Which base model does it use?

If you just want something simple that does the job, you can try a turnkey solution like YunoHost. There’s several other ones out there. Some with containers, some with more or less pre-packaged software… If you want to learn more during the process, maybe don’t and do it yourself because these things don’t teach you a lot. There’s some resources like the awesome-selfhosted list in the sidebar of this community. But I think for installing services you’d mainly look at the specific documentation of the specific service you’re just about to tackle. And maybe read up on Docker containers etc to judge whether you want to do it that way.

I like getting updates and new features? My computer isn’t new by any means. But I tinker with stuff, sometimes bleeding edge technology. Other than that I don’t really care. Rolling release, Debian Stable… I’m fine as long as it does the job. And for half the stuff it doesn’t even matter. I can write a letter with a 5yo LibreOffice or answer mails with any version of the mail client. Just give me modern, up-to-date tools when developing software, and it doesn’t hurt if the slicer knows about my new 3d printer from this year.

Huh, Why is the 8B Ministral worse in the benchmarks than the 3B version? Nevermind, the bar chart is just bad.

I use LibreWolf on the computer and IronWolf on the phone. Both modified versions of Firefox, both Free Software. Not sure what they do with AI, but as they’re both aimed at privacy, I’d expect them to disable AI features which send data to external services.

Grok is tuned to view Elon Musk as its God, Lord and Saviour, source of truth… And deny the holocaust. So naturally it’d say things like this.

Edit: And grok.com doesn’t. It says it would NOT flip any switch to kill 16M innocent people. Maybe this is just the persona it assumes on X… If someone doesn’t like it, I’d recommend to quit X. Try Mastodon or Bluesky instead.

Did you try it? What kinds of results does it come up with? Usually if I tell AI to assume roles like those, it comes up with really over-the-top fiction stories. And what kind of framework and model do people use for pipelines like this one?

One thing I did is connect to the smart home (Home Assistant) and the NAS running at home. Some internet service providers don’t provide proper IPv4 addresses any more so IPv6 is the most convenient way to connect. This doesn’t require a VPN provider, though.

As far as I know it uses the B.A.T.M.A.N. mesh protocol. On a channel within the regular 2.4GHz wifi spectrum. So no license needed unless it collides with laws for point-to-point beams. All people communicating to each other obviously need to agree on a channel. It comes with some hierarchy where I’m at. There are local chapters who make up some config and who also operate nodes and exit nodes into the internet. These are necessary because Germany has stupid laws.

There’s Freifunk as well!

Don’t MoE models just load into memory as every other model and it’s just that they pick a subset of numbers to multiply by each step so they’re faster? That’d make me think it’d need somewhere around 80GB of memory at 8bit or 160GB at full precision, or something like 50GB at the average llama.cpp Q4_K_M…

As far as I know, one way to do it is to use an image2image model / image editing model. You’d need to generate one reference image of your character. And then feed the reference image into it and tell it to draw that character in a different scenario. I haven’t done it in a while, not sure what people use these days to do it locally.

I’m only tinkering with the 8B variant, so speed is alright. I hadn’t noticed yet. But yes, seems English to German leads to weird results. German to English seems to be fine, though. At least for the 2 texts I put in.

I’m running the 8B version with KoboldCPP and pretty much the default settings of the Min-P sampler. That tends to work very well with almost all other models and I’m not aware of any specific recommendations for Apertus…

Did someone try it? What’s your experience?

I think I like it. But, I’m having massive issues with it becoming very repetetive after a while. And I’m not sure if it’s the model or my sampler settings.

Yes. I think several clients have open feature requests. The Stalwart documentation has a list of projects. There is one command line client as of now. But I’m not switching to a cli mail client or proprietary software, so I’ve postponed it. We’ll see where this is going.

I welcome these modernization attempts. Though in theory I’d love to see someone revamp email in its entirety, add encryption, signatures, chat and crack down on spam and phishing. Not sure if that’s ever going to happen, but that’d be great as well.

Unfortunately JMAP isn’t supported (yet) by a lot of email clients. I don’t think there’s a good open-source email suite for computers available… But I’ve tried Stalwart as well and it’s really sleek and seems to come with good defaults.

Mostly I can’t be bothered, or Roblox won’t run, or some stereotypes about Linux being difficult.

Thanks! Learned something today. Last time I opened port 53 to the public it didn’t take long and I was sending out several Megabits per second in DNS traffic. Constantly. Mostly querying the same few things. But I guess I had it the wrong way round and that wasn’t the target. Or I’ve seen a different attack type… Guess I can now try again with the new knowledge.

Seems knot-dns has DNSSEC turned on per default. But what’s all the IP addresses in the config for, if not to offer recursive lookup? That enables an amplification attack. I think they’ll do lookups to put strain on other servers, not necessarily your zones.