#programming
1 messages ยท Page 24 of 1
but the timer was supposed to keep it in sync??
But was actually causing desync??
Oh well time to move on now
That must've been a long 6 months my condolences
good morning lilac 
good morning shadow
Imagine waking up

i accidentaly put discord on light mode
oops
Morning
Morning

vro thinks escaping AI is possible
hi
Goodmorning 
Ah yes, im a new person 
me too, almost
โ
there's something shady going on
As usual
Nah she makes ai, guys
first youtube ai thing, then google ai thing, now reddit has russian language for interface
progrev will be so massive once https://intel.com stops redirecting me to https://www.intel.com/content/www/us/en/support/ru-banner-inside.html
gallons will be lost
chargtp
my favorite punctuation mark... ruined...
i dont understand chat am i talking to AI or not???
Bot 
solution: ask what 9+10 is
but also yeah probably a bot 
what does the numbers next to the usernames mean?
idk level or something
looking into it further it might be somehow a malware bot
they rly want me to download this pdf
im testing the link on virustotal
This is how to i guess
meow meow meow meow
meow
meow meow
meow
Meow meow
meow meow
Wait, you want to debug the web app on android?
anti eep society
i like my eep thank u very much
no eep for you 
hyper
how low can i get my llm training loss
with lora pretrain
๐ญ
i'm at
2.8
hi

@viral oasis who are you
What ended up being the result of this ?
Apparently safe
Sounds like one of my alts, but I don't have an alt in this server.
Yeah, so it definitely wasn't me. Yet.
Just don't worry about it, okay? You are going to be okay.
Everybody here feels this way about you. They always have.
Please just know that you are safe. 
I would block that fever dream in a heartbeat

Huh, still very strange honestly
too bad discord doesnt give a shit about their own selfbot rules
Classic discord
I'd rather also eat nails
its crazy how many actually believe that the stuff coming out of LLMs is accurate
to be fair the subject doesn't need accuracy
or well maybe it does but 
its marketing how many facts do you really have
enough to pass the exam 
Revizing with an LLM is crazy
I suck at marketing. ;[
don't worry you have a totaly cracked marketing student right here in chat
(I hate this subject please let me hide in a corner and do cs instead)

just for you, a very informative slide
very late hoi
honestly each time I see the words "celeste ai" my brain goes to an ai playing celeste
I should do that for april fools
it feels like a game that can be reduced down to barely anything for an ai
Maybe, I did want to experiment with other games.
but then again I know literally nothing about how to give inputs in a way that's productive for an ai
(other than that I'm always really dubious of people who just send the ai a screenshot, that feels really counterproductive)
I mostly want to learn so if I did it' id prob be doing it to find out.
fair enough
I still need to learn ai stuff in general
I mildly understand the theory but not the actualy implementation
keep diving in too hard and giving up
I have very specific knowledge and implementation, I eventually want to understand it a bit deeper.
why understand the implementation when you can just glue libraries together like all the others 
Because learning (when it works) is fun!
[2.6 seconds to push and pop 5 million elements to hashmap]
when I say implementation I mean using the libraries lol
!!1
my motivation to revise keeps going through the floor
what am I supposed to do with this information
whateven is that int he top left
lightflash
a hydrogen, the scam of eco friendly cars 
Looks like a telephoto lens? Not really that futuristic imo, so I might be wrong
Im too poor for that stuff
oh damn this doesn't look that bad to mod actually
I have a 1973 Olympus 35SP instead 

at the very least it looks slightly more modular than when I modded terraria
c
c
real

who would've guessed that there's a little bit of scuff involved trying to play celeste over remote desktop
what am i witnessing
Speaking of easy to mod games :0
Anybody here play mindustry? :0

