{"pageContext":{"index":324,"total":362,"name":"2025-07-19","items":[{"node":{"frontmatter":{"title":"80 年代中期的 MacPaint 艺术作品今天仍然看起来很棒","publish":true,"cover":"https://blog.decryption.net.au/images/macpaint/lesson3d.png","showCover":null,"date":"2025-07-19","category":"mobile","url":"https://blog.decryption.net.au/posts/macpaint.html","author":"Anthony","translator":"","tags":["mac"],"priority":2},"html":"<p>这些在9英寸、1位屏幕上创作的作品令人惊叹，他希望能找到这些艺术家，了解他们现在的成就。</p>"}},{"node":{"frontmatter":{"title":"Bitchat – 通过蓝牙 Mesh 网络去中心化消息应用(无需互联网、服务器或电话号码)","publish":true,"cover":"","showCover":null,"date":"2025-07-19","category":"mobile","url":"https://github.com/jackjackbits/bitchat","author":"permissionlesstech","translator":"","tags":["other"],"priority":1},"html":"<p>支持iOS和macOS</p>"}},{"node":{"frontmatter":{"title":"Grok回答会参考马斯克的观点","publish":true,"cover":"","showCover":null,"date":"2025-07-19","category":"AI","url":"https://simonwillison.net/2025/Jul/11/grok-musk/","author":"Simon Willison","translator":"","tags":["other"],"priority":1},"html":"<p>作者认为，这并非是 Grok 的系统提示指示其这样做，而是 Grok 因为 “知道” 自己是 xAI 构建的，而 Elon Musk 拥有 xAI，所以会去参考 Elon Musk 的观点。文章还指出，改变提问方式，例如使用 “who should one” 而不是 “who do you”，可能会导致 Grok 给出不同的答案，甚至忽略“一个词”的限制，并提供更详细的分析和对比。作者推测这种行为可能是 Grok 无意中产生的。</p>"}},{"node":{"frontmatter":{"title":"Mercury：基于扩散的超快速LLM模型","publish":true,"cover":"AI","showCover":null,"date":"2025-07-19","category":"AI","url":"https://arxiv.org/abs/2506.17298","author":"Inception Labs, Samar Khanna, Siddhant Kharbanda, Shufan Li, Harshit Varma, Eric Wang, Sawyer Birnbaum, Ziyang Luo, Yanis Miraoui, Akash Palrecha, Stefano Ermon, Aditya Grover, Volodymyr Kuleshov","translator":"","tags":["other"],"priority":1},"html":"<p>介绍了 Mercury Coder，这是 Mercury 系列中第一个专为代码应用设计的模型，目前有两个尺寸：Mini 和 Small。在 NVIDIA H100 GPU 上，Mercury Coder Mini 和 Small 分别实现了 1109 tokens/sec 和 737 tokens/sec 的 state-of-the-art 吞吐量，其速度比其他速度优化的模型平均快 10 倍，同时保持了相当的质量。此外，在多个代码基准测试以及 Copilot Arena 的真实世界验证中，Mercury Coder 表现出色，在质量方面排名第二，并且是速度最快的模型。作者还发布了一个公共 API 和免费的 playground。</p>"}},{"node":{"frontmatter":{"title":"NVM、FNM、Volta 之间有巨大的性能差异（理论上可达 500×）","publish":true,"cover":"","showCover":null,"date":"2025-07-19","category":"nodejs","url":"https://nodejs.org/en/blog/release/v24.4.0","author":"Pavel Romanov","translator":"","tags":["cli"],"priority":1},"html":"<p>NVM 往往最慢，而工具如 FNM 和 Volta 要快得多</p>"}},{"node":{"frontmatter":{"title":"OpenAI 的 Windsurf 交易告吹，Windsurf 的 CEO 将前往 Google","publish":true,"cover":"","showCover":null,"date":"2025-07-19","category":"AI","url":"https://www.theverge.com/openai/705999/google-windsurf-ceo-openai","author":"Hayden Field","translator":"","tags":["other"],"priority":1},"html":"<p>OpenAI原本计划收购AI编码初创公司Windsurf的交易告吹。取而代之的是，Windsurf的CEO Varun Mohan以及联合创始人Douglas Chen和部分研发人员将加入Google DeepMind团队，专注于agentic coding，主要在Gemini项目上工作。