<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>汉松札记 - AI Highlight</title><description>围绕 AI、技术领导力和高信号材料的阅读高亮与判断。</description><link>https://blog.xiaohansong.com/</link><item><title>Community Wisdom 299：先寻找真实拉力，再把反馈变成复利</title><link>https://blog.xiaohansong.com/posts/ai-highlight/community-wisdom-299-first-users-interview-debrief-career-compounding/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/community-wisdom-299-first-users-interview-debrief-career-compounding/</guid><description>这期三组讨论分别谈产品尚未完成时怎样找第一批用户、怎样复盘自己的求职面试，以及职业早期应该优化什么。</description><pubDate>Sun, 20 Sep 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>AI 产品与商业</category><category>职业与成长</category></item><item><title>Community Wisdom 298：先找痛点与瓶颈，再决定用 AI 还是换工具</title><link>https://blog.xiaohansong.com/posts/ai-highlight/community-wisdom-298-ai-product-validation-platform-metrics/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/community-wisdom-298-ai-product-validation-platform-metrics/</guid><description>这期四组讨论表面上分散在儿童篮球 AI 教练、Jira 与 Linear 迁移、平台团队的商业论证，以及工作中的 AI 用法，底层却在回答同一个问题：当一种新能力看起来可以加速工作时，怎样确认它击中了真正的约束。</description><pubDate>Sun, 13 Sep 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>AI 产品与商业</category><category>组织与工作流</category></item><item><title>AI 转型的瓶颈是习惯、激励与决策权</title><link>https://blog.xiaohansong.com/posts/ai-highlight/community-wisdom-200-ai-adoption-habits-incentives/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/community-wisdom-200-ai-adoption-habits-incentives/</guid><description>AI 工具部署完成后，团队采用率取决于真实任务收益、同伴示范、人机边界、试错成本和管理预期。</description><pubDate>Sun, 06 Sep 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>组织与工作流</category><category>人机协作</category></item><item><title>需求从文档变成团队能力之后，所有权反而要更清楚</title><link>https://blog.xiaohansong.com/posts/ai-highlight/community-wisdom-199-engineer-ownership-presales-pricing/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/community-wisdom-199-engineer-ownership-presales-pricing/</guid><description>这期最值得留下的是第一组讨论：工程师能否在 PM 没有给出清晰需求时直接开始构建。</description><pubDate>Sun, 30 Aug 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>组织与工作流</category><category>AI 产品与商业</category></item><item><title>AI 加速处理之后，判断系统必须升级</title><link>https://blog.xiaohansong.com/posts/ai-highlight/ai-%E5%8A%A0%E9%80%9F%E5%A4%84%E7%90%86%E4%B9%8B%E5%90%8E-%E5%88%A4%E6%96%AD%E7%B3%BB%E7%BB%9F%E5%BF%85%E9%A1%BB%E5%8D%87%E7%BA%A7/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/ai-%E5%8A%A0%E9%80%9F%E5%A4%84%E7%90%86%E4%B9%8B%E5%90%8E-%E5%88%A4%E6%96%AD%E7%B3%BB%E7%BB%9F%E5%BF%85%E9%A1%BB%E5%8D%87%E7%BA%A7/</guid><description>这期 Community Wisdom 的四组讨论指向同一个变化：AI 和自动化降低了生成、筛选、整理与表达的成本，系统瓶颈随之转向判断。</description><pubDate>Sun, 23 Aug 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>人机协作</category><category>组织与工作流</category></item><item><title>代码记录现实，文档保存意图</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E4%BB%A3%E7%A0%81%E8%AE%B0%E5%BD%95%E7%8E%B0%E5%AE%9E-%E6%96%87%E6%A1%A3%E4%BF%9D%E5%AD%98%E6%84%8F%E5%9B%BE/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E4%BB%A3%E7%A0%81%E8%AE%B0%E5%BD%95%E7%8E%B0%E5%AE%9E-%E6%96%87%E6%A1%A3%E4%BF%9D%E5%AD%98%E6%84%8F%E5%9B%BE/</guid><description>这期 Community Wisdom 最值得留下的几组讨论，都在区分表面问题与真正的约束。</description><pubDate>Sun, 16 Aug 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>软件设计与开发工具</category><category>组织与工作流</category></item><item><title>生成与构建变便宜之后，具体判断成为稀缺品</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E7%94%9F%E6%88%90%E4%B8%8E%E6%9E%84%E5%BB%BA%E5%8F%98%E4%BE%BF%E5%AE%9C%E4%B9%8B%E5%90%8E-%E5%85%B7%E4%BD%93%E5%88%A4%E6%96%AD%E6%88%90%E4%B8%BA%E7%A8%80%E7%BC%BA%E5%93%81/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E7%94%9F%E6%88%90%E4%B8%8E%E6%9E%84%E5%BB%BA%E5%8F%98%E4%BE%BF%E5%AE%9C%E4%B9%8B%E5%90%8E-%E5%85%B7%E4%BD%93%E5%88%A4%E6%96%AD%E6%88%90%E4%B8%BA%E7%A8%80%E7%BC%BA%E5%93%81/</guid><description>这期 Community Wisdom 的四组讨论指向同一个变化：AI 正在把生成文字、构建产品和制作材料的成本压低，通用产出因此迅速贬值。</description><pubDate>Sun, 09 Aug 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>AI 产品与商业</category><category>组织与工作流</category></item><item><title>能力变便宜之后，系统瓶颈决定价值</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E8%83%BD%E5%8A%9B%E5%8F%98%E4%BE%BF%E5%AE%9C%E4%B9%8B%E5%90%8E-%E7%B3%BB%E7%BB%9F%E7%93%B6%E9%A2%88%E5%86%B3%E5%AE%9A%E4%BB%B7%E5%80%BC/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E8%83%BD%E5%8A%9B%E5%8F%98%E4%BE%BF%E5%AE%9C%E4%B9%8B%E5%90%8E-%E7%B3%BB%E7%BB%9F%E7%93%B6%E9%A2%88%E5%86%B3%E5%AE%9A%E4%BB%B7%E5%80%BC/</guid><description>这期 Community Wisdom 最值得留下的内容来自四组讨论：如何选择开放权重模型、如何安排一次高成本商务旅行、如何从混杂客户中识别 ICP，以及产品交付速度超过市场承接速度时怎么办。</description><pubDate>Sun, 02 Aug 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>AI 编程</category><category>组织与工作流</category></item><item><title>问 AI 能不能做之前，先问系统会不会接受</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E9%97%AE-ai-%E8%83%BD%E4%B8%8D%E8%83%BD%E5%81%9A%E4%B9%8B%E5%89%8D-%E5%85%88%E9%97%AE%E7%B3%BB%E7%BB%9F%E4%BC%9A%E4%B8%8D%E4%BC%9A%E6%8E%A5%E5%8F%97/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E9%97%AE-ai-%E8%83%BD%E4%B8%8D%E8%83%BD%E5%81%9A%E4%B9%8B%E5%89%8D-%E5%85%88%E9%97%AE%E7%B3%BB%E7%BB%9F%E4%BC%9A%E4%B8%8D%E4%BC%9A%E6%8E%A5%E5%8F%97/</guid><description>这期 Community Wisdom 的主线很清楚：产出一个可用结果，和让结果穿过组织、市场与责任边界，是两件不同的事。</description><pubDate>Sun, 02 Aug 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>AI 产品与商业</category><category>组织与工作流</category></item><item><title>AI 交付需要一条可验证的证据链</title><link>https://blog.xiaohansong.com/posts/ai-highlight/ai-%E4%BA%A4%E4%BB%98%E9%9C%80%E8%A6%81%E4%B8%80%E6%9D%A1%E5%8F%AF%E9%AA%8C%E8%AF%81%E7%9A%84%E8%AF%81%E6%8D%AE%E9%93%BE/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/ai-%E4%BA%A4%E4%BB%98%E9%9C%80%E8%A6%81%E4%B8%80%E6%9D%A1%E5%8F%AF%E9%AA%8C%E8%AF%81%E7%9A%84%E8%AF%81%E6%8D%AE%E9%93%BE/</guid><description>这期 Community Wisdom 的四组讨论看似分散，实际都在处理同一个变化：AI 大幅压低了产出代码、界面和内部工具的成本，真正稀缺的部分随之转向意图传递、结果验证和责任承担。