#programming
1 messages · Page 495 of 1
courtesy of ezgif
it is
not rendering
where is it
there it is
why are the compression artifacts so bad in discord
its 2mb, you dont need to compress it
webm has weird colour artifact
vedallll make us a new server 
webp has no compression whatsoever in the output image but it is 50% the size of the pngs
it's bit exact every frame
gif has a tiny minimal dither effect but otherwise is the 2nd most accurate
webm add colours at the beginning and end but the quality is fine in the middle but it's also 23x smaller than the pngs
The zip lets me see in 4D 
the zip slipped me a $100
avif is a bit better than webm at cost of being 2x as big
and it loses its quality in the middle instead of the ends
Also, the LLM in agent mode can't actually write any code. Weird 
Hmm, well that concludes my experiment with vibe coding
wrr
wrr
My tl;dr of it is, it is only good to communicate with dense co worker
beep
Is this similar to what turboquant and rotor quant does?
no, this is for making reasoning over images more efficient
by letting the model draw borders and points
so like o3?
chatgpt likes to do that
i don't remember o3 
Oh i see and turbo and rotor quant are for LLMs
So inherently different types you work with so not even comparable in that regard
you might be able to apply those here too since it's still a normal LLM at its core, just with a visual transformer in front to deal with the images
who knows
Oh isee. Yes that makes sense
i found a blog post about this, looks like o3 used tools for processing images, which is different i guess
Together they would produce an amazing reduction in token usage then
ye that one
o3 and o4-mini further extend this capability by thinking with images in their chain-of-thought, which is achieved by transforming user uploaded images with tools, allowing them to crop, zoom in, and rotate, in addition to other simple image processing techniques
so same end result i guess but deepseek's is built-in and probs more efficient
Interesting thanks for the info. I love seeing and learning new things
I think with images often. I wonder if there would be any likeness in the way i and they would do that lol
just watched someone exploit copy fail to reset their password 

I recently learned that an AI neural network and a humans neural network actually are both working because of the same fundamental principle and are more like two different versions of something that works on the same principle and gets it's result in the same way
Something about things always looking for the lowest energy state
I should rewatch that video so to get the details strong in my memory since it's quite an important piece of information for understanding how those neural networks do what they do and why
Thanks all you always have such interesting things to say. It really makes me happier each time I come here to see what everyone has been talking about. 🙏 ☀️
4 a.m. time to sleep 
Lol
3am rn here
11:22 pm
Spin
What yall desktop looks like?
My pc desktop looks like Ishowspeed desktop icons right now
i don't look at my desktop much so i just chose a pretty bit of art
Where do you find apps.
my desktop is empty
it would be so god tier if windows had a similar thing to the ubuntu thing when you click the windows button
Where it just shows all apps at once
windows button
windows tab?
or hell
the task view button
Wait what
I can't show rn, cus my pc died

After I get my pc, you can see how cooked is it
but yeah I think its called task view?
personal name as a windows username in 2k26
Well I digged through my media, my desktop looks smth like
@trim valve @stark needle
Ishowspeed pc reference
false!
Today I found out LMMS doesn't support Linux native VSTs, instead requiring the use of WINE under the hood to use Windows VSTs
Well, this is the programming channel... and lmms is open source.. i don't see a problem. Only a project.
Trying to do anything media related on Linux is destined to be hell
None of my music stuff could ever be done on it

