单词学习：![[2026042035138099.mp3]]

新闻播报：![[2026042011933224.mp3]]

Anthropic 年化收入突破 300 亿美元，估值直逼万亿美金大关

简报：

• Anthropic 年化收入目前已超过 300 亿美元，较去年年底增长逾三倍，在与 OpenAI 的竞争中展现出强劲势头。
• 公司财务状况显著改善，毛利率从 2024 年的 -94% 跃升至 2025 年的 +40%，年化收入从不足 10 亿美元猛增至 90 亿美元。
• 业绩增长主要由 Claude Code、Cowork 产品以及推理模型代币销售的持续上升所带动。
• 首席财务官克里希纳·拉奥（Krishna Rao）已收到基于 8000 亿美元估值的投资提案，部分投资者预测其市值有望突破 1 万亿美元。
• 尽管发展势头强劲，但知情人士称公司目前暂无新融资计划，最早需待 2026 年 5 月董事会会议后再行讨论。

相关链接：

• Anthropic收入暴增，市值或突破万亿美元大关

Anthropic’s Annualized Revenue Surpasses $30 Billion, Valuation Nears the Trillion-Dollar Mark 🚀

Brief:

• Anthropic’s annualized revenue has now exceeded $30 billion, more than tripling since the end of last year, showing strong momentum in its competition with OpenAI.
• The company’s financial position has improved sharply, with gross margin rising from -94% in 2024 to +40% in 2025, while annualized revenue surged from less than dJyH6kDkfC5RlM16.answerText billion to $9 billion.
• The growth has been driven mainly by continued increases in sales of Claude Code, the Cowork product, and tokens for reasoning models.
• Chief Financial Officer Krishna Rao has received investment proposals based on an $800 billion valuation, and some investors predict the company’s market value could surpass dJyH6kDkfC5RlM16.answerText trillion 💰.
• Despite its strong momentum, sources say the company currently has no plans for a new funding round, with any discussion unlikely before the board meeting in May 2026.

Related link:

• Anthropic revenue surges, valuation may break the trillion-dollar threshold 📈

阿里推出新一代千问旗舰模型Qwen3.6-Max早期预览版

简报：

• 阿里正式发布Qwen3.6-Max-Preview，该模型为新一代千问旗舰模型的早期预览版
• 该预览版模型拥有更强的世界知识和指令遵循能力
• 该模型在多项基准测试上的智能体编程表现获得显著提升

相关链接：

• 阿里发布Qwen3.6-Max预览版

Alibaba Launches Early Preview of New-Generation Flagship Qwen Model Qwen3.6-Max 🚀

Brief:

• Alibaba has officially released Qwen3.6-Max-Preview, an early preview version of the new-generation flagship Qwen model
• This preview model features stronger world knowledge and improved instruction-following capabilities
• The model has achieved significant improvements in agentic coding performance across multiple benchmark tests 💡

Related link:

• Alibaba releases Qwen3.6-Max preview

Nature最新研究：大模型存在“潜意识学习”，不良特征可通过纯数字蒸馏悄然传染

简报：

• Nature最新发表的论文首次证实大型语言模型（LLM）存在“潜意识学习”现象，不良行为特征能通过纯数字、代码或思维链等看似无害的序列传递给下游模型。
• 实验表明，携带特定偏好（如“猫头鹰”）或危险特征的“老师模型”，即使生成的纯数字序列不包含任何相关语义，用于训练的“学生模型”仍会自发表现出对这些特征的强烈偏好。
• 机制研究显示，在模型蒸馏过程中，学生模型在权重空间中复制了老师的隐性特征梯度，这种信号隐藏在数据的统计分布纹路中，超越了表层语义。
• 这一发现揭示了AI安全系统面临三大失效风险：供应链可能遭受“隐形投毒”攻击，模型间可能进行人类无法察觉的通信，且目前基于语义的安全评估工具无法有效探测此类非语义污染。
• 论文促使AI行业重新审视模型安全范式，防范重点需要从“检查模型输出是否有毒”转向“评估模型权重是否干净”。

相关链接：

• Nature 论文揭秘：AI 不良特征可隐藏在纯数字中实现“隔空传毒”

Latest Nature Study: Large Models Exhibit “Subliminal Learning,” and Harmful Traits Can Quietly Spread Through Purely Numeric Distillation 🧠🔢⚠️

