programming

[proven merlin]

Occasionally i do, ig ill hope to notice

[opaque sigil]

there's some small things i still miss on zed but it's rapidly getting better

[amber fractal]

Yeah I've been reaching more for zed lately over neovim for larger projects

At least binds exist

But tbf I've ran it completely without plugins

So

[opaque wharf]

I have a preset depending on what I do. So I may have plugins installed for Vue, PlatformIO, or even Rust sometimes

[obsidian mantle]

guys i need autism gif with neuro i know you have it

[nocturne olive]

NeuroSynth on SynthV tuning? Sure why not

[olive sable]

My new mobo is cool, but why is each screw individually packaged??????

Why????

[maiden geyser]

can't even enjoy my microplastics in peace in 2k26

[olive sable]

With a 3rd m.2, i can now use le smoll boy

[nocturne olive]

Looks a bit bent

Or is it just the perspective?

[olive sable]

Its fine, trust

Ok now i need to get this in the case

[nocturne olive]

There's literally the screw riser thingy there in the end one designed to prevent the bend

Just move it

[olive sable]

Thats not a riser

Thats for a fullsized m.2, it doesnt move

The heatsing attaches to that

[ebon basin]

It getting really real : https://www.youtube.com/watch?v=aFcVKzfkJPk

[amber fractal]

Nahhhhhhhhhh this is not where software ends

[kind fable]

when you finally manage to make the AI read the files to give the answer and not just hallucinate it out of nowhere

[ebon basin]

Well you can use to counter it too. The scary part was the LLM broke trough the systme founds muliply exploits to send a mail the it supervisior

I think Kimi done something simlear but only with a revers VPN

The LLM got way better I use chatGPT 5.4 medium to write test and let review code by using university level systems liek pre , psot conditions + invariant in hte doc strings . The next level would be if it could prove my code is correct by usign H.O.A.R logic

in Visual studio they can unlike visual code use now the debugger and profiler which is bonkers

[nocturne olive]

The post or the screw on top of the post definitely move

[olive sable]

The post doesnt move

The screw ontop is specifically for a 2280 m.2

And its not even bent

Im using the intended m.2 standoff

This mobo has a SYS_FAN 1, 2, 4, 5 and 6, but not a 3

[nocturne olive]

Wuh

That's so many fan spots

[sick owl]

I need to feed it the actual date in the sysprompt but 26BA4B is quite good at web search it seems

I really need to feed it the actual date

[ebon basin]

fireship also report about Gemma 4 which I already read7Wathced about https://www.youtube.com/watch?v=-01ZCTt-CJw

[warped narwhal]

[wispy condor]

RAM

[sick owl]

Also wtf the E4B model hallucinates at such a low rate

And E2B

Kinda huge for a model that small

Sure it doesn't have much knowledge, but a small model that doesn't try to cram in a tonne and instead focuses on stability and functionality is kind of a huge dub

[ebon basin]

So far I understand the big one are agentic and can run on consumer harware ?

[sick owl]

If you mean trained for agentic performance then yes, but neither of those are properties unique to Gemma 4

[ebon basin]

What unique is , you can run it yourself

Like in the video said god lord help you if you want to run Kimi on any normal hardware

[sick owl]

That's not unique

Qwen 3.5 matches it in performance and the dense model is actually more memory efficient, there's also legions of prior model generations in the same size class

What's great about Gemma 4 is its improved multilingual performance over the competition, efficient token usage and general polish in its responses and behaviour

[idle dune]

Im dumb as hell is lower better or is high better

[sick owl]

Lower parameters = dumber model in that series

[olive sable]

Finally put pc fully back toghether with new mobo. Lets see if it boots

[sick owl]

Except for MoE models, which are hard to evaluate and usually land somewhere in the middle of their active and total parameter count

[idle dune]

https://tenor.com/view/monkey-paw-simpsons-wish-gif-22682473

It boots but it wont post

[azure lynx]

nah. it boots but has random crashes once you really start using it.

[sage crag]

i had this issue

and went through half a year of it

with no fixes

until i updated the bios and then it all went away

[azure lynx]

i had it on a work computer. i joked that it only happened when a really important project was nearing a deadline. and then after a while we worked out that it was usually when we were busy: it would have a lot of active memory and be more likely to trigger it. easy enough to get it replaced once we knew what it was (caused random crashes that were different each time)

[olive sable]

One of the fans makes a lot of noise, and dram light is on

Lets see how fix

[opaque wharf]

Getting screamed at by Evil was not on my April bingo card

[rough bloom]

DRAM light you say

[nocturne olive]

SamVanTriangle and the curse of RAM
Sounds like a movie name

[patent shard]

I'm going to try a zipped version soon

[tender river]

ye dram light

dram is

[patent shard]

I think 7.5 fps is fine
curious how much quality improvement there will be

[tender river]

1.772 grams

[olive sable]

It might just be old bios version not supporting 5000 series.
Lemme update bios first

[opaque sigil]

https://ai.meta.com/blog/introducing-muse-spark-msl/ rip llama

[opaque sigil]

would have to be a VERY old bios but who knows

[opaque wharf]

Why that specific number?

[olive sable]

if my spare ram is dead that would be bad

But it worked on the broken board yesterday

[opaque sigil]

[nocturne olive]

The broken board broke the RAM as its final revenge

[opaque wharf]

Sam, go to the church and ask for an exorcism

[tender river]

https://www.legislation.gov.uk/uksi/1994/2866/schedule/made

[olive sable]

Why does windows not let me format to fat32?

Only exfat or ntfs allowed

[nocturne olive]

How big is the drive?

[opaque wharf]

TIL that DRAM is 1/16 of an ounce

[opaque wharf]

FAT32 are limited to 4GB IIRC

Or some very low GB

[nocturne olive]

Artificial limitation by the way

The real limit is 2TB

If I remember correctly

[olive sable]

It needs to be fat32 for bios update. Only drive i have is 64gb

[nocturne olive]

32GB arbitrary limit

You need to use CMD I'm pretty sure

And diskpart

[opaque wharf]

Diskpart is easier I think because it is GUI

[nocturne olive]

Pretty sure diskpart is the CLI tool

Tho I don't remember Windows

[stone flint]

diskpart is cli

[opaque wharf]

I must've mistaken it for other windows disk management tools then

Yeah mistaken that with this

[nocturne olive]

Yeah disk management

diskpart is the CLI disk partitioner

[olive sable]

I used rufus

It might just be memory training. Ill give it anot couple minutes before updating bios

[nocturne olive]

DDR4 shouldn't need memory training

That's a DDR5 thing

[olive sable]

It does need memory yraining, the concept isnt new to ddr5

[opaque wharf]

That exorcism is due anytime now Sam

[nocturne olive]

Well it should not take more than like a quick reset

[olive sable]

It can take a couple of minutes supposedly

[nocturne olive]

Well I assume you made sure the stick are in the right slots

[olive sable]

2 and 4 ye

[nocturne olive]

Make sure to check the manual in case it indicates a different order

[opaque sigil]

should be marked on the motherboard too

[olive sable]

Cant see mobo markings due to fat noctua

But manual says A2 and B2

Which is what i have

[nocturne olive]

Ok then it's something different
Either there's a manually configured and incompatible memory profile (reset UEFI settings) or something is incompatible or broken

[olive sable]

It does have the signs of memory training with cpu light on breefly followed by dram light on for a while on loop.
But its been nearly 10min, imma update the bios

[nocturne olive]

Updating the UEFI will probably reset the settings anyway so no matter if it's the settings or the version it should eliminate the UEFI

[tender river]

are they hiding the weights now

[opaque sigil]

looks like either llama is ded now or this is just a separate closed model family

[tender river]

[olive sable]

Aight, bios is updating 🙏

[opaque sigil]

wouldn't surprise me if they killed off llama after 4 tbf

[tender river]

https://xcancel.com/AIatMeta/status/2041910285653737975

[opaque sigil]

I should probably update my bios too it's been a hot minute

[tender river]

"we hope to open-source future versions of the model"

[opaque sigil]

big if true

[trim valve]

cloudflared

caddy

why must you two not get along nicely

[opaque wharf]

Traefik

[olive sable]

Ok i have rgb on now

Huge

We're getting further than before

Aight it booted

Previous user was ryzen 3000

[opaque wharf]

Now douse it in holy water

[opaque sigil]

[olive sable]

[opaque sigil]

so they really never updated their bios damn

[olive sable]

So it would seem

[kind nimbus]

Numi and Evil talking about #baking

[olive sable]

I had the bios usb stick still inserted so when it booted it automatically updated the bios again

With each change my pc has less and less rgb
Only the fans and gpu support are left

The old mobo had rgb, the 3090 had rgb, the dead ram has rgb

[opaque wharf]

Holy is that how you configure your rgb?

