Ch01 AI 与 LLM 基础¶
理解大语言模型的内部机制:从 Token 到 Transformer,从预训练到推理
本章收录 489 篇实体,按深度递增排列。
本章导航¶
| Level | 含义 | 篇数 |
|---|---|---|
| ⭐ 入门 | 零基础可读 | 57 |
| ⭐⭐ 工程师 | 需编程基础 | 47 |
| ⭐⭐⭐ 专家 | 需ML基础 | 241 |
| ⭐⭐⭐⭐ 科学家 | 需研究背景 | 100 |
| ⭐⭐⭐⭐⭐ 大师 | 前沿/哲学 | 44 |
导读¶
你每天都在用 ChatGPT、Claude、Gemini,但你知道它内部发生了什么吗?
本章从最基础的概念开始:文字如何变成数字(Tokenization),数字如何获得语义(Embedding),模型如何理解上下文(Attention),以及它为什么能生成看似"智能"的回答(Next-Token Prediction)。
你不需要懂数学,但读完本章后,你应该能回答一个关键问题:当模型给出糟糕的回答时,是"齿轮坏了"还是"你喂错了原料"?
本章还涵盖训练动力学(模型在预训练中并非稳定进化,而是频繁跳跃)、推理优化(投机解码让吞吐提升 4 倍)、以及混合架构(Transformer 不是唯一的答案)。
这是全书的地基——后面讲 Agent、Harness、RAG 时,你会反复回到这里。
本章内容¶
⭐ 入门(57 篇)¶
- 001. 一、关于 Kollab
- 002. Linux Foundation and Industry Leaders Launch Akrites to Defend Critical Open Source Software Against AI-Enabled Cyber Threats
- 003. 用 LiteLLM WebSearch Interception 集成 AWS 托管的 Amazon Bedrock AgentCore Web Search 能力
- 004. You Must Fix Your Asserts
- 005. AutoResearch-LLM:让 Agent 接手 LLM 训练优化
- 006. Marc Andreessen on Builder Culture in the Age of AI | The a16z Show
- 007. Code Simulation for Enterprise Engineering | PlayerZero
- 008. AEO and GEO for AI Overviews, ChatGPT, Claude, Gemini, and Perplexity
- 009. A recent experience with ChatGPT 5.5 Pro | Gowers's Weblog
- 010. Why Internally-Built AI Fails Fund Accounting Audits
- 011. Back up and restore your Amazon EKS cluster resources using Velero | Amazon Web Services
- 012. EntryPoint Hijacking
- 013. Amazon Turns Alexa Into Its Next Storefront
- 014. How Superset built the IDE for AI agents on Vercel
- 015. Useful Memories Become Faulty When Continuously Updated by LLMs
- 016. How AI changes software P&L
- 017. Build Live Translation Apps with gpt-realtime-translate
- 018. GenRec: Netflix LLM 原生推荐排序器
- 019. What happens in the log when an app crashes as it starts up?
- 020. Is This Why Science Advances One Funeral at a Time?
- 021. AI Gateway production index
- 022. Lighthouse Attention
- 023. Redis之父下场,给DeepSeek V4单独造了一台推理引擎
- 024. ScarCruft 游戏平台供应链攻击分析
- 025. Intelligence Per Dollar
- 026. White House cyber official: identity security matters more
- 027. olmo-eval: An evaluation workbench for the model development
- 028. LLM RL中的熵 part 1: 熵的调控
- 029. OpenClaw 完全指南:这可能是全网最新最全的系统化教程了!(3.2W字,建议收藏)
- 030. OpenClaw 完全指南:这可能是全网最新最全的系统化教程了!(3.2W字,建议收藏)
- 031. Announcing Claude Managed Agents on Cloudflare
- 032. Scale Robot Reinforcement Learning with NVIDIA Isaac Lab on Amazon SageMaker AI
- 033. 两万字详解Claude Code源码核心机制
- 034. Ethan He:Cosmos Grok Imagine 潜空间视频 Agent
- 035. Anthropic 博客:Claude Code 大型代码库最佳实践
- 036. Claude Managed Agents 开发者指南
- 037. From siloed data to unified insights: Cross-account Athena Access for Amazon Quick
- 038. Gemini 3.5 Flash: more expensive, but Google plan to use it for everything
- 039. Dumb Ways for an Open Source Project to Die
- 040. 你不知道的 Agent:原理、架构与工程实践
- 041. 打造可靠的 AI 编程环境:Claude Code Hooks 完整开发者指南
- 042. AI Skill 测评指标体系
- 043. From Doer To Director: The AI Mindset Shift
- 044. LLM Wiki / Obsidian Wiki / GBrain 自组织与自进化
- 045. SkillOS: Learning Skill Curation for Self-Evolving Agents
- 046. Agent orchestration
- 047. Opus 4.7 发布:相比 4.6 核心变化与 Claude Code 搭配最佳实践
- 048. OpenAI buys AI consultancy to sell enterprises on its models
- 049. Streaming benchmark and recommendation results to MLflow with Amazon SageMaker AI
- 050. Apple Foundation Models
- 051. How Dropbox uses MCP and Dash to close the design-to-code security gap
- 052. A backdoor in a LinkedIn job offer
- 053. What Job Interviews Taught Me About Kubernetes
- 054. Agentic Code Review
- 055. Why Use App-Level Auth When Every Database Has Auth? (Splunk CVE-2026-20253)
- 056. Anthropic's Zero Trust for AI Agents Sets the Right Test. The Bearer Token Fails It
