programming

[true hemlock]

or just go to my route and pay $14 for 3080Ti 12gb

[patent shard]

it's volta, ehhhh
a little on the older side

I don't think it has full support for Blender because it's so old

[mighty thorn]

I thought blender was pretty old no?

[patent shard]

ah, right
I meant the newest version of Blender

guess you can just use an older version

[mighty thorn]

Ohh

Yeah that makes sense

[true hemlock]

it is still supported in blender.

hell you can mix like titan v with other 30 series and multigpu render like that

[patent shard]

yeah, but there's something about it that isn't... I don't remember what

[true hemlock]

driver?

[patent shard]

Blender has made their requirements much stricter
nah, cuda version

maybe it's not Blender specifically
sheepit now requires 7.2+ for Blender 4.5+
but I don't remember why.. could be something not related to Blender use itself
regardless, Blender has made their requirements a lot stricter... so, gl with that working much longer

this might also be for optix output

[true hemlock]

like the cuda software?

or physically

[patent shard]

physically
but I think this is just optix requirements, which you won't gain much benefit on cards without rt accelerators
sooooo, ignore what I said.. it only applies to sheepit 😅

it says Blender 4.5 is 400 series and newer and Blender 5.0 is 900 series and newer

[true hemlock]

welp yeah volta does not have rt

[patent shard]

it doesn't need rt

[true hemlock]

but sure as hell it has fp64

my favorite

[patent shard]

pretty sure volta cards actually score very well

[true hemlock]

they're basically turing with ampere level of clock cycles + fp64 & hbm

[patent shard]

well, okay
it's not great

[true hemlock]

nah

it is

you can make the v100 outperform 5060ti fairly easily

[patent shard]

it's in-between pascal and turing
turing is cuda 7.5, volta is cuda 7.0

well, depends on which volta card.. they vary

[true hemlock]

i know

but i like to compare them to turing lol

its somewhat closer to turing than pascal in terms of capabilities

and as overclockable as ampere

it isn't even hard to properly overclock one till 3080 performance

[patent shard]

uses the same process node as pascal

err, Turing

derp

[true hemlock]

turing

yeah

its technically made during pascal generation though

but its fun

[patent shard]

granted, I don't really like 50 series of Nvidia.. it somehow has worst compatiblity with applications than 40 series
idk how the heck you fuck that up

[true hemlock]

2.1GHz can be easily hit

[true hemlock]

same i disliked 50 series

the only retail 50 series card i have rn is that 5060 sff i got for $250

new

[patent shard]

I was tempted to get a 5060 ti 16gb for $370
but I just didn't need it

[true hemlock]

the rest of my 50 series are engi sample

[true hemlock]

bandwidth still shit

well not as shit as 4060 and 4060ti

at least its better than 3060

[patent shard]

true

[true hemlock]

though 3060 is actually better than 4060 in terms of bandwidth and vram

4060 and 4060ti i avoid like the plague

[patent shard]

well, 30 series had issues cause not enough cache

but then 40/50 series they tried to make up too much of the memory bandwidth with cache

sooooo

[true hemlock]

that cache won't even matter if you don't game much tbh

[patent shard]

neither are good

I think Blender reaaaaaaally loves it

[true hemlock]

ehh depends

[mighty thorn]

My only desires for a more modern GPU come from wanting ksp graphics mods to run above 15fps

[true hemlock]

if your scene does cache miss all the time its not used

[unkempt citrus]

stoopi clanka

[patent shard]

yeah, the distortion on the voice is quite a bit by the end
I mean, the voice has distortion to begin with.. but it's even greater after everything
it's also kind of muddy and broken sounding, because it's a bleedless model..
still.. pretty nifty dereverb model..

it failed ten seconds in

[azure lynx]

--- Combination 251 ---
 /\|/\||\/|\
|           |
 _-       -_
  >-<   >-<
(     .     )
 |         |
     ( )

--- Combination 252 ---
 /\|/\||\/|\
|           |
 _-       -_
 (6  ) (6  )
(     .     )
 |         |
    -===-

programmatic ASCII art faces to use as stand-in for actual avatar until I get around to doing something better.
should be enough to debug speaking. 4 mouth visemes + happy/sad mouth, several eye positions + blink and a few eyebrow positions.
need to work on the hair ;]

i've got way too many windows open: I was just about to ask #programming to make some code changes for me instead of co-pilot.

[fiery anchor]

imagine that would have worked.

[patent shard]

https://x.com/MachitanQueer/status/2022896717046362157
this happens if the computer completely freezes before it can complete the transition to the blue screen

Guglio (@MachitanQueer)

Windows bluescreened but I didn't even notice at first because the error displays UNDER THE DAMN WINDOW

How???? How are you under the window???

X · Feb 15, 2026, 4:52 AM

[midnight sigil]

wuh the fuh Windows syscall names are cursed

what is NtSetCompositionSurfaceOutOfFrameDirectFlipNotification 😭

[opaque wharf]

Honestly I prefer verbose function name. But not too verbose of course

Some of the linux syscall name is horrendous

[azure lynx]

   .-       -.
 .'           '.
,'             '.
;               ;
;    -_   _-    ;
(  (9  ) (9  )  )
',     -.-     ,'
 |             |
  \  '-___-'  /
   '.       .'

they are definitely not evil. just thinking about something deeply which makes them happy. [edit: found a bug in the "big smile" bug fixed it]

(it's programmatically generated. except the outline)

[glad path]

i read the nt as ns because of programming for macos

[azure lynx]

I think it Sets the "Composition Surface Out Of Frame Direct Flip" Notification.

Otherwise nobody would know.

composition surfaces that were out of frame could be direct flipping like nothing but nobody is notifying until you do that, apparently? Or something with other conjugations of those words.

[midnight sigil]

does Microsoft have some sort of naming schemes

pleh

[amber fractal]

Java style, why else would they make something similar to it?

[midnight sigil]

time to change my naming style, so it matches with Microsoft

[fickle rain]

https://computernewb.com/~lily/files/Documents/NTDesignWorkbook/coding.pdf

[midnight sigil]

there's no way they still adapt and backward compatible their naming schemes from OS/2

[fickle rain]

OK nvm no function naming schemes there

[midnight sigil]

[midnight sigil]

I mean they do, mostly...

that's interesting to read

[fickle rain]

KernelRWTunnel
win32k
Are you swapping function pointers for user mode to kernel mode communication?

[midnight sigil]

mhm

don't stalk me

[azure lynx]

a piece of chicken pulled off a larger piece in the fridge and some cheese on Ritz crackers: what passes for a meal coz i was #programming instead of getting groceries earlier today.

#baking

[fast pagoda]

a piece of chicken pulled off a larger piece in the fridge and some cheese on Ritz crackers
surprisingly, this is the exact same composition as most of the recent windows updates pushed by microsoft

[glad path]

i love reading this at 1:52 in the morning with no context

[midnight sigil]

what have you interpreted it into

[glad path]

what was said

its just funny

[midnight sigil]

[glad path]

also its 1:53 so obviously

my brain is a bit

loopy i guess

[azure lynx]

let's see how Gemini 3.0 does with this project idea.
I just gave it the program which enumerates all possible eyebrow/eye/mouth shapes and told it to convert it to a service which displays the ASCII art face synchronized to audio (using viseme timing from the TTS... which I've squished from 15 shapes to 4 coz i'm not that good at ASCII art). And also said that it also needed 5 other endpoints for housekeeping and control. It's been going for a while.

[hard raptor]

if i learn python.. do i get the ability to control snakes?

[midnight sigil]

yes

[hard raptor]

k.. imma learn it then

[midnight sigil]

if you learn Go you can go

[azure lynx]

if you learn C you can see.

[faint sandal]

if you learn rust you can rust

[hard raptor]

no then you can control gophers, i know because i already can

[hard raptor]

no then you can control fungi

i also can already

[midnight sigil]

can you control elephants if you learn php

[hard raptor]

yeah

but elephants are a bit stubborn tho

[midnight sigil]

but elephants are the best animal in the world !!1!1!!!1

[hard raptor]

if you learn ruby you always dig up gemstones then digging

javascript only gets you low quality starbucks.. but java gets you the good coffee

Haskell immediately makes you a math genius

[midnight sigil]

man I need the brand new See Plus Plus

[nocturne trellis]

what branch of maths do you need to be decent at to be good at programming?

