<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>Python on mayulu的AI笔记</title><link>https://mayulu.co/tags/python/</link><description>Recent content in Python on mayulu的AI笔记</description><generator>Hugo</generator><language>zhcn</language><copyright>Copyright © 2025–2026 mayulu All Rights Reserved</copyright><lastBuildDate>Mon, 13 Jul 2026 16:50:25 +0800</lastBuildDate><atom:link href="https://mayulu.co/tags/python/index.xml" rel="self" type="application/rss+xml"/><item><title>试验观察雅克比透镜 J-Lens</title><link>https://mayulu.co/posts/jacobian-lens-from-scratch/</link><pubDate>Mon, 13 Jul 2026 16:50:25 +0800</pubDate><guid>https://mayulu.co/posts/jacobian-lens-from-scratch/</guid><description>本文基于 Anthropic 开源 jacobian-lens 项目，完整复现 J-Lens 雅克比透镜，选用 Qwen3.5-4B 模型实操演示。文章讲解 uv 环境搭建、模型加载流程，拆解雅可比矩阵拟合、透镜推理核心源码，区分拟合反向传播与应用前向推理两套逻辑，提供拟合、推理、可视化完整可运行 Python 脚本，附 M3 Max 设备实测耗时，对比传统 Logit Lens 解码效果，适合想要落地 LLM 内部表征解读、复现 J-Space 全局工作空间实验的开发者。</description></item></channel></rss>