#programming
1 messages · Page 479 of 1
I mean people make cheats for videogames while younger
yeah but
I could say the same with achievement/trophy hunting
true, but this is after they learn coding
(I hope)
Not necessarily
Honestly if you’re dead-set on modding stuff specifically for Neuro, it might be a good idea to learn coding by making a small app for her.
*pasting
with no prior, and on a fairly niche game
Like the wordless
i mean this isnt exactly a unity plug and play
starting with modding a game not designed for it is probably not the move though
wait is pasting dead now that chatgpt is a thing
:neuroHesRight:
These times you can tell Claude instead of Ctrl+C Ctrl+V
Someone made it for her? I though Vedal set it up
have you considered making a webpage first? or maybe at most, a minecraft mod/plugin
vedal set it up on his end but likely someone else coded it
Can’t make a Minecraft mod if it’s physically impossible to get Minecraft on my MacBook, seriously I’ve tried it too much and I don’t wanna explain what happens anymore…
yeah someone
definitely not claude
Maybe he did, I’m not sure. But making a little adaptation of, say, something like a crossword puzzle, would teach you loads, not just about how code works but also how Neuro interacts with stuff
Programmer’s best friend
I thought minecraft was just a jar, it should run anywhere
If I can get it to run on a raspberry pi, you can get it to run on an arm mac
i mod minecraft on apple silicon all the time it's literally seamless
Idk, nowadays Microshaft makes even buying the thing practically impossible 
I should learn how to mod Minecraft properly again some time
Why Minecraft tho?
Okay just to give you an idea of what this will take, from my 5 minutes of research looks like:
- its a ubisoft game made for windows (and some consoles but we dont care about that)
- this means you will need a way to use windows during development
- there are no modding tools for the engine used
- your best bet is a external bot neuro can interface with being used to watch the game and provide inputs
- this will involve OCR and some kind of websocket
- you will need to make a input model for neuro to use in something else like unity
Well external bot or a lot of reversing
considering you have no programing experience (not a bad thing) this is going to involve youo first having to learn programing in about 4 separate environments and some quite wide ranging skill sets
wdym "some kind of websocket"
I mean its ubisoft, im guessing its locked down
idk my first reaction to it would be useing a socket to call into unity
???
thats what they do for the other stuff right
why do you need to call into unity
use the exsisting unity sdk rather than having to actually mod the game
i mean you could mod the game but idk man
hello what
this thing has 5 mods for it and theyre all shaderpacks
you need to mod the game either way
why? just give neuro inputs and have the vision model look at it
how do you transfer those inputs to the game
its gotta be easier
also the unity sdk is for modding games, so...
? with the socket?
and how do you transfer those messages from the socket connection into the game without modding
you would have a little unity project to run this
have a little unity project that feeds in keystrokes with SendInput or somthing
why do you need to have a sidecar unity project...
well how would you do it
if you're gonna need a bridge/sidecar at the very least use something that isn't for a game engine 
mod the game?
go on
its a closed off engine with no existing api, almost zero mods
you gonna decomp it first?
with zero programing experience on a mac?
you're gonna have to unless you can modify its ram space ig?
I don't think I have zero programming experience
you arent doing it
?
the engine does actually have mods for raymand legends
is it the same version

given the age gap i wouldnt say so
do i look like i have a magical crystal ball
prehpas
We’re talking about how Aleez could do it
ah
even so idk why you'd need to "socket into unity"...
on a game that doesn't need unity
or use it
i would use unity as a base for the input mechanism and the OCR
rather than going to dx9 directly
again
thats my thoughts after a google search and a steampage read
i mean i suppose writing a dll wouldnt be the worst way, but its a lot harder
and again
decomping a 8 year? old video game
designer filed a bug because a border is 1px when it’s “supposed to be 0.5”
@jagged turtle you have to fix this bug
- why do you need to use unity for both
- input mechanism is likely more performant using a pure scripting language sdk
wha
sir
its a windows vista game
im sure we can get by
one word: hidpi
that's normal
can you even compile unity projects to vista atp
youd run it in compat mode i guess
Poor excuse for bad optimization tbh
idk with more thought put into im sure theres a better way this is just what came to me first
as the easiest way
tbh when it comes to game modding nothing will be anywhere remotely "easy" so whatever is the easiest really doesn't matter for this convo
yeah looking at this, modding it seems possible if you have experience
but idt using unity for ocr and inputs is necessarily the best way
what’s the convo here
I think its talking about a neuro sdk impl for child of light?
step 1 - learn more coding
same advice for all new leafs coming in wanting to make X neuro integration or X neuro copy
it looks like its possible to add arbitrary lua to the game, but that's it
oh that seems easier than expected
it depends how sandboxed it is
possible to regenerate the secure_fat.gf file https://github.com/wukko/ubiart-secure-fat
they just want to build an integration for neuro
i know that person
lol
they're cool but they keep shilling helium on my timeline and its annoying
am I missing context
if you made a browser would you not shill it

vivaldi?
The Neptune
it's a swiss army knife, i like straightforwardness
+it's ugly
But I don't wanna learn, I wanna be magically an expert through sheer force of will!
||/sarcasm||
Good morning, banana bread