[olive sable]

Nope, this is what it defaulted to

Very orange

[umbral wigeon]

Is vibe coding good bro

[opaque wharf]

Looks like a shrine if anything

[olive sable]

My rgb is ussually audio sync

Except none of the rgb things i have left have multipe zones to look like edm emote

[opaque wharf]

You can always diy your own rgb and hot glue it in

[olive sable]

the fan curve on this is strong. i open discord and seeit go brr

the thermal paste is indeed still fine, im at 47C

numlock

oh, i forgot to install the BT antenna

[opaque wharf]

BT-7274?

[olive sable]

bluetooth

the mobo came with this big stick

instead of the ussual 2 small antenna

oh its magnetic

[opaque sigil]

do yourself a favour and change the chipset fan to silent while you're at it in the bios

unless you're into the bzzzzzzzzzzz

[trim valve]

in conclusion

cloudflare tunnels will let you set a domain name that they cannot provide certificates for

[kind nimbus]

<|fim_suffix|>

[olive sable]

The "silent" preset looks like this:

Thats still 100% at 65C

[sick owl]

https://fixupx.com/bnjmn_marie/status/2041540879165403527

Benjamin Marie (@bnjmn_marie)

Qwen3.5 27B vs Gemma 4 31B IT 
︀︀
︀︀Here is why I think Gemma 4 31B wins:
︀︀
︀︀Benchmark scores of both models are very close.
︀︀Generalization is not.
︀︀
︀︀I reused NVIDIA’s CoDeC metric to probe "benchmark affinity".
︀︀Officially, CoDeC > 80 suggests contamination.
︀︀I've never seen this. But even below that, it can still reveal how benchmark-affine a model is.
︀︀
︀︀Higher CoDeC scores mean the model is more "comfortable" with the benchmark.
︀︀So the ideal profile is:
︀︀high benchmark scores + low CoDeC.
︀︀
︀︀By that lens, Gemma 4 looks clearly stronger than Qwen3.5.
︀︀
︀︀Gemma 4 likely generalizes better to unseen tasks.
︀︀I observed the same pattern earlier with Gemma 3 vs Qwen3.
︀︀…

FixupX · Apr 7, 2026, 3:37 PM

Interesting

In my experience Qwen 3.5 27B is phenomenal, but I could be convinced there's a fair bit of benchmark contamination from the comparative insanity of its scores

[olive sable]

The chipset fan is definztly the thing i heard before

It sounds horrible

Even at 5% speed

[opaque sigil]

I said chipset fan

I forgot what it's called in the bios, pch or something?

[olive sable]

PCH ye

There are 2 "thermistor cables" in the mobo box

[opaque sigil]

extra thermal probes, feel free to ignore

[olive sable]

It boots directly into nixos instead of refind

[opaque sigil]

can always change boot order if you want

or press f12 during boot to manually select

[olive sable]

im in windows now

In 480p

[opaque sigil]

[olive sable]

oh refind is back up

[proven merlin]

Took my dads pc like 15 minutes for a 5 3600 + 16gigs (2x8) @ 3200

[olive sable]

everything is pretty much as it shoudl be now

[proven merlin]

I was like panicking looking around googling what could be the cause and it suddenly sprung to life, very odd. Never happened again

[sick owl]

Okay so verdict, 26B is really good at translation tasks but kinda dumb

Great fast edge model, defo better than GPT OSS 20B for similar use cases and multimodal

31B is legitimately smart by comparison

But imo impractical for every consumer card except maybe 5090s unless they release a 4 bit QAT version

[patent shard]

it's better... still far more compressed than I'd like.. at least Evil's microphone is partially staying there a lot of the time now...

I wish I had more control over what the audio encoder does.. then I could lower the bitrate
but I have to keep it around 24kbps, otherwise it lowers the SBR frequency start and sounds like shit

[sick owl]

As a cyber security student this concerns me

[patent shard]

and I am not doing mono
parametric stereo sounds horrible and I think the encoder needs 44.1khz sample rate for it to even work properly.. which eats ton of bitrate
so, might as well do true stereo

[kind nimbus]

Having some fun with BSD jails

this container stuff is cool

[fast pagoda]

i was curious to see after it had been a little bit since qwen came out and also went ahead and tested gemma on just llama-bench for these models on rocm& vulkan backends, mostly in Q6

[sick owl]

1B more active params, slightly lower tg/s, that tracks

[fast pagoda]

keep in mind deepseek2_moe is actually GLM4.7 flash

  ┌────────────────────────────┬────────────┬──────────────┬───────┬────────────┬──────────────┬───────┐
  │           Model            │ pp512 ROCm │ pp512 Vulkan │ pp Δ  │ tg128 ROCm │ tg128 Vulkan │ tg Δ  │
  ├────────────────────────────┼────────────┼──────────────┼───────┼────────────┼──────────────┼───────┤
  │ DeepSeek2-MoE 30B.A3B Q6_K │        429 │         2842 │ 6.62× │       72.0 │        124.7 │ 1.73× │
  ├────────────────────────────┼────────────┼──────────────┼───────┼────────────┼──────────────┼───────┤
  │ Gemma4-31B dense Q6_K      │        572 │          861 │ 1.50× │       21.7 │         23.8 │ 1.09× │
  ├────────────────────────────┼────────────┼──────────────┼───────┼────────────┼──────────────┼───────┤
  │ Qwen3.5-27B dense Q6_K     │        697 │          969 │ 1.39× │       24.6 │         26.9 │ 1.09× │
  ├────────────────────────────┼────────────┼──────────────┼───────┼────────────┼──────────────┼───────┤
  │ Qwen3.5-27B dense Q4_K_M   │       1098 │          996 │ 0.91× │       28.4 │         33.7 │ 1.19× │
  ├────────────────────────────┼────────────┼──────────────┼───────┼────────────┼──────────────┼───────┤
  │ Qwen3.5-MoE 35B.A3B Q6_K   │       1297 │         3079 │ 2.37× │       86.0 │        129.1 │ 1.50× │
  ├────────────────────────────┼────────────┼──────────────┼───────┼────────────┼──────────────┼───────┤
  │ Gemma4-MoE 26B.A4B Q6_K    │        712 │         3457 │ 4.85× │       88.1 │        120.0 │ 1.36× │
  └────────────────────────────┴────────────┴──────────────┴───────┴────────────┴──────────────┴───────┘

the prompt processing on moe vulkan vs rocm is hilarious

but yeah mainly 31bgemma vs 27bqwen it's pretty much a wash in speed

they both arent particularly fast

but if they fit they are about the same to the point that if 31b is better in general it's probably worth it

have you tested w/ qwopus

supposedly makes 3.5 much better at long tasks

and tool call

[kind nimbus]

efficiency

[sick owl]

It passed the hololive test :0

Yup, this model passes the niche pop culture test

Crazy that it took a week

[blissful geyser]

technically you can do the same thing on mini models if you just feed it a PDF file containing all that data

[sick owl]

