Report 2026-04-09 145 sources Input Tokens: 94,297 · Output Tokens: 4,648 · Total Tokens: 101,318

AI 戰情報告 (2026-04-09)

metaanthropicgooglesamsungtsmcarmsoftbankperplexity muse-sparkclaude-mythosgemini-proexaone-4.5glm-5.1 agentmanaged-infrastructurespatial-intelligencecybersecuritychip-strategy

AI 戰情報告 (2026-04-09)

昨日 AI 重點新聞

Meta 發表全新旗艦模型 Muse Spark，宣告 Llama 時代終結？ ref-3, ref-119, ref-77, ref-35, ref-47, ref-6, ref-122, ref-34

Meta Superintelligence Labs 推出首款專有模型 Muse Spark，這也是祖克柏投入數十億美元重組 AI 團隊後的首個成果。該模型不再採用 Llama 名稱，且目前僅提供私有 API 給特定合作夥伴，並深度整合至 WhatsApp、Instagram 與 Ray-Ban 智慧眼鏡。Muse Spark 提供「即時 (Instant)」與「思考 (Thinking)」模式，後者具備與 GPT-5.4 Pro 競爭的深度推理與長程規劃能力。

Anthropic 推出 Claude Managed Agents，將代理基礎設施「平台化」 ref-10, ref-11, ref-80, ref-114

Anthropic 正式進入基礎設施市場，推出 Managed Agents 公測版。該服務將開發 AI 代理所需的代碼沙盒 (Sandbox)、憑證管理與執行追蹤等複雜工程全數自動化，號稱可縮短 10 倍開發時間。這標誌著 Anthropic 從模型供應商轉型為 AI 代理的「作業系統」提供者，直接與 OpenAI Enterprise 競爭。

Claude Mythos 展現毀滅性資安實力，Anthropic 開啟 Project Glasswing 專案 ref-10, ref-103, ref-81, ref-123

延續昨日對 Claude Mythos 能力的報導，該模型被證實能自主修復數千個零日漏洞。為確保技術不被惡意利用，Anthropic 啟動 Project Glasswing，僅限 AWS、Apple 與 CrowdStrike 等防禦性夥伴使用。與此同時，美國上訴法院拒絕暫停該模型的「供應鏈風險」標籤，顯示政府對此強大技術的監管態度依然謹慎。

三星 Bixby 全面 Agent 化：導入大語言模型架構實現語意控制 ref-125

Samsung 宣佈將 Bixby 升級為基於 LLM 架構的 AI 代理。新版 Bixby 不再只是語音助手，而是具備跨 App 操作、理解模糊語意與執行複雜任務的能力，並將首發於 Galaxy 系列旗艦手機。這標誌著行動裝置端側 AI 代理正式進入「大規模普及期」。

TSMC CoWoS 產能呈現 80% 複合增長，Nvidia 幾乎鎖定所有產能 ref-55

台積電 (TSMC) 透露其最先進晶片封裝技術 CoWoS 的年複合增長率高達 80%。報導指出，Nvidia 已預訂了 2026 至 2027 年大部分的產能，這可能成為其他 AI 晶片廠（如 Intel, AMD）在算力競賽中的關鍵瓶頸。

Google Gemini 推出「筆記本」功能，強化長文檔與專案管理 ref-39, ref-66

Google 為其 Gemini App 增加 Notebooks 功能，允許用戶將特定文件、對話歷史與自訂指令整合至單一視圖中。此功能深度整合了 NotebookLM，能作為個人的 AI 知識庫，對標 ChatGPT Projects 的工作流管理能力。

軟銀與 Arm 聯手推動 Project Izanagi，Rene Haas 將兼任 AI 戰略要職 ref-56

軟銀 (SoftBank) 集團為推進其 1,000 億美元的 AI 晶片計畫 Project Izanagi，擬任命 Arm 執行長 Rene Haas 兼任集團高層職務。此舉旨在整合 Arm 的 IP 與軟銀的資本，打造能打破 Nvidia 壟斷的自主 AI 生態系。

