技术实验：开源仓库开盒，装起来跑起来压到出问题为止

这条线在做什么

GitHub 上每天都有新仓库冲上 Trending，README 写得漂亮，Star 涨得飞快，演示 GIF 流畅得不像话。但当你真的把它 clone 下来、装到自己那台不是 A100 的机器上、喂进自己那份不干净的真实数据——事情往往就不是那么回事了。

技术实验这条线补的就是 README 和现实之间的那段距离。

每一期的做法是固定的：

挑一个真实仓库——不是按热度挑，是按「拼一套 local-first AI 到底缺哪块」倒推着挑
在真实硬件上装起来——Mac、树莓派、旧笔记本、便宜的云主机，写清楚跑在什么上面
跑真实任务——用能代表实际使用场景的数据和负载，不是官方 demo 里那份
压到出问题——内存打满、并发拉高、断网、喂脏数据，看它在哪一步塌
给结论——推荐给谁、不推荐给谁、坑在哪、有没有更好的替代

为什么值得看

因为「能不能用」这个问题，只有真跑过的人才回答得了。

一个仓库有 3 万 Star，不代表它在 8GB 内存的机器上跑得动；文档里写着「开箱即用」，不代表你不用先花两小时解决依赖冲突；基准测试跑出漂亮数字，不代表换成你的数据还是那个数字。这些东西不会写在 README 里，只会出现在真正装过一遍的人的记录里。

和另外两条线的关系

选什么仓库来开盒，来自🧠 问题思考线维护的痛点地图——先确认这是个真问题，再去找解法
实测下来确实好用、又确实缺一块的，我们自己补的那部分会开源出去，进🌱 公共物品线

更新节奏

B 站和 YouTube 双平台同步更新，全部免费。长版 8-15 分钟讲完整部署过程，短版 8 分钟以内直给结论。

What This Line Does

New repos hit GitHub Trending every day. The README looks great, the star count climbs fast, the demo GIF is impossibly smooth. Then you actually clone it, install it on your machine — which is not an A100 box — and feed it your own messy real data. Things usually look different from there.

Tech Experiments exists to cover exactly that gap between the README and reality.

Every episode follows the same procedure:

Pick a real repo — chosen not by hype, but by working backwards from “what’s still missing to assemble a local-first AI stack”
Install it on real hardware — a Mac, a Raspberry Pi, an old laptop, a cheap VPS; we always state what it ran on
Run real work — data and load that represent actual use, not the ones shipped in the official demo
Push it until it breaks — exhaust the memory, raise the concurrency, cut the network, feed it dirty input, and see which step collapses first
Give a verdict — who should use it, who shouldn’t, where the traps are, and whether something better exists

Why It’s Worth Watching

Because “does it actually work” is a question only someone who ran it can answer.

30k stars doesn’t mean it fits in 8GB of RAM. “Works out of the box” doesn’t mean you won’t spend two hours on dependency conflicts first. A pretty benchmark number doesn’t mean you’ll see that number on your data. None of this makes it into a README — it only shows up in the notes of someone who actually installed the thing.

How It Relates to the Other Two Lines

Which repo gets torn down comes from the pain point map maintained by the 🧠 Problem Thinking line — confirm it’s a real problem first, then go looking for a solution
When something tests well but is still missing a piece, the piece we build gets open-sourced and moves into the 🌱 Public Goods line

Cadence

Published on Bilibili and YouTube simultaneously, always free. Long cut is 8-15 minutes with the full deployment walkthrough; short cut is under 8 minutes and goes straight to the verdict.