日期：2026-04-28
本期聚焦：重点关注模型发布与 release notes、官方 engineering blog、AI coding / agent / SRE、评测榜单变化、开发者实践博客、框架生态、开源模型与真实用户视角；当 HN、Reddit、Hugging Face 等社区源可访问时优先纳入。
-
Artificial Analysis 最新模型排名观察（Artificial Analysis）
中文摘要：Artificial Analysis 最新模型排名显示，GPT-5.5 (xhigh) 以 60 分位居智能指数榜首，GPT-5.5 (high) 以 59 分紧随其后，Claude Opus 4.7 (Max Effort) 与 Gemini 3.1 Pro Preview 并列第三（57 分）。速度方面，Mercury 2 以 687 tokens/秒领先，Granite 3.3 8B 达 333 tokens/秒。延迟最低的是 Ministral 3 3B（0.45 秒）。该平台的 Intelligence Index v4.0 涵盖 GDPval-AA、Terminal-Bench Hard、Humanity's Last Exam、GPQA Diamond 等 10 项评测，为开发者提供模型选型参考。
English Summary: Artificial Analysis' latest rankings show GPT-5.5 (xhigh) leading the Intelligence Index at 60 points, followed by GPT-5.5 (high) at 59, with Claude Opus 4.7 (Max Effort) and Gemini 3.1 Pro Preview tied at 57. For speed, Mercury 2 leads at 687 tokens/s, while Ministral 3 3B has the lowest latency at 0.45s. The Intelligence Index v4.0 covers 10 benchmarks including GDPval-AA, Terminal-Bench Hard, and Humanity's Last Exam.
-
Introducing Claude Opus 4.7（Anthropic News）
中文摘要：Anthropic 发布 Claude Opus 4.7，在多步骤任务效率上创下内部研究智能体基准新高，六项模块总分达 0.715，长上下文表现最为稳定。在 General Finance 模块得分从 Opus 4.6 的 0.767 提升至 0.813，数据披露与纪律性表现最佳。Quantium 评估显示其在推理深度、结构化问题框架和复杂技术工作方面进步显著；Databricks 的 OfficeQA Pro 测试表明文档推理错误减少 21%；Ramp 反馈称其在代理团队工作流中角色忠实度、指令遵循和跨工具调试能力更强，所需逐步指导大幅减少。
English Summary: Anthropic released Claude Opus 4.7, achieving a new high on internal research-agent benchmarks with 0.715 across six modules and the most consistent long-context performance. General Finance scores improved from 0.767 to 0.813. Quantium noted gains in reasoning depth and structured problem-framing; Databricks reported 21% fewer document reasoning errors on OfficeQA Pro; Ramp highlighted stronger role fidelity and reduced need for step-by-step guidance in agent-team workflows.
-
Featured An update on recent Claude Code quality reports（Anthropic Engineering）
中文摘要：Anthropic 工程团队发布 Claude Code 质量报告，复盘了 4 月 16 日 Opus 4.7 发布后出现的问题。团队为降低模型冗长度，在系统提示中加入了字数限制（工具调用间 ≤25 词，最终回复 ≤100 词），导致智能表现意外下降。另一缓存优化误删了历史推理内容，影响代码审查质量。问题已在 4 月 10 日的 v2.1.101 版本中修复。团队现已将 Opus 4.7 默认 effort 模式恢复为 xhigh，并计划为代码审查增加额外仓库上下文支持，以防止类似问题再次发生。
English Summary: Anthropic Engineering published a postmortem on Claude Code quality issues following the April 16 Opus 4.7 launch. A system prompt change adding word limits (≤25 words between tool calls, ≤100 for final responses) unexpectedly reduced intelligence, while a caching optimization dropped prior reasoning from context, degrading code review quality. Both issues were fixed in v2.1.101 on April 10. The team reverted Opus 4.7 defaults to xhigh effort and plans to add multi-repository context for code reviews.
