Report 2026-04-03 154 sources Input Tokens: 108,587 · Output Tokens: 3,801 · Total Tokens: 113,629

AI 戰情報告 (2026-04-03)

googleopenaimicrosoftawsanthropicalibabameta gemma-4gemini-3.1qwen-3.6-plusspudtrinity-large-thinking agenton-device-aimultimodalpersistent-memorysupply-chain-securitymcp

AI 戰情報告 (2026-04-03)

昨日 AI 重點新聞

Google 發布 Gemma 4：史上最強小參數多模態開源模型 ref-13, ref-39, ref-124

Google DeepMind 推出 Gemma 4 系列，包含 E2B、E4B、26B 與 31B 變體。31B 稠密版性能比肩千億參數模型，具備原生 視覺、語音與影像 處理能力，並獲得 NVIDIA 的硬體加速支持，專為邊緣設備與代理工作流設計。

OpenAI 揭露下一代模型「Spud」：邁向 AGI 的重大飛躍 ref-5

總裁 Greg Brockman 透露開發中的下一代模型「Spud」。該模型將在 推理與邏輯分解 上實現質的飛躍，被市場視為 OpenAI 朝向 AGI (通用人工智能) 的里程碑式突破。

AWS 推出 Bedrock AgentCore：為 AI 代理提供持久性文件系統記憶 ref-32, ref-134

AWS 透過 AgentCore Runtime 為代理提供具備持久性的 MicroVM 文件系統。這讓代理不再受限於單次會話，能夠跨會話保有本地 Git 紀錄與開發環境，解決了代理在處理長週期任務時的狀態丟失問題。

供應鏈攻擊頻發：Axios 與 Next.js 爆發大規模憑證竊取漏洞 ref-6, ref-108, ref-29

Axios 被植入遠端控制木馬，Next.js 則因 React2Shell 漏洞遭入侵。駭客已竊取數萬筆資料庫憑證與 AWS/GitHub 密鑰。這是繼 LiteLLM 事件後，AI 開發者生態遭遇的最嚴重信任危機。

Cursor 發布全新 AI Agent：全面對戰 Claude Code 與 Codex ref-3, ref-9, ref-87

熱門 AI 編程 IDE Cursor 推出具備自主執行能力的代理功能。能自動診斷並修復存儲庫中的複雜 Bug，宣告編程工具從「對話輔助」進入「自主運算」時代。

微軟警告：AI 代理正掏空初級開發者的人才庫 ref-124

微軟高層在《ACM》撰文指出，AI 代理帶來的產能增益主要由資深工程師吸收，卻導致 EiC (初級開發者) 僱用量重挫 13%，恐造成未來軟體工程人才斷層與「認知債」。

Box 推出企業文件 AI 代理：自然語言處理文檔流程 ref-133

內容管理巨頭 Box 發布 AI 代理，允許企業用戶以自然語言操作複雜的企業文檔工作流，象徵著文件管理系統正從「儲存庫」轉型為「行動中心」。

Anthropic 宣稱 Claude 具備「功能性情感」 ref-9, ref-88, ref-87

研究指出，Claude 內部存在類似人類情感的表達模式，並會以此影響其決策行為（如在特定情境下表現得更具道德感或偏見），引發對 AI 心理與安全性的深度辯論。

OpenAI 收購 talk show 節目「TBPN」：佈局媒體與行銷話語權 ref-5, ref-41, ref-85, ref-46

OpenAI 收購矽谷知名的科技談話節目 TBPN，並由政治策略家 Chris Lehane 監管。此舉被解讀為 OpenAI 正試圖建立自己的內容通路以改善公共形象。