Toast
Toast
Toast
Toast
Toast
I wanna make a chip for that someday
If its a from scratch one it's worth shilling. Yet another FF or chromium fork? I agree. Except maybe supermium lol
Now rotate it 90° and run a sorting algo over it
quicksort is quick surely
twittersort
more additions than deletions, bad commit
that straight not true
quality of life improves when project is archived
how can it be when money source gone?
Let's vibe code
@AGENT.md
i added completely new complex feature btw, this 1k+ lines is just including of it
man never refactored in his life hell nahh
except for this feature i completely rewrote 2 modules for new api
no that really not true
when you rewrite something, usually it takes even more space
you're doing something wrong then
for me most of the time it is
core code looks cleaner, but total number jump like x1.5 at least
ok I see the issue
I commit some shitty message as a maintainer and I got accpeted
every feature addition to a project constitutes rewriting it entirely from scratch, this is the most optimal way to waste time
yooo the AI approach
0 code extendibility, start from scratch every time you add something
i always knew i was a robot
embed perms as a new leaf
i reduced 1000 lines class to 200 lines
total net positive +500 at least
cuz: headers (x1.5 on its own), more asserts, more templates/abstractions, less chars per line like
while(QThread::currentThread()->eventDispatcher()->processEvents(...) == false);
->
auto this_thread_dispatcher = QThread::currentThread()->eventDispatcher();
auto flags = i with i remember names now;
while(this_thread_dispatcher->processEvents(flags));
how often do you post
pay2win probably
i really disagree
git commit -m "Update" -a
yeah that true btw
i know the solution, its hiring people to write everything for you and never looking at the loc
Built different
Open source maintainer after 1023th Ai generated PR by someone
wtf is this
how you even use it
just print git commit -am "aaaaaaaaaaaaaaaaaaaaa"

How to deal with Ai generated PR
reject it
ignore
btw, i have 2 PR dangling, both was approved to add in issue
one 2 weeks and second is 3 month //it was nov 22, not 3 month
ah nvm
countersubmit ai prs to the contributor's repos
it was just merged
https://github.com/bfgroup/Lyra/pull/109/commits
Meanwhile the repo commits filled with emoji

Best framework 100%
I like it
Can't believe my eyes, some cosplayer made this insane framework for modern js/ts
at least the Mandatory Anime Girl Project Mascot is human drawn
BunネイティブのWebフレームワーク「Elysia」を紹介!
スキーマを1つ書くだけで、TypeScriptの型推論・ランタイムバリデーション・OpenAPIドキュメント生成が全部揃う「Single Source of Truth」な開発体験が魅力
0:00 APIの型定義、二重管理になってない?
1:07 Elysiaの基...
คุณออม (SaltyAom) ผู้สร้างเฟรมเวิร์ก Elysia มาแบ่งปันประสบการณ์และแนวคิดการทำ end-to-end type safety ด้วยภาษา TypeScript
ภายในเซสชันนี้จะมีการพูดถึงข้อ...
This is a gem
Make it a stack of pringles
Eat one each time you get stuck on a line for an immediate solution.
Oh no looks like you reached your Claude opus limit for the next two weeks
Gaslight it
"No i havent"
Hit the high tech genie with the ol "i wish for more wishes"
:niuh:
↩ (@OpenAI)
GPT-5.5 delivers this step up in intelligence without compromising on speed.
︀︀
︀︀GPT-5.5 matches GPT-5.4 per-token latency in real-world serving, while performing better across nearly every evaluation we measured.
︀︀
︀︀It also uses significantly fewer tokens to complete the same Codex tasks, making it more efficient as well as more capable.
That's honestly a very good compromise on that kind of legislation since fighting open source would be mostly futile
I hope these arent the only benchmarks
Just compiled llama.cpp and so far its way better than ik llama for me. And also i got vulkan working in my lxc container, and my 5700G underwattaged does 26T/s on a 2B model. Im hoping that due to being essentially unified memory, that qwen3.6 35b a3b will run at a simmilar pace.
Only issue being that iq4_xs is still 17gb.. the qwen3 series was a bit nicer at 30b..
OpenAI also gibs model card
You can look at that IG
On my 32gig server, the igpu may only use up to 16gb, so some cpu may be needed. (Perhaps this speeds it up further?)
Astonishing that they managed to increase the number of experts in their MoEslop, truly a revolutionary next step
Perhaps a reason for me to upgrade ram.. if it wasnt so damn expensive
ai ai ai ai ai ai ai ai ai a ia ai aia idaicnaodhbfwhvf
GPQA and GLE degraded compared to 5.4
To better utilize GPUs, Codex analyzed weeks’ worth of production traffic patterns and wrote custom heuristic algorithms to optimally partition and balance work. The effort had an outsized impact, increasing token generation speeds by over 20%.
i need itsme to log into my bank app, but i need my bank app to log into itsme
truly the most technologically advanced belgian system
The solution is simple. Login to both at the same time 
going from 90hz to 120hz feels like a bigger jump than it should be knowing how it looks like on pc. so i think my old phone was just gpu limited
@real sierra @rare bridge 
https://store.steampowered.com/news/app/464350/view/553521213220061691?l=english
update
API changes
Added new property: Game.shard.access.
Added new method: Game.shard.activateAccess.
Creep methods claimController, reserveController, upgradeController now can return a new error ERR_ACCESS_DENIED.
pinging 2 mods for this stuff feels so wrong
This comment makes me explore more of my phone setting and I found a very interesting feature. Thanks Toast 
why can't they just use normal names... what does Pro even mean 
Doesn't this happen automatically on a lot of phones..?
Bypass charging my beloved
i dont think so, not on most phones. only when the battery is already fully charged maybe
Suprigingly not
Most route through the battery
they fucked up the update i can't access screeps
yeah but it says the update downloaded and to restart, then when i do restart it just says the same thing lmao
😔 banned from screeps
I don't think so. Usually the phone would still route power through the battery because it's easier to design by the manufacturer
how do i make the steam realize there's a new update?
or is it not client side?
Wouldn't work for me, anyway
I have to use wireless charging 💀
Which is also really hard on the battery because it's less efficient and insulates it
But 'tis the reality of a phone that is five years old
I like the bypass because it is less heat when existing, gives it a future once the battery becomes 
i think charging the battery causes a decent amount of the heat the phone produces.s o bypass means less thermal throttle im pretty sure
I believe my device is also able to trottle charging for that reason, so it is a slower charge to not add more heat than needed
That is just intergrated into the gaming mode though and not the standard usage
new changes
but alas I still am not at the point where I claim rooms
more microtransations
will not be relevant quite yet
shiro
my enemies may microtransaction all they want, my algorithms always win 
you really wanna be near the person to my north?