-
Scaling Managed Agents: Decoupling the brain from the hands（Anthropic Engineering）
中文摘要：Anthropic 工程博客发文阐述 Managed Agents 架构理念，提出将"大脑"（Claude 智能）与"双手"（具体工具执行）解耦。该元框架不预设特定工具链，而是提供通用接口支持多种代理框架（如 Claude Code 或领域专用代理）。核心创新在于会话日志作为持久化上下文存储，通过 getEvents() 接口允许模型按位置切片检索事件流，实现灵活的状态恢复、回溯和上下文重读。这种设计让代理能随模型能力提升而进化，无需频繁重写工具链假设。
English Summary: Anthropic Engineering introduced Managed Agents, an architecture decoupling the "brain" (Claude's intelligence) from the "hands" (tool execution). This meta-harness provides general interfaces accommodating various agent frameworks without encoding brittle assumptions. The key innovation is durable session logs serving as persistent context storage outside Claude's window, with a getEvents() interface enabling flexible event stream slicing for state recovery, rewinding, and context re-reading as models improve.
-
Physical AI that Moves the World — Qasar Younis & Peter Ludwig, Applied Intuition（Latent Space）
中文摘要：Latent Space 播客访谈 Applied Intuition CEO Qasar Younis 与 CTO Peter Ludwig，探讨物理 AI 在采矿、无人机、卡车、战舰等极端环境中的应用。该公司从 YC 时期的自动驾驶仿真工具起步，发展为估值 150 亿美元的物理 AI 平台，致力于成为"机器界的 Android"。访谈涵盖：车辆软件栈碎片化的整合愿景、Cursor 与 Claude Code 等 AI 编程工具在嵌入式和安全关键软件中的采用、端到端自动驾驶对仿真验证的新要求，以及神经仿真需足够快和便宜以支撑强化学习训练等话题。
English Summary: Latent Space interviewed Applied Intuition CEO Qasar Younis and CTO Peter Ludwig on physical AI powering mining rigs, drones, trucks, and warships. The $15B company evolved from YC-era autonomy tooling toward becoming "Android for machines," consolidating fragmented vehicle software stacks. Topics include AI coding tools (Cursor, Claude Code) adoption in embedded systems, how end-to-end autonomy changes simulation requirements, and why neural simulation must be fast and cheap enough to make RL practical for physical AI.
-
Automate repetitive tasks with Amazon Quick Flows（AWS ML Blog）
中文摘要：AWS 发布 Amazon Quick Flows，一款面向非技术用户的 AI 工作流自动化工具。用户可通过自然语言描述任务需求，系统自动生成包含数据收集、AI 分析、外部系统集成的完整工作流。文章以财务分析工具和员工入职自动化为例，演示了从文本输入、网络搜索、数据洞察到邮件/Slack 通知的全流程构建。Quick Flows 提供可视化编辑器，支持条件逻辑、循环、变量传递等高级功能，并可连接 SharePoint、S3、HR/IT 系统等外部服务，无需编写代码即可实现复杂的业务自动化。
English Summary: AWS introduces Amazon Quick Flows, an AI-powered workflow automation tool for non-technical users. It allows users to describe tasks in natural language, automatically generating complete workflows including data collection, AI analysis, and external system integrations. The post demonstrates building a financial analysis tool and employee onboarding automation, showcasing features like visual editing, conditional logic, loops, variable passing, and connections to external services like SharePoint, S3, and HR/IT systems—all without requiring code.
-
OpenAI ends Microsoft legal peril over its $50B Amazon deal（TechCrunch AI）
中文摘要：OpenAI 与微软达成修订协议，解决了此前因 OpenAI 与亚马逊高达 500 亿美元合作而引发的法律风险。根据新条款，微软对 OpenAI 知识产权的独家授权转为非独家，有效期至 2032 年；OpenAI 产品将优先在 Azure 发布，但可同时通过任何云服务商向客户提供。微软停止向 OpenAI 支付收入分成，但将继续获得 OpenAI 的收入分成至 2030 年（设有上限）。微软仍持有 OpenAI 约 27% 股份，双方继续保持"主要云合作伙伴"关系。此举消除了微软就 AWS 独家代理 Frontier 工具提起诉讼的可能性，使企业客户能够在多云环境中自由选择。
English Summary: OpenAI and Microsoft reached an amended agreement resolving legal risks from OpenAI's $50 billion deal with Amazon. Under new terms, Microsoft's exclusive license to OpenAI IP becomes non-exclusive through 2032; OpenAI products will launch first on Azure but can now be served across any cloud provider. Microsoft stops paying revenue share to OpenAI but continues receiving revenue share from OpenAI through 2030 (subject to a cap).
