宝玉@doteyAI score7171SemiAnalysis 实测 Anthropic、OpenAI 等九家 AI 订阅套餐后得出，同样 200 美元，Claude 订阅折算的 Token 用量约为 OpenAI 的 5 倍。
Awaiting translation
Anthropic Subscriptions Offer 5x+ More Value Than OpenAI Limit testing every AI subscription plan from Anthropic, OpenAI, Meta, SpaceXAI, MiniMax, Moonshot, Zdotai, Cursor, and Cognition https://newsletter.semianalysis.com/p/anthropic-subscriptions-offer-5x
宝玉@doteyAI score7171据 The Information 10 月 5 日报道，Meta 和微软都在减少员工内部使用 Anthropic 的 Claude，转向自家模型和工具。
Awaiting translation
🚨BREAKING: Microsoft and META are aggressively cutting employee use of Claude ahead of Anthropic's IPO Microsoft has cut internal claude spend by more than 33%, nuked per-employee token budget from $100k/month to $10k/month, and forced Copilot to auto-route to cheaper models META used Claude code to build Muse, then cut active users from 60,000 to 30,000 (50% decline) after launch, and replaced it with Muse Code Palantir and Nvidia are also scaling back claude over soaring prices and data privacy fears it’s OVER…
宝玉@doteyAI score6262OpenAI 宣布未来几周内对欧盟用户在 ChatGPT 和 Codex 生成的文字加入不可见水印，以满足欧盟《人工智能法案》要求，API 用户可在全球范围自行开启，默认关闭。
Awaiting translation
We're expanding our approach to content provenance to include text in response to EU regulatory requirements, while recognizing the significant limitations of current text watermarking technology. Our tools already help verify whether an image or audio file was created with our models. This work builds on those efforts to help people better understand when content may have been generated or edited with an OpenAI model. In the EU, we’ll start watermarking eligible text from ChatGPT and Codex over the coming weeks to comply with the EU AI Act. Customers using our API can turn on text watermarking for select models worldwide today.
宝玉@doteyAI score7070amontlabs/lcu 把 Codex 的 Computer Use 单独拆出，让 Claude Code、Codex CLI、Pi 等 Agent 工具通过 MCP 调用。
Awaiting translation
发现一个牛逼的东西，让任意 Agent 调用 Codex 的 Computer Use。 Codex 最强的就是 Computer Use。 但最近用 Claude Opus 5.5比较多，这样就强强联合了。 刚测试通过，安装后建议配置个 Hook，指定白名单可控制哪些 App 安装地址见评论区
AI Hero Skills v1.3 is out, adding three skills—/implement-spec, /pr, and /retro—and extending the main flow from /grill-with-docs → /to-spec → /to-tickets into implementation, PRs, and retros.
Why it matters: The author rounds out the skill set into a full path from writing specs to PRs and retros, and lays out the trade-offs at each step along with the rough edges they already know about.
宝玉@doteySelectedAI score8282i had a lot of fun chatting with @mattpocockuk today about how i was able to land 2,500 PRs last month! Matt is a wonderful interviewer so i think the interview turned out really interesting both of our skill plugins work great together, so i recommend giving both a try and picking the best skills that suit your workflow https://www.youtube.com/watch?v=MN9dGgmLyso
Why it matters: Using Lauren Tan's practice of merging 2500 PRs in a month, Baoyu explains that skipping individual reviews rests on validation Skills and rule constraints, and lays out the conditions under which he'd apply the same approach.
