#programming
1 messages · Page 403 of 1
14 
i thought that was per channel
no 
I reduced down into bitmask first and readded
idk wha tthat means but sure 
some combinations of resolution and refresh rate use different encoding to RGBA
hmmm
The lorge number is 570*1140*160/24 should be 32 but not useful for this.
the voluemtric refreshrate is supposed to be 30, so with 1080p30 we can ignore the refreshrate and jsut try to pack 570x1140x160 volumetric into a 1080p image somehow
i saw something using a 4:2:0 encoding i think in the config blob yesterday
i havent had good luck with 4:2:0 so far. but maybe thats cuz i was trying to do stuff that didnt work in general
oh boy youtube is somewhat down.
new idea.
instead of matching the binary representation of the voxels to 1080p image. we only send non-empty voxels?
it's a little bit completely not working. yeah
I was going to mention this
at least it's failing quickly
had a window still open, no issues with watching a already shown vid.
keep the spiral mask in memory and send anything in mask over.
that would sounds complicated.
complicated is fine if it works
then we're back at needing 180hz
which we cant
the spiral is only for converting one frame of volumetric data into a full rotation of the spiral
sending the volumetric data is compeltely independant of the spiral
the spiral is probably trivially computable from what the current time is if you need to
spiral itself can be cached
spiral is jsut this:
#version 450 core
const float PI = 3.14159265358979323846;
const float realRes = 570.5;
const uint WIDTH = 570;
const uint HEIGHT = 1140;
const uint DEPTH = 160;
const uint STRIDE = WIDTH * HEIGHT;
const uint WORDS_PER_Z = (DEPTH + 31) / 32;
const uint NUM_BITS = 24u;
layout(std430, binding = 0) readonly buffer VolumeBuffer {uint data[];} volume;
uniform float helixAngle;
uniform float deltaAngle;
out vec4 colour;
uint sampleZ(vec2 pos, float angle)
{
float a = atan(pos.y, pos.x) + angle;
float dist = fract(a / PI) * float(DEPTH);
return min(uint(dist), DEPTH - 1u);
}
void main()
{
uint x = uint(gl_FragCoord.x);
uint y = uint(gl_FragCoord.y);
uint baseIndex = x + y * WIDTH;
float offset = 0.5 * float(y & 1u);
vec2 newPos = vec2(float(x) + offset, floor(float(y) * 0.5) + offset);
vec2 centre = vec2(realRes * 0.5);
newPos = newPos - centre;
uint packedVal = 0u;
for (uint i = 0u; i < NUM_BITS; ++i)
{
float angle = helixAngle + float(i) * deltaAngle;
uint z = sampleZ(newPos, angle);
uint bitIdx = z & 31u;
uint wz = z >> 5;
uint wordIdx = baseIndex + (wz * STRIDE);
uint b = (volume.data[wordIdx] >> bitIdx) & 1u;
packedVal |= b << i;
}
uvec3 rgb;
rgb.r = packedVal & 0xFFu;
rgb.g = (packedVal >> 8) & 0xFFu;
rgb.b = (packedVal >> 16) & 0xFFu;
colour = vec4(vec3(rgb) / 255.0, 1.0);
}
VolumeBuffer is the shit we need to send
what would playing normal video on it look like?
well
you cant
it would only look good form the top i guess
complete garbage fromt he sides
it'd be interesting i think. not good, but strange to look at.
I'm trying to say something but also getting ready to leave. I'm not done with this point
i have no clue what you mean with caching the spiral
if you want to send only what the spiral hits, we still have to send everything
you only have to send the data which is currently displaying, not the whole volume
I'll explain but I need time
all of it is
most of it is empty space and can't be projected on
you only need to send a spiral of data per frame
Im playing RDR2 right now
if we send only the spiral we need 180hz instead of 30hz
where are you projecting the other data?
vertically
it can't hit the right place because it can only hit the spiral
we're sending 30hz volumetric, that gets converted to 176.041666z of 24 1-bit planes. aka 4225hz.
at 15rpm with a double helix, all the voxels will have been shown
30hz is the refreshrate of the full volume
you only need to send the data for where the spiral will be at the time the image is displayed. (you're stacking frames in a single value but the only get displayed one at a time on the position of the spiral at that point.)
if we only send data for where the spiral is rn
at that point
yes
then the pc would have to be the one to send the bitplanes
so we would need 176.0416666hz
the 30hz frames don't come in-between parts of rotation.
each frame the spiral will have already made a full rotation
I figured out a way around it! I unhid the second part of the form, filled it out, and it let me submit it
i think this is the point where a flowchart might help. and the point where the bed calls
true
you can overlay and delay the bitplanes, compressing the values into however many bits per pixel you have. but you only need a single sheet of pixels that wrap around the spiral per frame. not the whole voxel buffer
We cant fit 160 bitplanes in a single image tho. They're already packed to 24bitplanes at 176hz.
At 30hz we would need 140.8333 bits per pixel
Or something like that
I guess 141? 
4225/30 doesnt convert well
The only compression that might do anything is just again send nonzero. Just sending the omega sized 1d coords/color array to the pi and force it to remap.
you only need the pixels which hit.
no other pixels need sending and in fact, they're a waste of space!
bonus because we only have so many bit combos, we can make it a lot faster by doing passes and map all of these guys to a color
Instead of coords, you could maybe just send an int for distance to the nixt voxel
run length encoding would be fantastic for 1-bit data
And then just use some funky modulo math
Still need the color infomation 
The bitplane colours for the projector get calculated on the pi
Or you mean for the voxel encoding?
I guess we just have 3 bytes per pixel to work with. Splitting data between pixels gets messy
Nah we got enough for full, because we ain't sending most of the voxels
So we could do 3 bytes, each byte for distance? 
"On for x, of for y, on for z"
3 bytes is for every voxel sent
perhaps encoding marching cube faces?
That limits us to only 1 pixel per row tho
Or like
you could send and on and and off
2 pixels, gives you a block of arbitrary length
if you have more then send 2 more
how many to skip, how many to draw
Since our ground plane of 570x1140 can fit neay twice in 1080p, wed have about 2 voxels per row max if we try a full cube
I love compiling flash attention+cuda
I think that is good ye. But instead of per 2 pixels, id just do it in a pixel
We have 3 colour channels
8 bytes is more than enough in terms of distance
then you've also got color if you want in 3rd byte for lookup table
What is the lookup table for?
storing color? assuming you have 3 1-bit channels you could use for 1-bit color stuff?
or just store 3 values in 2 pixels
Now that i think about it
Worst case scenario
It is 570x1140x160 = 103968000
If we have a voxel at the very end it is going to be 103967999 away
Wont fit in 8bits
We also cant really do 27 bits
perhaps you could do a simpler encoding "top bit set, draw, top bit unset, skip" and then just do lots of 0xff skips of 127 voxels until you get the right distance in
I can't view the the site linked to in the verification email because Google 

