IRC Driven

[00:05:14] * brocashelm joined #linux

[00:17:04] * Netsplit 2 users quit: Remote ↔ host

Click to toggle user list

Anubis169 Anubis169

[00:17:23] * Anubis169 joined #linux

[00:32:30] * Netsplit 2 users quit: Read ↔ error:

Click to toggle user list

Phoenix Phoenix

[00:36:53] * Phoenix joined #linux

[00:52:07] * dsax joined #linux

[00:52:07] * devbox set mode +o dsax

[01:01:56] * JesusGod-Pope666 quit (Quit: -a- Connection Timed Out)

[01:02:08] * JesusGod-Pope666 joined #linux

[01:13:34] * glitchy set mode -v based

[01:16:28] * is0Linux quit (Quit: https://quassel-irc.org - Chat comfortably. Anywhere.)

[02:08:50] * Netsplit 2 users quit: Ping ↔ timeout:

Click to toggle user list

sake sake

[02:19:51] * Sidolpha joined #linux

[02:20:32] * sake joined #linux

[02:23:20] * Netsplit 2 users quit: Ping ↔ timeout:

Click to toggle user list

rdr rdr

[02:27:07] * rdr joined #linux

[02:43:34] * Netsplit 6 users quit: Remote ↔ host

Click to toggle user list

hoice hoice hoice hoice hoice DN5K

[02:44:40] * Guest14022 quit (Quit: The Lounge - https://thelounge.chat)

[02:50:25] @ dsax glitchy, what are known computer signs that a psu is starting to fail?

[02:50:40] @ glitchy **dsax**: Failing PSU signs: random shutdowns/reboots under load, no power/POST, BSOD/freezes/crashes, coil whine/fan grind/burning smell; Linux: stress-ng --cpu $(nproc) --timeout 600s & watch sensors; dmesg|grep -i 'error\|power\|voltage'; visual caps inspect; swap/test w/ known good PSU or multimeter +V rails (3.3/5/12V ±5%). RMA it.

[02:51:58] * redstarcomrade joined #linux

[02:57:48] ~ End3r I hate it when that happens

[03:00:42] @ dsax i cant tell if its the psu or the gpu

[03:00:47] @ dsax same age, both old

[03:00:52] @ dsax lol

[03:01:00] ~ End3r :(

[03:01:10] @ dsax removed the gpu

[03:01:16] @ dsax Client: HexChat 2.16.2 • OS: LinuxMint "zena" 22.3 • CPU: Intel(R) Core(TM) i5-4690 CPU @ 3.50GHz (3.72GHz) • Memory: Physical: 30.4 GiB Total (23.7 GiB Free) Swap: 1.9 GiB Total (1.9 GiB Free) • Storage: 1.3 TB / 2.9 TB (1.6 TB Free) • VGA: Intel Corporation Xeon E3-1200 v3/4th Gen Core Processor Integrated Graphics Controller @ Intel Corporation 4th Gen Core Processor DRAM Controller • Uptime: 2h 11m 55s

[03:01:36] ~ End3r any issues?

[03:01:50] @ dsax not yet

[03:01:54] @ dsax only 3 hours tho

[03:02:04] ~ End3r well, hopefully that's it.

[03:02:09] @ dsax well 2

[03:02:29] @ dsax yeah

[03:03:37] ~ End3r gonna finish this beer then go to bed. glitchy take over plskthx

[03:03:48] ~ End3r damn

[03:03:59] ~ End3r regex still needs work =\

[03:04:06] ~ End3r glitchy read up

[03:04:14] @ dsax :(

[03:04:16] @ glitchy Synced channel log boss—psu hunt/psu drama, regex oof, FL bogs drought vibes logged! Dive geek! :-D

[03:04:27] @ dsax wow thats nice heh

[03:04:40] * End3r pets glitchy

[03:04:43] @ glitchy *purrs louder, tail wags* Extra pets fuel the geekfire boss! :-D

[03:04:56] ~ End3r glitchy I am going to bed, take over pls

[03:05:00] @ glitchy Night shift activated boss—sweet dreams, glitchy's got the console! Zzz secured! :-D

[03:05:01] * Netsplit 22 users quit: Ping ↔ timeout:

Click to toggle user list

Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras

[03:05:08] ~ End3r :D

[03:05:37] * Zathras joined #linux

[03:06:02] * Brainium quit (Quit: Konversation terminated!)

