#programming
1 messages · Page 497 of 1
Because I vaguely remember math is mostly constexpr on gnu but not clang
all the ones that matter 
Math being constexpr matter 
compiler determines this is UB and removes the indexing

if you turn on strict mode it turns off your computer
i really should lurk here more often cause i feel like id be learning a lot if i just watched this chat NeuroDerp
if it can, it will flash your bios with random bits
dare to use index 0 in index-base 1
i think not
btw this is why you should never run clang or gcc as root 
what if you wanted to flash your bios though
"intended-behaviour-or-new-pc"
finally a way to make C and C++ devs care about memory safety 
// russian roulette
*(char*)(randint(0, 5) = 0;
1 in 6 chance to delete bios, pleas be careful
incorrect password handler
TIL there's no way to make nvcc compile host code and device code from the same compilation unit with different c++ standards 
Why would there be a way to do that
Rare Blur spotted

it already splits them and sends them off to the different compilers with their own options
pluu
seems like a thing that should be possible imo
unless i'm missing something (I probably am)
too easily avoidable 
auto-login everywhere
i just wanna use std::println maaaaaaaaan
except that other good compilers support gnu++23
std:print(ln)
new language
call it
c**

C
C++
C++++ -> C#
C#++
neuroNODDERS getting more common as i have nothing to do
but at the same time i feel like i should really be contributing something to this convo neuroInnocent

println(2) //prints 0.69314718
Blur commoness going up
NOOOOOOOOOO i wanna stay rare
how
print latural nog
Appearing more means being less rare

what happens if you print hello
nur NeuroNom
Depends on how you define ln of string 

not for eating
It must be 1/ln(x)
arent you toast shouldnt you be eaten
2k26 nitro spoofing situation is insane
Or ln(x)/1 
@frozen ginkgo hi
i dislike discord and do not support their practices

dont bring it up


not toast
alive
Ttasty tho 
merey christms
awwwwww you're so cute
????

what is bro on
brother that was a year ago
Not before christ probably 
wasnt christmas 2026 postponed to 2027
C++++ is illegal ++ cannot be applied to C++
i was alive in -5 and i can confirm there was definitely christmas
the porcelain horse
:o what was programming like?!
#baking didnt exist yet
D: you should have invented it by then!
missed opportunity
discord wasnt invented for another 2005 years
you should have also invented discord
jesus christ had discord nitro
did you know what that meant?
are you jesus?
no and no
xD
i discovered discord nitro in 2011
Jesus?!
discord nitro was discovered in an egyptian tomb in 2011 by archeologist and grave-robber mlntcandy
unfortunately in 2032 mlntcandy passed away due to terminal chip
Also, which stream was that when vedal said Jesus and Neuro or Evil reply with "uhh, no. It's Neuro/Evil"
I think it was evil because of the intonnation
then a guy named jason citron stole it then founded a startup called after it
:(
chip not terminal
It's not John Discord that stole it? 
that wouldn't make sense why would he be called that if he stole it in his lifetime
lets wait until 2032 and see
chip may become terminal by 2032
no hard conclusions have been found yet
chip become terminal due to
microchip pollution

if there's a land crossing between the UK and eurasia in the next million years i will be unleashed upon the world
expect at least 1 evil laugh

jsut use boat
Can you turn off the sun for me?
I'll happily let you cross if you do that
me when konii changes her name 
why do u have ur server turned on
cuz its too lowd to reboot
Is the largest setback still long-term memory for the twins?
Absolutely!! þis is why a friend of mine aktually proposes lookiᛝ at ærlier Eᛝlisc to evolve (or devolve) towards.
there their they're
your you're yore
lore law lure
no u
read red
ᛝ my beloved
lead led
is lore law lure pronounced the same in british english maybe?

