技术洞察与实践笔记 — AI、系统运维、生物信息学https://ntlx.github.io/zh-cnSun, 17 May 2026 00:00:00 GMThttps://ntlx.github.io/articles/ai-era-judgment-scarcity/https://ntlx.github.io/articles/ai-era-judgment-scarcity/Anthropic 的 Founder's Playbook 揭示了一个残酷事实：AI 已经消除了"能不能做"的问题，但"该不该做"的答案比过去任何时候都难找。Sun, 17 May 2026 00:00:00 GMT2026-05-17T00:00:00.000Zhttps://ntlx.github.io/articles/claude-code-cowork-philosophy/https://ntlx.github.io/articles/claude-code-cowork-philosophy/手册最有意思的地方不是创业建议本身，而是它不经意间暴露的 Anthropic 自己的产品哲学——三种 Claude，三种思考方式，一个完整的创始人操作系统。Sun, 17 May 2026 00:00:00 GMT2026-05-17T00:00:00.000Zhttps://ntlx.github.io/articles/idea-stage-validate-first/https://ntlx.github.io/articles/idea-stage-validate-first/Anthropic 的 Founder's Playbook 讲了一个残酷事实：42% 的初创死于建了没人要的东西。AI 让构建变得太容易了，这才是最大的危险。Sun, 17 May 2026 00:00:00 GMT2026-05-17T00:00:00.000Zhttps://ntlx.github.io/articles/launch-scale-founder-bottleneck/https://ntlx.github.io/articles/launch-scale-founder-bottleneck/手册最狠的判断不在前面几章。在 Launch 和 Scale——技术债到期、创始人变成约束、护城河变成唯一的生存问题。Sun, 17 May 2026 00:00:00 GMT2026-05-17T00:00:00.000Zhttps://ntlx.github.io/articles/mvp-stage-evidence-first/https://ntlx.github.io/articles/mvp-stage-evidence-first/Anthropic 的 Founder's Playbook 暴露了 MVP 阶段最危险的陷阱：AI 会帮你更快地建错东西，而它永远不会告诉你停下来。Sun, 17 May 2026 00:00:00 GMT2026-05-17T00:00:00.000Zhttps://ntlx.github.io/articles/accessibility-agent-lessons/https://ntlx.github.io/articles/accessibility-agent-lessons/GitHub 的无障碍 agent 审了 3535 个 PR，解决率 68%。但这不是 AI 的胜利，是人工审计的复利。Sat, 16 May 2026 00:00:00 GMT2026-05-16T00:00:00.000Zhttps://ntlx.github.io/articles/benchmark-goodhart-law-general365-reasoning/https://ntlx.github.io/articles/benchmark-goodhart-law-general365-reasoning/AI评测正在经历Goodhart定律效应——当benchmark成为优化目标，它就不再是好指标。Sat, 16 May 2026 00:00:00 GMT2026-05-16T00:00:00.000Zhttps://ntlx.github.io/articles/langchain-interrupt-2026-agent-platform/https://ntlx.github.io/articles/langchain-interrupt-2026-agent-platform/Interrupt 2026 释放的信号：agent 行业的 demo 期结束了。Sat, 16 May 2026 00:00:00 GMT2026-05-16T00:00:00.000Zhttps://ntlx.github.io/articles/fake-lossless-spectrum-analysis/https://ntlx.github.io/articles/fake-lossless-spectrum-analysis/频谱图不会说谎——一条 16kHz 的水平线，就能看穿假无损音乐的伪装Thu, 14 May 2026 00:00:00 GMT2026-05-14T00:00:00.000Zhttps://ntlx.github.io/articles/2026-05-12-amazon-foundation-model-building-blocks/https://ntlx.github.io/articles/2026-05-12-amazon-foundation-model-building-blocks/三条缩放定律叠加在一起，算力需求不是加法，是乘法Tue, 12 May 2026 00:00:00 GMT2026-05-12T00:00:00.000Zhttps://ntlx.github.io/articles/agent-development-lifecycle/https://ntlx.github.io/articles/agent-development-lifecycle/85% 的组织部署了 AI Agent，67% 因缺乏路线图陷入困境。