{ "@context": "https://schema.org", "@type": "ItemList", "name": "LZ AI NOTE 文章索引", "description": "供搜索引擎、问答系统和大模型检索使用的公开文章目录。", "url": "https://wizardpisces.github.io/ai-index.json", "numberOfItems": 190, "itemListElement": [ { "@type": "ListItem", "position": 1, "item": { "@type": "Article", "headline": "AiNotes", "description": "AiNotes AI 的摘要与思考 AI Coding Agent 的 IR 是什么 （这是所有 agent 共享的世界，或者 State，或者是 Agent 直接沟通的契约） 这可能不是一个 Agent 问题，而是一个 Computer Science 问题。 如果把历史串起来，会发现每一次大的工程能力提升，都伴随着一种新的\"中间表示\"（Represen…", "url": "https://wizardpisces.github.io/blog/AiNotes", "mainEntityOfPage": "https://wizardpisces.github.io/blog/AiNotes", "dateModified": "2026-06-30T10:05:29.127Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" } } }, { "@type": "ListItem", "position": 2, "item": { "@type": "Article", "headline": "DailyReflections", "description": "DailyReflections 简介：个人随笔 How To Earn A Billion Dollar 创业的关键在于深入了解特定用户群体，从而精准地打造他们真正想要的产品。 用户真正想要的是什么？你能为他们做些什么，从而显著改善他们的生活？ 这就是创业的同理心，也是我们在创始人身上寻找和培养的品质。 羽毛球正反手击球感悟 1. 击球本质 目标是提高拍…", "url": "https://wizardpisces.github.io/blog/DailyReflections", "mainEntityOfPage": "https://wizardpisces.github.io/blog/DailyReflections", "dateModified": "2026-06-26T03:16:52.174Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" } } }, { "@type": "ListItem", "position": 3, "item": { "@type": "Article", "headline": "LZ AI NOTE 是什么：中文 AI 公众号与知识站", "description": "LZ AI NOTE 是一个面向开发者、产品从业者和 AI 落地团队的中文 AI 公众号与知识站，覆盖 AI Agent、LLM、AI Coding、RAG 与 AI 工程实践。", "url": "https://wizardpisces.github.io/blog/Introduction", "mainEntityOfPage": "https://wizardpisces.github.io/blog/Introduction", "dateModified": "2026-08-24T03:11:58.243Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" } } }, { "@type": "ListItem", "position": 4, "item": { "@type": "Article", "headline": "TechNotes", "description": "TechNotes 简介：个人技术思考与总结 读 《埃隆 马斯克传》有感 擅长航天和电动车工程的人，也应该同样擅长AI和自动驾驶软件系统？ 但本质上，这几类系统的难度类型完全不同。 航天与电动车属于“物理工程问题”：目标明确（能飞、能跑、成本更低），系统边界相对清晰，失败模式通常可定位、可复现，工程迭代路径也较稳定。 而AI与自动驾驶属于“开放世界概率问题…", "url": "https://wizardpisces.github.io/blog/TechNotes", "mainEntityOfPage": "https://wizardpisces.github.io/blog/TechNotes", "dateModified": "2026-04-10T06:49:40.314Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" } } }, { "@type": "ListItem", "position": 5, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2024 09 19 Anthropic 工程团队发布的 《Contextual Retrieval》 （工程博客）提出了一种显著提升 RAG 检索准确率的方法。虽然这不是传统意义上的学术论文，但其技术影响力不亚于论文——它直指 RAG 系统最被忽视的瓶颈，并用极简方案给出了工程上可落地的答案。 简单来说，传统 RAG 把文档切成小块再嵌入，…", "url": "https://wizardpisces.github.io/blog/tech/ai-LLM/Anthropic%20-%20Contextual%20Retrieval", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-LLM/Anthropic%20-%20Contextual%20Retrieval", "dateModified": "2026-03-11T09:42:38.861Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-LLM" ] } }, { "@type": "ListItem", "position": 6, "item": { "@type": "Article", "headline": "最小 SSM：一维输入，一维状态", "description": "论文发布时间：2023 12 01 由 CMU 和 Princeton 团队发布的 Mamba （论文标题为《Mamba: Linear Time Sequence Modeling with Selective State Spaces》）是后 Transformer 时代最强有力的挑战者之一。 它不是对 Transformer 的修补，而是对其底层逻辑…", "url": "https://wizardpisces.github.io/blog/tech/ai-LLM/CMU%20%26%20Princeton%20-%20Mamba", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-LLM/CMU%20%26%20Princeton%20-%20Mamba", "dateModified": "2026-03-03T06:21:14.894Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-LLM" ] } }, { "@type": "ListItem", "position": 7, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2025 08 31 Cornell University 的研究人员发表了一篇题为《Can General Purpose Omnimodels Compete with Specialists? A Case Study in Medical Image Segmentation》的论文，在医疗 AI 领域引发了关于 “通才 vs 专才”…", "url": "https://wizardpisces.github.io/blog/tech/ai-LLM/Cornell%20-%20Medical%20Omnimodels%20vs%20Specialists", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-LLM/Cornell%20-%20Medical%20Omnimodels%20vs%20Specialists", "dateModified": "2026-01-30T03:36:41.548Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-LLM" ] } }, { "@type": "ListItem", "position": 8, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2025 07 11 Distyl AI 发布的论文 《How Many Instructions Can LLMs Follow At Once?》 探讨了一个非常实际且紧迫的问题： 大模型究竟能同时遵循多少条指令？ 这篇论文提出了一个新的基准测试 IFScale ，专门用于评估大语言模型（LLM）在高密度指令下的表现。 IFScale 解…", "url": "https://wizardpisces.github.io/blog/tech/ai-LLM/Distyl%20AI%20-%20IFScale%20-%20How%20Many%20Instructions%20Can%20LLMs%20Follow%20At%20Once", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-LLM/Distyl%20AI%20-%20IFScale%20-%20How%20Many%20Instructions%20Can%20LLMs%20Follow%20At%20Once", "dateModified": "2026-03-03T06:23:59.997Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-LLM" ] } }, { "@type": "ListItem", "position": 9, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2026 01 05 EverMemOS: A Self Organizing Memory Operating System for Structured Long Horizon Reasoning 由新加坡国立大学（NUS）和 Shopee 等机构联合发布。简单来说，这篇论文提出了一种 “自组织记忆操作系统” ，旨在解决大语言模型（LL…", "url": "https://wizardpisces.github.io/blog/tech/ai-LLM/EverMemOS%20-%20A%20Self-Organizing%20Memory%20Operating%20System%20for%20Structured%20Long-Horizon%20Reasoning", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-LLM/EverMemOS%20-%20A%20Self-Organizing%20Memory%20Operating%20System%20for%20Structured%20Long-Horizon%20Reasoning", "dateModified": "2026-01-21T03:47:11.286Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-LLM" ] } }, { "@type": "ListItem", "position": 10, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2022 10 11 Mind's Eye （论文标题为《Mind's Eye: Grounded Language Model Reasoning through Simulation》）是由 Google Research 和 DeepMind 等机构的研究人员提出的一种新范式。 简单来说，这篇论文提出了一种 “模拟增强推理” 的方法，旨…", "url": "https://wizardpisces.github.io/blog/tech/ai-LLM/Google%20-%20Mind's%20Eye%20-%20Grounded%20Language%20Model%20Reasoning%20through%20Simulation", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-LLM/Google%20-%20Mind's%20Eye%20-%20Grounded%20Language%20Model%20Reasoning%20through%20Simulation", "dateModified": "2026-01-20T02:47:42.682Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-LLM" ] } }, { "@type": "ListItem", "position": 11, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2023 10 03 Google DeepMind 与 UIUC 联合发布的 《Large Language Models Cannot Self Correct Reasoning Yet》 （arXiv:2310.01798）对 LLM 社区中广受追捧的\"自我纠正\"（Self Correction）能力进行了系统性的批判审视。 简单来说…", "url": "https://wizardpisces.github.io/blog/tech/ai-LLM/Google%20DeepMind%20%26%20UIUC%20-%20Large%20Language%20Models%20Cannot%20Self-Correct%20Reasoning%20Yet", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-LLM/Google%20DeepMind%20%26%20UIUC%20-%20Large%20Language%20Models%20Cannot%20Self-Correct%20Reasoning%20Yet", "dateModified": "2026-03-10T04:07:19.921Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-LLM" ] } }, { "@type": "ListItem", "position": 12, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2024 05 13 MIT CSAIL 团队发布的 The Platonic Representation Hypothesis （论文标题为《The Platonic Representation Hypothesis》）在 AI 理论界引起了深远思考。 简单来说，这篇论文提出了一个大胆的假说： 不同的 AI 模型（无论架构、模态或训练数…", "url": "https://wizardpisces.github.io/blog/tech/ai-LLM/MIT%20-%20The%20Platonic%20Representation%20Hypothesis", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-LLM/MIT%20-%20The%20Platonic%20Representation%20Hypothesis", "dateModified": "2026-03-03T06:24:13.973Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-LLM" ] } }, { "@type": "ListItem", "position": 13, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2024 04 24 微软研究院发布的 《From Local to Global: A Graph RAG Approach to Query Focused Summarization》 （arXiv:2404.16130）用知识图谱替代向量检索，解决了传统 RAG 在\"全局理解类问题\"上的根本缺陷。 