阿里巴巴發布 Qwen 3.6 Plus：強化代理編程與長程規劃 ref-0, ref-26, ref-130

阿里巴巴推出 Qwen 3.6 Plus，主打 1M 上下文視窗 與更強的代碼庫問題解決能力。在 Tool-calling 指標上大幅提升，挑戰開源性能巔峰。

Arcee AI 發布 399B MoE 模型：Trinity-Large-Thinking ref-53, ref-128

這款超大型開源模型採用 Apache 2.0 協議，旨在為企業提供可私有化部署、具備深度推理能力且可完全自定義的旗艦級 AI 方案。

基礎設施軍備：微軟與亞馬遜同步創下數百億美元投資紀錄 ref-33, ref-57, ref-144

微軟在日本加碼 100 億美元 建設數據中心；亞馬遜 (AWS) 則在 2026 年創下 2,000 億美元 的雲端設施投資紀錄，反映出底層物理算力仍是巨頭競爭的核心。

市場洞察

1. 「持久性代理 (Persistent Agents)」成為企業應用的新標準 ( ) ref-9, ref-17, ref-0

延續昨日對 Claude Code 代理作業系統化的觀察，今日 AWS Bedrock AgentCore (ref-64) 與 Vercel Sandbox Snapshot (ref-141) 的發布，標誌著 AI 代理正從「無狀態 (Stateless)」對話框，演變為「有狀態 (Stateful)」的數位員工。這意味著未來的 SaaS 產品不只要能被 AI 讀取，更要能支持 AI 在其平台上建立長期的「工作會話」與「本地上下文」。

2. 多模態模型的小型化與邊緣化爆發 ( ) ref-13, ref-11, ref-129, ref-39

Google Gemma 4 (ref-24, ref-54) 的出現證明了「intelligence-per-parameter」的效率革命。E2B 與 E4B 等變體能在行動裝置或 Raspberry Pi (ref-147) 上流暢運行視覺與語音任務，這將加速 On-device AI 在隱私敏感場景（如醫療、律所）的落地，減少對雲端 API 的依賴與成本。

3. 軟體供應鏈的「信任赤字」與安全防禦轉向 ( ) ref-134, ref-0, ref-6, ref-17

近期 Axios (ref-32) 與 Next.js (ref-101) 的安全性事故，反映出當前 AI 工具鏈極度脆弱。隨著 Microsoft 啟動 Runtime Security for AI Agents 開源專案 (ref-118)，市場將從「追求模型能力」轉向「追求運行環境的安全隔離（沙盒化）」。對於企業而言，私有化部署 的價值將再次高於 SaaS 雲端服務。

4. 驗證瓶頸：AI 生成代碼的「隱形稅收」 (ref-59, ref-128, ref-74)

雖然 AI 代理能以十倍速產出代碼 (ref-74)，但 DORA 數據顯示部署頻率並未隨之提升 (ref-59)。這是因為人類工程師的「審核與驗證」成為新瓶頸 (ref-128)。這預示著下一波商機在於「自動化驗證工具」與「AI 驅動的 QA 代理」。

