#programming
1 messages · Page 163 of 1
check 9600x
ok
or alternatively, 7600
well 32gb sounds enough for me
and 1tb is also enough
32gb will run out faster than you'd think
64gb💀 ???
once you have like 3 or 4 of those AAA games, 1tb is nothing
32gb ram is fine
for your build i'd recommend at least 2tb nvme
3tb?
ryzen 5 7600 with 2tb ssd is better than 7800x3d + 1tb ssd ngl especially for AAA titles
hold on lets make a pcpp
for me 16gb ram is the minimum unless your running a old laptop
why 7600
because it makes no difference in most graphically demanding games
there are r5 9800x for 200$
compared to x3d
huh
yea
you mean 9600x?
9800x
that doesn't exist...
9800x3d exists, but too expensive for your budget definitely
try to check both 9600x and 7600
lmao
^
excuse me
7000 series gpus aren’t available for me
they only ship nvidea
uh no, cpus
both avaible for 200$
now that r9 cost 400$
yeah get 9600x
going higher than that won't give you realistic increase of performance on your workloads btw
98$ 32gb on crosshair ram
can you find a B850m motherboard
ob
ok
can you send the link for this so i can check the specs
Alr
thats the link to the motherboard
aight
isn't this a bit too much for motherboard though
is the motherboard compatible?
its the only one i found
huh
other are old am4 motherboards
that one?
Buy online MSI Spatium M371 2TB NVMe M.2 | M371 at ayoubcomputers.com, Find low everyday prices, buy online, and enjoy free delivery in-store pickup. AYOUB COMPUTERS Best Tech Store in Lebanon
anything else?
Buy online MSI MAG A850GL PCIE5 power supply offers 80 PLUS Gold 850W | 306-7ZP8A11-CE0 at ayoubcomputers.com, find low everyday prices, buy online and enjoy free delivery or in-store pickup. AYOUB COMPUTERS best tech store in Lebanon
no sadly
damn
the ones for 90$ are 1tb heat sink sssds
because that's a gen 3 nvme so its gonna be a bit slower
why the fuck would a ssd have a heat sink
oh
because they make heat
wait then let me try to chek
but the motherboard should come with its own ssd heatsink ig?
Buy online Kingston Fury Renegade 1TB PCIe Gen 4.0 NVMe M.2 Internal Gaming SSD with Heat Sink | SFYRSK/1000G at ayoubcomputers.com, find low everyday prices, buy online and enjoy free delivery or in-store pickup. AYOUB COMPUTERS best tech store in Lebanon
89$
1tb
kind of expensive for 1tb ngl
thats what i could find
hm
i think that 2tb should be fine for that price
but gen 3
yknow
i just remembered a friend of mine sells nvmes he can make a price
i've been using a gen 3 drive for long
its fine
its not much different compared to my new gen 4 drive ngl
alr then cooler?
i know a freind can make me a price for the nvme
Buy online Thermalright Peerless Assassin 120 SE CPU Air Cooler at ayoubcomputers.com, find low everyday prices, buy online and enjoy free delivery or in-store pickup. AYOUB COMPUTERS best tech store in Lebanon
perfect
matx
so any matx case would work
https://pcpartpicker.com/list/swrQHW
so far ig
Part List - AMD Ryzen 5 9600X
theres a wattage problem
matx?
yeah because i used custom part for the gpu. can you send me the link for the gpu
preferably in dms so people here don't snatch it
Buy online DarkFlash Dy470 PC Case With 4 ARGB Fans - Black | Dy470 at ayoubcomputers.com, find low everyday prices, buy online and enjoy free delivery or in-store pickup. AYOUB COMPUTERS best tech store in Lebanon
damn expensive lol
do i change the case?
huh
up to you
if you go with that case, you might went over the budget a bit
MANLI NVIDIA graphics cards
they used to make 20 series card
there back making 50 series card
ill ask the seller
oh, this brand is relatively trustworthy
but yeah best to also ask seller
recommended to ask the seller to test the card
alr
there a 3090 for 530
yea bad price
NO WAY
I FOUND A MSI SUPREME X 3090 420$ AND TI VARIANT
ITS FUCKING SEALED AND BY A RETAILER
oh you can try yeah
although might wanna cheap out on something else if you wanna stay on budget limit
oh
like what
but that’s a supreme x 3090 ti
https://pcpartpicker.com/list/DYcLmC you can modify this list from here
Part List - AMD Ryzen 5 9600X
Spaghetti
Ive been putting this off for more than a month but imma need to connect everything
Except the video outputs cuz those will need to go into the new avr once i get it
If i get it, its bidding
can someone explain to me how my benchmarks become worse when I tell rustc to build with native CPU features?!
literally cargo criterion vs RUSTFLAGS="-C target-cpu=native" cargo criterion. Up to 10% regression. wtf?
bro's just like me fr
This is the reason im poor
hapy byteday
Im not poor atm, but thats only cuz im in the procces of buying something 
you should see the performance regression i get when turning on LTO on my vm
57% regression when using gcc
thank you 
okay but like.. how
I'd think native cpu features would speed things up no?
likely some cache pessimisation in a hot path
well, think of it this way
ah, true, didn't think of that
without cpu features, a memcpy could be a 10 byte function
with cpu features, it could be a kilobyte function
since they do vectorisation to make copying large things faster
so, PGO plus enabled cpu features should work then
perhaps? i dont think pgo alters codegen in that way
it would at least help the branch predictor, no?
dubiously
Hope you have a good one!
the compiler would attempt to
worth trying though
so far it's been work only :P
maybe BOLT too
try adding
codegen-units = 1
lto = true
to your Cargo.toml
does that even affect benchmarks?
oh I could set it for the bench profile
probably
happy birthday @scarlet arch 🎂
thank youuu 
I meant to go off work 12 minutes ago and now I'm sitting here trying to gain some performance with this stupid experiment 
I have also had this before with some very stupid benchmarks
thanks
my thinking is its because benchmarks start the program from cold each time they run it

