#programming
1 messages · Page 465 of 1
Depends mostly on if you need any apps that are windows on´y
Adobe, office, or games with easy anticheat are the big offenders
I mainly am griping cas I got a specific script Idk if I can be bothered to rewrite
cas I run a coding course
and I host that shit myself since where I host the infra is shit
Nixos is the natural next step up from arch. I think
mostly
for me rn its the effort am I awake enough to do it lol
tbf good thing abt it is that Its hard to fuck up because I am just copying configs
from laptop
pc
fuck
fml
I cant type
laptop to pc
there we go
You could have just eddited the message 
ye ik, my brain is not braining. I have been awke 18 hours barely 2 hours of sleep
was not the most meta experience in the world as I think you could imagine, I forgot I had stuff which required me to be alive early hence me not sleeping till 4 when I have to be up at 6
😭
Dont bother googeling this nonsense, its a rabbithole
I am passively aware of it

tbf in all reality it would be as simple as me plugging in my USB drive, running the installer , rebboting pc. Git clone , run script reboot
Straight back to pc stuff 
we do not worry abt it
Do I feel like rewriting multiple scripts tonight
tbf I dont have to tonight
I wont need till either later this week or , next week so technicalkly I have like a solid 4 - 7 days,

tbf in all logic I should probably just do it to make my life easier so I don't keep trying to do nix stuff on my pc and arch stuff on my laptop when brain decides to disable functioanlity
Kinda yea.
Next and probably last step
Just use a simple bash wrapper that directly installs your flake from GitHub and formats the drive 
Ming who?
I basically have that tbf
Its just not straight from intsaller
I have to boot but after that it does it
cas it installs hypr and stuff aswell
there is one of two correct paths after arch.
the first is gentoo, if you are a software pureist.
the second is nixos, if you want even more control
if you tell me you switched to LFS instead, I'm calling the police.
oh damn, I have a simple tokio + reqwest program deserializing json from rest endpoints, decided to try out smol + surf and it's like 50% faster, somehow?
to be fair, it just sends a couple of requests per call, it's not high throughput
but still
ain't nobody got time for emerge -uvDNa @world
I don't use gentoo, but I'm guessing this downloads and builds every package in existence?
it rebuilds everything that needs to get rebuilt
Basically what you need for updating your gentoo install.

I might be skipping LFS after moving to Arctix.. whoops
Might just go in without a handbook.. what could go wrong
Tbf id love to write some minimal linkers but i know Jack shit about toolchain work.
Imagine a ld where you can keep multiple versions of idk, any lib, lets say openssl, and force any app to use any other version simply with some config files-
If projects follow propper versioning, it would even allow for partial automatic updates i think
toast
Toast
is 8 network interfaces too many interfaces?
nah
my server currently has 10
❯ ifconfig -l
em0 em1 lo0 bastille0 em0.20 em0.99 vm-iot vm-home tap0 tap1
Install docker and random other junk that makes up more interfaces
😭
I just install docker in a different namespace to avoid that


I have returned from the ventures of installing nixos and for some reason me having to rewrite my hyprland display config 8 times because it didnt like it
I had stuff to do tongiht
I need to do that still
I am now not going to get to sleep early like I wanted fml
Have you seen the caveman thingy?
looks like normal youtube to me ?
it's a joke
mhm
Funnily enough, the embed isn't showing up for me
So I decided that I hate myself, and I implemented a brainfuck interpreter in raw assembly
but hey, it works 
Are you ok?
no
writing a program that is too big will crash the code apparently
the worst part
it's faster than the bf interpreter I wrote in C even when compiled with -O3
bf interpreters are boring
make a bf compiler
idiom recognition, jit, optimization passes
surely it will be fun 
Then write a C-to-bf compiler to see jit power 
@maiden geyser hi
AI Disclosure alert on the paper, claims are real but the code might not be up to par
At least if it is to be believed
I'm reserved in my opinion so far
I'll take a better look at it, after I finish eating.
But it sounds like the kinda thing I'd be looking for
@real sierra kill

SKULLGE_SUPREME
Let me know what you find, I'm also interested in what it is doing
who did we execute
Create AI app here with link
straight to the AI black hole for them
my ssd from 2012 may have died 
idk, got a massive amount of paging errors in event viewer and it feels warm
rip
I never really used it anymore, anyway
and it was only 240gb (which was $250 back in the day, oof)
I don't know what I had on it but hopefully nothing important
really need to back up all my stuff
really should of checked it sooner, I think it had been having issues for like a week now 
possibly just running some endless loop of whatever
I remember paying 500-600€ for a 1TB SATA SSD back in the days 
yeah, I did too
except USD
240gb obviously didn't hold me out for long and I ditched hard drives right when I got that ssd
I paid for 250gb for 150 back in 2013
and then 2tb got down to $63 and 4tb down to $150
dang you, discord.. need to resub to that channel to get that emote back..
SSDs for operating systems were probably one of the most significant speedups in computer history where you can immediately feel the impact
Toa

went to assist a robotics competition again this year and one of my juniors caught the top school cheating in broad daylight
still won first place after literally fully swapping out their robot
How did they cheat?

