#programming
1 messages · Page 485 of 1
yap that just compiled
but box can be moved, just &mut - no
Rolling Girl break
so to be like std::function it has to be heap alloc Box
mhm
since its heap allocated
std::function can be moved out of the scope 
so only Box one is std function
anyway it can be just
for most cases
i think
use std::{cell::Cell, rc::Rc};
fn main() {
let mut x = i32::MAX;
let y = Rc::new(Cell::new(x));
let yr = y.clone();
let function: Box<dyn Fn(&mut i32) -> i32> =
Box::new(|x| { y.set(4); *x = 1; *x + 1 });
assert!(function(&mut x) == 2);
assert!(yr.get() == 4);
}

all is fine
all is happy
yes...?
since rc can be & then it even can be Fn
but it actually FnMut i think
hard
if it doesnt need to be FnMut dont make it FnMut
AEAAAAAAAAAAAAA CHROMA MASKING IS DIFFICUUUUUULT
sleeping is counterproductive
rust supports downcasting anyway so
creating FnMut but using Fn should be fine

Fn - the button on laptops?
anyway, rust guys,
i not see anything else that type system related directly,
like i would describe rule about declaring something with foreign trait for local type by local trait,
or how iterators trait defined,
but it not directly about type system
this diagram is just entire rust type system?
YAYYYYYYY
the drugs are catching up to me I think
today i found out u can actually ctrl+v in linux shell but u need to hold shift
bruh
i always thought this didnt work
You didn't know?
same for copying
and idk if you know about it but you can usually right click or middle click to paste
wait what
middle click
what
im honestly just gonna fully switch to linux atp
and only keep windows for compatibility apps
most of my work is in wsl anyways
how did you just knew about this
man im a windows user mostly
i just run shit on wsl
Wsl is a gateway drug so you prepare yourself

Middle mouse paste is an interesting silly thing, it pastes whatever was the last text you had selected
i mean actual ml development runs in ubuntu
i alr used it on rented stuff and my raspberry pi and dgx and stuff
Yep
and i had a partition from years ago
Mhm
why rpi for ml 
also
must be tensorflow user
#baking reading comprehension
wym no pytorch
i mean i do know tensorflow i just dont use it
i am also a victim of zero reading comprehension
Jax only

it's good for web
and if u use tpus
but yeah honestly just use jax
i think TF is mostly for colab people
how big is the model
280B
active params?
I assume you mean flash if it's pro that's crazy
13b
ye flash
bro i aint got 8 spark setup
It was only optimized for ascend and shit I think
They did a funny where nobody got access to it to do anything preparation in the west
Cn labs have been doing that
should be getting ~30t/s and then deepseek said they reduced required flops by like 2x per token
bandwidth constrained still
qwen with 10b active gets 45t/s
iirc dgx spark are like
how big of a total model size?
122b
256GB/s
That's crazy it's not slower if it's that low
iirc there's a formula for bw requirement that also includes both active and full params
it's 250 tflops fp8
forgot what it is
need to wait for dflash
qwen 3.5 35b-a3b goes from 50t/s --> 90-120t/s with dflash
My ai pro is 383 tflop fp8 matrix and it's slow as balls I'm just used to that
With 620gb/s
ai pro what
what device
Radeon r9700 ai pro
susge
Damn dgx gpu is actually fairly weaker than I thought for some reason
The memory blinded me ig
also no one has run sparse 2:4 models yet on spark
which is what im gonna try
sparse 2:4 fp4
use actual 1000 tflops
5070 lol
I get , 1534 tflop on r9700 lol
actually makes sense tbh
Sparse fp4
for the power draw, but people are getting 80w ish
i mean yea it's a full desktop device
idle wattage lmao
My 3090 in inference uses full 370W I'm pretty sure
how do i check whole system
you don't. get a clamp meter

i aint got that lmao
and measure the AC or if it has a dc connector, do that
that's the best way
or
just get those AC power measurement
that you plug on the wall
that one uses shunt pretty sure
Surprisingly I'm only able to get to the 350 range
Kilawatt
beat my numbers
My 3090 has a slightly higher PL
if you wanna do software wattage reading
It comes with a factory preset PL of 370W instead of 350W
weird, the model default tdp is 350W

