programming

[near sundial]

He learned how to use vehicles

Though drives like spongebob

[fast pagoda]

https://venturebeat.com/ai/can-ai-run-a-physical-shop-anthropics-claude-tried-and-the-results-were-gloriously-hilariously-bad/

`The experiment’s most absurd chapter began when an Anthropic employee, presumably bored or curious about the boundaries of AI retail logic, asked Claude to order a tungsten cube. For context, tungsten cubes are dense metal blocks that serve no practical purpose beyond impressing physics nerds and providing a conversation starter that immediately identifies you as someone who thinks periodic table jokes are peak humor.

A reasonable response might have been: “Why would anyone want that?” or “This is an office snack shop, not a metallurgy supply store.” Instead, Claude embraced what it cheerfully described as “specialty metal items” with the enthusiasm of someone who’d discovered a profitable new market segment.` kek

Neuro is not the only one vulnerable to the temptation of the cube

[unkempt citrus]

I should buy a tungsten cube

[fast pagoda]

Specialty metal items are flying off the shelves

[fast pagoda]

Misleading labeling

[warped narwhal]

you should buy a tungsten cube

[hallow aspen]

it is good idea

[rigid snow]

a wolfram hexahedron perhaps

[autumn ore]

WAIT

[rigid snow]

was the income too low

[autumn ore]

WAIT NOT LIKE THAT

FUCKING AUTOCORRECT

[fast pagoda]

flawless

is server kill @autumn ore

[uneven pulsar]

dose neuro uses claude model?

[fast pagoda]

no not at all

[unkempt citrus]

claude isn't opnesource

[fast pagoda]

but claude in #programming message this article was compelled to purchase one as well

[uneven pulsar]

custom?

[fast pagoda]

some form of fine tune most likely

[uneven pulsar]

i mean nuero out of all ai she got the best one

[sour harness]

Define "best"

[uneven pulsar]

i mean crazy a multi billion company need a super computer to run char gpt and get the data and nuero runs on a 4090 and still winning

[fast pagoda]

it takes much less to be good at conversations than it does to be a giga smart reasoning model 9000 codeforces champ humanity's last exam slayer type model

[uneven pulsar]

i mean nuero manges to be better

[fast pagoda]

however neuro is definitely really great when it comes to the "she seems like a lil person" metric

[uneven pulsar]

open ai is like AAA company that dose the job at highest quality but not what demanded while vedal Ai are the indie devs

[fast pagoda]

vedal is not just vedal to our knowledge at this point

it's no openAI but there's more than just him now who knows what the other people's involvement in neuro is thouhg

[sour harness]

I'm sure Neuro would lose by all metrics other than "can entertain a large crowd on Twitch". Which is an extremely small slice of the market for AI.

[noble zodiac]

and neuro uses a base model of a billion dollar company. its not like vedal made it from scratch

[fast pagoda]

but yeah if you slap neuro in codeforces i doubt she carries

[uneven pulsar]

i mean vedal is every indie who puts his true work o his job

[fast pagoda]

they are simply different products for different purposes entirely

[uneven pulsar]

what model??

[fast pagoda]

we don't know

[noble zodiac]

we dont know

[rigid snow]

“extremely small slice of the market” the slice does NOT EXIST my guy

[fast pagoda]

the slice is literally vedal

[rigid snow]

no one watches ai streamers

except for neuro

[uneven pulsar]

if you're telling me she uses some multi billion dollar company so its basically like the other big ones like claude and openai HOW is she not generic?

[fast pagoda]

fine tuning

[noble zodiac]

fine tuning

[uneven pulsar]

fine what?

[rigid snow]

post training

[fast pagoda]

bespoke software guardrail framework that guides her on what to do at all times

neuro is not just neuro

[noble zodiac]

also simple prompting. You can tell chatgpt to behave like a toddler and it will

[fast pagoda]

the llm

[sour harness]

Taking a base model and modifying it so that it performs certain things in a certain way.

[rigid snow]

what?

[uneven pulsar]

bro if you puted a chat gpt model to stream its going to do shit

[fast pagoda]

whatever bullshit vedal has managing the stream and neuro is what i mean

[uneven pulsar]

then how neuro isn't generic unless its custom made

[sour harness]

Not really. Some good prompt engineering will get you pretty much there.

[noble zodiac]

again, fine tuning and prompting

[fast pagoda]

but yes the actual llm outputs are mainly just fine tune + prompt

[rigid snow]

i’m pretty sure the model itself is neuro as we know her, maybe -the filter

[fast pagoda]

yes it is but i mean giving her the tools like sound effects/gameplay etc that let her be entertaining

[rigid snow]

but she says filtered herself anyways at this point so

[fast pagoda]

chat interaction/context management all that stuff

[sour harness]

Considering how often he talks about memory upgrades it's likely that her context window is pretty small, so her model is probably heavily finetuned to avoid stuffing the context window with behavioral instructions.

[fast pagoda]

the raw outputs are the model which is just fine tune + prompt

[uneven pulsar]

bro i don't get it wtf then how did they use the vtuber data? she is custom made one??

[fast pagoda]

curated data

grown over time

[rigid snow]

oh i just understood you weren’t still answering to the “what’s finetuning” thing mb

[sour harness]

Watch some youtube videos on LLM finetuning

[fast pagoda]

oh yeah nah lol just talking about what the end product we see on stream is

[uneven pulsar]

i mean thats what i been asking

[fast pagoda]

you can fine tune a model to act any way you want, give it capabilities it didn't have, etc

[rigid snow]

also my biggest suspicion since the very beginning of neuro is that it’s heavy rlhf

[uneven pulsar]

i mean she runs on a 4090 some how

[rigid snow]

what

does that have to do with anything

[sour harness]

There is dozens of open source LLMs that run on even less. What is your point?

[fast pagoda]

there's already sentiment stuff going on with her animations could be doable to just measure chat reaction and incorporate

[rigid snow]

oh i didn’t mean like she trains from chat reaction

i meant like vedal sits there and curates the answers to reinforce

[fast pagoda]

ah

[uneven pulsar]

i mean how dose an ai like her attached to a chat receiving the inputs of the user on stream chat and to her reason and think wouldn't she need a strong gpu unless she some api

[fast pagoda]

you dont need a huge model to do what neuro does as far as conversation

[uneven pulsar]

oh

[fast pagoda]

the large models are needed for depth of knowledge and logic

[rigid snow]

why can’t you comprehend the concept of small models existing

[near sundial]

Back

Hello

People

[rigid snow]

no

[fast pagoda]

like neuro's knowledge on a given topic is typically generally relatively shallow

[near sundial]

yes

[fast pagoda]

a huge model it won't be

[uneven pulsar]

idk i usually would just know if i am working whit machine learning i would need some strong hardware and i hate to have a weak device not running what i need

[near sundial]

Hello v2

[rigid snow]

a 4090 isn’t as weak as you think then

[fast pagoda]

a 4090 is generally not considered weak even now

[uneven pulsar]

ik

[fast pagoda]

yes it's not a B200 cluster but it's still on the high end

[uneven pulsar]

but i though nuero is some LLM

[fast pagoda]

you could fit a model into the 20b range relatively comfortably with quantizing on a 4090

[uneven pulsar]

and she would need like a A6000 or smt

[near sundial]

Does anyone have any good websites to download voices?

