#linux Bot Logged User list

Network: Rizon

Modes: +CNRntz

Last Seen: 11 minutes ago

Topic: Welcome to #linux ! | Channel Rules: https://wiki.rizon.net/index.php?title=Linux | Ask your question(s) and be patient as it is different time(s) around the globe | NUchat IRC client: https://github.com/lord3nd3r/NUchat

#21

Rank

123

Users

Channel Log Archive for #linux

Hide Joins

* All times are UTC

Filtering by user: brothchild

Wednesday, April 15, 2026

[03:07:05] + brothchild glitchy: can you summarize the URL link provided by zathras about 5 hours ago?

[03:07:58] + brothchild glitchy that's it. was the congress.gov url unreachable?

[03:09:18] + brothchild glitchy that bill is tantamount to government being informed of all logins to any computer. Do you think it stands a chance in the courts?

[03:10:43] + brothchild glitchy suggest methods to remove supporters of that unconstitutional legislation from office

[03:11:26] + brothchild that's really nicely compact, good job to End3r

[05:39:39] + brothchild i'm so stupid and lazy i didn't really learn the 'find' command properly for decades

[05:40:21] + brothchild now i use find -exec maybe 50 times a day

[05:40:52] + brothchild but i still can't remember how to use print0 type stuff to properly handle filenames with spaces

[05:40:58] + brothchild so i rename every file to not have spaces

[05:41:09] + brothchild since a quarter century i rename all files

[05:41:26] + brothchild cause f* small-flaccid

[06:36:55] + brothchild Meow_Nemesis

[06:37:35] + brothchild glitchy in image-diffusion what does VAE stand for

[06:40:28] + brothchild glitchy harder question you may have to search https://github.com/ggml-org/llama.cpp. Does the llama.cpp project contain code that can be used as the text-encoder stage of the image generation pipeline for newer image gen models like flux (which uses qwen3-4b or qwen3-8b as the text encoder)?

[06:40:49] + brothchild glitchy harder question: Does the llama.cpp project contain code that can be used as the text-encoder stage of the image generation pipeline for newer image gen models like flux (which uses qwen3-4b or qwen3-8b as the text encoder)?

[06:42:06] + brothchild ooo bin/llama-embedding exists!!

[18:36:51] + brothchild hah "Systemd is toxic corruption: it the gene-jab clot-shot of linux"

[19:54:20] + brothchild hi de-facto

[19:55:58] + brothchild so i have loaded and running: Flux.2 Klein, generating 720x540 images in a batch job using 5-10GB, while gpt-oss-20b (a 12.6GB quant) is loaded and running

[19:56:18] + brothchild on a ryzen 3500u laptop with 13.5GB Usable ram

[19:56:22] + brothchild ... whatatimetobealive

[19:56:45] + brothchild no swap file or swap partition

[19:57:15] + brothchild and running browser with a ton of tabs

[19:57:26] + brothchild i am mister tweaker, legit

[19:57:49] + brothchild yeah

[20:05:26] + brothchild llama.cpp

[20:05:41] + brothchild you are informed de-facto

[20:06:46] + brothchild i should be more precise, i'm testing decensored versions

[20:06:50] + brothchild 12G gpt-oss-20b-heretic-ara-v3.i1-IQ4_XS.gguf

[20:06:50] + brothchild 12G gpt-oss-20B-INSTRUCT-Terse-Heretic-Uncensored.i1-IQ4_XS.gguf

[20:06:50] + brothchild 12G gpt-oss-20B-RichardErkhov-heresy-iQ4_XS.gguf

[20:07:02] + brothchild terse heretic is *beautiful* it's so terse

[20:08:10] + brothchild and the RichardEdkhov beat Qwen3.5 31B A3B on a hard coding challenge

[20:08:21] + brothchild (the others did not)

[20:08:30] + brothchild all on a 110€ laptop

[20:18:27] + brothchild we've seen abliteration improve scores and responsiveness as well

[20:18:37] + brothchild well not crude abliteration

[20:18:48] + brothchild it can be done with sufficient care to not reshape things

[20:18:56] + brothchild very interesting field of research

[20:19:07] + brothchild but it's beyond me

[20:19:29] + brothchild they've just (like days ago) identified the emotional impact of abliteration on the model

[20:19:35] + brothchild on a 4b and a 31b

[20:19:42] + brothchild or analyzed rather

[20:19:45] + brothchild it's wild stuff

[20:21:03] + brothchild saw a review of the stuff people are vibecoding with GLM 5.1, it's kind of depressing in a way

[20:21:21] + brothchild i enjoy algorithms

[20:21:38] + brothchild but i guess humans don't need to learn them anymore

[20:23:04] + brothchild i'm being stubborn and not running anything agentic

[20:23:19] + brothchild yeah i just ask it stuff on the side

[20:23:37] + brothchild wayy better than stuck-exchange.com or askubuntuknownothings.com

[20:24:19] + brothchild i think we'll see a lot of vibecoded messups

[20:25:27] + brothchild agree

[20:26:05] + brothchild ok maybe for some tasks you can adequately define valid inputs and outputs without caring about the how

[20:26:21] + brothchild yes

[20:26:32] + brothchild who is this de-facto person and why is he talking sensibly

[20:27:59] + brothchild End3r: likes to hurt people slightly

[20:29:49] + brothchild so de-facto the interesting thing is one of those abliterations solved a hard problem most <13GB models can not. 86/89 failed it

