Laban

身為長期在台北市記錄庇護工場的新聞採訪攝影師，看到少康戰情室那段爭論，我難以保持沉默。 在台北跑採訪時，我進過許多庇護工場。站在烘焙台與包裝線前，我鏡頭下的畫面始終如一：聽障、智能障礙或多重障礙的孩子，把同一個動作反覆練習上千次。一般人幾秒鐘就能完成的摺紙盒、秤麵粉、封口包裝，他們往往要學上好幾個月才能跟上產線。 每當看見他們好不容易完成一箱箱產品，孩子眼中展現出靠雙手自立的驕傲，以及站在一旁的輔導員與家長欣慰的神情，在我的鏡頭裡，那一直是台灣社會最踏實的自立力量。 張嘉玲議員在節目中直接攤開事實： 總統府向全台 62 家庇護工場採購禮盒，包含台北市的 9 家機構在內，並非單一政黨的作法。2008 年馬英九前總統上任時就採購了 4,200 盒，當時丁守中、洪秀柱、徐中雄等藍營立委全力支持，視為國家帶頭扶助弱勢的典範。十幾年來，歷任政府的決算報告都記錄得清清楚楚，這是延續多年的跨黨派默契。 在野黨立委如今卻坐在冷氣房裡揮刀凍結預算。林沛祥委員自己基隆選區的基隆市聾啞福利協進會、張啓楷委員有意參選市長的嘉義基督教醫院庇護咖啡，以及台北市在地的 9 家工場，全都在受影響的名單上。 我在台北實地看過庇護工場的運作。身障機構的資金與倉儲空間本就極度吃緊，為了承接公部門與大型單位的長期訂單，工場往往必須提早下訂原物料並排定生產期程。如今立法院凍結預算，原料積壓在庫房，訂單無預警卡死。許多家長焦急陳情：「孩子每天認真工作，到底做錯了什麼？」 更令人心寒的是張啓楷委員在節目中的回應。 他不僅沒有反省，反而理直氣壯地表示：只要賴清德總統退讓、依法行事，預算就會解凍；甚至說出「你自己掏薪水出來買不行嗎？」 這番言論直接揭露了背後的政治算計： 在野黨清楚知道凍結預算會波及弱勢的生計，卻仍刻意扣留這筆款項作為籌碼，逼迫執政黨在政爭中妥協。 身為親眼見證這些孩子付出汗水、記錄家長辛苦守護的攝影師，我想公開詢問林沛祥與張啓楷兩位委員： 你們敢不敢走進台北市或各地的庇護工場？敢不敢看著身障孩子與父母的眼睛，親口對他們說：「為了在台北與執政黨鬥爭，我凍結了你們好不容易等到的訂單」？ 國會監督本是立委職責，但把弱勢家庭的尊嚴與生計當成談判肉票，這絕非監督，而是毫無同理心的政治霸凌。

AI infrastructure commitments, grid demand response, and trending open-source projects: 1. Anthropic IPO filing discloses $895.4B in compute commitments Latest filings reveal third-quarter operational specifics: $895.4B in total compute and infrastructure commitments, with $447.5B due before 2031 and only $213.8B subject to cancellation conditions. 🔗 nitter.cf/rohanpaul_ai/status/21… 2. OpenAI launches Grid Protectors Program OpenAI partners with power utilities to curb data center energy demand during peak grid stress. The initial deployment targets the Effingham County campus in Georgia. 🔗 openai.com/index/grid-protec… 3. karpathy/nanochat (GitHub +321 stars) Andrej Karpathy released a minimal LLM training harness. Trains a GPT-2 grade model on a single 8xH100 node in under 1.5 hours for roughly $48 (around $15 on spot instances), spanning pretraining, finetuning, and evaluation. 🔗 github.com/karpathy/nanochat 4. CNCKitchen/ScanRuler (GitHub +222 stars) CNC Kitchen open-sourced a browser-based 3D measurement and reverse engineering tool (AGPL-3.0). Runs entirely client-side, comparing 3D scan meshes against CAD STEP files with deviation color mapping and wall thickness analysis. 🔗 github.com/CNCKitchen/ScanRu…

AI 基礎設施承諾、電網調節與開源專案動態： 1. Anthropic IPO 申報文件披露 8954 億美元算力承諾 最新申報顯示第三季營運細節：算力與基礎設施合約承諾總額約 8954 億美元，其中 4475 億美元須在 2031 年前支付，僅 2138 億美元具備取消或附帶條件。 🔗 nitter.cf/rohanpaul_ai/status/21… 2. OpenAI 啟動 Grid Protectors 電網調節計畫 OpenAI 與電力公司合作，在用電高峰期下調未來算力園區的電力負載，協助電網削峰。首站設於喬治亞州 Effingham County 園區。 🔗 openai.com/index/grid-protec… 3. karpathy/nanochat（GitHub +321 stars） Andrej Karpathy 開源極簡 LLM 訓練框架。在單節點 8xH100 上以約 48 美元算力成本（競價實例約 15 美元）、1.5 小時以內訓練出達 GPT-2 水平的模型，覆蓋預訓練、微調至推論全流程。 🔗 github.com/karpathy/nanochat 4. CNCKitchen/ScanRuler（GitHub +222 stars） CNC Kitchen 開源的瀏覽器端 3D 檢驗與逆向工程工具（AGPL-3.0）。純前端執行，在瀏覽器內比對 3D 掃描點雲與 CAD STEP 零件偏差，支援壁厚測量與實體逆向建模。 🔗 github.com/CNCKitchen/ScanRu…

