{
    "success": true,
    "msg": "51CTO请求成功",
    "title": "51CTO",
    "subtitle": "推荐榜",
    "update_time": "2026-10-15 10:55:58",
    "total": 50,
    "data": [
        {
            "index": 1,
            "time": "2026-10-15 13:54:27",
            "title": "给 HarmonyOS 6 柱状图加入场动画：一根一根蹦出来",
            "desc": "前言读一张柱状图，视线往往从左边走到右边。让柱子顺着这个方向依次出现，页面打开时就多了一点节奏。七根柱子不用等前一根彻底结束，稍微...",
            "url": "https://ost.51cto.com/posts/59729",
            "imgsrc": ""
        },
        {
            "index": 2,
            "time": "2026-10-15 10:05:28",
            "title": "DeepSeek服务器全球宕机！神秘Harness账号同时现身",
            "desc": "目前没有任何官方证据。但架不住时间点确实有点巧。一边是网页和API突然集体罢工，一边是Harness账号悄悄把B站主页搭了起来。",
            "url": "https://www.51cto.com/article/858265.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/122499c524c5a689645086eae8abb48d0b1a63.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 3,
            "time": "2026-10-15 09:43:30",
            "title": "鸿蒙ArkUI水波纹动画开发：从零打造数字键盘交互特效",
            "desc": "鸿蒙ArkUI水波纹动画开发：从零打造数字键盘交互特效在移动应用界面设计中，水波纹（Ripple）动效是一种经典的视觉反馈方式，能够为用户触...",
            "url": "https://ost.51cto.com/posts/59686",
            "imgsrc": ""
        },
        {
            "index": 4,
            "time": "2026-10-15 09:36:22",
            "title": "你的内部知识库，就是 LLM 最大的攻击面：用 STRIDE 给上下文感知 AI 做一次完整威胁建模",
            "desc": "如何用 STRIDE 方法，给一个上下文感知的 AI 流水线做一次完整的威胁建模。",
            "url": "https://www.51cto.com/article/858258.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/155ffcc0766738ba0fa314988b8cc7df41ac0c.png?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 5,
            "time": "2026-10-15 09:31:13",
            "title": "当视频模型回到原点，它还认得这里吗？Honeycomb用恒定大小记忆守住长视频的一致性",
            "desc": "来自哈佛大学、麻省理工学院、新加坡国立大学与 MIT IBM Watson AI Lab 的研究团队提出 Honeycomb，试图回答一个更根本的问题：视频世界模型能否在持续吸收新场景信息的同时，让特征记忆始终保持固定大小？",
            "url": "https://www.51cto.com/article/858255.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/b56e1422307d8d61d01833990e5e7dd015f290.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 6,
            "time": "2026-10-15 09:25:40",
            "title": "苹果LensVLM：长上下文压缩成图片，模型边看边「放大」",
            "desc": "LensVLM 的核心，是让模型在阅读压缩内容时，仍能按需查阅原文。模型先浏览全篇，再把阅读预算用在当前问题需要的页面上。",
            "url": "https://www.51cto.com/article/858254.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/d748ce32244e2d36c31724ebbe9db77cf4148f.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 7,
            "time": "2026-10-15 09:24:21",
            "title": "ACL 2026｜MMSearch-R1：让多模态模型学会「该搜就搜」，代价是什么？",
            "desc": "这篇论文值得关注的地方，不只是“搜索调用少了”，而是它展示了如何用奖励和数据共同塑造搜索行为。但读到成本分析时，还要追问一句：少搜省下来的开销，是否抵得上模型为了做决策而多花的计算？",
            "url": "https://www.51cto.com/article/858253.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/e42086303bbf08bdb3032870d906abf966404b.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 8,
            "time": "2026-10-15 09:20:59",
            "title": "清华提出Jev-LDE：ICL示例只改一次，1-shot平均准确率就能暴涨",
            "desc": "清华大学自动化系研究者提出了 Local Demonstration Editing（LDE），并训练了一个小型编辑模型 Jev-LDE。它不负责最终答题，而是在目标 LLM 推理前，对已经检索出的 demonstration set 做一次编辑，只允许 Keep、Delete 或 Replace 三种操作。",
