The result shows cheap small-model passes can beat LLM rerankers when split into selection and ranking stages, cutting model-call latency without losing recall on multi-hop questions. Teams building Graph RAG on Milvus can copy the two-stage pattern directly from the Zilliz repo. Compared with LLM-based reranking, the trade-off is a measurable recall gain for under a second of extra call time.
GitHub 的 Martin Woodward 聊了 AI 时代怎么学新东西：放开玩但给自己掐个表，做开发的朋友可以听听。
Hugging Face 花 80 万 GPU 小时做实验，把 MoE 训练的经验全写出来了，7 万词免费开放，还带能跑的代码，搞训练的直接收藏。
博主用 Claude Code 自动做完一整期 AI 日报视频，配音用 Gemini 3.8 Flash TTS，口型用 OmniHuman 1.5，一次只花几块钱，流程值得抄。
一场真实面试题：怎么接手一家技术债缠身的SaaS团队？作者把2-3个月的解题步骤拆得很细，连AI Coding怎么落地都讲了。
Screen Studio 中文转写一直拉胯，这位作者用国产本地模型加豆包云模型把它修好了，后面还要加倍速和 AI 配乐，做视频的可以看看。
一位产品工程师总结的 11 条设计清单，从字号到圆角都有具体规则，还能直接丢给 Agent 帮你 Review 代码仓库。
作者用 NewMax 的 Agent-im 搭了人生 Coach、阅读教练这些生活化 Agent，结果比效率工具用得还勤，思路挺有意思。
深圳 GEO 大会的两位嘉宾提前认识一下，一个讲企业怎么真正用 AI 落地，一个讲海外营销增长，11 月 21 日线下见。
有开发者实测用 Perplexity Teammates 和 GrokBot 配合跑销售线索流程，Notion 做共享记忆，思路可以抄。
Reviewing AI-generated code by reading every line no longer scales; the shift is toward humans auditing machine-produced verification artifacts like screenshots and recordings instead. This matters to engineering leads and solo developers deciding which AI changes are safe to merge, and the screenshot-evidence trick works with any agent workflow. The practice comes from Anthropic's own pipeline, which grew daily merges over tenfold in four months.
博主晒出用 Opus 做视频的真实账单：日均 5 亿 Token，卸载了剪映，纯靠代码渲染加视频生成，1 个月 100 万播放。想用 AI 做视频号的可以抄作业。
Vercel CEO 给出的独立开发者五步走：X 订阅起步，Cursor 里写代码一键部署 Vercel，X 上分发再用 X Money 收钱。
Cursor 的 Agent SDK 文档每页加 .md 就能变 Markdown，这条提示词直接扔给智能体就能跑通第一个项目，省去自己翻文档。
宝玉分享的逆向套路：先让 Claude 跑一遍 Skill，再让它照着复刻一个，能还原个大概，做 Agent 开发的可以试试。
The handbook distills hard-won engineering lessons on MoE efficiency that labs usually keep internal, making sparse-model training accessible to smaller teams. Engineers planning distributed pretraining or fine-tuning runs can use its profiling guidance and interactive figures to diagnose bottlenecks without burning hundreds of thousands of GPU-hours themselves. It complements Hugging Face's earlier open training playbooks rather than vendor whitepapers.
MoE architectures power many frontier open-weight models, yet training guidance at scale is scarce, so a step-by-step playbook with visual explanations fills a real gap for engineers. Practitioners planning pretraining or fine-tuning runs can use it alongside the team's earlier Ultra Scale Playbook. No regional angle applies; the closest comparison is existing distributed-training documentation from PyTorch and DeepSpeed.
有人用 Cursor Agent SDK 写了个 Slack 机器人，自动抓 X 反馈让智能体修，作者团队真在用这套流程分诊意见，思路可以抄。
The real shift is architectural: orchestrators no longer block on subagent completion, so chat-style agents can run long tasks without freezing the user interface. Developers building multi-step agents on LangChain should review the pattern and consider moving slow subtasks to background subagents. It contrasts with synchronous subagent designs common in other agent frameworks.
Figma 的 CFO 说团队一两个月建了 40-50 个 agents，面试时专考你会不会被 AI 替代思考，财务人值得听
Formal methods have rarely been applied to code that agents write, and PDD Kit shows a workflow where the tool both verifies and honestly reports coverage gaps instead of guessing. Developers building agent pipelines that ship unreviewed code can use the kit to set machine-checked intent specs. Microsoft's open position here contrasts with runtime-only testing approaches used by most coding assistant vendors.
Notion 出了个官方配置指南，几句话就能让 ChatGPT 或 Claude 把你的 Notion 当记忆库用，其他智能体粘贴提示词也行。
作者拿 Codex 和 Claude Code 对比，说 computer use 能直接操作那些人类都用不明白的老软件，比走 API/MCP 快多了。
Ethan Mollick 拿一句提示词测 Claude，Claude 真的在回复里把他 Rickroll 了，你可以复制这句去玩玩自己常用的模型。
OpenRouter 出了个 cookbook，教你把文档丢给 agent，让它自己学会调 OpenRouter 的接口，搞自动化的可以看看。
Figma 财务团队用 AI 复刻竞品来验证收购标的值不值钱，还造了 AI 版高管来审备忘录，这套打法挺新鲜的。
11月10日在SF，FactoryAI的David Roster手把手带你用真实项目练怎么安全把活儿丢给agents，想学 delegates 流程的可以看。
教你不改模型就能让 Haiku 干 Opus 5.5 的活：59 秒替代 7 分钟，token 省一半，思路挺实用。
有人嫌 ScreenStudio 导出慢、中文字幕差，直接拿开源的 Cap 改了个自用版，思路挺值得参考的。
The core implication is that official benchmark scores are increasingly unreliable signals for model selection, and community-run alternative evals are becoming the de facto quality check. Builders choosing models for production should cross-check lab claims against independent altbenches before committing to an API or fine-tuning pipeline. Following Artificial Analysis's rebuilds offers a practical sniff test without waiting for lab corrections.
Sourcegraph 的 Andrew Norrish 讲了怎么用代码搜索配智能体来查问题、修问题，评审跟不上生成速度的痛点很戳人。
有人拿腾讯 Hy4 的 API 做了个可拆装的 3D 耳机，Blender 里两次追问就搞定，17 个任务实测结果贴出来了。
有人用 GPT Image 2.6 一句话就把自己头像变成了吃米饭的“真香”GIF 表情包，提示词都贴出来了，照抄就能做自己的表情包。
Only the latest 2000 items are listed; for older ones see the daily archive
