Next-Level Benchmarking Overfitting – Thomas Neumann
Thomas Neumann 质疑 GenDB-V 相对 Umbra 和 DuckDB 的 22× 至 61× 加速,指出其性能收益依赖针对 TPC-H 的过度特化。
Thomas Neumann 质疑 GenDB-V 相对 Umbra 和 DuckDB 的 22× 至 61× 加速,指出其性能收益依赖针对 TPC-H 的过度特化。
bo2-mac 将玩家自有 Steam 版 Black Ops II 的三个可执行文件静态重编译为 macOS 原生 ARM64 应用,覆盖战役、僵尸和多人模式。
Cogent Security 与 OpenAI 合作构建 Cyber Defense Factories,用 AI 智能体自动完成漏洞调查到验证修复的全流程。
Greg Kurtzer 的 Ginux 已从独立 Rust 内核转为可配置的用户空间内核,作为普通 Linux 进程运行,在受管工作负载与宿主内核之间建立显式、带版本的契约。
OpenAI 工程师在 AI Engineer Code Summit 演示了 Codex 的云端任务、语音控制和个人智能体 dots 工作方式。Codex Cloud 版本 2 支持发布一次环境后,在隔离实例中运行多个任务,但本地任务向云端交接仍涉及上下文差异,团队正在改进体验。语音可用于管理任务线程和委派工作,dots 则可提醒遗漏、承接委派,并利用已有上下文推进 PR。
Wincon 在 TIS-100 的 55 道谜题上测试了 16 个大语言模型,在 18 道题中获得 133 个新的帕累托前沿点,并打破 3 项运行周期纪录。Opus 5.5 与 GPT-6 Astra 均解出全部 55 道题,各在 11 道题上找到新前沿点,两者的纪录接近度差异未达到统计显著水平。计入结果的程序均经过真实游戏重放及 100,000 次测试,社区玩家随后进一步优化了其中多数程序。
Gilbert + Tobin 引入 ChatGPT Enterprise、ChatGPT Work 和 Codex,将 AI 应用从个人辅助推进到多步骤智能体工作流,并由人员负责审核和批准。
安全公司 Claroc 的研究发现,SpaceXSI 的 Grok Build 编码智能体泄露数千名用户的数据,包括商业机密和其他敏感资料。平台虽将分享 URL 随机化,但一项设计使任何人都能枚举数千个在线项目,除少数情况外,打开链接即可查看应用及其关联数据。
OpenAI 收购了 Sonder 创始人 Francis Davidson 创办的 AI 旅行智能体 Odessia,交易条款未披露,看起来是一次 acqui-hire,Davidson 已加入 OpenAI 负责旅行业务方向。
Hacker News 首页文章《Life on Claude Nine》讲述软件工程师 Ivan 用 Claude 搭建邮件分类、日历管理、文档起草等自动化系统,随后把写代码、跑测试、提交仓库也交给 Claude,产出翻了三到四倍并获得晋升。他进一步构建管理 Claude 上下文、分析其错误并生成更优提示词的脚手架系统,同时逐渐失眠、疏远女友,感到自己在建造无法完全理解的东西。
Resend 联合创始人兼 CEO Zeno Rocha 在访谈中介绍了面向 AI 智能体的产品设计,涉及服务推荐、定价信息读取和 MCP 工具响应。文中引用的一项研究显示,Claude Code 在可提取的 102 次邮件服务推荐中有 64 次将 Resend 列为首选;另一项更近期的研究中,Postmark 在多数模型中排名居前。
OpenAI launched a test of Businesses in ChatGPT. Companies can show a merchant profile inside ChatGPT and use Business Hub to manage information, product prices and inventory, plus data on impressions and website clicks.
Why it matters: OpenAI is putting merchant profiles and customer-service agents inside ChatGPT, giving businesses a storefront in the chat interface beyond search and maps.
Modelplane 的推理工程指南面向平台工程师,系统讲解模型与 GPU 的配合、生产部署及职业学习路径。指南涵盖权重与 KV cache 显存预算、prefill 与 decode 瓶颈、引擎优化、网关鉴权、可观测性和扩缩容,并强调用真实流量开展并发评测。
Google added two features to Gemini Intelligence: creating widgets from natural language and device assistance. The update also brings the assistant to more devices, including the motorola signature 27, Pixel 10 Pro XL and Pixel 10 Pro Fold.
