bailixisu 的个人数字花园：记录学习、项目实践与持续思考。https://bailixisu.com/https://bailixisu.com/posts/knowledge/artificial-intelligence/qwen-vl-architecture-and-innovations/https://bailixisu.com/posts/knowledge/artificial-intelligence/qwen-vl-architecture-and-innovations/深入拆解 Qwen3-VL 的视觉编码器、MLP 连接器、DeepStack、位置编码与训练流程，并核对截至 2026-09-26 的 Qwen3.8-Next、Omni 最新公开报告，附论文与 PDF 入口。Sun, 27 Sep 2026 08:49:44 GMThttps://bailixisu.com/posts/knowledge/artificial-intelligence/computer-vision/vision-transformer-architecture-and-principles/https://bailixisu.com/posts/knowledge/artificial-intelligence/computer-vision/vision-transformer-architecture-and-principles/以 224×224 图像和 ViT-B/16 为例，逐步推导 patch、位置编码、CLS、多头自注意力及张量形状，解释 ViT 为什么有效，以及它与 CNN 相比的数据需求和计算局限。Fri, 25 Sep 2026 16:42:04 GMThttps://bailixisu.com/posts/knowledge/artificial-intelligence/agents/ai-agent-message-tools-skills-workflow/https://bailixisu.com/posts/knowledge/artificial-intelligence/agents/ai-agent-message-tools-skills-workflow/面向开发者拆解一条消息进入 AI Agent 后的完整链路：接入与规范化、状态和记忆、策略路由、Tool/Skill 装配、上下文构建、受限工具循环、审批、响应以及运行后治理。Mon, 21 Sep 2026 11:50:46 GMThttps://bailixisu.com/posts/knowledge/artificial-intelligence/machine-learning/gradient-descent-optimizers-and-gradient-boosting/https://bailixisu.com/posts/knowledge/artificial-intelligence/machine-learning/gradient-descent-optimizers-and-gradient-boosting/从参数空间与函数空间出发，系统讲解 SGD、Momentum、AdamW 与 GBDT、XGBoost、LightGBM、CatBoost 的公式、直觉、差异、示例和实战选型。Sat, 19 Sep 2026 18:13:41 GMThttps://bailixisu.com/posts/knowledge/artificial-intelligence/machine-learning/traditional-machine-learning-feature-engineering-workflow/https://bailixisu.com/posts/knowledge/artificial-intelligence/machine-learning/traditional-machine-learning-feature-engineering-workflow/以客户月度流失预测为贯穿案例，系统讲解审计边界、标签成熟、异常清洗、时间切分、特征工程、Pipeline、分类与回归评估、调参、交付和监控。Sat, 19 Sep 2026 12:26:49 GMThttps://bailixisu.com/posts/knowledge/artificial-intelligence/llm-infrastructure/vllm-fsdp-llm-training-and-inference-guide/https://bailixisu.com/posts/knowledge/artificial-intelligence/llm-infrastructure/vllm-fsdp-llm-training-and-inference-guide/系统讲解 vLLM 的分页 KV Cache、连续批处理与推理并行，以及 FSDP/FSDP2 的状态分片、集合通信和检查点，并给出从训练到 OpenAI 兼容服务的示例、选型表与排障清单。Fri, 18 Sep 2026 16:11:35 GMThttps://bailixisu.com/posts/knowledge/artificial-intelligence/agents/plan-execute-cot-react-workflow-comparison/https://bailixisu.com/posts/knowledge/artificial-intelligence/agents/plan-execute-cot-react-workflow-comparison/从抽象层级、控制流、成本、延迟、可控性与适应性比较 CoT、ReAct、Plan-and-Execute 和 Workflow，并用同一差旅任务展示执行轨迹与生产选型。