AI 搜尋巨巨 Perplexity 營收持續暴漲，ARR 突破 4.5 億美元 ref-129

Perplexity 在 2026 年第一季表現驚人，其年度經常性收入 (ARR) 突破 4.5 億美元，單月增長達 50%。其成功的關鍵在於從單一搜尋框轉向「代理搜尋 (Agentic Search)」，能自動為用戶執行多步驟的資料收集與報告撰寫任務。

AI 編程工具帶動 App Store 新應用激增 84%，市場進入「長尾應用」爆發期 ref-10

最新數據顯示，由於 Cursor、Claude Code 等 AI 編程工具的普及，App Store 新上架應用的數量成長了 84%。這反映出個人與小團隊開發 App 的難度已大幅降低，未來行動應用市場將面臨內容爆炸與分發挑戰。

Niantic 發表 Scaniverse 企業服務，打造 3D 世界地理空間模型 ref-72

Niantic Spatial 推出 Scaniverse 企業版，利用 3D 掃描、衛星影像與 GPS 數據建立精確的地理空間模型 (LGM)。執行長指出，目前僅 20% 的全球經濟數據已數位化，其餘 80% 實體經濟（如製造、能源）急需 AI 對實體世界進行「語意理解」以實現自動化。

LG 推出 EXAONE 4.5 多模態模型，發力企業級 AI 市場 ref-117

LG 發表其最新多模態模型 EXAONE 4.5，專注於處理專業領域的技術文件與圖像解析。此舉顯示韓國科技巨頭正在加強自研模型能力，以減少對美國基礎模型的依賴。

OpenAI 投資「機器人大軍」新創，佈局實體世界代理 ref-126

OpenAI 領投了一家專注於群體機器人 (Bot Army) 技術的新創公司，旨在突破 AI 代理在實體操作上的協作難度。這顯示 OpenAI 的野心已不再局限於聊天室，而是向物理實體代理邁進。

市場洞察

1. 「代理基礎設施」成為模型廠商的新戰場 ( ) ref-10, ref-11, ref-114

延續前日對 OpenAI 裝具工程 的討論，今日 Anthropic 推出 Managed Agents 顯示，領先的模型廠正在將「代理運行平台」產品化。企業不再需要自行架設複雜的沙盒環境，模型廠直接提供「一鍵託管」代理的功能。這意味著未來開發者的價值將從「編寫代碼」轉向「定義代理目標 (Agent Definition)」，也預示著 SaaS 模式將全面轉向 AIaaS (Agent as a Service)。

2. 硬體封裝與先進封裝技術成為算力競賽的終極瓶頸 ( ) ref-55, ref-139

TSMC CoWoS 的 80% 增長 (ref-27) 與 Nvidia 的鎖定產能行為，揭示了 AI 晶片的競爭點已從單純的架構設計轉向產能配置。軟銀與 Arm 的緊密整合 (ref-26) 則是試圖從最底層的 IP 授權到晶片代工進行全鏈條佈局，以避開未來幾年可能出現的算力供應斷裂。

3. 「空間智能」補足數位經濟與實體經濟的最後一塊拼圖 ( ) ref-139, ref-122

Niantic 提出的 地理空間模型 (LGM) (ref-56) 與 Edmund 在工業維護上的應用 (ref-8) 表明，AI 代理正在從單純處理文字與代碼，轉向理解「物理原子」。這對於製造、能源與物流業至關重要。當 AI 具備空間推理能力時，代理將能處理那些目前仍佔全球經濟 80% 但尚未數位化的傳統產業流程。