</description><pubDate>Sun, 19 Jul 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>评测与可观测性</category><category>AI 编程</category></item><item><title>执行成本下降之后，瓶颈转向治理</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E6%89%A7%E8%A1%8C%E6%88%90%E6%9C%AC%E4%B8%8B%E9%99%8D%E4%B9%8B%E5%90%8E-%E7%93%B6%E9%A2%88%E8%BD%AC%E5%90%91%E6%B2%BB%E7%90%86/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E6%89%A7%E8%A1%8C%E6%88%90%E6%9C%AC%E4%B8%8B%E9%99%8D%E4%B9%8B%E5%90%8E-%E7%93%B6%E9%A2%88%E8%BD%AC%E5%90%91%E6%B2%BB%E7%90%86/</guid><description>这期 Community Wisdom 收录了六组社区讨论。</description><pubDate>Tue, 14 Jul 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>组织与工作流</category><category>AI 产品与商业</category></item><item><title>负网络效应：当激励开始毒化社区</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E8%B4%9F%E7%BD%91%E7%BB%9C%E6%95%88%E5%BA%94-%E5%BD%93%E6%BF%80%E5%8A%B1%E5%BC%80%E5%A7%8B%E6%AF%92%E5%8C%96%E7%A4%BE%E5%8C%BA/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E8%B4%9F%E7%BD%91%E7%BB%9C%E6%95%88%E5%BA%94-%E5%BD%93%E6%BF%80%E5%8A%B1%E5%BC%80%E5%A7%8B%E6%AF%92%E5%8C%96%E7%A4%BE%E5%8C%BA/</guid><description>这期 Community Wisdom 的几组讨论都涉及同一个底层问题：系统中的参与者会根据真实激励行动，而团队经常只设计流程，行为后果成为被忽略的部分。</description><pubDate>Tue, 14 Jul 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>AI 产品与商业</category><category>组织与工作流</category></item><item><title>100 工具 Agent 是个陷阱</title><link>https://blog.xiaohansong.com/posts/ai-highlight/100-%E5%B7%A5%E5%85%B7-agent-%E6%98%AF%E4%B8%AA%E9%99%B7%E9%98%B1/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/100-%E5%B7%A5%E5%85%B7-agent-%E6%98%AF%E4%B8%AA%E9%99%B7%E9%98%B1/</guid><description>这是一场来自 AI Engineer 频道的演讲，主题是&quot;100 工具 Agent 陷阱&quot;：当你给 AI Agent 一次性提供过多工具时，准确率会随着工具数量增长而急剧下降。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>Agent 系统</category><category>上下文工程</category></item><item><title>Agent 构建 Agent</title><link>https://blog.xiaohansong.com/posts/ai-highlight/agent-%E6%9E%84%E5%BB%BA-agent/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/agent-%E6%9E%84%E5%BB%BA-agent/</guid><description>本演讲来自 AI Engineer 频道，主题为&quot;Agent 构建 Agent&quot;——介绍如何借助编程辅助工具与规格驱动开发，通过 AutoAgent 循环和实时数据分析流水线，迭代构建可靠、安全的 AI Agent。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>Harness Engineering</category><category>评测与可观测性</category></item><item><title>Agent 化 AI 工程师</title><link>https://blog.xiaohansong.com/posts/ai-highlight/agentic-ai-%E5%B7%A5%E7%A8%8B%E5%B8%88/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/agentic-ai-%E5%B7%A5%E7%A8%8B%E5%B8%88/</guid><description>本演讲介绍 Mutagent 团队如何用 Agent 化方式构建 AI Agent 生命周期管理系统，覆盖规格说明、构建、评测、部署、诊断、优化六个阶段，并现场演示了评测 Agent 和诊断 Agent 两款处于研究预览阶段的产品。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>Harness Engineering</category><category>评测与可观测性</category></item><item><title>AI 系统设计：从想法到生产</title><link>https://blog.xiaohansong.com/posts/ai-highlight/ai-%E7%B3%BB%E7%BB%9F%E8%AE%BE%E8%AE%A1-%E4%BB%8E%E6%83%B3%E6%B3%95%E5%88%B0%E7%94%9F%E4%BA%A7/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/ai-%E7%B3%BB%E7%BB%9F%E8%AE%BE%E8%AE%A1-%E4%BB%8E%E6%83%B3%E6%B3%95%E5%88%B0%E7%94%9F%E4%BA%A7/</guid><description>来自 MongoDB 的开发者倡导者 Apoorva Joshi 分享了一套四阶段 AI 系统设计框架：产品需求、系统设计、评估与监控、优化。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>Agent 系统</category><category>AI 产品与商业</category></item><item><title>HTML 就够了：让 Agent 生成图形</title><link>https://blog.xiaohansong.com/posts/ai-highlight/html-%E5%B0%B1%E5%A4%9F%E4%BA%86-%E8%AE%A9-agent-%E7%94%9F%E6%88%90%E5%9B%BE%E5%BD%A2/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/html-%E5%B0%B1%E5%A4%9F%E4%BA%86-%E8%AE%A9-agent-%E7%94%9F%E6%88%90%E5%9B%BE%E5%BD%A2/</guid><description>这场来自 AI Engineer 频道的短讲提出一个观点：让 Agent 生成幻灯片、文档、视频等图形化产物时，关键不是强迫它操作画布或直接写 SVG 坐标，而是给它使用更符合语言模型思维方式的媒介。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>Agent 系统</category><category>AI 产品与商业</category></item><item><title>Miranda 假说：Hamilton 如何毒化角色评测</title><link>https://blog.xiaohansong.com/posts/ai-highlight/miranda-%E5%81%87%E8%AF%B4-hamilton-%E5%A6%82%E4%BD%95%E6%AF%92%E5%8C%96-persona-%E8%AF%84%E6%B5%8B/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/miranda-%E5%81%87%E8%AF%B4-hamilton-%E5%A6%82%E4%BD%95%E6%AF%92%E5%8C%96-persona-%E8%AF%84%E6%B5%8B/</guid><description>这场演讲提出“Miranda 假说”：当文化作品覆盖历史记录时，角色扮演系统会生成流畅但时代错位的复合人格，而现有角色评测往往只衡量声音和人格一致性，无法发现这种失败。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>评测与可观测性</category></item><item><title>提示词就是平台</title><link>https://blog.xiaohansong.com/posts/ai-highlight/prompt-%E5%B0%B1%E6%98%AF%E5%B9%B3%E5%8F%B0/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/prompt-%E5%B0%B1%E6%98%AF%E5%B9%B3%E5%8F%B0/</guid><description>Dominik Tornow 认为，当 Agent 能够按需根据规范生成面向具体基础设施的实现时，软件平台的价值会从通用实现转向可复用的规范。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>上下文工程</category><category>AI 产品与商业</category></item><item><title>从研究到现实：把前沿 ML 研究带入生产</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E4%BB%8E%E7%A0%94%E7%A9%B6%E5%88%B0%E7%8E%B0%E5%AE%9E-%E6%8A%8A%E5%89%8D%E6%B2%BF-ml-%E7%A0%94%E7%A9%B6%E5%B8%A6%E5%85%A5%E7%94%9F%E4%BA%A7/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E4%BB%8E%E7%A0%94%E7%A9%B6%E5%88%B0%E7%8E%B0%E5%AE%9E-%E6%8A%8A%E5%89%8D%E6%B2%BF-ml-%E7%A0%94%E7%A9%B6%E5%B8%A6%E5%85%A5%E7%94%9F%E4%BA%A7/</guid><description>Higharc 的高级研究工程师 Vaidas Razgaitis 分享了他们如何把前沿 AI/ML 研究推进到生产环境：用研究原型分类文档提升研究可读性，用 Python 单体代码仓库和解耦微服务承接原型。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>AI 产品与商业</category><category>评测与可观测性</category></item><item><title>你无法提示整个房间：AI 最后无法取代的技能</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E4%BD%A0%E6%97%A0%E6%B3%95%E6%8F%90%E7%A4%BA%E6%95%B4%E4%B8%AA%E6%88%BF%E9%97%B4-ai-%E6%9C%80%E5%90%8E%E6%97%A0%E6%B3%95%E5%8F%96%E4%BB%A3%E7%9A%84%E6%8A%80%E8%83%BD/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E4%BD%A0%E6%97%A0%E6%B3%95%E6%8F%90%E7%A4%BA%E6%95%B4%E4%B8%AA%E6%88%BF%E9%97%B4-ai-%E6%9C%80%E5%90%8E%E6%97%A0%E6%B3%95%E5%8F%96%E4%BB%A3%E7%9A%84%E6%8A%80%E8%83%BD/</guid><description>这场演讲的核心判断是：当 AI 让写代码变得便宜之后，软件开发真正稀缺的能力，转向判断该做什么，以及把业务方、用户、决策者拉到同一个房间里，提炼出值得构建的需求。