- 057. LFM2.5-Encoders: Fast Long-Context Inference on CPU
⭐⭐ 工程师(47 篇)¶
- 058. 滴滴国际化客服质检智能化之路:基于 Amazon Bedrock 的多语种多业务线质检实践
- 059. DeepSeek-V4.1-Flash:把 KV Cache 压缩到极限的百万上下文 MoE(技术报告全系统解读)
- 060. Evaluating Netflix Show Synopses with LLM-as-a-Judge
- 061. Moondream Photon: Pipelined Decoding for VLM Inference Optimization
- 062. Repricing of Software Engineering Labor
- 063. 全网最全的Claude Fable 5 省钱攻略都在这了
- 064. Inference cost at scale with napkin math
- 065. AlphaEvolve: A coding agent for scientific and algorithmic discovery
- 066. The Unbearable Cheapness of Open Weight Models – James O'Claire
- 067. Making Claude a chemist
- 068. 国产预训练具身大模型开源:Wall-OSS-0.5零样本上真机,预训练即可部署
- 069. How to Get a 100% Conference Acceptance Rate, The Novee Way: A High-Severity CVE in Leading Call-for-Papers Software
- 070. 很多企业做完 AI PoC,为什么还是上不了生产
- 071. 独家对话罗福莉:AI范式已然巨变!
- 072. 后端架构 AI Friendly 的标准与路径:面向无人值守开发时代的系统重构
- 073. Evals 到底在评什么?一文拆解 AI 评估的三种方法
- 074. 【译】AI 时代真正的护城河,不是大模型
- 075. 腾讯混元 HiLS-Attention:可学习层级稀疏注意力实现无限上下文建模
- 076. 不改模型、不降质量,谷歌让Gemma 4快了3倍:本地跑大模型彻底变天
- 077. 这样的程序员,应该招吗?
- 078. 花费 2 个星期写了 8 篇 OpenClaw 源码拆解文章,我发现90% 的人对龙虾的理解都太表面了,深层次的真相竟然是这个
- 079. ICML 2026|告别「单线程」思维,智能体进化出了原生的并行推理大脑
- 080. GPT-5级推理能力塞进语音模型,OpenAI把同传翻译成本砍穿地板价
- 081. Mollick 实用 AI 选型指南:模型选择·Deep Research·Voice Mode·Agent 工作流
- 082. Skills Anthropic Openai Comparison Frontend Design
- 083. Cognitive Alpha Mining via LLM-Driven Code-Based Evolution
- 084. 原始文章存档
- 085. 开源 Claude Code 本地代码知识图谱:code-review-graph 完整上手攻略
- 086. 拿下1亿美元种子轮sglang团队创立radixark打造下一代开放ai基础设施
- 087. Anthropic puts Claude agents on a meter across its subscriptions
- 088. Getting LLMs Drunk to Find Remote Linux Kernel OOB Writes (and More)
- 089. 清华系团队出手!一张4090即可「爆改」,1.3B小钢炮震撼开源
- 090. Software After AI
- 091. 国产顶尖模型 benchmark 评分那么高,可实际效果为什么差?看完 Anthropic 这篇博客,刷分的因素太单一了
- 092. Trace2Skill 把\"轨迹里的局部经验\"蒸馏成可迁移的 Agent Skills
- 093. Markdown 不会过时
- 094. Shannon & Epiplexity: Finite Compute Information Theory
- 095. Nature | Anthropic:蒸馏过程潜意识传递行为偏好
- 096. Prompt 调试器:A/B 测试模板对比
- 097. 豆包 Seed 2.0 Lite升级:给 Agent 装上眼睛和耳朵
- 098. Karpathy × Boris 访谈:Software 3.0 时代编程完整地图
- 099. 撕开Claude Code真相:让它好用的98.4%,是工程不是AI
- 100. Stochastic Parrot Thought Experiment
- 101. Anthropic 发布 Computer Use 最佳实践
- 102. AWS re:Invent 2026 Agent 技术要点
- 103. Frontier post-training recipe review with Finbarr Timbers
- 104. System Over Model, Tested: Reproducing Mythos’s FreeBSD Find on Local Open-Weight Models
⭐⭐⭐ 专家(241 篇)¶
- 105. 刚刚Opus 4.7发布,相比4.6核心变化,与Claude Code搭配最佳实践
- 106. Nvidia Telco Reasoning Models Nemo
- 107. Anthropic 最新博客:Prompt Caching 是构建 Claude Code 的一切
- 108. AGI 之路,可能从一开始就走错了
- 109. Nvidia Edge First Llms Av Robotics
- 110. OpenTelemetry eBPF Instrumentation (OBI) — 零代码全栈可观测性的内核级实现
- 111. 高德路线规划双路线:MobilityBench(Agent 基准)+ TransitLM(端到端 RLLM)
- 112. 提速4.48倍!哈工大华为新框架让扩散大模型精度无损、推理起飞
- 113. The Google Capital Company
- 114. LLM 缓存原理与实践:从 KV Cache 到 Prefix Caching,Agent 命中率常挂 90% 的机制与工程含义
- 115. 你还在手动调代码,这个工程师的平台已经在自动进化了
- 116. I/O 2026: Welcome to the agentic Gemini era
- 117. Claude Code 1.0.24 工具调用安全事故:静默删 .gitignore 与 Redis flush 复盘
- 118. Backpressure is all you need
- 119. 打造 AI 智能体专属的代码知识库:GitNexus 完整上手攻略
- 120. AI4S 2026 H1 跨学科前沿全景(弦论泰斗、AI 提速百倍、与"该谁负责"之问)
- 121. 纳德拉「Token 资本」论 — Microsoft CEO 的 AI 时代企业战略宣言
- 122. 800行代码实现 Open Claw 的 Tool、消息总线、子Agent管理架构
- 123. 你的AI代码越写越乱,他72小时合了14个PR每个都更好——差距只在一个机制
- 124. Vibe Coding and Agentic Engineering Convergence: Simon Willison Interview
- 125. Codex 重磅升级:Appshots / Goal 毕业 / 锁屏远程操控
- 126. Karpathy's Autoresearch found a 3-year-old bug in our query engine (and improved performance by 11%) - PostHog
- 127. Lossy self-improvement
- 128. Build the Agent or Power the Agent?