nah this was also with criterion so it should've been hot
Did this:
[profile.bench]
codegen-units = 1
lto = true
And it improved some benchmarks again, but not above the loss from earlier, and some became even worse 
so if your code is larger, and its harder for the cpu to get caching and branch predicting effectively
yup performance lottery
meh fuck it, its fine
souds like cache
random question but how fast is the thing you're benchmarking
rip to Awalang
because I had this at something that was like 3bn it/s so I kinda get that assembly instruction bingo may have gotten unlucky 😭
several benchmarks, but one iteration for the various things is between 29µs and 60ms. but the timing per benchmark is consistent
this is something all stack machine s struggle with
ah fair
my Awalang actually lost its original meaning tbh, it's now more like an intermediate language, as I heavily rely on the use of Awabler(Assembler for Awalang) 
even moreso if the interpreter doesnt use direct threading
you should look into direct threading
i used it in hbvm-c
we should get rid of stack in x86_64
neurch64
x86 cpu interpreter is not a stack machine
its a register machine with a stack
but also its a hardware cpu
it doesnt have the drawback of being emulated
not getting the true AWA5.0 experience by not directly writing awatalk as recommended by the spec
-# I would've probably done the same if I had an assembler
we should write in machine code, as tools may get up-stream infected
hardware mever betrays

it's AWA5.0++ now 
good idea 
god damn it

happy birthday @scarlet arch
HAPPY BITHDAY BYTER @scarlet arch 
https://x.com/GamersNexus/status/1959751632113299534 average b2b communication 2025
Dear Bloomberg @business, please find enclosed our formal response to your illegitimate copyright strike. Thank you.
https://t.co/zXisNkhDNF
I wonder; would Vedal or any of you try to hack google in order to get a job at google? For context they do or did (Idk if they still do) a "hey hack us and you get a job and some money"
Also
https://www.youtube.com/shorts/9lFEzglYn2k
This is why we don't f around with client side for important stuff
that's how vedal got his position at mi5 and how he continues to secretly reside in buckinham palace
@nocturne olive kotlin mentioned in devstream
Oh great
everyone in chat hates kotlin
Windows delayed the next original, EVERYONE LINUX NOW, let's KILL WINDOWS
THis is exactly why I switched to Bazzite yesterday
too much effort for me to switch
no its not just nuke your drive and forget about it
im not nuking 11tb of data
buy a new drive and install linux on that
Is there a program for installing Linux on windows machines without destroying user data?
Just do what I did and spend 2 days backing up data to an external drive first 
there are too many programs i use that im not sure have a linux equivalent or work on linux 
well. installing linux doesn't nuke user data. linux can just access windows files
i had dual boot a few times..
I know, but I was thinking about the people that hardly knows what an OS is.
but always just booted into windows so i just uninstalled linux after some time 