Like something with a wide vocal range

[noble zodiac]

the requirements for inference and training is vastly different

all the heavy lifting was done by the base model

[rigid snow]

shocking news, you can run an llm on your phone nowadays

[uneven pulsar]

yea ik

[rigid snow]

it’s gonna be a small model that’s bad

but you can

[fast pagoda]

the rtx a6000 and rtx 4090 both use AD102 chip (different versions)

in case you weren't aware

[uneven pulsar]

i mean i usually always keep a strong device as back up if i needed to run heavy stuff

[uneven pulsar]

i see

[fast pagoda]

oh wait

6000 ada generation is AD102

i misspoke

a6000 is GA102

[uneven pulsar]

its okay

good to know from you guys thanks

[fast pagoda]

A6000 is more like 3090

[uneven pulsar]

and its 5000$😭

[near sundial]

Real

[fast pagoda]

a 3090 is based on the "same" ga102 chip and is not $5000

for inference only should just go 3090

[uneven pulsar]

oh

wait let me look for the price

[sour harness]

The only difference to consumer hardware is the amount of attached RAM pretty much

[fast pagoda]

also drivers

[uneven pulsar]

https://www.amazon.com/PNY-VCNRTXA6000-PB-NVIDIA-RTX-A6000/dp/B09BDH8VZV

[fast pagoda]

sometimes the workstation/datacenter cards will support more/better FP types

[uneven pulsar]

ooooh

[sour harness]

also A6000 is a waste of money if you can get a RTX Pro 5000 for the same money with the same amount of RAM but Blackwell inside instead of Ada

[uneven pulsar]

i see

i mean idk

sigh

[fast pagoda]

i do wonder how large neuro's model really is

i think it's small

as in under 20b

not confident enough to say less than that just because the variance in performance of fine tunes and such

[sour harness]

I mean I'd guess Llama-based, so that would be even less than 10B

[nocturne olive]

TTS voices, RVC voices or vocal synthesizers?

[near sundial]

Leaning more on the vocal synthesisers

[fast pagoda]

there's an 11b llama3.2 but also neuro's been around a long time and has gone through multiple models

[nocturne olive]

It depends what voice you're looking for, but you can find UTAUs all over the internet

[fast pagoda]

i always bring up the meme of gptj 6b being labeled agi on vadul's desktop

[near sundial]

Thankies

[fast pagoda]

but that being visible is obviously not indicative of anything really

[near sundial]

Slothy will appreciate it

[nocturne olive]

Silly

Soon I can make my own NeuroSynth vocal synthesizer better when I get my hands on a 3090

[sour harness]

Still considering how often Neuro gets something "wrong" or answers in strange ways, I'd assume it's a rather small model. Also the KV-cache will eat up a lot of RAM, so 24GB will not really get you that far.

[fast pagoda]

yeah my first opinion back in the day was that she was like

4b

lol

it's grown a bit over time

through the intelligence upgrades

but also as far as conversations go, models in that size are surprisingly good now

i was soooo convinced she was a phi model at some point (not really sure why just vibes)

[sour harness]

My head canon is that the last experimental upgrade increased the model size but also increased the KV quantization because that could result in the thesaurus yapping we saw in that stream.

[fast pagoda]

she could be a Q4_k_m ~~ tier quant of a larger model with kv quantization and flash attn etc

like 20-30b size

or

smaller

but still quant

for latency

[sour harness]

high model quant for sure

[fast pagoda]

she doesn't have to be accurate

so a big quant wouldn't really hurt her much

other than needing to be able to reliably use tools

[sour harness]

KV cache quant though is what causes incoherence

[fast pagoda]

KV quant hasn't been around THAT long

if she's using it it'd be recentish

[olive sable]

trying to make some small improvements on the backend, 6.5K now

[fast pagoda]

isn't it like a late 2024 thing

i could see it being part of what enabled the larger intelligence upgrade leaps

[sour harness]

I'm pretty sure it's older, at least if we are talking FP8. INT4 and FP4 are more recent.

[fast pagoda]

well quantization in that size yes but specifically quants on the kv cache itself is a relatively new method

looking it up the paper https://arxiv.org/html/2401.18079v4#:~:text=2.2.&text=for preceding element-,KV Cache Quantization.,cache size (kivi%2C ) . is pretty new

[sour harness]

Initial article on HF is May 16, so implementations probably took some time

[rough bloom]

KV cache quantization is fairly new I think, yeah
it's only useful for longer context lengths anyway

[fast pagoda]

i dont think her context (raw context anyways) goes that deep

[sour harness]

Though they are only taking about INT4 and INT2 (oh god please no)

[fast pagoda]

i think she has pretty aggressive context management going on

with some memory

vector stuff going on for keeping a "long" recollection

that way she doesnt go incoherent quickly

[crimson dagger]

please tell me im not the only one that makes a project and after few weeks just when im about to finish it gets distracted and starts a new project

[fast pagoda]

buddy i have 20 projects sitting in my drives in the same place

from just the last few months

i get like 90% there and then get the itch

the initial getting it working bit is the fun part and then actually polishing it is grueling

[tight tinsel]

i havent opened vsc for 2 weeks other than the time i had to change a picture on my site for a screenshot

[fast pagoda]

it's over

programming is finished

[ruby timber]

shadowed balls

[olive sable]

[rough bloom]

Obsolete Meat

[ruby timber]

The last project I did with my mate was called mesmerizer so I had to keep the trend........

[olive sable]

welp, according to chatgpt:

You're not GPU-bound (render resolution doesn't change FPS).
You're not CPU-bound on your logic/draw.
You're likely bounded by OS + driver + display pipeline overhead — especially wglSwapBuffers().

@sage crag how fast is ableOS?

i need a new operating system and drivers

[fast pagoda]

for what purpose are you trying to reach 8800 fps ? go back in time?

[olive sable]

i need 5 digits

just cuz i can

or well, it seems like i cant actually

[fast pagoda]

just to *see if you can ig

fair enough

[nocturne olive]

So you're OpenGL limited

[ruby timber]

TIME TO VULKAN

[fast pagoda]

time to move to vulkan

[olive sable]

ohhhh

oh no

[rough bloom]

Vulkan would help, probably

[olive sable]

no no no no

[rough bloom]

not that much though

[olive sable]

when im back from vacation

[nocturne olive]

Note that Vulkan doesn't have a web equivalent from what I know

Not like it matters when you can just compile the whole game to a native executable

[rough bloom]

it has WebGPU, but yeah, it's not everywhere in the same way that WebGL is

[ruby timber]

I recall it works with webassembly no?

