Notes on AI, agents, and building things that work.
2026-10-26 新增 Pi + OpenRouter 原创实验摘要：Flint Free 29/30 vs GPT-6 Luna 30/30；三个代码修复场景均通过160项检查；实测 Tool Call 遗漏、延迟长尾与 Reasoning 控制限制。结合独立 SWE-bench、Terminal-Bench、AA Index、Mac M4 32GB 量化部署，并严格说明原始实验文件尚待审计。
Continue reading从 Dots、ChatGPT Space、GPT-6.1 Sol、Ultrafast、Decisions API、Agents API、Codex Cloud、Computer Use 到 Plugin Extensions，本文不只列发布内容，而是用一张完整知识关系图解释 OpenAI 如何把 Chatbot 推向 Digital Employee、Multi-Agent Team 与真正的 Agent Platform。 另加入 Artificial Analysis 的第三方 benchmark snapshot，比较 GPT-6.1 Sol、Astra 与 Luna 的 intelligence / cost trade-off，并延伸到 Model、Reasoning Effort 与 Speed Routing。
Continue reading从 Computer Use、Coding、Professional Work、Science、Cybersecurity、Alignment、Long Context 与 ARC‑AGI 全面拆解 GPT‑6 Astra 的公开 benchmark，并区分真正的能力跃升、harness 带来的系统级提升，以及容易被 marketing 放大的数字。
Continue readingA dated, evidence-backed comparison of Google Gemini 3.8 Flash, Anthropic Claude Fable 5.1, and OpenAI GPT-5.6 Sol for reasoning, coding, AI agents, speed, cost, context, tool use, and real-world engineering workflows.
Continue readingWhy AI model comparisons should show Low, Medium, High, XHigh and Max effort separately instead of comparing only maximum scores.
Continue readingA practical comparison of GPT-5.6 Luna, Terra and Sol across reasoning effort, Artificial Analysis Intelligence Index, cost and first-response latency—and what the results mean for AI model routing.
Continue reading用程序员的 for-each 心智模型，逐步理解 PostgreSQL CROSS JOIN LATERAL、LEFT JOIN LATERAL、ON TRUE、Top 1 per parent，以及 ERP SQL 中最常见的使用场景与性能注意事项。
Continue readingA practical architecture review of Codex Harness: context engineering, deferred tools, sandboxing, verification loops, compaction, persistent execution, and what these ideas mean for building reliable AI agents.
Continue reading从 Codex Harness 的 Context Construction、Deferred Tools、Sandbox、Goal Loop、Compaction 与 App Server，理解真正的 Agent Engineering。
Continue readingWhy long-context AI needs externalized state, bounded recursion, better harnesses, and evaluator-driven learning—not just bigger context windows.
Continue reading长上下文不是终点。真正的变化，是把 Context 从“全部塞进 Prompt”改造成可查询、可计算、可递归、可验证的外部工作环境。本文拆解 RLM、Prime Agent、Continual Harness、Evaluator、自我改进与安全边界，并给出一套可落地的 Thinking Architecture。
Continue readingA practical role-by-role guide to GPT-5.6 Sol, Terra, and Luna: current API pricing, where each model fits across business functions, and how to design a cost-efficient AI routing and permission architecture.
Continue reading
