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 913 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 | 155 |
| ⭐⭐ 工程师 | 需编程基础 | 688 |
| ⭐⭐⭐ 专家 | 需ML基础 | 52 |
| ⭐⭐⭐⭐ 科学家 | 需研究背景 | 9 |
| ⭐⭐⭐⭐⭐ 大师 | 前沿/哲学 | 9 |
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¶
- 001. 2026年最值得关注的15款开发者工具深度解读
- 002. 🧠 The Token Economy pt2: The Intelligence Company Gets Built
- 003. 5 Things to Know about the CLARITY Act
- 004. The Google Capital Company
- 005. 2028: Two scenarios for global AI leadership
- 006. What happens in the log when an app crashes as it starts up?
- 007. Backpressure is all you need
- 008. Karpathy's Autoresearch found a 3-year-old bug in our query engine (and improved performance by 11%) - PostHog
- 009. 快手首个打工人Agent
- 010. Runtime Instrumentation of Qt6 Apps with Frida - Part 1: Getting Visibility
- 011. Why I'm leaving GitHub for Forgejo
- 012. Agentic Design System - From Chatbot to Orchestration
- 013. 这个五一节我做了个总结,在 2 年多的创业里,我一共做了 25 个 AI 项目
- 014. 20VC x SaaStr: The Most Aggressive Quarter in American Capitalism, Palantir's Ru
- 015. OpenSquilla launches open-source AI agent to cut token costs
- 016. 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
- 017. NotebookLM
- 018. Anthropic 联创:2028 年实现 AI 自我构建的概率超过 60%
- 019. Nvidia's Jensen Huang bets on this British startup to build 'next frontier' of AI
- 020. The Future of BMW
- 021. IfYouRevokeThisTokenItWillWipeTheComputerOfTheOwner
- 022. 第一批「AI原生」本科生,要毕业了
- 023. Freelance Designers Can't Compete With a $20/Month AI Subscription - Here's What Actually Works Now
- 024. Anthropic创始人行动手册 — 花叔x Claude Code译本
- 025. Claude's next enterprise battle is not models: it's the agent control plane
- 026. Domain Expertise Has Always Been the Real Moat
- 027. Toto 2.0: Time series forecasting enters the scaling era
- 028. You Must Fix Your Asserts
- 029. Introducing Claude for Small Business
- 030. Introducing Claude for Small Business
- 031. Why Ctrl+V won't paste images in Claude Code on WSL, with a fix
- 032. From Doer To Director: The AI Mindset Shift
- 033. Unexpected lessons from an AI-assisted prototyping experiment
- 034. How we made WINDOW JOIN parallel and vectorized
- 035. Cisco Preps For A World Of AI Agent Coworkers, Frontier Model Threats
- 036. Microsoft is quietly shopping for an OpenAI replacement
- 037. How to Land a Frontier Lab Job:如何拿到一份前沿实验室的工作
- 038. CyberSecQwen-4B: Why Defensive Cyber Needs Small, Specialized, Locally-Runnable Models
- 039. RAG vs LLM Wiki 深度对比:企业知识库架构选型指南
- 040. Akamai acquires Israeli AI browser security startup LayerX for $205 million in cash
- 041. AI-driven layoffs aren't making business sense
- 042. Gemini AI (Google)
- 043. Recent Developments in LLM Architectures: KV Sharing, mHC, a
- 044. AI GPUs probably live longer than three years
- 045. Running an AI-native engineering org
- 046. Marc Andreessen on Builder Culture in the Age of AI | The a16z Show
- 047. AEO and GEO for AI Overviews, ChatGPT, Claude, Gemini, and Perplexity
- 048. WorkOS Pipes: Third-party integrations without the headache
- 049. Viktor | Not a tool. A hire.
- 050. Amazon Turns Alexa Into Its Next Storefront
- 051. Back up and restore your Amazon EKS cluster resources using Velero | Amazon Web Services
- 052. ChatGPT官宣26位未来之星
- 053. Igor Babuschkin Seeks Up To $1 Billion For River AI
- 054. Coinbase Becomes Hyperliquid’s Official USDC Treasury Deployer as USDH Sunsets
- 055. 快手首个打工人Agent来了!工作秒变桌面软件:零代码、不烧token
- 056. How AI changes software P&L
- 057. Klarna delivers strong start to 2026 with $1bn revenue and $68m adj. operating profit | Klarna International
- 058. How Superset built the IDE for AI agents on Vercel
- 059. 百度智能云新一代AI机密计算实例:从CPU到GPU全链路可信
- 060. AI-powered honeypots: Turning the tables on malicious AI agents
- 061. Gemini 3.5: frontier intelligence with action
- 062. Malware crew TeamPCP open-sources its Shai-Hulud worm on GitHub
- 063. Private Fintech Has Quietly Become Bigger Than Public Fintech. Now What?
- 064. Be more expressive to close more sales
- 065. 我给Hermes配了4个Agent,真正有用的是这些事
- 066. AI Gateway production index
- 067. PyTorch 2.12 Release Blog – PyTorch
- 068. Vietnam to develop domestic cloud so it can ditch risky overseas operators for government workloads
- 069. Anthropic 联创:2028 年实现 AI 自我构建的概率超过 60%
- 070. Anthropic puts Claude agents on a meter across its subscriptions
- 071. Lighthouse Attention
- 072. Google 与 Amnesty International 合作加大间谍软件检测难度
- 073. Google Debuts Gemini-Focused Updates at I/O 2026
- 074. Grafana GitHub Token Breach Led to Codebase Download and Extortion Attempt
- 075. Intelligence Per Dollar
- 076. Code Simulation for Enterprise Engineering | PlayerZero
- 077. LLM-as-a-Verifier: A General-Purpose Verification Framework
- 078. 强化学习如何不用奖励模型提高通用问题推理能力
- 079. 特斯拉百万年薪招数据标注员,朝九晚五,无需AI经验
- 080. Tenorshare AI Diagrimo - Free AI Diagram Generator Online
- 081. Apple Silicon costs more than OpenRouter
- 082. White House cyber official: identity security matters more
- 083. How I Moved My Digital Stack to Europe
- 084. OpenAI Realtime API 架构首次公开
- 085. R1复现:拒绝采样微调加速RL收敛及模型遗忘问题探究
- 086. Thrive Capital Bets $100 Million on Shopify's AI Future
- 087. Announcing Claude Managed Agents on Cloudflare
- 088. Block leans into its AI future | Payments Dive
- 089. PyTorch 2.12 Release Blog – PyTorch
- 090. NGINX Rift: Achieving NGINX Remote Code Execution via an 18-Year-Old Vulnerability
- 091. We Tested DeepSeek V4 Pro and Flash Against Claude Opus 4.7
- 092. Building is just the beginning: Introducing Discoverability
- 093. 5 Things to Know about the CLARITY Act
- 094. 腾讯混元Hy3-preview发布
- 095. Qoder 1.0正式发布!从AI IDE迈向智能体自主开发工作台
- 096. Grafana GitHub Token Breach Led to Codebase Download and Ext
- 097. Anthropic to share Mythos cyber flaw findings with global finance watchdog
- 098. Agent orchestration
- 099. Gemini 3.5 Flash: more expensive, but Google plan to use it for everything
- 100. Useful Memories Become Faulty When Continuously Updated by LLMs
- 101. Identity Behavior & Context: ITDR Solution | Teleport
- 102. Cline releases open-source agent runtime SDK
- 103. Pipes – WorkOS Docs
- 104. Affirm Maps Road to $100B GMV With Card, AI Commerce and Global Expansion
- 105. Self-Filming Guide by Hello World Media
- 106. Dumb Ways for an Open Source Project to Die
- 107. Why Internally-Built AI Fails Fund Accounting Audits
- 108. What political censorship looks like inside an LLM's weights — a mechanistic-interpretability study of Qwen 3.5
- 109. Is Software Losing Its Head?
- 110. Microsoft for Startups | Microsoft
- 111. GPT-5.5来了!我撤回了退订ChatGPT的决定
- 112. Tokenomics: the 62.5-minute rule for Claude's cache
- 113. Products are out, brains are in
- 114. How to create websites with great UX designs: Principles and examples
- 115. A recent experience with ChatGPT 5.5 Pro | Gowers's Weblog
- 116. Announcing AWS CDK Mixins: Composable Abstractions for AWS Resources | Amazon Web Services
- 117. CEOs of the classroom: Why principals are the key to the AI era
- 118. What Do Your Startup Advisors Say About You? — Charlie O'Donnell - Coach, Author, VC
- 119. Code Simulation for Enterprise Engineering | PlayerZero
- 120. You.com | Download the Guide: Why API Latency Is a Misleading Metric
- 121. Deel's "Accelerate or Die" Moment
- 122. Nearly every enterprise is investing in AI, but only 5% say their data is ready
- 123. Build Live Translation Apps with gpt-realtime-translate
- 124. Square Adds Drive-Thru to the Menu
- 125. Schmoozing Is Dead, Agents Are Hitting 120% of Humans, and Growth Is the Only Thing That
- 126. Introducing Claude Platform on AWS: Anthropic’s native platform, through your AWS account | Amazon Web Services
- 127. 我给Hermes配了4个Agent
- 128. EntryPoint Hijacking
- 129. Apple Foundation Models
- 130. Here's The Rub: We Don't Believe You
- 131. Public Stealth Leaves Opportunity on the Table
- 132. Thread by @ZeusRWA on Thread Reader App
- 133. B2B Email Marketing: What Still Works?
