Series — Fine-Tuning, Smallest to Largest (same task, three techniques, smallest model to largest):
- Full Fine-Tuning (270M) ← you are here
- LoRA (1.5B)
- QLoRA (7B)
- If the small model wo...
128 lines of Python.
That's the entire core of a coding agent — the loop that powers tools like Claude Code and Cursor. I didn't believe it either, so I built one from scratch. Then I pointed it at a failing test, and it read the file, ran the test, saw the traceback, fixed the code, and re-ran it — choosing every step itself. No one hard-coded that.
It's open s...
APX has a web admin panel, but that does not make the browser the center of the system.
The design is stricter than that: the daemon owns the state, and the web UI is only a local client.
That split sounds small. It is not. It decides where truth lives, how much can drift, and whether th...
I spent the last few days doing something most content site owners never do: I bucketed every single one of my 1,684 indexed pages by their Google ranking position and asked a simple question — where are my impressions actually going?
The answer made me stop writing new content immediately.
Parallel agent demos look great right up until two tasks both try to use the same mouse.
One task is logging into a site. Another one opens a browser window. A third thing is only supposed to answer a status question, but now the whole system is clicking in the wrong place, cancelling the wrong run, or reporting a failure that is really just resource contention.
That was one of ...
我写过一个开源项目，叫 Knowledge-and-Memory-Management，v0.0.2。
它的定位很简单：帮 AI Agent 解决知识从哪来、怎么存、如何用的问题。 底座是一个叫 hermes-memory-installer 的记忆体系统，负责「记住」；KMM 在上面做了一层知识采集层——把碎片化的网页、视频、文档、公众号文章变成结构化的、可检索的知识节点。
说人话：你给 Agent 喂了一篇文章，它不只是读完就忘，而是自动生成笔记、录入知识图谱、同步到云盘，下次搜同类问题直接从本地命中。
项目里有个 AugmentedSearch 类，核心逻辑就一个 if：