Generally the app is better. Compatible with more container formats, audio formats (surround sound, Dolby digital, etc), and has hardware supported decoding for h265 video in addition to h264.

At least in the case of a Jellyfin server, you can download media locally when you know you’ll be without internet

If you switch the devices line to

- /dev/dri:/dev/dri

as other have suggested, that should expose the Intel iGPU to your Jellyfin docker container. Presently you’re only exposing the Nvidia GPU.

QSV is the highest quality video transcoding hardware acceleration out there. It’s worth using if you have a modern Intel CPU (8th gen or newer)

Jellyfin doesn’t need any particular setup to work directly from LAN because it doesn’t ever try to use a central login provider the way Plex does.

The only reason OP is struggling with it is because they set it up so that they can only connect to it via Tailscale.

So first of all, you shouldn’t involve yourself in your friend’s business. Fraud is generally frowned upon.

But secondly, you know that ChatGPT was trained on the entire internet, right? Like, every book. I don’t think “more books” is gonna help.

I hope you take your computer skills and make something of yourself. Try not to get any more involved in this scheme, seriously. You don’t need this crap marring your reputation.

Besides, there are better reasons/ways to fight the system than helping other people avoid learning.

Just that they’re no easier to use to fool an anti-AI system than using ChatGPT, Gemini, Bing, or Claude. Those AI detectors also give false positives on works made by humans. They’re unreliable in the first place.

Basically, they’re “boring text detectors” more than anything else.

I believe commercial LLMs have some kind of watermark when you apply AI for grammar and fixing in general, so I just need an AI to make these works undetectable with a private LLM.

That’s not how it works, sorry.

Quantized with more parameters is generally better than floating point with fewer parameters. If you can squeeze a 14b parameter model down to a 4-bit int quantization it’ll still generally outperform a 16-bit Floating Point 7b parameter equivalent.