Certificates 
The following packages were automatically installed and are no longer required:
caca-utils chafa dkms jp2a libchafa0t64 ocl-icd-libopencl1:i386 toilet toilet-fonts
why did my ubuntu come with toilet
obv for the caca-utils.
neofetch, yes i know that on top of my head
neofetch needs a toilet?

it needs caca-utils which needs toilet
Check neowofetch's package on debian sid package list if you want
i wonder how badly being interrupted constantly will affect my agent's ability to think.
I think perhaps making most events only visible to them when they've reached the end of a sentence instead of interrupting them whenever might help...
perhaps i could put non-urgent events on a timer that either waits a certain number of seconds or until the agent finishes generating a sentence or until the queue has N items in it (whichever comes first)
(spinning in my chair thinking about that and i just about fell off because i touched something pointy on the back of the chair and it made me move off center while spinning. ;/ should not do that.)
I was just joking dawg
nuru
I don't know what is worse, neurospin never animating or actually having it spin
if i had another monitor i could make a spinning chair simulator...
or better... a display i could read while spinning in my chair
VR headset
cable would get tangled.
I was just spinning in VRC not 3 days ago with a quest 2
i only have an old oculus rift (both no longer supported and would require a meta account to use)
I'm impressed the iGPU survived that experince
i should get a new one probably. i've got plenty of room in this apartment (see all my comment about spinning in my chair)
I'm pretty sure I can also get my rift working if I wanted to but linux
I borrow the quest 2 from sibling
Wtf, it doesnt work anymore because they stopped support?
What kind of support is needed for it anyway
entire stack

i think it probably works but they don't update the driver for that model?
it needed drivers for the camera tracking things and software to convert those images into 3d locations
and you had to sign in to meta for their software
there's probably a 3rd party software for the trackers but it's also ancient hardware.
i'm sure there's something better that's decently priced.
I've heard mixed things about 3rd party solutions, some people swapping in so they have less weight to run steamvr on windows
It exists but I don't have experince as when I would have wanted to I already had massive instabillity on windows for VR
Linux continues to work better than windows ever could (11)
i used to play a few games semi-regularly but stopped playing when i moved into a smaller apartment.
just had to frontload the suffering
i think i'd prefer someting made in the last couple of years over reviving my oculus.
quest 1 is out of support so I'd avoid that boi
can't even launch games itself, has to stram; hearing from a friend
tbf I'd be looking at either bigscreen beyond 2 or steam frame. Not sure which I'd prefer
looking at the prices, i'm not convinced I need a new VR setup right now coz I kept looking at all the cool add on parts.
i accidentally put multiple blankets that become like 60 lbs when full of water in my washer at once
so it was trying to do a spin cycle and would just start tubthumping right as it spun up
so it stopped and had gemini speak to me irl
"the washer needs your help"
excuse me what
Wait, your washing machine have smart function?
"it's too heavy. it needs you to help spin it." ;]
Damn, that sounds nice to know when you fucked up lol
dryer too and theyre linked
Only thing I might do is look into completly diffrent, but grab one of those mind probes that can read electrical signals so I can rig events to them.
My sensory is weird where I can feel where parts of my head are being used so I could at least hook into those exact spots to grab free activation points.
"The dryer wishes to inform you that it is not speaking to washing machine until they apologize properly for not doing a good job on the spin cycle yesterday."
i could see it
one of the breakups of all time
i have the dryer play pachelbel's Canon when it finishes a run
Should've used Tchaikovsky for actual cannon
ok I could've used that when I was washing a matress topper yesterday. It paused the wash to avoid self-destructing, and it took me half an hour to notice that I needed to re-balance it. 
They are unassuming in appearance
Wash clothes pretty good tho
If it has wifi
You might be able to configure it to
but after the AI revolution you'll have wash your own clothes!
cat sticker my beloved
I don't think it's very advanced but I did see some kind of port on there. Might have diagnostics.
damn my dryer done went on strike
"More socks! Less lint!"
tbf my parents had lg washers but got them when it was like 15+ years ago and even then they played a song when finishing a load
they could not configure theirs by talking to a ghost in the air tho
for some reason i think theyve like always played the jingle
lg
Lmao I never thought of voice assistant like that
Mine is in the basement and I live on the second floor. Couldn't hear it if it was making a noise.
"I'm throwing in 3 blankets that barely fit"
"bro get that out your going to kill me"
This is just yaoi
I'm going to perish now
you saw what you wanted to see in your heart