I definitely have not reconfigured the PL
-# Because the configuration daemon makes my PC stutter
it's just Gabox's fv7beta2 with Anvuew's dereverb
near the bottom here https://huggingface.co/GaboxR67/MelBandRoformers/tree/main/melbandroformers/experimental and yaml https://huggingface.co/GaboxR67/MelBandRoformers/blob/main/melbandroformers/vocals/voc_gabox.yaml
and here https://huggingface.co/anvuew/dereverb_bs_roformer/tree/main
I think they tried fv7beta2 and said it was a little noisy for them
it is on the fuller side.. not super full but can have some noise at times

You know, I wouldn't use it for watching tv or movies
But RTX vidoe enhancement is great at compensating for the shitty bitrate on 1080p YouTube streams
Do they even support Linux with that RTX video stuff?
Nah sadly
Guess no RTX video stuff for me then
The
The
yeah, it's decent for what it is
but low bitrate really hurts what it can do.. works a lot better with high quality content
below 1080p also is sharpened a little too much.. but not strong sharpening at 1080p and low sharpening at 1440p
for animation, you're probably better off with something like anime4k.. models specifically trained for it, you know
i tried "agentic engeenering" for the first time
just vibing a project without looking at all at what the AI is doing
it seemed to go quite well with lots of hiccups, but eh
and then i peaked behind the curtain and what the flip, how did the AI delete /dev/null, i don't even want to know
What agentic IDE do you use?
Or Ai
i used cto.new
with Gemini 3.1 Pro
this shit seems to be entirely free
or their free credits are big
also it's entirely remote and not connected to anything so i'm not worried about security and stuff
I think I gonna use gemini as autocomplete instead of agentic coding, because it can make mistake without good archit
honestly i don't think the agents idea is the right one, but it's interesting
i think AI is just good at being a better autocomple than just an LSP
LSP is fast, but dumb, AI is slow, but knows more
Goodmorning 
hi
No 
ahahahah
i literally
deleted dev null
tried to screen it
and it spectacle cant start

Was it a good idea?
you know, just everything that try to use /dev/null will fail
its evening btw
it's night btw
The morning:
marble hornets lost episode
no it's when i wake up
am i going schizo or youtube removed clips feature?
why i cant find it anywhere in ui
is it because i deleted dev/null? 
vibe mathing... (they had to clean up the output, but got there nonetheless)
damn, that was too powerful
(solved an erdos problem btw)
vanor staring right at me too
Experts believe it may have further uses
first use for LLM found
the very first one
didn't they train a gemini model specifically for this a couple months ago
am i tripping
LLMs can actually have a use?
that one is probably still looping trying to load mathjax or something in the terminal
since it's gemini
Good thing it's 40 minutes left then 
DEepmind does produce useful things
my brother is a meteorologist for NOAA and they use weathernext 2
as one of their tools
Man, I sure hate LLM being called AI in marketing everywhere 
It drowns out the actual useful AI
oh yeah, deepmind produces pretty cool stuff. i think of all the companies around, google is the only one making sustainable ai
they just have bothered to make non-llm shit as well
google is also so far the only company that i assume will actually still make AI stuff 10 years from now
I mean, they are the one that came up with transformer
the rest are VC backed and money sinks
they were doing it before transformer and will be doing it after, if this all implodes
Nvidia ceo can speak thai
I stand with nvidia ceo
does google even use nvidia stuff? i thought they use their own hardware
they have some nvidia stuff for cloud basically but they are not beholden to it for ai
Pretty sure they do for training stuff and colab
And also cloud offering
mainly colab and gcloud