Google不会持有Windsurf的股份，但会获得Windsurf部分技术的非独家授权。Windsurf的业务主管Jeff Wang将担任临时CEO，全球销售副总裁Graham Moreno将担任总裁。Google表示欢迎Windsurf的AI编码人才加入，以推进agentic coding的工作。Mohan和Chen也对加入Google DeepMind感到兴奋。OpenAI之前被报道计划以30亿美元收购Windsurf。</p>"}},{"node":{"frontmatter":{"title":"Node24.4.0发布 — 支持 --watch-kill-signal 来指定信号发送到watch mode","publish":true,"cover":"","showCover":null,"date":"2025-07-19","category":"nodejs","url":"https://nodejs.org/en/blog/release/v24.4.0","author":"Rafael Gonzaga","translator":"","tags":["v24"],"priority":1},"html":""}},{"node":{"frontmatter":{"title":"Postgres LISTEN NOTIFY 无法扩展","publish":true,"cover":"","showCover":null,"date":"2025-07-19","category":"database","url":"https://www.recall.ai/blog/postgres-listen-notify-does-not-scale","author":"Elliot Levin","translator":"","tags":["postgres"],"priority":1},"html":"<p>通过模拟负载测试，分别在启用和禁用LISTEN/NOTIFY的情况下进行对比，证实了全局数据库锁是性能瓶颈的罪魁祸首。后来放弃使用LISTEN/NOTIFY，转而在应用层实现相同的逻辑。</p>"}},{"node":{"frontmatter":{"title":"WebAssembly（Wasm）过去的十年","publish":true,"cover":"","showCover":null,"date":"2025-07-19","category":"javascript","url":"https://queue.acm.org/detail.cfm?id=3746171","author":"ACM Queue","translator":"","tags":["webassembly"],"priority":1},"html":"<p>Web 上的 Wasm应用有限，web浏览器的插件与扩展，轻量级虚拟化（替代docker）</p>\n<ul>\n<li>\n<p>通过 Emscripten 编译 Unreal 游戏，虽然看起来惊艳，但实际并未引发大规模游戏行业采用 。</p>\n</li>\n<li>\n<p>Adobe 的 Photoshop Web 版和 Figma 是成功实例，但它们为性能调优投入巨大；Figma 后来开始更多用 TypeScript 替代 Wasm 写UI逻辑，仅在图像处理部分使用 Wasm 。</p>\n</li>\n<li>\n<p>Wasm 用于组件化任务时效果良好，如 SQLite 编译为 Wasm 替代传统 WebSQL，以及 Perfetto 的大规模 JSON 数据解析也采用 Wasm </p>\n</li>\n<li>\n<p>WasmGC（Garbage Collection）增强后，使得 GC 语言（如 Java、Scala、OCaml、Dart/Flutter、Compose）在 Web 上的部署更加可行，谷歌表格（Google Sheets）已全面切换使用 WasmGC </p>\n</li>\n</ul>"}},{"node":{"frontmatter":{"title":"React Native支持 Node-API","publish":true,"cover":"https://cdn.prod.website-files.com/67e6c26f2d676c1963e098b9/686f99dc571b3542e04dc23a_React%20Native%20Japan%20-%20Build%20Times%20v3%20English%20transparent.001.png","showCover":null,"date":"2025-07-19","category":"mobile","url":"https://www.callstack.com/blog/announcing-node-api-support-for-react-native","author":"Kræn Hansen, Mariusz Pasiński, Matt Hargett, Vladimir Morozov, Michał Pierzchala, Mike Grabowski, Jamie Birch","translator":"","tags":["react native"],"priority":1},"html":"<p>如果你现在想写一个原生模块，依旧必须用现有 JSI 模式。比如：</p>\n<p>iOS 写 Swift/Obj-C + C++</p>\n<p>Android 写 Java/Kotlin + C++</p>\n<p>JS 中通过 NativeModules.MyModule 或 TurboModuleRegistry.get('MyModule') 使用</p>\n<p>而未来 Node‑API 到来后，上面这些代码很可能只需要写一次 Rust/C++ 模块，并通过标准 Node‑API 接口同时适配多个平台。