Sources 145 items

ref-0 Hacker News US cities are axing Flock Safety surveillance technology ref-1 Hacker News Show HN: Is Hormuz open yet? ref-2 Hacker News I've been waiting over a month for Anthropic to respond to my billing issue ref-3 Hacker News Muse Spark: Scaling towards personal superintelligence ref-4 Hacker News MegaTrain: Full Precision Training of 100B+ Parameter LLMs on a Single GPU ref-5 Hacker News John Deere to pay $99M in right-to-repair settlement ref-6 Hacker News Show HN: Skrun – Deploy any agent skill as an API ref-7 Hacker News Show HN: Unicode Steganography ref-8 Hacker News Show HN: TUI-use: Let AI agents control interactive terminal programs ref-9 Hacker News 'There's a lot of desperation': older workers turn to AI training to stay afloat ref-10 Hacker News Claude Managed Agents Overview ref-11 Hacker News Process Manager for Autonomous AI Agents ref-12 Hacker News You Can Just Print an Air Purifier ref-13 OpenAI Blog The next phase of enterprise AI ref-14 Google Developers Blog Boost Training Goodput: How Continuous Checkpointing Optimizes Reliability in Orbax and MaxText ref-15 Google Developers Blog ADK Go 1.0 Arrives! ref-16 Google Developers Blog Build a smart financial assistant with LlamaParse and Gemini 3.1 ref-17 Hugging Face Blog ALTK‑Evolve: On‑the‑Job Learning for AI Agents ref-18 Cursor Blog Bugbot now self-improves with learned rules ref-19 LangChain Blog Better Harness: A Recipe for Harness Hill-Climbing with Evals ref-20 Vercel Blog Zero Data Retention on AI Gateway ref-21 Vercel Blog How FLORA shipped a creative agent on Vercel's AI stack ref-22 Vercel Blog SERHANT.'s playbook for rapid AI iteration ref-23 Vercel Blog Use and manage Vercel Sandbox directly from the Vercel CLI ref-24 Vercel Blog Summary of CVE-2026-23869 ref-25 Vercel Blog Vercel Sandbox now supports up to 32 vCPU + 64 GB RAM configurations ref-26 Vercel Blog Chat SDK adds Liveblocks support ref-27 Vercel Blog Transfer Marketplace resources between teams ref-28 Vercel Blog Activity Log now available in Vercel CLI ref-29 Vercel Blog View specific error codes in runtime logs ref-30 AWS ML Blog Customize Amazon Nova models with Amazon Bedrock fine-tuning ref-31 AWS ML Blog Human-in-the-loop constructs for agentic workflows in healthcare and life sciences ref-32 AWS ML Blog Building intelligent audio search with Amazon Nova Embeddings: A deep dive into semantic audio understanding ref-33 AWS ML Blog Reinforcement fine-tuning on Amazon Bedrock: Best practices ref-34 Latent Space [AINews] Meta Superintelligence Labs announces Muse Spark, first frontier model on their completely new stack ref-35 Simon Willison Meta's new model is Muse Spark, and meta.ai chat has some interesting tools ref-36 Simon Willison Quoting Giles Turnbull ref-37 TechCrunch AI Poke makes using AI agents as easy as sending a text ref-38 TechCrunch AI AWS boss explains why investing billions in both Anthropic and OpenAI is an OK conflict ref-39 TechCrunch AI Tubi is the first streamer to launch a native app within ChatGPT ref-40 TechCrunch AI Astropad’s Workbench reimagines remote desktop for AI agents, not IT support ref-41 TechCrunch AI OpenAI releases a new safety blueprint to address the rise in child sexual exploitation ref-42 TechCrunch AI Databricks co-founder wins prestigious ACM award, says ‘AGI is here already’ ref-43 TechCrunch AI Final 3 days to save up to $500 on your TechCrunch Disrupt 2026 pass ref-44 TechCrunch AI Atlassian launches visual AI tools and third-party agents in Confluence ref-45 The Verge AI Gemini gets notebooks to help you organize projects ref-46 The Verge AI OpenAI made economic proposals — here’s what DC thinks of them ref-47 The Verge AI Meta is reentering the AI race with a new model called Muse Spark ref-48 The Verge AI The vibes are off at OpenAI ref-49 The Verge AI Unionized ProPublica staff are on strike over AI, layoffs, and wages ref-50 Ars Technica Iran-linked