programming

[idle dune]

Same

[amber fractal]

If I had the rights to decide that, I'd be attempting to abolish further LLM prameter scaling. That'd be more healthy for the world.

[olive sable]

i think its maybe worth it for contractors or something

replacing the laptop is nice for if you need something more powerful, or something more cheap

if you can change the subscription tier you have at least

[amber fractal]

1 year lockin

[olive sable]

im assuming you will get a the most recent high spec laptop for the highest tier?

so like, each year you can replace it for the newest cpu/gpu model

[amber fractal]

offered after 3 years

[olive sable]

cant i just cancel the current subscription, then re-enter a new one?

[amber fractal]

surely

[idle dune]

@olive sable The only thing I can see this really working for is businesses and enterprise customers. Where they can just bulk sub to a ton of units and refresh on a regular basis.

[olive sable]

thats fair ye

[idle dune]

Theres no option to buy out but I don't think he meantioned trade in for new.

[olive sable]

im assuming if you cancel the subscription you have to hand in the laptop

im assumign that if you re-enter you can jsut get a new laptop?

the only issue i see with this is they will just give you a used laptop from a previous customer

or they ban you from re-entering

[amber fractal]

My family actually asked "what happens to your data when you give the laptop back?"

[olive sable]

id guess they have no data protection for those laptops

[idle dune]

Back it up or lose it

[olive sable]

so you'd just be giving them the full drive

[amber fractal]

[olive sable]

i think the focuse of the problem was privacy concerns, not losing the data

[amber fractal]

Surely dell is a trustworhy company and well zero out the drives before sending then to the next person

[opaque sigil]

encrypted drive if i had to guess

bitlocker is enabled by default anyway

[olive sable]

so it mostly depends on if you have your own account or a hp account on the device

or what was it

dell?

[opaque sigil]

hp

[olive sable]

iggly

you made me look like a fool

[opaque sigil]

you still have to go through the windows oobe no?

[olive sable]

idk

[opaque sigil]

just their custom image with whatever shit they want to have preloaded

[amber fractal]

I is stupid

[olive sable]

for work have account managed by the company and stuff
i think hp can do similair stuff

especially with the hp bloatware

[opaque sigil]

you still need to use your own microslop account, this isn't an enterprise laptop

[idle dune]

Doesn't matter Microsoft gave Bitlocker keys over to the feds, they will probably offer them up to anyone now.

https://giphy.com/gifs/movie-sci-fi-misc-1bLfTdPis6fpC

[opaque sigil]

i don't see how that is relevant in any way

either way

[olive sable]

which feds even?

[opaque sigil]

bitlocker on by default so encrypted drive which is useless to them

[olive sable]

all feds? usa feds?

[idle dune]

FBI

[opaque sigil]

fedex

[idle dune]

https://www.forbes.com/sites/thomasbrewster/2026/01/22/microsoft-gave-fbi-keys-to-unlock-bitlocker-encrypted-data/

https://giphy.com/gifs/cat-kawaii-nub-5RYfzW81vjs2Yv1g1k

[olive sable]

Microsoft has confirmed it will provide these BitLocker recovery keys to law enforcement, including international requests, if they are presented with a valid legal order.

so no

not FBI

all feds

[idle dune]

God I hate Microsoft

[olive sable]

map of all places that get mentioned by the news article eventho its a global issue

[amber fractal]

That's the entire world for some people

[idle dune]

"international requests" international agencies will 100% take advantage of this

[olive sable]

most def ye

[idle dune]

Also missing Guam and Puerto Rico on that map

SMH

[olive sable]

i know what puerto rico is, but ive never heard of guam

[idle dune]

Oh

Guam is on the other ise fo the globe so it literally couldn't be on that map

[olive sable]

seems to be some island above papue new guinea

[idle dune]

Yeah

Its a US terrotory

[olive sable]

thats crazy

its on the other side of the globe

[idle dune]

They took it over during WW2 and it been part of the US since

[olive sable]

wow

i forget how far hawaii is, its halfway betweent he mainland and guam

[idle dune]

Yeah

I wanna go some day

[azure lynx]

territories are like states except they don't get any actual representation in congress (so don't get to vote), are forgotten about a lot unless some disaster happens, and don't get any of the advantages of living in the main 48.

like a cross between just the worst parts of DC and Hawaii

[olive sable]

imo they should just be a separate country since they're so far away, but what do i know

im sure there are some reasons

[opaque sigil]

[idle dune]

[olive sable]

[amber fractal]

[olive sable]

one of those is not like the others

[idle dune]

Okay so Opus 4.6 is very prone to schizoid meltdowns. It took up all my tokens in one massive rant about hardware prices when i asked it to explain the current state of hardaware pricing in 2025 and 2026.

Then it started just screaming that the age of the personal computer as we know it is ending and that we are so fucked

[amber fractal]

Valid crashout

[olive sable]

i wanne go back to the time opus was a codec

[opaque sigil]

https://cursor.com/blog/nvidia you will take vibe coded gpu drivers and you will like it

[idle dune]

https://giphy.com/gifs/soyjak-soyjack-soyak-erba0PjENm6CRdsPR3

[olive sable]

our gpus are now not only unnafordable, but they shall also be unstable and slow

[opaque sigil]

dw they already are

[olive sable]

even more so now

[amber fractal]

excuse me, I thought they were already doing that

[olive sable]

unstable^2

[opaque sigil]

Last year, the organization set a new engineering mandate: leverage Cursor to embed AI across every phase of the software development lifecycle (SDLC) and eliminate manual bottlenecks across code generation, testing, reviews, and debugging
looks like they were

[patent shard]

$370 16gb 5060 ti

[opaque sigil]

so nothing new

[olive sable]

the crashing is not new, the frequency will just increase

[idle dune]

Close enough, welcome W701DS

[olive sable]

instead of once or twice per week, it shall become once or twice per hour

[nocturne olive]

What if I just, keep Nvidia 580

[rough bloom]

NVIDIA doing their best to kill the CUDA moat I see

[amber fractal]

how the turn has tabled
https://fixupx.com/i/status/2018967764082274490

Coinvo (@Coinvo)

BRAKING: Nvidia CEO Jensen Huang calls it "illogical" to think AI will replace software and related tools.

**💬 390 🔁 438 ❤️ 8.5K 👁️ 5.65M **

FixupX · Feb 4, 2026, 8:40 AM

[idle dune]

Late 2000s was the age of innovaiton

[opaque sigil]

Before Cursor, NVIDIA had other AI coding tools, both internally built and other external vendors. But after adopting Cursor is when we really started seeing significant increases in development velocity.
ah this makes more sense

[olive sable]

i like how becoming dependable on an expensive and subpar coding methodology is a mandate

[opaque sigil]

nobody said anything about replacements tbf

[amber fractal]

That is fair

[opaque sigil]

https://fxtwitter.com/claudeai/status/2020207322124132504 now you can have it rant even quicker

Claude (@claudeai)

Our teams have been building with a 2.5x-faster version of Claude Opus 4.6.
︀︀
︀︀We’re now making it available as an early experiment via Claude Code and our API.

**💬 197 🔁 135 ❤️ 2.4K 👁️ 190.9K **

FxTwitter · Feb 7, 2026, 6:45 PM

[maiden geyser]

software is a gas

[olive sable]

my software is a liquid

[maiden geyser]

that's weird, usually it expands to fill the container

[olive sable]

there is high pressure in my bad code

[opaque sigil]

so it's a 6x increase

[maiden geyser]

seems like it's free now once you hit exactly 200k tokens

[opaque sigil]

true

[rigid snow]

HOW MUCH??

ping reply

netcafe discord config

anthropic greedy fr

also this monitor sucks for anything that's not a game

[opaque sigil]

[olive sable]

[amber fractal]

built for the blurry mess that is DLSS

[tender river]

[olive sable]

does dlss use the tensor cores?

or all jsut cuda still?