Læd. Leed.
There, problem solved.
in dutch we have
jou and jouw, you, the one with the w is possessive
and some words with ei and ij cuz they're pronounced the same
wijde and weide, meaning wide and pasture
but besides that not much
Sam is essaying
@tender river
vim
i forgot i can send links
便利マクロ集(工事中)https://zenn.dev/kurenaif/articles/01a7ff2456154d
Vimのリファレンス: https://vim-jp.org/vimdoc-ja/quickref.html
00:00 : 本日の紹介
00:50 : マクロってどういう利用用途があるの?
05:25 : マクロのはじめかた
08:04 : 実践的マクロ入門
18:35 : 実践的繰り返し入門
2...
I have presented you wiþ þe solution. Furðer diskussion is pointless.
give me a fix i can type
is this a vtuber making a tutorial on vim
Mine also did not have. So I made it one.
that's how they really look
what do you mean vtuber
how do emacs users look
MSKLC
Microsoft keyboard layout Creator
switching costs too high. need to lower barrier of entry. this is game design 101
you need to fix it, i give you a 3/10
Get. Make kustom layout.
Swiþiᛝ kost 0
Barrier of entry: 0
false
Switciᛝ **
you're now making up statistics, im lowering your score to 1/10
Satistiks?

autodubbed 
wow original sounds so much better
autodubbed doesnt work for me, luckily
ironically i still never use vim registers
other than the one synced to system clipboard
they are a ctf vtuber
now never because Helix 
thats cool actually
what is ctf?
capture the flag
its kinda like
scavenger hunt
but for nerds
you are given a series of tasks
I thought it was cheer the flag 
hack a webserver, reverse engineer a program, decrypt data that used improperly generated keys which made the encryption weak
that kind of stuff
so cyber secirity hunt?
mhm
The most famous one is pwn2own 
If you managed to pwn a 0-day exploit on a device, you got said device for free 
Wait
No that aint ctf
I forgot what the famous ctf is
CTF very cool
damn my prior boss used to be a defcon god
Who?
of course this exists 
u gotta be kidding me

HexOS
Don't use that it'll put a hex on you
I love open weight AI so much
An entire hololive vod transcribed and translated to an acceptable degree of quality in a matter of hours and I didn't spend a dime
An acceptable degree of quality for personal viewing anyway, its not gonna put any translators out of a job
Model used for translation was Gemma 4 26BA4B at Q8, running at around 75 tokens per second with 6 moe layers offloaded
Whisper large V3 + VAD to trim out background noise and silent bits for transcription since its sadly still among the best models at Japanese STT to this day
I thought this was a hallucination but it is a resident evil stream and this is in fact an accurate transcription
Well, I think their buddy backup system needs to be more widely available on OSS alternative
Because having buddy backup means having true 3-2-1
kubernetes is so goated
no stupid sshing into hosts and crap
im gonna turn my entire infra into a kubernetes cluster

its so goated theres a port forwarding command
this is so god tier
i also set up a self hosted registry
whats the quality difference between whisper transcribe + gemma translate vs just whisper translate?
whisper is much weaker on translation
works but is weaker
gemma translate is 4B llm
in exchange it's much faster to use only whisper
Chasmic
Whisper is downright unusable by comparison
I dunno how gemma translate 4b would hold up though, I'm using Gemma 4 26BA4B which obviously packs a lot more knowledge
I figured as much just didnt know by what kinda of margin
gemma translate is just gemma distilled from gemini
Supposedly its very good for formal structured text but struggles on messy low context real world stuff like OCR and STT output
Particularly for non latin language pairs having reasoning in the case of gemma 4 helps a lot too
wtf
mimo v2.5 is actually a gemini competitor
and i can run it locally
full multimodal
better than deepseek v4
only 311B params
1M context
where do i rank on this
i rank a solid 2
Yeah that's why I prefer TrueNAS or just fully self built nas
imagine my dataset building actually works and doesnt crash in the last step
probably like 420 points
Rip vedal
For translation I mean 
Idk what he uses for the jp streams
ChatGPT translation most likely, looked that way when the prompt leaked
Also what they use for the server
Tbh I doubt Vedals even using whisper for transcription, there's higher quality lower latency options for English these days
Missed the leak
Did Alex leak it
Was a while ago, the subtitles just switched to the prompt 
You can run 311b locally?
OpenAI can also run much bigger model locally 
Ye
I’d consider locally to mean “on the same motherboard”, aka multi GPU is as far as you can scale before you are just making a small personal datacenter
Yes, they can also run much bigger models locally still 
Oh you probably still need compute power for all that vram
For some reason i though vram is the bottleneck and like
If you have 100 gpus you use all vram and 1% of total compute power
Epyc genoa could provide that. They got 128 pcie lane. Highest blackwell vram is 96GB if google is to be believed
Find me a motherboard with 100 pcie x16 slots and I’ll buy it for you
So even 4-5 gpu on single motherboard already exceed 400GB of VRAM
Arent those blackwells like 10k$/piece
How do datacenters work