简单来说，传统 RAG 只能回答\"找…", "url": "https://wizardpisces.github.io/blog/tech/ai-LLM/Microsoft%20-%20GraphRAG%20-%20From%20Local%20to%20Global", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-LLM/Microsoft%20-%20GraphRAG%20-%20From%20Local%20to%20Global", "dateModified": "2026-03-11T09:41:25.324Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-LLM" ] } }, { "@type": "ListItem", "position": 14, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2024 12 27 南洋理工大学（NTU）与新加坡管理大学（SMU）联合发布的 《Long Context vs. RAG for LLMs: An Evaluation and Revisits》 （arXiv:2501.01880）系统对比了长上下文（Long Context, LC）与 RAG 两种策略，并指出此前评测中的方法论缺陷。…", "url": "https://wizardpisces.github.io/blog/tech/ai-LLM/NTU%20%26%20SMU%20-%20Long%20Context%20vs%20RAG%20for%20LLMs", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-LLM/NTU%20%26%20SMU%20-%20Long%20Context%20vs%20RAG%20for%20LLMs", "dateModified": "2026-03-11T09:43:07.106Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-LLM" ] } }, { "@type": "ListItem", "position": 15, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2020 01 23 OpenAI 发布的 Scaling Laws （论文标题为《Scaling Laws for Neural Language Models》）是深度学习时代最具影响力的实证研究之一。 简单来说，这篇论文发现了一个令人震惊的事实： 语言模型的性能（交叉熵损失）与模型大小、数据量、计算量之间存在精确的幂律关系，横跨七个数量…", "url": "https://wizardpisces.github.io/blog/tech/ai-LLM/OpenAI%20-%20Scaling%20Laws%20for%20Neural%20Language%20Models", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-LLM/OpenAI%20-%20Scaling%20Laws%20for%20Neural%20Language%20Models", "dateModified": "2026-03-10T09:57:40.801Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-LLM" ] } }, { "@type": "ListItem", "position": 16, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2025 09 04 OpenAI 和 Georgia Tech 联合发布的 《Why Language Models Hallucinate》 （arXiv:2509.04664）从统计学和博弈论的角度，对大模型的“幻觉”现象给出了一个颠覆性的解释。 简单来说，这篇论文认为 幻觉（Hallucination） 不仅仅是模型能力的缺陷，更是 …", "url": "https://wizardpisces.github.io/blog/tech/ai-LLM/OpenAI%20-%20Why%20Language%20Models%20Hallucinate", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-LLM/OpenAI%20-%20Why%20Language%20Models%20Hallucinate", "dateModified": "2026-03-03T06:23:45.756Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-LLM" ] } }, { "@type": "ListItem", "position": 17, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2023 12 18 同济大学、复旦大学联合发布的 《Retrieval Augmented Generation for Large Language Models: A Survey》 （arXiv:2312.10997）是 RAG 领域最具影响力的综述之一，系统梳理了 100+ 篇 RAG 研究。 但这篇文章不打算复述综述本身——而是提…", "url": "https://wizardpisces.github.io/blog/tech/ai-LLM/Tongji%20University%20-%20Retrieval-Augmented%20Generation%20for%20Large%20Language%20Models%20-%20A%20Survey", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-LLM/Tongji%20University%20-%20Retrieval-Augmented%20Generation%20for%20Large%20Language%20Models%20-%20A%20Survey", "dateModified": "2026-03-10T09:57:51.454Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-LLM" ] } }, { "@type": "ListItem", "position": 18, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2025 05 09 阿里巴巴 Qwen 团队发布的 Gated Attention （论文标题为《Gated Attention for Large Language Models: Non linearity, Sparsity, and Attention Sink Free》）提出了一种 “大道至简” 的架构改进。 简单来说，这篇论文…", "url": "https://wizardpisces.github.io/blog/tech/ai-LLM/paper-transformer/Alibaba%20-%20Gated%20Attention", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-LLM/paper-transformer/Alibaba%20-%20Gated%20Attention", "dateModified": "2026-01-27T10:59:10.424Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-LLM", "paper-transformer" ] } }, { "@type": "ListItem", "position": 19, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2025 03 27 这是一篇来自 Anthropic 的重磅研究（[原文链接），旨在通过“显微镜”级别的工具，解剖大语言模型（LLM）内部运作的 生物学机制 。核心概念是 “思维映射”（Mapping the Mind） 。] [归因图 (Attribution Graphs)] 解决了什么问题？ 核心问题：LLM 是一个不可解释的“黑盒”…", "url": "https://wizardpisces.github.io/blog/tech/ai-LLM/paper-transformer/Anthropic%20-%20On%20the%20Biology%20of%20a%20Large%20Language%20Model", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-LLM/paper-transformer/Anthropic%20-%20On%20the%20Biology%20of%20a%20Large%20Language%20Model", "dateModified": "2026-01-22T04:07:00.543Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-LLM", "paper-transformer" ] } }, { "@type": "ListItem", "position": 20, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2024 12 27 DeepSeek 团队发布的 DeepSeek V3 （论文标题为《DeepSeek V3 Technical Report》）是开源大模型领域的一枚重磅炸弹。 简单来说，这是一个拥有 6710 亿参数 （激活参数 370 亿）的混合专家（MoE）模型。它在性能上对齐了 GPT 4 和 Claude 3.5 Sonnet…", "url": "https://wizardpisces.github.io/blog/tech/ai-LLM/paper-transformer/DeepSeek%20-%20DeepSeek-V3-MLA", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-LLM/paper-transformer/DeepSeek%20-%20DeepSeek-V3-MLA", "dateModified": "2026-03-03T06:21:16.684Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-LLM", "paper-transformer" ] } }, { "@type": "ListItem", "position": 21, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2024 01 11 DeepSeek 发布的 DeepSeekMoE （论文标题为《DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture of Experts Language Models》）是对现有混合专家模型（MoE）架构的一次重要革新。 简单来说，它通过 把大专…", "url": "https://wizardpisces.github.io/blog/tech/ai-LLM/paper-transformer/DeepSeek%20-%20DeepSeekMoE%20-%20Towards%20Ultimate%20Expert%20Specialization", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-LLM/paper-transformer/DeepSeek%20-%20DeepSeekMoE%20-%20Towards%20Ultimate%20Expert%20Specialization", "dateModified": "2026-03-03T06:21:00.256Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-LLM", "paper-transformer" ] } }, { "@type": "ListItem", "position": 22, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2026 02 25 DeepSeek 联合北大、清华发布的 DualPath （论文标题为《DualPath: Breaking the Storage Bandwidth Bottleneck in Agentic LLM Inference》）直击当下 AI Agent 推理的核心瓶颈。 简单来说，这篇论文发现在多轮 Agent 推理场…", "url": "https://wizardpisces.github.io/blog/tech/ai-LLM/paper-transformer/DeepSeek%20-%20DualPath%20-%20Breaking%20the%20Storage%20Bandwidth%20Bottleneck%20in%20Agentic%20LLM%20Inference", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-LLM/paper-transformer/DeepSeek%20-%20DualPath%20-%20Breaking%20the%20Storage%20Bandwidth%20Bottleneck%20in%20Agentic%20LLM%20Inference", "dateModified": "2026-03-03T06:20:12.224Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-LLM", "paper-transformer" ] } }, { "@type": "ListItem", "position": 23, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2026 01 12 DeepSeek 发布的 Engram （论文标题为《Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models》）在 AI 圈引起了很大关注。 简单来说，这篇论文提出了一种 \"条件存储\"（Conditi…", "url": "https://wizardpisces.github.io/blog/tech/ai-LLM/paper-transformer/DeepSeek%20-%20Engram%20-%20Conditional%20Memory%20via%20Scalable%20Lookup", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-LLM/paper-transformer/DeepSeek%20-%20Engram%20-%20Conditional%20Memory%20via%20Scalable%20Lookup", "dateModified": "2026-03-03T06:07:44.009Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-LLM", "paper-transformer" ] } }, { "@type": "ListItem", "position": 24, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2025 02 17 DeepSeek 团队发布的 NSA (Native Sparse Attention) （论文标题为《Native Sparse Attention: Hardware Aligned and Natively Trainable Sparse Attention》）是长文本模型效率优化的一次重要突破。 简单来说，这篇…", "url": "https://wizardpisces.github.io/blog/tech/ai-LLM/paper-transformer/DeepSeek%20-%20Native%20Sparse%20Attention", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-LLM/paper-transformer/DeepSeek%20-%20Native%20Sparse%20Attention", "dateModified": "2026-03-03T06:21:23.608Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-LLM", "paper-transformer" ] } }, { "@type": "ListItem", "position": 25, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2025 01 22 DeepSeek 发布的 DeepSeek R1 （论文标题为《DeepSeek R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning》）在 AI 圈引起了很大关注。 