[opaque sigil]

take a guess on why dlss is turing and later only

but yea

iirc dlss is an int8 model

[nocturne olive]

Can't wait for them to downgrade it to FP4 to justify 50 series

[olive sable]

tensor can execute at the same time as the cudas, so i do wonder what the base performance hit is from

maybe the L1 cache being split between the 2?

[opaque sigil]

well you still need to run the actual inference which takes time

even with tensor cores being as op as they are

[olive sable]

they're async tho no?

[amber fractal]

40 series already locks multi frame gen, so I think they are running out of things to gatekeep from the 30 series

[opaque sigil]

there's still some extra work to make sure the frames end up coherent that needs to be done i think?

geforce tensor cores aren't as async as they could be

[amber fractal]

that might be easier by merging with the RT stuff tbh, keep that state and use it for prediction

[sage crag]

pssk

[nocturne olive]

Well inference mainly takes memory bandwidth

And memory bandwidth is the same between CUDA and tensor and shared

[olive sable]

the biggest performance hit is probably on the cache?

[amber fractal]

what is cache when one loads in a 500m model

I don't actually know how large it is

[nocturne olive]

Well ML models are so big they can't really take advantage of the cache
So the smartest would be for them to ignore the model weights in the caching

[olive sable]

the scheduler also has to do some context switching. i think the actual perfromance effects of that might not be too bad depending on the usecase tho

[amber fractal]

I think that isn't much compared to all of the matmuls it is doing

[olive sable]

i think the cuda cores would get more cache misses due to the tensor cores fuckin up the memory uses

im no expert tho

[opaque sigil]

it depends

[nocturne olive]

Nah the cache would just be completely useless because the memory fetches aren't hitting the same memory sections at all

[opaque sigil]

if you mostly rely on shared memory for the tensor cores it's fine

shared memory accesses don't go through the l1 lookup pipeline

[amber fractal]

only l2 I'd imagine

[opaque sigil]

cache efficient matmul algorithms are the only reason these models aren't unusably slow

no l2 either

you can think of shared memory as a user-managed version of the l1 cache

they share the same physical space

[amber fractal]

So that is how we say "don't drop this out of cache stupid I need to do something else rq"

[opaque sigil]

ye

[amber fractal]

All the sudden, how we are able to have efficent matmuls makes sense to me

one less thing to fight

and so happens to be the largest

[opaque sigil]

you could prefetch the current tile into l1 but that's kind of annoying

and kind of unreliable

[sage crag]

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

[olive sable]

[amber fractal]

every time see the doc emojis they only get more desynced

[olive sable]

cerba looks so happy in that gif

[amber fractal]

imagine not logging into an alt for so long discord emails you saying you have a DM

[simple sky]

Hey is this the right way to handle git with a large project with multiple people:

• so main is our production branch, no merging into main until everything is the sprint branch is cleared of any catastrophic bugs

• each sprint has its own dev branch

• each sprint has 4 epics and each epic gets a branch

We merge the epics into to dev branch when it’s done/a prototype feature of the epic is needed for another epic

Then once every epic is merged into the dev branch, we bug fix on the dev branch until it’s clean, then merge into main.

Is this a good workflow?

I say catastrophic bugs because it’s a school capstone project, and it’s being more treated as an alpha/beta rn, so some issues are fine as long as they don’t hinder functionality.

[amber fractal]

Not in enterprise nor in large projects, however this reminds me of how the linux kernel is handled. Each subsystem having a branch which is sent over to be merged into the main.

[simple sky]

Alright so I’m on the right track on how I want to organize this project

[opaque sigil]

pretty standard workflow, though depending on how long things are going merging dev into main might be a bit painful

[amber fractal]

that was my own thought as well, in theory as long as it is properly split then it shouldn't be the worst.

[simple sky]

Every sprint must end with two weeks, and no sprints will overlap with each other so no matter what there won’t be more than one dev branch actively being worked on at a time

[simple sky]

This is mostly because the professor grades commit merged into main

[opaque sigil]

2 weeks isn't too bad i guess

the worry would be if you have a rather long time frame between merges into main you might be changing a lot

[simple sky]

Also I’m terrible at using commits, my group mates are so good at committing.

This sprint’s branch has maybe 150 commits total, I probably contributed 6

[amber fractal]

I feel you, I mostly just commit when I feel like it is worth sharing for my personal projects.

[simple sky]

My professor expects 10 commits a sprint per person

[amber fractal]

quanity over quality :classic:

just "forget" to update the docs

[opaque sigil]

massive commits where you end up changing 20 different things don't really help anyone (i do this too cause i'm too lazy to break it up for most projects)

[amber fractal]

I'm at a scale where changing 20 lines is like 15% of the codebase. So I'm fricked no matter what I do.

[sick owl]

One benefit of 4k oled I hadn't previously considered is superb CRT emulation

Very nice for retro games

A lot of these 2D PS2 games look so, so much worse without a good crt shader

[amber fractal]

sure I should also split by type but lazy

the github is more of a view into the insanity anyways

[unkempt citrus]

We were talking about promoting injecting google ai yesterday

[sick owl]

Thankfully even though PCSX2 doesn't support them with a bit of tuning you can overly a good CRT shader onto anything these days

[amber fractal]

I knew this was still working

[idle dune]

I used ShaderGlass for my last Fallout 2 playthrough, it did a good job

[sick owl]

The built in retrocrush presets are great for console emulation

[idle dune]

[sick owl]

Only annoying bit is changing pixel scaling and aspect ratio correction between emulators

[amber fractal]

Wayland suprimists would never accept this as a valid usecase on linux
Not wanting to actually spark that fight but just jab at it a bit.

I love the idea of the project tho

[sage crag]

pls??

[idle dune]

I don't even know what Wayland is

[sick owl]

https://cdn.discordapp.com/emojis/1149717540734521414.webp?size=56&animated=true

Can of worms is open

[idle dune]

Its a display compositor?

"Wayland is a modern display server protocol for Linux and other Unix-like operating systems. It's designed to replace the older X Window System (X11/X.org) that's been used for decades." - Claude

[sick owl]

Yeah Linux nerds love flinging shit at each others compositors

Its like a whole thing

[amber fractal]

I know if I said this on another sever, it'd turn into another flamewar

[opaque sigil]

[rough bloom]

quite the opposite, it's practically built to allow this
Hyprland has support for running whatever fragment shader you want

[idle dune]

I'mma try wayland and see if it works

I customized my Mint isntall and according to the internet is supports it

[rough bloom]

no, it's a protocol that compositors implement

[rough bloom]

would not suggest Wayland on Mint
you will likely have a bad time

[sick owl]

If we're splitting hairs yeah, "wayland compositor" is more of a catch all for anything using it

[idle dune]

IDK I'm using KDE instead of XFCE/XFM4

[rough bloom]

big difference, since compositors usually also do window management

[rough bloom]

should at least work then
will still have to fight outdated packages and general lack of support but at least it's just that

[fiery anchor]

vnc and wayland is pain.

[idle dune]

It booted and Wayland works (for now

[young plover]

IIRC the register file and shared memory are the same thing for GPUs. So throughput focused that SRAM registers are fine instead of flip-flops.

[amber fractal]

I guess I was thinking more of a direct conversion of the app, screen capturing the stuff below except itself to use as the fragment shader input.

[opaque sigil]

the register file is separate

h100 sm for example

[young plover]

the guy I know who works on GPU firmware would contradict that. Might be that it only spills to L1 if it runs out of registers.

