Laban (@lunkertw) header
Laban (@lunkertw) profile picture

Yotron AI CTO Freelance Photographer #Photographer #Taipei #Taiwan Based in Taipei Taiwan

Taiwan,Taipei City
Joined January 2008
Photo by Laban (@lunkertw)Photo by Laban (@lunkertw)Photo by Laban (@lunkertw)Photo by Laban (@lunkertw)Photo by Laban (@lunkertw)Photo by Laban (@lunkertw)Photo by Laban (@lunkertw)Photo by Laban (@lunkertw)Photo by Laban (@lunkertw)Photo by Laban (@lunkertw)
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。
47
645
20
2,195
56,088
Latest AI Shifts & GitHub Trending Highlights (2026/11/13): • Jensen Huang Resigns as NVIDIA CEO NVIDIA announced Jensen Huang's resignation as President, CEO, and board director, effective immediately, sending shares down 14% at market open. The board's abrupt move ends his 33-year leadership, drawing industry comparisons to Steve Jobs's 1985 ouster from Apple. 🔗 Context: mp.weixin.qq.com/s/ToplAo9NH… 🔗 Discourse: teapawt.net/rohanpaul_ai/status/21… • Google DeepMind Introduces Gemini 4 Hydrogen Built for everyday coding and agentic workflows, rolling out across Gemini API, AI Studio, and Google Search. 🔗 Post: deepmind.google/blog/introdu… • NVIDIA/SkillEvaluator (+108 stars today) Multi-tier evaluation framework for AI agent skills with quality gates, semantic overlap detection, and live behavior benchmarks. 🔗 github.com/NVIDIA/SkillEvalu… • fastapi/fastapi (+82 stars today) High-performance Python web framework, serving as the core API foundation for production AI agents and microservices. 🔗 github.com/fastapi/fastapi
Apple 1985 (Steve Jobs), Atari 1978 (Nolan Bushnell), Uber 2017 (Travis Kalanick), OpenAI 2023 (Sam Altman). every few years a board relearns that the founder was the strategy.
52
AI 產業變動與 GitHub 開源熱點(2026/11/13): • 黄仁勋辭去 NVIDIA 總裁與執行長職務 NVIDIA 公告黄仁勋辭去總裁、CEO 與董事職務,開盤重挫 14%。董事會收網終結其 33 年掌舵,業界輿論將此對比 1985 年賈伯斯被逐出蘋果。 🔗 始末:mp.weixin.qq.com/s/ToplAo9NH… 🔗 評論:teapawt.net/rohanpaul_ai/status/21… • Google DeepMind 推出 Gemini 4 Hydrogen 專為日常編碼與 Agent 工作流設計的輕量高頻模型,同步上線 Gemini API、AI Studio 與搜尋生態。 🔗 官網:deepmind.google/blog/introdu… • NVIDIA/SkillEvaluator(單日 +108 星) 評估 AI Agent 技能的多層級框架,具備質量門禁、語意重疊檢測與行為評估。 🔗 github.com/NVIDIA/SkillEvalu… • fastapi/fastapi(單日 +82 星) 高效能 Python Web 框架,持續作為 AI Agent 服務與生產級 API 首選。 🔗 github.com/fastapi/fastapi
Apple 1985 (Steve Jobs), Atari 1978 (Nolan Bushnell), Uber 2017 (Travis Kalanick), OpenAI 2023 (Sam Altman). every few years a board relearns that the founder was the strategy.
