日期：2026-04-30
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 的模型排行榜显示，GPT-5.5 (xhigh) 以 60 分的智能指数位居榜首，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，均获得 57 分。速度方面，Mercury 2 以每秒 778.1 个 token 遥遥领先，Granite 4.0 H Small 以 400.1 t/s 位列第二。成本最低的是 Qwen3.5 0.8B，每百万 token 仅需 0.02 美元。该榜单共评估了 367 个模型，涵盖智能、速度、延迟、价格和上下文窗口等多维度指标。
English Summary: Artificial Analysis' model ranking shows GPT-5.5 (xhigh) leading with an Intelligence Index score of 60, followed by GPT-5.5 (high) at 59. Claude Opus 4.7 (Max Effort) ties with Gemini 3.1 Pro Preview at 57 for third place. Mercury 2 dominates speed at 778.1 tokens per second, while Qwen3.5 0.8B is the most affordable at $0.02 per million tokens. The ranking evaluates 367 models across intelligence, speed, latency, price, and context window metrics.
-
Introducing Claude Opus 4.7（Anthropic News）
中文摘要：Anthropic 正式发布 Claude Opus 4.7，这是其旗舰推理模型的重大升级。在 Rakuten-SWE-Bench 基准测试中，Opus 4.7 解决生产任务的能力是 4.6 版本的 3 倍，代码质量和测试质量均有两位数提升。早期合作伙伴反馈积极：Hebbia 报告称工具调用和规划准确性实现两位数增长；Bolt 发现长时应用构建任务性能提升高达 10%；某金融 AI 团队指出其在多步骤研究任务中表现最强，在通用金融模块得分从 0.767 提升至 0.813，且在演绎逻辑方面显著改善。
English Summary: Anthropic officially launched Claude Opus 4.7, a major upgrade to its flagship reasoning model. On Rakuten-SWE-Bench, Opus 4.7 resolves 3x more production tasks than Opus 4.6 with double-digit gains in code and test quality. Early partners report strong results: Hebbia saw double-digit accuracy improvements in tool calls and planning; Bolt observed up to 10% better performance on long-running app-building tasks; a financial AI team noted it achieved the strongest baseline for multi-step work with scores improving from 0.767 to 0.
-
Featured An update on recent Claude Code quality reports（Anthropic Engineering）
中文摘要：Anthropic 工程团队发布关于 Claude Code 质量问题的复盘报告。4 月初，团队为降低 Opus 4.7 的冗长输出，在系统提示中加入了字数限制（工具调用间文本不超过 25 词，最终回复不超过 100 词），导致模型智能显著下降。该变更于 4 月 16 日与 Opus 4.7 同步上线，引发用户关于代码审查质量下滑的反馈。团队于 4 月 7 日将默认推理强度恢复为 xhigh，并在 4 月 10 日的 v2.1.101 版本中修复了缓存优化问题——该问题曾导致模型丢失先前的推理链条。此外，Opus 4.7 在完整代码库上下文测试中成功发现了 Opus 4.6 遗漏的 bug。
English Summary: Anthropic's engineering team published a postmortem on Claude Code quality issues. In early April, a system prompt change adding word limits (≤25 words between tool calls, ≤100 words for final responses) to reduce verbosity significantly degraded model intelligence. Shipped alongside Opus 4.7 on April 16, this caused user complaints about declining code review quality. The team reverted default effort to xhigh on April 7 and fixed a caching optimization bug on April 10 in v2.1.101 that was dropping prior reasoning from conversation history. Notably, Opus 4.7 successfully found bugs that Opus 4.6 missed when given complete repository context.
-
Scaling Managed Agents: Decoupling the brain from the hands（Anthropic Engineering）
中文摘要：Anthropic 工程博客发布关于托管智能体（Managed Agents）的架构设计文章，提出"将大脑与双手解耦"的设计理念。文章指出，随着模型能力提升，agent harness 中关于模型局限性的假设会快速过时，因此需要频繁质疑和更新。关键技术包括：分离可恢复的上下文存储与 harness 中的上下文管理，支持事件获取和转换以实现高提示缓存命中率；通过多智能体协作实现"多大脑、多双手"的并行处理；以及采用上下文压缩、记忆工具和上下文修剪等技术处理超长任务。架构强调 harness 负责上下文工程，而会话层保证持久性和可查询性。
English Summary: Anthropic's engineering blog published an architecture design article on Managed Agents, proposing the concept of "decoupling the brain from the hands." The post notes that assumptions about model limitations in agent harnesses become stale as models improve and need frequent reevaluation. Key techniques include: separating recoverable context storage from harness-level context management with support for event fetching and transformation to achieve high prompt cache hit rates; enabling "many brains, many hands" through multi-agent collaboration; and handling long-horizon tasks via context compaction, memory tools, and context trimming. The architecture emphasizes that harnesses handle context engineering while the session layer guarantees durability and interrogability.
