OpenAI launched GPT-5.5 on April 23, 2026 for paid subscribers (Plus, Pro, Business, Enterprise), simultaneously rolling it into ChatGPT and Codex. The model matches GPT-5.4 per-token latency while delivering higher intelligence. It excels at writing and debugging code, web research, data analysis, document creation, and autonomous multi-step task completion — making it OpenAI's closest offering to an AI "super app" agent yet. API access opened April 24. GPT-5.5 is priced above GPT-5.4 but achieves significant token efficiency gains.
The Brief
6 stories · April 29, 2026
OpenAI Models and Codex Arrive on Amazon Bedrock
OpenAI 模型與 Codex 登上 Amazon Bedrock
AWS put OpenAI frontier models, Codex and Bedrock Managed Agents powered by OpenAI into limited preview. The move lets enterprises use OpenAI capabilities inside Bedrock with familiar controls such as IAM, PrivateLink, encryption, guardrails and CloudTrail logging.
AWS 將 OpenAI 前沿模型、Codex 與由 OpenAI 驅動的 Bedrock Managed Agents 推出有限預覽。企業可在 Bedrock 內使用 OpenAI 能力，同時沿用 IAM、PrivateLink、加密、guardrails 與 CloudTrail 記錄等治理控制。
Read source ->Anthropic Brings Claude Into Adobe, Blender, Ableton and More
Anthropic 將 Claude 接入 Adobe、Blender、Ableton 等創作工具
Anthropic released Claude connectors for creative software including Adobe Creative Cloud, Blender, Autodesk Fusion, Ableton, Affinity, SketchUp, Splice and Resolume. The strategy shifts Claude from a chatbot into a workflow layer that can read tool context, write scripts and automate production work.
Anthropic 發布 Claude 創作工具連接器，涵蓋 Adobe Creative Cloud、Blender、Autodesk Fusion、Ableton、Affinity、SketchUp、Splice 與 Resolume。這代表 Claude 正由聊天機器人轉向工作流程層，可讀取工具脈絡、撰寫腳本並自動化製作任務。
Read source ->Google Reportedly Signs Classified Pentagon AI Deal
Google 據報簽署美國五角大樓機密 AI 協議
The Verge, citing The Information, reported that Google signed a classified agreement allowing the U.S. Department of Defense to use its AI models for any lawful government purpose. The report highlights unresolved tension between national-security demand, employee concerns and enforceable AI safety limits.
The Verge 引述 The Information 報導，Google 已簽署機密協議，允許美國國防部將其 AI 模型用於任何合法政府用途。報導凸顯國安需求、員工疑慮與可執行 AI 安全限制之間的緊張關係。
Read source ->OpenAI and Anthropic Brief Congress on Cyber-Capable Models
OpenAI 與 Anthropic 向國會簡報具網攻能力的前沿模型
Axios reported that OpenAI and Anthropic held classified briefings with House Homeland Security staff on advanced models with cybersecurity implications. Anthropic has held back Mythos Preview from public release, while OpenAI is using a tiered approach for GPT-5.4-Cyber.
Axios 報導，OpenAI 與 Anthropic 已向眾議院國土安全委員會幕僚進行機密簡報，說明具網路安全影響的前沿模型。Anthropic 暫緩公開發布 Mythos Preview，OpenAI 則對 GPT-5.4-Cyber 採取分級發布。
Read source ->AWS Launches Quick and Agentic Connect Apps for Enterprise Workflows
AWS 推出 Quick 與 Agentic Connect 應用，攻入企業流程
AWS used its What’s Next event to launch Amazon Quick as an AI work assistant and expand Amazon Connect into agentic AI solutions for supply chains, hiring, customer experience and health care. The message: cloud providers want the application layer, not just model hosting.
AWS 在 What’s Next 活動推出 Amazon Quick AI 工作助理，並將 Amazon Connect 擴展為面向供應鏈、招聘、客戶體驗與醫療的 agentic AI 解決方案。訊號很清楚：雲端供應商不只要託管模型，也要進入應用層。
Read source ->Microsoft Says Copilot Adoption Is Moving From Productivity to Growth
Microsoft 稱 Copilot 採用正由效率工具轉向業務增長
Microsoft framed customer AI adoption around Intelligence + Trust, citing Copilot, Fabric, Agent 365 and Microsoft IQ deployments across companies such as PepsiCo, Tata Realty and Tru Cooperative Bank. The emphasis is shifting from time saved to decisions and operating outcomes.
Microsoft 以「Intelligence + Trust」概括企業 AI 採用，列舉 PepsiCo、Tata Realty、Tru Cooperative Bank 等公司部署 Copilot、Fabric、Agent 365 與 Microsoft IQ。重點正從節省時間轉向決策品質與營運成果。
Read source ->
Microsoft Loosens OpenAI Exclusivity as the Alliance Enters a New Phase
Microsoft 放寬 OpenAI 獨家權，雙方 AI 聯盟進入新階段
Microsoft is cutting its revenue share and giving up exclusive rights to sell access to OpenAI's models, while remaining OpenAI's primary cloud partner. The change gives OpenAI more room to distribute models through other channels and gives Microsoft more freedom to pursue a broader multi-model strategy.
Microsoft 正下調其收入分成，並放棄獨家銷售 OpenAI 模型存取權，但仍保留主要雲端合作夥伴角色。這讓 OpenAI 可透過更多渠道分發模型，也讓 Microsoft 更容易推動多模型 AI 策略。
Read source ->China Blocks Meta's $2B Acquisition of AI Agent Startup Manus
中國叫停 Meta 20 億美元收購 AI 代理新創 Manus
Chinese regulators blocked Meta's acquisition of Manus, an AI agent startup, in a sharp reminder that frontier AI deals now sit inside national-security and technology-transfer politics. The move interrupts Meta's agent strategy and raises the risk premium for cross-border AI acquisitions.
