{"total":8,"items":[{"id":1,"type":"article","title":"DeepSeek-R1 推理模型技术报告深度解读：强化学习如何涌现推理能力","summary":"DeepSeek-R1 证明了纯强化学习（RL）可以在不依赖监督微调的情况下涌现出强大的链式推理能力。本文拆解其 GRPO 算法、奖励设计与蒸馏策略，并分析 R1-Zero 的\"顿悟时刻\"对开源社区的意义。","keyPoints":["R1-Zero 完全基于 RL 训练，无需 SFT 冷启动即涌现自我反思能力","GRPO 算法省去 Critic 网络，显著降低训练成本","蒸馏小模型（1.5B-70B）在多项基准上超越 GPT-4o","全部权重 MIT 协议开源，可自由商用"],"category":"llm","tags":["DeepSeek","强化学习","推理模型","开源"],"sourceName":"公众号 · 机器之心","sourceUrl":"https://mp.weixin.qq.com/s/example-deepseek-r1","cover":null,"projects":[{"name":"deepseek-ai/DeepSeek-R1","url":"https://github.com/deepseek-ai/DeepSeek-R1","stars":90200,"language":"Python","description":"DeepSeek 推理模型官方仓库"}],"publishedAt":"Sun Oct 11","views":3241,"likes":289},{"id":2,"type":"video","title":"实测：用 Claude Code 10 分钟从零构建一个全栈应用","summary":"视频号作者现场演示了 Claude Code 的 Agent 编程能力：从需求描述到数据库设计、API 开发、前端页面生成全程无手写代码。本摘要由 AI 从视频语音中自动提取。","keyPoints":["自然语言描述需求即可生成完整项目骨架","Agent 自动执行终端命令、读写文件、运行测试","适合快速原型验证，生产代码仍需人工审查","与传统 Copilot 补全模式相比，Agent 模式自主性更强"],"category":"agent","tags":["Claude","AI 编程","Agent","实测"],"sourceName":"视频号 · AI 产品黄叔","sourceUrl":"https://channels.weixin.qq.com/example-claude-code","cover":null,"projects":[],"publishedAt":"Sat Oct 10","views":1876,"likes":156},{"id":3,"type":"article","title":"LangGraph 1.0 发布：构建生产级多智能体系统的新范式","summary":"LangGraph 1.0 带来稳定 API 与持久化执行能力，将 Agent 工作流抽象为状态图。文章通过客服多智能体案例讲解节点、边、检查点与人机协作中断机制。","keyPoints":["状态图模型让复杂 Agent 流程可视化、可调试","内置持久化检查点，支持长任务断点续跑","Human-in-the-loop 中断机制保障关键节点人工审批","与 LangSmith 深度集成实现全链路观测"],"category":"agent","tags":["LangGraph","多智能体","LangChain","工作流"],"sourceName":"公众号 · 量子位","sourceUrl":"https://mp.weixin.qq.com/s/example-langgraph","cover":null,"projects":[{"name":"langchain-ai/langgraph","url":"https://github.com/langchain-ai/langgraph","stars":12800,"language":"Python","description":"用状态图构建可控 AI Agent"}],"publishedAt":"Fri Oct 09","views":2154,"likes":178},{"id":4,"type":"video","title":"Sora 级视频生成平替？国产开源模型 CogVideoX 实测对比","summary":"作者用 20 组提示词对比测试 CogVideoX 与主流商业视频模型，从运动连贯性、物理规律、文字渲染三个维度打分。结论：开源方案在短镜头上已接近可用水准。","keyPoints":["CogVideoX-5B 在消费级显卡可本地推理","复杂物理交互仍是所有模型的共同短板","开源生态（ComfyUI 工作流）快速成熟","商业视频模型在一致性上仍有明显优势"],"category":"vision","tags":["开源","实测","视频生成","CogVideoX"],"sourceName":"视频号 · 数字生命卡兹克","sourceUrl":"https://channels.weixin.qq.com/example-cogvideo","cover":null,"projects":[{"name":"THUDM/CogVideo","url":"https://github.com/THUDM/CogVideo","stars":11300,"language":"Python","description":"清华智谱开源视频生成模型"}],"publishedAt":"Thu