01

史上最大 IPO 敲钟!Anthropic 登陆纳斯达克,开盘暴涨 34%,募资 813 亿美元改写资本史 Largest IPO in History! Anthropic Debuts on NASDAQ, Opens Up 34%, Raises $81.3 Billion

北京时间 11 月 20 日凌晨(美东 11 月 19 日),Claude 母公司 Anthropic 以代码 ANT 挂牌纳斯达克。发行价 83 美元,开盘冲至 111.50 美元(涨 34%),盘中最高 115.10 美元。募资 813 亿美元超 SpaceX,成史上最大 IPO。 Anthropic debuted on NASDAQ as ANT overnight Beijing Time (Nov 19 US). Priced at $83, shares opened at $111.50 (+34%) and hit $115.10 intraday. The $81.3B raise tops SpaceX to become the largest IPO ever.

XINHUANET / TECHCRUNCH 21:30 发布 21:30 (UTC+8)
世纪IPOHistoric IPOAnthropicAnthropic纳斯达克NASDAQ
详情Details 收起Close
02

谷歌七周内第四款模型!Gemini 4 Neon 发布:100 万上下文、速度翻倍、价格击穿旗舰 Google's Fourth Model in Seven Weeks! Gemini 4 Neon Lands: 1M Context, 2x Speed, Undercuts Flagship

Google DeepMind 于 11 月 18 日(美东)发布 Gemini 4 Neon,为七周内第四款模型。支持文本/图像/视频,100 万 token 上下文,Artificial Analysis 智能指数 61 分排全球第三,速度约 110 tokens/s(Carbon 为 57)。 Google DeepMind shipped Gemini 4 Neon on Nov 18, its fourth model in seven weeks. It handles text, images and video with a 1M-token context, scores 61 on the Artificial Analysis Index (#3 globally), and runs at ~110 tokens/s.

GOOGLE DEEPMIND 20:45 发布 20:45 (UTC+8)
Gemini 4Gemini 4谷歌GoogleNeonNeon
详情Details 收起Close
03

Gemini 住进你的 Mac!谷歌上线电脑操控:跨文件、桌面应用与网页自主干活,步步需你批准 Gemini Moves Into Your Mac! Google Adds Computer Use: Autonomous Control Across Files, Apps and Web

谷歌为 macOS 版 Gemini 桌面应用加入 computer use 能力,可跨本地文件、原生 Mac 应用和网页代为操作,并在聊天中实时预览。每项操作均需用户批准,先面向美国英语用户开放。 Google added computer use to the macOS Gemini app, letting it operate local files, native Mac apps and the web with live chat previews. Every action requires user approval; available to US English users first.

THE VERGE / TECHCRUNCH 20:10 发布 20:10 (UTC+8)
GeminiGemini电脑操作Computer UsemacOSmacOS
详情Details 收起Close
04

ChatGPT 实时听写上线!说话即出字、10 分钟录音上限被彻底取消,全平台全套餐推送 ChatGPT Live Dictation Is Live! Speak and See Text Instantly, 10-Minute Cap Removed, All Platforms

OpenAI 为 ChatGPT 上线 Live Dictation 实时听写:用户说话时输入框同步显示转写文字,彻底取消原有 10 分钟录音时长限制。覆盖 Web、桌面、iOS、Android,全球所有套餐(含免费用户)逐步开放。 OpenAI launched Live Dictation for ChatGPT: text appears in the input box as you speak, and the old 10-minute recording cap is gone. Rolling out across Web, desktop, iOS and Android to all tiers, including free users.

OPENAI / ADAPTD.AI 19:30 发布 19:30 (UTC+8)
OpenAIOpenAI实时听写Live DictationChatGPTChatGPT
详情Details 收起Close
05

世界模型能上手玩了!谷歌 Genie 4 正式向开发者开放 API:720p 实时交互、原生立体声 Playable World Models! Google Opens Genie 4 API to Developers: 720p Real-Time Interaction and Native Stereo Audio

谷歌 11 月 19 日通过 Gemini API 与 AI Studio 向付费开发者开放世界模型 Genie 4。支持文本或图像生成 720p/24fps 连续可交互世界,同步输出环境音与对白,键盘操控延迟低至 0.2 秒。 Google opened its Genie 4 world model to paid developers on Nov 19 via the Gemini API and AI Studio. It generates 720p/24fps interactive worlds from text or images with native stereo audio, and keyboard control latency as low as 0.2s.