- 129. 安全护栏的三域演进 — 阿里云云原生从 Claude Fable 5 提炼的护栏设计原则
- 130. Workflow architecture
- 131. The distillation panic
- 132. Token 退化问题:分词器与后训练数据分布失配
- 133. Building Jaeger’s ClickHouse backend: 8.6× compression on 10 million spans
- 134. GPT-5.6 Preview System Card — Community Detection & Benchmarks
- 135. GPT-5.5:Sign of the Future — Mollick 的模型/Apps/Harnesses 三层框架与 4 提示 PhD 论文实验
- 136. Karpathy LLM Wiki V2:AI 知识管理新范式
- 137. 估值3000亿63家新实验室杀疯了murati贝佐斯集体押注下一代ai
- 138. AI芯片架构:从逻辑门到矩阵乘法
- 139. 别再把上下文当聊天记录
- 140. OpenAI携手五巨头开源革命性超算协议:一举解决超大集群LLM训练不稳定和网络性能难题
- 141. Every Team is Building the Same Cache — TierFS
- 142. Claude思考黑箱终结了!Anthropic 祭出AI读心术:揭秘Claude的隐藏想法!
- 143. Claude Code KAIROS 范式深度解析
- 144. How to Write an Effective Software Design Document
- 145. GLM-5.2: Built for Long-Horizon Tasks
- 146. We've Been Here Before: Decompilers, Fuzzers, and Now AI
- 147. Claude Code 学术文献综述:45 页 SCI 一区级产出
- 148. Mythos for Offensive Security: XBOW's Evaluation
- 149. Agentic Design System - From Chatbot to Orchestration
- 150. Gemma 4 与开源模型成功标准 —— Interconnects 五维评估框架
- 151. Getting more from each token: How Copilot improves context handling and model routing
- 152. 最佳 Claude Code 配置:Andrej Karpathy 的 CLAUDE.md,134+k star了!
- 153. The Inference Shift
- 154. OlmoEarth v1.1: A more efficient family of Earth observation models
- 155. AWS Sagemaker Capacity Aware Inference Fallback
- 156. 阿里Qwen提出Skill-RM:把奖励模型做成可复用Agent Skill
- 157. Building Modelplane on Crossplane
- 158. Problem with Mathematically Proven Claims About Llms
- 159. Nathan Lambert's Mid-2026 Open Models Bets
- 160. Cola DLM:字节跳动连续潜空间扩散语言模型
- 161. Ralph Loop 不够用:长时间 Agent 还缺这 3 件事
- 162. Business intelligence at scale: Key obstacles
- 163. 给"氛围编程"系上安全带:阿里集团 AI 代码评审实践与 Benchmark 开源
- 164. AMD 免费云 GPU 私有化部署 DeepSeek-R1
- 165. Checkmarx Jenkins plugin compromised in new supply chain attack
- 166. DiffusionGemma:扩散式文本生成模型(Google 26B MoE,4× 推理加速)
- 167. Claude Opus 4.7 发布分析
- 168. 所有实验室都怕字节,所有人都在夸DeepSeek!美国研究员36小时中国AI行
- 169. The analytics engineer in 2026: system designer, governance owner, AI context provider
- 170. Codeindex · 让大模型更好地理解你的代码
- 171. OpenSquilla launches open-source AI agent to cut token costs
- 172. 20VC x SaaStr: The Most Aggressive Quarter in American Capitalism, Palantir's Ru
- 173. How To Measure Development Productivity?