gpg in 2026 
(gnupg getting post quantum crypto)
me w/ rsa keys
can even deploy Gemini on-prem with NVIDIA GPUs apparently
that guy has so much money and power right now i would be unsurprised if he came around handing out full nvl72 rackservers to passers by
Again, google just published their finding about new more efficient quantum computer that would require far less qubit to crack rsa
Isn't that the one with killswitch?
i just saw a provider offering this
thing is like a SCIF
if power goes out the whole model is gone and you need to ship it to google to reload it onto the gpus
like for real 
oceans 14 will be heisting gemini 3.1
thats their tactic
they advertise an air-gapped deployment as being possible
if you can afford this you will have backup power anyway
so ye, keeping the model in VRAM only would make sense
The thing is, people that has air gapped and need gemini would actually buy it
me carefully weighing the server to deduce the number of micron bga packages of hbm in there
yeah its meant for governments and hospitals etc. so.. its suppose to be secure
if it were loading to a disk or something someone might inject something into the model or something.. but i guess you could do the same for a loaded model 
I mean, nothing is stopping anyone from reading from the gpu memory either if they REALLY want those weight
confidential computing 
the literal locked fortress of a container it's in
If state actor can replicate hdmi signal just from the EMI emission alone, that fortress better be a faraday cage
unironically im sure it is
I'm not sure if they encrypt anything or how exactly it's deployed, but ye, it's going to be secure
you won't have any access to those GPUs at all
its gonna be "secure"
i'm sure some people that will have it will literally just have a unlocked frontdoor straight to the server room 
colos are already in locked cages in regular dc
at the very least the connection between the cpu and gpu is encrypted (iirc memory pages too optionally) + if it notices you trying to anything weird it just insta shuts down
Yeah, data center security is no joke
if nothing else they have access control just to know who to bill the shit out of if something gets fucked with
ye, even if you can physically access the hardware (already very difficult for anyone not contracted to do so), good luck actually getting the model out of the server without disturbing anything 
That'll be $100k fine, thank you very much
"yeah todd here farted on the gemini server so we've gone bankrupt. thanks todd"
for some reasons
so many things breaks when you delete dev/null
i prob should restart system
mknod or udevadm
Ok, i installed artix dinit tty only.. im kinda wilded out that that uses 300mb of ram just doing nothing..
Now im heavly considering alpine + xlibre repos + zfs.. buut muslc.. ig i do have zo go the Route of lxc.. i wonder how Nvidia-open deals with it
dentge
Nvidia open does not deal with alpine.
Welp there goes that.
Opensuse zfs? Openmandriva? Why must linux suck nowadays..
Well nvm apparantly they dont Support that either
Ubuntu Out of all things support it, but no thanks
konii not toast
that is why im no longer toast

welp, i guess the real toast will be free now

Intel datacenter GPUs 
real ones stay Toast
gaudi is closer to TPUs right
chat no way i managed to create a robotics llm dataset generator it's running now
yes

looks like it
no graphics stuff to be seen here
except the media engine for some reason
its been fun, but you gotta know when to quit
was it all just a larp?
terrible idea is forming in my head
how bad of an idea is it to use a usb card reader as quick auth
its prooobably fine
isn't that quite common in some industries?
probably, but i feel like those industries do it properly rather than hacking together random oss
This is why you don't vibe code a PR
when some vibe coders come in and break working projects
ai should be banned
allowing ai in the kernel is a joke
you say that like you can ban it

the only thing you can do is tell people to disclose it, they'll use it either way
It's a permanent fixture now
For better or worse
Issue is some people are trying to use a power drill as an angle grinder and are shocked when shit breaks
QA has become mandatory and the vast majority of ai vibecoders don't know what quality looks like
They just assume the tool can do it flawlessly - and worst of all, in a large number of cases I bet it does... so when it goes wrong, it super goes wrong, making the tool look unreliable when it's operator incompetence
At least, such is my hot take
Warmup
monitor consumed by the void 
i can tell you from my week of trying out vibecoding... it 100% can't create quality code
atleast not the majority of the time and without guidelines.
i've started to write novels to describe the refactor and issues i want it to fix just so it doesn't try to get creative and ruin my codebase.
and the fully vibecoded project i tried in c++, works... but well. it looks vibecoded to heck
and i'm too lazy to go through 10000+ lines of code to fix it
so far the best usecase i've got are refactors and spelling/grammar issues. maybe the odd feature that is described in detail. but generally the smaller the scope of the usecase the better it actually works
Technically first ever fully complete NeuroSynth version 1 song
"Narrower use case works better"
Wild, the thing that excels at specialized tasks does better when the task is specialized
i know right
I can't imagine why that is 
yet people hail it as it can replace senior devs... i have yet to see it actually do a better job
AGI is a lie
it can create things that work... but is it maintainable and can grow? no
except for neuro
?
also...
screeps update. my automatic building logic works. and my first extension is finished
Our oshi is built stupid different
I've been focusing very hard on NeuroSynth stuff so I haven't gotten to opening it yet
What's neurosynth? Aeris go open source?
The NeuroSynth in question
Ah assumed was a piece of software
NeuroSynth is a vocal synthesizer
The power of vocal synthesis or the power of Neuro vocal synthesis?
Let's start with just the vocal
I've had a design in mind for a whole month and wanna fiddle with it
Also
Well cooked, sir
You mean vocal synthesis from scratch
If you're going for that, I suggest using Japanese, and starting with just a simple consonant-vowel pair system (CV as the UTAU folks call it)
Or if you just want a custom vocal synthesizer, there's UTAU