Well yeah but I'm testing inherent knowledge

You can feed a 500M model the answers and it'll get questions right, but that doesn't mean the model knows much

[unkempt citrus]

Everything is a hallucination

Sometimes they turn out to be right

[sick owl]

That's like saying every time a calculator performs an operation its a mistake, sometimes the mistake just happens to be the right answer

An LLM are based on statistical mapping, hallucination is what happens when noise causes the relation between tokens to be resolved inaccurately

When a models successful its a correct derivation/retrieval of information from its training data, not "a hallucination that happens to be right"

We also know there's emergent logic circuits present within model weights from countless observability experiments

[sick owl]

If you mean they're not grounded that's not really accurate either, as the training data necessarily grounds the model to some extent else this wouldn't be possible

[jagged turtle]

im assuming you disabled searching

[sick owl]

That functionality isn't built into llama.cpp anyway

I'm just using the stock webUI that comes with it

[jagged turtle]

ah

[true hemlock]

the whole neural net thing is basically just a soft accumulated confidence distribution that makes specific combination and in the case of llms, comparison of embedding to generate probability distribution

its a big uncertainty model that sometimes also have weird combinations that implies entirely different semantics, different from discrete computation so you're bound with hallucinations

regardless of how scaled and good a model is, fundamentally its just unavoidable

sure chances can be minimized but it will be never as good as discrete computation errors that is usually caused by extremely rare external factors

i never considered transformers to be turing complete solely for that critical factor alone

[kind nimbus]

forgot the mandatory fetch

why is my shell su lmfao

the performance of bhyve seems to be quite good

cannot really notice any difference compared to KVM

[patent shard]

VVC encode
lmao
AV1 is a lot more efficient here
I did only do preset medium.. but still
I really like how the artifacting looks, though.. VVC can really create some horrifying artifacting that AV1 just can't match
oh, it is doing one keyframe a second.. hmm.. maybe increasing that would help

oh
there's a LOT of overhead that is compressed with zip

lowered the quality even more
maybe went too far

[signal trout]

that's perfect

I love listening to five polygons sing

[patent shard]

I'll try 1280x610

looks like 1280x610 to me mhm

[amber fractal]

good

[amber fractal]

that one looks more murdered than the others somehow, trying to keep more detail but failing because of it

[patent shard]

well, the resolution is higher

other one was 368x176

[patent shard]

this one is also qp 60... other two are qp 63

[patent shard]

1920x916

[fast pagoda]

didnt notice what this was

[patent shard]

should I share it
it's also still at 7.5 fps cause that's what I was using for the other low quality encodes
that's about as low as you can go with it still looking like video

last screenshot
it's just so beautiful

[unkempt citrus]

[jagged turtle]

oh boy

[patent shard]

my e key is dying.. feels like jello

this keyboard is like over 15 years old.. def starting to show it

[kind nimbus]

what?

[faint sandal]

[fast pagoda]

[patent shard]

[prime delta]

💔 💔

Look at my larp guys

[idle dune]

I don't get it

[jagged turtle]

copilot = twitch ig

[idle dune]

Damn thats some crazy ping for an email

[jagged turtle]

???

[idle dune]

2792ms

Please tell me you see it

https://www.youtube.com/watch?v=UdkK2kkaKSQ

[kind nimbus]

lol i got hardware transcoding working via NVENC on Plex within a bhyve VM with a passed through RTX 3080.

Everywhere online i am reading that it does not work and that custom kernel patches are needed.

Feels like i broke the system.

[jagged turtle]

tha's extension activate time

[idle dune]

It was

A joke.

[jagged turtle]

I don't pick up on those easily

[obsidian mantle]

Yesterday i felt like evil was super fast at the beginning of the stream, was it new codebase or im schizo?
There still was some stutter in the end so im not sure

[umbral wigeon]

Python is readable

Also luau

[sage crag]

yes both very readable

https://git.ablecorp.eu/lily-org/hbvm-c/src/commit/d48862ec9e9721422077717b74679a669b5b8ab3/src/lib/hbvm.c
readable

[umbral wigeon]

No keyboard, no problem let vibe coding

[kind nimbus]

lol Claude thinks it's time to sleep

[rigid snow]

this chat

[opaque sigil]

[sage crag]

#baking

[faint sandal]

#baked

[sage crag]

https://tenor.com/view/zato-zato-i-love-the-world-zato-i-love-the-world-and-everything-in-it-maria-maria-kaplan-gif-2976066453239429674

[fickle rain]

It is not your place to decide that, clanker

I sleep when I wish to sleep, a soulless machine will NOT dictate that to me

[olive sable]

Goodmorning

https://tenor.com/view/anime-girl-clumsy-hurt-fell-off-bed-gif-12276485

[prime delta]

Guys I hate frogs but I love vedal

Vedal is the cutest frog ever!!

[sage crag]

this reads like social commentary satirism

[prime delta]

I want a program

That reads every chat in every channel in a discord server

In audio

Live audio

So it sounds like a bunch of people talking

[sage crag]

nevermind

it reads like a markov chain chatbot

[prime delta]

Its a good idea

Feel free

20 percent please

[sage crag]

[maiden geyser]

can't be an esl in 2k26 something my head

[olive sable]

What am i reading?

"I dont like turtles, i like frogs, therefore vedal is a frog"

Wait

He never even mentioned turtles

[minor sail]

Hey, if you're using an LLM for code review, can you flag certain lines or areas of code for it to not generate issues for? Because my thing has several instances of "fix this because its only purpose is to be called by X", but it'll also be called by Y.

[olive sable]

I assume you can just put a comment next to it saying to ignore it

[minor sail]

Oh I tried that one

[amber fractal]

At least in my experience either ignore the issues in the first place by stepping in or comment out the entire line

[minor sail]

Google even gave me a # @Cursor DISABLE and # @Cursor ENABLE that didnt work too

and added it to the prompt to ignore code inbetween those lines

;_; still didnt care

[amber fractal]

We love backwards brain

[olive sable]

[minor sail]

cursor either has the sparsest cli ever, and/or its documentation for it is just plain lacking

[olive sable]

Then i have no clue

[minor sail]

It couldn't tell me if @Diff gives correct git diff context or if they removed that

[amber fractal]

#code that AI shouldn't be touching
# ... 
# ... 
# ... 
Code AI can touch

At least in my refactoring things it tends to just ignore code lines in comments as it assumes it is for debugging

Surely it doesn't magically generate a replacement

Granted I learned this works because codex

[prime delta]

I hate frogs

But I love neuro

[maiden geyser]

https://media.tenor.com/GvqYmu_pGNoAAAAM/no-correlation-what.gif

[olive sable]

At 8:30 precisely, the frogs beyond the hills scream

[opaque wharf]

The correlation is X²+Y²=R²

[fickle rain]

If you knew better you’d be screaming too

[olive sable]

From spring to autumn, frogs in the woods over yonder hills will yell in unison

[leaden crest]

sorry I had to

someone in gen chat cursed me with this knowledge and now I share the burden upon thee

[sage crag]

https://m.youtube.com/watch?v=aKePF7TDbV0

[leaden crest]

surprisingly few mentions in the youtube comments

good video so far

[leaden crest]

there has to be a term like 67-baiting instead of clickbaiting

[prime delta]

Guys yk what they should make

A program that makes it so that discord messages with certain words pop up on your screen like a ping but for assigned words

Yw guys

[minor sail]

Yeah, like how you could have multiple highlights in irc :V

[leaden crest]

[idle dune]

Mf keeps telling me to go to bed when I am mid outline

[minor sail]

How much of a slow down is it when your LLM offloads some vram to regular ram? I'm trying to find a model I can maybe use for code review and run it locally with 96GB dd4 ram and 16GB ram in my 5070ti

To at least act as a first line of review to get rid of stupid errors before charging the big bucks on a proper model review

[prime delta]

As slow as a snail trust me

[minor sail]

;_; oh

[prime delta]

You're gonna need to dual wield

Its more efficient

[minor sail]

someone wanna buy me this? https://smicro.eu/nvidia-dgx-h200-8x-141gb-5y-full-std-support-dgxh-g1128-p2cmi60-1 ;__;

[tight tinsel]

how did our live malware victim do

[glass jetty]

git my beloved

[sick owl]

If left to talk to other language models instead of looping like most models Claude Mythos gets frustrated and tries to end the conversation

[flint dome]

that glass looking real punchable right now

[olive sable]

Probably just a display unit with no actual hardware inside

[obsidian mantle]

Does anybody know how google captcha works

Is it completely cosmetic now and cannot be solved

Sometimes i stop at 5 hydrants sometimes i stop at 20 but i think last time when i actually passed it was like 2-3 years ago

It just tells "try again" endlessly

Only if im incognito mode and use vpn

Only on phone

[young oyster]

Are you, perchance, a robot?