i feel like there was a term for this
ai wash
before we had ai everything
auto
lol
auto was the term
fuzzy logic
that was big for a while... meant "thermometer"
like
this is just
normal wash
it seems to just change the length and like how long it does each cycle part
that's literally what auto did i think
where is the cleaning the body wash?
the AI part is that it can talk to your phone though
zojirushi still got dat fuzzy neuro
the cleaning time isnt changeable like that i guess
so there's nothing there since i cant change it
it just says how long it's gonna be till it's done lol
it probably has a strain gauge for the weight and measures RPM vs the percent duty cycle it's powering it and does some math on that
might not even have a strain gauge...might just use inertia estimates
im surprised it didnt complain about the weight once the blankets were soaked
i guess technically it did
probably couldn't get up to speed
well it was because it was unbalanced
100 lbs of blanket on one side of the tub
make spin cycle go boom
yeeting myself out the window energy
bruh it has like this gap at the top
where technically you could reach down into the body fo the washer
it's just open
so
i put shit in there but if i put too much
invariably some shit gets yeeted in there
right at the start of the cycle
because it spins right at the start to measure the weight
spins right at the start to measure the weight
I knew i'd learn something watching youtube videos about random stuff.
yeah you were on the money lol
it will tell me when i should clean the tub and such
so i guess that's a nice "smart" feature
and apparently i have used 852 Wh in the last month with it
probably just has a timer for that... at least the air filter thing I had just waited until it had a runtime of like 4 weeks and lit up the "Check filter" icon, even if I'd just cleaned it.
it does, it literally just counts runs lol
i dont think it adjusts or anything
it just reminds me every 30 runs
gemini will probably tell me
scare the shit out of me again
it's the type of thing they probably tell you to do in the manual but nobody used to do it after the first time because they didn't keep track
i clean my garbage disposal every week like a good boy too
im very astute
and humble
(i liek the funny foam that it makes)
i've had to run a clean cycle on my dishwasher twice since i've moved in even though i never use the dishwasher because things keep leaking off the bench and dripping into it.
if it was smart it'd know it wasn't wanted and leave.
liquid spills, drips off the counter onto the door of the machine and somehow gets inside
oh lol
that's ominous for the seal
i guess if it's at the top it probably doesnt hurt anything
it seems to seal fine when i run the cleaning cycle ;/
maybe if it's not locked
i think mine pushes in a bit when i lock it to run it
from the locking action
mine is old as shit tho
i just cant make myself not use it because of how much more efficient it is
except of course from me spilling stuff
than my stupid ass playing in the sink
i just clean stuff as i use it. mostly i have sandwiches or cold food, or the food is in a paper/cardboard bowl for the microwave
and if i have to clean the air fryer, i need to run hot water anyway so may as well do the rest of the stuff.
and i guess i just got used to not using it.
(spinning is a perfectly normal thing. even in the kitchen without a chair. defnitely grew out of any autism i had as a child.definitely. definitely.)
i had a roommate who paced. he seemed very anxious. was quite high strung.
I do also pace, my family just accepts that I pace
helps to not be in one place as I'm thinking a lot
i don't pace, but i do get up and do things with purpose. repetitively. i find myself looking in the same 2 cupboards and the fridge 10x a night without actually feeling hungry or getting any food.
with apparent purpose
im trying to catch this fucking thing playing a tune and dancing at the end of thee cycle
it's said 1 min left for like 10 mins now
very smort
probably the weight of the water in blanket
fuzzy logic... "It's close, ok.. not that close...."
probably tries to get it within a % of the weight witout water from the start or something
maybe as it dries out near the edge it's able to pull a wet part from the middle
and that adjusts the loading
if it used a % of the weight you wouldn't be able to take something out... i think it just measures how much water is coming out and once it's basically stopped it'ill stop
i feel like pretty much everything it wants to know is handled by weight and like a flow sensor of some kind
yeah like that
behold
well
if i trim size lule
the dryer will have a sensor for humidity, but a washing machine will always leave the clothes wet
Google's certificate problem is over and this still doesn't work, so the fact that I had to find a way to bypass a completely broken form in the wacom center app to get to the send verification email button in the first place is starting to make me think that Wacom doesn't actually have the software offers they advertise and are just pretending for some reason
i changed the jingle to bethoven's 5th
it's a virtual host and it's showing the host's cert and not the one that it resolves to.
the speed controllers for my robots let you play a tune on startup.
with 4 speed controllers for 4 motors you can have 4 part harmony
reminds me of floppotron
(assuming they start up together)
Patreon: https://www.patreon.com/floppotron
Donations: https://paypal.me/floppotron
Eurythmics - Sweet Dreams cover by computer hardware orchestra.
Which song would you like to hear next?
Like it? Share it! https://www.facebook.com/floppotron/
More info on how it works: http://silent.org.pl/home/ .
if those drives weren't already dead, that drum beat would soon kill them
lol
yeah i dont think drives do very well exposed to air these days with how the platters are so dense now
som ehave nitrogen and shit in them
helium
kek
nitrogen
the 24tb drive i have has helium
not exactly sure why but i imagine friction
once you get over a certain speed you want as little friction as possible probably
yeah agree they cram them in there as close as possible
i'm going to have to make a list of tech i need to replace. ;/
i haven't got any drives large enough to backup my main drive and that's 81% full.
Claude gets it
i'd give them a day to fix it (in case it's related to some other outage and they need someone on their side now to fix it) and then look for an 800 number to call?
i just spent like 10 minutes raging and wandb claiming im not logged in
gave it like 4-5 tokens
until i realized
i was giving it hf tokens


maybe you can just download the software from their site without filling anything in?
Well the software isn't free, you need to do it through the wacom center so they know you have the device
If you have the device it's free
for 6 months
only for 6 months??
after that it's one first born child per month
(you could probably use someone else's)
look, I get if the software was locked behind the purchase of a wacom tablet, but to then only be able to use it for 6 months is crazy
👆 at least do that
still a bit shitty but you can do better than not letting consumers use the in-house app for configuring their tablets at all
The perpetual license for clip studio paint is $258 so at least you get it for some time
aintneurway
after 6 months, what happens to your art? can you still open it?
Idk, I assume you still have the project files and can convert them somehow
the "and can convert them somehow" is doing a lot of lifting
if you have something which can open them already, why are you using the program?
"""somehow"""
Although to be fair, that's the perpetual license, it's $0.99/month
$1/month
honestly not that bad
what like, how long would you normally use it for?
When compared to Adobe that's amazing
are there any open source or free illustrator clones?
I mean, I'm shit at art, I'd just want to try it out if it's possible to get for free for some time
you could probably look around and find out, every adobe product should have some oss alt
Plus if it's included with the tablet it should be easy to get, I shouldn't be having to use devtools to complete forms because the site is so broken
(i've only every used pixel type tools like Photoshop and GIMP)
i know that blender has a grease pen tool
for animation
You can try Inkscape, but it's apperently a lot harder to use then illustrator
Affinity Studio by Canva is free
How have I never heard of this, it actually looks cool
It was affinity [product] for a while. And they even have perpetual license for up to Affinity 2. The Studio version does not have the same license but you only need it if you use their AI stuff
they say it's free and they'll never sell your data coz they want people to use their AI services and the more users of the basic tool, the more chances to upsell them to that service.
but then why the "Sign up to download" buttons?
why do they need my details if it's that free?
I can give you the installer and it will install just fine
I don't even signed in when opening their app IIRC
Lemme actually spin up my VM real quick
You know windows is bad when even a fucking VM takes ages to boot 
they made VMs worse (imo) when they made all of windows a VM (sorta) and stopped vmware from working. ;/
stupid hyperv stuff
VMWare works for me
it broke like for a version or 2 unless you did workarounds when they made the change
like years ago, sometime in windows 10

used to need vmware for work stuff
It sure is taking it's time
no network?
Lemme restart it lol
btw, I don't know how tf you're supposed to even download VMWare at this point, the only way I've been able to figure out how to do it is downloading the installer from the internet archive since I can't find it on broadcom's website
i wouldn't be surprised if VMs are just like normal files that the OS opens like all the compressed stuff
i try avoid the gui in windows unless it's helpful for something
Uhhh, one sec. I actually forgot how to share FileSystem 
You can use vmware without the gui?
i used to run an ESX or GSX server? or both maybe at different times
headless VMware server