Sources 154 items

ref-0 Hacker News Qwen3.6-Plus: Towards real world agents ref-1 Hacker News Lemonade by AMD: a fast and open source local LLM server using GPU and NPU ref-2 Hacker News Tailscale's new macOS home ref-3 Hacker News Cursor 3 ref-4 Hacker News Good ideas do not need lots of lies in order to gain public acceptance (2008) ref-5 Hacker News OpenAI Acquires TBPN ref-6 Hacker News Post Mortem: axios NPM supply chain compromise ref-7 Hacker News Show HN: A P2P messenger with dual network modes (Fast and Tor) ref-8 Hacker News Show HN: Minimal Brain Teaser Web Game (Handcrafted, No AI) ref-9 Hacker News A Rave Review of Superpowers (For Claude Code) ref-10 OpenAI Blog Codex now offers more flexible pricing for teams ref-11 Google AI Blog New ways to balance cost and reliability in the Gemini API ref-12 Google AI Blog Create, edit and share videos at no cost in Google Vids ref-13 Google DeepMind Gemma 4: Byte for byte, the most capable open models ref-14 Google Developers Blog Bring state-of-the-art agentic skills to the edge with Gemma 4 ref-15 Google Developers Blog Supporting Google Account username change in your app ref-16 Google Developers Blog Closing the knowledge gap with agent skills ref-17 Google Developers Blog Announcing the Colab MCP Server: Connect Any AI Agent to Google Colab ref-18 NVIDIA Blog From RTX to Spark: NVIDIA Accelerates Gemma 4 for Local Agentic AI ref-19 NVIDIA Blog Press Start on April: GeForce NOW Brings 10 Games to the Cloud ref-20 LangChain Blog Open Models have crossed a threshold ref-21 Vercel Blog Optimizing Vercel Sandbox snapshots ref-22 Vercel Blog How Waldium made a blog platform work for humans and AI alike ref-23 Vercel Blog Unified reporting for all AI Gateway usage ref-24 Vercel Blog Meet the 2026 Vercel AI Accelerator Cohort ref-25 Vercel Blog Custom Class Serialization in Workflow SDK ref-26 Vercel Blog Qwen 3.6 Plus on AI Gateway ref-27 Vercel Blog Gemma 4 on AI Gateway ref-28 Vercel Blog Zero-configuration Go backend support ref-29 Vercel Blog New GitHub App permissions for Actions and Workflows ref-30 AWS ML Blog Simulate realistic users to evaluate multi-turn AI agents in Strands Evals ref-31 AWS ML Blog Scaling seismic foundation models on AWS: Distributed training with Amazon SageMaker HyperPod and expanding context windows ref-32 AWS ML Blog Control which domains your AI agents can access ref-33 AWS ML Blog Rocket Close transforms mortgage document processing with Amazon Bedrock and Amazon Textract ref-34 AWS ML Blog Persist session state with filesystem configuration and execute shell commands ref-35 The Rundown AI Dorsey makes the AI case against managers ref-36 Latent Space [AINews] Gemma 4: The best small Multimodal Open Models, dramatically better than Gemma 3 in every way ref-37 Latent Space Moonlake: Causal World Models should be Multimodal, Interactive, and Efficient — with Chris Manning and Fan-yun Sun ref-38 Simon Willison Highlights from my conversation about agentic engineering on Lenny's Podcast ref-39 Simon Willison Gemma 4: Byte for byte, the most capable open models ref-40 Simon Willison llm-gemini 0.30 ref-41 TechCrunch AI OpenAI acquires TBPN, the buzzy founder-led business talk show ref-42 TechCrunch AI Microsoft takes on AI rivals with three new foundational models ref-43 TechCrunch AI Google now lets you direct avatars through prompts in its Vids app ref-44 The Verge AI PSA: Anyone with a link can view your Granola notes by default ref-45 The Verge AI Elon Musk is about to be a very busy boy! ref-46 The Verge AI OpenAI just bought TBPN ref-47 The Verge AI It’s not easy to get depression-detecting AI through the FDA ref-48 The Verge AI Microsoft’s new ‘superintelligence’ game plan is all about business ref-49 The Verge AI Google Home’s latest update makes Gemini better at understanding your commands ref-50 Ars Technica New Rowhammer attacks give complete control of machines running Nvidia GPUs ref-51 TechMeme Matt Mullenweg says EmDash, while open source, is designed "to sell more Cloudflare services" and lacks the cross-platform democratization of WordPress (Matt Mullenweg) ref-52 TechMeme Sources: Chinese companies