they had a performance where they would do a "magic trick" with their robot behind a curtain and during that time they completely swapped out their robot for a new one
only reason they could even pull this off is because reputation and also because their school genuinely has infinite budget
Damn
But how was it caught
i assume both robots look the same otherwise there is no sense
they did it in the open hidden from the judges but my juniors were recording it all
welp the competition has basically no prizes and its only used for their portfolio so not that big of a deal
Go sleep 
No
I have to go to Amsterdam

This thing arrives today
But im not home till friday
Free wifi card yoink
interesting acronym there
Haha
I need to get a card like that for my Thinkpad
Toast!
Enterprise Resource Planning
Economical RAM Pricing
computable_fn 
hey hey ktrain
i wouldn't know, no birthday today
but if it were
it would've been a good birthday



just doing a quick walk-around of the server to make sure everything is in order
Don't worry. Next time you wake up the whole server would be toast 
don't check #neurotic-neurons if you don't want to stay up later 
happy birthday and good night shiwo
what happening
The new Socket, labeled as "Socket V," is not only next-gen CPU-ready, with reports claiming Razor Lake, Titan Lake, and Hammer Lake, all landing on the same package, but will also make it easy for DIY users to reuse existing coolers without needing to change their cooling solutions.
who are you
okay, imma just list it all
(eq means full core equivalent based on 2e ~ 1p)
also bLLC is big last level cache so huge cache but apparently not stacked like amd so its a 154mm² tile instead of 98mm² tile. probably for the better TBH allows for us xOC people to have fun with lower thermal density overall
Core Ultra X(?)
- 48 core - 2x (8p+16e bLLC) - (32p eq)
- 40 core - 2x (8p+12e bLLC) - (28p eq)
Core Ultra 9
- 24 core - (8p+16e bLLC) - (16p eq)
- 24 core - (8p+16e) - (16p eq)
- 18 core - (6p+12e) - (12p eq)
Core Ultra 7
- 20 core - (8p+12e bLLC) - (14p eq)
- 20 core - (8p+12e) - (14p eq)
- 12 core - (4p+8e bLLC) - (8p eq)
Core Ultra 5
- 18 core - (6p+12e) - (12p eq)
- 8 core - (4p+4e) - (6p eq)
- 4 core - (4p)
Core Ultra 3 (i felt bad so i'll actually include lp core this time)
- 6 core - (2p + 4lp) - (2p eq)
intersting.
yes i ignored the lp cores most of it
because i don't consider them compute core they're more like i/o cores residing in the i/o tile
I wonder how they will perform.
i just need more info about the pcie but i haven't gotten time to look into that
but well
also, i hope they bring back SMT
I forget what SMT is
simultaneous multi threading
can have benefits on compute workloads that fully utilise all the cores you give them
hardware level thread concurrency
so it switch between 2 threads per cores
this have benefit
but can have negative consequences on some latency sensitive single core tasks, typically the tradeoff is good though
if your first thread is doing vector heavy compute
it can do it while also switching to the 2nd thread
so you can get overall more multi threaded instructions executed
usually minimal tbh
i never feel any difference between SMT on and off for single thread programs
but it does feel different on multi
so i just kept SMT on all times
ye, on my context switching interpreter, which is pretty much the worst case for smt, the difference in runtime is about 15% at most
also SMT doesn't waste space its just extra registers and pointer hardwares and switching module
maybe some predictions component aswell
in some games i get worse jitter with smt enabled but not noticeable.
if i have a lot of programs open, its actually preferable to have smt enabled, which makes sense i guess
oh right, triangle was one with big performance impact of smt on vs off
i got much higher fps with smt disabled on triangle
it pins one core to 100% usage and has another thread doing who knows what
the fact i dont have a dgpu may be related
Does this mean intel will support multiple cpu gen with the same socket?
yep
They finally realize the error of their ways lol
lga 1851 is ass
extreme oc people still prefer 14900KS for core oc
but ig they prefer 285K for mem oc
Their price 
im so glad i don't rely on x3d chip
AMD hardware is still good
and preferred high core count
Just not budget friendly
amd has the best balance with efficiency btw
5.2ghz avx512 all 16 core at 179W
but their power state suck
Why do you even hit that
Did you hit that poor CPU on a regular basis?
meanwhile my intel
0.9W
power state wise is far superior
but shit efficiency at high load
because it pulls 270W
for the 5.2GHz all core
...on avx2
:(
that's nothing tbh
frog hit 800W on her threadripper 3990x
0.9W 
i think just recently 920W after we uh, mess around with the OCP a bit

burst though
that much through only 2* 8pin eps is 🔥
they can handle it fine, just not ideal
without this ocp trips