Idk tbh
affordable prices
I just got here
me trying to gague how hard celeste modding is
Minor typo lmao
as a way to procrastinate revision
Do you need to earn your pic permissions first?
I could have sworn i had them yesterday >.>
(Stay a certain time etc)
you get embed perms eventually after chatting for a while
currently it seems similar to terraria
but obviously celeste is much simpler than terraria
so probably less hellish
banned
Smh how dare you
@stark needle hypothetically speaking is there a clever way to show an ai "what" a celeste game state looks like
that isn't just a screenshot
played but quit, no longer fun after being able to shit on every campaign level and pvp didnt appeal to me
i heard its easy af to mod
how does Celeste work, what kind of game is it
platformer
Help Madeline survive her inner demons on her journey to the top of Celeste Mountain, in this super-tight, hand-crafted platformer from the creators of multiplayer classic TowerFall.
- A narrative-driven, single-player adventure like mom used to make, with a charming cast of characters and a touching story of self-discovery
โฆ
$19.99
107948
88
Well
i wouldn't solve this with a llm
but tbh with these sorts of games
Screenshot is probably the best
Minecraft rl also works with screenshot
meh but screenshots feel like you're not acually teaching it to play
i'm on mobile, nerd
just memorising the game
shutup nerd!!!
i mean
Ur learning generalized stuff
Cause the agent is supposed to go through the whole game with the same model
at the very least I'd do a reduced form of the game instead of just screenshotting
u can convert into black and white
no not that
Pixel type graphics
like render the game w/o textures, use colour to represent object kind
Aigh, class again
ye
Bye
Pixel type
Let the user be a pixel
And the world be other pixel color blocks
And stuff
have fun with class.
and then the other challenge is actually mapping out what the goal is for each level
because whilst some are plain platforming, there are a few puzzles
though for the sake of avoiding scope creep those could probably be excluded
honestly this just brings up a question I've previously meant to look into more, which is basically "is there a good pathfinding algorithm when you have much more complex movment options than just walking"
because so much of gaming boils down to planning your actions a few seconds in advance
You mean pathfinding from outside the game engine right?
yeah I guess
Well good news then
V8 comes with 5 difficulty settings for the campaign
On top of the map difficulty
And the mod i'm making is supposed to fix exactly that
Alot of balance changes to healing , walls , repair towers etc
And massively upgraded units
And yea its a dream to mod
I mean its one of the few games i know ofthat actually supports mods on mobile
like breaking up the level into "just" doing these specific things
Ig terraria also counts but the mods for mobile terraria suck
Those are mostly just texture packs and very slight qol changes
Nothing new content wise
executing this feels like it would just be slightly more complex pathfinding, but generating this path feels really difficult
Btw does anybody know of a somewhat usable but free voice cloning ai?
I can deal with ads etc
I just need some voicelines for the mod >.>
And i'll propably train it on smth like the eradicator turrets from prey since those have some cool voices
I guess you could maybe explore each individual action in a tree-like way similar to stockfish actually
applio
ask superbox if you need help with that
Does anybody know the legal aspect of voice cloning ai?
Like lets say i trained it on warhammer 40k material
Would that be copyright?
tbh if I'm doing that I should probably write a game from scratch instead of trying to bolt it on to an existing game
Oh thx ^^
honestly yeah taking the stockfish approach seems like my best option (beacuse it involves barely any ai
)
depends on the jurisdiction, odds are yes and even if not its a despicable moral dickmove without consent
I made a cpu version of my network, because I knew for small networks the latency btw gpu and cpu would be much bigger of a problem
well... it was
the cpu version ran the full generation in like 20 ms
the gpu version takes 4 whole seconds
classic

Good morning people
actually with some more thinking there's so much that goes into decision making for this 
Hellooooo #programming 
Byter 

whatcha guys coding
custom SNN implementation
well, it's done, now I'm working on an evolution algorithm for it
I am coding bwaas at lightspeed