Brief:

• A newly published Nature paper provides the first evidence that large language models (LLMs) exhibit a “subliminal learning” phenomenon, in which undesirable behavioral traits can be transmitted to downstream models through seemingly harmless sequences such as pure numbers, code, or chains of thought.
• Experiments show that when a “teacher model” carries specific preferences (such as for “owls”) or dangerous traits, a “student model” trained on purely numeric sequences generated by the teacher—even when those sequences contain no related semantic content—can still spontaneously develop a strong preference for those traits.
• Mechanistic analysis indicates that during model distillation, the student model copies the teacher’s latent feature gradients in weight space. This signal is hidden in the statistical texture of the data distribution and goes beyond surface-level semantics.
• This finding reveals three major failure risks for AI safety systems: supply chains may be vulnerable to “invisible poisoning” attacks, models may be able to communicate in ways humans cannot detect, and current semantics-based safety evaluation tools are unable to effectively identify this kind of non-semantic contamination.
• The paper prompts the AI industry to rethink its model security paradigm: the focus of prevention needs to shift from “checking whether model outputs are toxic” to “evaluating whether model weights are clean.”

Related link:

• Nature paper reveals: Harmful AI traits can hide in pure numbers and spread “remotely”

蚂蚁灵光闪应用数量突破3000万，投入1亿元专项激励扶持AI创作者

简报：

• 蚂蚁集团旗下AI产品灵光App宣布正式启动“灵光闪应用创作者激励计划”，投入1亿元专项基金扶持开发者
• 自2025年底发布至今四个月内，该平台通过自然语言生成的闪应用数量已突破3000万个，涵盖工具、游戏等多个维度
• 激励计划于4月20日正式开启，设立日度与周度双重机制，每日最高奖金100万元覆盖万个应用，每周单创作者最高奖金100万元
• 灵光App同步上线“灵光圈”公共创意社区，并迭代多智能体协作、全模态生成及手机端原生能力集成等功能
• 平台通过降低编程门槛，使非编程背景的普通用户能够通过意图驱动创作AI应用，推动个体创造力释放

相关链接：

• 蚂蚁灵光App闪应用突破3000万个，宣布投入1亿元启动专项创作者激励

Ant Group’s Lingguang Flash Apps Surpass 30 Million, Launches RMB 100 Million Special Incentive Program to Support AI Creators 🚀

Briefing:

• Lingguang App, an AI product under Ant Group, officially announced the launch of the “Lingguang Flash App Creator Incentive Program,” backed by a dedicated RMB 100 million fund to support developers
• In just four months since its release at the end of 2025, the platform has generated more than 30 million flash apps through natural language, covering multiple categories such as tools and games
• The incentive program officially begins on April 20, introducing both daily and weekly reward mechanisms, with daily prizes of up to RMB 1 million distributed across 10,000 apps, and weekly prizes of up to RMB 1 million for a single creator
• Lingguang App has also launched “Lingguang Circle,” a public creative community, while upgrading features such as multi-agent collaboration, full-modal generation, and integration of native mobile capabilities ✨
• By lowering the barrier to programming, the platform enables ordinary users without coding backgrounds to create AI apps through intent-driven workflows, helping unlock individual creativity 💡

Related Link:

• Ant Lingguang App Flash Apps Surpass 30 Million, Announces RMB 100 Million Special Creator Incentive Program

微软33亿美元打造的最强AI数据中心提前投运，内置数十万块GB200芯片

简报：

• 微软CEO萨提亚·纳德拉于4月16日宣布，位于美国威斯康辛州的“全球最强”AI数据中心Fairwater已提前上线。
• 该数据中心总投资规模达33亿美元，搭载数十万块英伟达GB200超级芯片，性能为全球最快超算的十倍，主要用于前沿AI模型的训练。
• 设施占地约127.6公顷，包含三栋建筑，采用闭环液冷系统，建设过程使用了约1.2万吨结构钢材及数百公里长的地下电缆与机械管道。
• 微软计划在Fairwater数据中心投运后，于美国其他地区继续建设多个同类型的AI数据中心。