简单来说，这篇论文揭示了如何仅通过 强化学习（RL） 而非大…", "url": "https://wizardpisces.github.io/blog/tech/ai-LLM/paper-transformer/DeepSeek-R1%20-%20Incentivizing%20Reasoning%20Capability%20in%20LLMs%20Via%20RL", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-LLM/paper-transformer/DeepSeek-R1%20-%20Incentivizing%20Reasoning%20Capability%20in%20LLMs%20Via%20RL", "dateModified": "2026-03-03T06:21:29.439Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-LLM", "paper-transformer" ] } }, { "@type": "ListItem", "position": 26, "item": { "@type": "Article", "headline": "输入: 已经过 Embedding 的词向量", "description": "论文发布时间：2017 06 12 Google 团队发布的 Transformer （论文标题为《Attention Is All You Need》）是 AI 发展史上的一座里程碑。 简单来说，这篇论文提出了一种全新的网络架构 Transformer ，彻底抛弃了当时主流的循环神经网络（RNN）和卷积神经网络（CNN），完全依赖 注意力机制（Atten…", "url": "https://wizardpisces.github.io/blog/tech/ai-LLM/paper-transformer/Google%20-%20Attention%20is%20all%20you%20need", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-LLM/paper-transformer/Google%20-%20Attention%20is%20all%20you%20need", "dateModified": "2026-03-23T03:58:11.262Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-LLM", "paper-transformer" ] } }, { "@type": "ListItem", "position": 27, "item": { "@type": "Article", "headline": "可学习权重：所有输入用同一组权重", "description": "论文发布时间：2026 03 16 Kimi Team 发布的 Attention Residuals （论文链接）是 Transformer 架构演进的一次重要创新。这篇论文针对现代大语言模型中标准残差连接的固有缺陷，提出了一种基于注意力机制的动态残差连接方案。 核心问题：PreNorm 残差连接的固有缺陷 标准残差连接的问题是什么？ 在现代 LLM 中…", "url": "https://wizardpisces.github.io/blog/tech/ai-LLM/paper-transformer/Kimi%20Team%20-%20Attention%20Residuals", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-LLM/paper-transformer/Kimi%20Team%20-%20Attention%20Residuals", "dateModified": "2026-03-23T07:29:44.076Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-LLM", "paper-transformer" ] } }, { "@type": "ListItem", "position": 28, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2025 08 13 这篇由 Shanghai AI Laboratory 等团队发布的综述 《Speed Always Wins: A Survey on Efficient Architectures for Large Language Models》 是对后 Transformer 时代架构演进的一次全面“阅兵”。 简单来说，随着大模…", "url": "https://wizardpisces.github.io/blog/tech/ai-LLM/paper-transformer/Shanghai%20AI%20Lab%20-%20Speed%20Always%20Wins", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-LLM/paper-transformer/Shanghai%20AI%20Lab%20-%20Speed%20Always%20Wins", "dateModified": "2026-03-03T06:24:18.020Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-LLM", "paper-transformer" ] } }, { "@type": "ListItem", "position": 29, "item": { "@type": "Article", "headline": "Agent 系统的五个核心范式", "description": "Agent 系统的五个核心范式 从 deepagents 源码出发，提炼 AI Agent 系统设计的通用范式 引言 分析 deepagents 源码后，我发现 Agent 系统的核心挑战可以归结为 五个范式问题 。这些问题不依赖于具体框架——无论你用 LangGraph、CrewAI 还是自建系统，都必须面对它们。 本文不是框架教程，而是试图回答： 构建…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/Agent%20Framework%20%E6%A0%B8%E5%BF%83%E6%9C%BA%E5%88%B6%E6%B7%B1%E5%BA%A6%E8%A7%A3%E6%9E%90%20-%20deepagents%20%E6%A1%88%E4%BE%8B%E7%A0%94%E7%A9%B6", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/Agent%20Framework%20%E6%A0%B8%E5%BF%83%E6%9C%BA%E5%88%B6%E6%B7%B1%E5%BA%A6%E8%A7%A3%E6%9E%90%20-%20deepagents%20%E6%A1%88%E4%BE%8B%E7%A0%94%E7%A9%B6", "dateModified": "2026-04-01T09:19:27.258Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent" ] } }, { "@type": "ListItem", "position": 30, "item": { "@type": "Article", "headline": "目标不是免罪牌：AI 为什么要学会“不该赢”", "description": "文章来源：Anthropic Research 本文讨论 Anthropic 的两篇对齐研究文章： Agentic Misalignment: How LLMs could be insider threats Teaching Claude why 目标不是免罪牌：AI 为什么要学会“不该赢” 过去谈 AI 安全，很容易把问题想成“用户提出危险请求，模型…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/Anthropic%20-%20%E7%9B%AE%E6%A0%87%E4%B8%8D%E6%98%AF%E5%85%8D%E7%BD%AA%E7%89%8C%20-%20Agentic%20Misalignment%20%E4%B8%8E%20Teaching%20Claude%20Why", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/Anthropic%20-%20%E7%9B%AE%E6%A0%87%E4%B8%8D%E6%98%AF%E5%85%8D%E7%BD%AA%E7%89%8C%20-%20Agentic%20Misalignment%20%E4%B8%8E%20Teaching%20Claude%20Why", "dateModified": "2026-05-19T03:50:49.510Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper" ] } }, { "@type": "ListItem", "position": 31, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2026 04 24 中国科学院软件研究所等团队发布的 TraceElephant （论文标题为《Seeing the Whole Elephant: A Benchmark for Failure Attribution in LLM based Multi Agent Systems》）研究的是一个很工程化的问题： 多 Agent 系统失…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/CAS%20-%20TraceElephant%20-%20Seeing%20the%20Whole%20Elephant", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/CAS%20-%20TraceElephant%20-%20Seeing%20the%20Whole%20Elephant", "dateModified": "2026-05-08T07:37:01.600Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper" ] } }, { "@type": "ListItem", "position": 32, "item": { "@type": "Article", "headline": "CMU & Allen AI - Self-Refine - Iterative Refinement with Self-Feedback", "description": "论文发布时间：2023 03 30（NeurIPS 2023） 卡内基梅隆大学和 Allen AI 联合发布的 Self Refine （论文标题为《Self Refine: Iterative Refinement with Self Feedback》，项目地址 selfrefine.info）提出了一个极其简单但有效的思路： 让同一个 LLM 先生成…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/CMU%20%26%20Allen%20AI%20-%20Self-Refine%20-%20Iterative%20Refinement%20with%20Self-Feedback", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/CMU%20%26%20Allen%20AI%20-%20Self-Refine%20-%20Iterative%20Refinement%20with%20Self-Feedback", "dateModified": "2026-03-06T10:19:53.443Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper" ] } }, { "@type": "ListItem", "position": 33, "item": { "@type": "Article", "headline": "EPFL & Copenhagen - Fleet of Agents - Coordinated Problem Solving with LLMs", "description": "论文发布时间：2024 05 10 EPFL 和哥本哈根大学联合发布的 Fleet of Agents（FoA） （论文标题为《Fleet of Agents: Coordinated Problem Solving with Large Language Models》）提出了一种基于 遗传粒子滤波（Genetic Particle Filtering）…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/EPFL%20%26%20Copenhagen%20-%20Fleet%20of%20Agents%20-%20Coordinated%20Problem%20Solving%20with%20LLMs", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/EPFL%20%26%20Copenhagen%20-%20Fleet%20of%20Agents%20-%20Coordinated%20Problem%20Solving%20with%20LLMs", "dateModified": "2026-03-05T07:48:27.390Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper" ] } }, { "@type": "ListItem", "position": 34, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2023 03 31 KAUST（沙特阿卜杜拉国王科技大学）发布的 CAMEL （论文标题为《CAMEL: Communicative Agents for \"Mind\" Exploration of Large Language Model Society》，NeurIPS 2023）是多 Agent 协作领域的早期奠基性工作之一。 简单来…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/KAUST%20-%20CAMEL%20-%20Communicative%20Agents%20for%20Mind%20Exploration%20of%20LLM%20Society", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/KAUST%20-%20CAMEL%20-%20Communicative%20Agents%20for%20Mind%20Exploration%20of%20LLM%20Society", "dateModified": "2026-03-10T07:25:36.117Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper" ] } }, { "@type": "ListItem", "position": 35, "item": { "@type": "Article", "headline": "[From Mind to Machine: The Rise of Manus AI](https://arxiv.org/abs/2505.02024)", "description": "论文发布时间：2025 05 04 From Mind to Machine: The Rise of Manus AI 这篇文章介绍了 Manus AI ，一个试图解决 AI 从“思考（Mind）”到“执行（Machine）”断裂问题的全自主数字智能体。 核心问题：执行赤字 目前的 LLM 虽然推理能力（IQ）很强，但缺乏真正的 执行力（Agency） …", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/Manus%20-%20From%20Mind%20to%20Machine", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/Manus%20-%20From%20Mind%20to%20Machine", "dateModified": "2026-03-05T07:37:40.037Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper" ] } }, { "@type": "ListItem", "position": 36, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2025 12 03 MemOS: A Memory OS for AI System 由上海交通大学、复旦大学等机构联合发布。