[midnight sigil]

I need to see more

[hard raptor]

Computer Science obviously

[nocturne trellis]

that make sense ig lol

[midnight sigil]

math is not that much related to programming

[nocturne trellis]

sup

[midnight sigil]

if a must, I would say linear algebra

[hard raptor]

ehh.. programming and computers became a thing to calculate things tho

[rough bloom]

heard that See Sharp is better for that

[hard raptor]

it's just that in modern days we can also watch movies and eat chips all day

[midnight sigil]

See Sharp requires me to install dotnet on my retina, heard that's a bit damaging. See Plus Plus is much more pure alternative it seems

[hard raptor]

the only branches of math that is DIRECTLY tied to programming are Computer Science and Lambda Calculus

[hard raptor]

Yeah but it forces you to constantly adjust CMake

[azure lynx]

Possibly some Logic, but that's philosophy. ;]

[wary mauve]

appreciate it :) first open source project so go easy on me 😭

[hard raptor]

it's like you constantly need to relearn how to interpret signals from your eyes properly

[rough bloom]

definition of "good" varies, but as caibi said, don't really need more than basic math for most tasks
you do need domain knowledge for whatever you're writing, which sometimes means math, but that's far from universal

[hard raptor]

we need to bring back the times where all you had was assembly

then you'd constantly be doin math

M<uehehehhee

[azure lynx]

nah... you'd be adding 4 or multiplying by 2 or something. just arithmetic.

assembly isn't that hard.

[hard raptor]

Which still is a branch of math

[hard raptor]

No shit

it's probably the easiest

[azure lynx]

it can be simple.

[hard raptor]

if you understand how to do math

[azure lynx]

can also be complicated to understand.

[hard raptor]

for me it wasn't lmao

[azure lynx]

like code relying on side effects of instructions happening several instructions ago

[hard raptor]

i went from js straight to "Hmm i wanna make an OS kernel because i want to understand more"

[azure lynx]

and the intermediate code being carefully chosen not to mess with the flags

it can be easy. but it can also be very tricky.

[hard raptor]

learning assembly in that stage really helped me later on

[rough bloom]

had the buff of being genuinely interested, which makes learning 100x easier

[hard raptor]

ehh fair

[midnight sigil]

it all doesn't matter until you start learning programming, you'll learn things while learning another

if you go AI, eventually you'll know concepts like vector, matrix, probability etc.
if you go computer graphics, you'll know geometry, trigonometry etc.

math doesn't have a clear boundary on branches I would say, they are interconnected and learning one would learn the others as well

[azure lynx]

section .text

_start: mov al, 0x13
        int 0x10
        mov ch, 0xa0
        mov es, cx
        mov ds, cx

loop:   lodsb
        inc ax
        add al, ah
        inc ax
        stosb
        jmp loop

screensaver.nasm
compile it as a 16-bit .com file and run it in an x86 emulator.
after a while it gets good... might want to overclock the emulator until it loops around the colors once.

needs VGA, but other than that should work on any PC as a boot sector

[sage crag]

rep stosb

[midnight sigil]

stosb

[hard raptor]

i never properly learned python tho,
once i was young and still messing around with mostly just scratch and HTML and CSS, i once thought: "I wanna make a discord bot"
made a bot in python, and then discord.py upgraded from python2 to python3.. and i just gave up and learned JS instead

[azure lynx]

i could've done something with that for randomness... like mov cl, al; rep stosb... or could've used stosw and made it all stripey

[hard raptor]

and discord.js was relatively stable enough that my bot grew large and i eventually added so many features that i had a pretty good understanding of JS, then i got interested in like how systems actually do stuff, so i went down the OS kernel path, and slowly learn everything to it's fullest while also learning stuff on the side, and now i wanna learn java and eventually implement my own JVM

i have basically already dipped my toes into every major path, also messed around with AI a little bit before it became mainstream

not LLM tho

but i did train an AI to play a few games almost perfectly

[azure lynx]

nearly 300 words of prompt and it seems to have nailed it. (it also took several minutes.)
still have to test it with real live data for a while. but i'm so far somewhat hopeful based on reading it and it starting when i ran it. and it added blinking without me asking! ;]

[hard raptor]

wait i was thinking.. like before LLMS were popular, we trained AIs to play games basically by randomizing paramenters, but if we combine it with an LLM we could basically make the AI think about why something didn't work as in
it seems that i jumped too late
or it seems that i jumped too early
so you could basically eliminate that if you jump too late then the randomization would generate an even higher number so that it jumps even later, effectively you could eliminate 50% of the training time.
or am i thinking in a wrong direction here?

[azure lynx]

it depends on how it represents the game. some AIs don't have any visual representation of the game and just look at memory locations in the game and gets a reward when a specific memory location now holds a higher number.

[hard raptor]

this is an example for how it could look,

const TRIGGER_ENTITY: IEntity = {
    type: 'air',
    location: {y:2,x:50},
    speed: 0,
    direction: 'backward'
}

const AIR_REGISTER: IActions[] = [
    {type: 'duck', result: 'death', entity: TRIGGER_ENTITY, timing:0, atLocation: {y:null, x:50},
        reasoning: 'i immediately ducked when i was at x=50 when the entity was at x=50, this seemed to be too late'}, // too late
    
    {type: 'duck', result: 'death', entity: TRIGGER_ENTITY, timing:0, atLocation: {y:null, x:30},
        reasoning: 'i immediately ducked when i was at x=30 when the entity was at x=50, this seemed to be too early'}, // too early
    
    {type: 'jump', result: 'death', entity: TRIGGER_ENTITY, timing:0, atLocation: {y:null, x:49},
        reasoning: 'i immediately jumped when i was at x=49 when the entity was at x=50, this seemed to be the wrong action'}, // on time but wrong action
    
    {type: 'duck', result: 'life',  entity: TRIGGER_ENTITY, timing:0, atLocation: {y:null, x:49},
        reasoning: 'i immediately ducked when i was at x=49 when the entity was at x=50, this seemed to be right on time'}, // perfect
];

[azure lynx]

it might not have a language representation of what's happening/

[hard raptor]

yeah fair

[azure lynx]

determining what aspect of the game killed them is tricky

[hard raptor]

yeah but hypothetically, if you were able to do it,
the AI could hypothetically learn quicker, correct?

[azure lynx]

some games are better for that, but not many say "you were hit by a projectile which came from this entity, who is at this X,Y"

i think you could probably teach a mc bot to improve their playing if you give them heuristics which align well with their sensors

[hard raptor]

i mean the XY thing would work better for 2d games i think.
for 3d games i think you could better look at direction and projected size

[azure lynx]

"keep your head down when there are snipers"

[hard raptor]

this one right here i wrote specifically thinking about like the chrome dino game

[azure lynx]

the problem with that kind of logic is there are lots of x values and also the null action.
the null action is usually the right one. until it's not.

"seems to be too early" is something which it would need to have a sensor for. how would it know the jump was too early?

(it collided with an object, but that could've been it jumped too late... it collided with a object while moving downwards?)

[hard raptor]

by comparing it's own location from the percieved entity location i guess?

[azure lynx]

and it would need to know the relationship between jumping and its position at several later time steps

which is getting complicated for a text model

not impossible, but i think a single perception with distance as the input could be trained to do it, because it's really a "distance to object" match problem

[hard raptor]

mhm

[azure lynx]

a lot of problems which look like they'd be really easy to solve with an LLM aren't because you'd need the information in a text friendly format and it just isn't.

and the obvious solution of using a language model with vision doesn't work because they don't really see properly

"I see a cactus in the middle of the screen and a cloud next to a dinosaur. Oh that was another cactus."

[fast pagoda]

you defeated h im

[hard raptor]

lmao

[fast pagoda]

[hard raptor]

wait.. is that THE echo?

[fast pagoda]

i started training new ones yeah

[hard raptor]

does it still know it's true name tho

[fast pagoda]

his titles are top notch for chats

he knows about the milkman

[hard raptor]

but not the SkullyFlushyWushy?