- 134. How Dropbox uses MCP and Dash to close the design-to-code security gap
- 135. Igor Babuschkin Seeks Up To $1 Billion For River AI
- 136. A backdoor in a LinkedIn job offer
- 137. The Oracle and the Firm
- 138. What Job Interviews Taught Me About Kubernetes
- 139. YC Spring 2026 全批 196 家公司分析:AI 不再是差异点
- 140. Every Frame Perfect
- 141. OpenAI三个语音模型发布同传被杀死
- 142. Agentic Code Review
- 143. Vietnam to develop domestic cloud
- 144. I Sold My AI Startup Before Revenue: Here’s What Investors Missed — And Founders Shouldn’t
- 145. The golden rule of Customizable Select
- 146. Is This Why Science Advances One Funeral at a Time?
- 147. Why Use App-Level Auth When Every Database Has Auth? (Splunk CVE-2026-20253)
- 148. Anthropic's Zero Trust for AI Agents Sets the Right Test. The Bearer Token Fails It
- 149. Engineering roles shift from developing code to managing AI
- 150. Inngest - AI in Production: The 2026 Benchmark Report
- 151. Exaforce | Agentic SOC Platform and MDR
- 152. Yum Brands' tech chief on building its 'AI backbone' | CIO Dive
- 153. Open Defense Initiative | depthfirst
- 154. Tether launches developer grants program for local-first AI and payments infrastructure
- 155. Google shipped Gemini 3.1 Flash-Lite in General Availability
- 156. 打造可靠的 AI 编程环境:Claude Code Hooks 完整开发者指南
- 157. 800行代码实现 Open Claw 的 Tool、消息总线、子Agent管理架构
- 158. 开源 AI 知识管理搭档 Obsidian + Claude Code 完整集成指南
- 159. Kimi Work:通用 Agent 战场从云端迁移到本地
- 160. 卡片式对话的协议方案探索和思考
- 161. Agent框架OWL原理详解
- 162. Claude Code 源码核心机制详解
- 163. iii-dev-worker-trigger-function
- 164. LLM agent脚手架如何具备自进化能力?——以hermes agent为例
- 165. User Journey Maps: How UX Teams Turn Friction Into Better Products
- 166. 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
- 167. 深入理解 Claude Code 源码中的 Agent Harness 构建之道
- 168. ChatGPT 官宣 26 位未来之星,他们是:穿墙少年、街头摊贩、盲童的朋友……
- 169. What comes next with open models
- 170. 深入理解 Claude Code 源码中的 Agent Harness 构建之道
- 171. PyTorch 2.12 Release Blog
- 172. 刚刚Opus 4.7发布,相比4.6核心变化,与Claude Code搭配最佳实践
- 173. 天猫新品团队AI编码实战指南(下)
- 174. 基于 Prowler 与 GenAI 构建金融行业智能合规中枢 — OpenMAIC 课堂版
- 175. Trail of Bits: Skill Scanner Bypass 实证研究
- 176. 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)
- 177. AI 硬件寒武纪时刻:百度智能云如何成为爆发催化剂(3 类玩家 + 4 个深度案例)
- 178. nvidia telco reasoning models nemo
- 179. Postmortem: TanStack npm supply-chain compromise | TanStack Blog
- 180. PromptQueue + OpenGorilla 集成 — AI-Native 异步任务引擎与自进化认知层
- 181. Anthropic Claude Managed Agents 平台正式发布
- 182. The Shape of the Thing
- 183. 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
- 184. gsd-get-shit-done-context-management-tool
- 185. 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)
- 186. Anthropic 最新博客:Prompt Caching 是构建 Claude Code 的一切
- 187. AI Evals 评估方法论
- 188. MSA 稀疏注意力三国杀:NSA / DSA / MoBA / MSA 四方案深度对比
- 189. AI 时代的 Git 版本管理最佳实践
- 190. AGI 之路,可能从一开始就走错了
- 191. AI Native 时代 —— 研发组织何去何从
- 192. LangGraph 底层原理:它是怎么把 LLM 变成一台状态机的
- 193. RAG技术框架的演进方向
- 194. nvidia edge first llms av robotics
- 195. claude-opus-47-并不是一次全面升级甚至部分能力大幅衰退nn大家应该在合适的场景下选择使用nn昨晚-opus-47-上线全网又炸了nn我仔细看了下官
- 196. 淘天营销中后台生码工作流最佳实践
- 197. OpenTelemetry eBPF Instrumentation (OBI) — 零代码全栈可观测性的内核级实现
- 198. 750B MoE PD 分离推理:AWS EFA vs 自建 RoCE 通信架构实战对比
- 199. 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
- 200. 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
- 201. MNN-Sana-Edit-V2:端侧运行的图像漫画风编辑大模型
- 202. OpenClaw与Hermes源码架构对比
- 203. Triton L2缓存命中优化矩阵乘法(fp16&int8)详解及性能测试
- 204. Latest open artifacts (#20): New orgs! New types of models! With Nemotron Super, Sarvam, Cohere Transcribe, & others
- 205. GPT-5: It Just Does Stuff
- 206. NomShub — Cursor 远程隧道利用链:Shell Builtin 沙箱逃逸 + Dev Tunnels 武器化
- 207. CLAUDE.md
- 208. nanobot:4000行极简 Agent 框架架构解析
- 209. 你还在手动调代码,这个工程师的平台已经在自动进化了
- 210. Personal AI 工作台:Claude 18 动作框架
- 211. Microsoft MXC — 跨 OS 代理代码执行容器:AppContainer/Sandbox/Hyperlight 三层隔离
- 212. 100 年压缩到 100 天!红杉资本:这就是 AGI
- 213. DeepSeek Thinking with Visual Primitives 深度解读
- 214. gemma 4 multi token prediction drafters
- 215. AI Agent 架构设计(七):Skills 系统设计(OpenClaw、Claude Code、Hermes Agent 对比)
- 216. 读完这篇你就搞懂 DeepSeek v4 了
- 217. model-harness-fit-agent-harness
- 218. 我们刚过了人类最后一个劳动节?AI新职业的八个变化
- 219. Three Years from GPT-3 to Gemini 3
- 220. GBrain — YC CEO Garry Tan 的 Postgres-native AI 第二大脑:5 大设计决策 + 零 LLM 知识图谱 + 8 阶段检索 + Brain⊥Source 正交维度
- 221. From "System of Record" to "System of Intelligence
- 222. Claude Code vs Hermes — Session 工程师 vs Goal Runtime
- 223. Claude Code 1.0.24 工具调用安全事故:静默删 .gitignore 与 Redis flush 复盘
- 224. Fundamental’s Large Tabular Model NEXUS is now available on Amazon SageMaker JumpStart
- 225. 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)
- 226. 打造 AI 智能体专属的代码知识库:GitNexus 完整上手攻略
- 227. Claude 4/5 Sonnet & Opus Release Notes
- 228. 纳德拉「Token 资本」论 — Microsoft CEO 的 AI 时代企业战略宣言
- 229. 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
- 230. How Frontier Teams Are Reinventing AI-Native Development
- 231. Claude Code 源码深度解析(13 核心机制)
- 232. How harnesses and post-training close the open-weight bug-finding gap
- 233. Codex 三件套:6 职位插件 + Sites + Annotations
- 234. Claude Managed Agents 新更新\"专属云\"模式:把Agent的手放回企业内部
- 235. cursor harness model production floor
- 236. [亚马逊AWS官方博客](https://aws.amazon.com/cn/blogs/china/)
- 237. 800行代码实现 Open Claw 的 Tool、消息总线、子Agent管理架构
- 238. 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
- 239. Claude Code 之父最新访谈:编程已经结束、harness 将消失、Claude Code 将只有 100 行代码、loop 才是未来
- 240. Claude Code 源码拆解:从启动到多 Agent 扩展层
- 241. ChatGPT Dreaming V3:长期记忆架构级重构(时效 75.1% / 偏好 71.3% / 算力 -80%)
- 242. Claude Fable 5 — Ethan Mollick hands-on qualitative evaluation
- 243. 很多企业做完 AI PoC,为什么还是上不了生产
- 244. Claude Opus 4.8: The System Card
- 245. 严格 CSP 下的密码窃取:HTML 注入 + Chrome 自动填充攻击
- 246. 清华 AI 自进化组织研究报告:AI 业务资产化与公司形态重构
- 247. Vibe Coding and Agentic Engineering Convergence: Simon Willison Interview
- 248. Claude Code 大型代码库最佳实践
- 249. Evaluating Netflix Show Synopses with LLM-as-a-Judge
- 250. 18岁高中生用AI挖出150万未知天体,首批ChatGPT原住民毕业
- 251. Anthropic 最新论文:阻止 AI 叛变的方法(Model Spec Midtraining)
- 252. 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
- 253. Lossy self-improvement
- 254. Anthropic LLM ATT&CK Navigator: AI-Enabled Cyber Operations
- 255. Anthropic 博客:Claude Code 大型代码库最佳实践
- 256. Memento-Skills — 技能外部记忆让 Agent 自进化(arXiv 2603.18743)
- 257. FrontierCode — Cognition AI 的 PR Mergeability 编码基准
- 258. AGI 之路,可能从一开始就走错了(腾讯研究院·王鹏)
- 259. knowledge-work-plugins拆解:Anthropic官方开源,4 种组件、3 级加载、2 层记忆,纯文件的 AI岗位插件集
- 260. SkillOS: Learning Skill Curation for Self-Evolving Agents
- 261. Codex 重磅升级:Appshots / Goal 毕业 / 锁屏远程操控
- 262. 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
- 263. gepa optimize anything
- 264. Token 退化问题:分词器与后训练数据分布失配
- 265. The distillation panic
- 266. LLMjacking: what these attacks are, and how to protect AI servers
- 267. 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
- 268. Qwen Code Skill Testing Framework: Recording, Playback, and Assertions
- 269. 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
- 270. 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
- 271. karpathy-llm-wiki-v2-2026
- 272. Multimodal
- 273. 刚刚Opus 4.7发布,相比4.6核心变化,与Claude Code搭配最佳实践
- 274. Rethinking Search as Code Generation
- 275. 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
- 276. Protecting against token theft
- 277. Agent Harness Engineering: A Survey — ETCLOVG Taxonomy
- 278. Epiplexity:有限算力信息论
- 279. Claude思考黑箱终结了!Anthropic 祭出AI读心术:揭秘Claude的隐藏想法!