That's insane thing to do with expensive HW 
Optimize it further
We demand optimized triangle


I misread that as One Prick 

I did not expect to find another chez viewer here
creature
000 - 5v
001 - 15v
010 - 5v
011 - 9v
100 - 5v
101 - 20v
110 - 5v
111 - 9v
i think im dentge
what a weird
I'm feeling dentege as I'm attempting to deal with the EML thing
Currently it seems like I can't do it in raw python because mystical things called errors
and the other versions seem to disallow the error checking
Have anyone tried both gemma 4 and gpt oss?
Can you give me some recommendations? The online one always throw me to paid tier (im poor)
one round of dentege later, and I can confirm the paper's contents are actually real

its a known fact that all of these functions can be calculated using exp and ln
its not very useful though
for cheap computations, there are often better ways than using the exponential form of the functions i.e. using polynomials
theyre accurate and much faster
Makes sense
analogue circuits can also be used to compute them, so the eml function is a bit useless there too
plus with things like floating point, exp and log have such huge error margins for certain values
especially when ln is computed as eml(..., eml(...)) or something

the function is literally exp minus log, why would you want to calculate a huge value and then subtract a tiny value from it if you value accuracy

That I did pick up on early. I feel like the only useful thing might just to take advantage of the composition, but even then there are surely better methods
the composition is actually a flaw rather than a benefit imo
I can see that as well
its like adding unnessecary polymorphism to a programming project
you dont need to use this function to calculate other functions in a worse way than just calculating those functions straight up
for simplicity of implementation sake it could be considered useful i guess
The only way I could see it being useful is if on the silicon there was exactly this one op for floating point and it just attempts to use as many copies of it via some pipeline
the point of the matter is, exp and ln are nontrivial functions, with exp creating huge values and ln creating small ones, so not only is the computational cost high to do in software but the accuracy is awful
i'd say having an exp and an ln instruction would be more useful than an eml instruction
Agreed there
fused ops like fma are useful for improving pipeline efficiency, but i doubt such a huge instruction would be very efficient anyway, it definitely wouldnt be a 1-cycle op
i can see it being useful in certain asics, but they probably have more specific needs that can be catered to more directly

I'd say the only way it works is someone is insane enough to all in and bruteforce it being useful. Aka composing all other ops as said eml functions.
its cheaper to just not do that
But even then, not the most useful

It opens an interesting door only in terms of thought and not much else
reminds me of how flash attention uses an approximation of 2^x and ln(2) to compute e^x because just relying on the SFU for it is too slow 
That thought being, just hardware support for fused opts in general. But that's far too complex to consider as is.

I'm not the most invested in it, I just thought it was a interesting idea. But clearly it isn't that easy otherwise it'd already be done
guys what is possibility that it is lldb fault? i found no any issue, it is single even never repeated, and it is completely inside trt10
logger do literally nothing btw

that is literally it
as narrow as possible
there literally nothing else done with trt before this call, it is first trt call in entire application (only getter cuda calls before)
my reaction to this sigsegv
do you have more context
no, that all
since it is literally first trt call there no any other context to provide
i still have it in debugger btw, if some one have other guesses
fck that was 3 hours already wtf
no way i already spend 3 hours for this
it's a bit inconvenient that there's no symbol names
nah it is expected
it is .so (obviously with obfuscation)
i not doing cuda code, i only use what libs already provide
most low lvl is just memory reordering, only task for gpu here i neural network so
there 0 cuda code and i prob can compile app without any cuda binary in path
that is cursed
that absolutely random 0x0 address