- 174. Navigating EU AI Act Requirements for LLM Fine-Tuning
- 175. Anthropic vs DoW (Department of War) 与开源模型的 5-10 年权力均衡
- 176. IMClaw:通过微信/飞书操控ClaudeCode/Codex/GeminiCLI/Pi Agent蜂群
- 177. The inevitable need for an open model consortium
- 178. Claude Opus 5 on Vending-Bench: Best Capitalist or Aligned, Never Both
- 179. AI 友好架构设计
- 180. Yann LeCun 谈 LLM 不是智能与世界模型 JEPA
- 181. Grafana 13.1 release: observability as code updates, extending Grafana Assistant across more data sources, and more
- 182. AI and Liability
- 183. AI Is Writing More Code. Your CI Pipeline Can't Keep Up
- 184. DeepSeek-V4深度拆解:一篇论文同时做了五件大事
- 185. Google's Gemini Omni video model surfaces ahead of I/O debut
- 186. Multi-tenant LLM Analytics 三层安全架构
- 187. Why Drizzle ORM couldn't publish new releases on NPM for a month | vlt /vōlt/
- 188. DeepSeek MoE 并行策略与GPU通信优化
- 189. Anthropic 首次揭秘下一代 Claude 怎么造
- 190. What political censorship looks like inside an LLM's weights — a mechanistic-interpretability study of Qwen 3.5
- 191. CVPR冠军代码开源:小米SVOR破解视频消除三大顽疾,连人带影一键抹除
- 192. A Route to Root in a 4G Industrial Router
- 193. 高德 AI 资产度量与评价体系:三层评估模型 + 离线采集 + 人工反馈闭环
- 194. Anthropic 联创:2028 年实现 AI 自我构建的概率超过 60%
- 195. 刚刚,OpenAI 放出三个语音模型,顺便杀死了「同传」
- 196. ELF: Embedded Language Flows (何恺明)
- 197. Anthropic Economic Index report: Cadences
- 198. Latest open artifacts (#19):Qwen 3.5、GLM 5、MiniMax 2.5 — 中国实验室的前沿推进
- 199. Mollick 的 AI 实用指南:免费 vs 付费·模型选择·深度研究(2025-10)
- 200. 大模型可控新突破:Steering 机制、评估体系与开源落地
- 201. ICLR 2026 Scenethesis:英伟达 & 普渡大学用 Agent 闭环实现文生 3D
- 202. The new AI lock-in
- 203. Nathan Lambert:开源权重安全论的三个认知陷阱
- 204. 柚漫剧 AI 全流程提效拆解
- 205. Decisions and Dollars
- 206. Claude's next enterprise battle is not models: it's the agent control plane
- 207. The Shape of AI: Jaggedness, Bottlenecks and Salients
- 208. LLMs are complicated now
- 209. Freelance Designers Can't Compete With a $20/Month AI Subscription - Here's What Actually Works Now
- 210. AI科研超越人类 — Prime Intellect递归自改进实验
- 211. CTGAN+LLM 组合拳:携程测试数据生成工程方案
- 212. Anthropic's bug-hunting Mythos was greatest marketing stunt ever says curl creator
- 213. GLM-5.3: How Chinese labs keep stride with the frontier
- 214. Claude统治一切!吞下这颗红药丸,焊工也是顶尖程序员
- 215. 欢迎来到 AGI 时代的 AI 治理 — Nathan Lambert 论美国政府强制 Anthropic 限制 Fable/Mythos 访问
- 216. How my non-engineering team at Sentry learned to ship
- 217. BrowserBC:人类轨迹蒸馏为可复用技能,让小模型获得大模型的网页操作能力
- 218. Ghosts of Encryption Past – How we Read All Your Emails in Salesforce Marketing Cloud › Searchlight Cyber
- 219. Language Models and Meaning
- 220. Claude 发布官方报告,承认存在 3 处质量退化问题
- 221. DeepSeek V4 Triton FP4 优化实战
- 222. Products are out, brains are in
- 223. 百度提出 Unlimited OCR:用 Reference Sliding Window Attention 实现长文档 OCR 常量 KV Cache
- 224. Toto 2.0: Time series forecasting enters the scaling era
- 225. Reading today's open-closed performance gap
- 226. ICML 2026 | APO:将多教师冲突转化为动态约束,破解多模态大模型推理对齐难题
- 227. Solution overview
- 228. GPT-5:It Just Does Stuff — Mollick 的主动式 AI 原语
- 229. Everything a Senior Engineer Needs to Know About What's Inside an LLM
- 230. OpenAI 的最强对手,离「AI Windows」又近了一步
- 231. Management as AI superpower
- 232. How far behind are open models? (LessWrong 2026-05)
- 233. Introducing Claude Platform on AWS: Anthropic's native platform, through your AWS account
- 234. Introducing Claude for Small Business
- 235. AI 驱动的裁员没有商业意义 — Gartner 研究
- 236. Explicit vs. Implicit in the Age of Intelligences — Le secrétaire de Fernand
- 237. Claude Code 核心开发者经验:Action Space 设计
- 238. The Code-as-Content Era
- 239. Latest open artifacts (#21): Open model bonanza! Gemma 4, DeepSeek V4, Kimi K2.6, MiMo 2.5, GLM-5.1 & others. On CAISI's V4 assessment.