            "url": "https://www.51cto.com/article/858252.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/080ee8e061f66312a028112278de404a345b84.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 9,
            "time": "2026-10-15 09:12:24",
            "title": "Anthropic震撼研究：AI竟能通过电视盒子自我复制",
            "desc": "Anthropic 的前沿红队发布了一项新研究，探讨 AI 能否像计算机病毒一样，在不同设备之间自主复制、运行，甚至不断传播。",
            "url": "https://www.51cto.com/article/858251.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/e29f63348847886efb3013c0dbf14d7eba71ed.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 10,
            "time": "2026-10-15 09:09:07",
            "title": "开源文档解析MinerU从3.0到4.0：统一文档模型、API接口、本地文档库和更友好的Agent阅读支持",
            "desc": "在AI应用特别是 ​​#RAG知识库​​ 开发中 ​​#文档解析​​ 是比较重要的前置任务，​​#MinerU​​ 是其中比较主流的项目之一，它主要负责把将PDF、Office文件等复杂文档转换为 LLM 可直接使用的 Markdown/JSON格式。",
            "url": "https://www.51cto.com/article/858250.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/26a0f4e5392c344392a9242d321566eebc8f11.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 11,
            "time": "2026-10-15 09:06:53",
            "title": "Agent一直在行动，任务为何没有进展？PoS识别并打破Belief Trapping",
            "desc": "研究团队提出 PoS（Progression of States）。它持续维护显式信念状态，记录当前世界与未解决要求。",
            "url": "https://www.51cto.com/article/858249.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/56e01492634b8cd17be445b557ad8138f6c889.png?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 12,
            "time": "2026-10-15 08:58:48",
            "title": "最新ego视角手部重建框架，单模型输出世界坐标系下手部运动",
            "desc": "上海人工智能实验室联合上海交通大学等高校团队提出InfiniHand，用一个流式模型联合估计手部位置、手部形态和相机轨迹，将连续动作恢复到共享的世界坐标系中。",
            "url": "https://www.51cto.com/article/858243.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/b5ebba5827d2a7216425222d3380af9267556f.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 13,
            "time": "2026-10-15 08:57:00",
            "title": "Claude Code桌面版暴改，一人指挥AI军团飙代码！",
            "desc": "就在今天，Anthropic正式放出了Claude Code桌面版史诗级更新！",
            "url": "https://www.51cto.com/article/858242.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/d68eceb59501912d6410585a343ba9f617b562.png?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 14,
            "time": "2026-10-15 08:55:00",
            "title": "Claude写代码太快，Anthropic连源码系统都重造了！七人只用五个月",
            "desc": "一个7人团队仅用5个月就自建了全新的版本控制系统Scion，并将公司的单体代码库从Git完整迁移了过去。",
            "url": "https://www.51cto.com/article/858239.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/e4a59e712611e0abe9a944f149e26acefbf6ca.png?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 15,
            "time": "2026-10-15 08:53:00",
            "title": "OpenAI高管会议室惊现窃听器！35岁工程师涉案被AI举报送进FBI",
            "desc": "就在刚刚，OpenAI自曝一起离奇的办公室窃听案。自家高管会议室的电视机背后，被人偷偷插上了一台加密录音设备。",
            "url": "https://www.51cto.com/article/858238.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/b8ad3b6898bd960fb67476ad1063fc72d2e232.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 16,
            "time": "2026-10-15 08:52:00",