that ran on its own server and i used it for testing... work adjacent stuff.
15 aur/vmware-vmrc 12.0.5-1 [+15 ~0.04]
VMware Remote Console
14 aur/vmware-keymaps 1.0-3 [+55 ~1.13]
Keymaps required by some VMware packages
13 aur/vmware-horizon-usb 2406-4 [+56 ~0.01]
VMware Horizon Client connect to VMware Horizon virtual desktop - USB device redirection
12 aur/vmware-horizon-tsdr 2406-4 [+56 ~0.01]
VMware Horizon Client connect to VMware Horizon virtual desktop - folder sharing
11 aur/vmware-horizon-smartcard 2406-4 [+56 ~0.01]
VMware Horizon Client connect to VMware Horizon virtual desktop - smartcard authentication
10 aur/vmware-horizon-rtav 2406-4 [+56 ~0.01]
VMware Horizon Client connect to VMware Horizon virtual desktop - Real-Time Audio-Video (webcam and audio-in)
9 aur/vmware-horizon-mmr 2406-4 [+56 ~0.01]
VMware Horizon Client connect to VMware Horizon virtual desktop - multimedia redirection
8 aur/vmware-horizon-integrated-printing 2406-4 [+56 ~0.01]
VMware Horizon Client connect to VMware Horizon virtual desktop - integrated printing
7 aur/vmware-horizon-html5mmr 2406-4 [+56 ~0.01]
VMware Horizon Client connect to VMware Horizon virtual desktop - HTML5 MultiMedia Redirection
6 aur/vmware-horizon-client 2406-4 [+56 ~0.01]
VMware Horizon Client connect to VMware Horizon virtual desktop
5 aur/vmware-workstation 25H2-4 [+236 ~3.94]
The industry standard for running multiple operating systems as virtual machines on a single Linux PC.
4 chaotic-aur/vmware-workstation 25H2-4 [633.70 MiB 1.05 GiB]
The industry standard for running multiple operating systems as virtual machines on a single Linux PC.
3 chaotic-aur/vmware-keymaps 1.0-3.3 [6.77 KiB 57.99 KiB]
Keymaps required by some VMware packages
2 extra/xf86-input-vmmouse 13.2.0-2 [14.16 KiB 41.02 KiB] (xorg-drivers)
X.org VMWare Mouse input driver
1 extra/open-vm-tools 6:13.0.10-1 [993.47 KiB 4.67 MiB]```

pacman : The term 'pacman' is not recognized as the name of a cmdlet, function, script file, or operable program.

I was going say that
Beat me to it 
Also why does scoop take like 10 minutes to search for shit
there are package managers for windows like chocolatey i think?
and scoop
And the built-in one winget
yeah.
I don't think pacman for the deprecated Windows Subsystem for Android is going to help me here
I did actually find an archive of WSA on GitHub at one point and used it for a while until I ran out of storage for it, it's like 11GB
And I am not good at storage management
Yeah it installs just fine with no internet
Let's see if I need to login
Oh yeah, first time only to get the license. Let's test it
it says they might use it to communicate with me. not interested in being communicated with.
I forgot how tedious it is to login to my account 
No internet required after the first time login

Also, I did have it on my linux machine
By following this guide
https://github.com/seapear/AffinityOnLinux/blob/main/Guides/Bottles/Guide.md
Tried the same for Fusion360 but they being an ass about the QT lib version
So WINE didn't save it hence why I have VM
Oddly enough, I don't need to sign in on linux 
fusion is unfortunately something i had to sign up for coz i needed something decent to design robots with.
Yeah, it WAS working before, but the latest installer broke it. So if you have an older installation and has logged in before, you are safe
Because only the log in is borked
Everything else works fine
I follow this guide
https://github.com/cryinkfly/Autodesk-Fusion-360-for-Linux
i think you just need to log in every 90 days or something to keep it active. i had to re-sign in once last year when i took a break for a while.
Nah, not even. The problem is not the actual login, it is the use of web browser redirect. Because once you've logged in, the auth mechanism doesn't depend on the browser anymore
And for those browser, they somehow depends on QT
Which they ship a custom version
ah. QT is really really cool when it works.
when it doesn't, it can be awkward
based response on the right
How it feels like talking in #programming when there's the legendary 100 hash
first code change in days that broke in a non-trivial way. ;/
now i gotta work out where my ASR output is going. i see it in the new GUI but it doesn't make it out the door to the agent's /event endpoint.
should be relatively easy to fix though. i suspect it just replaced the 'output this text' with 'send the text to the part where it displays it'... without also outputting it
This is me for any new 400b+ param model
-# knowing the pramater count implies I can run it, for anyone who will fight me on pedantics
yep. that's what happened. it made a new event queue and the consumer just adds the text to the log window.
VUE WHY, IT HAS BEEN 3 YEARS 
https://github.com/rudyxu1102/rfcs/blob/feat/infer_attrs/active-rfcs/0000-infer-attrs.md
"discord has a properly made UX system" 
I was looking for a toggle, I did not recall you existng
3 lines from the log file, 3 bugs to fix:
[2026-02-18 03:04:11.324] [warning] Avatar service call failed: TIMEOUT/ERR
[2026-02-18 03:04:11.338] [debug] Unhandled event type: 9
[2026-02-18 03:04:11.447] [info] ThoughtModel: Streaming TTS chunk (End Block): ' '
wrong address for avatar.
it doesn't use type 9 any more... that was the avatar call.
and wtf is it doing TTS on a trailing space for??
in advanced settings this is the boi in question.
what a joke of overlapping binds
I love it when this channel is becoming the equivalent of rubber duck lol
One of my favorite times to pop in
Especially if it is Shiro that is crashing out 
that one is a feast that should be celebrated
you can't be activated with either setting, it literally only works because the other bind is disabled while scrolling
Sometimes I wish I didn't like to test so I can remain oblivious, maybe then I would have failed demos of NN out sooner.
Is ES6 doesn't add newline on string template? 
https://github.com/hanszhoou/endent
Also, I think someone at discord needs to be fired for writing dog shit code that enable sentry console logging on prod
I'd post more here but it is mostly either NOWAYING IT WORKS to FUCK THE IMPL IS FLAWED and also a large crashout.
i enjoyed sending extra (malformed) data to analytics providers. extra commas, quotes and parentheses in the fields. that kind of thing.
I will accept NNv4's win in managing its goal on the match side. After all of this time it is cool to see
I'm no longer starting from nothing
shallowdream204/BitDance-14B-16x · Hugging Face https://share.google/YoDWEw8FocpbUDgUu
weird for it to be a google share
This may interest you
It's cuz I'm on my phone and stupid
While maintaining the standard language modeling paradigm for text tokens, BitDance employs a next-patch diffusion paradigm for visual tokens to predict multiple tokens in parallel—up to 64 per step. This unified multimodal framework is simple, scalable, and capable of efficiently generating high-resolution, photorealistic images.
Seems very similar to what I'm doing. Minus feeding into an embedding layer
time to begin reading this boi
oh shit it's llm based image
"we scale the entropy of binary representations,
expanding the vocabulary size up to 2^256" Yep that sounds exactly like what is being done with NN
I guess I'll add something like this onto the list of items to train on.

