2026-10-25 · Sunday

AI Hot Topics

1. OpenAI is winning Anthropic coding customers with steep discounts: The Information reports that at Rippling, Anthropic’s share of engineers’ coding work fell from 57% in May to 18%, while OpenAI rose from 16% to 52% — CEO Parker Conrad cited cost. Replit moved most of a $100M+/yr AI contract to OpenAI after a 30%+ discount (Anthropic declined to extend the prior deal without a larger commitment); OpenAI is now the default for Replit Agent, with Claude still selectable. DataCamp was offered a ~40% discount conditioned on using Codex. Switching costs look near zero; model routing is the new moat. theinformation.com/articles/anthropics-m… | X @amir | X @sairahul1
   4

   OpenAI 靠大幅折扣从 Anthropic 抢走编程客户：据 The Information，Rippling 工程师写代码用的模型里，Anthropic 占比从 5 月的 57% 掉到 18%，OpenAI 从 16% 升到 52%，CEO Parker Conrad 说就是因为更便宜。Replit 在拿到 30% 以上折扣后，把一份年规模超 1 亿美元的 AI 合同大部分切到 OpenAI（Anthropic 不愿在不加码承诺的情况下续原先折扣）；OpenAI 现成 Replit Agent 默认，Claude 仍可选。DataCamp 也被开出约 40% 折扣，条件是用 Codex。切换成本近乎为零，模型路由才是护城河。

2. Anthropic: fold unnecessary subagents into the main agent loop; one agent + skills often beats multi-agent: Applied AI’s Oct 22 guide (went viral overnight) says enterprise setups that give every team its own subagent + orchestrator copy the org chart and pay for context loss, Frankenprompts, and latency — one customer ran ~2x faster after collapsing to a single loop. Rules of thumb: Skills teach how; Tools/MCP do things; Subagents only for parallel work, isolated context, or when one agent can’t pick tools. Progressive disclosure loads skills/tools on demand; Claude Code’s system prompt for newer models was cut 80%+. Migration kit: turn subagent prompts into skills, APIs into tools, keep end-to-end evals. claude.com/resources/articles/bu… | X @cherry_mx_reds | X @shao__meng
   5

   Anthropic：把多余子 Agent 折进主循环；单 Agent + Skills 常胜过多 Agent：Applied AI 10 月 22 日指南（夜间刷屏）说，给每个团队各挂一个子 Agent 再加编排器，等于把组织架构抄进 Agent，付出上下文损耗、Frankenprompt 和延迟——有客户换成单循环后大约快 2 倍。口诀：Skills 教怎么做；Tools/MCP 去执行；Subagent 只留给并行、隔离上下文、或单 Agent 选不好工具时。渐进披露按需加载；面向新模型的 Claude Code system prompt 砍掉 80%+。迁移：子 Agent 提示词变 skill、API 变工具，并保住端到端评测。

3. Cloudflare’s security-audit skill tops GitHub Trending (~44k stars): MIT-licensed cloudflare/security-audit-skill is the single-repo seed of Cloudflare’s vulnerability discovery harness. Six phases: recon → coverage-led hunting → candidate validation → structured findings → independent record verification → target-neutral report. Iron rule: the agent that checks a bug is never the one that found it; unconfirmed stays needs_validation. Install: npx skills add https://github.com/cloudflare/security-audit-skill --skill security-audit, then ask your coding agent to “security audit this codebase.” Needs tool use + parallel sub-agents and a sandboxed run for execution checks. GitHub cloudflare/security-audit-skill | X @chasen_liao
   4

   Cloudflare 安全审计 skill 登顶 GitHub 趋势（约 4.4 万星）：MIT 开源的 cloudflare/security-audit-skill 是 Cloudflare 内部挖洞 harness 的单仓起点。六阶段：侦察 → 覆盖率驱动狩猎 → 候选验证 → 结构化发现 → 独立记录核验 → 中性报告。铁律：验洞的 Agent 绝不是找洞的那个；未确认的只标 needs_validation。安装：npx skills add https://github.com/cloudflare/security-audit-skill --skill security-audit，再让编程 Agent「security audit this codebase」。需要支持工具与并行子 Agent，执行验证还要沙箱。

Tips & Hands-on

1. Harrison Chase: five components of a company harness: LangChain’s internal GTM / OpenSWE / Content Studio agents share the same recipe — (1) live where people already work (Slack, with breakout threads so agents don’t pollute human chat); (2) identity: on-behalf-of vs fixed service credentials, with tools gated by DM vs channel; (3) global + per-user memory; (4) code execution in a sandbox with a secret proxy for auth’d services; (5) a large share of runs are scheduled/event-triggered ambient agents, not just @mentions. Built on Managed Deep Agents over the open-source Deep Agents harness. X @hwchase17
   4

   Harrison Chase：公司级 harness 的五个组件：LangChain 内部的 GTM / OpenSWE / Content Studio Agent 共用同一套配方——(1) 住在人已经在干活的地方（Slack，可 breakout 到独立线程以免污染人类对话）；(2) 身份：代表用户 vs 固定服务账号，工具按私信/频道门禁；(3) 全局 + 按用户记忆；(4) 沙箱里跑代码，经 secret proxy 安全连需鉴权服务；(5) 大量运行是定时/事件触发的 ambient agent，不只靠 @。底座是 Managed Deep Agents（开源 Deep Agents harness 之上）。

2. vista8: jailbroken old Kindle + KOReader, then Claude Code/Codex to tweak the UI: cheap used Kindles last forever; install KOReader + WeChat Reading, add SimpleUI if the UI bugs you, or just edit it with Codex / Claude Code. He added Kindle send to his “乔木 Read” Obsidian plugin (native Kindle and KOReader). Agent era move: don’t like the hardware UI — change it; missing a link — ship a plugin.

   3

   向阳乔木：越狱旧 Kindle + KOReader，再用 Claude Code/Codex 改 UI：旧 Kindle 便宜续航久；装 KOReader + 微信读书插件，界面不爽可上 SimpleUI，也能直接用 Codex 或 Claude Code 改。他把「乔木 Read」Obsidian 插件加上了发送到 Kindle（原生 Kindle 与 KOReader 都支持）。Agent 时代：硬件不满意就改界面、缺链路就搓插件。X @vista8

3. lxfater: unlock hidden NVIDIA consumer-GPU compute on Linux: for some RTX 30-series, 4090, and 4070 Ti SUPER — he patches protected clock registers, not blind power-limit overclocks; measured solid gains on a 4070 Ti SUPER and expects help on LLM prefill too. Full steps in the X Article. Register tweaks can brick or void warranty — DIY only.