GOOGLE / 凤凰网科技 18:50 发布 18:50 (UTC+8)
谷歌GoogleGenie 4Genie 4世界模型World Model
详情Details 收起Close
06

AI 开始设计芯片!OpenAI 模型改写 AMD 高端 GPU 底层代码,5 个月干完一年设计进展 AI Designs Chips! OpenAI Model Rewrote AMD's High-End GPU RTL, Delivering a Year of Progress in 5 Months

OpenAI 披露其模型修改了 SystemVerilog RTL 代码,参与 AMD 一款高端 GPU 设计。经专业工具评估找到最高 14% 的优化空间,不到 5 个月完成约一年的设计进展,AMD 已将部分修改纳入送交制造的设计。 OpenAI revealed its model modified SystemVerilog RTL code for an AMD high-end GPU, finding up to 14% optimization per pro tools and compressing about a year of design work into under five months. AMD has folded some changes into its tape-out.

OPENAI / 钛媒体 18:15 发布 18:15 (UTC+8)
OpenAIOpenAIAMDAMD芯片设计Chip Design
详情Details 收起Close
07

端到端取代级联方案!火山引擎豆包 SeedRealtime 1.0 全量上线:边看边听还能主动接话 End-to-End Replaces Cascade! Volcano Engine Launches Doubao SeedRealtime 1.0: See, Listen and Speak Proactively

火山引擎 11 月 19 日正式上线豆包实时音视频通话模型 SeedRealtime 1.0 的 API 服务,把声音、画面、时序与表达统一进一个端到端模型,取代 ASR、VLM、TTS 级联方案,音视频对话节奏问题减少一半。 Volcano Engine launched the API for its SeedRealtime 1.0 real-time audio-video model on Nov 19, unifying sound, vision, timing and expression in one end-to-end model to replace ASR/VLM/TTS cascades, halving conversational rhythm issues.

IT之家 / 网易 17:40 发布 17:40 (UTC+8)
火山引擎Volcano Engine豆包Doubao实时音视频Realtime AV
详情Details 收起Close
08

把 1200 亿参数塞进桌面!英伟达 RTX Spark 主机上市:128GB 统一内存,断网也能跑智能体 120B Params on Your Desk! Nvidia RTX Spark Desktops Launch: 128GB Unified Memory, Agents Run Offline

华硕、戴尔、微星 RTX Spark 紧凑型桌面主机于 11 月 18 日上市,配 20 核 Grace CPU、Blackwell RTX GPU 及最高 128GB 统一内存。官方称可本地运行 1200 亿参数大模型,外出时主机仍可持续运行智能体与定时任务。 ASUS, Dell and MSI launched compact RTX Spark desktops on Nov 18 with 20-core Grace CPUs, Blackwell RTX GPUs and up to 128GB unified memory — said to run 120B-parameter models locally and keep agents going while you're away.

NVIDIA 官方博客 17:00 发布 17:00 (UTC+8)
英伟达NvidiaRTX SparkRTX Spark本地大模型Local LLM
详情Details 收起Close
09

跨城市训练大模型!谷歌 Streaming DiLoCo 上线:流量暴降逾 60 倍、步速狂飙 9.57 倍 Train LLMs Across Cities! Google's Streaming DiLoCo Cuts Traffic 60x+ and Boosts Step Speed 9.57x

谷歌将 Streaming DiLoCo 加入 MaxText 训练框架。测试显示,在 10 Gbit/s 带宽下跨 slice 通信流量减少逾 60 倍、训练 step 加速 9.57 倍,且 loss 与数据并行方案相当,为跨地域分布式训练扫清瓶颈。 Google added Streaming DiLoCo to its MaxText framework. Tests show 60x+ less cross-slice traffic at 10 Gbit/s and a 9.57x step speedup, with loss on par with data parallelism — clearing a key cross-region training bottleneck.

