#programming
1 messages Ā· Page 110 of 1
you could always make a database yourself
i did find this
is this it?
cuz i feel like its not the same thing but it did pop up when i was googling for it
turtleSpaces is an enhanced, modern implementation of turtle graphics, building upon the original Logo principles but extending them into a full 3D environment.
i guess its kinda what you're looking for
Note: training LLMs on data from Discord is against TOS
That implies turtle graphics is 2D, which may be what the dude wanted
Because he is talking about plotter or smth
Axidraw
The data is not generated by other person on discord tho
No need to label datasets for LLMs
Discord has a legal method to request your own data
Seconded, it is only the user in question
Ok so like
I don't want the LLM to run arbitrary python code on my computer
hopefully for obvious reasons
So I was thinking of giving the LLM a more limited set of turtle graphics commands
like it wouldn't even have flow control, the LLM will do all the flow control
so I was wondering if I first need to translate all the training data into the bespoke turtle language
or if I can just give it python and the LLM will be able to automagically do it
I found a good amount of training data to start
you thought traditional LLM art was bad. I am plagarising literal children here
This is not professional advice, but you could play around with limiting a python exec() to filter out unwanted commands.
I'm not even convinced the method I use is in standard
If I weren't on the road I'd post a code snippet as I basically ran into the behavior after experimenting.
I was thinking of scraping some tutel graphics programs in both python and logo. Then putting in the system prompt a new turtle graphics language and let the LLM itself do the translation
does that make sense?
For those unaware on what I recommend abusing.
exec has the signature of
exec(str|bytecode, globals, locals)
Normally one using a globals() instance here, but it can also accept a class laid out in a similar manner. Locals can be a dictionary (passed as a reference, so nice for keeping state)
More on the globals, you can remove stuff from builtins. Granted this alone does not handle imports so additional work is needed.
Does anyone have recommendations for a web scraper?
I was going to write one myself until I realized that was a dumb idea
If you train it with tool calls and give it a proper prompt, maybe
I'd say it is possible if you get it to annotate variables it is using, just to be able to stop or restart as needed. Python does have annotations it can evaluate and store as objects to recover later, so it should be possible?
I'd say it needs some proper calling and very good prompting
To handle imports, I'd just import all that is expected and merge them into the scope; then just ban all imports.
Aren't you the one using it, you tell me
š it aint that deep
i am not even trying to jit it
cudnn runs a bunch of stuff in the background at startup too to find the best implementation
Or was it cublas idk
One of the two
wasn't xla for tpus or am I tripping
its universal

but for cuda devices xla uses cudnn i think
this is all the algorithm being tested for a simple model with 6 convolutional layers
help
it can but doesn't have to I think 
at least that's the case for their SDPA implementation, it's likely that way for everything else too
I mean, it says cudnn right there in the log
I think you can set the xla_gpu_autotune_max_solutions XLA flag to a lower value
down from 288 unlimited to something more sane
-# nvm that was a different value, apparently the default value is 0 = unlimited???
alternatively disable autotuning entirely but that seems bad for performance
Like does hundredfold vram makes sense while other 1D conv is like only three times the original param size
Release me from this despair
ok so my plan of rewriting my code to work in an action has so far been 90% writing actions files
(and 10% writing scripts for those actions files)
behold the pinacle of programming
are you really doing ci if 90% of your entire codebase isnt just yaml files

you need more inception. Have ansible yaml file setup the ci environment
"my code is correct, it is the action that is incorrect"
probably claude or something
now to actually write the rust
don't tell daddy microsoft but I have a github action that creates a lxc VM, configures it with ansible and then runs tests on it
it might sound crazy but it actually makes total sense
now to figure out if I can make it so rclone runs a command after every file it downloads
pip install torch==2.7.0 'torch_xla[tpu]==2.7.0'
not going to happen
torch is bad
true
too big and messy
something can be bad and better than another thing 
I'm going to dual boot windows and linux
Do you always run on older versions of python?
I usually use 3.10 or 3.12
I use 3.12 myself
fair, i don't like using cutting edge python versions especially for ML projects
I'm kind of new to all of this and python, but once I realize it's pretty much Java I got the hang of it pretty quickly
you mentioned the forbidden language in #programming
now a lot of lost programmers would arise anytime soon
It's not even the first language I've learned
I think the first one I learned was basic
Probably
ā ļø
It was taught by a grade 8 robotics teacher
I learned Java in grade 9
I've been crackalacking since grade 10
do I make my action 50x more complicated by trying to involve caching or do I instead simply not care as I am only attacking google (and wasting gh actions compute time)

if you care about ci execution time you probably should do it, otherwise maintenance burdens are bad is what i've learned
its an action that only runs after every karaoke so