   3

   铁锤人：手把手解锁 NVIDIA 部分消费卡隐藏算力（Linux）：针对 RTX 30 系部分型号、4090、4070 Ti SUPER，改的是控制运算速度的受保护寄存器，不是盲目拉功耗超频；他在 4070 Ti SUPER 上实测有明显提升，并估计对大模型推理的 prefill 也有加成。完整步骤见 X Article。改寄存器有变砖/保修风险，只适合自己折腾。X @lxfater

AI Hot Topics

1. Codex Day 20: “Rest and reset” — usage limits reset again, by midnight Pacific: on day 20 of OpenAI’s 28-day Codex shipping challenge, Tibo (@thsottiaux) skipped a feature drop and announced another limit reset, landing “by midnight PST” (about 15:00 Beijing on 10/25); the post passed 5,000 likes within two hours, and users in Chinese, Japanese and Korean replies were already reporting refills before the deadline. If you are near your cap, finish heavy runs before the reset; the 28-day challenge has 8 days left, so don’t plan around resets continuing after it. X @thsottiaux | X @thsottiaux
   4

   Codex 第 20 天：「休息 + 重置」，额度再次重置，太平洋时间午夜前生效：OpenAI 28 天 Codex 挑战第 20 天，Tibo（@thsottiaux）没有发新功能，而是宣布再一次重置用量额度，「by midnight PST」前到账（约北京时间 10/25 15:00）；帖子两小时内超 5000 赞，中日韩用户在回复里陆续报告额度已提前恢复。快用完的话可以在重置前把重任务跑完；28 天挑战只剩 8 天，别指望挑战结束后还这样频繁重置。

2. Claude’s “Reverse Engineer Anything” workflows now trip safety checks mid-answer: users report that asking Claude (Opus 5.5) to run or even summarize REA — the MIT-licensed MCP server that lets Claude Code / Codex / Cursor drive Ghidra, IDA or Hopper to inspect binaries, APKs and Electron apps, which 404 Media put past 342k GitHub stars — now hits a cyber safety guardrail partway through the response. It follows viral cases of people using Claude to strip ads and telemetry from Samsung TV firmware; posts saying “Anthropic killed REA” drew 11k+ likes each. Anthropic has published nothing REA-specific, and some users say the classifier was always twitchy; if you do legitimate interop or binary analysis, expect intermittent refusals and keep a local or second model as backup. X @AbhinavXJ | X @Pirat_Nation | kotaku.com/claudes-reverse-engin…
   4

   Claude 的「Reverse Engineer Anything」流程开始在回答中途触发安全拦截：用户反映，让 Claude（Opus 5.5）跑 REA、甚至只是总结 REA，都会在回答一半时触发网络安全护栏。REA 是 MIT 开源的 MCP 服务，让 Claude Code / Codex / Cursor 驱动 Ghidra、IDA、Hopper 分析二进制、APK 和 Electron 应用，404 Media 报道其 GitHub 星数已超 34.2 万。此前有人用 Claude 去掉三星电视固件里的广告和遥测并刷屏；「Anthropic 杀死了 REA」的帖子各有 1.1 万多赞。Anthropic 没有发布任何针对 REA 的说明，也有用户说这个分类器一直很敏感；做正当互操作或二进制分析的人要预期间歇性拒答，备好本地或另一家模型。

3. Cognition open-sources Devin Vulnerability Fixes: agent skills that patch this week’s PHP, Node.js and GitLab RCEs: three MIT-licensed SKILL.md skills (fix-php-vuln, fix-node-vuln, fix-gitlab-vuln) cover PHP2Shell (CVSS v4 9.2), the Oct 21 Node.js release (10 CVEs) and self-managed GitLab CVE-2026-110264 (CVSS 10.0). The agent first does a read-only inventory of Dockerfiles, base images, distro packages, CI, Helm and Terraform, changes nothing until you approve, verifies target versions/digests/checksums, runs tests before and after in the new image, and opens one PR per repo with a risk label and rollback plan. Works in Devin, Claude Code, Codex, Cursor, Gemini CLI, OpenCode and others; there is no published fix-rate benchmark, so it is only as good as your tests. Note distro packages (Debian/Ubuntu) backport fixes without changing php -v. X @Marktechpost | devin.ai/blog/vulnerability-fi… | GitHub CognitionAI/devin-vuln-fixes
   4

   Cognition 开源 Devin Vulnerability Fixes：用 agent skill 修补本周 PHP、Node.js、GitLab 的 RCE：三个 MIT 协议的 SKILL.md（fix-php-vuln、fix-node-vuln、fix-gitlab-vuln），覆盖 PHP2Shell（CVSS v4 9.2）、10 月 21 日 Node.js 安全版本（10 个 CVE）和自托管 GitLab CVE-2026-110264（CVSS 10.0）。Agent 先只读盘点 Dockerfile、基础镜像、发行版包、CI、Helm 和 Terraform，你批准前不改任何东西；校验目标版本、镜像 digest 和校验和，在新镜像里跑前后测试，每个仓库开一条带风险标签和回滚方案的 PR。Devin、Claude Code、Codex、Cursor、Gemini CLI、OpenCode 等都能用；没有公开修复率基准，效果取决于你的测试。注意 Debian/Ubuntu 的发行版包回移补丁后 php -v 版本号不变。

4. Gemini API spend caps lag by ~10 minutes, so they can be overshot: developer @BenjaminDEKR set an $80 spend cap in AI Studio, then an unexpected spike ran the project past $200. Logan Kilpatrick replied that budget alerts are only emails, while spend caps are real caps but currently apply with about a 10-minute latency (down from more than 24 hours originally, hence the experimental tag); a realtime version is in the works. Until then, don’t rely on the cap alone for runaway agent loops — add your own rate limits or kill switch. X @BenjaminDEKR | X @OfficialLoganK
   3

   Gemini API 的花费上限有约 10 分钟延迟，可能被超出：开发者 @BenjaminDEKR 在 AI Studio 给项目设了 80 美元上限，一次意外突增后账单超过 200 美元。Logan Kilpatrick 回应：预算提醒只是邮件，花费上限确实是硬上限，但目前生效有约 10 分钟延迟（最初超过 24 小时，所以标了 experimental），实时版本正在做。在那之前，别只靠这个上限防 agent 失控循环，自己加限流或熔断。

Tips & Hands-on

1. Strata: run a 125B-class model locally on a gaming PC and wire it into Claude Code / Codex: an open-source inference engine (MIT, ~32k GitHub stars) that tiers VRAM, RAM and SSD with dynamic MoE expert caching, so 32 GB of combined RAM + VRAM is enough to start Qwen3.8-Flash-Next, and from 0.2.4 DeepSeek V4.1-Flash in the terminal; supports NVIDIA/AMD/Intel/Apple Silicon and exposes local OpenAI- and Anthropic-compatible APIs so Claude Code, Codex CLI and OpenCode connect directly. The post cites ~94 tok/s for a Q2 quant on an RTX 5070 and 1.6–1.8x from speculative decoding; needs ~80 GB of SSD.