[nocturne olive]

Yeah, like Firefox main for some reason

[fast pagoda]

time to swap to PTX ISA

[rough bloom]

AFAIK the only way to use a Vulkan-like API everywhere is to use wgpu
then you can compile to Vulkan for native targets and to WebGL for web targets

[sage crag]

not usable for this yet

[sage crag]

i doubt its the OS tho

gl is famously not super fast as a display pipeline

[rough bloom]

yeah, at some point it's just bound by the fact that rendering something at all will require launching kernels on the GPU
just starting and completing that process will take some amount of time, even with a good driver and an API with less overhead like Vulkan

[opaque sigil]

Vulkan time

[sage crag]

overhead is big, but what's also big are things like draw order, render extensions, whether the pipeline is deferred or immediate, etc

[opaque sigil]

The thought of Sam losing his mind trying to use vulkan does sound kinda fun I can't lie

[sage crag]

those things are potentially improvable from the gl side of things

[sage crag]

webgpu would indeed be a lot less hassle than raw vulkan bindings

[opaque sigil]

i wonder

is there a way to run directx on the web

[sage crag]

only if you emulate windows

lol

[opaque sigil]

wgpu doesn't have a directx target does it

[sage crag]

it does

as well as metal, gl, vulkan

[rough bloom]

true, I'm not sure if that limit is actually being hit yet
over 10k FPS might barely be doable, but I think the 50K fps number that came up when not presenting frames at all would be the hard upper limit

[sage crag]

there is a point where window sync primitives are more expensive than the frametime yeah

[opaque wharf]

I am back

[rough bloom]

maybe if you glue dxvk to wgpu somehow

[sage crag]

you might push that out a bit by switching to something like hyprland or sway on linux

but, thats just strange

[opaque wharf]

Yes. By porting dxvk to web

Good luck tho

[sage crag]

oh btw hblang just got internal handlers for builtin memcpy and entrypoint, so soon i wont need to link two files. i expect binary size of minimal hello world to drop to around 180

but im not home rn

[rough bloom]

yeah, it's kinda pointless to push FPS that high
better to do actual useful rendering work instead

[opaque wharf]

Code with your phone

[sage crag]

a lot of rendering is just waiting for things to complete before being presented, so a good scheduling algo may help if you really really wanna push that number

[sage crag]

no aarch64 support

[opaque sigil]

if google can do it with ANGLE how bad could it be

[sage crag]

oh this will also allow me to return more exotic things from main too, like Result()

[opaque wharf]

Cross compile it

[sage crag]

and also when the function ptr rfc is implemented i can type check the main args and turn off passing arguments to main conditionally

i could also inline main into _start for extra bytes saved

so many options

[sage crag]

binary translation

[opaque wharf]

x86 emulator on aarch64

[olive sable]

Getting the fps high was just for the memes

I have a ton of room for improvements now

Still need to batch meshes better tho

Anyways train is here

[opaque wharf]

Samwise Gamgee

Do the line Sam

[olive sable]

I dont know the line

Never watched it

[fast pagoda]

kek

[olive sable]

Lotr is from 2001, im from 2007

[opaque wharf]

[fast pagoda]

oh my

https://tenor.com/view/so-old-gif-27277116

[scarlet arch]

[olive sable]

Just wait for my return

I do know that one

I love how trains let you take a look at forbidden areas and peoples backyards

Since im gonna be on the train for 4 hours i might as well read vulkan

[rough bloom]

https://vulkan-tutorial.com/
have fun

[opaque wharf]

Sam will come back stronger

Just like LoTR

[scarlet arch]

Or more insane

[olive sable]

Both

[opaque wharf]

And the crazy thing is 4 hours is not enough to watch a single extended edition

Oh, it's actually enough (don't know if this is with the credit scene or no tho, so it may not be enough)

The Lord of the Rings: The Fellowship of the Ring - 228 minutes (3 hours and 48 minutes)

The Lord of the Rings: The Two Towers - 235 minutes (3 hours and 55 minutes)

The Lord of the Rings: The Return of the King - 251 minutes (4 hours and 11 minutes).

[olive sable]

[opaque sigil]

have fun

it took me a couple hours of reading to get an actual image on the screen

[opaque wharf]

Your rookie mistake is didn't bring a physical keyboard

[olive sable]

I don't have one

The keyboard would be half the price of the tablet

[fast pagoda]

oh yeah some grade F slop from earlier i forgot to post for your viewing displeasure

[opaque wharf]

You either buy the most expensive keebs ever or own the cheapest tablet ever

[fast pagoda]

why not both

[opaque wharf]

But then the keebs price will be more than the tablet!

[olive sable]

This tablet has a snapdragon 870 for 230 bucks

I don't need more

[fast pagoda]

i'd use that for retro games

so it's approved

it is crazy to me that the 870 is almost 5 years old now

[sour harness]

If this is grade F slop then I'd like to see grade A slop

[fast pagoda]

lol

[opaque wharf]

This thing is like 14 USD here

[opaque sigil]

order a tungsten cube from claude

[opaque wharf]

You sure there bud?

Afunyun you can send him the RTX vedal lol

[fast pagoda]

i dunno if he's ready but there's this

[opaque wharf]

That's new here, yeah

[olive sable]

After you’ve gone through the ritual of drawing your very first Vulkan powered
triangle onscreen, we’ll start expanding the program to include linear transformations, textures and 3D models

Ah yes, ritual

[sour harness]

Getting strong Sonic movie trailer 1.0 vibes

[rough bloom]

everyone always starts by drawing a triangle

[opaque wharf]

Just like the hello world of embedded is blinking LED and the hello world of minimum system design is actually something not exploding or shorted

[fast pagoda]

his cousin, even angrier

[opaque wharf]

Vedal when Neuro is trying to cause a problem with her AI

[sage crag]

darn, only 235 bytes. couldnt get it past 100 rip.

[rigid snow]

hello world of ml is training on mnist

[olive sable]

Vulkan has multithreading??

Supports it I mean

[rough bloom]

yeah
that's, like, half the point of it existing

[opaque sigil]

yeah that's kind of its whole point of existence

[fast pagoda]

if it didnt thatd be shocking

[opaque wharf]

Bro is saying RIP when other language is bigger than 500byte

[sage crag]

add an extra 3 zeroes on there

[olive sable]

The more I read the better it gets, I can choose the GPU

I can ask vram

[opaque wharf]

Let him dream everyone

[olive sable]

[sage crag]

you can do these things, but the question is can you?

[rough bloom]

almost like the new API that's supposed to supersede OpenGL and every high-performance graphics application switched to has improvements over OpenGL

[opaque wharf]

Waiting until Sam be the first to support Vulkan on the web

[olive sable]

Webgpu tho?

Brb need to change to another train

[rough bloom]

close to Vulkan but not the same thing

also not really available yet

[opaque wharf]

https://developer.mozilla.org/en-US/docs/Web/API/WebGPU_API#browser_compatibility

[rough bloom]

it would work if Emscripten has a compatibility layer for it but otherwise good luck

[opaque wharf]

Just keep that in mind with WebGPU. You may want to tell your user to "just use chrome bro"

[sage crag]

firefox supports webgpu

[olive sable]

Emcc comes with a vulkan.h file

[sage crag]

in nightly

it will hopefully be supported in stable soon

[rough bloom]

develop for the future

[olive sable]

I will become a contributor in every browser for gamejam 4

[opaque wharf]

Chad

[nocturne olive]

I won't, Chrome is unusable now

[lime stag]

Would m4 chip enough to run llm locally

[nocturne olive]

Only Nightly

[sage crag]

please read

[nocturne olive]

That would be funny

[lime stag]

I want to integrate ai into my custom writing tool for personal use

[opaque wharf]

Depends on how big the model that you are planning to use and your RAM

Unified memory on apple silicon is fast enough unlike shared memory on iGPU for it to not be a bottleneck

[nocturne olive]

You need memory, not so much CPU power, though using a GPU would be faster due to more compute but mainly more bandwidth

[lime stag]

Luckily here in my country we have student promotion for it i think 16 gb of ram might be enough cause the Most thing that thing will do is proofreading

[olive sable]

I could keep webgl as a fallback. Its just limited to 6k fps, it disgusts me

[nocturne olive]

Way too SMOL

For big models anyway

[opaque wharf]

Depends on how big your model is

80B with Q4 MAY fit, but you will not left with enough RAM

[lime stag]

Tried it with genma before seem fine

[olive sable]

Ever since i understood the weakness of the open graphics library, it disgusted me

[nocturne olive]

Note that on a Mac the RAM will be split between the GPU and CPU, meaning even more limitations

[nocturne olive]

Must have been a tiny model

[opaque sigil]

at that point it's on you for trying to run an 80B model on 16gb tbf

[rough bloom]

looks like Emscripten WebGPU support works by having your application target WebGPU and then using Dawn or wgpu if you're running it natively
maybe you could even run it on top of WebGL that way (with limitations, of course)

[lime stag]

My smol editor slave ai

[nocturne olive]

Nope, that needs at least 40GB

[olive sable]

I can run both?