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>人机协作</category><category>组织与工作流</category></item><item><title>你的 Agent 在生产环境失败了。祝你好运复现它</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E4%BD%A0%E7%9A%84-agent-%E5%9C%A8%E7%94%9F%E4%BA%A7%E7%8E%AF%E5%A2%83%E5%A4%B1%E8%B4%A5%E4%BA%86-%E7%A5%9D%E4%BD%A0%E5%A5%BD%E8%BF%90%E5%A4%8D%E7%8E%B0%E5%AE%83/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E4%BD%A0%E7%9A%84-agent-%E5%9C%A8%E7%94%9F%E4%BA%A7%E7%8E%AF%E5%A2%83%E5%A4%B1%E8%B4%A5%E4%BA%86-%E7%A5%9D%E4%BD%A0%E5%A5%BD%E8%BF%90%E5%A4%8D%E7%8E%B0%E5%AE%83/</guid><description>这场来自 AI Engineer 频道的演讲讨论了 AI Agent 在生产环境中失败后为什么很难复现，以及为什么团队应该把目标从追求模型的逐位确定性（bitwise determinism）。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>评测与可观测性</category><category>Agent 系统</category></item><item><title>你的 Agent 正在浪费 Token，而你还不知道</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E4%BD%A0%E7%9A%84-agent-%E6%AD%A3%E5%9C%A8%E6%B5%AA%E8%B4%B9-token-%E8%80%8C%E4%BD%A0%E8%BF%98%E4%B8%8D%E7%9F%A5%E9%81%93/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E4%BD%A0%E7%9A%84-agent-%E6%AD%A3%E5%9C%A8%E6%B5%AA%E8%B4%B9-token-%E8%80%8C%E4%BD%A0%E8%BF%98%E4%B8%8D%E7%9F%A5%E9%81%93/</guid><description>AWS 开发者布道师 Eric Hanchett 介绍了降低 Agent token 成本的五种做法：缓存系统提示词，按任务难度路由到不同模型，把大型工具结果移出上下文，限制工具循环次数，以及裁剪多轮对话历史。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>上下文工程</category><category>Agent 系统</category></item><item><title>别再写语气指令了。要分层组织它们</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E5%88%AB%E5%86%8D%E5%86%99%E8%AF%AD%E6%B0%94%E6%8C%87%E4%BB%A4%E4%BA%86-%E8%A6%81%E5%88%86%E5%B1%82%E7%BB%84%E7%BB%87%E5%AE%83%E4%BB%AC/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E5%88%AB%E5%86%8D%E5%86%99%E8%AF%AD%E6%B0%94%E6%8C%87%E4%BB%A4%E4%BA%86-%E8%A6%81%E5%88%86%E5%B1%82%E7%BB%84%E7%BB%87%E5%AE%83%E4%BB%AC/</guid><description>演讲者 Isadora Martin Dye 分享了她在经营婚礼场地、并为场地构建 AI Agent 的过程中总结出的四层提示词架构：不可变身份层、情境模式层、示例锚定语气层、生成后否决层。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>上下文工程</category><category>AI 产品与商业</category></item><item><title>在生产工作流中使用规格驱动开发</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E5%9C%A8%E7%94%9F%E4%BA%A7%E5%B7%A5%E4%BD%9C%E6%B5%81%E4%B8%AD%E4%BD%BF%E7%94%A8%E8%A7%84%E6%A0%BC%E9%A9%B1%E5%8A%A8%E5%BC%80%E5%8F%91/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E5%9C%A8%E7%94%9F%E4%BA%A7%E5%B7%A5%E4%BD%9C%E6%B5%81%E4%B8%AD%E4%BD%BF%E7%94%A8%E8%A7%84%E6%A0%BC%E9%A9%B1%E5%8A%A8%E5%BC%80%E5%8F%91/</guid><description>AWS 高级开发者布道师 Erik Hanchett 介绍规格驱动开发的概念、动机和实践方式，并以 AWS 的 AI 编程工具 Kiro 为例，演示如何在生产工作流中用规格文档引导 AI 编程助手构建更高质量的代码。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>AI 编程</category><category>Harness Engineering</category></item><item><title>当所有上下文都重要：扩展型缓存增强生成</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E5%BD%93%E6%89%80%E6%9C%89%E4%B8%8A%E4%B8%8B%E6%96%87%E9%83%BD%E9%87%8D%E8%A6%81-%E6%89%A9%E5%B1%95%E5%9E%8B%E7%BC%93%E5%AD%98%E5%A2%9E%E5%BC%BA%E7%94%9F%E6%88%90/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E5%BD%93%E6%89%80%E6%9C%89%E4%B8%8A%E4%B8%8B%E6%96%87%E9%83%BD%E9%87%8D%E8%A6%81-%E6%89%A9%E5%B1%95%E5%9E%8B%E7%BC%93%E5%AD%98%E5%A2%9E%E5%BC%BA%E7%94%9F%E6%88%90/</guid><description>Luis Romero Sevilla 讨论了一个约束很强的知识表示场景：所有文档都与用户问题相关，文档之间关系密集，并且数据会频繁整体替换。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>上下文工程</category><category>AI 基础设施</category></item><item><title>手中的 OpenClaw：构建一个实体 AI 终端</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E6%89%8B%E4%B8%AD%E7%9A%84-openclaw-%E6%9E%84%E5%BB%BA%E4%B8%80%E4%B8%AA%E5%AE%9E%E4%BD%93-ai-%E7%BB%88%E7%AB%AF/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E6%89%8B%E4%B8%AD%E7%9A%84-openclaw-%E6%9E%84%E5%BB%BA%E4%B8%80%E4%B8%AA%E5%AE%9E%E4%BD%93-ai-%E7%BB%88%E7%AB%AF/</guid><description>这场来自 AI Engineer 频道的演讲介绍了 Lech Kalinowski 如何构建一个实体 AI 终端 OpenClaw：它用 OLED 与电子纸双屏、ESP32 微控制器、本地后端和开源 LLM。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>Agent 系统</category><category>AI 产品与商业</category></item><item><title>我们用本地代码索引减少了 94% 的 AI 编程 token</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E6%88%91%E4%BB%AC%E7%94%A8%E6%9C%AC%E5%9C%B0%E4%BB%A3%E7%A0%81%E7%B4%A2%E5%BC%95%E5%87%8F%E5%B0%91%E4%BA%86-94-%E7%9A%84-ai-coding-token/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E6%88%91%E4%BB%AC%E7%94%A8%E6%9C%AC%E5%9C%B0%E4%BB%A3%E7%A0%81%E7%B4%A2%E5%BC%95%E5%87%8F%E5%B0%91%E4%BA%86-94-%E7%9A%84-ai-coding-token/</guid><description>Rajkumar Sakthivel 分享了他们如何通过本地代码索引层减少 AI 编程过程中的输入上下文，把每次查询发送给模型的 token 大幅压缩，同时用混合搜索、结果压缩、调用关系追踪和评分过滤维持代码检索准确率。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>AI 编程</category><category>上下文工程</category></item><item><title>把 10,994 条笔记变成记忆</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E6%8A%8A-10-994-%E6%9D%A1%E7%AC%94%E8%AE%B0%E5%8F%98%E6%88%90%E8%AE%B0%E5%BF%86/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E6%8A%8A-10-994-%E6%9D%A1%E7%AC%94%E8%AE%B0%E5%8F%98%E6%88%90%E8%AE%B0%E5%BF%86/</guid><description>这是一场来自 AI Engineer 频道的演讲或访谈，主题围绕“把 10,994 条笔记变成记忆”。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>上下文工程</category></item><item><title>日志就是 Agent</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E6%97%A5%E5%BF%97%E5%B0%B1%E6%98%AF-agent/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E6%97%A5%E5%BF%97%E5%B0%B1%E6%98%AF-agent/</guid><description>Ishaan Sehgal 主张，Agent 的真正身份不在模型、运行时或执行环境里，而在只追加的会话日志里。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>评测与可观测性</category><category>Agent 系统</category></item><item><title>有失忆症的天才</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E6%9C%89%E5%A4%B1%E5%BF%86%E7%97%87%E7%9A%84%E5%A4%A9%E6%89%8D/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E6%9C%89%E5%A4%B1%E5%BF%86%E7%97%87%E7%9A%84%E5%A4%A9%E6%89%8D/</guid><description>Nx 的 Victor Savkin 用“有失忆症的天才”比喻 AI Agent 的两大核心局限：只能看到代码库的一小部分（空间限制），以及每次会话重新开始时都没有跨会话记忆（时间限制）。