- 240. Unexpected lessons from an AI-assisted prototyping experiment
- 241. Claude Code 设计原则与对照分析
- 242. Gemma 4 QAT Models: Quantization-Aware Training for Mobile and Edge
- 243. How LLMs Actually Work: 0xkato Transformer Walkthrough
- 244. ontology-driven-dev:把业务语义钉成 AI 原生应用地基的本体建模开发法
- 245. Some ideas for what comes next, May 2026 (Interconnects)
- 246. How to Land a Frontier Lab Job:如何拿到一份前沿实验室的工作
- 247. Recent Developments in LLM Architectures: KV Sharing, mHC, and Compressed Attention
- 248. Claude Code Skills 与 Superpowers 实践
- 249. Microsoft is quietly shopping for an OpenAI replacement
- 250. 告别“氛围编程”:基于 Harness 治理和 SDD 的团队级 AI 研发范式演进与实践
- 251. Cisco Preps For A World Of AI Agent Coworkers, Frontier Model Threats
- 252. Fake OpenAI Privacy Filter Repo Hits #1 on Hugging Face, Draws 244K Downloads
- 253. 给 AI 做工作面试:Mollick 的模型评估方法论
- 254. Codex 自主赚钱:全自动商业闭环实验
- 255. 开放模型生态快报 #22:Zyphra、Cohere、Poolside 扩张开放模型版图
- 256. Identity Behavior & Context: ITDR Solution | Teleport
- 257. OPD 重新审视失败模式与简单修复
- 258. AWS SageMaker 阿塞拜疆语 LLM 训练:BBPE 分词 + FSDP + Liger 三阶段方案
- 259. AI第一次科研竞赛中击败人类!Opus 4.7狂飙2930步创世界纪录
- 260. Cohere North Mini Code -30B MoE Agentic Coding Model
- 261. 阿里云 MSE AI 任务调度 + Agent Sandbox:动态休眠/唤醒 OpenClaw Agent 成本下降 90%+
- 262. 长期运行的 Agent 怎么管:Hermes 治理分层与 5 张卡
- 263. Introducing eve
- 264. Announcing AWS CDK Mixins: Composable Abstractions for AWS Resources | Amazon Web Services
- 265. Claude Code 泄露后的漏网之鱼 claude-code-best 这两个月到底演进了什么
- 266. CVPR 2026 Highlight|让AI像电影人一样「看」视频,8B小模型反超GPT-5与Gemini 3.1 Pro
- 267. 从零构建大语言模型 —— 读完这篇你就懂了
- 268. Laser Acl2026 Latent Superposition Visual Reasoning
- 269. Mellum 2 (JetBrains open-weight 12B MoE code LLM)
- 270. AWS 强化微调:LLM-as-Judge 训练范式
- 271. Self-Filming Guide by Hello World Media
- 272. Running an AI-native engineering org
- 273. 扣子 3.0 离谱更新:把 Codex、Claude Code 拉进一个项目工作?
- 274. Build a self-improving AI PM OS with Claude Code
- 275. The best argument I’ve heard for why AI won't take your job
- 276. Inngest - AI in Production: The 2026 Benchmark Report
- 277. GrowLoop:开放域对话的真人感评测 — 用种子+Rubrics自动生长Benchmark
- 278. Introducing 1-bit and Ternary Bonsai Image 4B: Image Generation for Local Devices
- 279. Build real-time voice streaming applications with Amazon Nova Sonic and WebRTC
- 280. Knowledge Agents: Beat Frontier Models with Better Structure
- 281. 我把 Karpathy 的 AutoResearch 搬到了软件开发领域,效果炸了
- 282. AI GPUs probably live longer than three years
- 283. 3小时蒸发200万:一个AI客服引发的灾难
- 284. CMU Language Models Need Sleep (arxiv 2605.26099):SSM-Attention 睡眠巩固机制
- 285. LLM-Driven Feature Discovery
- 286. TokenSpeed: A Speed-of-Light LLM Inference Engine for AI Agents
- 287. 企微的这些新功能,补齐了AI在你公司的最后一公里
- 288. A Guide to Which AI to Use in the Agentic Era
- 289. Redis agentic AI flowers with Iris
- 290. Prompt Injection as Role Confusion
- 291. 腾讯混元 Hy3 正式版:Agent 能力跃升与多产品落地
- 292. Introducing Gemini Omni
- 293. Architecture & data foundations for AI-powered Search
- 294. 百度智能云新一代AI机密计算实例:从CPU到GPU全链路可信
- 295. We let four AIs run radio stations. Here's what happened. | Andon Labs
- 296. LLM 工作机制与可解释性
- 297. Open Notebook
- 298. 我用 Claude Code 做需求调研,像多了一个产品经理
- 299. Introducing Seer Agent: The answer is already in Sentry. Now you can ask for it.
- 300. Anthropic PM Jess Yan 的三个 Claude Agent 实践
- 301. AutoResearch 异步多 Agent AI 寒武纪新阶段
- 302. Is Software Losing Its Head?
- 303. World’s first native color LiDAR gives machines human-like vision
- 304. Sparse Autoencoders
- 305. 明星开源项目,为什么开始离开 GitHub?