doesn't tensorrt require cuda to build
you don't build tensorrt
build what, it may be require some binary but prob that should be only .so lib
all cuda, cudnn and trt goes already compiled obviously
and you're 100% certain that the logger is the problem?
in that case it would show trace to not only logger overridden function but also qt inner and custom handler
surely they wouldn't send you a nullptr message right 
I mean a nullptr message should be valid
idk how qt handles that
the only other thing i could think is that somehow the logger reference gets invalidated by something by the time it's trying to be used but surely that can't be happening right
what does your logger implementation look like
i just send above just in case
this is absolutely valid, every variable match and logger itself is nothing but overridden class
so only field in logger is vtable ptr
wrrr
wrrrr
idk exactly how it works, but have you tried throwing in debug prints 
because at the very least, a printf("log called") might be a good idea at the start of the log function
erm log works perfectly it only broke now
and since i for the first time released debugger and tried to restart...
that is fucking weird shit:
runs perfectly all the time in main application
fail always in automatic test executable
so i may want to try look at what i linking now
btw idk what have changed but not error is even more ridiculous
wait a minute... 👀 this layout looks familiar
guys that how my reset looks like
and somehow only reason i see why it breaks is model being swapped
it is only cause, cuz init apparently called for every single slot controller receive, just cuz there will always be error about no such file
but changing model so fast can cause errors inside trt
just weird fuck idk guys that shit is cursed
it is only way i see it happens
i can show test case that cause that (since previously there was no way to swap file naturally cuz i used run() function instead of object-in-thread, so i had to wait for last model to generate, now it happens automatically just because signal which trigger model name change can happen only inside thread that either idle, or already run model, gen or inference
literally only reason i see is swapping name too fast
but i have no idea how it would cause memory errors
and btw no its not threads issue with race on model name
test case that cause error
open model operate in special mode that allow from code change instead of messing with DE native window
remove here additional two clicks and there no error
-# (actually there error but not trt error, just wrong old test condition)
I GOING FUCKING INSANE
nice
bullshit
i found an issue
but that bullshit
some check in completely different thread btw
it cause exit.c and therefore error in different thread
that fucking awful, i debugged that for entire day
any with all other cout noise i just havent noticed this cout for this whole time
just figured out that it is weird one, removed all other couts i added temporarily, and searched for cout/cerr in entire app
only then i found this one
that is fucking awful
that is complete shit
I HATE DEBUGGING
std::cout << __PRETTY_FUNCTION__ <<std::endl; \
std::exit(1)```
-> 284 if (not QFile(filename).exists()) { testing_abort(); return; }
nowadays I only really debug because someone else's code doesn't do what it says it does
i prefer to use __builtin_debugtrap (clang only) and __builtin_trap (gcc and clang) for exiting in debuggers
it doe

i mean that "abort" was written long ago, i forgot it is exit and not common abort
nvm
did some digging and benchmarking, turns out that surf::get reuses the same client under the hood, while reqwest doesnt. stashing a reqwest client to a oncelock and using that solves the issue
Kasane Teto vs genuine FINloid Aino Erufu, Cha Cha Cha
it was my code from long ago
anyway that code actually do what it designed to - exit with 1 that cause auto-runner to propagate error
just only wrong thing here is that it cause other thread to error in sigsegv
-#
Taking a detour from NeuroSynth to explore Finnish vocal synthesis
if it causes an error then it's either not a complete implementation or the inerface isn't clear enough to avoid misuse
it cause error cuz it is error (surely abort is error)
me from that long ago wasnt aware that for debug i should only use aborts and not random exit
this line is 2 years old looking at git
buy more ram

Download more ram
oops.
This now works only with play money instead and everyone is given 500 nEuros to click stuff 
Also no need in any wallet extensions, no need for email too (when signing in click "Guest").
My last update as there's no interest, onto my CV as a pet-project it goes 
Bro there's no way
At my new job
Literally the entire source code is dumped into the root folder
I'm not even kidding i could share my screen in fucking video call😭 😭
They use folders just for extra artifacts that isn't code like configs or images or other data
Maintainability is gonna be so friggin ass
I found out why
I stalked linkedin
Seems like the main current maintainer had 0 experience or education in computer science
She has a masters but in linguistics
And somehow then got software engineer job
She earns like 2x my salary rn btw


Also the other new hire was doing something and i was like ??????
He was trying to ssh into the prod db server and set it up as a vscode workspace via ssh

That seems very stupid
Trying to launch a vscode server on a prod server
Cause he said he did it at his previous company
I literally must be dreaming

worst possible server to do that with
There's no way this is real
Yeah
The entire infra is living by hopes and dreams
at least if it was some random app server with nothing saved locally it'll probably get obliterated as part of some big upgrade anyway
but the DB lives forever, now with that guys VS Code setup 
The 10 year legal data retention server is living on fucking hostinger ssd with no backup
It's literally a dump of zip filed csvs
Via FTP
that seems... somewhat reasonable at least
other than the no backup part
I can totally believe this btw
no cold storage cloud bucket on tape with data redundancy
????
no
you could, but the main advantage of cold storage is lower cost, if you're willing to pay for it to stay on an SSD then whatever 
It's a 1tb ssd and they said they don't know how to handle future data anymore cause it's at 850gb full
ideally you'd have some proper archival system but if not then I think putting things on some random host and not letting anyone touch it is reasonable
if it had backups, that is 
The host has public ssh password login
U can literally ssh into it from anywhere
And the password is stored in a shared teams chat

I'm happy that my workplace usually doesn't do bs like that
acquisitions are a different story though 
and poghd usage
random Linux 2.6 hosts with 3 years uptime and no backups my beloathed
i dont on my phone
i know 
Toast

Someone who plays Minecraft
It said on a sign to ping you
NOT toast
you called? 