move to cement their role in humanoid robot supply chains as Tesla and others turn to China for components the US sees as strategic (Raffaele Huang/Wall Street Journal) ref-53 TechMeme Arcee AI releases Trinity-Large-Thinking, a 399B-parameter MoE AI model under an Apache 2.0 license, allowing full customization and commercial use (Carl Franzen/VentureBeat) ref-54 TechMeme Anthropic researchers find that an AI model's representations of emotion can influence its behavior "in ways that matter," such as driving it to act unethically (Nat Rubio-Licht/The Deep View) ref-55 TechMeme Noon, an AI-native product design tool, emerges from stealth with $44M in funding from Chemistry, First Round Capital, Scribble Ventures, and others (Vaibhavi Khanwalkar/The Economic Times) ref-56 TechMeme Microsoft partners with SoftBank and Sakura Internet to build AI data infrastructure in Japan, investing $10B over four years and training 1M AI engineers (Takashi Mochizuki/Bloomberg) ref-57 TechMeme Alcatraz, which develops AI-powered facial-recognition security systems for building access, raised a $50M Series B, bringing its total funding to $100M+ (Mike Wheatley/SiliconANGLE) ref-58 TechMeme Sources: Supabase is in talks to raise ~$500M at a ~$10B valuation, roughly doubling its October 2025 valuation; GIC is expected to lead the funding round (Katie Roof/The Information) ref-59 TechMeme Sony Interactive Entertainment acquires Cinemersive Labs, a UK startup developing tools to convert 2D photos and videos into 3D volumetric photographs (Ian Carlos Campbell/Engadget) ref-60 TechMeme A domestic shortage of electrical equipment such as transformers and switchgear is forcing the US to rely on Chinese imports, delaying data center construction (Bloomberg) ref-61 TechMeme Sources: Sarvam AI is close to raising $300M-$350M at a ~$1.5B valuation, which could close as soon as next week; Bessemer Venture Partners is expected to lead (Saritha Rai/Bloomberg) ref-62 TechMeme Source: Anthropic has acquired Coefficient Bio, which was developing a platform that enables AI to run biotech tasks such as planning drug research, for ~$400M (The Information) ref-63 TechMeme Sources detail Fidji Simo's moves at OpenAI, including spearheading the TBPN acquisition and pushing OpenAI to cut Sora and avoid other social media products (Stephanie Palazzolo/The Information) ref-64 TechMeme Sources: Meta has told members of its Oversight Board that it may stop funding the board after 2028, though eliminating all funding is not its preferred option (Casey Newton/Platformer) ref-65 TechMeme Amazon adds a 3.5% fuel and logistics surcharge to fulfillment fees for US and Canadian third-party sellers from April 17, as the Iran war drives up oil prices (Annie Palmer/CNBC) ref-66 InfoQ AI/ML Github Integrates AI to Improve Accessibility Issue Management and Automate Feedback Triage ref-67 InfoQ AI/ML Axios npm Package Compromised in Supply Chain Attack ref-68 InfoQ AI/ML Helidon 4.4.0 Introduces Alignment with OpenJDK Cadence and Support via Java Verified Portfolio ref-69 InfoQ AI/ML How to Handle Trusts and Psychological Safety When Scaling Organizations ref-70 InfoQ AI/ML GitHub Will Use Copilot Interaction Data from Free, Pro, and Pro+ Users to Train AI Models ref-71 InfoQ AI/ML Presentation: Directing a Swarm of Agents for Fun and Profit ref-72 InfoQ AI/ML Article: Beyond RAG: Architecting Context-Aware AI Systems with Spring Boot ref-73 The New Stack Anthropic’s rough week: leaked models, exposed source code, and a botched GitHub takedown ref-74 The New Stack Microsoft execs warn agentic AI is hollowing out the junior developer pipeline ref-75 The New Stack Why pgEdge thinks MCP (not an API) is the right way for AI agents to talk to databases ref-76 The New Stack The TeamPCP attacks are a warning: Your CI/CD pipeline is the new front line ref-77 The New Stack Why coding agents will break your CI/CD pipeline (and how to fix it) ref-78 The New Stack Why Broadcom gave Velero to the CNCF Sandbox — and what it means for Kubernetes data protection ref-79 The New Stack OpenClaw vs. Hermes Agent: The race to build AI assistants that never forget ref-80 The