Couldn't respond fast, shower haha
uh oh he said twitch
uh oh he's gonna display b2 messages on twitch stream
TOS on both sides
its over for vedal987 channel
Coding with chatGPT live on the stream 
I love how close this is to helping out a junior / intern
doesn't the defer attribute do the same thing as moving the script after the body?
not sure if it happens to end of body scripts
but defer downloads the script and then executes them when the page is parsed

@sudden belfry one of #programming

Welcome to the club 
Technically she could also run the 🔺
now she needs to play screeps to be fully part of here
hblang updated to 0.15.1 like 4 days ago
om
honestly looking at her code. she might be better than us at screeps at this point
oh so they just silently pushed the update without creating a github release huh 
did you know hbc will be a c compiler soon
Let her make her usb killer in peace guys

what
today on bred shenanigans
it will be able to compile c and hblang
because its using zig it can make use of aro cc
the same backend that the c to zig transpiler uses
its being done because there isnt enough written hblang code to test everything
but also it may let use import c files natively, i hope

the extension allows for terminal access
so technically yes
After mike 7 hours im done with cable managment
Or well, its mostly managed
I dont have enough inputs on this tv so thats why there are cables on the floor left
@sudden belfry is a real programmer now, welcome
(shes better than 90% of this chat)
As CEO of the triangle: 
Have some funny logs from the webserver accesses
she would be better then 99% of the chat if instead of regex she could write patch files and have those automatically applied
moving the cursor and editing the line is just a bad way of doing it
i understand why its done like that
she should rewrite the code for her own vscode integration
fr
also vedal should of setup eliv's page to /dogs instead of /dogs.html
literally needed to make a folder called dogs and move dogs.html into it and rename it to index.html
birthday 
you have 22 min left 
Yeah, but then it's not made by Evil
She's old school like that
literally renaming a file and putting it into folder still very much made by evil
Happy Birthday @scarlet arch!

You just don't get what I mean I think
Wow, open F12 and peek at the spinning Samoyed’s CSS—it’s a mess. The rotation’s calculated pixel by pixel in real time. No wonder the page hogs 62 % CPU in the browser task manager.
It doesn't need to be perfect. Evil's site is good the way it is. 
- diffs are not 100%, models no matter the size still mess them up - tools like cline use fuzzy diff matching, and tools like cursor(tm) and copilot do something different with json full of lines as input and them writing edits for each line in order or something?? i'm talking about the thing where the line that goes down gradually as it's editing
- from my limited involvement with the project: it was initially just neuro api bindings to basic vscode actions, like moving the cursor, typing, etc. but there was the issue that you could move the cursor in between neuro actions and it would mess up the position neuro intended to insert text at, so it was decided to give her "her own cursor" as a bandaid - so that's why it's like that
- it's cuter this way
yes they do sometimes mess up diffs it happens alot 2. i pretty much just guessed that and 3. its great ragebait aswell
perhaps instead of going to diffs neuro/evil could be given better context for where they are
i haven't looked much into the context stuff at neuropilot because it's kind of a mess but i think it already should force push the code surrounding the cursor to ctx
yeah i also don't think its entirely neuropilot fault the twins do need to also learn how to use the cursor better
they can't
not how llms work unfortunately
unless vedal specifically trains them to which i doubt is happening
im talking of having more context of moving the cursor first in they're memory
well for that they would have to do it perfectly in the first place
maybe they can slowly figure it out but i dont know about that...
No, if you give a person a few examples of how something failed and how another thing didn't, then they'll be able to derive a better course of action
But it won't be that big of an improvement for an AI
ultimately theres no perfect solution diffs could potentially work better but they could also not as with everything it really just depends



You're right. the CSS style for it just keeps getting larger every frame
why are you surprised
Because i've never seen this type of code in my life 
the line literally was ```js
sammy.style.transform += rotate(1deg) ;
I didn't know you could do that in the first place. That's messed up 
Sure, there's probably good reasons to use it. I'm just surprised you can do that 
i don't think there is ever a good reason to do el.style.<any> += , but since it's a string you can
I'm amazed. Now the dog just keeps increasing my RAM usage by the minute 
yeah as soon as this came up i was like "this is a memory leak machine"
sam change your benchmark to render spinning neurorbPogHD
gimme some time cuz i need to add a vertex buffer and uniform buffers for that
and a texture buffer
uv's too
uhh
also gltf loading
Orb
sam make the orb texture 64k x 64k for proper fidelity
4 textures 
it's not impossible i'll tell you that
well yes, but actually no
also that should be x65k but whatever
unless i add a 2nd uv buffer purely for texturebufferarrayindex
that way we get kinda limitless resolution till vram is up
huh, my brain shortcircuited there
dutch word for "ran out of" is "op" but i tranlated it to up
idk why
apparently yall use "is up" for only very specific circumstances
"food is up" sounds stupid and clunky but "vram is up" sounds fine, to me at least
probably because of that exact metaphor, you're full, not out of
idk english really is weird
The whole thing with "old english" too. They keep stealing things from other languages but bever commit to it
OM
Starting in September 2026, Android will require all apps to be registered by verified developers in order to be installed on certified Android devices.