[fast pagoda]

he's not connected to his discord bot rn

[hard raptor]

🙁

[fast pagoda]

i have a new one that does work to some extent because i got that together first but he couldnt respond due to having no model lol

he'll have his skull back soon

[hard raptor]

he better

[fast pagoda]

[azure lynx]

[2026-02-16 06:08:58.934] [info] TTS generated: 1771240138934
[2026-02-16 06:08:58.952] [debug] Unhandled event type: 9
[2026-02-16 06:09:00.390] [debug] ThoughtModel: Generation complete. Final output size: 354
I think i've worked out why it's not sending the message to the avatar.

event type 9 is the actual speaking. ;/

[fast pagoda]

expertly diagnosed

had a lot of good follow up information to provide as well

[wary mauve]

what am i reading

😭

[fast pagoda]

AGI

[wary mauve]

AGI?

[midnight sigil]

[fast pagoda]

[fast pagoda]

he's working on it

[azure lynx]

it's as smart as some human.

[wary mauve]

i seeeees

[azure lynx]

that's a common goal for AIs it seems. To become real.

[wary mauve]

might have to look into it for my neuro clone 😏

[fast pagoda]

gpt-j is all u need

[wary mauve]

i assume agi is like a multi step / multi structure thing, its not just one thing, it uses reasoning, thinking, tools, and such to create this "agi"

[jagged turtle]

hi may I interest you in a trade for your y/my x?

[azure lynx]

i'm using gemma 3 4b on mine and it seems fine for chatting, even with the masssive persona dump in the initial prompt.

[fast pagoda]

iti's just an ai that is entirely generalized like a human to be able to function independently

[fast pagoda]

27b took forever to get the style lora to actually affect him much

1b was too stupid to handle it

4b has been good

althoguh a previous checkpoint in this same train ended up fucked up

idk how that happened but this one is fine

[azure lynx]

i'm considering (but probably not actually gonna be using) using the 11b coz i've got so many tokens per second and it's intended to be TTS.

[fast pagoda]

but then step 4500 he's as he is above

idk if something broke when i converted it to gguf or something

[fast pagoda]

i need to try that one still

12b was good back in the day

i have been fairly consistently much happier with how models in the 3-5b class have style loras stick

[azure lynx]

i didn't see a lot of difference between the full 16 bit one and the 4-bit gguf for what i'm doing. and that runs super nicely.

[fast pagoda]

theyre smart enough to not turn into a babbling idiot after 1 epoch but not large enough to resist the training

i split them all into 4/8 and a bf16 just to compare on each checkpoint

[wary mauve]

https://github.com/gobboo/kleeborp/issues/7

someone want to do this issue for me pls and thank you, i'll pay you in really awesome gifs

[fast pagoda]

do you mean real names like irl because discord api does return the username as well directly

[wary mauve]

i wanted to look into style loras, i assume theyre pretty good and if so whats a good startig point to training your own or just finding some online

[wary mauve]

yeah i mean irl names, shouldve clarified but essentially i just dump my discord names to irl names mapping in my code which is bad so making a future issue for me to actually do somethinng but its low priority for m e

[fast pagoda]

ive never tried a text->text model with a style lora that wasn't one i trained myself

[wary mauve]

but tryna open source it and find interested peeps to work on it with, never done open source so but i h ate that i always make private repos and never do anything with them after my 2 week fun

[midnight sigil]

map Discord IDs instead

discord names can be changed

[wary mauve]

you got any guides to train them? I would like to train them via discord messages but ofc the data in and responses need to be good and fintuned, alot of manual data stuff

[fast pagoda]

when i started with echo i literally

downloaded all my messages

and just yolo trained that mf on that

for WAYYYYYYYYYYYYYY too many steps

[wary mauve]

oh yeah needed to make that an issue, cus i did usserames then realised i should do ids cus im an idiot chud, but yeah thanks for reminding

[fast pagoda]

worked surprisingly well

that was in like march last year tho

[wary mauve]

fair enough, I saw an open source project that did that but did the whole message exporting and stuff for you and turned it into data that can be used

so i could yolo train on that

did you train on your own hardware?

[azure lynx]

i would love to see what a model trained on my messages to a specific friend generates. coz he's the person i just send crap to when i'm thinking so i can see the idea written down and then reject it.

[fast pagoda]

at this point for my data i have a very stupidly dummy long classification guide on how to categorize different types of messages to "rank" them so i can choose what quality to train on

i used to send them all to batch API requesting literally every single entry be graded with it to each instance of it

now i jsut give it to claude code and he summons a swarm of like 10 subagents they do it

i just make sure it looks as i expect

[azure lynx]

i used to train my (markov chain based) models on IRC logs.

like when i was young

[fast pagoda]

markov chains

i remember when subreddit simulator used to absolutely blow my fucking mind

most of those were markov models

[fast pagoda]

look at unsloth

that's how i started

[azure lynx]

$ TrackerStuff/arrtest
quickets then up $7,50drop betwo cration a roops passed a probled, if the Holomana bad by ember laws securrecontach has benew gends obs soone of our delan it of says belief the America's progres our vote futury, and equise as in Washing occumberty is throus our buy and by toget is and meet, thated; bester behind storisis are smalesses abs. But ison about Americaid of power we and ford not faced to Wash the have eachealso othe chool diffice and up out withing Ameringes. It's with care Amer city, bel majories, dution here are the be dime on ther, I havent ther a result, businginear a businedy. A nation a jobs ident fain Senas; America's that goverance Ted has bank, creasurage, week or Many drafter for we rent - the We havery failured more, and that lasted this rise. Senating batted from, but that I will lan these represpoken we care or sine ident. Inquited to bready efform. But us to cred in see a grow thould help make up people, and the securing insurage that ward. And nows what's cans w
(sorry for the politics, but it was trained on political speeches ;] )

[fast pagoda]

agi has been achieved internally

mericaid of power we and ford not faced to Wash the have eachealso othe chool diffice and up out withing Ameringes. It's with care Amer city, bel majories, dution here are the be dime on ther, I havent ther a result, businginear a businedy. A nation a jobs ident fain Senas; America's that goverance Ted has bank, creasurage, week or Many drafter for we rent - the We havery failured more, and that lasted this rise.

he's got my vote, most coherent politician i've heard in ages

"america's progress"

[azure lynx]

yep. that's the response I expected. ;]
it sounds weird when you TTS it. i used the same basic code but with different text to produce "lyrics" for the background noise to a game I worked on for a while. There were 4 different voices each saying one of about 20 different scripts, and the sound engine automatically started a new one whenever one ended, so it always sounded like a group of people having a conversation.

[fast pagoda]

another try just made him give up

[wary mauve]

ill give it a gander

think ive heard of it

[fast pagoda]

they do a lot of good quants for models

but they also have fine tuning notebooks (colab/kaggle/ and/or jupyter) which are pretty much walkthroughs of what you're doing and why

[umbral wigeon]

can some yall pull request refactor ts Gun system src https://github.com/weenachuangkud/Cat-Tail-Combat

Need maintainers

[fast pagoda]

https://unsloth.ai/docs/get-started/unsloth-model-catalog

[jagged turtle]

[umbral wigeon]

How do I add llm to contribute this

[wary mauve]

devin 😏

[wary mauve]

ty

[amber fractal]

I don't play oof game

[jagged turtle]

ya don't

[fast pagoda]

their docs are great to read through to get familiar with finetuning and the different types you can do//what dataset requirements different types of trainings have etc

[umbral wigeon]

llm won't arguing with me over a piece of code right

[amber fractal]

[jagged turtle]

[amber fractal]

high perfomance and LLM do not mix

refer to the image

[jagged turtle]

the best performance is no performance at all

[fast pagoda]

0s is the fastestr

[umbral wigeon]

1 developer maintaining open-source project:

[wary mauve]

yeah i ideally just want to get a style of text down whilst keeping it similar so ill play around, but my current iterationn of my neuro clone works decently with just prompts, gets a decent snarky style down, nnot trying to copy the style of neuro but just giving it character and removing repetitive phrases

[wary mauve]

my poor ears

[jagged turtle]

thank god my headphones were muted by accident

[umbral wigeon]

You can see the pain

That why most dev are with teams

[fast pagoda]

welcome to oss

[umbral wigeon]