中國監管機構叫停 Meta 收購 AI 代理新創 Manus，凸顯前沿 AI 交易已進入國安與技術轉移政治範圍。此舉打斷 Meta 的代理策略，也提高跨境 AI 併購的不確定性。
Read source ->DeepSeek Cuts V4 Prices Up to 75%, Turning Long-Context AI Into a Price War
DeepSeek V4 最高降價 75%，長上下文模型戰轉為價格戰
After releasing V4-Pro and V4-Flash, DeepSeek moved quickly to discount V4-Pro by up to 75% and slash cache-hit pricing across its lineup. The model family pairs a 1 million-token context window with a Huawei-optimized deployment story, intensifying competition on both cost and compute sovereignty.
DeepSeek 發布 V4-Pro 與 V4-Flash 後，迅速將 V4-Pro 最高降價 75%，並大幅下調快取命中價格。該系列主打 100 萬 token 上下文與華為晶片優化部署，令模型競爭同時轉向成本與算力自主。
Read source ->EU Tells Google to Open Android to Rival AI Assistants
歐盟要求 Google 開放 Android，讓競爭 AI 助手深度接入
The European Commission proposed measures requiring Android to give competing AI services effective access to apps and device functions, from sending email to sharing photos. Google called the intervention unwarranted, but the DMA process could reshape how Gemini, ChatGPT, Claude and future agents compete on mobile.
歐盟委員會提出措施，要求 Android 讓競爭 AI 服務有效接入應用與裝置功能，包括發送電郵、點餐及分享照片。Google 批評此舉屬不必要干預，但 DMA 程序可能重塑 Gemini、ChatGPT、Claude 等行動端 AI 代理競爭格局。
Read source ->Meta Reserves Up to 1GW of Space-Beamed Solar Power for AI Data Centers
Meta 預留最高 1GW 太空太陽能，支援 AI 數據中心用電
Meta signed a reservation agreement with Overview Energy for up to 1 gigawatt of space-based solar capacity. The plan is to collect sunlight in orbit and beam energy to ground solar projects, a futuristic answer to the very real electricity bottleneck behind large-scale AI infrastructure.
Meta 與 Overview Energy 簽署預留協議，取得最高 1GW 太空太陽能容量。方案是在軌道收集陽光並將能量傳送至地面太陽能項目，反映大型 AI 基礎設施正面臨真實且迫切的電力瓶頸。
Read source ->Sereact Raises $110M to Scale a World-Model Robotic Brain
Sereact 融資 1.1 億美元，擴展世界模型機器人大腦
German robotics software company Sereact raised $110 million to scale Cortex 2.0, a robotic brain that combines vision-language-action models with a world model. Trained on more than a billion real warehouse picks, the system aims to make industrial robots adapt before errors become expensive.
德國機器人軟件公司 Sereact 融資 1.1 億美元，用於擴展 Cortex 2.0。該機器人大腦結合視覺語言動作模型與世界模型，並以逾 10 億次真實倉儲抓取資料訓練，目標是讓工業機器人在錯誤變昂貴前先行適應。
Read source ->
Anthropic's Claude Mythos: Most Capable AI Model Ever Built, Restricted to 50 Orgs
Anthropic 推出史上最強 Claude Mythos，僅限 50 家機構試用
Anthropic confirmed that Claude Mythos — its most capable model ever — will not be publicly available. Fifty organizations receive gated access under 'Project Glasswing.' Meanwhile, Anthropic's annualized revenue has surpassed $30 billion, up from ~$9B at end of 2025.
Anthropic 確認 Claude Mythos 是其有史以來最強的模型，但不對外公開，僅透過「Project Glasswing」計劃讓 50 個組織試用。Anthropic 年化營收已突破 300 億美元，較 2025 年底增長超過三倍。
Read source →Google Cloud Next 2026: AI Agents, A2A Protocol, and Full-Stack Bet Against OpenAI
Google Cloud Next 2026：AI 代理、A2A 協議全面上陣，挑戰 OpenAI
Google unveiled a suite of AI agent tools at its annual Cloud Next conference, including the production-grade Agent2Agent (A2A) protocol, managed MCP servers, a no-code agent builder for Workspace, and a redesigned developer platform with 200+ models including Anthropic's Claude.
Google 在年度 Cloud Next 大會上推出一系列 AI 代理工具，包括 Agent2Agent (A2A) 協議、MCP 管理伺服器、Workspace 無代碼代理建構器，以及支援超過 200 個模型的開發者平台。
Read source →OpenAI Raises $122B in Historic Funding Round Led by Amazon, Nvidia, SoftBank
OpenAI 完成 1220 億美元融資，由亞馬遜、輝達、軟銀領投
OpenAI closed a $122 billion funding round — the largest in startup history — led by Amazon ($50B), Nvidia ($30B), and SoftBank ($30B). Combined with Anthropic's $30B Series G and xAI's $250B SpaceX acquisition, Q1 2026 saw $267.2B in AI venture deal value.
OpenAI 完成 1220 億美元融資，創創業公司歷史紀錄，由亞馬遜（500 億）、輝達（300 億）及軟銀（300 億）領投。加上 Anthropic 的 300 億 G 輪及 xAI 的 2500 億併購，2026 年 Q1 AI 風投總額達 2672 億美元。
Read source →Google's TurboQuant Cuts LLM Memory 6× with 3-Bit KV Cache, No Accuracy Loss
Google TurboQuant：3-bit KV 快取將 LLM 記憶體降低 6 倍，準確率零損失
Google unveiled TurboQuant at ICLR 2026 — an algorithm combining PolarQuant vector rotation and Quantized Johnson-Lindenstrauss compression to reduce KV cache to just 3 bits with zero accuracy loss. Result: 6× memory reduction, 8× speedup in attention computation.