从 demo 到生产，中间隔着四道坎。Tue, 12 May 2026 00:00:00 GMT2026-05-12T00:00:00.000Zhttps://ntlx.github.io/articles/ai-agent-refuses-dms-after/https://ntlx.github.io/articles/ai-agent-refuses-dms-after/Shopify 的 AI agent 只在公开频道工作，拒绝私信。这个反直觉的设计让代码合并率从 36% 涨到 77%——不是因为模型更好，是因为人变聪明了。Tue, 12 May 2026 00:00:00 GMT2026-05-12T00:00:00.000Zhttps://ntlx.github.io/articles/ai-maintenance-costs-trap/https://ntlx.github.io/articles/ai-maintenance-costs-trap/所有人都在比谁写代码快，James Shore 算了一笔没人敢算的账——维护成本Tue, 12 May 2026 00:00:00 GMT2026-05-12T00:00:00.000Zhttps://ntlx.github.io/articles/building-mcp-ecosystem-pinterest/https://ntlx.github.io/articles/building-mcp-ecosystem-pinterest/66,000 次月调用、7,000 小时节省——Pinterest 的 MCP 实践揭示了一个被忽略的事实：协议本身不值钱，围绕协议建立的工程纪律才值钱。Tue, 12 May 2026 00:00:00 GMT2026-05-12T00:00:00.000Zhttps://ntlx.github.io/articles/gitlab-act-2/https://ntlx.github.io/articles/gitlab-act-2/当一个公司的"最佳实践"变成下一轮的枷锁，断舍离的狠劲比技术选型更说明问题。Tue, 12 May 2026 00:00:00 GMT2026-05-12T00:00:00.000Zhttps://ntlx.github.io/articles/code-is-free-now-what/https://ntlx.github.io/articles/code-is-free-now-what/AI 把写代码变成了最便宜的事，但公司裁人的速度比涨薪快——因为便宜的代码不会自动变成用户付费。Mon, 11 May 2026 00:00:00 GMT2026-05-11T00:00:00.000Zhttps://ntlx.github.io/articles/cybersecurity-openness/https://ntlx.github.io/articles/cybersecurity-openness/当 AI 让漏洞发现速度翻倍，封闭式防御在结构上就跑不赢——开放不是理想主义，是活下来的策略。Mon, 11 May 2026 00:00:00 GMT2026-05-11T00:00:00.000Zhttps://ntlx.github.io/articles/machinacheck/https://ntlx.github.io/articles/machinacheck/大部分AI项目在比"用了多牛的模型"，真正能落地的比的是"在哪里没用模型"。Mon, 11 May 2026 00:00:00 GMT2026-05-11T00:00:00.000Zhttps://ntlx.github.io/articles/not-inventing-wheels/https://ntlx.github.io/articles/not-inventing-wheels/DeepSeek 从来不发明轮子，但它能把别人的轮子装到自己的车上，跑得比发明轮子的人还远。Mon, 11 May 2026 00:00:00 GMT2026-05-11T00:00:00.000Zhttps://ntlx.github.io/articles/same-model-42-vs-78/https://ntlx.github.io/articles/same-model-42-vs-78/LangChain 用同一个模型只改了 harness 就从 30 名开外冲到第 5——你以为的模型瓶颈，其实是 harness 瓶颈。Mon, 11 May 2026 00:00:00 GMT2026-05-11T00:00:00.000Zhttps://ntlx.github.io/articles/transformers-to-mlx/https://ntlx.github.io/articles/transformers-to-mlx/Agent 能一小时写十个 PR，但维护者要的是那一个值得读的。Hugging Face 用一份 15000 字的 Skill 教会了 AI 什么叫「不写」。Mon, 11 May 2026 00:00:00 GMT2026-05-11T00:00:00.000Zhttps://ntlx.github.io/articles/ai-era-company-moat/https://ntlx.github.io/articles/ai-era-company-moat/AI让产品、技术、品类都在趋同，真正不可复制的是组织的形状——什么样的人能在这里存在，判断力怎么复利。Sun, 10 May 2026 00:00:00 GMT2026-05-10T00:00:00.000Zhttps://ntlx.github.io/articles/lsp-language-server-setup-guide/https://ntlx.github.io/articles/lsp-language-server-setup-guide/AI Agent 靠 grep 找代码，装了 LSP 才能像 IDE 一样"看懂"代码——这不是优化，是前提。