            "title": "Linux之父彻底入坑AI！Claude Code写内核，性能暴涨3倍",
            "desc": "谁敢信，平时喷烂代码骂得最凶的Linus，这回居然自己带头用AI写代码了！",
            "url": "https://www.51cto.com/article/858232.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/974664e19bedbd2c58a55723efcf6434571817.png?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 17,
            "time": "2026-10-15 08:50:00",
            "title": "证明大爆炸，数学界变天！AI把陶哲轩逼出「数学 2.0」",
            "desc": "陶哲轩的这场演讲，表面谈数学，实则揭开了AI时代最危险的悖论：人类正在以前所未有的速度获得答案，却可能逐渐失去理解答案的能力。",
            "url": "https://www.51cto.com/article/858228.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/c780e5764605811a7fd1385982236d9da33ae4.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 18,
            "time": "2026-10-15 08:47:33",
            "title": "情绪不是标签，而是链条：TRACE框架揭示情绪的隐藏逻辑",
            "desc": "TRACE 对情绪计算领域的长期影响是深远的。它不仅提供了一个统一的情绪链条结构，也提供了一套完整的任务体系与推理方法。它让情绪计算从表层识别迈向认知推理，从碎片化任务迈向链条化理解，从单模态识别迈向多模态融合。这样的转变将推动情绪计算进入新的时代，也将为下一代智能系统提供新的基础能力。",
            "url": "https://www.51cto.com/article/858241.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/a10cb1a83ab4909e81b9419eef151c677b1b44.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 19,
            "time": "2026-10-15 08:45:08",
            "title": "\"骂了 MCP 一整年\"的AI 编程神器 Pi agent 为何肯拥抱MCP了？",
            "desc": "新框架 Pi Durable 专治“长跑 Agent 猝死”：进程挂了，重启后从检查点精确续跑，约 1.2 万行代码。争议也随之而来，极简派在评论区留下一句：所以 Pi 现在也开始堆积垃圾了。",
            "url": "https://www.51cto.com/article/858240.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/e892b8f773e68cc7a5c612cba68b80bd2b0d14.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 20,
            "time": "2026-10-15 08:37:30",
            "title": "Anthropic Scion 重做 Git：Agent 时代的版本管理系统",
            "desc": "今年 5 月，Anthropic 平均每天合并约 2,500 次代码变更；到了 9 月，这个数字已经超过 30,000。它的主干历史接近 100 GB，压缩后的工作副本超过 25 GB，仅 CI 每天就要做约 2,500 万次 checkout。",
            "url": "https://www.51cto.com/article/858237.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/214678915494055f0cb3592ddbd616fc87f393.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 21,
            "time": "2026-10-15 08:32:53",
            "title": "WorkBuddy 到底有多强大？",
            "desc": "WorkBuddy 是腾讯推出的 AI 桌面智能体，支持自然语言驱动本地任务执行。通过配置自定义模型，可以在 WorkBuddy 中接入 GPT、Claude、Gemini 等主流模型，充分发挥其文档处理、数据分析、自动化办公等能力。",
            "url": "https://www.51cto.com/article/858236.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/0479da886214965c302163ba83038d52aeccb8.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 22,
            "time": "2026-10-15 08:25:00",
            "title": "如果不是被起诉，我们永远不会知道Anthropic在2021年就完美预言了今天的AI",
            "desc": "五年前，他们在思考如何让AI成为软件工程师的同事。五年后，我们已经开始讨论如何让AI独立完成整个软件工程任务，甚至参与改进自己的模型和研发系统（RSI）。\nAI的发展有时看起来快得不可思议，但仔细回头看，你会发现：有些人其实很早就知道自己要往哪里走了。",
            "url": "https://www.51cto.com/article/858214.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/14/36625f031f9405b3e889722f1995db08183b05.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 23,
            "time": "2026-10-15 08:25:00",
            "title": "NVIDIA让1T模型权重同步提速35倍：从87.5分钟压缩到150秒",