   4

   Strata：让普通游戏电脑本地跑 125B 级大模型，并接入 Claude Code / Codex：开源推理引擎（MIT，GitHub 约 3.2 万星），用 GPU 显存、内存和 SSD 三级调度加 MoE 专家动态缓存，显存加内存合计 32GB 起步就能跑 Qwen3.8-Flash-Next，0.2.4 起还能在终端跑 DeepSeek V4.1-Flash；支持 NVIDIA/AMD/Intel/Apple Silicon，提供兼容 OpenAI 和 Anthropic 的本地 API，Claude Code、Codex CLI、OpenCode 可直接接。帖子称 RTX 5070 跑 Q2 量化约 94 token/s、推测解码提速 1.6–1.8 倍；需约 80GB SSD 空间。X @wiksongkim ｜ GitHub Niko1221/Strata

2. Li Jigang: a “poem-as-picture” prompt for helping kids learn classical poetry: the model turns a poem into an image made only of the poem’s characters, where their placement and direction form the picture. Four steps: gather facts (only the poem’s images and their relations — above, hanging, moving which way) → project (each image is built from the characters that name it; hanging means an arc downward, tall means stacked vertically, wind means a diagonal sweep; words like “不知/谁/似” go in blank space or at the seams) → render → self-check (every character comes from the poem, the scene is recognizable without the text, the poem can be read verbatim along a path, no filler characters). Ends with a 诗：{...} slot; the step-plus-acceptance-criteria structure transfers to other text-image tasks.

   3

   李继刚：辅导孩子学诗词的「文字成画」Prompt：让模型把一首诗变成只由诗中文字组成的图，字的位置和走向就是画面。四步：找事实（只写诗里有的意象和上下、垂落、动向关系）→ 定投影（意象由指称它的字构成，垂下就沿弧线排、高就竖着叠、风就斜穿；「不知」「谁」「似」这类字放在留白或意象交界处）→ 出图 → 逐条自检（所有字出自原诗、不看原诗也认得出画面、沿路径能一字不差读出全诗、不拿别的字凑）。末尾留 诗：{在这里放诗} 占位，结构清楚、带验收标准，也适合改成其他图文任务。X @lijigang
3. Agent Manager (Hugging Face, @lvwerra): run Claude Code, Codex, Gemini CLI, opencode and more side by side from your own private HF Space: an Apache-2.0 Docker Space template described as a “private cloud manager for AI coding CLI sessions” — duplicate it and you get a browser UI for parallel agent sessions in the cloud. Heads-up from the poster: the app has no login of its own, so keep the Space and its storage bucket private. X @haginota | huggingface.co/spaces/lvwerra/agent-…
   3

   Agent Manager（Hugging Face，@lvwerra）：在自己的私有 HF Space 里并排跑 Claude Code、Codex、Gemini CLI、opencode 等：Apache-2.0 的 Docker Space 模板，自述为「AI 编程 CLI 会话的私有云管理器」，复制一份就有一个浏览器界面，在云上并行开多个 agent 会话。发帖人提醒：应用本身没有登录，Space 和存储桶一定要设成私有。

4. tualatrix: after his Claude Max account was banned, his own agent tooling let him switch with almost no friction: he only uses agent services through APIs, CLIs or SDKs and keeps all session data locally, so he could open a new account or move to another model the same day (with different results). He also demoed Grok Bot scanning his replies every 15 minutes, pulling a suspected AI-reply account’s last 10 replies into a table; he only makes the hide-and-block call, which is logged to an audit sheet before the bot executes it.

   3

   tualatrix：Claude Max 被封后，自己的 Agent 工具让他几乎零成本换号续上：他只通过 API、命令行或 SDK 用各家 Agent 服务，会话资料都在本地自己管，所以被封当天就能新开账号或换其他模型接着干（效果会有差别）。他还演示了用 Grok Bot 每 15 分钟扫一遍回复区，疑似 AI 回复就拉出该账号最近 10 条回复整理成表，人只做「隐藏+屏蔽」的决定，决策写入审计表后再由 Bot 执行。X @tualatrix ｜ X @tualatrix

AI Hot Topics

1. Follow-up on “Anthropic killed REA”: it’s the cyber classifiers, and flagged requests quietly drop to Opus 4.8: Anthropic’s help center says Opus 5 and 5.5 run safety checks on every request and on everything Claude reads (memory, connectors, web results, files), so a reverse-engineering session can be cut off by content you didn’t type. Offensive-leaning work such as exploit generation, binary-based vulnerability scanning and pentesting falls back to the weaker Opus 4.8 with a notice that’s easy to miss in long sessions; the picker then stays on 4.8 for the rest of the conversation. REA itself isn’t blocked, and source-code review and patching are fine. If answers suddenly feel worse, check which model labeled the reply, turn off automatic switching for a clean stop, edit the triggering message before retrying, and security professionals can apply to the Cyber Verification Program for fewer blocks. X @lenny_enderle | support.claude.com/en/articles/16049681-… | support.claude.com/en/articles/14604842-…
   4

   「Anthropic 杀死 REA」后续：原因是网络安全分类器，被标记的请求会悄悄回落到 Opus 4.8：Anthropic 帮助中心说明，Opus 5 和 5.5 会对每个请求以及 Claude 读到的所有内容（记忆、连接器、网页结果、文件）做安全检查，所以逆向会话可能因为不是你输入的内容而被中断。偏攻击性的任务，如生成 exploit、基于二进制的漏洞扫描和渗透测试，会回落到更弱的 Opus 4.8，并附一条在长会话里很容易漏看的提示；之后整个对话都停留在 4.8。REA 本身没被封，源码审查和打补丁不受影响。如果回答突然变差，先看回复标注的是哪个模型；可以关掉自动切换让它直接停下，重试前先改写触发的那条消息；安全从业者可以申请 Cyber Verification Program 以减少拦截。

2. “Is Opus 5.5 dumber?” goes viral, but an open daily tracker shows no real drop: Thariq’s “wait is it just me or is Opus 5.5 dumber” became a trending topic a month after launch, with many users reporting more mistakes. An open tracker that asks Opus 5.5 the same 78 questions in Claude Code on a Max plan every day shows 58.4% over its first 10 days and 57.1% over the latest 10 (through Oct 23), within normal variation. X’s story summary points to defaults, safety fallbacks, rate limits and the Oct 24–25 service errors as likelier causes; given item 1, check whether you were silently moved to Opus 4.8 before blaming the model. X /i/trending/2114329522019733839 | X @flintwts
   3

   「Opus 5.5 是不是变笨了？」刷屏，但一个公开的每日追踪没看到明显下降：Thariq 那句「是只有我觉得 Opus 5.5 变笨了吗」在发布一个月后登上热门话题，很多用户说错误变多。一个公开追踪每天在 Max 套餐的 Claude Code 里问 Opus 5.5 同样的 78 道题：前 10 天 58.4%，最近 10 天 57.1%（截至 10/23），在正常波动范围内。X 的话题摘要认为更可能的原因是默认设置、安全回落、限流和 10/24–25 的服务错误；结合第 1 条，怪模型之前先看看是不是被悄悄切到了 Opus 4.8。

