<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>ZAC / NOTES</title><link>https://blog.zacw.top</link><description>关于 AI 系统、基础设施、工程实践与长期学习的技术笔记。</description><language>zh-CN</language><item><title>DeepSeek 一周年进展：从 R1 影响力到 V4 的 Agent 化</title><link>https://blog.zacw.top/articles/deepseek-one-year-evolution/</link><guid>https://blog.zacw.top/articles/deepseek-one-year-evolution/</guid><pubDate>Tue, 28 Jul 2026 00:00:00 +0000</pubDate><description>沿着模式统一、Agent、稀疏注意力、长上下文与开放部署五条线，重建 DeepSeek 从 V3.1 到 V4 Preview 的一年演进。</description></item><item><title>一次可回滚的服务迁移，是怎样设计的</title><link>https://blog.zacw.top/articles/reversible-service-migration/</link><guid>https://blog.zacw.top/articles/reversible-service-migration/</guid><pubDate>Tue, 28 Jul 2026 00:00:00 +0000</pubDate><description>迁移不是复制文件，而是管理状态、流量与回滚边界。</description></item><item><title>为 Agent 建立可验证的执行边界</title><link>https://blog.zacw.top/articles/verifiable-agent-boundaries/</link><guid>https://blog.zacw.top/articles/verifiable-agent-boundaries/</guid><pubDate>Fri, 24 Jul 2026 00:00:00 +0000</pubDate><description>让自动化可靠的关键，不是增加更多能力，而是明确权限、证据和停止条件。</description></item><item><title>从日志到证据：诊断复杂系统的工作方法</title><link>https://blog.zacw.top/articles/logs-to-evidence/</link><guid>https://blog.zacw.top/articles/logs-to-evidence/</guid><pubDate>Mon, 20 Jul 2026 00:00:00 +0000</pubDate><description>日志只是观察材料；诊断需要时间线、对照实验和可以被推翻的假设。</description></item></channel></rss>