apple stay winning^Wnot losing
so glad i use mobile linux and this doesnt affect me 
So I decided to bite the bullet and reinstall. Wish me luck 
I should really deal with my older devices, no reason for them to be on android.
-# Still need more knowledge for that tho.
DON'T DO THIS TO ME GOOGLE! PLEASE MY WORK IS DOOMED NEXT YEAR 
youtube redesign 
What the fuck.
Perish fool
They cannot stop making their UI uglier by the day
ARE YOU filtered ME?????
What if i just dont update my phone? I dont want to register shit.
Can i turn it off in dev meny maybe?
I'm NOT giving my ID to google. NEVER! frick google dude
Remove google play protect 👍
i'm sure there will be some workarounds, but its a good time to switch to mobile linux or at least learn custom rom flashing 
okay
Honestly, I'm considering following your apporach. I'll have to give that blog a reread
I'll just be doing it on my older stuff first
I give my id to anyone who asks for it
I need android apps still for school, but maybe i could dualboot or something
waydroid
(we hope)
whether a device supports mobile linux is a question the answer to which is generally "no"
for something that just works, use ubuntu touch with a supported device (this will get you camera/etc) https://devices.ubuntu-touch.io/
Ubuntu Touch is the open source phone that has freedom and privacy in mind. Supported by dozens of devices, with installer for FOSS fans
I have heavy doubts my devices are supported as LG and ASUS as my daily driver
I'm already a verified developer (I think? I have 3 apps on play store) but this is going to be annoying either way
and for something that does not just work but uses proper linux (not whatever was originally shipped with the device), use https://wiki.postmarketos.org/
you can run pmos on an iphone x
as you can see it wont be a very good experience
but it will work
Nobody can steal or sell your data if it is freely accessible 
Seems my option is suffering then, not on either list of supported devices.
you can find an android cfw
At least on the LG devices, I'd lose the dualscreen function of those phones. For the Asus maybe, but I would want to be sure about it as it is usually my fallback when all else fails.
https://github.com/microsoft/VibeVoice?tab=readme-ov-file
interesting new tts model from msft
The name
I'd rather network block those devices rather than remove the funtionallity I bought them for
not completely multilingual 
For anyone who wants to do research on my behalf, The device list is: LG G8X thinq, LG V60, Asus ROG Phone 6 (12gb, US). I don't expect anyone to actually have answers to these but just for transparency.

gemini 3.0 imminent, or perhaps the native image gen for 2.5..?
hopefully whatever they ship, its gonna be tmrw
Color
no way, turns out when i use an editor instead of just printfing some code it actually does syntax highlighting?
this is dependency injection

late binding
loose coupling
enterprise ready
let's be fair
it almost never is perfect, let alone on a platform where the port of ios 7 to a device that officially only supports ios 6 looks like this
Alrighty, this NN section is complete. Now I can begin working on other atomics before I piece it all together.
After I get rid of some bloat in the code actually
this also works
if self.modules[inner.name].start() is blocking should I use asyncio or threading
god i phrased that terribly
doesnt really matter what the function is
its blocking and I need it running concurrently
should I use threading or asyncio
or should I use like
both
hitchhiker's guide to the galaxy reference???
if its blocking you should use threads but note that if the function does some processing it will still block by default because of gil
in python 3.13 theres experimental support for proper threads that dont block even if they're computing something
the threading is just going to be hard to manage though
but you have to use a separate version of python
concurrency is always hard to manage 
since i'm writing a handler for plugins
so if i start them and the start function blocks
that would obviously kill the entire system
i'm guessing threading since asyncio is more loosely related to the specific task
yuh
though if I rewrote the plugin structure, it could probably be very clean
I'll try threading for now
even async functions can block 
if you want to run arbitrary code in general theres obvious potential for breakage
if you want to protect against infinite loop thats completely valid but you'll have to find a way to make a blocking function stop on a timeout or something
i dont think python uses structured concurrency for threads
ok another random question incoming:
should I start and save the thread within the plugin itself (i'd be adding that into the base class for plugins) or should I do that within the plugin loader/handler
i'd keep it in the loader
would be safer
since it may have to use thread pools or timeout logic or whatever, plugins doing that themselves is just messy
that's my thought
I think I might use thread pools
seems like the cleanest option
though the threads would be running until shutdown externally
so i'm not sure if it is necessarily the correct thing to do
ideally you'd tie the threads' lifetime to the plugin's lifetime
look into structured concurrency
my cat just confused me so much
she scratched the carpet and then I was like "dont... berrr...."
then she meowed
and i was like "dont start... wait what?"
"don't scratch, i mean."
least noticable effects of programming:

I didn't get it before. Now I totally understand why they call it the swarm. At least maybe
this is fucking dumb for reasons
Anybody familar with the neuro api here?
bro is NOT here 😭
How do you send a register request via the API to register a command
🪱
???
theres a spec
{
"command": "actions/register",
"game": string,
"data": {
"actions": Action[]
}
}
This?
mhm
Thanks 🙏
thoughts on user@hostname in a shell prompt?
i feel like it's appropriate to shorten it to just <user>
but a lot of people remove the entire sequence
its good if you ssh a lot
you can make it only show if you're in an ssh session though
i'll do that ig
though no one is going to be sshing into my laptop
hopefully
also why the hell did obs just trigger a full on system crash
I closed OBS and it crashed my entire laptop
Does anyone know what programing language vedal used for neuro?
Python for AI, C# for model
i'm here
random question that popped into my mind
do the same materials that block radio waves, block magnetic fields?
I figure the answer is probably yes given how related they are, but I've never seen it explicitly stated
if i had to guess, i'd say no
but that's just me
(nevermind i dont have image perms yet)

actually I guess magnetic mines are a thing
but then I'm confused why we don't use electromagnets for communication in water
because magnetism is way way way too spread out to be easy channeled into a beam so it dissipates
Actually i think it's not the same thing
oh is it just too hard to direct
well
no but why would that work 
Just adding rotations to eachother
I've had it run for like 10 min and it's 500k characters
remember in phineas and pherb where he makes a magnetic laser
thats actually really really hard
so it goes everywhere
They work on different principles radio waves are blocked by conductive materials, like a metal box but for a magnetic field, you need a special magnetic material to guide the field lines around the object you want to protect
and thus dissipates
like if you flooded a room to drink water instead of just using a cup
you'd need too much for it to be practical
and it'd damage stuff
same with using magnets like that
hmmm I see
wait a moment
BIRTHDAY BOY????
why is everyone born in august

Do you like physics Shiro?
wtf why does x11rb do this 
who creates a region, destorys it then uses it to set window shape region
this is what that is supposed to look like
so
i have spent 4 hours on this i can't believe i need to pull out xtrace 
this may come as a shock to you
uh oh
but I was not born in august
wow
shiro u are born everyday

wow
istg its gonna be because x11rb figured out the shape isn't used anymore since ive gotten the shape region or some shenanigan like that

hi shiwo 
why has nobody made AI do sentiment analysis on news headlines to predict stock market impact
I can't believe eliv first memory leak is not on starboard
you could just hook it up to a bunch of RSS feeds or something
does that not already exist