I actually have a very high entropy dataset I can test with for NN
i'd like to see something which can actually keep the style of drawings. models have converted my rushed sketches into beautifully inked anime characters too many times.
i think it's because it loses too much in the encoding
it goes "i can see what you were trying to do" and does that instead
only thing I'd recommend is a hack to extract the style, give a smol lm the goal of combining your prompt with what it sees and pray to frick it works.
binary hypercube looks so bad out of context but that is the state I'm thinking in most of the time
"it reveals that the model effectively learns the
discrete nature of binary tokens without any manual constraints." And this is NNv4
you'd then match it to a embedding which is crossed via normal transformers
patching is the only way this scales, so makes sense there
we are doing well because we get free cross attention from the bits, yeah only makes sense if your abstracting once for image noise then another for general context.
SIlently proving that patch matching can stack
That is about all I know, starting to go into the generation part which I still don't fully understand yet. Sure it is a transformer model, but it is still weird to me.
as an open letter to anyone who might try and impl something similar, you need patching for this to work; trying to binary match a single token's worth of infomation isn't enough to generalize
do all text2image and image2image models use CLIP? coz it feels like i see CLIP's 77 token limitation in a lot of other models
I will correct myself and say NNv4 doesn't act in the same way as this paper, they are still using transformers just with bit activations; meanwhile I'm pattern matching.

I only picked image based as a stress test, I'm not knowledgeable on what the field is doing.
no they do not
a lot do though
it's small and good-ish at what it does
since around sd 2.0 (flop) it used openclip and clip
and then after that you had clip + t5
which is an llm
imagen doesnt use clip
it really has trouble when you start using prepositions though
it just uses t5
relative locations and stuff
flux and sd3 type DiT typically use multiple clips which concat embeds
and t5
but in later gens recently clip is kinda fading out finally
larger more recent DiT use t5
some of the recent chinese models use GLM
which i believe is how GLM (the precursor to like, GLM5) started
clip is not an LLM
so it cannot grasp anything like what an LLM can
all it does is contrastive pairs
the models that use clip + t5 or similar generally just use clip as a styyle guide
aesthetic stuff
it's good at that
just nothing as far as precise placement or anything
autoregressor iamge gen i dont think use clip in general
because that's part of their wheelhouse
clip also gets the wrong homonym at chance levels, but it's consistent with what it thinks you mean.
coz it's just a word -> concept match
clip is a truly brainless encoder
that's it
you can replace CLIP with the encoder portion of just about any LLM
it will change how the model "acts" though
77 tokens hasnt come up for me in at least 2 years
as far as being something i've even remembered was an issue
i hardly ever use such big prompts, but i generated some prompts using some other data and some were really long and i got it on a couple of different models. they might've been old. i mostly based it on download numbers.
i wanted a small model so i could try live updating an avatar directly from the model... was still too slow to do it live without caching.
i think if you go on like civit you'll find the most popular by raw downloads are still like sd1.5
because those were very easy to run and it was brand new so they were huge
with lightning distills and stuff peple got to almost real time with these models
with like 4090s and very small sdxl lightning checkpoints with distils making the gens in like 1-4 steps possible
sdxl turbo i think was one i played with for that. was borderline ok but i couldn't get enough image stability. couldn't provide latents for image2image because it would overwrite the image
you have to limit the amount it can change the image otherwise it'll immediately poof it into dust
running something with like .2 weight on the init image can get subtle cahnges
without the latents it would also just blit between different modifications as i traversed the embedding space
yeah there were a lot of fun little videos possible with that
i slopped out a controlnet guided (lineart) shukusei video in like 2023
that used that in the background
with the text2image i had good stability of the characters, but they looked nothing like my art.
but with image2image and a character still, i had the character looking like my art but they weren't stable when they morphed.
there's got to be a different source for the latents during the image generation.
first lora of any kind i ever trained were sd1.5 ones
coz it produces different images with the same prompt, which means it's not just using the image itself as the latent
it does, but the samplers introduce noise
by design
they are deterministic in general if you use the same seed and init
when i pass in a latents argument, it converts the image to noise.... what argument should i pass in?
lower the strength of the sampling
how's the hash silencer
banned from here
what argument gives me the same seed each time? coz i've tried it with different values of strength.
image = pipe(
prompt_embeds=interp_embeds,
pooled_prompt_embeds=interp_pooled,
image = mindy,
# latents=latents, # Crucial: use the SAME latents to keep character identity... or use image
num_inference_steps=4,
strength = 0.25,
guidance_scale=0.0
).images[0]
i imagine if you're not specifying a seed it will simply use a random
if i have image as well as latents, it always draws a noisy gray image
i do a slerp over the embeds to get the interpolated bits
the idea was to generate every possible expression using facial expression descriptions with the base avatar... but it didn't work well.
that kind of thing is not generally what diffusion image models are good at because as you have found their outputs are very uncontrollable with that kind of consistency and granularity
it's inherent to the way they work
I guess enough of you reported the account
this is why it was a big deal with gpt4o finally was allowed to generate images
it's autoregressive
works completely differently
was much more consistent from that
plus the strength of its encoder portion was being a massive llm
when i change it to use the text2image stuff it just works how I expected. it's because image2image is a hack, isn't it?
depending on what you're using it could be just the way it's initializing the image and what you're using as arguments, idk
it's all a hack
this subsection of the ai space is true slop and woo woo half the time
"lets see what happens if we put 50 of those things in a row! no! 500!"
"awwww. out of memory. 50 then."
it's been the lack of memory holding us back from truly amazing models. we should be happy there is a RAM shortage. ;/
I do actually find the article interesting, the researchers were comited to giving their thing a shot; testing aganist models they had no way of accessing
unsolicited friend requests generally get ignore by me. especially from a serial vedal pinger.
https://aisparkup.com/posts/9220 not in english but translate is giving enough of an idea
Most of the time I just see what they want
then cast judgement
"Serial vedal pinger" lol
pinged twice in 2 days
Just ping his evil twin, ladev
so yes
yes
I can't stop winning losing
they were just about to win... you can feel it
Just... one more round
ok. fixed one issue. it should no longer leave a trailing " " in the buffer. It was tacking it on to the beginning of the next sentece because the logic looking for ([?.!]\s) didn't increment the pointer for the space.
and TTS of the " " by itself resulted in a sound blip. coz the tts is stupid.
i do like to call my drill stupid when i screw up making a hole in the wall
dont use cfg7 on flux krea btw it's not a good idea
This is what the TTS thinks " " sounds like.
{
"audio" : <delete base64 encoded sample coz it's above>
"audioEncoding": "wav",
"phonemes": [],
"vdurations": [],
"visemes": [],
"vtimes": [],
"wdurations": [
46
],
"words": [
" "
],
"wtimes": [
225
]
}```
weird.
it's saying "there is a word ' ' which starts at 225 ms and ends 46 ms later. it has no phonemes and also no visual representations of those on the face.
it's 1/2 right. unfortunately not the important 1/2.
"di bloh"
can't use a thing I obtained, my use case fights aganist the saftey features
YAYYYYYYYYYY
colorful.
another colorful array.
AEAAAAAAAAA D:
That is crazy.. that is actually crazy.
what you get when you process a file into source so you don't have to read it every time you start up.
and i probably only ever used like 6 colors, max. could've just used to RGBs really.
i had "content" instead of "payload" in a json blob. ;/
i think that means they can hear me now.
Speaking of sticking to rgb, there is always truecolor depending on the terminal support
also known as not windows terminal
the windows terminal has gotten much better.
i actually use them for most of my terminals. (i still have to remember which windows to middle-click paste and which to right click paste.)
I thought I made a really really COOL colour gradient. It's 3 colours; cyan, magenta, and grey.
Magenta might be purple, I'm not really sure.
they work well together.
it's magenta if it's 100% red and 100% blue. coz that's how i learned colors.
I honestly don't know what these colours are.
very cyberpunk/DOS game aesthetic
add some yellow maybe to make it pop? ( #ffff00 )
will be a different feel though
I'm not controlling the colours, though.
public static void WriteOscillatingGradient(string text, double time)
{
Console.Write('\r');
for (int i = 0; i < text.Length; i++)
{
double t = (Math.Sin(time + i * 0.3) + 1) / 2;
var r = (byte)LinearInterpolate(255, 0, t);
var g = (byte)LinearInterpolate(0, 255, t);
var b = (byte)LinearInterpolate(128, 255, t);
Console.ForegroundColor = ClosestConsoleColor(r, g, b);
Console.Write(text[i]);
}
Console.ResetColor();
}
public static double LinearInterpolate(double a, double b, double t)
{
return a + (b - a) * t;
}
It was meant to interpolate between these, but I messed up somehow.
there aren't many close colors and you've fixed the red and green values and the blue value is only 0-1
But it should go between red-red and blue-blue in a scaling.
I think it's because console snaps to 'closest' very rigidly.. but yeah, I forgot this (original was used in a non-console environment and allowed for very smooth gradients)
i'd just do:
r = 255 * (Math.Sin(time + i * 0.301) + 1) / 2;
g = 255 * (Math.Sin(time + i * 0.2941) + 1) / 2;
b = 255 * (Math.Sin(time + i * 0.311) + 1) / 2;
so that way they don't just fade in and out in sync
that should make sparkly colors that drift out of sync
to make it more shimmery, multiply the time value by a small number, like 0.1 or 0.025 maybe
then it'll blend more smoothly (at the expense of starting grayer)
might want to make it 16 + 240 * instead of 255 * so it doesn't go completely black
might even need to be 64 + 192 * coz of the quantization
Yeah I gotchu!
erg.. if there's only 2 levels of blue it might be 255/3. 85 + 170 * might work if the others don't
i've done this kind of thing with RGB lights and a microcontroller, but they don't have the quantization either
cann i have some 
Weeeee
barber sign
That'd be red and white, but I like the idea.
Bing just taught me I don't have to deal with console not having full colours.
YAYYYYYYYYY
there's a package called lolcat on linux that does basically this
you can just pipe output thru itt
not that it's not cooler to make it yourself cuz it is
looks like the trans flag or am i just colorblind
you're colorblind
it's closer to the bisexual flag 
i dont see it at all
it literally has like
white
white? 
is that green
one thing is that the custom boi is animated
SO_LINGER with a timeout of 0 is such a nuclear way to close sockets... i wonder if i should feel bad for it.