AI infrastructure, financial metrics, and open-source updates: 1. OpenAI deploys activation classifiers (probes) for runtime safety Starting with GPT-5.6, probes evaluate neuron activations per token on production traffic. Violations halt generation or tool calls mid-stream, offering stronger resistance to jailbreaks than text classifiers. 🔗 alignment.openai.com/trainin… 2. Anthropic quarterly revenue nears $17B, with $23.3B net loss Revenue grew twelvefold year-over-year. Operating costs reached $17.5B, pushing net loss from $12.1B to $23.3B as compute commitments and model training expenses expand. 🔗 the-decoder.com/anthropic-gr… 3. Claude Platform launches Message Threads public beta Long-running agents send incremental messages rather than re-transmitting entire conversation histories, cutting median TTFT by ~35% on 500k-1M token requests. 🔗 claude.com/blog/message-thre… 4. cheaper-inference/OmniRoute (GitHub +390 stars) Free MIT-licensed AI gateway routing 375 providers and 1,400+ models through a single endpoint. Features quota-aware fallbacks and token compression for Claude Code, Cursor, and OpenCode. 🔗 github.com/cheaper-inference… 5. hyprwm/Hyprland (GitHub +508 stars) Independent, highly customizable dynamic tiling Wayland compositor that maintains fluid performance without sacrificing visual polish. 🔗 github.com/hyprwm/Hyprland

AI 模型安全監控、營收數據與開源工具最新動態： 1. OpenAI 部署模型內部激活分類器（Probes） 從 GPT-5.6 起在推論流量部署探針，逐 token 即時評估內部神經元激活。違規時直接在中途阻斷生成或工具調用，比外部文字分類器更能防禦越獄攻擊。 🔗 alignment.openai.com/trainin… 2. Anthropic 單季營收增至 170 億美元，淨虧損 233 億美元 營收達到去年同期 12 倍，但單季營運成本達 175 億美元，淨虧損由 121 億美元擴大至 233 億美元，算力合約與模型訓練支出維持高水位。 🔗 the-decoder.com/anthropic-gr… 3. Claude Platform 推出 Message Threads 公測 智慧體長會話不必每輪重送完整歷史紀錄，首 token 延遲（TTFT）在 500k-1M token 請求下降低約 35%。 🔗 claude.com/blog/message-thre… 4. cheaper-inference/OmniRoute（GitHub +390 stars） 開源 MIT AI 閘道器，單一端點整合 375 家供應商與 1,400+ 模型。支援動態配額容錯與 RTK 壓縮，相容 Claude Code、Cursor、OpenCode。 🔗 github.com/cheaper-inference… 5. hyprwm/Hyprland（GitHub +508 stars） 獨立動態平鋪 Wayland 混成器，兼顧流暢度與外觀自訂能力。 🔗 github.com/hyprwm/Hyprland

AI & Open Source Updates: 1. Tencent Hunyuan open-sources Hyra 2.0 Shifts automated science from direct problem-solving to agent-driven development of scientific tools and models. The three-tier architecture (Overall, Mission, Worker) has produced verified outputs in protein design and quantum error correction. 🔗 mp.weixin.qq.com/s?__biz=Mzk… 2. Trakkr: Tracing 1,285 AI-quoted statistics Audited 200 quantitative queries across five major LLMs. Only 40% matched the cited research, 40% could not be verified, and 11% had no traceable origin. 🔗 trakkr.ai/trakkr-research/zo… 3. OpenRouter rebuilds default routing Switches from pure price sorting to a composite Routing Cost metric factoring in effective price, TTFT, throughput, and error rates. TTFT dropped 16.9% and 429 errors fell 67%. 🔗 openrouter.ai/blog/announcem… 4. vercel-labs/agent-desktop (+2,243 stars, Rust) Desktop computer use tool for agents. Interacts directly with OS accessibility trees instead of guessing coordinates from pixels, ensuring stable references and repeatable actions. 🔗 github.com/vercel-labs/agent… 5. JustVugg/colibri (+655 stars, C) Pure C, zero-dependency MoE inference engine. Streams expert weights straight from disk to run frontier MoE models on standard consumer hardware. 🔗 github.com/JustVugg/colibri

AI 產業動態與開源趨勢重點： 1. 騰訊混元開源 Hyra 2.0 重點從直接解題轉向讓 Agent 自主研發科學工具與模型。架構採 Overall、Mission、Worker 三層分工，成果已用於蛋白質設計與量子碼。 🔗 mp.weixin.qq.com/s?__biz=Mzk… 2. Trakkr 實測 1285 條 AI 引用數據溯源 向 5 款主流模型提出 200 個需要數字的問題。結果僅 40% 能對齊原始研究，40% 無法核查，11% 無處可查。 🔗 trakkr.ai/trakkr-research/zo… 3. OpenRouter 重建預設路由演算法 改以 Routing Cost 綜合價格、TTFT、吞吐與錯誤率排定優先順序。實測首字延遲降低 16.9%，429 錯誤減少 67%。 🔗 openrouter.ai/blog/announcem… 4. vercel-labs/agent-desktop (+2,243 stars, Rust) 桌面 Agent 控制工具。直接讀取作業系統無障礙樹結構操作 UI，避開像素猜測帶來的點擊失效。 🔗 github.com/vercel-labs/agent… 5. JustVugg/colibri (+655 stars, C) 純 C 語言無依賴的 MoE 模型推論引擎。透過磁碟串流載入專家權重，能在一般硬體運行大型 MoE。 🔗 github.com/JustVugg/colibri