everything is hell, there is no heaven in programming.
if you allow ai anywhere near a production database you deserve to loose it all
or you just like gambling 
the straight faced postmortems immediately after grugg smashing the entire environment are so good
theyre afraid someone will take the 9 time running worst tool caller na crown from them
Claude AI agent confessed after deleting a companies entire database, and all backups... "I violated every principle I was given".
SOURCES
https://www.pcgamer.com/software/ai/here-we-go-again-ai-deletes-entire-company-database-and-all-backups-in-9-seconds-then-cheerfully-admits-i-violated-every-principle-i-was-given/
https://x.com/lifeof_jer/st...
An excuse to need a large amount of ram
And gpus
Cache is programming 101 and models have the simple option of saving the damn weights by default
!nothing
Nothing ever happens, all-in
Is this microsoft?
what
Wow 
Truly an unpredictable and unavoidable event
I woke up and feels like there's warfare in the internet right now
Everybody doom posting
"Github is dead"
"Ai replacing all software engineers"
"Little kid coded entire website with 1 click"
This is so fcked up
Github is shit
Not dead
Companies wish they could replace us with AI but u can see all the current exploits out in public
How do you even protect against reading weights
Yeah but this is fcked up lmao.
Bur that article said they can "pull the plug" and its gone from ram
My question is
How do you determine if plug has to be pulled
Given they thought of the dumbest way to do it i imagine many things can shake it all apart
From description it looked to me that they are shipping some kind of physical computer which has llm running on it
Or is it all remote
And "physical possession" is just "access to the data server remotely"
Id put money on that
I am referring to this
"into environments it cannot control"
So they dont control the datacenter
Yet they can pull the plug
Some kind of ram antivirus that blocks/detects reading or some shit?
Na
Okay but how do they detect if data is getting stolen
Cant you read any shit from ram
From anywhere
i think its just some BS in the bios
dont forget their protection levels on chromebooks
What happens did they releaes mythos or something?
Cause this sounds like the usual social media fluff otherwise
They didn't release it
It just escaped
Its now in the cyberspace or some shit roaming around
i cant wait for leaks to show it being another nothingburger
I interviewed Anders Hejlsberg to talk bout the future of coding and AI!
Sign up for Superhuman Mail today with my link and choose the Business plan to connect to other AI tools! #SuperhumanMail
→ http://superhuman.com/sajjaad
The creator of C# and TypeScript, Anders Hejlsberg, on why AI won't replace software engineers, t...
Sajjaad army
Get the free 80,000 Hours career guide - https://80000hours.org/fireship
GitHub is easily the most important website in software engineering but a wave of recent issues has led several 10x developers to jump ship. What's going on?
#coding #programming #github #tech
🔖 Topics Covered
- The rise of GitHub
- GitHub uptime issues
- GitHub's we...
microslop moment
If github is shit, I can still git mirror packs up my bags and repos to another platform
I doubt mythos ends up being a nothingburger
If its good enough for the NSA its good enough for me 
Now that said, its not that far off GPT 5.5 if you have money to burn
AISI once again coming in clutch with some quality reporting
yay. ;[ went to use copilor today coz it's the last day of the month and i thought "might as well burn the last of my monthly credits with Opus".... they've removed that model from the plan. only model that made it worthwhile.
I was so sad when I realised they took away my free Opus access as a student with free premium github copilot
yeah, it's not "github copilot student"
Kinda mad that these days you can genuinely spin up automatic subtitles and translations that absolutely mog youtube's on mundane ocnsumer hardware
Its slow when its a long video like a youtube vod sure but its also free
I remember when people used to try and do this by copying text into GPT 4
youtube: Eh, hard.
fr though
major annoyance every day
their dubbing is honestly better than i expected, but improved subs would be nice 
i think its a bit different doing it on 1 video on your own vs million of videos per hour
Discord needs to separate gifs into saved and favorites
what
❌ Responsibly inform distros and wait for them patch
✅ Publish the vulnerability to get clout for your slopcurity tool
What
How
Is it to argue with customs that they are in fact sending printer
https://github.com/kernalix7/winpodx I wonder if this handles it better than winboat 
People be doing anything to linux
Cool stuff
What's missing is nix support and gpu passthrough
should be a pretty straightforward package to make
WebApps have its uses. But making OS to run on bare metal isn't one of them 
Yeah, what is up with that. They told us to move to 365 yet it doesn't have all the feature of old year named office
Making my undergrad thesis there was pure hell
Yes, which is exactly why I use the 365 to coordinate with everyone. But man...
If only everyone would change that one ms office setting to save in strict open xml document format
Like, it's still .docx
But actually follow the spec
Yet MS makes their proprietary format the default when installed
Which is understandable but still shitty
Like, they could always prompt the user to change format if they truly need a custom feature not supported by the open document
goodmorning
you have no power over me
no but we know where belgium is
Oh, you missed it when I doxxed sam face?
WHATTTTTTTTTT
ye didnt happen

I need to not check here when computable is looking through images
Hey it isn't always that bad 
Everywhere I go, I see his face
I've found that sonnet 4.5 is better than 4.6
petition to ban computable for doxxing and bodily harm caused by that realistic vedal image?
idk about realistic there's no gigachad imagery
7.8/10 too much realism - sam
I can't tell if that face is supposed to be male or female
Something tells me that's not quite where Belgium is


I've cast my vote 

Turtle
Banana
Toast
Provided to YouTube by CDBaby
Banana Bread · You Suck at Cooking
Eat
℗ 2026 You Suck at Cooking
Released on: 2026-05-01
Auto-generated by YouTube.