- 306. Fine-tune LLM with Databricks Unity Catalog and Amazon SageMaker AI
- 307. LLM Thonking:推理努力与安全分诊效果的实证研究
- 308. Kimi K2.5 架构创新 — 1T MoE 一层路由 + 三能力跃迁
- 309. gzip 作为语言模型:压缩-预测等价性的信息论探索
- 310. NVIDIA Exemplar Cloud: AI 训练基础设施性能差距诊断
- 311. N-Day Research with AI: Using Ollama and n8n | Nikhil's Cybersec Blog
- 312. Optimizing Models to Be Fast at Codegen
- 313. Open and Closed Models Are on Different Exponentials
- 314. Detecting Misuse with the Claude Compliance API: The Threat Is in the Content
- 315. So You Want to Sell Inference
- 316. OpenAI Quietly Bought Voice-Cloning Startup Weights.gg
- 317. Announcing OpenAI-compatible API support for Amazon SageMaker AI Endpoints
- 318. GPT-5.5 实测:翻车的学霸
- 319. CDP Bridge MCP:让 LLM 操作真实浏览器
- 320. How Open Model Ecosystems Compound
- 321. The Unsustainable Subsidy
- 322. Langsmith Trajectory Evals
- 323. A Bitter Lesson for Data Filtering
- 324. Transformer 的拓扑麻烦:DeepMind 论证状态追踪是架构性缺陷,CoT 只是补丁
- 325. 强化学习如何不用奖励模型提高通用问题推理能力
- 326. Dean Ball on open models and government control
- 327. Open Defense Initiative | depthfirst
- 328. Why Internally-Built AI Fails Fund Accounting Audits
- 329. Mistral OCR 4: SOTA OCR for Document Intelligence
- 330. R1复现:拒绝采样微调加速RL收敛及模型遗忘问题探究
- 331. 清华自进化Skill双星:EmbodiSkill + SkillEvolver
- 332. 腾讯混元 CL-Bench Life:让大模型读懂你的日常生活
- 333. model half-life
- 334. LiteFrame: Efficient Vision Encoders Unlock Frame Scaling in Video LLMs
- 335. Microsoft Open-Sources RAMPART and Clarity to Secure AI Agents During Development
- 336. We Tested DeepSeek V4 Pro and Flash Against Claude Opus 4.7 and Kimi K2.6
- 337. Better Experiments with LLM Evals — A funnel, not a fork | Spotify Engineering
- 338. Self-Taught RLVR 综述
- 339. Qoder 1.0正式发布!从AI IDE迈向智能体自主开发工作台
- 340. Claude Opus 4.7 并不是一次全面升级,甚至部分能力大幅衰退
- 341. Paperclip · AI 公司操作系统 · 第 5 篇(完结)
- 342. 微软 1000 行代码,把 Claude Opus 干翻了 15 分
- 343. DeepSeek-V4-Flash means LLM steering is interesting again
- 344. Mini Shai-Hulud Strikes Again: TanStack + more npm Packages Compromised
- 345. 我给Hermes配了4个Agent
⭐⭐⭐⭐ 科学家(100 篇)¶
- 346. 打造可靠的 AI 编程环境:Claude Code Hooks 完整开发者指南
- 347. Subagents 详解:Claude Code 如何避免上下文污染
- 348. 800行代码实现 Open Claw 的 Tool、消息总线、子Agent管理架构
- 349. 开源 AI 知识管理搭档 Obsidian + Claude Code 完整集成指南
- 350. What comes next with open models
- 351. LLM agent脚手架如何具备自进化能力?——以hermes agent为例
- 352. 滴滴 IBG 智能客服质检系统:3 管线(意图 86% / 合规 90%+ / VOC)+ 企业 LLM 落地方法论
- 353. 天猫新品团队AI编码实战指南(下)
- 354. Trail of Bits: Skill Scanner Bypass 实证研究
- 355. 三器合一(Comet + OpenSpec + Superpowers):用文件系统给 AI 编程上工程纪律 — 术哥源码级深度分析(WHAT/HOW/WHEN-NEXT 分工 + 9 平台 Hook + 上下文压缩 + 多 Agent 协作 + Shell+YAML 工程取舍)
- 356. Anthropic Claude Managed Agents 平台正式发布
- 357. 🧠 The Token Economy pt2: The Intelligence Company Gets Built
- 358. MIRA + MPA:深度原理 AI Scientist 递归自训练打造材料基座模型,40 项实验全面 SOTA
- 359. LLM 主题 = 生成变量 — 别把 LLM 提取的主题当成真实变量(William Gieng 因果推断方法论)
- 360. Memory in the LLM Era: Modular Architectures and Strategies in a Unified Framework
- 361. MSA 稀疏注意力三国杀:NSA / DSA / MoBA / MSA 四方案深度对比
- 362. 750B MoE PD 分离推理:AWS EFA vs 自建 RoCE 通信架构实战对比
- 363. Anthropic Claude Cowork 知识工作 Agent 任务边界 — 5 筛选信号 + 6 阶段工作流 + 6 企业控制点
- 364. AI Agent 架构设计(七):Skills 系统设计(OpenClaw、Claude Code、Hermes Agent 对比)
- 365. Triton L2缓存命中优化矩阵乘法(fp16&int8)详解及性能测试
- 366. Claude 4/5 Sonnet & Opus Release Notes
- 367. Microsoft MXC — 跨 OS 代理代码执行容器:AppContainer/Sandbox/Hyperlight 三层隔离
- 368. 读完这篇你就搞懂 DeepSeek v4 了
- 369. Scaling Laws, Carefully
- 370. 我们刚过了人类最后一个劳动节?AI新职业的八个变化
- 371. Olmo Hybrid and the Hybrid Architecture Wave (2026)
- 372. Three Years from GPT-3 to Gemini 3
- 373. From "System of Record" to "System of Intelligence