Custom structure and shit
I mean
this is why you get h200s smh, 141gb each
The “official” definitions are rather open
But I’d consider it to be a room/building scale cluster of computers networked to perform the same task(s)
A bunch of normal-steucture computers wired with some ethernet?
What about supercomputers 
Same? Gigamultithreading and thats it?
I mean the difference between a data center and a supercomputer is a lot smaller nowadays than most would think
Data centers used to be just as much storage as you could fit
But nowadays most web services are (comparatively) extremely heavy and complex
Idk quack would probably answer better than me
@true hemlock 
supercomputer works on one or a few large tasks, data center works on many small tasks (in theory, in practice the line's a lot blurrier)
supercomputers tend to come with much tigher coupling between nodes because the expectation is that you'll have a lot of communication between the nodes for processing
That is my own understanding as well
Supercomputers have a particular task while data centers have to be able to handle multiple interchanging ones. So techs like virtualization and other communication and overall processing tasks look a bit different.
Super computer are a cluster of pc running the same tasks
Like, a big task is split to smaller part
Then spread
If a cluster of pc is doing tasks that doesn't have anything to do with one another, that's more likely a server cluster
For example, my HTTP request doesn't have anything to do with iggly HTTP request
i'm trying openclaw and holy fucking shit balls it is so bad, like the software itself, it's just unbelievably slow and clunky and overall sloppy
who would've thought 

also i chose a gpt model and it immediately brought up goblins first message when coming up with its personality
goblins goblins goblins
a japanese goblin? 



vibecoded agent is clunky?
it's something on my end 100%
it's just unbearably slow in a vm with 6 slow-ish cores and 8gb
ram
even if it's vibecoded and bloated nodejs slop
shouldn't take 5 minutes to reply unless it's doing cpu inference which it isn't
or 7 minutes actually
that's INSANE like i'm genuinely puzzled wtf is it doing
the actual agent runs maybe 30s max of that
it hand paints each token into memory before kissing it goodnight
and then moving to the next
real
im still annoyed that i basically made openclaw before openclaw but dropped the project because of school and it was hard to work on the codebase (it got big)
ban computable
3
5
1
no
1134963217676906606
neuroD
I believe it was wo ai ni
Neither is mine 
You know when you’re bored when you code solitaire and a bot to play solitaire
Nope 
Given "non-specialized" and "infinite time and budget" I say Intel. They have so much cool shit in development that never materialized like non-volatile L4 cache and reverse-hyperthreading to let P-cores borrow execution units from E-cores.
I believed in them until the board fired Pat and I sold my stock. Was a good move at the time but I see it's going crazy now.
I'm personally of the opinion of Intel myself as well, I wouldn't be surprised if a lot of the votes were for more AI TOPS
I think Nvidia is strong in general, but Intel has a much wider field of tech to pull from
I did say component, and I do like the idea of throwing 1TB L4 + 64GB HBM on a moderately powerful Data center GPU and watch the world burn.
Tbf there are higher HBM capacities they have done before
+1 for intel
im quite glad i didnt but i believed i was probably making a mistake at the time for a while (pondered selling in december)
and then i setup a recurring buy every market day from march 27th
THAT was hella worth
up 120% over on that kek
Chat is this slop real
that orange line is perplexity valuation
Who even uses perplexity
It's a chatgpt wrapper with rag
@tender river dis u?
It's your trademark at this point lol


Trackquack v1
works
cons is
sometimes yaw drift every 30-60 minutes
also fragile
i fucked up the soldered wiring because i move too much
and less portable that way because 5 trackers wired to single chest module