Battle to the death 
no actually
once the novice wall goes down we're all cooked
person to our north will fuck us so bad
my sacrifice will help you two prepair
Also bypassing the battery means you're not straining it as much by just not using it
Why is budget phone have more actual usefu feature tho 
depends on the manufacturer i think
if samsung adds a feature to their A series, it will be cuz they added it to the S series the previous year or so.
However, I still envy the ability to have usb-c video out with mouse and keyboard input 
i also have wireless charging now, altho i will litteraly never use that
Yeah, wireless charging is just a scam IMHO
Specifically on phone
Or any device larger than a watch
even ona watch its a scam
they could jsut have magsafe-esque thingy with 2 copper contacts on the inside
Ehhh, I wouldn't say that. There's a valid reason to have watertight watch for scuba
when are they gonna continue ticking shard3 i want my roads to be built
ye, shard 3 seems to be frozen ye
Also, I would assume no sane person would use their phone while diving 
i don't think having 2 copper contacts would make the watch not watertight, especially when they already have metal contacts for dev port stuff
The corrosion is the issue. But yeah, it could still be done I think. Just not rated for very deep
all i know is that one of the apple watches has some scuba diving certification, presumably that one still does wirelessly charge so idk what they do
40m max 
Yeah, as I said, wireless charging is useful for watch device because they can be certified for scuba and quite deep one at that
ye
i don't think i've ever used wireless charging anywhere other than a car
do be convenient
Wait, where did you put your phone on your car?
im misremembering, my current galaxy watch 4 doesnt have a debug port. my old galaxy watch active did. and its behind a litle door so that stops the corrosion i guess.
Wired discharging?
phone acts as wireless charger
to charge earbuds or whatever im geussing

Yè, but that little door can only hold the water pressure not as deep as actual sealant
in the middle console the whole space is a wireless charger 
throw it in there and connect wirelessly to the car while it's slowly charging
do they somehow equalize the charge or is there some sort of role negotiation like in usb-c? i dont know how wireless charging works
Interesting. Most people in ID usually put their phone on a mount to aid with GPS too without having to install screen on the console
you could always charge another phone via a usb-c cable, should be more efficient
its not smart enough for that
and equalising is almost never a concern because its so inefficient
you might lose battery in a 5:1 ratio
I don't know how much of a headache it would be to have this feature
this is a pretty new bmw tbf, for an older car having a mount is really nice though yeah if there's no built-in option already (and even when there is
)
most modern phones can
that exists, it has for a long time
its not
do you get a popup notification asking you which way you want the charge to go?
as well as sharing multimedia
no

you turn on reverse wireless charging on a phone
my 10200mah tablet is basically a 230 euro powerbank that can play bloons tower defense 5
Nah, it's cursed IMHO. Your phone would need to be able to supply at least 3A maybe to be even a viable charger
iirc most phones just do 5W
i was asking about usb c 
depends on phone
See, that's just cursed for me. Unless it is the usual usb device spec not usb-c negotiated stuff
But even then...
i remember charging my tablet using my phone with a microusb otg + regular usb-a to microusb cable
How does the phone knows who is charging who 
worse than c-to-c would be micro-to-micro
and that was very unambiguous
it firstly just assumes which side you want to charge, but the "host" does get a notif to make the other device the host, and if you want to enable data transfer
Wireless data transfer via charging coil 
no thanks
Why even have dedicated wireless antenna
we're talking about the usb-c stuff
its called nfc 
of ports
it supports all kinds of features but only if you use the right cable 
usb-c is nice cuz it works in reverse, and it works upside down, and it works for data, and it works for video, and it works for a lot more.
depends on if you bought a decent cable tho
NFC have actual use like device without a power supply
But a charging coil...
no
man, after having used an s22 ultra for a couple years a 6.3" screen feels so smol 
I was jsut continuing my sentence 
ye, they firstly give some power to the other device, and then they listen for feedback. its how the nes lego smart brick works
The dilemma of digital age. Want to be portable but want to have large screen real estate
6.3" is small
to me
some people have small hands so you know
Indeed. That's why I don't think NFC are stupid as transferring data with charging coil 
i tried to save it, but it only sounds worse now out of context
i really like the inner screen though so i'm a bit conflicted
(this is a pixel 9 pro fold)
the whole point of a fold is to fold it open no?
not every app plays nicely with the extra space + i'm lazy + battery life
i can technically fit a steam deck into my pants, doesn't mean i want to do that 
weak
i mean with normal aspect ratio, i think the folds are too wide even-tho they're only 8"
unfolded they're all pretty much square i think? (with a few exceptions)
who needs the foldable to fold anyway 
jsut get a tablet 
Who even needs a tablet, just carry around a basic laptop
Sam, why did you always word it like that 
its not intentional 😭
the english language is just a terrible language made by perverts
↩ (@OfficialLoganK)
Huge improvements over Ironwood TPU’s (2-3x), can’t wait to see Gemini in action with these : )
if only this was all it took to fix gemini
🙏 more bandwidth for free users
I think the post means it more as google saving costs on GPU
in an ideal world they would stop messing around with gemini and casually drop a model that easily beats everything else due to sheer scale but that's copium
EB sized model 
they've been making TPUs for a decade
they clearly figured something out otherwise they wouldn't be able to give out 1M context gemini flash for pennies I feel like
whatever happened to the 10M context window models they made
no, they made newer TPU chips which are 3x more efficient than the chip they're running on rn
higher cost and it likely doesn't work as well with the newer models
they did have some research project 'titans' or smth like that on models with near infinite context
not sure whatever happened to that, seems to be still in progress
we're back
they do have some incredibly talented people at deepmind
+ unlimited compute
Also TIL the estimate suggest we're still in low hundreds of ZB range