简单来说，这篇论文提出了一种 “AI 记忆操作系统” 的概念，试图通过像管理电脑内存一样管理大模型的记忆，让 AI 拥有可进化、可迁移的长期记忆能力。 MemOS 解决了什么问题？ 核心问题：当前的大模型缺乏一个统一…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/MemOS%20-%20A%20Memory%20OS%20for%20AI%20System", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/MemOS%20-%20A%20Memory%20OS%20for%20AI%20System", "dateModified": "2026-03-03T06:23:43.595Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper" ] } }, { "@type": "ListItem", "position": 37, "item": { "@type": "Article", "headline": "Meta - Toolformer - Language Models Can Teach Themselves to Use Tools", "description": "论文发布时间：2023 02 09（NeurIPS 2023） Meta AI 发布的 Toolformer （论文标题为《Toolformer: Language Models Can Teach Themselves to Use Tools》）解决了一个看起来矛盾的问题： LLM 在语言理解上碾压传统模型，但在算术、事实查询这些\"简单\"任务上却不如一…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/Meta%20-%20Toolformer%20-%20Language%20Models%20Can%20Teach%20Themselves%20to%20Use%20Tools", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/Meta%20-%20Toolformer%20-%20Language%20Models%20Can%20Teach%20Themselves%20to%20Use%20Tools", "dateModified": "2026-03-06T08:35:47.811Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper" ] } }, { "@type": "ListItem", "position": 38, "item": { "@type": "Article", "headline": "NVIDIA & Caltech - Voyager - An Open-Ended Embodied Agent with LLMs", "description": "论文发布时间：2023 05 25 NVIDIA、Caltech、UT Austin 和 Stanford 联合发布的 VOYAGER （论文标题为《Voyager: An Open Ended Embodied Agent with Large Language Models》）是 LLM 驱动的 Embodied Agent（具身智能体） 领域的标杆性…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/NVIDIA%20%26%20Caltech%20-%20Voyager%20-%20An%20Open-Ended%20Embodied%20Agent%20with%20LLMs", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/NVIDIA%20%26%20Caltech%20-%20Voyager%20-%20An%20Open-Ended%20Embodied%20Agent%20with%20LLMs", "dateModified": "2026-03-04T04:15:32.375Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper" ] } }, { "@type": "ListItem", "position": 39, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2022 10 06 Princeton 和 Google Brain 联合发布的 ReAct （论文标题为《ReAct: Synergizing Reasoning and Acting in Language Models》）是 AI Agent 领域的奠基性工作之一。 简单来说，这篇论文提出了一种让大语言模型（LLM） 同时生成推理轨迹…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/Princeton%20%26%20Google%20-%20ReAct%20-%20Synergizing%20Reasoning%20and%20Acting%20in%20LLMs", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/Princeton%20%26%20Google%20-%20ReAct%20-%20Synergizing%20Reasoning%20and%20Acting%20in%20LLMs", "dateModified": "2026-03-04T02:58:38.060Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper" ] } }, { "@type": "ListItem", "position": 40, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2023 05 17 Princeton 和 Google DeepMind 联合发布的 Tree of Thoughts（ToT） （论文标题为《Tree of Thoughts: Deliberate Problem Solving with Large Language Models》）是 LLM 推理范式的里程碑式工作。 简单来说，这…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/Princeton%20%26%20Google%20-%20Tree%20of%20Thoughts%20-%20Deliberate%20Problem%20Solving%20with%20LLMs", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/Princeton%20%26%20Google%20-%20Tree%20of%20Thoughts%20-%20Deliberate%20Problem%20Solving%20with%20LLMs", "dateModified": "2026-03-04T03:30:01.499Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper" ] } }, { "@type": "ListItem", "position": 41, "item": { "@type": "Article", "headline": "Princeton & Northeastern - Reflexion - Language Agents with Verbal Reinforcement Learning", "description": "论文发布时间：2023 03 20 Princeton 和 Northeastern University 联合发布的 Reflexion （论文标题为《Reflexion: Language Agents with Verbal Reinforcement Learning》）提出了一种全新的 Agent 学习范式—— 不通过梯度更新权重，而是通过\"语言…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/Princeton%20%26%20Northeastern%20-%20Reflexion%20-%20Language%20Agents%20with%20Verbal%20Reinforcement%20Learning", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/Princeton%20%26%20Northeastern%20-%20Reflexion%20-%20Language%20Agents%20with%20Verbal%20Reinforcement%20Learning", "dateModified": "2026-03-04T08:28:18.702Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper" ] } }, { "@type": "ListItem", "position": 42, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2023 09 05 Princeton 大学发布的 CoALA （论文标题为《Cognitive Architectures for Language Agents》）是 AI Agent 领域最重要的 理论框架性 工作之一。 简单来说，这篇论文没有提出新的模型或算法，而是做了一件更有价值的事： 从认知科学和符号 AI 的数十年历史中提取精…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/Princeton%20-%20CoALA%20-%20Cognitive%20Architectures%20for%20Language%20Agents", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/Princeton%20-%20CoALA%20-%20Cognitive%20Architectures%20for%20Language%20Agents", "dateModified": "2026-03-10T10:46:57.866Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper" ] } }, { "@type": "ListItem", "position": 43, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2024 05 06 Princeton 发布的 SWE agent （论文标题为《SWE agent: Agent Computer Interfaces Enable Automated Software Engineering》，NeurIPS 2024）是 AI Agent 应用于真实软件工程任务的里程碑式工作。 简单来说，这篇论文提…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/Princeton%20-%20SWE-agent%20-%20Agent-Computer%20Interfaces%20Enable%20Automated%20Software%20Engineering", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/Princeton%20-%20SWE-agent%20-%20Agent-Computer%20Interfaces%20Enable%20Automated%20Software%20Engineering", "dateModified": "2026-03-10T09:50:42.422Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper" ] } }, { "@type": "ListItem", "position": 44, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2026 01 13 RUC（中国人民大学）等机构联合发布的 Memory in the Age of AI Agents 是一篇关于智能体记忆的重量级综述。 简单来说，这篇论文为 AI Agent 的 记忆系统（Agent Memory） 建立了一套 全景式的“世界地图” ，它不再局限于传统的“长短期记忆”分类，而是从形式、功能和动态三个维…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/RUC%20-%20Memory%20in%20the%20Age%20of%20AI%20Agents", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/RUC%20-%20Memory%20in%20the%20Age%20of%20AI%20Agents", "dateModified": "2026-01-20T09:12:12.647Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper" ] } }, { "@type": "ListItem", "position": 45, "item": { "@type": "Article", "headline": "Stanford & Google - Generative Agents - Interactive Simulacra of Human Behavior", "description": "论文发布时间：2023 04 07（UIST'23 最佳论文） Stanford 和 Google Research 联合发布的 Generative Agents （论文标题为《Generative Agents: Interactive Simulacra of Human Behavior》）是 AI Agent 领域的奠基性工作之一。 简单来说，这…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/Stanford%20%26%20Google%20-%20Generative%20Agents%20-%20Interactive%20Simulacra%20of%20Human%20Behavior", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/Stanford%20%26%20Google%20-%20Generative%20Agents%20-%20Interactive%20Simulacra%20of%20Human%20Behavior", "dateModified": "2026-03-04T04:13:22.938Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper" ] } }, { "@type": "ListItem", "position": 46, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2023 08 07 清华大学联合俄亥俄州立大学、UC Berkeley 发布的 AgentBench （论文标题为《AgentBench: Evaluating LLMs as Agents》）是 AI Agent 评测领域的里程碑工作。 这篇论文提出了 首个系统性的多维度基准测试 ，覆盖 8 个不同的交互式环境，对 29 个 LLM 进行…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/THU%20-%20AgentBench%20-%20Evaluating%20LLMs%20as%20Agents", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/THU%20-%20AgentBench%20-%20Evaluating%20LLMs%20as%20Agents", "dateModified": "2026-03-16T03:26:26.368Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper" ] } }, { "@type": "ListItem", "position": 47, "item": { "@type": "Article", "headline": "UC Berkeley & Microsoft - Gorilla - Large Language Model Connected with Massive APIs", "description": "论文发布时间：2023 05 24 UC Berkeley 和微软研究院联合发布的 Gorilla （论文标题为《Gorilla: Large Language Model Connected with Massive APIs》）解决了一个核心问题： LLM 在调用 API 时严重幻觉——编造不存在的 API、传错参数、用错库。 