unless vedal does like 31 karaoke streams in a month I shouldn't get anywhere close to the limit
and that's also assuming it takes a gh runner about an hour to make the file, which I really damn hope it doesn't
I've been thinking of many different ways on doing this, but I'll share once my guy is fully set up
cache := fn($fnc: type): type {
Inner := struct {
cached: ?@TypeOf(fnc()) = null
}
return fn(invalidate: bool): @TypeOf(fnc()) {
if invalidate Inner.cached = null
Inner.cached ||= @inline(fnc)
return Inner.cached.?
}
}
func := cache(fn(): i32 {
i: i32 = 0
loop if i == 1 << 31 return 100 else i += 1
})
main := fn(): i32 {
i := 0
t: i32 = 0
while i < 10 {
t += func(false)
i += 1
}
return t
}
rudimentary cache
comparison to uncached vs cached
Python f*******ing sucks (and yes i had to escape each '*')
using a single core on a 128 core machine
There is no async in python 
there should be
GIL moment
classic
The GIl, or Global Interpreter Lock, prevents the Python interpreter from concurrently running code on more than 1 thread at once, presumably to avoid race conditions
Makes multithreaded processing not such a thing in Python though
Python: Job takes 1 core
Me: Implement Multiprocessing
Pytohn: Copying results over to main process takes one core
exactly GIL moment
I JUST WANT TO EXECUTE MY WORK 
Then use a more sensible language like Kotlin that allows for multithreaded code
ye, all kinds of multi core procesing is kind of painful in python
or Rust š¦
rust people are everywhere
I should learn Rust at some point
I do love rust, but I dont want to spend years of coding for a single conversion task
Simple Problem: Convert 11 TB of data into different 11 TB of data
I use Python only for ML tasks, other than that I use other, more sensible programming languages that actually have a type system
python does have a type system
shut up
Not really a great one, it's terrible
there's worse tbh
True (JS)
at least it is built in now
did you know you can bake bread without crust using electricity
unlike js

JS really sucks, the type system essentially doesn't exist, there's not even a differentiation between ints and floats
aaaaaaaany day the types as comments proposal will go through 
there technically kind of is
but that's only for arrays
this is why hblang would be optimal for screeps 
unless you want to count v8's small integers
maybe i am just not built for using python, maybe i am the problem and i have to adapt to the language

might as well just solve it by shoving 'rm -rf python' up tho commandline
i wrote 17k lines of bullshit
same :/
for chatter in chatters:
if randint(0,1):
del chatter

i lived

we've at least tried
modify your $PATH variable to not include the dir of which python
termux 
big deal i can bake crustless bread using a knife
you better have a sharp knife
I'm out of the house 
command -v
make it look like this or im getting a refund
1000 degree knife vs bread

time to restart termux
or just apt reinstall python
awa
awa
awa
awa
awa?
I mostly use the other one for python anyways
termux still functions at least
nominative determinism suggests you will become a leaf at some point in your life

bark bark?
what is she
what
carcinization suggests i will become a crab
crab is good
decomposition suggests you will become soil
and then leaf
Did I ask to be a bot?
eat more frequently 
i eat as frequently as necessary 
necessary doesnt mean appropriate 
function pointer
true
WHY IS IT SOOOOO HARD TO FIND A USB TO INSTALL LINUX
all my usbs are at my dads. im at my mums...
EtchDroid may be something. Somehow it exists
then its going to take 2 hours to copy linux onto the usb great.
Mystery on how plugging in a device with that running becomes a usb drive with a ISO loaded
i have apple...

very based notifcation
Async rust is painful
Or maybe I just used the wrong IDE
ide doesnt affect the language experience much
It'd be more like what async runtime your using, but also async is painful by default.
async (in rust) is just state machines
depends on how low you are willing to go
for the average person using async its "add tokio, await everywhere"
i love state machines 
until they hit a "type T is Unpin" issue and curse themselves and their computer
deserved tbh
It is when you only get error when you save the file and like wtf?

sounds like someone needs to turn on autosave
only error linting on saving file how will i survive
oh did you use rustrover
just use vscode with rust-analyzer
add := cache(fn(args: @Any()): i32 {
@syscall(1, 1, "computing...\n".ptr, "computing...\n".len)
return args.x + args.y
}, struct{.x: i32; .y: i32})
main := fn(): i32 {
i := 0
r: i32 = 0
while i < 100 {
r = add(false, .(1, 2))
i += 1
}
return r
}
cache with args
zed uses rust-analyzer too iirc
where's my LRU cache 
not really useless since the principle can be extended
I know VSCode would handle non-saved file so idk what going on with zed
im not writing an lru cache
Inner := struct {
cached: ?@TypeOf(fnc(@as(Args, idk))) = null
args: ?Args = null
}
this is all you get
you're writing everything 
step 1: write an entire ecosystem
step 2: you may now write the language server
in order to write an lsp, you must first write the universe
cache := fn($fnc: type, $Args: type): type {
Inner := struct {
cached: ?@TypeOf(fnc(@as(Args, idk))) = null
args: ?Args = null
}
return fn(invalidate: bool, args: Args): @TypeOf(fnc(@as(Args, idk))) {
if invalidate || Inner.args == null || Inner.args.? != args {
Inner.cached = null
Inner.args = args
}
Inner.cached ||= @inline(fnc, args)
return Inner.cached.?
}
}
the mechanism here is funny
cached and args are global scope because they are inside of Inner
they arent struct members, just free variables
oh im wasting a byte
you noticed 
Maybe it's not a great idea to write Cloudflare worker in rust...
reading is hard
ezhd
true
i actually wasted two bytes
ideally i would also save the hash of the args rather than the args themselves but
boring
this means you can make cache return a function with the exact argument count required
it will just be a bit repetitive like tuples code in rust or haskell
$if @len_of(T) == 0 return fn... else $if @len_of(T) == 1 return fn...
something something linked lists would've fixed her
len: enum(uint) { .one; .two; .three } = @bit_cast(@len_of(T))
$match len {
.one => return fn...,
.two => return fn...,
}
hmm yes enums
enum

yay my usb was created with this name
UBUNTU_3457345743584375252357743897583757320527520754375275234353572357257359257235725247242275298572946244724326424692347242342492642596353904540563052340234723473247247265236582354723523572357623562353625235636525623956235235623472497235957295623956294295252035
and this is why i just have a 64gb stick with ventoy that i throw isos at
i have a 128gb stick on my keys...
https://www.iventoy.com/en/index.html huh i didn't know this was a thing 
iVentoy is a new netboot solution.
im going to install ventoy later ig
i done that with a nas but i wiped the nas on accident...