-
QCon San Francisco 2026: 12 Tracks Announced（InfoQ AI/ML）
中文摘要：QCon San Francisco 2026 公布 12 个技术专题，会议定于 11 月 16-20 日举行。其中 4 个专题聚焦 AI 生产化实践：智能体架构设计、AI 系统工程、评估与安全护栏、数据平台重构；其余 8 个涵盖分布式系统、架构剖析、弹性工程、平台工程、开发者体验、现代 API 设计、Staff+ 工程师技能以及非工程师角色的代码能力。大会强调实战案例而非产品路线图，由资深从业者组成的委员会筛选议题，旨在帮助高级工程师应对自主智能体故障模式、流量峰值下的 P99 延迟保障、API 架构演进等现实挑战。
English Summary: QCon San Francisco 2026 announced 12 tracks for its November 16-20 conference. Four tracks focus on production AI: Architecting for Agents, Engineering AI Systems, Guardrails & Safety Nets (Evals), and Data Platforms Reimagined. The remaining eight cover Distributed Systems, Architecture Teardowns, Resilience Engineering, Platform Engineering, Developer Experience, Modern API Design, Staff+ Engineering Skills, and Code Beyond Engineers.
-
OpenAI available at FedRAMP Moderate（OpenAI News）
中文摘要：OpenAI 宣布 ChatGPT Enterprise 和 API Platform 获得 FedRAMP 20x Moderate 授权，成为美国联邦机构可合规使用的 AI 服务。FedRAMP 20x 是 GSA 于 2025 年 3 月推出的快速授权路径，采用云原生安全证据、关键安全指标（KSI）和自动化验证。获得授权后，联邦机构可在内部运营和任务支持场景中使用 GPT-5.5 等前沿模型，未来还将通过 FedRAMP 环境访问 Codex Cloud。OpenAI 已在 FedRAMP Marketplace 上架，政府机构可通过 Carahsoft 等授权经销商采购。
English Summary: OpenAI announced FedRAMP 20x Moderate authorization for ChatGPT Enterprise and API Platform, making frontier AI available to U.S. federal agencies. FedRAMP 20x, launched by GSA in March 2025, uses cloud-native security evidence, Key Security Indicators, and automated validation for faster authorization. Federal agencies can now use GPT-5.5 and soon access Codex Cloud through FedRAMP environments for internal and mission-support use cases. OpenAI is listed on the FedRAMP Marketplace and available through authorized resellers like Carahsoft.
-
The next phase of the Microsoft OpenAI partnership（OpenAI News）
中文摘要：OpenAI 与微软宣布修订合作协议，为双方长期合作提供更大确定性。核心条款包括：微软保持 OpenAI 主要云合作伙伴地位，OpenAI 产品优先在 Azure 发布，但可跨任意云服务商交付客户；微软对 OpenAI 知识产权的授权从独家转为非独家，有效期至 2032 年；微软停止向 OpenAI 支付收入分成，但继续获得 OpenAI 的收入分成至 2030 年（设有总额上限）；微软作为大股东继续参与 OpenAI 增长。双方将继续在数据中心扩容、下一代芯片、网络安全等 ambitious 项目上合作。
English Summary: OpenAI and Microsoft announced an amended partnership agreement providing long-term clarity. Key terms: Microsoft remains OpenAI's primary cloud partner with products shipping first on Azure, but OpenAI can now serve customers across any cloud provider; Microsoft's license to OpenAI IP becomes non-exclusive through 2032; Microsoft stops paying revenue share to OpenAI but continues receiving revenue share from OpenAI through 2030 (subject to a cap); Microsoft remains a major shareholder.
