APU board? They are going EOL soon, but these devices are built like a tank. Full Linux x86_64 support, coreboot bios. https://www.pcengines.ch/apu.htm A few sellers in the EU still have them.

I use a mixture of systemd-nspawn and different user logins. This is sufficient for experimentation, for actual use I try to package (makepkg) those tools to have them organized by my package manager.

Also LVM thinpools with snapshots are a great tool. You can mount a dedicated LV to each single user home to keep everything separated.

After many years of using multiple devices and even servers with Archlinux installed it never broke because of an update (spoiler: I use systemd-boot instead of grub). If a system is to be used by a less experienced user, just install linux-lts Kernel instead.

Unstable does not mean it crashes/breaks often, it just means it does not guarantee to not bump to the newest upstream version and that it does not do backports. This can be a problem when using unmaintaned software that does not like using a recent python/php.

This is also great because if you find a bug in a software you can report it to upstream directly. Debian maintainers only backport severe bugs, not every one of them. It can take over a year for new features to arrive - especially painful with applications like gimp, krita, blender, etc. You can use debian-unstable of course, which is close to upstream as well.

Do not expect you can offer this service for a competive price against cloud prices. Caring for a company IT system is a big challenge and requires more work the more users there are.

For a company this size: make a clear contract. Consider how much time you need for setup/installation, monthly hours for maintenance, monitoring and at least daily(!) backups. Let them choose if they want it with a failover and charge for the required hours and material. Also put in the contract when they can expect support from you, including a clause for a holiday substitute admin (if needed). Then put a pricetag on support hours for holding people’s hands when they “can’t find that file they uploaded a week ago and it is surely a server issue” and put a pricetag on engineering hours for any modifications they might want, like installing any plugins they deem useful for themselves. Hardware prices, traffic, rack space and power should be included as well. Have a good plan for updates, choose your distro wisely, do not rely on autoupdates.

Play all this through in your head, add up the hours, choose a fair rate and then you have your pricetag.

Cloud will always be cheaper, because they have their infrastructure already deployed. Building from the ground up is more expensive, but I think it is worth it. Will they?

But…isn’t unsupervised backfeeding the same as simply overtraining the same dataset? We already know overtraining causes broken models.

Besides, the next AI models will be fed with the interactions from humans with AI, not just it’s own content. ChatGPT already works like this, it learns with every interaction, every chat.

And the generative image models will be fed with AI-assisted images where humans will have fixed flaws like anatomy (the famous hands) or other glitches.

So as interesting as this is, as long as humans interact with AI the hybrid output used for training will contain enough new “input” to keep the models on track. There are already refined image generators trained with their own but human-assisted output that are better than their predecessor.