            "desc": "在两个 AWS 区域之间的测试中，1T 检查点以 3% 元素变化率通过中继树同步，耗时 150 秒；全量检查点的传输参照为 87.5 分钟，相差 35 倍。这里的 1T 检查点由 Nemotron-3-Ultra-550B-A55B 加倍层数构造，大小为 2,242 GB；3% 则是人工注入的测试条件，不是该规模模型真实训练时测得的变化率。",
            "url": "https://www.51cto.com/article/858202.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/14/c2681d047c3452d791180267ec3bbc089988de.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 24,
            "time": "2026-10-15 08:24:09",
            "title": "三天12万 Star 的 REA，架构到底牛在哪？",
            "desc": "这个周末我把仓库的 src 目录结构、官方架构图（docs/architecture.mermaid）和 MCP 工具设计文档（docs/tool-design.md）完整读了一遍。说实话，比\"能拆 App\"更值得学的，是它的工程架构。今天这篇，带你一层一层拆。",
            "url": "https://www.51cto.com/article/858235.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/03da89f31d7189b993a6702ca473af4ebdacdb.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 25,
            "time": "2026-10-15 08:18:45",
            "title": "大厂面试放行 AI 之后：程序员的能力模型，正在被重写",
            "desc": "这篇文章不制造焦虑，只拆解三个工程问题：大厂现在到底怎么考人，公司为什么还要招程序员，以及普通人该补哪四种能力。文中的能力模型、goal 模板和两份清单，都可以直接拿走用。",
            "url": "https://www.51cto.com/article/858234.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/48c82f795c623a70417198233b4757c8cb50aa.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 26,
            "time": "2026-10-15 08:18:43",
            "title": "Redis的九种数据结构，如何选择？",
            "desc": "本文我们分析了 Redis中常见的 9种数据类型，从最基本的字符串、列表、集合到复杂的有序集合、哈希，再到位图、HyperLogLog、流和地理空间索引，每种数据类型都有其独特的特性和应用场景。熟练掌握这些数据类型的使用及其底层实现，能够帮助我们设计高效、可扩展的系统架构，充分发挥Redis的性能优势。​",
            "url": "https://www.51cto.com/article/858233.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/1612c96892f7a3637159708454df9f3a64b749.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 27,
            "time": "2026-10-15 08:13:01",
            "title": "Claude Code 史上最大升级来了：Claude MODS 正式登场",
            "desc": "对于团队，原文介绍了一个叫 ​​sec-default​​ 的内置保护 mod：它会在 Team、Enterprise 方案，以及使用托管设置的机器上优先加载，防止用户安装的 mods 覆盖权限拒绝规则。管理员也可以安排自己的 mods 在用户 mods 前后运行，只允许批准的 mods，或者完全关闭这项能力。",
            "url": "https://www.51cto.com/article/858231.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/81bd5a908868a047279606ab2fedf3165bd5b7.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 28,
            "time": "2026-10-15 08:08:18",
            "title": "明明两个线程互不干扰，为什么伪共享拖慢程序？",
            "desc": "伪共享是高性能多线程开发里很隐蔽的性能坑。线程逻辑上没有共享数据，但是变量内存排布让它们落在同一个缓存行，MESI 缓存一致性协议带来大量缓存失效开销，拖慢并行程序。​",
            "url": "https://www.51cto.com/article/858230.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/32216c83841e9835d530624a8e0f031a9f8b9f.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 29,
            "time": "2026-10-15 08:04:20",
            "title": "Java实战：用决策模型增强RAG检索效果",
            "desc": "传统的向量检索与重排模型主要评估文本间的语义匹配度，缺乏针对恶意指令的审查机制，容易让带有攻击意图的切片直接渗透到大模型的上下文中。",
            "url": "https://www.51cto.com/article/858229.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/957aa8854d34422327c195b1eadb3fe1a6f619.png?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 30,
            "time": "2026-10-15 07:57:40",
            "title": "刚刚，Claude彻底解禁第三方应用！",