Wonder if I can run a model on my 1928 adding machine
Wait
I can’t help but notice
I feel as though something is amiss here
If I didn’t know any better
I’d say a majority of recent chatters had been converted into a heated crunchy bread product
good thing you are not forced to use gnome and can switch to normal UI
Idk
Shouldn't everyone have basic human rights
That is silly, though 
define normal
https://www.youtube.com/watch?v=M1si1y5lvkk This is awesome. How to make HTTPS not secure. (Do a packet capture of https://tom7.org/ to see it in action. See that it says "Connection is secure". ]
In this pre-paid self-own, I study security vulnerability through the lens of self-flagellation. I go through a lot of trouble to make life worse for me and my antediluvian website's users (AS USUAL).
Get it or try it:
https://tom7.org/httpv/
Or read the 25-page (!) SIGBOVIK 2026 paper: https://tom7.org/httpv/httpv.pdf
Keywords: https, tls 1....
like, usable
I'm not a linux user but considering it for my next pc
and I've been mostly recommended kde
I mean gnome is technically usable
I think you might be getting delussions from hunger

Gippity is manly and looks at stolen code?
Kewl
gnome is technically usable like magma is technically a drink
(minor exaggeration)
banana bread is technically not toast
azaka is technically not toast either
but you dont see anyone complaining about it (right now)
(you will later) (its how they operate)
that's right I do not wish to be toast

ohisee
this is #baking but cooking and restaurant meals are also welcome
just dont post like, a plate of half eaten toast
what about buns
bnuuys?
bun
yes
might be time for pfp change
deep fried between two sticks of butter.

british
give it to me before you dip and i will give two cookies for one stick
insane how this took months before first ping
I think putting up a sign in minecraft is even more insane lmao
make sure you toss it in the dishwasher after this
The RAM or the mug?
it's equivalent to asking "would you wash the toast or the plate"
Have you seen the episode of hells kitchen where the lady wash pasta from the trash can?
You don't handwash your toast before eating?
british cooking at its finest
ngl it probably at least has some taste after sitting in the bin for a bit
oh fuck I forgot to buy fish
was gonna make some fish and chips this friday
EWWWWWWWWWWWWWWW
how dumb is that cook oml