Google 在 ICLR 2026 發表 TurboQuant，結合 PolarQuant 向量旋轉與 QJL 壓縮技術，將 KV 快取壓縮至 3-bit，記憶體用量減少 6 倍，注意力計算速度提升 8 倍，且準確率零損失。
Read source →Cognition AI in Funding Talks at $25B Valuation as Coding Agents Surge
Coding AI 新創 Cognition 估值洽談 250 億美元，AI 程式代理市場急速成長
Cognition AI — maker of the Devin autonomous coding agent — is in early talks to raise a new funding round that would more than double its valuation to $25 billion. The deal reflects soaring demand for AI software engineering tools. Snap separately disclosed AI now writes 65%+ of its new code.
自主程式代理 Devin 的開發商 Cognition AI 正洽談新一輪融資，估值有望超過 250 億美元。與此同時，Snap 透露 AI 現已負責撰寫超過 65% 的新程式碼，顯示 AI 程式代理需求急速攀升。
Read source →Human Scientists Still Outperform Best AI Agents on Complex Research Tasks
Nature 研究：面對複雜科研任務，人類科學家仍勝過最強 AI 代理
A Nature study benchmarked AI agents against human scientists on complex, multi-step research tasks. Despite rapid progress, the best AI agents still trail human experts on tasks requiring deep domain reasoning, novel experimental design, and long-horizon planning.
《自然》期刊一項研究將 AI 代理與人類科學家在複雜多步驟科研任務上進行比較，結果顯示，儘管 AI 進展迅速，在需要深度領域推理、創新實驗設計與長期規劃的任務上，人類專家仍具明顯優勢。
Read source →
OpenAI Releases GPT-5.5: Smarter, More Efficient, Real-Work Agent
OpenAI 發布 GPT-5.5：更聰明、更高效，可自主完成真實工作任務
Anthropic Ships Claude Opus 4.7 and Mythos Preview Targets Cybersecurity
Anthropic 推出 Claude Opus 4.7 與 Mythos 預覽版，鎖定網路安全應用
Anthropic released Claude Opus 4.7, outperforming its predecessor across agentic coding, multidisciplinary reasoning, and computer use — at the same price. Separately, the Claude Mythos Preview (April 7) demonstrated exceptional cybersecurity capabilities, prompting Anthropic to launch Project Glasswing: a programme sharing Mythos with 11 organisations to find and fix critical software vulnerabilities proactively. Anthropic also introduced Managed Agents, a hosted Platform service for stable long-horizon agent work with durable state and safer tool access.
PwC Study: Top 20% of Firms Capture 74% of All AI Economic Value
PwC 研究：頂尖 20% 企業囊括 74% AI 經濟收益，差距持續擴大
PwC's 2026 AI Performance Study (1,217 senior executives, 25 sectors) reveals that 74% of AI-driven economic gains flow to just 20% of organisations. These "AI leaders" generate 7.2× more revenue and efficiency gains than average peers. The decisive differentiator: leaders pursue AI-enabled growth and industry convergence, not merely productivity. They are 1.9× more likely to run fully autonomous AI and 2.8× faster at removing human decision bottlenecks. The majority of companies remain stuck in pilot mode, harvesting little economic benefit.
Brain-Inspired Hafnium Oxide Chip Could Cut AI Energy Use by 70%
劍橋大學仿腦氧化鉿芯片可將 AI 能耗降低 70%，突破硬件瓶頸
University of Cambridge researchers (led by Dr. Babak Bakhit) engineered a modified hafnium oxide memristor that mimics neuronal memory-processing, combining storage and compute in one device. Unlike conventional metal oxide memristors — whose unpredictable conductive filaments require high voltages — the new device operates at extremely low currents with excellent stability and multiple distinct switching states. Benchmarks show up to 70% energy savings versus current AI hardware. Published in collaboration with PNAS, the advance could reshape AI inference hardware.
Bipartisan AI Transparency Act Demands Foundation Model Disclosure to FTC
跨黨派 AI 透明法案要求基礎模型向 FTC 揭露訓練資料與算法細節
Representatives Beyer (D-VA), Lawler (R-NY), and Jacobs (D-CA) introduced H.R. 8094 — the AI Foundation Model Transparency Act of 2026 — on March 26, 2026. The bill directs the FTC (consulting NIST, Commerce, and OSTP) to set disclosure standards for high-impact foundation models. Covered entities include models with 10M+ monthly users, trained with 10²⁶+ compute operations, or posing risks to security, civil rights, or public health. Fully open-source models are exempt. The FTC must issue final rules within one year of enactment.
"No Steering Wheel": Hinton Urges Global AI Governance at UN Conference
「沒有方向盤」：Hinton 在 UN 數字世界大會疾呼建立 AI 全球治理框架
Nobel laureate Geoffrey Hinton addressed the 2026 Digital World Conference in Geneva via video link, warning that unregulated AI is "like a very fast car with no steering wheel." He urged immediate global governance frameworks, noting that massive investments are being channelled into convincing the public that regulation impedes progress. Hinton expressed alarm that only ~1% of AI resources go to safety research, while humanity faces the existential question of whether "we can co-exist with super-intelligent AI." He called the current moment urgent yet dangerously under-resourced.
Nature Study: AI Agents Score Half of PhD Experts on Complex Research Tasks
《自然》研究：AI 代理在複雜科研任務上的得分僅為博士專家的一半
A state-of-the-industry report published in Nature (April 13, 2026) finds that the best AI agents perform only half as well as human PhD experts on complex research tasks — even as AI mentions in natural science publications grew nearly 30-fold (2010–2025). The paradox: researchers are heavily adopting AI tools despite little evidence of productivity improvement, and some argue that scientific quality has suffered. The report concludes that scientific norms have not had time to adapt to the pace of AI adoption.