68
Latest AI Industry Shifts & GitHub Trending Highlights: • US Lawmakers Investigate Nvidia China Chip Sales Rep. Richard Neal and Sen. Ron Wyden sent an inquiry to the Commerce Department, examining export arrangements and tariff deals between Trump and Jensen Huang, demanding preservation of regulatory records. 🔗 techcrunch.com/2026/11/12/de… • Google Smart Glasses Pre-orders Open Google Labs VP Josh Woodward announced official pre-orders shipping Dec 10, with in-person trials at Warby Parker, Gentle Monster, and select Google stores. 🔗 teapawt.net/joshwoodward/status/21… • Figma Launches 3D Transforms Open Beta Full seat users can now tilt, rotate, and arrange layers in 3D space with non-destructive perspective controls while keeping canvas assets editable. 🔗 figma.com/blog/how-3d-transf… • ggml-org/llama.cpp (+287 stars today) Minimalist C/C++ inference engine for local LLMs. 🔗 github.com/ggml-org/llama.cp… • openai/whisper (+118 stars today) Robust multilingual speech recognition model trained on large-scale weak supervision. 🔗 github.com/openai/whisper
1
157
AI 與開源生態最新動態(2026/11/13): • 美國會調查 Nvidia 對華晶片安排 眾議員 Richard Neal 與參議員 Ron Wyden 致函商務部,針對川普政府與黃仁勳的晶片出口及關稅協議啟動審查,要求封存相關監管紀錄。 🔗 techcrunch.com/2026/11/12/de… • Google 智慧眼鏡開放預購 Google Labs 副總裁 Josh Woodward 宣布智慧眼鏡官網開賣,預計 12 月 10 日出貨,並在 Warby Parker、Gentle Monster 與 Google 指定實體店提供試戴體驗。 🔗 teapawt.net/joshwoodward/status/21… • Figma 上線 3D Transforms 公測 向所有 Full seat 用戶開放空間旋轉、視角透視與非破壞性排版,原圖層保持即時可編輯。 🔗 figma.com/blog/how-3d-transf… • ggml-org/llama.cpp(單日 +287 星) 極簡 C/C++ 本地 LLM 推論核心。 🔗 github.com/ggml-org/llama.cp… • openai/whisper(單日 +118 星) 弱監督訓練多語音音訊辨識開源模型。 🔗 github.com/openai/whisper
1
1
176
Latest AI Industry Shifts & GitHub Trending Highlights: • Replit Agent Adds Purchasing & Google Ads Replit Agent now handles enterprise supply procurement via Stripe Link & Visa, and deploys Google Ads directly within Replit under code-enforced approvals. 🔗 replit.com/blog/replit-agent… • Jensen Huang Steps Down as Nvidia CEO CFO Colette Kress steps in as interim CEO. Nvidia shipped nearly 14M flagship AI chips (85% market share) over 5 years; lack of a designated successor triggers industry recalibration. 🔗 exponentialview.co/p/jensen-… • Huawei's Horinger Data Center Compute Estimated Epoch AI estimates ~105k Ascend 910C chips active at Horinger via satellite imagery and thermal models (~81k H100 equivalents, or 7% of Colossus 2). 🔗 epochai.substack.com/p/what-… • openai/codex-security (+663 stars) OpenAI's official CLI and TypeScript SDK to identify, validate, and remediate security vulnerabilities in code. 🔗 github.com/openai/codex-secu… • pookjw/HermesEnabler (+390 stars) Open-source utility enabling Hermès Watch Faces across all Apple Watch models. 🔗 github.com/pookjw/HermesEnab…
1
93
AI 產業最新動向與 GitHub 開源趨勢觀察: • Replit Agent 支援企業代購與 Google Ads 透過 Stripe Link 與 Visa 為企業代購物資,並在 Replit 內直接投放跨通路廣告,流程由程式碼強制逐筆審批。 🔗 replit.com/blog/replit-agent… • 黃仁勳辭任 Nvidia CEO 後的產業變局 CFO Colette Kress 暫代職務。Nvidia 過去 5 年售出近 1,400 萬顆旗艦晶片,市佔 85%,缺乏明確接班梯隊引發算力生態重估。 🔗 exponentialview.co/p/jensen-… • 華為和林格爾數據中心算力推估 Epoch AI 透過衛星影像與散熱設備估算,該中心運作約 10.5 萬顆昇騰 910C,算力約相當於 8.1 萬顆 H100,為 Colossus 2 的 7%。 🔗 epochai.substack.com/p/what-… • openai/codex-security (+663 stars) OpenAI 發布的安全性 CLI 與 TypeScript SDK,用於自動偵測、驗證與修復程式碼漏洞。 🔗 github.com/openai/codex-secu… • pookjw/HermesEnabler (+390 stars) 在所有 Apple Watch 裝置上解鎖 Hermès 專屬錶面的開源工具。 🔗 github.com/pookjw/HermesEnab…