Oct 08","views":4102,"likes":402},{"id":5,"type":"article","title":"RAG 工程实践：从 60 分到 95 分的检索优化清单","summary":"一篇面向落地的 RAG 优化指南，覆盖文档切分策略、混合检索、重排序、查询改写与评估体系搭建，附完整开源工具链选型建议。","keyPoints":["语义切分 + 父子索引显著优于固定长度切分","BM25 与向量检索混合召回是性价比最高的提升","引入 Reranker 模型可将 Top-5 准确率提升 15%+","建立自动化评估集是持续优化的前提"],"category":"tools","tags":["RAG","向量检索","工程实践","LLM 应用"],"sourceName":"公众号 · 夕小瑶科技说","sourceUrl":"https://mp.weixin.qq.com/s/example-rag","cover":null,"projects":[{"name":"langgenius/dify","url":"https://github.com/langgenius/dify","stars":65800,"language":"TypeScript","description":"LLM 应用开发平台"},{"name":"infiniflow/ragflow","url":"https://github.com/infiniflow/ragflow","stars":42100,"language":"Python","description":"开源 RAG 引擎"}],"publishedAt":"Wed Oct 07","views":5687,"likes":621},{"id":6,"type":"article","title":"AI 编程重塑软件团队：一家 30 人公司的提效复盘","summary":"某创业公司引入 AI 编程工具 6 个月后的量化复盘：人均代码产出提升 2.3 倍，但 Code Review 与架构设计能力成为新瓶颈。文章给出团队落地的三条路径。","keyPoints":["初级工程师提效最显著，高级工程师价值向设计迁移","AI 生成代码的 Review 成本约占节省时间的 30%","建立团队级 Prompt 与规则库是关键动作","测试覆盖率要求反而需要提高"],"category":"industry","tags":["AI 编程","工程管理","落地案例"],"sourceName":"公众号 · InfoQ","sourceUrl":"https://mp.weixin.qq.com/s/example-team","cover":null,"projects":[],"publishedAt":"Tue Oct 06","views":1932,"likes":145},{"id":7,"type":"video","title":"OpenAI o3 发布会 8 分钟速览：编程与数学能力再跃迁","summary":"视频快速梳理 o3 模型发布要点：ARC-AGI 得分突破、编程竞赛排名、推理成本分析，以及对开发者生态的潜在影响。","keyPoints":["ARC-AGI 测试得分大幅超越前代模型","Codeforces 排名进入全球前 200","高推理模式下成本仍是规模化应用的约束","API 分阶段开放，安全评估同步推进"],"category":"news","tags":["OpenAI","o3","发布会","模型评测"],"sourceName":"视频号 · 量子位 QbitAI","sourceUrl":"https://channels.weixin.qq.com/example-o3","cover":null,"projects":[],"publishedAt":"Mon Oct 05","views":7643,"likes":834},{"id":8,"type":"article","title":"本地大模型完全指南：Ollama + 国产模型的隐私计算方案","summary":"面向个人与小团队的本地 LLM 部署教程，对比 Qwen、GLM、DeepSeek 蒸馏版在 Mac 与国产显卡上的实际表现，给出按硬件选型的决策表。","keyPoints":["16GB 内存 Mac 可流畅运行 8B 量化模型","Ollama + Open WebUI 十分钟搭好私有 ChatGPT","敏感数据处理场景本地推理是刚需","MoE 架构小模型是端侧部署的趋势"],"category":"tools","tags":["Ollama","本地部署","隐私","开源模型"],"sourceName":"公众号 · 开源中国","sourceUrl":"https://mp.weixin.qq.com/s/example-ollama","cover":null,"projects":[{"name":"ollama/ollama","url":"https://github.com/ollama/ollama","stars":132000,"language":"Go","description":"本地运行大模型的最简单方式"},{"name":"open-webui/open-webui","url":"https://github.com/open-webui/open-webui","stars":84600,"language":"Svelte","description":"自托管 AI 对话界面"}],"publishedAt":"Sun Oct 04","views":4521,"likes":512}],"source":"mysql"}