hackers disrupt operations at US critical infrastructure sites ref-51 Ars Technica Thousands of consumer routers hacked by Russia's military ref-52 MIT Technology Review AI Mustafa Suleyman: AI development won’t hit a wall anytime soon—here’s why ref-53 TechMeme As AI transforms white-collar work, executives across sectors say trust, not AI, is the key differentiator where accuracy, accountability, and regulation matter (Financial Times) ref-54 TechMeme A look at a global scramble to protect submarine cables vulnerable to potential sabotage, as new monitoring tech like distributed acoustic sensing is developed (Mike Cherney/Wall Street Journal) ref-55 TechMeme TSMC says its most advanced chip packaging tech, CoWoS, is growing at an 80% CAGR as it ramps up capacity; Nvidia has reportedly reserved most of the capacity (Katie Tarasov/CNBC) ref-56 TechMeme Sources: Arm CEO Rene Haas is in line for an additional role at SoftBank Group to advance Project Izanagi, the Japanese group's AI chip strategy (Financial Times) ref-57 TechMeme UAE's leading AI company G42 says its data center campus and overseas plans are on track despite regional tensions and Iranian attacks on UAE infrastructure (Bloomberg) ref-58 TechMeme Memo: UK NHS data chief Ming Tang says Palantir-built Federated Data Platform is delivering "outstanding results" as she pushes deeper rollout of Palantir tech (Chris Smyth/Financial Times) ref-59 TechMeme A look at Neo, a VC firm that made early bets on startups including Kalshi and Cursor, pushing the value of its first two funds far above industry averages (Yuliya Chernova/Wall Street Journal) ref-60 TechMeme Poke, an AI agent that lets users automate tasks via text message, raised $10M, on top of a $15M seed last year, at a $300M post-money valuation (Sarah Perez/TechCrunch) ref-61 TechMeme Nava, which develops cloud infrastructure for AI workloads by combining data centers, GPUs, and software tools, raised a $22M Series A led by Greenoaks Capital (Samreen Ahmad/Tech in Asia) ref-62 TechMeme An analysis of the 200 most-recent tweets on X by 18 big publishers indicates links seem to hurt engagement, after debate arose over X's value for sharing news (Laura Hazard Owen/Nieman Lab) ref-63 TechMeme John Deere agreed to pay $99M into a fund to settle a 2022 right-to-repair class action lawsuit, and make digital repair tools available to farmers for 10 years (Reuters) ref-64 TechMeme Sources: Anthropic completes an employee tender offer at a $350B valuation, falling short of the $6B investors wanted, as employees held shares ahead of the IPO (Bloomberg) ref-65 TechMeme Memo: xAI is reorganizing its engineering team, as SpaceX SVP Michael Nicolls says xAI is "clearly behind"; source: Nicolls has taken the title of xAI president (Grace Kay/Business Insider) ref-66 TechMeme Google introduces "notebooks" in the Gemini app for deeper NotebookLM integration and a space to organize chats and files, after adding sources in December 2025 (Abner Li/9to5Google) ref-67 TechMeme Yuga Labs settles its 2022 lawsuit against artist Ryder Ripps and Jeremy Cahen over their alleged copycatting of its BAYC NFTs; the terms were not disclosed (CoinDesk) ref-68 InfoQ AI/ML Tailwind CSS 4.2 Ships Webpack Plugin, New Palettes and Logical Property Utilities ref-69 InfoQ AI/ML Cloudflare and ETH Zurich Outline Approaches for AI-Driven Cache Optimization ref-70 InfoQ AI/ML GitHub Actions Custom Runner Images Reach General Availability ref-71 InfoQ AI/ML Presentation: Local First – How To Build Software Which Still Works After the Acquihire ref-72 The New Stack Niantic Spatial wants to map the 80% of the economy AI can’t see ref-73 The New Stack In the AI Age, Java is More Relevant Than Ever ref-74 The New Stack With Claude Managed Agents, Anthropic wants to run your AI agents for you ref-75 The New Stack Microsoft wants to make service mesh invisible ref-76 Wired AI Conflicting Rulings Leave