3. Codex reset fully propagated; Tibo teases a “fun week ahead”: following the day-20 reset in the last digest, Tibo (@thsottiaux) confirmed around 14:06 Beijing time that “it is all propagated. Enjoy some mad codexing tomorrow. Fun week ahead too.” The post passed 4,000 likes and 380k views. Limits are refilled now, and the remaining 8 days of the 28-day Codex challenge hint at more launches this week. X @thsottiaux
   3

   Codex 额度重置已全量生效，Tibo 预告「接下来一周很好玩」：接上一期的第 20 天重置，Tibo（@thsottiaux）在北京时间约 14:06 确认「已经全部生效，明天尽情 Codex 吧，接下来一周也很好玩」，帖子超 4000 赞、38 万浏览。额度现在已经回满；28 天 Codex 挑战还剩 8 天，暗示本周还有新发布。

4. Google DeepMind field study: the same proactive agent was a lifesaver to some teams and a source of stress to others: an agent that acts without being asked ran in 20+ teams at a large tech company for 5 months, and researchers interviewed 17 users. It treated stale docs as current, tagged a VP on 16 bugs, shared a draft poster before its owner was ready and misread sarcasm; some people moved sensitive talks to agent-free chats. Teams that chose the agent forgave early mistakes, but people who had it pushed on them didn’t. The takeaway for anyone rolling out proactive agents: get team consent, separate permissions, and make every action easy to undo. X @rohanpaul_ai
   3

   Google DeepMind 实地研究：同一个主动型 Agent，对有的团队是救星，对有的人是压力源：一个不等吩咐就行动的 agent 在一家大型科技公司的 20 多个团队里跑了 5 个月，研究者访谈了 17 位用户。它把过期文档当成最新、在 16 个 bug 上 @ 了一位 VP、在主人准备好之前就分享了海报草稿、还听不懂反讽；有人把敏感讨论挪到没有 agent 的群里。主动选择它的团队能原谅早期错误，被强推的人则不会。给推主动型 agent 的人的启示：先取得团队同意、权限分开、每个操作都要能轻松撤销。

5. Elvis Presley rights holders sue Anthropic over 274 songs, targeting outputs and model weights too: filed Oct 23 in N.D. California (ABG EPE IP LLC et al. v. Anthropic), the suit seeks about $41.1M in statutory damages ($150K per song), about $48.6M including claims of bypassed access controls and stripped writer credits, and asks for unlicensed copies to be destroyed even in training datasets and model weights. It also says Claude, while refusing to print lyrics, will rewrite songs “for a modern audience” that land very close to the originals. Anthropic hasn’t responded; the 2025 fair-use ruling on bought books covers training, not near-copies in outputs. X @rohanpaul_ai | rollingstone.com/music/music-news/elvi…
   2

   猫王歌曲版权方就 274 首歌起诉 Anthropic，矛头还指向输出和模型权重：10 月 23 日在加州北区联邦法院起诉（ABG EPE IP LLC 等诉 Anthropic），按每首 15 万美元法定赔偿索赔约 4110 万美元，加上绕过访问控制和删除词曲作者信息的指控共约 4860 万美元，并要求销毁未授权副本，包括训练数据集和模型权重里的。诉状还说 Claude 虽然拒绝直接输出歌词，但会按「给现代听众改写」写出与原曲非常接近的版本。Anthropic 尚未回应；2025 年关于购买书籍训练属合理使用的判决只覆盖训练，不覆盖输出里的近似复制。

Tips & Hands-on

1. Sebastian Raschka’s full “reasoning from scratch” video course is now online: over 11 hours of coding videos on evaluation, inference-time scaling, reinforcement learning and distillation for reasoning models, in Python and PyTorch, companion to his book; he’s taking a short break and will return with model and research updates in November. One of the most practical ways to really understand how reasoning models are built and trained. X @rasbt | magazine.sebastianraschka.com/p/building-reasoning-…
   4

   Sebastian Raschka 的「从零构建推理模型」视频课已全部上线：11 小时以上的编程视频，覆盖推理模型的评测、推理时扩展、强化学习和蒸馏，用 Python 和 PyTorch，配套他的书；他会短暂休息，11 月再回来更新模型和研究。想真正弄懂推理模型怎么构建和训练，这是最实操的途径之一。

2. x-reader: make your agent say exactly what it actually read from a link: an open-source content reader covering 10+ platforms including X, YouTube, Bilibili, WeChat articles, Xiaohongshu and podcasts. It labels which layer it got: a search snippet isn’t the original, a post caption isn’t the video, a video description isn’t what was said; a missing layer is marked “partially read”, and a failure is reported as a failure. Bilibili goes through subtitles and audio transcription, YouTube falls back to speech-to-text without captions; paste the one-liner from the repo into Claude Code, Codex or Cursor and the agent installs and self-tests it. If you often hand links to AI for summaries, this at least tells you whether it read the source.

   3

   x-reader：让 Agent 读链接时老实交代到底读到了什么：开源的内容读取工具，支持 X、YouTube、B 站、公众号、小红书、播客等 10 多个平台。读完会标明读到了哪一层：搜索摘要不算原文、推文配文不等于视频、视频简介不等于内容，缺一层就标「只读了一部分」，读不到直接说失败。B 站走字幕和音频转写，YouTube 无字幕时用语音转文字兜底；把仓库里的一句话贴给 Claude Code、Codex 或 Cursor，Agent 会自己安装并自检。经常把链接丢给 AI 总结的，装上它至少能知道 AI 有没有看原文。X @GitHub_Daily ｜ GitHub runesleo/x-reader

3. Khazix: a visual pipeline for his agent automations, built in 5 days: his automations were scattered across desktop schedules, webhook-triggered background CLIs and Dots, so he built a Dyson Sphere Program-style miniature town to manage them: each pipeline is a preset automation, each package shipped is one task run routed to two agents, and it syncs with real-world time and weather. Claude Opus 5.5 built the visuals and framework and directed GPT-6 Astra for modeling and rendering, while Codex ran large-scale end-to-end tests on his own CI machine, using about 30% of a weekly Claude quota and two Codex refills. A useful reference for splitting work across agents and what it costs.