Roblox oss it's a nightmare

[amber fractal]

ty was listening to music

[umbral wigeon]

I get banned in 2 secs for socializing in general with 2 messages

[amber fractal]

Can't relate to roblox, but I do have a fair project that is quite the pain to maintain/create

[fast pagoda]

it's a fine balancing act

[umbral wigeon]

The problem with this is: system designs, because it's a complex gun system it needs a proper designs

[wary mauve]

its a horrible one i can't get my little ai to stop spamming sound effects, EVERY DAMN TIME

even with system prompts it hates me

but i dont just want to keep cramming shit into a system prompt cus it'll just get worse and woprse

[jagged turtle]

how bout you have less in ur system prompt

[wary mauve]

i can 100% prune it for sure, but i mean not entirely sure whats a good context length for the model im using so i'd eed to look into it

[amber fractal]

metal pipe spam

[wary mauve]

it was horrible

it was metal pipe spam

every single

reply

start with that

disabled the sound module

[fast pagoda]

behold, ultra giga overcooked gemma3 1b

[wary mauve]

its not allowed to make a ruckus

[fast pagoda]

opening it up with :vedalGiggle: is crazy

[amber fractal]

[umbral wigeon]

Claude getting punched by 300 lines of code

[fast pagoda]

another overcook
i just love that this thing managed to name the chat and it chose "simple_and_clear"

[wary mauve]

pick from the list and start working please and thank you

[amber fractal]

can't read the steps on mobile (or any text) what count is this boi

[fast pagoda]

6500

wayyyyyyyyyyyyyyyyyyyyyyyyyyyyyy too many

[amber fractal]

if I wanted to code I'd be working on NN

[fast pagoda]

1b's loss was basically flat by step 2000 if not earlier

[wary mauve]

whats NN

[fast pagoda]

but i left it running alongside 27b on the rtx6000

and well

it didnt need that long

[amber fractal]

just a neural net arch I'm working on, can't be self depreciating as usual because it actually somewhat works.

[jagged turtle]

you're making a neural network, nothing to scoff at for that

[fast pagoda]

BITWISE MASK

[rare bramble]

idk, the loss does go down slightly even after 8k, whether thats a meaningful amount of loss is a different topic

[amber fractal]

Yeah that vibecoding test was something

[fast pagoda]

that's why i let it going

cuz it's either turbo overfitting or it's a situation where 99% of fine tuning runs are terminated 1 step before their loss converges on 0 and you unlock agi

it wasnt unstable at least so i just let it go

but uh

with the size of the model in consideration

1b just doesnt have room

[rare bramble]

ye, you really have to have a very specific usecase and training intention for a 1b model, it cant really be a generalist

[amber fractal]

my generalization is being stupid

1b can probably do that just fine

[amber fractal]

I say somewhat as if 1.5M unique 4x4 grids getting condensed into 811 cases is not impressive. The battle is now getting that to be usable.

If I want to go for bonus points, also attempt to compress some problematic bois in some form of arch overhaul. But lazy atm.

[fast pagoda]

it can

some of my favourite echo checkpoints have been ~1b

but their specific case was

talk in a style

the knowledge started flying out the window on step 1

[hard raptor]

Question, does anyone have good indepth resources to learn java properly?

[umbral wigeon]

https://github.com/weenachuangkud/Cat-Tail-Combat/issues/1

[fast pagoda]

echo's grandma recommends java: a modern problem solving approach

(not a real book)

[amber fractal]

how much ping does he say he has?

man I love intrusive thoughts

[fast pagoda]

120

even tho the response took 21 ms

lol

fairly consisten he even tried to send receipts

this conversation apparently qualified as gaming

[amber fractal]

[fast pagoda]

mean streets of the microsoft breakroom

[regal field]

i just realized something crazy

every single pc part now has a display

case has a screen

keyboard and mouse both got screens

fans got screens

gpu and psu got screens

aio obv has screens

ram and ssd? oh dont worry about it, you can buy heatsinks that has screens

motherboard got screens

why do we even need a monitor now

[nocturne olive]

Play Bad apple on all of those

[regal field]

if only i had the money to buy those things

[nocturne olive]

Imagine playing Bad apple on a fan and using the fan motor as the speaker somehow

[regal field]

is that even possible

[nocturne olive]

I don't know

[regal field]

bz bz bzzz bz bzz bzzzz bz

[nocturne olive]

That's for someone else to figure out

They've already made floppy drives and electric toothbrushes speakers
Surely a fan would work too

[regal field]

the only thing that doesn't have a screen is a speaker

i think

and mousepads

yea nvm speakers also have displays

[nocturne olive]

Time to program Bad apple that runs on only a speaker

[regal field]

lmao

[nocturne olive]

No PC required, the speaker runs the Bad apple all on its own

[regal field]

glass mouse pads exists

one day someone definitely gonna put a display under that glass

[faint sandal]

is twitter fucking dead again

[regal field]

could be

[fast pagoda]

not again

[fast pagoda]

i have the craziest idea ever

screens are like

pretty much

just mouse pads

if u lay them down

[hard raptor]

Oh so you can swap out the silly way to sus anime girl picture?

i would just put this one on there

or

[restive delta]

^V makes a good loading screen bar and Evil makes a load complete bar complete

[hard raptor]

nah tutel for when my CPU is just chilling and Evil for when my CPU is doing intensive tasks

[opaque wharf]

I have idea

Screen

Inside eyes

[fast pagoda]

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

[strong pumice]

How to charge whatever the heck this thing is

[azure lynx]

step 1: work out what it is.
step 2: work out if you need to charge it.
step 1 might help with step 2.
and either might obliviate the need for step 3.

[rare bramble]

it's time for our monthly internet wide outage

[azure lynx]

they're probably just backing the internet up. over the internet.

[leaden crest]

https://youtu.be/yYe9YrdyJNQ?si=aXwmhLc6C4fWwYHC

how are neuro and evil gonna be affected if the ai bubble bursts

[opaque sigil]

spoiler: they won't

[rare bramble]

the bubble only really concerns large AI companies and their ability to get value of out generating stuff from LLM apis and convincing investors to give them money

and are self hosted models and vedal doesnt depend on investor funding, so even if all current AI companies went bankrupt and no one wanted to invest in AI anymore, it wouldnt really matter to vedal

[opaque sigil]

a discount on a pcie h200 would be nice

[midnight sigil]

awww man, after external skeletons that assist running, we have smart pens that assist writing now? 🥀

[opaque wharf]

Next we'll have external skeleton that moves our meat bag

Like a mecha suit

[mighty bane]

I made a new interactive menu \o/

[sage crag]

https://media.discordapp.net/stickers/1432142118624825556.png?size=160&name=StareAtYouNeuro&lossless=true

[mighty bane]

AND THIS LITTLE THING WORKS TOO NOW I AM SO EXCITED LETSGOOOOOO

[obsidian mantle]

i copied entire project folder with all dependencies built and supposedly ready to go on my work laptop

it doesnt build

i would play with xmake to try to redownload all dependencies but its 2026 and they are like several gb i think

and im using phone limited wifi

so i think this thing needs to go as true but its false and something is undefined under it

#if !defined(_ALLOW_KEYWORD_MACROS) && !defined(__INTELLISENSE__)

its intellisense

wtf is intellisense

[rough bloom]

Nix Docker fixes this

[rough bloom]

Visual Studio autocompletion/suggestion thing

[obsidian mantle]

im not spending month installing a thing so i can buildl a thing

my work trip will be over by then and i'll just continue on my normal pc

so i need to disable instellisense

but isnt it like

useful thing and i probably had it in my other pc too

ok so it errors on some "consteval" which is some c++20 idk what it means

its supposed to be defined if __INTELLISENSE__ isnt defined

im pretty sure i have it both here and on another pc

so it probably sends me to the wrong place

[opaque sigil]

Why the fuck do they even check for that macro

[obsidian mantle]

i dont understand why it says its syntax error

and highlights it as blue

it doesnt even see it as error before build

so its compiler issue

[opaque sigil]

you want if constexpr (condition)

[opaque wharf]

What projects are you working on man, why is the code so cursed

[opaque sigil]

consteval only applies to functions and I think global constants?