-
[AINews] DeepSeek V4 Pro (1.6T-A49B) and Flash (284B-A13B), Base and Instruct — runnable on Huawei Ascend chips（Latent Space）
中文摘要：DeepSeek 正式发布 V4 系列模型，包括 V4 Pro（1.6T 总参数 / 49B 激活）和 V4 Flash（284B 总参数 / 13B 激活），采用 MIT 开源协议。这是自 2024 年 12 月 V3 和 2025 年 1 月 R1 以来的首次重大版本更新。V4 系列支持 100 万 token 超长上下文，通过 CSA（压缩稀疏注意力）和 HCA（重度压缩注意力）技术，KV 缓存相比 V3.2 减少约 10 倍。模型采用 FP4/FP8 混合精度，训练数据量达 32T tokens。独立评测显示 V4 Pro 在开放权重模型中排名第二，仅次于 Kimi K2.6，在 Agentic 任务上表现领先。值得注意的是，DeepSeek 同时适配华为昇腾芯片，计划下半年部署 Ascend 950 超级节点以进一步降低价格。
English Summary: DeepSeek released the V4 family including V4 Pro (1.6T total / 49B active params) and V4 Flash (284B total / 13B active), under MIT license. This marks the first major release since V3 (Dec 2024) and R1 (Jan 2025). V4 supports 1M token context via CSA and HCA attention mechanisms, achieving ~10x KV cache reduction vs V3.2. Trained on 32T tokens with FP4/FP8 mixed precision, V4 Pro ranks #2 among open-weight models per independent benchmarks, behind only Kimi K2.6, with leading agentic performance. DeepSeek also announced Huawei Ascend chip compatibility, with plans to deploy Ascend 950 supernodes in H2 to reduce pricing.
-
Reading today's open-closed performance gap（Interconnects）
中文摘要：文章探讨了开放模型与闭源模型之间的性能差距评估问题，指出单一综合评分掩盖了能力覆盖的复杂性。以 Artificial Analysis Intelligence Index 为例，作者分析了基准测试如何随时间演变、与实际使用场景的相关性变化，以及不同训练范式对评分的影响。当前行业焦点已从简单对话和数学转向复杂代码与 Agentic 任务，而闭源实验室正投入巨资掌握这些领域，并开始向会计、法律、医疗等专业领域扩展。文章指出，开放模型实验室（尤其是中国实验室）在追赶过程中面临数据获取成本和环境构建的挑战，但 RLVR（可验证奖励强化学习）训练方法的普及使它们仍能保持竞争力。作者认为，随着任务难度增加和数据私有化趋势，开放模型能否持续追赶仍存不确定性。
English Summary: The article examines how the open-closed model performance gap is often oversimplified into a single number, masking nuanced capability coverage dynamics. Using the Artificial Analysis Intelligence Index as an example, the author discusses how benchmarks evolve, their correlation with real-world usage, and how training paradigms shift scores. Industry focus has moved from chat and math to complex coding and agentic tasks, with closed labs investing heavily while expanding into specialized domains like law and healthcare. Open labs (especially Chinese ones) face challenges with data acquisition costs and environment building, but RLVR training methods help them remain competitive. The author questions whether open models can sustain this catch-up as tasks grow harder and data becomes more proprietary.