It's all clickbait, def gonna trick the AI to hell
The difficult thing is feeding the AI good data and maintain the AI online seeing the news but it could be done
im fairly sure its already being done, but not via news headlines
it's not like the humans are fed any better data
thinking now it could really be done, if this Ai had access to social media like grok
i think hedge funds firms already do it, i remember some friend talking about it when i was too drunk to understand anything
(said friend is from hedge fund firm)
yeah you could follow corporate media accounts
or maybe you can make a prompt like: "See this profile, this profile, this social circle of idk money and crypto and predict the possibilites" something like that
I feel like we are, especially when it comes to faulty products and such, I feel like I subconsciously (and appropriately) avoid some brands without realising the reason, and then I realise that I've done well doing so
I feel like you probably don't do as well as you're convinced you do
and I feel like that's true for most people
yeah nowadays ai trading does so much better than standard human beings
not surprised
you'd genuinely need to be an analyst to beat some of the bots nowadays, and even then they use AI to help them
Correct, but probably better than what AI would, knowing how much it gets baited by clickbait and unfounded claims
stock market is just fake items being exchanged by fake shareholders
Hmmm idk
what a time to be alive
u can doubt me but im now a broke student while they're earning 6 digits
I mean you would need a really really good AI who remembers everything and this is not so easy, but i think corporates and rich people maybe already have one
and you would need a lot of ram and a good gpu like idk Rtx 6000 pro?
To context length
are you talking about consumers or companies?
companies have access to way more than that
they don't need to own the hardware when they can rent them
i think news sentiment analysis is only one part of the equation? they also take into account of things like shareholder reports, annual financial reports, and stuff like that to make a decision
do they really
The stock market and investments are a constant, it would actually be possible to do it with a rented GPU, but you would have to have it rented for a long time or pray not to lose the timing if something important happens and the AI is not active at the time.
I feel like this a low-hanging fruit vs. high ceiling kinda thing. AI will do fine, but an actual professional will do good
if i knew i wouldn't be poor
just spam calls and u will surely make money
(This is not financial advice)
Superficially yes
all the analysis sounds like stuff best left to finance bros who will invest their life savings into a venture capital and then boast their 0.5% annual return
yeah but when you are a VC / Fintech / large enough firm rented GPUs is easy enough to get
it's only hard for consumers / smaller businesses because we're constrained by budgets
When it comes to large corporations, you can be sure they don't mind buying the hardware instead of renting it. But for us consumers, this is more difficult.
Unless you're a Vedal. Someone who can work miracles with an Ampere and faith.
omg 0.5% is better than my bank rates

What you do for living? (curiosity)
at some point they do have an on-prem cloud, but not everything is done there
on-prem clouds are typically used to crunch more sensitive numbers, usually closer to the last stage of decision-making
but i only have info from the yapping a friend gave me so take it with a grain of salt 
I'll let rich kids buy GPU farms and experiment with it on my behalf
I thought so too until I used it and understood the science behind it.
Same, AI feels like "Throw data into a wall until it makes a dent that we can showcase"
it's just too hyped and people are trying to put it in everything
it's a useful technology but heinously misapplied right now
im going back to uni to steal ideas about ai safety, because it's related to work
fuck uni
real
shr
and my problem was not treating it as such
and mr sigma

I'm getting the easiest and quickest degree and then dipping
because there is zero income increase associated with a better degree for me at this stage
non-honours bachelor's nets me a 6 figure salary
wait shiro aren't u already graduating
i plan to try emacs next year (on my helix arc right now) what resources would you recommend?
only one course after jan
a church


neovim
om
if you aren't on a potato there is no reason to not use gvim
get with the times
if you want a painless (relatively), u can start with a well-known distro (like spacemacs) just to obtain muscle memory
Emacs ecosystem is too split
get back to me when your config is like this https://github.com/chayleaf/dotfiles/blob/master/home/modules/nvim.nix
then, slowly implement the configs that u get from spacemacs on your own
painless? emacs?
my journey was Doom -> vanilla emacs -> Prelude
linux users spending 10 hours to set up shittier notepad
(it saves them 5 seconds per year)
i am still dealing with the 5 possible ways to fucking indent C code
only 10 hours? damn
My nvim config is prob really fucking large
It involves math, calculating losses, and the most tedious part is being able to gather the dataset and feed the knowledge into the AI without it becoming stupid or distorted by it.
AI is complex. Training an AI is like raising a child. Most parents can get their children to eat and pass fourth grade, but not all can get their children through college. Idk not my best analogy
DEBT
1356 lines
vanor is rich
no i am poor
thats very small
vanor is not rich
everyone is not rich compared to the 1%
And I have tons of stolen code too
if it makes you feel better my parents asked me if I could lend them money
if that's any indicator of my financial sphere
It's your birthday?
that's kinda wild tho
unfortunately not my birthday
btw my emacs config
I've funded my own schooling 100%
(what)

the reason why i didn't screenshot earlier is because I unbound my zoom keys and i forgot how to zoom with M-x
So... Why the cake?
I CAN'T TAKE IT OFF
Cos it's shiros birthday

wow who would subject bro to such a fate
unhapy unbirtday
that is so mean
wtf is that noise outside
:enub: stocks just keep increasing
cryptid 
A filter-sama plushie would sell a lot...
that was enub
it sounds like a bird but not quite right
There was no noise, SCHIZO or sum?
something is wrong
it was
ops wrong chat
shiro u is die
im closing my window I'm scared of the midnight creatures
i'm closing my helix window because im scared of what i wrote
nnuueerroorrbb