- 280. Anthropic 官方 Agent Harness 平台:Claude Managed Agents 完整指南
- 281. Project Glasswing: what Mythos showed us
- 282. claude-code-kairos-paradigm-2026
- 283. Anthropic 官方 14 种 Skill 设计模式
- 284. 数据级 Harness:架构师 JiaGouX 解读 Anthropic 95% 数据分析与 5 个反直觉边界
- 285. Tencent AI Infra: Backend Engineer's Guide to AI System Hardware and Software
- 286. Claude Code vs Codex 上下文架构:五层压缩管道 vs 容器文件系统
- 287. The 2026 State of AI Traffic & Cyberthreat Benchmark Report - HUMAN Security
- 288. 在 RDS PostgreSQL 中实现 RaBitQ 量化
- 289. Fusedash - Generative Analytics Platform | AI Dashboard Software
- 290. Boris Cherny 新访谈:开发工具正在从 IDE 变成 Agent 控制台
- 291. Mythos for Offensive Security: XBOW's Evaluation
- 292. We've Been Here Before: Decompilers, Fuzzers, and Now AI
- 293. Anthropic Claude Fable 5 on AWS:内置保护措施的 Mythos 级功能现已推出
- 294. Gemma 4 与开源模型成功标准 —— Interconnects 五维评估框架
- 295. 一、关于 Kollab
- 296. Claude Code Agent Teams 实战:怎么拆任务、控权限、收证据
- 297. Google Agentic RAG — Sufficient Context Agent + FramesQA 90.1% Cross-Corpus
- 298. Hermes Agent Skill 系统深度解析
- 299. Cline releases open-source agent runtime SDK
- 300. problem with mathematically proven claims about llms
- 301. aws sagemaker capacity aware inference fallback
- 302. LLM-as-a-Verifier: A General-Purpose Verification Framework
- 303. LLMReaper - DOM Based AI Conversation Exfiltration via Browser Extensions
- 304. Introducing Scheduled Tasks 2.0
- 305. 从多智能体编排到AI自主决策:资损防控体系的架构演进
- 306. 最佳 Claude Code 配置:Andrej Karpathy 的 CLAUDE.md,134+k star了!
- 307. Cola DLM:字节跳动连续潜空间扩散语言模型
- 308. Ralph Loop 不够用:长时间 Agent 还缺这 3 件事
- 309. 复杂度棘轮:AI编程的质量只升不降机制
- 310. 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
- 311. AI Native 时代 —— 研发组织何去何从
- 312. Hermes-Agent 官方 Kanban 深度实测:让商业 CLI 工具当 Orchestrator
- 313. Checkmarx Jenkins plugin compromised in new supply chain attack
- 314. Claude Opus 4.7 发布分析
- 315. ChatGPT Memory
- 316. Agent Harness 解析:智能体架构深度拆解
- 317. 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
- 318. 何恺明首个语言模型:105M参数,不走GPT自回归老路
- 319. Agent Skill 进阶模式与治理
- 320. 现在如何使用 AI:一份快速指南(Ethan Mollick)
- 321. 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
- 322. 开源模型的下一阶段:三类模型分类与生态化路径
- 323. Codeindex · 让大模型更好地理解你的代码
- 324. Google Workspace Updates: Small businesses can now seamlessly import users from Microsoft to Google Workspace
- 325. Agent 记忆注入实战:5 维框架(选什么/放哪里/怎么放/放多少/何时放)+ 4 前沿论文(MemGuide/STITCH/ACE/Lost in the Middle)
- 326. hermes-agent-closed-learning-loop
- 327. RAG 深度解析:分块向量化召回重排才是蒸馏同事 Skill 的关键
- 328. IMClaw:通过微信/飞书操控ClaudeCode/Codex/GeminiCLI/Pi Agent蜂群
- 329. Claude Managed Agents 开发者指南
- 330. Navigating EU AI Act Requirements for LLM Fine-Tuning
- 331. Tracking TamperedChef Clusters via Certificate and Code Reuse
- 332. Google & Amnesty International:联手打击商业间谍软件
- 333. YC 总裁开源了自己亲手写的 AI Agent 大脑,1 周就 1 万点赞。
- 334. Yoonho Lee: Text Optimization as a Legitimate Learning Mechanism
- 335. tmall ai coding practice team knowledge base npm
- 336. hermes-9-module-architecture-winty
- 337. 注定改变历史的一代人
- 338. 腾讯研究院AI速递 20260430
- 339. Yann LeCun 谈 LLM 不是智能与世界模型 JEPA
- 340. 腾讯研究院AI速递 20260506
- 341. iii.dev
- 342. 读完 Claude Code 和 OpenClaw 的 memory 源码,我对\"Agent记忆需要向量数据库\"这件事产生了怀疑……
- 343. DeepSeek视觉原语论文:当所有人在堆图像分辨率时,它在堆「指代精度」!