i held peace talks with my code and decided against the nuclear option. now trying to reuse connections for certain operations.
I hate that Microsoft might be vibecoding Windows, but it's inevitable
︀︀
︀︀microsoft laid off everyone who knows how c++ works so now they just prompt gpt 5 to fix the codebase. 30% of windows is written by ai. that is why your printer drivers were deleted to make room for 4gb of copilot telemetry
︀︀
︀︀they rewrote office in typescript. file explorer and the notification center are now just bloated electron instances that take 3 seconds to render a right click menu
︀︀
︀︀the taskbar and start menu were rebuilt from scratch in react just to shove ads and "recommended" bloatware in your face. it uses more ram than world of warcraft did in 2004
︀︀
︀︀copilot is being forced into notepad and paint. they are forcing you to test it in your basic tools
︀︀
︀︀windows search isn't looking for your files. it's a bing wrapper designed to sell you a microsoft 365 subscription while you're desperately trying to find a local pdf
︀︀
︀︀the widgets…
i see it
and i’m not colorblind
the teal/cyan looks like white if you don’t pay attention at all
it’s kinda similar to that one blue and black image of a coke can that appears red
white coke can
🏳️⚧️
looks the same enough to me
colour scheme wise it looks more like the gay pride flag, with a hint of purple
I don't know how the flags work. I just wanted to oscillate between cyan and magenta.
Except for bool flags I guess.
does anyone have experience writing fnaf 1 mods? would it be done by decompiling and adding stuff to clickteam fusion or is it possible to write it seperately and have it hook into the game?
While schizo rambling is cool and all, is there any actual proof to, say, Office being written in TS?
https://xcancel.com/forloopcodes/status/2023845548063944889#m
Oh OK so this person’s opinion can be safely discarded
why wouldn't the games take lesser storage as they used to before?
modern game uis are literally just websites now. modern warfare 3 runs its menus on react and electron. you are giving up 16gb of ddr5 ram just to render a battle pass button
nvidia dlss 3 is a software scam to let unreal engine 5 devs skip optimization. why write clean c++ whe...
It’s another schizo
I wouldnt classify what i see from the embed as shizo, rather just fairly rightfully upset
Maybe a bit much, but one has to bully corporations
nvidia dlss 3 is a software scam to let unreal engine 5 devs skip optimization. why write clean c++ when you can force gamers to buy a 5090 to fake 60fps using a 480p base render?
Will clean C++ give you 60 FPS in pathtraced Cyberpunk 2077
clean c++ 
they've been at it for a while
disliking Win11 for being slow and broken is fine and all but just say the actual reasons instead of making up random bs 
Some cool new training optimisation research from Deepmind
Well the dlss claim is one of the main parts that id describe as a bit much, i played with dlss 4.5 preset k at stuff like 360x240 upscaled to 1080p and its like Witchcraft. However i can absolutely understand that theyre upset at the lack of just simple optimizations just not being done, so many tricks in old games because they were needed, and they had to the average player no real visual impact.. and the decline in actually meaningful settings is kinda sad. Games like battlefield 1 and such, if you set the quality to low, vs medium vs high vs ultra actually had a difference that was visible.. nowadays the only really noticeable difference is between low and medium, and Performance wise, many games barely Care about what settings youre at, it all runs roughly the same.. which defeats the point of lower graphic settings.. an example of this would be "It takes Two", probably not the best example but a game that i recently played and noticed that
and if office is written in typescript, that'll be because it runs in a webbrowser now too.
I personally do think that dlss is overused, but it has a valid usecase and is valid that it exists. Once dlss can render and then upscale from a lower resolution to a higher one faster than it takes to render natively without AA, and then looks as good as lets say 120% render scale, thats where imo it shines. Its surprisingly close to that
Sure, but the decline of actually native programs is kinda redicoulus. Expected from corpos, but redicoulus
The worst offender is the w11 taskbar being run on nodejs afaik
if you have a web version the easiest way to keep it in sync is make it the actual version
it's not being run on Nodejs 
the recommendations thing in the start menu is using React Native AFAIK
the taskbar itself had a design change but I don't think the underlying stack has really changed
you can turn those off, you know. (I've never seen them)
Well sure that then.. they had UWP and interactive tiles, all super speedy, look at windows mobile. That stuff ran on shitboxes. But they axed it
The issue is the web version is garbage
Desktop Office is generally less garbage
i didn't say it was the best for customers. i said it was easiest.
Or yknow, webassembly + a layer that can either use some react nonsense or gdi or whatever os interface.. only because you offer a web version, that many people would rather just not use, doesn't mean that the main office should just turn into a local copy of their server + a browser..
Also a neat insight I missed from vending bench
And for crossplatform, we had a thing called java.. and Microsoft has its own C#
it's harder to keep a web version up to date with a native version.
It's easier to keep a native version up to date with the web version if they are they same version.
that was all i was saying
i understand why the big companies all rewrite thier products to go web-based. On the sruface, it's OS agnostic. Doesn't matter if Mac, Android, Winslop, iOS, Linux, whatever. They have the Browser as abstraction, which also does a lot of sandboxing security things. And to them, internet is always available. Performance is only a cost to the consumer, his Time isn't thier Cost.
Yeah thats true, and thats why it makes sense from a corpo view. But that doesn't mean its not valid to be very upset at it, no?
Checks out 
once networks got fast enough and browsers powerful enough, it was bound to happen.
Yeah, optimization got thrown out for convenience. That's the whole point the dude was trying to make. Doesnt seem shizo to me
This one (for any late joiners)
ranting about it multiple times might be an indicator though