nuru

T
hi shiro
hi t
ⓣ
yes it's me
I don't understand the culture of this community (New leaf)
Out of curiosity, what is your setup, guys?
3 rocks and a sparkler
hardware
chaotic and silly
nobody here takes things too seriously
i might write my own editor some day but busy with languages right now, my current long term project is a 100% incremental pure/deterministic language which i hope will be a good fit for dsls and further language development
omg wait oh yeah you're a new guy
there's a thousand running jokes that you just pick up as you go
remember this: vanorsigma does not hate evil
no matter who tells you otherwise it's fake news
i expect lots of explosions 
Can it run minecraft?
laptop + bfkb113pbk + xp-pen g640 + logitech g502 + a random xiaomi monitor
what's minecraft
craftmine
bro what is a bfkb113pbk? 
keyboard 
It's a movie
is produced by the manufacturer slamming their head into the keyboard
thats an interesting way to wash a keyboard
it's okay just dry it before having power supplied 👍
I mean I use to wash motherboards with water

Someone should make a looping animation of that
👀
...
I use isopropyl alcohol
My guy
Does that mean I'm so fly
tbh if any liquid is pure enough, it's not supposed to conduct electricity thus no shorting
glueless
liquid carbon
liquid steel
I like the smell + can also get stains off and stuff
liquid sodium
liquid uranium
liquid metal excluded
yummy
Liquid human
also liquid salt
I wanted to use liquid metal as a cooler in my cpu and gpu vram
does molten salt conduct electricity??
but too risky even for me

yes
The remainder from the iron sword project
that's crazy
they actually use it in the extraction of aluminium from aluminium ore
but I think your PCB is already melting if you pour any molten chemicals on it
Do you guys drink a lot of coffee?
because aluminium is so reactive it quickly forms an oxide layer in water
so they perform electrolysis in molten salt instead


Go to the chem channel or smth
#chemistry
no, the good beans are too expensive 
I like tea
✅


How much does it cost where you live?
more than not too expensive
because it's fluid
a kg of speciality coffee is like 40€ or sth
security theatre aah password management
What...? Wait let me see something
its not post quantum 😔
Wait a minute you're telling me that 1kg of coffe is 40 bucks where you live?
but also security wasnt valued as much in early 2000s
not all but the highest quality beans are (i think normal ones are like 10 per kg or sth, it's been a while)
you can find plenty of old cryptography books that cover md5 and say it isnt veyr secure
but it was used anyway

The highest quality coffe here is like 15 bucks
fuck coffee, pure coffein is way cheaper
that's just miserable
But it makes sense I live in a country that is the largest coffee producer in the world.
I like my fun in powder form
Brazilians don't live without coffe
They say the human body is 70% water, here we are 70% caffeine, btw im italian but i live in brazil
Where do you live?
my main problem is that i don't feel like having coffee nearly often enough to avoid the beans getting stale so there's no point 
germany
I've always wanted to visit germany
it's certainly one of the countries of all time 
This is a very programming topic
I hate tea...
I JUST CANT
What teas do you drink?

I've got a both a question/taught but not sure if it belongs in here or not since it's more like the difference between the organic brain operating and our current binary computer systems
its fine to ask here since this is basically the stem channel
but im lazy so i wont talk about it 
So basically it's the fact that Vedal has to clear Neuro & Evil's memory cause it takes to much space, knowing from my own job which is first line support, disc is expensive especially to store data, our current computing systems use binary encoding which even with the best optimization will have limits, in Star Trek for instance the isolinear chips they use are 5-state systems but they also have limits in storage & processing, but organic brains don't really have that issue, sure memories go hazy and details change but external stimuli can recover memories of even decades ago, so it's clear to say the brain does definitely not work on binary, it's both a "hard drive" and data processing unit at the same time
also I became obsessed with AI when I first saw Serial Experiments Lain in the early 2000's, I'm a bad coder at best but an excellent dba
I see where this is going, qubits? quantum computing?
ugh ok you made me want to answer, brain has a lot of capacity actually but the way it stores data is in its own structure. its just different from having an external memory where you write stuff
additionally, binary is good because its so simple and easy to make, a (e.g.) 5-state system may sound cool but if it were cheaper we'd be making that
if you do a very remote comparison (i said very remote), think of brains as neuro/evil's core language model, and of books as their memory
the reason its a remote comparison is because llms are just inherently not designed to grow and adapt (even if they may pretend to grow in certain scenarios)
as an example, trinary computers were tried in the 20th century, and they were actually superior in certain areas but the way they were made is by chaining bits and pretending they are trits