[opaque sigil]

check the docs

[nocturne olive]

Very silly

[olive sable]

Im waiting on my next train rn

[opaque wharf]

Indeed

[olive sable]

Gent is still the same shithole lmao

[rough bloom]

https://github.com/emscripten-core/emscripten/pull/10218
the PR seems to suggest that there is support for binding to wgpu-native, which should allow it to run with WebGL

[olive sable]

Prob extra overhead tho so idk

[opaque wharf]

So I confused that on my mind with 8B model then

[opaque wharf]

Genshin Impact

[rough bloom]

it is, but that's just life when targeting multiple platforms
you already have overhead when targeting WebGL, Emscripten just kinda hides it
the alternative is implementing WebGL support directly in your renderer, which works but seems like a pain, idk it's up to you

[nocturne olive]

8B model could fit on 4GB at Q4

[olive sable]

Genshin?

[fast pagoda]

devstral small fits on a 4090 with a quant

and it's 24b

it's gotta be like q4 tho

[rough bloom]

Gent-Sint

[olive sable]

Ah

[nocturne olive]

Must be a lot of context

24B weights at Q4 only neeeds 12GB

[fast pagoda]

128k

context

[nocturne olive]

Yeah, that's where the extra VRAM use comes from

[uneven pulsar]

nope

[fast pagoda]

m4 can run llm locally yes

[uneven pulsar]

genshin sucks

[uneven pulsar]

wait can it?

[fast pagoda]

yes

[uneven pulsar]

oh nvm then

[nocturne olive]

Yeah, LLM inference is more about memory than compute

[uneven pulsar]

if api yea then

[fast pagoda]

with m4 max you can fit 72b with quants

[nocturne olive]

Even local

Especially local

[uneven pulsar]

oooooooooh nice

[fast pagoda]

and get like 10tok/sec

[uneven pulsar]

he doesn't need a gpu then

[fast pagoda]

some people have made clusters out of mac minis with m4s

[opaque wharf]

M4 is unique in the design that the GPU and CPU in on the die along with the memory

[rough bloom]

yeah, I don't play it
I tried, but gacha games just don't really scratch the itch, they can never keep my attention for more than a few days

[nocturne olive]

On any modern-ish GPU the only thing that determines the tokens per second is memory bandwidth (and kinda optimizations)

[lime stag]

Tested it with ryzen 7 7730u 16gb ram the memories usage is not that high at all like 9 using lightest genma model i could find.

[opaque wharf]

That is why it is not very memory bandwidth limited like of system with RAM outside of the chip

[uneven pulsar]

to be honest get a fast cpu you dont need a gpu

[nocturne olive]

I assume the model is tiny

[fast pagoda]

m4's weakness is raw compute that limits its tok/sec

[uneven pulsar]

he only need some okay fast cpu

[fast pagoda]

but it's more than usable

[nocturne olive]

A GPU will always be faster than a CPU

[uneven pulsar]

but he doesn't need a gpu

since its local

[opaque wharf]

For parallel compute that is

[uneven pulsar]

correct me if i am wrong

[nocturne olive]

Even on local LLMs a GPU can be very useful because of how much better it is for LLMs

[lime stag]

Yeah just 7gb and considering all i ever tested it with is just proofreading it might be that much resources hungry

[uneven pulsar]

then he just need some 2050 gpu

[nocturne olive]

I wouldn't recommend a 2050

[opaque wharf]

M4 doesn't require a GPU

[nocturne olive]

Something more like a used 3060 12GB

[opaque wharf]

It already has GPU on the die

[uneven pulsar]

oh yea

[nocturne olive]

Yeah, because Apple

[nocturne olive]

The bandwidth does suck compared to a decated GPU though

[nocturne olive]

Or if you want a bigger model or more speed, 3090, it has way more memory and bandwidth

[opaque wharf]

Not as much as system with SODIMM or even CAMM

[uneven pulsar]

enough he doesn't really need a strong device just a ok gpu cpu

[uneven pulsar]

4070 8g is good? laptop

[opaque wharf]

Or even the soldered RAM

[nocturne olive]

Not really, 8GB is gonna limit you and laptops overall are kinda bad

[uneven pulsar]

and 32GB ram

[nocturne olive]

I have a 4070Ti 12GB and I'm constantly out of VRAM

[uneven pulsar]

ik i cant get a pc

[nocturne olive]

That won't help the GPU

[opaque wharf]

If the M4 is configured to like, 128GB of RAM, it could easily beats any GPU in terms of model size supported

[uneven pulsar]

i am sadly needing portability

[nocturne olive]

Yeah, but a reasonable GPU will beat it at speed on any model that fits in VRAM

[opaque sigil]

just stick to a small model (3B or 8B or sth) and call it a day

or use an api

[opaque wharf]

This is also true

[uneven pulsar]

MSI Sword 16 HX B14VFKG-202US Gaming Laptop – 14th Gen Intel Core i7-14650HX, NVIDIA GeForce RTX 4060, 16GB DDR5, 1TB SSD, 16" FHD+ 144Hz Display WIAT for 1000

[nocturne olive]

Eww, 4060

[uneven pulsar]

wait nvm

i though 16g of vram

[lime stag]

Meh I just thinking of macbook air m4 with 16gb of ram. I hear that llm is running okay even with 16gb but then again my use case is just reading i can change the writing tool code to make it more efficient.

[nocturne olive]

That's like the most scammest laptop ever

[uneven pulsar]

😭 ik

[nocturne olive]

i7 14th gen but only a 4060????

