bind: Address already in use、40ms 的请求延迟、断线 75 分钟才被发现——这三个经典故障分别对应 SO_REUSEADDR、TCP_NODELAY、SO_KEEPALIVE 三个 socket 选项。本篇从内核语义逐个拆解：REUSEADDR 到底放行了什么、Nagle 与延迟确认的 40ms 之约、keepalive 默认 7200 秒的空洞，以及缓冲区选项与自动调优的互斥陷阱。

多个进程等待同一个监听 socket，新连接到来时全部被唤醒却只有一个能 accept——这就是惊群。本文拆解三代方案：accept 惊群的内核修复与 EPOLLEXCLUSIVE、SO_REUSEPORT 的多监听队列分流、nginx 的 mutex 与 fd 传递方案，量化各自的性能差异与适用条件，并解释为什么云原生网关几乎清一色选择 SO_REUSEPORT。

Last month, we shared word of a Stanford study that found entry-level employment in so-called "AI-impacted" occupations …

Sleeping on the Orion is like flipping your pillow to find “the cool side,” except that your entire bed is the cool side…

服务器到底要不要 swap？这个争论持续了三十年。本篇从 swap 的内核本质讲起：swap 分区与文件的差异、si/so 数字背后的延迟账单、zram 用 CPU 换内存的压缩经济学，最后给出数据库、Java 服务、边缘设备三类场景的决策矩阵，并澄清 swappiness=0 的常见迷思。

A latent recurrent language model applies a shared block of layers repeatedly, so one checkpoint can trade inference com…

Large language models (LLMs) and vision-language models (VLMs) are usually frozen after deployment, so they do not learn…

Web agents complete user requests by reading and acting on pages that third parties write, so an instruction planted on …

Agentic reinforcement learning (RL) disaggregates training from rollout, so each policy update must reach the rollout cl…

Unified multimodal models can both look at and render images, so in principle they can repair their own generations: dia…

Early adopters of Meta’s Muse have been speculating that the reason the AI works so well is because it’s OpenClaw under …

When Sakeena Fiza describes her work as a validation engineer at NVIDIA, she does so in terms more befitting a detective…

Degradations vary widely across images, so a practical restoration system has to handle many degradation types with one …

Google is refocusing its CC AI agent on household coordination, letting families share emails, schedules, and tasks so t…

A radiology report can already answer a clinical question, so it is hard to tell whether a vision-language model also us…

Streaming omni-modal models must decide what and when to answer from the video chunks and synchronized audio observed so…

Video understanding has rapidly evolved toward video large language models (VideoLLMs): systems that couple video repres…

刚果民主共和国的埃博拉疫情死亡人数超过 3500 人，确诊病例超过 7400 例，逾 1700 人康复。WHO 称，疫情最严重的 Ituri 省的病毒传播速度在下降，上周报告了约 300 例新增病例和 160 例死亡病例，占全国的一半；So…

拆开 gdb 和 strace 的引擎盖：ptrace 系统调用如何支撑断点、单步与系统调用追踪，它的性能代价、权限边界与现代替代方案。

PSI 改变了内存监控——从「还剩多少」到「浪费了多少时间」。上一节正文已完整阐述 PSI 的机制、读取、触发器、对比与监控接入。

网上流传的「Linux 调优脚本」大多来路不明。本篇讲透 sysctl 机制本身：/proc/sys 的树形结构、/etc/sysctl.d 的加载链与持久化、 somaxconn/port_range/file-max/overcommit 四组高频参数的真实语义与推导，容器里哪些参数进了 namespace，以及「改之前先测基线」的验证方法论。

P99 尖刺与内存回收的关系比你想的紧密。本篇拆解 zone 水位线触发的两条回收路径、kswapd 的节律与 direct reclaim 的代价、active/inactive LRU 链表的页面老化机制，以及 workingset refault 如何衡量「回收错了页面」的成本，并给出应用工程师的观测清单。

ip 命令怎么知道网卡的实时状态？容器运行时如何感知网络事件？答案都是 netlink。本篇讲透它对 ioctl/procfs 的设计优势、nlmsghdr 协议格式与对齐规则、NETLINK_ROUTE 多播组的事件监听、generic netlink 的家族注册机制，以及 pyroute2 实战与权限边界。