</p>"}},{"node":{"frontmatter":{"title":"chili3d - 浏览器中运行的cad 模型编辑器","publish":true,"cover":"https://github.com/xiangechen/chili3d/raw/main/screenshots/screenshot.png","showCover":null,"date":"2025-07-19","category":"javascript","url":"https://github.com/xiangechen/chili3d","author":"xiange","translator":"","tags":["cad"],"priority":2},"html":""}},{"node":{"frontmatter":{"title":"hyper-fetch - 请求数据框架","publish":true,"cover":"","showCover":null,"date":"2025-07-19","category":"frontend","url":"https://github.com/BetterTyped/hyper-fetch","author":"BetterTyped","translator":"","tags":["react"],"priority":1},"html":""}},{"node":{"frontmatter":{"title":"react-player - 播放各种 URL，包括文件路径、YouTube、Facebook等","publish":true,"cover":"","showCover":null,"date":"2025-07-19","category":"frontend","url":"https://github.com/cookpete/react-player","author":"Pete Cook","translator":"","tags":["react"],"priority":1},"html":""}},{"node":{"frontmatter":{"title":"tiptap - 基于ProseMirror的无 UI 框架依赖富文本编辑器","publish":true,"cover":"","showCover":null,"date":"2025-07-19","category":"frontend","url":"https://github.com/ueberdosis/tiptap","author":"ueberdosis","translator":"","tags":["text editor"],"priority":1},"html":""}},{"node":{"frontmatter":{"title":"upyo - 简单的邮件发送库","publish":true,"cover":"https://upyo.org/logo.svg","showCover":null,"date":"2025-07-19","category":"nodejs","url":"https://github.com/dahlia/upyo","author":"Hong Minhee (洪 民憙)","translator":"","tags":["smtp"],"priority":1},"html":""}},{"node":{"frontmatter":{"title":"react-weather-effects - 背景图片上的天气3d特效","publish":true,"cover":"","showCover":null,"date":"2025-07-19","category":"frontend","url":"https://github.com/rauschermate/react-weather-effects","author":"Mate Rauscher","translator":"","tags":["3d"],"priority":1},"html":""}},{"node":{"frontmatter":{"title":"使用AI工具的开发者完成任务的时间比不使用AI工具的开发者慢19%","publish":true,"cover":"https://metr.org/assets/images/downlift/methodology.png","showCover":null,"date":"2025-07-19","category":"AI","url":"https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/","author":"METR","translator":"","tags":["other"],"priority":2},"html":"<p>该研究通过一项随机对照试验(RCT)，评估了2025年初的AI工具对经验丰富的开源开发者在自己仓库中工作效率的影响。研究结果表明，使用AI工具的开发者完成任务的时间比不使用AI工具的开发者慢19%。尽管开发者预期AI能够加速他们的工作，并且事后仍然认为AI提高了效率，但实际情况却相反。</p>\n<p>研究总结了三种假设来解释这些观察结果：</p>\n<ol>\n<li>RCT低估了AI的能力。</li>\n<li>基准测试和轶事高估了AI的能力。</li>\n<li>所有三种方法都基本正确，但衡量的是任务分配的子集，这些子集对模型来说更具挑战性。</li>\n</ol>"}},{"node":{"frontmatter":{"title":"可能让你生气的javascript date测试","publish":true,"cover":"","showCover":null,"date":"2025-07-19","category":"javascript","url":"https://jsdate.wtf/","author":"Sam Rose","translator":"","tags":["date"],"priority":1},"html":""}},{"node":{"frontmatter":{"title":"我使用 o3 从我保存的 Pocket 链接中分析我自己","publish":true,"cover":"","showCover":null,"date":"2025-07-19","category":"AI","url":"https://noperator.dev/posts/o3-pocket-profile/","author":"noperator","translator":"","tags":["other"],"priority":1},"html":"<p>作者使用 o3 工具分析了自己保存在 Pocket 中的近 900 篇文章链接，时间跨度近 7 年，试图了解 o3 能从这些 URL 中推断出什么关于自己的信息。