Fri, 18 Sep 2026 08:20:52 GMThttps://bailixisu.com/posts/knowledge/artificial-intelligence/agents/agent-tool-use-reliability/https://bailixisu.com/posts/knowledge/artificial-intelligence/agents/agent-tool-use-reliability/资料核查至 2026-09-18：系统解释 Agent 的工具选择、函数调用、状态闭环与规划执行，梳理按需工具发现、代码化编排、MCP 协议演进及状态化与安全评测，并给出生产可靠性架构与实施清单。Fri, 18 Sep 2026 02:55:33 GMThttps://bailixisu.com/posts/knowledge/artificial-intelligence/positional-encoding-transformers/https://bailixisu.com/posts/knowledge/artificial-intelligence/positional-encoding-transformers/从自注意力的置换等变性出发，推导正弦/余弦位置编码，比较可学习位置嵌入、相对位置、RoPE 与 ALiBi，并给出带形状说明的 PyTorch 实现。Thu, 17 Sep 2026 03:00:54 GMThttps://bailixisu.com/posts/knowledge/agents/memory/claude-code-memory-mechanism/https://bailixisu.com/posts/knowledge/agents/memory/claude-code-memory-mechanism/从 CLAUDE.md、自动记忆、JSONL 会话记录和上下文压缩四层拆解 Claude Code 的跨会话记忆，并说明配置核验、子 Agent 记忆与每日归档边界。Wed, 16 Sep 2026 16:32:30 GMThttps://bailixisu.com/posts/projects/deepseek-harness/platform-and-package-index/https://bailixisu.com/posts/projects/deepseek-harness/platform-and-package-index/解释 profile/preset 装配、Web Host 与 Client、Typert RPC、TypeScript/Python SDK、工程验证，并提供 49 组 219 包的中文职责索引。Wed, 16 Sep 2026 06:37:58 GMThttps://bailixisu.com/posts/projects/deepseek-harness/project-overview/https://bailixisu.com/posts/projects/deepseek-harness/project-overview/基于本地 47f9438 版本，梳理 DeepSeek Harness 的 49 个模块组、219 个包，以及插件、Agent、工具、会话和 Web UI 之间的关系。Wed, 16 Sep 2026 06:37:58 GMThttps://bailixisu.com/posts/projects/deepseek-harness/runtime-and-state/https://bailixisu.com/posts/projects/deepseek-harness/runtime-and-state/沿 ReactLoopAgent 的真实调用顺序，解释 Inbox、Turn、Step、模型流、工具并发、事件日志、持久化与上下文压缩。Wed, 16 Sep 2026 06:37:58 GMThttps://bailixisu.com/posts/projects/deepseek-harness/tools-and-orchestration/https://bailixisu.com/posts/projects/deepseek-harness/tools-and-orchestration/逐层解释文件、Shell、PTY、LSP、Web、MCP、Skills、沙箱、子 Agent、Workflow、Goal 和会话内提醒的职责与边界。Wed, 16 Sep 2026 06:37:58 GMThttps://bailixisu.com/posts/knowledge/python-backend/interviews/06-api-reliability/https://bailixisu.com/posts/knowledge/python-backend/interviews/06-api-reliability/用重复下单和接口变慢等场景，解释幂等键、重试、消息可靠性、连接池、可观测性、部署与测试。Tue, 15 Sep 2026 07:00:09 GMThttps://bailixisu.com/posts/knowledge/python-backend/interviews/05-redis-caching/https://bailixisu.com/posts/knowledge/python-backend/interviews/05-redis-caching/从缓存读写链路出发，拆解常见类型、穿透击穿雪崩、删除缓存的竞态、锁租约、淘汰持久化和事务。Tue, 15 Sep 2026 06:58:40 GMThttps://bailixisu.com/posts/knowledge/python-backend/interviews/04-mysql-transactions/https://bailixisu.com/posts/knowledge/python-backend/interviews/04-mysql-transactions/以 MySQL 8.4 InnoDB 为例，解释联合索引、执行计划、隔离级别、快照读、扣库存、死锁和游标分页。