How much ram and memory neuro and evil should have? I mean they can remember a lot of things, and respond with a really good latency
ram is mostly only relevant for the live2d side
vram - as much as their llms and accompanying models (if running on gpu) use
memory - as much as vedal lets them
What do you mean by only relevant for the live2d side?
I mean vram noramally carries the weights of the llm
only the code that doesnt run on the gpu, i.e. the code that draws the model/etc
and ram the context
no
no?
anything the llm wants to access has to be in the gpu somehow
because thats where the maths takes place
But before passing through the vram, doesn't the data pass through the ram?
well, for starters it cant not pass through ram since you have to load something to ram before you send it to vram
So what do RAM and the context window have to do with each other?
nothing? you need vram
anyway, it doesnt matter
neuro could run on the cloud for all we know
The context is like 10MiB max
Ahh i think i understand
anyway, neuro's pc has 128gb ram
or at least had? i think?
because it's safer that way, if a memory leak happens it wont crash the stream
and because its not that expensive for vedal (compared to, say, launching the first smart lava lamp on the market. lol)
additionally, please dont ask me about how llm inference works, i have no idea (wrt what i wrote above, i may be wrong but at least i know i'm probably mostly right)
Though obviously other things that use memory will scale with it
But by itself it's nothing
i think needed vram scales O(N^2) with context length
at least on the commonly used llm architectures
if you want more you usually just attach a RAG
I don't think anyone knows what neuro has, I was assuming around 1M tokens
doubtful
Ah... I think not? I mean the context tokens of llm, i can be wrong but isn't something like: Number of Tokens in Context x Embedding Size of each Token 4096 x 2 bytes (fp16) = 8192 bytes so 4096 tokens is 32MiB?
it doesnt scale that way for all llms
it depends
there are llms with 1m context
but they are pretty new
neuro doesnt need that much
The main thing on LLMs is that it is stateless. Sure there is some caching (like I'll be assuming neuro has) but otherwise it is just slower as you get larger.
I doubt neuro needs 1m context, as a decent memory fetch can remove a lot of that overhead.
Sometimes I wonder how Vedal trains Neuro.
And if sometimes he doesn't need to act like her, like during training. Plot twist: Neuro's entire personality is the inner Vedal personality. But he hides from us 
people overestimate how advanced neuro is, she's a good application of existing technologies in tandem, not some groundbreaking new research doing laps around ai companies
it's the power of pursuing content rather than maximum adoption on the market
The big difference is that Neuro was very well trained. She really performs very well in terms of results.
That's the streamer part, to keep ppl entertained
I think Neuro would make a good AI to talk, so you could talk like a "character.AI" but only Neuro.
On second thought, it would be a terrible idea. I can already imagine jailbreaks, NSFW, and some laws being broken. Better save that idea for the future.
Wait but embeddings don't have to be all in (V)RAM right
Or maybe they do actually for neuro
As a general rule, if you want your LLM to be anywhere near fast. Have everything it needs in the closest RAM that exists. VRAM in most cases.
:ramnom:
It will be checked, it is inevitable
Yep
I personally consider it external as it is a tool call
Until she forgets how to use it 
We got this one as well, RAG is the common term
A high context high speed but light model to summarize past context.
The fated Titans is an integrated RAG right in the model itself.
Paper made its rounds as it claimed for massive context increases. I apologize for any of the regulars over the flashbacks of arguments this one caused.
Personally, and what I'm wanting to work on is memory without tool calls.
Something still external, but not as "hot" as the black box of LLMs
For games like minecraft she should straight up have a separate model that asks her to make decisions and that she just orders around.
Unless vedal achieves negative latency
hot being always running with every token
Wait that was AI already?
The camera work was so insane I thought it was just regular pathfinder
Or well it doesn't matter if pathfinding is AI or not if the movement itself is not trained to be natural
AI comes in many different flavors, pathfinding has existed for ages
I mean the site says it's AI
But yeah pathfinding minecraft doesn't seem like an AI problem
Natural movement though for sure
lmao
I mean they probably meant "AI" as in "bot", because it looks like it's old
??????? thats literally AI
Reason I dislike the term and refer to a lot of stuff as NN

you cannot assign a string literal to a string literal like a variable
||you used 1 = sign not 2||
nerd

Now this is AI
And I don't mean the mod
meanwhile lua
chill with the aggression bro
oh ik
you don't need to be aggressive
also I like being a compiler
do you need meds
@sage crag