I had the random instinct to rewrite a very ineffeicent back end service in one of my apps in Go
Will I regret this decision in future , hopefully not, am I fucking my sleep schedule more , defiantly
From what language tho?
TypeScript? Then probably not regret it that much since MS is also doing that
Typescript [idk if this was a good decision but I made it]
gg
Bro guessed
I wrote it in typescript then about a month later am regretting it after seeing the usage
I also just realised I meant to go hunt down my phone like 30 minutes ago but got sidetracked
its so jover
tbf I have not been using it for a similar reason
and partially cas I lowk forgot abt it for a solid while
dont even ask me how idrk, then I was like how tf am igonna fix this then I was like oh wait I just write it in go, that owuld make alot more sense
I also found my phone
finally
It had somehow eneded up under the middle of my bed I am still unsure of how that happened
Who knew AI company would hype up their AI
https://blog.robbowley.net/2026/04/14/ai-watershed-moment-or-expensive-pen-tester-the-aisi-mythos-data/
ehh
sounds like a skill issue to me
i mean if you're gonna rewrite regardless sure, do in a different lang
Its an expensive pentester sure, but the significant point is that its an expensive pentester with no steadfast will of its own
That makes it more or less a vending machine for zero days
for finding zero days.... right?
Depends how malicious you are
As you can see, the problem is variance. Out of all the 10 run, only 3 managed to actually solve the 32-step benchmark. And that benchmark has a lot of caveat that real world software wouldn't have
Y'all! I wanna make a video game, so I can study for my NREMT is a new way! 😄
Plus it is indeed an improvement but not the watershed moment that they hype it up
To expect one successful outcome you’d need 3-4 runs on average – and each run is likely comparable in time to the Opus 4.6 runs.
That’s approximately $2,900-3,500 per successful outcome.
A human expert completing the same range: 14 hours, once, reliably. At $125-190 per hour (UK rates) that’s $1,750-2,625.
Meaning you're better off hiring people and buying exploits like every government has been doing
You're missing the point a bit here
You don't need to be a government, and you don't need to be accountable to anyone if you have an LLM with those capabilities
And the classic saying of "this is the worst it will ever be" feels more relevant than ever
Even open weight releases will catch up with this in time
Sure it might not be a financial improvement, but its accountability and not cost that makes this significant
This is of course also probably the most expensive this level of capability will ever be
Its also very good at picking out exploits humans miss, that Linux privilege escalation zero day was like 27 years old
I'm not sure where else to go with this, I'm looking for anyone who's familiar with complex information systems or HCI as it relates to alignments. I'm not knowledgeable and I could really use some help figuring out what I'm calling an anti meme.
I may be misunderstanding, but how can a zero day be 27 years old?
a zero day just means it has not been found before and its been in the code for 27 years.
Doesn't that mean that every exploit is a zero day?
I am currently making an AI
no. since some exploits get introduced with new features and such
thats atleast how i would describe it. zero days are usually old bugs that have never been found
hypothetically, is there a way for me to make three mobs in minecraft to become truly sentient without LLM but with Javascript/node.js and make them able to read the chat?
Something like Neuro-sama?
normally since the first release of a program
If found in live software when the vendor isn't aware then yes
what
The "zero day" moniker refers to there being "zero days" left for the vendor to put out a fix
what I mean is if there is a way for me to make an AI just like Neuro
yes you use an LLM
oh
ig so, to me it makes more sense by framing it as "zero days' notice to vendors before wild exploitation"
can you help me with it?
why
Someone post the gif
Because I need help with it.
That's a better way to word it actually
what specifically
Since it remains a zero day once responsibly disclosed but with no fix available, your framings clearer
Half of the coding
no
you didnt let me finish
the answer was always gonna be no anyways
Okay but atleast let me finish
ill help you
thank you
fine
I'd argue Whisper is far from the best option for asr these days
notlike that
the AI is meant to be a personalized AI
it's the OG so there's a lot of resources for it if you need help
but yeah maybe not the "best"
One that studies human phsyce to change the said game or behavior
Also, for native voice input + output in an LLM you have the Qwen 3 Omni series and for just native input you also have Voxtral and Gemma 4 E2-4B
Please refer to the "just works":"all of it" box in the table above
I've sat here for the past couple of minutes trying to decipher what "human phsyce" even means
I'm assuming he means a soul
I forgor how to spell somehow
I mean, I suppose a model that spews random garbage could be considered some kind of soul
But I'd just call that a very inefficient random number generator
basically yeah
Is joke innit
ur 14 just go back to school
Hey i was wondering if someone knows what TTS programm Vedal uses or has he created his own?
neuro's voice is one of the microsoft azure ones but pitched up i think
evil's is more complicated than that
do u know if he uses an API? I realised that atleast the websites i checked are hella expensive
so was wondering if there is a better option
inb4 sued by Macroslop for v3
not sure
Thanks btw
i think thats just the one its based on. but it isn't actually microsoft azure that is being used 
is that so?
probably not, it seems slightly too fast for it to be a remote server
have you heard microsoft's tts?
its not even close to as good as neuro
hmm makes sense
how the helli do u create a local tts tho...
I assumed evil’s is some sort of rvc but maybe im confusing it with smth else
using a tts library
tbh im kinda dumb for not even searching up anything about local tts :/ thanks 😅
might wanna look into kokoro
if it’s your first tts
yea that was the first thing that popped up
"# PyGame is a team management system designed to extract data from a large text file and save it as a JSON object"
- deepseek-r1:8b, completely unprompted


kinda yea tried using elevenlabs while expirimenting with stuff but after i saw the price tags i preferably wanna switch lol
i toldit to monologue about lua, back in august
and it didn't stop
and degraded to madness
Man, Pygame is so heavy for a glorified data extractor library
what am I reading
Dent
"I need to write a code snippet of 1088-bit floating point number"
this is my favorite quote
Reminds me of needing to update the FAQ
what's a 2018 Forda shortstory
dentge
What if the FFN weights in a transformer aren't a black box — but a graph database you can query directly?
LARQL is a query language for transformer internals. DESCRIBE an entity. SELECT edges by relation. SHOW the relation types the model taught itself. INFER a prompt and watch the graph walk produce the answer.
No matrix multiply. The vind...
I don't have the energy, rn, to come up with a sarcastic enough comment, that is any form of concise, so I'll just drop a TL;DR:
Not that surprised.
how do i kill john connor?

