均拒絕評論。TechCrunch 指出,這將是 Meta 至今最大一筆對外 AI 投資。
閱讀文章 ↗Reddit 控告 Anthropic 未經授權抓取訓練資料
2025年6月4日,Reddit 在加州舊金山法院起訴 Anthropic,指控其在無授權協議下抓取平台內容訓練 Claude,並在宣稱封鎖爬蟲後仍存取逾10萬次。這是大型平台首度直接控告AI模型開發商的訓練資料訴訟。
閱讀文章 ↗
2026
9 ARTICLESCloudflare's Disallow AI Training Setting: What Changes for Your Crawl Policy
Cloudflare's new setting lets mixed-use crawlers keep indexing your site while refusing AI training use.
READ POST ↗Bot Preference Sync: Cloudflare Syncs Your AI Bot Policies to robots.txt
Cloudflare's Bot Preference Sync writes your AI bot settings to robots.txt, keeping stated preferences and enforced rules aligned. Learn how it works, publisher defaults, and…
READ POST ↗DharmaOCR vs Newer Models: Why Specialized OCR Training Still Pays
Dharma-AI benchmarked DharmaOCR against Mistral OCR4 and Unlimited-OCR on Brazilian Portuguese. The two-stage training recipe and token-level drift mechanics carry the real lessons for builders.
READ POST ↗General Intuition Raises $320M at $2.3B for World Models
General Intuition raised $320M at a $2.3B valuation to train one model on action-labeled gameplay that plays games, runs simulations, and controls physical robots.
READ POST ↗Anthropic's $30B Run Rate and Multi-Gigawatt TPU Deal: What It Means for AI Product Builders
Anthropic's new Google-Broadcom TPU deal signals a shift to compute supply chains. Learn what it means for cost, resilience, and your AI product strategy.
READ POST ↗Rakuten AI 3.0: Japan's Largest Model Goes Open-Weight
Rakuten open-sources Rakuten AI 3.0 under Apache 2.0: about 671B total parameters, 40B active per token, GENIAC-backed and tuned for Japanese. Benchmarks and the reality of running it.
READ POST ↗Meta Signs Multibillion-Dollar Deal to Rent Google TPUs
Meta signed a multibillion-dollar deal to rent Google Cloud TPUs for next-generation models, and is negotiating to buy millions more for its own data centers. The AI chip market just changed shape.
READ POST ↗Letting AI Talk to Itself Boosts Learning, OIST Study Finds
OIST researchers added multi-slot working memory and self-directed inner speech to an active-inference AI. It generalized and multitasked better on sparse data — no giant training sets required.
READ POST ↗EaPU Cuts AI Training Energy Nearly a Million-Fold vs GPUs
Zhejiang Lab and Fudan University researchers published EaPU: probabilistic weight updates cut memristor writes by over 99% and training energy by nearly six orders of magnitude versus GPUs.
READ POST ↗
2025
4 ARTICLESOpenAI drops Scale AI after Meta deal; Google to follow
After Meta's $14.3 billion Scale AI deal in June 2025, OpenAI confirmed it was winding down Scale work and Reuters reported Google planned to follow, shaking trust in the AI data supply chain.
READ POST ↗Meta's $14.3 billion Scale AI bet lands Alexandr Wang
On June 12, 2025, Meta finalized roughly $14.3 billion for 49% of Scale AI, valuing it at $29 billion. Alexandr Wang left the CEO seat for Meta's superintelligence effort; Jason Droege took over.
READ POST ↗Meta in talks to invest billions in Scale AI
Bloomberg, June 8, 2025: Meta is in talks to invest billions in data-labeling firm Scale AI — possibly over $10B, among the largest private funding events; terms not final.
READ POST ↗Reddit sues Anthropic over AI training data scraping
On June 4, 2025, Reddit sued Anthropic over unauthorized scraping of Reddit content to train Claude, the first Big Tech platform suit against an AI model developer over training data.
READ POST ↗