Anthropic in ‘Supply-Chain Risk’ Limbo ref-77 Wired AI Meta’s New AI Model Gives Mark Zuckerberg a Seat at the Big Kid’s Table ref-78 Wired AI The US Army Is Building Its Own Chatbot for Combat ref-79 Wired AI 5 Burning Questions About Elon Musk’s Terafab Chip Partnership with Intel ref-80 Wired AI Anthropic’s New Product Aims to Handle the Hard Part of Building AI Agents ref-81 Stratechery Anthropic’s New Model, The Mythos Wolf, Glasswing and Alignment ref-82 Tom's Hardware Intel and SambaNova team up on heterogenous AI inference platform — different hardware performs different workloads ref-83 Tom's Hardware PCI Express roadmap: The path to 1TB/s with PCI 8.0, the challenges of integration, and beyond ref-84 Tom's Hardware AMD reveals $899 price tag for Ryzen 9 9950X3D2 — first dual-cache X3D CPU is $200 more expensive than the Ryzen 9 9950X3D ref-85 Tom's Hardware Bain Capital's data center unit removes disgraced tenant suspected of smuggling Nvidia GPUs to China — Megaspeed previously alleged to have spent roughly $2 billion on AI processors for illicit distribution ref-86 Tom's Hardware British cryptographer Adam Back is the secret creator of Bitcoin, claims new report — Back refutes investigation, says parallels to Satoshi are just a coincidence ref-87 Tom's Hardware Tech industry lays off nearly 80,000 employees in the first quarter of 2026 — almost 50% of affected positions cut due to AI ref-88 Tom's Hardware Asus ROG Xbox Ally review: The cheapest Windows handheld gets points for showing up ref-89 Tom's Hardware A brief history of Denuvo DRM and the new hypervisor bypass — inside the cat-and-mouse game between Denuvo and the piracy scene ref-90 Tom's Hardware Taiwanese chip makers call on government to stockpile helium, liquid natural gas — TSIA pleads for strategic supplies as US and Iran sign ceasefire in Middle East ref-91 Tom's Hardware Grab an entire RTX 5090 gaming PC for just $8 more than the GPU itself and score a whopping $1,600 off — huge HP discount requires a $39 controller or monitor to secure you a 4K powerhouse with a 9800X3D, 32GB DDR5, and a 1TB SSD ref-92 Tom's Hardware Get $100 off the ROG Xbox Ally handheld — Ryzen Z2 A-powered version with 16GB of RAM is under $500 again ref-93 Tom's Hardware Snap up 32GB of Corsair Vengeance RAM for just $192 when you pair it with AMD's 9800X3D processor and Asus X870E motherboard for $1,054 — bundle also includes a free 240mm AIO cooler and a copy of Crimson Desert ref-94 Tom's Hardware Russian state hackers are hijacking TP-Link and MicroTik routers to steal Outlook credentials, cybersecurity center warns — APT28 group targets DNS and redirects traffic to attacker-controlled servers ref-95 Tom's Hardware be quiet! Pure Power 13 M 1200W supply review: Platinum-level efficiency, premium pricing ref-96 Tom's Hardware China intensifies efforts to poach semiconductor talent from Taiwan, claims report — international restrictions motivate illicit efforts to obtain talent and equipment ref-97 Tom's Hardware Corsair's new customizer tool lets buyers trick out the Frame 4000D case — configurator lets you explore every possible option for this highly modular enclosure ref-98 Politico Tech Anthropic loses appeals court bid to pause supply chain risk label ref-99 The Hacker News (Security) New Chaos Variant Targets Misconfigured Cloud Deployments, Adds SOCKS Proxy ref-100 The Hacker News (Security) Masjesu Botnet Emerges as DDoS-for-Hire Service Targeting Global IoT Devices ref-101 The Hacker News (Security) APT28 Deploys PRISMEX Malware in Campaign Targeting Ukraine and NATO Allies ref-102 The Hacker News (Security) Shrinking the IAM Attack Surface through Identity Visibility and Intelligence Platforms (IVIP) ref-103 The Hacker News (Security) Anthropic's Claude Mythos Finds Thousands of Zero-Day Flaws Across Major Systems ref-104 EU Startups Czech startup Edmund raises €2.5 million to expand its AI debugging platform