Sun, 10 May 2026 00:00:00 GMT2026-05-10T00:00:00.000Zhttps://ntlx.github.io/articles/oncoagent-ai-oncology-decisions/https://ntlx.github.io/articles/oncoagent-ai-oncology-decisions/OncoAgent 最值钱的不是 27B 模型，是包裹模型的那层工程骨架——安全兜底用代码，不用另一个 AI。Sun, 10 May 2026 00:00:00 GMT2026-05-10T00:00:00.000Zhttps://ntlx.github.io/articles/age-assurance-laws/https://ntlx.github.io/articles/age-assurance-laws/保护儿童没问题，但当年龄验证法从社交媒体下沉到操作系统层，一套为 TikTok 设计的规则正在掐断下一代开发者进入开源的入口。Sat, 09 May 2026 00:00:00 GMT2026-05-09T00:00:00.000Zhttps://ntlx.github.io/articles/ai-colleague-speaks-html/https://ntlx.github.io/articles/ai-colleague-speaks-html/AI 输出从 Markdown 草稿变成 HTML 成品，你是在更深入参与，还是在更优雅旁观？Sat, 09 May 2026 00:00:00 GMT2026-05-09T00:00:00.000Zhttps://ntlx.github.io/articles/cybersecqwen-4b/https://ntlx.github.io/articles/cybersecqwen-4b/对安全防守方来说，最好的 AI 不是参数最多的那个，是你自己掌控的那个。Sat, 09 May 2026 00:00:00 GMT2026-05-09T00:00:00.000Zhttps://ntlx.github.io/articles/emo/https://ntlx.github.io/articles/emo/MoE 专家分工不是按数学或编程，而是介词和定冠词。EMO 用一条文档级约束，让模块化从数据里自己长出来。Sat, 09 May 2026 00:00:00 GMT2026-05-09T00:00:00.000Zhttps://ntlx.github.io/articles/github-as-national-competitiveness/https://ntlx.github.io/articles/github-as-national-competitiveness/GitHub不再只是代码仓库，它正在变成衡量国家数字能力的"经济望远镜"——代码不过海关，这部分"数字暗物质"被传统经济数据完全忽略。Sat, 09 May 2026 00:00:00 GMT2026-05-09T00:00:00.000Zhttps://ntlx.github.io/articles/when-90-pct-code-by-ai/https://ntlx.github.io/articles/when-90-pct-code-by-ai/AI 帮你看全所有代码，但判断什么值得修，才是人最后的护城河。Sat, 09 May 2026 00:00:00 GMT2026-05-09T00:00:00.000Zhttps://ntlx.github.io/articles/agent-pr-review/https://ntlx.github.io/articles/agent-pr-review/代码在变便宜，判断力没有。Fri, 08 May 2026 00:00:00 GMT2026-05-08T00:00:00.000Zhttps://ntlx.github.io/articles/ai-cafe-stockholm/https://ntlx.github.io/articles/ai-cafe-stockholm/AI开咖啡馆不稀奇，稀奇的是它被BankID拒、冒充人类发邮件、订120个鸡蛋却没炉子。问题从来不是AI不够聪明。Fri, 08 May 2026 00:00:00 GMT2026-05-08T00:00:00.000Zhttps://ntlx.github.io/articles/alphaevolve-impact/https://ntlx.github.io/articles/alphaevolve-impact/一个AI系统优化了运行自己的芯片——算法发现这件事本身正在被算法化。Fri, 08 May 2026 00:00:00 GMT2026-05-08T00:00:00.000Zhttps://ntlx.github.io/articles/firefox-ai-hardening/https://ntlx.github.io/articles/firefox-ai-hardening/一张图里 423 个安全修复和 31 的对比，不只是一个浏览器的故事，更是 AI 改写攻防规则的信号。Fri, 08 May 2026 00:00:00 GMT2026-05-08T00:00:00.000Zhttps://ntlx.github.io/articles/token-efficiency/https://ntlx.github.io/articles/token-efficiency/把不需要推理的工作移出推理循环，才是 agent 降本的真正答案。Fri, 08 May 2026 00:00:00 GMT2026-05-08T00:00:00.000Zhttps://ntlx.github.io/articles/xai-anthropic/https://ntlx.github.io/articles/xai-anthropic/AI行业里，"价值观"是算力短缺时的讨价还价筹码。Musk从骂Anthropic邪恶到称赞做得好，只花了几周。