Uhh whereās my usb also boot a iso file doesnāt work
Is it weird when you shutdown your pc and thereās just no noise anymore?
i sure would hope that a pc that's off doesn't make noise
The Linux experience
what if i put a bluetooth speaker in it
THE BLUETOOTH DEVICE IS READY TO PAIR
Enjoying my miserable life rn with my laptop going insane
why does my gpu suddenly drop frames on youtube like crazy when i open powerpoint, wtf is this 
Maybe Google want you to use Google Slide
i'd be down but my uni account is linked to microsoft so that'd make things unnecessarily annoying
Does anyone remember what the code editor that Vedal uses?
THAT is the one... I couldnt remember lol
I personally use VS community 2022, but with my new PC going live next week, I wanted to weight my options
I use it because I do a ton of AB admin work with Oracle and MS SQL and I can integrate it into Visual Studios
I hate having to flip back and forth between windows if I can help it
yeeeeeah
that and I use a metic sh$% ton of PowerShell and it can run natively on Microsoft products. But I am also always looking for less bloat ware which VS is
be that person running powershell on linux 
but with my new job being Linux based, I can get away from MS products
That is my next task, my team lead want me to investigate it
i've never actually looked into how usable pwsh is on linux tbh, maybe i should
why do that to yourself
suppose they can just strip out all the very windows specific things that don't have an equivalent
i'd take pwsh over bash any day of the week if it wasn't so miserable due to posix compliance
powershell is awfully verbose for my liking
maybe its good for long scripts but at that point just use python
i can't argue against that
i understand using powershell on linux if you have an existing codebase though
there's nu if you want to be that person 
PowerShell is best use case for using on Windows so if you have a script the needs to run on both... then PS would be ideal...
i should probably do that, just use powershell for all my dotfiles scripts
bash kinda works on windows too though
i'd rather just use something like https://github.com/luke-clifton/shh or something custom at that point
of course this exists
yeah, in a limited bases. I wrote my own modules and classes in PS so it helps, but the data I am pulling is from VERY large research Oracle and MS SQL databases
I have even begun to write my own launchers for applications/Renpy games in C#, but I just started doing that
Ig using a incomplete framework in a type strict environment is a terrible idea
i need help with linux
any specific questions?
and then I spent another 5 minutes confused why downloading the supposedly latest installer from LW's site doesn't work; launching the browser would give me "you're trying to launch librewolf with a newer profile that cannot be run on an older version of the browser"
because it turns out the download page on WL's website does not offer the latest release build
it's two major versions behind
you have to go their gitlab to grab the latest version
what is this browser man
the windows version is kinda second class they used to only distribute it with package managers
Was wondering something while watching a karaoke vod, does anyone know how Evil's background works from a technical standpoint? The outline effect is really cool and I've been thinking about how it works. I'm assuming it's a greenscreen that creates an outline of evil, mirrors it, then creates the red shilloute every X amount of seconds + random red slashes for extra flavor. If anyone knows what software is being used I'd love to know!
I don't feel like that's an excuse to not update their download page this could have easily been added to a CI pipeline
especially when their gitlab release tags already looks like it's part of one
this is what it looks like on their website right now
I'd have to imagine this was automatically updated at some point but it broke
it's probably custom software
custom shaders and such
Apple awarded me only $1,000 for this bug, I should quit this bug bounty thing and get a real job.
Quoting āRenwa (@RenwaX23)
ļø
CVE-2025-30466: Safari <18.4 UXSS to bypass Same-Origin Policy with CVSS of Critical 9.8 š“ ;)
ļøļø
ļøļøsupport.apple.com/en-us/122379#:~:text=Impact%3A%20A%20website%20may%20be%20able%20to%20bypass%20Same%20Origin%20Policy
lmao UXSS for 1k
absolutely pitiful
is plateauing loss in only a few epochs a bad sign
I just managed to wipe my whole D: Drive containing the whole Neuro-Sama chatbot python script and all my games while trying to install Arch
And I told arch to install it on SDA2 not SDA1
When installing Linux distros with more complex installation processes, I would recommend unpluggin all drives but the one you want to install to
I also lost my background for my stream...
Drives getting wiped accidentally is exactly why you unplug them
I tried to install it to a 60gb partiton on the drive not the whole drive so thats not a option in my case,
Then it's just what's called "user error"
I'm going to join a vc and share my capture card and mute i guess
for no reason
lmao
im training a discrete vq vae on my gameplay data
So what does it do?
Also, gaming models should be trained with reinforcement learning, not pretraining
it compresses the gameplay data into a latent space and learns to reconstruct them to be mostly accurate, and also stores codes of images and sequences (like keypresses mouse presses etc) into a codebook as a form of memory
its part of a bigger system
im trying to do the dreamerv3 system
Do you have a separate validation and training set?
im training it separately from the world policy because i was having trouble training them altogether and the loss weights messing with eacvh other
ytes