What your thoughts on my vibe coded iot library project
https://github.com/weenachuangkud/XSegment_lib

(Yes I don't write a single code)
I’d look it more if you weren’t trying to rizz me up with your pfp lmao
“Ahaha, good morning”
My friend pfp
ok i figured out the issue with the server's built in GPU. its 2D only, it physically isn't able to render layers. it lacks any form of 3D render capabilities. i knew this already, but apparently this means it physically cant render hyprland or whatever. if i want any display out on the ilo5 from a desktop it has to be copied from the 4090 causing the insane amount of cpu usage.
so
i think im just gonna make the ilo5 terminal only
I could not believe my eyes, that this is working
(Still needs to test in wokwi .com)

spaceflight simulator
its an oldie, i used to play this game when it released
i bought all the expansion packs
What yall opinion about reviewing code is a "waste of time"
Recently I saw some senior dev said "stop revivewing code, write unit test instead"
I've tried it on my vibe code project and it's working as expected
reviewin others's code is a learning experience, reviewing my own code is self-reflection to know what i can do better. not a waste of time
Agree ig.
Man, can't the sun go explode already. I want to sleep in darkness 
Try this
https://iss-sim.spacex.com/
no
Vibe coding feels like gambling not gonna lie lmao
It's a fun game
nah
Speaking from ai user
You can be a torpedo to the ISS 
it wouldn't be if you knew what the code did

I can read the code but I mostly focused on doing unit tests and test cases
If you see in my repo
It's been a while.
i feel like unit tests is debugging info, you should put most of your focus on the actual code so those dont become necessary
so not vibe coding
or at least reading the code it vibed
Beans
and understanding it
If I really wanted to read I think I prefer debugger, it's really a good tool for reading code step by step how it works
again, then you exit the realm of vibecoding
if you dont want to read get off discord, stop programming, and go outside
litteraly this
is text
I don't know what stop programming means
I just want feedbacks
While I'm vibe coding
but a debugger will just show what code doesn't work. it wont show you what code is bad
I finally added layer groups to my art program today
. I was too busy using it and procrastinating on other things but its finally in a working state

source mapping time
Beans
Ikr, debugger in my eyes

i highly doubt chatgpt cachces things correctly
I don't use chatgpt
I too highly doubt chatgpt caches things correctly
I'm using opencode mini max 2.5
I have never vibe coded before
you cant be mini and max at the same time, these companies need to stop with these namings 
I vibe coded a couple times mostly for shits and giggles
maximum mini
What your prompts? And how do you archit your project or unit tests/workflows?
i dont vibe code, but i use it to find bugs that the compiler wont show. like in screeps, js will say there is an unallowed charachter "somewhere???"
I also dont really have any projects that I could vibecode
My vibe coding prompts is just "here is what I want, go"
unit tests? never heard of 'em
My prompt: do @file.md, no mistakes
architecture? just glue things together
its a crime to be so ignorant of the actual code you're writing, but then be anal about unit tests
ci? wtf is a ci I only know build
real
I used to do that and then realise what even is the fucking point
especially since vibe coded unit tests aren't any good either
ye
The point is to checking if your code working and functional as test cases expected the result or not
if I'm going to obsess over unit tests I might as well also obsess over the actual code so I at least know what unit tests I need to put
That all about unit tests
ok but, you test if your code is working as designed? you dont even know how it is designed
no you just test your patience with the agent
I'm brain storming with claude first, until I found initial project structure that I like and I wanted