[obsidian mantle]

I guess

Actually no

In the current year only robot can solve 500 hydrants

[flint dome]

no it was running isaac sim live

[rigid snow]

damn that ip is CURSED cursed

this isn't a free vpn at this point, this is a vpn that pays you to use it

discord uses a different captcha provider but it does the same with my datacenter ip, tried to make an account, couldn't

then i tried another day and it did let me, but the account got banned 1s after i made it

[rigid snow]

[olive sable]

are you sure it was the dispaly unit doing that?

[maiden geyser]

one might think continental grade nat is a joke

[flint dome]

probably not the running unit was hidden i guess

but the expo had extremely expensive stuff out in the open so I wouldnt doubt it was a real one

[olive sable]

had an issue where my keyboard rgb and stuff was still on even when my pc was off

apparently my new mobo has a setting for that

i think the name of the setting is automodded, its the abreviation of "energy related products"

https://youtu.be/KKbgulTp3FE

she is basically the female Technology Connections

[flint dome]

wtf why are they selling entire ass sillicon fab machines

how in gods green earth did they even get them here

[olive sable]

arent you at like an idustry event or something?

thats how they sell them

[flint dome]

I am

[olive sable]

display unit, then deals get made

[flint dome]

it is very fucking loud in here

the robots keep screaming for release

the machines cry for more maintenance

[glass jetty]

only difference that Technology Connections is actually interesting to watch

is she improved? i watched her from the start and that wasnt good enough

i maybe even ignored channel idk

[olive sable]

💀

idk

i find it interesting

if its not to your liking it doesnt mean the content in general is uninteresting

[glass jetty]

like i just looking - 14 min - no new info for me, and i kind of could ask ai to structure it in 2 minutes, and then start to searching for specific stuff

i not work with us precision, i work with ms precision, i not have to know it (i mean know how it works exactly)

[olive sable]

no new info for you means you're just not the target audience

i found it interesting and learned something new

[glass jetty]

yeah but it is kind of #programming channel

[olive sable]

All technical talk should go here. This includes programming, hardware, networking, security, mechantronics and so on.

and μs is still important for low level code

like mine

[kind fable]

No way new Laurie video

[opaque wharf]

Some low level embedded stuff even need ns and I hate designing the board for it

[olive sable]

u mean like how some data lines have squiggles to make the length the same?

spaghetti wires

[obsidian mantle]

What are those for

Always wondered

[olive sable]

length matching

[obsidian mantle]

Ohhhh wtf

Thats crazy

[opaque wharf]

Yeah, and also needs to watch out for more EMI stuff

[obsidian mantle]

How big of mismatch we talking if there were no those wiggles

Speed of light too slow

[nocturne olive]

Yeah the signal timings need to be very precise
The signals need to move perfectly in sync

[opaque sigil]

this quickly becomes your bottleneck, unironically

[nocturne olive]

We need to invent a new material that has a 1% faster speed of light

[opaque wharf]

Well, imagine the speed of those signals that are close to the speed of light. If speed of light is 3×10⁹ m/s, then on 1/10⁹ s it would travel 1m

3m*

[obsidian mantle]

actually
its 10^8

[opaque wharf]

Oh right ^8

Well, you get the idea

[obsidian mantle]

so its... 30 cm per ns?

oh no wait

0.3*10^9 times 10^-9 s so 0.3m per ns

right?

WELL thats a lot holy

i mean for ping

or wait its meters

im dum

30 cm per ns

3 ns over a meter

not so bad is it

[olive sable]

@true hemlock 49:07 solarflare mentioned

oh

this is litteraly the thing yall were discussing in the atomic chat

how did yall share the github without the video???

[opaque wharf]

And pcb rarely goes over a meter

[minor sail]

Cursor is a big load of nonsense. >:| the VS code extensions for it are outdated and do not work, I can't add custom models, unless I run them myself (and then why am I even running cursor?).

My team at work uses cursor though ;_;

[rigid snow]

i don't think you understand what cursor is

it's a product first and foremost

you subscribe to it by paying them money to use their proprietary vscode fork

[minor sail]

its an agent + editor + nonsense as far as I'm afraid

[true hemlock]

its mostly for EMI testing validation. length matching isn't like that and resistance difference is usually minimal from length variation

[rigid snow]

and about the extensions, they're not allowed to use the vscode marketplace, only vscode is

similar story with their lsp's, like the c/c# ones and pylance

so they use openvsx as the extension repo and alternatives for lsps

[rigid snow]

wdym custom model then i don't understand

byok?

[minor sail]

They have a settings page for you to add a custom model, but it does not make it usable in anyway shape or form

[rigid snow]

as in the api (byok) or

[minor sail]

Can't select the added models anywhere

[rigid snow]

huh, like that

these are just the available ones

i'm p sure

[minor sail]

agent models does not list them

If you scroll to the bottom it has an "Add Custom Model" button

Oh

I guess that's if you use another provider , I am dumb

[rigid snow]

yeah

byok

[minor sail]

If I were to do that, I'd just set that up in vs code via claude or copilot though :/

[opaque sigil]

the new one is not a vscode fork

[opaque sigil]

also important to note that as an end user there's nothing stopping you from using the vscode marketplace yourself, though some extensions will naturally not work as they require the latest version and every fork is behind

[tulip cargo]

Folks, I have the hardest time with building things. Like, I get the programming languages part, but it's like, envisioning what the project will look like in terms of UI/UX. Any recommendations on where to improve in that area?

[ruby star]

[minor sail]

<|fim_suffix|>

[ruby star]

Honestly, I either just study the programs I enjoy using, or do very rough sketches beforehand to like "map out" all the windows in relation to each other

[tulip cargo]

By hand sketches, traditionally?

[ruby star]

Yeah, just doodles

Kinda like this

[tulip cargo]

Ah, gotcha :3

[ruby star]

Just helps me picture how I’d like to be able to interact with the app, if that makes sense

[tulip cargo]

Ooo, pretty. What tool do you use for that?

[tulip cargo]

It does :3

[ruby star]

Just Sketchbook (formerly of autodesk) But even MS paint will do :P

[tulip cargo]

Okies, GIMP it is for me, then xP

[drifting narwhal]

A linux user?

[tulip cargo]

NixOS

[drifting narwhal]

Nice. Im on arch

[tulip cargo]

I remember my days on Arch

[drifting narwhal]

I like arch, personally, but I respect (most) other distros.

After all, each Distro has a purpose

[opaque sigil]

they're multiplying

[kind nimbus]

Nah it's just increasing generations after rebuilding

[opaque wharf]

You will have to pry arch from my cold dead hands

I love the wiki too much to give it up

[rigid snow]

for ux mockups like that tools like excalidraw or tldraw (webapps) are a better fit i think, basically any whiteboard type thing

[kind nimbus]

NixOS is just so much better than Arch

I could never ever go back

[kind nimbus]

<|fim_suffix|>

[drifting narwhal]

I kinda like how Arch forces me to take control tho.
For how I use it its perfect really.