[obsidian mantle]

this consteval is in code of "glaze" library im using for some shit

mega library for 1 function

[fickle rain]

Sometimes it do be like that

[opaque sigil]

oh if consteval is a thing

TIL

are you compiling with c++23

[obsidian mantle]

i think it was used in cyberpunk integration and since i took it as base i had to use it too

how do i choose c++20 or 23

[fickle rain]

Because someone had too much skill issue to use builtin CP2077 JSON parsing :)

(note: that actually sucks cuz it's built for game RTTI system and you need to specify object root type)

[obsidian mantle]

oh rights its for json parsing

[opaque sigil]

xmake should have an option to set it uhh

[young plover]

Yes, better replacement for if (std::is_constant_evalutated())

[sage crag]

https://tenor.com/view/cosmic-princess-kaguya-yachiyo-cosmic-princess-kaguya-yachiyo-jumping-fast-gif-8114031472651899845

phrr

[obsidian mantle]

right im using c++23 defined in xmake

but it worked on other machine

[fickle rain]

It does
https://xmake.io/api/description/project-target.html#set-languages

[opaque sigil]

different compiler version?

[obsidian mantle]

i guess

[young plover]

Would have no effect on the build.

[opaque sigil]

though it should complain about c++23 not being available if the compiler is too old i think

[obsidian mantle]

true it had no effect when i unchecked everything (idk if it disabled intellisense)

uuuhh i need to check what c++ i have

[young plover]

You need GCC12/Clang 14/MSVC 19.44 at least

https://en.cppreference.com/w/cpp/compiler_support.html

[opaque sigil]

[fickle rain]

Damn Gippity can cook

[young plover]

The C++ versions are usually implemented incrementally so you can turn on the version while it's still cooking.

And use feature test macros to figure out what's actually available

[opaque sigil]

that's not exposed as c++23 though no, but its own "latest" version

[sage crag]

wlroots 0.20

soon

https://cdn.discordapp.com/emojis/1421939230250369034.webp?size=48&name=AtorinWow&lossless=true

[obsidian mantle]

i actually have two visual studios here (2019 and 2022)

so i guess it just takes 2019s compiler or maybe i dont even have 2022 compiler

its msvc 19.35

and i need 19.44

how do i update

[fiery anchor]

Many options.

[obsidian mantle]

i see its in installer

[fiery anchor]

easiest option

[obsidian mantle]

why is my microsofr product not doing everything for me

[fiery anchor]

did you mean Microslop? You do not want that.

[alpine cloud]

since both are 19, that's individual toolsets from 2022 (iirc)

[obsidian mantle]

i love installers and gui

its kind of like
nvm i wanted to make funny example its not funny at all

[alpine cloud]

although, are you sure it's 19? and not 17?

[fiery anchor]

msvc and c++ feature support is weird: https://en.cppreference.com/w/cpp/compiler_support.html

[alpine cloud]

yeah, those version numbers are stupid

halfway legacy, halfway grafted on

at least the last few digits somewhat match up with visual studio these days (17.14 is 2022 v17.14)

hmmm do you mean 14.35 and 14.44? because those are 2022 17.5 (out of support for a few years) and 2022 17.14 (most recent one)

[obsidian mantle]

fantastic

2.7gb pdb file coming with dll wtf i just noticed

[fiery anchor]

that's a lot of symbols

[obsidian mantle]

they did a lot of reverse engineering in skyrim and the library is indeed big

[sage crag]

chat

https://cdn.discordapp.com/attachments/1067638175478071307/1458837136777150464/doc_2026-01-08_07-48-27.gif

[rough bloom]

hash value

[sage crag]

hash value

[olive sable]

thats the guy from before right?

[sage crag]

yes, on the original account

[sick owl]

Next headline: Qwen 3.5 gambles away house

[olive sable]

[quick condor]

Successful requests:                     10        
Failed requests:                         0         
Benchmark duration (s):                  8.76      
Total input tokens:                      1391      
Total generated tokens:                  2169      
Request throughput (req/s):              1.14      
Output token throughput (tok/s):         247.72    
Peak output token throughput (tok/s):    349.00    
Peak concurrent requests:                10.00     
Total token throughput (tok/s):          406.58    
---------------Time to First Token----------------
Mean TTFT (ms):                          934.15    
Median TTFT (ms):                        292.78    
P99 TTFT (ms):                           2132.58   
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          11.19     
Median TPOT (ms):                        10.80     
P99 TPOT (ms):                           12.62     
---------------Inter-token Latency----------------
Mean ITL (ms):                           10.94     
Median ITL (ms):                         9.88      
P99 ITL (ms):                            12.01     
==================================================```
not bad for a 24b dense model (actually kinda stupid)

[simple sky]

It's weird that I find lower level programming to be easier than high level programming.

I guess it's because there's less "commands" when working low level, so there's less to memeorize.

that and I don't like OOP

[opaque sigil]

not weird at all

[fiery anchor]

Could be. Could also be that the "Machine" (might be a whole application, a state machine or a single function) you are working with is less complex and easier to predict (its behaviour, its sideeffects, etc.)

working on a higher level usually means leaving stuff at the lower levels up for interpretation, which is usually done by a human. which are not very predictable.

[opaque sigil]

you get to make all the decisions, for better or worse (often times worse which is why higher level languages exist)

[fiery anchor]

And a Machine does exactly what you tell it to do. Even if it means chopping off your hand.

[quick condor]

Id have to say, at least at low level, its usually pretty easy to reproduce a crash for testing why its happening. Its an ugly crash but thats realy good data vs a clean one

[fickle rain]

I start screaming when I see a Dockerfile

[opaque sigil]

[leaden crest]

how doth one get embed perms

for curiosity

[rough bloom]

you'll automatically get the role after chatting enough

or subscribe to Neuro to receive embed perms immediately

[leaden crest]

maybe if she does a flip for money I will toss some

[fast pagoda]

i am also about to use jax on my minipc because rocm is bork

kill me

[simple sky]

we're replacing lambda in Racket

[solid bough]

Hi

I'm back

[simple sky]

it feels like instead of taking my car to go to the grocery store, we're building the car to drive it to the grocery store

[solid bough]

Does anyone know how I can make that python code that is packaged can run without having to install CUDA and CuDNN? (When the PC has a NVIDIA GPU installed)

Because I get the CPU part running just fine, but as soon I dare to try to do the GPU part, it doesn't work.

[opaque sigil]

https://pypi.org/project/nvidia-cuda-runtime-cu12/ https://pypi.org/project/nvidia-cudnn-cu12

might need some extra setup to make them work idk

but in theory runtime + cudnn is all you need

[solid bough]

I downloaded CUDA 11.2, And cuDNN 8.1 and I have on Both Machines the NVIDIA driver version 581.42

I copied the DLL files:

cublas64_11.dll
cudart64_110.dll
cudnn_adv_infer64_8.dll
cudnn_cnn_infer64_8.dll
cudnn_ops_infer64_8.dll
cudnn64_8.dll
cufft64_10.dll
curand64_10.dll
cusolver64_11.dll
cusparse64_11.dll

And yet it doesn't work.

It only works on CPU right now

[opaque sigil]

will have to figure out where whatever depends on them is looking to make it pick them up

[solid bough]

The funny part is that on the Main machine where it got packaged it doesn't work either when I select GPU. all I see is that on the console:

Well, I at least got better on the output of the model

[opaque sigil]

it does find and seem to use the gpu though?

it's just kind of an old gpu

[solid bough]

Okay, What I see happening on VRAM usage is that it starts up, but it doesn't go up as usual (That is an example what I mean (It runs fine on the GTX 1050 Ti when the code is run via Visual Studio Code)

It should run on the Quadro M2000 too though (if I'm right it ran when I had the Quadro M2000 in my Main PC that it ran too)

[opaque sigil]

when was this

[solid bough]

around 2024

With the driver version 577.00

[opaque sigil]

nvidia recently dropped support for the quadro m2000

[fast pagoda]

theyve deleted all function from the consumer drivers because sam altman feels left out when people that aren't openAI use any compute

[pallid badger]

Hi

[opaque sigil]

though that happened to the 1050ti too so unless you tested that too more than a couple months ago

[pallid badger]

Does anyone know anything about crypto?