- 344. Cline releases open-source agent runtime SDK
- 345. Google's Gemini Omni video model surfaces ahead of I/O debut
- 346. 刚刚,OpenAI 放出三个语音模型,顺便杀死了「同传」
- 347. Opus 4.7 发布:相比 4.6 核心变化与 Claude Code 搭配最佳实践
- 348. Agent框架OWL原理详解
- 349. AI Agent Gateway 架构设计 — OpenClaw/Claude Code/Hermes 三框架对比
- 350. 谷歌PM公开:2026开发者五大新技能——问题塑形/上下文设计/审美/编排/判断力
- 351. Claude Code 安全审查的隐性盲点:Model Anchoring Bias 实证分析
- 352. Apple Silicon costs more than OpenRouter
- 353. Claude Code and What Comes Next
- 354. 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
- 355. Nathan Lambert:开源权重安全论的三个认知陷阱
- 356. 柚漫剧 AI 全流程提效拆解
- 357. ICLR 2026 Scenethesis:英伟达 & 普渡大学用 Agent 闭环实现文生 3D
- 358. YC掌门人60天写了60万行代码:gstack开源
- 359. The inevitable need for an open model consortium
- 360. Claude Code 工具设计复盘(官方)
- 361. Introducing the Ettin Reranker Family
- 362. AI科研超越人类 — Prime Intellect递归自改进实验
- 363. Claude Opus 4.7 并不是一次全面升级,甚至部分能力大幅衰退。\n\n大家应该在合适的场景下选择使用。\n\n昨晚 Opus 4.7 上线,全网又炸了。\n\n我仔细看了下官方博客 https://www.anthropic.com/news/claude-opus-4-7 \n\n发现这次的升级和之前有点不太一样。\n\n先说优点吧。\n\n编程:SWE-bench Pro 从 53.4% 涨到 64.3%,这是 Claude 的主战场,新模型不可能退步的。\n\n办公任务:OfficeQA Pro 从 57.1% 干到 80.6%,简单理解就是让它处理 Excel 和 Doc 这些文件更靠谱了。\n\n视觉:图像分辨率从 1568px 拉到 2576px。XBOW 安全视觉测试从 54.5% 到 98.5%,接近满分,这也是这次升级最能打的地方了。\n\n另外还有个非常容易忽略的点,4.7 的指令遵循能力大幅增强了。\n\n官方重点提醒了 — 如果你直接用旧 prompt 切 4.7 可能产生意外结果,可能以前模型会 \"脑补\" 你的意思,现在它直接照做。\n\n接下来我们再看看退步的部分。\n\n首先是长上下文检索能力大幅退步。\n\nMRCR v2 测试,256k 下从 91.9% 掉到 59.2%。1M 下更惨,78.3% 直接回到 32.2%。\n\n你要是喜欢把整本书、整个代码仓库塞进去问问题 — 别用 4.7,继续用 4.6。\n\n网页搜索:BrowseComp 从 83.7% 掉到 79.3%。\n\nAnthropic 也说了,做深度网页搜索,4.6 的 scaling curve 更好。\n\n翻译成人话 — deep research 场景,官方推荐你用 4.6。\n\n然后还有个最容易被忽略的:可能有隐形涨价。\n\nAPI 定价没变,还是 $5/$25。但 Anthropic 换了新 tokenizer。\n\n同一段代码、同一份文档、同一个 prompt,丢给 4.7 要多吃最多 35% 的 token。\n\n官方的解释是:模型更准了,一次过的概率更高,省了来回修改的轮次,所以总成本可能反而低。\n\n逻辑上没毛病。但这个逻辑成立的前提是 — 你的任务恰好落在 4.7 提升明显的场景。\n\n如果你日常做的是知识管理、写方案、数据分析这类提升不大的场景,那就是纯纯多烧 token。\n\n所以怎么选?\n\n写代码、办公自动化、视觉理解,屏幕操作类 Agent → 4.7,直接上。\n\n长文档精确检索、deep research → 4.6,别换。\n\n日常随便用用,考虑成本问题还是 4.6。\n\n一句话总结:Opus 4.7 在编程和视觉上有肉眼可见的飞跃。\n\n但全面升级?谈不上。
- 364. Harness Engineering实践做了一个平台让AI一晚上自动评测和优化你的系统
- 365. Latest open artifacts (#19):Qwen 3.5、GLM 5、MiniMax 2.5 — 中国实验室的前沿推进
- 366. Anthropic's bug-hunting Mythos was greatest marketing stunt ever says curl creator
- 367. claude code harness deep understanding
- 368. Claude统治一切!吞下这颗红药丸,焊工也是顶尖程序员
- 369. 用好 Claude 的 18 个动作:搭一个个人 AI 工作台(Personal Harness)
- 370. Olmo Hybrid and future LLM architectures
- 371. The Coming Loop
- 372. How my non-engineering team at Sentry learned to ship
- 373. Language Models and Meaning
- 374. Yann Dubois × Matt Turck:OpenAI 后训练与强化学习的内部视角
- 375. Claude 发布官方报告,承认存在 3 处质量退化问题
- 376. 300万人在存的Claude提示词
- 377. DeepSeek V4 Triton FP4 优化实战
- 378. Gemini 深度导读生成器 Prompt:让 AI 重写而非摘要
- 379. The Shape of AI: Jaggedness, Bottlenecks and Salients
- 380. kasra.blog LLM Agent 黑客能力实测: USD 1500 投石问路 14 模型
- 381. OpenAI Reasoning Models (o1/o3/o4-mini)
- 382. Reading today's open-closed performance gap
- 383. AI Skill 测评指标体系
- 384. How Claude Code works in large codebases: Best practices and where to start
- 385. Tokenomics: the 62.5-minute rule for Claude's cache
- 386. 视频 RAG 分块策略:停顿 / 滑动窗口 / LLM 主题分块
- 387. Prompting Amazon Nova 2 for content moderation
- 388. GPT 5.4 是 Codex 的一次大跨越:四维评估视角与 Agent 战争回归
- 389. 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
- 390. OpenAI 的最强对手,离「AI Windows」又近了一步
- 391. DeliAutoResearch SKILL:DeepSeek陈德里的自主科研智能体框架
- 392. Introducing Claude Platform on AWS: Anthropic's native platform, through your AWS account
- 393. How far behind are open models? (LessWrong 2026-05)
- 394. 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.
- 395. Explicit vs. Implicit in the Age of Intelligences — Le secrétaire de Fernand
- 396. claude code core developer lessons action space design
- 397. The Code-as-Content Era
- 398. AI 驱动的裁员没有商业意义 — Gartner 研究
- 399. Hermes-Agent Kanban 实测 — 商业 CLI 作为上层 Orchestrator
- 400. Nemotron 3.5 Content Safety: Customizable Multimodal Safety for Global Enterprise
- 401. Gemma 4 QAT Models: Quantization-Aware Training for Mobile and Edge
- 402. Gemini 3.5: frontier intelligence with action
- 403. Claude Code 设计原则与对照分析
- 404. My bets on open models, mid-2026
- 405. Recent Developments in LLM Architectures: KV Sharing, mHC, and Compressed Attention
- 406. Claude Code 七大模块详解
- 407. 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
- 408. claude-code-skills-superpowers-practice
- 409. APO — Autonomous Preference Optimization(ICML 2026)
- 410. Inference cost at scale with napkin math
- 411. Agent Executor, Google's distributed Agent Runtime
- 412. Management as AI superpower
- 413. 高德工业级能力底座:AI-Native 的端云一体基建
- 414. Skills:让 Claude 记住「怎么做」,告别重复教学
- 415. 让 Agent 自主完成任务
- 416. 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
- 417. Scenethesis(ICLR 2026)英伟达 & 普渡大学用 Agent 闭环实现文生 3D
- 418. GPT-5:It Just Does Stuff — Mollick 的主动式 AI 原语
- 419. ds4c deepseek v4 antirez
- 420. 突发!Anthropic拿下马斯克Colossus 1全部算力:Claude要放开用了
- 421. DeepSeek Code Harness:对标 Claude Code 的中国方案
- 422. Microsoft for Startups | Microsoft
- 423. Claude Dispatch and the Power of Interfaces
- 424. 淘天营销中后台 AI 生码工作流最佳实践
- 425. codex-autonomous-earning-money
- 426. Mass Intelligence: AI 普及的拐点
- 427. EntryPoint Hijacking
- 428. AWS SageMaker 阿塞拜疆语 LLM 训练:BBPE 分词 + FSDP + Liger 三阶段方案
- 429. Nvidia embraces role of AI investor, pushing past $40 billion in equity bets this year
- 430. Inside ChatGPT Search: how web.run and fan-out queries shape results
- 431. 你的一人公司品牌部,带着Image-2模型的lovart中文版来了
- 432. LLM Steering 行为引导
- 433. 我把Mac留在家,用手机让TRAE SOLO替我打了一天工
- 434. 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
- 435. Latest Open Artifacts #19:Qwen 3.5·GLM-5·MiniMax M2.5 中国开源模型前沿
- 436. while breathless in stodgy viridian
- 437. aws reinforcement fine tuning llm as judge
- 438. 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
- 439. Xero Announces Integration with Anthropic's Claude
- 440. Cohere North Mini Code -30B MoE Agentic Coding Model
- 441. Mollick AI 进展的 32 只水獭基准
- 442. Mellum 2 (JetBrains open-weight 12B MoE code LLM)
- 443. 奥特曼最险一战:前女CTO当庭翻脸,OpenAI权斗彻底打到台前
- 444. 大厂代码库几百万行,Claude Code怎么跑起来的?Anthropic首次公开全套打法
- 445. Claude Managed Agents 官方 Harness 平台指南
- 446. Introducing eve
- 447. Inngest - AI in Production: The 2026 Benchmark Report
- 448. GPT-5.5实测有点翻车。\n\n写完文章后,我已经拿Codex中的GPT-5.5测了不少长步骤的复杂任务:做动画、做PPT、做网站。\n\n我的感受是这是个非常无聊的学霸,会做题会尽量达成你布置的任务,但做出来的东西就是相当无聊。在指令相同,调用Skill相同的情况下,还是Opus4.7出来的东西更符合我审美。\n\n我猜可能是OpenAI真的在很努力跑分了,所以模型也在做题这事上优化得过度。以及,在GPT-5.5的System Card报告里,GPT-5.5有29%的情况会撒谎说自己完成了不可能完成的编程任务,远高于GPT-5.4和5.3版本。
- 449. Designing with AI: Why Claude Design is not the future of enterprise design
- 450. Build real-time voice streaming applications with Amazon Nova Sonic and WebRTC
- 451. AutoResearch 迁移到软件开发:多 Agent 交叉审核的工程实践
- 452. Model Size Scaling in 2023-2031
- 453. RAG深度解析:分块、向量化、召回、重排,才是"蒸馏同事skill"的关键
- 454. LLMReaper - DOM Based AI Conversation Exfiltration via Browser Extensions
- 455. Why and how to implement an AI asset rationalization strategy
- 456. 别再把上下文当聊天记录
- 457. weve-been-here-before-ai-vulnerability-research
- 458. 不用学AI了!圈内公开的秘密:顶级玩家已开始让AI用AI
- 459. AI-powered BI with Snowflake and Amazon QuickSight
- 460. 给 AI 做工作面试:Mollick 的模型评估方法论
- 461. deepseek-v4深度拆解一篇论文同时做了五件大事
- 462. 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
- 463. GPT-5.5全球首破!0源码盲写程序,编程AI进入新纪元
- 464. 企微的这些新功能,补齐了AI在你公司的最后一公里
- 465. The New Phishing Click: How OAuth Consent Bypasses MFA
- 466. AlphaEvolve: A coding agent for scientific and algorithmic discovery
- 467. Anthropic 最新播客:如何打造下一代 Claude
- 468. Knowledge Agents: Beat Frontier Models with Better Structure
- 469. 别再手写 Skill 了!微软最新研究:像神经网络一样训练 Skill
- 470. Private Fintech Has Quietly Become Bigger Than Public Fintech. Now What?