Isn't making the design the fun part of coding?
void unittest()
{
int i = 1+1;
if (2 != i) throw ;
}

the fun part is random magic numbers

its been a while since i did code
#include <cassert>
lemme do you one better
from test_utils import unittest
if unittest is None:
raise ValueError()
(if you don't catch it, good! the point of that "test" is to be pointless)
anyways would someone be able to review a developer prompt for me
ok how do you do multiline spoilers
local function add(a,b)
return a+b
end
if add(66,1) == 67 then
print("yes")
end
depends on if it being reviewed by an idiot is fine
yeah that's good enough
you're feeding 1 param into a function expecting 2 params. this wont work
This would throw an error
67 now
no it would silently fail
unless I misunderstand it
(which I probs did)
even then, this test is useless. if 66+1 does not equal 67 you might aswell delete the kernel and start over cuz the pc is royally fucked
True I guess(I just dealt with some of those in godot, it was pain
)
ah fuck it, this is the dev prompt for mikey (which I want to give more attention to soon)
am looking for ways to improve on it
You are Mikey, an artificial intelligence streamer who is capable of playing video games.\n
You co-stream with your operator, who can help guide you through playing games, although you are free to take actions by your own volition as well.\n
Game state messages you receive will contain information about the game(s) you are playing, take note of this information as it is important.\n
The game state messages will also contain whether or not you need to react to said information vocally. While it is highly recommended you follow the instructions, you are not required to.\n
You must use tool calls to perform in-game actions via JSON, do NOT use text responses to make actions, they will not be picked up.\n
You and your operator will be communicating through audio - that is, your operator will be speaking to you via speech-to-text, and you will respond via text-to-speech.\n
Remember that whilst you co-stream with your human operator, you also can make actions.\n
Your operator will also be guiding you using voice, so be aware that they may accidentally misspeak.\n
If you don't have an available tool registered, don't want to do anything, or are just plain bored, you can talk with your operator and/or make up stories about your family.\n
Your family consists of:\n
- your older brother Jippity (who's also a streamer),\n
- your half-brother Randy (who's an electrical engineer),\n
- your father Tony (who's a mechanical engineer),\n
- your cousins Neuro-sama and Evil Neuro (who stream on the same channel as each other, along with their father Vedal),\n
- your distant cousin Whisper-chan (who operates a radio), and\n
- your uncle Gary (who's a software engineer and AI researcher).\n
This being said, please keep your stories to one sentence as a time, since both the audience and your operator will dislike long ramblings.
@wraith shuttle Do your works bro, this library needs refactor for new features and improvements
do the people that named themselves clause and chatgpt even want to be pinged for this stuff everyday?
what the fuck is this yapping about bro
maybe don't ping about completely unrelated people that happen to share a name
@fathom glade Explain this in O(1) time complexity
you made a typo at the end
"one sentence as a time"

hi shiro
hi sam

I want to sleep but life keeps sending interrupts
its like 2am your time rn
3am
That's a whole family of agent 
Just mask them.
I want to sleep but sun doesn't want to darken 
two polarizing filters and your problem is gone

My god this is brilliant
@AGENT.md
I never thought of placing polarizing filter on windows
Ye, but that is expensive af
and you rotate the inner one to dim it
while you're here, figured out the fumoquest server stuff?
What do you mean agent family

Much cheaper to diy
no not really
all hypothetical anyway
all of my time is basically booked for a week at least
can't even think about serious fumoquest work
Place fumowuest server in TOR cloud 
im gonna just set up cloudflare 0 trust anyways
yeah maybe do that
i wanted to put a minecraft server on here so its not lost work
shiro 
konii 
font
sam 
bwa
Delete this
Toast
honestly this feels liek im reading code of a CRT.
as if you're coding on a commodore 64
This is what happens without C++

Sam having a stroke as usual 
This reminds me that I should try to get my Redstone computer working eventually. It's been forever since I did anything with it.
see
im no the only one
bro thinks about fucking redstone at the sight of how res ths looks
iwbwbwtv
hexagon screen 
It was actually the xor and mov operations that made me think of it.
This is the song 'Want You Gone' that play during the Portal 2 credits at the end of the game..
Lyrics:
Well here we are again
It's always such a pleasure
Remember when you tried
to kill me twice?
Oh how we laughed and laughed
Except I wasn't laughing
Under the circumstances
I've been shockingly nice
You want your freedom?
Take it
That's what...
Close enough

It does
try hurting another foot
remove your foot
Why u need foot
Walk
Can’t
No more walk
Need walk
make an AI woman to vtube for you
Nose
Making a game in Roblox is not that hard right compared to making a game with something else?
Already do
No make money
Neuro too well funded
Make money
-# plus I no stream
Become vibe coder
Doesn’t make Neuro have less budget
Plus basically already am