Google 发布 Gemini 4 Argon，其 Argon 智能体团队通过分析全集群性能遥测数据自主识别并应用内存优化，上线后释放超 300 TiB 内存，预计总节省 500 TiB 至 1 PiB。
Awaiting translation
宝玉@doteyAI score6666Awaiting translation
New in The Atlantic: @dgrobinson resigned this week. He was among the longest-tenured employees at OpenAI—and oversaw safety reports on 12 frontier launches. He is very worried: “The time for trial and error is over.” You can read his essay here: https://www.theatlantic.com/technology/2026/10/openai-safety-team-resignation/688881/?gift=1ga2TvL-DbuHDQIcYF7oR4o908Fsjxr4NFLlsptkfP8
宝玉@doteyAI score6262剑桥大学 AI 科学与政策项目（CASP）9 月发布一篇论文，Hinton、Bengio 等 20 多位研究者联名讨论 AI 研发被自动化后是否会引发智能爆炸，Hinton 在 X 上推荐了该论文。
Awaiting translation
The idea of an intelligence explosion caused by recursive self improvement has been around for a long time but until very recently it did not seem imminent. Now many leading researchers think it may happen quite soon. You can read our paper about it here: https://casp.ac/reports/intelligence-explosion
Boris Cherny@bchernyAI score6060Awaiting translation
Claude Tag in Slack can now use your personal connectors! You can now securely access that Drive doc, Salesforce account, or Warehouse table that you have personal access to right where the work happens. Avail. on Teams today and Enterprise next week https://claude.com/blog/claude-tag-now-supports-personal-connectors-in-channels
Boris Cherny@bchernyAI score5858Boris Cherny 用 Opus 5.5 配合 Lean 对 Claude Agent SDK 做形式化验证，几段简短提示词换来 16 个 PR，修复了多个 bug 和竞态条件。
Awaiting translation
Lovable has shipped Opus 5.5, which the company says matches Opus 5 in results while cutting the number of steps by one-third to one-half. On Lovable's internal benchmarks, Opus 5.5 ties Opus 5 on 0-to-1 builds and iterative code changes, and comes out 4% to 6% ahead on validation discipline; across all reasoning effort levels, steps per task drop by 26% to 57% and input tokens fall by 21% to 59%, with the differences significant at the 95% confidence level.
Why it matters: Lovable shares official comparison data between Opus 5.5 and Opus 5 on step counts and tokens, so readers can judge the real change in build efficiency.
Cline 把 VS Code 扩展从约 76,000 行单体核心迁移到 Cline SDK，并自建灰度发布机制：一个安装包内打包 loader、legacy 和 next 两套扩展，由 PostHog 功能开关按百分比决定激活哪套，崩溃时自动回退到 legacy，开关可随时降到 0% 作为 kill switch。
Awaiting translation
Why it matters: Cline 官方复盘如何把 1100 万用户的 VS Code 扩展迁到新 harness，含灰度机制与前后指标对比。
斯坦福大学研究人员主导、Terminal-Bench 团队联合全球科研机构专家打造的 Terminal-Bench-Science 0.1 发布，首批含生命、物理、地球、数学和工程科学领域的 70 项任务。
Awaiting translation
AI Hero Skills v1.2 发布，整套 Skill 打包为 Claude Code 插件，并为每个 SKILL.md 添加 Codex 元数据，使同一套 Skill 在 Claude Code 和 Codex 中通用。
Awaiting translation
Andrej Karpathy@karpathyAI score6262Awaiting translation
The Terminal-Bench team releases Terminal-Bench 3.0, whose first version spans 7 domains and 74 tasks, with the strongest model passing about 34%. Building on Terminal-Bench 2.1, this release broadens task diversity and adds CI/CD, semantic versioning, and result migration to keep improving the benchmark.
Why it matters: Terminal-Bench 3.0 rebuilds the benchmark with 74 tasks and CI/CD-based versioning, so readers can see how the new benchmark separates models.
Lovable 在内部搭建了一套进攻性安全程序，让 AI 智能体集群像人类攻击者一样探测系统入口，直到拿到可验证的漏洞证据。它用夺旗赛的思路做验证：把 flag 散布在基础设施和权限最高的产品界面中，不预埋任何漏洞，智能体取到 flag 就说明找到了真实入侵路径，而不是模型猜测。
Awaiting translation
Why it matters: Lovable 公开了用夺旗机制验证漏洞的内部攻防智能体编排方法，可迁移到自家安全测试流程。
Lovable 上线 agent integrations，任何公开已发布的 Lovable 应用都能添加 MCP server，从而被 ChatGPT、Claude 等 AI 工具直接调用。
Awaiting translation
Why it matters: Lovable 把已发布应用接入 MCP，读者可据此判断自家应用如何被 ChatGPT、Claude 直接调用。
Andrej Karpathy@karpathyAI score6060Awaiting translation
Introducing Claude Tag, a new way for teams to work with Claude. In Slack, Claude joins as a team member with access to the channels and tools you choose. Tag Claude in and delegate tasks to it while you focus on other work.
Andrej Karpathy@karpathyAI score7575Andrej Karpathy 评价 Claude Fable 5 发布，指出它与 Mythos 是同一底层模型，只是增加了安全防护，在几乎所有基准上以明显优势达到 SOTA。
Awaiting translation
Fable 5 is state-of-the-art on nearly all tested benchmarks, with exceptional performance in software engineering, knowledge work, scientific research, and vision. The longer and more complex the task, the larger Fable 5’s lead over our other models.
