Jinspire
Writing about curiosity, insights, and reflection.
记录好奇、洞察与思考。
Artifacts → 项目 →2026年中的编程
个人浅谈最新的代码模型、智能体工作流、何时复刻并提取第三方代码，以及为什么人类的学习如今成为了真正的瓶颈。
Coding in Mid-2026
A personal look at the latest coding models, agent workflows, when to fork and extract third-party code, and why human learning is now the real bottleneck.
开源大语言模型的现状：前沿领域的趋同
前沿开源大模型正趋同于标准混合专家底座。真正的差异化已转移至注意力机制设计如MLA与GQA之争，以及长上下文处理能力。
The State of Open-Source LLMs: Convergence at the Frontier
Frontier open-source LLMs are converging on standard MoE backbones. True differentiation has shifted to MLA versus GQA attention and context length.
2025年年终总结
历经创业失败与无数拒信，作者最终加入心仪的AI公司。他拥抱高强度工作与项目开发，感悟到痛苦能激发动力，并发现对创造事物的热爱。
Summary of the Year 2025
Surviving startup failure and rejections, the author joined Reflection AI. Embracing intense work to build AI, they found pain fuels deep growth.
顶级AI领袖 - 能力雷达
本文利用Gemini 3，从模型架构、基础设施、产品及财富等六大维度，在1至10分范围内对顶级AI领袖的能力进行了量化评估。
Top AI Leaders - Ability Radar
The author uses Gemini 3 to evaluate top AI leaders across six dimensions including model architecture, infrastructure, and wealth on a 1-10 scale.
AI仍然无法读取Google Scholar的数字
测试表明，在查找谷歌学术最高引用学者时，多数AI出现严重幻觉。仅Gemini表现优异，这意味着通用人工智能依然遥远。
AI still can't read Google Scholar's numbers
Testing AI models to find Google Scholar's most cited researchers reveals widespread hallucinations. Gemini 2.5 won, but AGI still seems far away.
LLM训练对我自身学习过程有哪些实际启示？
像大模型一样，人类需将终身学习作为预训练，在职场中有意地进行微调，并严格控制日常的奖励模型，从而避免廉价多巴胺，实现长远目标。
What are some practical implications of LLM training for my own learning process?
Like LLMs, humans need lifelong learning as pretraining, intentional workplace fine-tuning, and strict control over daily reward models to thrive.
从零开始花100小时构建嵌入搜索引擎，我学到了什么
嵌入将数据转为向量匹配相似度，但无法处理负面提示。推荐pgvector与Qwen，建召回与重排双阶段架构，迈向智能体搜索。
What I learned from using 100 hours to build an embedding search engine from ground up
Embeddings map data to vectors for similarity matching but fail on negative prompts. Use pgvector, Qwen, and a two-stage recall-rerank pipeline.
半年用AI写了一百万行代码，我学到了什么？
生成百万行代码表明，AI编程节奏极快，开发者需专注架构设计与上下文管理。其虽擅长基础代码，但须严防过度设计与静默失败。
What I learned from using AI to write ONE MILLION lines of code in half a year?
Writing a million lines of AI code shows vibe coding demands intense architectural thinking and strict context management to prevent overengineering.
创业日记 - 持续更新
作者离开谷歌加入AI初创公司，技术能力快速提升并晋升为联合创始人。但随着产品进展停滞，他失去热情，最终在九个月后离职。
Startup Diary - Continuously Updated
Leaving Google for an AI startup brought technical growth and a cofounder role. Stalled progress killed his passion, so he quit nine months later.
为什么说世界的本质是DP？
思考的本质是动态规划，依靠调用记忆经验避免从零计算。学习就是加载前人算好的缓存表，省去试错成本，站在巨人肩膀上寻找自己的路。
Why the Essence of the World is DP?
Human thought mirrors dynamic programming by reusing past experiences. Learning is downloading calculated solutions from others to build our paths.
机器学习资源索引
本文精选了核心人工智能资源，重点梳理了斯坦福大模型课程、前沿模型可解释性论文、底层训练架构指南以及行业顶尖专家名单。
ML Resources Index
This index curates essential AI resources, detailing Stanford LLM courses, Anthropic interpretability papers, training guides, and industry experts.
图像生成101：入门指南
图像生成技术经历了VAE、GAN、扩散模型到VAR的演进。SD模型采用VAE编码、U-Net潜在去噪与CFG引导。
Image Generation 101: an Introduction
Image generation evolved from VAE and GAN to Diffusion and VAR eras. Stable Diffusion uses VAE encoding, U-Net latent denoising, and CFG guidance.
Gemini 2.5 绝对令人惊叹
作者认为Gemini 2.5性能已超越OpenAI。凭借其卓越的思考与长上下文能力，他已将编程和深度问答工具全部替换为该模型。
Gemini 2.5 is strictly amazing
Gemini 2.5 clearly outperforms OpenAI and Anthropic. The author now uses it exclusively for Cursor coding and deep QA due to its superior thinking.
并行计算101：LLM并行性入门
为将超大语言模型载入显卡，本文结合多篇经典论文，详细解析了数据并行、流水线并行、张量并行与专家并行的核心原理及其实际应用。
Parallelism101: Introduction to LLM Parallelism
To fit massive LLMs into GPUs, this post details data, pipeline, tensor, and expert parallelism using foundational papers like GPipe and Megatron.
用Deepseek读《物演通论》
作者开启用Deepseek读物演通论的Gitbook系列，因排版便捷考虑将博客迁至该平台，并分享了提供该书及相关哲学课程的网站。
Reading 'The Evolution of Things' with DeepSeek
The author launched a Gitbook series reading The Theory of Universal Evolution with Deepseek, prefers its formatting, and shared a resource site.
可视化LLM生成过程
本文通过追踪从词元嵌入、注意力机制到多层感知机的矩阵变化，详细可视化了大型语言模型生成最终单词的完整内部计算过程。
Visualize Generation in LLM
This post details LLM word generation, tracking matrix changes from token embeddings through QKV attention and MLP layers to predict the final word.
人工智能, 创业, 与反思
作者加入初创团队重获创造之乐。他将人工智能视作将取代人类的硅基后代，认为在无意义的宇宙中，意识涌现与创造本身即是存在的全部意义。
AI, Startup, and Reflections
Leaving Google for a startup brought creative joy. The author embraces AI as our silicon offspring destined to replace us in a meaningless universe.