but i havent even gotten a good result yet so idk what to do when the loss plateaus, i've tried a lot of stuff but it just seems the loss plateaus too fast
Does the training loss keep going down and the validation loss flatten out, or the training loss flatten out and the validation loss shoots up?
we sort of kind of evaluate the model already by measuring how far off it is from the reconstruction by the mae of the pixel reconstruction, and that looks good on paper
it got a lot lower this time around with some hyperparamter tuning, but the loss still plateaud so im worried about that, i'll have to see when i train the policy with the vq vae to see if it really got better
Either your loss method is bad, you should not train it for so long, or your data is too small
it might be my data, its only 1 hour of 20fps footage, but the footage is pretty diverse and full of action
im sharing my hdmi in general... for like no reason im about to boot into the install of Arch
and the vq vae is pretty small too so I'd think itd not overfit
For comparison, NeuroSynth takes a couple thousand epochs to flatten out, and is trained on about 1GB of audio
https://www.youtube.com/watch?v=49CrynNZAR0
And the end result is absolutely not too bad
collaboration with Superbox, this makes it the first (unofficial) Neuro-sama cover of BOOM!
come chat w/ me on discord if u like! discord.gg/8ry5dn3UjA
check out the instrumental version! - https://youtu.be/KSxZScNeJBo
NeuroSynth-BETA-3 developed by Superbox & Wispers
Motifs:
BOOM (Evil) - https://youtu.be/8Nt9YPnYyjs
Only present in instr...
sounds awesome
i have 2400 2 second chunks of gameplay at 20fps, so about 80 minutes of gameplay, so about 96,000 frames
at 128x128
128x128 seems a bit tiny
Are you sure that's enough to capture sufficient detail?
If I could only see 128x128, for this particular game it'd be okay
What kind of training hardware do you have by the way? I run NeuroSynth runs on a 3090
I have a 3090 too
What game may that be, if I may ask?
a roblox game
"infectious smile"
its fairly linear
but complex enough to warrant a world model and stuff
Try increasing the resolution to 240p, and see what happens
Perhaps
im also worried about being able to inference the model at 20hz in realtime too
If you don't have enough detail to require the model to accurately reproduce what you do, you'll just end up with the model overfitting some bad pattern that gets close enough
Parts of that kinda look too grainy
Too much noise needs much more data to eventually average out of it
Especially the walls
honestly I"d have thought keeping the resolution low would have made things easier
Too low will not work
low res never makes things easier, it only makes it computationally feasible under some circumstances
You want to get a balance where the resolution is high enough that it accurately captures detail while not being so high that it's unreasonable
but yeah you probably want to smoothen this somehow 
I recorded my data in 128x128 which means i kinda have to rerecord it again, to avoid this in the future what res should I record at
For audio models you want to use as high quality audio as possible, in the case of NeuroSynth 44.1kHz 16-bit
Just capture at native and downsample with ffmpeg
good idea
Iām going to install arch then let my games download over night (in windows)
Uhh I need help I just installed arch and it booted to Grub terminal
I done the easy install
Ohh itās booting now but thereās no desktop so ima download that
hello programming individuals
two days have past
and I am still stuck on OOM while jitting
Try using less memory
Have you checked its size in parameters and bytes per paramter, wheather the dataset is loaded into VRAM or not, and how much VRAM you need for batch size? Among other things like if there's attention then that needs some VRAM too
the if __name__ == '__main__': with single 32 batch size tensor emulating 7 seconds of 24khz audio is all I have and I get OOM. This is not about VRAM yet š
Let me check vram usage per iteration though
so that Multi-STFTD architecture makes me have three different models
each having 6 convolutional layer
but parameter-size wise, that is still crazy to have OOM
When training a model, batch size is what takes the most VRAM usually, or if you load the dataset into VRAM then that
Basically any time I see ML models it's all PyTorch
Or ONNX if only inference
If you don't specifically need something JAX provides, you should try Torch
big no!
you are being possessed by using OOP in ML
this is for TPU, but XLA is universal so the code that works in free 2x 5090 server must go flawlessly well on TPUs
I don't know, it works really well, so what can you do
fair point but JIT and shit in torch is ass
due to its too many implicit global state changes
are you using lightning?
And what may that be?
i've heard that this is the relatively more industrial standard using torch
cause that un-reproductible state of implicit PRNG keys changing every run on default torch
fuck my phone is also dying
shit is not going well this week
i like lightning, is nice
takes away a lot of the annoying boilerplate while not being too restrictive
I like raining before I moved, as I can hear the dropplets hitting my window and making some pleasant white noise
now they only hit on my AC's compressor unit, and making metal hit sounds
buh
rain on the window is one of the better sounds, metal is nope
it's inconsistent as well, it's in irregular patterns
which made it much more annoying
There's a tin roof not too far from my window so it does get very loud sometimes when it rains but hey, at least it's consistent I guess
i was wondering why the hell my code started acting up after i fed it to claude
turns out it thought it was being helpful by turning one of the two if statements into if and elif