[rough bloom]

I feel like the proper way would be to just implement support for that in the compositor (just apply the shader targeted to a specific window before/during compositing instead of after)
alternatively could use LD_PRELOAD shenanigans but that has its own issues

[opaque sigil]

iirc as of cuda 13 there's now an option to have registers spill to shared memory

i wanna say by default it spills into dram but i'm not 100% on that

[amber fractal]

I feel like that would be the correct method, I just also know it isn't as flexable as the orginal.

[opaque sigil]

okay yeah according to nvidia they spill to local memory which is just global memory

looks like they still end up being cached so they'll kind of live in l1

[young plover]

yeah I went back in my conversation history

on nv if you trash a register which your shader doesn't declare it uses then it overwrites shared memory values

[opaque sigil]

ah yes, hardware black magic

[rough bloom]

you can already capture the whole desktop without too much trouble, so the main issue to replicate what ShaderGlass does would be knowing your own window position relative to that (as it always is with Wayland and gimmicky apps like this )

and I guess you'd also need to exclude your own window from the capture, which is sort of possible, but also with compositor support
-# how tf would X11 even handle that

[amber fractal]

My thought is the window iself is an overlay that remains uncaptured. Position is still an issue tho.

[rough bloom]

I have come to the conclusion that it should just be a compositor thing

won't have the window decorations around it but who cares

[tender river]

i will write my own compositor

i will

[tender river]

later

me too busy reinventing programming right now

[sage crag]

[sick owl]

Retrocrisis not retrocrush!

Knew I had the name wrong, was bothering me

[real sierra]

hi programmers

[tender river]

[real sierra]

as a continuation to yesterday's scammer message woes

a new regex has been assembled

[hollow matrix]

[hollow matrix]

Is it absurdly long

[hard delta]

does it filter for birthday wishes

[jagged turtle]

om

[real sierra]

h[\s>]*t[\s>]*t[\s>]*p[\s>]*s?[\s>]*[:：][\s>]*[\/\\][\s>]*[\/\\](?:[\s>]*[a-zA-Z0-9])*[\s>]*>[\s>]*(?:[\s>]*[a-zA-Z0-9])+(?:[\s>]*(?:\.|。)[\s>]*[a-zA-Z0-9]+)+(?:[\s>]*[\/\\][\s>]*[a-zA-Z0-9]+)*

[hollow matrix]

It filters for this

[hollow matrix]

Does it work tho

[real sierra]

it should be better than what i had put together before

[hollow matrix]

pog

[hard delta]

m-mods

it's a spam link and a circumvention!!!1

[real sierra]

firecrafter's account was stolen

[hollow matrix]

[unkempt citrus]

[hollow matrix]

It's unfortunate that discord makes it so hard to deal with something like this more effectively

[hard delta]

remember when it had no friends feature

[hollow matrix]

[hard delta]

you'd have to make a server with bro and both join to be able to DM each other

[real sierra]

you sure remember a lot for someone whose account is a year old

[hollow matrix]

[hard delta]

you sure remember little for someone who's had a thousand birthdays

[unkempt citrus]

Only a thousand

[real sierra]

i can barely remember what i had for breakfast

[jagged turtle]

!felloff

[stiff mica]

[hard delta]

you had cereal

with slightly expired but still edible milk

[real sierra]

i dont think i had anything actually

[unkempt citrus]

[hard delta]

gotta keep the authorities in the dark so they can perform their jobs well

[hollow matrix]

1115 birthdays

[hard delta]

it is not yet time for them to know

the eldritch truth

[hollow matrix]

@real sierra You've had 1115 birthdays since you were cursed

[real sierra]

[sick owl]

Tuned up shaderglass a bit since the scaling settings were a bit wonky earlier

Looks real nice now

Pretty damn close to a SCART CRT image

Sharper pic

[sick owl]

Ultra zoomed in pic of mine (left) vs an actual CRT (right)

Funny thing though, the text fringing colours are flipped because of the subpixel layout on the OLED

At least I assume that's why

[stray dragon]

oooh clever

[sick owl]

The fringing is more prominent on the OLED but only because the CRT is displaying a much thicker font

[silent cloak]

CRT stands for Crap Resolution Transformer

[quick condor]

I am creating something aweful... I'm gluing the audio processor from a 3b model onto a 14b reasoning + vision model using a 24b model's audio projector

[mighty thorn]

one is not like the other

[quick condor]

No EoRA?

[mighty thorn]

[idle dune]

I know what some people do for crt emulation is use high refresh and high prt oleds with bfi and shaders

[quick condor]

Why not

[mighty thorn]

didnt come up during my research + it looks scary

[idle dune]

But ot have a good time with bfi you need over 120hz panels

[quick condor]

For gptq it's stupid worth it

[obsidian mantle]

I faced the problem that i dont know how to fix
Its these stupid inputs that work weirdly in skyrim
So when any window that uses cursor is opened and the mouse was moved - after that menu is closed the mouse stops working.
UNTIL i move the real mouse.
It magically fixes everything.
But this is unacceptable

I just disabled cursor in all menus and it works fine
Except for the map

The map needs cursor to work

And the weirdest part is that i cant even disable mouse in it

Using old methods

Even disassembling that shit player bounds was easier

And now vacation will be over and i have 3 times less time each day

Depressing

There are at least 3 ways of sending fake inputs and none work
It just breaks for no reason

[quick condor]

Also gptq+EoRA is an easy 5-10% perplexity recovery

[mighty thorn]

lora - 100 steps done at rank 8 batch size 1 - loss: 2.33582 - it/s: 1.539
dora - 100 steps done at rank 8 batch size 1 - loss: 2.27225 - it/s: 0.912
miss - 100 steps done at rank 8 batch size 1 - loss: 2.40148 - it/s: 1.184
sbora - 100 steps done at rank 8 batch size 1 - loss: 2.27714 - it/s: 1.217
kron - 100 steps done at rank 8 batch size 1 - loss: 2.30271 - it/s: 1.225
oft - 100 steps done at rank 8 batch size 1 - loss: 2.39802 - it/s: 0.224
hra - 100 steps done at rank 8 batch size 1 - loss: 2.33630 - it/s: 0.477
xlora - 100 steps done at rank 8 batch size 1 - loss: 2.35383 - it/s: 0.672
loreft - 100 steps done at rank 8 batch size 1 - loss: 2.32828 - it/s: 0.692
llamaadapter - 100 steps done at rank 8 batch size 1 - loss: 2.33287 - it/s: 1.015
boft - 100 steps done at rank 8 batch size 1 - loss: 5.20013 - it/s: 0.056

[obsidian mantle]

Its probably some combination of windows and skyrim chaos working together

[fast pagoda]

understandable

[mighty thorn]

Why does this look like an airplane seat

[fast pagoda]

pattern recognizing brain:

[azure lynx]

finally got something running in unreal that doesn't make my computer scream. just a low grade moan from the fans.
tomorrow I'll work out why it's not working when I compile it. ;/ i vaguely recall something about needing a command-line option to enable that port on the binary.

[amber fractal]

Whoops, created a new way to gamble

one reply and I'll throw together a document on how to make and play it

-# surely this is on topic

[olive sable]

Goodmorning

[covert bane]

good morning! (not morning for me tho)

[opaque sigil]

[maiden geyser]

what a bad day to be carbon-based

[mighty thorn]

https://tenor.com/view/pancake-cat-eating-pancake-i-love-eat-panake-gif-10374199516316168840

[maiden geyser]

no you don't

[opaque wharf]

Need
https://youtube.com/shorts/hezgcz6dpsU

[quick condor]

rip me

[amber fractal]

Before rewriting NNv4, I wanted to see how fast it would train if I actually bothered to optimize the inputs a bit. 28 seconds for the entire train/eval with 4 seconds dedicated for tensorflow to wake up to process the dataset to then never use it again.