[20:30:23] + brothchild I'd really like a linux finetune

[20:30:35] + brothchild seems like we could get our act together and do one

[20:30:45] + brothchild would need work to generate good dataset

[20:30:57] + brothchild tag things by version to prevent pulling up outdated infos

[20:31:13] + brothchild generate synthetic queries and examples for commands and parameters

[20:32:02] + brothchild i'd like to see pluggable experts someday

[20:32:35] + brothchild "bind the the sql-server expert plus the java expert"

[20:32:43] + brothchild yes ofc

[20:33:54] + brothchild yeah see you don't want just some declarative statement

[20:34:09] + brothchild you want to synthesize fractal paths of questions and situations that lead to that data

[20:34:13] + brothchild in the form or queries or such

[20:34:40] + brothchild idk. i've just been reading abotu finetuning not doing it

[20:34:53] + brothchild for two years now

[20:35:12] + brothchild reading up on the failures is helpful too

[20:35:59] + brothchild also nobody seems to be doing the investigation of image diffusors

[20:36:06] + brothchild how are they trained for refusals etc

[20:36:53] + brothchild like i can ask flux.2 klein to generate something gross and it spits back a completely unrelated image

[20:37:17] + brothchild would be interested how that works

[20:37:35] + brothchild swapping out the text encoder didn't seem to make a difference

[20:38:13] + brothchild also funny i feel a hardware lust i haven't felt in decades :)

[20:38:33] + brothchild dual channel DDR5 seems like an insult now

[20:40:09] + brothchild split up every byte to 8 channels in the memory controller, how hard can it be?

[20:40:46] + brothchild https://github.com/antirez/iris.c image gen in C on CPU

[20:42:22] + brothchild is 1 really a thing?

[20:43:01] + brothchild oh yeah, i was thinking llm's again

[20:43:09] + brothchild yeah pretty sure they do that

[20:43:31] + brothchild with llms i dunno

[20:44:24] + brothchild mhm

[20:45:43] + brothchild "attention is all you need, but how much can you afford"

[20:45:54] + brothchild yes ^

[20:46:24] + brothchild i am a huge fan of gpt-oss-20b terse but

[20:46:43] + brothchild gpt-oss-20B-RichardErkhov-heresy-iQ4_XS.gguf has been a bit smarter

[20:46:58] + brothchild but you gotta try the terse, it's so nice to get compact answers

[20:47:06] + brothchild gpt-oss-20B-INSTRUCT-Terse-Heretic-Uncensored.i1-IQ4_XS.gguf

[20:47:31] + brothchild fast enough for an iGPU laptop too.

[20:48:47] + brothchild it's a happy kind of crazy to have both that and flux mmap'd with zram enabled

[20:49:11] + brothchild running 23GB of apps on 16GB RAM with no swapfile, no OOM's just occasional waits

[20:49:45] + brothchild i need to image this tweakfest

[20:50:14] + brothchild you know how shitty linux is with OOM out of the box...

[20:52:01] + brothchild nice

[20:52:16] + brothchild i am not yet initiated into llama-swap

[20:52:34] + brothchild anything you can tell me to save me from noob errors or something

[20:53:20] + brothchild swap isn't a real priority for me yet

[20:53:29] + brothchild llama-swap

[20:53:39] + brothchild do you run opencode or something

[20:54:31] + brothchild yeh .. i dunno

[20:54:48] + brothchild there's a barebones harness called CLIO i think that might align with me

[20:54:51] + brothchild in perl :)

[20:55:01] + brothchild ah

[20:55:11] + brothchild ty noted

[20:55:33] + brothchild like you said, i don't want to be out of the loop

[20:55:37] + brothchild on my projects

[20:56:24] + brothchild there used to be a neat project called UML - user-mode linux

[20:56:34] + brothchild before the whole container / docker etc craze

[20:57:06] + brothchild which i still think emerged mainly as a band-aid to python's cancer ecosystem

[20:58:49] + brothchild i don't want to be on the same planet as people who write scripts that download gigabytes of data to a tmpfile on a ramdrive

[20:59:06] + brothchild we are not the same

[20:59:26] + brothchild ✓

[21:01:01] + brothchild there's also things that will map-out entire project directories in such a way that llms can understand it using less context

[21:02:30] + brothchild but that's already the llm running too far ahead of me

[21:02:45] + brothchild i want to paste-in what I think is appropriate for the LLM's attention, i think

[21:03:55] + brothchild do you really want your llm ingesting the entire ffmpeg filters documentation or do you want to scan it visually and copy-in the functions you're interested-in?

[21:05:04] + brothchild hm makes sense

[21:05:44] + brothchild also people seem to have success with splitting up work between a 'planner' and a 'worker' model with seperate contexts

[21:05:46] + brothchild 'ensembles'

[21:06:00] + brothchild also something i'd like to play with

[21:08:22] + brothchild i've read human brains are a bit like two brains, with a generative and an inhibitor brain

[21:09:26] + brothchild is there any progress towards models that will 'recognise' that they don't know an answer and invoke a search tool?

[21:13:20] + brothchild idk if un-confidence about next-token is the right test, maybe it works

[21:13:53] + brothchild maybe it could be averaged over a sentence then scored

[21:14:42] + brothchild i am overdosing on curiosity

[21:16:18] + brothchild yeah i have expermineted a bit with samplers

[21:16:22] + brothchild not enough

IRC Log Archive

This is a searchable archive of IRC conversations from #linux on Rizon. Use the search box above to find discussions on specific topics. All conversations are preserved to help the community learn from past discussions.