            "desc": "就在刚刚，Anthropic祭出王炸，Claude即将大爆发！简而言之：Anthropic彻底扫清了政策障碍，正式确认Claude订阅用户的额度，可以无缝接入第三方App和Agent！",
            "url": "https://www.51cto.com/article/858227.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/03e2b412654c84e0615946033cc25df2d63ef3.png?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 31,
            "time": "2026-10-15 07:54:58",
            "title": ".NET中一行 Hello World，凭什么能从 76 MB 瘦到 0.73 MB？",
            "desc": "天下没有免费的午餐：AOT 产物不能在运行时生成新代码，动态代理、运行时代码生成这类玩法，都得先确认兼容 AOT。Windows 上编译还要装 Visual Studio 的 C++ 构建工具，头一回配环境得多花点功夫。",
            "url": "https://www.51cto.com/article/858226.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/983ff0d99e63135be580756d5ff21d48c154f1.png?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 32,
            "time": "2026-10-15 07:45:00",
            "title": "Claude Code的Harness到底值多少钱？论文发现通过率接近，Token开销却可能差很多",
            "desc": "同模型配对、相同配置重跑和轨迹核查给出的答案是：两个 Qwen 模型上，Claude Code 与 mini-SWE-agent 在±5个百分点范围内等价；45道难题不足以分辨几分的稳定差距。能够明确观察到的 harness 差异包括部分运行被提前终止，以及固定输入与步数带来的费用变化。",
            "url": "https://www.51cto.com/article/858216.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/14/244c81b8530a626be91998a2fd6ef9066f912f.png?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 33,
            "time": "2026-10-15 07:31:12",
            "title": "面试官：大促整点发券，峰值 100 万 QPS，一共 1 万张券，你怎么设计？",
            "desc": "今天把整条漏斗从上到下捋一遍：每一层各自削掉什么、代价是什么，以及这套设计里最值钱的那条原则——少卖可以补，超卖没法补，所有兜底都要往少卖那侧倒。",
            "url": "https://www.51cto.com/article/858224.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/417d55576ba9fedb7eb11639ec4bc8ee578d2a.png?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 34,
            "time": "2026-10-15 07:15:00",
            "title": "Google 面试开始允许用 AI 了，刷题时代要结束了吗？",
            "desc": "面试官会评估候选人的 AI 熟练度。先说第一个，你怎么写提示词，能不能把问题准确地描述给 AI。然后是第二个，AI 给你返回了答案，你能不能判断它对不对。第三个，当 AI 输出有问题的时候，你怎么调试、怎么修正。",
            "url": "https://www.51cto.com/article/858199.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/14/58f5ae477190a6046d7716e7bf804f5ccf935f.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 35,
            "time": "2026-10-15 07:00:00",
            "title": "面试官：生产环境 Agent 如何做任务优先级调度？",
            "desc": "调度只面向就绪步骤，先按资源池划分，再分配等级份额，等级内做租户公平。任务成本差异大时按成本记账，低优保留最低服务份额，避免高优持续到达造成饥饿。",
            "url": "https://www.51cto.com/article/858189.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/14/f55760b71df6d1827f9134f1cd03c4d9fafcb8.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 36,
            "time": "2026-10-15 06:15:00",
            "title": "Dropout 深度解析：从随机子网络、Inverted Dropout 到 Ridge Regression",
            "desc": "听起来这应该会让情况变得更糟。但事实恰恰相反，它成为了深度学习中应用最广泛的技术之一。它是帮助 AlexNet 在 2012 年赢得 ImageNet 的关键因素之一，如今仍然内置于现代语言模型背后的 Transformer 架构中。",
            "url": "https://www.51cto.com/article/858207.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/14/d84b3a428be5d5fac43575d126561bbf64a56d.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 37,
            "time": "2026-10-15 06:00:00",
            "title": "Anthropic自研Scion替代Git，检出从分钟级变秒级、CI多出约25% CPU",