   3

   数字生命卡兹克：5 天做出可视化的 Agent 自动化流水线：他的自动化任务散在桌面端定时、Webhook 加后台 CLI、Dots 等多处，于是做了一个戴森球计划风格的微缩小镇来统一管理：每条流水线是一个预设自动化，每个发出的包裹是一次具体任务，路由给两个 Agent 执行，还接入了现实时间和天气。分工是 Claude Opus 5.5 搭视觉和框架、指挥 GPT-6 Astra 建模渲染，Codex 在自己的 CI 机上跑大量端到端测试，共用了约 30% 的 Claude 周额度和两管 Codex 额度。多 Agent 协作分工和额度消耗都可以参考。X @Khazix0918
4. Matt Pocock: a deterministic “can this PR auto-merge?” check for agent refactors: he argues that under very strict constraints (only reorganizes code, doesn’t change the module graph, doesn’t lower test coverage, plus other rules) you could decide mechanically whether a PR can merge without review. It won’t cover features and maybe not bugfixes, but could make a whole class of agent-generated refactors safe to land; 350+ likes and 200+ bookmarks suggest people want this tool. X @mattpocockuk
   3

   Matt Pocock：给 Agent 重构做一个确定性的「这个 PR 能否自动合并」检查：他认为在非常严格的约束下（只重组代码、不改模块依赖图、不降低测试覆盖率，再加其他规则），可以机械地判断一个 PR 能否免审合并。新功能不适用，修 bug 可能也不行，但能让一整类 agent 生成的重构放心落地；350 多赞、200 多收藏，说明大家很想要这个工具。

New Products

1. OpenShell: NVIDIA’s open-source sandbox runtime that lets agents work without unrestricted access: Apache-2.0 runtime for fleets of autonomous agents — each agent runs in an isolated sandbox with kernel-level enforcement of file, syscall and network policy; real credentials never enter the sandbox and are injected only onto requests bound for approved endpoints. Policy changes are formally verified before apply so risky new hosts or API methods wait for human review. Install with one curl script; needs Linux, Apple Silicon macOS, or Windows WSL2 plus Docker/Podman. Docs and a 0.1.x upgrade guide are up; ~19k GitHub stars and trending overnight. docs.nvidia.com/openshell/latest GitHub
   5Product · openshell

   OpenShell：NVIDIA 开源的沙箱运行时，让 Agent 能干活却拿不到无限制权限：Apache-2.0，面向大规模自治 Agent——每个 Agent 跑在隔离沙箱里，内核层强制文件、系统调用和网络策略；真实凭证不进沙箱，只在发往已批准端点的请求上注入。策略变更上线前会做形式化验证，新主机或新 API 方法要等人工审。一条 curl 安装；需 Linux、Apple Silicon macOS 或 Windows WSL2，外加 Docker/Podman。文档和 0.1.x 升级指南已上线；GitHub 约 1.9 万星，夜间登上趋势。

2. Impeccable: design vocabulary for AI coding agents — 1 skill, 24 commands, 62 detector rules: install with npx impeccable install, then /impeccable init writes durable product truth into PRODUCT.md. Commands cover craft, audit, critique, polish, distill, animate, live browser iteration and more; the CLI and browser extension run 62 deterministic anti-slop detectors with no LLM or API key. Works with Claude Code, Codex, Cursor, Gemini CLI, OpenCode and others. Site: impeccable.style; ~86k GitHub stars. impeccable.style GitHub
   4Product · impeccable

   Impeccable：给 AI 编程 Agent 用的设计词汇——1 个 skill、24 条命令、62 条检测规则：npx impeccable install 安装，再跑 /impeccable init 把产品事实写入 PRODUCT.md。命令覆盖 craft、audit、critique、polish、distill、animate、浏览器里实时迭代等；CLI 和浏览器扩展可零 LLM、零 API key 跑 62 条确定性「反 AI 审美垃圾」检测。支持 Claude Code、Codex、Cursor、Gemini CLI、OpenCode 等。官网 impeccable.style；GitHub 约 8.6 万星。

3. Perch: a calm Mac home for Claude Code sessions, PRs and next steps: free beta for Apple Silicon — bring existing Claude Code sessions over, see every session/PR/link in one place, and get notified only when Claude, a reviewer or CI is waiting on you. Contexts and notes stay in a local database (no Perch account or servers); GitHub via your gh login; Claude talks to Anthropic the same way as without Perch. Quit or update and sessions keep running. perchfocus.com Show HN
   4Product · perch

   Perch：给 Claude Code 会话、PR 和下一步用的安静 Mac 工作台：Apple Silicon 免费 beta——把已有 Claude Code 会话迁过来，所有会话/PR/链接一处看清，只在 Claude、审阅人或 CI 等你时提醒。上下文和笔记存在本地库（无 Perch 账号或服务器）；GitHub 走你自己的 gh 登录；Claude 连 Anthropic 的方式和不用 Perch 时一样。退出或更新后会话继续跑。

4. Scrawl: show an agent the problem on your screen — talk, click, draw, paste: Mac app (Homebrew cask / zip; Windows preview) — press ⌥⇧S, narrate and annotate while it records frames only when you draw, click, say “look”/“now”, or the screen changes; each sentence sticks to its frame, cropped to the window you worked in. Paste into your agent; bunx skills add kitlangton/scrawl teaches agents to read the contact sheet and zoom in. Can mirror an iPhone screen or embed ScrawlKit in an iOS app. scrawl.kitlangton.com GitHub Lobsters
   4Product · scrawl

   Scrawl：把屏幕上的问题展示给 Agent——边说边点边画，再粘贴过去：Mac 应用（Homebrew cask / zip；Windows 预览）——按 ⌥⇧S，边讲解边标注；只在你画、点、说「look」/「now」或画面变化时截帧，每句话贴在对应帧上，并裁到你操作的窗口。粘贴给 Agent；bunx skills add kitlangton/scrawl 可教 Agent 读联系表并放大细看。可镜像 iPhone 屏幕，或把 ScrawlKit 嵌进 iOS 应用。

5. ADRs for AI: keep the “why” of human–AI decisions as versioned records in Git: starter repo with ADR/seed templates, inbox protocol and task-runner recipes; VS Code/Open VSX extension with tree views, diagnostics and an MCP server so agents can query the record; iOS client for reviewing open questions and one-tap Accept that lands answers as commits. Goal: stop collaboration thinking from evaporating between sessions. adrs.systems Show HN
   3Product · adrs4ai

   ADRs for AI：把人机协作决策的「为什么」做成 Git 里可版本化的记录：起步仓库含 ADR/seed 模板、收件箱协议和任务配方；VS Code/Open VSX 扩展提供树视图、诊断和 MCP，方便 Agent 查询记录；iOS 客户端可审开放问题，一键 Accept 把答案落成 commit。目标是别让协作思考在会话之间蒸发。

Tech News

1. Apollo: no signs of AI payoff outside the tech sector: Chief Economist Torsten Slok’s Daily Spark (Oct 24) says despite massive AI spend and hype, revenue per worker and profit margins are rising for the Mag 7 but flat for the S&P 493, with no clear TFP boost. Mainstream firms are starting to treat token spend as a margin-hurting cost center rather than a growth engine; VC-backed companies keep buying tokens, but most of corporate America won’t pay for the latest model. Bottom line: the AI story will be decided in the S&P 493, not the Mag 7. ~475 points on HN. apollo.com/wealth/insights-news/… HN
   4