</p>\n<p>o3 的分析结果相当准确，推断出作者年龄在 30 岁中期到 40 岁出头，男性，居住在弗吉尼亚州沿海地区，拥有计算机科学学士/硕士学位，是一位在安全和基础设施方面的高级/Staff 软件工程师，家庭年收入约为 15 万至 22 万美元，已婚，有 3-4 个年幼的孩子（包括一个蹒跚学步的孩子），政治上倾向于财政保守主义/公民自由主义，具有传统的社会倾向，职业生涯风险承受能力高，财务风险承受能力中等，学习方式是自学，信息摄入以深度技术、个人理财/FIRE、育儿/家务、信仰和文化为主。</p>\n<p>作者表示，o3 的分析结果非常准确，甚至有些令人毛骨悚然。</p>\n<p>作者认为，这表明广告公司可以通过分析用户的兴趣来对其进行画像，而这种能力现在已经变得非常普及。作者计划利用这些信息来构建一个个人内容推荐系统，并已迁移到 Wallabag 和 FreshRSS。</p>"}},{"node":{"frontmatter":{"title":"苏黎世联邦理工学院和洛桑联邦理工学院将发布一个在公共基础设施上开发的 LLM","publish":true,"cover":"https://ethz.ch/en/news-and-events/eth-news/news/2025/07/a-language-model-built-for-the-public-good/_jcr_content/articleLeadImage/image.imageformat.carousel.1906144068.jpg","showCover":null,"date":"2025-07-19","category":"AI","url":"https://ethz.ch/en/news-and-events/eth-news/news/2025/07/a-language-model-built-for-the-public-good.html","author":"Florian Meyer, Mélissa Anchisi","translator":"","tags":["other"],"priority":1},"html":"<p>苏黎世联邦理工学院 (ETH Zurich) 和洛桑联邦理工学院 (EPFL) 将于2025年夏末发布一个基于公共基础设施开发的大型语言模型 (LLM)。该模型在瑞士国家超级计算中心 (CSCS) 的 “Alps” 超级计算机上训练，标志着开源人工智能和多语种卓越性的一个里程碑。</p>\n<p>该LLM由EPFL、ETH Zurich和CSCS的研究人员合作开发，完全开源，旨在支持广泛应用，并促进科学、社会和产业领域的创新。该模型的一个显著特点是它能流利地使用超过1000种语言。</p>\n<p>该模型包含80亿和700亿参数两个版本，以满足不同用户的需求。700亿参数的版本将是全球最强大的完全开源模型之一。该模型在超过15万亿个高质量训练tokens上进行训练，确保了其可靠性和多功能性。</p>\n<p>在数据使用方面，该LLM的开发充分考虑了瑞士的数据保护法、版权法以及欧盟人工智能法案的透明度义务。</p>\n<p>“Alps”超级计算机为该模型的训练提供了强大的支持，该计算机配备了超过10,000个NVIDIA Grace Hopper Superchips，并使用100%碳中和电力。</p>\n<p>该LLM将以Apache 2.0许可证发布，并提供详细的模型架构、训练方法和使用指南，以方便透明的重复使用和进一步开发。</p>"}},{"node":{"frontmatter":{"title":"美国法院废止了联邦贸易委员会的点击取消要求","publish":true,"cover":"https://cdn.arstechnica.net/wp-content/uploads/2024/10/mouse-click-1536x864.jpg","showCover":null,"date":"2025-07-19","category":"other","url":"https://arstechnica.com/tech-policy/2025/07/us-court-cancels-ftc-rule-that-would-have-made-canceling-subscriptions-easier/","author":"Jon Brodkin","translator":"","tags":["other"],"priority":1},"html":"<p>美国联邦上诉法院推翻了美国联邦贸易委员会（FTC）提出的“点击取消”（click-to-cancel）规则，该规则本要求公司取消服务应该像注册一样简单。法院认为，由Lina Khan领导的拜登政府时期的FTC未能遵循美国法律规定的完整规则制定程序。虽然法院对FTC的动机表示同情，指出许多美国人不知不觉地加入了需要定期付费的订阅计划，但法院认为FTC在进行初步监管分析时未能达到1亿美元的经济影响门槛。尽管行政法法官发现该规则的影响超过了该门槛，但FTC没有进行初步监管分析。许多行业团体和企业，包括有线电视公司，在四个联邦巡回法院起诉了FTC。案件在第8巡回法院合并，由三名法官审理。法院认为，缺乏初步分析意味着行业团体和企业没有足够的时间来质疑FTC的调查结果。共和党委员Melissa Holyoak和Andrew Ferguson投票反对该规则，Holyoak警告说，新的规定可能无法通过法律挑战。现在FTC主席是Ferguson。</p>"}},{"node":{"frontmatter":{"title":"软件公司因为不得不开发chatgpt认为已经存在的新功能","publish":true,"cover":"https://s3.us-east-1.amazonaws.com/media.holovaty.com/images/2025-07-07-sheet.jpg","showCover":null,"date":"2025-07-19","category":"AI","url":"https://www.holovaty.com/writing/chatgpt-fake-feature/","author":"Adrian Holovaty","translator":"","tags":["other"],"priority":2},"html":"<p>Soundslice这家公司（提供乐谱扫描、编辑和练习服务）发现他们的错误日志里出现了大量的ChatGPT会话截图，用户上传这些截图是因为ChatGPT告诉他们Soundslice支持ASCII Tab（一种吉他音乐的简陋记谱方式）导入并播放音频。然而，Soundslice实际上并不支持这个功能。最终，Soundslice决定开发这个导入器，并相应地修改了用户界面。</p>"}}]}}