绑 80 端口要 root、ping 要 setuid——特权就该这样粗粒度吗？本篇讲透 capabilities 把 UID 0 拆成 38 项特权的设计、permitted/effective/inheritable/bounding/ambient 五个集合的运算规则、文件能力与 ping 的进化史、容器 --cap-drop 的经济学，以及 setcap 后动态库失效等三个高频坑。

!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…

Nov 3, 2026 GPT-Image-2.5 prompted by THE DECODER Stanford researchers Owen Queen and James Zou tested 50 widely cited c…

!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…

!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…

MCP 生态的 server 数量爆炸之后，「怎么发现与信任好的 server」成了基础设施问题：注册中心、目录、分发与安全审计正在成形。本篇拆解工具发现层的构成（官方 registry、企业私有注册中心、包管理器的类比）、信任与安全的机制设计，以及这一层对 Agent 生态的意义。

Sponsored In partnership withSnowflake In just a few years, generative AI has changed the face of enterprise analytics, …

!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…

This is today's edition of* *The Download*,*our weekday newsletter that provides a daily dose of what's going on in the …

!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox fir…

A new approach known as “reverse game theory” aims to control powerful AI models by designing systems that incentivize t…

!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…

服务器不是一台均匀的内存池：NUMA 让跨 socket 内存访问慢 40% 以上，巨页则用 2MB/1GB 页面把 TLB 压力降一个数量级。本文讲清 NUMA 拓扑、本地/远端内存与分配策略（default/bind/preferred/interleave），巨页的两种形态（hugetlbfs 与 THP）、收益与陷阱，并给出 Redis、数据库、JVM 的落地参数。

「AI 编程是资深工程师专属」的知乎高赞论断，恰好可以拿一年来的硬数据逐条检验：METR 随机对照实验里资深开发者被 AI 拖慢 19% 且自认为快了 20%，Uplevel 追踪 800 名开发者发现缺陷率上升 41%，GitClear 遥测显示重构行为塌缩到 3.8%、遗留维护骤降 74%。本文梳理 2025–2026 年九项重量级研究，论证真正的门槛不是「会不会用工具」，而是验证 AI 输出的能力——它恰恰由经验构成。

sk_buff 是内核网络栈的通用货币：一个结构体串起从驱动到协议栈到 socket 的全部路径。本文拆解它的线性数据区与缓冲区头尾指针设计、clone 与 copy 的区别、各层协议头如何被依次剥掉，并算清一次收发的真实成本账（内存分配、拷贝次数、缓存行），解释 GRO/TSO 为什么靠「攒大包」来摊薄这份成本。

内存不足时 Linux 不崩整机，而是派 OOM Killer 按启发式评分杀进程——但被杀的常常不是元凶。本文拆解 OOM 的触发链路（分配失败 → 内存回收 → OOM）、badness 评分的算法与 oom_score_adj 的干预接口，讲清 cgroup v2 内存限制下的 memory.oom.group 与 oom_kill_disable，最后给出容量水位、监控告警与优雅降级三道防线。

socket、管道、设备、普通文件能用同一套 read/write 接口操作，靠的是 VFS 虚拟文件系统层。本文拆解 VFS 的四个核心对象（superblock/inode/dentry/file）、dentry 缓存与 path lookup 的成本、页缓存与 inode 的绑定关系，以及 procfs/sysfs/devtmpfs 三个「伪装成文件」的内核接口——读完能看懂 /proc 里那些指标的真正来源。

This is today's edition of* *The Download*,*our weekday newsletter that provides a daily dose of what's going on in the …

!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…

Sponsored Provided byHPE When a developer gives an AI agent a coding task, the agent does more than answer once. It plan…

In 2001, a decade after he invented the World Wide Web, Tim Berners-Lee articulated a new vision. Together with two coll…

!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…

Rapid advances in cowork agents are shifting human-AI collaboration toward real-world automation, delivering software, d…

P99 飙升、CPU 总量不高、却有一个核的软中断逼近 100%——问题几乎总出在收包路径。本文沿「网卡 DMA → 硬中断 → NAPI 轮询 → 软中断预算 → GRO → 协议栈 → socket 队列 → epoll 唤醒」的完整数据通路走一遍，讲清每一站的成本、丢包计数器与调优参数，最后给出可照抄的排查清单与分层定位表。

Coding agents solve repository issues through long sequences of exploration, editing, and validation. These trajectories…

On-policy distillation (OPD) trains a student on its own generations by matching the teacher's token probabilities. Thes…