I wanna tilt it toward spanish, actually
If such is possible
SynthV does Spanish, but it does seem like it would be a phonetically much more complex language so I don't recommend starting with that
Oh no like
It'll speak english
I just want the accent
Kinda in the way you can sorta hear neuro go british in her otherwise monotone-ish voice
Accent, then you need to either do phonetic trickery or have your base voice have a Spanish accent
Still recommend starting off with Japanese if doing fully from scratch though
Lotta cooking to do, then
i really wanna start playing arena ontop of just world, but there are so many gamemodes idk which ones are the best. And also setting up a monorepo for it sounds like pain because different seasons and gamemodes have different type definitions 
but also you should definitely start playing
you still have 32 days in the novice area left. after that it appears somewhere else and you might not be able to join us
I have no idea what any of that means but
I guess
novice area just means noone outside can attack you. basically. after that time is done. any player can come into your room and kill you 
In the game, right? 
there are always new novice areas... so if you can't you can't 
the enemies hack into your pc and take it over when you loose your room 
Yo irl netrunners before gta6
true
i tried using utau once
it didnt work
i dont know if it was utau or the voicebank
but it couldnt make any noise aexcept "Aa" and "ii"
Classic UTAU or OpenUTAU?
I tried using Microsoft Sam is that close enough
utau synth
openutau
a teto one
Teto JA integration for OpenUTAU?
idk
That's the one you should use for OU Japanese Teto synthesis
Looks a whole lot like the one I mentioned
Probably SHIFT_JS
boy i need 64 cores and 144gb vram on desktop linux
😭
who knew robotics needed that much vram
and cpu compute
Day 9 of making my own game engine
Hardware issue with vga not working
Byeeeeeeee :3
what are you even doing that uses that much vram for robotics?
don't modern robotics use ML models?

look
looking

basically i'm procedurally generating environments and mining for expert paths from random spawn to finding a specific object
the llm generates controls
each example is ~50 turns
and the environments are huge
wild that it needs that much vram :o
each instance uses ~1.5gb vram
i need to generate ~20000 examples
D:
i already have in theory 256gb vram locally
O:
but i cant use it cause arm64
that's disappointing
this is only step 1
the bad part is then training the actual models
huh, how come
the environment was built for x86
source code not available
dang
emulate it 
this is what im thinking
iirc docker has support for emulation
on linux
would that be decently fast though?
good question
assuming the cpu isn't involved too much
should be fine in theory since the gpu has its own isa anyway
it's a whole Docker container
probably not horrible then
the cpu is at full load
right now
oof

cpu generates monte carlo best path
given specific constraints
since if i just use a* or similar it will give weird paths
wait what
the source code IS AVAILABLE????
this might actually be it
wtf
wait can u cross compile for arm64 in unity
Today I learned: Python executable isn't compiled with PIE
what is this paywall slop
(admittedly that might just be my build :D)
Nvm checked on Ubuntu system Python now
There was a pwn chal I was too dumb to figure out until the submission time ran out already :D
It turned out it was, uh, real easy