OpenAI Ships GPT-5.5: 1M Context, 82.7% Terminal-Bench, Agentic Leap
OpenAI 推出 GPT-5.5：百萬 Token 上下文、82.7% Terminal-Bench、強化自主代理能力
OpenAI released GPT-5.5 on April 23, 2026 — the first fully retrained base model since GPT-4.5. It ships with a 1M-token context window and tops the Artificial Analysis Intelligence Index at 60, scoring 82.7% on Terminal-Bench 2.0. API pricing is $5 / $30 per million tokens (input/output); a Pro tier at $30/$180 targets enterprise workloads. OpenAI President Greg Brockman highlighted the model's autonomous problem-solving: "It can look at an unclear problem and figure out just what needs to happen next." GPT-5.5 is rolling out to Plus, Pro, Business and Enterprise subscribers in ChatGPT and Codex.
Amazon Doubles Down on Anthropic: $25B Investment + $100B AWS Pledge
Amazon 再押注 Anthropic：$250 億美元投資 + $1,000 億 AWS 承諾
Amazon has agreed to invest up to $25 billion in Anthropic — on top of the $8 billion previously committed — as part of an expanded AI infrastructure deal. The structure: $5 billion upfront plus up to $20 billion tied to commercial milestones. Alongside the investment, Anthropic commits to spending over $100 billion on AWS over the next decade and will secure up to 5 GW of chip capacity for model training. The full Claude Platform becomes available directly within AWS accounts. Amazon made a similar $50B deal with OpenAI just weeks earlier, positioning AWS as the cloud backbone of frontier AI.
OpenAI Tops $25B Revenue, Eyes 2026 IPO at $852B Valuation
OpenAI 年化營收突破 $250 億，估值 $8,520 億、籌備 2026 年 IPO
OpenAI has crossed $25 billion in annualized revenue as of February 2026, up from $21.4B at year-end 2025. The company closed the largest private funding round in history: $122 billion at an $852 billion post-money valuation (Amazon: $50B, NVIDIA: $30B, SoftBank: $30B). Despite booming revenue, OpenAI projects ~$14 billion in losses — driven by massive compute costs. The company is preparing for an IPO, hiring its first head of investor relations, with an H2 2026 filing target and potential 2027 listing. Rival Anthropic is approaching $19 billion in annualized revenue.
Google April Gemini Drop: 3.1 Pro, Flash TTS, Gemma 4, macOS App
Google 四月 Gemini 更新：3.1 Pro、Flash TTS、Gemma 4 及 macOS 原生應用
Google's April 2026 Gemini drop brings several upgrades: Gemini 3.1 Pro for complex reasoning, Gemini 3.1 Flash TTS — a text-to-speech model with 70+ language support and SynthID watermarking — and two new open-weight Gemma 4 models (26B-A4B and 31B). Computer Use is now enabled in gemini-3-pro-preview and gemini-3-flash-preview. A native Gemini macOS app launches with keyboard-shortcut access for desktop workflows. Separately, Google confirmed a context-aware Siri rebuilt on Gemini will debut in 2026 as part of its Apple partnership.
Arizona Bans AI as Final Decision-Maker for Medical Claim Denials
亞利桑那州立法禁止以 AI 作為醫療保險拒保最終裁定者
Arizona Governor Hobbs signed HB 2175, becoming one of the first states to prohibit health insurers from using AI as the final decision-maker for medical claim denials. Effective July 1, 2026, the law requires a qualified physician (medical director) to individually review all denials, while AI tools may still assist in processing. Arizona joins five states (also CT, MD, NE, TX) with such laws, building on California's 2024 precedent. The broader context: Arizona's legislature adjourns Saturday April 25, with three additional AI bills still in play — covering AI transparency, liability and general-purpose AI guardrails.
April 2026 AI Landscape: Anthropic Leads Benchmarks, $122B OpenAI Round
2026 年 4 月 AI 全景：Anthropic 登頂基準測試，OpenAI 完成 1,220 億融資
A sweep of April 2026's AI landscape reveals a rapidly consolidating field. Anthropic now leads frontier model benchmarks — Claude Opus 4.7 scores 87.6% on SWE-bench Verified and 94.2% on GPQA, with 1M token context and 3.3× higher-resolution vision. OpenAI closed a record-breaking $122 billion Series E at $852B valuation. Infrastructure investment is soaring: combined cloud commitments by major AI labs to Amazon, Google and Microsoft now exceed $500 billion over the decade. MIT Technology Review notes AI's economic footprint is tracking faster than any previous technology wave.
DeepSeek V4 Pro: 1.6T Params, 93.5% LiveCode, $3.48/M Output
DeepSeek V4 Pro 登場：1.6 兆參數、93.5% 代碼基準，輸出僅 $3.48/M
China's DeepSeek dropped its V4-Pro and V4-Flash preview models on April 24 with no warning, posting a technical report and API access simultaneously. V4-Pro carries 1.6 trillion total parameters, 49 billion active per token, and a 1 million-token context window trained on 33 trillion tokens. On LiveCodeBench it scores 93.5% — beating every closed model — and 80.6% on SWE-bench Verified, within 0.2 pts of Claude Opus 4.6. Input pricing is $1.74/M and output $3.48/M, roughly 7× cheaper than Claude at near-identical coding performance.