for industrial maintenance ref-105 EU Startups Meet the experts joining our “Equity Crowdfunding & Future of Venture Deals” panel at this year’s EU-Startups Summit! ref-106 EU Startups Spain’s Golden Owl raises €1.4 million to advance AI-powered anticipatory intelligence operating system ref-107 EU Startups Zurich’s Herbert Ventures launches €32.5 million Fund I to back European founders at pre-Seed and Seed ref-108 EU Startups British FinTech firm MillTech secures €51 million investment at €277 million valuation ref-109 EU Startups Estonia’s Handhold raises €3 million to give every software buyer an AI account manager ref-110 EU Startups Accenture acquires Spain’s Keepler to deepen AI and data capabilities in Spain and EMEA ref-111 EU Startups Inside the exhibition hall of the EU-Startups Summit 2026 ref-112 EU Startups London-based photo-organising startup Picnic acquired by French app publisher MWM ref-113 Google News Introducing Our First Proprietary AI Model to Deliver Personalized, Clinically Grounded Women’s Health Guidance ref-114 Google News Anthropic launches Claude Managed Agents to speed up AI agent development ref-115 Google News Meta Unveils New A.I. Model, Its First From the Superintelligence Lab ref-116 Google News Visa Rolls Out AI Agent Shopping Infrastructure Globally ref-117 Google News LG unveils EXAONE 4.5 multimodal AI model ref-118 Google News USC team develops new AI model to advance rare disease therapies ref-119 Google News Goodbye, Llama? Meta launches new proprietary AI model Muse Spark — first since Superintelligence Labs' formation ref-120 Google News App Store Flooded as AI Coding Tools Fuel 84% Surge in New Apps ref-121 Google News Z.ai unveils GLM-5.1, enabling AI coding agents to run autonomously for hours ref-122 Google News Meta AI head Alexandr Wang to everyone calling Muse Spark AI model 'dissapointment': We are quite upfront ref-123 Google News ‘Claude Mythos in cybersecurity’: How Anthropic’s new AI model can find hidden software vulnerabilities ref-124 Google News Anthropic Says Its New AI Model Is So Good at Finding Security Risks, You Can't Use It ref-125 Google News Samsung's Bixby Goes Full AI Agent With LLM Architecture ref-126 Google News OpenAI backs new AI startup seeking bot army breakthroughs ref-127 Google News Nebius in Talks to Buy Israeli AI Startup AI21 After Nvidia Deal Fizzles ref-128 Google News The Publicis-Microsoft Deal is Bigger Than You Think ref-129 Google News AI startup Perplexity’s revenue soars 50%, ARR hits $450M: FT ref-130 Google News AI startup Nava raises $22 million in round led by Greenoaks Capital ref-131 Google News Chinese robot AI startup D-Robotics raises $150m for expansion ref-132 Google News Anthropic Reports $30 Billion Run Rate, Expands Broadcom-Google AI Partnership ref-133 Google News AI startup Sarvam close to raising $300 million at $1.5 billion valuation ref-134 Google News Project Prometheus: Jeff Bezos’s AI Startup Hires xAI Co-Founder Kyle Kosic To Lead Infrastructure for ref-135 Google News At David Sacks’s Behest, White House Barrels Forward on Industry-Friendly AI Policy ref-136 Google News AI regulation is stalling as governments clash over control ref-137 Google News OpenAI opens applications for an external AI safety research fellowship ref-138 Google News Sam Altman accused of misleading board on AI safety breach ref-139 Google News China rolls out first comprehensive AI ethics framework with review rules ref-140 Google News Oracle: The Most Underrated AI Infrastructure Play (NYSE:ORCL) ref-141 Google News Alibaba launches data center with 10,000 of its own chips as China ramps up AI push ref-142 Google News Supply chain challenges risk delaying Nvidia's Rubin GPUs ref-143 Google News Intel (INTC) Is Up 19.9% After Joining Musk’s Terafab AI Chip Initiative With Tesla and SpaceX ref-144 Google News GPU Rowhammer Attack Enables Privilege Escalation and Full System Compromise