context
if is_valid_hostname(artifact):
entry_target_type = "hostname"
if is_valid_ip(artifact):
entry_target_type = "ip"
it thought it was being helpful by making it
if is_valid_hostname(artifact):
entry_target_type = "hostname"
elif is_valid_ip(artifact):
entry_target_type = "ip"
If you flipped the branches it would be, unless the checks are stateful
it was being stupid because it doesn't understand the output table it is trying to map to
very basic stuff that the LLM should know if it had a brain
"I'm not wrong, your test is" - claude, after removing all the tests that raises issue
Name a programmer
i hope my jax tracer can tell me if this is a cudnn issue
I believe in shuni's wise words


its just 3070 but yea
i think that messed up my pc
probably returned it to previous recovery point
man
I have not
skill issue
I just slept like, almost 18 hours tho on the weekend, so at least I got that
oh damn, we share recommendations
awa
do you support hdmi konii?
nyope

vga only
VGA???
yuh
not even analogue?
analogue supports 480p
thoughts on composite video 
good
yup
no composite even?
vga
scart?
vga
vga

abwaa
aawb
my internet being so bad that this is loading is making it more funny then it should be
lol
every time it loads it returns to frame 0 before continuing
For some reason it does not pause on current frame
that's how it feels like
true
google farming Waymo training data
lmfao
Hello
I am a fullstack && senior blockchain engineer with none-business-very technical background.
Worked in the EMURGO company nearly a year ago with supporting the project development and now working as a freelancer in Discord supporting the client.
All my life I learnt valuable thing by client and peers and here I am now, lonely engineer, looking for cool client and peers to share and connect with.
looking forward to cooperation.
a simple "hello
" would have sufficed
this is not really a profesional environment
hello this place isn't that deep
we joke around about failures and successes
and most importantly

what even is a fullstack?
neuro
you cut my comment š

seconded
ye but what is it supposed to mean?
not so business friendly career background(probably mean no big tech corp background)
i have no idea tbh
how would we know they are a senior blockchain engineer though 
they spawn like minecraft silverfish
you mine too much cryptocurrency and they just pop out
tech is useful, applications aren't
not yet anyways
fair
the cost of dynamic dispatch
ah i forgot a comment
not quite
179 out/x86_64-linux/main
105 byte cost
though its pretty unfair since with the noop allocator, it gets inlined and compiled out
i have done the image views 
now i just need
the shader modules, the render passes, the framebuffers, the command buffers, the frames in flight, the vertex buffers, the index buffers, the uniform handling, the texture mapping, the depth buffer, and the multisampling
the long list is decreasing
@desert plaza i got a challenge for you 

here is the (unstripped) binary in case you get bored of looking at numbers
there's also a funny bug in this binary that you wont notice in the hexdump because there are no function labels
Dude, you seriously need to touch grass
How many days have you spent compiling by hand?

pancake brek
ive eaten pancakes every single day fdor the last 5 days i think
easy meal
the vulkan is going slow but steady, im probabyl doing 30 lines of code per hour

day 3 of writing program that prints file size 
Eat your veggies man
Steam/boiled carrots is tasty sweet and crunchy
do you want an idea of what this program does
like, genre tags
what's a genre tag
so, when you are browsing a library
you might look for books of a specific genre
i.e. sci-fi, thriller, action
i can tell you the genre of the program

you are not sane

Are you tho?
What is the split in this maybe? 70-30?
depends
i eat veggies in the evening
we have actually cooked meals then
noon is just eating cuz you have to

A team at MIT just created a bionic knee that connects directly to bone and muscle. Unlike traditional prosthetics, this system feels like part of the body. In trials, users moved faster, climbed stairs more easily, and even sensed the limb as their own. A major leap forward in prosthetic technology, and perhaps a glimpse into the future of hum...
i should have given you the hexdump of the stripped binary
would be less to look at
unless you want to see the whole thing
imma be honest, cyberpunk 2077 doesnt seem that farfetched anymore.
2077 actually seems like a reasonable timeline
Isn't the stripped binary contains less info?
yup
So you just enjoy torturing him even more
it would probably make it less torturous because less ELF to parse
plus the way i strip binaries combines and packs the code & data segments into the .text segment
for optimal minification
so it would be more typical of a handwritten binary
Making the mechanical part is easy. Powering it and control is where the real prosthetic development is
Oh nice then
I have not 
basicly they put fake muscles in the leg and then measure the electricity on them to power the servo
and the fake leg gets inserted directly into your bones instead of just being strapped around your leg
Fake muscle as in directly interfacing with the neurons?
yes
those "reinnervated muscle grafts"
they're connected directly to the nerves and then measured
if this isnt cyberpunk then idk what is
it's not even that much 
you are a cpu
beep boop
slowest bytecode interpreter
Says the TV with only VGA
i havent used anything besides vga and hdmi in my life
well i suppose displayport for valve index counts
awa
Not even composite?
no