Gorilla 通过在 …", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/UC%20Berkeley%20%26%20Microsoft%20-%20Gorilla%20-%20Large%20Language%20Model%20Connected%20with%20Massive%20APIs", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/UC%20Berkeley%20%26%20Microsoft%20-%20Gorilla%20-%20Large%20Language%20Model%20Connected%20with%20Massive%20APIs", "dateModified": "2026-03-06T08:39:15.097Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper" ] } }, { "@type": "ListItem", "position": 48, "item": { "@type": "Article", "headline": "UC Berkeley - MemGPT - Towards LLMs as Operating Systems", "description": "论文发布时间：2023 10 12（ICML 2024） UC Berkeley 发布的 MemGPT （论文标题为《MemGPT: Towards LLMs as Operating Systems》，项目地址 research.memgpt.ai）提出了一个核心类比： LLM 的上下文窗口就像计算机的物理内存，大小是固定的——但操作系统通过虚拟内存 +…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/UC%20Berkeley%20-%20MemGPT%20-%20Towards%20LLMs%20as%20Operating%20Systems", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/UC%20Berkeley%20-%20MemGPT%20-%20Towards%20LLMs%20as%20Operating%20Systems", "dateModified": "2026-03-09T04:06:22.228Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper" ] } }, { "@type": "ListItem", "position": 49, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2023 10 06 UIUC（伊利诺伊大学）发布的 LATS （论文标题为《Language Agent Tree Search Unifies Reasoning, Acting, and Planning in Language Models》）是 AI Agent 领域的一项重要整合性工作。 简单来说，这篇论文提出了 第一个将推理（R…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/UIUC%20-%20LATS%20-%20Language%20Agent%20Tree%20Search%20Unifies%20Reasoning%20Acting%20and%20Planning", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/UIUC%20-%20LATS%20-%20Language%20Agent%20Tree%20Search%20Unifies%20Reasoning%20Acting%20and%20Planning", "dateModified": "2026-03-10T06:48:54.140Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper" ] } }, { "@type": "ListItem", "position": 50, "item": { "@type": "Article", "headline": "USTC & Meituan - Skill1 - Unified Evolution of Skill-Augmented Agents via RL", "description": "论文提交：2026 05 07，v3 修订：2026 05 12 Skill1 （论文标题为《Skill1: Unified Evolution of Skill Augmented Agents via Reinforcement Learning》）研究的是： 如何让带技能库的 LLM Agent 真正自我进化 。 一句话说，它不是只让 Agent 学…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/USTC%20%26%20Meituan%20-%20Skill1%20-%20Unified%20Evolution%20of%20Skill-Augmented%20Agents%20via%20RL", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/USTC%20%26%20Meituan%20-%20Skill1%20-%20Unified%20Evolution%20of%20Skill-Augmented%20Agents%20via%20RL", "dateModified": "2026-05-14T10:08:36.505Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper" ] } }, { "@type": "ListItem", "position": 51, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2025 05 19 Uniphore 发布的 Pre Act （论文标题为《Pre Act: Multi Step Planning and Reasoning Improves Acting in LLM Agents》）是对经典 ReAct 范式的重要升级。 简单来说，这篇论文指出 ReAct 的\"一步一想\"模式在复杂多步任务中力不从心…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/Uniphore%20-%20Pre-Act%20-%20Multi-Step%20Planning%20and%20Reasoning%20Improves%20Acting", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/Uniphore%20-%20Pre-Act%20-%20Multi-Step%20Planning%20and%20Reasoning%20Improves%20Acting", "dateModified": "2026-03-04T02:58:01.155Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper" ] } }, { "@type": "ListItem", "position": 52, "item": { "@type": "Article", "headline": "Xu et al - ReWOO - Decoupling Reasoning from Observations for Efficient ALMs", "description": "论文发布时间：2023 05 24 ReWOO （论文标题为《ReWOO: Decoupling Reasoning from Observations for Efficient Augmented Language Models》）解决了一个被普遍忽视的效率问题： ReAct 等主流 Agent 范式在每一步推理时都要把完整的历史 prompt 重新发…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/Xu%20et%20al%20-%20ReWOO%20-%20Decoupling%20Reasoning%20from%20Observations%20for%20Efficient%20ALMs", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/Xu%20et%20al%20-%20ReWOO%20-%20Decoupling%20Reasoning%20from%20Observations%20for%20Efficient%20ALMs", "dateModified": "2026-03-06T10:10:14.151Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper" ] } }, { "@type": "ListItem", "position": 53, "item": { "@type": "Article", "headline": "ZJU & Microsoft - HuggingGPT - Solving AI Tasks with ChatGPT and Hugging Face", "description": "论文发布时间：2023 03 29（NeurIPS 2023） 浙江大学和微软亚洲研究院联合发布的 HuggingGPT （论文标题为《HuggingGPT: Solving AI Tasks with ChatGPT and its Friends in Hugging Face》，项目代号 JARVIS）提出了一个核心理念： 让 LLM 当\"大脑\"，调…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/ZJU%20%26%20Microsoft%20-%20HuggingGPT%20-%20Solving%20AI%20Tasks%20with%20ChatGPT%20and%20Hugging%20Face", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/ZJU%20%26%20Microsoft%20-%20HuggingGPT%20-%20Solving%20AI%20Tasks%20with%20ChatGPT%20and%20Hugging%20Face", "dateModified": "2026-03-06T08:24:10.730Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper" ] } }, { "@type": "ListItem", "position": 54, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2025 09 03，v2 修订：2025 09 04 AgenTracer （论文标题为《AgenTracer: Who Is Inducing Failure in the LLM Agentic Systems?》）研究的是： 如何训练一个模型，自动判断 Agent 系统失败时到底是谁的问题、哪一步的问题。 它和 TraceElepha…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/trace-AgenTracer%20-%20Learning%20Failure%20Attribution%20Through%20Counterfactual%20Traces", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/trace-AgenTracer%20-%20Learning%20Failure%20Attribution%20Through%20Counterfactual%20Traces", "dateModified": "2026-05-08T09:11:40.294Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper" ] } }, { "@type": "ListItem", "position": 55, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2026 03 26 这篇论文的全名是《Rethinking Failure Attribution in Multi Agent Systems: A Multi Perspective Benchmark and Evaluation》。 一句话结论 这篇论文反驳了一个隐含假设： 多 Agent 系统失败时，不一定只有一个唯一根因。 过去…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/trace-Rethinking%20Failure%20Attribution%20in%20MAS", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper/trace-Rethinking%20Failure%20Attribution%20in%20MAS", "dateModified": "2026-05-08T10:08:52.073Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper" ] } }, { "@type": "ListItem", "position": 56, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2026 04 24 金山办公 AI Product Center 发布的 STEM （论文标题为《STEM: Structure Tracing Evidence Mining for Knowledge Graphs Driven Retrieval Augmented Generation》）是一篇面向 KG RAG / 多跳知识图谱问…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper-RAG/Kingsoft%20-%20STEM%20-%20Structure-Tracing%20Evidence%20Mining%20for%20KG-RAG", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper-RAG/Kingsoft%20-%20STEM%20-%20Structure-Tracing%20Evidence%20Mining%20for%20KG-RAG", "dateModified": "2026-04-30T07:23:24.725Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper-RAG" ] } }, { "@type": "ListItem", "position": 57, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2024 01 中国科学技术大学、UCLA 和 Google DeepMind 联合发布的 CRAG （论文标题为《Corrective Retrieval Augmented Generation》）是 RAG 领域的重要进展。 