日期：2026-04-01
本期聚焦：重点关注AI coding、AI SRE、AI辅助生活产品与工作流。
-
Salesforce announces an AI-heavy makeover for Slack, with 30 new features（TechCrunch AI）
中文摘要：Salesforce 宣布对 Slack 进行大规模 AI 升级，推出 30 项新功能，旨在将 Slack 从企业通讯工具转变为多功能业务平台。核心亮点包括：Slackbot AI 代理获得可复用的 AI 技能功能，用户可创建自定义任务并在多种场景下应用；Slackbot 现支持作为 MCP（Model Context Protocol）客户端，可与外部服务和 Salesforce 的 Agentforce 平台无缝协作；新增会议转录与摘要功能，帮助用户快速回顾会议要点；Slackbot 还能监控桌面活动，基于用户的交易、对话、日历等数据提供智能建议。这些功能将在未来几个月陆续上线，标志着 Salesforce 正通过 AI 深度整合，将 Slack 打造成企业核心业务流程的不可或缺平台。
English Summary: Salesforce announced a major AI overhaul for Slack with 30 new features, transforming it from a messaging tool into a versatile business platform. Key updates include reusable AI skills for Slackbot, allowing users to create custom tasks; MCP client support for seamless integration with external services and Agentforce; meeting transcription and summarization capabilities; and desktop activity monitoring for intelligent suggestions. These features will roll out in the coming months as Salesforce aims to make Slack central to enterprise workflows.
-
OpenAI, not yet public, raises $3B from retail investors in monster $122B fund raise（TechCrunch AI）
中文摘要：OpenAI 完成新一轮 1220 亿美元融资，估值达 8520 亿美元，为其史上最大融资轮。本轮由软银、Andreessen Horowitz 等领投，亚马逊、英伟达和微软参与投资。其中约 30 亿美元来自散户投资者，OpenAI 还将被纳入 ARK Invest 管理的 ETF，为 IPO 铺路。公司月营收达 20 亿美元，拥有超 9 亿周活用户和 5000 万订阅者，企业业务占比升至 40%。此次融资不仅补充了资金弹药，更被解读为 OpenAI 在向公开市场讲述其增长故事的信号。
English Summary: OpenAI closed a $122 billion funding round at an $852 billion valuation, its largest to date. Co-led by SoftBank, Andreessen Horowitz, and others with participation from Amazon, Nvidia, and Microsoft, the round included $3 billion from retail investors. OpenAI will also join ARK Invest ETFs ahead of its expected IPO. The company reports $2 billion monthly revenue, 900 million weekly active users, and 50 million subscribers, with enterprise revenue now at 40%.
-
PyPI Supply Chain Attack Compromises LiteLLM, Enabling the Exfiltration of Sensitive Information（InfoQ AI/ML）
中文摘要：PyPI 上的 LiteLLM 包遭遇供应链攻击，被下载超过 4 万次。恶意版本 1.82.8 可窃取 SSL/SSH 密钥、云服务商凭证、Kubernetes 配置、Git 凭证、API 密钥、加密货币钱包等敏感信息。FutureSearch 研究员 Callum McMahon 发现该攻击，他在通过 Cursor 启动本地 MCP 服务器时触发下载，导致系统被感染。幸运的是，恶意代码存在递归 fork 的缺陷，导致系统崩溃从而暴露了攻击。PyPI 团队在约 40 分钟内隔离了受影响的包。此次攻击源于 Trivy 安全扫描器的漏洞，使攻击者得以入侵 LiteLLM 的发布管道。
English Summary: A supply chain attack compromised the LiteLLM package on PyPI, with over 40,000 downloads of the malicious version 1.82.8. The malware could exfiltrate SSL/SSH keys, cloud credentials, Kubernetes configs, Git credentials, API keys, and crypto wallets. Discovered by FutureSearch researcher Callum McMahon after his system crashed due to a recursive fork bug in the malware, the package was quarantined within 40 minutes. The attack exploited a vulnerability in Trivy to gain unauthorized access to LiteLLM's publishing pipeline.