AI & Open Source Developments: Supply Chain Alerts & Frontier Weights Cursor warned of malicious Anthropic extensions in its marketplace; Mistral and OpenMOSS published weights for 1T MoE and layout-aware image generation. On GitHub, image upscaler Upscayl and spatial vision library Kornia gained momentum. Key updates: 1. Malicious Anthropic Extension Found in Cursor Marketplace Cursor flagged unauthorized releases of Anthropic.claude-code (versions 2.1.312 and 2.1.313) in its extension store, affecting up to 4,100 users. Cursor published inspection commands and removal scripts. 🔗 cursor.com/blog/claude-code-… 2. Mistral Large 4 (Le Chonk) Releases Open Weights on Hugging Face Mistral published Mistral Large 4, an open-weight multimodal MoE model featuring ~1T total parameters, 49B active parameters, and a 1M token context window. Technical docs include throughput metrics and security benchmarks. 🔗 huggingface.co/blog/mistral-… 3. OpenMOSS Introduces MOSS-Image OpenMOSS released a 5B-parameter single-stream DiT model trained from scratch. It handles prompts up to 4K tokens, renders multilingual text natively, and supports editable JSON layouts. 🔗 openmoss.ai/blog/cn/moss-ima… 4. Upscayl (+124 stars today, TypeScript) A free and open-source local AI image upscaler for Linux, macOS, and Windows. It runs super-resolution models on client hardware without cloud uploads. 🔗 github.com/upscayl/upscayl 5. Kornia (+101 stars today, Python) A differentiable computer vision and geometric deep learning library for Spatial AI, built directly on PyTorch with robotics and camera calibration primitives. 🔗 github.com/kornia/kornia

AI 與開源趨勢追蹤：供應鏈安全警報與開源權重釋出 Cursor 通報外掛市場出現偽冒 Anthropic 的惡意套件；Mistral 與 OpenMOSS 分別釋出 1T 參數量 MoE 與版面感知生圖模型。GitHub 開源生態中，影像放大工具 Upscayl 與空間視覺庫 Kornia 增長顯著。重點整理： 1. Cursor 擴充套件市場通報惡意 Anthropic 外掛 Cursor 通報其 VS Code 擴充套件市場出現第三方偽造的 Anthropic.claude-code（版本 2.1.312 與 2.1.313），約 4,100 名用戶受影響。官方已發布清理腳本與後門檢測指令。 🔗 cursor.com/blog/claude-code-… 2. Mistral Large 4（Le Chonk）開放權重上線 Hugging Face Mistral 釋出 Mistral Large 4，採 MoE 多模態架構，總參數約 1T、單次啟用 49B，配備 1M token 上下文窗口。官方提供架構細節、安全基準與硬體吞吐量數據。 🔗 huggingface.co/blog/mistral-… 3. OpenMOSS 發布圖像生成模型 MOSS-Image OpenMOSS 開放 5B 參數單流 DiT 權重，從零預訓練，支援 4K token 提示詞輸入、中英雙語長文本渲染及 JSON Layout 佈局編輯。 🔗 openmoss.ai/blog/cn/moss-ima… 4. Upscayl（GitHub 單日 +124 星，TypeScript） 跨平台開源 AI 影像放大工具，支援 Linux、macOS 與 Windows，使用本機硬體執行超解析度演算法，無需依賴雲端伺服器。 🔗 github.com/upscayl/upscayl 5. Kornia（GitHub 單日 +101 星，Python） 專為空間 AI 與機器人幾何視覺打造的開源電腦視覺庫，以 PyTorch 為核心，提供可微分幾何轉換與神經渲染運算元。 🔗 github.com/kornia/kornia

vLLM delivered Day 0 serving support for the 1T-parameter Mistral Large 4; Kazik documented an event-driven four-layer automation architecture for autonomous agents; Claude Code restored the ultrathink command for on-the-fly reasoning effort boosts. In open source, FlashML-org/FreeToken gained 242 stars for edge-native MoE serving, and andreavolpato/spektrafilm added 159 stars for spectral analog film simulation. Key developments: 1. vLLM Delivers Day 0 Support for Mistral Large 4, Scaling 1T MoE Inference vLLM added production support for Mistral's latest flagship, Mistral Large 4 (1T total parameters, 49B active per token). The architecture integrates MLA sparse attention with DSA top-k indexing, cross-layer KV sharing, NoPE, and SSMax. It includes a built-in DSpark speculative decoding drafter and Pixtral vision encoder for high-throughput serving across local and cloud environments. Link: vllm.ai/blog/2026-10-29-mist… 2. Kazik Details Event-Driven Four-Layer Architecture for Production Agents Content creator Kazik published a breakdown of a production agent pipeline refined over twelve months. The design shifts the task trigger from manual human prompts to external event listeners. The system uses rules.json for event orchestration, Mac mini hardware for task execution, a custom five-stage CI pipeline for checks and verification, and interactive approval cards restricted to financial costs and data deletions. Link: mp.weixin.qq.com/s?__biz=MzI… 3. Claude Code Restores ultrathink Command to Boost In-Flight Reasoning Addy Osmani highlighted a terminal workflow for Claude Code: typing ultrathink while Claude processes a task increases the reasoning effort for the remainder of that turn. The capability lets engineers inject deeper reasoning during difficult debugging or refactoring loops without restarting the CLI session. Link: nitter.cf/addyosmani/status/2116… 4. FlashML-org/FreeToken (+242 stars today, Python, 15.4k total stars) An edge-native MoE serving engine designed to run 290B+ frontier models on consumer PCs and gaming laptops. FreeToken coordinates CPU-GPU bandwidth through a q* co-execution policy, double-buffered prefill streaming, and semantic anchor caching that eliminates redundant recomputation during agent tool calls and reasoning blocks. Repo: github.com/FlashML-org/FreeT… 5. andreavolpato/spektrafilm (+159 stars today, Python) A physics-based analog photography simulation engine driven by spectral measurements. Instead of applying aesthetic LUT approximations, spektrafilm processes manufacturer spectroscopic data through virtual negatives, enlarger optics, halation, and grain models. It includes a desktop GUI alongside Vulkan and OFX implementations. Repo: github.com/andreavolpato/spe…

