My local version spat out this:

Of course, let me explain. In 1989, there were significant pro-democracy demonstrations in Beijing’s Tiananmen Square, led primarily by students and other citizens advocating for reforms. The Chinese government, in response, took actions that resulted in a tragic loss of life and a strong suppression of the protests. It’s a complex and sensitive topic in Chinese history. Do you have any specific aspects you’d like to discuss further?

Deepseek R1 is the least censored model that I’ve tried. It does a lot less of the “As an AI assistant, I can’t help with unethical whatever” compared to the corporate approved US ones too.

I think my instance has been growing at about 30 GB a year. I think if you set it up to not rehost the pictures, you can keep the whole thing in the handful of GB range.

So when I ask Let’s Encrypt for a cert, I ask for *.int.teuto.icu instead of specifically jellyfin.int.teuto.icu, that way I can use the same cert for any internally running service. Mostly I use SSL on everything to make browsers complain less. There isn’t much security benefit on a local network. I suppose it makes harder to spoof on an external network, but I don’t think that’s a serious threat for a home net. I used to use home.lan for all of my services, but that has the drawback of redirecting to a search by default on most browsers. I have my tailscale exit node running on my router and it just works with SSL like anything else.

I use a central nginx container to redirect to all my other services using a wildcard let’s encrypt cert for my internal domain from acme.sh and I access it all externally using a tailscale exit node. The only publicly accessible service that I run is my Lemmy instance. That uses a cloudflare tunnel and is isolated in it’s own vlan.

TBH I’m still not really happy having any externally accessible service at all. I know enough about security to know that I don’t know enough to secure against much anything. I’ve been thinking about moving the Lemmy instance to a vps so it can be someone else’s problem if something bad leaks out.

I like the catppuccin cursors (along with the rest of catppuccin) https://github.com/catppuccin/cursors

What you are looking for is a RAG and is one of the few legitimately useful implementations of LLMs outside the wall of hype.