Simplified pseudocode (in Python cuz I am lazy)
def vulnerable_func(s):
split = s.split(" ")
n = []
if len(split) > 1:
for i in split[1:]:
a = int(i, 0)
n.append(a)
return Py_BuildValue(split[0], split[1:]) # pass numbers to be interpreted as Python objects
return Py_BuildValue(split) # Picks up regs/stack crap/whatever else depending on length
Slopping has dulled me
At least I learned something
frick it
i'm just gonna tell codex to rewrite the entire engine in arm64 native game engine
Based and true
how did that happen....
his final job is to get recycled, so i guess i dont mind the scenic route
sam I think you lost a triangle in the toast arc somewhere
nah we stopped with letterswap
is his path cached? like he doesn't recalculate it? because then maybe all other paths were blocked by other creeps so he just took this one and saved it
its just this
const spawn = creep.pos.findClosestByPath(FIND_MY_STRUCTURES, {filter: s => s.structureType === STRUCTURE_SPAWN});
if(spawn.recycleCreep(creep) == ERR_NOT_IN_RANGE)
{
creep.moveTo(spawn, {visualizePathStyle: {stroke: '#ffffff'}});
}
no caching
wtf
i don't think you really hit the cpu limit until you reach like rcl 5 or something and have 3 rooms 
or have a very very complicated automated build script
i only ever ran out of cpu because of the wasm interop of c# 
well you see. im not caching cuz these litle guys really really really like blocking eachother's paths
or initial loading of the wasm file
unrelated but here is my OS, it makes the screen blue and it is 536 bytes large 
have a navigation manager that has traffic controls and police creeps that resolve issues
this seems very useful
you overrun the vga buffer by quite a bit there
i want to make a 2 lane system of one-way roads, so they never try to go into each other and block themselves
probably sufficient is the only truth

also make sure to run cli before hlt
i mean creeps can swap places so technically if you coordinate it correctly using the navigation manager there shouldn't be any blocking going on
interrupts in my halt is inconceivable
maybe, but screeps are a bit dumb
ill see what i can do
i will get there when i get there 
i would need to place roads in a way where nothing parks in the middle of a road
and run this after your clear
mov ah, 0x02
xor bx, bx
xor dx, dx
int 0x10
definitely not me
my brain is soft and squishy, like sponge
it's so much more than that
against?

vampires
dont try using a hoodie against a vampire
it doesnt work as well as you might think
hoodies are defene against cold. not vampires
also agaisnt rain if the wind is not too bad
hoodies are a gamble for the vampire 
remember how i bought clothes a weekago
apparently one of the hoodies they have me is a size too big
(real)
im not a vampire but if i were i would definitely say that hoodies are not useful
its men clothes 
more common than you may think!
what contradiction?
like i really wonder how ppl getting interest in touching stuff like this, isnt its all is arch specifics and just specification
why bother, i only learned how this things supposed to work and its all
wdym?
btw, was debugging why my tun wont open just to realize i made pacman -Suy and i have to restart os, not just DE 
reinventing wheels and making them worse has been a programmer hobby since the advent of programming
but all the sudden when it isn't a CPU it is fine 
The past month I reinvented lodash's omit function :/
not for me i guess 
i prefer something practical
thats why i have almost no code to show that i know programming 
when the project was already using lodash
if you want to make soemthign that hasnt been made before you might aswell close up shop
not if i have anything to say about it
random program generator
go
fuzzing x86_64 1 instruction at a time
As someone who's pet project is in this catagory, by the time you make it someone else is already going to beat you to it or it is useless 
with games you only get stuff that hasnt been made before if the concept of the game is very obscure, or if the budget is through the roof
nobody wants another minecraft clone, but in the category of voxel games there have been interesting variations
Spend all of your think thinking until you deflate into
and then vibecode the game 
that was fast
make game about rewriting in rust
The thing I'd want to try eventually is actually realize the universal item concept that the blockchain enjoys were hyping about.
But instead of items, your throwing yourself from server to server
your 🥈 emote reaction was 🥈 because i already posted 🥈 before you reacted
you mean the thing that would literally never work?
because seeing the fruits of your labour is fun :D
no, text was first
fals
tru
yep, sure love attempting to make something that is physically impossible unless you just build the entire game engine around it and even so it'll never work because of a thing called balencing
i mean thats just roblox 


don't let the bed creeps bite
its almost 27th april
I'd be more afraid of konii randomly appearing, I'll be dead by morning
the problem with konii appearing is that we dont know what she looks like
she could be anyone
anywhere
if it is Sam then all the sudden some new photos of someone in a dress suddenly appears
i cant find it, gimme a sec