OpenAI Launches GPT-5.5: 1M Context, Super-App Push, $5/$30
OpenAI 發布 GPT-5.5：百萬 Token 上下文，邁向超級 App 戰略
OpenAI released GPT-5.5 on April 23, its first fully retrained base model since GPT-4.5. It hits 82.7% on Terminal-Bench 2.0, 58.6% on SWE-Bench Pro, and 78.7% on OSWorld-Verified. The 1M-token context window matches DeepSeek V4-Pro. Pricing doubled to $5/$30 per million tokens. Available to Plus, Pro, Business, and Enterprise ChatGPT subscribers, GPT-5.5 anchors OpenAI's 'super app' strategy combining ChatGPT, Codex, and browser capabilities into a single interface.
Tencent & Alibaba Circle DeepSeek at $20B+ Valuation
騰訊與阿里洽談入股 DeepSeek，估值逾 200 億美元
Tencent Holdings and Alibaba Group are in talks to invest in DeepSeek at a valuation above $20 billion — double the $10 billion figure floated earlier this year. DeepSeek, owned by High-Flyer Capital Management, has yet to generate meaningful revenue as its models are open-source and its chatbot is free. Investors cite its track record of building frontier models at a fraction of Western costs and surging developer adoption following today's V4 launch.
Google Workspace Intelligence: 9× Faster Sheets, Cross-Tool Gemini
Google Workspace Intelligence：試算表速度提升 9 倍、跨工具 Gemini 代理
At Google Cloud Next '26, Google launched Workspace Intelligence — a semantic context layer connecting Gmail, Drive, Docs, Sheets, and Calendar so Gemini can reason across all data. Sheets gains prompt-based population 9× faster than manual entry; Docs generates data-driven infographics; Slides builds full decks from company templates. Expanded connectors link to Asana, Jira, and Salesforce. Rolls out across Business and Enterprise tiers.
Musk's Terafab Locks In Intel 14A — $3B, 100K Wafers/Month
馬斯克 Terafab 採用 Intel 14A 製程，30 億美元、每月 10 萬片晶圓
Elon Musk confirmed Terafab will use Intel's 14A process node — making Tesla the first major customer for Intel's most advanced fab technology. The $3 billion Austin project targets 100,000 wafer starts per month, producing Tesla's AI5 chip with pilot production in 2026 and volume in 2027. SpaceX handles high-volume manufacturing under a likely technology licensing deal. A pivotal moment for Intel's foundry ambitions against TSMC.
Idaho Enacts 4 AI Bills: K-12 Literacy, Chatbot Safety, Data Centers
愛達荷州通過 4 項 AI 法案：K-12 素養、聊天機器人安全、資料中心規範
Idaho Governor Brad Little signed four AI bills effective July 1, 2026. SB 1227 directs the education department to build a K-12 AI literacy framework. SB 1297 (Conversational AI Safety Act) bars AI from claiming to be a licensed mental health professional and requires chatbots to route suicidal ideation prompts to crisis hotlines. Two additional bills address AI-generated content privacy and data center environmental accountability. Idaho joins the wave of states acting ahead of any federal framework.
Yann LeCun's AMI Labs Raises $1.03B Seed Round at $3.5B Valuation
圖靈獎得主 LeCun 創辦的 AMI Labs 完成逾 10 億美元種子輪融資
Advanced Machine Intelligence (AMI) Labs, founded by Turing Award winner Yann LeCun, has raised $1.03 billion in seed funding at a $3.5 billion valuation, backed by Nvidia and Bezos Expeditions. The startup is building AI architectures designed to overcome the limitations of current large language models, focusing on world-model reasoning and energy-efficient inference. LeCun has long argued that current LLMs are architecturally insufficient for human-level intelligence.
Jeff Bezos Nears $10B Funding Round for Physical-World AI Startup
貝佐斯為實體世界 AI 新創公司籌集近百億美元
Jeff Bezos is close to finalizing a $10 billion funding round for his AI startup that is developing models capable of understanding and operating in the physical world. The company focuses on building AI that bridges the gap between digital reasoning and real-world physical environments—an area considered the next major frontier in AI research. The round would represent one of the largest AI-focused raises in history.
DeepSeek V4 Released With Fully Open Weights, Rivals US Frontier Models at $5.2M Training Cost
DeepSeek V4 開源發布，以 520 萬美元訓練成本媲美美國頂尖模型
Chinese AI lab DeepSeek has released V4, a one-trillion parameter model with fully open weights that achieves performance competitive with leading US frontier models. Remarkably, the model was trained for an estimated $5.2 million—a fraction of the billions spent by OpenAI and Anthropic. The release reignites debate about the efficiency of AI development and whether massive capital expenditure is truly necessary for frontier performance.
AI Energy Breakthrough: New Method Cuts Power Use 100× While Boosting Accuracy
AI 能耗突破：結合符號推理使功耗降低百倍、精度同步提升
Researchers have unveiled a hybrid approach that could slash AI energy consumption by up to 100 times while simultaneously improving model accuracy. The method combines traditional neural networks with human-like symbolic reasoning—a technique long advocated by critics of purely statistical AI. As energy costs become one of the biggest constraints on AI scaling, this breakthrough could fundamentally reshape how next-generation models are built and deployed.
Nebraska Lawyer Suspended After AI-Generated Brief Contains 20 Hallucinated Citations
內布拉斯加州律師因 AI 生成摘要含 20 條幻覺引文被吊銷執照
A Nebraska attorney has been suspended from legal practice after his appellate brief was found to contain 57 defective citations, including 20 fabricated by AI. US courts imposed at least $145,000 in sanctions against attorneys for AI citation errors in Q1 2026 alone, signaling an escalating crackdown. The incident underscores the growing legal risk of deploying AI in high-stakes professional contexts without rigorous human verification.