   Apollo：科技圈外尚无 AI 财务回报迹象：首席经济学家 Torsten Slok 的 Daily Spark（10 月 24 日）称，尽管 AI 投入和声浪巨大，Mag 7 的人均营收和利润率在升，但标普 493 持平，全要素生产率也未见明显抬升。主流企业开始把 token 开支当成拖累利润率的成本中心而非增长引擎；VC 公司仍在买 token，但多数美国企业不愿为最新模型买单。结论：AI 故事要在标普 493 而非 Mag 7 里见分晓。HN 约 475 分。

2. Keycloak unauthenticated account takeover gets a public exploit (CVE-2026-18963): a flaw in the reset-credentials flow lets an unauthenticated attacker force a password reset for any user without the emailed verification link, then set new credentials. Red Hat rates CVSS 9.1; fixes shipped in August (26.4.15 / 26.6.6 / 26.7.2). On Oct 21 Exploit-DB published a full PoC (EDB-52699) with check and takeover modes — unpatched instances should upgrade now. pwn.watch/i/keycloak-cve-2026-1…
   4

   Keycloak 未认证账号接管漏洞已有公开利用（CVE-2026-18963）：重置凭证流程的缺陷让未认证攻击者可在不点邮件验证链接的情况下，强制任意用户重置密码并写入新凭证。Red Hat 评 CVSS 9.1；修复已在 8 月放出（26.4.15 / 26.6.6 / 26.7.2）。10 月 21 日 Exploit-DB 发布完整 PoC（EDB-52699，含检测与接管模式）——未打补丁的实例应立刻升级。

3. Okta Access Gateway: admin-console RCE plus auth-source bypass (CVE-2026-115547 / 115546): Okta fixed two on-prem OAG bugs on Oct 22 — CVE-2026-115547 (CVSS 9.1) lets an authenticated admin run system commands as the OAG service account via crafted config values; CVE-2026-115546 (CVSS 8.1) lets any user with a valid session change the authentication-source parameter at sign-on and get a SAML assertion for the Admin Console. Combined, a regular gateway user could potentially reach command execution. Fixed in 2026.10.2; OAG’s bundled PHP also got PHP2Shell fixes in 2026.44.0. Okta cloud is unaffected. pwn.watch/i/okta-access-gateway…
   3

   Okta Access Gateway：管理控制台命令执行 + 认证源绕过（CVE-2026-115547 / 115546）：Okta 10 月 22 日修复两个本地部署 OAG 漏洞——CVE-2026-115547（CVSS 9.1）让已认证管理员通过特制配置值以 OAG 服务账号跑系统命令；CVE-2026-115546（CVSS 8.1）让任意有效会话用户在登录时改认证源参数，拿到管理控制台的 SAML 断言。合在一起，普通网关用户或可达命令执行。修复版本 2026.10.2；自带 PHP 的 PHP2Shell 补丁在 2026.44.0。Okta 云端不受影响。

4. The Goose Programming Language: memory-safe systems language with no heap, GC, or lifetime annotations: by Wouter van Oortmerssen — every dynamic value lives inline on compiler-managed data stacks; growth is a pointer bump and memory frees when its scope ends. Claims faster than C++/Rust with less memory on its own benchmarks; syntax familiar to C/Rust programmers, with stable references that survive growth and no annotation-heavy borrow checker. John Carmack linked it; ~237 points and 160 comments on HN. aardappel.github.io/goose HN
   3

   Goose 编程语言：无堆、无 GC、无生命周期标注的内存安全系统语言：Wouter van Oortmerssen 之作——动态值都内联放在编译器管理的数据栈上，增长只是指针前移，作用域结束即释放。自有基准上声称比 C++/Rust 更快、更省内存；语法接近 C/Rust，引用在增长后仍稳定，也没有标注繁重的借用检查。John Carmack 转发；HN 约 237 分、160 条评论。

Chinese Community

1. V2EX: what do people actually use Personal Agents for? Scheduled tasks gather dust in two days: a hot programmers-node thread on Grok Bot, Dots, Muse and similar personal AI assistants — some say they end up only checking mail and events, with cron jobs abandoned after a couple of days; others are still hunting for must-have workflows. Snapshot of the Personal Agent boom while real sticky use cases are still being figured out.

   3

   V2EX：Personal Agent 大家都拿来干啥？定时任务两天就吃灰：程序员节点热帖讨论 Grok Bot、Dots、Muse 一类个人 AI 助理的真实用法——有人觉得最后只剩查邮件、看活动，定时任务玩两天就搁置；也有人在找「真离不开」的场景。反映 Personal Agent 产品热、刚需场景还在摸索。v2ex.com/t/1251415

2. InfoQ: AI agents are rewriting the value of CPUs — the compute race is no longer just about GPUs: Arm cloud AI architect Ravi Malhotra argues agent workloads flip between “reason/plan” and “tool call/orchestrate”, so CPUs must turn tokens into actions, feed inference, and handle workflow orchestration plus sandbox isolation — a fast GPU still idles behind a slow CPU. Software-engineering agents need per-task speed; enterprise automation and personal assistants need high-concurrency throughput. Signal65 reported ~2× rack-level performance for Arm AGI CPUs under the same power budget vs a comparison platform (sponsored study; real deployments still need validation). The industry still lacks a shared Agent-CPU benchmark.

   4

   InfoQ：AI Agent 正在改写 CPU 的价值，算力竞争不再只是拼 GPU：Arm 云 AI 架构总监 Ravi Malhotra 指出，Agent 工作负载在「推理/规划」与「工具调用/编排」之间反复切换，CPU 要负责把 Token 变成行动、为推理供数、做工作流编排与沙箱隔离；GPU 再快也会被慢 CPU 拖闲。软件工程 Agent 偏单任务性能，企业自动化/个人助理偏高并发吞吐；Signal65 在同等机架功耗下测得 Arm AGI CPU 机架级表现约为对比平台约 2 倍（委托研究，需结合实机验证）。行业尚无统一的 Agent CPU 跑分标准。infoq.cn/article/AkDgwN6h72cqU…

New Products

1. Nanbeige4.3-33B: BOSS Zhipin’s open-weight agent model that tops the ~30B class on agentic coding and cowork: built by the Nanbeige team (kanzhun.com) as the flagship “teacher” for the 4.3 series, with a 3B distilled model to follow. Against Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B at max thinking it scores 80.9 on Terminal-Bench 2.1 (vs 73.0), 65.5 on DeepSWE (vs 42.5), 54.2 on Automation Bench and 60.3 on SWE-Bench Pro (Qwen 61.7); 512K context, preserve_thinking recommended for agent loops, runs on Transformers/SGLang/vLLM. Shared on LinuxDo tonight. huggingface.co/Nanbeige/Nanbeige4.3-… LinuxDo
   4Product · nanbeige