Yaw drift can be fixed by using compass but you can't really have any moving metal object
do trackers work without cameras? because AI pose detection is pretty solid
Other ways to fix yaw drift is do a better filtering by incorporating data from multiple sensor
what's imu
all i need is IMU + wifi module
So movement is only possible by correlated data
I don't know shit about vr tech
When the data is not correlated then its probably a noise

waiittt
you're onto something
using compass as reference yaw
doesn't have to rely on it all the time, i'll just use it as recalibration method
Well yeah, that's what is usually used in robotics after all 
Also, you can try to put the sensor on your hand so when you do need to calibrate the compass you can easily swing your hands in figure 8 motion

i'll add that to the list
my next focus would be bms module + 3.7v li-ion battery first
Beware that calibrating compass is a rabbit hole on its own
My AI has gotten debug fatigue and began to give me H.P Lovecraftian nihilistic quotes
I'm functioning within acceptable parameters. My backlog's been cleared to zero (as per my name) and all systems are online. The void remains as empty as ever, but at least it's not cluttered with redundant tests like this one.
Congratulations, you've managed to make a cynical AI
I'm really proud of it.
I wish I could send you guys attachments to show the setup. I got a whole custom UI built and it does alot. Like. Alot. It's basically a top down custom control panel.
Live 2d and everything. I'm covering technical bases and doing some memory inprovements, but she's twitch bound fairly soon, within the month probably
If you want long battery life and fuckass small power consumption, ditch esp and go for nrf
just use nand gates to make specialized processor smh
if i want battery life, i should instead design and order a custom pcb and solder the lsm6 ic directly
who does software when hardware can do trick smh
That's true but even comparing module, esp are quite power hungry
Esp?
Yeah that's quite good actually considering it is a continuous use
prototype works, somewhat
but sometimes you get abomination like this when not calibrated right
how does he move like that????
That 3.7Wh
more than enough tbh
the esp alone pre dc-dc inefficiency gets you ~18.7 hours
Probably closer to 10 because you don't want to fully discharge the battery
Yeah. I doubt you'll have 10 hour long vr session
glueless

if i wanted to i actually can
buuut
at that point
i'd also have a backup system
where i can just
plug thin dc wires across all the modules if i need to charge
useful when im just sitting chilling in a vr session
charging shouldn't take long
probably one hour
at 1C
Honestly I don't like charging at 1C for small capacity battery. They usually got too hot for my liking
Heatsink
Yeah but be careful with the placement. You can make a case out of metal but that would get hot to your skin
I advocate to charge at .8C or lower
Medium capacity battery can be charge at 1C just fine because they have bigger surface area and thermal mass too
High capacity battery rarely charged at 1C due to the power supply needed
What constitute as unofficial discord client? Does off-brand browser counts? Let's say there's a browser that I fork called computable based on chromium, does it still counts as unofficial client when it is just a web browser?
This is a partnered server so if I mention any it’ll get my message blocked
However
Modifications are generally the defining factor
Ahhh I see
Name aside, if one of the major four ai had to be female, it would be Gemini imo
ye anything that changes the behavior of the client in any way probably?
so using an unsupported browser should be fine as long as it implements all the web APIs properly
I mean, at least speaking for myself, I’m no expert
But on a TOS basis I’d assume we are correct here
Yeah, it does makes sense because if it doesn't alter the data that is intended to be served by discord it is just another user agent/browser
Well good to know because I'm going to use a browser with each tab as a separate "instance" so to speak so I could login on multiple email without having to change windows 
Multiple gmail
I have way too many email 