Anthropic Unveils Claude Mythos Preview to 50 Partner Organizations
Anthropic 向 50 個合作機構發布 Claude Mythos Preview
Anthropic has unveiled Claude Mythos Preview to roughly 50 partner organizations under a restricted research program called Project Glasswing. The model is described as a "step change" above Claude Opus 4.6 and features an unprecedented 10-trillion-parameter architecture. Anthropic says a broader public release is planned for later in Q2 2026, pending safety evaluations.
PwC: 75% of AI Economic Gains Captured by Just 20% of Companies
普華永道：AI 經濟收益的 75% 被僅 20% 的企業所獨享
PwC's 2026 AI Jobs Barometer finds that three-quarters of AI's economic gains are being captured by just 20% of companies — a group PwC calls "AI leaders" focused on growth rather than pure cost-cutting. The study analysed 5,000 companies across 45 countries and found AI leaders are growing revenue 2.4× faster than laggards, with productivity gains of 14% in customer service and 26% in software development.
Meta Plans $115–135 Billion in AI Infrastructure Spending for 2026
Meta 計劃在 2026 年投入 1,150–1,350 億美元於 AI 基礎設施
Meta CEO Mark Zuckerberg confirmed the company plans to spend $115–135 billion on AI infrastructure in 2026 — nearly double last year's capital expenditure. The announcement came alongside the reveal that Meta's AI products now serve more than 1 billion monthly active users. The spending reflects an all-in bet on AI becoming the foundation of Meta's advertising, social networking, and emerging hardware businesses.
SoundHound AI Acquires LivePerson for ~$250M in All-Stock Deal
SoundHound AI 以約 2.5 億美元全股票方式收購 LivePerson
SoundHound AI has announced an all-stock acquisition of LivePerson and its Conversational Cloud platform for approximately $250 million. The combined entity will serve 25 Fortune 100 companies and 12 of the world's top 15 global banks, creating a dominant player in enterprise conversational AI. SoundHound shares fell on the announcement as investors weighed dilution concerns against the strategic rationale of the deal.
Northwestern Engineers 3D-Print Artificial Neurons That Talk to Real Ones
西北大學工程師 3D 列印能與真實神經元通訊的人工神經元
Engineers at Northwestern University have achieved a breakthrough in brain-computer integration by 3D-printing artificial neurons that can communicate bidirectionally with biological neurons. The printed neurons mimic the electrochemical behaviour of real nerve cells and successfully exchanged signals with live neural tissue in laboratory conditions. Researchers say the advance could eventually enable next-generation neural implants and treatments for neurological disorders.
MIT Tech Review Unveils '10 Things That Matter in AI Right Now' at EmTech AI
MIT 科技評論於 EmTech AI 大會發布《現在 AI 最重要的 10 件事》
MIT Technology Review today released its inaugural '10 Things That Matter in AI Right Now' list, debuted live at the EmTech AI conference on MIT's campus before going online. The list highlights the most consequential trends shaping AI in 2026 — from the race between open-source and proprietary models to the economic divide between AI winners and laggards. As of April 2026, Anthropic leads performance rankings, trailed closely by xAI, Google, and OpenAI, while China has nearly erased the U.S. lead in multiple benchmarks.
Snap Cuts 1,000 Jobs as AI Now Generates Over 65% of Its New Code
Snap 裁員逾 1,000 人：AI 現已生成其超過 65% 的新代碼
Snap CEO Evan Spiegel announced the layoff of approximately 1,000 employees — roughly 14% of the company's workforce — citing rapid AI advancements that have transformed software development at the company. AI now generates more than 65% of Snap's new code, and the shift is expected to deliver over $500 million in annualised cost savings. The move is one of the most concrete examples yet of AI displacing software engineering roles at a major consumer tech company.
Cursor AI in Talks to Raise $2 Billion at Over $50 Billion Valuation
Cursor AI 洽談融資 20 億美元，估值逾 500 億美元
AI coding assistant Cursor is in advanced talks to raise a $2 billion funding round at a valuation exceeding $50 billion, according to sources cited by CNBC. The deal would represent a dramatic increase from Cursor's previous valuation and would make it one of the most valuable AI startups outside of the frontier model labs. Cursor's growth has been fuelled by surging demand from software teams looking to boost developer productivity with AI pair-programming.
Zhipu AI's GLM-5.1 — 744B MoE — Beats Claude Opus 4.6 and GPT-5.4 on SWE-Bench
智譜 AI 發布 GLM-5.1（7,440 億參數 MoE）：在 SWE-Bench 超越 Claude Opus 4.6 與 GPT-5.4
Chinese lab Zhipu AI has released GLM-5.1 under the MIT licence — a 744-billion-parameter mixture-of-experts model with 40 billion active parameters per forward pass and a 200K context window. Benchmarks place GLM-5.1 above both Claude Opus 4.6 and GPT-5.4 on SWE-Bench Pro, marking the first time a fully open-source Chinese model has topped this flagship software-engineering leaderboard. The release is the latest evidence that China's AI gap with the U.S. is narrowing at a rapid pace.
Nature Study: Human Scientists Still Outperform the Best AI Agents on Complex Research Tasks
《自然》研究：在複雜科研任務上，人類科學家仍勝過最先進的 AI 代理
A new study published in Nature finds that despite dramatic AI progress, human scientists consistently outperform even the most advanced AI agents on complex, open-ended research tasks. Researchers tested leading AI systems — including models from Anthropic, OpenAI, and Google — on real scientific problems requiring multi-step reasoning, hypothesis generation, and experimental design. While AI excels at narrow benchmarks and literature synthesis, human scientists retain a decisive edge in creative problem-solving and adaptive reasoning under uncertainty.