-
Agentic AI Patterns Reinforce Engineering Discipline（InfoQ AI/ML）
中文摘要：Paul Duvall 在其 AI 辅助开发模式库中探讨了智能体 AI 如何强化软件工程规范。随着 AI 生成代码量激增，主干开发、频繁提交、自动化测试等实践变得更加关键。Duvall 观察到开发者不再逐行审查 AI 生成的代码，而是依赖自动化验证和智能体护栏。规范驱动开发成为新趋势，通过前置定义行为、约束和验收标准来指导 AI 生成和验证输出。Paul Stack 和 Gergely Orosz 也讨论了类似转变，包括通过 GitHub Issue 工作流替代 PR、以及使用"remixing"模式让智能体按项目标准重建贡献代码。这些实践表明，AI 时代软件工程正从人工检查转向自动化质量保障。
English Summary: Paul Duvall discusses how agentic AI patterns reinforce engineering discipline as AI-generated code volumes surge. Practices like trunk-based development, frequent commits, and automated testing become essential. Duvall notes developers now rely on automated validation and agent guardrails rather than line-by-line code review. Specification-driven development is emerging, defining behavior and constraints upfront to guide AI output.
-
Yupp shuts down after raising $33M from a16z crypto’s Chris Dixon（TechCrunch AI）
中文摘要：众包 AI 模型反馈初创公司 Yupp 在成立不到一年后宣布关闭，尽管此前已获得 a16z crypto 的 Chris Dixon 领投的 3300 万美元种子轮融资，以及来自 45 位知名天使投资人的支持。Yupp 允许用户免费测试和比较 800 多个 AI 模型的输出，并收集反馈数据出售给模型开发商。公司称拥有 130 万用户，每月收集数百万条偏好数据。然而，由于 AI 模型能力在短时间内突飞猛进，以及行业正转向专家反馈和智能体系统，Yupp 未能找到足够的产品市场契合度。部分员工将加入一家知名 AI 公司。
English Summary: Yupp, a crowdsourced AI model feedback startup, is shutting down less than a year after launching, despite raising a $33 million seed round led by a16z crypto's Chris Dixon and investments from over 45 angels including Jeff Dean and Biz Stone. The platform allowed users to test and compare 800+ AI models for free, generating anonymized feedback data for model makers. With 1.3 million users and millions of monthly preferences collected, Yupp failed to achieve product-market fit as AI models rapidly improved and the industry shifted toward expert feedback and agentic systems.
-
Alexa+ gets new food ordering experiences with Uber Eats and Grubhub（TechCrunch AI）
中文摘要：亚马逊为 Alexa+ 推出与 Uber Eats 和 Grubhub 的整合，用户可通过自然对话方式点餐。用户只需关联外卖账户，即可通过语音指令如"我想点意大利菜外卖"来浏览餐厅、询问菜单、定制餐品，并随时修改订单。订单确认前会显示完整的价格和数量摘要。该功能首先向使用 Echo Show 8 及以上设备的 Alexa+ 用户推出。亚马逊表示这是实现自适应交互模型的重要一步，未来还将扩展到杂货购物和旅行预订等领域。这一升级正值快餐行业探索 AI 点餐之际，尽管此前麦当劳和 Taco Bell 的尝试曾遭遇准确性挑战。
English Summary: Amazon launched food ordering integration for Alexa+ with Uber Eats and Grubhub, enabling users to order through natural conversation. After linking delivery accounts, users can browse restaurants, ask menu questions, customize meals, and modify orders via voice commands like 'I want to order Italian for delivery.' A comprehensive summary appears before confirmation. The feature rolls out first to Alexa+ users with Echo Show 8 devices and larger.
-
Nomadic raises $8.4 million to wrangle the data pouring off autonomous vehicles（TechCrunch AI）
中文摘要：NomadicML 完成 840 万美元种子轮融资，投后估值 5000 万美元，由 TQ Ventures 领投。该公司专注于解决自动驾驶汽车和机器人领域的数据管理难题，通过深度学习模型将视频素材转化为结构化、可搜索的数据集。其平台帮助客户从海量存档数据中提取有价值的边缘案例，用于强化学习和模型迭代。Nomadic 采用视觉语言模型实现"智能体推理"，用户描述需求后系统自动寻找相关内容。Zoox、三菱电机等客户已在使用该平台。公司还获得了 Nvidia GTC 创业大赛一等奖。
English Summary: NomadicML raised $8.4 million in seed funding at a $50 million post-money valuation led by TQ Ventures. The startup tackles data management challenges for autonomous vehicles and robots, converting video footage into structured, searchable datasets using deep learning models. Its platform helps customers extract valuable edge cases from archived data for reinforcement learning and model iteration. Using vision language models for 'agentic reasoning,' users describe needs and the system automatically finds relevant content. Customers include Zoox and Mitsubishi Electric, and the company won first prize at Nvidia GTC's pitch contest.