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>上下文工程</category><category>Agent 系统</category></item><item><title>未来属于领域专用 Agent</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E6%9C%AA%E6%9D%A5%E5%B1%9E%E4%BA%8E%E9%A2%86%E5%9F%9F%E4%B8%93%E7%94%A8-agent/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E6%9C%AA%E6%9D%A5%E5%B1%9E%E4%BA%8E%E9%A2%86%E5%9F%9F%E4%B8%93%E7%94%A8-agent/</guid><description>这是一场来自 AI Engineer 频道的演讲，演讲者 Justin Schroeder 来自 StandardAgents，主题围绕&quot;未来属于领域专用 Agent&quot;——为什么通用 Agent 通过继承式堆叠上下文终将遇到瓶颈。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>Agent 系统</category><category>AI 产品与商业</category></item><item><title>构建一个自主工程组织</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E6%9E%84%E5%BB%BA%E4%B8%80%E4%B8%AA%E8%87%AA%E4%B8%BB%E5%B7%A5%E7%A8%8B%E7%BB%84%E7%BB%87/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E6%9E%84%E5%BB%BA%E4%B8%80%E4%B8%AA%E8%87%AA%E4%B8%BB%E5%B7%A5%E7%A8%8B%E7%BB%84%E7%BB%87/</guid><description>Block 工程主管 Angie Jones 分享如何在两年内将 3500 人的工程组织转型为自主式工程组织：从定义成熟度模型、发起 AI Champions 计划、实现代码库 AI 就绪，到多 Agent 并行与跨代码库编排器。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>Harness Engineering</category><category>组织与工作流</category></item><item><title>构建优秀 Agent 技能：缺失的手册</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E6%9E%84%E5%BB%BA%E4%BC%98%E7%A7%80-agent-%E6%8A%80%E8%83%BD-%E7%BC%BA%E5%A4%B1%E7%9A%84%E6%89%8B%E5%86%8C/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E6%9E%84%E5%BB%BA%E4%BC%98%E7%A7%80-agent-%E6%8A%80%E8%83%BD-%E7%BC%BA%E5%A4%B1%E7%9A%84%E6%89%8B%E5%86%8C/</guid><description>这是来自 AI Engineer 频道的一场演讲，主题是如何构建高质量的 Agent 技能。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>Harness Engineering</category><category>上下文工程</category></item><item><title>构建系统，而不是代码</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E6%9E%84%E5%BB%BA%E7%B3%BB%E7%BB%9F-%E8%80%8C%E4%B8%8D%E6%98%AF%E4%BB%A3%E7%A0%81/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E6%9E%84%E5%BB%BA%E7%B3%BB%E7%BB%9F-%E8%80%8C%E4%B8%8D%E6%98%AF%E4%BB%A3%E7%A0%81/</guid><description>核心论点是：Agent 的构建模块变了，但工程纪律没有变。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>Harness Engineering</category><category>AI 编程</category></item><item><title>浏览器 Agent 需要的不是更好的模型，而是更好的眼睛</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E6%B5%8F%E8%A7%88%E5%99%A8-agent-%E9%9C%80%E8%A6%81%E7%9A%84%E4%B8%8D%E6%98%AF%E6%9B%B4%E5%A5%BD%E7%9A%84%E6%A8%A1%E5%9E%8B-%E8%80%8C%E6%98%AF%E6%9B%B4%E5%A5%BD%E7%9A%84%E7%9C%BC%E7%9D%9B/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E6%B5%8F%E8%A7%88%E5%99%A8-agent-%E9%9C%80%E8%A6%81%E7%9A%84%E4%B8%8D%E6%98%AF%E6%9B%B4%E5%A5%BD%E7%9A%84%E6%A8%A1%E5%9E%8B-%E8%80%8C%E6%98%AF%E6%9B%B4%E5%A5%BD%E7%9A%84%E7%9C%BC%E7%9D%9B/</guid><description>Kushan Raj 认为，浏览器 Agent 的主要瓶颈不在于模型能力，而在于模型看到网页、理解网页状态变化、规划长序列操作的环境表示太差。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>Agent 系统</category><category>上下文工程</category></item><item><title>生产环境中的 Agent：OpenGov 如何构建并扩展 OG Assist</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E7%94%9F%E4%BA%A7%E7%8E%AF%E5%A2%83%E4%B8%AD%E7%9A%84-agent-opengov-%E5%A6%82%E4%BD%95%E6%9E%84%E5%BB%BA%E5%B9%B6%E6%89%A9%E5%B1%95-og-assist/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E7%94%9F%E4%BA%A7%E7%8E%AF%E5%A2%83%E4%B8%AD%E7%9A%84-agent-opengov-%E5%A6%82%E4%BD%95%E6%9E%84%E5%BB%BA%E5%B9%B6%E6%89%A9%E5%B1%95-og-assist/</guid><description>OpenGov 工程师 Gabe de Mesa 介绍了他们如何在生产环境中构建并扩展 OG Assist：从 Effect 原生 Agent 循环、A2A 协议、评测、人在环路中的审批、安全沙箱、长上下文记忆，到可观测性、工具与技能。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>Agent 系统</category><category>AI 产品与商业</category></item><item><title>用 TurboQuant 加速你的 Agent 检索</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E7%94%A8-turboquant-%E5%8A%A0%E9%80%9F%E4%BD%A0%E7%9A%84-agent-%E6%A3%80%E7%B4%A2/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E7%94%A8-turboquant-%E5%8A%A0%E9%80%9F%E4%BD%A0%E7%9A%84-agent-%E6%A3%80%E7%B4%A2/</guid><description>这场演讲介绍了 TurboQuant 如何把 Agent 检索中的嵌入向量和 KV 缓存从 32 位压缩到约 3 到 4 位，从而在基本保持检索质量的前提下降低约 5 倍内存占用。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>上下文工程</category><category>AI 基础设施</category></item><item><title>用 RL Agent 检测并修复 ETL 管道故障</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E7%94%A8-rl-agent-%E6%A3%80%E6%B5%8B%E5%B9%B6%E4%BF%AE%E5%A4%8D-etl-%E7%AE%A1%E9%81%93%E6%95%85%E9%9A%9C/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E7%94%A8-rl-agent-%E6%A3%80%E6%B5%8B%E5%B9%B6%E4%BF%AE%E5%A4%8D-etl-%E7%AE%A1%E9%81%93%E6%95%85%E9%9A%9C/</guid><description>Anna Marie Benzon 展示了一个面向云端 ETL 故障的 RL 引导型管道健康 Agent：用确定性规则建立可观察事实，用表格型 Q learning 在有限动作集合中选择修复策略，再用外部安全层约束自动化权限。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>Agent 系统</category><category>模型训练与推理</category></item><item><title>用户信号死在检索边界</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E7%94%A8%E6%88%B7%E4%BF%A1%E5%8F%B7%E6%AD%BB%E5%9C%A8%E6%A3%80%E7%B4%A2%E8%BE%B9%E7%95%8C/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E7%94%A8%E6%88%B7%E4%BF%A1%E5%8F%B7%E6%AD%BB%E5%9C%A8%E6%A3%80%E7%B4%A2%E8%BE%B9%E7%95%8C/</guid><description>这场 AI Engineer 演讲讨论 Agent 为什么会在检索阶段反复失败：传统记忆系统多数只按语义相似度保存偏好和历史，缺少来自执行结果的反馈。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>上下文工程</category><category>AI 产品与商业</category></item><item><title>端侧前沿成果</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E7%AB%AF%E4%BE%A7%E5%89%8D%E6%B2%BF%E6%88%90%E6%9E%9C/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E7%AB%AF%E4%BE%A7%E5%89%8D%E6%B2%BF%E6%88%90%E6%9E%9C/</guid><description>演讲者 Rachel Lee Nabors（来自 Arize）介绍了如何通过评估和选用端侧小语言模型（SLM）替代 GPT、Claude 等大型云端模型，从而降低推理成本、提升隐私安全与响应速度。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>AI 基础设施</category><category>模型训练与推理</category></item><item><title>结构化非结构化内容</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E7%BB%93%E6%9E%84%E5%8C%96%E9%9D%9E%E7%BB%93%E6%9E%84%E5%8C%96%E5%86%85%E5%AE%B9/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E7%BB%93%E6%9E%84%E5%8C%96%E9%9D%9E%E7%BB%93%E6%9E%84%E5%8C%96%E5%86%85%E5%AE%B9/</guid><description>Red Hat 开源工程师 Cedric Clyburn 介绍 Docling——一个隶属于 Linux Foundation 的开源文档处理工具，演示如何将 PDF、图片等非结构化数据准确转换为 Markdown/JSON。