简单来说，这篇论文提出了一种 \"纠正式检索增强生成\" 框架，在检索结果和生成器之间插入一个 轻量级的\"质检员\" …", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper-RAG/USTC%20%26%20UCLA%20%26%20Google%20DeepMind%20-%20CRAG%20-%20Corrective%20Retrieval%20Augmented%20Generation", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper-RAG/USTC%20%26%20UCLA%20%26%20Google%20DeepMind%20-%20CRAG%20-%20Corrective%20Retrieval%20Augmented%20Generation", "dateModified": "2026-03-13T09:09:07.892Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper-RAG" ] } }, { "@type": "ListItem", "position": 58, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2023 10 华盛顿大学、Allen AI 和 IBM Research 联合发布的 Self RAG （论文标题为《Self Reflective Retrieval Augmented Generation: Learning to Retrieve, Generate, and Critique through Self Reflec…", "url": "https://wizardpisces.github.io/blog/tech/ai-agent/paper-RAG/UW%20%26%20Allen%20AI%20-%20Self-RAG%20-%20Learning%20to%20Retrieve%20Generate%20and%20Critique%20through%20Self-Reflection", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-agent/paper-RAG/UW%20%26%20Allen%20AI%20-%20Self-RAG%20-%20Learning%20to%20Retrieve%20Generate%20and%20Critique%20through%20Self-Reflection", "dateModified": "2026-03-13T09:08:31.977Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-agent", "paper-RAG" ] } }, { "@type": "ListItem", "position": 59, "item": { "@type": "Article", "headline": "Claude Code 没装 Langfuse SDK，运行记录为什么仍会出现？", "description": "Claude Code 没装 Langfuse SDK，运行记录为什么仍会出现？ 先在终端里执行： 几秒后打开 Langfuse，页面里出现了一棵完整的调用树：Claude Code 发起了几次模型请求、用了哪些工具、每一步花了多久。 奇怪的是，Claude Code 没有安装 Langfuse SDK，也没有和我们的 Python 应用运行在同一个进程里…", "url": "https://wizardpisces.github.io/blog/tech/ai-%E5%B7%A5%E7%A8%8B%E6%8B%86%E8%A7%A3/Claude%20Code%20%E6%B2%A1%E8%A3%85%20Langfuse%20SDK%EF%BC%8C%E8%BF%90%E8%A1%8C%E8%AE%B0%E5%BD%95%E4%B8%BA%E4%BB%80%E4%B9%88%E4%BB%8D%E4%BC%9A%E5%87%BA%E7%8E%B0%EF%BC%9F", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E5%B7%A5%E7%A8%8B%E6%8B%86%E8%A7%A3/Claude%20Code%20%E6%B2%A1%E8%A3%85%20Langfuse%20SDK%EF%BC%8C%E8%BF%90%E8%A1%8C%E8%AE%B0%E5%BD%95%E4%B8%BA%E4%BB%80%E4%B9%88%E4%BB%8D%E4%BC%9A%E5%87%BA%E7%8E%B0%EF%BC%9F", "dateModified": "2026-08-18T08:22:49.782Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-工程拆解" ] } }, { "@type": "ListItem", "position": 60, "item": { "@type": "Article", "headline": "AAR - 自动化对齐研究的启示与陷阱", "description": "文章发布时间：2026 04 17 Anthropic 在 2026 年 4 月发布了一篇名为 《Automated Alignment Researchers: Using large language models to scale scalable oversight》 的研究博客（同步发表在 Alignment Science blog 的 Aut…", "url": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/AAR%20-%20%E8%87%AA%E5%8A%A8%E5%8C%96%E5%AF%B9%E9%BD%90%E7%A0%94%E7%A9%B6%E7%9A%84%E5%90%AF%E7%A4%BA%E4%B8%8E%E9%99%B7%E9%98%B1", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/AAR%20-%20%E8%87%AA%E5%8A%A8%E5%8C%96%E5%AF%B9%E9%BD%90%E7%A0%94%E7%A9%B6%E7%9A%84%E5%90%AF%E7%A4%BA%E4%B8%8E%E9%99%B7%E9%98%B1", "dateModified": "2026-04-22T04:21:12.578Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-笔记" ] } }, { "@type": "ListItem", "position": 61, "item": { "@type": "Article", "headline": "AI 越省力，人越累：从帕金森定律到 Token 浪费", "description": "AI 越省力，人越累：从帕金森定律到 Token 浪费 1958年，英国历史学家帕金森观察到了一个令人啼笑皆非的现象：大英帝国的官僚机构，正在以一种不可遏制的速度膨胀着。一个不称职的官员，宁可雇用两个更平庸的助手来分担工作，也不愿退位让贤。层层效仿后，每个人都忙得不可开交，但整个机构的效率却江河日下。帕金森由此提炼出一条讽刺的真理——工作会自动膨胀，直到占…", "url": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/AI%20%E5%B8%95%E9%87%91%E6%A3%AE%E5%AE%9A%E5%BE%8B", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/AI%20%E5%B8%95%E9%87%91%E6%A3%AE%E5%AE%9A%E5%BE%8B", "dateModified": "2026-07-29T07:34:42.656Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-笔记" ] } }, { "@type": "ListItem", "position": 62, "item": { "@type": "Article", "headline": "AI 的关键是语料", "description": "AI 的关键是语料 观点：模型行为不是由架构、超参数或优化器选择决定的。它是由你的 数据集决定 的。其他一切只是有效地将计算交付给逼近该数据集的手段。 如果模型的架构、功能、参数，作用是有限的，真正决定性的因素是训练模型的语料。那当发布了一个大模型的时候，也许重点应该放在告诉大家模型用什么语料训练模型。（模型完美，但是语料垃圾，一样不行；反之，模型很平常，…", "url": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/AI%20%E7%9A%84%E5%85%B3%E9%94%AE%E6%98%AF%E8%AF%AD%E6%96%99", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/AI%20%E7%9A%84%E5%85%B3%E9%94%AE%E6%98%AF%E8%AF%AD%E6%96%99", "dateModified": "2025-06-20T11:25:12.917Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-笔记" ] } }, { "@type": "ListItem", "position": 63, "item": { "@type": "Article", "headline": "AI 阻碍新技术传播", "description": "AI 阻碍新技术传播? 推导： 如果大家都使用大模型解决技术问题，而大模型又是基于已有数据（旧技术）训练，新技术的应用会落后，没人使用，会导致网上有关新技术的信息很少，导致 AI 缺乏新技术的训练数据，那新技术就很难被普及？ 模型训练出来的模型进一步偏向于旧技术 这实际上描述了一个 技术固化陷阱 （Technology Lock in by AI），或者说…", "url": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/AI%20%E9%98%BB%E7%A2%8D%E6%96%B0%E6%8A%80%E6%9C%AF%E4%BC%A0%E6%92%AD", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/AI%20%E9%98%BB%E7%A2%8D%E6%96%B0%E6%8A%80%E6%9C%AF%E4%BC%A0%E6%92%AD", "dateModified": "2025-06-20T11:25:12.918Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-笔记" ] } }, { "@type": "ListItem", "position": 64, "item": { "@type": "Article", "headline": "AICR", "description": "AICR 使用 AI 进行 code diff review 的 Prompt 框架思考 \"弱规划，强框架\" 让 Agent 自主决定会更好吗？ 方案 A：强框架 方案 B：自主规划 核心权衡：可靠性 vs 灵活性 为什么 AICR 选择强框架？ 什么场景适合自主规划？ 混合方案：在框架内保留灵活性 最佳实践是 在强框架内嵌入有限的自主性 ： 可视化对比 …", "url": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/AICR%20Agent%20Prompt%20%E5%93%B2%E5%AD%A6", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/AICR%20Agent%20Prompt%20%E5%93%B2%E5%AD%A6", "dateModified": "2025-12-01T04:03:49.926Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-笔记" ] } }, { "@type": "ListItem", "position": 65, "item": { "@type": "Article", "headline": "Agent 进化论：从规则驱动到强化学习的范式跃迁", "description": "Agent 进化论：从规则驱动到强化学习的范式跃迁 在构建 AI Agent（智能体）的实践中，我们常常陷入“Prompt Engineering”和“Hard code Logic”的泥潭：为了让 Agent 即使在边缘情况下也能正确调用工具，我们不得不编写冗长的 System Prompt 和防御性代码。 然而，这种 “基于规则（Rule based）…", "url": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/Agent-Reinforcement-Learning", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/Agent-Reinforcement-Learning", "dateModified": "2025-12-25T07:11:11.649Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-笔记" ] } }, { "@type": "ListItem", "position": 66, "item": { "@type": "Article", "headline": "参考资料", "description": "论文发布时间：2025 07 15 Chain of Thought Monitorability （论文标题为《Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety》）由 Anthropic、Google DeepMind、OpenAI 等多家顶级 AI …", "url": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/Chain%20of%20Thought%20Monitorability%20-%20A%20New%20and%20Fragile%20Opportunity%20for%20AI%20Safety", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/Chain%20of%20Thought%20Monitorability%20-%20A%20New%20and%20Fragile%20Opportunity%20for%20AI%20Safety", "dateModified": "2026-03-03T06:23:45.402Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-笔记" ] } }, { "@type": "ListItem", "position": 67, "item": { "@type": "Article", "headline": "详细解析", "description": "Claude 的 JIT 检索其实就是“在对话时临时做索引”，因此有冷启动成本； 而类似 zilliz/claude context 的 MCP 插件提供“预构建、持久化、高质量索引”，让 Claude 可以跳过 JIT 的昂贵推理步骤，直接走“高质量捷径”。 换句话说： JIT = 模型现算现查（token 贵、慢、有概率错） MCP = 给模型准备好的…", "url": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/Claude%20code%20JIT%20%E7%B4%A2%E5%BC%95%E5%AF%B9%E6%AF%94%E9%A2%84%E6%9E%84%E5%BB%BA%E7%B4%A2%E5%BC%95", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/Claude%20code%20JIT%20%E7%B4%A2%E5%BC%95%E5%AF%B9%E6%AF%94%E9%A2%84%E6%9E%84%E5%BB%BA%E7%B4%A2%E5%BC%95", "dateModified": "2025-11-19T09:30:01.836Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-笔记" ] } }, { "@type": "ListItem", "position": 68, "item": { "@type": "Article", "headline": "ControlNet 与 LoRA 插件", "description": "ControlNet 与 LoRA 插件 ControlNet 和 LoRA 并不是完全相同的东西，尽管它们都与神经网络和图像生成有关。 ControlNet： 作用：用于控制 AI 图像生成。 特点：允许用户对生成的图像进行精细的控制。 应用：在计算机视觉、艺术设计、虚拟现实等领域中非常有用。 