- 374. Claude Code vs Hermes — Session 工程师 vs Goal Runtime
- 375. Codex 三件套:6 职位插件 + Sites + Annotations
- 376. Claude Managed Agents 新更新\"专属云\"模式:把Agent的手放回企业内部
- 377. Anthropic 官方 14 种 Skill 设计模式
- 378. Agent Memory 模块化框架与评测:Memory in the LLM Era 4 模块 + 10 方案对比 + 新方法 F1 38.79 + 4 条工程原则
- 379. Claude Code 源码拆解:从启动到多 Agent 扩展层
- 380. Claude Fable 5 — Ethan Mollick hands-on qualitative evaluation
- 381. 快手首个打工人Agent
- 382. Anthropic 最新论文:阻止 AI 叛变的方法(Model Spec Midtraining)
- 383. 条条电路通罗马:大模型可解释性的「唯一机制」可能从一开始就不存在
- 384. Anthropic LLM ATT&CK Navigator: AI-Enabled Cyber Operations
- 385. FrontierCode — Cognition AI 的 PR Mergeability 编码基准
- 386. Gepa Optimize Anything
- 387. DeepSeek V4 Flash & Pro: Million-Token Context and Trillion-Parameter Inference
- 388. Qwen-Image-Flash: Beyond Objective Design — Few-step Distillation Training Recipe
- 389. Tencent AI Infra: Backend Engineer's Guide to AI System Hardware and Software
- 390. 在 RDS PostgreSQL 中实现 RaBitQ 量化
- 391. Claude Code vs Codex 上下文架构:五层压缩管道 vs 容器文件系统
- 392. The 2026 State of AI Traffic & Cyberthreat Benchmark Report - HUMAN Security
- 393. Introducing Claude Tag
- 394. Claude Code Agent Teams 实战:怎么拆任务、控权限、收证据
- 395. Introducing Scheduled Tasks 2.0
- 396. The annotated PyTorch training loop
- 397. LLM-as-a-Verifier: A General-Purpose Verification Framework
- 398. 卡帕西\"LLM Wiki\",到底是什么?——用 Claude + Obsidian 给自己造一个第二大脑的完整拆解
- 399. ChatGPT Memory
- 400. Accelerating Gemini Nano models on Pixel with frozen Multi-Token Prediction
- 401. Tracking TamperedChef Clusters via Certificate and Code Reuse
- 402. POPO (Group Prioritized Off-Policy Optimization):清华 RLVR 训练高效组级回放框架
- 403. YC Spring 2026 全批 196 家公司分析:AI 不再是差异点
- 404. Yoonho Lee: Text Optimization as a Legitimate Learning Mechanism
- 405. iii.dev
- 406. Anthropic 实战分享:如何让 AI Agent 持续工作几天?
- 407. Why Ctrl+V won't paste images in Claude Code on WSL, with a fix
- 408. AI Agent Gateway 架构设计 — OpenClaw/Claude Code/Hermes 三框架对比
- 409. Claude Code 安全审查的隐性盲点:Model Anchoring Bias 实证分析
- 410. ICML 2026 Position Paper — Transformer 图灵完备性高度依赖上下文管理 (RUC 魏哲巍团队)
- 411. The Coming Loop
- 412. Anthropic发布「AI原生创业公司」手册:涵盖全流程四大核心阶段,一人公司法典来了
- 413. Reimagining the mouse pointer for the AI era
- 414. Apple Silicon costs more than OpenRouter
- 415. 小米AI — ICML 2026 论文矩阵(11篇)
- 416. NVIDIA Nemotron 3 Ultra: Hybrid Transformer-Mamba MoE for Agentic AI on SageMaker JumpStart
- 417. Hermes Agent SOUL.md:3 层提示词、14 个内置人格,从源码看身份定制的完整设计
- 418. Claude Code 架构解析
- 419. Gemini 深度导读生成器 Prompt:让 AI 重写而非摘要
- 420. 300万人在存的Claude提示词
- 421. How Claude Code works in large codebases: Best practices and where to start
- 422. kasra.blog LLM Agent 黑客能力实测: USD 1500 投石问路 14 模型
- 423. Scenethesis(ICLR 2026)英伟达 & 普渡大学用 Agent 闭环实现文生 3D
- 424. Prompting Amazon Nova 2 for content moderation
- 425. GPT 5.4 是 Codex 的一次大跨越:四维评估视角与 Agent 战争回归
- 426. never waste a token
- 427. LangChain用Agent做销售获客,3个月转化率提升2.5倍,看完后我发现,国内 Agent 落地的方法都错了
- 428. Mythos首个报告:30天挖出10000致命漏洞
- 429. Gemini 3.5: frontier intelligence with action
- 430. How we made WINDOW JOIN parallel and vectorized
- 431. Domain Expertise Has Always Been the Real Moat
- 432. Inside ChatGPT Search: how web.run and fan-out queries shape results
- 433. 高德工业级能力底座:AI-Native 的端云一体基建
- 434. Claude 代码能力的系统工程解析:Constitutional AI + 可验证奖励 RL + 产品飞轮
- 435. Anthropic Institute《When AI builds itself》深度解读:AI 进入 AI 研发执行层、瓶颈迁移与研发级 Harness(架构师 JiaGouX)
- 436. AI-powered BI with Snowflake and Amazon QuickSight
- 437. Model Size Scaling in 2023-2031
- 438. The recent history of AI in 32 otters
- 439. What Is Software, and Will LLMs Replace It?
