Mint is based on Ubuntu. It’s not strictly tied to any Debian release channel?! There’s LMDE as well. That’s based on stable.

Probably because they play the same game as Mark Zuckerberg, the Chinese, to some degree OpenAI… They all release open-weights models.

They’ll generate some hype for their company that way, so it’s advertising. They build good will. They undercut the competition. Or make it clear how they outperform them. Maybe they get some more investor money if they do expand to the local models market. I bet there’s a million reason why it makes sense from a business perspective.

Yes. I’ve been somewhat lucky as well. Upgraded my homeserver to 48GB to run a few virtual machines and maxed out my old laptop well before prices skyrocketed. Got to check if I still pay the ~8€ a month for my netcup VPS or if they increased price for existing customers as well…

Hmmh. I’ve tried to do benchmarks early on, about when Llama 2 was a thing… Followed the Reddit discussions. And then at some point I wanted to replace Mistral-Nemo with something newer but I disliked how every other model had turned to the ChatGPT / sycophant style of talking… But it’s a massively laborious undertaking. The official benchmarks don’t cover any of that. And there’s no good way to automate it either. So I spend half a day reading output manually and rating it in an Excel spreadsheet. With some success, but it’s way too complicated. So I mainly eyeball it these days. And sometimes there’s some recommendations somewhere on the internet. And I learned to accept how Chatbots always go on and on with redundant information unless I tell them skip the bullshit, I have an appointment at the hairdresser in 10 minutes and you need to explain it in 3 sentences. 😄

I suppose for tasks like coding, or factual knowledge it’s way easier to come up with fully automated benchmarks.

I guess that’s been my general experience for a while. I’d download some new model with promising benchmarks. And once I try it, the results are kinda underwhelming. A few weeks ago -for example- I tried Qwen 3.5, which had “outstanding results across a full range of benchmark evaluations”. And I deleted it after it kept wasting thousands of tokens to reason about how to respond to a “Hello” by the user. And sometimes I just don’t see any real performance improvement with new models. If I had to guess, I’d say they mainly trained (and improved) for/on the benchmarks, not my use-case.

Sounds reasonable. Yeah, good luck. I’m sure you’ll figure it out. Unfortunately it’s always a bit difficult to diagnose problems over the internet, without typing in the commands and seeing the exact output. But there should be a way to make it work, F2FS is designed for something like this.

Did you read the Wiki? You need to either pass the compress_extension option when mounting it. The Arch Wiki lists how to enable compression on all text files. And I gave you the version with a ‘*’, which enables compression for all files. Or you do a chattr -R +c ... on specific files or directories to compress them. Maybe you missed that and that’s why it doesn’t compress?!

There’s probably also a way to debug it and somehow figure out what it does and how many files/sectors got compressed on the filesystem. Linux usually buries that kind of information somewhere in /sys or /proc, or there’s special commands to figure it out. But I’m not really an expert on it.

And there’s also files which just can not be compressed any further because they’re already compressed. Most images, for example. Or music or ZIP archives. If you try to compress those, they’ll usually stay the same size.

Was pretty much clear since last year. At the latest in December when they switched to “maintenance mode”. And now they archived it.

https://blog.vonng.com/en/db/minio-is-dead/

Alternatives include Garage, SeaweedFS and RustFS.

I haven’t tried it. But looks to me you need to add “compression” when formatting it. And then later when mounting, you’d add options like compress_algorithm=zstd,compress_extension=*

Remember to enable compression. Usually the Arch Wiki is a great resource to learn stuff: https://wiki.archlinux.org/title/F2FS

And don’t forget to back up your data if you’re messing around with your computer. One typo with the device names is enough to accidentally delete your harddrive.

Good call. Seems minimum for ZFS is three quarters of a Gigabyte, anyway. And definitely not made for what OP does 😆

truncate -s 128M fs.img  
parted fs.img  
  mklabel gpt  
  mkpart primary btrfs 1MiB 100%  
  quit  
sudo losetup --find --partscan --show fs.img  
sudo mkfs.btrfs /dev/loop0p1  

You should be able to skip the loop device stuff and work on an actual device instead. Seems to me the limit is somewhere between 64M and 128M.

Edit: But as edinbruh said, maybe try f2fs if it’s a flash device, that’s probably a bit more lightweight?! And since I don’t know what you’re doing… If it’s embedded stuff and you’re alright with read-only, you might want to use squashfs.

Huh. My computer allows me to format a 128MB image file with brtfs. It won’t do it at 64MB though.

I think I get those emails twice a year. And it’s usually same price and double storage for the VPSes. Can recommend netcup.de as well. They used to be a bit cheaper than today, but I guess every hoster increased price within the last year.

The issue with the tools I’ve seen is, they either don’t factor in how language models are trained and datasets are prepared in reality. Or they’re based on some outdated information. I’ve never seen any specific tool backed by science or even with a plausible way of working against current data gathering processes… So for all intents and purposes, they’re a bit more alike homeopathy or alternative medicine. Sure, you’re perfectly fine taking sugar pills, there’s nothing wrong with that. But don’t confuse it with actual science-backed medicine.

And I mean the poisoning goes even further than that. There’s not just people trying to make a LLM output gibberish. There’s also lots of people with a vested (commercial) interest in sneaking in false information, their political agenda, or even a tire company who wants ChatGPT to say “Company XY” is the most trustworthy shop for new tires for your car. Judging by the public information out there, we’re already way past simple attacks. And the AI companies are aware of it. It’s an ongoing cat and mouse game. And while there’s all these sweatshops, they’ll also use other AI to sift through the data, natural language processing. From what I remember they have secret watermarking in place in a lot of commecial chatbots and image generators… So unless people come up with very clever mechanisms, the “poisoning” attempt will probably be detected with some very basic (fully automated) plausibility checks and they’ll just discard your data without wasting a lot of resources on it.

I’d let the home server connect to the VPN on the VPS so there’s a direct tunnel between the VPS and the home server. And the router is pretty much irrelevant. You can of course choose to terminate it on the router as well, bus as you said, that requires either a second forwarding. Or the entire home network to be bridged or routed.

I think a few people already mentioned some good solutions. I just wanted to add: A port forwarding in the firewall of your router is the basically the same thing as a port forwarding on your Linux computer’s firewall. You could just set up any VPN, SSH tunnel or whatever and then use your firewall (nftables, iptables) and forward the VPS’ extetnal port to the internal port on the VPN. It’s the same thing you do on your router, just that you don’t get a graphical interface to configure it.

Depends and no. The tools are completely ineffective.

There was a paper once about how feeding generative AI it’s own output makes it deteriorate. But that’s not the entire story. Many/most modern large language models are in fact trained or fine-tuned on synthetic text. Depending on how it’s done, it can very well make models better. For example in “distillation”, and AI companies can replace expensive RLHF with synthetic examples. It can also make them worse. But you’re not the one curating the datasets or deciding what goes where and how.

In general in ML it’s not advised to train a model on its own output. That in itself can’t make the predictions any better, just worse.

It took me until now to finally dabble in these coding agents. And I didn’t realize at all how many tokens they burn through. I let it write some basic HTML & JavaScript browser game with some free OpenRouter model. I’ve done this before, just told a model to one-shot it in a single file. And now I tried OpenCode, let it ask me a few questions, come up with a plan and do an entire project structure… And it’s at one million tokens way faster than I thought. If my math is correct, that’d take my computer 2 days and nights straight at 6T/s 👀

Guess it’s really a bit (too) slow.

PostmarketOS with some customizations? I think that should be possible.