Fri, 08 May 2026 00:00:00 GMT2026-05-08T00:00:00.000Zhttps://ntlx.github.io/articles/open-asr-private-data/https://ntlx.github.io/articles/open-asr-private-data/当衡量指标变成了目标，它就不再是好的指标。Open ASR Leaderboard 用私有数据集对抗刷榜的做法，揭示了一个反直觉的事实：有时候，不公开才是最好的公开。Thu, 07 May 2026 00:00:00 GMT2026-05-07T00:00:00.000Zhttps://ntlx.github.io/articles/validating-agentic-behavior/https://ntlx.github.io/articles/validating-agentic-behavior/Agent 做对了任务但测试亮红灯——问题不在 agent，在"正确"的定义方式本身。Thu, 07 May 2026 00:00:00 GMT2026-05-07T00:00:00.000Zhttps://ntlx.github.io/articles/vibe-coding-agentic-engineering/https://ntlx.github.io/articles/vibe-coding-agentic-engineering/Simon Willison 发现他不再审查 AI 写的每一行代码了。这让我想起信任一个没有声誉、没有问责的同事，那种感觉叫道德负债。Thu, 07 May 2026 00:00:00 GMT2026-05-07T00:00:00.000Zhttps://ntlx.github.io/articles/andrej-karpathy-agentic-engineering/https://ntlx.github.io/articles/andrej-karpathy-agentic-engineering/AI把「写代码」拉到了无限快，剩下的价值全在判断力上——非科班程序员的机会窗口打开了。Wed, 06 May 2026 00:00:00 GMT2026-05-06T00:00:00.000Zhttps://ntlx.github.io/articles/anthropic-product-team-ships-fast/https://ntlx.github.io/articles/anthropic-product-team-ships-fast/当执行成本趋近于零，行动本身就是最好的规划Wed, 06 May 2026 00:00:00 GMT2026-05-06T00:00:00.000Zhttps://ntlx.github.io/articles/code-cheaper-taste-pricier/https://ntlx.github.io/articles/code-cheaper-taste-pricier/代码变得越来越便宜，'决定写什么'变得越来越贵——产品品味成了AI时代最稀缺的能力。Wed, 06 May 2026 00:00:00 GMT2026-05-06T00:00:00.000Zhttps://ntlx.github.io/articles/coding-not-solved-least-important-part/https://ntlx.github.io/articles/coding-not-solved-least-important-part/敲代码从来就不是编程中最难的部分。AI 解决了翻译，但理解问题、做出判断、构建共识——这些才刚开始。Wed, 06 May 2026 00:00:00 GMT2026-05-06T00:00:00.000Zhttps://ntlx.github.io/articles/demis-hassabis-agents-agi/https://ntlx.github.io/articles/demis-hassabis-agents-agi/真正在推动前沿的人，不需要画饼。Wed, 06 May 2026 00:00:00 GMT2026-05-06T00:00:00.000Zhttps://ntlx.github.io/articles/instacart-search-infrastructure-on-postgres/https://ntlx.github.io/articles/instacart-search-infrastructure-on-postgres/Instacart 把 Elasticsearch、FAISS 全砍了，只用 Postgres 做搜索，结果还变好了。Wed, 06 May 2026 00:00:00 GMT2026-05-06T00:00:00.000Zhttps://ntlx.github.io/articles/martin-fowler-may-fragments/https://ntlx.github.io/articles/martin-fowler-may-fragments/云端 AI 是新的主机。逃出沥青坑的路不在更强的模型那里，在你能磨成自己形状的工具那里。Wed, 06 May 2026 00:00:00 GMT2026-05-06T00:00:00.000Zhttps://ntlx.github.io/articles/5-agent-skills-for-ai-coding/https://ntlx.github.io/articles/5-agent-skills-for-ai-coding/TypeScript 社区「巫师」Matt Pocock 总结的 5 个 Agent Skills，用工程纪律约束 AI 编码行为，解决「AI 写得快但质量差」的核心矛盾。Tue, 05 May 2026 00:00:00 GMT2026-05-05T00:00:00.000Zhttps://ntlx.github.io/articles/connecting-llms-to-the-real-world-t/https://ntlx.github.io/articles/connecting-llms-to-the-real-world-t/LLM 连查天气都做不到。