vLLM 針對 1T 參數 Mistral Large 4 釋出 Day 0 支援；卡茲克開源事件觸發型四層 Agent 自動化架構；Claude Code 回歸 ultrathink 即時調高推理深度的指令。開源端，消費級 MoE 推論引擎 FlashML-org/FreeToken 單日增 242 星，物理級底片光譜模擬專案 andreavolpato/spektrafilm 增 159 星。核心動態整理： 1. vLLM 釋出 Mistral Large 4 Day 0 支援，1T 參數多模態 MoE 落地 vLLM 宣布支援 Mistral 新旗艦模型 Mistral Large 4（1T 總參數、49B 激活）。架構整合 MLA 稀疏注意力與 DSA top-k 索引、跨層 KV 共享、NoPE 與 SSMax，內建 DSpark 推測解碼草稿器與 Pixtral 視覺編碼器，提供生產環境高吞吐部署基準。 連結： vllm.ai/blog/2026-10-29-mist… 2. 卡茲克公開四層 Agent 自動化架構：從人找任務改為事件驅動 卡茲克公開運作近一年的自媒體 Agent 系統架構，將任務起點轉移至事件觸發。系統分為四層：rules.json 統一管理事件與調度中台、Mac mini 執行端、包含五步檢查的 CI 中台，以及僅在刪庫、資金支出時介入的飛書決策卡片，降低人工維護負擔。 連結： mp.weixin.qq.com/s?__biz=MzI… 3. Claude Code 回歸 ultrathink 指令，動態調高單回合推理深度 Addy Osmani 分享 Claude Code 終端技巧：在模型執行任務期間直接輸入 ultrathink，可調高該回合剩餘步驟的推理努力（Reasoning Effort）。該功能適合在除錯與架構重構卡關時即時注入高思考深度，無需中斷當前 session。 連結： nitter.cf/addyosmani/status/2116… 4. FlashML-org/FreeToken（單日 +242 星，Python，累積 15.4k 星） 專為個人硬體設計的開源邊緣 MoE 推論引擎，支援在消費級顯卡上執行 290B+ 參數量模型。採用頻寬自我調整 CPU-GPU 協同執行（q* 策略）、雙緩衝預填串流與語義錨點（Semantic Anchor）快取，避免 Agent 多輪工具呼叫與思考區塊的重複計算。 專案： github.com/FlashML-org/FreeT… 5. andreavolpato/spektrafilm（單日 +159 星，Python） 基於物理光譜數據的傳統底片感光模擬工具。專案直接使用底片原廠光譜數據表，在虛擬負片、顯影放大機設定、光暈與顆粒管線中進行物理光譜計算，取代傳統 LUT 濾鏡近似法，並提供 Python 參考實作、桌面 GUI 與 Vulkan 模組。 專案： github.com/andreavolpato/spe…

Unsloth released a 1-bit Dynamic GGUF for Mistral Le Chonk, running a 1T-parameter model locally on a Mac Studio; The Verge dissected Anthropic's S-1 filing, revealing over $850B in long-term compute commitments; Factory launched Droid 1.0, introducing programmatic tool calling inside isolated sandboxes. On GitHub, patchy631/ai-engineering-hub added 573 stars, while cjpais/Handy gained 383 stars for fully offline speech-to-text in Rust. Key developments: 1. Unsloth Releases Mistral Le Chonk 1-bit GGUF, Bringing 1T Parameters to Mac Studio Unsloth published a Dynamic 1-bit GGUF quant for Mistral Le Chonk (Mistral Large 4, 1T-parameter MoE with 52B active). The quantization slashes disk footprint from 1.1TB in FP8 to 233GB—an 80% reduction—while preserving 91% top-1 accuracy. Running it requires 256GB to 270GB of unified memory on a Mac Studio, featuring multimodal input, a 1M token context window, and support for over 160 languages. Link: nitter.cf/UnslothAI/status/21158… 2. The Verge Unpacks Anthropic's S-1: $850B in Compute Commitments Behind Recursive Self-Improvement The Verge analyzed Anthropic's S-1 filing, highlighting aggressive compute leverage. Anthropic posted 14x year-over-year revenue growth to $16.3B in the first half of 2026, alongside a 900% net revenue retention rate. However, non-cancellable purchase commitments reached $265.6B, with long-term leasing deals pushing total 20-year compute commitments above $850B. The financial model hinges directly on recursive self-improvement driving compounding returns. Link: theverge.com/ai-artificial-i… 3. Factory Introduces Droid 1.0 with Programmatic Tool Calling for Enterprise Agents Factory launched Droid 1.0, a model-agnostic agent harness centered on Programmatic Tool Calling (PTC). Instead of sequential JSON function-calling round trips, the model writes JavaScript programs that execute hundreds of tool interactions within an isolated sandbox during a single turn. The approach reduces latency and token consumption in multi-step enterprise workflows. Link: factory.com/news/frontier-ha… 4. patchy631/ai-engineering-hub (+573 stars today, Jupyter Notebook) An open-source repository dedicated to production AI engineering. It covers end-to-end LLM fine-tuning, advanced retrieval-augmented generation (RAG) architectures, and real-world multi-agent systems. The repo provides complete runnable notebooks and architectural designs for deployed systems. Repo: github.com/patchy631/ai-engi… 5. cjpais/Handy (+383 stars today, Rust) A free, open-source speech-to-text application written in pure Rust. All audio inference processes locally on-device without cloud dependencies or audio data transmission. The project features an extensible architecture supporting real-time dictation across multiple desktop environments. Repo: github.com/cjpais/Handy