I think best use for ai is autocomplete while you're manaully coding.
Take vedal coding in python as an example, he did have autocomplete turn on while coding manaully and finish llm in not even an hour
The Twitch streamer Vedal has created an AI called Neuro-Sama and at the end of the subathon they built a language model from scratch while following this tutorial from Andrej Karpathy: https://www.youtube.com/watch?v=kCc8FmEb1nY
To watch Neuro&Vedal live: https://www.twitch.tv/vedal987
And also got it working + he knows what he's doing
Unlike telling ai to do it, after it finished
You're gonna read/review something you don't expected ai to write or you don't know why ai write code like this, and what the purpose
(Unless you're some god reader, you can use ai to vibe code it's fine)
within limits
And then random ai bro in the comment: "Ai replacing all developers, SWE is cooked(emoji)"
Ok i need os suggestions.. anything but windows is ok. I need zfs and xlibre, i need support for my rtx 2060, id like JWM or MATE but others are fine aslong as it aint gnome, and id like to sit on desktop around 100-300mb ram used, lower = better.. oh and no snap/flatpak..
I normally run ArchLinux, but i wanna experiment with something else.. any suggestions?
I was thinking freebsd but idk
Maybe Haiku
Just use anything you like ig
Debian also comes into question
Ive been looking around and didn't find anything i liked.. thats why my question
I like mx linux cus I want performance and stable + old looking desktop, logo looks cool
It's based on pain point
well i've been using AI to autocomplete since the very first time github copilot got into closed beta. i'm a OG AI user
back in my days there was no agents, or claude
Mx does look interesting, do they support zfs rootfs?
I still have not installed it, cus my pc is dead and on life support and I don't have any money
Window 10/11 is so bad that my pc just died
I unfortunately have to daily Win 10 because of Win95 Development
Luckly Win 10 Enterprise IoT exists
Coding in smasung phone using Acode apline linux is pain lmao
Touch kb coding, yikes
I have a Nokia Communicator E90 which i use ssh on..
Physical keyboard symbian goodness
Virutal keyboard cover 20% of my screen
Only 20%? For me its like 33%.. do you have it set super small?
I couldn't deal with that ngl
I'm an unpaid homless open source maintainer lmao
0$ per month
20$ per week from other job
Yikes
Small, it is good for me as I've been using something this small for awhile.
OOF YEA NO
i feel like the value of that is different in the thai economy
Just my pinky would hit 3 buttons lmao
Oh actually I get free money for being a good boy in my country
what does "being a good boy in your country" entail exactly?
I can see it, I don't really mistype anymore after getting used to it, but slim figure ig 
Probably. I get super low pay due to learning a job ("Ausbildung"/trainee) like pure Minimum and im sitting at 600€ a month
Fighting the autocorrect, still not used to it existing
Some old man will pay me if I working or doing something for them
Like selling things in street
oh, so its just a normal job, you said it in a way that made it sound like a government thing
I mean, im by no means not underweight, i just dont have the hand coordination to Hit that while Walking or not having physical keys-
Ahh, yeah that is a skill 100%
A lot of ppl are very kind, you can make money for free if you have connections with ppl on streets
These days it is second nature though I prefer not to
Yeah absolutely. Funnly enough, im fairly ok at soldering small SMD stuff with horrible tools lol
"Some old man will pay me if I working or doing something for them"
that is called a job. its like how people mow the lawn or babysit here
Jumpscare word- crazy.

goodnight konii
oh no im not sleeping

It's 4 AM for me
everytime soneone says their time everyone else immediately says their own time too 
herd mentality
its 55:29 here
Swarm for a reason
I dislike the arrow keys being constantly there, so I'm removing those
Even smaller now
-1st row of keys
there are arrow keys, jsut didnt see them
As for posting kb vs text ratios, i present, 1:3
on my smasnug, the L/R arrow keys are baked into the spacebar by dragging the spacebar
How do I sleep, feels high-enegry
Gboard moment
Yes, stock pixel6a lol
I don't mind it, wouldn't know how to make it prettier
My material you is.. questionable
I forget how large something close to stock is
another day, another instance of clangd that acts up and uses all system resources 
you're keyboard seems to be corrupt, its all squiggles 
1:3 yippie
That's closer to half 
it is pretty much half ye
Id still fatfinger so much lmao
hi

