How to run useful AI on hardware you already own, or can buy second-hand for the price of a year’s subscriptions. Every number in these was measured on my own machines on the day of writing — not quoted from somebody’s benchmark.
Jellyfin on a Pi 5 is up in three seconds, costs 238MB of memory and scans a library in seconds. It is also 100 times slower at transcoding than at not transcoding, and the settings page will cheerfully offer you hardware acceleration this machine cannot do. Here is the whole build, the one dropdown to leave alone, and the query that found a library of 19,935 items of which three existed.
Two database tables, one shell hook and about fifty lines of PHP will tell you what every AI agent you run has been doing. Mine has recorded 57,875 heartbeats since April for 76MB. Here is the whole build, command by command — then the six ways it lied to me, every one of which reported success, and the six queries that catch them.