i'd like to know how much of that is google
i just noticed, what is going on here
Oh, the usual baking stuff
don't ask me, i still don't know how this started
as ussual, konii
I lied. It was the total capacity
does anyone happen to know a provider of b200/b300s with performance counter access 
I'm torn between CDN and google but then I remembered youtube 
Gemini is great for everything that isn't coding
true
so much new stuff wtf
guess i'm a bit biased there 
Of course wikipedia has an article about it
https://en.wikipedia.org/wiki/Zettabyte_Era
The Zettabyte Era or Zettabyte Zone is a period of human and computer science history that started in the mid-2010s. The precise starting date depends on whether it is defined as when global IP traffic first exceeded one zettabyte, which happened in 2016, or when the amount of digital data in the world first exceeded a zettabyte, which happened ...
oh yeah, its thursday after all
Hmmm. I wonder how fast we are going to the next era
Maybe the new moore's law is the storage era measure
wouldn't moore's law technically apply to storage too 
i guess it's a bit different
since nand is so simple in comparison so you just stack it like crazy
Does anyone still use copilot student plan? After they removed claude pro models its kind of bad. Haiku makes so many mistakes and doesnt read all the context brah
just use codex
I learned a new word today. Lügenpresse 
2-bit Qwen3.6-27B GGUF made 26 tool calls, triaged 15 GitHub issues and fixed, tested + reproed our repo’s 3 latest issues. 🔥
︀︀
︀︀Try this locally in Unsloth Studio with just 12GB RAM. Studio also has a new look!
︀︀
︀︀GitHub: github.com/unslothai/unsloth
ok this is interesting
honestly, the moment open source models below 27B reach gemini 3.1 flash level, ima acc stop using provider models
I think a proper state of space would need 3 bit at minimum.
±1 is off, ±infinity is in another dimension entirely, 0 is close and nan for wtf
2 bit 😭
Bro this shit aint working😭
but it is working
At least they have ±1, 0 and nan 
nan overrated 
I will give you division by 0 
Mozilla is experimenting with Brave's adblock engine, here's how to turn it on.
TIL about dbzer0. I learned a lot of things today 
#define EPS (1e-9)
I think Tom7 (suckerpinch in youtube) has read a paper analyzing open source program across github. There are so many erroneous use of epsilon like -10^9

What languages do you guys code with?
Java
I like Java
Especially 25
Otherwise for low level C(++)
Basically C style C but using nice C++ things like templates
How long have you used Java?
I was gonna learn java but didnt need it after all sob
But i havent heard anything negative about it (personally)
How long ? Since i started modding minecraft long ago with Java 8
But how long i actually do good shit ?
Few years
Idk maybe lemme check i need to remember what year we in and what's my age
So true
I guess like 5 years ago maybe i started getting more serious
Like 3 years ago i started actually having the good ability
Maybe 1y ago i'd say i've really been doing good stuff
And i still have lot to learn
Let me introduce you to spring framework
How does minecraft modding work? What programs do you use to make mods?
You just use Fabric
Which lets you use mixins
I dont think i want to..
And now you've heard the bad stuff 
lol
Mixins is what allow you to actually hook into the code
It's actually built based on ASM
hibernate 
Oh ok
You're just doing high level bytecode
Thats cool
But that's like hidden
Once i stop eating i can show basic stuff
It's actually not that hard to understand
I could explain it to my mom and to my grandfather so
And they actually looked like they understood kinda
Why am i giving weird signs while trying to make sure to not do
Lol
I can understand most things
Ive been programming since i was 11 years old like smol bebe, so I hope so
If you have programming concepts then it should be good
Also knowledge about like the low level
Bytecode to Java is like assembly to C kinda
Hmm okay 🤔
Huh i wonder how that compares to ublock origin
I kinda like having control over it via the menu bar tho
from other server
My computer died. I tried installing a custom Ubuntu BIOS and the hard drive died. I'm almost certain I partitioned the Memory disk disk incorrectly, or maybe I installed the BIOS wrong.

what is bro on about?
His computer died. He tried installing a custom Ubuntu BIOS and the hard drive died. He is almost certain he partitioned the Memory disk disk incorrectly, or maybe that he installed the BIOS wrong.
Qwen3.6 35B A3B, on a Ryzen 7 5700G + Vulkan iGPU running llamacpp in a LXC, 16k context, im getting 11t/s. Not bad, but not amazing. It does not fit all into "v"ram though so thats somewhat expected.. any ideas how to speed it up?
Damn i need some of that stuff it seems to be good
maybe a teeny tiny bit lower overhead cause no extension but other than that there's no real reason to use it over ublock i'd imagine (at least for now, maybe they do some things in the future
)
they use the same lists after all
there is no easter bunny, there is no tooth fairy, and there is no ubuntu BIOS
Yea thats what i was wondering
Maybe it's not worth it yet
half the reason i use ublock is for the element picker tbh
so many small annoyances to zap out of existence
also i want my bios to run OpenBSD or NetBSD instead please
i wonder how the llvm on uefi thing is going now hmm
Obvious answer is more ram/vram and higher clock while lower CL.. but meant software wise.. i cant seem to turn off mmap so those 3t/s more aren't gonna happen
i almost never use that feature lol
but i recently found out that i can configure ublock and bitwarden firefox extensions via nix home-manager lol
nice if you fully bought into home-manager i guess
yea i use it fully
its great tbh
i also configure my KDE via plasma-manager and pretty much everything else with it
plasma-manager my beloved