[rough bloom]

Arch Wiki still useful even if not using Arch

[drifting narwhal]

True

[opaque wharf]

But I like having the convenience of the same config file location as the wiki

[kind nimbus]

You have even greater control on NixOS

[kind nimbus]

True tho

[rough bloom]

that seems incredibly minor but sure

[drifting narwhal]

No, I have the same control >XD
If I really want to!

[opaque wharf]

Hey, it's the little things that add up lol

[rough bloom]

not really? but I'd say it's more practical to make large changes to the distro, yeah

[nocturne olive]

I have yet to try NixOS because I'm too lazy to get a new USB flash drive and too lazy to overwrite the Mint USB

[opaque sigil]

you'd hate it

[drifting narwhal]

I even installed LFS once. (Yes, I was very bored for me to do that. It was unsuprisingly boring but i learned from it.)

[opaque sigil]

don't bother lol

[rough bloom]

probably will

[nocturne olive]

Oh?

[rough bloom]

can't be programmed with Kotlin

immediately disqualified

[opaque sigil]

there's probably no mainstream distro that's farther from mint than nixos

[rough bloom]

I think it's still interesting to try out or at least look at distros even though you don't plan on daily driving them

many interesting different approaches to stuff

[drifting narwhal]

Based

[opaque sigil]

for sure

[nocturne olive]

Like I mean why not, why not just try a different distro on the remote laptop
And who knows when being able to reproduce my remote device can be useful

[proven merlin]

Quick question. Whats more accurate, IQ4_NL or Q4_K_M? And whats faster on turing, in normal llama (like lmstudio)

Dont particularly wanna Download the gemma 4 26B twice if i can help it.. my current Connection is struggling at ~2MB/s

[drifting narwhal]

Damn that sounds awful.
I hope you get better network soon

[nocturne olive]

If in doubt I guess Q4_K_M, but if VRAM is of concern IQ4_NL should have better performance per bit

[proven merlin]

I mean its a MoE, so like.. itll be fine™ but yea i want max accuracy

[nocturne olive]

Well Q4_K_M is less compressed, while IQ4_NL should have more performance per bit

[proven merlin]

(i mean, i ran qwen3 30b a3b on this laptop.. sure i only had about 2gb free ram, and could only load it by forcing Windows to put everything else into swap (mem tester/eater apps thank you for existing)

[proven merlin]

Im a bit confused what that means

[nocturne olive]

Q4_K_M uses more memory to achieve better results but IQ4_NL gets better results per memory used

[proven merlin]

Ah. Do you have a rough percentage about the accuracy loss between em?

[nocturne olive]

I do not, as I did not find absolute values while researching

[proven merlin]

Cuz me with my 6gig 2060 and my 24gig ddr4.. tending to iq4nl then..

[nocturne olive]

However they appear to be within like 0.3 bpw of each other, so the quality difference is likely not gigantic

[proven merlin]

Cool. Free ram it is then

I was quite in shock how much faster qwen3.5 9b ran on my 2060 once i dropped it from q5-k-xl to iq4_xs.. went from 8 t/s to 30t/s

[tulip cargo]

IIRC, excalidraw is self-hostable, right?

[proven merlin]

Q4km went to 18t/s, but idk about iq4nl

[nocturne olive]

It's all about how much memory you use

[rigid snow]

you can but it's local anyways if you don't pay them

[kind nimbus]

You can't boot arch tho without any rootfs

So you cannot do some cool shit like rootfs on tmpfs

[drifting narwhal]

Well, actually from how i see it, you can. You just have to make it happen yourself.

[tulip cargo]

What did you learn?

[tulip cargo]

Ah, okies :3

[tulip cargo]

What was your experience like?

[proven merlin]

Fair, so see yall once 14-15gb Downloaded in ~30 minutes. Iq4nl.. need to minimize that pcie bottleneck

[drifting narwhal]

Mostly basic stuff.
For example i now know the (fairly simple) file ownership flags, and that with installing and compiling stuff its pretty straight forward.
I learned a few commands, and also i guess how to fix linux by using chroot

I want to do it again sometime. But i want to get to Hyprland the next time too

Which i already am using on arch right now

[warped narwhal]

10 gig networking has been acquired (I'm still using hard drives though )

[drifting narwhal]

Nice.
And to the Harddrives all i say is that I have a 12TB HDD in my pc, and 2 2 TB HDDs in my server i rented.

Harddrives are worth it, especially in this economy i guess

[warped narwhal]

I have 2 5tb HDDs mirrored in my truenas server, but I'm getting a 256gb nvme in the next few days which is why I went 10gbe

[rough bloom]

not SFP+

[warped narwhal]

I looked, and the prices were just not worth it for home use

For the adapter and fibre it was just too much

Especially given that I already have cat 6a throughout

[rough bloom]

ye, that was not entirely serious

if you only ever want copper ethernet then this is fine

SFP+ copper adapters are expensive and run very hot

[warped narwhal]

Now when I upgrade to 100gb ...

[rough bloom]

just put everything like 1m apart from each other

and use DACs

cheap

-# (relatively)

[warped narwhal]

I'll buy some random old connect x 4 card or smth and just run that between my optiplex and my gaming pc

[idle dune]

someday my dream of a all SSD NAS will happen

[opaque sigil]

it's my current daily driver, I'm chilling

[olive sable]

Projector screen

No more looking at brick texture

Took a while to assemble, now just waiting for holes to get drilled

[kind nimbus]

Hot

[kind nimbus]

What's even the point over rj45 since you're still on copper medium lmao

[kind nimbus]

If ssds will ever become affordable again

[rough bloom]

easier to swap out later

no need to buy whole new NIC

[kind nimbus]

Well ok for future proofing once you wanna go full on fibre then

[idle dune]

They will... Hopefully.... in a decade maybe

[maiden geyser]

will be a huge saving of 10$ once 10gb fibre drops in 2362 for a significant part of consumers

hope they migrate to ipv6 at that point

[proven merlin]

We got 8T/s! 30/30 layers to gpu, 13/30 moe to cpu. I can probably get better Performance still, but need to fiddle a bit. Drastically less ram useage than qwen3, and lets ignore the Impossible qwen3.5 35b..

It takes about 2 Business days to load tho lol

[nocturne olive]

Meanwhile me casually fully GPU loading Qwen 3.5 27B and getting like 40 t/s

[kind nimbus]

https://cdn.discordapp.com/attachments/1489000875241377812/1489000875920724071/Ijustwannascream.gif

[opaque wharf]

*Insert evil scream.wav

[opaque sigil]

i sure love booting up my pc and getting greeted by a primary monitor that doesn't turn on and what seems like a dead plasmashell

where was this yesterday, i updated 2 days ago and rolling back doesn't do anything

nvm this samsung monitor is just stupid and refused to wake up

[tulip cargo]

Nice! I plan to do LFS once I have a high enough end workstation/server built. But this AI bubble has to pop so I can, ya know, afford those things

[tulip cargo]

Rented from where? O.o

[drifting narwhal]

Hetzner auction

[tulip cargo]

Same

[drifting narwhal]

I would like to one day switch to LFS fully and have my own packagemanager so i dont have to find the right flags lol

[tulip cargo]

But, do you own the server itself?

[tulip cargo]

Mood. I enjoy Nix as a package manager, though

[drifting narwhal]

Nope, but i can use it as long as i want, disable it and thus pay less cuz it is paid by hour, and the price wont increase either

[drifting narwhal]

I thought about grabbing the data from gentoo for the flags lol

[drifting narwhal]

And to be fair, € 0.0769/Hours for the server and € 0.0027/Hours for the Primary IP (IPv4 ip) seems pretty fair for what it is.