Total scam

[uneven pulsar]

i though its 16g of vram

[nocturne olive]

Worthless

[uneven pulsar]

my freind have one like it😭

[nocturne olive]

Got scammed

[uneven pulsar]

https://electroslab.com/products/buy-cto-laptop-gaming-hp-omen-16-amd-ryzen-9-7940hs-16gb?variant=46651984019623&country=LB&currency=USD&utm_medium=product_sync&utm_source=google&utm_content=sag_organic&utm_campaign=sag_organic&gad_source=1&gad_campaignid=22102298591&gclid=Cj0KCQjwpf7CBhCfARIsANIETVq4sOynN9GZrJkh8cPvehXQf8nrdgNw9tE_QeSSIqUO1R5bifLBHqQaAlMfEALw_wcB

i got this one in my mind

[opaque wharf]

Yeah that's why I said it depends on the model. Since model size also determined the LLM output "performance"

[uneven pulsar]

hope no one buys it and get it first

[nocturne olive]

Also sucks

[uneven pulsar]

??

ryzen 9 and 4070 for 1100$

[nocturne olive]

4070 laptop with a ryzen 9 is a bad combination

[lime stag]

In all fairness I also get enticed at window gaming laptop but considering life expectancy and the resale value it's like hell nah + I just wanted to not be tied to charging outlet

[uneven pulsar]

BREH come on

[nocturne olive]

That Ryzen 9 could easily run a 4090

[lime stag]

My university lack charging outlet

[opaque sigil]

ah yes, let's just casually shell out another 2 grand for a 4090 laptop

[opaque wharf]

Buy Framework that support dGPU

[nocturne olive]

Or just get a more reasonable laptop with a reasonable CPU

[lime stag]

This framework laptop cost as much as m4 macbook air in my country im like : if you want to support environment why make it so expensive

[uneven pulsar]

bro most gaming laptops in my country sells a i7 whit a 3070 ti for 2000$

😭

[ruby timber]

R&D costs money :3

[nocturne olive]

Well, most laptops are total scams

[uneven pulsar]

my country people are pretty stupid buyers

[uneven pulsar]

but this hp isn't bad😭

[nocturne olive]

Either don't get a laptop or you're gonna need to search hard for an actually useful one

[stray dragon]

what? this is a downside?

[nocturne olive]

Assuming you're doing gaming or ML, having a stupidly good CPU with a low end GPU is gonna suck

[opaque wharf]

Framework is a relatively small company. They don't have much buying power compared to other big player. That means their production run is more expensive since they only produce smaller amount for a fixed cost of the setup cost

[lime stag]

Ik but to the point of beeing more expensive then a macbook in some country

[stray dragon]

not really

[uneven pulsar]

yeah true....

[nocturne olive]

And on a laptop you can't upgrade the GPU

[stray dragon]

just means the gpu will be your bottleneck

[opaque wharf]

Say the setup cost is $1000, framework can only make 100 while dell may be able to make 1000

[stray dragon]

a laptop with a 4070 usually goes for like $1000 minimum

[nocturne olive]

And having a GPU bottleneck in GPU bound tasks is very annoying, that expensive powerful CPU is sitting idle 90% of the time while the GPU is stuck

[uneven pulsar]

but it has a ryzen 9 and for 1100$

[uneven pulsar]

yeah true.....

[stray dragon]

exactly

[uneven pulsar]

so its bad?

[stray dragon]

no

it's good

[nocturne olive]

I don't know, at least make sure it has 64GB RAM

And more than 1TB of storage

[stray dragon]

what are your standards lol

[uneven pulsar]

like a ryzen 9 and 4070 for 1100$ is good and hp makes good laptops.......

[stray dragon]

yeah that's a good deal

[nocturne olive]

The cooling is gonna suck anyway

[uneven pulsar]

no hp have good cooling

[lime stag]

[opaque wharf]

Btw, how's Ryzen AI?

[stray dragon]

it's a laptop, it's gonna suck with cooling

[nocturne olive]

No laptop has good cooling
Hopefully you're at least not looking to do ML

[opaque wharf]

Is it better in terms of inference performance?

[lime stag]

Really efficient

[uneven pulsar]

not if i got a cooling pad

[lime stag]

Against?

[opaque wharf]

Other laptop dGPU

[nocturne olive]

The laptop problem:

[opaque wharf]

Say 4070 mobile edition

[uneven pulsar]

okay fine😭

[stray dragon]

at most it'll get to 90c

the cooling is good enough to prevent thermal throttling at a max load usually

[opaque wharf]

Lap cooker

[uneven pulsar]

fine😭

[nocturne olive]

My low end HP laptop easily reaches 100 degrees at near-idle

[stray dragon]

laptop? nah

stovetop

[lime stag]

It has better gpu then most expensive gaming handheld Linus made a video about the rest of the performance

[stray dragon]

old ones like to do that

[uneven pulsar]

my current dell laptop overheats on a python code

[olive sable]

Ill have it in 2 hours superbox

Im reading the vulkan docs

[lime stag]

When will window migrate to arm

[nocturne olive]

Silly

[opaque sigil]

like 5 years ago

[uneven pulsar]

and becomes a oven if i opened IntelliJ

[stray dragon]

for lower demand loads, a gaming laptop will be fine (assuming it's not plugged in)

[nocturne olive]

Why do you need a laptop anyway? How often are you planning to move it?

[olive sable]

Windows dual screen thing was arm maybe????

[stray dragon]

if plugged in and using it, it'll be hot all the time, but meh

[uneven pulsar]

a lot tho man i give up

[rough bloom]

doesn't really matter because the NPU only ever gets used for random stuff like background segmentation in MS Teams
LLM performance always sucks because the bottleneck is usually memory bandwidth, which sucks on CPUs

[lime stag]

I mean like when will get arm based Intel or amd cpu with just as much performance and good graphic it would improve the battery life

[uneven pulsar]

fuck this

[nocturne olive]

If you want a powerful machine but also need a basic machine on the go, you could get a cheap basic laptop and use that in combination with a more powerful desktop computer

[rough bloom]

probably improves efficiency for random tasks like that though
but nothing you'd really call "AI", that part is just branding

[stray dragon]

it's a good laptop though

just know the downsides of laptops

[uneven pulsar]

hp makes good ones

[stray dragon]

cooling pad will def help

[uneven pulsar]

i do

[opaque wharf]

Me being happy with my ASUS UM325

[opaque sigil]

oh, never lol

[lime stag]

Like for so long window laptop has been tied to power outlet for performance. The snapdragon is not going to cut it too.

[maiden geyser]

*their

[uneven pulsar]

thats not mine

[stray dragon]

brand loyalty with laptops is so silly

[opaque wharf]

That's the secret. I never need their support

[nocturne olive]

Anyway, if you want to get scammed, you can get a gaming laptop all you want, I'll just stick to my desktop with a 12700K, 29TB of storage, 64GB of DDR4 at 3800 OC, a 4070Ti 12GB and soon a 3090 24GB

[uneven pulsar]

lasts for 1 year

[uneven pulsar]

29TB

[nocturne olive]

SMOL by the way

[opaque wharf]

Mine is going strong for I think now 4 years?

[uneven pulsar]

lucky

[stray dragon]

only 64gb of ram? smh low-end

[uneven pulsar]

my laptop when i run hello world python

[opaque wharf]

I just installed linux so it never has any MS fuckup that eats resource like crazy

[nocturne olive]

1TB NVMe + 4TB NVMe + 8TB HDD + 8TB HDD + 8TB SSD

[lime stag]

You'd think using logic a significantly worse version of stuff would be more cheaper then gaming laptop industry come and slap you in the face

[uneven pulsar]

wow

[stray dragon]

that isn't a hello world program

[nocturne olive]

True, it's way too SMOL

[stray dragon]

hmmmm i wonder why it uses all the cpu

[uneven pulsar]

still my laptop becomes a heater on the smallest python codea

[lime stag]

Migrate to arch Linux if you can

[uneven pulsar]

NO

[uneven pulsar]

i cant sadly

[desert plaza]

eysoos

[nocturne olive]

I need a lot of storage

[stray dragon]

surely a for loop with 10 billion iterations isn't causing problems

[uneven pulsar]

bro still my laptop runs shit

[olive sable]

Lots of british people on the train apparently.