I mean, i personally if i had the time would go all out, make my own software and spread anti corpo propaganda all day long.. but i dont have the time. Seems just like common unemployment to me not shizo?
hitting some of the same points each time on two diferent rants
except that's not what they said at all 
they didn't say "Win11 is slow because MS doesn't care about optimization" (true), they said "Win11 is slow because they rewrote the taskbar in React" (false)
Ok.. he miiight be VERY upset lol
Its just a bad attitude to rant about devs being stupid when its more likely the work environment and industry standards dictating things
Win11 is one step ahead of the game in certain things innit
I mean, the details arent correct, but more or less more and more being turned into own chromeium based browser and server surely isnt helping
Im mostly scared about Windows 365 or whatever they are calling it.. Windows Link?
WTF is this?
That shit makes me wanna go on a rampage
it's not, but the whole post is just incorrect details that people will actually believe now 
thats the cloud based pc thing right?
it doesn't help you if you convince people by not telling the truth. they can be unconvinced by the truth
I can respect that.. not enough people are hating on Microsoft.. do the details actually matter aslong M$ suffers for being ass?
Yea
that makes me so angry. "You will own nothing and be happy"
including your pc....
im so glad i switched over to linux
they do
if you don't care about the technical details then just don't mention them and focus on your experience with the OS instead
How to become not addicted to windows tho?
True, if they do research sure. The average consumer doesn't do that tho so.. seems effective. Who doesn't love spreading misinformation on the Internet? (I dont have anything to add from here on out, just somewhat replying at this point)
if you make your whole rant seem to be based on a specific fact and people find that fact is false then your whole rant will look silly.
If you need to make up fake reasons for people to hate it then clearly its not hate worthy to those people
believe it or not, different people have different needs and requirements
always take your vitamins in alphabetical order.
for safety.
It’s like, kernel people are cooking 24/7 but userland people have some major issues with UX
I mean, in the end we found out the idea of the points werent wrong.. just the factual infos were.. which yeah is a shame, is just uninformed, not shizo.. or what am i missing now?
"What are vitamins?"
old people candy i think
time to do non-computer things for a few hours.
Will Windows 11’s UX issues ever be unfucked?
Microsoft is vibecoding their OS now, so.
its not guaranteed but pretty possible
actually a decent summary 
i know, very cool 
well thought out, examined both sides, unbiased
alright, new day. new sin, im thinking of mixing amd and nvidia again... and this time its 1 Sapphire Radeon RX 7800 XT 16GB for actual gpu cards. and 6 NVIDIA RTX 4060 Ti with 16 GB for ai training tasks and such... this feels dirty, but id like your toughts on the idea (and yes i will physically block those rtx-4060's from getting used for anything else)
aww
not gonna mention the fact you have 6 4060s but its probably fine as long as they are seperated, at least thats my experience on linux
wait thats so cute
i have a 3080 alongside my radeon pro r9700 rn
no problems
arch linux
they're gonna be in one 7 full pcie mobo
same here, 🤝
1660 super alongside an rx580, arch linux
wow 3 arch users here
hmm
all i did was install mesa first then put the nvidia card b ack in and put the nvidia & cuda shit i needed to use for compute without putting it as a display out at any time
works like a dream honestly
coming from windows i was terrified
wich one of you messes around with the nvidia/amd open source drivers then?
noveau
because im working on kernel specific things
yeah noveau
im using the cuda card for pytorch and such alongside rocm7.2/hip without issues
other than the issues with rocm lol
but that's independent of nvidia
i just use nvidia and the cuda drivers
i just put cuda related shit in a venv normally
just in casde
probably dont have to
yeah cuda related shit belongs in a venv, but im arraying the 4060's to be one beefy card actually
im also hoping to somehow getting myself some good high-bandwith ram
but uhh... yeah, either super expensive or too small for me to use
i had like a whole backup disaster recovery plan ready to go turning it on
and then it's just fine
also, is it a sin if im using claude code to update my wiki's... report bugs etc on my private gitea server?
no i use claude 95% for bulk file and organization/doc shit lmao
good,
it's his best ability beyond just vibing things
haiku by the way im not using opus 4.6 for that
too expensive
what i do have is an activity runner that fixes my bugs automatically, has been working fine honestly
@fast pagoda what max plan do you have btw?
this is annoying.
the bad decisions tier
not as annoying as hitting your weekly limit in one day
thanks for reminding me to go downgrade it
dw, i have the highest max plan
then again its a plan 8 people use at once so... what do you expect?
i accidentally have that for a 2nd month rn i meant to downgrade it last month and for]gor
i have been using it a decent amount but i havent hit any limit
within this tier anyways
well, i have some programs that use claude haiku or sonnet/opus for cleaning up things and sorting, so i shouldnt be suprised im hitting my limit often
such a comical amount of this was vacuum related
You aren't kidding when you said it was a bad decision tier 
yeah thats not that bad
i mean considering i didnt pay $446 i guess it worked out
but yeah i have a batch request thing that i've pivoted to batch shit out to a bunch of subagents now
whats the monitor thing you use?
claude-monitor
thanks
ooooo
predict when you're gonna run out of use and such
thats useful
so i just like to have a split terminal with this in it
i got extremely butthurt at rocm like 2 days ago because it just refuses to run training reliably with the drivers out for rdna4 rocm
so i have had a dangerously-skip-permissions claude just smashing his head into the issue
for hours
well he's stopped apparently atm
and i found out the reason is that system is totally hung frozen kekw
surely that's fine
he slopped too close to the sun i guess
oh dam
yeah that looks about right
minecraft java vulkan