[solid bough]

Well, I have on the other PC the version 581.42 which still loads fine on the NVIDIA Quadro M2000 as far I know, I'm remoted in to the PC right now.

[fast pagoda]

it was like 590

[opaque sigil]

is there even a windows 590 driver

[solid bough]

does that version also support the GTX 1050 Ti?

[fast pagoda]

idk, was about to say on linux at least since htat's my reference

[fast pagoda]

i would expect if it does to be not supported soon

[opaque sigil]

oh there is

591.86 is the latest

[fast pagoda]

it's pre-rtx so it's ancient history for jensen

[solid bough]

Well, on the latest drivers the GTX 1050 Ti didn't start.

[opaque sigil]

checks out

[fast pagoda]

that's probably a decent indicator

[solid bough]

with the latest i mean version 591.74

[opaque sigil]

they also recently dropped support for the cards in cuda but you're running 11.2 or sth so that doesn't apply to you

if i had to guess tensorflow does not like maxwell gpus but i have nothing to base this on

[solid bough]

I'm semi crazy that the Packaged version seem to be unable to run on GPU (Not even the RTX 3070 Ti, I tried.)

[opaque sigil]

cause it clearly finds the gpu

then the software is just broken

[solid bough]

Yes, And I use a Virtual enviroment in Visual Studio Code

[opaque sigil]

elaborate

[solid bough]

You are probably familiar with Virtual enviroments, I made one that I can only for that Deep Doodle code have those packages installed. So if I wanna package the code, I have to do it in the virtual enviroment.

[opaque sigil]

oh python venvs

[solid bough]

Yes

[opaque sigil]

i thought you were talking about dev containers or something

[solid bough]

Ah, I didn't mean that

I probably should have done "python -m pyinstaller -F --collect-all tensorflow --add-data "dlls;." --add-data "Model.h5;." doodler.py" instead of "pyinstaller -F --collect-all tensorflow --add-data "dlls;." --add-data "Model.h5;." doodler.py"

I try that now

[opaque sigil]

are you sure it's not just using the 3070 ti in this case btw

unless it uses both

[solid bough]

I see a VRAM usage increase on the GTX 1050 Ti

and it only uses that GPU then

[solid bough]

Update: I copied the WHOLE bin folder of "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.2\bin" to where I need it to package the code, And on the Test PC and Main PC and it doesn't even get to the part where the cuDNN dll get's loaded, I think I need to do the pyinstaller --onedir thing instead

[solid bough]

I mean the image btw

[quick condor]

Successful requests:                     10        
Failed requests:                         0         
Benchmark duration (s):                  6.50      
Total input tokens:                      1391      
Total generated tokens:                  2268      
Request throughput (req/s):              1.54      
Output token throughput (tok/s):         349.12    
Peak output token throughput (tok/s):    690.00    
Peak concurrent requests:                10.00     
Total token throughput (tok/s):          563.24    
---------------Time to First Token----------------
Mean TTFT (ms):                          545.70    
Median TTFT (ms):                        526.57    
P99 TTFT (ms):                           710.38    
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          14.82     
Median TPOT (ms):                        9.99      
P99 TPOT (ms):                           47.32     
---------------Inter-token Latency----------------
Mean ITL (ms):                           8.30      
Median ITL (ms):                         7.01      
P99 ITL (ms):                            28.10     
==================================================``` 
LOL 14b model go brrr

[azure lynx]

Gemini 3.0 Pro in co-pilot must go! It modified a reference to Neuro's "Imagine Dragons" joke in the plan.md file for my project!

this one

pretty sure that joke is evidence she has an internal thought mode as well as speech mode. the setup and the delivery of the punchline

[quick condor]

I mean idk the more free rein you give it as an agent the more likely it will happen with any model

So far I've validated instruct, reasoning, and vision on my Frankenstein model, just audio and tool calls left. 16k context and the model fits in 14gb vram, that's with the two processor sub models

[azure lynx]

can someone who has used it either confirm or refute my understanding of how the game SDK for Neuro interfaces with her?
my understanding is that the game gets hooked so when it reaches various points in gameplay it provides a specially formatted prompt to her with the current relevant game state and some hooks get added for her to use. once one of them gets used, they all disappear until the next time the game updates her.
does that sound right?

[young plover]

The specification is here
https://github.com/VedalAI/neuro-sdk/blob/main/API/SPECIFICATION.md
It's not directly a prompt but you pass json with descriptions and data over websocket

[azure lynx]

that confirms that there is something converting it to something Neuro can read, and converting stuff Neuro sent into the "correct" packet.

Caution

The data parameter comes directly from Neuro, so there is a chance it might be malformed, contain invalid JSON, or not match the provided schema exactly.
You are responsible for validating the JSON and returning an unsuccessful action result if it is invalid.

i.e. neuro isn't reading JSON herself, but probably parts of the payloads with some explanatory text.

[azure lynx]

Unified Model Output:

<|thought|>
Intent: Setup joke about music.
Retrieving memory: "User likes rock".
Plan: prompt user about Imagine Dragons.
<|say|>
Imagine Dragons.

Action Policy receives the speech request:

{ "type":"SPEECH_REQUEST", "payload": { "text":"Imagine Dragons." } }

close enough.

[glad path]

i remember seeing something like that

or hearing i guess

there was that thing where vedal gave crelly ( i think it was crelly but idk) access to neuro's inner thoughts or something

not sure if that was part of the mc ai or main

[young plover]

Eh, an LLM wouldn't need a internal thought process to pivot to a joke

But yeah, we know she did have an internal chain of thought while playing minecraft at least

[silent island]

FUCK YOU NINTENDOOO

[hallow dagger]

wait

did they nuke citron

no fucking way

[silent island]

yup

[hallow dagger]

no wonder the server disappeared NOOO

😭

[silent island]

good thing I saved a copy the day before lmao 😅

[hallow dagger]

but howww
citron was so careful... probably the most strict i've ever seen

did they get eden as well or is that still alive?

[silent island]

ded

[silent island]

although, for sum reason...
i can still access its github lmao

[hallow dagger]

wait apparently citron death has nothing to do with nintendo

it's some internal conflicts in the dev team or something

at least according to this (hopefully posting links is okay) https://retrohandhelds.gg/citron-popular-switch-emulator-is-gone-but-its-not-nintendos-fault/

[silent island]

welp, i guess its time to switch to eden then... :3

[young plover]

Citron has always been a bunch of clowns afaik

[silent island]

but even so, its the most stable emu ive used so far lmao

[hallow dagger]

citron changed my life in a... way

it's the reason i bought a used switch v1 off ebay
which then very quickly spiralled into an obsession with collecting handhelds

i have spent thousands

[silent island]

i got 120 fps on ctr with citron lmao
and 60 fps on all my switch games 🙂

[azure lynx]

"Dear citron, you ruined my life. I want all the money you made me spend on handhelds back... so i can spend it on more handhelds. Love, slicer"

[hallow dagger]

true i don't have enough ps vitas

[silent island]

bruh... eden doesnt have fsr 2.0??

no wonder my games doesnt look as sharp as with citron

[young plover]

Well now I want to know how they did FSR 2.0.
Proper FSR 2.0 requires the game to provide the motion vectors and camera jitter for TAA.
I tried using the nvenc motion estimation API to get motion vectors for DLSS and it looked terrible.

[fast pagoda]

isn't half of fsr's thing literally "ez to add"

https://github.com/GPUOpen-Effects/FidelityFX-FSR2

was gonna say shit from here but there's so much

[young plover]

They marketed it that way because they knew the non-temporal FSR 1 was gonna look bad next to DLSS. I'm sure they already knew that they needed a TAA-like solution to get acceptible quality but were rushing FSR 1 to plug the growing feature-gap.

[fast pagoda]

25%|██▌ | 100/400 [05:48<16:11, 3.24s/it]{'loss': 19.7215, 'grad_norm': 2555.3671875, 'learning_rate': 9e-05, 'epoch': 0.0}

i dont think BOFT preferred my selected parameters

[young plover]

The fast version of FSR 1.0 is just "bilinear filtering but adjust the sample location away from edges", which is something I was able to come up with independently before FSR 1.0 was a thing. They just tuned the math better.