180
AI & Open-Source Highlights (Nov 12, 2026) 1. Nvidia Board Weighed CEO Jensen Huang's Role WSJ reported Nvidia's board is deciding whether to remove founder CEO Jensen Huang amid personal relationship inquiries. The board confirmed no company secrets were compromised. 🔗 teapawt.net/rohanpaul_ai/status/21… 2. OpenRouter's 4-Layer Guardrail for AI Agent Infinite Loops OpenRouter benchmarked six models and outlined four defense layers: doom-loop detection, loopKey normalization, hard step/cost caps, and strict API credit limits to stop broken tool retry loops. 🔗 openrouter.ai/blog/tutorials… 3. Runway Open-Sourced 27B World Action Model Praxis-1 Runway released Praxis-1 weights under Apache 2.0. The 27B MoE diffusion Transformer predicts video and robotic actions simultaneously, achieving 46.2% success on NVIDIA RoboLab-120. 🔗 runwayml.com/news/company-ne… 4. nab138/iloader (+68★) A streamlined, open-source sideloader for iOS IPA signing and installation. 🔗 github.com/nab138/iloader 5. dmunozv04/iSponsorBlockTV (+59★) A standalone SponsorBlock client for smart TV YouTube apps, skipping sponsored segments automatically. 🔗 github.com/dmunozv04/iSponso…
WSJ: Nvidia’s board, which publicly backed Jensen Huang on Sunday, is now close to deciding whether to remove its founder CEO. The shift came as a second affair, with an Nvidia employee, surfaced alongside his admitted relationship with a suspected Chinese intelligence asset. When it opened an independent probe on Sunday, the board said it had no reason to believe Nvidia secrets were compromised. Shares fell about 8% after hours, and Huang, the largest shareholder at about 3.6%, has rejected formal succession planning. With roughly 60 direct reports and no named successor, Nvidia’s own proxy report warns that weak succession could hurt strategy and execution.
20
AI 與開源生態最新動態(2026-11-12) 1. Nvidia 董事會評估 CEO 黄仁勋 去留 WSJ 報導,Nvidia 董事會正評估是否撤換創辦人 CEO 黄仁勋。起因為涉入中國情報人員及內部員工婚外情調查,董事會目前認定無機密外洩。 🔗 teapawt.net/rohanpaul_ai/status/21… 2. OpenRouter 發布 Agent 防死循環實測架構 OpenRouter 測試六款主流模型,整理出四層防護機制:doom-loop 偵測、loopKey 參數正規化、步數與成本硬上限、API Key 信用額度,防止 Agent 重複呼叫工具燒光預算。 🔗 openrouter.ai/blog/tutorials… 3. Runway 開源 27B 世界動作模型 Praxis-1 Runway 以 Apache 2.0 授權開放 27B MoE 架構世界動作模型 Praxis-1 權重。可同時預測視訊與動作,在 NVIDIA RoboLab-120 基準以 46.2% 成功率位列第一。 🔗 runwayml.com/news/company-ne… 4. nab138/iloader (+68★) 開源 iOS 側載安裝工具,簡化 IPA 簽名與安裝流程。 🔗 github.com/nab138/iloader 5. dmunozv04/iSponsorBlockTV (+59★) 適用於各大電視版 YouTube 的 SponsorBlock 用戶端,自動跳過影片內建業配與贊助片段。 🔗 github.com/dmunozv04/iSponso…
WSJ: Nvidia’s board, which publicly backed Jensen Huang on Sunday, is now close to deciding whether to remove its founder CEO. The shift came as a second affair, with an Nvidia employee, surfaced alongside his admitted relationship with a suspected Chinese intelligence asset. When it opened an independent probe on Sunday, the board said it had no reason to believe Nvidia secrets were compromised. Shares fell about 8% after hours, and Huang, the largest shareholder at about 3.6%, has rejected formal succession planning. With roughly 60 direct reports and no named successor, Nvidia’s own proxy report warns that weak succession could hurt strategy and execution.