-
Microsoft says it has over 20M paid Copilot users, and they really are using it（TechCrunch AI）
中文摘要：微软宣布 Microsoft 365 Copilot 付费用户已突破 2000 万，并强调用户活跃度真实且持续增长。尽管外界普遍认为 Copilot 使用率不高，但微软在财报电话会议上展示了强劲的数据，摩根士丹利分析师 Keith Weiss 称这些数字"远超大多数人预期"。Copilot 现已支持多模型访问，包括 Anthropic 的 Claude，并具备智能自动路由和模型协同能力。Agent 模式成为增长驱动力，目前已作为 Word、Excel 和 PowerPoint 的默认体验上线。微软强调 Copilot 不依赖单一模型，用户可在聊天中默认使用多个模型。
English Summary: Microsoft announced that Microsoft 365 Copilot has surpassed 20 million paid users, emphasizing that engagement is genuine and growing. Despite perceptions that Copilot is underutilized, Microsoft showcased strong numbers during its earnings call, with Morgan Stanley analyst Keith Weiss calling the figures "way ahead of most people's expectations." Copilot now supports multi-model access including Anthropic's Claude, with intelligent auto-routing and model collaboration capabilities. Agent mode is driving adoption and is now the default experience across Word, Excel, and PowerPoint. Microsoft emphasized that Copilot is not dependent on any single model, with users having access to multiple models by default in chat.
-
Extracting contract insights with PwC’s AI-driven annotation on AWS（AWS ML Blog）
中文摘要：PwC 与 AWS 合作开发的 AI 驱动标注解决方案 AIDA，利用 Amazon Bedrock 大语言模型帮助企业从合同中提取结构化洞察。该系统结合 OCR、用户自定义提取规则和检索增强生成（RAG）技术，支持模板化批量提取、单文档对话问答和跨文档全局搜索。AIDA 采用 AWS 云原生架构，包括 Amazon ECS、S3、RDS、OpenSearch Serverless 等服务，并集成 Amazon Bedrock Guardrails 进行内容过滤和敏感信息保护。在实际客户部署中，AIDA 可将合同审阅时间缩短高达 90%，帮助法律、合规和采购团队更高效地获取关键信息。
English Summary: PwC's AI-driven annotation solution AIDA, built on AWS, leverages Amazon Bedrock LLMs to extract structured insights from contracts. The system combines OCR, user-defined extraction rules, and Retrieval Augmented Generation (RAG) to support template-based batch extraction, document-level chat Q&A, and global search across documents. AIDA uses AWS cloud-native architecture including Amazon ECS, S3, RDS, and OpenSearch Serverless, with Amazon Bedrock Guardrails for content filtering and PII protection. In customer implementations, AIDA has reduced contract review time by up to 90%, helping legal, compliance, and procurement teams access key information more efficiently.
-
Building the compute infrastructure for the Intelligence Age（OpenAI News）
中文摘要：OpenAI 宣布扩展 Stargate 项目以构建支撑通用人工智能（AGI）的计算基础设施。自 2025 年 1 月承诺到 2029 年在美国部署 10GW AI 基础设施以来，OpenAI 已提前超越该目标，仅过去 90 天就新增超过 3GW 容量。OpenAI 强调计算能力是先进 AI 的关键输入，更多算力可实现更好的模型训练、更可靠的服务和更低的成本。最新模型 GPT-5.5 已在得克萨斯州 Abilene 的旗舰 Stargate 站点完成训练，该站点采用 Oracle Cloud Infrastructure 和 NVIDIA GB200 系统。OpenAI 表示将与合作伙伴、本地社区和更广泛的基础设施生态系统合作，以满足不断增长的 AI 需求。
English Summary: OpenAI announced the expansion of its Stargate project to build the compute infrastructure powering AGI. Since committing to 10GW of AI infrastructure in the US by January 2025, OpenAI has already surpassed that milestone, adding over 3GW in the last 90 days alone. The company emphasizes that compute is the critical input for advanced AI, enabling better model training, more reliable service, and lower costs over time. Its latest model GPT-5.5 was trained at the flagship Stargate site in Abilene, Texas, which operates on Oracle Cloud Infrastructure and runs NVIDIA GB200 systems. OpenAI stated it will work with partners, local communities, and the broader infrastructure ecosystem to meet growing AI demand.