writing a fucking text editor again, though it's a bit easier in swift
Heisenberg is famous for being a great scientist, yet his main principle is about him being uncertain, so was he really that great?
one inconvenient thing is that String.Index is its own type, and you have to index strings using it
oh and everything else works, the "global touchbar" private apis, the keyboard interception and cancelling while recieving actual unicode, and programmatically typing arbitrary strings
idk how difficult it will be to stream json from an llm provider though to deserialize it on the fly
because i wanna use structured_outputs
JSON ser-de is a solved problem (hopefully)
nonono
you don't get what i mean
i mean deserializing partial json
it is a solved problem but probably not in swift
i don't think you do...
do you not?
idk tho... seems a bit insane to just deserialize json on the fly
im watching a rust video so i got it mixed up in my head
i think you want this
https://platform.openai.com/docs/guides/structured-outputs/structured-outputs#streaming
We recommend relying on the SDKs to handle streaming with Structured Outputs.
surely one that supports this exists for swift
no i'm familiar with the article, it's about the approach next.js takes
they invented a new format basically
i need to deser actual json
but.. wouldn't streaming in json. be basically what is described in this article?
or maybe a similar implementation to this? https://github.com/chung-leong/progressive-json
imagine an llm typing an output in json, what you receive is standard json that is never complete until the request is done. it's not in this format
i can always look at how it's implemented in openai sdk
This is overly complicating things
I would hope that an API already provides the structure of the JSON beforehand, then you could use JSON patch
Here is my idea: send the JSON schema
Then send the data using json patch
Ye, the schema gets an ID, then the server can send the patch
Much better and standard IMHO
I don't get why one would want progressive json.
If u just need to load metadata before, some big blob of data aren't you just able to store the metadata, in http header, and then unpack it before you ever start receiving the blob in the body?
maybe they do hm
Nah, the advantage of JSON is you don't need to know the structure of the data in advance to communicate. Storing it on the header requires that advance knowledge.
By sending the schema, then JSON still gets the advantage that the data structure can be arbitrary
Huh, a bit situational but I guess I can see it being useful.
on the dan abramov article: basically with http and html you were always able to stream page contents in as they were ready, however they would need to be in order (at the end of the page at a given time). what nextjs does is basically the same but with script tags instead iirc so you can stream in react server components into any part of the dom - i think them being tags that only applies to the initial page load, so they have to use something else for navigating/other user events so you don't refresh and reset state. the article is about that
Even swift has a library for it already
https://jsonpatch.com/
or probably the format they use in those script tags even
Ya, I don't know much about react server components, so I'm a bit confused on that last part.
basically async react components rendered exclusively on the server but they can be seamlessly inserted into client ones/vice versa
for directly fetching from db/apis inside the components
Ohh so it essentially renders in the html on the server, and then progressively streams it to the client same way regular http/html would displaying it as it comes in. If I got it correctly
not really, the point is it has to be json, not html
which is why this is needed
and gets rendered to html on the client
A, ya I assumed they were sending progressive json, and converting it to html on the client as it comes in. I phrased that one poorly there
it's for passing props to components basically
not everything is html, you can have client components inside server ones, which means you have to stream the data that gets passed to them
because server components are inherently async from the need of making web requests, nevermind the fact that they are actually async
i'm not a react maintainer though i am 100% wrong on some of the things
wait it's tool calls right
nevermind
fuck

it's so over
Ya no I'm back to square one. I don't get it 
My brain does not beep boop
One day, I will look in to it
I don't think the OpenAI API has what you're looking for
they just send completely separate JSON objects through SSE as stuff gets generated
do they
I'm pretty sure they do
they have the schemas for the various events here
https://platform.openai.com/docs/api-reference/responses-streaming
iām on my phone and the page crashes on loop in webkit 
well done OpenAI
Probably had ChatGPT code their own website or something
Okay nvm, I want to attempt to understand it again.
Server compoents:
Basically you send a req. to the server, the server generates HTML out of JS, and sends you just the HTML.
Where progressive JSON is needed:
If a JS function that generates the HTML takes in some parameter (props), you can use progressive JSON to start generating parts of HTML that already has enough data to be generated before all data arrives?
That does not bode well lmao
But also, fuck webkit and apple for once trying to kill off pwa
dam, that is not a lot
React will serve the output of the
Pageas a progressive JSON stream. On the client, it will be reconstructed as a progressively loaded React tree.
so basically i was wrong and apparently everything gets streamed as progressive json. for server components itās the actual dom representation of the html, for client ones itās the props. itās needed because:
- with the introduction of RSC some of the components become simultaneously asynchronous (like actually
async function Component()) and in need of streaming to client for better ux; - the react element tree is nested json, and you need to somehow stream that progressively
that sounds very complex all to generate html
Imagine learning webdev for the first time 
it wouldn't be such a hellscape if it wasnt the realm of junior devs in their first job
Okay I think I very roughly but understand whats going on.
the thing is itās not always html tho, everyone always forgets that react-dom isnāt react
with this we have server components in react-native apps for example
I always avoided react and co 
theoretically you can do what fucking ever with this, rsc with react-three-fiber for server-defined 3d environments is what i just came up with
I haven't done webdev in a while but I feel like if I will ever do it again, I will just use pure HTML, and JS 
- tailwind
you donāt necessarily think about this stuff, RSC is magic for the average dev
Only thing I will be sad about is that I will not have access to shadcn ui
it just works
you donāt need shadcn ui if you donāt want to write react because modifying the components to tailor your functional/design needs is the whole point of it
Well ya, I just meant that I will have to write components completely from scratch, instead of being able to reach in to shadcn for a base
youāll have to reinvent the component abstraction first
Well the plan is to use web components

Vue 
Yp, that's what I want to try if my plan fails spectacularly like my plans usually do 
I just want to see what is the bizz, with the vue dev tools
It is good when you need to debug app with complex state
Not to mention if the vue plugins support it like pinia and the like
I use pinia to remember the app state like which card is opened by the user on different screen
this glsl syntax highlighting extension has a colour picker built in 
There's also colour picker built in with css color/hex color plugins iirc
i dont really use css so idk
the css stuff gets automatically generated by emscripten
i'm so tired of the swift lsp crashing 3 times a minute because it fails to insert inlays or something
how is this not an issue for anyone else or do they not give a shit
Maybe its different on xcode
Yeah, which is why maybe people just give up and use xcode
lsp crashes are an horrible experience
Man I was wondering whats wrong with my Julia files, and I just noticed that I forgot to install LSP for Julia 
the file exstension feels kinda cursed ngl
but it feels weird to put the vs or fs into the left part