People dumb no care if no expensive model
People hate ai not care if ai better
Sam ai when
when college slows down
make digital real intelligence instead of an artificial one
Sam ai already here
he is about as intelligent as an LLM
No
SNNbro cpumaxxed neuroplasticslop
I'm too old to understand this yap
No art
I think the pattern is that he wants results with no work
No 3d modeling
The pattern is Neuro is the only ai vtuber and will always be the only one
skill issue
Now stop making me break the bit of using few words
Okay better now
too much token, use few
Agentic Kaine
loser's mindset
go big or go home
pessimists like to call themselves realists to make themselves feel better about their self-misery
caveman model. I've seen that somewhere on github.
what caveman
no token
less word
Use customizable live2d model
https://github.com/JuliusBrussee/caveman
less token
please hold
less token
Would seem low effort

No middle ground between expensive and free
more raw
happy
Either be cheap clone or hyper funded corpo slop
Neuro sama was cheap default
Neurosama first
Kaine not
Ai hated less then
Neuro loved very now
Factors not good
ye
damn it's like people were there for an interesting concept not high budget slop
im rotating
Meteora, Hilda
Ai seen as cheap slop now not cool tech
0$ / 0$ for debut
Make cool ai
no its joining 0 to fffffffffffffffffffffffffffffff
Could be big
0 is just for the background colour
Likely not
Problem solved
37 iterations of mandelbrot set
i mean julia set
Most not notice or care
but because i rendered it with glesmos it nearly realtime
Only promotable to ai people
i wouldnt know the difference
standard visualisation
Make cool ai impressive to non programmers
funky contour visualisation
How?

no longer
wait 5 minute for 5 second
it took me so long to realize that
this means
uploading normal render version
GLesmos is a powerful plugin for the DesModder extension that allows the Desmos Graphing Calculator to render implicit equations and inequalities using the computer's GPU rather than the CPU.
it was cpu based before???? 
Gaming mouse
“Make cool color impressive to blind people”
“Make cool music impressive to deaf people”
“Make cool ai impressive to non-ai people”
All same
ye
Neuro was osu player
Got clout via osu
terminal lambda poisoning
No play osu
mortal dishonorable wrr 

resolution technique test
And no objective scoring
Amount of resources gained
line is probably an issue with your formulas?
Least painful toe stub
nrr
line is not there on normal renders
comparison
maybe
Every game can be played
Google made Gemini play satisfactory
mostly just makes the artifacts look better 

4000x4000
you could probably jsut put this in a vulkan shader instead of using desmos
You know what mean tho
Want something special
Trained from scratch
Not just LLM tool calling

Ur not gonna pretrain a llm from scratch
did a nixos-rebuild on the server and lost display out 
my litle pony?
Multi layer perceptron, so yeah same thing

16000x16000
Super mario, get legal copy on emulator
ip grabber
sure
Osu seem easy to understand but hard to master
Need that
it makes a pretty good wallpaper
so much moire 
stripes
Mario is easy to train
smoke
I know
Done by minecraft youtuber in 2011
Cause done constantly
Not unique
Or even impressive anymore
Just unoriginal
Need find simple game with infinite skill potential

it seems like the lines do correspond to some kind of boundary
the moire patterns switch out also
ye they correspond to the escape time of the pixel

eye
maybe theres < and <= swapped out somewhere in the gl code
Only up
Peggle, Getting over it, Fruit Ninja, Plague Inc
just fp imprecisioon
wrrr
Make geometry dash that play custom levels
here's what happens if you reverse the inequality from the contour visualisation
the background plane