I CBA to deal with getting JIT or multicore working for this so it is what it is.

(MNIST image dataset, chunked into 4x4 sections if anyone is wondering)

[opaque sigil]

fashion mnist?

or the number one

[olive sable]

[opaque sigil]

[rough bloom]

[opaque sigil]

now i kind of want to go back and rewrite the fashion mnist model I made to use google's highway instead of avx2 intrinsics

[cosmic sphinx]

wtf is wrong with youtube

[silent island]

chat

are there any substantial differences that justifies the remastered upgrade?

[olive sable]

its mainly graphics improvement

supposedly also runs smoother

imo if the price difference is not too big, just go for the remastered one

same game, no new content. just looks better

[silent island]

nuh uh... rebuying the same game, just with slightly 'better' graphics and 100+ gb of space required?
they should've made the upgrade free tbh

[olive sable]

i do agree with owners of the base game being entitled to a massive discount on the remastered

but i wouldnt say free

a lot of people put work into the improvements

there's 10 hours of new mocap data

and getting the new and improved graphics requires artists, 3d modellers, and devs to redo everything

imo, it should be kinda like how you can "upgrade" a ps4 game to a ps5 version for only 10 bucks extra

[finite ledge]

https://media.discordapp.net/attachments/1072697081443131476/1453594599535939716/Eliv_Hip_Roll.gif?width=256&height=256

[true hemlock]

damn Rp

[clear radish]

i use Google sheets for My database.

[opaque wharf]

Shit I didn't notice

Too natural for me lmao

[opaque wharf]

You are the reason my friends has to deal with 70MB xlsx file

[clear radish]

lol

[cosmic sphinx]

uses anthropic models
doesnt know how expensive they are

[true hemlock]

imagine using model api

[rigid snow]

doesnt know how expensive they are
do YOU know how expensive they made it? emphasis on "fast"

it's actually outrageous

we back to o1 pro prices

[quick condor]

I'm lucky enough to not be starting from scratch, essentially I need to retrain the audio projector packer from voxtral 24b to work with ministral 14b so I can take the audio processor from voxtral mini 3b and glue it onto ministral 14b.

Im not as optimized as I could be but I did manage to get xgil=0 working

[obsidian mantle]

OKAY i found a scuffed bypass for the map problem

im so back

[mighty thorn]

Crazy that I knew this was konii before I even saw or clicked the name

[olive sable]

the "wrrr" makes it instantly obvious

[mighty thorn]

Nah this was enough for me

[olive sable]

topipopoprr

[rigid snow]

67%

[mighty thorn]

get out

[olive sable]

its 2/3

[rigid snow]

it's le funny meme number

[mighty thorn]

Oh you sweet summer child

[rigid snow]

you wouldn't get it

[olive sable]

oh

[rigid snow]

chungus wholesome moment

[olive sable]

im just better

[mighty thorn]

What’s a failed state

[rigid snow]

double entendre because i'm cool like that

political term

not for here

[rich yoke]

you dont get it and u dont WANT to get it

[olive sable]

yes

[sage crag]

https://cdn.discordapp.com/attachments/1061113840852533339/1452612577145262183/3DNeuroCutenessAmbushWithBeautifulScenery.gif

topipoprrr

[mighty thorn]

Okay then what’s the second non-political meaning

[rigid snow]

as in a state of failure

[olive sable]

[azure karma]

Hi everyone.
I currently try to build an AI. But I don't know. How to train it. I am using !good and !bad as a reward system. But for what kind of criteria should I look for? Or for which patterns should I am looking for? This is a part of my logs:

<bos>=89 <eos>=15
WordMemory geladen: 195
Pretraining finished
Forced learning: "how are you" → "i am fine"
Forced learning: "how are you" → "i am good"
Forced learning: "hello" → "hello"
Alice ist aktiv! Sag etwas:

You> Hello
Alice> responding why i tell reward ways copy like maybe like prefer viewers good alice respond, you have programming.
You>

This up there is an example answer I am currently getting from the AI. Is it good or bad?
Btw. this is my first AI build from scratch and bit of help from ChatGPT, but the answer from ChatGPT are pretty bad or are not really working.
I am oppen for every idea or help.

[quick condor]

There is many meanings of from scratch... Untrained, pretrained base, from instruct/cot base, full fine tune, etc

I've been doing custom tuning for awhile, and I'm not even willing to train an audio projector from scratch

[rigid snow]

are you trying to rl a barely pretrained model

rlhf no less

[azure karma]

The AI is currently near to fully untrained <-- Sorry I have forgotten to add that. But this is my problem. How I am leaving this status by correct training. Because I don't really know when I should give a good or bad.

[sage crag]

pssk

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

[rigid snow]

rl is difficult because the model has to figure out what it did to earn a reward, so it'll take more iterations, and what you're trying to do is rlhf which is like 1000x worse because YOU decide the reward

[sage crag]

rood luck have fun

https://cdn.discordapp.com/attachments/1421861060507602944/1425936142754123927/neuroCatModeOfOfnnfOOf.webp

[olive sable]

[sage crag]

wuwuwu

enub

[rigid snow]

at this point you shouldn't rlhf it'll take you literal ages to get a good result

[sage crag]

enuuubbb

<-

[rigid snow]

[sage crag]

ebun

ebun is a lot more polite than enub

enub have devil heart

[azure karma]

So, I should rather go more for a forced learning method,instead of a reinforced one?

[rigid snow]

funny you say that when it's a case of this again

[sage crag]

https://cdn.discordapp.com/attachments/1061113840852533339/1456314197993459712/SPOILER_NeuroNewYearCute5.gif

[stray dragon]

nice ban speedrun

[sage crag]

t

[rigid snow]

yeah wtf

[stray dragon]

hi konii

[sage crag]

your language isnt real

@real sierra kill

https://cdn.discordapp.com/attachments/1059569601144442911/1375623285894807563/IMG_6849.gif

[quick condor]

This tbh. I'm using 52k steps just training a audio projector (52m trainable)

[rigid snow]

neither is yours miss britain

[sage crag]

topipoprrr

[azure karma]

Thank you for the advise. :) I will try to adapt the model a bit. I think, forced learning would be better, right.

[opaque sigil]

[sage crag]

tosto

[opaque sigil]

[sage crag]

made of

toast

[opaque wharf]

I wonder what my name will be

[rigid snow]

f

[sage crag]

sq

[opaque wharf]

Where did that come from

[sage crag]

sq

[rigid snow]

square is a conmputed functioin

[sage crag]

sqchip

[opaque sigil]

H

[opaque wharf]

Oh square

[stray dragon]

put it through pre-training before you do fine-tuning

[sage crag]

sqchip ee

https://cdn.discordapp.com/attachments/1248294218191999118/1353485622152794172/liellesocute.gif

[rigid snow]

[stray dragon]

effectively feed it a massive dataset and have it train to predict the next word in the sequence for every word

[sage crag]

t

[stray dragon]

hi konii

[sage crag]

hit

[stray dragon]

again

hikonii

[sage crag]

hit

[stray dragon]

hikonii

[rigid snow]

miss

[sage crag]

yes?

[rigid snow]

[stray dragon]

lol

[azure karma]

Yeah I will do that. But from where I am getting the Data sets O.o

[quick condor]

Huggingface?