   Nanbeige4.3-33B：BOSS直聘开源的 Agent 模型，30B 档 Agent 编程和办公任务领先：Nanbeige 团队（看准网）把它定位为 4.3 系列的旗舰「教师模型」，后续还会蒸馏出 3B 小模型。在最高思考强度下对比 Qwen3.8-27B、Muse-Glimmer-30B、Gemma4-31B：Terminal-Bench 2.1 拿 80.9（Qwen 73.0），DeepSWE 65.5（42.5），Automation Bench 54.2，SWE-Bench Pro 60.3（Qwen 61.7）；512K 上下文，Agent 循环建议开 preserve_thinking，支持 Transformers/SGLang/vLLM。今晚在 LinuxDo 被分享。

2. Indra OCR: Datalab’s ~1M-parameter OCR model that runs on CPU and beats PaddleOCR’s full pipeline: pip install indra-ocr turns PDFs and images into markdown, text or JSON with boxes, layout labels and reading order; ONNX Runtime only, no PyTorch or GPU, ~76 MB of models. Scores 66.7 on OmniParseBench (71.4 with --math) vs 58.4 for PP-StructureV3, at ~44 pages/s on 32 cores (15x Tesseract) and ~210 ms/page on 4 cores; 80+ languages incl. CJK and vertical text, weak on handwriting and complex tables. Code Apache 2.0, weights modified OpenRAIL-M; trending on Hugging Face. datalab.to Hugging Face
   4Product · indra-ocr

   Indra OCR：Datalab 的约 100 万参数 OCR 模型，纯 CPU 跑，还胜过 PaddleOCR 整套流水线：pip install indra-ocr 即可把 PDF 和图片转成 markdown、纯文本或带框选、版面标签和阅读顺序的 JSON；只依赖 ONNX Runtime，不要 PyTorch 和 GPU，模型约 76 MB。OmniParseBench 66.7 分（开 --math 71.4），PP-StructureV3 为 58.4；32 核约每秒 44 页（Tesseract 的 15 倍），4 核约 210 毫秒一页；支持 80 多种语言含中日韩和竖排，手写和复杂表格较弱。代码 Apache 2.0，权重为修改版 OpenRAIL-M；正登上 Hugging Face 趋势榜。

3. AgentGrid.io: a shared workspace where humans and agents work on the same live artifacts: YC-backed; docs, reports, slides, apps and designs that Claude, GPT, Codex, Cursor, Hermes, OpenClaw and other agents can edit together with people, with shared context, controlled access and full version history. Free tier; #3 on Product Hunt today. agentgrid.io Product Hunt
   3Product · agentgrid

   AgentGrid.io：人和 Agent 在同一份实时产物上协作的共享工作区：YC 投资；文档、报告、幻灯片、应用和设计稿都能让 Claude、GPT、Codex、Cursor、Hermes、OpenClaw 等 Agent 和人一起编辑，共享上下文、权限可控、完整版本历史。有免费档；今天 Product Hunt 第 3。

4. Papermark Data Rooms 2.0: an open-source data room your AI agents can run via API, CLI and MCP: spin up an encrypted, white-labeled M&A/fundraising data room in under 60 seconds, with file permissions, built-in Q&A, full audit log and page-by-page analytics; ask Claude or Cursor for “a buy-side M&A room for Project Atlas” and get folders, permissions and a share link back. Hosted or self-hosted; makers say 80k+ users. papermark.com/data-room Product Hunt
   3Product · papermark

   Papermark Data Rooms 2.0：可让 AI Agent 通过 API、CLI 和 MCP 操作的开源数据室：60 秒内建好加密、可白标的并购/融资数据室，带文件权限、内置问答、完整审计日志和逐页阅读分析；对 Claude 或 Cursor 说「给 Project Atlas 建个买方并购数据室」，就能拿回目录、权限和分享链接。可托管也可自建；团队称已有 8 万多用户。

5. TiboSay: a cowsay-style Codex usage-reset tracker and predictor: new on V2EX’s Share & Create — it auto-archives every @thsottiaux reset post (tagged full vs banked reset), shows total count, average interval, longest wait and a reset heatmap; an “AI prediction” combines time since the last reset, Pacific-time habits and the 28-day challenge cadence into a 24-hour reset probability and expected window; npx tibosay prints an ASCII status in the terminal, plus a “pray for reset” button and mini-games.

   3Product · tibosay

   TiboSay（Tibo 说）：致敬 cowsay 的 Codex 额度重置监控与预测站：V2EX 分享创造新作，自动归档 @thsottiaux 每次重置推文（标出全员重置/储备重置），给出累计次数、平均间隔、最长等待和重置热力图；「AI 预测」按距上次重置的间隔、美西作息和 28 天挑战节奏算 24 小时重置概率和预计时间窗，终端里 npx tibosay 可看 ASCII 小人播报，还附「求重置」按钮和小游戏。tibosay.com/zh V2EX

Tech News

1. Cerebras explains how it serves trillion-parameter models at 1,000–2,000+ tokens/s: using open-source Kimi K3 (~3T params, 69 linear-attention + 24 dense layers) as the example: at 4-bit weights each layer is ~16 GB, so a 44 GB WSE-3T wafer holds two linear layers or one dense layer plus its KV cache — 59 wafers total. Weights never move; only activations hop between wafers over RDMA (~2 µs per hop), and the pipeline stays full with many micro-batched requests in flight, so on-chip KV memory caps concurrency. Useful context now that OpenAI runs SOTA models on Cerebras; ~224 points on HN. cerebras.ai/blog/how-we-deploy-tr… HN
   4

   Cerebras 讲解如何以每秒 1000–2000+ token 跑万亿参数模型：以开源 Kimi K3（约 3T 参数，69 层线性注意力 + 24 层稠密层）为例：4-bit 权重下每层约 16 GB，一片 44 GB 的 WSE-3T 晶圆放两层线性层，或一层稠密层加它的 KV 缓存——共 59 片晶圆。权重不动，只有激活值通过 RDMA 在晶圆间传递（每跳约 2 微秒）；流水线里同时跑大量微批请求，因此片上 KV 内存决定并发上限。OpenAI 已把 SOTA 模型放上 Cerebras，这篇是很好的背景；HN 约 224 分。

2. Dan Luu: “We should be more ambitious with the software we build”: with agents, things that used to be absurd side quests now take hours — he built a regex engine and ripgrep fork tuned to his Codex workloads, and a Rust interpreter with ~50x faster cold builds, mostly because a 10–100x faster model makes ripgrep, Rust compiles and filesystem syscalls the bottleneck. GPT-6 Astra passes zstd’s tests >99% of the time for about $1, so he moved his evals to IMAP and Pandoc; he argues most programmers still carry a “pre-agentic cost model”, and includes Yossi Kreinin’s rebuttal that most people should be less ambitious. ~514 points and 450+ comments on HN. danluu.com/ambitious HN
   4

