Mapping music genres to AI:
classical = traditional ML, where the textbooks point but nobody's looking
pop = agentic AI, where the hype is, no barrier
indie = open weights, you self-host it in your bedroom
jazz = bayesian, who insist everyone else is just a special case of them
Playing Oliver Qiu with SUB/WAVE & Navidrome hosted locally, where my local LLM generates DJ narrative and my local TTS speaks it.
最近 youtube music 的订阅要停了,开始琢磨新的听音乐解决方案,坚决不想用 spotify,考虑 bandcamp 但是还是不想捆绑在单个平台。然后反思我其实想要的是往 winamp 里拖我到处下载的 mp3 的那种感觉。为什么不 self host 一个音乐服务在我的 mini pc 上?然后就掉进了navidrome 这个兔子洞。
这是我看到基于 navidrome 最酷的一个产品:一个从 self-hosted 的音乐库拿音乐出来不间断播放的电台,可以连 local llm & tts 来生成电台 DJ 的声音。 getsubwave.com/
Nowadays the only criteria of buying a baby cam is whether it gives you free RTSP stream access. It's the same reason people choose Obsidian over Notion.
Started using pi calling my self-hosted qwen 3.6 35b. ~20 tokens/s from my Beelink. Experience has been quite good so far.
My baby cam has built-in cry detection but it only works inside its own app. So I:
- pulled the RTSP stream from the camera
- ran YAMNet on my mini PC
- piped alerts to Telegram
Once you have the RTSP, the camera becomes a sensor. Not just sound - image + ML opens up a lot more
I built a pipeline that takes random posts from my old Chinese tweets, translates them to Dutch, and sends them to me as voice messages on a schedule. It's technically a language learning tool but it feels more like a time machine that speaks in Dutch.