True. 2:5
Return to sane size for me
@sage crag how has ableos been going? havent heard of it for ages
I'm sleepy that my tear dropping

exit 0
Sometimes i wonder how an ai can be so adhd/autistic.. then i remember the internet
john able cancelled 
I love it
Sam is getting executed 
Try exit() or return 0;
Gn
not cancelled 
exit()
it was bound to happen, surprised i lasted this long
exit(255)
keep the opensauce money
-> segfault
"meh, close enough, ship it"


so whats up with ableos?
not much going on recently
Surely I'll be fine now that I slightly changed the keyboard size
You did worse-
You moved it down slightly
Your entire hand positioning is now off
Just barely.
i still need to finish setting up my new phone. i think its mostly there but you always notice some stuff only after weeks of usage
I should backup all my data on my phone and flash graphene os
And then not copy any of my old data back-
Makes sense to me
The current install is very much a "just gotta last until i flash graphene".. for the last 3-4 years...
Its running with 3gb free xD
About 1.2TB useage of my 8tb hdd are just duplicates of my backed up data
I have to at some point really really sort my data..
More HDD space than I have
I don't have much to back up 
I need a :neuroDread:
Or something

True, despair works
In total my main machine has 9.75TB of storage...
My main laptop has 3TB (1.5Tb formatted)
Largest of any system I have is 2TB NVME
Then i have like 35 other laptops and desktop with varying disks
And uh idk a few servers with 2-16tb each
of course
.. although the 16tb server is running on dead disks
Im essentially feeding zfs to the chances
Its my apt, steam, pacman, and so on mirror/cache :D
I have too much hardware.. ^^"
I'd have 2, 2, 1.5 then whatever other storages I can find
Id like to save all the hw and disks from work that were throwing away but like i cant
At some point ill own a 200 disk jbod
And ill need a 2000 disk jbod
Absolutely Reasonable.
I just couldn't say no to a 120€ 8tb hdd, and well, the shitton of hardware we are getting at my job that we gotta dispose off
(i have a chair dedicated for laptops.. and a room.. and a drawer type thing
The chair is stacking 8 laptop atm...
Makes sense, I don't nearly have that much. But I also don't own a desktop
Oh dont get me started on desktops.. please..
I have no desktop no server easy 
brotherrrrrr, that's $1500
I essentially have a laptop and desktop for every version of Windows ;-;
You could buy a framework for that
the pro even
Probably with the 2tb ssd lol
Surely this isn't the drive I'm thinking of
But to be fair its an Enterprise one.. so you have one of two chances:
Dead on arrival, or dead in 4 weeks.
Nope it is not that one
i almost went into shock when i saw that on my first search
You dont wanna know how many hdds we've received that were either DOA or started throwing errors 2 days in..
Only enterprise tho. Consumer hdds were fine.
Gigabyte AI TOP 100E 1 TB 
So very useful name
some people might actually get some use out of it
but I wouldn't be paying for that cost
Nope, this boy has a trick
it is called "just a bit" of write endurence
Endurance: 109500 TBW
So.. a decent enterprise ssd, not muuuch more?
Samsung 980 Pro (1TB):
Endurance: 600 TBW
It's use case is mainly for storing model checkpoints if I had to guess
as good luck managing to wear it out in any other usage
It has so much write that you could use it as swap without reservation
if only intel wasn't incompetent we would have optane still
I do wish
I mean.. you should see the Synology nas a customer of ours has.. thing eats high Enterprise tbw ssds like its cereal
Most of the high end ones are 8000+ but not the 100K this is at
It burns through one within two months essentially
Theyre like 30000tbw or so iirc
Makes sense
The Samsung enterprise with power outage protection
Or whatever
500gb is like 300€ or something
Fair cost considering what you get
Nah for the TBW, if you tried to do consumer which is 1/30th most of the time then there would be a lot more downtime
Well sure, the tbw yes, but most of the price comes from the POP when we bought em.. whichs stupid
And then we use em as root disk + ZFS SLOG lmao
In a mirror ofc.
Held down by ducttape and the cooler..
22110 nvme in a 2280 port :D
as all gen 5 drives surely are, not that this one is gen 5
I think it's gen3.. or 4? Idk. But yes, the thing is very much unevenly cooled
Its lucky the cooler was able to get one screw in-
The other side i had to disassemble the screw-hold-in-placer
To avoid short circuits
Well.. what do you wanna do, had to get the mirror working and we only had one 22110 port..
Enterprise professional grade IT :)
better not know how the sausage is made
Honestly it just gives me respect for the effort people do to get stuff working
the entire internet is filled with botches, it's just a matter of accepting it
My workplace pc's ssd is help in place with packing tape because there wasn't a standoff screw to be found.. -1 pcie x1 slot lol
True, but that filles me with rage
I try my very best to offer and implement the best solution.. cuz i don't wanna hate myself for making subpar ones
Lol
I tend to pick what makes sense at the moment, I'll botch to get something working or I'll spend time to get a solution working
Fair, and i mean makes sense
Quick and dirty hack, until you can get it fixed propperly
Welp.. was a nice chat, enjoyed it- i still don't know what Linux i should try, but oh well. Imma sleep.
what do you need compute for?
Hi, I'm compute
synthetic robotics ML teacher trajectories generation
does it run on cpu? or do you need gpu?
How complex is your robot to the point that it struggles to generate forward kinematic
i need everything
> compute starved
> has 2 3090s
Isaac sim?
no custom procedural simulator
man if this was compatible with arm64
i'd have access to 256gb extra vram
why don't you spin up an aws instance for a couple hours for extra compute?
which bit isn't compatible
the simulator is written in unity