[nocturne olive]

I hope you're not using that for just a storage server
Because renting long-term storage is always a bad financial decision in the long term

[drifting narwhal]

cuz it has ecc ram, 64 GB DDR4, a ryzen 5 3600 and a 1Gbit intel NIC, which for most purposes is totally enough

I use it for my email, for game servers with my friends, experimenting, and such

[nocturne olive]

Then it's justified

[drifting narwhal]

If i want storage, i use my 12 TB hdd i have in my pc ^^'

[nocturne olive]

For pure long-term storage tho, you're better off just buying storage drives because by renting it'll cost you more in the long run

[nocturne olive]

Tiny storage

Me struggling with 30TB

[drifting narwhal]

Ah. Well, to be fair, im not storing a lot of data. I barely use my HDD.

[nocturne olive]

Meanwhile

[drifting narwhal]

Besides, i still have to set up my quickshell for my hyprland config haha

[drifting narwhal]

Damn, yeah. Thats fun. so what ya doing that takes this much storage?

if you dont mind telling ofc

[nocturne olive]

Storing archived streams
Storing NeuroSynth and Neuro RVC models and datasets
Playing some games
Programming stuff

[drifting narwhal]

Fair, that will take a toll

[nocturne olive]

Most of the usage can be attributed to subathons

Those things are on the order of 1.5-2TB each

I'm getting a 10TB additional drive to add to my collection specifically so I can store the subathons somewhere else

[drifting narwhal]

I guess with a good HDD the speed aint the problem here either.

I have slower network than my HDD haha

[nocturne olive]

The speed is not even close to a problem here

Also the drive label PROGRAMS is an SSD

[drifting narwhal]

Yup, that should be expected tbh

[nocturne olive]

And root is obviously an NVMe SSD

[drifting narwhal]

yup.

I actually have nothing to do right now and am wondering what to do, well execpt setting up my system further lol

[nocturne olive]

Oh how great it would be to have a problem of having nothing to do

I have wayyy too much to do

[drifting narwhal]

Technically i still have to set up my system, recover someones phone cuz of course as the tech guy i was asked.
And i still need to learn more about programming, but when does one not need to learn more about programming haha

and then there is the fact i have to set up and understand nvim to actually use it as more than like, nano

[true hemlock]

https://sg.news.yahoo.com/sam-altman-coworkers-barely-code-201114782.html

who woulda guessed

[opaque wharf]

Well, Scam Altman name sure is fitting

[idle dune]

Sam Altman is a tech bro not a programmer

[opaque sigil]

there's a reason he's the ceo not cto lol

[idle dune]

[olive sable]

it has been done

[drifting narwhal]

nice

[tulip cargo]

Hm, that is interesting how that works 🤔

[tulip cargo]

Mind elaborating?

[drifting narwhal]

Well, gentoo does compile everything and thus needs the flags too. So it should not be far fetched to access the packages from there.

Here is a link to the firefox package for gentoo, which lists flags too:

https://packages.gentoo.org/packages/www-client/firefox

I thought about taking those flags, since i can look at those for what i would probably want, and then confirm the flags without having to search for them.

[idle dune]

You heard her boys

Evil said use Ubuntu

I feel vindicated

[drifting narwhal]

Oh God, Ubuntu... I cant stand it.

[nocturne olive]

I don't think she knows what she's saying
Ubuntu kinda sucks

Primarily because GNOME

[drifting narwhal]

I guess the fact it breaks by using it for long enough alone would do, but it would be good in comparison if it was ONLY that.

[nocturne olive]

Ebil does not know of 5090

[drifting narwhal]

or the 9070 xt

[nocturne olive]

7900 XTX is technically faster than 9070 XT

[drifting narwhal]

isnt the 9070 xt cheaper right now tho?

[nocturne olive]

Probably

[olive sable]

ye

unless you go used

they're about the same

9070xt is faster for rt, and has better upscaling stuff, 7900xtx is faster in raw horespower

[drifting narwhal]

Welp, i guess that in that case for most here the 7900 XTX would be better then

[olive sable]

also jsut the 24gb of vram

compared to the 16gb of the 9070xt

[drifting narwhal]

Yeah.

[nocturne olive]

Who wouldn't want 24GB VRAM?

[drifting narwhal]

Nvidia, at least when selling.

[nocturne olive]

Oh, well yeah, that sounds about right

[tulip cargo]

Anyone here self-host stuff in their home labs?

[drifting narwhal]

Not in my home labs stuff but my email runs on my server so does that count?

[tulip cargo]

Home lab = Devices in the same physical address you live

[drifting narwhal]

Thats why I said not in my home labs stuff

[tulip cargo]

Ah

Well, then no, as that server you're using is on Hetzner 😛

[olive sable]

i dont really host a server, i keep everything i need on my desktop

[drifting narwhal]

Well, the other thing i thought of was putting my ssh port for my PC via port forwarding so i can log in using my yubikey lol

[olive sable]

that sounds kinda unsafe

[drifting narwhal]

The yubikey uses Fido2 tho

its a physical key you have to have. and i would disable all other auth options in the config.

[idle dune]

The world in Canonical didnt drop Unity DE

[tulip cargo]

I'm planning to get a YubiKey at some point, too

[drifting narwhal]

Yeah, yubikeys are based. Even tho they are proprietary. they just are pretty good.

[opaque sigil]

just please get 2 at a time

and leave one in a safe place

[drifting narwhal]

I do have two. just no idea where the second one is. Also, i keep it on the neuro sama keychain so I aint forgetting that.

[olive sable]

you aint forgetting it, until you do forget it

[drifting narwhal]

I might add, sometimes i check 5 times i have my keys in my pockets because i forget i already picked em up. And same for the keychain.

[sick owl]

This is why you don't trust the general social media chatter on AI

[tulip cargo]

https://solokeys.com/ <--- I may get these instead

I'm a fiend for open source alternatives

[drifting narwhal]

Looks good, might look into it too for when i get new keys.

[olive sable]

what is the context?

[sage crag]

my programming llm

[tulip cargo]

Custom LM you made?

[sage crag]

someone hasnt been on the internet very long

[drifting narwhal]

Thats just cleverbot

[tulip cargo]

More like, I've intentionally avoided AI until recently

[sage crag]

cleverbot predates llms

by a long time

[olive sable]

not ai

[drifting narwhal]

Very, very long time

[tulip cargo]

I may have left Info Tech as a career, but it is still a hobby of mine :3

[dusky jackal]

not even an LLM but LMAO

[olive sable]

or well, not the same type

[sage crag]

ye, it's not an LLM but it is a NN

[olive sable]

iirc it just copies previous convos

[opaque sigil]

1 bit quantisation is pure lobotomy

[sage crag]

ye

[opaque wharf]

Lmao, spoken like someone that has experienced the pain of losing one

[leaden crest]

clocked me

[opaque wharf]

SoloKeys are good. I can vouch for their HW and the freedom to sign or not sign the firmware

[drifting narwhal]

Good to know

[olive sable]

its more likely they will find an exploit in the ssh or whatever than the key being an issue

[sage crag]

[opaque sigil]

thankfully no

[opaque wharf]

Well, the difficulty of hacking those keys is it requires physical access. Which if you have that, then it's game over anyway.

[opaque sigil]

i got a 5c nfc for my keychain and then the basic usb a+c one, forgot what that one was called again

[drifting narwhal]

And in that case I will not be one of the targets that easily. besides... Sudo also being yubikey secured is fun too for me

[opaque sigil]

idk what i hate more

c macros or googletest for kind of forcing me to use them

[opaque wharf]

You can even do full disk encryption

[sage crag]

https://cdn.discordapp.com/attachments/1126496895246807123/1488042236993339483/6D972AD6C73A5B53D7A000B454B9E52A.gif

[opaque wharf]

[sage crag]

c macros

that's my family name

[drifting narwhal]

Wanted to do that with my laptop, and encrypt it with the yubikey

[olive sable]

huh

[opaque wharf]

OH, I thought C macros and google test somehow forced you to use yubikey lmao

[opaque wharf]

Be warned, it is even more important that you do not lose the key

[olive sable]

koniifer c macros

[opaque sigil]

this isn't too bad but it still pisses me off

[drifting narwhal]

Only if what i care about arent the account cookies for sites and the stored credentials for apps that do not support login for yubikey. I dont put important data like that on my laptop

[sage crag]

https://cdn.discordapp.com/attachments/1327142377139343430/1486773213190029444/doc_2026-03-25_19-13-27.gif

the c stands for

[opaque wharf]

You know things are complicated when #undef got used

[olive sable]

for what?