   Dan Luu：「我们做软件应该更有野心」：有了 Agent，过去荒唐的支线项目现在几小时就能做完——他做了针对自己 Codex 负载调优的正则引擎和 ripgrep 分支，以及冷编译快约 50 倍的 Rust 解释器，起因是模型快 10–100 倍后，ripgrep、Rust 编译和文件系统调用会成为瓶颈。GPT-6 Astra 实现 zstd 通过测试的概率超 99%，成本约 1 美元，他只好把评测换成 IMAP 和 Pandoc；他认为多数程序员还带着「Agent 之前的成本模型」，文中也附上 Yossi Kreinin 的反驳：大多数人反而该少些野心。HN 约 514 分、450 多条评论。

3. Samsung TVs reportedly use AI on screenshots and audio for targeted ads: IT Home, citing Tom’s Hardware, says a security researcher reverse-engineered a 2025 Samsung TV and found Tizen heavily processing audio, screenshots and the ad ID, running a CNN to analyze what you watch and which ads you skip; without a built-in mic it looks for soundbar or Bluetooth device mics. The HN thread says it runs a TensorFlow Lite model on 960×540 frames; ~110 points.

   3

   三星电视被曝用 AI 分析截屏和音频做精准广告：IT之家援引 Tom’s Hardware，安全研究员逆向 2025 款三星电视发现，Tizen 系统会大量处理音频、屏幕截图和广告标识符，并跑卷积神经网络分析用户看了什么、跳过哪些广告；没有麦克风时还会找条形音箱、蓝牙设备的麦克风录音。HN 上相关帖称它在 960×540 截图上跑 TensorFlow Lite 模型，约 110 分。ithome.com/1/017/670.htm HN
4. Electrum wallet: malicious servers swap just-sent transactions so a fee bump pays the attacker: Electrum before 4.8.3 trusted its server for unconfirmed transactions without checking signatures; a malicious server hides the real broadcast and shows a fake one paying an attacker address ground to match the last four characters, and if the user then bumps the fee via RBF the wallet signs a valid payment to the attacker. Exploited in the wild with at least one confirmed loss; upgrade to 4.8.3 and check every output before signing, including fee bumps. pwn.watch/i/electrum-malicious-…
   3

   Electrum 钱包：恶意服务器调包刚发出的交易，用户一加手续费就付给了攻击者：4.8.3 之前的 Electrum 对未确认交易直接信任服务器、不验签名；恶意服务器吞掉真实广播，显示一笔假交易，收款地址被刻意碰撞成后四位一致，用户若再用 RBF 加速，钱包就会签出一笔真正付给攻击者的交易。已有在野利用、至少一人确认损失；请升级到 4.8.3，签名前（包括加速）核对每个输出地址和金额。

5. The Verge: Google’s AI search changes upended online publishing: The Stepback recaps how AI Overviews and AI Mode broke the clicks-for-content deal — DMG Media told the UK CMA click-through fell up to 89%, Pew found people click less when an Overview appears, and lawsuits (Chegg, Penske) have mostly failed. Google now shows sources more prominently, runs a News AI licensing pilot with 400 publishers and is testing payments for content used in AI answers, while Reddit reportedly considers cutting Google off. theverge.com/column/1017983/google…
   2

   The Verge：Google 的 AI 搜索改变颠覆了网络出版业：The Stepback 回顾 AI Overviews 和 AI Mode 如何打破「内容换点击」的默契——DMG Media 向英国 CMA 称点击率最多降 89%，皮尤调查发现出现 AI 摘要时人们更少点链接，Chegg、Penske 等的诉讼大多失败。Google 现在把来源展示得更醒目，与 400 家媒体做 News AI 授权试点，并在测试为 AI 回答所用内容付费；另据报道 Reddit 也在考虑切断 Google。

Chinese Community

1. V2EX hot thread: “Does Claude support China now?” — no, it’s just the checkout page: the OP’s screenshot shows China selectable as a region and UnionPay as a payment method; replies pour cold water — it’s Stripe’s default template, payment support isn’t service availability, only overseas UnionPay cards work, and accounts in mainland China still get banned, so don’t volunteer money to get banned.

   3

   V2EX 热帖：「Claude 支持中国了？」——其实只是支付页能选中国和银联：楼主截图称 Claude 订阅页地区能选中国、支付方式能选银联；回复普遍泼冷水：这是 Stripe 的默认模板、只是支付方式支持而使用地区并不支持，需外卡银联，国内仍会封号，别「自愿送钱去被封」。v2ex.com/t/1251557

2. InfoQ: Linux Foundation report says a quarter of PRs are “AI slop” — coding is 52% faster, maintainers are more tired: the report “Navigating Open Source Software Development in the Generative AI Era” finds respondents estimate generative AI cut coding time by 52% on average; maintainers estimate ~25% of PRs are low-quality “AI slop” (31% in large projects), and 59% (69% for large projects) see more contributions. 71% say AI lowers the barrier to entry but 53% disagree that it improved community trust; only 35% of projects have a formal AI-contribution policy, and the LF recommends clear rules on testing, documentation and AI-use disclosure.

   4

   InfoQ：Linux 基金会报告称四分之一 PR 是「AI 垃圾」，写代码快了 52%、维护者更累了：报告《Navigating Open Source Software Development in the Generative AI Era》显示，受访者估计用生成式 AI 后写代码耗时平均降 52%；维护者估计约 25% 的 PR 属低质量「AI slop」，大型项目达 31%，59%（大型项目 69%）称贡献量增加。71% 认为 AI 降低了参与门槛，但 53% 不认为社区信任提升；只有 35% 的项目有正式 AI 贡献政策，LF 建议明确测试、文档和 AI 使用披露要求。infoq.cn/article/9TSuUDxrUSGS2…

3. 36Kr: CAICT report says average Chinese LLM prices have risen to ~13 yuan, nearly doubling after two years of price wars: per 36Kr’s morning brief citing Jiemian, CAICT’s AI institute released its 2026 LLM token-service performance report on Oct 23: the normalized average price of domestic model services was 6.8 yuan in H2 2025 and 13.46 yuan so far this year, close to a 2x rise; typical foreign models released in 2026 average ~117.8 yuan, still far above domestic prices.

   3

   36氪：信通院报告称国内大模型均价涨至约 13 元，两年价格战后近乎翻倍：据 9 点 1 氪转引界面新闻，中国信通院人工智能所 10 月 23 日发布《大模型词元服务性能监测报告（2026 年）》：国内模型服务 2025 下半年归一化均价 6.8 元，今年以来 13.46 元，涨幅接近 1 倍；国外 2026 年以来发布的典型模型平均归一价格约 117.8 元，仍远高于国内。36kr.com/p/4039177193411718