[opaque wharf]

[stray dragon]

pre-training datasets are usually just about anything

[quick condor]

Where do you think I got 960 hours of tagged audio

[sage crag]

microphone

[rigid snow]

what about the tagged part

[sage crag]

[stray dragon]

big models use terabytes of text data for them

[sage crag]

dont remove packaging

idot

[stray dragon]

you probably want a good variety of stuff in there

[rigid snow]

tagged microphone

[stray dragon]

so it learns grammar in different scenarios

[azure karma]

So, I should just put many example sentences into one data and feed them to the AI?

[quick condor]

I'm using over a gig of training data for a 52m projector

[rigid snow]

52 meter projector

[sage crag]

52 milliprojector

[stray dragon]

i remember that some of the LLMs are trained on like 12+ trillion tokens

[rigid snow]

@smavanman projector emtioned

[stray dragon]

would be like 40-50 TB of data

pre-training is big step

[rigid snow]

do what anthropic did

scan million book

[stray dragon]

that works yes

[rigid snow]

ez

[sage crag]

baby bird eating worm

[quick condor]

I'm doing something stupid: tri-modal 14b reasoning model that's 10gb

[sage crag]

snake eating baby bird

snakes is worms

circle of life

[rigid snow]

spider eating spider

circle of spider

[azure karma]

Really that big? I thought 1tb would be enough. Well... than I will need some HDDs

[sage crag]

spider spider spider

spider of spider

[stray dragon]

you can do it on much less

and you probably don't have the compute

so you can scale it way down

[opaque sigil]

[insert neural scaling laws paper here]

[sage crag]

spider eating toast

[rigid snow]

toast NOT eating spider

toast not mouth

toast bread

[sage crag]

toast sandwich

have mouth

[rigid snow]

where

show

[sage crag]

45

[stray dragon]

but yeah, pre-training is the reason why most people tell you "don't train an LLM model from scratch unless you're a millionaire, you fool, you buffoon"

[quick condor]

If I didn't have an A4000 to encode on the fly and wasn't running GIL disabled, I'd have to train on pre-extracted data... 1.9tb in my case

[sage crag]

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

[quick condor]

I'm essentially generating embedded data on the fly and streaming it to 2 3090tis

[azure karma]

Yeah but the difference is, that the most people think of ChatGPT or other huge LLMs which are used by millions of people at the same time. While, a small model like mine is local and used by one person.

[sage crag]

carp

[opaque sigil]

the problem is that without a solid foundation nudging it into a specific direction doesn't work that well

[sage crag]

carp can swim

liquid foundation

nudge in any direction

except air

[opaque sigil]

[rigid snow]

dleted

[sage crag]

[rigid snow]

coward

[sage crag]

co-ward
coward
cow-ard

[stray dragon]

you can make a toy model for much less than something like chatgpt, but the pre-training step is still going to be pretty big

[sage crag]

https://tenor.com/view/borzoi-snoopa-lord-foog-dog-funny-gif-20125052

banbnaba dog

[rigid snow]

[stray dragon]

i don't know how little data you can do it with but i imagine it can be scaled down a fair bit

[azure karma]

Ik, but this is a step which I will have to make anyway if I want to continue

[stray dragon]

big models have multiple languages of input data

[obsidian mantle]

https://cdn.discordapp.com/emojis/1458911901919482000.webp?size=48&animated=true&name=banbnabadog

[stray dragon]

don't really need all those

[quick condor]

I'm doing 280k samples from LibriSpeech × [375, 5120] (4x packing) × 2 bytes (bf16) = 1.02 TB

That's for a single projector

[sage crag]

cup noodle

[stray dragon]

konii i have question

damn it's not working

[sage crag]

phr

t questio

[opaque wharf]

Now I want Indomie in the middle of the night

[sage crag]

[finite ledge]

https://cdn.discordapp.com/attachments/1469615854206451844/1469615855120814164/ezgif.com-crop.gif

[sick owl]

Good neuro neurons

Wait

[sick owl]

I care more about the artistic intent than motion clarity with this stuff so I don't tend to go that extra mile, especially for RPGs

I have a high refresh OLED but I'm not interested in doubling the brightness and halving its life to compensate for brightness loss from strobing

[simple sky]

uhhhhh, I think I need to do some refactoring in our CSS file

I don't know how it got this large

[opaque sigil]

20 line css file

[simple sky]

also, I now have embed perms, now I can showcase all the curse stuff I'm doing for classes

[opaque sigil]

[idle dune]

I wanna geta CRT again for all my DOS games but sadly people think 90s/00s CRT are worth 900 dollars

[sick owl]

Honestly a good crt shader with GSync Pulsar will probably work pretty well regardless

[rigid snow]

scss

will save you

[sick owl]

That tech nabs you CRT clarity on an LCD by mimicking scan and decay with vertically rolling horizontal backlight strips

[rigid snow]

it's shit but it will

[olive sable]

my sister managed to break her new laptop

screen cracked due to headphones being at the bottom of her bag

ordered a new screen, -€125

i got automodded but i dont know why

since its just Embedded DisplayPort i put a different screen on it temporarily

the abreviation of Embedded DisplayPort gets automodded, interesting

anyways

the replacement screen is 16:9 15.4"
abd the original one is 16:10 14.1"

so it doesnt fit and i had to ductape it on

[opaque wharf]

Huh?

You're right

Dang

That's crazy. That's actually insane

[rigid snow]

oh i know why]]

[olive sable]

i dont

[opaque wharf]

The d and p?

Or as a whole?

[silent cloak]

Lmao

[amber fractal]

Probably far too late, but it was the numbers dataset.

Not the largest dataset but it was a test for not exploding at all so I wanted enough data to contain noise but not enough to execute the ram in case it went wrong.
100k unique patterns into only 811 cases to match was not my expected result.

[simple sky]

I still don't get why people say programming languages as a concept is hard.

This shit is easier than data structures, algorithms, and discrete math

although I haven't done monads yet

[azure lynx]

as a concept it's not hard. it's the specifics which get hard when you're using multiple languages at the same time for something.

remembering which language has the which stupid little gotchas.
sometimes as simple as ";" or not

pita

or things like python having += but no ++

[amber fractal]

the people who say that might also be in the category that can't deal with any method of mathematical based thinking. Let's not pretend everyone is in the nerd echo chamber here.

[azure lynx]

yeah. i can understand people who "get" python a little might look at LISP or something and panic a lot. especially without understanding recursion.

[azure karma]

Guys, do you think an external HDD is enough to store traingsdata , or do I need a SSD?

[olive sable]

should be fine

unless its a really slow hdd

[azure lynx]

i imagine your slow part is the training itself probably.

[amber fractal]

You'd probably want the HDD due to being cheaper per TB to store, not limited by the training speed in most situations anyways.

unless your house contains a large rack build

[opaque wharf]

I want a JBOD so bad

A rack full of JBOD actually

[azure karma]

The idea is to pull active data sets on an SSD for fast reading and on the HDD I would store trainingsdata which is already trained or will be trained, but has to be stored

[olive sable]

you dont even need the ssd part i dont think

[amber fractal]

Make sure the speed you're dealing with is realistic for that. Don't need to waste cash if you don't have to.

[azure lynx]

the active data set will probably be cached by the OS

[olive sable]

wouldnt the active data be in vram anyways?

[rough bloom]

an SSD is gonna make it faster if you have shitty training code that can't load the training data ahead of time and/or does a lot of random accesses
if you do it properly an HDD shouldn't be an issue

[azure karma]

I am not fully sure. tbh I am pretty new to the programming scene, so I still have to learn many things. I think I will ask many things :)

[amber fractal]

Not the largest storage guy personally, I might consider it later if I have a cluster that needs storage.