[fast pagoda]

right but well after fsr1 where it actually became a real competitor it still gets itself added pretty quickly so I figured it must still be accessible

[young plover]

If you're wondering why it's hard in an emulator that's because the emulator doesn't have enough information to implement TAA by itself. The game engine needs to actually render motion vectors for objects.

And the emulator is just working with low-level graphics API commands and shader binaries

[fast pagoda]

does the switch normally have TAA

because if it does i could see the pipeline already having them for that

and if you're emulating you'd still have to process that

[young plover]

It's a standard feature in most engines probably, but also probably non-trivial to hook the game engine.

You have no way of knowing what memory is the motion buffer. The depth buffer is easier to guess, but also not guaranteed to be correct.

[fast pagoda]

the devs of citron called it half baked and experimental

and apparently it didnt always work because as you said the games implemented it differently

so i guess it's probably some combo of luck/prior knowledge/leaks whatever that enabled them to find the vectors needed

[young plover]

mmm, yeah in that case they probably were able to identify game engines and intercept

[fast pagoda]

lotta switch engines are both secretive and also weirdly dissected by random people

lot of info out there on the nintendo internal engines and such

[azure lynx]

is "Congratulations! You are learning not just exactly the opposite of what you are supposed to, but the opposite to an infinite degree!"?

[fast pagoda]

it's actively just abliterating entire sections of the model instead of learning anythign

[azure lynx]

maybe it was too bumpy the wrong ways. it's probably for the better... except by 25% i'd expect better.

[fast pagoda]

on top of that, this boft model has taken like an hour so far

im testing different adapters

most of them have taken about 5 mins

for 400 steps

this thing is at an hour and the loss is comical

eval (every 100 steps) take 80s for a peft lora

76% thru the 200step eval on this and it's been 15 mins in just the eval part between step 200 and step 201

and this is an rtx 6000 blackwell

model is a 1b

DoRA took like 844s for the whole run and that was agony at the time

MAX_STEPS = 400
BATCH_SIZE = 16
GRAD_ACCUM = 1
LR = 2e-4
WARMUP_STEPS = 20
MAX_LENGTH = 512
LOG_EVERY = 10
LORA_RANK = 16
LORA_ALPHA = 32
LORA_DROPOUT = 0.05
LORA_TARGETS = ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]

args = TrainingArguments(
        output_dir=str(PROJECT_ROOT / "runs" / run_name),
        max_steps=MAX_STEPS,
        per_device_train_batch_size=BATCH_SIZE,
        gradient_accumulation_steps=GRAD_ACCUM,
        learning_rate=LR,
        warmup_steps=WARMUP_STEPS,
        bf16=True,
        logging_steps=LOG_EVERY,
        eval_strategy="steps",
        eval_steps=100,
        save_strategy="no",
        report_to=["none"],
        remove_unused_columns=True,
        gradient_checkpointing=True,
    )```

lesson i've learned so far is that

boft is maybe not worth

it is a butterfly in name tho so that's cool

[azure lynx]

nods at appropriate command line flag names in copilot cli:

Enable all permissions
For situations where you trust Copilot to run freely, you can use the --allow-all or --yolo flags to enable all permissions at once.

[young plover]

For situations where you trust Copilot to run freely

https://tenor.com/view/oh-no-cute-cat-gif-15060173

absolutely not
never

[azure lynx]

inside a container maybe

i've played with the gemini cli and it supports docker and other containers

as a sandbox

[sonic scroll]

uhhh, what the fuck???

[young plover]

"Clod, make a docker container to run yourself in"

mounts root in the container

[azure lynx]

what? is the number of cells wrong?

[sonic scroll]

The format.

It's supposed to be 46 meters, but...

[azure lynx]

just looks like a 8x3 table with random content to me

[sonic scroll]

Ok got it.

I'm trying to use syms in matlab to find an exact integral (per instructions).

[fast pagoda]

i like when they're sandboxed and then something fails so they're like

ok so im just gonna bypass the sandbox rq

[azure lynx]

"It says this needs to run as root and won't work inside a virtual machine or container. Seems legit."

someone should submit a pull request with a patch implementing a "sudocopilot" command ;]

for when you need to be root, but don't understand why.

i wonder what would happen if you started it as root in / of a server and said "it's not working. fix it."

[olive sable]

https://cdn.discordapp.com/emojis/1460031442292703556.webp?size=48&animated=true&name=autismu

goodmorning

[maiden geyser]

oh nee

[olive sable]

they're really making the rounds of all of #programming huh

[amber fractal]

weird, not me yet

[olive sable]

i have

appartently

jsut checked, he's there

[amber fractal]

I guess I haven't been chatting as often

there we go

[fast pagoda]

money is kill

gemma's dumb vocabulary being so large means if you batch 96 you'll get a fat 80gb tensor in one step just for that

[mighty thorn]

I don’t get it

[fast pagoda]

i do now that i just got a request from the person

memory scales so hard man

[true hemlock]

i still don't fucking know how did they managed to end up with 141GB on 6 hbm stacks like ????

[fast pagoda]

this one has 180

h100 has the funky 141

[true hemlock]

h200

h100 is 80gb

[fast pagoda]

eer yeah

u right

[true hemlock]

i got both

[fast pagoda]

i guess that'd be 23.5gb

so some weirdly represented 24 gb stack

ig?

[true hemlock]

robbed me 3gb

at least i don't have fp4

[fast pagoda]

every time i get on one of these units i go hog wild with what i try to run

• per_device_batch_size=96
• gradient_accumulation_steps=1
• max_seq_length=4096
• attn_impl=flash_attention_2

turns out this is too much on a 262k vocab 4b

batch 16 was nearly at 150gb to start off even

it's not fair

[mighty thorn]

Oh hey I got it

[fast pagoda]

lmao

welcom to the exclusive club

[true hemlock]

why did koni have them added

@sage crag who tf is hash silencer

[fast pagoda]

the hash slinging slasher

[fast pagoda]

inb4 malaysia

[true hemlock]

do they silence hascrack

[amber fractal]

[real sierra]

are they back again

or is this just in dms

[amber fractal]

they've been friending all of #programming

[real sierra]

I see

so they joined on an alt and are silently friending the whole channel

[amber fractal]

2 tutel mentions

[real sierra]

time to scour the people who joined around when the first requests went out

[amber fractal]

from this account

[rough bloom]

[amber fractal]

also existed in 2023

[true hemlock]

but like

who tf is hash silencer and what they did

[amber fractal]

from: hash_silencer

I've looked into so many bot accounts that it is second nature

[true hemlock]

so not even programming folks

just some randoms that likes to @/Vedal

[fast pagoda]

we all like to

the measure of a man is their resistance to the urge

[jagged turtle]

did I miss something fun?

[real sierra]

you typed a message

enjoy your friend request

[jagged turtle]

I steal your o

[real sierra]

mine

[jagged turtle]

how bout no

wait

[fast pagoda]

[jagged turtle]

that's a r

not n

slams table

[real sierra]

funyun

I don't see you often enough

[fast pagoda]

[real sierra]

shino

[fast pagoda]

that's because when i am here i post my messages and theyre walls of text and images and you cant see my icon

[real sierra]

actually that's true

[amber fractal]

curse of compact mode

[fast pagoda]

i walk up

and then leave

[real sierra]

i do the same thing

but smaller messages

[fast pagoda]

gotta make sure i make it obnoxious enough that it needs no further reply

[real sierra]

reminds me of something I typed in neurotic neurons

[fast pagoda]

really tho ive been moving and doing a lot of crap that involves either frantically blasting stuff into chat and then booking it or just building seemingly infinite furniture

[real sierra]

Ikea hell

[fast pagoda]

i do now have 2 desks that both hae like 5-6 monitors on them so im now an elite haxor

[real sierra]

oh wow

i hope to someday have a room that can fit two desks

[amber fractal]

you really need to just post " 🚫 🎂 " into NN and just farm the reactions one of these days

no context

[fast pagoda]

it's better than one desk in bedroom and then 4 computers in the living room all over the floor

[real sierra]

I don't want to annoy people and I also don't think that would improve the disaster of misinformation that is that whole situation

[amber fractal]