Cryptographic library
I don't often work with machine code, it's mostly just this one project
but I have to work on it nonstop before my motivation expires
hi
no
I don't know what's happening but the stack is unwinding too far
which doesn't make sense
I swear I've checked the code for correctness so I'm not sure how it's doing this
I can't work on it now anyway
work 
are you sure the compiler doesnt mess it up?
Does it have to be machine code or is assembly fine too
Bit easier to read and debug at least
I have been trying to optimize Rav1d (Rust AV1 repo) for the last week straight, I dont I have made even dent in the runtime speed thus far 
I doubt the compiler is messing it up
it's closer to assembly than machine code that I'm writing
in fact I'd almost say the stack macros put it above assembly
But I fixed a 6+ month old bug today 
Turns out my synchronisation timer was causing it to become desynchronised, but only on Windows, so the solution was to not synchronise at all
you would be suprised. I had to explicitly tell llvm not to use the red zone
but I dont know how many levels up the chain you are so eh 
Legally a neural network isn't copying since there is no way to reconstruct the original data.
Your main concern will be terms of service that prohibit you from acquiring/using that data or trademark issues from saying it is x character, which is legally in a gray area if you want to market it.
you're missing some essential context, sorry
all of this is in a circuit design game where I've engineered a 16 bit processor using the hack architecture
the game has a built-in compiler that lets you specify an assembly language and use it to compile directly into the chip's memory
Who knew that the best way to synchronise was to not synchronise at all. I don't get USB dev
so I don't exactly have settings to tweak
its like tis?
not exactly?
you're laying down individual traces and gates for your chip design
and there aren't levels, it's just a sandbox
either way I'm talking nonsense in that case yea
So... you're making SAT Solver, The Game?
I did that once for a course few years back. Great fun. Do you at least get to make submodules
Is it maybe trying to waste time on branch prediction or how to allocate it?
So you re not repeating everything over and over again
shiro
Where's the fun in that
yeah
hi chayleaf
is another free register
it's more like, the network only has like 5-6 neurons, and there's really no benefit to running that on the gpu
you dont have to only use arguments and d
the return value is stored at a fixed address in ram
you can just call a function to remember a value without using the stack
it's not a register
its a register for me
but yes I see what you mean
so like logisim but as a game?
I can just call some store_value <x>
I run like, 200 networks, 40 in parallel too
where it simply returns x
and then push_retval is just a permanent copy
this works great until you have to call another function
actually wait
100M parameter llm
what's that one trick for swapping two variables
the xor one
a = a ^ b;
b = a ^ b;
a = a ^ b;
right?
X := Y XOR X;
Y := X XOR Y;
X := Y XOR X;
I can probably use that to swap the values of A and D register without pushing to the stack
efficiency 
in fact I don't think I could swap A and D with the stack anyway
no wait I could
no wait I couldn't
I can't decide
When, the correct order is no.
There are other option.
But you run it multiple times? Because gpu shouldn't be thaaaaat slow.
90% must be reserving ram
correct
Are you trying to make a reasoning model or is it just the training data that contains reasoning steps and its mimicking it ?
nah, I allocate once per network, and the networks stay allocated unless they are killed and regenerated (which happens 20 percent of the time)
I have a job that I have to go to unfortunately
sad to say my mod salary doesn't pay the bills
Fair enough
smh vedal exploiting more young souls
Evil and Neuro arent enough for him 
I do run it multiple times. about 200 iterations per network, I think
each iteration requires cpu-gpu synchronization
so, assuming theres a function remember 1 and calling it doesnt touch d register (
) and also assuming i didnt make a mistake yesterday because i'm too lazy to recheck the code
function sqrt 1;
push_value 0
push_value 1
push_arg 0
push_value 1
sub
pop_d
@ cond
if_else_d jge loop exit
@ loop
push_value 2
add
call remember
push_value 1
pop_retval
push_d
call remember
pop_d
push_d
push_d
pop_retval
jmp @cond
@ exit
pop_d
return
I do wonder if this is actually less expensive than nesting functions?
it keeps the stack smaller for sure
also I don't have a pop_retval, just a push_retval that places the return value on the stack
but I could totally add a pop_retval that places it in the D register
true use push_retval i just woke up
probably not that fast since it still has to call functions and return from them
but a bit faster like maybe 1.5x
or 2x
it is possible to have a remember macro that doesn't touch the d register, and it's quite fast too:
macro remember{LOAD RETVAL; LR a?}
just skip all the function nonsense and modify the return value slot directly
I think this would speed up your algorithm a lot too
you were hiding this power from me and claiming you nee dmalloc smh
I need malloc 
This is what the model outputs after training on reasoning from a larger model
the function i wrote above is wrong btw
but its close enough
I have not read on how distills like the qwen 3 0.6B are trained to be reasoning models, i wonder if this is the approach they take or if they do make an intermediate model trained without reasoning first
chills
you can probably do some crazy optimizations if you work with the bare assembly
Iirc, you can disable reasoning on that model so its probably the latter approach
an under-utilized feature of this architecture is that computation results can be written to multiple registers at once
i.e. you can save the results of an instruction to any combination of A, D, and RAM simultaneously
i'm not using a pretrained model
In this case
This model is pretrained from scratch
Yeah i noticed, im just wondering how feasible your approach is
Especially at that size
it's barely feasible
it also depends on your usecase
if you want to deploy to a production environment
of course its not feasible, but for fun
it might be usable
Yeah of course, i mean you should not be using a 100M param llm in prod regardless of how good it's made 
it's actually for "production" but with a specific use case
looks like they use distillation from the larger models for the post-training of the smaller ones
the base models are trained independently though
Yeah i guessed it would be like this, still that 0.6 is very impressive for its size
Just a little bit more efficiency and we'll soon have llms small enough to be shipped in malware
And smart enough to deal damage and self replicate
The 2026 bonzi buddy remake will be wild
Ah ok thx
Yeah, this will kill the GPU time 
its nonrecursive in the first place with tail recursion
I think your LLM is a bit broken
copyright isnt about data but about artistic expression
if i cosplay as a character and take a selfie its still copyright infringement
but if someone, without knowing of that work, makes literally the exact same character, then the two works have independent copyright
technically yes, whether it holds up in court is a different matter
mhm
how laws are being applied in reallife somehow often seems to be ignored
Well, yes but kinda no. And it gets into individual works as well.
The simple clear and cut copyright is "did you steal this data" and in that AI is a definitive no.
The wider copyright issue is "are you copying a piece of art so extensively and intentionally that it is a derivative work" and that is a much much higher standard.
In the specific case of training on voice data it isn't a copy period and the voice likely isn't distinctive enough to be an independent copyrightable work, unless you are claiming it is x character voice from x game from x voice actor. Which, don't be an idiot, of course you can't do that (commercially)
Iirc there was some drama about a famous persons voice being recreated for a chatbot.
Didn't matter whether or not it sounded like her, only that it wasn't exclusively trained on her voice to recreate it. (And even then it'd be uncertain if it could be proven that the voice is unique)
Copyright, trademark, and patent are all distinct from each other. Sometimes people confuse one another and hence why it is better to listen to a lawyer if you truly need legal advice/protection
Fair
Yeah, now if we want to discuss what is the legal status, I think it depends a lot on the jurisdiction. I know some countries didn't use the precedence so ruling will be different for every trial
Here is my legal advice: Do you have the funds to sustain a legal battle with whatever company you are dealing with? If the answer is no don't tempt it
Lego sued a company producing those 100 tons concrete blocks for protecting pedestrian zones
is any kid gonna confuse the two and get a concrete block for christmas boosting their sales? no and yet they lost in court to lego
can I get a concrete block for christmas
Ye ik
concrete requires lime, gravel, and sand, i would know i play modded factorio
so is lego
get some duplo
XDFCVGKHBKJVXDT implementing DFT in Blender compositor is frying my brain