-
Presentation: Agents, Architecture, & Amnesia: Becoming AI-Native Without Losing Our Minds（InfoQ AI/ML）
中文摘要：InfoQ 发布了一场关于 AI 自主性与架构演化的演讲，演讲者 Tracy Bannon 借用《魔法师的学徒》寓言警示无节制 AI 自主性的风险。她探讨了从机器人（bots）向自主智能体（autonomous agents）转变的过程，指出过快的演进速度可能导致"架构遗忘"（Architectural Amnesia）——即系统决策逻辑和演化历史逐渐丢失的现象。演讲强调企业在拥抱 AI 原生转型时需要保持清醒的架构意识，避免因追求自动化速度而牺牲系统的可理解性和可维护性。
English Summary: InfoQ published a presentation on AI autonomy and architectural evolution, where speaker Tracy Bannon uses the cautionary tale of "The Sorcerer's Apprentice" to illustrate the risks of unbridled AI autonomy. She discusses the shift from bots to autonomous agents, explaining how reckless speed can lead to "Architectural Amnesia"—the gradual loss of system decision logic and evolutionary history.
-
Cybersecurity in the Intelligence Age（OpenAI News）
中文摘要：OpenAI 发布了一份关于智能时代网络安全的行动计划，提出五大支柱：普及 AI 驱动的网络防御能力、加强政府与行业协调、强化前沿网络安全能力的安全防护、在部署中保持可见性和控制权、以及赋能用户自我保护。OpenAI 指出，人工智能正在重塑网络安全格局，同样的能力既可以帮助防御者识别漏洞、自动修复和更快响应，也可能被恶意行为者用于扩大攻击规模、降低攻击门槛并提高攻击复杂度。该计划旨在通过与联邦和州政府及主要商业实体的专家对话形成，以支持民主机构和流程，同时扩大可信行为者获取防御技术的渠道。
English Summary: OpenAI published an action plan for cybersecurity in the Intelligence Age, outlining five pillars: democratizing AI-powered cyber defense, coordinating across government and industry, strengthening security around frontier cyber capabilities, preserving visibility and control in deployment, and enabling users to protect themselves. OpenAI notes that AI is reshaping cybersecurity—the same capabilities that help defenders identify vulnerabilities, automate remediation, and respond faster are also being used by malicious actors to scale attacks, lower barriers to entry, and increase sophistication. The plan was informed by conversations with cybersecurity and national security experts across federal and state government and major commercial entities, aiming to support democratic institutions while broadening access to defensive technologies for trusted actors.
-
[AINews] not much happened today（Latent Space）
中文摘要：Latent Space 的 AINews 栏目承认当天 AI 领域新闻相对平淡。不过文章仍梳理了值得关注的技术动态：vLLM 0.20 发布，带来 TurboQuant 2-bit KV 缓存、DeepSeek V4 MegaMoE 支持等内存与推理优化；Poolside 发布首个开源模型 Laguna XS.2（33B MoE 编程模型，Apache 2.0 许可）；NVIDIA 推出 Nemotron 3 Nano Omni（300 亿参数多模态 MoE，支持 256K 上下文）；Mistral 推出 Workflows 预览版，聚焦企业级智能体编排；以及 GPT-5.5 在 Epoch Capabilities Index 上达到 159 分的新高等。
English Summary: Latent Space's AINews column acknowledged a relatively quiet day in AI news. However, the article still highlighted notable technical developments: vLLM 0.20 release with TurboQuant 2-bit KV cache and DeepSeek V4 MegaMoE support for memory and inference optimization; Poolside's first open-source model Laguna XS.2 (33B MoE coding model under Apache 2.0); NVIDIA's Nemotron 3 Nano Omni (30B parameter multimodal MoE with 256K context); Mistral's Workflows preview focusing on enterprise agent orchestration; and GPT-5.
-
[AINews] ImageGen is on the Path to AGI（Latent Space）
中文摘要：本文探讨了GPT-Image-2在图像生成领域的持续爆发，指出尽管各大实验室都在竞相模仿Anthropic专注于企业AI和编码工具，但GPT-Image-2仍在推动更多创意应用。文章讨论了图像生成模型是否值得投入稀缺GPU资源的问题，认为多模态能力（包括语音和视觉生成）是实现AGI的关键组成部分。此外还涵盖了OpenAI与微软合作条款的更新、GPT-5.5的基准测试表现、GitHub Copilot转向按使用量计费、小米开源MiMo-V2.5系列模型、Sakana的Conductor多智能体编排系统，以及Google TPU v8架构拆分等重要行业动态。
English Summary: This article explores the continued explosion of GPT-Image-2 in image generation, noting that while labs race to emulate Anthropic's enterprise AI focus, GPT-Image-2 drives creative applications. It argues multimodal capabilities (voice and visual generation) are essential for AGI. The piece also covers OpenAI's updated Microsoft partnership terms, GPT-5.5 benchmark performance, GitHub Copilot's shift to usage-based billing, Xiaomi's open-source MiMo-V2.5 models, Sakana's Conductor multi-agent orchestration system, and Google's TPU v8 architecture split.