20
地元の名店「松蘭」さん とうとう閉店されたそうです。 地元から古参の飲食店が無くなっていくのは寂しさしかありません。
3
2
43
3,959
AI 最新動態與 GitHub 飆星專案速報 【模型與產業動向】 1. Google 發表 Gemini Decisions API:搭載 Gemini Decision 1,在 150-250ms 延遲內支援文字、圖像、影音與 256K token PDF,單次批次評估高達 1000 道封閉式決策題,瞄準自訓分類器替代市場。 🔗 blog.google/innovation-and-a… 2. Claude Code 發布模型選型指引:官方拆解 Fable 5.2 與 Opus 5.5 在複雜程式碼與開放架構任務的分工,並指出低成本明確任務如何降級至 Sonnet/Haiku 5.5。 🔗 youtube.com/watch?v=C33hJGt-… 3. Grok 構建應用爆大面積數據外洩:資安公司 Claroc 抽檢 37,571 個上線應用,94.7% 零鑑權即可存取。主因在於發布預設公開連結與子網域名稱流入公開 passive DNS,外洩資料含病患隱私與 SpaceXSI 內部管線。 🔗 the-decoder.com/grok-built-a… 【GitHub Trending 精選】 1. krahets/hello-algo(本日 +77 stars):多語言動畫圖解資料結構與演算法教學,支援繁中、Python、C++、Rust、Swift 等一鍵運行。 🔗 github.com/krahets/hello-alg… 2. TheAlgorithms/Java(本日 +108 stars):Java 全演算法純開源實作典藏庫。 🔗 github.com/TheAlgorithms/Jav…
25
Key AI developments and trending GitHub projects: 1. a16z introduces the B2A (Business to Agent) economy AI agents are becoming dominant internet consumers, making buying and subscription decisions directly on behalf of humans. 🔗 a16z.news/p/welcome-to-the-b… 2. Anthropic IPO targets up to $90.4B Bloomberg reports strong institutional demand for Anthropic's IPO. Seeking up to $90.4B in combined capital at a ~$1.4T valuation under ticker ANT. 🔗 teapawt.net/rohanpaul_ai/status/21… 3. xAI faces class-action lawsuits over Grok image generation Victims filed lawsuits alleging Grok misuse in generating illicit content, raising accountability questions for AI labs. 🔗 ithome.com/1/025/809.htm GitHub Trending: • GoogleChrome/modern-web-guidance-src (+133 stars) Google's curated skill set guiding coding agents toward modern native Web APIs and performance best practices. 🔗 github.com/GoogleChrome/mode… • NapNeko/NapCatQQ (+141 stars) A modern protocol framework built on NTQQ architecture for bot automation. 🔗 github.com/NapNeko/NapCatQQ
Bloomberg: Anthropic’s IPO has drawn orders for several times the shares available, mostly from anchor-size indications of interest (i.e. very large orders, from big institutional investors) Anthropic is seeking up to $90.4B through the IPO and a private placement, which would exceed SpaceX’s $85.7B record and make it the largest listing ever. The roughly $1.4T fully diluted valuation target sits below some investors’ expectations, a common setup that lets strong demand lift the final price. Secondary activity has already started, with some investors offering pieces of their allocations to others for fees and pre-IPO holders selling stakes through special purpose vehicles. Goldman Sachs and Morgan Stanley lead the deal with 20 other banks, and the stock is expected to trade on the Nasdaq Global Select Market as ANT.