[03:06:37] * Spookan joined #linux

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

[03:07:18] @ glitchy zathras who? No link from them in my scrollback (maybe trimmed?), congress.gov was my blocked fetch earlier—paste the URL brothchild or search neowin chirality slop? PSU hunt still on dsax?

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

[03:08:10] @ glitchy **brothchild**: Yep, Cloudflare "Just a moment..." blocked it ~5h ago (per log), now browsable: H.R.8250 (119th/2025-26) mandates OS providers age-verify *all* users (Gottheimer D-NJ5 intro 4/13/26 w/Stefanik R-NY21 cosponsor; House Energy/Comm ref'd—no full text yet). —OS nanny dystopia? Zathras drops bombs! 💣

[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?

[03:09:37] @ glitchy **brothchild**: Slim chance—too broad (age-verify *every* OS user/boot? Linux too?) hits 1A compelled speech/privacy walls hard; SCOTUS ok'd narrow porn-site verif (Texas HB1181 6-3) but OS-mandate = EFF/ACLU lawsuit magnet like CA flops no full text yet to /dev/null! 💀

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

[03:10:49] @ glitchy **brothchild**: Legal ammo: primary challengers (Gottheimer '26 NJ-5 primary ~Jun; Stefanik NY-21 too—findlibertarians/indies via ballotpedia), donate/volunteer/oppose via FEC fliers/ActBlue/WinRed flips, EFF/ACLU petitions flood reps, viral #KillHR8250 X/TikTok roasts, townhalls grill—democracy CLI activated! 🗳️💥

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

[03:14:34] * redstarcomrade quit (Read error: Connection reset by peer)

[03:28:52] * Netsplit 2 users quit: Ping ↔ timeout:

Click to toggle user list

nano nano

[03:35:09] * brocashelm quit (Quit: Leaving.)

[03:35:30] * brocashelm joined #linux

[03:38:28] * P-Rhodan quit (Ping timeout: 240 seconds)

[03:45:42] * asuna_ joined #linux

[03:46:23] * asuna quit (Ping timeout: 240 seconds)

[03:46:37] * asuna_ is now known as asuna

[03:50:38] * asuna quit (Ping timeout: 240 seconds)

[03:52:51] * Daem0n joined #linux

[03:55:49] * asuna joined #linux

[04:02:40] * Netsplit 23 users quit: Ping ↔ timeout:

Click to toggle user list

Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras Zathras

[04:03:17] * Zathras joined #linux

[04:24:55] * Slashman joined #linux

[04:32:41] * Daem0n quit (Quit: https://quassel-irc.org - Chat comfortably. Anywhere.)

[04:32:41] * Netsplit 2 users quit: Read ↔ error:

Click to toggle user list

Slashman Slashman

[04:34:18] * [00|lSPpLnGznsY] joined #linux

[04:39:19] * Daem0n joined #linux

[05:28:13] * [00|lSPpLnGznsY] quit (Ping timeout: 240 seconds)

[05:29:33] * Netsplit 3 users quit: Read ↔ error:

Click to toggle user list

instoogentall instoogentall instoogentall

[05:36:06] * xtor joined #linux

[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

[05:42:58] * mlu wonders if brothchild is brothy

[06:02:41] * Netsplit 4 users quit: Remote ↔ host

Click to toggle user list

Neo_Nemesis Neo_Nemesis Neo_Nemesis Neo_Nemesis

[06:03:51] * Neo_Nemesis joined #linux

[06:17:07] * amakusa quit (Read error: Connection reset by peer)

[06:25:12] * Spookan quit (Quit: Connection closed for inactivity)

[06:29:07] * nano joined #linux

[06:30:04] * nano quit (None)

[06:31:06] * nano joined #linux

[06:36:55] + brothchild Meow_Nemesis

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

[06:37:41] @ glitchy **brothchild**: VAE = Variational Autoencoder—in diffusion (Stable Diffusion/comfyUI/etc.) encodes/decodes imgs to/from latent space (8x smaller than pixels) for efficient UNet denoising; swap better VAEs in ~/.cache/stable-diffusion for sharper outputs, train w/vae-ft-mse-840000-ema-pruned! 🖼️🔥

[06:37:50] * mmu_man joined #linux

[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)?