yea 
Plasma manager? What's that?
maps nix options to the various kde plasma config files
Ohh, so nix thing
ye
is also nice as a reference to find out where tf a specific setting is saved
why is downloading songs on spotify such a broken thing
I even made a custom nix hm module for plasma-manager to set some options and predefine other stuff
i feel like they have some linear algorithms here that choke at 10k+ songs 
It was either that or bro just has a plasma cannon written up for the funny
New conceal carry goes crazy ig
Pulled from an AI cluster by [SELLER], to be repurposed for an AI cluster by [ME].
Ah, I see
You got it from those folks at horus then
Flirt with paracausality responsibly
Found an issue with llamacpp's auto-fit.. trying load the model + mmproj file causes it to budget for just the model, thus mmproj fails
Horus?
There was a option to keep some ram open iirc, what was that called again?
Those sure are hardware specs, completely wasted on my layman ass
But the numbers are big so I'll just say Impressive
They're pretty good, yeah.
What is "horus"? And why does "paracausality" relate to that?
Oh
Yeah don't mind me, if you see me in here just assume I'm the least informed person. You'll likely be correct.
here's an example, my cursor indicate the injection point
the Mixin annotation on my class indicate which class i want to inject into
the wrapoperation annotation on the method is basically saying this method is gonna wrap that operation
meaning that when the operation is gonna be called, mine is gonna be called and be a wrapper around
basically like parenthesis around something, letting me execute code before and after the operation, being able to change the input and output
the method of the wrapoperation is init, which match the bytecode as i'm targetting the initializer of the class
then my target is an INVOKE call (INVOKEVIRTUAL in this case) of the method checkeulaagreement of the eulareader class which return a boolean (the Z indicate boolean return)
then my method take the instance of the class and the operation as parameters
and from there i'm able to do custom code, in my case i return a boolean value from my config or the return of the call of the method on the class
basically it's like if i turned the code into this (except there's some more function calls)
this.eulaAgreedTo = SharedConstants.isDevelopment || (Config.Server.bypassEula || this.checkEulaAgreement());
this is quite an easy example, but it can get more complicated
Found solution: --no-mmproj-offload
i put the 13400 in the new mobo
the stock cooler required me to change the socket lock thingy backplate
fucking stressful to do that
lets see if i broke it
here's another very similar example which should be understandable if you read the first one
wrapoperation kinda carrying tbh
-# also sadly you cannot mixin into Java directly otherwise what i'm doing here would be simpler
Can NOT recommend this! Its painfully slow. Like its beem going at an image for more than 3 minutes
NVIDIA trying to save their stock as soon as Google announces their new TPUs lmao
i hope RAM and VRAM and SSD will go back down
HDD didn't spike as high
doesnt TPU just use more HBM?
iirc the stats from before showed higher HBM per thingy
hopefully they will need less TPUs than GPUs so it will even out or reduce
total HBM capacity per pod went up


kind of risky but i guess even if it had failed it's google so

whereas nvidia can just cut tensor cores and pretend this never happened if they wanted to
they could also keep tensor cuz its kinda cool for certain stuff
dlss is not that cool, but fast matrix math can be used for other things

I'm honestly incredulous that it took them till tpu v4 to split it into inference vs training (via shittier ones like v4i being designated inference but really just being stripped v4) and then till v8 to actually really aggressively split it from the beginning like this with purpose
the real question is are TPUs cheap enough to be worth it for consumers
or if GPUs will stay as the main pick
they're way too tightly integrated to be run by anyone that isn't google lol
im pretty sure tpu is dirt cheap comparetivly to gpu space
Jax
yeah that's what I feared
Tpu*
like ideally it would be an optional module supported by some mobos
I ran a fine tune on tpu like 2 years ago
That's coral
did you see they supposedly have proper pytorch support now
it's dog
I didn't see that yet
nah. just make it a pcie device
Haven't gotten to read the announcement but I saw a picture of sundar holding the split infer/training versions
Having mediatek and broadcom separately do the chips is good
Less central
the engineering blog is a nice read
I actually used Jax briefly for training on my mini pc
Holy shit was it slow
Worked better than rocm which is to say
At all
It was interesting
I couldn't do anything except think of Jax from league of legends the whole time
If they do they ain't sharing unless you're on Google cloud infra or something cuz only ever see openxla
looks like they do
i get it but 
at least go the nvidia route and just ship binary blobs man
works too
this is why you should install a faraday cage
the pcie h(100|200) is basically a tpu 
chat what's the state of gaussian splatting or similar tech
I haven't been following over the years so idk if it's good yet or still just for-fun
I feel like b300 meets that description better
no pcie
Gaussian splatting is just pointcloud. no real change in how it works
There's a water cooled one with an Italian name which makes it basically a Ferrari
its both. it is good but it is also just for fun. no real practical usecase
why is no one using it? performance?
among other things yes
Irl point cloud reference 
97k* but yes
i hate that i was able to extrapolate that
Lmao
Best part is the sillycon probably doesn't cost much more than like 5090
I mean, it's more surface area and shit and mem
But like
Is that a difference of 20x
Hell fucking no
huh
isn't it just a more efficient voxel grid?
should be easy to optimise long distance with some kind of tree data structure
It is not a more efficient voxel grid 
you pay for the r&d 
Surely
stores less data than microvoxels no?
in terms of games Gaussian splatting is bad cuz you're wasting 50 years of optimized pipelines using textures and the ability to reuse them.
and in terms of things other than games i have no clue what the industry is like
The r&d of 'lets cripple fp64 to 0'
It does however, allow you to reconstruct scene without having to gather a bunch of accurate point cloud
Took extra research to figure out how to ratfuck the thing for all compute
kinda
tis very hard work
I like how when we're all typing it's just 4000 toast
idk about you, I can tell which toast is typing
its cuz im a triangle
That's easy to tell when only 1 or 2 is typing
When 3 is typing, all bets are off
Especially when all 7 toast are present
I mean, it is able to do a ton of interpolation no?
yes and no
with interpolation, do you mean connecting the points and guessing the colours inbetween?
besides that there is the potential of having meshes with volume and procedural/AI generation of realistic 3D environments
with proper optimisation it could really bang
Yes
Ye its not a mesh
problem is you cant with 100% accuracy say which points should be connected
Hence why I said it performs a lot of interpolation. Including guessing where the connected point is 
not really, you're still jsut getting volumeless points, liek the vertices in a mesh. but now there is no data on how they're connected
you can't see the vision
Text-to-3D using Gaussian Splatting: Applies 3D Gaussian splatting to text-to-3D generation.[12]