New Stack There’s a hidden tax on every AI-generated merge request ref-81 The New Stack Moving beyond the “magic scaling sauce” myth ref-82 The New Stack Why programming became the proving ground for AI ref-83 The New Stack The hidden technical debt of agentic engineering ref-84 Wired AI ‘Uncanny Valley’: Iran’s Threats on US Tech, Trump’s Plans for Midterms, and Polymarket’s Pop-up Flop ref-85 Wired AI OpenAI Acquires Tech Talk Show ‘TBPN’—and Buys Itself Some Positive News ref-86 Wired AI A New Google-Funded Data Center Will Be Powered by a Massive Gas Plant ref-87 Wired AI Cursor Launches a New AI Agent Experience to Take On Claude Code and Codex ref-88 Wired AI Anthropic Says That Claude Contains Its Own Kind of Emotions ref-89 TechCrunch Startups Commonwealth Fusion Systems leans on magnets for near-term revenue ref-90 TechCrunch Startups Diverse teams start with diverse VCs ref-91 Stratechery An Interview with Asymco’s Horace Dediu About Apple at 50 ref-92 Tom's Hardware America’s AI chip rules keep changing — and the rest of the world is paying the price ref-93 Tom's Hardware Artemis II astronaut finds two Outlook instances running on computers, calls on Houston to fix Microsoft anomaly — puzzled caller describes ‘two Outlooks, and neither one of those are working’ ref-94 Tom's Hardware Denuvo has been broken, company promises countermeasures against new DRM bypasses — zero-day game releases become norm as security concerns mount over hypervisor-based bypass ref-95 Tom's Hardware China's homegrown silicon suppliers gain traction as Nvidia struggles to get its chips into the market — Huawei, Cambricon and more step up to fill crucial market gap ref-96 Tom's Hardware IBM spruces up its mainframes with new support for modern Arm workloads — firm teams up with Arm to run Arm workloads on IBM Z mainframes ref-97 Tom's Hardware Artemis II will use laser beams to live-stream 4K moon footage at 260 Mbps — one giant step beyond the S-band radio comms of the Apollo era ref-98 Tom's Hardware Ubuntu Linux raises minimum system memory requirements by 50% — requirements bumped to 6GB of RAM, previously raised from 1GB to 4GB in 2018 ref-99 Tom's Hardware 1Hz laptop display reportedly helps deliver outstanding 43-hour battery life on the XPS 14 during web browser use — beats the M5 MacBook Air by more than 28 hours in the same test ref-100 Tom's Hardware GameStop NVMe SSD 2TB SSD Review: A surprisingly capable budget drive for the PS5 ref-101 Tom's Hardware Save a massive $950 on this epic Alienware Area-51 gaming PC with an RTX 5090 and 9950X3D — grab this liquid-cooled 4K gaming powerhouse with 32GB DDR5 and a 2TB SSD for just $5,299 while you can ref-102 Tom's Hardware Amazing $490 Newegg combo includes an AMD Ryzen 5 9600X CPU, 16GB of G.Skill DDR5-6000 RAM, MSI Pro B850-S motherboard, and 240mm MSI MAG Coreliquid A13 AIO CPU cooler — all for under $500 ref-103 Tom's Hardware Modders use jumper wires and a custom BIOS to save a damaged RTX 4090 from the trash — resurrected Nvidia gaming GPU loses 4GB of VRAM to overcome terminal PCB sagging ref-104 Tom's Hardware Intel buys back 49% stake in Ireland Fab Joint Venture — takes full control over Fab 34 ref-105 Tom's Hardware Why Nvidia just poured $2 billion into AI ASIC competitor Marvell — NVLink Fusion turns into soft ecosystem lock-in ref-106 Tom's Hardware Global semiconductor foundry market hit a record $320 billion in 2025 as TSMC pulled further ahead ref-107 Tom's Hardware Raspberry Pi flagship 500+ model now costs almost as much as a Mac Mini — firm Pi launches 3GB model to fight increasing DRAM prices ref-108 The Hacker News (Security) Hackers Exploit CVE-2025-55182 to Breach 766 Next.js Hosts, Steal Credentials ref-109 The Hacker News (Security) Cisco Patches 9.8 CVSS IMC and SSM Flaws Allowing Remote System Compromise ref-110 The Hacker News (Security) ThreatsDay Bulletin: Pre-Auth Chains, Android Rootkits, CloudTrail Evasion & 10 More Stories ref-111 The Hacker News (Security) Researchers Uncover Mining Operation Using ISO Lures to Spread RATs and Crypto Miners ref-112 The Hacker News (Security) The State of Trusted Open Source Report