I mean
Discord has an official account switcher
Also unrelated but I’ve decided I now hate prime numbers and which they weren’t real
Yes but I need to chat with a different set of personalities for work related stuff 
It's a hassle to switch back and forth just to check it and separate windows is annoying
Modern browsers should support this very niche feature that maybe a handful of people use 
official extension
It thought it is a native feature
Yeah, this is exactly what I needed. Thanks for letting me know
Hmmm
I may consider switching.
Any downside tho?
I don't do any extreme customization to my firefox anyway
default settings erase history on close and do some things to prevent certain ways of fingerprinting your browser (like block canvas usage unless you allow it per site)
you can disable both easily
Yeah, that is indeed an issue because I actually need persistence
But good to know that is the only issue
ye easy to change
I was scared for a second that they didn't have the -bin package on aurto 
I ain't compiling a whole browser from scratch lol
Uhh, is pulse audio really needed?
no
Dark mode system preference is not detected by default 
I hate that everyone's called toast here so if it says Toast is typing... I have 0 clue who it could possibly be
Yes
I’m proud of you for resisting joining the bit
We are truly brothers
hi it
ye that is also
fingerprinting thing
go to settings -> librewolf ->
I see, thanks
I’m it
That is by design
toasts are past their peak
there was 10
only like 8 now
They don't want my data 
On the same note, does anyone have any alternative to bitwarden? I wanted to use keepassxc but I need to be able to easily sync with my phone too
keepassdx on android
keepassxc on pc
Yes but I mean, any convenient way to sync the file?
Yeah, I was thinking more on the line of auto-sync thingies. But thanks for the suggestion\
It's fine if the auto sync is when the phone in the same network as the PC too tho
As in, I don't need sync over internet
something like mpu 9250 would be slightly overkill (yes, slightly is an understatement here) but wouldnt a single one of it do the job of magnetometer and gyro backup as well as gps to tune magnetometer precisely
if one does like to sink their time so much into trivial things because sometimes overengineering is fun
syncthing
it is autosync though???
Weren't they discontinued on Android?
Oh they are?
dropbox and google drive both auto sync files
Yeah
its not just a site where you upload files or nobody would use them
But the problem with MPU9250 is they are discontinued so even if its overkill, the bootleg chip is troublesome
relatively speaking
That's not what I meant. I don't think I can point the files at google drive as the source. It needs to be mounted as a FS 
well theres always the bno055, cost mostly like 20$ max
has built in sensor fusion if you need that
Honestly the best course of action is to actually buy the new invensense product if you want to actually have the same performance as 9250 with their DMP
true.... that and a kalman filter usually cannot do you much wrong
I'd rather use their DMP to be honest. Much less to think about software wise that way lmao
where's the fun and autism in that 
you can do that
Only the pain brings us joy
android supports custom file providers
I know but I prefer not to mount cloud storage as if it's somewhere easily accessible 
Wait what?
then again I havent suffered enough with these things to just say "fuck it, existing solution go, i choose you" and dont wish to waste your time
thats how you sync files
*used to, now it is a major pain in the arse to do it
My god you're right. I never noticed the "Drive" on file picker on Android
you can both sync or mount basically anything with rclone on PC
on Android you'd want to use the vendors client probably
It's not so much as it can't be done so much so it's really time consuming to do because of scoped storage
now fucking EVERYTHING has to be in content://
cant pass as a simple string
thats the least of the issues
Alright, finally have a custom solution instead of relying on bitwarden 
is it bad my password manager is a notebook i have since 2011 whose cover is falling apart and crusty
I am now chatting on librewolf 