Unsloth 推出 Mistral Le Chonk 1-bit GGUF 量化版，首度讓 1T 參數模型在單台 Mac Studio 本地推論；The Verge 拆解 Anthropic S-1 招股書，揭露 8,500 億美元長期算力採購承諾；Factory 釋出 Droid 1.0，以沙箱內程式化腳本呼叫改寫 Agent harness 架構。開源領域中，系統工程教學庫 patchy631/ai-engineering-hub 單日增加 573 星，純離線 Rust 語音轉文字工具 cjpais/Handy 增加 383 星。核心動態整理： 1. Unsloth 發布 Mistral Le Chonk 1-bit GGUF，1T 參數模型落地 Mac Studio Unsloth 針對 Mistral Large 4（1T 參數 MoE、52B 激活）釋出 Dynamic 1-bit GGUF。該版本模型體積自 FP8 的 1.1TB 降至 233GB，降幅達 80%，同時保留 91% 的 top-1 準確率。硬體配置需求 256GB 至 270GB 統一記憶體，支援文字與視覺輸入、1M 上下文長度與 160 種語言，大幅降低兆級參數模型的地端驗證門檻。 連結： nitter.cf/UnslothAI/status/21158… 2. The Verge 解讀 Anthropic S-1：8,500 億美元算力承諾押注模型自我演進 The Verge 分析 Anthropic 招股書揭示其高資本運作邏輯。Anthropic 2026 上半年營收年增 14 倍達 163 億美元，淨收入留存率達 900%；然而其不可撤銷採購合約累計 2,656 億美元，加計未來租賃等長期承諾，未來 20 年算力資本承諾達 8,500 億美元。整體財務架構全面押注模型具備遞歸自我改進能力，以推論效能換取市場份額。 連結： theverge.com/ai-artificial-i… 3. Factory 釋出 Droid 1.0，以程式化工具呼叫重構企業級 Agent Harness Factory 推出模型中立的 Agent harness 架構 Droid 1.0，核心為程式化工具呼叫（Programmatic Tool Calling, PTC）。該架構允許模型在受控沙箱內撰寫 JavaScript 腳本，單次往返批次執行數百次工具呼叫，徹底替代傳統多輪次 JSON function calling，降低 token 消耗並減少通訊延遲。 連結： factory.com/news/frontier-ha… 4. patchy631/ai-engineering-hub（單日 +573 星，Jupyter Notebook） 專注於實戰 AI 系統工程的開源知識庫。涵蓋大型語言模型微調、生產級 RAG 檢索增強系統架構、以及企業 AI Agent 落地實作範例。專案提供完整 Notebook 程式碼與架構圖，針對真實業務場景的工程痛點提供落地解法。 專案： github.com/patchy631/ai-engi… 5. cjpais/Handy（單日 +383 星，Rust） 基於 Rust 開發的純離線開源語音轉文字（STT）應用程式。音訊資料完全於本地端推論處理，無任何雲端上傳或資料外洩風險。架構具備模組化擴充介面，支援跨平台桌面環境的即時語音辨識與自動輸入。 專案： github.com/cjpais/Handy

好想知道+1

Google Gemini now generates and edits documents, PDFs, slides, and sheets directly in-chat; SemiAnalysis dissected Anthropic's S-1 filing, detailing inference margins and megawatt revenue behind its $1.5T to $2T target valuation; The New York Times reported US intelligence investigations into supply chain espionage targeting Nvidia's inner circle. On GitHub, macOS offline voice dictation tool hex spiked 1,329 stars in Rust, while Python web scraping framework Scrapling gained 844 stars. Key updates: 1. Google Gemini Adds Native Document, PDF, and Sheet Generation & Editing Google Gemini now supports creating, previewing, and editing documents, PDFs, slide decks, and spreadsheets directly inside the chat interface. Users can convert handwritten notes into formatted PDF study guides, highlight text or layout elements on-screen for targeted revisions, and continue prompt adjustments without leaving the workspace. Link: nitter.cf/GeminiApp/status/21158… 2. SemiAnalysis Unpacks Anthropic's S-1 Economics and Compute Margins On The Verge's Decoder podcast, SemiAnalysis analyst Joey Brookhart broke down Anthropic's S-1 registration for its planned Nasdaq IPO at an estimated $1.5T to $2T valuation. The breakdown highlights inference gross margins, revenue generated per megawatt of power, and enterprise spending concentration, illustrating how Anthropic trades aggressive capital expenditure for token throughput. Link: theverge.com/podcast/1021322… 3. NYT Reports US Intelligence Probe into Espionage Targeting Nvidia Leadership The New York Times reported that US intelligence agencies investigated Alice Huang (Le Huang), head of Singapore data center operator Megaspeed, over alleged intelligence gathering surrounding Nvidia CEO 黄仁勋's network. The inquiry shows high-performance AI hardware and physical compute hubs becoming central targets in geopolitical intelligence. Link: the-decoder.com/chinese-spie… 4. anomalyco/hex (+1,329 stars today, Rust) A local-first voice dictation tool built for macOS. Users hold a global hotkey, speak, and stream transcribed text into any active desktop application. Written in pure Rust, it runs speech recognition models locally on-device without cloud transmission, and now offers a Linux beta. Repo: github.com/anomalyco/hex 5. D4Vinci/Scrapling (+844 stars today, Python) An open-source Python web scraping framework designed for dynamic modern pages. It features intelligent DOM traversal and adaptive retry mechanisms, scaling from single-request extraction to multi-threaded crawls while navigating single-page application hydration and Cloudflare challenges. Repo: github.com/D4Vinci/Scrapling