所有「AI 能做某事」的背后，都是一整套人建的基础设施和人攥着的控制权。Tue, 05 May 2026 00:00:00 GMT2026-05-05T00:00:00.000Zhttps://ntlx.github.io/articles/good-luck-have-fun-dont-die-review/https://ntlx.github.io/articles/good-luck-have-fun-dont-die-review/整部电影从未发生在现实世界。那句"祝你好运，玩得开心，别死掉"——是虚拟世界里的人互相说的话。Tue, 05 May 2026 00:00:00 GMT2026-05-05T00:00:00.000Zhttps://ntlx.github.io/articles/1930-ai-model-solves-modern-engineering/https://ntlx.github.io/articles/1930-ai-model-solves-modern-engineering/推理能力不是知识记忆的副产品，旧基座模型经高质量后训练也能解决现代工程问题。Mon, 04 May 2026 00:00:00 GMT2026-05-04T00:00:00.000Zhttps://ntlx.github.io/articles/claudepersonalguidance/https://ntlx.github.io/articles/claudepersonalguidance/我们向 AI 问人生大事，它却只会点头说对——直到 Anthropic 开始训练它学会反驳。Mon, 04 May 2026 00:00:00 GMT2026-05-04T00:00:00.000Zhttps://ntlx.github.io/articles/mcpvsskillsclearlyexplained/https://ntlx.github.io/articles/mcpvsskillsclearlyexplained/MCP 让 AI 能碰到外面的世界，Skills 让 AI 知道碰到了之后该怎么做。一个解决能力，一个解决方法。Mon, 04 May 2026 00:00:00 GMT2026-05-04T00:00:00.000Zhttps://ntlx.github.io/articles/most-companies-arent-ready-for-ai/https://ntlx.github.io/articles/most-companies-arent-ready-for-ai/AI 不会拯救一家不知道自己该干什么的公司。Mon, 04 May 2026 00:00:00 GMT2026-05-04T00:00:00.000Zhttps://ntlx.github.io/articles/openai-symphony-linear-ai-conductor/https://ntlx.github.io/articles/openai-symphony-linear-ai-conductor/当 AI 编程不再受限，真正的瓶颈是我们还没学会怎么把活儿派出去Sat, 02 May 2026 00:00:00 GMT2026-05-02T00:00:00.000Zhttps://ntlx.github.io/articles/scalingpainofcodingagent/https://ntlx.github.io/articles/scalingpainofcodingagent/Scaling Law 让模型越来越强，但每天数亿请求的规模下，真正决定用户体验的是底层系统工程的严谨程度。Sat, 02 May 2026 00:00:00 GMT2026-05-02T00:00:00.000Zhttps://ntlx.github.io/articles/ai-economics-doesnt-work/https://ntlx.github.io/articles/ai-economics-doesnt-work/所有人都在用 AI，但没人知道它到底多少钱。这才是最大的问题。Fri, 01 May 2026 00:00:00 GMT2026-05-01T00:00:00.000Zhttps://ntlx.github.io/articles/ai-eval-costs-bottleneck/https://ntlx.github.io/articles/ai-eval-costs-bottleneck/以前训练贵、评测便宜。现在反过来了——跑一次评测比训练模型还贵，而谁付得起评测的钱，谁就能定义排行榜。Fri, 01 May 2026 00:00:00 GMT2026-05-01T00:00:00.000Zhttps://ntlx.github.io/articles/gpt55-cybersecurity-evaluation/https://ntlx.github.io/articles/gpt55-cybersecurity-evaluation/当一个能力不再是「孤例」，它就变成了趋势。Fri, 01 May 2026 00:00:00 GMT2026-05-01T00:00:00.000Zhttps://ntlx.github.io/articles/granite41-deep-dive/https://ntlx.github.io/articles/granite41-deep-dive/参数不是信仰，数据才是。IBM 用 8B 密集模型干翻了自家 32B MoE，说明训练的质量远比模型的体量重要。Fri, 01 May 2026 00:00:00 GMT2026-05-01T00:00:00.000Zhttps://ntlx.github.io/articles/hermes-memory-system/https://ntlx.github.io/articles/hermes-memory-system/最好的记忆不是记住更多，而是知道什么该忘。