uhm does anyone know how Vedal managed to make Neuro and Evil so "realistic" i highly assume he uses an ai like llama or gemma right? or did he somehow make his own?
wdym
wheres the chart thing
They dont sound like AI. They ask random stuff they learn they sound more "human" if that makes sense.
they were trained on more human sounding data simple
I can tell they are AIs though from how they talk...
gemini claude etc can sound human but it’s a corporate decision to be more neutrally sounding
that.. actually would make sense
ah hell nah
other ones exist but theyre mostly filtered bait
Do yall know if u can "train" any ai or only specific ones?
what do you mean
wheres the CHAARTT
As he said the ai vedal uses has trained on human sounding data. can i train an ai like Llama3.1 Darkstorm Aspire 8B the same way or are only some specific ais designed to be "trained"
Any neural network can be trained
In the context of what you are talking about, it's called "fine tuning"
huh Thanks a lot man il look into it 👍
where the helli did u find that
pinned
damn
The technology itself is extremely simple, the hard part is the human element of designing things to be cohesive enough to be interesting
tell your gpu how simple it is
who tf wrote "Object detection" for vision 💀
Not how VLM vision works lmao
yeah i pointed that out here
Not difficult, just time consuming
Even high school students can multiply matricies
i'm joking around
who isn't 🧠💡
hey you’re k0ntrol guy right
big fan
Yeah thanks! It's a very enjoyable project to work on
I gotta get the website up soon for people to help collect data
always data with these dang ais 
It'd be really cool if Vedal looked in the K0NTR0L chat because he could implement some of those methods to improve 3d
Everytime I watch the stream I am SCREEMING to myself because I know how to fix it to be more natural. Although... it is very funny to see them make nonsensical movements
Yeah
do you know why this happens
whatever that means 🤷♂️
Thing is, I can reduce L*tency too
she looks like she's trying to frown
this is in between two normal frames of her just smiling
Since u said it is extremely simple do u mind helping me out a sec? 😅 i dont exactly train AIs on a daily basis lol.
Do you know Python?
My gpt does xD
Bro 😭
i have a hole setup already
then it's the worst twitch ever
its just the personality thats kinda... ruff on the edges lmao
You won't be able to easily finetune an LLM without knowing programming
Def a skill worth learning
im working on it... lets just say py wasnt on the list yet
Dont worry about the coding i got that
i know some people :D
i genuenly dont have anything better to do with my life
not even close
Then learn programming lol
I work at IT so its in training :)
i just need to know how i can "teach" the ai. The way i imgain it is making it read thro a crap ton of conversations but i feel like im wrong arent i?

ais arent really taught anything they’re all just good at probability
more data better probability (not necessarily)
they'll also take any and all shortcuts they possibly can
Makes sense
better data is better than more data
But how do i feed them the data 🙃
Do i have to manually paste conversations to it or can i "automate" it
I know the answer but you won't like it. It's not easy
I worry I will only mislead you
is your answer lora my answer was lora
uuh
you really should learn python before you fiddle with any packages though
you should learn programming before touching any packages
like, all the basics with modular programming
LoRA/Q-lora yeah. You can use unsloth to automate it a little better. q-lora is better than standard LoRA because 8-bit weight and a sufficiently large rank will consume a good deal of VRAM
i got plenty... maybe
a neuro clone really shouldnt be your first project
luckily i upgraded from my 4060ti 8g lol
You are gunna need a 5090 or enterprise
unless you use like a 1b parameter model lol
which isn't going to be very good
my 4090 trains 8b just fine
i got 24gb vram il install a second gpu soon that has to be at least a bit right?
that's pretty nice
:)
probably enough for a small model. That being said, as KTrain said. It might not be a good idea to do this as your first project
you might get more discouraged
and confused
starting small builds more confidence
not to mention the ai hallucinations
Im a quick learner :) trust me
I'm all for full sending it... but this stuff requires a lotta background knowledge
“quick learner” should justify learning programming more no?
its a work in progress 
I highly suggest learning it
Not only will it help with LLM projects but also a great skill period
Even though it's less "exciting", it's good to have a solid foundation
I mean i can reat py i just dont write it. i man html css and js so i have a good understanding of what does what
I can read it but cant write it. that makes sense right?
I can train 10M at decent batch on 6gb. Do I recommend this action? Hell no
I understand. I still stand on what I said before
Lol I trained a 30m model on 4gb
bs=1
pretrained 💀
I think it was batch 256, the vram optimization strategies went crazy
But also slow