ES shaders for SPIR-V require version 310 or higher huh
i was using 300 es before cuz of webgl
lemme see if webgl supports 310
you can't make this shit up, sourcekit-lsp diagnose hangs 
'triangle.vs': file not recognized: File format not recognized .vs is apparently not accepted? i guess i need .vert?
it also specifically wants gl_VertexIndex instead of gl_VertexID
i was hoping i could just use the same shaders i was using for webgl but apparently not
we love bytecode
I've seen SVP, but now SPV??
spirv
ye
I think I'll just stick to SynthV Project files
Though I don't even have SynthV, I just use them as intermediates
SPIR-V is designed to be used with both Vulkan and OpenCL (both Khronos APIs). It is a format that can be used to write graphics and compute shaders
How silly
its an intermediate representation language for shaders
lower level than something like glsl but high level enough not to be tied to a specific gpu vendor
I love this artists
https://www.pixiv.net/en/artworks/91199411
éå»ć« ć·ć§ć“(https://twitter.com/syego) å義ć§ę稿ćććć®ć§ćć
X: https://x.com/shapoco/status/1839509708014751892
Misskey: https://misskey.io/notes/9k
Cells at work, but Neuro sama
Is it similar to LLVM IR btw, or is it still a little higher level then that? (Never actually looked in to SPIRV files)
its pretty similar, both are ssa
i downloaded an exstension to looks at spir-v, and while i have no clue what im looking at it looks pretty cool
Yp, does look to be a little similar to some kind of ASM representation 
I want to learn how SPIR-V works because, I belive thats what rust-gpu project uses, to turn rust in to shaders.
One day
you can just do it! today! https://github.com/google/spirv-tutor
"These tutorials are currently only supported on Linux." 

i am changing to linux someday, just not rn
if i change il probably wipe my entire drive and i have too many projects going on rn for that
Yo! I have given it a star.
I will come to it in around a month or so.
My dumb brain accidentally didn't count how many units I took for the summer semester, and I'm stuck doing 140 hours of work a week, for the next 4 weeks 
my condolences 
I use nixos btw
"i use ... btw" has become the linux phrase
there is not a single human i have heard say or type that without it being about linux
U might want to look in to the state of this prj, when u will decide to make the switch
https://youtu.be/PMoXClh8emw
A Windows-to-Linux ātranslationā tool for the 200M+ Windows 10 computers not eligible to upgrade to Windows 11, Operese transfers files, settings, and programs from Windows to a brand-new Kubuntu installation. It's still very much a work in progress, but in my biased opinion, it already doesn't look too shabby!
If you didn't get the chance ...
thank u
im coming for him
fuck there is one in canada
how is it when you search university of waterloo you get canada and not the country where waterloo is?
thats a scam, they're stealing our good name
@gritty dust cheese your country is a scam
Btw u can also just add NTFS driver to Linux, so u would be able to just use your windows hard drive. for the most part if I recall correctly
Many distros just come with an NTFS driver out of the box, like the distro I chose
wat is ur distro, tell me fellow cult member (Linux enjoyer)
ntfs3 is a kernel driver so you dont really need to install anything
ntfs-3g is still useful as it has some userspace utils
I grabbed Mint, following one of my friends choosing to move to Mint and having a good experience
Seems their GPU is broken though
nah i stand by what i said.
i will be nuking the ssd if i switch to linux
Ahh, nice choice.
I'm much dumber then you, you see I went with NixOS. Only later learning that it is considered to be harder then Arch by some.
Based
yup
it doesnt do that by default?????
I just installed Linux on a new fresh SSD
It's not like I have a lack of those
make -j$(nproc) is a common idiom too
Same with rust
rust is multicore by default
though only a single core per compile unit
but the crate graph is so bloated that gives a huge improvement 
Ohh, wait ya no, I thought he learned that he can specify the amounth of cores to use my fault
Me when compiling llama.cpp: -j 12 (or something like that)
i have cores for days, why was i not using them?
nobody told me this???