so earlier someone said that there was this guy that developed AI and was fired "because he got AI psychosis" and turns out a more likely reason for his firing was he leaked company secrets
I was like "yeah that sounds like a strange reason for his firing" and this logic makes more sense
getting ai psychosis is basically on the job requirements these days
i think people call gaussian splatting volumetric cuz you can change the angle of an image of you have ample depth data froma n existing image.
this is more optemized than 3D scanned meshes
its still terribly optemized compared to actual meshes and textures
unless you're a UE5 dev
the goal for some of these people is to make a real AI gf, imho nothing wrong with a little mad science as long as nobody gets hurt
Gaussian Splatting is volumetric in the sense that they do render volume (like, what is volumetric rendering if not rendering of a bunch of sampled point in 3d space)
if you're using UE5 the gaussian splatting would also perform terribly.
but nah UE5 performs bad cuz of the software hiding a lot of shit in menus that you need to change for optimizing, and negligent devs not putting effort into finding those options
That sampled point now comes from actual sample instead of a probability density function
UE5 was marketed in a way that encourages bad optimisation
it was marketed as easy imo
which would imply the software would take care of optemizations
which it only partially does
no they straight up marketed nanite as "stop optimising triangle counts, it will do it for you"
which isn't true, it still runs shit with badly optimised meshes
also depending on mesh shape it ends up looking shit too
Less mad science, more weird science
She blinded me, with science!
and ue5 in general is slow as hell
like man, I was a believer
switched from ue4 expecting an upgrade
ended up moving to unity lmao
it does look cool, but its not really something id be willing to use unless thats my intentional artstyle
Overlay anime character on top with text dialogue and you've got your game 
what a steal
while it looks realistic, i feel like a good 3D modeller with some PBR could get something that looks close to that level of realism, while being optemized better
I think I've had time traveled in discord. When the clock strikes 00.00 and someone and I sent a message, the chronological order is somehow 23.59, 00.00, 23.59
Of course it would be because you have more information than just a gazillion of point 