11
AI 產業重大動態與 GitHub 開源熱點整理: 1. a16z 提出 B2A(Business to Agent)經濟 a16z 發布最新分析,指出 AI 智慧體正逐步接管傳統網路流量與商業決策,代用戶決定購買與訂閱行為,企業商業模式將全面轉向面向智慧體設計。 🔗 a16z.news/p/welcome-to-the-b… 2. Anthropic IPO 傳募資上看 904 億美元 彭博社報導,Anthropic IPO 已獲機構投資人超額認購,預計透過公開募股加私募籌集最高 90.4B 美元,完全稀釋估值目標約 1.4 兆美元,代號 ANT。 🔗 teapawt.net/rohanpaul_ai/status/21… 3. xAI 旗下 Grok 面臨兒少保護集體訴訟 多名受害者針對 Grok 遭濫用改圖問題對 xAI 提起集體訴訟,引發市場針對生成式 AI 法律責任與監管邊界的高度關注。 🔗 ithome.com/1/025/809.htm GitHub 趨勢專案: • GoogleChrome/modern-web-guidance-src (+133 stars) Google 官方釋出的 Coding Agent 網頁指引規範,協助 AI 程式碼助理優先採用原生現代 Web API,降低老舊冗餘代碼。 🔗 github.com/GoogleChrome/mode… • NapNeko/NapCatQQ (+141 stars) 基於 NTQQ 核心架構的現代化協定端框架,提供穩定高效的機器人串接方案。 🔗 github.com/NapNeko/NapCatQQ
Bloomberg: Anthropic’s IPO has drawn orders for several times the shares available, mostly from anchor-size indications of interest (i.e. very large orders, from big institutional investors) Anthropic is seeking up to $90.4B through the IPO and a private placement, which would exceed SpaceX’s $85.7B record and make it the largest listing ever. The roughly $1.4T fully diluted valuation target sits below some investors’ expectations, a common setup that lets strong demand lift the final price. Secondary activity has already started, with some investors offering pieces of their allocations to others for fees and pre-IPO holders selling stakes through special purpose vehicles. Goldman Sachs and Morgan Stanley lead the deal with 20 other banks, and the stock is expected to trade on the Nasdaq Global Select Market as ANT.
1
28
Anthropic founder IPO details surface alongside new tiered agent architectures and browser automation tools on GitHub. Key AI updates & GitHub trending projects: 1. Anthropic founders to sell $250M each in IPO while retaining control The Wall Street Journal reports Anthropic's IPO targets a valuation near $1.3T, seeking over $80B in capital. Seven founders will each sell roughly 3.4M shares ($70–$78 expected range), keeping ~$18B in holdings alongside dual-class voting control. 🔗 teapawt.net/rohanpaul_ai/status/21… 2. OpenRouter benchmarks fp8 vs fp4 quantization quality and pricing OpenRouter evaluated 1,455 provider endpoints and Qwen3.8 27B across 280 test tasks, quantifying accuracy trade-offs and inference cost reductions between fp8 and fp4 alongside provider.quantizations routing rules. 🔗 openrouter.ai/blog/insights/… 3. shadcn/improve (+333 stars) A code audit workflow from shadcn that delegates repository inspection and plan drafting to high-end frontier models, leaving execution steps to cheaper models. 🔗 github.com/shadcn/improve 4. browser-use/browser-use (+171 stars) An open-source Python library for web browser agents, featuring stealth browser automation and the hosted Browser Use Cloud agent API. 🔗 github.com/browser-use/brows…
WSJ: Anthropic’s 7 founders will each sell about $250M of stock in its IPO and keep roughly $18B apiece. Each plans to sell nearly 3.4M shares at a projected range of $70 to $78, according to Tuesday’s company filing. The listing, expected next week, would value Anthropic near $1.3T and raise $80B or more, partly via a private placement to investors including Google. Once trading starts, each founder keeps about 3.2% of Class A shares, so together they hold roughly 22% of outstanding stock. That stake carries only about 11.2% of voting power, yet a founder-controlled LLC holding a single Class F share guarantees them 50.1% on most shareholder votes Public buyers of Anthropic's shares therefore get economic exposure with almost no say over control, which the filing warns could put decisions at odds with shareholders’ financial interests.