Tue, 15 Sep 2026 06:58:39 GMThttps://bailixisu.com/posts/knowledge/python-backend/interviews/03-web-frameworks/https://bailixisu.com/posts/knowledge/python-backend/interviews/03-web-frameworks/从 WSGI/ASGI 到请求校验、依赖生命周期、数据库会话、N+1、鉴权和 CORS，整理八组常见问答。Tue, 15 Sep 2026 06:55:42 GMThttps://bailixisu.com/posts/knowledge/python-backend/interviews/02-concurrency-asyncio/https://bailixisu.com/posts/knowledge/python-backend/interviews/02-concurrency-asyncio/讲清并发与并行、可选无 GIL 构建、协程调度、TaskGroup、阻塞调用、锁、超时取消和并发限制。Tue, 15 Sep 2026 06:55:41 GMThttps://bailixisu.com/posts/knowledge/python-backend/interviews/01-python-core/https://bailixisu.com/posts/knowledge/python-backend/interviews/01-python-core/用八组问答和可运行例子理解 Python 容器、对象引用、默认参数、闭包、装饰器、生成器、内存与类型标注。Tue, 15 Sep 2026 06:52:42 GMThttps://bailixisu.com/posts/knowledge/python-backend/interviews/00-interview-roadmap/https://bailixisu.com/posts/knowledge/python-backend/interviews/00-interview-roadmap/面向实习与初中级 Python 后端的复习路线，串起语言、异步、Web、MySQL、Redis、接口与排障六篇面经。Tue, 15 Sep 2026 06:52:41 GMThttps://bailixisu.com/posts/knowledge/llm-foundations/reinforcement-learning/09-dpo/https://bailixisu.com/posts/knowledge/llm-foundations/reinforcement-learning/09-dpo/从 chosen/rejected 答案对出发，计算 reference 校准的偏好分差、sigmoid 损失与梯度，解释 DPO 与 PPO、SFT 的区别。Tue, 15 Sep 2026 05:45:08 GMThttps://bailixisu.com/posts/knowledge/llm-foundations/reinforcement-learning/08-gspo/https://bailixisu.com/posts/knowledge/llm-foundations/reinforcement-learning/08-gspo/用四个 token 的概率比解释 GSPO 的几何平均、序列级裁剪与梯度，区分它与 GRPO，并说明 MoE 稳定性的适用边界。Tue, 15 Sep 2026 05:45:07 GMThttps://bailixisu.com/posts/knowledge/llm-foundations/reinforcement-learning/07-dapo/https://bailixisu.com/posts/knowledge/llm-foundations/reinforcement-learning/07-dapo/逐步解释 Clip-Higher、动态采样、token 级损失聚合和超长奖励处理，配完整链路与能手算的小例子。Tue, 15 Sep 2026 05:45:06 GMThttps://bailixisu.com/posts/knowledge/llm-foundations/reinforcement-learning/06-remax/https://bailixisu.com/posts/knowledge/llm-foundations/reinforcement-learning/06-remax/解释随机生成与贪心生成的双路流程，手算奖励差和策略更新，区分贪心基线、最优答案、RLOO 和 critic。Tue, 15 Sep 2026 05:45:05 GMThttps://bailixisu.com/posts/knowledge/llm-foundations/reinforcement-learning/05-rloo/https://bailixisu.com/posts/knowledge/llm-foundations/reinforcement-learning/05-rloo/用四个回答计算 leave-one-out 基线，说明 RLOO 如何构造 REINFORCE 梯度，以及它与 GRPO、PPO 的关系和差异。Tue, 15 Sep 2026 05:45:04 GMThttps://bailixisu.com/posts/knowledge/llm-foundations/reinforcement-learning/04-grpo/https://bailixisu.com/posts/knowledge/llm-foundations/reinforcement-learning/04-grpo/从同一道题采样四个答案开始，手算组内奖励标准化与 token 裁剪，解释 GRPO 的完整流程、零优势组和 KL 正则。