示例：用户可以上传线稿，让 AI 帮助填色渲染、控制人物姿态…", "url": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/ControlNet%20%E4%B8%8E%20LoRA%20%E6%8F%92%E4%BB%B6", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/ControlNet%20%E4%B8%8E%20LoRA%20%E6%8F%92%E4%BB%B6", "dateModified": "2025-07-25T02:20:30.119Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-笔记" ] } }, { "@type": "ListItem", "position": 69, "item": { "@type": "Article", "headline": "Ego Reasoning（自我推理）：AI 认知能力的边界与突破", "description": "Ego Reasoning（自我推理）：AI 认知能力的边界与突破 核心概念 字面含义 ：\"ego\" = 自我，Ego Reasoning = 从自我视角出发的有限推理 本质 ：基于\"我知道、相信、感受到\"的信息进行推理，而非从全知视角推理 认知科学层面：Theory of Mind 的前置阶段 三个递进层次 推理类型 定义 示例 Ego Reasonin…", "url": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/Ego%20%E6%8E%A8%E7%90%86", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/Ego%20%E6%8E%A8%E7%90%86", "dateModified": "2025-10-16T10:56:52.715Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-笔记" ] } }, { "@type": "ListItem", "position": 70, "item": { "@type": "Article", "headline": "GPT与开发-TDD", "description": "GPT 是什么？ GPT 是“Generative Pre trained Transformer”的缩写，即生成式预训练变换模型，是一种语言模型，可以执行非常复杂的任务，如回答问题、生成文章和代码，或者翻译文章内容等。 实质：文字接龙（bert 是完形填空） 如何训练： Pre trained （无监督学习） 自动学习各种资料，熟悉各种语义关系 好处：学…", "url": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/GPT%E4%B8%8E%E5%BC%80%E5%8F%91-TDD", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/GPT%E4%B8%8E%E5%BC%80%E5%8F%91-TDD", "dateModified": "2025-06-20T11:25:12.908Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-笔记" ] } }, { "@type": "ListItem", "position": 71, "item": { "@type": "Article", "headline": "LLM的智能来源：核心概念与扩展", "description": "LLM的智能来源：核心概念与扩展 核心三要素： 压缩（Compression） : 大语言模型（LLM）通过训练将海量数据中的语言模式、统计规律和世界知识“压缩”并内化到其参数中。其 泛化能力 和 抽象能力 都是这种高效压缩的直接体现。 组合（Composition） : LLM 能够将学到的语言元素和概念重新组合，生成连贯、有意义的新文本。这种能力是其语…", "url": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/LLM%20%E6%99%BA%E8%83%BD%E6%9D%A5%E6%BA%90", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/LLM%20%E6%99%BA%E8%83%BD%E6%9D%A5%E6%BA%90", "dateModified": "2025-08-04T07:29:38.570Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-笔记" ] } }, { "@type": "ListItem", "position": 72, "item": { "@type": "Article", "headline": "MCP 升级后，AI 产品如何走出聊天框？", "description": "MCP 升级后，AI 产品如何走出聊天框？ MCP 在 2026 年 7 月 28 日发布了新版规范。核心变化是：协议从双向、有状态的交互，转为无状态的请求/响应。 这次升级更像一次架构纠偏：MCP 把原本可选的无状态实践变成协议默认，并补齐了无状态情况下向客户端索取输入的标准流程。 特别关注点： Session 消失：每个请求都要把话说清楚 假设你对 A…", "url": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/MCP%20%E5%8D%87%E7%BA%A7%E5%90%8E%EF%BC%8CAI%20%E4%BA%A7%E5%93%81%E8%83%BD%E7%9C%9F%E6%AD%A3%E8%B5%B0%E5%87%BA%E8%81%8A%E5%A4%A9%E6%A1%86%E5%90%97%EF%BC%9F", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/MCP%20%E5%8D%87%E7%BA%A7%E5%90%8E%EF%BC%8CAI%20%E4%BA%A7%E5%93%81%E8%83%BD%E7%9C%9F%E6%AD%A3%E8%B5%B0%E5%87%BA%E8%81%8A%E5%A4%A9%E6%A1%86%E5%90%97%EF%BC%9F", "dateModified": "2026-07-30T10:16:19.754Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-笔记" ] } }, { "@type": "ListItem", "position": 73, "item": { "@type": "Article", "headline": "[Context Engineering for AI Agents: Lessons from Building Manus](https://manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus)", "description": "Context Engineering for AI Agents: Lessons from Building Manus 这篇文章是 Manus 团队基于构建通用 AI Agent（Manus）的实战经验，分享关于 上下文工程（Context Engineering） 的深度技术文章。 核心思想 ： 在构建复杂的 Agent 系统时， Context …", "url": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/Manus%20-%20Context%20Engineering%20for%20AI%20Agents", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/Manus%20-%20Context%20Engineering%20for%20AI%20Agents", "dateModified": "2026-01-20T02:24:12.544Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-笔记" ] } }, { "@type": "ListItem", "position": 74, "item": { "@type": "Article", "headline": "Prompt 工程的本质", "description": "Prompt 工程的本质：概念空间对齐 核心论断 Prompt 调整不是\"教会 LLM\"，而是 激活 LLM 训练时已固化的最优推理路径 。本质是用最小信息量实现概念空间的精确对齐。 三层对齐机制 表层对齐 （语法触发） 特定术语和格式模式： \"let's think step by step\" 、 \"act as an expert\" 效果：激活特定推理…", "url": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/Prompt%20%E5%B7%A5%E7%A8%8B%E7%9A%84%E6%9C%AC%E8%B4%A8", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/Prompt%20%E5%B7%A5%E7%A8%8B%E7%9A%84%E6%9C%AC%E8%B4%A8", "dateModified": "2026-07-28T01:49:52.328Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-笔记" ] } }, { "@type": "ListItem", "position": 75, "item": { "@type": "Article", "headline": "Promptions", "description": "文章 promptions helps make ai prompting more precise with dynamic ui controls介绍了微软研究院推出的一个名为 \"Promptions\" (Prompt + Options) 的新框架，旨在通过动态的用户界面（UI）控件，让 AI 的提示（Prompting）过程更加精准和高效。 文章核…", "url": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/Promptions", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/Promptions", "dateModified": "2025-12-29T09:38:10.275Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-笔记" ] } }, { "@type": "ListItem", "position": 76, "item": { "@type": "Article", "headline": "问题有多难？一个例子", "description": "跨文档知识更新 场景：两个不同 PRD 中的某些段落描述的是 同一个业务事实/规则 ，但后来新 PRD 更新了描述。RAG 已经把旧文档切成 chunk 存入向量库，这时如何 精准替换或抑制旧知识 。 这与普通\"文档版本更新\"不同，本质问题是： RAG 的索引单位是 chunk，而不是\"知识点\"。 向量数据库只知道\"这个向量和查询向量有多近\"，它不知道两个…", "url": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/RAG%20-%20%E8%B7%A8%E6%96%87%E6%A1%A3%E7%9F%A5%E8%AF%86%E6%9B%B4%E6%96%B0", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/RAG%20-%20%E8%B7%A8%E6%96%87%E6%A1%A3%E7%9F%A5%E8%AF%86%E6%9B%B4%E6%96%B0", "dateModified": "2026-03-12T03:14:39.474Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-笔记" ] } }, { "@type": "ListItem", "position": 77, "item": { "@type": "Article", "headline": "《The Bitter Lesson》", "description": "《The Bitter Lesson》 \"The Bitter Lesson\"是一篇由计算机科学家理查德·萨顿（Richard Sutton）撰写的论文。它提出了一个重要的观点，即在人工智能和机器学习的发展过程中，最重要的教训是利用计算资源和数据量的增长。 文章的几个观念 传统的人工智能方法受限：在过去的几十年中，人们在构建人工智能系统中主要依赖于手动设计…", "url": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/the%20bitter%20lesson", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/the%20bitter%20lesson", "dateModified": "2026-05-29T07:22:27.063Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-笔记" ] } }, { "@type": "ListItem", "position": 78, "item": { "@type": "Article", "headline": "为什么RAG正在变得越来越像“多模型调度系统”？", "description": "这篇文章讲的是： RAG 里用户问题不要直接丢给检索和生成，而要先被解析成一张“执行计划表”，再由 dispatcher 决定后面的检索方式、chunk 处理方式、模型档位和审计记录。 用一个例子串起来： 用户上传一份保险合同，问： What is the annual premium and what are the main exclusions? 年…", "url": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/%E4%B8%BA%E4%BB%80%E4%B9%88RAG%E6%AD%A3%E5%9C%A8%E5%8F%98%E5%BE%97%E8%B6%8A%E6%9D%A5%E8%B6%8A%E5%83%8F%E2%80%9C%E5%A4%9A%E6%A8%A1%E5%9E%8B%E8%B0%83%E5%BA%A6%E7%B3%BB%E7%BB%9F%E2%80%9D%EF%BC%9F", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/%E4%B8%BA%E4%BB%80%E4%B9%88RAG%E6%AD%A3%E5%9C%A8%E5%8F%98%E5%BE%97%E8%B6%8A%E6%9D%A5%E8%B6%8A%E5%83%8F%E2%80%9C%E5%A4%9A%E6%A8%A1%E5%9E%8B%E8%B0%83%E5%BA%A6%E7%B3%BB%E7%BB%9F%E2%80%9D%EF%BC%9F", "dateModified": "2026-06-29T03:43:25.892Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-笔记" ] } }, { "@type": "ListItem", "position": 79, "item": { "@type": "Article", "headline": "可优化的 AI Agent：用 3 个例子讲清 DSPy、TextGrad、AdalFlow 与 Trace/OPTO", "description": "可优化的 AI Agent：用 3 个例子讲清 DSPy、TextGrad、AdalFlow 与 Trace/OPTO 很多团队第一次把 LLM 接进业务时，都会走一条熟悉的路： 1. 写一版 Prompt（“你是一个专业助手…”） 2. 线下跑几条 case，看起来不错 3. 一上线，开始“玄学翻车”：边界条件、长文本、工具失败、格式不稳 如果你经历过这…", "url": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/%E5%8F%AF%E4%BC%98%E5%8C%96%E7%9A%84%20AI%20Agent%EF%BC%9ADSPy%E3%80%81TextGrad%E3%80%81AdalFlow%20%E4%B8%8E%20Trace-OPTO", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/%E5%8F%AF%E4%BC%98%E5%8C%96%E7%9A%84%20AI%20Agent%EF%BC%9ADSPy%E3%80%81TextGrad%E3%80%81AdalFlow%20%E4%B8%8E%20Trace-OPTO", "dateModified": "2026-01-04T04:12:18.969Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-笔记" ] } }, { "@type": "ListItem", "position": 80, "item": { "@type": "Article", "headline": "核心概念：什么是“上下文工程”？", "description": "这篇文章来自 Anthropic 的工程团队，标题为 《Effective context engineering for AI agents》（AI Agent 的高效上下文工程） 。 