cant compile unity to arm64 without fuckass license
thats what they wrote the simulator in
i didnt make it myself
buying a license might be cheaper than buying hw to provide that compute
im mildly curious if https://box86.org/ or similar would work at all
it's estimating between 14-30 hours till completion
looks like it needs the enterprise license
oh...
so a "call to ask for a quote" type of price
basically i generate these chat converstaions
with image+depth image -> tool call
each trajectory is ~50 turns long
unity for arm has a 30 day trial
you could maybe use it just to run this computation and then ditch it afterwards
arm linux?
yes
can't you just make new accounts to get that 30 day trial over and over
wtf where
is this for that vrchat thing i tried to help with ages ago before promptly forgetting
i just need 1 compile
tied to a card it seems
agh
yes
😭
https://unity.com/products/unity-industry is what is needed for the premium runtimes
sorry for disappearing with that
and it has a trial
the free trial might be it
i finally got a good approach working i just need to generate this
according to some papers i need mroe like 100k episodes
im prob gonna generate this over a week
did you end up using that fuckass shader?


but yea im just gonna compile the thing and then ditch the subscription lmfao

lilac vision shader for vrchat

holy shit this is gonna generate so many tokens
pain
each example is ~16448 tokens
this is a while ago
i was still doing python back then

i have 328 million tokens to train 1 epoch
damn this is gonna take forever
1 dgx spark is ~8k tokens/second
so in total 16k t/s
it believes me but the first thing it does is check if i lied
wait at 16k t/s it's just 6 hours
classic
as it should 
why must half the label tagging companies either be super expensive or "request a quote" pricing
Holy benchmark
It looks alright, i still wouldn't call it practical tho
Well it was meant as a PoC anyway
ye
I propose a benchmark if LLM is truly capable of thinking. Writing KDE wayland in Holy C
i mean we do require more complex/ unique benchmarks to actually measure the intelligence of models. every current benchmark that is used is either already benchmaxxed by the models or the tests itself are broken and don't actually work correctly
We convert the kde test suite to holy c, "Do a test driven development" 
We need more software written in Holy C
We make the LLM useless by benchmaxxing them to holy c 
https://caseys-evals.com/esobench close enough 
LLMs never gonna pass the notepad++ test
how does one remove a node from a binary search tree? 🙏
tree.remove(node)
you ask it nicely to leave
how does ts work tho
use an axe

the tree holds the nodes, when you suply the remove function with the node you want to remove, it removes the node


have to call the lumberjack function
an iq too high?
Nice. It goes to shows that the LLM is not smart enough yet
its my favorite shade of purple


@high brook




todd