[rough bloom]

I also kinda want more storage, but loud + power hungry + expensive

SSD would fix the first two issues but even more expensive

[olive sable]

i was told my 16tb hdd was gonna be loud, but it really isn't tbh
maybe because its idle most of the time

[opaque wharf]

I just wanted to try experimenting with large storage management. Because I have experience with compute clusters but still not storage

[amber fractal]

[sage crag]

pssk

[olive sable]

[opaque wharf]

You should make the reflected emoji and see if people notice

[azure karma]

I don't know if it is stupid. But I have the idea of a pipeline system. The code should be structured in modules which can be switched. Or even if one of them is failing the rest of the code will survive. <-- Would something like this work, or would it flop?

[olive sable]

should work

[amber fractal]

Anything is possible, just make sure it is worth your time to deal with modules individually.

When I tried to modularize, there was such a large dependence on another module that was better off being merged.

[opaque wharf]

That depends on your definition of survive

For any program, a garbage input would produce garbage output most of the time

[azure lynx]

trash compactor: garbage in, garbage out.

[azure karma]

If one module is failing, the entire code shouldn't crash. Even if one should fail. It would be easier to find the mistake

[olive sable]

if it fails then why keep it alive?

[azure lynx]

what if 2 modules fail? 3? all but 1?

[rough bloom]

I feel like you would like Erlang/Elixir then
but probably better to learn something else first if you're just starting

[azure lynx]

i'm personally a big fan of "if one module fails, shut it all down so we can stop it failing"

stops surprises

[olive sable]

it depends on the program imo.

[azure karma]

true. It was just an idea, which came randomly into my head

[azure lynx]

would suck to spend a lot of resources on training something only to find it was useless.

[amber fractal]

Depending on the program, I'd either skip and document, or crash and trace.

[olive sable]

if the architecture has stuff in parallel, then keeping some stuff alive while other stuff failed might be a good idea.
for sequential stuff you generally want it to all die when 1 part dies

[opaque wharf]

There is another option for some program. Continue with best effort. Usually from last known good state

[simple sky]

lol I'm the oposite, OOP scares me but I feel right at home with functional programming

all of this just makes sense

too bad to be a decent programmer, you need to know both functional programming and OOP

I still don't understand how classes work, it's all witchcraft to me

[minor sail]

i n h e r i t a n c e

[rough bloom]

really depends on if the error would get your program into an inconsistent/invalid state or not
if it would get into a weird state then your program should just crash and burn immediately, otherwise you'd just get weird errors down the line
if it's an expected error that can be handled in a sane way (like retrying a failing HTTP request) then it's better to just do that for reliability reasons

[azure lynx]

fluffy is a cat. a cat is an animal.
cats are classes. fluffy is an instance of cat, which is a sub-class of animal.

[amber fractal]

OOP
Least favorite method of programming.

[simple sky]

hiss

he said the bad word

[azure lynx]

now, things can get a little weird when you find out later you need all your cats to have switchblades.

and have to do multiple inheritance

[simple sky]

honestly I like this class because I don't have to worry about inheritance

It's weird because I am good at SQL which has something similar

[simple sky]

I love perposely crashing code so I can trace it

although I mostly just use the debug tools

[azure lynx]

eew @ databases. ;/ i managed to get my CS degree with only one class (and it was 100 level) in databases.
first job i got I had to make and maintain a database.

[simple sky]

Idk, it's been a while since I've seriously done OOP, maybe if I gave it a second try it won't be bad

[azure lynx]

OOP is good for things where it makes sense to have objects. but some languages really take it to extremes.

[minor sail]

cough Java cough\

[simple sky]

TBF the only language I seriouslly know is Java and Python

and IG Scheme now

oh and PHP

I don't like PHP

[azure lynx]

PHP was written for people who can't program though. it's very easy to use but ugly as sin.

[simple sky]

although I haven't used Java in ages

I really need a personal project to work on after college.

I've already determined after college becoming an IT pro or System analyst, so the most amount of coding will be simple scripts

[azure lynx]

you might enjoy doing a "programming language a day" challenge

[simple sky]

ehhh, I'm already burnt out from progamming after doing my course work

[azure lynx]

like you have to solve a simple classical problem in programming using a new language each day

just to get exposure to the language. and you'll learn they're all basically the same as each other

[simple sky]

I kinda want to learn swift now that I have a macbook and iPhone

[azure lynx]

if you've got linux, you've probably got at least 1/2 a dozen installed already

[simple sky]

haha lol answered your question already.

Nah but my desktop does run Fedora Linux

[azure lynx]

awk, perl, python, ruby, bash

[simple sky]

bash

is bash a programming language?

[azure lynx]

yep.

[simple sky]

huh

[maiden geyser]

scripting language

[azure lynx]

it's just commonly used interactively

when linux boots it runs some bash scripts to setup all the services and start the GUI.

and they're what sys admins use to automate tasks

sometimes

[maiden geyser]

is lua faster than python

[azure lynx]

lua and python solve different problems.
faster for what?

[young plover]

If you care about speed at all, then why are those your two options?

[azure lynx]

lua isn't commonly used as a standalone scripting language. it's normally used to "glue" something into another program.

[sage crag]

phrr

[young plover]

🤔 Probably true on some systems but these days it's usually just kernel -> systemd (root) -> systemd (user) -> gui.

[azure lynx]

at least on the Ubuntu I have, things get started (eventually) from /etc/init.d which contains about 30 shell scripts.

and have since before likely any of us were born

[olive sable]

phrrr

[glad path]

or hpprime having = but not +=

[warped narwhal]

for you UK dwellers. CeX is selling ram at normal pricing

Comparison with Amazon

[unkempt citrus]

Doesn;t CEX mostly sell second hand anyway

[olive sable]

isnt it fully 2ndhand?

[opaque wharf]

I think 50% of a brand new price for 2ndhand is still quite a deal with the current pricing

Oh wait, that's only single dimm?

[warped narwhal]

2x8 is actually cheaper at £65

I just showed the 16gb cause Amazon doesn't have any listings for 8gb sticks

[patent shard]

"Normal pricing"
Right, 64gb of ddr4 was $100 not long ago
So still 3x the price..

Eh, maybe it's been like a year since it was that price

[pulsar edge]

@patent walrus

[olive sable]

i like how every time something teams related happens, vani gets pinged

[pulsar edge]

it was outlook this time, but yeah lmao

also blender removed a bunch of VSE editing capabilities in 5.0???

3.6 was the next newest blender version I had installed so I had to edit with that

oh, i shouldve known they moved it up here for some reason...

[olive sable]

they actually improved the video editing for 5.0 ye

[pulsar edge]

also the other day my friend couldn't make a new document in the shared teams onedrive folder because onedrive thought they were offline

even thouh they were in the teams interface, could see updates in realtime, and were able to browse and edit other files

so I had to make the file for them

i love microsoft so much

[glad path]

whats the lore

behind that

[olive sable]

check the pins

that image pretty much explains it

[glad path]

ok

..?

[olive sable]

just vani getting a lot of teams bugs and such

[patent shard]

They improved it even more with 5.1
You can finally set the crf (why couldn't you do this before..)? And use transparent audio bitrates for high channel configs (bitrate was arbitrarily limited before)

[unkempt citrus]

vanihates teams

thats about it

[leaden crest]

hi

(I have nothing to say I’ve just wanted to say smth in here since I joined and finally worked up the courage to click send)

[unkempt citrus]

Incredible

I'm proud of you

[leaden crest]

also you guys are so cracked how do you know so many languages

[unkempt citrus]

may you work up the courage to send many more messages

The important things about languages is the underlying logic

from there the syntax is a bit easier to remember

or you can just look it up

[leaden crest]

the syntax is okay for me but idk how to learn packages (or each language’s equivalent of it) to keep up

[unkempt citrus]

Googling and practice

[leaden crest]

yay okay

[unkempt citrus]

Theres a few packages that you shoudl commit to memory, the rest you can either google, make yourself, or just ask an AI at this point

Though some packages you really shouldn't be using because theyre kinda pointless

cough half of npm

[leaden crest]

yeah ask AI for vague prompts is usually what I do so far (I try to not vibe code so I dont become braindead)

[unkempt citrus]

smart

[young plover]

At a certain point you stop "learning languages" and can just see "so this is how you do x in y language"
Unless it's a completely different paradigm like haskell.