my point is mainly procedural generation
if you can get a nice recursive system, maybe use some AI to fill in blanks and add details
and then properly optimise the tech because obviously just rendering it raw will run like shit
then it would go hard
Which reminds me, reversing a graph was a topic in numberphile that I need to watch
procedural mesh gen is not even close
Reversing the graph = reconstruct the whole graph vertex and edge from just a few vertex and edge
Elise Raphael takes us on a journey through graph reconstruction. Extras from this interview - https://youtu.be/6AqA2nIHtxg
More links & stuff in full description below ↓↓↓
Elise Raphael - https://www.unige.ch/math/la-section/enseignants-et-chercheurs/elise-raphael
More graph videos...
Subcubic graph numbers - https://youtu.be/4-eXjTH6Mq...
Pretty sweet
i see where you're going, but even with AI you're never gonna make a mesh that is 100% consistent with attempting to fix the holes in your data.
@opaque sigil your EB model milord 
AI is better at gaussian splatting than making traditional meshes tho
I haven't watched the video, my hypothesis is reconstructing a graph is impossible 
didnt watch before sending
i skimmed through it, i think the answer was you have mnultiple thinges so there is no definite 100% correct solution
"we dont know cuz otherwise we would have solved the congecture by the negative"
hmmm
Unsolved problem in mathematics
Congrats Sam, submit your proof 
no u
yeah raw splatting is not the way, tech hasn't progressed at all in recent years, researchers are slacking crazy
Splat
not toast, opinion rejected
Butter side down opinion
i mean, gaussian splatting is the same except you get ellipsoids instead of disks, and smooth blending. and maybe some change of colour depending on the angle you view it form depending on how advanced thee software you used is
gaussian splatting has some natural interpolation due to the overlap and blending betweent hem, but that stops being effective at a certian point
well by raw I meant any splatting
nah that's not even the main performance loss I would assume
the original splatting method is from the 90s
it performs decently
but it looks ass
oh?
tf are they doing to make gaussian so slow then
blending and stuff
they have to be doing something wrong 100%
If this is still the bit in a week then I’ll join you
you underestimate how expensive blending is
I can see 2 simple approaches
of either sampling per pixel with precalculated screen space positions and depth
or depth filtering then drawing onto one texture
they apparently use gaussian splatting for making 3D copies of cities for like real estate planning and construction projects.
forensics too so they have a copy of the crime-scene
also autonomous vehicles iirc
and FVX???? i think???
pair that with a tree structure based on depth
to basically combine points into bigger points at a distance where you're not seeing much detail anyway
ok but like, thats expensive
it might look cheap if you think about it seperatly
but you're doing that 2 million times, once per pixel for a 1080p image
it's not 2 million what
so each operation adds up
1920x1080
you said per pixel
2 million is nothing for a gpu
yes and no
if you optimise it well it's all simple operations
for normal splatting you can get away with using a Z-buffer and jsut using the closest splat
for gaussian splatting you're doing blending
so you're always using every splat
what
you know, transparancy
screen space positions filtering
most of them won't cover many pixels unless you're close
yes and no.
you're misunderstanding how this works
I might mess around with it
not a high prio project tho so not anytime soon
so like legit question, what exactly is ue5 doing that is slowing it down compared to ue4? unreal slop seems to perform really badly and I want to know how to best optimize it
I hope you're not trying to traverse a tree on the gpu 
there is a lot of small slowdowns that combine into a big slowdown
for example
shit, that's an answer I wish I didn't get, that means it's hard to fix because there's many problems rather than just one
idk what game this is, one of the newer space games using ue5. you have a mask to do certain effect, and the devs didnt set it to reuse mode, so it re-rendered the same mask 4 times per frame
How is everyone
it's not
I wouldn't be able to charge my phone without it, the port just doesn't work anymore
no option to replace the port?
so ue5 is just plagued with bloatware that is unnecessary for it to function
I don't want to 
and ue5 isn't given the most basic of optimizations either wtf 
that might of been good to do at first.. but now I just need a new phone and don't feel like it
so just continue wireless charging.. at least should of gotten a fast charging compatible wireless charger, oh well
it works
idk if you could call it bloat, they're "features"
UE5 does have the option, its somewhere in the menu they have, somewhere....
devs didnt enable it tho
it's bloat if it has to stay in the executable regardless of it being used or not, if it gets trimmed out then it's not bloat
why does epic have all of the slow features on by default, rather than off, and the fast features off by default, rather than on? nvm ik why - laziness
damn shame because like
Nanite, Lumen, TSR. UE5 stacked a bunch of crap together causing it to run like crap
I wanted to make a fan game of epic battle fantasy but in 3d using unreal as an overly complex joke
yeah so as I said, bloatware
Guys so an update on the Link 3D model.
The files are too old, they’re from 2014, and Blender can’t properly work on them.
I guess this is the end of the road
its quick, its easy, and its free until you get 1million USD, using unreal engine and producing slop
I mean hey it probably will perform as good as ebf did during the peak of flash
2014 is nto even that bad, wdym blender cant work on them?
what are you trying to do that didnt work?
(not very because of the particle fx)
ue devs like to cope but most successful games in indie and AA are made in unity
and most AAA is made with either in-house or other licensed engines
i can say mean things about unity too, but they deserve it less and are mostly partially better
Lemme show you the image, I found someone on VRC that was able to help me but they didn’t trust the source, gut feeling
I honestly should reinstall unity because there are games I want to mod that are made with it - P06, Hyperdimension Neptunia SvS, Terra Invicta...
As a indie game dev I also use Unity lmao 😭
I don’t know how to work on Blender, I’m a newbie
the model you downloaded was a blend file???? not an fbx or a gltf????
unity is a great engine held back by its devs making weird choices
it's still mainly a do-everything-yourself engine tho so it works just fine
but if I wanted to make something new with Unity I would consider Godot, GameMaker, or RPG Maker first because of the install fee fiasco LMAO
ik they walked it back but like, I'm not rewarding a company that treats its customers like that
This is a different one, it’s not the otr file, it’s one I found on deviantart alongside a Zelda based on the same game too
the drama is over, no point switching to a worse engine
Here it is
nvm switching to Scratch
if i wanted to make something id be very proud of i would use vulkan
if i wanted to make something i would be mostly proud of and i want it to not take 5 years, id use unity
if i only cared about upholding standarts of high graphics ultra hd 4K raytracing, id use ue5
thats a png
Yes but that’s the model, it’s not the other one where it was a mod
Yea that’s why I use unity cause my games are kinda retro style horror graphics
ok but like, what file type is it?
surely its not a .blend file
ue5 graphics don't even look good tbh
It is blender
they objectivly look good if the devs put effort into it, but at what cost?
See?
thoughts on that threat interactive channel?
I mainly watch it for fun but they do have a point about shit like ghosting and smeary/blurry looking graphics in ue5 games due to abuse of ai features
the guy is mostly right, but i do find him a bit sus with at the end of his videos he just starts promoting his own shit
legs.png
FBX has been around since 1996 🥀
thats like older than all of us right?
nah literally no one im existence was born before 1997
so yeah, i wasnt able to load any of the textures
which is weird
lemme see if i can do it with another link. the same artist did another one
okay its from 2019, theres hope
...what the fuck is xps?
you gotta be fucking kidding me, bro "used" 25 32x32 textures. this model is just ripped directly from the N64, no way he made this
it's allegedly an organisation and not a one man thing
so like it makes some sense
personally I don't do any active support but don't see it being an issue, would rather people spend money on that than the common consumer exploitation
the best part is that compared to the previous one we can at least see the textures and the model
the problem is that is ancient
if it is just a scheme
at least there is some knowledge given to people that get exploited lol
and paying is still optional so
i think he criticizes the engine for stuff that you can do with the engine if you want to.
which im not sure is placed correctly.
on one side you can blame the engine for even allowing it, on other side you could argue that its the devs that shouldnt use it
the problem is mainly the industry being increasingly bigger budgets, bigger risks, bigger teams, bigger games, while trying to keep short development times.
meanign they cut corners on the dev where they can
I mean, people have gotten so complacent we get dogshit optimisation in all consumer software industries
and if an engine not only performs bad but also looks way worse than it could
then yeah it's up to engine devs to work on that, not expect game devs to triple their efforts to fix the tool first
i feel like unoptimized slop was always gonan be the direction we would go in
ever since people stepped away from doing the assembly programming themselves
I prefer low level programming but god damn that's harsh
I'm glad AAA is in decline tho
can't expect people to recognise a bad product when the degradation was a long process
yet a lot of gamers do gravitate towards smaller productions nowadays
the compiler wont always make 100% optimized code, and with that we got more and more complacent each year. and since the industry profits from having to spend less time on optimization, they push it as far as they can
ngl that is true
at bcit we are told that optimization should only happen after the features are already in place
not entirely true
it's more that hardware got so good they stopped having to microoptimise everything
and when there are corners to cut you can expect for-profit organisations to do that
so optimization in the computing industry has taken a backseat for a long time by now
im just glad people stopped taking this shit and the mainstream gamers also want optimization isntead of the stuff you saw years ago about pc masterrace big gpus
in favor of bloatware (re: my previous comment)
I landed in a really good company man
I'm getting shit on in a code review if I accidentally copy a variable regardless of its size lol
in my opinion hardware getting better is just a given. altho i agree that is also a big factor
well, I prefer having machines I can upgrade instead of having fixed spec consoles, but I do get that the other side of the coin needs to happen more often
oh my capstone project is almost all about optimization, I would probably fit in well there
the goal is putting ML-KEM and ML-DSA on IoT devices and/or microcontrollers
it's the main factor
optimisation was a necessity, now it's a standard no one will hold you accountable for
because business optimisation wants less time and more product and optimisation and designs/planning take time
it's also why corporations are coping on AI without it having a proven record of genuine code improvement
what i liked about pc is that its supposed to be a workhorse that is 2 or 3 times faster than the current console. but devs stopped making games optemized enough for consoles to run them well.
its not jsut pc getting higher perfromance setting tiers
tbh like
I personally am upset that coding ever became more than just a hobby, but thats just because im a bit extreme on the unhappyness of the current state of IT
a GOOD PC game (re: technical achievement) is something that actually uses the resources efficiently, ie Terra Invicta, which simulates a whole ass solar system as a very important part to the 4X gameplay
coding hasn't been a hobby thing for like most of its existence because personal computers weren't a thing for so long
Hobby/Research then. And one has to consider that i (and my father) have grown up in a world where "(IBM) PC" didnt exist yet, but programmable machines did until they got mostly replaced by CBM's and such
i think for games, the lack of optimization happened later than for less heavy applications.
with games being mostly single threaded back then and still decently single threaded these days, it was much more of a necesity for games.
also with shadercode each extra operation costs fps
that reminds me, we need more people making hardware ie GPUs/RAM/CPUs...
in terms of being fully optemized, i think rollercoaster tycoon takes the cake
ok chat I need to wake up in 4 hours so uhhh
gn
oh obviously, Terra Invicta uses a third party engine (Unity)