            "desc": "Anthropic 自己把边界划得很清楚：Scion 绝不是 Git 的通用替代品，只有规模到了超大 monorepo、且客户端以 agent 和自动化任务为主时才有意义。目前没有公开发布的计划，他们接下来会和 GitHub、GitLab 以及开源社区合作，推动 Git 服务端提供能力相近的服务端接口。",
            "url": "https://www.51cto.com/article/858198.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/14/81c1d7072312ab931cd49592ec733002e1fbeb.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 38,
            "time": "2026-10-15 05:25:00",
            "title": "<calc-input> 能算公式不重要，重要的是它怎么接进原生表单",
            "desc": "要么把原生 ​​<input>​​ 整个架空，换成框架里的受控组件，于是提交、校验、回车换行、Tab 顺序、剪贴板粘贴、中文输入法全得自己重新接一遍；要么就直接引个重一点的第三方依赖，把表单变成它自家的私有结构。结果原生表单那套现成的提交和约束校验，反而被丢在了一边。",
            "url": "https://www.51cto.com/article/858213.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/14/b3420528740db95a99e2765e7fc478aa869da9.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 39,
            "time": "2026-10-15 05:00:00",
            "title": "动态规则的 Agent 实践：用确定性工程驾驭 LLM 的资产智能验收",
            "desc": "让LLM负责「读懂材料、按规则比对」，工程系统负责「流程编排、质量保障、可追溯」：规则以文本形式配置化下发，不再硬编码进代码。",
            "url": "https://www.51cto.com/article/858215.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/14/237c8a620b59da3a935043b5e4044a1132c959.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 40,
            "time": "2026-10-15 05:00:00",
            "title": "AACL最新RAG，复旦META赢麻了",
            "desc": "今天介绍的文章是UBC、中科大、复旦、浙大提出META：首个把RAG式情景记忆直接嵌进高频交易决策回路的多智能体框架——决策前用指标向量检索历史相似行情的先例，交易后把结果写回记忆库增删改。",
            "url": "https://www.51cto.com/article/858212.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/14/95000c830f03d415ddd982c0c138ed96660f39.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 41,
            "time": "2026-10-15 04:45:00",
            "title": "2029 年大限将至！Go 前密码学掌门人更新“密码学正确答案”：所有非对称算法都得换",
            "desc": "2026 年还有另一个重要变化：LLM 现在非常擅长发现 Bug。“让你买得起的最强模型帮你审查一遍”，现在已经正式成为一条“密码学正确答案”，和老牌经典“使用内存安全语言”并列。先看结论：一张速查",
            "url": "https://www.51cto.com/article/858217.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/14/e32b1a833080ff8893b24514eb28c9d612119f.png?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 42,
            "time": "2026-10-15 04:45:00",
            "title": "REA 刷屏，大模型要逆向吞噬一切，桌面付费软件真的无了？",
            "desc": "传统软件行业防侵权，靠的是净室开发，一组人拆解竞品，只写功能说明书，另一组从没见过竞品代码的人照着说明书重写，用来证明新代码是独立创作的。现在好了，同一个 Agent 又读反编译结果，又写新代码，隔离墙没了。",
            "url": "https://www.51cto.com/article/858186.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/14/a6fb472945834ffd6ff06416cd53140ca02105.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 43,
            "time": "2026-10-15 04:00:00",
            "title": "面试官：“线上 Agent 突然变慢，你怎么定位？”，我：“先换个更快的模型”，他：“连慢在哪都不知道，就开始优化？”",
            "desc": "排查 Agent 延迟，最怕一上来凭感觉优化。正确思路是先把一次任务「拆开看」，再沿着真正拖慢结果的路径动手。",
            "url": "https://www.51cto.com/article/858211.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/14/65db97b457cc1de7180681b9edf6189cd34b98.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 44,
            "time": "2026-10-15 04:00:00",
            "title": "不会写代码，也能做微信小程序，把想法说出来，让 WorkBuddy 帮你生成并发布小程序",
            "desc": "如果前面走的是自有账号路线，体验版试用满意后，打开发布面板的“小程序设置”，按实际功能填写备案、隐私保护等所需信息。再选择“正式版”提交，WorkBuddy 协助处理代码上传、提交审核和版本发布，你可以在面板里查看进度。",