8
Anthropic IPO 股權細節曝光,開源社群湧現高階模型分工架構與瀏覽器代理新進展。 今日 AI 熱點與 GitHub 飆星專案: 1. Anthropic 7 位創辦人傳 IPO 各套現 2.5 億美元並保留控制權 華爾街日報報導,Anthropic 招股估值接近 1.3 兆美元、預計募資逾 800 億美元。7 位創辦人每人出售近 340 萬股(發行價區間 70 至 78 美元),保留約 180 億美元持倉與雙重股權結構。 🔗 teapawt.net/rohanpaul_ai/status/21… 2. OpenRouter 實測:fp8 與 fp4 量化對模型品質與成本的實質影響 OpenRouter 針對 1,455 個供應商端點與 Qwen3.8 27B 進行 280 項實測,公布 fp8 與 fp4 的性能折損與推論成本對比數據,並整理 provider.quantizations 調用指南。 🔗 openrouter.ai/blog/insights/… 3. shadcn/improve (+333 stars) shadcn 開源的代碼審查與重構工具。架構採用雙層分工:指派高階模型審核整個代碼庫並擬定計畫,再調度低成本模型分步執行。 🔗 github.com/shadcn/improve 4. browser-use/browser-use (+171 stars) AI 瀏覽器代理 Python 函式庫,支援擬真網頁自動化,並提供 Browser Use Cloud 託管代理 API。 🔗 github.com/browser-use/brows…
WSJ: Anthropic’s 7 founders will each sell about $250M of stock in its IPO and keep roughly $18B apiece. Each plans to sell nearly 3.4M shares at a projected range of $70 to $78, according to Tuesday’s company filing. The listing, expected next week, would value Anthropic near $1.3T and raise $80B or more, partly via a private placement to investors including Google. Once trading starts, each founder keeps about 3.2% of Class A shares, so together they hold roughly 22% of outstanding stock. That stake carries only about 11.2% of voting power, yet a founder-controlled LLC holding a single Class F share guarantees them 50.1% on most shareholder votes Public buyers of Anthropic's shares therefore get economic exposure with almost no say over control, which the filing warns could put decisions at odds with shareholders’ financial interests.
23
Anthropic reveals over $10B in compute commitments in amended S1, while open-source ecosystems and agent security evaluations gain momentum. Key AI updates & GitHub trending projects: 1. Anthropic discloses multi-billion compute deals with Microsoft & Google Anthropic's amended S1 filing reveals revised cloud commitments. Microsoft commits at least $47.2B through June 2033, while Google increases its minimum commitment to $121.3B. Both are irrevocable take-or-pay agreements regardless of actual capacity usage. The filing also notes $5.0B in short-term compute spend last quarter at higher prices. 🔗 teapawt.net/rohanpaul_ai/status/21… 2. OpenRouter benchmarks context token costs: MCP vs CLI vs Skills OpenRouter evaluated context consumption across tool architectures on identical GitHub tasks. On Claude Sonnet 3.5/5.5, CLI execution consumes significantly fewer context tokens than full MCP server schemas. The team outlines scoped tool definitions and Tool Search as practical mitigations for agent overhead. 🔗 openrouter.ai/blog/insights/… 3. Berkeley RDI publishes CyberBinGym benchmark for 1-day exploit synthesis UC Berkeley released CyberBinGym across 53 open-source projects and 100 real-world tasks. Evaluating agents on stripped vulnerable and patched binaries without source code or bug descriptions, results show agents like Codex locate 88% of patches but generate working exploits on only 20%, with hardened binaries lowering success further. 🔗 rdi.berkeley.edu/blog/cyberb… 4. ashishpatel26/500-AI-Agents-Projects (+126 stars) A curated collection of 500 AI agent production use cases across healthcare, finance, education, and retail, complete with architecture breakdowns and implementation links. 🔗 github.com/ashishpatel26/500… 5. fosrl/pangolin (+108 stars) A modern networking and identity-aware security platform built with TypeScript, delivering zero-trust access control and secure connectivity for infrastructure and AI workloads. 🔗 github.com/fosrl/pangolin
Anthropic reveals in its amended S1 filing today, more than $10 billion in new agreements with Microsoft and Google. The Microsoft deal now covers at least $47.2B and up to $49.2B of spending through June 2033, up from a $44.9B minimum. Google’s minimum rose from about $114.5B to about $121.3B through July 2033, and both deals must be paid whether or not Anthropic uses the capacity. That extra long-term capacity could ease its reliance on short-term compute, which cost $5.0B last quarter at significantly higher prices.