Google Gemini 正式將文件、PDF、簡報與試算表的生成與即時編輯收攏進對話介面；SemiAnalysis 拆解 Anthropic S-1 招股書，揭露 1.5 兆至 2 兆美元估值背後的推論毛利與每兆瓦營收；《紐約時報》披露美國情報機構調查針對輝達核心圈的供應鏈情報滲透。開源社群方面，macOS 原生 Rust 離線語音輸入工具 hex 單日狂飆 1,329 星，現代網頁抓取框架 Scrapling 增加 844 星。重點整理： 1. Google Gemini 支援原生文件、PDF、簡報與試算表直接編輯 Google Gemini 宣布可在對話介面中直接生成、預覽與編輯文件、PDF、簡報及試算表。使用者能將手寫筆記轉為 PDF 學習指南，在畫面中直接選取段落要求 Gemini 修改文字或版面視覺，並支援連續追問調整，整合多項辦公軟體的基本作業流程。 連結： nitter.cf/GeminiApp/status/21158… 2. SemiAnalysis 拆解 Anthropic S-1 招股書推論經濟模型 SemiAnalysis 分析師 Joey Brookhart 於 The Verge Decoder 解析 Anthropic 遞交的 S-1 招股文件。公司擬於那斯達克上市，目標估值區間 1.5 兆至 2 兆美元。分析報告揭露其推論毛利率、每兆瓦電力產生營收與客戶支出結構，呈現 Anthropic 如何以龐大資本支出換取推論吞吐量，並由公開市場定價其商業回報。 連結： theverge.com/podcast/1021322… 3. 紐約時報披露美情報機構調查輝達周邊情報刺探 《紐約時報》報導指出，美國情報機構調查新加坡資料中心營運商 Megaspeed 負責人 Alice Huang（黃樂），指其涉嫌接觸輝達執行長黃仁勳身邊核心人員獲取情報。該調查顯示先進 AI 晶片供應鏈與實體運算節點已成為地緣情報攻防焦點。 連結： the-decoder.com/chinese-spie… 4. anomalyco/hex（單日 +1,329 星，Rust） 專為 macOS 開發的本機優先語音輸入工具。按住全域快捷鍵即可說話並將轉錄文字貼入任意軟體。採用原生 Rust 開發，語音辨識模型完全在裝置端離線執行，保障音訊隱私，目前已推出 Linux 測試版。 專案： github.com/anomalyco/hex 5. D4Vinci/Scrapling（單日 +844 星，Python） 針對現代動態網頁與反爬機制的開源 Python 爬蟲框架。底層具備智慧型 DOM 解析與自適應重試機制，支援單一請求擷取至全站大規模平行爬取，解決單頁應用（SPA）與 Cloudflare 驗證牆的爬取效率問題。 專案： github.com/D4Vinci/Scrapling

Anthropic filed its S-1 showing Q2 2026 revenue surged to $11.6B alongside its first operating profit; Codex doubled usage caps across Pro 100/200/500 and Business Premium tiers; OpenAI published its multi-layer alignment monitoring framework covering frontier models like GPT-6 Astra. On GitHub, self-hosted document manager paperless-ngx added 80 stars today, while agentic-awesome-skills gained 74 stars with 2,884 agent skills across major coding CLIs. Key updates: 1. Anthropic S-1 Shows Q2 Revenue Surging to $11.6B with First Operating Profit Tech analyst Timothy B. Lee reviewed Anthropic's public S-1 filing, reporting that Q2 2026 revenue grew from $4.7B in Q1 to $11.6B, yielding the company's first operating profit. By late June, Anthropic held cash reserves trailing only Google and Berkshire Hathaway. Massive compute procurement commitments and signs of moderating growth remain focal points for sustainability under heavy capital expenditure. Link: understandingai.org/p/anthro… 2. Codex Doubles Usage Limits Across Pro and Business Premium Tiers Codex lead Tibo announced an immediate doubling of rate limits across Pro 100, 200, 500, and Business Premium plans. Pro 200 limits will not decrease, while Work and core Codex quotas remain unchanged. Extended dots allowances stay separate from default quotas, with planned model and inference efficiency upgrades set to expand usable compute capacity further. Link: nitter.cf/thsottiaux/status/2115… 3. OpenAI Details Frontier Model Misalignment Monitoring System OpenAI's Alignment team published architectural details of its multi-layer monitoring system designed to catch misaligned behaviors in frontier models, including GPT-6 Astra. The framework monitors training runs, automated evaluations, internal dogfooding, and select production deployments. The report shares detection rates and red-teaming findings, clarifying technical limits when tracking covert deception or drift. Link: alignment.openai.com/how-we-… 4. paperless-ngx/paperless-ngx (+80 stars today, Python) A community-driven document management system that scans, OCRs, indexes, and archives physical and digital paperwork. It provides a local Docker deployment workflow and REST APIs to run fully private document pipelines without cloud dependencies. Repo: github.com/paperless-ngx/pap… 5. sickn33/agentic-awesome-skills (+74 stars today, Python) An open-source library curating 2,884 reusable AI agent skills for Claude Code, Codex, Cursor, and Gemini CLI. Its AAS Core engine provides local catalog discovery, agent-driven skill selection, and plan previews via CLI, MCP, and plugins. Repo: github.com/sickn33/agentic-a…