Nowaying echo
9+10=21 computation
not online
The 2+2 example lives rent free in my head from that one chatter
A lotta chatters may as well be llms lol
My man was more stupid than an llm
a 10m parameter llm 💡
ITS EVEN FUNNIER THE THIRD TIME ITS EVEN FUNNIER THE FOURTH TIME ITS EVEN FUNNIER THE FIFTH TIME ITS EVEN FUNNIER THE SIXTH TIME ITS EVEN FUNNIER THE SEVENTH TIME ITS EVEN FUNNIER THE NINTH TIME
an llm who generates all the time has nothing to think but hallucinations
nah u can't be leaking my chats
that's just a normal llm
ik it’s a spinoff of that one meme
pure hallucinations
Where is the gender calculus post I need to grab my laptop to properly search chat history
It was during the harpooning I thought you knew 
Makes sense
Gotta farm
If I wasn't dead I would have unloaded until that man was harpooned
mr beast scam inbound, chatter got hacked

to the ellie ver
Are you sure you want to 
#programming message here's the start of the day
this went on for 4 hours
I dont trust those random capitalizations
either they’re being intentionally sarcastic or its part of the neuro arg
True
in: programming from:braintemple for your own search
let me find the 2+2 post
I was just reading a paper yesterday about using purely monad algebra homomorphisms instead of Geometric representations
that is defently not this
no yeah ik
Too dumb to dumb it down
I don't understand it well enough to explain it lol
It was a very short paper too
maybe they dont understand it either 
peak humanity
they just states their opinion
math paper without proof?.?
yeah it's a position paper
most definitely an arg.
Trust me bro citations
like a… conjectureeee…?
it was just an idea
not even a conjecture
what is this philosophy major 💔 💔 💔
If researchers wanted to hear opinions they would just make Twitter accounts
my position paper is that NNv5 is real and is not a figment of my imagination
They just guessed that using monads would be a more interesting and useful way to understand deep learning compared to geometric manifolds

I don't know if I agree or not really
Funny to see catagory theory applied to DL
"1.1. Our Opinion"
Its up there with the old 11D hypercube for AI models paper
The references take up double the actual paper
on a math paper
its so neurover
There are barely any references in that paper
Ai generated
oh whoops ur right nvm
Not AI 😭
#programming message
No one tell them who sent in the report, despite quack openly saying he was going to report it
I looked at page 9 and didnt scroll down lol
I knew it... 
@real sierra happy birthday here’s a math paper, thoughts?

Math proof to show its always his birthday

better be a good paper
mabe for entertainment
Transformer based ANNs innovation challenge impossible
i knew someone who would preach monads
TRUE

Procedural programming only 💡
We only code in assembly
i only code in llm prompts
Which assembly
Damn the museum called
I really enjoy 6502 assembly
It wants its exhibit back
less bloat
8 bit is nice cuz then you never need to confuse a word and a byte 💡💡
yeah
RICOH 8 bit cpu
128 bit adder 🤨
I played a lil with armv7 assembly too
kinda cool
anything but x86_64 
Write bare metal
@real sierra’s birthday is known to be some day this year
there are up to 366 possible birth days in a given year
shiro is a person online and thus he could be lying about his true birthday
we use 366 iterations of shiro assigned to a different day of the year
and by the pigeonhole principle at least one of those must be correct, where the pigeon is the true birthday and the holes are the days
alternatively going off the assumption that shiro could be lying about his birthday
say that in a given year there is a 1/n chance that we correctly guess his true birthday
by wishing him a happy birthday every day of the n day year, we will have an expected value for the correct number of birthdays as 1
holdup im cooking more
holy drop
References, lemmas, results, methods? Not seeing those sections buddy
it was his birthday a few days ago
also needs examples of usage
like actually
Latex formulas only
experimental applications and future work
If its his birthday everyday does that mean he just has one big birthday or multiple smaller ones?
If so when does one end and the next start?
Does it depend on timezones?
So its constantly in the present
thats why he wants gifts
Wait a second that was his birthday gift
He wasnt given a present he was given the present
So his birthday is always
but note lemma 1.1:
present is the present for all real numbers
let the set we use for birthday be called \mathbb{B} = {0, 1, 2, … 365}
let today be day t \in \mathbb{B}
the sequence
t, t+1, t+2, …, t+365 \pmod{366} will include every day once
let b equal his true birthday
the sequence
t-b, t-b+1, t-b+2, …, t-b+365 \pmod{365}
will also cover every day once
by the pigeonhole principle if we choose one birthday b to be is birthday b \in \mathbb{B} then exactly one value in the sequence will be 0 \pmod{366} , regardless of whether it is in front or behind. Every value in each sequence is unique, so we are again guaranteed to hit shiro’s birthday once if we choose 366 pigeons into 366 holes where each pigeon occupies one hole
LaTex my beloathed
For math and research i do prefer it over word or the others
Mostly due to formatting options
I think i learned it originally back in a linear algebra class in uni
ah yes, great times