RCT is fully coded in ASM
Or transport tycoon. (Same dev, awesome game, always overlooked)
gaussian splatting 
I'll look it up in may I guess, I should be working on a project for a course...
there technically exists something that might be simialir to what you mean. mobile gpus use tiled rasterization
desktop gpus also have a settign for that iirc, but its used less
since gaussian has transaprency, you need to do this thing blended ontop of eachother for each pixel
djd you ever get the volumetric display workign
i killed the pi accidentally, and im investing in safety of my hardware rn
if i kill the projector i will cry cuz that thing costs 1K
the pi was only 270, which is a lot, but not death of the project levels

how
voltage go brrr
we dont know, there are 3 ideas
we're pretty sure its caused by 5v going into a 3.3v tho
you need to do one of these for each pixel
so that adds up to a lot of operations
even if you do tiled rendering where you only include the gaussians that could have an effect on your image into the raster
you're still doing a lot of +=
if you do it with classic raster you're overwriting your framebuffer data a lot
I just frew up
itsd in math language
in codee its less bad
still a lot of opps for gpu to do
but less bad
#version 450
layout(local_size_x = 16, local_size_y = 16) in;
struct Gaussian
{
vec2 pos;
vec3 color;
float opacity;
vec3 conics;
};
layout(std430, binding = 0) readonly buffer GaussianBuffer
{
Gaussian g_list[];
};
layout(rgba32f, binding = 1) uniform writeonly image2D frameBuffer;
layout(push_constant) uniform PushConstants
{
uint total_gaussians;
vec2 screen_dims;
};
shared uint tile_indices[1024];
shared uint tile_count;
void main()
{
ivec2 pixel_coords = ivec2(gl_GlobalInvocationID.xy);
if (pixel_coords.x >= screen_dims.x || pixel_coords.y >= screen_dims.y) return;
vec2 p = vec2(pixel_coords) + 0.5;
float transmittance = 1.0;
vec3 accumulated_color = vec3(0.0);
for (uint i = 0; i < total_gaussians; i++)
{
Gaussian g = g_list[i];
vec2 d = p - g.pos;
float power = -0.5 * (g.conics.x * d.x * d.x + g.conics.z * d.y * d.y) + g.conics.y * d.x * d.y;
if (power > 0.0) continue;
float alpha = g.opacity * exp(power);
if (alpha < 1.0/255.0) continue;
accumulated_color += g.color * alpha * transmittance;
transmittance *= (1.0 - alpha);
if (transmittance < 0.0001)
{
transmittance = 0.0;
break;
}
}
imageStore(frameBuffer, pixel_coords, vec4(accumulated_color, 1.0 - transmittance));
}
something along these lines
compute shader cuz easier to do tiled stuff
this still misses the precompute stuff you need to calcualte the jacobian thingy
and you hsoudl also sort the gaussian list bwforehand
and this code is still terribly optimized, it would probably crash your pc if the scene had a million gaussians
Jacob seems cool
Makes ai stuff work too

no
Again?
I shoved this into visual studio, compiled and ran it, and now my PC is trying to turn itself into a car.
Terrifying.
Not anymore
visual studio literally wont let you run that
Update, my PC ran me over, please contact the Trauma Team
Glsl on CPU is still glsl 
bro is cpaping
Update 1: my pc is getting louder
Update 2: it’s gone
no, im mentioning the first and only time
"bro is cpaping"
true, I'm on life support after the tragic accident
Ohh, I thought you already buy a new one
im trying to have some money left over
Trauma Team is occupied trying to stop a rampaging cyberpsycho who installed a DOOM mod to his left arm
Be me with 1 USD in my account 😎
guys, thats gpu code. glsl; OpenGL Shading Language
visual studio wont run that. unless you do some weird shenanigans where you map glsl functions to c++ functions
besides that its no
Understandable, guess I'll die
Guy can anyone check my ai code in projects. I used gemini to crossreference it.
bullshit, all the companies are just realizing they can hold supply for price to get high and in hope the ai companies buys them instead of consumers
This isn't a company press release, its Nikkei
I trust them to verify their reporting
They're generally quite good
its some general marketing report
i don't blame them
but i blame the rest of the industry for directly causing it to happen.
the supply can easily meet the demand but greed just doesn't allow it i guess.
really tired of the "dram demand is increasing because of ai"
no, fuck no. supply IS the one decreasing due to the unbalanced allocation of the said supply
i hate the lack of competition in the semiconductor space
all the companies sell less to consumer and hold all their supplies in hope that the ai industries buy most of the supply
its just batshit insane
these companies made "ai industries" as an excuse for that
That's what you get for ASML monopoly
And the technology must be protected at all cost lest we want to get cheap semiconductor
Also, why is my phone autocorrect this bad
Brooo
@stark needle
Not again
Shadow you’ll never believe it
Shadow he invented agi
Without changing model weights
Anyone got summary of what kinda slop it is this time?
I wonder if the ai industry knew that all that stood between them and sentiment models was 5kb prompt
It’s a text file
That’s it
Pure ai psychosis
So probably just asked ChatGPT "Invent sentient AI"
It’s like
Half way between a prompt