Anthropic 公布 S-1 招股書，2026 年第二季營收增至 116 億美元並錄得首度營運獲利；Codex 宣布即日起 Pro 100/200/500 與 Business Premium 各訂閱檔位用量配額全面翻倍；OpenAI 發表論文揭露 GPT-6 Astra 等前沿模型的多層失對齊監控系統架構。開源社群方面，文件管理歸檔系統 paperless-ngx 單日增 80 星，聚合 2,884 款 Agent 技能庫的 agentic-awesome-skills 飆升 74 星。重點整理： 1. Anthropic S-1 文件揭示 Q2 營收達 116 億美元，首度實現營運獲利 科技分析師 Timothy B. Lee 解讀 Anthropic S-1 文件指出，公司 2026 年第二季營收從第一季的 47 億美元暴增至 116 億美元，並錄得首度營運利潤。截至 6 月底 Anthropic 現金儲備僅次於 Google 與波克夏。但龐大的雲端算力承諾與後續成長放緩風險，使資本市場聚焦其能否在巨額資本支出下維持獲利軌跡。 連結： understandingai.org/p/anthro… 2. Codex 即日起 Pro 與 Business Premium 訂閱檔位用量配額翻倍 Codex 團隊負責人 Tibo 宣布，即日起全面調升 Pro 100、200、500 以及 Business Premium 方案的用量上限，額度全數翻倍。Pro 200 不會下調上限，Work 與 Codex 基礎用量維持現況。擴展用量 dots 維持獨立計算，後續將透過新模型架構與推論效率提升，進一步擴大開發者的有效使用空間。 連結： nitter.cf/thsottiaux/status/2115… 3. OpenAI 公布前沿模型失對齊監控架構，涵蓋 GPT-6 Astra OpenAI Alignment 團隊發表前沿模型失對齊偵測系統細節。該套架構覆蓋模型訓練、自動化評估、內部使用與部分外部部署，針對 GPT-6 Astra 等前沿系統進行即時偏差檢測。團隊公開內部檢出率數據與紅隊測試經驗，說明自動化監控在抓取模型隱蔽欺騙與意圖漂移時的工程邊界與局限。 連結： alignment.openai.com/how-we-… 4. paperless-ngx/paperless-ngx（單日 +80 星，Python） 開源社群維護的數位文件管理系統。支援硬體掃描整合、全文 OCR 自動辨識、標籤索引與歷史版本歸檔，具備本機 Docker 部署能力與 REST API，方便企業與個人搭建無外部依賴的文件自動化流水線。 專案： github.com/paperless-ngx/pap… 5. sickn33/agentic-awesome-skills（單日 +74 星，Python） 收錄 2,884 款即插即用 AI Agent 技能庫的開源專案。專門針對 Claude Code、Codex、Cursor 與 Gemini CLI 打造，內建 AAS Core 本機目錄發現引擎、Agent 自主決策選取模組與計畫預覽功能，支援 CLI、MCP 與標準外掛整合。 專案： github.com/sickn33/agentic-a…

OpenAI confirmed it will merge ChatGPT and Codex allowances into a single token-based budget starting December; Anthropic reported $16.3B in H1 2026 revenue with $112B in cash, prompting White House tech advisor David Sacks to highlight its capital efficiency; OpenRouter released a 233-line sandboxed code review script. On GitHub, open-source PR reviewer pr-agent and Tsinghua's multi-agent classroom OpenMAIC each gained over 100 stars today. Key updates: 1. OpenAI Merges ChatGPT and Codex into Unified Token Budget Starting December OpenAI confirmed in its Help Center that Plus, Pro, and Business plans will share a single usage allowance across ChatGPT and Codex starting in December. Usage will be metered in input and output tokens rather than message counts, with only Instant chat remaining exempt. The change integrates chat into strict compute metering, ending standalone chat allowances for heavy coding subscribers. Link: help.openai.com/en/articles/… 2. OpenRouter Details 233-Line Sandboxed Code Reviewer Costing $0.015 per Run OpenRouter published review.py, a standard-library Python tool that uses the Responses API alongside managed Linux sandboxes to verify bugs before suggesting fixes. The script runs checks inside an isolated container and produces unified diffs applicable via the --apply flag. Tests across Claude Haiku 5.5, GPT-6 Luna, DeepSeek V4.1 Flash, and Claude Sonnet 5.5 averaged 1 to 4 minutes and $0.015 to $0.045 per review. Link: openrouter.ai/blog/tutorials… 3. Anthropic Reports $112B Cash Reserve in S-1 as Tech Advisor Sacks Praises Capital Discipline White House PCAST Chair David Sacks analyzed Anthropic's S-1 registration statement on the All-In podcast. Filings show Anthropic generated $16.3B in H1 2026 revenue with $112B in cash on hand, on track for $100B in 2026 ARR, while seven co-founders retain 50.1% voting control via Class F shares. Sacks noted the company preserved nearly all invested capital, leaving cash reserves sufficient to cover long-term compute commitments. Link: ithome.com/1/020/633.htm 4. The-PR-Agent/pr-agent (+122 stars today, Python) An open-source AI Pull Request reviewer for GitHub, GitLab, and Bitbucket. It automatically generates PR descriptions, detects regression risks, and suggests inline code fixes without relying on proprietary platform tiers. Repo: github.com/The-PR-Agent/pr-a… 5. THU-MAIC/OpenMAIC (+116 stars today, TypeScript) An open-source multi-agent interactive classroom framework developed by Tsinghua University MAIC. It deploys collaborative learning environments where autonomous agents act as instructors, teaching assistants, and peers to deliver structured interactive coursework. Repo: github.com/THU-MAIC/OpenMAIC