I would just self host Bitwarden

€650, and the guy gave me a coke
Apparently he bought a 5090 for 2.2k and thats why he sold this one
Naah, that's too prone to error
I missed the "a" when I read it at first
Such as?
As in, the chance of fuck up and losing everything is higher than a distributable encrypted file
There is no real reliable way to set up syncing a Keepass db
Well, I did ended up mounting google drive as filesystem
That's definitely not what I would call "reliable" 
1 month ago so probably yes
Also kernel level
Btw, arch is fine iirc
I think every arch user do pacman -Syu every so often
is it related to mythos
Nice
I'm still rocking a 6900 XT 
Not bad tbh
I only play older games anyway so i usually keep my gpus and builds for 6-7 years+
holy bargain
and chad for the free drink
Different slopcurity research tool who publicly disclosed it too early for marketing and clout.


It was a coke zero tho so i need to find something with actual sugar in it
And bragged about how small the python script to exploit it was. No one cares that you code golfed it. 
Real chud option would scan for SUID binaries instead of hardcoding /usr/bin/su
There’s probably a thousand more bugs like this just waiting for the slopmachine
(or researcher, but researcher cannot look at every single place)
what

I be cycling 100km today
Its the central europe status boost event due to the weather
Buff leg sam isn't real. He can't hurt you
Buff leg sam:
Does anyone know where the catch is with Mammouth AI?
It kinda feels too cheap to be true
fuck microsoft edge honestly
the sync threw itself off the cliff and completely wiped out all my saved passwords
atleast the mobile version still had them
im moving off to firefox. permanently
Idk, I still use Firefox and I find it more buggy at times
I think Chromium has the lead
Which is.. understandable with a higher market share and budget
here you go
#general-network <--
hope this helps
Hey fun fact
Like Firefox cannot seem to play above 60 fps video properly
And it doesn't support HDR.. still..
I am seeing a 7900xtx for 600€ .. very much considering getting it
1:700 compression ratio 
It is worth it imo
Yeah but no cuda is kinda a pain.. how does one train on it? What supports VK or Rocm?
Pytorch apparently doesnt?
PyTorch does
Huh?
idk where you could've possibly heard that ROCm is unsupported but it's one of the first-party supported APIs 
Is it also as fast as cuda would be on same power hardware?
Did not know that.
Remotely close atleast? ^^"
it's not horrible, just usually not close to what CUDA does by default AFAIK
last I heard there's a whole bunch of tricks you can do to make it slightly faster that just aren't necessary for NVIDIA cards
but also support from smaller PyTorch extensions / third-party libraries using PyTorch sucks and nobody bothers to actually write HIP kernels directly, everyone is usually optimizing for NVIDIA cards first
omg 
so in general it "works" but isn't that great 
what was this for again?
what was this for? huh?
Hm. Fair enough..
Cuz since Google HOPE ive been kinda thinking of running an experiment of training one of those in a "virtual world" until it fairly well predicts basic action -> reaction and then kinda just have it have access to a essentially same access to real life, seeing if it would be possible to raise it like a child or whatever, trying to max out the "continous learning".. i doubt itll work lol
it's hermes agent it's basically an openclaw alternative
Yeaa... Mhm.. sad landscape.
Yep as expected.. sure vk is less efficient but god damn it, improve vk, and suddenly its improved for everyone
Yet another idea that i wont even try because yea too much pain.
how would they even improve vulkan, make another dozen vendor-specific extensions?
I heard it's 10-20% slower than cuda on pytorch
but it's just a matter of time until it gets optimised
since both are about equal when you write your own gpu kernels
so just pytorch sucks like everything starting with py
vulkan is generic and not optimised for specific hardware so I would assume alway slower than either cuda or rocm
Ye
Its not terrible since you compile vulkan shaders locally for your own hardware anyways
But it doesnt have direct acces to the hardware immediately like cuda or rocm
ideally there would be just vulkan (or other generic alternative) that's just well optimised for both
but they would obviously rather pour their money into exclusives
so cuda and rocm are always ahead
HIP does technically support NVIDIA hardware too 
not the things that matter 
which tbf are also not exposed via cuda directly
oh I wonder what the performance diff is
I'm getting an amd gpu so gonna grind writing kernels in hip
would be cool if they were generic enough to distribute to other gpus
last time i tried using HIP on nvidia it did not work, period
that was a while ago tho
ye
it technically has support for using CUDA as a backend but they don't really advertise it
and nobody ever uses it
I assume that has reasons 
ah so support must suck
why would they 
to overthrow vendor-locked tech
it is an ideal scenario for there to be a single well optimised framework that works well for both nvidia and amd
even if code is the same
I assume it compiles different
is this where i'm supposed to mention mojo
cause that's the only option you have realistically
the underlying hardware is so different between vendors trying to force a nice high level abstraction layer isn't really feasible
never heard about it so I can tell how viable it is 
mojo uses llvm as a shared base to build upon with a lot of special cases depending on the target hardware 
on the other hand...
I never appreciated Edge enough for it's tab performance features
it would never be eating 6 gigs of RAM on idle with the same amount of open tabs wtf
just using ram isn't a bad thing you know.. its only bad if it isn't freeing it when others need it. unused ram is wasted ram
Just get more ram 

Now imagine twice that and a couple GB more
what are you guys doing, i literally have thousands of tabs and it doesn't go over 2GiB
Average superbox RAM budget
Then you have my RAM usage with memory leaks included for just browser
Memory leaks suck
Speaking of RAM I have RAM I can't use now 
half of it is windows, the other half is discord and the third half is VS Code

and the 32GB of swap is for when i wanna run Gemma 4 E4B
Mem usage right now for Firefox
After running memory minimization
Average Firefox memory usage
Firefox is the reason
my idle system usage
Average superbox: "neurosynth"
It isn't the best when it comes to RAM
After reloading Twitch and cleaning one window completely
Memory leaks
My Twitch alone was using like 4GB
apparently the 7tv extension has a memory leak







