Yarrr! But it really mostly is Yarr these days. So don't go firing up Trumpet winsock to check Forte Agent :)
SuspiciousCarrot78
joined 2 weeks ago
I was tempted to say $0, but then I thought harder about the problem.
Technically I do have ongoing costs
- PAYG costs for Usenet-news (iirc, $22USD for 500GB block)
https://usenet-news.net/index1.php?url=home
- News indexer (I think...$60 every 5 years?)
Electricity (whatever tiny amount raspberry pi sips). At a guess, maybe $50/yr.
So, amortised over time - very low but not zero. In theory, if I dropped Usenet, it would even lower. And theoretically, I could run the pi off a single solar panel and a diy solar kit but I'm not busy pretending to be Robinson Crusoe just yet. Though... It might be a cool project.
No good deed goes unpunished. The sense of self entitlement some people display is staggering. FOSS project? Well, you should have done x y or z.
Also, I gave you $3 via Ko-fi, so you need to provide customer support in perpetuity and come to my house and install it. And heaven forbid you try to recoup costs!
Projects don't just die out - a lot of them are killed (one way or another). For example, I had a fully specced out FPGA design that would capture the signal from Wii GPU and do internal upscaled resolution (think: like what dolphin emulator does but with actual hardware) not just post process sharpening. Total cost under $100 and some know how.
The amount of flack I copped for it made me shut down the github and work on it for myself. Once it's perfected, I may post about it again but I sure as shit am not compelled to deal with the fucking peanut gallery anymore.
There are many excellent options - far too many to list. So I will briefly say - there are some really nice 4B models (like Qwen3-4B HIVEMIND, Nanbeige, IBM Granite 3B) which you should be able to run at higher quants (Q6 and up) quite nicely. Of course, there are always newer models (Gemma, Qwen3.6 - soon 3.7) etc.
Best bet is to poke around hugging face, on TheBloke, Unsloth or DavidAUs archives and see what they have in the 3-7B range that tickles your fancy. Don't immediately jump for the newest releases - the old ones are still good. Qwen3-4B 2507 instruct is still a favourite of mine and more recently Qwen3.5-2B shows promise.
The Jellyfin vs Plex thing always struck me as odd. As in - why are we holding JF to a different standard to (say) Immich, Syncthing, Pi-hole or any one of a thousand different programs people self host?
Yes, JF ships multi-user accounts and client apps etc. I get it, "multi-use" is implied, so the comparison isn't totally unfair. But there's a difference between 'this feature exists' and 'this is the primary purpose of the tool'.
The fact that you CAN share it externally doesn't mean everyone running JF is doing that, or that it should be the benchmark the whole project is judged by.
To me, self host means "I host it, myself" not "I host it and then pretend to be Netflix for family and friends". If that's the use case, then of course, Plex away.
It's cool that you CAN share JF externally, and it's cool that Plex does that differently / better. We shouldn't hold one to the standards of the other.
Well, it IS French. All the best evil comes from France :P
Funny, I was just thinking last night how stupid Haiku seems compared to Sonnet...and all the while, that chipper little fuck was our only hope of avoiding AI apocalypse LOL
Do wish they had used more LOCAL models, instead of cloud based. I'm pretty sure Granite would have told em to go pound sand. That thing is straight laced to point of absurdity.
There use to be one called PDFDRIVE. I mean, there still is, but there use to be too :)
I actually have a theory here...I think there's a bare basement level that a model needs to be...anything above which, deterministic tooling can do the rest. We've just been yeeting into a black box.
Why that matters is this - if you can make a 450M model do what a 7B model does...that has a huge set of implications (see above examples), not least of which is for use GPU poors.
I'm doing some smoke testing on this idea right now for what I'm calling an 'expert system', where the model is treated like a squawk box and the infrastructure around it provides the brains (not RAG, per se. More like sidecars or tool calling). I'm liking what I see so far but there's lots of fucking work to go. There may yet be a cheat code for some of the NVIDIA tax, if we take the work outside of the magic parrot :)
No? Just me then. How about this - 99% accurate COPD cough count...with a itty bitty convolutional model, on a $30 Adurino.
https://www.edgeimpulse.com/blog/ai-dont-like-the-sound-of-that-cough/
Why this might be cool. Different coughs correlate to different conditions (aka there is work going on in cough acoustics as a diagnostic signal / proxy for spirometry and breath sounds).
The above was trained on his coughs...it's not far from there to "was that a healthy cough, wet cough, dry cough, wheeze? Is this a Blue Bloater or Pink Puffer?"
I've long suspected PoC (Point Of Care) systems could be adapted to use language models. Imagine - Qwen3.5-2B (with --mmproj) that lives on your phone...and you can point at mole or freckle and ask "hey...is this fucky or what" - and it actually KNOWS because it has access to DermNZ and can classify based on ABCDEs
Debatable :) Torrents rely on seeders. I've downloaded movies and TV shows >5 yrs since initial upload via Usenet. Yes, things expire there too (eventually), but when the getting is good, it's uniformly good / fast.
OTOH, 1337 has been pretty decent to me of late.
It's tricky. On one hand, Jellyfin and the arr stack are what got me into self hosting. OTOH...torrents are simpler - I can plug my external SSD directly into my router, which streams to NovaPlayer on any android device - nothing else needed. Want a new show / movie? Grab the torrent, punt it across to ssd via samba share. It auto populates.
https://github.com/nova-video-player/aos-AVP
It's...simpler. Arguably more elegant / less moving parts.
Dunno.