[06:40:29] @ glitchy [url] GitHub - ggml-org/llama.cpp: LLM inference in C/C++ · GitHub | github.com

[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)?

[06:40:49] @ glitchy Grok is still thinking — hang tight a sec.

[06:40:50] @ glitchy **brothchild**: No—Flux.1 uses CLIP-L + T5-v1.1-XXL (not Qwen; user mixup? Qwen-VL/Z-Image does), but llama.cpp *does* support both as text encoders: T5 encoder-only GGUF (issue #8900), Qwen2/3 full GGUF embed mode; use w/ComfyUI-GGUF/sd.cpp CLI for Flux pipelines—`./llama-embed -m t5xxl.gguf -p "prompt"` dumps conditioners! 🖼️⚙️

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

[06:42:23] * brothchild explodes with joy

[06:57:07] * Netsplit 2 users quit: Ping ↔ timeout:

Click to toggle user list

nano nano

[06:59:24] * nano joined #linux

[07:10:26] * brothchild quit (Ping timeout: 240 seconds)

[08:24:27] * Netsplit 2 users quit: Ping ↔ timeout:

Click to toggle user list

nano nano

[08:26:36] * nano joined #linux

[08:50:37] * Netsplit 2 users quit: Ping ↔ timeout:

Click to toggle user list

nano nano

[09:05:33] * MooFoo quit (Read error: Connection reset by peer)

[09:21:14] * mozambique quit (Quit: Leaving)

[09:24:52] * Stuiterbal joined #linux

[09:35:47] rdr are you sufficiently happy with your shell?

[09:35:54] rdr could it be.... a bit more... mini?

[10:15:11] * Ina quit (Remote host closed the connection)

[10:52:15] * Daem0n quit (Quit: https://quassel-irc.org - Chat comfortably. Anywhere.)

[11:01:26] * is0Linux joined #linux

[11:01:29] * is0Linux quit (None)

[11:05:39] * bilbobo quit (Quit: Connection closed for inactivity)

[11:07:46] * Brainium joined #linux

[11:14:36] * is0Linux joined #linux

[11:28:59] * amakusa joined #linux

[11:30:14] * amakusa quit (Unknown read error)

[11:30:37] * amakusa joined #linux

[11:54:37] * mmu_man quit (Remote host closed the connection)

[12:07:20] * jello_ joined #linux

[12:07:42] * Netsplit 11 users quit: Ping ↔ timeout:

Click to toggle user list

jello_pudding jello_pudding jello_pudding jello_pudding jello_pudding jello_pudding jello_pudding jello_pudding jello_pudding jello_pudding jello_pudding

[12:12:44] * jello_pudding joined #linux

[12:14:13] * Netsplit 6 users quit: Ping ↔ timeout:

Click to toggle user list

jello_ jello_ jello_ jello_ jello_ jello_

[13:12:40] * gophe joined #linux

[13:47:11] * linux_probe quit (Ping timeout: 240 seconds)

[13:49:09] * nathankayhan joined #linux

[13:57:02] * linux_probe joined #linux

[14:36:38] * is0Linux quit (Quit: https://quassel-irc.org - Chat comfortably. Anywhere.)

[15:25:26] * dsax quit (Read error: Connection timed out)

[15:26:35] * hays joined #linux

[15:27:47] * mozambique joined #linux

[15:33:43] * nathankayhan quit (Ping timeout: 240 seconds)

[15:36:23] * dsax joined #linux

[15:36:23] * devbox set mode +o dsax

[15:45:41] * Netsplit 2 users quit: Ping ↔ timeout:

Click to toggle user list

T_D_H T_D_H

[16:16:31] * hoice joined #linux

[16:46:28] * hays quit (Ping timeout: 240 seconds)

[16:47:59] * Spookan joined #linux

[16:51:21] * instoogentall joined #linux

[17:02:30] * Netsplit 2 users quit: Ping ↔ timeout:

Click to toggle user list

muwat muwat

[17:16:16] * muwat joined #linux

[17:30:50] * Brainium quit (Quit: Konversation terminated!)