and I even set a goal to achieve the Cooley-Tukey FFT algorithm
this is harder than my SSGI in compositor man
summation is so fucking confusing(in the context of Blender compositor)
is it 2D DFT
WHAT IS THAT MESSAGE
I think some dude named Acerola has an implementation for Unity
Oh, yep. Here is the repo
https://github.com/GarrettGunnell/Water
Imagine if Neuro or Evil
Me mounting 5 linux VM disk images on a single 6th VM having to rename every volume group and then fix 5 different grub and fstab configs 
just a typical day at work
replace the model with opensama and now the smart home has sass 
man i should really update my gpu setup, right now i'm on a 3060+1050 ti and inference on LM Studio is fine, but I do want better VR performance
i guess i can go team blue/red this generation
And also doesn't know what it's saying because it's not been trained with the appropriate info
which adds to the entertainment value
i had to do something like that once https://github.com/elitak/nixos-infect/issues/94
Whar is thar
The monthy hall problem
Never heard of it
So you have a 1/3 chance to pick the right door
The host opens a wrong door
So the other remaing door gets a chance of 2/3 of being correct
My physics teacher and classmatrs all said it becomes 50/50
NO IT DOESNT
AAAAAรAAAAAAAH
SoSilly
yea it's counter-intuitive
Its 2/3 goddamnit
I had a 30min discussion about this in class, and they still think its 50/50
Im so mad
Im actually fumint
Yeah, it become 2/3, didn't Mythbusters even make a re-enactment of it and got to roughly the same?
Did I just dox my age 
I have watched an explanation on it so many times and I still dont really understand why the odds change, like i kinda get it, but I still dont get it
Kinda ye
The one you chose initialy keeps being 1/3
So the only remaing door is now 2/3
the problem is, they probably don't know the rules
cuz "host will always pick a door with dear" is a big rule
without it the thing will be different
Nah they knew the host always chooses a wrong door
I think it's easier to comprehend if you imagine a billion doors, you pick 1 of them with 1 / billion chance, now you reveal all the other doors except 2, the price is most definitely behind the other door that wasnt opened
it makes more sense to me when I imagine it like that, but 3 doors is kinda hard to comprehend
This isnt even the worst theyve done. 6 months ago they tried to tell me rockets cant propel themselves without an atmosphere.
YOU'RE WRONG, THE PROPELL PARTICLES
AAAAAAH
this is why he teaches in a middle school and not a university
Sorry for the schizo rant lol
fuck Newton's third law
But i liked that one 
I get being confused about the monty hall problem, it's mostly theoretical and just kinda hard to think about imo, but rockets???
do they all believe that nobody was ever on the moon?
like, how do you reconcile the belief that an atmosphere is necessary with the fact that rockets clearly exist and work? 
leaving the atmosphere is the most fuel consuming part of the whole thing

that went about as well as ACME contraptions LMAO
maybe they were thinking of jet engines?
i mean there is not a single jet engine capable of reaching the stratosphere
tbf the moon does have an atmosphere, the density is just really low
even then realistically we still need liquid fuel or something else to actually get to orbit
real life is not KSP after all
i felt like neurocord has taught me more than school did this year
good morn
jet engines will still work outside the atmosphere, it's just insufficient
it's jet, netherless
no eep society 
i need my eep
It fucking works sooo good holy, it can actually detect things well with a better camera that's not 420p yooo