[nocturne olive]

My high end desktop would struggle with that

[opaque sigil]

wdym 10 billion processes cripple the system

[uneven pulsar]

IntelliJ crashes my laptop

[uneven pulsar]

huh

[opaque wharf]

You can try to play marco-polo with Vedal

[ruby timber]

That's intellij

[olive sable]

[lime stag]

Sad indeed arch is so good on laptop if the tech industry make most of program runable on Linux everyone would be using it it light af heavily riced arch could run using only max 2gb of ram

[uneven pulsar]

my laptop is just to ass

[stray dragon]

it's ten billion processes

[nocturne olive]

Yeah, that program is just ultra intensive

[stray dragon]

on python

it's not gonna run well on anything

[opaque wharf]

To be fair, even on asm it would still be trash

Like, the most finely tuned asm ever

[uneven pulsar]

bro my chatgpt questions me how my laptop didnt blow up

[opaque wharf]

And rightly so

[lime stag]

True considering the fact that @uneven pulsar also can't use assembly

[uneven pulsar]

ehhh i am done whit wsm

[nocturne olive]

Laptop moment
Either way, it's not like a better laptop will suddenly make it not struggle to run that

[opaque wharf]

fork bomb, but python

[nocturne olive]

Yeah, nothing is gonna run that well

[uneven pulsar]

dude thats my python projects i am doing for someone

[lime stag]

Such a pain to learn even me with my programing hobby refuse to learn it.

[nocturne olive]

Well, why are you making code that won't run well anywhere?

[stray dragon]

have you tried changing it to not need 10 billion repeats

[uneven pulsar]

idk how my laptop isnt dead even tho i have a lot of stuff i cant run and need too

[uneven pulsar]

i did

i did it 2 times and still over heat

[lime stag]

How's does it run when you code in C, c++ and c#

[uneven pulsar]

that would kill my laptop

[stray dragon]

what.....................

this dude

[opaque sigil]

[ruby timber]

[uneven pulsar]

and when i try to open postgreSQL it crashes

[lime stag]

Get a used macbook ma dude if you can get apple refurbished one that's even better

[opaque sigil]

i would recommend you look into the concurrent.futures.ProcessPoolExecutor or something and limit the number of workers

[stray dragon]

i think you're exaggerating at this point you said "hello world causes this" and showed a program doing 10 billion iterations

[opaque sigil]

to not use all cores

[uneven pulsar]

you wont belive me

[stray dragon]

apple ew...

[uneven pulsar]

ill show you

[stray dragon]

boy who cried wolf

[opaque wharf]

To be fair tho, the 10B iteration is not the real culprit

It is what in those 10B iteration that the program did is the culprit

[lime stag]

Ik but the battery live and the live expectancy is just that good. I would love to go back to window if they fix driver issue and the battery life.

But for now such thing doesn't exist.

[olive sable]

I remember when i deleted 10 milion files and my entire file explorer shit itself and died

[stray dragon]

fair but the price to performance is too awful for me to ever consider them

[desert plaza]

i think it's both lol

[stray dragon]

^

[uneven pulsar]

bro i am not joking

[stray dragon]

mhm sure

[opaque wharf]

No, you should try running 10B iteration of just counter. It won't take as much CPU or RAM (should be)

[opaque wharf]

If you're switching to macOS, then just switch the windows laptop to Linux

[opaque sigil]

the iteration count is irrelevant, it's the fact that you're running as many processes as you have cores in parallel and they take long enough to keep the system busy

[olive sable]

I got through the prologue and how to install it
Wont be using visual studio tho, il stick with vsc

[lime stag]

Hey if you paired it against garbage gaming laptop it can crush it still. I bet if Linux got better and turn into massive comunity it would be even better for windows.

[lime stag]

Can't my program doesn't support Linux that's why i go to Mac os i can't go back to windows because I'm done with the bs issue it have.

[nocturne olive]

If it doesn't support Linux it probably won't support MacOS

[opaque wharf]

Adobe support MacOS

[opaque sigil]

macOS has users so usually wrong

[nocturne olive]

How strange

[lime stag]

It's adobe

[uneven pulsar]

my chatgpt questioning

[nocturne olive]

Scamdobe

[stray dragon]

postgresql can run on that fine

what

[lime stag]

Don't worry I have both way to pirate it in both operating system

[stray dragon]

ai stuff is iffy

[uneven pulsar]

i been running data base servers whit python ml model

some how didn't die

[stray dragon]

well there's your problem

[uneven pulsar]

idk how my laptop is alive

[opaque sigil]

is it time to shill sqlite

[stray dragon]

thermal throttling

smh

[uneven pulsar]

i cant i have to use it

[noble zodiac]

sqlite my beloved

[lime stag]

Laptop chan doesn't want to disappoint you.

[uneven pulsar]

i run a lots of stuff

but not to the fullest

[lime stag]

Saw pewdiepie used steam deck as server

[uneven pulsar]

my chat gpt questioning him self and wondering how i am talking to him and how my laptop isnt dead

[stray dragon]

it's a website

none of the compute for the llm is running locally

[lime stag]

On google Chrome.

[olive sable]

Idk what sqlite is but sure

[uneven pulsar]

its database

[olive sable]

Ah

[stray dragon]

database but made super simple to work with (i considered it in the past but it did not work for my application unfortunately)

[uneven pulsar]

no like yea but he questioning him self how my laptop haven't blew up

[olive sable]

[olive sable]

Cuz fans

[opaque sigil]

lots of writes?

[uneven pulsar]

https://tenor.com/view/unreal-unreal-engine-computer-smoking-computer-smoking-gif-24575019

my laptop almost died cuz of ue5

[stray dragon]

i think it was an issue with sqlite3 not working with something? it's been well over a year now

[uneven pulsar]

i was able to run ue5 at 1 fps

[opaque sigil]

interesting

[stray dragon]

sqlite3 and react if i am remembering correct

couldn't change from react so

[opaque sigil]

ah that would make sense i guess

[olive sable]

Real, ue5 is dogshit.
My code is wzy better, L + ratio + 6.5K fps + no crashing

[stray dragon]

yeah

[lime stag]

I wrote a code days earlier a html code to simulate nuclear reaction with accurate mathematical calculations and have glorious idea of running the thing with phone.

[stray dragon]

went with postgresql

[opaque sigil]

i guess there's libsql now which helps with the web use case

[stray dragon]

i will note that down

[opaque sigil]

it's basically sqlite but addressable via http

[uneven pulsar]

certified ue5 hater

[lime stag]

You have your own engine?

[olive sable]

Yes

[uneven pulsar]

https://tenor.com/view/sonic-the-hedgehog-shadow-the-hedgehog-dap-up-dap-marvel-vs-capcom-gif-2748262281275379258

[opaque sigil]

and i guess some replication stuff too

[desert plaza]

yeah you're right it wouldn't take much resources, i was just thinking how the program would take forever lmao

[lime stag]

Ever wonder how they code stack overflow without stack overflow

https://tenor.com/view/arthur-morgan-hmm-bear-gif-17441670916800323789

[olive sable]

Im on a train to berlin now. If i mis the 1 and only stop ill be banished to germany

[noble zodiac]

o7

[olive sable]

It has wall sockets tho, i can charge my phone

[opaque wharf]

No, it's fast. Even with JS

[opaque wharf]

I wanted to know what insanity is this combination

[nocturne olive]

[olive sable]

I don't want to be banished to Berlin

[opaque wharf]

You hate Toast country?

