I’ve been using it for years and this has never been an issue for me.

Very cool. I’ll need to sit down and read the full spec, but I’ve got some FastAPI apps where I could see this being useful.

Absolutely there’s a difference. LLMs, when it comes to this specific task, are better. That’s why they’re being used here. It’s a job they are uniquely well suited to. They do indeed come with high hardware requirements, which is why you’re not forced to use them, and why they provide the option to offload the work to a cloud service.

Personally I would absolutely not want to ever feed my documents into an off-device model, but the point of self-hosted software is that it does what you tell it to and they absolutely should include letting you make bad decisions.

I think you’re under the impression that the difference between those things is far greater than it actually is. Large Language Models work by developing statistical maps of associations. That’s discrimination. They’re a direct evolution of categorization models. The ability to associate a hash of a JPEG with “cat” is the same as the ability to associate “How are you?” with “Great, how about you?” It’s all associative mapping. LLMs are just the current leading edge of that technology. If you want to, for example, generate a list of tags that describe a document, an LLM is the best tool we currently have for doing that.

To put it another way, what you term “discriminative AI” is “generative AI.” It’s generating a category or list of categories in response to an input. That’s not functionally different than generating a sentence in response to a sentence, it’s just an order of magnitude less complex. You can argue terminology but the technology exists on an evolutionary curve, with no real hard boundaries.

Paperless and Papermerge have always done a lot more than just OCR. If that’s all they were, most people can do that already on the software that comes with their scanner. The core selling point of these applications is automatic categorization, sorting and tagging, and those have always relied on machine learning tools. Literally the first thing you do after setting Paperless up is start training the AI.

what’s the best alternative?

You’re asking for suggestions for an automatic document categorization tool that doesn’t use any automatic document categorization?

It reads as especially hysterical in this context, because Paperless is an automatic document categorization system, and I’m really sure what they think the automatic part of that is if it’s not “AI” of some broad description. Paperless, Papermerge et al are basically wrappers for machine learning tools and have been for as long as they’ve existed. LLMs are a natural and obvious fit for the kind of work these applications exist to do.

This just feels like someone reading “Improved AI pathfinding” in the patch notes for a video game and screaming “OH MY GOD IS NOWHERE SAFE?!”

Modern routers have a feature that’s typically known as “hairpin NAT” or “NAT loopback”. This means that if the DNS lookup resolves to the router’s public IP, it routes the packet internally instead of passing it out to the internet. So for your subdomains all of the traffic should stay local. Typically this feature is enabled by default, but you may want to check just to be sure.

You’re not only the one who kame according to those numbers.

Interesting. If I’m following this correctly, the architecture is very similar to how services like Tailscale and Zerotier work; every client contacts a central server to say where it can be reached, and then the central server shares that information with authorized parties, but the actual communication is all peer to peer.

Assuming I’ve grasped this correctly, this sounds like a very smart use of a proven architecture. I guess my main question would be, how does the central server determine when to share contact info between two clients? Is everyone essentially aware of everyone else (on the backend at least) and then some client side component restricts who can actually communicate, or is there some pre-shared secret that would authorize the server to connect your client to someone else’s?

Edit to add: The other thing I was going to note is that obviously removing the central server means there’s no way to retrieve your messages if you lose or wipe your phone, but I see you’ve already included robust backup options. Thumbs up for that.

Edit 2: Just to clarify my question, I see that theres a system for scanning or importing QR codes in order to add someone as a contact (I do want to suggest that you should also be able to use a code; not everyone always wants to be sending an image in place of, effectively, a phone number), but my question is more about how this is architectured on the backend. Does the key encoded in the QR code authorize the TURN server to connect you with the other user?

It’s cool, I bought one for you.

“LOL, you guys thought you had privacy? Get rekt, scrubs.”

I think an absolutely great use for an old server at a youth center would be whatever the kids decide to use it for. Seriously, put it to a vote. There are lots of fun options, like a Minecraft server, one of the various Discord alternatives, a private wiki, or whatever else you find on the Awesome Selfhosted List . Just let it be their choice.

And, more importantly, set it up with them. Let anyone who wants to be involved in the process join in. If you’re kind of a newbie to Linux yourself, that’s even better; let them see you fuck up. Let them help you search for answers when you run into problems. Make it into a collaborative project.

Now, as for the actual hardware you have there… Man, that thing is ancient. If it’s what you’ve got it’s what you’ve got. As I’ve already indicated I think the process is worth more than the result here. That said, if you look around you can almost certainly find a refurbished old corporate PC for a hundred bucks or so that will be more powerful. Liquidation auctions are especially good for that kind of thing.

100%, Cockpit is what you’re looking for OP.

Locally? You’d need a VERY powerful GPU to really be able to match the capabilities of Opus 4.6 online. I’ve played around with this stuff for the same reasons and while you can absolutely run a model with all of Claude’s capabilities offline, very few people will have the hardware to let it actually run at an acceptable speed and with a sufficient context window. That last part is the most important thing for coding because it’s what allows the model to operate across an entire project and not just a few functions at a time.

You are gonna love Deadlock.

I’m a league refugee that switched to Deadlock. My god it is so much better. Both the game and the community.

A little of both. I think it’s good for software to be customizable, but layering on plugins often tends to lead to instability or other issues, so the ideal for me is where the program does 90% or more of what I want out of the box, and plugins fill the gaps.

So, yes, you’re basically correct.

There are search layers that remove the need to access radarr / sonarr directly when searching for shows (someone mentioned jellyseer, for example), so that part of the process can be streamlined, and once you’re watching a show it’s generally very good at pulling new episodes as soon as they’re available, so you’re typically, at most, a day behind actual airing dates. But if you’re trying to just bounce around and try a bunch of different shows it wouldn’t be the best for that. The biggest constraint is generally the speed of your internet and the popularity of what you’re watching. With a high speed connection and a well seeded torrent it’s often only a a couple of minutes to download a pilot episode, and you could have the whole season done by the time you finish watching that.

The other question is one of storage. If you’ve got plenty of hard disk space then you can probably afford to just throw anything that sounds interesting on your pull queue and work your way through it when you actually have time to sit down and watch. Basically you sort of pre-emptively build your “Netflix at home” library and then do your bouncing around channel hopping stuff with the five or so vaguely interesting shows that you added while you were at work.

Is it a replacement for Netflix et al? Not strictly speaking, but if you don’t mind changing up your habits a little it’s probably close enough.

I do, but it’s more out of laziness than anything else. I hate having to remember sixteen different ways of doing things, so I tend to configure all my stuff as identical as reasonably possible. Is this the best way of doing things? Probably not. But it keeps my blood pressure down.