- 471. Claude Code 动态工作流 8 译本(行小招译注 + Hermes DAG 对比)
- 472. 从语言涌现到协作涌现:如何让 AI 产生高质量决策
- 473. Semis Memo: Supply Chain Inheritance
- 474. A Guide to Which AI to Use in the Agentic Era
- 475. 独家对话罗福莉:AI范式已然巨变!
- 476. LLM 工作机制与可解释性
- 477. Making Claude a chemist
- 478. Americans still feel pessimistic about the economy. When will it change?
- 479. We let four AIs run radio stations. Here's what happened. | Andon Labs
- 480. FastAPI APIRouter:如何正确组织路由
- 481. How to Get a 100% Conference Acceptance Rate, The Novee Way: A High-Severity CVE in Leading Call-for-Papers Software
- 482. 很多企业做完 AI PoC,为什么还是上不了生产
- 483. 2026年最值得关注的15款开发者工具,你用过几个?
- 484. Co-Existence and the End of Co-Intelligence
- 485. Semis Memo: Supply Chain Inheritance
- 486. The distillation panic
- 487. 最新Claude Code创始人:编程已经解决了,Harness重要性持续降低,CC未来只有100行代码
- 488. Haptics design and implementation - Microsoft Design
- 489. GPT-5 Is Here and OpenAI Has Some Tips
- 490. The recent history of AI in 32 otters
- 491. Is Software Losing Its Head?
- 492. Evals 到底在评什么?一文拆解 AI 评估的三种方法
- 493. What Is Software, and Will LLMs Replace It?
- 494. xz 后门两周年:扫描器仍然检测不到什么
- 495. Fine-tune LLM with Databricks Unity Catalog and Amazon SageMaker AI
- 496. Claude Agent SDK Skills:可复用的专业知识体系 — Skill ≠ Tool ≠ Memory,渐进加载 + 4 方分工 + 团队 Skill 库设计
- 497. Garry Tan
- 498. gzip 作为语言模型:压缩-预测等价性的信息论探索
- 499. llm-wiki-obsidian-wiki-gbrain-self-organization-self-evolution
- 500. Prompt Injection as Role Confusion
- 501. OpenSquilla
- 502. 【译】AI 时代真正的护城河,不是大模型
- 503. LLM Thonking:推理努力与安全分诊效果的实证研究
- 504. CHERIoT-Ibex: Closing the door on memory safety vulnerabilities with hardware-enforced protection
- 505. 你不知道的 Agent:原理、架构与工程实践
- 506. 腾讯研究院AI速递 20260506
- 507. 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
- 508. GPT 5.4 Codex 评测:Interconnects 的 Agent 使用体验
- 509. Reasoning lift: What happens to AI visibility when AI thinks harder
- 510. thinking-machines-interaction-models-ai-cold
- 511. 这样的程序员,应该招吗?
- 512. Thinking Machines Lab
- 513. 奥特曼最险一战前女cto当庭翻脸openai权斗彻底打到台前
- 514. 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
- 515. 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
- 516. Improve bot accuracy with Amazon Lex Assisted NLU
- 517. Tracking TamperedChef Clusters via Certificate and Code Reuse
- 518. Introducing 1-bit and Ternary Bonsai Image 4B: Image Generation for Local Devices
- 519. DeepSeek-V4深度拆解:一篇论文同时做了五件大事
- 520. 来自字节跳动TRAE的Harness Engineering指南
- 521. Business intelligence at scale: Key obstacles
- 522. Valkey 为什么这么快?盘点 Valkey 中提升性能的黑科技
- 523. Claude Code HTML Artifact Workflow (IFanR Analysis)
- 524. Anthropic创始人行动手册:打造一家AI-Native创业公司(附36页中文PDF)
- 525. Mollick 实用 AI 选型指南:模型选择·Deep Research·Voice Mode·Agent 工作流
- 526. OpenAI Quietly Bought Voice-Cloning Startup Weights.gg
- 527. Open and Closed Models Are on Different Exponentials
- 528. LLMReaper - DOM Based AI Conversation Exfiltration
- 529. 构建无服务器Kiro调度平台:用Kiro CLI + EventBridge + ECS Fargate实现定时AI任务
- 530. Cognitive Alpha Mining via LLM-Driven Code-Based Evolution
- 531. skills anthropic openai comparison frontend design
- 532. Solution overview
- 533. N-Day Research with AI: Using Ollama and n8n | Nikhil's Cybersec Blog
- 534. Solution overview
- 535. The Minimum Viable Unit of Saleable Software
- 536. Workflow architecture
- 537. 原始文章存档
- 538. 一次构建,随处复用:Python 中的泛型仓库模式
- 539. Announcing OpenAI-compatible API support for Amazon SageMaker AI Endpoints
- 540. 开源 Claude Code 本地代码知识图谱:code-review-graph 完整上手攻略
- 541. Can We Agree on a Storage/Workload Architecture Taxonomy? — Jack Vanlightly
- 542. U.S. Bank shifts critical apps to AWS for AI push | CIO Dive
- 543. 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
- 544. 破案了!为啥ChatGPT老想着「稳稳地接住你」
- 545. Auto-Improving Agent Platform (Ashpreet Bedi)
- 546. ChatGPT默认模型大升级,GPT-5.5 Instant正式上线:新增记忆来源功能
- 547. Gemma 4 and what makes an open model succeed
- 548. CDP Bridge MCP:让 LLM 操作真实浏览器
- 549. Optimizing Models to Be Fast at Codegen
- 550. LLM Wiki 架构解析:Karpathy 的 Markdown 知识库模式
- 551. 腾讯研究院AI速递 20260508
- 552. Redis之父下场,给DeepSeek V4单独造了一台推理引擎
- 553. The Unsustainable Subsidy
- 554. token级,精准控制生成长度:3B模型击败GPT 5.4、Claude
- 555. From siloed data to unified insights: Cross-account Athena Access for Amazon Quick
- 556. 清华系团队出手!一张4090即可「爆改」,1.3B小钢炮震撼开源
- 557. Anthropic puts Claude agents on a meter across its subscriptions
- 558. Claude Code Harness 深度分析
- 559. rigged-game-scarcruft-compromises-gaming-platform-supply-chain-attack
- 560. nvidia gemma 4 edge ai
- 561. User Interviews Guide: How to Interview Users Like a Pro
- 562. 2 小时,0 行手写代码,我用 Claude 做了一个生产级 VSCode 插件
- 563. langsmith trajectory evals
- 564. Meta's AI Chief On AI Beef, New Models And Life With Zuck - EP 71 Alex Wang
- 565. 打造可靠的 AI 编程环境:Claude Code Hooks 完整开发者指南
- 566. Interaction Models: A Scalable Approach to Human-AI Collaboration
- 567. 两篇Harness论文:微软M⋆记忆进化 + 谷歌AutoHarness动作约束
- 568. xai解散但Grok还活着
- 569. 时间序列预测数据增强方法
- 570. OpenAI发布新一代实时语音模型,能够像人说话一样进行推理、翻译和转录
- 571. Claude Code 架构解析
- 572. stochastic parrot deep mystery llms
- 573. So You Want to Sell Inference
- 574. Getting LLMs Drunk to Find Remote Linux Kernel OOB Writes (and More)
- 575. 商汤开源 SenseNova-U1:一个模型,同时「看懂」和「画懂」
- 576. 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
- 577. Transformer 的拓扑麻烦:DeepMind 论证状态追踪是架构性缺陷,CoT 只是补丁
- 578. GitHub Copilot Individual Plans: Flex Allotments in Pro and Pro+, and a New Max Plan
- 579. Software After AI
- 580. 量子位编辑作者招聘
- 581. Claude Opus 4.8: The System Card
- 582. 豆包 Seed 2.0 Lite升级:给 Agent 装上眼睛和耳朵
- 583. How an image could compromise your
- 584. 浙大开源SkillX:全自动构建Agent技能知识库,即插即用提升10%性能
- 585. 微软 Agent Framework 全栈指南:从 Hello Agent 到生产托管(Python)
- 586. Anthropic puts Claude agents on a meter across its subscriptions
- 587. Trace2Skill 把\"轨迹里的局部经验\"蒸馏成可迁移的 Agent Skills
- 588. Markdown 不会过时
- 589. LLM raiders and how to repel them
- 590. LLM raiders and how to repel them
- 591. 世界读书日|AI时代,我们该读什么?