Tue, 15 Sep 2026 05:45:03 GMThttps://bailixisu.com/posts/knowledge/llm-foundations/reinforcement-learning/03-ppo/https://bailixisu.com/posts/knowledge/llm-foundations/reinforcement-learning/03-ppo/沿完整 RLHF 链路解释 PPO-Clip，手算正负优势的概率比与裁剪，讲清 old policy、reference、critic 和奖励的不同职责。Tue, 15 Sep 2026 05:45:02 GMThttps://bailixisu.com/posts/knowledge/llm-foundations/reinforcement-learning/02-actor-critic-gae/https://bailixisu.com/posts/knowledge/llm-foundations/reinforcement-learning/02-actor-critic-gae/区分 actor、critic、reward model 和 reference，借三步生成例子手算 TD 误差与 GAE，解释策略与价值函数怎样协同训练。Tue, 15 Sep 2026 05:45:01 GMThttps://bailixisu.com/posts/knowledge/llm-foundations/reinforcement-learning/01-reinforce/https://bailixisu.com/posts/knowledge/llm-foundations/reinforcement-learning/01-reinforce/用一个只有两个候选答案的模型，手算 REINFORCE 的概率、损失、梯度和 SGD 更新，理解基线、方差与完整在线训练流程。Tue, 15 Sep 2026 05:45:00 GMThttps://bailixisu.com/posts/knowledge/llm-foundations/reinforcement-learning/00-llm-rl-roadmap/https://bailixisu.com/posts/knowledge/llm-foundations/reinforcement-learning/00-llm-rl-roadmap/从 prompt、rollout、奖励与优势到参数更新，串起 REINFORCE、Actor-Critic、PPO、GRPO、RLOO、ReMax、DAPO、GSPO 与 DPO 的入门学习路线。Tue, 15 Sep 2026 05:44:59 GMThttps://bailixisu.com/posts/knowledge/llm-foundations/fine-tuning/lora-ab-matrices-initialization/https://bailixisu.com/posts/knowledge/llm-foundations/fine-tuning/lora-ab-matrices-initialization/从低秩增量的逻辑出发，推导 LoRA A/B 矩阵形状、参数量和首步梯度，用可运行 PyTorch 示例解释一随机一为零的初始化，并介绍 PEFT、QLoRA、PiSSA 与 MoE 适配。Mon, 14 Sep 2026 12:11:00 GMThttps://bailixisu.com/posts/knowledge/llm-foundations/architectures/moe-reinforcement-learning-vs-dense/https://bailixisu.com/posts/knowledge/llm-foundations/architectures/moe-reinforcement-learning-vs-dense/解释 MoE 与 Dense 的参数、计算和显存差异，拆解大模型 RL 中的路由不一致、专家失衡、探索不足与通信瓶颈，并梳理 R3、PR2、RoutePack 和 ESRL 的解决思路。Mon, 14 Sep 2026 12:10:00 GMThttps://bailixisu.com/posts/knowledge/agents/memory/2026-agent-memory-beyond-rag/https://bailixisu.com/posts/knowledge/agents/memory/2026-agent-memory-beyond-rag/超越上下文、对话 RAG 和用户画像，解释 2026 年 MemRL、MemSkill、MemoryLACE、RuleMem、MemForest、FocusMem、TMEM 与环境核验记忆的机制、代码现状及实现路径。Mon, 14 Sep 2026 09:59:31 GMThttps://bailixisu.com/posts/knowledge/agents/frontiers/2026-agent-frontiers-and-open-problems/https://bailixisu.com/posts/knowledge/agents/frontiers/2026-agent-frontiers-and-open-problems/基于截至 2026 年 9 月 14 日的一手资料，梳理 Agent 的八条前沿路线，解释可靠性、记忆、强化学习、多智能体协作、电脑操作、自动科研和安全的进展与边界。Mon, 14 Sep 2026 08:38:13 GMThttps://bailixisu.com/posts/knowledge/llm-foundations/transformer/06-efficient-attention-variants/https://bailixisu.com/posts/knowledge/llm-foundations/transformer/06-efficient-attention-variants/区分稀疏、低秩、核线性注意力与 FlashAttention，推导因果状态更新，并用手算与可运行代码理解 Performer、RetNet、GLA、DeltaNet 和 Gated DeltaNet。