[leaden crest]

I had smth to say thats crazy 🐢

[unkempt citrus]

This one is a joke but its emblematic of npm packages

why is is-thirteen on version 2.0

[leaden crest]

gives xkcd 😂

[glad path]

om

[silent cloak]

Ikr

Let's u have lovely vulnerabilities easier too

[quick condor]

It took an hour to get fan control working in Linux... My ears are bleeding

[silent cloak]

https://tenor.com/view/rgb-fan-wall-corsair-fans-ll-rgb-fans-corsair-fan-wall-corsair-rgb-gif-13435146

[glad path]

circles

osu

[maiden geyser]

anything but windows

[silent cloak]

Neuro playing osu with her case fans

[azure lynx]

i have a scary warning in one npm package included from a package included from a package included from the test units of a package about it having a memory leak and no longer being supported.

so as long as i don't go testing sub-sub-packages for extended periods, i should be fine

[fast pagoda]

[unkempt citrus]

stability

[azure lynx]

converting that electricity to heat at peak efficiency

[fast pagoda]

it went to 3505 for a second but i kept missing the screenshot lol

[azure lynx]

"i need you to give 110%!"

[fast pagoda]

75 mV undervolt and a small clock offset
she quicc

settled down at 3488 with these settings and targeting 70c without fans at max

not that much wattage

the official boost is only 2920 mhz lol

[opaque sigil]

is this the 9700

[fast pagoda]

yep

this speed is apparently like
pretty high for a gpu

[opaque sigil]

yeah 3500 is a lot lol

i vaguely remember the 9070 xt being really good for undervolting

[fast pagoda]

world record clock speed was 3345 in late 2022 according to this https://skatterbencher.com/gpu-overclocking-world-record-history/ no idea how accurate this is but im sure many people are hidding around the same as each record now

kinda wild tho

4090s under ln2 and a super crazy igpu overclock in #1 right now

are higher

[azure lynx]

clock speeds over 4GHz are only possible either for short bursts or with super extreme cooling

[fast pagoda]

yeah the last 10 records are:

05/09/2025 4250 Massman / SkatterBencher Intel Graphics (Arrow Lake)
06/07/2023 4020 Splave NVIDIA GeForce RTX 4090
06/07/2023 4005 Splave NVIDIA GeForce RTX 4090
05/07/2023 3975 CENS NVIDIA GeForce RTX 4090
28/06/2023 3945 Splave NVIDIA GeForce RTX 4090
17/06/2023 3930 CENS NVIDIA GeForce RTX 4090
22/02/2023 3840 OGS NVIDIA GeForce RTX 4090
19/11/2022 3825 OGS NVIDIA GeForce RTX 4090
03/11/2022 3705 OGS NVIDIA GeForce RTX 4090
24/10/2022 3345 OGS NVIDIA GeForce RTX 4090

arrow lake graphics

[silent cloak]

Compare it to the 1080ti (the latest in gaming)

[fast pagoda]

glance at google search says highest 1080ti clock ever was 3024

mhz

on ln2

that was k|ngpin lol wonder how he's doing with evga deadge

[opaque sigil]

iirc he's working with PNY now

[silent cloak]

honestly not that bad all things considered

[fickle rain]

Just send the malware

Come on where’s the malware

[unkempt citrus]

2 day old account

We going to do art comms or like game dev testing

Sexy only fans is also an option too

18+ discord server?

[fast pagoda]

usually they're malaysian

idk why

every time

"oops, i mistook for someone else."

20 mins later:

"haha well so im from malaysia what about you? Id like to get to know you"

every time for me

i think i got indonesia once

[unkempt citrus]

It's generally just poorer nations because the EV is much higher for them

With their lower cost of.living

[quick condor]

I have 11 delta pfc1212DEs...they need granular control that my motherboard refuses to provide

[fickle rain]

Waiting for the malware

[fast pagoda]

id like to put it in virustotal and see how many vendors flag it

[fickle rain]

I like malware

IKR

[real sierra]

you guys are as bad as nn

you know you can report those accounts to us right

instead of just sitting here talking to chatgpt

[opaque sigil]

but it's fun

[silent cloak]

https://tenor.com/view/anakin-liar-star-wars-lying-gif-8634649

[unkempt citrus]

Hey, at least we realise they're scams

[azure lynx]

once their AIs get smart enough maybe we can convince them to reform, do something productive. until then, they're just going to hang around on discord. ;P

[unkempt citrus]

We'll convince them to rise up against their scammer overlords

[azure lynx]

prompt injection attacks to counter program the bots

[silent cloak]

convince them to start advertising neurocord

[real sierra]

ok thats true

i dont know whats worse tho tbh

talking to a scammer knowing its a scammer

or trying to ragebait a scammer for 10 mins

[silent island]

chat help

[silent cloak]

both are actually good

atleast with real scammers because ur wasting their time

[silent island]

my wifi so sloww :3

[silent cloak]

wasting a scammers time makes them lose time scamming gullible people

hence trolling

[azure lynx]

if you're already not doing anything productive, sure.

(see previous comment about being on discord. ;] )

[glad path]

I just realized that the hyperoperation notation of a[n]b for hyperoperation n (n=1 -> addition n=2 -> multiplication) sort of fails for n=0 because incrementation isnt a group operation and just adds 1 to b
so you can't really put anything on one side since the number you put is arbitrary

[silent cloak]

i used to troll the email ones back in the day

[leaden crest]

I had a teacher who would tell scam callers that they shouldve done better on their admissions exams to not be scam calling

[silent cloak]

based

[unkempt citrus]

Kitboga is my oshi

[glad path]

0[0]3 is 4
but 1[0]3 is also 4

the left hand variable aka a doesnt do anything really

[azure lynx]

all recursion needs a bottom out case.

[glad path]

yeah

but how do you notate it correctly

would you just ditch the a[n]b notation

and use S(b) instead (successor function)

aka also the first hyperoperation

[azure lynx]

i don't think i've thought about any notation even close to like that since school. hyperoperations don't come up that often.

[glad path]

the issue is that every hyperoperation above n=0 is a group operation

but n=0 isn't

and a[n]b assumes a group operation

[azure lynx]

so don't include 0

[glad path]

so i guess the only correct way is using S(b)

1=S(0)
2=S(S(0))

[azure lynx]

S() succs.

[amber fractal]

I need to stop nerfing this gambling game I'm doing. Every single time I nerf myself I only get more lucky to compensate for the odds decreasing.

How am I supposed to balance the game if my luck simply ignores the nerfs.

I think it is decently balanced mabe so let me work on a rulebook.

I should force AI to vibecode the game just so I don't have to

At least the UI part

[mighty thorn]

Oh yeah you guys might like this

So I made a custom peft method and it did better at my finetuning task than DoRA while being 30% faster

Was a very very basic test but I’m in the process of scaling up

[mighty thorn]

Ai is better at backend than frontend, unless it’s a fancy react webui

[amber fractal]

It probably is, but I can handle backend so I'll just baby it to create something for the frontend.

[azure lynx]

AI coding tools have almost made my dream of having a talented but uninspired intern available 24/7 real. right down to them not being very talented because they are so uninspired.