            "url": "https://www.51cto.com/article/858171.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/14/a22981620a274daa36d688f1f51881d48e0835.png?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 45,
            "time": "2026-10-15 03:00:00",
            "title": "淘天三面被P9问：当多条记忆内容在Agent互相冲突，你会怎么解决？我脱口而出“用最新的”，他接下来的3个追问把我问懵了",
            "desc": "这篇文章，就是把那场面试里被问住的问题，好好想清楚。仔细拆解下来，所谓的“记忆冲突”其实分为两大类型：一类是系统底层的版本冲突，另一类是内容层面的语义冲突。",
            "url": "https://www.51cto.com/article/858210.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/14/2188b9a780f97fa63d28066ad36f3ffe43646d.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 46,
            "time": "2026-10-15 02:00:00",
            "title": "从海量用户行为中挖掘 LLM 智能，十项基准全线提升",
            "desc": "大模型的能力越来越强，也越来越多地在搜索、广告、推荐等核心业务的环节中被使用。它要求模型跨越长历史、识别有效线索，容忍噪声数据，并依据证据形成和更新判断。Vela 将这些要求转化为训练任务，探索了真实用户行为作为监督来源的价值。",
            "url": "https://www.51cto.com/article/858209.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/14/a11e4076986bafbbdb729903cd00e947db8c08.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 47,
            "time": "2026-10-15 01:00:00",
            "title": "RSI：AI如何改进自己？",
            "desc": "RSI值得关注的，不是“AI开始无限升级”的口号，而是一个可检验的问题：我们能否把AI研发中的有效改进，积累成系统自身的能力，并让这些能力帮助下一轮研发？",
            "url": "https://www.51cto.com/article/858173.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/14/e46a4c72508c4fed1cb6757805bb095df0fd0c.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 48,
            "time": "2026-10-15 00:24:56",
            "title": "Ontology 工程化实战：别再把 Ontology 当知识图谱，真正难的是让 AI 安全地\"动起来\"",
            "desc": "很多人第一次听说 Ontology，脑子里想到的是 Object 连 Link 画成一张图。但真正进了生产环境你会发现，这远远不够。完整的 Ontology 应该是一个闭环：从企业世界里的对象和关系，走向意义，走向可计算的能力，走向受权限约束的执行，走向结果，再走向反馈，反过来完善 Ontology 本身。",
            "url": "https://www.51cto.com/article/858220.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/38a4ed0526c955055ac5216d2a3df52c5c82d4.png?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 49,
            "time": "2026-10-15 00:10:17",
            "title": "教师知道正确答案，却未必知道学生该怎么纠错：DCE提出动态共演化蒸馏",
            "desc": "论文训练使用 OPSD 所采用的 14,717 道 OpenThoughts 数学推理题，并在 Qwen3 的 1.7B、4B、8B 等规模以及一个 Gemma-4-12B-IT 迁移实验中报告结果。主结果说明，在这组数学训练数据、非思考解码和指定采样预算下，随轮次刷新特权教师的 DCE 与经过验证的短改写可以共同改善 8B、4B 的准确率—长度组合。",
            "url": "https://www.51cto.com/article/858219.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/15/866d12a39120fb47ada36853d347df4e480d84.png?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        },
        {
            "index": 50,
            "time": "2026-10-15 00:10:00",
            "title": "从聊天到执行：Ailo 个人 AI 助手的设计与实践",
            "desc": "Ailo 是一个源码公开的桌面个人 Agent 实践项目：用户配置自己的模型，把材料、项目、执行状态和成果连接起来。本文以 v0.1.62 的设计与实现为例，先看它怎样组织工作，再动手创建一个每日 AI 学习助手。",
            "url": "https://www.51cto.com/article/858172.html",
            "imgsrc": "https://s2.51cto.com/oss/202610/14/f98c78e58b7ab706eef440c9daf196238f78d9.jpg?x-oss-process=image/format,jpg,image/resize,m_fixed,w_416,h_232"
        }
    ],
    "copyright": "鬼鬼API——聚合热搜榜"
}