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>AI 产品与商业</category><category>上下文工程</category></item><item><title>绕过多模态税：混合 RAG、SQL RRF 与 UI 遥测</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E7%BB%95%E8%BF%87%E5%A4%9A%E6%A8%A1%E6%80%81%E7%A8%8E-%E6%B7%B7%E5%90%88-rag-sql-rrf-%E4%B8%8E-ui-%E9%81%A5%E6%B5%8B/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E7%BB%95%E8%BF%87%E5%A4%9A%E6%A8%A1%E6%80%81%E7%A8%8E-%E6%B7%B7%E5%90%88-rag-sql-rrf-%E4%B8%8E-ui-%E9%81%A5%E6%B5%8B/</guid><description>本演讲由 Ogilvy 高级后端开发者 Abed Matini 在 AI Engineer World Fair 2026 线上专场发表，主题是&quot;绕过多模态税&quot;：介绍如何构建一套无需付费框架的混合 RAG 系统。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>上下文工程</category><category>AI 基础设施</category></item><item><title>语音输入，视觉输出：痛苦与狂喜</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E8%AF%AD%E9%9F%B3%E8%BE%93%E5%85%A5-%E8%A7%86%E8%A7%89%E8%BE%93%E5%87%BA-%E7%97%9B%E8%8B%A6%E4%B8%8E%E7%8B%82%E5%96%9C/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E8%AF%AD%E9%9F%B3%E8%BE%93%E5%85%A5-%E8%A7%86%E8%A7%89%E8%BE%93%E5%87%BA-%E7%97%9B%E8%8B%A6%E4%B8%8E%E7%8B%82%E5%96%9C/</guid><description>Allen Pike 分享了 Forestwalk Labs 在构建语音输入、视觉输出 AI 体验中的经验：语音是更高带宽的人类输入方式，但实时语音交互受延迟限制很重；把输出切换为视觉界面后，系统可以在更宽松的响应窗口内提供自然体验。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>人机协作</category><category>AI 产品与商业</category></item><item><title>面向金融合规的 AI 多文档关联分析</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E9%9D%A2%E5%90%91%E9%87%91%E8%9E%8D%E5%90%88%E8%A7%84%E7%9A%84-ai-%E5%A4%9A%E6%96%87%E6%A1%A3%E5%85%B3%E8%81%94%E5%88%86%E6%9E%90/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E9%9D%A2%E5%90%91%E9%87%91%E8%9E%8D%E5%90%88%E8%A7%84%E7%9A%84-ai-%E5%A4%9A%E6%96%87%E6%A1%A3%E5%85%B3%E8%81%94%E5%88%86%E6%9E%90/</guid><description>这场演讲介绍了一个面向企业金融合规与欺诈检测的 AI 多文档关联分析框架。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>AI 产品与商业</category><category>上下文工程</category></item><item><title>递归式编程 Agent</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E9%80%92%E5%BD%92%E5%BC%8F-coding-agent/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E9%80%92%E5%BD%92%E5%BC%8F-coding-agent/</guid><description>Raymond Weitekamp（OpenProse）在 AI Engineer 大会演讲，阐述如何把递归语言模型（RLM）的思想应用到编程 Agent：以&quot;被误管理的天才&quot;为切入点。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>AI 编程</category><category>Agent 系统</category></item><item><title>面向非确定性 AI Agent 的确定性基础设施</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E9%9D%A2%E5%90%91%E9%9D%9E%E7%A1%AE%E5%AE%9A%E6%80%A7-ai-agent-%E7%9A%84%E7%A1%AE%E5%AE%9A%E6%80%A7%E5%9F%BA%E7%A1%80%E8%AE%BE%E6%96%BD/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E9%9D%A2%E5%90%91%E9%9D%9E%E7%A1%AE%E5%AE%9A%E6%80%A7-ai-agent-%E7%9A%84%E7%A1%AE%E5%AE%9A%E6%80%A7%E5%9F%BA%E7%A1%80%E8%AE%BE%E6%96%BD/</guid><description>Nishant Gupta 认为，随着组织从聊天机器人走向自主 AI Agent，核心挑战正在从模型智能转向生产可靠性：模型可以是概率性的，但承载 Agent 的基础设施必须提供确定性的验证、编排、观测、安全与恢复能力。</description><pubDate>Tue, 30 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>AI 基础设施</category><category>评测与可观测性</category></item><item><title>Community Wisdom 191：动力、流程和增长角色的真实约束</title><link>https://blog.xiaohansong.com/posts/ai-highlight/community-wisdom-191-career-slump-team-structure-growth-role/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/community-wisdom-191-career-slump-team-structure-growth-role/</guid><description>这期从职业低谷、新团队 1:1、成熟团队流程改造和增长角色演化里，提炼出同一个判断：先识别真实约束，再谈努力、流程和自动化。AI 让执行动作压缩之后，人的价值更集中到判断、定义和编辑。</description><pubDate>Sun, 28 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>组织与工作流</category><category>职业与成长</category></item><item><title>AI 安全灰天鹅：Agent 时代的提示注入与身份边界</title><link>https://blog.xiaohansong.com/posts/ai-highlight/ai%E5%AE%89%E5%85%A8%E7%81%B0%E5%A4%A9%E9%B9%85-agent%E6%97%B6%E4%BB%A3%E7%9A%84%E6%8F%90%E7%A4%BA%E6%B3%A8%E5%85%A5%E4%B8%8E%E8%BA%AB%E4%BB%BD%E8%BE%B9%E7%95%8C/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/ai%E5%AE%89%E5%85%A8%E7%81%B0%E5%A4%A9%E9%B9%85-agent%E6%97%B6%E4%BB%A3%E7%9A%84%E6%8F%90%E7%A4%BA%E6%B3%A8%E5%85%A5%E4%B8%8E%E8%BA%AB%E4%BB%BD%E8%BE%B9%E7%95%8C/</guid><description>Agent 安全不是给模型加一句更严厉的系统提示，而是重构整条执行链：区分可信与不可信上下文，训练专门的策略与红队模型，拦截工具调用，并重新设计 agent-native identity。</description><pubDate>Tue, 23 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>评测与可观测性</category><category>Agent 系统</category></item><item><title>AI 算力输出最大化：FLOPs 电网与 Anthropic 的 P0</title><link>https://blog.xiaohansong.com/posts/ai-highlight/ai%E7%AE%97%E5%8A%9B%E8%BE%93%E5%87%BA%E6%9C%80%E5%A4%A7%E5%8C%96-flops%E7%94%B5%E7%BD%91%E4%B8%8Eanthropic%E7%9A%84p0/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/ai%E7%AE%97%E5%8A%9B%E8%BE%93%E5%87%BA%E6%9C%80%E5%A4%A7%E5%8C%96-flops%E7%94%B5%E7%BD%91%E4%B8%8Eanthropic%E7%9A%84p0/</guid><description>这期访谈最有价值的起点，是把 AI scaling 从买更多 GPU 改写成如何把已有 GPU 变成真实训练进展。</description><pubDate>Sun, 21 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>AI 基础设施</category><category>模型训练与推理</category></item><item><title>Lenny 社群智慧：付费不用、Fractional CPO 与团队上下文治理</title><link>https://blog.xiaohansong.com/posts/ai-highlight/lenny%E7%A4%BE%E7%BE%A4%E6%99%BA%E6%85%A7-%E4%BB%98%E8%B4%B9%E4%B8%8D%E7%94%A8-fractional-cpo%E4%B8%8E%E5%9B%A2%E9%98%9F%E4%B8%8A%E4%B8%8B%E6%96%87%E6%B2%BB%E7%90%86/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/lenny%E7%A4%BE%E7%BE%A4%E6%99%BA%E6%85%A7-%E4%BB%98%E8%B4%B9%E4%B8%8D%E7%94%A8-fractional-cpo%E4%B8%8E%E5%9B%A2%E9%98%9F%E4%B8%8A%E4%B8%8B%E6%96%87%E6%B2%BB%E7%90%86/</guid><description>本期高亮提炼 Lenny 社群关于付费不用、早期产品顾问、Referral 反作弊、自由职业工具栈，以及 Claude Code 团队上下文治理的讨论。核心判断是：AI 让个人和团队更快执行，也更快暴露行为嵌入、判断质量和上下文治理的缺口。