-
Exclusive: Runway launches $10M fund, Builders program to support early-stage AI startups（TechCrunch AI）
中文摘要：AI 视频生成公司 Runway 推出 1000 万美元风投基金和 Builders 计划，支持早期 AI 初创企业。基金将投资种子至 C 轮公司，单笔最高 50 万美元，重点关注三个方向：推动 AI 前沿的技术团队、在基础模型之上构建应用层的开发者、以及探索新媒体创作和叙事的实验性公司。Builders 计划则为符合条件的初创企业提供 50 万 API 积分和 Characters API 访问权限。Characters 是 Runway 的实时视频智能体 API，支持创建具有面部和语音的交互式 AI 角色。Runway 希望借此探索自身无法单独追求的用例，推动"视频智能"生态发展。
English Summary: AI video generation startup Runway launched a $10 million venture fund and Builders program to support early-stage AI startups. The fund will invest up to $500,000 in pre-seed to Series C companies across three areas: technical teams pushing AI frontiers, builders creating application layers on foundation models, and companies experimenting with new media creation. The Builders program offers eligible startups 500,000 API credits and access to Characters, Runway's real-time video agent API for creating interactive AI characters with faces and voices. Runway aims to explore use cases it cannot pursue alone and build a 'video intelligence' ecosystem.
-
Article: Optimization in Automated Driving: from Complexity to Real-Time Engineering（InfoQ AI/ML）
中文摘要：本文深入探讨了自动驾驶汽车的技术架构，重点关注从原始传感器数据到安全控制命令的实时优化技术。作者 Avraam Tolmidis 详细介绍了感知管道的动态资源分配，包括基于场景（高速公路 vs 城市环境）的上下文感知传感器优先级调整，以及使用卡尔曼滤波器的动态传感器权重管理。在轨迹规划方面，文章阐述了模型预测控制（MPC）求解器如何将路径生成框架化为优化问题，通过成本函数平衡跟踪精度与乘客舒适度。文章还讨论了实时计算预算管理、确定性调度策略，以及用于可解释性的数据层设计。这些优化技术是将自动驾驶系统延迟控制在毫秒级的关键。
English Summary: This article explores the technical architecture of autonomous vehicles, focusing on real-time optimization techniques from raw sensor data to safe control commands. Author Avraam Tolmidis details dynamic resource allocation in perception pipelines, including context-aware sensor prioritization based on scenarios (highway vs urban) and dynamic sensor weighting using Kalman filters. For trajectory planning, the article explains how Model Predictive Control (MPC) solvers frame path generation as an optimization problem, balancing tracking accuracy and passenger comfort through cost functions. The piece also covers real-time compute budget management, deterministic scheduling, and data layer design for explainability.
-
Google Unveils AppFunctions to Connect AI Agents and Android Apps（InfoQ AI/ML）
中文摘要：Google 推出 AppFunctions，旨在将 Android 转变为"智能体优先"的操作系统。这一 Jetpack API 允许开发者在应用中暴露自描述功能块，供 AI 智能体或助手调用以完成用户目标。与 MCP 云服务器类似，AppFunctions 在设备本地执行，提供更好的隐私保护和更低的网络延迟。例如，用户可要求 Gemini 助手"显示我在三星相册里的猫照片"，助手将解析请求、检索照片并在自身界面展示。对于尚未集成 AppFunctions 的应用，Google 还推出了 UI 自动化平台作为回退方案，可处理复杂任务如多站拼车或家庭披萨订购。该功能目前处于早期测试阶段，首先在 Galaxy S26 系列上推出，计划随 Android 17 更广泛推广。
English Summary: Google unveiled AppFunctions to transform Android into an 'agent-first' OS. This Jetpack API allows developers to expose self-describing capabilities within apps for AI agents or assistants to fulfill user goals. Similar to MCP cloud servers but executing on-device, AppFunctions offers improved privacy and lower latency. For example, users can ask Gemini to 'show me pictures of my cat from Samsung Gallery,' and the assistant retrieves and displays them. For apps not yet integrated, Google introduced a UI automation fallback that handles complex tasks like multi-stop rideshares. The feature is in early beta on Galaxy S26 devices, with wider rollout planned for Android 17.