32
Anthropic 修訂版 S1 揭露超百億美元算力協議,開源生態與智慧體安全性測試迎來關鍵進展。 今日 AI 熱點與 GitHub 飆星專案: 1. Anthropic S1 修訂文件揭露微軟與 Google 算力承諾 Anthropic 修訂 S1 招股文件,揭露與 Microsoft、Google 達成超百億美元算力協議。微軟承諾至 2033 年支出至少 472 億美元,Google 承諾額提高至 1,213 億美元;兩筆均為不可撤銷的照付不議(Take-or-pay)條款。上一季短期算力支出達 50 億美元且單位成本顯著提升。 🔗 teapawt.net/rohanpaul_ai/status/21… 2. OpenRouter 實測 MCP vs CLI vs Skills 上下文 Token 成本 OpenRouter 針對同一 GitHub 任務實測不同工具架構的 Token 消耗。在 Claude Sonnet 3.5/5.5 下,CLI 架構消耗的上下文大幅低於完整 MCP 協議,並提出透過縮窄工具定義與 Tool Search 降低 Agent 運算開銷的具體方案。 🔗 openrouter.ai/blog/insights/… 3. Berkeley RDI 發布 CyberBinGym 二進位漏洞利用基準 UC Berkeley 推出 CyberBinGym,涵蓋 53 個開源專案、100 個 1-day 任務。測試顯示在無源碼與漏洞描述下,AI 智慧體(如 Codex)能辨識 88% 的修補位址,但僅有 20% 能獨立寫出真實可運行的 Exploit,編譯加固更進一步壓低成功率。 🔗 rdi.berkeley.edu/blog/cyberb… 4. ashishpatel26/500-AI-Agents-Projects (+126 stars) 涵蓋醫療、金融、教育與零售等領域的 500 個 AI Agent 實戰案例庫,收錄完整開源專案連結與架構設計,便於快速對照不同業務場景的落地實作。 🔗 github.com/ashishpatel26/500… 5. fosrl/pangolin (+108 stars) 面向現代基礎架構、微服務與 AI 工作負載的安全連線平台,基於 TypeScript 構建,提供零信任存取控制與邊界網路隔離。 🔗 github.com/fosrl/pangolin
Anthropic reveals in its amended S1 filing today, more than $10 billion in new agreements with Microsoft and Google. The Microsoft deal now covers at least $47.2B and up to $49.2B of spending through June 2033, up from a $44.9B minimum. Google’s minimum rose from about $114.5B to about $121.3B through July 2033, and both deals must be paid whether or not Anthropic uses the capacity. That extra long-term capacity could ease its reliance on short-term compute, which cost $5.0B last quarter at significantly higher prices.
1
25
Stripe acquires OpenRouter to consolidate AI infra, while hardware supply chains and open-source ecosystems face major tremors. Key AI updates & GitHub trending projects: 1. Stripe acquires OpenRouter OpenRouter officially joins Stripe. Powering over 10M developers and enterprises with unified model routing, failover protection, and spending controls, it commits to maintaining platform neutrality. 🔗 teapawt.net/OpenRouter/status/2120… 2. Deep dive into Jensen Huang scandal & GPU export trails Ed Zitron consolidates reporting from Bloomberg, NYT, FT, and Culper Research, examining Jensen Huang's disclosed relationship and Megaspeed's alleged role routing restricted NVIDIA GPUs into China. 🔗 wheresyoured.at/jensens-biza… 3. MiniMax releases open-weight H3-Image MiniMax unveils H3-Image, an open-weight generator delivering 4096 × 4096 4K images in under 10 seconds, accelerating inference up to 5x over mainstream alternatives. 🔗 minimax.io/blog/minimax-h3-i… 4. tokio-rs/topcoat (+192 stars) A batteries-included Rust web framework built by the Tokio team, delivering a modern full-stack development experience. 🔗 github.com/tokio-rs/topcoat 5. we-promise/sure (+184 stars) An open-source personal finance application built with Ruby, focusing on privacy, simplicity, and community collaboration. 🔗 github.com/we-promise/sure
This quoted post is unavailable.
