GDPR for Dummies. GDPR is DSGVO in German, “Datenschutz-Grundverordnung”.

bummer.

Ah. maybe hand it over to the next person? I suppose people still need to switch painlessly? But I get it. We used to host lots of stuff in my university years. A forum, chat, classifieds, filesharing… A big photo album for all our pictures and events… As far as I know all of that has gone. Either due to lack of interest or nobody was able and willing to pick it up.

Why both? Isn’t one enough?

Alright. I think I’m fine with that. I’ll read the documentation and see if I can find some time over the weekend to set it up. I’ll report my success in case I remember. It’s going to take me some time to notice missing information, though. And I’m not sure how caching of images works. We’ll see. I hope I don’t get banned, but I don’t really care that much. Thanks for the insights.

What does it do with threads, just display the messages sequentially? Or not at all?

Guess it’s mautrix then and I have my answer. Thanks!

Thanks. Why do you say appservice is old? Because of the used technology? Or because their last commit is from 9 months ago?

I don’t really need double puppeting. I’m just interested in funnelling out data from discord and it’s just me and my account. I should clarify that in my post…

Thanks for the suggestions. I’m gonna take a closer look at the features of both and probably choose mautrix, since I already run their WhatsApp bridge. Maybe someone else chimes in and can contribute their experience with the out-of-your-element bridge.

I think most people use something like exllamav2 or vllm or use GGUF to do inference and it seems neither of those projects have properly implemented multimodality or this specific model architecture, yet.

You might just be at the forefront of things and there isn’t yet any beaten path you could follow.

The easiest thing you could do is just use something that already exists, be it 4bit models, wait a few weeks and then upgrade. And I mean you can also always quantize models yourself and set the parameters however you like, if you have some inference framework that supports your model including the adapters for vision and has the quantization levels you’re interested in…

Well, I’d say there is information in language. That’s kinda the point of it and why we use it. And language is powerful. We can describe and talk about a lot of things. (And it’s an interesting question what can not be described with language.)

I don’t think the stochastical parrot thing is a proper debate. It’s just that lots of people don’t know what AI is and what it can and cannot do. And it’s neither easy to understand nor are the consequences always that obvious.

Training LLMs involves some clever trickery, limit their size etc so they can’t just memorize everything, but instead are forced to learn concepts behind those texts.

I think they form models of the world inside of them. At least of things they’ve learned from the dataset. That’s why they can for example translate text. They have some concept of a cat stored inside of them and can apply that to a different language that uses entirely different characters to name that animal.

I wouldn’t say they are “tools to learn more aspects about nature”. They aren’t a sensor or something. And they can infer things, but not ‘measure’ things like an X-ray.

I’m currently reading the paper. I occasionally debate here on Lemmy whether LLMs are just stochastic parrots, or if they actually grasp the concepts they’re talking about. There’s also evicence for that.

Ultimately I wonder if and when we’ll get LLMs that address ‘hallucinations’ and expose a setting to adjust the factuality of the answer. I suppose that’s somewhere in the model or at least possible to learn for the model. But certainly not controlled or factored in in the current generation of LLMs.

services.tabby.enable = true;
services.tabby.acceleration = "cuda";

? Could be another way.

https://github.com/YellowRoseCx/koboldcpp-rocm

That will be optimized for AMD and as far as I know has the same / a very similar user interface.

(The 8GB of VRAM on your graphics card will be some limitation. So maybe stick with smaller and quantized models.)

And share your success stories on !ChatbotsNSFW@lemmynsfw.com

I recently listened to this German language podcast episode about the social cost and how life is for a few clickworkers in Africa: Das Wissen | SWR: Clickworker – Ausgebeutet für künstliche Intelligenz

Yeah, doesn’t really work. I mean it has a rough idea of that it needs to go east. And I’m surprised that it knows which interstates are in an area and a few street names in the cities. I’m really surprised. But I told it to get me from Houston to Montgomery as in your example. And in Houston it just tells random street names that aren’t even connected and in different parts of the city. Then it drives north on the I-45 and somehow ends up in the south on the I-610-E and finally the I-10-E. But then it makes up some shit, somehow drives to New Orleans, then a bit back and zig-zags it’s way back onto the I-10. Then some more instructions I didn’t fact check and it gets that it needs to go through Mobile and then north on the I-65.

I’ve tested ChatGPT on Germany. And it also gets which Autobahn is connected to the next. It still does occasional zig-zags and in between it likes to do an entire loop of 50km (30 miles) that ends up 2 cities back where it came from… Drives east again and on the second try takes a different exit.

However: I’m really surprised by the level of spatial awareness. I wouldn’t have expected it to come up with mostly correct cardinal directions and interstates that are actually connected and run through the mentioned cities. And like cities in between.

I don’t think I need to try “phi”. Small models have very limited knowledge stored inside of them. They’re too small to remember lots of things.

So, you were right. Consider me impressed. But I don’t think there is a real-world application for this unless your car has a teleporter built in to deal with the inconsistencies.

Thanks. Yeah I know most of the story/history of Matrix. I’m just now making the decisions for the years to come. And Dendrite has been the announced successor to Synapse for quite some time now… I’m not sure what to make of this. If it’s going to happen soon, I’d like to switch now. And not move again and relocate my friends more times than necessary.

Judging by the graphs on my Netdata, Synapse plus the database are currently eating more resources than I’d like for just chat. Afaik the other projects were meant to address that. But I’ve never used anything else. And I’ve always refrained from joining large rooms because people told me that’d put considerable load on the server. If there’s a better solution I’m open to try even if it’s not the default choice… It just needs to work for my use-case. I don’t necessarily need feature-completeness.

Yeah, with the multiple domains: I meant I have 1 VPS and like 3 domain names for different projects. I have a single email-server, one webserver and they just handle all three domains. Even Prosody (XMPP) has “VirtualHost” directives and I only need to run it once to provide service on all the different domains. With Matrix this doesn’t seem to be the case… I’d need to launch 3 different instances of Synapse simultaneously on that one server and do some trickery with the reverse http proxy. That’d be more expensive and take more time and effort. I don’t really care about how the identities are handled internally, I can provide them in a format that is supported. And the users are seperate anyways. It’s just: I’d like to avoid running the same software three times in parallel.

Out of curiosity: Do you have to deal with that much spam? If so: Is there a specific reason?

Because I only get some bot join one of the public rooms and start spamming every few months or so. And we deal with that pretty quickly. My own account has been perfectly safe for years… So my experience is different. Might be my usage-pattern vs yours?!

Which model(s) did you try? I’m willing to test it later. Downside is, I mainly use smaller LLMs, live in Germany, in an urban region with lots of streets and different Autobahnen and it’s kind of a hassle to deal with textual driving instructions anyways. 😆

Quite some AI questions coming up in selfhosted in the last few days…

Here’s some more communities I’m subscribed to:

• !localllama@sh.itjust.works
• !fosai@lemmy.world
• !stable_diffusion@lemmy.dbzer0.com

And a few inactive ones on lemmy.intai.tech

I’m using koboldcpp and ollama. KoboldCpp is really awesome. In terms of hardware it’s an old PC with lots of RAM but no graphics card, so it’s quite slow for me. I occasionally rent a cloud GPU instance on runpod.io Not doing anything fancy, mainly role play, recreational stuff and I occasionally ask it to give me creative ideas for something, translate something or re-word or draft an unimportant text / email.

Have tried coding, summarizing and other stuff, but the performance of current AI isn’t enough for my everyday tasks.