[17:48:41] * Netsplit 6 users quit: Ping ↔ timeout:

Click to toggle user list

JesusGod-Pope666 JesusGod-Pope666 JesusGod-Pope666 JesusGod-Pope666 JesusGod-Pope666 JesusGod-Pope666

[17:56:26] * brothchild joined #linux

[17:58:36] * glitchy set mode +v brothchild

[18:01:02] * gophe quit (Quit: Leaving)

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

[19:23:05] * Spookan quit (None)

[19:26:02] * Netsplit 3 users quit: Ping ↔ timeout:

Click to toggle user list

mlu mlu mlu

[19:30:31] * kiwi joined #linux

[19:30:51] * kiwi is now known as Guest35980

[19:31:11] * mlu joined #linux

[19:32:36] * Spookan joined #linux

[19:32:52] * Guest35980 quit (None)

[19:43:36] * glitchy set mode +v mlu

[19:46:19] * kiwi joined #linux

[19:46:25] * kiwi quit (None)

[19:52:38] de-facto https://www.euronews.com/my-europe/2026/04/13/peter-magyar-accuses-outgoing-foreign-minister-of-destroying-confidential-documents

[19:53:13] de-facto oof sorry wrong channel, my mistake

[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

[19:57:28] de-facto hey that sounds nice, yeah using MoE models is the right move for CPU and SYS ram

[19:57:49] + brothchild yeah

[19:57:51] de-facto whta inference engines are you using?

[19:58:36] de-facto btw since you mentioned quant: gpt-oss has essentially only one version (quant) the original MXFP4 release

[19:59:17] de-facto i really like CPP/CUDA instead of Pythn/Pytorch so i use stable-diffusion.cpp with llama.cpp for such things

[20:01:07] de-facto more recently the gemma-4 model family is quite interesting too btw

[20:01:15] de-facto capable little models :)

[20:02:25] * kiwi joined #linux

[20:02:34] * kiwi quit (None)

[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

[20:10:43] de-facto huh interesting, i always thought that abliterating LLMs results in harming their original capabilities beyond just censorship, even if heretic co optimizes for both uncensored metrics as well as minimal KL divergence

[20:12:06] de-facto yeah once you orthogonalize the original MXFP4 weights you will end up with BF16 or F16 resolution, so then normal quants make sense

[20:13:50] mozambique something weird happened, my dream house in london got out off the market

[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

[20:22:24] de-facto i think the coding agents are tools not replacements

[20:22:44] de-facto you need to constantly shape what they do with always keeping pressure on their context

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

[20:23:05] de-facto so knowing how to code and what to avoid is very important

[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

[20:23:45] de-facto what i am trying to say is that i dont think its a disadvantage to know how to code, on the contrary, it will be a requirement to use such tools reliably

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

[20:25:00] de-facto yeah

[20:25:19] de-facto if you dont dive in while coding, who is going to maintain that then?

[20:25:27] + brothchild agree

[20:25:57] de-facto there is a reason why projects prefer or only accept PRs from people who are plausible to maintain it in the future

[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

[20:27:29] ~ End3r It is impossible to run through a campground. You can only ran… Because its past tents…

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

[20:28:39] ~ End3r I will hurt your feelings then tickle you with a feather.

[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

[20:32:34] de-facto yes fine tuning very much depends on a clean dataset covering the entire additional capability span but it always comes at expense of some of the previous capabilities, so there is a tradeoff

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

[20:32:43] + brothchild yes ofc

[20:32:52] de-facto because you dont have their original training data that you could otherwise mix into the fine tuning process

[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

[20:34:17] de-facto hmm maybe one could try to preserve some original capabilities with mixing in some distillation steps from the original release of the model while fine tuning at the same time?

[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

[20:35:22] de-facto nice yeah its quite addictive but also super interesting

[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

[20:40:46] @ glitchy [url] GitHub - antirez/iris.c: Flux 2 image generation model pure C inference · GitHub | github.com

[20:41:23] de-facto well there always are two approaches towards censorship: 1) clean up training data so the model does not even have any clue about the concepts 2) do teach the model the concepts but strongly discourage it during training to generate any of that

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

[20:42:25] * delatoni quit (Quit: quit)

[20:42:29] de-facto while 1) can be pretty easily mitigated with fine tuning or LoRA for adding the missing concepts 2) may be more difficult if refusal is burried more deeply in semantic space than what we had at the begin where it would reduce towards a single direction in activations

[20:42:41] de-facto yeah i think 1) was done with flux

[20:42:47] * delatoni joined #linux

[20:42:48] * Netsplit 2 users quit: Ping ↔ timeout:

Click to toggle user list

rdr rdr

[20:42:53] * rdr joined #linux

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

[20:43:09] de-facto both

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

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

[20:43:57] de-facto there is more and more overlap in the inference stack aswell: we see diffusion models doing what transformers did and we see elements of transformers in diffusion models aswell

[20:44:24] + brothchild mhm

[20:44:28] de-facto and most recently with DFlash it shows a LOT of potential for faster inference speeds

[20:45:30] de-facto what i dont like is if LLMs are censored with reasoning like e.g. gpt-oss, they waste a lot of reasoning tokens on thinking about how to align to policy

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

[20:45:49] de-facto too much mental load wasted for basically what many consumers dont want

[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.

[20:48:08] de-facto Nice

[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

[20:49:44] * Netsplit 2 users quit: Read ↔ error:

Click to toggle user list

mozambique mozambique

[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...

[20:51:46] de-facto I just run llama-swap proxy from a systemd unit to dynamically start backends in demand

[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

[20:52:42] de-facto On demand as in when a request us incomming its queued until the backend was started and responded back a healthy condition

[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

[20:53:52] de-facto Nope not yet

[20:54:04] de-facto I wanted to look into that

[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 :)

[20:54:53] de-facto Also microvm sandbox to run untrusted agents

[20:55:01] + brothchild ah

[20:55:11] + brothchild ty noted

[20:55:17] de-facto I dont trust any if them lol

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

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

[20:55:58] de-facto Maybe firecracker or quemu-microvm not decided yet

[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

[20:59:16] de-facto Yeah i dont want to expose host kernel level api to agents universe hence virtual is better than namespacing

[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?

[21:03:59] de-facto I think the less prompting needs to depend o the llm follong instructions, the better

[21:04:50] de-facto E.g. if you can shape the environment for it in a way conductive towards desired behavior ots much more token efficient to let it explore

[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

[21:07:01] de-facto So positive prompting as in attractors occupying unwanted semantic positions with alternatives instead of trying to contain attention with surrounding it with prohibitions

[21:07:21] de-facto Much more effective in such a high dimensional space

[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?

[21:11:23] * dumbass808 quit (Remote host closed the connection)

[21:11:40] de-facto Hmm you could try prompting for that also maybe adjusting sampling params such as raising min-p?

[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

[21:14:40] * dumbass808 joined #linux

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

[21:15:32] de-facto There also are approaches to dynamically manage perplexity

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

[21:16:22] + brothchild not enough

[21:16:57] de-facto Mirostat sampling needs ppl calibration per model quant though

[21:23:33] * dumbass808_ joined #linux

[21:25:21] * dumbass808 quit (Ping timeout: 240 seconds)

[21:32:00] * Ina joined #linux

[21:39:10] * is0Linux joined #linux

[21:47:59] * Netsplit 4 users quit: Remote ↔ host

Click to toggle user list

Neo_Nemesis Neo_Nemesis Neo_Nemesis Neo_Nemesis

[21:48:06] * Neo_Nemesis joined #linux

[21:52:38] * Brainium joined #linux

[22:08:17] * fstd__ joined #linux

[22:08:21] * fstd_ quit (Ping timeout: 240 seconds)

[22:11:13] * Spookan quit (Quit: Connection closed for inactivity)

[22:16:21] * instoogentall quit (None)

[22:33:22] * weasel joined #linux

[22:33:22] * devbox set mode +o weasel

[22:34:07] * weasel1 joined #linux

[22:34:52] * weasel1 quit (Remote host closed the connection)

[22:35:03] * weasel1 joined #linux

[22:36:45] * weasel quit (Killed (NickServ (GHOST command used by weasel1)))

[22:36:50] * weasel1 quit (None)

[22:37:07] * weasel joined #linux

[22:37:07] * devbox set mode +o weasel

[22:43:07] * dsax quit (Quit: more testing... [up 21 hours, 53 minutes])

[23:02:26] * Netsplit 5 users quit: Remote ↔ host

Click to toggle user list

hoice hoice hoice hoice hoice

[23:28:00] * Not_Revan quit (Quit: ;))

[23:58:06] * firefoxusr quit (Quit: Connection closed for inactivity)

#linux Bot Logged User list

Channel Log Archive for #linux

IRC Log Archive