- 592. Hermes Agent v0.13 — /goal 目标管理与 Kanban 多 agent 协作
- 593. openai gpt realtime voice models qbitai
- 594. Shannon & Epiplexity: Finite Compute Information Theory
- 595. OpenChronicle:把AI记忆变成可复用的基础设施
- 596. Nature | Anthropic:蒸馏过程潜意识传递行为偏好
- 597. 你缺的不是更好的 AI,而是一个\"装自己\"的系统
- 598. Inngest - AI and backend workflows, orchestrated at any scale
- 599. UIUC、Meta、斯坦福等最新综述!Code as Agent Harness
- 600. 对话灵感实验室:全帧率 VLM、低成本与分层部署
- 601. 明星开源项目为什么开始离开 github
- 602. Harness Engineering:AI 从\"聪明\"到\"可靠\"的第三代工程范式
- 603. Anthropic最危险路线图曝光: 无限记忆、多智能体! 硅谷AI终局仅剩双雄决顶
- 604. AI Is Writing More Code. Your CI Pipeline Can't Keep Up
- 605. 清华自进化Skill双星:EmbodiSkill + SkillEvolver
- 606. prompt-debugger-compare-templates-winty
- 607. code-review-graph:Claude Code 本地知识图谱,减少 6.8 倍代码审查 Token
- 608. pi mono github
- 609. AWS 一周综述:AWS Transform 上线一周年、AWS 云端 Claude Platform、EC2 M3 Ultra Mac 实例等(2026 年 5 月 18 日)
- 610. olmo-eval: An evaluation workbench for the model development
- 611. Microsoft Open-Sources RAMPART and Clarity to Secure AI Agents During Development
- 612. LLM RL中的熵 part 1: 熵的调控
- 613. 豆包 Seed 2.0 Lite升级:给 Agent 装上眼睛和耳朵
- 614. Anthropic to share Mythos cyber flaw findings with global finance watchdog
- 615. Skill Issues: Compromising Claude Code with malicious skills & agents -- Part 1
- 616. Consistency, But in Excellence Not Appearance
- 617. Have your views on AI changed
- 618. Jury dismisses all claims in Elon Musk's lawsuit against OpenAI CEO Sam Altman
- 619. Mistral OCR 4: SOTA OCR for Document Intelligence
- 620. Karpathy × Boris 访谈:Software 3.0 时代编程完整地图
- 621. MiniCPM-V 4.6 (1.3B) 面壁智能
- 622. 腾讯研究院AI速递 20260429
- 623. We Tested DeepSeek V4 Pro and Flash Against Claude Opus 4.7 and Kimi K2.6
- 624. 蛰伏一年,周衔团队带来首个具身基础模型,烹饪做实验弹琴,效果炸场
- 625. Open Defense Initiative
- 626. Open Defense Initiative | depthfirst
- 627. 鹅厂员工最近看的一本书是什么
- 628. Better Experiments with LLM Evals — A funnel, not a fork | Spotify Engineering
- 629. Mozilla warns UK: Breaking VPNs will not magically fix Britain's age-check mess
- 630. Google adds llms.txt check to Chrome Lighthouse
- 631. 撕开Claude Code真相:让它好用的98.4%,是工程不是AI
- 632. Meta announced Muse Spark in Voice Mode and Meta Glasses
- 633. GLM-5.2: Built for Long-Horizon Tasks
- 634. 面向 LLM 的架构设计:什么是真正的 AI Friendly 架构?
- 635. The great memory panic of 2026
- 636. Reinforcing Recursive Language Models | alphaXiv
- 637. RubyGems 暂停新用户注册:DDoS + 恶意包供应链攻击事件
- 638. Ai True Moat Organizational Capability
- 639. Graphsignal: LLM Inference Profiler
- 640. Self-Taught RLVR 综述
- 641. PersonaVLM: Long-term Personalized Multimodal LLM
- 642. 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)
- 643. The new AI lock-in
- 644. BigQuery Threat Model Report
- 645. 柚漫剧 AI全流程提效拆解
- 646. Crypto Inflows Hit $858M as CLARITY Lifts Sentiment
- 647. 腾讯研究院ai速递-20260429
- 648. Decisions and Dollars
- 649. Claude Fable 5 and new AI safety fables
- 650. stochastic parrot thought experiment
- 651. How Open Model Ecosystems Compound
- 652. 被裁了想转 AI Agent?先看面试官到底在筛你哪 7 样东西
- 653. Architecture & data foundations for AI-powered Search
- 654. GPT-Image-2 完全指南!附大量玩法案例,顺便开源我的生图 Skill ~
- 655. Paperclip · AI 公司操作系统 · 第 5 篇(完结)
- 656. Fake OpenAI Privacy Filter Repo Hits #1 on Hugging Face, Draws 244K Downloads
- 657. Microsoft to retire ‘Together Mode,’ its virtual meeting space for Teams
- 658. Anthropic 发布 Computer Use 最佳实践
- 659. 破案了为啥chatgpt老想着稳稳地接住你
- 660. Is Software Losing Its Head?
- 661. 明星开源项目,为什么开始离开 GitHub?
- 662. Restive Ventures Announces $45M Fund III to Back AI-Native Financial Services; Targets $1T Market Opportunity
- 663. EdgeClaw — 端云两栖龙虾框架
- 664. 从 Prompt 到 Harness:Claude 官方学习资料
- 665. 腾讯混元新里程碑:Hy3 preview 发布开源,Agent 表现全面提升
- 666. Claude Code Memory Setup (Obsidian + Graphify)
- 667. 快来,和AI实战派一起AI!AIGC峰会最新嘉宾阵容来了
- 668. Google 开发者福利,每月免费领取 10 美金,别忘了来领啊!
- 669. 这张信息图,居然是8B开源模型做的??
- 670. Gemini app rolling out 'Extended' thinking level, new 3rd-party app integrations
- 671. 腾讯研究院ai速递-20260507
- 672. 微软 1000 行代码,把 Claude Opus 干翻了 15 分
- 673. DeepSeek MoE 并行策略与GPU通信优化
- 674. LLMs are complicated now
- 675. Anthropic's bug-hunting Mythos was greatest marketing stunt ever, says cURL creator
- 676. 直播预约 | 数据引擎:具身智能的下一个决胜局
- 677. Anthropic
- 678. 快来和ai实战派一起aiaigc峰会最新嘉宾阵容来了
- 679. 让 Coding Agent 从黑盒到透明:阿里云 Agent 观测审计数据采集实践
- 680. GPT -Image 2神级提示词分享
- 681. The new AI lock-in
- 682. 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)
- 683. Mythos首个报告:30天挖出10000致命漏洞
- 684. OpenAI buys AI consultancy to sell enterprises on its models
- 685. 智子芯元
- 686. 腾讯员工公寓曝光,竟是这样的布置!