这不仅是一篇技术文章，更是大模型应用开发范式的一次重要转变宣言： 从“提示词工程（Prompt Engineering）”向“上下文工程（Context Engineer…", "url": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/%E6%9C%89%E6%95%88%E7%9A%84%20AI%20Agent%20%E4%B8%8A%E4%B8%8B%E6%96%87%E5%B7%A5%E7%A8%8B", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/%E6%9C%89%E6%95%88%E7%9A%84%20AI%20Agent%20%E4%B8%8A%E4%B8%8B%E6%96%87%E5%B7%A5%E7%A8%8B", "dateModified": "2025-11-24T09:33:51.667Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-笔记" ] } }, { "@type": "ListItem", "position": 81, "item": { "@type": "Article", "headline": "机器与人类的视角", "description": "机器与人类的视角 人类视角中一切都是符号的排列组合，通过符号沟通并理解世界；万物（人类能够理解的）的符号化都是逐步抽象演变出来的，比如从早期的象形文字，到当代的丰富多彩的多民族语言文字；再比如从 0/1 计算机编码规则，到更加抽象的编程语言；都是从底层的规则定义开始逐步演化组合出更加庞大的抽象符号 机器视角中一切都是像素的分布排列组合，人类通过投喂足够多的…", "url": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/%E6%9C%BA%E5%99%A8%E4%B8%8E%E4%BA%BA%E7%B1%BB%E7%9A%84%E8%A7%86%E8%A7%92", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/%E6%9C%BA%E5%99%A8%E4%B8%8E%E4%BA%BA%E7%B1%BB%E7%9A%84%E8%A7%86%E8%A7%92", "dateModified": "2025-06-20T11:25:12.919Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-笔记" ] } }, { "@type": "ListItem", "position": 82, "item": { "@type": "Article", "headline": "机器学习实践记录", "description": "简介 初次探索解决一个实际 AI 问题，走了很多弯路，后来发现很多方法论都是现成的 业务目标是解决自动化场景遇到的验证码识别 看完本文的收获：你将了解到机器学习在业务中的一种应用场景。分享将从非专业的机器学习开发的视角重现了以业务场景为核心，逐步介绍工具调研、模型调研、训练和部署等过程中所遇到的问题以及解决思路。 问题具体化 识别变长不规则的图片验证码（包…", "url": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/%E6%9C%BA%E5%99%A8%E5%AD%A6%E4%B9%A0%E5%AE%9E%E8%B7%B5%E8%AE%B0%E5%BD%95", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/%E6%9C%BA%E5%99%A8%E5%AD%A6%E4%B9%A0%E5%AE%9E%E8%B7%B5%E8%AE%B0%E5%BD%95", "dateModified": "2025-12-29T09:28:39.671Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-笔记" ] } }, { "@type": "ListItem", "position": 83, "item": { "@type": "Article", "headline": "LSP（Language Server Protocol）", "description": "LSP（Language Server Protocol） Cursor 是在 LSP 基础上构建了智能中间层 Cursor 的 LSP 增强架构 1. 传统 LSP 架构 2. Cursor 的增强架构 具体的智能处理机制 1. 请求预处理（发给 LSP 前） 2. 响应后处理（LSP 返回后） 3. 上下文增强 实际优化示例 1. 智能补全优化 2. …", "url": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/%E7%90%86%E8%A7%A3Cursor%20Indexing", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/%E7%90%86%E8%A7%A3Cursor%20Indexing", "dateModified": "2025-06-30T04:00:18.014Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-笔记" ] } }, { "@type": "ListItem", "position": 84, "item": { "@type": "Article", "headline": "理论", "description": "理论 MIT的这篇论文通过详细的实验和案例研究，挑战了语言是思维基础的传统观点，为人工智能的发展提供了新的思路。未来的AI研究可能需要更多地关注非语言形式的数据和方法，探索语言之外的思维形式。这不仅为大语言模型的改进提供了新的方向，也为实现更强大、更智能的人工智能系统奠定了基础。 as a powerful tool for transmitting kn…", "url": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/%E8%AF%AD%E8%A8%80%E6%98%AF%E4%BA%A4%E6%B5%81%E5%B7%A5%E5%85%B7%E8%80%8C%E4%B8%8D%E6%98%AF%E6%80%9D%E7%BB%B4%E5%B7%A5%E5%85%B7", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/%E8%AF%AD%E8%A8%80%E6%98%AF%E4%BA%A4%E6%B5%81%E5%B7%A5%E5%85%B7%E8%80%8C%E4%B8%8D%E6%98%AF%E6%80%9D%E7%BB%B4%E5%B7%A5%E5%85%B7", "dateModified": "2025-06-20T11:25:12.919Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-笔记" ] } }, { "@type": "ListItem", "position": 85, "item": { "@type": "Article", "headline": "软件2.0", "description": "软件2.0 原文 概念：软件2.0通常指的是那些在其核心集成了机器学习模型的软件系统。这些系统的行为不仅仅是由传统的编程代码决定的，而是由数据驱动的模型生成的。 以下是对文章一些观点的理解 融合模块以求最优 Modules can meld into an optimal whole 原文：普通软件通常被分解成多个模块，各个模块中间通过共有函数、API 或…", "url": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/%E8%BD%AF%E4%BB%B6%202.0", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/ai-%E7%AC%94%E8%AE%B0/%E8%BD%AF%E4%BB%B6%202.0", "dateModified": "2025-06-20T11:25:12.920Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "ai-笔记" ] } }, { "@type": "ListItem", "position": 86, "item": { "@type": "Article", "headline": "Go 简介", "description": "Goroutine Goroutine 可以看作对 thread 加的一层抽象，它更轻量级，可以单独执行。因为有了这层抽象，Gopher 不会直接面对 thread。 Goroutine vs threads 内存占用 1. 创建一个 goroutine 的栈内存消耗为 2 KB，实际运行过程中，如果栈空间不够用，会自动进行扩容。创建一个 thread 则…", "url": "https://wizardpisces.github.io/blog/tech/backend/Go%E7%AE%80%E4%BB%8B", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/backend/Go%E7%AE%80%E4%BB%8B", "dateModified": "2025-06-20T11:25:12.930Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "backend" ] } }, { "@type": "ListItem", "position": 87, "item": { "@type": "Article", "headline": "JSON序列化", "description": "问：easyjson 跟 golang 原生 JSON 方法对比 easyjson是一个用于快速进行json序列化和反序列化的工具包，它通过为结构体生成方法来避免使用反射，从而提高性能和效率。 与golang原生的json方法相比，easyjson的性能可以提高2～3倍。 easyjson的使用方法是先安装它，然后为需要序列化的结构体添加注释，然后运行ea…", "url": "https://wizardpisces.github.io/blog/tech/backend/JSON%E5%BA%8F%E5%88%97%E5%8C%96", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/backend/JSON%E5%BA%8F%E5%88%97%E5%8C%96", "dateModified": "2025-06-20T11:25:12.930Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "backend" ] } }, { "@type": "ListItem", "position": 88, "item": { "@type": "Article", "headline": "napi 宏的理解", "description": "napi 宏的理解 宏与 DSL Rust 宏的核心之一是通过生成代码来简化重复任务，而 DSL（领域专用语言）则是在宏的帮助下，为特定领域问题设计的一种更易读、易用的语法。通过宏，可以把 Rust 的代码变得像定制的小语言一样，让开发者专注于问题本身，而不是细节的实现。 以下是几个示例，展示如何用 Rust 宏构建 DSL： 示例 1：构建简单的 HTM…", "url": "https://wizardpisces.github.io/blog/tech/backend/napi%20%E5%AE%8F%E7%9A%84%E7%90%86%E8%A7%A3", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/backend/napi%20%E5%AE%8F%E7%9A%84%E7%90%86%E8%A7%A3", "dateModified": "2025-06-20T11:25:12.931Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "backend" ] } }, { "@type": "ListItem", "position": 89, "item": { "@type": "Article", "headline": "rust Arena内存分配", "description": "Arena 在 Rust 中， arena 是一种内存分配策略，主要用于高性能场景中高效地分配和管理一组对象。相比 Box 、 Rc 、 Arc 和 RefCell 等智能指针， arena 的优势在于它可以集中管理对象的内存，从而避免频繁的分配与释放操作。 什么是 Arena？ Arena （内存池）是一种分配器，通过在一块连续的内存区域中分配多个对象来…", "url": "https://wizardpisces.github.io/blog/tech/backend/rust%20Arena%E5%86%85%E5%AD%98%E5%88%86%E9%85%8D", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/backend/rust%20Arena%E5%86%85%E5%AD%98%E5%88%86%E9%85%8D", "dateModified": "2025-06-20T11:25:12.931Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "backend" ] } }, { "@type": "ListItem", "position": 90, "item": { "@type": "Article", "headline": "作用域规则", "description": "作用域规则 Rust 的作用域规则是其内存安全性的基础之一，主要围绕 所有权 、 借用 和 生命周期 来确保内存的管理安全且高效。以下通过几个简单例子说明 Rust 的作用域规则。 所有权 例 1：变量的所有权与作用域 变量 s 的所有权存在于定义到作用域结束之间。 离开作用域时，Rust 会自动调用 drop ，释放内存。 例 2：所有权的转移（Move…", "url": "https://wizardpisces.github.io/blog/tech/backend/rust%20%E4%BD%9C%E7%94%A8%E5%9F%9F%E8%A7%84%E5%88%99", "mainEntityOfPage": "https://wizardpisces.github.io/blog/tech/backend/rust%20%E4%BD%9C%E7%94%A8%E5%9F%9F%E8%A7%84%E5%88%99", "dateModified": "2025-06-20T11:25:12.932Z", "inLanguage": "zh-CN", "isAccessibleForFree": true, "author": { "@type": "Person", "name": "LZ", "url": "https://wizardpisces.github.io/blog/Introduction" }, "publisher": { "@type": "Organization", "name": "LZ AI NOTE", "url": "https://wizardpisces.github.io/" }, "about": [ "backend" ] } }, { "@type": "ListItem", "position": 91, "item": { "@type": "Article", "headline": "参考", "description": "Rust 的智能指针是一种封装了指针行为的数据结构，它不仅能像普通指针那样访问数据，还附加了更多的功能，比如所有权管理、引用计数或内存自动回收。常见的智能指针有 Box 、 Rc 、 Arc 和 RefCell 。下面通过具体例子讲解这些智能指针及其使用场景。 智能指针 1. Box