https://sinimite.work/tags/llm/Recent content in LLM on SiBloghttps://sinimite.work/images/og-default.svg?v=20260525-210321https://sinimite.work/images/og-default.svg?v=20260525-210321Hugo -- 0.156.0zh-cnWed, 29 Jul 2026 00:00:00 +0900https://sinimite.work/posts/rag-reranking-best-practices-2026/Wed, 29 Jul 2026 00:00:00 +0900https://sinimite.work/posts/rag-reranking-best-practices-2026/系统讲解 RRF、Cross-Encoder、Pairwise、Listwise、Late Interaction、Learning-to-Rank、MMR 与 Set-wise Selection，并说明如何在候选召回、质量、延迟、成本和 Token Budget 之间设计生产级 RAG Rerank 模块。https://sinimite.work/posts/ai-agent-context-compression-best-practices-2026/Sun, 26 Jul 2026 00:00:00 +0000https://sinimite.work/posts/ai-agent-context-compression-best-practices-2026/一份面向生产级 AI Agent 的上下文压缩指南，覆盖上下文退化、工具结果管理、Artifact、任务检查点、供应商原生 Compaction、Hermes Agent 实现与工程评估。https://sinimite.work/posts/gradient-descent-for-llm-training/Sun, 26 Jul 2026 00:00:00 +0000https://sinimite.work/posts/gradient-descent-for-llm-training/面向没有机器学习背景的普通程序员，用直观例子解释大语言模型训练中的梯度下降、损失函数、反向传播、学习率与优化器。https://sinimite.work/posts/opentelemetry-ai-agent-observability-guide/Thu, 16 Jul 2026 21:08:52 +0900https://sinimite.work/posts/opentelemetry-ai-agent-observability-guide/面向 AI 应用开发者的 OpenTelemetry 入门指南，涵盖 Trace、Metric、Log、Profile、Context Propagation、Collector、GenAI 语义约定和生产实践。https://sinimite.work/posts/points-of-the-podcast-language-agents-from-reasoning-to-acting/Sun, 24 May 2026 16:05:10 +0900https://sinimite.work/posts/points-of-the-podcast-language-agents-from-reasoning-to-acting/一篇关于 Language Agents: From Reasoning to Acting podcast 的观后感，从 ReAct、Reflexion、Tree of Thoughts、memory、benchmark 和 ACI 出发，解释 LLM agent 为什么是模型、工具、记忆、环境、评估和 UX 共同构成的工程系统。https://sinimite.work/posts/ai-engineer-skill-value-map-2026/Sun, 03 May 2026 20:09:15 +0900https://sinimite.work/posts/ai-engineer-skill-value-map-2026/一篇面向 AI 应用工程师转型的技能价值地图，说明基础 RAG、prompt engineering 与框架 API 的价值变化，并解释 eval、governance 和 agentic workflow 为什么更值得投入。https://sinimite.work/posts/llm-from-hidden-state-to-token-output/Sat, 02 May 2026 00:00:00 +0000https://sinimite.work/posts/llm-from-hidden-state-to-token-output/为什么 Chain of Thought 有效，为什么 prompt 不是越长越好，为什么模型会幻觉？这些看似无关的工程现象，背后其实是同一个核心机制在不同侧面的体现。本文用一把“万能钥匙”打开 LLM 工程的五扇门。https://sinimite.work/posts/llm-training-for-ai-engineers/Sat, 04 Apr 2026 12:00:00 +0900https://sinimite.work/posts/llm-training-for-ai-engineers/一篇面向 AI 应用工程师的大模型训练全景指南，覆盖预训练、后训练、蒸馏、Reward 设计、Agent 训练与 Harness Engineering，帮助你理解模型能力从何而来，以及这些训练决策如何影响真实应用落地。https://sinimite.work/posts/llm-agentic-evolution/Sat, 28 Mar 2026 00:00:00 +0000https://sinimite.work/posts/llm-agentic-evolution/LLM 正在从'被动回答问题的工具'演化为'自主完成目标的 agent'。这不是模型能力的线性提升，而是使用范式的根本转变。本文从一个转型 AI 工程师的视角，梳理这个转变的本质、技术栈的演化、以及它对工程师意味着什么。https://sinimite.work/posts/entropy-in-llm/Fri, 27 Mar 2026 00:00:00 +0000https://sinimite.work/posts/entropy-in-llm/熵（Entropy）贯穿 LLM 的整个生命周期——训练用交叉熵做 loss，推理用 temperature 控制输出熵，评估用 perplexity 衡量质量，产品层面用 entropy 信号检测幻觉。本文从一个转型 AI 工程师的视角，梳理熵在 LLM 中的完整图谱。https://sinimite.work/posts/agent-model-harness/Wed, 18 Mar 2026 00:17:00 +0900https://sinimite.work/posts/agent-model-harness/从 Agent = Model + Harness 这个公式出发，重新理解 AI 应用工程师的核心工作其实是 harness engineering。https://sinimite.work/posts/prompt-engineering-best-practices-guide/Wed, 11 Mar 2026 11:00:00 +0900https://sinimite.work/posts/prompt-engineering-best-practices-guide/系统梳理 2025-2026 年 Prompt Engineering 与 Context Engineering 的生产级最佳实践，涵盖 Claude、GPT、长上下文、Prompt Chaining、Tool Use 与 Eval 工作流。https://sinimite.work/posts/apple-m5-vs-m4-practical-comparison-ai-engineers/Fri, 06 Mar 2026 08:30:00 +0900https://sinimite.work/posts/apple-m5-vs-m4-practical-comparison-ai-engineers/M5 与 M4 的工程化对比：单核/多核、缓存、统一内存、GPU AI 加速与本地推理体验。https://sinimite.work/posts/llm-api-kv-cache/Thu, 05 Mar 2026 11:09:00 +0900https://sinimite.work/posts/llm-api-kv-cache/KV Cache 是连接「Transformer 理论」和「LLM 工程部署」的一个关键概念。理解它，你就打通了从「模型怎么算」到「模型怎么跑」的最后一环。https://sinimite.work/posts/llm-chain-of-thought-cot/Wed, 04 Mar 2026 21:51:00 +0900https://sinimite.work/posts/llm-chain-of-thought-cot/理解什么是LLM Chain-of-Thought (CoT)，以及如何进行 prompt engineering触发 LLM的Chain-of-Thought (CoT)https://sinimite.work/posts/prompt-engineering-from-concept-to-implementation/Tue, 03 Mar 2026 19:50:00 +0900https://sinimite.work/posts/prompt-engineering-from-concept-to-implementation/提示词工程？也许没有你想的那么简单。https://sinimite.work/posts/understanding-prompt-injection/Tue, 03 Mar 2026 19:50:00 +0900https://sinimite.work/posts/understanding-prompt-injection/了解什么事 LLM的 Prompt注入，以及了解一些最近本的防御措施https://sinimite.work/posts/understanding-transformer-intuition/Tue, 24 Feb 2026 12:00:00 +0900https://sinimite.work/posts/understanding-transformer-intuition/我的第一篇正式博客文章。