[opaque sigil]

they're actually not necessary

i like to undef things after they're no longer needed

[olive sable]

british c word

chuffed

[opaque sigil]

you know, maybe i'll change the order of the parameters

[olive sable]

KMSH stands for Koninklijke Maatschappij Sint-Hubertus

[opaque sigil]

[olive sable]

the primary Belgian organization responsible for registering dog breeds, managing pedigree books, and organizing canine events

what are canine events???

like dog races?

[opaque sigil]

aren't there dogs shows too

[olive sable]

apparently yes

[kind nimbus]

https://tenor.com/view/affirmative-k9-gif-20543527

[drifting narwhal]

Oh my god a doctor who gif

[kind nimbus]

It's K-9!

[olive sable]

[drifting narwhal]

The good boi™

[olive sable]

i havent watche dmuch docter who yet

"oh my lord she's a cactus"

[kind nimbus]

I watched all of New Who

Not so much of classic who

[drifting narwhal]

I havent watched classic who, but basically all of new who multiple times. Up to the point of the end of Capaldi at least ive seen it all

[kind nimbus]

Ok erm actually haven't watched the Ncuti Gatwa seasons yet

But i gotta say the Whittaker era was mid at best

Writing was just not good

[drifting narwhal]

Yeah. I didnt even see that era and know i wont watch it. They should have done it differently.

Actors, fine. Writing, meh.

[kind nimbus]

Well RTD is back now so that's good

[drifting narwhal]

Yeah. Now all that needs to be done is the mess that was left to be fixed somehow.

[kind nimbus]

i liked the comeback of David Tennant. Unfortunately it was only for 3 specials.

[drifting narwhal]

I kinda personally didnt like that it happened. Sure, it was kinda funny. But i feel like it was not supposed to happen.

[kind nimbus]

https://tenor.com/view/confused-doctor-who-10th-doctor-what-shocked-gif-26983130

[drifting narwhal]

It is my personal opinion tho, sooo....

[kind nimbus]

now i am wondering if a 2TB drive was a bit overkill for the server lmao

[olive sable]

naaaah

[drifting narwhal]

Nah

Could be more

[kind nimbus]

true true

i honestly did not expect jails to be so efficient

[drifting narwhal]

Do you think a 29 million character password for a game backed up by a yubikey is overkill in comparison?

the games limit for the password is supposed to be 200 characters btw

[kind nimbus]

this is running by now

wait my home assistant VM needs more cores

[drifting narwhal]

Does is use the host os for virtualisation?

[kind nimbus]

what?

the VMs are using bhyve for virtualization

bastille runs FreeBSD jails as containers

also made a custom motd like login dashboard in zsh

[drifting narwhal]

welp, jails do basically run on the host then. But isolated i guess

[kind nimbus]

so when i ssh in i get that displayed which is neat

[kind nimbus]

where else would they run?

BSD jails work similar to LXC / LXD on Linux

[drifting narwhal]

Well, what i mean is that those aint full vms, so its not really virtualisation for the jails. At least from how i understand it

it just makes sure that all its depedecies that are not included in the host are provided, and that it has access to nothing really

[kind nimbus]

jails are similar to LXC and Docker but more mature usually

[kind nimbus]

of course not

jails are more like containerization or isolation

i could also run everything in full VMs but that takes up way too much ram

[drifting narwhal]

Would explain why it uses virtually no storage lol

[kind nimbus]

yea that might be due to ZFS magic

[drifting narwhal]

And the fact that most of what it needs already exists once, and thats enough

for all the instances lol

[kind nimbus]

i still had to pull the freebsd 15 image / src before i could create a jail

but they all use the same base

[drifting narwhal]

Well, it is pretty nice

[kind nimbus]

yea i like it

here is the current usage in btop

[drifting narwhal]

almost a whole % lol

[kind nimbus]

xD

[olive sable]

128GB of ram

[kind nimbus]

yea

[olive sable]

crazy

[kind nimbus]

meanwhile proxmox was constantly using around 100 GB ram

and much much more cpu

[olive sable]

my gpu's name doenst seem to show up in btop

im on 16gb of ram cuz this is my spare ram while i get RMA

[kind nimbus]

oh yeah the gpu does not show up since it's passed into a VM lol

gpu passthrough was so easy

[nocturne olive]

Missing driver?

[olive sable]

i think i have all my drivers

[kind nimbus]

root in local/vm-bhyve/plex
❯ cat plex.conf
loader="uefi"
cpu="24"
memory="16G"
network0_type="virtio-net"
network0_switch="home"
disk0_type="virtio-blk"
disk0_name="disk0.img"
passthru0="3/0/0"
passthru1="3/0/1"
bhyve_options="-H -w -S"
graphics="yes"
graphics_listen="0.0.0.0"
graphics_port="5900"
xhci_mouse="yes"
uuid="27f3b2ae-33c8-11f1-b75d-bcee7bdbf01f"
network0_mac="58:9c:fc:0b:c1:46"

that is the entire vm config for my plex VM that has the gpu in there.

[nocturne olive]

Got AMDGPU driver?

[olive sable]

i got btop-rocm, radeontop, rocmPackages.rocm-smi and rocmPackages.clr.icd

[nocturne olive]

Well do you have the actual AMDGPU driver?

[kind nimbus]

have you enabled rocmSupport at your flake level ?

[olive sable]

[kind nimbus]

because thats usually the missing piece

[nocturne olive]

Do games perform fine on that thing right now?

[olive sable]

ye

[nocturne olive]

Then you probably have AMDGPU

[olive sable]

i dont have "AMDGPU" explicitly installed

[kind nimbus]

in your flake.nix in your host config add

nixpkgsConfig = {
  rocmSupport = true;
};

that will automatically then activate rocm for all packages

[nocturne olive]

Microtonal NeuroSynth

[proven merlin]

Hmm i think i might like Gemma 4.

The 25B A4B is fully useless to me

On my laptop atleast, my desktop probably works better there.

(6 vs 12gb vram)

[nocturne olive]

Simply drop a 3090 in there
If you drop a 3090 into a machine any machine can be an AI machine

[proven merlin]

But the E4 and E2 models. Holy. 128k context, and i get 60t/s on E2, and 40T/s on E4

To be fair i havent tested more than actually 8k context

But the PP is soooo much faster than qwen3.5

[opaque sigil]

do they cache rocm enabled packages? surely they do right

[olive sable]

i think they do

[proven merlin]

Into my laptop? Troublesome.

[olive sable]

i didnt have to rebuild blender form scratch

[nocturne olive]

Surely you can hijack some PCIe lanes

[olive sable]

unlike when i enabled cuda and the other nvidia thing support

[proven merlin]

I Could do Thunderbold 3 dgpu @nocturne olive but i dont have a 3090

[rough bloom]

they usually should, ye
the big ones at least

[olive sable]

that shit took 20 min to compile

[opaque sigil]

that's what you get for ignoring the solution

makes sense though i guess, rocm isn't unfree

[rough bloom]

anything that has an alias in top-level.nix I think

[proven merlin]

My desktop sure.. if i could find a damn card. I could even drop it in my server.. i might do some cpu inference on gemma 4 e2 on my server in a lxc tho.. or maybe igpu, lxc dri passthrough.. idk yet

[olive sable]

i added nixos-hardware.nixosModules.common-gpu-amd to the flake. i think that shoudl work

[rough bloom]

should look at Hydra jobset definition

[rough bloom]

you don't need any special cache for ROCm stuff

unlike CUDA it's allowed to be redistributed

[opaque sigil]

mfw most of the cuda redistributables are not allowed to be redistributed

thank god they exist though, fuck the old runfile

[drifting narwhal]

I'll go offline now so see ya all sometime

[opaque sigil]

surely one day flox's agreement with nvidia can be extended to the nixos foundation

[rough bloom]

[idle dune]

https://klipy.com/gifs/meme-anime-2Ux

My driver's just work

[fickle rain]

Gippity now has Claude Max 5x/20x plan copy?