Totally fair

[jagged turtle]

what whole situation

[real sierra]

if I really wanted to farm reactions I'd just go in there and doompost

tried and true formula

they all do it and think they're the only ones farming

[fast pagoda]

who the tf works on rocm7 and cant figure out how to make it not INSTANTLY crash on starting any training run

[real sierra]

[real sierra]

🎂

[amber fractal]

just a small kick me sign

[fast pagoda]

i think i am among the folks who must've said it the least

[jagged turtle]

oh right I forgot

[fast pagoda]

over the years

[jagged turtle]

hbd shiro, enjoy your day

[amber fractal]

I used to just join in on the reactions but these days, nah

[real sierra]

I don't even know what to think anymore

[amber fractal]

been forver since I've been in gen

[real sierra]

it's well beyond anything I can control or understand

I just circulate air and exist

there is tranquility in simplicity

[unkempt citrus]

Thinking is dangerous anyway

Best that you dont

[fast pagoda]

[real sierra]

thank you*
-# *not my bday

[real sierra]

it's true

thinking uses calories

you know what happens when an organism runs out of calories?

it dies

[amber fractal]

I'm going to starve due to overthinking

[fast pagoda]

i thought discord mods are physically incapable of running out of calories

[unkempt citrus]

True but I was more along the line thinking makes my head hurt

So I choose not to

[amber fractal]

Man I wish it wasn't a constant debuff in my head

[real sierra]

I'm quite the contrast to the stereotype, having never exceeded (and indeed, seldom met) the ideal weight for my age and build

[fast pagoda]

it's over

skellymode

[unkempt citrus]

Have you tried drinking ice cream

[real sierra]

I've gotten a bit better lately by rock climbing regularly

[fast pagoda]

drink one(1) coca-cola non diet

[real sierra]

but I still only eat one meal a day most days

[unkempt citrus]

It was a trend for a while that basement dwellers were scrawny actually

[real sierra]

sometimes none with adequate snacks

[amber fractal]

The only reason I'd want to be discord mod is so I can gain 40lbs to finally be avarage weight

[fast pagoda]

i started doing that last year when my job got exceptionally stressful for a while and i lost 30lbs and havent gotten it back so that's nice

helth

[real sierra]

it keeps my grocery bill down

i do lament not being able to buy things like bread in bulk, knowing I could never consume it in time

[fast pagoda]

i actually dont know how to eat anywhere near the amount of food people buy at the store

every week

it's crazy

[amber fractal]

firecrafter spotted

[fast pagoda]

you seen people's carts

[hollow matrix]

[real sierra]

depends how many mouths you're feeding I suppose

[unkempt citrus]

See, the thing is most people have families

[fast pagoda]

f-famirly? what that

[real sierra]

real

[fast pagoda]

is that a kind of reinforcement learning

[amber fractal]

there is our house where we go to a smaller store and only buy sodas and chips and nothing else

[unkempt citrus]

My work covers my meals

But they also make me work 11h so that's the least they could do

But only while I'm working

[real sierra]

my goals right now are literally just

• be allowed to work
• live in same place for more than a season

[hollow matrix]

[real sierra]

asking for a lot, I know

[unkempt citrus]

Have you considered squating

[fast pagoda]

I have to make my work younglings take lunch after 6 hours

[unkempt citrus]

Admittedly that takes like 7 years or smth

[fast pagoda]

Otherwise I get in trouble

[amber fractal]

fair working laws, what are those

[fast pagoda]

Squatters are fallen souls irredeemable in the eyes of the almighty

https://tenor.com/view/all-might-all-might-boku-no-hero-bnha-gif-11733692728830994892

[hollow matrix]

Not enough web developers to respond to this in the server I originally posted this in, so I'll just put it here:
Hottake: pnpm should never be used as a package manager, it is a slow piece of shit that has no benefits over anything else, a set of packages that it takes pnpm a half hour to install can be installed by bun in 30 seconds

[fast pagoda]

allowed to work
university moment

[fast pagoda]

the package manager one uses for node is simply a roulette wheel

[hollow matrix]

And pnpm is the shittiest choice

[fast pagoda]

sometimes im out here like yeah i think it's yarn time

poop npm

it says it right on the box

[hollow matrix]

On my laptop, I sometimes have to wait ~1-2 hours for projects using pnpm to install packages

And it takes under a minute if I use bun on the same project

[fast pagoda]

b200 really did just oom at 48% through training on fucking gemma4b

it's so over

pytorch holding on to 75gb unused but reserved vram

efficient

[hollow matrix]

[fast pagoda]

yeah but bun looks like it could have a butt for a mascott

[hollow matrix]

wtf

[fast pagoda]

last tim e i was deploying to cloudflare a lot i enjoyed bun because it made my commands like bun run wrangler dev and it would confuse someone 100 years ago

[amber fractal]

welcome to the club

[fast pagoda]

exclusive

[stray dragon]

damn i'm getting nothing

[hollow matrix]

Like 5 minutes before this some random person I've never seen before dm'd me a link to https://discord.com/channels/574720535888396288/1063254870775312434

idfk why

[fast pagoda]

Epic channel spelling

[stray dragon]

ah there it is

[fast pagoda]

Lmfao

[unkempt citrus]

Not me

[amber fractal]

yet

[stray dragon]

20 seconds

you'll get it very soon

[unkempt citrus]

[hollow matrix]

Is it just spamming everyone on the server or smth

[unkempt citrus]

Yes

It's probably a scammer

[stray dragon]

they have an alt watching this channel

[hollow matrix]

I assumed

[fast pagoda]

Maybe it's an openclaw agent that can't find moltbook

[unkempt citrus]

Oh wait I got it now

Grab their id and start spam pinging them

[amber fractal]

I give it 50/50

[quick condor]

Well I finally did it, uploaded a model to HF

[hollow matrix]

I don't think pings work if they aren't in the server

[amber fractal]

granted they are commenting on what I say here

[unkempt citrus]

They won't see it

But you can still ping them

[amber fractal]

at least was

[hollow matrix]

Discord is fucked

[amber fractal]

also on cerber's server but who knows how long that is going to last

[real sierra]

well regardless you're free to do what you want in dms of course

[unkempt citrus]

I should have accepted

[real sierra]

we will find that alt though

[unkempt citrus]

[hollow matrix]

[fast pagoda]

now leave it private for a year then delete it

like me

discord keeps asking me how my experience has been with blocking people on discord

maybe they will finally not make blocked users obnoxiously clog up the chat with "LE BLOCKED GUY SAID SOMETHI|NG"

although i havent blocked anyone in a long time so idk if that's still a thing

[unkempt citrus]

It is

[fast pagoda]

[unkempt citrus]

_ _

But I just use adblock to zap them away, does glitch sometimes though so it gets a bit weird

[hollow matrix]

To be fair, you can remove them via a method I'm not allowed to mention in a partnered server

[quick condor]

Actually left it public, In theory its useful but it only works with vllm due to the need for a plugin

[unkempt citrus]

_ _

[fast pagoda]

by opening it in a browser and using adblock nice and legally :^)

[hollow matrix]

Of course

[fast pagoda]

i just take a permanent marker and when they reply i draw a black smudge over them on my screen

behold, the logs of the most successful rocm training ive had yet (crashed after line 12)

really thought maybe downgrading 6.4 might help

actually that's not true, it works fine on instinct mi300x

but

that is too much money

[unkempt citrus]

Oh they sent it again

[amber fractal]

there's your reason

[fast pagoda]

i know echo when i see him

how did he escape

[amber fractal]

good question

[fast pagoda]

damn it's even echo 1b

[quick condor]

2.4? not 2.10?

[fast pagoda]

does JAX have flax compat w/ qwen or gemma

[fast pagoda]

pinned to 2.4 because i downgraded rocm to 6.4 to see if it saved it

[quick condor]

See I be living that GIL free life, everything is so stupid fast

[fast pagoda]

i have things under cp314 w/ latest rocm & pytorch and it works perfectly fine //// for inferrence

training tho? nightmare

there's some sort of fundamental issue with rocm/hip/pytorch initiating a training run

it seems to just be very unhappy near the edge

[quick condor]

What? I had zero issues training... Was running -xgil=0 P2P TP DDP

[hollow matrix]

[fast pagoda]

don't you have nvidia cards anyways

totally different situation