Neuro's memory is done using assisting systems, not part of the model itself
Neuro is 100% not custom pretrained, only heavily finetuned for being a Twitch streamer, based on some open-source base model
I play ksp with real solar system in realistic progression 1 (RP-1), I landed on Mars and stuff with humans, was cool
...2
android studio has installed kotlin on my pc 
Peak
Kotlin is the Android language and by far my favorite language
i will actually not be using it
i just needed to install adb, and that comes in android studio
Silly
Did you use windows? There exists another download page separate from the android studio
Kotlin is peak though
it doesnt you can just install platform tools
android studio is gigantic
people used to get jumped for saying shit like this
why keep it around if you don't need it
i implemented this library just to avoid depending on android studio for one little debugging thing https://github.com/chayleaf/android-trace-log
Rust Android trace log parsing library. Contribute to chayleaf/android-trace-log development by creating an account on GitHub.
android studio doesnt have adb?

android studio comes with a sdk installer that can install android sdk or ndk or platform tools for you
Nope
fuuuck
but you can just install them without android studio
why would chatgpt lie to me 
paid by google
Have you not seen our cute AI overlord?
imagine watching streams in 2k25
And just as that one person in the other chat was trolling brutally lol
Chatgpt lies???
WHAT
My openai investment money๐ญ๐ญ๐ญ๐ญ๐ญ
Agi was not real :(((
Advanced Gaslighting Imitator๐ข ๐ข ๐ก
It's not imitating if it's real gaslighting 
at least the adb is working now 
adb should not install android studio
what the fu
Aaaah
I love
and
but dev streams are definitely my favorite, surely we can get a dev stream next monday


Whar is thar?
candy
That's crazy
do yall not have haribo?
No clue
you know what a gummy bear is right?
Maybe
we have haribo in finland
Whuh
ye i think so
damn it, can I have some please, for research purposes
i went to the haribo hq as a kid once, it was so fun 
so i have found all the plqatform vzariable savailable to me in the web
the problem is that the console overflowwed
hopefully theres something like 'tablet' here
look at all this yummy metadata
You're cooked
target spotted
most of this seems to be emscripten/webgl stuff
but i can get device orientation and cool stuff like that from here
'screen', 'screenLeft', 'screenOrientation', 'screenTop', 'screenX', 'screenY' 
niceee
... Titan sub be like


i may or may not be going through your town on google maps streetview
Om
10/10 ice cream shops btw
not too shabby

too bad im never gonna visit
Damn bro sadge

The only issue left is these pups on mobile saying something broke
cuz it didnt broke
its actually working prett well now
its beating python so who cares that it isnt beating rust yet?
Have you seen strager video? He made a perfect hash table in C++ and it beats rust
is a foldable a phone or a tablet?
or does it switch inbetween the 2 when folding?
i mean like what google reports to my program, not the definition
I found it why Kotlin coroutines are a must.
Found the video
https://www.youtube.com/watch?v=DMQ_HcNSOAI
he made a perfect hash table for one very very specific scenario
you can't compare that lol
He has assumption about the data, yes. Is the data inserted into the hashmap on his case truly random?
comparing a hash table that makes assumptions about the data that is going to be stored within against one that doesn't is kind of pointless
is a nice showcase of how much you can gain for making assumptions though

Look, I don't know the problem that the other dude is trying to solve. But the step that strager take applies when processing data in general
Yeah, and looking at that video again reminds me of the old PHP days when string length is indeed being used as an index.
(also for some context konii is working on a general purpose hash table for a stdlib so can't make too many assumptions sadly)
Ahhh, then yeah, making hash table ahead of time is not possible and collision is a part of the assumption
google colab should rly put a cache
between pip and the instances
cause the same n packages are constantly downloaded

OOBE\BYPASSNRO my beloved ๐ฅฐ
I like switching code editors every two months
This time it's zed
as long as they have vim support
I still have to try emacs evil mode
hii ash
!!! shadow's alt ego is here
zed is nice but extensions being in rust means there's just not that many unfortunately
And I guess it doesn't quite work under wsl yet but that'll be solved soon enough I'm sure 
Yeah the biggest downside so far is the lack of extensions
ye
Oooh and wsl is something else yeah
Last time I wanted to try it I think they didn't even have a windows client
You have to compile it from source for windows but even then it's far from perfect
i hate compiling stuff from source for windows
You can definitely tell that it was made for macOS and linux primarily
It's fine for rust
Usually
I seeeee
not the floating point precision
yeah-
Did you not use the same code for both platform?
You can technically connect to wsl via ssh and that works
@stark needle why are shadow and light different quantities
But paths get fucked up
it's been a while since I last tried tbf maybe they've improved it
I know they have their headless server now sort of which is nice
it happens
Did you code in assembly? If yes, then understandable (especially considering you are after highly optimized solution). But you can cross-compile on C, so technically one code can be compiled to different target
Niuh