Mon, 14 Sep 2026 05:39:09 GMThttps://bailixisu.com/posts/knowledge/llm-foundations/transformer/02-self-attention-deep-dive/https://bailixisu.com/posts/knowledge/llm-foundations/transformer/02-self-attention-deep-dive/逐步拆解 Q、K、V 投影、缩放点积、Mask、Softmax、Value 聚合与 Multi-Head Attention，并用可手算案例解释训练和推理中的真实数据流。Sat, 12 Sep 2026 04:17:00 GMThttps://bailixisu.com/posts/knowledge/llm-foundations/transformer/01-transformer-foundations/https://bailixisu.com/posts/knowledge/llm-foundations/transformer/01-transformer-foundations/从序列建模难题出发，拆解 Transformer 的 Encoder–Decoder、Self-Attention、位置编码和残差路径，并建立现代 Transformer 变体的完整坐标系。Sat, 12 Sep 2026 03:50:00 GMThttps://bailixisu.com/posts/knowledge/agents/pi-agent/13-mini-pi/https://bailixisu.com/posts/knowledge/agents/pi-agent/13-mini-pi/用 pi-ai 与 pi-agent-core 构造最小 Agent，并用确定性 Fake Provider 验证两轮模型调用、一次 Tool Use、结果回灌和事件顺序。Fri, 11 Sep 2026 10:52:00 GMThttps://bailixisu.com/posts/knowledge/agents/pi-agent/12-agent-harness/https://bailixisu.com/posts/knowledge/agents/pi-agent/12-agent-harness/分析 Pi experimental AgentHarness 的 Session、Branch、AgentLane、Operation、事务存储、Effect Gate 与崩溃恢复，并和稳定 CLI v3 对照。Fri, 11 Sep 2026 10:51:00 GMThttps://bailixisu.com/posts/knowledge/agents/pi-agent/11-reliability-security/https://bailixisu.com/posts/knowledge/agents/pi-agent/11-reliability-security/分析 Pi 的 Abort、Retry、Overflow Recovery、事件结算、Project Trust、凭据与工具权限，并给出容器化和 Tool Policy 的分层防护方案。Fri, 11 Sep 2026 10:50:00 GMThttps://bailixisu.com/posts/knowledge/agents/pi-agent/10-session-tree-compaction/https://bailixisu.com/posts/knowledge/agents/pi-agent/10-session-tree-compaction/深入 Pi JSONL v3 的 append-only 会话树，解释 branch、fork、clone、Branch Summary、Compaction、split turn 与上下文恢复。Fri, 11 Sep 2026 10:49:00 GMThttps://bailixisu.com/posts/knowledge/agents/pi-agent/09-modes-and-tui/https://bailixisu.com/posts/knowledge/agents/pi-agent/09-modes-and-tui/比较 Pi 的四种运行模式和直接 SDK 集成，解释同一个 AgentSession 如何被终端 UI、脚本、事件流与外部应用复用。Fri, 11 Sep 2026 10:48:00 GMThttps://bailixisu.com/posts/knowledge/agents/pi-agent/08-extensions/https://bailixisu.com/posts/knowledge/agents/pi-agent/08-extensions/系统梳理 Pi Extension 的加载、事件、Hook、命令、工具、Provider、Session 状态与 TUI 扩展能力，以及它们真正的安全边界。Fri, 11 Sep 2026 10:47:00 GMThttps://bailixisu.com/posts/knowledge/agents/pi-agent/07-model-provider/https://bailixisu.com/posts/knowledge/agents/pi-agent/07-model-provider/拆解 pi-ai、ModelRuntime、模型目录、认证解析、Provider 协议、thinking 映射、流式事件与跨模型上下文转换。