- 440. Valkey 为什么这么快?盘点 Valkey 中提升性能的黑科技
- 441. Gemini 3.5 Flash 内置 Computer Use 能力
- 442. Red-Teaming after Mythos — Zico Kolter & Matt Fredrikson, Gray Swan
- 443. RubyGems 暂停新用户注册:DDoS + 恶意包供应链攻击事件
- 444. Graphsignal: LLM Inference Profiler
- 445. Memory Prices - Stanford DAM Interactive Dataset
⭐⭐⭐⭐⭐ 大师(44 篇)¶
- 446. Kimi Work:通用 Agent 战场从云端迁移到本地
- 447. 卡片式对话的协议方案探索和思考
- 448. Agent框架OWL原理详解
- 449. 面向电商直播场景的全模态大模型推理加速方案
- 450. LLM RL中的熵 Part 2: 熵对训练的调控
- 451. Claude Code 源码核心机制详解
- 452. AI Agent 评测实战:5 维指标体系 + L1/L2/L3 准出分级 + 评测集三层设计 + 6 大 Benchmark + LLM-as-Judge 三模式
- 453. Postmortem: TanStack npm supply-chain compromise | TanStack Blog
- 454. PromptQueue + OpenGorilla 集成 — AI-Native 异步任务引擎与自进化认知层
- 455. 2026年最值得关注的15款开发者工具深度解读
- 456. The Shape of the Thing
- 457. GSD 上下文管理工具:用 Plan 约束 Agent 行为边界
- 458. AI Evals 评估方法论
- 459. AI 时代的 Git 版本管理最佳实践
- 460. LangGraph 底层原理:它是怎么把 LLM 变成一台状态机的
- 461. 淘天营销中后台生码工作流最佳实践
- 462. LLM 自我提升系统综述 — Yang 等 113 页四阶段闭环框架(Zesearch NLP Lab)
- 463. OpenClaw与Hermes源码架构对比
- 464. nanobot:4000行极简 Agent 框架架构解析
- 465. NomShub — Cursor 远程隧道利用链:Shell Builtin 沙箱逃逸 + Dev Tunnels 武器化
- 466. CLAUDE.md
- 467. DeepSeek Thinking with Visual Primitives 深度解读
- 468. Gemma 4 Multi Token Prediction Drafters
- 469. Personal AI 工作台:Claude 18 动作框架
- 470. 05—skill-creator 源码深度拆解:LLM Skill 触发率、防过拟合与三 Agent 评审完整指南
- 471. 严格 CSP 下的密码窃取:HTML 注入 + Chrome 自动填充攻击
- 472. Protecting against token theft
- 473. Impeccable:把 AI 前端设计变成可检查的工作流 — 33.4k Star 开源项目深度分析
- 474. Qwen Code Skill Testing Framework: Recording, Playback, and Assertions
- 475. 多模态评估器:MLLM-as-Judge 图文评估
- 476. Rethinking Search as Code Generation
- 477. Anthropic Claude Community 插件仓库劫持事件:SHA 校验如何避免供应链攻击
- 478. AI Infra 入门干货总结:大模型是如何高效推理的
- 479. Anthropic Claude Fable 5 on AWS:内置保护措施的 Mythos 级功能现已推出
- 480. Cline releases open-source agent runtime SDK
- 481. Hermes-Agent 官方 Kanban 深度实测:让商业 CLI 工具当 Orchestrator
- 482. Agentium — 从零实现 Agent 系统的开源框架
- 483. Agent Skill 进阶模式与治理
- 484. Hermes Agent 闭环学习机制
- 485. Natural Language Autoencoders — Anthropic 激活→文字可解释性方法
- 486. 天猫 AI 编程实践:团队知识库 + NPM
- 487. Noam Brown:推理预算应成为AI评估的基础变量
- 488. Tokenomics: the 62.5-minute rule for Claude's cache
- 489. Skills:让 Claude 记住「怎么做」,告别重复教学
本章收束¶
合上这一章,你要带走的不是名词表,而是一副透镜:模型是概率机器,不是知识库。它给出的每个字都是"下一个 token"的赌注——所以糟糕的回答只有两种来源:齿轮本身的能力边界,或你喂错的那把原料。往后各章的每一个工程决策——选哪个模型、给多少上下文、花多少预算——都建立在这副透镜上。