also hi floofy
yeah @olive sable
or well... pay for the api
he needs to see that
nightmare
i make idle money, its fine
besides, most of the shit im running is heavy rn
setting up SVS/TTS on a custom voice model, and vevo is somewhat slow
so im letting claude manage that in the background
i had him do a runpod wandb sweep yesterday and that mf charged me $50 for b200
“him” 
i like to personify the alien math blob
yeah
that way i have no problems being absorbed into the rehoboam later
thats not too bad
mhm
i probably would've gone for like
i dunno
rtx 6000 pro
for the speed
it didnt need that gigantic of batches
mfw eval round immediately reserves 140 gb of vram
jesus
also... im thinking of making a vram card
not a gpu, a pcie card completely slapped full of vram modules
all i want for christmas is for pytorch to properly work on rdna4
all i want for christmas is not having to use onnx

SAM
that
ikr
if they're changing the renderpipeline they might aswell do it properly
true
not gonna be a claude code prompt to rewrite with vulkan
i skulled this while knowing i have this haunting my drives
You're being too optimistic lol
oh boy, thats a lot of paperweight
blegh
i can’t stop thinking about how this is going to be a nightmare for mods to migrate
i dont have to worry about things like that
the amount of fails on step 0
...
it's just the rocm kernel is cooked for rdna4 as far as i can tell
you know, maybe you should deploy claude code to fix this
from my experience the actual shaders and such are similair. it only some extra OOP or so that hopefully mojang, or else a mod-launcher, will need to handle
i DID he's looked at all these and slammed into the issue for hours with no success either
like after 2 hours of debugging or more id let it have a go
ehem
Interesting. Mods have shown that they don't even need to. You can speed it up just by using modern OpenGL properly. I suppose Vulkan is better for current mobile drivers though.
may i?
it works fine on instinct
it's just rdna
like the same exact train setup and everything
o,o
oh no shit they mention this in the article
honestly, i want to debug this for rdna
cant be that hard
discord for feedback from modders

it's a driver issuethat ive not actually tried to address, just sidestep cuz lazy
damn it’s so hard to not like them
and that's been worse
then fix the driver issue
"surely ill just use an alternative to pytorch
"
jarvis, bring me up to speed on the entire rocm driver stack ive never looked at5
xd
yeah no im making a pytorch alternative for rust
since im sick and tired of pytorch
what mobile drivers
it’s java
i suppose you’re right because apple silicon and macos but that’s explicitly called out in the article
actually i wonder if this does this on windows i literally dislike booting to windows so much i havent even thought to try
i mean it works dont get me wrong but python is so god damm slow
i find with pytorch and whatnot it's usually not the bottleneck, the actual work is being done not at all in python
depends on the person ig
Wait no way what??
fast, but horror
lol
ive used libtorch
sounds accurate
ive made ai projects in c++
i do not recommend it
like seriously it might be worth it regarding speed but if you wanna use libtorch make sure your project is simple
technically by using pytorch ultimately you're using c++ because that's what it wraps
it uses libtorch itself
it's just precompiled extensions
not again
looks fine
maybe i should try looking into sxm boards
all of my active claude instances crashed
me
i hate restarting docker containers ;w;
i need docker
they are, unfortunately, still up 
-# sample size 1
my docker keeps houndreds of projects running
sample size 2
Cloudflare status page not very meaningful
ye
that is true
specifically for the sp5 socket so ig budget is gonna be fucked
ive had a lot of issues with cloudflare that arent ever listed
why SP5
good, my most heavy claude instance didnt crash
i was like wtf is happening and i had to use a vpn to be routed to another server
becauyse epyc is epyc
ye Epyc is nice but why the newest and most expensive one 
you're questioning as if i don't have a 9655 running with bunch of supposedly-sxm-GPUs-that-uses-sxm-to-pcie-adapters
My bad guys. I just deploy around that time
gemini says
create music
now
did they release lyra
oh dont worry that might havent been because of you
im running a lot of instances myself
I know dude, I'm just joking lol
It's just so happen that I also just deploy around that time
yeah I know
but IIRC there's not a great difference in lane count between SP5 and older platforms like SP3
like, you could go dual-socket SP3 and probably end up paying less than for SP5
SP5 just denser and more performant
cloudflare founder and ceo's personal clawdbot deleted prod
Lemme get one of their repo
look, as long as this instance keeps running i wont have to be sared, ngl
this one's been running for four days straight (technically)
i thought that was titled "worst-oauth-provider" and i believed it