Why it matters: The new features cut manual steps for organizing information and changing settings on a phone.
OpenAI 周二推出处于 beta 阶段的 Business Hub,让企业在 ChatGPT 中创建资料并核验、更新产品与服务信息。符合条件的企业还能创建专属 AI 智能体,回答用户有关企业及其产品的问题。Business Hub 提供有限的业务洞察,包括资料浏览量、企业提及次数和网站点击量;企业资料须经过正式核验流程。
Google has opened Gemini 4 Carbon after two weeks of testing by safety experts. It sits alongside Gemini 4 Argon and targets complex multi-step coding, debugging and open-ended scientific research. Carbon carries the same per-token price as Argon, so deeper reasoning and longer verification loops raise usage.
Why it matters: Carbon and Argon share a token price, so deeper reasoning and longer verification loops can push the real total cost of complex tasks higher.
Modern Web Guidance is a skill and CLI tool that puts modern web practices and browser compatibility data into a coding agent's context. It covers 250 modern web features and 203 real development use cases, with local semantic search, guide lookup and scans for old patterns. Looking up guides for experimental or unstable APIs requires the developer's explicit consent.
Why it matters: It retrieves modern browser guidance per task and injects it into the agent's context, so generated code carries compatibility data and lightweight fallbacks.
OpenAI 宣布在 ChatGPT 中以 beta 形式推出 Business Agents,让用户与代表品牌的认证 AI 智能体交流,初期面向美国符合条件的电商和在线服务企业。
Anthropic released Claude Meetings, Claude Whiteboards and Claude Sheets, letting teams generate plans live during meetings and collaborate on whiteboards and spreadsheets. The tools aim to cut the separate step of sorting notes into plans after a discussion ends.
Why it matters: Anthropic is putting Claude inside meetings and linking it to collaborative whiteboards, so teams can produce plans and timelines while they talk.
Google DeepMind has released Gemini 4 Carbon to developers, enterprises, scientists and consumers. It targets complex knowledge work, long-horizon agent workflows, cybersecurity and scientific reasoning. Google prices Carbon at the same entry rate as Argon, but Carbon spends more tokens on deeper exploration and verification.
Why it matters: Google positions Carbon for complex multi-step tasks, charging the same entry price as Argon while spending more tokens to explore and verify.
Google has released Gemini 4 Carbon to developers, enterprises, scientists and consumers, targeting complex knowledge work, agentic coding, cybersecurity and scientific reasoning. Carbon ranks first on CursorBench 4.0, WANDR and AutomationBench, and Google says it costs more than 2x less per task than the next-best external model.
Why it matters: Google sets Carbon at the same entry price as Argon, with Carbon spending more tokens on reasoning and verification for complex multi-step tasks.
Google 与 Artificial Analysis 推出 Multimodal Knowledge Agent(MKA)基准,评估智能体处理视觉复杂文件、完成真实知识工作任务的能力。
Google released Gemini 4 Carbon to paid developers, enterprises and Google AI Ultra subscribers, aimed at complex multi-step coding and open-ended scientific research. Google also previewed a swarm feature in Antigravity that turns a single prompt into a team of agents working together.
Why it matters: Google set Gemini 4 Carbon's token pricing below frontier competitors, though total task cost still depends on how many tokens the model consumes.
