HackerNews中文版

每个人都在将大型语言模型（LLMs）应用于代码。数十亿美元投入到下一个标记的预测中。然而，他们所能做到的最好的就是自动补全和聊天。他们无法告诉你在你的代码库中实际发生了什么。他们无法绘制架构图。他们无法检测漂移。他们无法解释为什么一个文件的更改会导致三层深的内容出现问题。他们只是猜测，自信地猜测。而且他们的错误频率足够高，以至于你仍然需要验证一切。我采取了完全不同的方法。没有语言模型，没有嵌入，没有训练数据，没有云计算，没有GPU。我构建了一个系统，能够自动读取原始源代码并理解其架构。每一个决策都是可追溯和可解释的。 *45,476个函数。6个真实世界的开源代码库。4种语言。89.3%的准确率。* 它在你的笔记本电脑上运行不到一秒钟。它不需要联网。而且每次扫描都会变得更好——无需人工标注。人工智能行业存在盲点。每个人都在追逐生成技术。没有人在构建理解能力。你无法管理你看不见的东西，而现在，没有人能清晰地看到自己的代码库。我正在构建一个能够看清这一切的系统。我在寻找早期设计合作伙伴和投资者，他们理解下一波开发者工具的重点不是“人工智能编写你的代码”，而是“人工智能理解你的代码”。

查看原文

Everyone's throwing LLMs at code. Billions of dollars on next-token prediction. And the best they can do is autocomplete and chat.They can't tell you what's actually happening in your codebase. They can't map the architecture. They can't detect drift. They can't explain why a change in one file breaks something three layers deep. They guess. Confidently. And they're wrong often enough that you still have to verify everything.I took a completely different approach. No language model. No embeddings. No training data. No cloud. No GPU.I built a system that reads raw source code and understands the architecture. Automatically. Deterministically. Every decision is traceable and explainable.*45,476 functions. 6 real-world open-source codebases. 4 languages. 89.3% accuracy.*It runs on your laptop in under a second. It doesn't phone home. And it gets better with every scan — without human labeling.The AI industry has a blind spot. Everyone's chasing generation. Nobody's building understanding. You can't govern what you can't see, and right now, nobody can see their own codebase clearly.I'm building the system that sees it.Everyone's throwing LLMs at code. Billions of dollars on next-token prediction. And the best they can do is autocomplete and chat.<pre><code> They can't tell you what's actually happening in your codebase. They can't map the architecture. They can't detect drift. They can't explain why a change in one file breaks something three layers deep. They guess. Confidently. I took a completely different approach. No language model. No embeddings. No training data. No cloud. No GPU. I built a system that reads raw source code and understands the architecture. Automatically. Deterministically. Every decision is traceable and explainable. 45,476 functions. 6 real-world open-source codebases. 4 languages. 89.3% accuracy. It runs on your laptop in under a second. It doesn't phone home. And it gets better with every scan — without human labeling. The AI industry has a blind spot. Everyone's chasing generation. Nobody's building understanding. You can't govern what you can't see, and right now, nobody can see their own codebase clearly. Looking for early design partners and investors who understand that the next wave of developer tooling isn't "AI writes your code" — it's "AI understands your code." </code></pre> twoelf47@gmail.com

无大型语言模型，无训练数据，无云计算 – 理解架构的引擎