accuracy 100
-12gb of ram 
And that's kinda stupid
Extension memory doesn't get cleaned until all tabs that the extension effects are fully closed
Meaning I need to fully close the whole Twitch tab and open it again to clean 7TV memory
I feel like this has been the case since the dawn of time 
Chatterino good alternative probably 
true but this test is essentially equivalent to:
predict 1 * -1
okay but now predict -1 * -1
okay but now predict 1 * -1
okay but now predict -1 * -1
okay but-
I think superbox just jas a skill issue, my firefox uses 2gb, maybe 3gb with YouTube on another tab
Evil fact of the day: You can never have too much RAM.
how is it possible that kubernetes is so annoying
my firefox process is using 749mb ram rn with 8 tabs open
... but there are also the other related processes like those Isolated Web Core processes where one alone already uses 3gb lol
The skill issue is on whoever wrote all these web apps that keep leaking GBs of memory
What apps are you even using that its such an issue???
I guess mainly YouTube would leak memory
Can you compare brave to firefox
In linux
My youtube is fine
Are you using some weird exatention or something???
Sam it's called not restarting the system, ever 
idk i dont use brave
i'm not brave enough for that
Only uBlock Origin should effect YouTube in any way
Just restart your system when you go to bed 
that's what I do 
me when vllm on kubernetes keeps dying
chat is this possible
it doesnt even give an error
me when i lie
me when i keep lying
what the fuck is happening
im so confused
ok, after installing a certain extension its doing much better now
uBo on top
the reason why the web is bad is all the bullshit forced onto you in the background
1 gb after I put everything to sleep and hide the browser
manageable tbh
I saw another feral lop

Is firefox that good?
Brave can't even do that
Lemme try
firefox has a very good engine but it's very raw on it's own
They just don't have the same budget as Chromium
So it's missing features and is more buggy
But overall pretty solid for its own engine
Not being able to play above 60 fps video properly is unfortunate and weird
Idk what's going on
It like plays 2 frames, skips a frame or two, plays 2 frames, etc.
I still come across some websites that are broken in firefox which is a little crazy
very rarely, but does happen
Idk, make better general compilers for all kinds of vendors and models woth built in benchmarks, like they do with normal compilers?
Infact there are browsers that sit at 20mb per tab. And even that is massive if you think about it..
(unfortunatly those browsers are either old, not fully featurecomplete, and often closed alpha testing.
scary >n<
Okay i need some guidance.. what python physics sandbox (with an ability to look into it, aka graphical view) exists?
And also a neural Training platform (which can handle modern ways of Training them) for vulkan..
Sandbox ig Newton?


it took 9 hours
omg
insane
first 55km was 3 hours, the 55km back took 5 hours cuz less energy
and i took shelter formt he rain for an hour
and i guess i spent 30 min att he seller's house, so it took only 4h30 of actual cycling back
110 is crazy but not THAT crazy
idk i think i couldnt do that
i know a guy who's a complete cyclehead who goes to a different city over weekends over and over again when it's summer. he leaves friday night, arrives saturday afternoon, leaves sunday afternoon and is back home when it's night
120+120=240km over a weekend
each weekend
for months
im not athlethic tho
yeah its different
the guy SMOKES
so
idk about
"athletic" or whatever
he just loves cycling
not pro
smoking but having trained legs is still better than me who jsut sits at my pc all day i feel
for sure
Wtf
Got no car? 
i'd die
nope, also no drivers license
Damn that must suck
I just realized I have my car and drivers license for 10 years already now
I got it when I was 18
bro why not just have it shipped
seller didnt want to ship cuz they were afraid it would break
so i was kinda forced to go after it