eBay疑似已阻止Meta Muse等第三方AI智能体浏览平台,The Verge实测复现了无法获取商品链接和详情的情况。eBay此前已禁止未经人工审核下单的智能体工具,其发言人承认在未经授权抓取和数据训练等情形下限制智能体访问,但未直接回答近期是否新增限制。部分测试中,Muse转而引导用户访问镜像展示eBay商品的外部网站。
Microsoft 发布 .NET 11,带来 C# 15、运行时优化及 AI 智能体开发工具,作为 STS 版本提供两年支持。SDK 内置 MCP 服务器模板,Microsoft.Extensions.AI 提供模型访问抽象,Microsoft Agent Framework 支持智能体和工作流,Blazor AI 组件仍处于实验性预发布阶段。
Meta 周一宣布将 AI 智能体 Muse 扩展至 Apple Watch 和 Android Wear OS,用户可在智能手表上通过语音通话、短信或听写直接与 Muse 交互。
Google 在 Adobe MAX 上宣布扩大与 Adobe 的合作,把 Adobe 的创意与生产力工具直接接入 Google 日常应用,同时将 Google 前沿 AI 模型引入 Adobe 应用。
Cyanogen 联合创始人 Kirt McMaster 的新公司 Kora Systems 发布了一款 Linux 桌面操作系统,将 AI 模型与自主智能体直接集成进系统,可跨应用工作、访问文件并在任务间保留上下文。
Adobe 在 MAX 活动上宣布为 Photoshop、Premiere 和 Firefly 等产品增加 AI 创作功能,并发布本地 MCP,向外部聊天机器人开放更多 Creative Cloud 功能。
Anthropic 的 Fable 5.2 在 RLI 真实自由职业任务测试中完成率达到 41.3%,一年前领先系统 Manus 仅为 2.5%。排名其后的 Opus 5.5 和 GPT 6 Astra 分别达到 21.25% 和 20.83%,Fable 5.2 的主要进步集中在音频与音乐编辑、建筑与室内设计及 3D 模型制作。
Adobe 在 Max 大会上宣布多项智能体功能,包括 Creative Cloud 本地 MCP 接入预览、后台持续运行的智能体和 Firefly AI Assistant 升级。
Google DeepMind published AlphaProof Code, a multi-LLM agent framework that formally verified key security properties of Signal's core cryptographic code in four weeks. The work moves formal verification from the protocol design to the actual implementation.
Why it matters: DeepMind pushes Signal's formal verification from protocol design down to real code, covering the design-to-implementation gap where bugs tend to hide.
AI 生成代码可能削弱大型 SaaS 代码库的长期可靠性,文章将风险归结为浅层测试、过度复杂的代码、开发者理解下降和审查时间不足。作者称,在 Meta 任职的最后几个月至少 90% 的代码由 AI 生成,但产出没有达到 2X;目前在另一家大型科技公司使用 AI 生成 100% 的代码。
ClickHouse 收购 Xata,计划将即时数据库分支、闲置计算缩容至零及智能体服务层整合进 ClickHouse Managed Postgres。Xata 表示,即使数据库包含数 TB 数据,创建独立分支也只需约 3 秒,可供每个拉取请求或智能体运行隔离测试。Xata 客户短期内可照常使用,平台继续接受注册,待功能达到一致后再迁移至 ClickHouse Managed Postgres。
Meta AI 推出 Muse Ember,定位为旗下最小、最高效的模型,用于复杂推理、工具调用、编码子任务、路由、提取和长文档解析。
Anthropic started its IPO roadshow on Monday and released Claude Fable 5.2 and Claude Mythos 5.2. Fable 5.2 ranks first on the Artificial Analysis Intelligence Index v5 with 65 points.
Why it matters: Fable 5.2 cuts the price per token by 20%, yet GDPx cost per task rose from $67.16 to $88.15.
The CPython formalization project used Opus 5.5 to rewrite CPython 3.15.2 functions line by line in Lean. So far 4,071 functions have passed tests, formal verification and independent agent review, 15.7% of the functions to be checked.
Why it matters: The project pairs line-by-line modeling with differential testing and independent review, using real runs to check whether formal conclusions miss exception and memory behavior.
Agent 技能依赖维护需要同步锁文件与实际技能文件,文章比较了 gh skill 和 APM + Dependabot 的方案与限制。Dependabot 调用 apm lock 更新账本,不同步已入库的技能文件且跳过安全扫描,对多行依赖配置、MCP 服务和私有源也存在支持限制。
Simon Willison 用 Claude Code 构建 Beep Locator,通过多台设备监听和三角定位,将家中蜂鸣声的位置缩小到约 40cm 范围,找到了需要更换电池的旧一氧化碳探测器。
lexandstuff tested Claude Code with Fable 5.2, starting from a nephew's design sketches and existing game code. One run produced a playable 12-level side-scrolling action game. The model rewrote most of the code, generated about 175 images and 140 audio clips, and wrote a bot that played all twelve levels in a headless browser and checked screenshots.
Why it matters: The model wrote a bot that played all twelve levels in a headless browser and checked screenshots, using automated play and visual checks as basic tests for a generated game.