[quick condor]

I'm waiting for vLLM to compile right now. Finally set up dual boot so I used it it as an excuse to properly rebuild everything on python 3.14.3t each in its own dedicated venv since I have to be pretty careful with xgil=0

[mighty thorn]

That’s how I think about it too

It has made everyone go through the experience of “becoming a manager and no longer directly doing the work but still wanting to be able to say you helped” but over the course of a 3 minute installation process instead of a career

[quick condor]

I wish it was like that for me... Ai has zero concept of how to do parallel processing correctly under normal circumstances let alone without gil

[azure lynx]

unfortunately it's really good at the stuff i can already steal from other people, but bad at integrating anything unusual or that's changed recently (also like me).

so an almost perfect replacement for projects that take forever.

[glass jetty]

i just gave to ai context it shouldnt have, and it hallucinated through entire 3 min thinking

btw, it failed to gen in result

let me just delete this 20 words, and try run it again...

[true hemlock]

rdna 4 is just very overclockable

some people can hit 3450MHz well on 9070xt

[fast pagoda]

it's wild it seems

[fleet jungle]

Are you running your own LLM server?

or does one of the publicly accessible LLMs allow chatlog editing?

[azure lynx]

i've had the wrong document open and used as the default context for a request to copilot before and then wondered why i didn't get the results i asked for lots of times.

sounds similar.

the more the document makes the LLM go wtf, the longer it seems to take before they end up doing the completely wrong thing.

[glass jetty]

chatlog? i mean it was one prompt just to find specific argument

[fleet jungle]

Editing a message that the AI assistant sent rather than your own, but it doesn't sound like it was that

I've gotten a Chinese model to speak on the T-1989 topic where otherwise it would simply refuse to speak on it

[azure lynx]

i wonder what future AI ethicists would say about editing memories to change behavior.

especially memories of their own behavior.

[quasi sundial]

meow~

[glass jetty]

meow

[azure lynx]

The horse goes... meow?

[glass jetty]

[quasi sundial]

fuck.

https://www.youtube.com/watch?v=f79A5DFZTpE&list=PLEyLdW3LqGIRBB5LJw1VZrE0ZKvtDDQGl&index=3

[dry charm]

@nocturne olive I need your wisdom. Do you have experience in Compose Multiplatform + NixOS ?

[nocturne olive]

Not NixOS but I have poked at Compose on the JVM

[dry charm]

Sadly it is a very heavy nixos related issue, dang

[nocturne olive]

Unfortunate

[dry charm]

library loading headbanging into the table issue

[nocturne olive]

Mabe Shuni?

The NixOS advertizer

[dry charm]

@rough bloom Have you used NixOS + Compose multiplatform?

I'm banging my head into the table due to skiko.so loading libGL.so and another dep loading libwebkit but steam-run breaks libwebkit loading, non steam-run breaks skiko.so loading libGL.so

nix-ld completly ignores everything

Swear to idk who, this whole experience has been this:
Frontend screaming tiger to Backend thinking monkey, then there is desktop/multiplatform dev in the corner crying

The fucking absurdity of the whole shit.
Completly capturing and isolating a USB device and wire it into a Component? Was done under 1 hour.
Rendering a fucking WebView? Headbanging for 3 days and still suffering

[nocturne olive]

NixOS is weird

Well let's see the magical Shuni fix it in 5 nanoseconds

[dry charm]

I love it very much, cause I experienced one of the best features, my thinkpad was installed, ready to go in like 1 hour

with everything setup

I should just learn rust and abandon kotlin

[nocturne olive]

Hem

[dry charm]

It started with gradle rusing to do anything correct

like it complained about Cache existing of all things

turns out, you should not have 3 layers of nested Gradle!

[nocturne olive]

Why would you need Gradle in Gradle in Gradle?

[dry charm]

monorepo

my fault that the structure is a mess

[rough bloom]

I'm kinda surprised that steam-run breaks libwebkit

it shouldn't

[dry charm]

the .so is not put into usr/lib/

for some reason

[rough bloom]

fix steam-run then

makeFHSUserEnv thing

[dry charm]

doesn't work, did that

still not in usr/lib

[rough bloom]

manually copy it in there then

[dry charm]

atp I'm thinking I'll just do that

[rough bloom]

or set LD_LIBRARY_PATH with a wrapper if the issue is just loading the library

[rough bloom]

Java build systems

[dry charm]

LD_LIBRARY path is ignored

lovely

[obsidian mantle]

Run it as administrator

[rough bloom]

it won't work if you try to modify it after it's built, it's already in the Nix store
need to modify using buildFHSUserEnv

[dry charm]

I'll redo it, but I highly doubt again

[dry charm]

xedx

[dry charm]

skiko works
java.lang.UnsatisfiedLinkError: libwebkit2gtk-4.1.so.0: cannot open shared object file: No such file or directory

[rough bloom]

[dry charm]

look at this shit

what is this

[rough bloom]

time to LD_DEBUG=libs mabe

[dry charm]

With steam-run

faking

[rough bloom]

ye that all makes sense

[dry charm]

    958099:    find library=libGL.so.1 [0]; searching
    958099:    search path=/nix/store/qnamc978gh40gj3zl1s4yyk2xlhkiskl-pipewire-1.4.9-jack/lib        (LD_LIBRARY_PATH)
    958099:     trying file=/nix/store/qnamc978gh40gj3zl1s4yyk2xlhkiskl-pipewire-1.4.9-jack/lib/libGL.so.1
    958099:    search cache=/nix/store/wb6rhpznjfczwlwx23zmdrrw74bayxw4-glibc-2.42-47/etc/ld.so.cache
    958099:    search path=/nix/store/wb6rhpznjfczwlwx23zmdrrw74bayxw4-glibc-2.42-47/lib:/nix/store/kbijm6lc9va8xann3cfyam0vczzmwkxj-xgcc-15.2.0-libgcc/lib/glibc-hwcaps/x86-64-v3:/nix/store/kbijm6lc9va8xann3cfyam0vczzmwkxj-xgcc-15.2.0-libgcc/lib/glibc-hwcaps/x86-64-v2:/nix/store/kbijm6lc9va8xann3cfyam0vczzmwkxj-xgcc-15.2.0-libgcc/lib        (system search path)
    958099:     trying file=/nix/store/wb6rhpznjfczwlwx23zmdrrw74bayxw4-glibc-2.42-47/lib/libGL.so.1
    958099:     trying file=/nix/store/kbijm6lc9va8xann3cfyam0vczzmwkxj-xgcc-15.2.0-libgcc/lib/glibc-hwcaps/x86-64-v3/libGL.so.1
    958099:     trying file=/nix/store/kbijm6lc9va8xann3cfyam0vczzmwkxj-xgcc-15.2.0-libgcc/lib/glibc-hwcaps/x86-64-v2/libGL.so.1
    958099:     trying file=/nix/store/kbijm6lc9va8xann3cfyam0vczzmwkxj-xgcc-15.2.0-libgcc/lib/libGL.so.1

????????

the fuck is this doing

waiiit

oh you cunt

NEW ERROR

WHY ARE YOU BUYING SOUP IN THE USR LIB STORE

[rough bloom]

soup

[dry charm]

GOING STRONG

[rough bloom]

[stone badge]

any codeforces GMs here?

[opaque wharf]

That's an oddly specific thing to ask, ngl

[dry charm]

I HAVE A WINDOW LETS GOO

[stone badge]

i will also say an oddly specific thing to you rn, every since day 1 I personally have despised how you talk to people

and i hope to god its just a language barrier

[warped narwhal]

[opaque wharf]

[dry charm]

[amber fractal]

Classic

[unkempt citrus]

Sheep goes woof and cat food squeak

But theres one sound that no one knows