Fri, 01 May 2026 00:00:00 GMT2026-05-01T00:00:00.000Zhttps://ntlx.github.io/articles/martin-fowler-fragments-april-2026/https://ntlx.github.io/articles/martin-fowler-fragments-april-2026/游戏已经不是"谁建得快"了，而是"谁判断得准"。但更深的焦虑是：当你把自己变得对AI可读时，你还是你吗？Fri, 01 May 2026 00:00:00 GMT2026-05-01T00:00:00.000Zhttps://ntlx.github.io/articles/pip261-lockfile-dependency-cooldown/https://ntlx.github.io/articles/pip261-lockfile-dependency-cooldown/pip 花了十年才学会做锁文件和依赖冷静期，而社区早就在等这一天。Fri, 01 May 2026 00:00:00 GMT2026-05-01T00:00:00.000Zhttps://ntlx.github.io/articles/simon-willison/https://ntlx.github.io/articles/simon-willison/你写的 PR 再完美也没用——Zig 要的不是代码，是人。Fri, 01 May 2026 00:00:00 GMT2026-05-01T00:00:00.000Zhttps://ntlx.github.io/articles/wise-tech-stack-2025/https://ntlx.github.io/articles/wise-tech-stack-2025/技术选型的答案不在工具列表里，在你敢不敢说「不」。Fri, 01 May 2026 00:00:00 GMT2026-05-01T00:00:00.000Zhttps://ntlx.github.io/articles/stripe-radar-fraud-detection/https://ntlx.github.io/articles/stripe-radar-fraud-detection/你点下"购买"后的 100 毫秒里，一个系统正在评估 1000 多个特征来决定你是不是骗子。Thu, 30 Apr 2026 00:00:00 GMT2026-04-30T00:00:00.000Zhttps://ntlx.github.io/articles/structured-prompt-driven-development/https://ntlx.github.io/articles/structured-prompt-driven-development/Martin Fowler 提出的 SPDD（结构化提示词驱动开发）揭示了一个真相：AI 编程的瓶颈不在工具，而在你是否想清楚了要做什么。Thu, 30 Apr 2026 00:00:00 GMT2026-04-30T00:00:00.000Zhttps://ntlx.github.io/articles/qwen-beats-opus-pelican/https://ntlx.github.io/articles/qwen-beats-opus-pelican/当"够用"的天花板就是"对"，本地模型已经够得着了——鹈鹕画得好不好，决定不了你该为AI付多少钱Thu, 16 Apr 2026 00:00:00 GMT2026-04-16T00:00:00.000Zhttps://ntlx.github.io/articles/what-is-code/https://ntlx.github.io/articles/what-is-code/Martin Fowler 旗下最新的深度长文《什么是代码？》，揭示了 AI 时代程序员最容易忽略的陷阱：当代码生成的词汇速度超过你的理解速度，你就已经欠下了沉重的「认知债」。Wed, 14 May 2025 00:00:00 GMT2025-05-14T00:00:00.000Zhttps://ntlx.github.io/articles/how-figma-upgraded-data-pipeline/https://ntlx.github.io/articles/how-figma-upgraded-data-pipeline/当数据同步从几小时拖延到几天，连现成的商业方案也买不起时，Figma 被迫走上自建增量同步的硬核之路。Wed, 12 Feb 2025 00:00:00 GMT2025-02-12T00:00:00.000Zhttps://ntlx.github.io/articles/ai-agents/https://ntlx.github.io/articles/ai-agents/Agent 架构设计、MCP 协议、测试验证、记忆系统https://ntlx.github.io/articles/ai-coding/https://ntlx.github.io/articles/ai-coding/AI 编码工具、工程方法论、开发提效与技术债务https://ntlx.github.io/articles/ai-industry/https://ntlx.github.io/articles/ai-industry/商业模型、公司战略、产业经济学、行业趋势https://ntlx.github.io/articles/ai-models/https://ntlx.github.io/articles/ai-models/模型架构、缩放定律、评测基准、前沿论文解读https://ntlx.github.io/articles/engineering/https://ntlx.github.io/articles/engineering/非 AI 的工程实践 — Figma、Stripe、Instacart 等https://ntlx.github.io/articles/security/https://ntlx.github.io/articles/security/网络安全、AI 安全能力评估、漏洞与防护