Why's the print target specific actually
Or is it the data that's being printed that's different
Ah
Got it
windows target when 
i dont think anyone cares about that
but its there in the compiler, probably not that much work (you just have to use the correct image format) but who would use it
can I pretend that phase does not exist
Makes sense
have you tried changing step to 1 again?
probably still not worth it
i keep winning
swisstables are very hash sensitive but i guess changing step size alleviates that
google just wrapped the default hasher with their own in their impl i think to make all hashes good
it would be best if step size = 1 worked since thats the easiest to optimize further, but i guess the hashing problem needs to be solved for that
please reboot more often unless you want kernel exploits
If it's windows, don't bother. Microsoft will reboot it for you when you least expect it ;)
(I deleted the reboot scripts)
Whuh
Haven't heard of any important vulnerabilities recently
no 
I'll probably reboot at some point I guess
And try to get the RAM to 4000 MT/s
RAM SMOL?
all for rendering photos of anime women
Man, I need to sleep. Anyway, good luck optimizing the hash map. Also since you are developing cross platform, do take in mind that different ISA will have different performance for the same algorithm implementation. And also the clock cycle for instruction does not tell the whole story
@stark needle what is happening to your name
why did google removed cache search feature
it is like the only reason why i was selling my soul to them
yeah crypto is super easy to get wrong
you mistype a 36 digit long number bam your impl is now broken
clearly must use bad hash for optimal performance
i tried making it do more batched reads but it didnt really help no matter what i did, i'm surprised it isnt slower though
i even did stupid stuff like precompute 8 shifts of short_hash, which ended up making performance worse 
probably because it forces stack to be used rather than registers, or something
havent looked at a single line of assembly in the process of working on this
note that i also did this
Meta := union{.i: uint; .b: [8]u8}
HashMap := fn($K: type, $V: type, $A: type, $H: type): type return struct {
.metadata: ^u8;
.metadata2: ^Meta;
and made metadata be set to the same value as metadata2 for old code to work
it looks like there arent big improvements lying around without compiler improvements
well, except for evil stack overrun to arbitrary code of course
maybe if there was a way to do the comparisons faster...
like i tried doing meta.b[j], still no dice
also removing the entire hashmap only takes a second so its hard to benchmark
with the variance being around 100ms
actually what does hashbrown do (probably simd
)
it uses quadratic probing apparently
it does
probably depends on autovectorization?

the other approach is this which is fairly easy actually
n & 0x8080808080808080
but, you need to be able to find the trailing one or something
you need to be able to find the lowest 1 bit in the integer
and then we have to write an order of magnitude more code and we suffer 
okay i give up on sleeping this night i wanted to "fix" my sleep schedule (enough to go to sleep after neuro streams) but i think the more realistic option is staying awake (now that i have neuro stream as motivation i wont go to sleep before then i think)

zig is easier to translate to
since hblang is basically zig
if you want C it makes sense to specialize to ints only since thats what you're benchmarking anyway
I cry every day looking at Nvidia 5080 specs
wanted to get the text in the reply instead of the emote. ๐
also i have a useless and pointless but doable optimization for you - replace two allocations with one
meta at the start and the rest after it, or vice versa
yeah
it does technically help but who is gonna use like a billion hashmaps for it to matter anyway 
real
the thing is it doesnt save memory but it does theoretically help with fragmentation
and make the allocator use like 8-16 less bytes for tracking allocations
Wjo pinged me I'm lazy to scroll up
There's quite literally an inbox that lists your pings
I'm a complex mixture
where
Top right on desktop, bottom on mobile
Doesn't that just show global role pings

in rehash you dont have to deal with it because you just copy the bytes
all you need to do is not be C++ and not have move constructors
yup
does deinit free pointers?
thats maybe a deinit api issue
oh but pointer.deinit() wont compile so its fine i think
unless v is a pointer, right
wait wait
if T is a pointer type
then T.deinit wont exist

well, i guess you can just add a check that doesnt run deinit if its a pointer
thats surely the only special case where deinit exists on a type not explicitly declaring it
(thats why implicit dereferences can be dangerous, oh well)
yup
or create a wrapper
yes but that doesnt mean (^T).deinit is a thing does it
it means T.deinit takes a pointer
i was in fact right
this doesnt compile
and @compiles((^A).whatever) returns false