</description><pubDate>Sun, 21 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>组织与工作流</category><category>上下文工程</category></item><item><title>Lenny 社群智慧：AI 改变产品运营模型、微型团队 GTM 与 AI 作品集</title><link>https://blog.xiaohansong.com/posts/ai-highlight/lenny%E7%A4%BE%E7%BE%A4%E6%99%BA%E6%85%A7-ai%E6%94%B9%E5%8F%98%E4%BA%A7%E5%93%81%E8%BF%90%E8%90%A5%E6%A8%A1%E5%9E%8B-%E5%BE%AE%E5%9E%8B%E5%9B%A2%E9%98%9Fgtm%E4%B8%8Eai%E4%BD%9C%E5%93%81%E9%9B%86/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/lenny%E7%A4%BE%E7%BE%A4%E6%99%BA%E6%85%A7-ai%E6%94%B9%E5%8F%98%E4%BA%A7%E5%93%81%E8%BF%90%E8%90%A5%E6%A8%A1%E5%9E%8B-%E5%BE%AE%E5%9E%8B%E5%9B%A2%E9%98%9Fgtm%E4%B8%8Eai%E4%BD%9C%E5%93%81%E9%9B%86/</guid><description>这组讨论问的是 AI 如何改变产品运营模型，真正有价值的答案不是 PM 能不能写代码，而是交付加速以后，团队哪里会变成新的约束。</description><pubDate>Sun, 14 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>组织与工作流</category><category>AI 产品与商业</category></item><item><title>GitHub 的 AI 编程压力测试：Agent 时代的软件平台重构</title><link>https://blog.xiaohansong.com/posts/ai-highlight/github-coo-kyle-daigle%E8%AE%BF%E8%B0%88-ai%E6%97%B6%E4%BB%A3%E4%BB%A3%E7%A0%81%E5%B9%B3%E5%8F%B0%E5%8F%98%E9%9D%A9/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/github-coo-kyle-daigle%E8%AE%BF%E8%B0%88-ai%E6%97%B6%E4%BB%A3%E4%BB%A3%E7%A0%81%E5%B9%B3%E5%8F%B0%E5%8F%98%E9%9D%A9/</guid><description>这篇访谈的底层问题是：AI coding agent 让代码产量、提交频率、构建次数和权限检查同时放大，GitHub 这类平台开始承受一种新的系统压力。</description><pubDate>Sat, 13 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>AI 编程</category><category>AI 基础设施</category></item><item><title>Lenny 社群智慧：创业融资、Vibe Coding 与 Agent 数据语义层</title><link>https://blog.xiaohansong.com/posts/ai-highlight/lenny%E7%A4%BE%E7%BE%A4%E6%99%BA%E6%85%A7-%E8%9E%8D%E8%B5%84-vibe-coding%E4%B8%8Eagent%E6%95%B0%E6%8D%AE%E8%AF%AD%E4%B9%89%E5%B1%82/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/lenny%E7%A4%BE%E7%BE%A4%E6%99%BA%E6%85%A7-%E8%9E%8D%E8%B5%84-vibe-coding%E4%B8%8Eagent%E6%95%B0%E6%8D%AE%E8%AF%AD%E4%B9%89%E5%B1%82/</guid><description>Lenny 社群关于 bootstrapping、Vibe Coding 产品验证、Agent 数据语义层和 APM 基本功的讨论，核心线索是：构建成本下降后，稀缺性迁移到约束选择、分发验证、语义一致性和判断质量。</description><pubDate>Sat, 13 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>AI 产品与商业</category><category>AI 编程</category></item><item><title>Satya Nadella 2026：AI 平台战略与企业 Agent 的价值重构</title><link>https://blog.xiaohansong.com/posts/ai-highlight/satya-nadella-2026-ai%E5%B9%B3%E5%8F%B0%E6%88%98%E7%95%A5%E4%B8%8E%E4%BC%81%E4%B8%9Aagent/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/satya-nadella-2026-ai%E5%B9%B3%E5%8F%B0%E6%88%98%E7%95%A5%E4%B8%8E%E4%BC%81%E4%B8%9Aagent/</guid><description>Satya Nadella 谈 AI 时代的平台战略：企业护城河不在单一模型，而在 harness、context layer、private eval、trace 和组织级 agent memory。</description><pubDate>Sat, 13 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>AI 产品与商业</category><category>Agent 系统</category></item><item><title>异步 Agent 时代：从本地补全到后台编码工厂</title><link>https://blog.xiaohansong.com/posts/ai-highlight/the-age-of-async-agents-cognition/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/the-age-of-async-agents-cognition/</guid><description>这组摘录给出了全文最重要的演化框架：AI coding 先是补全工具，再是本地 Agent，下一阶段是能在后台独立推进任务的异步 Agent。</description><pubDate>Sat, 13 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>AI 编程</category><category>Agent 系统</category></item><item><title>Video Agents：视频生成的下一站是 Agent</title><link>https://blog.xiaohansong.com/posts/ai-highlight/video-agents%E8%A7%86%E9%A2%91%E7%94%9F%E6%88%90%E7%9A%84%E4%B8%8B%E4%B8%80%E7%AB%99%E6%98%AFagent/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/video-agents%E8%A7%86%E9%A2%91%E7%94%9F%E6%88%90%E7%9A%84%E4%B8%8B%E4%B8%80%E7%AB%99%E6%98%AFagent/</guid><description>这期 Latent Space 采访 Ethan He，表面主题是 Grok Imagine、视频生成和世界模型，真正有价值的是一个更大的判断：视频生成正在走 AI 编程已经走过的路。</description><pubDate>Thu, 04 Jun 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>Agent 系统</category><category>模型训练与推理</category></item><item><title>Cursor Composer 2 训练基础设施与高性能 RL</title><link>https://blog.xiaohansong.com/posts/ai-highlight/cursor-composer-2%E8%AE%AD%E7%BB%83%E5%9F%BA%E7%A1%80%E8%AE%BE%E6%96%BD%E4%B8%8E%E9%AB%98%E6%80%A7%E8%83%BDrl/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/cursor-composer-2%E8%AE%AD%E7%BB%83%E5%9F%BA%E7%A1%80%E8%AE%BE%E6%96%BD%E4%B8%8E%E9%AB%98%E6%80%A7%E8%83%BDrl/</guid><description>这期访谈把编程 Agent 的能力提升拉回到系统工程：真实环境、rollout、reward、权重同步、数值一致性和上下文压缩，决定模型能否从会写代码变成能完成软件工程任务。</description><pubDate>Thu, 28 May 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>模型训练与推理</category><category>AI 基础设施</category></item><item><title>Abridge：医疗 AI 领域的临床智能层构建者</title><link>https://blog.xiaohansong.com/posts/ai-highlight/abridge%E5%8C%BB%E7%96%97ai%E9%A2%86%E5%9F%9F%E7%9A%84%E4%B8%B4%E5%BA%8A%E6%99%BA%E8%83%BD%E5%B1%82%E6%9E%84%E5%BB%BA%E8%80%85/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/abridge%E5%8C%BB%E7%96%97ai%E9%A2%86%E5%9F%9F%E7%9A%84%E4%B8%B4%E5%BA%8A%E6%99%BA%E8%83%BD%E5%B1%82%E6%9E%84%E5%BB%BA%E8%80%85/</guid><description>Abridge 展示了垂直 AI 产品如何从临床文档切入，逐步演化为医疗系统的上下文层、评估系统和 Agent 工作环境。</description><pubDate>Tue, 26 May 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>AI 产品与商业</category><category>评测与可观测性</category></item><item><title>Replit CEO 访谈：AI 原生开发者与后提示时代</title><link>https://blog.xiaohansong.com/posts/ai-highlight/replit-ceo%E6%B7%B1%E5%BA%A6%E8%AE%BF%E8%B0%88-ai%E5%8E%9F%E7%94%9F%E5%BC%80%E5%8F%91%E8%80%85%E4%B8%8E%E5%90%8E%E6%8F%90%E7%A4%BA%E6%97%B6%E4%BB%A3/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/replit-ceo%E6%B7%B1%E5%BA%A6%E8%AE%BF%E8%B0%88-ai%E5%8E%9F%E7%94%9F%E5%BC%80%E5%8F%91%E8%80%85%E4%B8%8E%E5%90%8E%E6%8F%90%E7%A4%BA%E6%97%B6%E4%BB%A3/</guid><description>这篇访谈最值得保留的地方，是它把 AI 编程从工具效率问题推到了生产关系问题：谁有资格创造软件，谁最懂真实问题，组织内部的建设者如何被重新定义，以及 Agent 产品为了接住这种变化需要具备什么样的系统能力。</description><pubDate>Tue, 26 May 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>AI 编程</category><category>人机协作</category></item><item><title>Demis Hassabis：AGI 倒计时与 AI Agent 时代</title><link>https://blog.xiaohansong.com/posts/ai-highlight/2026%E8%B0%B7%E6%AD%8Cio-demis-hassabis%E8%AE%BF%E8%B0%88-agi%E5%80%92%E8%AE%A1%E6%97%B6%E4%B8%8Eagent%E6%97%B6%E4%BB%A3/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/2026%E8%B0%B7%E6%AD%8Cio-demis-hassabis%E8%AE%BF%E8%B0%88-agi%E5%80%92%E8%AE%A1%E6%97%B6%E4%B8%8Eagent%E6%97%B6%E4%BB%A3/</guid><description>这篇访谈把 Google DeepMind 的几条路线放在同一个框架里：模型能力、Agent 化、模拟环境、科学发现、持续学习和可靠性共同推进 AGI。