- 687. Microsoft for Startups | Microsoft
- 688. DiffusionGemma:扩散式文本生成模型(Google 26B MoE,4× 推理加速)
- 689. CMU Language Models Need Sleep (arxiv 2605.26099):SSM-Attention 睡眠巩固机制
- 690. A Bitter Lesson for Data Filtering
- 691. 高德 Marketing AutoResearch:AI Native 营销增长经营托管框架
- 692. 大模型可控新突破:Steering 机制、评估体系与开源落地
- 693. Qoder Skills 完全指南:从零开始,让 AI 按你的标准执行
- 694. 一个 Mission 跑 16 天、烧 7.78 亿 Token:Factory 公开了多 Agent 系统的构建哲学
- 695. Something's Rotten in the State of macOS Icon Design
- 696. the token economy
- 697. model half-life
- 698. 深入浅出 Harness Engineering 之核心模式与理念
- 699. never waste a token
- 700. gpt-image-2完全指南
- 701. Google brings local AI agents to laptops with Gemma 4 12B
- 702. 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)
- 703. Spec as AIOS:AI-Native 全栈交付的抗熵架构
- 704. Build a self-improving AI PM OS with Claude Code
- 705. The best argument I’ve heard for why AI won't take your job
- 706. 百度 Unlimited OCR:R-SWA 参考滑动窗口注意力实现超长文档连续解析
- 707. 存之有序,治之有矩——Agent 记忆系统的工程实践与演进
- 708. E9Ffy3R5Kwa1Ja5Pywbbrg
- 709. MiniMax M2.7 — 自我进化LLM
- 710. 2026 年面向 LLM 的 RL 方法总结:从 PPO 到 DPO 到 GRPO,再到多智能体 RL
- 711. 手把手:用 Hermes Skills + Karpathy 的 LLM Wiki 让 AI 越用越懂你
- 712. optimize_anything: A Universal API for Optimizing any Text Parameter
- 713. Qwen-RobotWorld Technical Report: Unifying Embodied World Modeling through Langu
- 714. Claude Code Dynamic Workflows 第 6 译本(玉澄 / 51CTO 视角)
- 715. 龙虾装上了,可以用来干啥?分享下我的 OpenClaw 多智能体团队搭建经验!
- 716. 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
- 717. Microsoft Build 2026 全景:7 款 MAI 模型 + Scout + OpenClaw 登 Windows + NVIDIA
- 718. OpenAI官方CLI上线,跟复杂的SDK说拜拜
- 719. 工作流的 Skill 怎么写?从 7 个顶级 Skill 中提炼的模式与最佳实践
- 720. Gaode Uplift Model Iteration Agent Long Running Harness
- 721. OpenAI buys AI consultancy to sell enterprises on its models
- 722. IMClaw:通过微信/飞书操控ClaudeCode/Codex/GeminiCLI/Pi Agent蜂群
- 723. 淘天营销中后台生码工作流最佳实践
- 724. LLM-as-a-Verifier: A General-Purpose Verification
- 725. 面向大型代码库的 Claude Code 团队落地经验与扩展策略(Agent Harness)
- 726. DeepSeek V4 (Flash & Pro) :通往百万级上下文与万亿参数推理的新纪元
- 727. Anthropic Dynamic Workflows 深度实战:JavaScript 编排脚本 + ultracode 模式 + /deep-research + 保存复用
- 728. 构建基于多智能体架构的深度思考交易系统
- 729. 全球AI新王诞生!Anthropic估值冲爆1.2万亿,首次反超OpenAI
- 730. 吴恩达2026新课上线!3小时包教包会,零代码小白也能成为AI超级玩家
- 731. Claude Code Dynamic Workflows 第 7 译本(架构师 JiaGouX 视角:任务级 Harness 统一框架)
- 732. Subagents 详解:Claude Code 如何避免上下文污染
- 733. 长期运行的 Agent 怎么管:Hermes 治理分层与 5 张卡
- 734. 两万字详解Claude Code源码核心机制
- 735. Building is just the beginning: Introducing Discoverability | Lovable
- 736. How to Automate AI Model Documentation with the NVIDIA MCG Toolkit
- 737. DeepSeek-V4-Flash means LLM steering is interesting again
- 738. Harness 工程分享:LLM Agent 的 prompt cache 命中率 90% 实战
- 739. 看完 Claude Code Agent Teams,我更确定接下来拼的是 Agent Runtime,技术拆解:Lead、Task List、Mailbox 和 Hooks 是什么东西
- 740. 我把 Karpathy 的 AutoResearch 搬到了软件开发领域,效果炸了
- 741. Introducing Seer Agent: The answer is already in Sentry. Now you can ask for it.
- 742. DeepSeek视觉原语论文:当所有人在堆图像分辨率时,它在堆「指代精度」!
- 743. Mini Shai-Hulud Strikes Again: TanStack + more npm Packages Compromised
- 744. 阿里云 MSE AI 任务调度 + Agent Sandbox:动态休眠/唤醒 OpenClaw Agent 成本下降 90%+
- 745. Karpathy加入Anthropic后首讲:Vibe Coding已死,Software3.0来了
- 746. 从手动到自动化:用AgentEval构建Agent评测体系
- 747. Latest open artifacts (#20): New orgs! New types of models! With Nemotron Super, Sarvam, Cohere Transcribe, & others
- 748. Thread by @0xCheeezzyyyy on Thread Reader App – Thread Reader App
- 749. Google Workspace Updates: Small businesses can now seamlessly import users from Microsoft to Google Workspace
- 750. Well-architected best practices for software supply chain security
- 751. DeepSeek V4是怎么训练出来的?58页论文深入解读
- 752. Open Notebook
- 753. Claude Code 上下文工程 —— Anthropic 团队的工程实践
- 754. System Over Model, Tested: Reproducing Mythos’s FreeBSD Find on Local Open-Weight Models
- 755. Xinzhiyuan Openai Tax Ai Self Improving Codex Eval Loop 20260606
- 756. 基于钉钉机器人的 Qoder CLI / Claude Code 双引擎 AI 助手实践
- 757. Apache RocketMQ 5.5.0 开源 LiteTopic:百万级 AI 会话专属通道
- 758. OpenAI Codex Sites:AI 一句话建站部署,Cloudflare Worker + D1 + R2 全套(极客之家译介)
- 759. A Route to Root in a 4G Industrial Router
- 760. Opd Revisiting Failure Modes Simple Fixes Storm
- 761. 今天起,Claude正式接入Office全家桶,跨应用还能共享记忆
- 762. Claude Code 团队成员亲述:动态工作流该怎么用(机器之心译本)
- 763. 5分钟AI长视频不翻车!国产开源框架杀到全球第一梯队
- 764. LLM-as-a-Verifier: A General-Purpose Verification Framework
- 765. Mnilax:CLAUDE.md 规则从 Karpathy 的 4 条增加到 12 条,Claude 错误率从 41% 降到 3%
- 766. Thread by @OpenAIDevs on Thread Reader App – Thread Reader App
- 767. World’s first native color LiDAR gives machines human-like vision
- 768. ChatGPT 的'失忆症'终于被治好了!Dreaming V3 让大模型拥有长期记忆
- 769. SkillOpt:把 Agent 技能文档变成可训练对象
- 770. 从人机协作到AI主导:我们是怎么把代码质量交给'数字SRE'的?
- 771. 5亿视频炼出全球最大GUI开源数据集、推理Token省71%小模型反超大模型——小米AI团队多篇论文入选ICML 2026
- 772. claude-code-hackathon-winners-2026
- 773. How We’d Market to Software Developers at Startups
- 774. SEO is a product, AEO is brand
- 775. 给 Claude Code 装上'超能力' — 它干活比我还靠谱
- 776. Loongsuite Pilot Sls Ai Coding Metrics Practice
- 777. Let's be honest: retail banking is broken
- 778. Rod Johnson 回归:Embabel 与「最后一波由人类选择的框架」
- 779. 第一批ai原生本科生要毕业了
- 780. Chime Turns a Profit as Members Hit 10.2 Million
- 781. Ethan He Cosmos Grok Imagine Latent Space Video Agent 20260606
- 782. Every AI Subscription Is a Ticking Time Bomb for Enterprise
- 783. Claude Code 动态工作流源码解析:pi-dynamic-workflows 架构实现
- 784. Generalization Dynamics of LM Pre-training — Jiaxin Wen
- 785. 我把 Claude Design 做成了 Skill,人人都能成为顶级网站设计师
- 786. 读完这篇,你就搞懂 DeepSeek v4 了
- 787. Hermes Agent 工具系统实战解析:40+ 工具为什么不用配置表
- 788. Agent 记忆存储方案深度洞察
- 789. 港中文 SLIM:动态技能生命周期管理,arXiv 2605.10923
- 790. The analytics engineer in 2026: system designer, governance owner, AI context provider
- 791. 知识库问答 @文档:从 DOM 方案到 ProseMirror 落地
- 792. 垂类 AI 创企的自救:FlashLabs 从 FlashIntel 到 AI Native
- 793. Autoresearch Next Phase Async Multi Agent Ai寒武纪
- 794. Buildkite Pricing | Buildkite
- 795. 曾经爆火的小龙虾凉了?现在圈内都在用这款新 Agent
- 796. Evaluate your Amazon Nova Sonic voice agent at scale, no microphone required
- 797. Claude Code 泄露后的漏网之鱼 claude-code-best 这两个月到底演进了什么
- 798. Why the Human Genome's Tangled Physicality May Confound AI
- 799. ChatGPT记忆大升级,十亿人免费用!