i will now break the universe
i'm sometimes wrong
the laws of logic forbid me from existing, goodbye 
probably hblang
Does anyone here know anything about pygame
depends, a couple people probably do
Ok
I donโt have well functioning computer so I canโt download anything on it so Iโm stuck using codehs for pygame. But when I try using audio files it just sends a link instead of the audio
I can try
its hblang
I swear you're the only people I've ever seen using hblang
true!
I thought it was a custom built language
if it's on a website use JavaScript, otherwise if it doesn't need to be fast use python, otherwise use rust
do you really need anything else
thats because konii recruited me into her cult 
Use wasm and cut out the js
Also I hate dynamic typing
actually I say that but I've been writing assembly for the last few days
wasm is mostly a compilation target
oh thank god
Try ot the examples
want to point out more? go on you'll be here all day
that's how you know the language is good 
prompt injecting Google AI?
oops, looks like it tried to say something it wasn't supposed to
chills
so heres a fun question - if i wanted to record an RGB value in gpu memory, how computationally expensive would it be to take the MSB of each RGB value, multiply it into a single digit, and writeback a position/vector/object value to the same address space later?

Reason behind that would be to simplify the image, group pixels of relatively similar values, categorize each group as a different object, generate a simple vector outline of each object, and then compare frame by frame to determine motion between lists of closely matching objects
you might want to look into convolution filters, as they can do this much better (at least in terms of edge detection
CNNs basically?
https://en.wikipedia.org/wiki/Kernel_(image_processing) wikipedia has one for edge detection
In image processing, a kernel, convolution matrix, or mask is a small matrix used for blurring, sharpening, embossing, edge detection, and more. This is accomplished by doing a convolution between the kernel and an image. Or more simply, when each pixel in the output image is a function of the nearby pixels (including itself) in the input image,...
oh no, nothing that complicated, just an image kernel
aka "take the surrounding 8 pixels, multiply them by the given numbers, divide by 9 and set this pixel as the result"
Ahh, now that I think about it, that's a lot faster than trying to do every pixel, even with the color values severely truncated
Although I was thinking about ways to implement fast tracking of lots of things, and being able to build up object permanence / position tracking
Although my programming skills are basically "Hello World" with a Heathkit
ff374f727ce9d6de93be0793733c321 
its 3am
sleep bro
i have slept
That range is wild
programming time
no, sadly not
In the context of hours slept
its 8 AM here

its 9pm here
Oh, you're in the same timezone as me
damn the timezones
ohhh canada lol
wdym, im in asia
im talking about me lol
oh
my code isnt working anymore... I hate vision ai now
ai slop 55% time doesnt work
its yolo ya dummy, custom trained and custom pain
If your model is not invariant to lighting, rotation, and translation then you're cooked
45% it works but have no idea why and changing anything breaks XD
oh wait mb
I added that, I mean my coordinate system to gather the position of objects and translate that some how to something a robot arm can do to pick that up...
because yolo being a mean guy rn
Did you use a controlled light environment?
yeah it detects it... kinda, im using a old windows 7 laptop tho to run it lol

Oh, does not work as in throwing error and not in the sense that it stopped detecting?
more so yeah it randomly stops detecting it, my code requires a stable detection of like 4 seconds to confirm its an actual puck incase of misdetections... and my laptop can barely run it so its not working
(the puck is the thing the robot arm is trying to pick up btw)
Yeah, I've seen your setup. Just like my old mechatronics assignment. From my experience, the main culprit is usually the light condition. Also, you are better off using image processing for simple object. Only use YOLO when needing to detect actually complex object
If you want it to be consistent, install a consistent light source that will be unobstructed. Usually that does the trick
yes ma'am
But yeah, do check it first if it is the light condition or not since you're using YOLO.
Im using yolo since there are multiple objects it has to detect and stuff, I can shine a massive light to keep it consistent, I think my teacher has a old broken flashlight I can steal the LED from and hook up to a battery
Yeah, the pucks are in different color right? YOLO is a bit overkill IMHO. BUT, if the target is an actual object later, then maybe yeah, having setup YOLO may be beneficial for the long run
its not just pucks, its tennis balls and other objects
Ahh alright
its due in a week and a half im cooked
I have most of the code done I think, I could maybe try running a contrast corrector and run each image through that before sending it to the yolo model to get its detection?
Do you need an object that is not on the YOLO dataset? If not, that is plenty of time, don't worry
my data set is 6k images, 2k of each object I need
its just debugging code and stuff
Yeah, that would help. It should be tolerant to brightness variant to a certain degree. But in general it helps if the light stay consistent
Oh, so you do fine-tuning. Then yeah, a bit worrying
time to use a flashlight and dangle it from the ceiling lol
Its not working my code... I have 40 errors, Ill fix it tomorrow
a wise man once said