</description><pubDate>Sat, 23 May 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>模型训练与推理</category><category>Agent 系统</category></item><item><title>Anthropic Claude Code 深度访谈：Agent 工程化的下一步</title><link>https://blog.xiaohansong.com/posts/ai-highlight/anthropic-claude-code%E6%B7%B1%E5%BA%A6%E8%AE%BF%E8%B0%88-ai-agent%E5%B7%A5%E7%A8%8B%E5%8C%96%E5%AE%9E%E8%B7%B5%E4%B8%8E%E8%A1%8C%E4%B8%9A%E5%8F%98%E9%9D%A9%E5%89%8D%E7%9E%BB/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/anthropic-claude-code%E6%B7%B1%E5%BA%A6%E8%AE%BF%E8%B0%88-ai-agent%E5%B7%A5%E7%A8%8B%E5%8C%96%E5%AE%9E%E8%B7%B5%E4%B8%8E%E8%A1%8C%E4%B8%9A%E5%8F%98%E9%9D%A9%E5%89%8D%E7%9E%BB/</guid><description>Boris 谈 Claude Code 的真实重点不是增长，而是 Agent 工程化：真实操作权、流程重构、长任务监督、并行调度，以及人类从执行者转向顶层驾驶者。</description><pubDate>Fri, 22 May 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>AI 编程</category><category>Harness Engineering</category></item><item><title>从 OpenClaw 到 Hermes Agent：Agent 不是聊天框，而是新的工作系统</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E4%BB%8Eopenclaw%E5%88%B0hermes-agent/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E4%BB%8Eopenclaw%E5%88%B0hermes-agent/</guid><description>这期播客真正有价值的地方，不是介绍 Hermes Agent 或 OpenClaw 的功能，而是把 Agent 行业的几个底层分歧讲清楚了：模型与 harness 谁更重要，Agent 应该厚约束还是薄封装，人的价值会迁移到哪里。</description><pubDate>Fri, 15 May 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>Agent 系统</category><category>Harness Engineering</category></item><item><title>Claude Managed Agents：AI 平台正在变成结果交付层</title><link>https://blog.xiaohansong.com/posts/ai-highlight/claude-managed-agents-%E4%BA%A7%E5%93%81%E5%AE%9A%E4%BD%8D%E4%B8%8Eai%E5%B9%B3%E5%8F%B0%E6%BC%94%E8%BF%9B/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/claude-managed-agents-%E4%BA%A7%E5%93%81%E5%AE%9A%E4%BD%8D%E4%B8%8Eai%E5%B9%B3%E5%8F%B0%E6%BC%94%E8%BF%9B/</guid><description>Anthropic 团队把 Managed Agents 定位为一套可运行、可部署、可拥有身份、可被组织接管的执行基础设施。真正的竞争点开始落在状态、工具、沙箱、权限、长期运行和多 Agent 编排上。</description><pubDate>Wed, 13 May 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>Agent 系统</category><category>AI 基础设施</category></item><item><title>Anthropic 前沿 Agent 记忆系统与 Dreaming 机制高亮</title><link>https://blog.xiaohansong.com/posts/ai-highlight/anthropic%E5%89%8D%E6%B2%BFagent%E8%AE%B0%E5%BF%86%E7%B3%BB%E7%BB%9F%E4%B8%8Edreaming%E6%9C%BA%E5%88%B6/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/anthropic%E5%89%8D%E6%B2%BFagent%E8%AE%B0%E5%BF%86%E7%B3%BB%E7%BB%9F%E4%B8%8Edreaming%E6%9C%BA%E5%88%B6/</guid><description>Mahes 把 Anthropic 这一年多的产品演进放在一条清晰链路里：MCP 给 Agent 外部工具和数据，Claude Code 与 Agent SDK 提供执行 harness，Skills 让 Agent 获得可复用能力。</description><pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>上下文工程</category><category>Agent 系统</category></item><item><title>Claude Code 团队如何重构研发管理</title><link>https://blog.xiaohansong.com/posts/ai-highlight/claude-code%E5%9B%A2%E9%98%9F%E7%AE%A1%E7%90%86%E7%BB%8F%E9%AA%8C-fiona-fung/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/claude-code%E5%9B%A2%E9%98%9F%E7%AE%A1%E7%90%86%E7%BB%8F%E9%AA%8C-fiona-fung/</guid><description>Fiona 的核心判断很直接：过去多年，软件工程流程默认编码带宽昂贵，所以规划、设计文档、排期、所有权和评审都围绕节省工程时间展开。</description><pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>组织与工作流</category><category>AI 编程</category></item><item><title>对话姚顺宇：从理论物理到 AI 的跨界探索与行业洞察</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E5%A7%9A%E9%A1%BA%E5%AE%87%E8%AE%BF%E8%B0%88-%E4%BB%8E%E7%89%A9%E7%90%86%E5%88%B0ai-%E5%90%8E%E8%AE%AD%E7%BB%83%E4%B8%8E%E9%95%BF%E7%A8%8B%E4%BB%BB%E5%8A%A1/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E5%A7%9A%E9%A1%BA%E5%AE%87%E8%AE%BF%E8%B0%88-%E4%BB%8E%E7%89%A9%E7%90%86%E5%88%B0ai-%E5%90%8E%E8%AE%AD%E7%BB%83%E4%B8%8E%E9%95%BF%E7%A8%8B%E4%BB%BB%E5%8A%A1/</guid><description>姚顺宇反复用物理史解释当下大模型研究：我们对模型内部机制的理解仍然粗糙，但这不妨碍工程前进。</description><pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>模型训练与推理</category><category>Agent 系统</category></item><item><title>Dario 与 Daniela：Agent 组织化与产品节奏</title><link>https://blog.xiaohansong.com/posts/ai-highlight/dario%E4%B8%8Edaniela-amodei%E5%AF%B9%E8%AF%9D-%E6%8C%87%E6%95%B0%E5%A2%9E%E9%95%BF-%E7%BB%84%E7%BB%87%E7%BA%A7agent%E4%B8%8Eai%E4%BA%A7%E5%93%81%E8%8A%82%E5%A5%8F/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/dario%E4%B8%8Edaniela-amodei%E5%AF%B9%E8%AF%9D-%E6%8C%87%E6%95%B0%E5%A2%9E%E9%95%BF-%E7%BB%84%E7%BB%87%E7%BA%A7agent%E4%B8%8Eai%E4%BA%A7%E5%93%81%E8%8A%82%E5%A5%8F/</guid><description>Dario 对指数增长的描述很值得保留：他们十年前就用 scaling laws 画出了曲线，但当曲线真的落到组织、收入、算力和产品节奏上，人的体验仍然是不稳定的。</description><pubDate>Thu, 07 May 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>AI 产品与商业</category><category>组织与工作流</category></item><item><title>Claude Code 创始人 Boris Cherny 访谈：AI coding 的下一阶段</title><link>https://blog.xiaohansong.com/posts/ai-highlight/claude-code%E5%88%9B%E5%A7%8B%E4%BA%BAboris-cherny%E8%AE%BF%E8%B0%88/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/claude-code%E5%88%9B%E5%A7%8B%E4%BA%BAboris-cherny%E8%AE%BF%E8%B0%88/</guid><description>Claude Code 的起点很值得看：它并不是在模型已经完全成熟后才被设计出来的，而是 Anthropic 团队意识到产品形态落后于模型能力。</description><pubDate>Wed, 06 May 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>AI 编程</category><category>Agent 系统</category></item><item><title>Demis Hassabis（DeepMind）访谈：AGI 技术进展、AI 在科学领域的应用与创业建议</title><link>https://blog.xiaohansong.com/posts/ai-highlight/demis-hassabis%E8%AE%BF%E8%B0%88-agi%E6%8A%80%E6%9C%AF%E8%BF%9B%E5%B1%95-%E7%A7%91%E5%AD%A6%E5%8F%91%E7%8E%B0%E4%B8%8Eagent%E8%B7%AF%E5%BE%84/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/demis-hassabis%E8%AE%BF%E8%B0%88-agi%E6%8A%80%E6%9C%AF%E8%BF%9B%E5%B1%95-%E7%A7%91%E5%AD%A6%E5%8F%91%E7%8E%B0%E4%B8%8Eagent%E8%B7%AF%E5%BE%84/</guid><description>这次访谈的价值不在于 Demis 又说了一遍 AGI 很快，而在于他把 DeepMind 路线里的几条线串起来了：记忆与持续学习、AlphaGo 式搜索、Agent、科学发现、多模态世界模型，以及深科技创业在 AGI 时间线下的防御性。