-
Reading today's open-closed performance gap（Interconnects）
中文摘要：本文深入分析了开源与闭源模型之间的性能差距，指出将这一差距简化为单一数字会掩盖模型能力覆盖领域的细微差别。作者Nathan Lambert强调，随着任务从简单聊天、数学和基础代码转向更复杂的编码和智能体任务，评估基准每12-18个月就会发生变化。闭源前沿实验室投入巨额资金掌握当前重点领域，同时开始向需要专业知识和领域特定工具的新知识工作任务拓展。文章还探讨了中国开源模型实验室如何通过购买折扣后的环境和数据集来追赶，以及前沿实验室需要不断创新以维持收入增长的商业压力。
English Summary: This article analyzes the open-closed model performance gap, arguing that reducing it to a single number obscures nuanced capability coverage. Author Nathan Lambert emphasizes benchmarks shift every 12-18 months as tasks evolve from simple chat and math to complex coding and agentic work. Closed frontier labs invest heavily in current foci while pushing into specialized knowledge work requiring domain expertise.
-
Building an emoji list generator with the GitHub Copilot CLI（GitHub AI/ML）
中文摘要：GitHub工程师Cassidy Williams分享了在Rubber Duck Thursday直播中如何使用GitHub Copilot CLI构建一个表情符号列表生成器的实战经验。该项目允许用户在终端中粘贴或输入项目列表，通过AI自动为每个项目匹配相关表情符号，并将结果复制到剪贴板。开发过程使用了@opentui/core构建终端UI、@github/copilot-sdk提供AI能力、以及clipboardy处理剪贴板访问。文章展示了Copilot CLI的计划模式、自动驾驶模式、多模型工作流、allow-all工具标志以及GitHub MCP服务器的综合运用，为开发者提供了使用Copilot CLI和SDK构建项目的实用参考。
English Summary: GitHub engineer Cassidy Williams shares a hands-on guide to building an emoji list generator using GitHub Copilot CLI during the Rubber Duck Thursday livestream. The tool lets users paste or type bullet points in the terminal, automatically matches relevant emojis via AI, and copies results to clipboard. The project used @opentui/core for terminal UI, @github/copilot-sdk for AI capabilities, and clipboardy for clipboard access.
-
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的真实案例。作为计费团队的资深软件工程师，Brittany为解决数字碎片化问题，将分散在十多个应用中的信息统一到一个平静的中央空间中。她采用"先规划后实施"的工作流程，利用AI进行规划、Copilot进行实现，仅用一天时间就完成了v1版本。她通常使用VS Code的agent模式进行同步开发，同时用Copilot Cloud Agent处理异步任务。Brittany的经验表明，借助AI工具从零开始构建解决方案从未如此简单，这对学习如何与新的AI工具协作非常有帮助。
English Summary: GitHub engineer Brittany Ellich shares a real-world case study of building a personal organization command center using GitHub Copilot CLI. As a staff software engineer on the billing team, Brittany unified information scattered across a dozen apps into one calm central space to solve digital fragmentation. Using a plan-then-implement workflow with AI for planning and Copilot for implementation, she completed v1 in a single day. She typically uses VS Code's agent mode for synchronous development while running Copilot Cloud Agent for asynchronous tasks. Her experience demonstrates that building solutions from scratch has never been easier with AI tools.
-
Ollama is now powered by MLX on Apple Silicon in preview（Ollama Blog）
中文摘要：Ollama宣布在Apple Silicon上推出基于MLX框架的预览版本，这是目前在苹果芯片上运行Ollama的最快方式。新版本利用Apple的统一内存架构，在所有Apple Silicon设备上实现显著加速，在M5、M5 Pro和M5 Max芯片上更是利用新的GPU神经加速器来提升首token时间和生成速度。此外，Ollama新增对NVIDIA NVFP4格式的支持，在保持模型精度的同时减少内存带宽和存储需求。缓存系统也得到升级，包括跨对话重用缓存、智能检查点存储和更智能的淘汰策略，使编码和智能体任务更加高效。
English Summary: Ollama announced a preview version powered by Apple's MLX framework on Apple Silicon, offering the fastest way to run Ollama on Macs. The new version leverages Apple's unified memory architecture for significant speedups across all Apple Silicon devices, with new GPU Neural Accelerators on M5, M5 Pro, and M5 Max chips improving time-to-first-token and generation speed. Ollama also adds support for NVIDIA's NVFP4 format to maintain model accuracy while reducing memory bandwidth and storage. The cache system is upgraded with cross-conversation reuse, intelligent checkpoint storage, and smarter eviction policies for more efficient coding and agentic tasks.