-
Building an emoji list generator with the GitHub Copilot CLI（GitHub AI/ML）
中文摘要：GitHub 团队在 Rubber Duck Thursday 直播活动中展示了如何使用 GitHub Copilot CLI 构建一个表情符号列表生成器。该项目允许用户在终端中粘贴或输入列表，通过 AI 自动为每条项目匹配相关表情符号，并将结果复制到剪贴板。开发过程中使用了 @opentui/core 构建终端 UI、@github/copilot-sdk 提供 AI 能力、clipboardy 处理剪贴板功能。团队采用了 Plan 模式与 Claude Sonnet 4.6 进行规划，然后使用 Claude Opus 4.7 实现功能。文章还介绍了多模型工作流、Autopilot 模式、allow-all 工具标志以及 GitHub MCP 服务器等 Copilot CLI 特性的实际应用。
English Summary: GitHub's Rubber Duck Thursday livestream demonstrated building an emoji list generator using GitHub Copilot CLI. The tool lets users paste or type bullet points in the terminal, automatically matches relevant emojis to each item via AI, and copies the result to clipboard. The build used @opentui/core for terminal UI, @github/copilot-sdk for AI capabilities, and clipboardy for clipboard access. The team employed Plan mode with Claude Sonnet 4.6 for planning, then Claude Opus 4.7 for implementation. The article showcases practical applications of Copilot CLI features including multi-model workflows, Autopilot mode, the allow-all tools flag, and the GitHub MCP server.
-
Build a personal organization command center with GitHub Copilot CLI（GitHub AI/ML）
中文摘要：GitHub 工程师 Brittany Ellich 分享了如何使用 GitHub Copilot CLI 构建个人组织指挥中心，以解决数字信息分散在多个应用中的问题。该项目是一个 Electron 应用，将分散在不同平台的任务、日程和信息统一到一个集中的可视化界面中。开发采用"先规划后实现"的工作流程：使用 Copilot 进行需求访谈和规划生成 plan.md，然后由 Copilot 根据规划实现功能。Brittany 同时使用 VS Code 的 Agent 模式进行同步开发，以及 Copilot Cloud Agent 处理异步任务。她强调，尽管这是她的第一个 Electron 应用，但借助 AI 辅助开发，从想法到可用工具仅用了不到一天时间，同时她也手动简化了代码库以提高可维护性。
English Summary: GitHub engineer Brittany Ellich shared how she built a personal organization command center using GitHub Copilot CLI to solve digital fragmentation across multiple apps. The Electron app unifies tasks, schedules, and information from various platforms into one centralized visual interface. The development followed a "plan-then-implement" workflow: using Copilot to interview her requirements and generate a plan.md, then having Copilot implement based on the plan. Brittany used VS Code Agent mode for synchronous development alongside Copilot Cloud Agent for asynchronous tasks. She noted that despite being her first Electron app, AI-assisted development took her from idea to working tool in under a day, though she manually simplified the codebase for maintainability.
-
Ollama is now powered by MLX on Apple Silicon in preview（Ollama Blog）
中文摘要：Ollama 发布基于 Apple MLX 框架的预览版本，为 Apple Silicon 设备带来显著性能提升。新版本利用 Apple 的统一内存架构，在 M5、M5 Pro 和 M5 Max 芯片上借助 GPU Neural Accelerators 加速首 token 生成时间和解码速度。测试显示，使用 Qwen3.5-35B-A3B 模型时，prefill 性能可达 1851 tokens/s，decode 性能达 134 tokens/s。Ollama 新增对 NVIDIA NVFP4 格式的支持，在降低内存带宽和存储需求的同时保持模型精度，使用户获得与生产环境一致的结果。此外，缓存系统升级包括跨会话复用缓存、智能检查点存储和更智能的淘汰策略，显著提升了编码和 Agentic 任务的响应速度。
English Summary: Ollama released a preview version powered by Apple's MLX framework, delivering significant performance improvements on Apple Silicon devices. The new version leverages Apple's unified memory architecture and GPU Neural Accelerators on M5, M5 Pro, and M5 Max chips to accelerate time-to-first-token and decode speeds. Testing with Qwen3.5-35B-A3B showed prefill performance reaching 1851 tokens/s and decode at 134 tokens/s. Ollama now supports NVIDIA's NVFP4 format, maintaining model accuracy while reducing memory bandwidth and storage requirements for production parity. Cache system upgrades include cross-session cache reuse, intelligent checkpoint storage, and smarter eviction policies, significantly improving responsiveness for coding and agentic tasks.