什么时候该坚持,什么时候该放弃?从 The Dip 到 Strategic Quitting 的决策框架
坚持并不总是美德,放弃也不一定是失败。本文从 Seth Godin 关于 The Dip、沉没成本、阻力、目标受众与 AI 时代职业价值的讨论中,提炼出一套可实际用于工作、项目、创业和职业选择的 Continue → Pivot → Quit 决策框架。
Continue readingNotes on AI, agents, and building things that work.
坚持并不总是美德,放弃也不一定是失败。本文从 Seth Godin 关于 The Dip、沉没成本、阻力、目标受众与 AI 时代职业价值的讨论中,提炼出一套可实际用于工作、项目、创业和职业选择的 Continue → Pivot → Quit 决策框架。
Continue readingA practical guide to Retrieval-Augmented Generation: how RAG connects language models to external knowledge, improves traceability, and supports reliable AI Agents.
Continue readingRAG(检索增强生成)如何让 AI 结合外部知识回答问题?本文从索引、检索、重排、生成、权限、评估和 AI Agent 实战讲清楚 RAG。
Continue reading2026 年的提示工程不只是改写一句提示词,而是一起设计上下文、工具边界、输出契约、评估和恢复机制。
Continue readingPrompt engineering in 2026 is less about magic wording and more about designing reliable context, tool contracts, output schemas, evals, and feedback loops for AI agents.
Continue reading深入比较 Qwen3.8-Max、Gemini 3.1 Pro、GPT-5.6 Sol 与 Claude Fable 5 的综合能力、编程表现、上下文窗口、API 成本、Agent 工具生态及企业数据隐私风险,并给出适合 ERP 与 AI Agent 系统的多模型路由建议。
Continue readingA practical August 2026 comparison of four frontier AI models, covering human-preference capability, coding and agent strengths, context windows, API pricing, multimodal support, and the best model-routing strategy for enterprise and ERP AI systems.
Continue readingOpenAI has confirmed Astra as its next major model, but it is not yet a public GPT-6 product. This deep dive separates confirmed facts from speculation, explains Astra’s ten mathematical advances, Lean verification, long-horizon reasoning, safety risks, and its implications for coding agents, MCP, RAG, IDP, and AI-native ERP.
Continue reading深入分析 Claude Fable 5、Opus 5、Sonnet 5 与 Haiku 4.5 的能力、价格、Claude Code、MCP、PDF/IDP、企业应用、隐私限制,以及开发者应该如何选择模型。
Continue readingGPT‑5.6 的真正升级,不只是更会写代码,而是更擅长理解大型代码库、操作终端与工具、持续调试、验证结果,并完成端到端工程任务。本文加入 Coding Agent Index、SWE-Bench Pro、DeepSWE 与 Terminal-Bench 分数图表,分析它在 JavaScript、SQL、ERP、MCP、RAG 和 Coding Agent 场景中的实际价值与限制。
Continue reading2026年7月,网络上出现大量低价 Claude Max 20x 账号。事件背后并非官方降价,而是有人利用 SEPA 自动扣款的结算时间差,在付款确认前取得高额度权限并转售。本文拆解事件机制、真假信息、平台风险,以及 SaaS 与 AI 产品应该吸取的工程教训。
Continue readingWhat Anthropic's “Map vs Territory” mindset teaches us about building better AI systems, stronger engineering workflows, and safer enterprise implementation.
Continue reading