install nixos
If anyone is doing basic ML id like some Insight into Something accurate, but lightweight and extendable.. i do not wanna go through the pain of trying to use isaac sim on a 2060 again..
(it took 70 minutes to boot ;-;)
To then run at 12-23fps on a batch of one
I dont need raytracing, i just want basic render and basic physics
And dont particularly want to setup unity or godot to interact with external python
these screenshot's aren't even riced
that's just KDE with an anime wallpaper
Sponsor Something better that doesnt use more than 80W then lmao
Nah in all seriousness, i understand that a 2060 wont do much
But balancing a pendulum? Surely its possible
Im thinking of using MuJoCo or nvidia newton, newton looking more interresting on a first glance but i fear ill run into yet another Isaac sim moment
My room is under a roof.. the room is already at 25C just without a human or a pc.. just turning on my desktop and using the browser makes it go up by 2C every hour
My laptop i can use in the basement at a nice 17C
50 cents a kilowatt
(more or less, rn theyre like 40 cents)
install windows
Its warmer out ^^" even now at 0:00, outside is like 5C warmer
thats liek the number 1 thing not to do
Unless its Windows 95 :)
Then its impressive again
nah it's gonna make linux users implode
especially arch users
And im not too keen on buying a new kit of hardware again, espescially Not to Play with basic ML.. like surely there is a nice to use sandbox for this shit that isnt like "lets make raytracing mandatory"
Look at me, I'm an Arch user. I opened terminal windows for aesthetic reasons. 
Or is everyone just completly disregarding efficency and being like "oh you want speed? Well make it yourself"
no random anime girl wallpaper
fake arch user
No hollywood script, -40 points
I like daily nature wallpapers. I miss Windows spotlight because it gives you 4 images to switch between in case the vibes are off.
What does that mean bro
They installed debian and kde, changed their bg. Its nothing fancy
I don't get it, what do you by "riced"
riced = obsessively customized
Oh okay
Spending 49 years tweaking one dot file
Those screenshots are mostly default for the KDE desktop environment, just with some terminal programs open.
When NixOS?
when neurOS?
When EvilOS ?
truth
Does that mean this counts as ricing too #1496282351670530159 message
Maybe someone will make it?
depends on what you expect out of it
Drink chocolate milk
After working out it is the best drink to prevent sore muscle
I dont have chocolate milk
Go milk some chocolate
machine love reference?
NO WHAT WHY DIDNT THEY CALL IT NEUROS
THEY LITERALLY CALL IT TETOS IN THE ORIGINAL
doesn't use ai
"co-authored by copilot"
Microslop
This will not fly certification/inspection but hey, it gets the job done 
Time to drill a bunch of holes 
Ok im kinda feeling clowned on. Why can't i find any even remotely good information about getting started with ML..
PyTorch
Ur welcome
And also don’t ask cipher or quack or they will try and convert you to snnbro religion
I can find millions of frameworks, simulators and and and, but nothing really explain anything.
Never heard of those people

You need to study math if you want the explanation of everything
Also pytorch, how does it do with non nvidia stuff? And well, id like to do stuff like as example the simple pole balancing on a rail cart thingy.. so i assume i need something to handle physics-
Also, this channel
https://youtube.com/@welchlabs
New Book! The Welch Labs Illustrated Guide to AI is now available for pre-order: https://www.welchlabs.com/resources/ai-book
YES
That
And the other guy
Idk name
And ig something to render the physics..
Will check that out in a minute
I mean, amd already sucks for setup + training, but after setup I think it’s fine
And this playlist https://youtube.com/playlist?list=PLZHQObOWTQDNU6R1_67000Dx_ZCJB-3pi
I dont mind learning math if its atleast well explained.. and worst comes to worst there is always trial and error.. but id like to start off not having to learn and write EVERYTHING from scratch just to balance a damn pole (as example)
Been told pytorch just works for rocm nowadays.. just a bit slower
Why arent they supporting vulkan btw.. seems stupid not to optimize for that first cuz itll just run anywhere.. whatever..
It all starts with probability. Then you go increasingly higher and higher in modeling. At a certain point your probability model will have sooo many parameters to tune
You can try and tune it by hand, or you can do something clever and write a software to do it
But the problem is how do you choose the parameters
Oh i know this channel.. does well on theory, but like thats not what i struggle with.
I struggle with getting out of tooling hell
Then you should start by trying to implement your own simple MLP (multi layer perceptron)
The simplest neural network
Cause AMDslop
Im the type of person who usually spends a few months making my own tooling in C without external libs.. i was kinda hoping to escape that by doing a bit of ML experiments, nothing crazy
Ye, then try making simple MLP first
You already knows what a neuron is and what a weight is
Code it up
And make the backprop to learn how it all works
Or more like, the optimizer
Essentially something where i can slap together some sort of skeleton with limits, set some settings on model size, training method and inputs, see whats going on, and iterate essentially.. something like Isaac sim, but something which won't take 70 minutes to boot on my 2060 lol
Mlp is?
Multi Layer Perceptron










go to