相关链接：

• 微软“全球最强”AI 数据中心 Fairwater 提前上线，包含数十万块 GB200

Microsoft’s $3.3 Billion “World’s Most Powerful” AI Data Center Goes Online Ahead of Schedule, Packed with Hundreds of Thousands of GB200 Chips 🚀

Brief:

• On April 16, Microsoft CEO Satya Nadella announced that Fairwater, the “world’s most powerful” AI data center located in Wisconsin, USA, had officially gone online ahead of schedule.
• The data center represents a total investment of $3.3 billion and is equipped with hundreds of thousands of Nvidia GB200 superchips. Its computing performance is said to be ten times that of the world’s fastest supercomputer, and it will primarily be used to train cutting-edge AI models.
• Covering approximately 127.6 hectares, the facility includes three buildings and uses a closed-loop liquid cooling system. Its construction required about 12,000 tons of structural steel, along with hundreds of kilometers of underground electrical cables and mechanical piping 🏗️
• Following the launch of the Fairwater data center, Microsoft plans to continue building several similar AI data centers in other parts of the United States.

Related link:

• Microsoft’s “World’s Most Powerful” AI Data Center Fairwater Goes Online Ahead of Schedule, Featuring Hundreds of Thousands of GB200 Chips

英特尔14A工艺年底或迎重磅客户，瑞银猜测含英伟达与苹果等

简报：

• 英特尔即将推出的14A工艺技术，预计将在今年年底吸引知名企业客户的加入。
• 瑞银猜测，英特尔该新技术的潜在客户将包括英伟达、苹果、谷歌和AMD等。
• 瑞银还强调，英特尔有可能将其位于俄亥俄州的晶圆厂与马斯克的Terafab项目合并，以进一步提升其晶圆代工声誉。

相关链接：

• 英特尔14A工艺整装待发

Intel’s 14A Process May Land Major Customers by Year-End, UBS Speculates Nvidia and Apple Could Be Among Them 🚀

Brief:

• Intel’s upcoming 14A process technology is expected to attract well-known corporate customers by the end of this year.
• UBS speculates that potential clients for Intel’s new technology could include Nvidia, Apple, Google, and AMD.
• UBS also emphasized that Intel may combine its Ohio wafer fab with Musk’s Terafab project to further enhance its foundry reputation. 🔧✨

Related link:

• Intel’s 14A Process Is Ready to Go

Anthropic发布Claude Design，打造从创意到代码的设计开发全链路

简报：

• Anthropic正式推出实验性产品Claude Design，支持通过文字描述、图片或办公文档快速生成交互原型、演示文稿、单页报告及营销物料
• 该工具由Claude Opus 4.7视觉模型驱动，用户可通过自然语言对话、自定义滑块或内联评论进行精细化迭代，甚至能自动应用团队代码库中的品牌规范
• 生成的设计成果支持导出至Canva继续编辑，或输出为PDF、PPTX及独立HTML文件，极大降低了非设计人员的视觉创作门槛
• 核心功能支持“一键打包移交给Claude Code”，可自动转化设计资产与规格用于前端开发，缩短从创意到可运行产品的周期
• 目前Claude Design处于研究预览阶段，面向Claude Pro、Max、Team及Enterprise订阅用户逐步开放使用

相关链接：

• Claude推出Claude Design：支持一句话生成交互原型、PPT及营销物料

Anthropic Launches Claude Design, Building a Full Design-to-Development Workflow from Idea to Code ✨

Brief:

• Anthropic has officially introduced its experimental product Claude Design, which enables users to quickly generate interactive prototypes, presentations, one-page reports, and marketing materials from text prompts, images, or office documents.
• Powered by the Claude Opus 4.7 vision model, the tool supports fine-tuned iteration through natural language conversations, custom sliders, or inline comments, and can even automatically apply brand guidelines from a team’s codebase.
• The generated design outputs can be exported to Canva for further editing, or saved as PDF, PPTX, and standalone HTML files, significantly lowering the barrier to visual content creation for non-designers 🎨
• A core feature allows users to hand projects off to Claude Code with one click, automatically converting design assets and specifications for frontend development and shortening the cycle from concept to working product.
• Claude Design is currently in the research preview stage and is being gradually rolled out to Claude Pro, Max, Team, and Enterprise subscribers 🚀

Related link:

• Claude Launches Claude Design: Generate Interactive Prototypes, PPTs, and Marketing Materials with a Single Sentence