GOOGLE DEVELOPERS 16:20 发布 16:20 (UTC+8)
谷歌GoogleDiLoCoDiLoCo分布式训练Distributed Training
详情Details 收起Close
10

具身大脑杀进汽车产线!谷歌 Gemini Robotics 2.5 亮相:奔驰工厂 96% 任务成功率 Embodied Brains Hit the Assembly Line! Google Gemini Robotics 2.5 Debuts: 96% Task Success on Mercedes Factory Floor

谷歌 DeepMind 在 11 月中旬 CoRL 2026 会议期间发布 Gemini Robotics 2.5,已在奔驰汽车产线实现 96% 任务成功率,Apollo 2 人形机器人基于该模型自主完成零件分拣,具身智能临界点渐近。 Google DeepMind unveiled Gemini Robotics 2.5 around CoRL 2026 in mid-November, achieving 96% task success on a Mercedes production line, with the Apollo 2 humanoid autonomously sorting parts — signaling the embodied-AI tipping point nears.

科创板日报 / 中华网 15:45 发布 15:45 (UTC+8)
具身智能Embodied AI谷歌GoogleGemini RoboticsGemini Robotics
详情Details 收起Close
11

芯片与云端双重断供!美国新规封锁中国 AI:月之暗面取消 1.5 亿美元 B300 订单 Double Hit on Chips and Cloud! New US Rules Squeeze Chinese AI: Moonshot Cancels $150M B300 Order

美国新出口规则要求先进 AI 芯片在用于服务中国企业的海外数据中心时须取得许可,波及新加坡、马来西亚、印尼等地。月之暗面取消印尼 1.5 亿美元英伟达 B300 容量交易,台积电与三星停止部分对华出货。 New US export rules require licenses for advanced AI chips used in overseas data centers serving Chinese firms, hitting Singapore, Malaysia and Indonesia. Moonshot canceled a $150M Nvidia B300 deal in Indonesia; TSMC and Samsung halted some China shipments.

THE DECODER / ABC NEWS 15:10 发布 15:10 (UTC+8)
出口管制Export Controls英伟达Nvidia中国AIChinese AI
详情Details 收起Close
12

AI 电耗翻倍!IEA《世界能源展望》:2030 年数据中心用电或达 1000 太瓦时,电网成扩张死穴 AI Power Surge! IEA World Energy Outlook Sees Data-Center Demand Doubling to 1,000 TWh by 2030

国际能源署(IEA)在 11 月 19 日发布的《世界能源展望 2026》中预测,到 2030 年数据中心年用电量可能翻倍至约 1000 太瓦时。到 2035 年需新建 2600 万公里线路,未来十年电网支出或增加 75%(现行政策情景)。 The IEA's World Energy Outlook 2026 (Nov 19) sees data-center electricity use doubling to ~1,000 TWh by 2030. By 2035, 26 million km of new lines are needed and grid spending may rise 75% this decade under current policies.

CNBC / THE GUARDIAN 14:30 发布 14:30 (UTC+8)
IEAIEA数据中心Data Centers能源Energy
详情Details 收起Close
13

纳德拉:6 万人买断转岗,未来每份工作都变成 AI 代理管理 Nadella: 60,000 Buyouts as Every Job Turns Into AI Agent Management

微软 CEO 萨提亚·纳德拉 11 月 18 日在 DealBook 峰会表示,公司以买断方式让员工选择是否加入 AI-first 重组,10 月已向 6 万名员工提供买断。他认为 AI 不会取代工程师,但每项工作都会变成某种形式的 AI 代理管理。 Microsoft CEO Satya Nadella said at the Nov 18 DealBook Summit that the firm offered buyouts in October to 60,000 employees as part of an AI-first restructuring. He argues AI won't replace engineers, but every job will become some form of AI agent management.