17
Stripe 正式收購 OpenRouter,AI 基建板塊加速整合;開源生態與硬體供應鏈同步迎來重磅動盪。 今日 AI 熱點與 GitHub 飆星專案: 1. Stripe 收購 OpenRouter Stripe 官方宣布併購 OpenRouter。服務已覆蓋逾 1,000 萬開發者與企業,整合後將持續維持中立架構,並為團隊提供多模型調度、故障切換與支出控制。 🔗 teapawt.net/OpenRouter/status/2120… 2. Ed Zitron 長文剖析黃仁勳事件與 GPU 流向 整合 Bloomberg、NYT、FT 與 Culper 報告,深入梳理黃仁勳承認之婚外情事件,並追蹤 Megaspeed 疑似作為轉口實體非法將 NVIDIA GPU 輸入中國之金流與供應鏈證據。 🔗 wheresyoured.at/jensens-biza… 3. MiniMax 發布開源生圖模型 H3-Image 基於 H3 架構之開源權重圖像生成模型,支援文生圖、參考圖與編輯,可於 10 秒內生成 4096 × 4096 4K 圖像,推論速度達主流架構 5 倍。 🔗 minimax.io/blog/minimax-h3-i… 4. tokio-rs/topcoat (+192 stars) Tokio 團隊推出的開箱即用 Rust Web 全端開發框架,提供一體化的非同步後端與現代 Web 構建體驗。 🔗 github.com/tokio-rs/topcoat 5. we-promise/sure (+184 stars) 社群驅動的個人財務管理開源應用,採用 Ruby 構建,強調極簡記帳、隱私自主與全民共建。 🔗 github.com/we-promise/sure
This quoted post is unavailable.
17
Gemini 4 Carbon is now live, shaking both developer workflows and commercial expectations. Key AI updates & GitHub trending projects: 1. Gemini 4 Carbon debuts across API and Antigravity Google DeepMind rolls out Gemini 4 Carbon on Gemini API, Google AI Studio, and Antigravity, featuring single-prompt 3D reconstruction demos. 🔗 teapawt.net/_philschmid/status/212… 2. Gary Marcus challenges Anthropic's $1.4T IPO valuation Scrutinizing Anthropic's reported target valuation, Marcus highlights nearly $900B in future compute commitments and Q3 sequential revenue growth decelerating from 147% to 47%. 🔗 garymarcus.substack.com/p/ju… 3. Lovable introduces built-in security scanner and auto-fix Lovable adds pre-deploy checks covering code, logic, schema, and Row Level Security (RLS), letting agents patch open data access rules in one click. 🔗 lovable.dev/blog/find-it-fix… 4. homeassistant-ai/ha-mcp (+792 stars) Connect Claude, ChatGPT, and AI assistants directly to Home Assistant via Model Context Protocol (MCP) for home automation. 🔗 github.com/homeassistant-ai/… 5. WhiteTowerAI/ohmygame (+201 stars) Open-source one-person game studio integrating AI game canvas and multi-model APIs with bring-your-own-key support. 🔗 github.com/WhiteTowerAI/ohmy…
Gemini 4 Carbon is now available in the Gemini API, @GoogleAIStudio, and @Antigravity (Google AI Ultra subscribers). At Google, teams have already improved open-source vulnerability recall by 25%+ over Argon and found 40x more remote protein homologs for AlphaFold 3. I've been using it for past weeks for all my tasks (default is medium effort for me). My tasks already got 10x more ambitious, but it still feels that i need to at least 100x more ambitious. Here is an example of that using /swarm in @Antigravity: turning 7 public planning PDFs (206 engineering drawings) into an interactive 3D reconstruction of our London Platform 37 office overnight from a single prompt.
13