Fri, 11 Sep 2026 10:46:00 GMThttps://bailixisu.com/posts/knowledge/agents/pi-agent/06-prompt-context-skills/https://bailixisu.com/posts/knowledge/agents/pi-agent/06-prompt-context-skills/从 buildSystemPrompt() 与 AgentSession.prompt() 出发，分析工具说明、项目规则、Skills、Prompt Templates 和 Slash Commands 如何进入模型上下文。Fri, 11 Sep 2026 10:45:00 GMThttps://bailixisu.com/posts/knowledge/agents/pi-agent/05-memory/https://bailixisu.com/posts/knowledge/agents/pi-agent/05-memory/从工作上下文、JSONL 会话树、项目规则、Skills、Custom Entry 到 Compaction，建立 Pi Memory 的分层模型，并解释它为什么不是向量长期记忆。Fri, 11 Sep 2026 10:44:00 GMThttps://bailixisu.com/posts/knowledge/agents/pi-agent/04-tool-use/https://bailixisu.com/posts/knowledge/agents/pi-agent/04-tool-use/拆解 Pi 工具的定义、参数准备、TypeBox 校验、Extension Hook、并行策略、文件写入队列、输出截断与 Tool Result 回灌。Fri, 11 Sep 2026 10:43:00 GMThttps://bailixisu.com/posts/knowledge/agents/pi-agent/03-agent-loop/https://bailixisu.com/posts/knowledge/agents/pi-agent/03-agent-loop/深入 packages/agent，解释 prompt、continue、turn、消息流、工具批次、steering 与 follow-up 如何组成 Pi 的低层执行循环。Fri, 11 Sep 2026 10:42:00 GMThttps://bailixisu.com/posts/knowledge/agents/pi-agent/02-startup-and-resources/https://bailixisu.com/posts/knowledge/agents/pi-agent/02-startup-and-resources/沿着 main()、createAgentSessionServices() 与 DefaultResourceLoader，拆解 Pi 如何合并配置、判断项目可信度、发现资源并组装可运行的 AgentSession。Fri, 11 Sep 2026 10:41:00 GMThttps://bailixisu.com/posts/knowledge/agents/pi-agent/01-overall-architecture/https://bailixisu.com/posts/knowledge/agents/pi-agent/01-overall-architecture/以 Pi v0.85.1 的官方文档和源码为依据，拆解 CLI、AgentSession、Agent Loop、工具、记忆、模型适配与 TUI，并还原一次请求的完整调用链。Fri, 11 Sep 2026 10:40:00 GMThttps://bailixisu.com/posts/knowledge/agents/architecture/workflow-vs-skills/https://bailixisu.com/posts/knowledge/agents/architecture/workflow-vs-skills/从控制流、状态、知识封装和执行约束出发，解释 Agent Workflow 为什么存在、Skills 如何实现，以及什么时候单独使用或组合使用。Fri, 11 Sep 2026 08:31:39 GMThttps://bailixisu.com/posts/knowledge/agents/architecture/react-plan-execute-multi-agent/https://bailixisu.com/posts/knowledge/agents/architecture/react-plan-execute-multi-agent/从反馈控制、显式规划、状态所有权和组织协作四个维度，解释三种 Agent 形态为什么不同、何时适用，以及如何通过对照实验评判。Fri, 11 Sep 2026 08:04:24 GMThttps://bailixisu.com/posts/note-to-article/https://bailixisu.com/posts/note-to-article/一套从问题、证据、实践到结论的轻量整理方法。Fri, 11 Sep 2026 05:10:00 GMThttps://bailixisu.com/posts/how-this-blog-works/https://bailixisu.com/posts/how-this-blog-works/Astro、Markdown、GitHub 与 Vercel 组成的轻量写作工作流。Fri, 11 Sep 2026 05:05:00 GMThttps://bailixisu.com/posts/hello-learn-everything/https://bailixisu.com/posts/hello-learn-everything/为什么我要建立这个数字花园，以及这里将会记录什么。Fri, 11 Sep 2026 05:00:00 GMT