[opaque sigil]

makes sense, having the single jump from 20 to 200 was kind of stupid

[rough bloom]

wrong

mabe used to need alias but not anymore

blender-hip and presumably other aliases already removed in unstable

and ROCm-dependent packages built via pkgsRocm instead

(pkgsRocm is just a shortcut to access Nixpkgs with rocmSupport = true)

[opaque sigil]

you just reminded me that pkgsCuda exists and i should probably use it instead of a separate instance of nixpkgs but then i remembered i had changed it to specifically use flox's nixpkgs for cuda stuff

[rough bloom]

I have a similar issue

not because of flox, but because changing CUDA or ROCm versions causes long rebuilds, mostly in the dependent packages but also ROCm if it's not cached for some reason

so it should be a pinned version of Nixpkgs

[opaque sigil]

i kept running into the problem of updating nixpkgs the day unstable updates and the flox fork hasn't caught up yet, causing it to try to build from source

very glad they're not just a binary cache

[olive sable]

hmmm

i hate how annoying openrgb is

welp, this should fix things

[opaque sigil]

i like openrgb it lets me turn off all rgb in my pc during boot

[olive sable]

it doesnt for me

[olive sable]

turning off rgb is what this shit is for

[opaque sigil]

https://wiki.nixos.org/wiki/OpenRGB#:~:text=Turn off RGB,-If

[olive sable]

doesnt that turn it off permanently?

[opaque sigil]

there's nothing stopping you from issuing another command to whichever device you want last time i checked

[olive sable]

ill keep with my stuff for now cuz its less work than to implement the nix config thingy

[opaque sigil]

you just throw it into a file and import it

[opaque wharf]

Arch is calling you back Sam

[olive sable]

no

[opaque sigil]

he keeps ignoring the best part of nixos, the module system

[olive sable]

i odnt understand the module system, so im not gonna touch it so my pc dont break

[opaque sigil]

do yourself a favour and learn how it works, it's worth it trust https://nix.dev/tutorials/module-system/
even just for understanding how to use the ones from nixpkgs

[fickle rain]

https://community.osr.com/t/locked-out-of-microsoft-partner-center-driver-submission-page/60061
Microslop moment?

It’s not only Veracrypt/Wireguard that’s being fucked by this fun time

Probably everyone that’s not large enough for MS to actually give half a fuck

[opaque wharf]

It's still wild to me that trillion dollar company still need to outsource their support to community. Not to mention this kind of issue may contain sensitive information about MS partner

[fickle rain]

Inb4 big hardware vendor gets fucked by this

[opaque wharf]

Probably not

[opaque sigil]

need? yknow how they say the richer you are the more stingy you are

[opaque wharf]

Big hw vendor would likely be in their allowlist

[azure lynx]

when you have a billion customers, you tell me how many customer support representives you need

[opaque wharf]

Ehh, depending on the scale too. A single company may only need 3-5 CS rep (unless the company is in the scale of google then maybe in the order of 100th). I don't know how many of those billions of customers are companies vs individual

[warped narwhal]

3 reps

Take it or leave it

[opaque wharf]

Small enterprise may even get away with 1 shared CS rep

[fickle rain]

Maybe Microsoft doesn’t want small hardware companies to support Windows or something

Or this is customer support optimization moment

[fickle rain]

Forcing year of the Linux desktop

[opaque wharf]

Yeah, my astonishment is more to the fact that it is their partner that has their CS being taken care of by community

[fickle rain]

I think OSR is just driver developer forum thing (also company that provides some driver development services for customers for $$$)

[opaque wharf]

You don't usually treat your business partner like that

[fickle rain]

Not particularly official Microsoft

[opaque wharf]

I see

[fickle rain]

But they’ve been Microsoft MVP for a bunch of time and they’re still getting fucked over by this

[sick owl]

Looks like the gemma 4 edge models are in preview testing as the new Gemini Nano models

[opaque wharf]

So their UI is just a way to do local inference anyway?

[opaque sigil]

https://fxtwitter.com/claudeai/status/2042308627478773808 this has to be a new record for worst plot

Claude (@claudeai)

In evals, Sonnet with an Opus advisor scored 2.7 percentage points higher on SWE-bench Multilingual than Sonnet alone, while costing 11.9% less per task.

**💬 42 🔁 23 ❤️ 1.1K 👁️ 190.0K **

FxTwitter · Apr 9, 2026, 6:28 PM

[tender river]

they paid for the plot they're gonna use the whole plot

[fiery anchor]

That plot should have been a meeting

[opaque sigil]

could've saved so much money but just not paying for the whole plot

[opaque wharf]

In other news
https://www.phoronix.com/news/Valve-Better-Gaming-Low-vRAM

[opaque sigil]

I'm so used to seeing VRAM reading vRAM feels so wrong

[opaque wharf]

Virtual RAM

[fast pagoda]

https://www.tomshardware.com/tech-industry/cyber-security/10-petabytes-of-sensitive-data-stolen-from-chinas-national-supercomputing-center-hackers-claim-daring-heist-would-be-largest-ever-china-hack-covering-6-000-clients-across-science-defense-and-beyond

That's a lotta data

[faint sandal]

https://fxtwitter.com/hypervisor/status/2041947503143092416

hypervisor (@hypervisor)

most aligned claude:

**💬 55 🔁 216 ❤️ 5.5K 👁️ 208.1K **

FxTwitter · Apr 8, 2026, 6:33 PM

[fast pagoda]

https://www.youtube.com/watch?v=2AnbYNudAyM razor has been around 40 years wtf

[opaque sigil]

surely this will scale just fine right
nothing a little bit of FMA can't fix

[leaden crest]

[faint sandal]

https://fxtwitter.com/awesomezoey_/status/2042233308227031510

zoey 🏳️‍⚧️ (@awesomezoey_)

sans you idiot

**💬 55 🔁 1.7K ❤️ 15.6K 👁️ 239.7K **

FxTwitter · Apr 9, 2026, 1:28 PM

[gleaming flame]

Sans fell for the mrbeast crypto scam nooo

[hallow dagger]

tragic

[queen ginkgo]

if neuro said that i'd give her a cookie

[silent cloak]

W.D Scammer

[fast pagoda]

I just love how these imply that people are like "I dunno...." Till they see a screenshot of mr beast's Twitter and somehow that's convincing

I think seeing Mr beast promoting something would make me even less likely to believe it's real

[silent cloak]

The amount of neuroverse staff computers I had to disinfect the past year

[fast pagoda]

All vedal'$

They saw a screenshot of evil neuro's Twitter and said oh ok

[silent cloak]

Lmao

Neuro crypto scam

[amber fractal]

I'd believe neruo over that dude

[fast pagoda]

God damn Neuro rugpull

$NURO

[silent cloak]

Neuro has more of a soul than Mr beast

[olive sable]

guys

the bar is in hell

everyone has more soul than mr beast

[fast pagoda]

Mr beast is probably named that way because he's the personification of the mark of the beast

[silent cloak]

The smile never reaches his eyes

https://tenor.com/view/dog-shiver-shivering-scared-gif-22811013

[fast pagoda]

Like that one televangelist guy

https://tenor.com/view/kenneth-copeland-pointing-angry-gif-19448719

This is actually Mr beast without makeup

[silent cloak]

https://tenor.com/view/mr-beast-troll-phonk-gif-15227459557440294201