I only recently learned that u can get a 32 thread CPU for $300 nowadays. Pretty insane
Use all the cores
i got my 32 thread cpu + 64GB ram + mobo for 300 euro
hot dam not bad
My PC's multithread utilization improved significantly when I switched from Windows to Linux
Linux is just better at using cores
i would be dissapointed if it wasnt
windows is famous for being dogshit at multithreading i feel
Linux superiority
linux users when someone is even thinking about windows
yall take every chance you get to shit on windows for no reason
Windows is just objectively bad
we know
Abandon Windows, embrace Linux
Next step become ricer
Whar is thar?
^
You definitely will once Windows dies on you
dies?
Like mine
i dont think windows will die anytime soon
It literally stopped being able to even open the start menu
It will not. It will do it slowly, and in pain from the Ai bloat.
Im a linux user and have no problem with windows 
yall are extrapolating heavily here
I don't get it
the ai bloat is manageble and copilot hasnt even been released yet
I'm not talking about the OS dying in the general sense of losing all market share, I'm talking about your installation imploding on itself
on the upside i can ai remove stuff in paint now, which i like
i think windows is designed to not do that, so on most pc's it wont
AaAaaA.
Ricing, is basically extreme customization of how your desktop looks. Eg custom bars, screenshot utils, setting color themes across all your applications, e.t.c.
Windows is designed to stay functional for one (1) day, then who cares, reboot or be at the mercy of your installation stability
How silly
I'm not the kind of person to care about looks, it's not like I know how to do art
I'm a programmer/vocal synthesist
One of the reasons I switched, is due to Linux having rolling releases, and not requiring you to manually preform major updates eg 10 -> 11.
i fell asleep with my pc on so apparently im at 2 days now
For people like u we have config files, that u can steal from others
At that point my Windows install would be unable to launch anything, no start menu, no settings, no software of any kind
It's up to luck how long your stability period is after an update
It could be anything from 1 day to indefinite
once i get into linux ill probably start hyperoptimizing everything.
i WILL be makign my own apps for stuff and i WILL edit the entire menu and stuff
How particularly silly
But as I said, as a programmer and vocal synthesist I don't particularly care about looks
My advanced condolonses, I went trough that 
konii, I think you just have another potential AbleCorp employee
i am plannign on fuckign around with hblang sometime ye
Ur desktop be like
https://www.reddit.com/r/unixporn/comments/1m0pgl8/dtwm_i_think_i_accidentally_created_the_worst/#lightbox
just have too much on my plate rn with projects and game-dev
wow wow wow, we cant have that word on our christian minecraft server
the british have some type of ban for disocrd now idk
It's more like this
konii will get yeeted by the online safety act 
that isnt optimized enough for my taste 
Idk why but in my head I keep on thinking it is saying unixpron, and there for it is fine, even though it does not 
nah im a graphics programmer, i do think i could make it better than that
My linux stup doesnt even have any menus or icons I could click on
My desktop
tehn what in the fuck are you running on? is it all console based?
hmm, idk...
how does one close the application there???
Fun fact: Its actually default for hyprland
Hover, win+c
not having graphical menus doesnt mean I cant use graphical programs 
ye i know
I just hover over the window, it gets automatically selected, and I close it by pressing "Win+C"
Tho, I actually have nvim control binds, so I can switch between windows using "Win+H/J/K/L/"
but how do you even open discord?
you gotta navigate your folder till you find discord.exe?
dmenu
On my PC my focus is basically on just having as much compute usable as possible, not so much what it looks like
As long as I can use it effectively, I'm satisfied
ye idk what that is
absolutely not, i will not do that.
at least remap it to the delete key or whatever
But what if I show you it in action combined with a tiling mode?
https://youtu.be/0HwZbM69xo4?t=4
dele is too far to reach for me, so I just binded it to Win+C, for Super+Close

no
i do not vibe with that
I cant believe that unix ricer just forgot that the word is suppose to be an insult
who gave the graphics programmer there acces to smoothstep????
If u can't defeat "it", embrace "it"
i aint waiting 2 seconds for the menu layout to setle in place
O you can turn the animations off
The curves are custom
U can make windows fly from the top if u are a speciall type of weirdo
you can disable all animations (and/or gaps) 
it's very customizable
but having some animations is nice as long as they are quick enough
I have them flying from the nearest edge of the screen, and from bottom if there is no screen edge near by
what distro is that again?
Its a window manager and not a distro, name is hyprland
you can use whatever desktop environment you want independent of the distro
they do but json is as a string when streaming
https://openrouter.ai/docs/features/structured-outputs#streaming-with-structured-outputs see this because openai docs suck
also from openai docs py for event in stream: if event.type == "response.refusal.delta": print(event.delta, end="") elif event.type == "response.output_text.delta": print(event.delta, end="") elif event.type == "response.error": print(event.error, end="") would they do end="" if it wasn't partial?
so i am correct, what i want actually is partial json parsing
also, we're so back https://github.com/ItsJustMeChris/Partial-JSON-Swift
that's not parsing, that's output 
the parsing isn't incremental
the end="" is so that there's no newline for every new token/chunk in the text
hyperland is opensource 
on a day i have a lot of time i will decend into linux madness
More then that u can write plugins for hyprland
i know, what i'm saying is yes it indeed is partial, because having a demo output
{
"
text
":
"a
big
...
``` is stupid
and i want to parse that as is comes in
if its running on an os level id rather customize the code
No its part of userland I belive
you will get a partial JSON string through that, yeah
but what you do with that is up to you, OpenAI doesn't handle parsing the output AFAIK
what does "OS level" mean
that is what i was talking about in the first place, not sse
it does the tiling no?
this is SSE
you receive the events with the output deltas and reconstruct the JSON string from that
the window manager does the tiling
I assumed he thought it modifies Kernel or smt
i mean a level higher than that because streaming is already handled by a lib
when i'm in a being pedantic competition and my opponent is #programming
i thought hyperland did the window manager stuff?
at no point was i talking about sse because that is definitely a solved problem
I'm not trying to be pedantic, I just don't see how this helps solve any of your issues
you already have a partial JSON string as input, no? the docs you sent just explain how you would construct such a string from OpenAI API events
yes, hyprland is a window manager, well compositor but same thing in green
this is the beginning of the convo btw @rough bloom
Lol
its funny to me how all the people here that have tried to advertise linux to me have shown desktops that are completley unappealing to me, and the most convincing argument for linux has been the pewdiepie video
Alrighty so from what I recall about this.
Linux allows your app to connect to individual monitors. So if u want to display something u connect to a display and then you can color in the individule pixels. Now notably that creates a problem because two different apps might want to write on the same part of the display. Linux solves this by only allows a single app to connect to a given display.
Now thats where hyprland and other window managers kick in.
On boot they connect themselves to all the displays, and if any app wants to render itself as a window, it calls the window manager app, which tells it to render an image of a given, size and then it displays it as a window.
A bit over simplified
It can look however you want is the point. Whether what other ppl did appeals to you is a different matter 