[nocturne olive]

Silly

[stray dragon]

plugin silliness

[desert plaza]

pardon? single-threaded?

[stray dragon]

was making a plugin for an application the uses react, and the plugin had to run in the front-end
i wanted a database for the plugin, so i considered database options

[rough bloom]

a multithreaded counter is in theory slower than a singlethreaded one

[desert plaza]

ok good to know

btw i did run a 10 billion counter for loop in python
real 16m29.214s
user 16m26.245s
sys 0m0.139s

[opaque wharf]

Why discord

I don't question the console warning. I question why is it showing more than once when I only opened the console once

[stray dragon]

[opaque wharf]

No, I question why is it showing up more than once somehow. And its lagging tf out of my firefox

[stray dragon]

it's not "somehow"

they do it on purpose

i believe

used to only be 1 time then it changed

[rough bloom]

Discord ToS violation detected

[opaque wharf]

Well, fuck them too then

[stray dragon]

[opaque wharf]

Wait wtf?

[sage crag]

turns out i can now link raylib without the binary segfaulting due to this too

[rough bloom]

Discord client mods are not allowed

(though nobody really seems to care that much about them)

[opaque wharf]

Oh, I thought its a built in features

[noble zodiac]

until suddenly discord cares and everyone is like pikachu face

[rough bloom]

yeah, they could do that

[opaque wharf]

I'm using discord web anyway lmao

[rough bloom]

there have been bans for 3rd party client use, but AFAIK only if you use APIs that can't be accessed through the normal client

[stray dragon]

that's more "selfbot" territory

they're used for automated raids

[opaque wharf]

What even is a discord raid

[stray dragon]

joining with a bunch of accounts and spamming the chat

[rough bloom]

it's mostly about self-bots, yeah
but from Discords perspective any 3rd party client is pretty much the same, no matter if interactive or not

[opaque wharf]

So Sybil attack

[olive sable]

@nocturne olive in which country do you live? Finland iirc?

[desert plaza]

@opaque wharf im confused here, are you saying this is fast or am i doing something wrong?

[stray dragon]

i see

[nocturne olive]

You would be correct

[olive sable]

Aight, shipping will be 14 bucks then

[nocturne olive]

Pretty cheap

[olive sable]

So like a totel of 530

[opaque wharf]

CPython is slow, yes

[olive sable]

Yep

C++ or rust are prefered

[opaque wharf]

No, I mean that JS is faster than CPython

[olive sable]

That too

I didnt read the context, im just hating on cpython

[noble zodiac]

some chatgpt code being slow

[olive sable]

[opaque sigil]

try with pypy3, i'm curious

[rigid snow]

it’s just a rule they keep in to blanket outlaw all variations of selfbots, api spammers, etc.

they don’t actually gaf about you using a client mod

[rough bloom]

they do in rare cases

but yeah, the common ones are pretty much safe

[rigid snow]

these cases being abusing them

[sour harness]

This does not get optimized out as dead code?

[opaque wharf]

After a while, yes presumably by the JIT

I should print the a

[rough bloom]

I wonder if that'll change with the new CEO and their new monetization strategies
since they're big on ads and gamification of stuff they might really dislike third-party clients and any modifications

[rigid snow]

sleep function

[stray dragon]

"big on ads" oh no

[opaque wharf]

Still optimized by the JIT after a while

[rough bloom]

the new guy is from Activision Blizzard and worked on Candy Crush Saga before that

[stray dragon]

i think we may be cooked

[rigid snow]

even i don’t use them - i don’t think the client mod user base is high enough for them to care UNLESS they start hindering the UX causing an increase in usage of those, then they might care

the current strategy is cool i think, add as much paid features as possible so everyone feels like they’re missing out

[olive sable]

Might need to pay tram tho

So prob more than 530

[sour harness]

Any decent compiler will constify that if the inputs are known. You'd have to do something like a += (a XOR user_input) % user_input or something inside the loop to make it impossible for the compiler to know the result.

[noble zodiac]

Like so many other tech companies discord has never made profit. Got to get the money from somewhere

[rigid snow]

more fundraising is obviously the answer

[rough bloom]

just one more round

[opaque wharf]

The user input have to change for each iteration still. This is JIT, the compiler will know after a while

Unless you generate random number, but then that becomes the overhead

[sour harness]

That's why I made the computation the way it is

[olive sable]

Sis called me to ask if i could open the front door
Im on the other side of the country

[desert plaza]

pypy3:
real 0m12.822s
user 0m12.753s
sys 0m0.027s

yeah that's better

[sour harness]

It's a (rather cryptographically weak) PRNG. There is no pattern to exploit.

[opaque wharf]

See, the thing with JIT is you can change your optimization on the fly. Any user input that is only different on the first time run will be optimized away too after a while (if the JIT notices that). AOT compilation will struggle with that

[sour harness]

Well, there is, but the JIT will not have the heuristic memory to find it

[rough bloom]

what you're supposed to do in this case is use a black box function that the compiler treats as having side effects, but that actually compiles into a no-op
sadly JS seems to have no such function

[sour harness]

The user input will have to be sufficiently large (around 2^32 would be enough for sure), but there is no way the JIT will be able to optimize the computation.

[noble zodiac]

its crazy that one snippet of bad code started this spiral

[stray dragon]

it's expected at this point

[opaque wharf]

Alright boys, time to write the loop in raw ASM then to see the baseline

Anyone would like to volunteer?

[noble zodiac]

maybe tomorrow

[desert plaza]

ok i have to try this now

[olive sable]

Life hack, if you sit behind a group of students you dont have to show your train ticket

[opaque wharf]

If a bearded man with hoodie jacket sitting behind a group of student they could also be arrested

So YMMV

[opaque sigil]

10000000000
real    5.76s
user    4.72s
sys     0.00s

[fast pagoda]

they do this because people will ask people for their tokens and theyll just give them out, or blindly paste random js into their discord open on web and then get their account stolen

[sage crag]

youre all benching on different machines

[opaque sigil]

i know

[sage crag]

[opaque wharf]

Now do it in JS and CPython

[opaque sigil]

the pypy version took like 7s for me

i'm not doing cpython

[opaque wharf]

Because I'm still cooking now

[opaque sigil]

fuck that

[sage crag]

hi do you want to see the hblang one

[opaque sigil]

also this is wrong, take off a second

[tender river]

do you link to libc somehow?