Stanford AI Index 2026: China Has Nearly Closed the U.S. AI Lead
史丹福 2026 AI 指數：中國幾乎追平美國 AI 領先優勢
Benchmark Report · SUN 19 APR 2026
Stanford's Artificial Intelligence Index Report 2026 landed this week with a striking finding: China has nearly erased the United States' lead in artificial intelligence. The country now outpaces the U.S. in AI patents, academic publications, and the deployment of industrial robots. Measured by model performance benchmarks, the gap has narrowed to a matter of percentage points.
The report also found that 53% of the global population is now using generative AI — a milestone the internet took over a decade to reach. AI capabilities in natural sciences, drug discovery, and code generation continue to advance faster than human baselines, raising both excitement and concern across research communities worldwide.
Meta Unveils Muse Spark — First Model From Its New Superintelligence Lab
Meta 推出 Muse Spark——超級智能實驗室首款旗艦模型
Model Release · SUN 19 APR 2026
Meta has unveiled Muse Spark, its first flagship AI model produced by the newly-formed Meta Superintelligence Labs — the division Mark Zuckerberg built around Scale AI founder Alexandr Wang, who joined Meta as part of a $14.3 billion deal. On writing and reasoning benchmarks, Muse Spark significantly outperforms all previous Meta models and comes close to the top results from Google, OpenAI, and Anthropic.
The launch marks Meta's most serious attempt yet to close the gap with frontier AI labs. Muse Spark is available to users across Meta's apps, with enterprise API access rolling out this quarter. Zuckerberg framed it as the first step in building AGI and bringing it to everyone.
Q1 2026 Shatters All VC Records: $300B Raised, 80% Flowing Into AI
2026 年第一季打破所有創投紀錄：$3,000 億融資，八成流入 AI
Venture Capital · SUN 19 APR 2026
The first quarter of 2026 set an all-time record for global venture capital activity. Investors poured $300 billion into roughly 6,000 startups worldwide — a figure that surpasses any previous quarter in venture history. Of that total, $242 billion (80%) went directly to AI companies, cementing artificial intelligence as the defining capital magnet of this era.
The quarter included four of the largest venture rounds ever recorded: OpenAI closed $122 billion at an $852 billion post-money valuation; xAI raised $20 billion; Waymo secured $16 billion; and Anthropic completed a $30 billion Series G in February. Eclipse VC raised $1.3 billion specifically for AI infrastructure and robotics in April, reflecting growing investor appetite beyond pure software.
Anthropic Releases Claude Opus 4.7; Mythos Preview Scores 93.9% on SWE-bench
Anthropic 正式推出 Claude Opus 4.7；Mythos 預覽版在 SWE-bench 取得 93.9% 高分
Model Release · SUN 19 APR 2026
Anthropic made two major announcements this week. First, Claude Opus 4.7 is now generally available, positioned as Anthropic's strongest model across coding, agentic tasks, vision, and complex multi-step reasoning. It extends the Opus 4 architecture with improved instruction-following and faster inference.
Second — and more striking — Anthropic quietly moved Claude Mythos Preview into controlled early access under Project Glasswing, limiting it to 50 enterprise partners. Mythos Preview scores 93.9% on SWE-bench Verified and 94.6% on GPQA Diamond, the highest published scores on both benchmarks. According to internal testing, the model also autonomously identified thousands of zero-day vulnerabilities across major operating systems and browsers, raising significant questions about responsible AI deployment at this capability level.
PwC Study: Top 20% of Companies Capture 75% of AI's Economic Gains
PwC 研究：頂尖兩成企業獨攬 AI 帶來的七成五經濟效益
Industry Research · SUN 19 APR 2026
PwC's 2026 AI Performance Study reveals a stark divide in who is benefiting from the AI boom. Three-quarters of artificial intelligence's economic gains are being captured by just 20% of companies — those leading firms that are focused on growth rather than mere productivity improvements. The remaining 80% of businesses adopting AI see only marginal operational benefits.
The study surveyed over 5,000 organisations across 25 countries. Leading companies share three common traits: they embed AI deeply across business functions rather than in isolated pilots; they treat AI as a growth driver, not a cost-cutter; and they invest heavily in talent and governance alongside technology. PwC warns that without deliberate strategy, AI adoption risks widening the competitive gap between industry leaders and followers to an unprecedented degree.
The quiet protocol that became the backbone of every AI agent on the planet
悄悄成為全球 AI 代理骨幹的那個協議
There's a piece of software you've probably never heard of that now runs underneath almost everything happening in AI. And the number that just came out made me stop and re-read it twice.
Anthropic's Model Context Protocol — MCP for short — hit 97 million installs in March 2026. Per month. That's not total downloads. That's monthly.
MCP is the plumbing that lets AI agents connect to tools, databases, APIs, and apps in a standardised way. Think of it like USB — before USB, every device used a different plug. MCP is the universal plug for AI agents.
Anthropic launched it quietly in late 2024. By March 2026, it had:
→ 97 million monthly SDK installs
→ 10,000+ public MCP servers indexed across registries
→ 300+ compatible clients — editors, chat apps, enterprise platforms
→ Every major AI lab — OpenAI, Google, Meta, Microsoft — now ships MCP-compatible tooling
And then Anthropic did something unusual: they gave it away. In December 2025, they donated MCP to the Agentic AI Foundation under the Linux Foundation — co-founded with OpenAI and Block. A competitor-backed open standard.
This is what infrastructure dominance actually looks like. Not the loudest model, not the biggest funding round — but the protocol that everyone else quietly builds on top of.
🔗 Source: https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation
#Anthropic #MCP #ModelContextProtocol #AIAgents #OpenSource #ArtificialIntelligence #TechInfrastructure
Microsoft just made a $10 billion bet on Japan — and it tells us exactly where AI is heading next
微軟剛在日本押注 100 億美元——這告訴我們 AI 下一步往哪走
Something important happened at the start of April that I think got buried under the usual AI model news cycle.