im currently writing an interview about fictional characters in a book i only read 50 pages of 

i have 2 hours left to complete this
time for spark notes
chatgpt is my friend
om
fair usage of chatgpt
i do genuinly like the book, its jsut that i havent had any time to read it
fair tbh
the underlying themes are a bit dark but its quite funny
I think for my class i spark noted like half the stuff for books.. because I was too busy coding
the book is called "In the Lives of Puppets" i think
is not being in school old? 
it can be i guess
or is not having used AI in scohol old
that lol
anyway when a cashier sold me alcohol without checking my ID i died a little inside

is it just me that hates emojis in code in this chat?
or is it a universally hated thing?
...what?

const hello = 'Hello  !'
I sometimes comment and say "good luck optimizing this :)"
HOW DID YOU EVEN DO IT?
so when i compare the codebases it ends up being 55% emojis being different 
Oh, I thought you were talking about emoji in discord chat code block
emojis in an actual codebase
thats some psycho stuff bro
Yeah, that would be a nightmare
you monster
and some of the logging statements are so hilariously overengineered that they rival terminal graphics XD
not mine, someone else 
XD
Death, only for strings that use emoji otherwise do not use those in comments
I use those a lot in my code
those are fine
๐ is not alright to be using
i got forced to configure a UTF-8 locale when i only wanted to stick with ASCII only because the logs contained emojis

Bro, why do you need ASCII only...
cuz a, ima lazy, and b, its lighter
Understandable
also the codebase comments are entirely in chinese 
OOOHH guys my teacher has to read my code, do I comment it all in chinese or something stupid?
zalgo
Honestly, that's fine
Comment in chinese if the person in chinese just means computer is working for everyone around the world
peak
But THIS is not fine
that is actually compilable swift
swift is the only language i know off the top of my head that lets u code like that
i mean those emojis are valid UTF-8 characters
I mean, any language that allows UTF-8 for identifier should theoretically allow that behavior
Emojicode is a full-blown programming language consisting of emojis.
this person is going to hell
Bubble sort

******** languages usually use XID_Start and XID_Continue unicode classes (is that the right word?) for determining what characters identifiers are allowed to start and continue with, and i dont think emojis are in there

As in programming language. Unless otherwise noted in the spec, if the spec only said "identifier must be a valid utf-8 byte sequence", then even whitespace could be allowed. That's why if you want to know a language could use emoji as identifier, just take a look at the spec
yes these are character classes, or properties
note that identifier and how it's represented in source code are technically different things, like a string "a\n" doesn't actually contain a backslash
I know. What i mean by identifier is like variable name. For a programming language, unless the spec said otherwise, the internal representation for identifier doesn't usually care about how the identifier is named. Just that it has a name
This is what I mean by identifier
https://en.wikipedia.org/wiki/Identifier_(computer_languages)
And about the string escape character (the "\") is also another matter entirely
it isnt, because there can be similar syntax for identifiers
There could be, yeah. But then again, most compiler or interpreter doesn't care about the specific value
"none of it worked but boy it was beautiful"
Vibe Coding in a shellnut.
Also "none of it worked but it was beautiful" could also describe sanding your tires to get rid of the grooves. It'll look nicer, but boy will you regret it.
mhm caring about it is the job of the parser
there's something wrong with this piet
@scarlet arch
(not mine)
Exactly. After parsing, there is not difference between ' ' and 0x20
maybe i should start low level langs
saying it as if there are more than 1
theres a billion of low level languages but its hard to define low level language because you can do high level or low level programming in most languages
its more useful to learn different programming paradigms and apply them in any language
because even though python uses an inteperter, it can do low level
i have seen python code doing system calls
i wrote code before that creates virtual input devices at kernel level using python
there do be having bindings for everything
message passing (smalltalk/erlang), imperative (C/Python), functional (Scheme/ML/Haskell/Idris/a billion research languages cause functional languages are easy to reason about), logic programming (this one is not as immediatley useful but knowing how datalog and prolog work can be very useful since a lot of stuff is just computing more data from some input data which is basically what datalog does), stack-based programming (assembly and forth), metaprogramming (mostly lisp but also most languages have some metaprogramming facilities)
yes, python bidings for everything
lmao hyprland has a python sdk for me to screw around with
but most languages offer some kind of mix, like java is imperative + message passing (oop)











low level code