TECHCRUNCH / CNBC 13:55 发布 13:55 (UTC+8)
微软Microsoft纳德拉NadellaAI重组AI Restructure
详情Details 收起Close
14

45 亿美元押注开源模型!Palo Alto 收购 Applied Compute,财报营收超预期暴涨 39% $4.5B Bet on Open Models! Palo Alto Acquires Applied Compute as Earnings Beat With 39% Revenue Growth

Palo Alto Networks 财报超预期,并同意以股票方式斥资 45 亿美元收购 Applied Compute,扩展其在开放权重 AI 模型训练与服务领域的能力。季度收入 34.4 亿美元,同比增长 39%,盘前股价一度涨约 8–10%。 Palo Alto Networks beat earnings and agreed to acquire Applied Compute in a $4.5B all-stock deal to expand open-weight AI model training and serving. Quarterly revenue hit $3.44B, up 39% YoY, with shares up 8–10% premarket.

CNBC / SILICONANGLE 13:20 发布 13:20 (UTC+8)
Palo AltoPalo Alto并购Acquisition开源模型Open Models
详情Details 收起Close
15

具身智能的 GPT 时刻?浙大宇树开源 500 万条机器人动作数据,总时长超 1 万小时 Embodied AI's GPT Moment? ZJU and Unitree Open-Source 5 Million Robot Motion Sequences, Over 10,000 Hours

浙江大学与宇树科技具身智能联合实验室开源预训练数据集 Movens-5M,收录 500 万条经仿真验证的机器人动作、总时长超 1 万小时,覆盖 297 种动作与 16 类场景,配套的 Movens-Zero 可据文字、语音、音乐生成全身动作。 Zhejiang University and Unitree's joint lab open-sourced Movens-5M, a pretraining dataset of 5 million simulation-verified robot motions totaling over 10,000 hours across 297 actions and 16 scenarios; its Movens-Zero generates full-body motions from text, speech or music.

机器之心 / 网易 12:40 发布 12:40 (UTC+8)
具身智能Embodied AI宇树Unitree开源数据Open Dataset
详情Details 收起Close
16

AI 攻防升级!Sakana AI 成立独立安全研究组,推集体网络防御与 Fugu-Cyber AI Cyber Arms Race! Sakana AI Forms Independent Security Research Group for Collective Cyber Defense and Fugu-Cyber

Sakana AI 今日宣布成立独立的安全研究组(Security Research Group),聚焦 AI 驱动的集体网络防御,并将 7 月发布的 Fugu-Cyber 纳入该组工作。不过其 Fugu-Cyber 得分由 Sakana 自行报告,尚未提供基线对比。 Sakana AI announced an independent Security Research Group today focused on AI-driven collective cyber defense, folding in the Fugu-Cyber project released in July. Its Fugu-Cyber scores are self-reported and lack baseline comparisons.

SAKANA AI / EXPLAINX 12:05 发布 12:05 (UTC+8)
Sakana AISakana AI网络安全CybersecurityAI防御AI Defense
详情Details 收起Close
17

苹果要造自己的 AI 智能体!招聘信息曝光:跨 App、浏览器、命令行与 API,还能调度子代理 Apple Building Its Own AI Agent! Job Listing Reveals Cross-App, Browser, CLI and API Control With Sub-Agents

彭博社援引招聘信息报道,苹果正招聘工程师牵头打造 agentic AI 助手,可跨应用、浏览器、命令行、代码与 API 工作,并使用子代理。报道称该项目与 Siri、UI Frameworks、Foundation Models 相关,但仅为招聘意向。 Bloomberg reports Apple is hiring an engineer to lead an agentic AI assistant that works across apps, browsers, command lines, code and APIs with sub-agents. Tied to Siri and Foundation Models, it's a hiring signal, not a product.

BLOOMBERG / THE VERGE 11:30 发布 11:30 (UTC+8)
苹果AppleAI AgentAI Agent招聘Hiring
详情Details 收起Close
18

苹果开源 Depth Pro 2!四个深度模型可测真实米制距离,照片视频实时相机都能用 Apple Open-Sources Depth Pro 2! Four Depth Models Measure True Metric Depth Across Photos, Video and Live Camera

苹果发布 Depth Pro 2,包含 S、B、L、H+ 四个开放权重深度模型与一个焦距估计器,可从照片、视频和实时相机画面估计以米为单位的逐像素深度。采用研究用途许可,尚未公布基准数字。 Apple released Depth Pro 2 with four open-weight depth models (S, B, L, H+) and a focal-length estimator, producing per-pixel metric depth in meters from photos, video and live camera. Research-use license; no benchmarks yet.