Microsoft committed $10 billion to Japan — over four years, spanning AI infrastructure, cybersecurity, and workforce development. It's the largest single AI infrastructure commitment by any Western tech company in Asia, building on the $2.9 billion they invested in Japan just two years ago.
Here's what the money actually goes toward:
→ Building cloud and AI data centres partnered with SoftBank and Sakura Internet
→ Training 1 million engineers and developers on AI platforms by 2030
→ A $1 million research grant program for domestic Japanese AI researchers
→ Deepening national cybersecurity partnerships
The move partners Microsoft with SoftBank — the same company that just put $30 billion into OpenAI. That's not a coincidence. This is the world's largest companies betting not just on AI models, but on the physical and human infrastructure that will run those models for decades.
Japan is particularly interesting here. Despite being a global tech power, Japan only ranks mid-tier in AI adoption right now. Microsoft is essentially betting that the country is about to enter its AI S-curve — and they want to own the infrastructure underneath it when it does.
When you see $10 billion going to build the pipes, the real question isn't whether AI will take off — it's who owns the pipes when it does.
🔗 Source: https://news.microsoft.com/source/asia/2026/04/03/microsoft-deepens-its-commitment-to-japan-with-10-billion-investment-in-ai-infrastructure-cybersecurity-workforce/
#Microsoft #AI #Japan #AIInvestment #TechInfrastructure #ArtificialIntelligence #SoftBank #Asia
Someone just bet $122 billion that AI is the biggest thing since the internet — and they might be right
有人剛押注 1,220 億美元賭 AI 是繼互聯網後最重大的事——而他們可能是對的
I spend a lot of time thinking about what it means when really smart people with really large amounts of money make a bet. And last month, some of the world's sharpest investors made the biggest bet in private market history.
OpenAI just closed a $122 billion funding round — the largest private fundraise ever recorded — at a valuation of $852 billion. To put that in perspective: OpenAI is now worth more than the entire GDP of most countries in the world. And it's not even public yet.
Here's who put money in and how much:
→ Amazon: $50 billion (yes, billion)
→ Nvidia: $30 billion
→ SoftBank: $30 billion
→ Regular retail investors (first time ever): $3 billion
And here's why they did it — the business is genuinely exploding:
→ 900 million weekly active users — nearly 1 in 8 people on Earth
→ $2 billion in revenue every single month
→ Enterprise now makes up 40%+ of revenue and growing
→ IPO expected before end of 2026
What I find most striking isn't the number — it's who is investing. Amazon and Nvidia aren't venture capitalists making long-shot bets. They're infrastructure companies making strategic investments in something they believe will become as essential as electricity or the internet.
When the companies selling shovels start buying the gold mine, pay attention.
🔗 Source: https://openai.com/index/accelerating-the-next-phase-ai/
#OpenAI #AI #ArtificialIntelligence #TechInvesting #Startup #FutureOfWork #GenerativeAI #Innovation
Half the world now uses generative AI — and it happened faster than the internet
全球一半人口已在使用 Generative AI——而且普及速度比互聯網還快
I've been telling people for a while now that AI is moving faster than anything we've seen before — and Stanford just proved it with hard data.
Their 2026 AI Index dropped this week, and the headline number stopped me in my tracks: 53% of the global population now uses generative AI. That's over half the world. In just three years. For context, the internet took about 10 years to reach the same milestone. The PC took 16.
Here's what else jumped out at me:
→ 88% of companies are now using AI in at least one core business function
→ Global corporate AI investment hit US$581.7 billion last year — up 130% in a single year
→ AI models now match or beat humans on PhD-level science and math questions
→ US consumers are already getting $172 billion in annual value from AI tools, most of which are free
And the plot twist that nobody in Silicon Valley likes to talk about? The US ranks 24th in AI adoption — behind Singapore (61%), UAE (54%), and a dozen other countries. The countries actually using AI most are not necessarily the ones building it.
The revolution is already here. The only question is whether you're riding it or watching it.
🔗 Full report: https://hai.stanford.edu/ai-index/2026-ai-index-report
#AI #ArtificialIntelligence #StanfordAI #FutureOfWork #TechTrends #AIAdoption #GenerativeAI #Innovation
Google just made AI 6× cheaper to run — without losing a single bit of quality
Google 剛讓 AI 運行成本降低 6 倍——而且沒有犧牲任何一點品質
Here's something that doesn't get enough attention: one of the biggest barriers to AI being everywhere is simply memory. Running a large language model requires an enormous amount of RAM — which is why it still costs so much to run AI at scale, and why truly smart AI on your phone feels far away.
Google may have just changed that equation.
Their research team unveiled something called TurboQuant at ICLR 2026 — a new algorithm that compresses the "working memory" of AI models (technically called the KV cache) from 16 bits down to just 3 bits. Think of it like discovering a much better ZIP format specifically for AI's brain.
The numbers are genuinely impressive:
→ 6× less RAM needed to run large AI models
→ Up to 8× faster processing speed on high-end GPUs
→ Zero accuracy loss — the AI answers just as well, just uses a fraction of the memory
→ No need to retrain or modify existing models — it works as a drop-in
What this means in plain English: AI that currently needs a data centre could soon run on your laptop. AI assistants that feel slow could feel instant. And the cost of running AI services could drop dramatically for everyone.
The open-source community has already started building with it — which means we'll likely see real products using this within months.
🔗 Google Research blog: https://research.google/blog/turboquant-redefining-ai-efficiency-with-extreme-compression/
#AI #GoogleAI #TurboQuant #LLM #MachineLearning #TechInnovation #EdgeAI #ArtificialIntelligence