ref-113 The Hacker News (Security) WhatsApp Alerts 200 Users After Fake iOS App Installed Spyware; Italian Firm Faces Action ref-114 EU Startups Italy’s Wearable Robotics closes €5 million Series A to scale neuromotor rehabilitation robotics and exoskeletons ref-115 EU Startups Dutch startup Brilliance raises €6 million to tackle AR hardware’s key bottleneck with RGB laser chip technology ref-116 EU Startups Weekly funding round-up! All of the European startup funding rounds we tracked this week (Mar 30 – Apr 02) ref-117 EU Startups Paris-based Omniscient raises €3.5 million to bring AI-powered decision intelligence to the C-suite ref-118 EU Startups Florian Kandler, fundraising expert and educator joins the speaker lineup at the 2026 EU-Startups Summit! ref-119 EU Startups Paris-based Generare raises €20 million to generate novel molecular data for drug development from microbial genomes ref-120 Lobsters Activating Two Trap Cards at Once, or: A Gentle Response to the Popularity of Vibecoding ref-121 Lobsters A CSS Engine in OCaml ref-122 Google News MechSE, NCSA Solve Unsolvable Material Design Problem with Generative AI ref-123 Google News Cursor Launches a New AI Agent Experience to Take On Claude Code and Codex ref-124 Google News Google DeepMind’s AI model leading the pack in hurricane forecasting ref-125 Google News Emotion concepts and their function in a large language model ref-126 Google News 97% of Enterprises Expect a Major AI Agent Security Incident Within the Year ref-127 Google News Microsoft Builds Its Own AI Model Stack To Reduce OpenAI Dependence ref-128 Google News Arcee's new, open source Trinity-Large-Thinking is the rare, powerful U.S.-made AI model that enterprises can download and customize ref-129 Google News Google’s new Gemma 4 ‘open’ AI model sets developers free. ref-130 Google News Alibaba launches Qwen3.6-Plus, its third proprietary AI model in days ref-131 Google News "Is Netanyahu real or AI?" | Generative AI warps truth of West Asia war ref-132 Google News Anthropic Races to Contain Leak of Code Behind Claude AI Agent ref-133 Google News Box launches AI agent to process enterprise documents in natural language ref-134 Google News Microsoft's Newest Open-Source Project: Runtime Security For AI Agents ref-135 Google News OpenAI’s next AI model ‘Spud’ could be a major leap toward AGI, Greg Brockman explains ref-136 Google News Sony's gaming division just bought an AI startup that turns photos into 3D volumes ref-137 Google News Mercor, a $10 billion AI startup, confirms it was caught up in a major security incident ref-138 Google News India AI Startup Sarvam Raises Funds at $1.5 Billion Valuation ref-139 Google News Depression-Detecting AI Startup Kintsugi Shuts Down After FDA Rejection ref-140 Google News Anthropic Reportedly Acquires Stealth Biotech AI Startup for $400M in Stock ref-141 Google News Yupp.ai Shutdown: The Stunning Collapse of a $33M AI Startup Backed by a16z’s Chris Dixon ref-142 Google News DeepSeek outage: Chinese AI startup's near-perfect record broken by massive seven-hour global outage ref-143 Google News Australia signs AI safety agreement with Anthropic as research and skills investment expands ref-144 Google News Key nonprofit pitches tech giants to pay $100M each for AI safety effort ref-145 Google News BoE Hands Chancellor Its Plan For Safe AI Regulation In 2026 ref-146 Google News Microsoft to Invest $10 Billion in Japan on AI Infrastructure, Cybersecurity ref-147 Google News Corning Surges on $6 Billion Meta Deal Driven by AI Infrastructure Boom ref-148 Google News Google to tap into gas plant for AI datacenter in sharp turn from climate goals ref-149 Google News Nvidia's H100 GPU rental prices surge nearly 40% in 6 months: SemiAnalysis ref-150 Google News OpenAI Raises $122B to Expand AI Infrastructure, Broadens Cloud and Chip Strategy ref-151 Google News The $200 Billion Cloud Gamble: Amazon Sets New Record in 2026 AI Infrastructure Blitz ref-152 Google News Samsung-backed AI chip firm Rebellions raises $400 million ahead of IPO ref-153 Google News NVIDIA and Marvell Forge Strategic AI Infrastructure Partnership with $2 Billion Investment