</description><pubDate>Tue, 05 May 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>模型训练与推理</category><category>Agent 系统</category></item><item><title>OpenClaw 与 Agent 时代：AI 驱动的工作范式与软件生态变革</title><link>https://blog.xiaohansong.com/posts/ai-highlight/openclaw%E4%B8%8Eagent%E6%97%B6%E4%BB%A3-ai%E9%A9%B1%E5%8A%A8%E7%9A%84%E5%B7%A5%E4%BD%9C%E8%8C%83%E5%BC%8F%E4%B8%8E%E8%BD%AF%E4%BB%B6%E7%94%9F%E6%80%81%E5%8F%98%E9%9D%A9/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/openclaw%E4%B8%8Eagent%E6%97%B6%E4%BB%A3-ai%E9%A9%B1%E5%8A%A8%E7%9A%84%E5%B7%A5%E4%BD%9C%E8%8C%83%E5%BC%8F%E4%B8%8E%E8%BD%AF%E4%BB%B6%E7%94%9F%E6%80%81%E5%8F%98%E9%9D%A9/</guid><description>这篇访谈最值得保留的地方，不是 OpenClaw 或 Claude Code 哪个工具更好，而是它把 Agent 从工具讨论推进到了工作界面、公司形态、商业模式和人的角色变化。</description><pubDate>Tue, 05 May 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>人机协作</category><category>Agent 系统</category></item><item><title>李光耀领导力访谈：从殖民到建国</title><link>https://blog.xiaohansong.com/posts/ai-highlight/%E6%9D%8E%E5%85%89%E8%80%80%E9%A2%86%E5%AF%BC%E5%8A%9B%E8%AE%BF%E8%B0%88-%E4%BB%8E%E6%AE%96%E6%B0%91%E5%88%B0%E5%BB%BA%E5%9B%BD/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/%E6%9D%8E%E5%85%89%E8%80%80%E9%A2%86%E5%AF%BC%E5%8A%9B%E8%AE%BF%E8%B0%88-%E4%BB%8E%E6%AE%96%E6%B0%91%E5%88%B0%E5%BB%BA%E5%9B%BD/</guid><description>这次高亮只筛选领导力相关内容。</description><pubDate>Sat, 02 May 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>职业与成长</category><category>组织与工作流</category></item><item><title>Andrej Karpathy 访谈：AI 时代的编程范式、可验证性与人类价值</title><link>https://blog.xiaohansong.com/posts/ai-highlight/andrej-karpathy%E8%AE%BF%E8%B0%88-ai%E6%97%B6%E4%BB%A3%E7%9A%84%E7%BC%96%E7%A8%8B%E8%8C%83%E5%BC%8F-%E5%8F%AF%E9%AA%8C%E8%AF%81%E6%80%A7%E4%B8%8E%E4%BA%BA%E7%B1%BB%E4%BB%B7%E5%80%BC/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/andrej-karpathy%E8%AE%BF%E8%B0%88-ai%E6%97%B6%E4%BB%A3%E7%9A%84%E7%BC%96%E7%A8%8B%E8%8C%83%E5%BC%8F-%E5%8F%AF%E9%AA%8C%E8%AF%81%E6%80%A7%E4%B8%8E%E4%BA%BA%E7%B1%BB%E4%BB%B7%E5%80%BC/</guid><description>Karpathy 这里讲的重点，不是 AI 让旧式编程变快，而是可操作对象变了。</description><pubDate>Thu, 30 Apr 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>AI 编程</category><category>人机协作</category></item><item><title>OpenAI 的 Ryan Leopo 谈 Harness Engineering：当代码免费时，如何构建软件</title><link>https://blog.xiaohansong.com/posts/ai-highlight/openai%E7%9A%84ryan-leopo%E8%B0%88harness-engineering-%E5%BD%93%E4%BB%A3%E7%A0%81%E5%85%8D%E8%B4%B9%E6%97%B6%E5%A6%82%E4%BD%95%E6%9E%84%E5%BB%BA%E8%BD%AF%E4%BB%B6/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/openai%E7%9A%84ryan-leopo%E8%B0%88harness-engineering-%E5%BD%93%E4%BB%A3%E7%A0%81%E5%85%8D%E8%B4%B9%E6%97%B6%E5%A6%82%E4%BD%95%E6%9E%84%E5%BB%BA%E8%BD%AF%E4%BB%B6/</guid><description>Ryan Leopo 的核心判断很直接：编码 Agent 已经让实现本身变得不再稀缺。</description><pubDate>Thu, 30 Apr 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>Harness Engineering</category><category>AI 编程</category></item><item><title>OpenAI 的 Ryan Leopo 谈&quot;Harness Engineering&quot;：当代码免费时，如何构建软件</title><link>https://blog.xiaohansong.com/posts/ai-highlight/openai%E7%9A%84ryan-leopo%E8%B0%88harness-engineering/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/openai%E7%9A%84ryan-leopo%E8%B0%88harness-engineering/</guid><description>Ryan 在演讲开篇给出整个框架的支点，不是&quot;AI 能帮你写代码&quot;，而是从稀缺资源角度重新定义了软件工程的约束条件。</description><pubDate>Thu, 30 Apr 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>Harness Engineering</category><category>AI 编程</category></item><item><title>Shopify CTO Muel Parkin 访谈：内部 AI 工具落地、自研 ML 基础设施与前沿技术应用</title><link>https://blog.xiaohansong.com/posts/ai-highlight/shopify-cto%E8%AE%BF%E8%B0%88-%E5%86%85%E9%83%A8ai%E5%B7%A5%E5%85%B7%E8%90%BD%E5%9C%B0-%E8%87%AA%E7%A0%94ml%E5%9F%BA%E7%A1%80%E8%AE%BE%E6%96%BD%E4%B8%8E%E5%89%8D%E6%B2%BF%E6%8A%80%E6%9C%AF%E5%BA%94%E7%94%A8/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/shopify-cto%E8%AE%BF%E8%B0%88-%E5%86%85%E9%83%A8ai%E5%B7%A5%E5%85%B7%E8%90%BD%E5%9C%B0-%E8%87%AA%E7%A0%94ml%E5%9F%BA%E7%A1%80%E8%AE%BE%E6%96%BD%E4%B8%8E%E5%89%8D%E6%B2%BF%E6%8A%80%E6%9C%AF%E5%BA%94%E7%94%A8/</guid><description>访谈一开始就在回应 Jensen Huang 关于工程师应大量消耗 token 的观点。</description><pubDate>Thu, 30 Apr 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>组织与工作流</category><category>AI 基础设施</category></item><item><title>https://www.latent.space/p/unsupervised-learning-2026</title><link>https://blog.xiaohansong.com/posts/ai-highlight/aie-europe-debrief-agent-labs-thesis/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/aie-europe-debrief-agent-labs-thesis/</guid><description>这组摘录来自 AIE Europe 会后的判断，讨论的是 agent infra 是否开始稳定，以及稳定到底体现在哪里。</description><pubDate>Tue, 28 Apr 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>AI 基础设施</category><category>Agent 系统</category></item><item><title>🧠 Community Wisdom Sharing Claude Code skills across the org, Why did lennysjobs.com shut down, opening an LLC for side work, AI for SMB sales, and more.md</title><link>https://blog.xiaohansong.com/posts/ai-highlight/community-wisdom-sharing-claude-code-skills-across-the-org-why-did-lennysjobs/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/community-wisdom-sharing-claude-code-skills-across-the-org-why-did-lennysjobs/</guid><description>这组讨论围绕 50 人创业公司内部如何共享 Claude Code skills。</description><pubDate>Sat, 25 Apr 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>组织与工作流</category><category>AI 编程</category></item><item><title>2026-04-09-[Podwise] 新文章：We Gave Every Employee an AI Agent. Here&apos;s What Happened.</title><link>https://blog.xiaohansong.com/posts/ai-highlight/2026-04-09-podwise-%E6%96%B0%E6%96%87%E7%AB%A0-we-gave-every-employee-an-ai-agent/</link><guid isPermaLink="true">https://blog.xiaohansong.com/posts/ai-highlight/2026-04-09-podwise-%E6%96%B0%E6%96%87%E7%AB%A0-we-gave-every-employee-an-ai-agent/</guid><description>这篇对谈里最有价值的一层，不是 Every 把 OpenClaw 跑起来了，而是他们讲清了为什么个人代理和通用聊天模型在组织里会导向完全不同的关系结构。</description><pubDate>Thu, 09 Apr 2026 00:00:00 GMT</pubDate><category>AI Highlight</category><category>组织与工作流</category><category>人机协作</category></item></channel></rss>