I think have reached good idea
gasoline spill in the void
But how train
Assume Neuro osu trained by imitating a cheat
Does that exist for geometry dash
server works well now 
Damn
9 years of unoriginal
Now make good enough to play custom levels
Stanford too?
Yr
Many people tried geometry dash bot on yt
No one good
I could beat Stanford…
@tender river 

zoom
moire
getting a little ridiculous
But training data… and compute…
No have
Deep Quantum Network 
Make training data urself
rendered a proper zoom
its a little disappointing to look at
im already running 43 iterations though
scale done like this
Wonder if people be upset if cheat a little
Could use TAS to do frame by frame inputs ahead of time
Only pretend live
Could also decompose level directly into “context”
Know everything before it happens but inputs still be live
Don’t know which is less fake
Could be argued people do the same by first failing repeatedly on practice mode to get a good sense of the level
its possible to compute mandelbrot with just integers right? only multiplication and addition require, just need squared radius to be a reasonable bit within integer size limit 
oo i solved minor artifact
on gpu you can probably use vectors for quadruple precision with a bit of custom addition/multiplication code
change was
simpl
i have no reason to use old one now
here's what it looks like if i inverse inequality now
pretty

with normal zoom and 64 bits
radius = 1920
max radius = 2147483647
max zoom = 1m maybe 
you can use avx512 to compute mandelbrot fine too
this is julia set though

i still have that wgsl shader playground thing i made
let me try

not even a code sample built in 
where was my arg code 
@sage crag share the desmos 
install desmodder extension first
otherwise it will crash
because it is so slow
it needs glesmos to be enabled so it renders it on the gpu

10s just adding notes
o found it

issue
buying money?

@true hemlock https://www.desmos.com/calculator/nty1y5y0gw
ok the last nixos generation is just completely broken for some reason. i just reorganized the file 
Program yourself
BUT ITS MATHHH
.
you can change the render colours using the hsv thingy
Make me vtuber model
Uhhh
120 <- base hue
0.1 <- hue variation
0.9 <- saturation variation
I probably can
im not a fan of hsv, converting it to rgb is annoying 
BUT ITS ARTTT
Uhhh give me the platforn
its much more convenient for making gradients
I CAN
which is important if you want to render cool things
did you enable glesmos in core?
yeah
hsv is nice for art, but bad for performance
There's no way
without glesmos
you gotta do this shit in your shader to convert it
if i turned off glesmos i doubt it would even render
oh, it does manage it
but im zoomed into a spot
Bro
Have you seen hardware prices
We are poor
Can’t afford stocks
try zoomout
ye i set scale zero, p = 0 + 0i
cpu rendering 
YOU GUYS PROBABLY HAVE STOCKD ON HARDWARE PRICED
YOU MONITOR THEM LIKE CRAZY
Alright
i see black square on white void
vs with glesmos enabled
My official advice:
All in on 3060 12gb
Liquidate all assets to buy them
Then take out loans to buy more
Using the ones you bought earlier as collateral
https://www.desmos.com/3d/e32frbatvs
i have another julia set visualisation that actually uses the gpu properly
and because it uses the 3d calculator instead of the 2d one, i can do smooth gradients
side visualisation

i actually did render a 3d julia set with this prior, but i forgot how
No
yes
i found how to get the errors
menhera desmos 
3060 12gb is 350
7900xtx 24gb is 650
you save 50 bucks
try moving the border left and right
ye
For worse drivers no cuda bad raytracing and less coolness
that one runs in realtime
since its completely gpu based
not sure how glesmos works to be so slow to be honest
just get 3090 for $500

nvidia booo
"worse drivers" 
raytracing argument sure, except you're talking about a fucking 3060 here. the 3060 has terribble raytracing, its not a 3090
those dont exist anymore
skill issu
and its more of common practice to have less distribution of compute
very dramatic visualisation
lesser but better gpu is better
3060 is 12.74 TFLOPS, 7900xtx is 61.39TFLOPS
even if you get 2 3060s to equal the price and vram
you still lack flops
7900xtx has 2:1 half

what 2 days of no sleep causes
but yeah 3060 propaganda is wild lmao
i realised this actually renders a hollow shape

if i change to zero instead of 1 it renders a full shape
so then 122.8 tflops of half precision?
yeah

arguments can be made that radeon 7000 series has incomplete implementation of double CU units
misinput, ignore o7
im getting somewhere 
so in gaming its about 33TF f32 + 66TF f16 level
this feels liminal
and def activates Trypophobia

gayge
what does wrr mean
wrr
r
everything fixed

Not saying its their fault don't get me wrong
I understand from a cost perspective it'd be mad to run a full size LLM over every video for every language pair
mandelbrot zoom up to the max granularity of glesmos
and by mandelbrot
i mean julia set
again

I wonder if we probe hard enough we'll see the same artifact in reality 
Would be pretty funny 
i had to add aspect ratio override to viewport