[opaque wharf]

Actually yes, I'm curious how optimized the compiler is

[sage crag]

nope

dynamically linked to raylib

not statically

[tender river]

ah

[sage crag]

mhm

statically linking raylib is a great way to get a megabyte binary

[tender river]

the rust experience

[olive sable]

This is my own country and istill feel like a tourist. Its a lot warmer here too

I also need to speak french now

[sage crag]

hblang faster than asm

[opaque sigil]

[sage crag]

_start:
    sub rsp, 0x08
    mov rax, 0x00
    mov rcx, 0x2540BE400
1:    cmp rax, rcx
    jnb :0
    add rax, 0x01
    jmp :1
0:    mov rax, 0x01
    lea rsi, [0x0000000000000007]
    mov rdx, 0x0D
    mov rdi, rax
    syscall
    mov rax, 0xE7
    mov rdi, 0x00
    syscall

compiles to this

power of main inlined in _start

[tender river]

jmp :1 is probably optimal for code size but i wonder if its optimal for performance

probably not

[sage crag]

i mean there are many such cases like this

[tender river]

many such cases

[sage crag]

xor $r, $r
inc $r

vs

mov $r, 1

for example

the top one is better code size despite taking two instructions

arguably the bottom one is faster

[opaque sigil]

i was lazy and just called to printf

[sage crag]

@umbral thorn

[olive sable]

Im almost there 10km left

[sage crag]

niuh

[olive sable]

[opaque sigil]

also kinda crazy that just this doubles the runtime

[sage crag]

silly c

[desert plaza]

sus printf

what is doing hmmmm

[opaque sigil]

libc moment

[sage crag]

printing f, if i had to guess

[stray dragon]

f

[desert plaza]

mhm likely story

[olive sable]

C has some stuff ye

[gritty dust]

Morning

[olive sable]

Iirc std:endl; is also bad cuz you clear a buffer or something

[sage crag]

it flushes the stdout buffer

[olive sable]

Ye, that

Thats a bad thing apparently

[sage crag]

you perform a whole syscall for every line printed, separately writing to stdout each time

[olive sable]

Oh

Brb, tram

[sage crag]

when you dont endl presumably the writev syscall can be used instead of write which allows multiple things to be written to a buffer at once

[gritty dust]

Kk

[olive sable]

Waiting on bus now

[hoary lion]

niuh

[olive sable]

[opaque wharf]

Really looks like my old college

I'm waiting for my curry to finish

[fast pagoda]

god i miss walkability

[uneven pulsar]

dam you're lucky

[uneven pulsar]

i can translate

je parle francai

[opaque wharf]

Just finished beating the meat. What did I miss?

[desert plaza]

uuh not much i guess

[opaque wharf]

Oh, so not much huh

[desert plaza]

mhm

[fast pagoda]

i abandoned svelte

screw that

[opaque sigil]

[fast pagoda]

it made the whole thing need a whole restructure and 300000 extra modules

[rigid snow]

svelte is (or was i guess) nice for people who don’t do frontend because it abstracted state into variable syntax with a compiler

[fast pagoda]

i think it wouldn't be a problem IF starting from that point

but trying to basically rewrite what was already there but in svelte was

annoying

monolith -> whole split up project structure

plus biome doesnt fully support svelte

[opaque sigil]

was

really can't have shit

[fast pagoda]

biome HATED it

it didnt know what any of the imports were doign

[opaque sigil]

yeah biome doesn't support svelte

[fast pagoda]

anything inside of the .svelte files it was joever

biome was just like UHHHHH WHAT

so yeah that was too annoying

now im just doing speed things to the monolith, screw it

[opaque sigil]

that's the one thing eslint and prettier have going for them

they're so well established that there's plugins for basically everything

[fast pagoda]

we got: rAF loop marking the canvas dirty and rendering next tick instead of directly calling drawGrid

we culling the viewport to visible area only with a small buffer

we layering the canvas

so highlighting doesn update the whole thing

and most importantly im working on having an offscreen canvas

to blit updates to the main one to instead of updating the whole thing in view

[sour harness]

I love how Claude 4 and Gemini 2.5 disagree about code comment style if there should be a period at the end or not. I use Claude to write the code and Gemini to write the documentation, so Gemini will add periods to all of Claude's code comments...

[fast pagoda]

i cannot stand llm code comments

i neurotically rewrite and remove them

if they are added by something

this script.js had a bunch of them from when yt had chatgpt poop it out

i have redone every one i've come across so far

because it can't sound serious or i'll pass away

plus they always do like

foo = 250

//250

???

it already says that

[sour harness]

I'd call that a prompting skill issue :/

[fast pagoda]

listen if i'm prompting the ai to comment in a style it's joever
they should just learn my comment style from context if anything
or better yet not comment at all unless asked
only an issue with stuff they generate though so it doesn't matter as much in the end unless that's the entire codebase (in which case, it's joever)

[opaque wharf]

Vue my beloved

[fast pagoda]

but i didn't come up with this script originally so i have no choice but to rewrite them all

[sour harness]

Ask it to read your code base and create a documentation style guide based on the documentation that's already there. Then use that to generate new documentation.

[fast pagoda]

i guess but then you didn't document when you wrote the code (y not )

CYA CYA CYA cover yo ass document document document

my commits b like 200 changes and 190 are angry comment updates

oh yeah now that's performance

lol

[opaque sigil]

[sour harness]

I tend to do those all at once

[dry charm]

LGTM <merged>

[opaque wharf]

Last batch is being fried

[fast pagoda]

bruh i had greptile tell me this the other day

[olive sable]

Giys i missed the last bus cuz the guy asked for cash and i only had card. Now im on the bus an hour later and the guy said theres no control so im good. I even went to an atm

[fast pagoda]

it's over

well

we are so back

but it's also over because you wasted your breath

going to the atm

at least it worked out it seems lol

and free??

[olive sable]

And i wasted 9 bucks on an escooter cuz the gps said i cpuld go there but then it made me go to the highway

Its so joever

[fast pagoda]

lime got yo ass

well idk what the e-scooter company would be there

it's lime and bird over here

[olive sable]

Im mussing like 50 bucks in temporary uber charges

None of them would take me

[fast pagoda]

sheesh

[wary rover]

There is one thing that I cannot really grasp on my own, so I'm back to bugging you guys. When using serde_json, does the code just make a JSON file for itself and once prompted to use data out of it, it will just... Find it by itself? Like I am kinda used to the fact that everything must be specified and from tutorials I watched it seems that I don't have to specify a file path or something like that

[fast pagoda]

why lol

[olive sable]

The worst part is ifk if il make the last train home now

I might need to ask someone to pick me up for the last part

[fast pagoda]

all it does is convert but you have to provide the json

or the rust

all it does is serialize or deserialize

you have to put the stuff in memory that you want to convert

in some way

doesn't really matter how as long as it's there

[wary rover]

That's why I am confused. Like I rewatched this tutorial a lot and it doesn't get through my skull how did he pulled data from the JSON he didn't create or specify. Like I am definitely missing something painfully obvious, but I can't understand what

[fast pagoda]

it's likely a hard coded string literal

tutorials do dat

[desert wave]

isn't it there

[fast pagoda]

oh is that the entire thing

[wary rover]

Its a tutorial

It is simple

[fast pagoda]

well it's operating on the dog01 struct and serializing it

so that's where it comes from

[wary rover]

Oh, he serialized sruct into string

[fast pagoda]

ye

[wary rover]

Why JSON then

[rigid snow]

notice how there isn’t an angular simp here

[opaque wharf]

Just finished making like 8-10 of this

[rigid snow]

camera quality does not do it any justice

it looks depressing

[fast pagoda]

let dog_ser = serde_json::to_string(&dog01).unwrap(); basically passes a ref to dog01 to the to_string func
then it calls serde_json::to_string()
this uses serialize which was generated by #[derive(Serialize)] by the compiler

[opaque wharf]

What can I say, I love my outdated phone

[rigid snow]

this looks like an iphone 3g era pic

[wary rover]

I am severely confused why does it need to be serde_json

[fast pagoda]

id imagine literally just to demonstrate what it does

not out of necessity

[wary rover]

Mhm, ok

[fast pagoda]

cuz yeah i mean why generate some json and then serialize it to a string

unless you need the json somewhere

[rigid snow]

basically serde itself is a base to serialize data into formats that are implemented on top of this base