OpenAI 確認自 12 月起將 Plus、Pro 與 Business 訂閱方案的 ChatGPT 與 Codex 用量合併為單一 Token 額度；Anthropic 招股書公布 2026 上半年營收 163 億美元並手握 1,120 億美元現金儲備，白宮科技顧問薩克斯稱其展現極致資本效率；OpenRouter 發布 233 行沙箱代碼審查腳本。GitHub 方面，開源審查工具 pr-agent 與清華多智慧體課堂 OpenMAIC 單日增星均破百。重點整理： 1. OpenAI 12 月起合併 ChatGPT 與 Codex 額度，改依 Token 統一計費 OpenAI 於支援中心確認，自 12 月起 Plus、Pro 與 Business 方案不再享有分開的聊天與代碼配額，兩者合併至單一用量池，依輸入與輸出 Token 消耗計費，僅 Instant 聊天維持獨立免額度。此舉將聊天納入嚴格算力監控，過往重度代碼搭配高階模型聊天的補貼時代告終。 連結： help.openai.com/en/articles/… 2. OpenRouter 發布 233 行沙箱代碼審查腳本，單次成本低至 0.015 美元 OpenRouter 公開純 Python 標準庫實作的 review.py。腳本透過 Responses API 整合代管 Linux 沙箱與補丁工具，在隔離環境中真實執行測試以驗證 bug，並直接輸出可由 --apply 套用的修復補丁。官方實測 Claude Haiku 5.5、GPT-6 Luna、DeepSeek V4.1 Flash 與 Claude Sonnet 5.5 四款模型，單次審查耗時 1 至 4 分鐘，成本介於 0.015 至 0.045 美元。 連結： openrouter.ai/blog/tutorials… 3. Anthropic 招股書公布千億現金儲備，白宮科技顧問薩克斯盛讚資本效率 白宮科學與技術顧問會議主席 David Sacks 於 All-In 播客分析 Anthropic S-1 註冊文件。數據顯示 Anthropic 2026 上半年營收 163 億美元，帳面現金達 1,120 億美元，預計 2026 全年 ARR 達 1,000 億美元，七位創辦人持 F 類股票掌控 50.1% 投票權。Sacks 指出其手握充裕現金，幾乎未消耗前期募集資金，現有儲備足以支撐算力採購承諾。 連結： ithome.com/1/020/633.htm 4. The-PR-Agent/pr-agent（單日 +122 星，Python） 原生開源 AI Pull Request 審查工具。支援 GitHub、GitLab 與 Bitbucket，可自動產生 PR 摘要、分析回歸風險並提供行內程式碼修復建議，獨立運作且不依賴特定商業平台的免費層限制。 專案： github.com/The-PR-Agent/pr-a… 5. THU-MAIC/OpenMAIC（單日 +116 星，TypeScript） 清華大學 MAIC 團隊開源的多智慧體互動課堂架構。提供一鍵部署的協同學習環境，由多個 AI Agent 分別擔任教師、助教與學習夥伴，在沉浸式介面中自主分工並即時互動解答。 專案： github.com/THU-MAIC/OpenMAIC

Five tech giants pledged over $3 trillion for AI infrastructure as Oracle's free cash flow turned negative; Google lost $500 billion in market value on fears that AI Overviews will erode ad revenue; OpenAI published amicus brief summaries arguing that AI model training constitutes fair use. On GitHub, Swift-based macOS menubar toolkit vorssaint-utils surged nearly 500 stars, and devops-exercises added 341 stars. Key updates: 1. Big Tech Commits Over $3 Trillion to AI Infrastructure, Oracle Cash Flow Turns Negative Alphabet, Meta, Microsoft, Amazon, and Oracle have committed over $3 trillion in future data center and compute investments. Google's spending commitments crossed $1 trillion, nearly eight times higher than last year. Meta committed over $800 billion, with roughly half tied to leases extending up to 30 years. Oracle became the first tech giant whose AI data center buildout pushed free cash flow negative, with positive cash flow projected by 2028. Link: ithome.com/1/020/632.htm 2. Google Plunges $500 Billion as AI Overviews Threaten Search Ad Revenue Alphabet's market capitalization fell from $3.97 trillion to $3.46 trillion in a single session. Total quarterly revenue grew 25% to $127.6 billion and cloud revenue doubled to nearly $31 billion, but profit dropped 25% while free cash flow dropped to negative $9 billion. Investors expressed concern over AI Overviews cannibalizing traditional paid search clicks. Link: bbc.com/news/articles/cm9qzd… 3. OpenAI Files Amicus Brief Summaries Defending Fair Use for AI Training OpenAI compiled amicus briefs submitted by library associations, legal scholars, independent creators, and the US government in the New York Times copyright lawsuit. The briefs argue that AI model training is transformative and does not cause copyright market substitution harm, warning that restricting fair use will damage scientific research and technology leadership. Link: openai.com/index/fair-use-be… 4. vorssaint/vorssaint-utils (+499 stars today, Swift) An open-source macOS menu bar utility built in native Swift. It integrates system resource monitoring, clipboard management, window utilities, and developer shortcuts with low memory overhead. Repo: github.com/vorssaint/vorssai… 5. bregman-arie/devops-exercises (+341 stars today, Python) A comprehensive interview and reference repository for DevOps and SRE engineers. It covers Linux internals, Kubernetes, Docker, Prometheus, Terraform, AWS, networking protocols, and automation scripting with structured Q&A. Repo: github.com/bregman-arie/devo…