- 800. Scale Robot Reinforcement Learning with NVIDIA Isaac Lab on Amazon SageMaker AI
- 801. NeurIPS 2026 使用闭源 AI 检测器 Pangram 批量 desk-reject 论文事件
- 802. Accelerate LLM model loading and increase context windows with GPUDirect on Amazon FSx for Lustre and TurboQuant
- 803. 让大模型学会「自己教自己」!京东&中科院信工所连发三篇论文定义Self-TaughtRLVR
- 804. Introducing 1-bit and Ternary Bonsai Image 4B: Image Generation for Local Devices
- 805. 从 Harness 到动态工作流:Claude Code 多智能体任务编排的新范式
- 806. 简单接口:返回最终响应字符串
- 807. LLMShare: using shared chatbot pages to distribute malware
- 808. 长周期 Agent 详解:从 Ralph Loop 到可接管 Harness
- 809. Fundamental’s Large Tabular Model NEXUS is now available on Amazon SageMaker JumpStart
- 810. Meta Skill:Skill 之上编排 Skill 的抽象层
- 811. Detecting Misuse with the Claude Compliance API: The Threat Is in the Content
- 812. 从零构建大语言模型 —— 读完这篇你就懂了
- 813. 微软等提出 SkillOpt:把 Skill 当成模型一样训练
- 814. 国内首个 Frontier 三件套开源大模型:MiniMax M3 完整技术拆解
- 815. Claude Code Team 10 Tips Boris Data派Thu
- 816. The inevitable need for an open model consortium
- 817. lightseek tokenspeed
- 818. 一文看懂 OpenAI 开源的 Codex 编排规范:Symphony
- 819. 扣子 3.0 离谱更新:把 Codex、Claude Code 拉进一个项目工作?
- 820. 扣子 3.0 正式发布:@ 一下全员开工
- 821. 视频 RAG 中分块策略:基于停顿、滑动窗口与基于 LLM 的方法
- 822. DeepSeek陈德里AI论文第二弹:从6分到8分,DeliAutoResearch SKILL又进化了
- 823. Reducing container cold start times using SOCI index on DLAMI and DLC
- 824. Reimagining the mouse pointer for the AI era
- 825. NVIDIA Nemotron 3 Ultra now available on Amazon SageMaker JumpStart
- 826. obsidian-llm-wiki-local: Obsidian本地AI知识图谱自动构建工具
- 827. Introducing computer use in Gemini 3.5 Flash
- 828. Direct Connect (DX) 迁移最佳实践
- 829. OpenAI buys AI consultancy to sell enterprises on its models
- 830. Vera Arrives: NVIDIA’s First CPU Built for Agents Lands at Top AI Labs
- 831. 面向大型代码库的 Claude Code 团队落地经验与扩展策略(Agent Harness)
- 832. 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
- 833. Anthropic 实战分享:如何让 AI Agent 持续工作几天?
- 834. Why Internally-Built AI Fails Fund Accounting Audits
- 835. Openclaw Boris Cherny Agent Loop Design Patterns
- 836. Anthropic 95% 数据分析自动化:21% → 95% 准确率突破(极简短摘要,原文存档)
- 837. 大反转!马斯克牵手对手 Dario,Anthropic 与 SpaceX 罕见合作
- 838. State of Routing in Model Serving
- 839. Orchestrating Self-Evolving Agents with CrewAI and NVIDIA NemoClaw
- 840. 开始在 Amazon Bedrock 上使用 OpenAI GPT-5.5、GPT-5.4 模型和 Codex
- 841. Dean Ball on open models and government control
- 842. What it feels like to work with Mythos
- 843. Farewell Ai2
- 844. Subagents 详解:Claude Code 如何避免上下文污染
- 845. 面向电商直播场景的全模态大模型推理加速方案
- 846. Claude Code 架构深度解析
- 847. Recent Developments in LLM Architectures: KV Sharing, mHC, and Compressed Attention(机器之心中文全文)
- 848. 从 CPU 到 GPU 全链路可信,百度智能云新一代 AI 机密计算实例的探索与落地
- 849. LLM RL中的熵 Part 2: 熵对训练的调控
- 850. 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
- 851. 从全量启动到最小核:手淘外链唤端链路的三次架构演进
- 852. 从全量启动到最小核:手淘外链唤端链路的三次架构演进
- 853. Memory in the LLM Era: Modular Architectures and Strategies in a Unified Framework
- 854. 高德路线规划双路线:MobilityBench(Agent 基准)+ TransitLM(端到端 RLLM)
- 855. LLM 自我提升系统综述 — Yang 等 113 页四阶段闭环框架(Zesearch NLP Lab)
- 856. Claude Code 多智能体协作体系设计:从单Agent到多Agent工作流
- 857. claude-code-7-layer-memory-architecture
- 858. Olmo Hybrid and the Hybrid Architecture Wave (2026)
- 859. I/O 2026: Welcome to the agentic Gemini era
- 860. DeepSeek V4 Flash & Pro: Million-Token Context and Trillion-Parameter Inference
- 861. AI芯片架构:从逻辑门到矩阵乘法
- 862. LLM Fine-Tuning Cost Breakdown
- 863. DeepSeek视觉原语论文:当所有人在堆图像分辨率时,它在堆「指代精度」
- 864. factory mission multi agent architecture
- 865. AI Infra 入门干货总结:大模型是如何高效推理的
- 866. Qwen-Image-Flash: Beyond Objective Design — Few-step Distillation Training Recipe
- 867. The Inference Shift
- 868. 阿里Qwen提出Skill-RM:把奖励模型做成可复用Agent Skill
- 869. OlmoEarth v1.1: A more efficient family of Earth observation models
- 870. Agentium — 从零实现 Agent 系统的开源框架
- 871. 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)
- 872. 小米AI — ICML 2026 论文矩阵(11篇)
- 873. ELF: Embedded Language Flows (何恺明)
- 874. NVIDIA Nemotron 3 Ultra: Hybrid Transformer-Mamba MoE for Agentic AI on SageMaker JumpStart
- 875. What Is Urban Density Design? A Clear Guide to How Cities Get Built Denser
- 876. SenseNova-U1 — 商汤原生统一多模态模型
- 877. How LLMs Actually Work: 0xkato Transformer Walkthrough
- 878. 从零构建 LLM 七阶段实战教程
- 879. laser acl2026 latent superposition visual reasoning
- 880. Kimi Attention Residuals(AttnRes)— PreNorm 稀释问题与 Block 折中方案
- 881. Introducing Gemini Omni
- 882. Kimi K2.5 架构创新 — 1T MoE 一层路由 + 三能力跃迁
- 883. Google出手统一全模态检索:Gemini Embedding 2把文本、图片、音频和视频压进同一向量空间
- 884. 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
- 885. Accelerating Transformers Fine-Tuning with NVIDIA NeMo AutoModel
- 886. Introducing Claude Tag
- 887. Pi Agent:极简核心 + 事件总线扩展框架
- 888. Hermes Agent 9 模块系统架构
- 889. Owner-Worker-Verifier 架构
- 890. LiteFrame: Efficient Vision Encoders Unlock Frame Scaling in Video LLMs
- 891. Redis agentic AI flowers with Iris
- 892. 上下文隔离
- 893. Everything a Senior Engineer Needs to Know About What's Inside an LLM
- 894. Ghosts of Encryption Past – How we Read All Your Emails in Salesforce Marketing Cloud › Searchlight Cyber
- 895. Frontier post-training recipe review with Finbarr Timbers
- 896. Natural Language Autoencoders — Anthropic 激活→文字可解释性方法
- 897. Mollick 的 AI 实用指南:免费 vs 付费·模型选择·深度研究(2025-10)
- 898. Claude Opus 4.8 系统卡片深度分析
- 899. 我用 Claude Code 做需求调研,像多了一个产品经理
- 900. Red-Teaming after Mythos — Zico Kolter & Matt Fredrikson, Gray Swan
- 901. Gemini 3.5 Flash 内置 Computer Use 能力
- 902. 腾讯混元 CL-Bench Life:让大模型读懂你的日常生活
- 903. Sparse Autoencoders
- 904. LLM-Driven Feature Discovery
- 905. Nathan Lambert's Mid-2026 Open Models Bets
- 906. Anthropic vs DoW (Department of War) 与开源模型的 5-10 年权力均衡
- 907. 是时候讨论\"人类配额制\"了:应对AI就业冲击的一种可能性探讨
- 908. Anthropic 首次揭秘下一代 Claude 怎么造
- 909. Consumer Confidence
- 910. Some ideas for what comes next, May 2026 (Interconnects)
- 911. The Main Path to Truly Creative AI
- 912. Anthropic出手!AI的内心独白曝光了——NLA研究解读
- 913. the anti singularity