Thorsten Ball 的 Joy & Curiosity 通讯在写了三年后进入两三周、最多四周的暑期休更。本期链接包括 Anthropic 发布的《When AI builds itself》文档、Ted Chiang 关于 AI 是否具有意识的文章，以及 Ladybird 浏览器宣布不再接受公开 pull request，理由是 AI 工具已迅速改变开源信任的经济学。
Awaiting translation
Inside Claude Code, Anthropic has already built up hundreds of Skills in active use. The team sorts them into nine categories—library and API references, product validation, data fetching and analysis, business process automation, code scaffolding, code quality and review, CI/CD and deployment, runbooks, and infrastructure operations—and notes that the best Skills should fall cleanly into one of them.
Why it matters: Anthropic’s internal framework for categorizing hundreds of Skills, along with its experience writing them, can carry over to a team building its own Skill library.
Anthropic has shipped dynamic workflows in Claude Code. Claude can write its own harness on the fly for a specific task, and these workflows can be shared and reused. Workflows orchestrate subagents through functions like agent(), parallel(), and pipeline(), and you can specify which model each agent uses and whether it runs in its own worktree. If a session is interrupted, resuming it picks up where it left off.
Why it matters: The Anthropic team breaks down six orchestration patterns for dynamic workflows and where each one fits, and these patterns carry over to multi-agent task design.
Augment wires the Incident Investigator expert from its internal Cosmos platform into Slack and PagerDuty. It automatically triages every alert and runs root-cause analysis, then suggests one of four actions: fix the code, roll back, upgrade, or just keep monitoring. Humans only review the RCA and make the call.
Why it matters: Augment has shared the full playbook for putting Cosmos Expert on alert triage, along with a month of before-and-after data, so you can adapt it to your own on-call process.
Amp 本周宣布成立 Amp Labs，团队已与多家公司合作，新阶段从澳大利亚悉尼起步。Amp 官网同期发布《Software After Software》，阐述其对软件未来的判断以及 Amp 与 Amp Labs 的存在理由。联合创始人还做客 Mayank Gupta 播客，聊了自己如何进入编程、从练 Vim 到成为 Amp 联合创始人的经历。
Awaiting translation
Anthropic 工程师 Thariq Shihipar 提出用 HTML 替代 Markdown 作为 Claude Code 的输出格式，理由是 HTML 信息密度更高、更易阅读和分享，还能做双向交互。
Awaiting translation
Why it matters: Anthropic 工程师分享用 HTML 替代 Markdown 承载 Claude Code 输出的做法，附常见场景的提示词与模板。
Augment Code 在模型选择器中新增 Prism，由一个小而快的规划模型在每个用户轮次前决定用哪个底层模型，官方称相比前沿推理模型每任务成本低约 20-30%，质量差异可忽略。
Awaiting translation
In AGENTS.md, Augment Code front-loads roughly 2.5k characters of Karpathy-style coding rules, then runs 40 OpenClaw PRs through Auggie, Claude Code, and Codex for comparison.
Why it matters: A head-to-head test of three coding agents on the same set of PRs shows that prompt constraints mainly cut costs rather than improve quality, and it also surfaces differences between the harnesses.
Claude Opus 4.7 已接入 Lovable，Lovable 基准测试显示其日常任务效率明显提升。相比 Opus 4.6，它完成任务所需轮次减少 40%，token 用量减少 10–20%，速度提升 15%，性能得分高 2–3%。对开发者而言，这意味着更快的迭代和更少的来回返工。
Awaiting translation
Anthropic 的 Thariq Shihipar 复盘了 Claude Code 工具设计中的取舍，核心主张是工具要贴合模型自身能力，而判断能力边界只能靠观察输出和反复实验。
Awaiting translation
Why it matters: Anthropic 工程师复盘 Claude Code 工具设计的取舍，给出可迁移到自建智能体的判断方法。
Lovable 的基础设施团队公开了其 LLM 供应商负载均衡方案，用于在峰值每分钟超过十亿 token 的流量下避免“model provider unavailable”。
Awaiting translation
Why it matters: Lovable 公开了每分钟十亿 token 规模下的多供应商负载均衡方案，可借鉴其用 PID 控制器和项目级粘性保住 prompt caching 的做法。
Anthropic 将 Terminal-Bench 列为 Claude 4 模型卡七项基准之一，Claude 4 Opus 在 Terminal-Bench-Core 上取得 43.2% 的 SOTA 成绩。Dario Amodei 在 Code with Claude 主题演讲中也提及该基准。Terminal-Bench 团队表示将在未来几天验证 Claude 4 的表现并更新官方排行榜。
Awaiting translation