APPLE RESEARCH / HUGGING FACE 10:50 发布 10:50 (UTC+8)
苹果AppleDepth ProDepth Pro深度估计Depth Estimation
详情Details 收起Close
19

英伟达罕见推迟 10-Q 财报!黄仁勋内部调查发酵,SEC 就信息披露展开民事调查 Nvidia Unusually Delays 10-Q Filing! Huang Internal Probe Deepens as SEC Launches Civil Disclosure Investigation

英伟达 11 月 19 日宣布将季度报告 10-Q 推迟至「未来数周」提交,一改与财报同日发布的惯例且未说明理由。UBS 指与黄仁勋内部调查及新增风险因子披露有关,《华尔街日报》称 SEC 正就披露是否误导投资者展开民事调查。 Nvidia said on Nov 19 it will delay its 10-Q filing to 'the coming weeks,' breaking with its practice of filing alongside earnings and offering no reason. UBS linked it to the internal probe of Jensen Huang and possible new risk factors, while the WSJ reported a SEC civil investigation into disclosures.

CHIPDISPATCH / WSJ 10:15 发布 10:15 (UTC+8)
英伟达NvidiaSEC调查SEC Probe黄仁勋Jensen Huang
详情Details 收起Close
20

IMF 警告:AI 需求将推高全球利率,AI 硬件已占全球商品贸易逾十分之一 IMF Warns AI Demand Will Push Up Global Rates as AI Hardware Tops 1/10 of World Goods Trade

IMF 总经理格奥尔基耶娃表示,AI 需求与中东能源冲击将推高全球通胀,要求多地采取更紧货币政策。预计 2026 年全球通胀达 4.7%,AI 硬件已占全球商品贸易超过十分之一。 IMF chief Kristalina Georgieva said AI demand and energy shocks will lift global inflation and prompt tighter policy. Global inflation is forecast at 4.7% for 2026, with AI hardware now over one-tenth of world goods trade.

THE GUARDIAN / FXSTREET 09:40 发布 09:40 (UTC+8)
IMFIMF全球通胀Global InflationAI硬件AI Hardware
详情Details 收起Close
21

AI 眼镜别买!荷兰监管机构测试 6 款后警告:误报救护车、乱读药盒、录像灯可被胶带遮住 Don't Buy AI Glasses! Dutch Watchdog Tests 6 and Warns of False Ambulance Alerts, Misread Medicine and Tape-Covered Lights

荷兰数据保护机构 AP 测试 6 款 AI 摄像眼镜后建议消费者不要购买,理由是准确性、隐私和安全风险突出。测试中出现误报救护车将至、误读药品说明、向未成年人给出不当约会建议等问题。 Dutch data watchdog AP tested six AI camera glasses and advised consumers against buying them over accuracy, privacy and security risks — including false ambulance alerts, misread medicine labels and leaks to non-EU servers.

BBC / REUTERS 09:00 发布 09:00 (UTC+8)
AI眼镜AI Glasses隐私Privacy监管Regulation
详情Details 收起Close
22

合成生物首个大模型!天津大学发布「天华开物」:560 亿参数,一周给出酶设计方案 First Foundation Model for Synthetic Biology! Tianjin University Releases 'Tianhua Kaiwu': 56B Params, Enzyme Designs in a Week

天津大学 11 月 19 日发布「天华开物」酶设计大模型,这是合成生物领域首个公开发布的大模型,由中国自主研发,参数规模达 560 亿。该模型与自动化实验设备结合形成工业酶闭环迭代优化系统,一周内可给出酶设计方案。 Tianjin University released 'Tianhua Kaiwu' on Nov 19, the first publicly released foundation model in synthetic biology, developed in China with 56 billion parameters. Paired with automated lab equipment, it forms a closed-loop enzyme optimization system delivering designs within a week.

中国青年报 / 新浪 08:20 发布 08:20 (UTC+8)
合成生物Synthetic Biology天津大学Tianjin University酶设计Enzyme Design
详情Details 收起Close
← 浏览往期存档 · BROWSE ARCHIVE →