#programming
1 messages · Page 439 of 1
that's for sure
a lot of new tech has been invented, plus he's of course learned more
there were implementations of adapters around the same-ish time but idk about before then because i didnt care more at the time
I dont doubt that vedal can make an llm from scratch but I wonder if the twitch chat datasets alone are big enough to encompass every possible scenario
they arent
i'm sure he's basing it off of one of the general datasets
she has enough world knowledge that she's either a finetune but a deep one or it's a full ass big ass from scratch llm
it costs a fuckload of compute to train a base model though
yeah and considering vedal seems like he has a life outside of neuro I would say neuro’s just a finetune
surely… he wouldnt spend years of his life formatting data to make a vtuber before knowing it would be this big
like if i were to set out to train a 3b llm from scratch, and neuro is likely (not CERTAINLY but she doesnt act like she's this small at least for a while now) is easily a 500k+ investment
you can do things for fun you know lol
yeah but not 500k for fun
you don't need to know that something will be popular
well yeah but considering his speculated age is early 20s he would be broke when he was making neuro
that kind of blind investment could only come from soul
you need trillions of tokens of data and clusters of gpus to do it even with smaller models
llama 3.2 3b took 460,000 hours on h100
neuro originally was an osu NN also
separate and much easier to train
the early llm neuro i don't think was that general, it was pretty firmly twitch ranting from the fragments i saw of the early versions
you can train a twitch chat replica far easier
not one that you can continue to grow into what neuro is now
yall are thinking about generalist models which are so much more difficult to do
pretty sure he's swapped her out a few times, wasn't there a few times where people complained that she's noticeably different?
i have been watching neuro since early 2023 (february) so about 2 months after her debut as a vtuber in llm form (non osu)
she has always been a small but general language model
holy og wtf
she was very dumb
so was every small model
yeah i havent been in seerver that long because i got banhammered once kekw
yo im tryna find clips of neuro v1
she has changed noticeably a few times yes tho
why is youtube tweaking
look at the clip of airis
she's clearly coherent enough that it's definitely a normal llm
staz does NOT have 0 views 💔
oh shoot that might be a lil controversial
ehh wtv it’s programming
Has she been sentient all along?
March 24, 2023:
https://youtu.be/whEtJ5DEZBY?si=KJdsqeWDo0C8mugj&t=7341
#vedal987 #neurosama #vtuber #ai
2023 already had some decent compact models, if vedal was/is somewhere in early 20s and working in software he can plausibly get hardware to run or train one locally
pretty sure he isn't that broke and only doing twitch
training a full model even then (in fact especially then) takes hardware on a scale that it needs an organization
like
clusters
to have it take anything less than years
i don't think it does
1b

okay so he could rent them out just for training time
but again that's only if you are building a production grade general model
subathon to fund the vera rubin rack
training llama 2 put out 539 tons of CO2 and took 3.3 million gpu hours
to get a model as stupid as llama 2
vedal is just using 1B (if statements) trust
the very first GPT-2 which was a 124m parameter model took 7 days on 32 TPUs
which sounds very small
but
it cost 43 thousand dollars
at the time
and yeah gpt2 is far dumber than neuro was
neuro was chaotic and janky like any smaller model then
when you feed it twitch chat
she was never 124m dumb tho
there is also the fact that she knows a little bit about everything
twitch chat can literally scheme to make her talk about molecular chemistry or smth
the fact that vedal would have to prepare neuro for literally everything despite being intended for one purpose implies a generalist model
it is also part of what helps her generalizxe to be able to do her thing
https://www.youtube.com/watch?v=L4nfzJx7uDI&list=PLCkrLkA-HmjPhhRjJMaZBbhQIHRz8ALzF january 2023, she'd been vtubing less than 2 months
not to mention neuro used to say “as an ai” a bit
this is definitely a regular language model in there
which ofc isnt a very solid argument but the alternative is that vedal really wanted to drill into his daughter that she is a clanker
actually I dont doubt that either 
ignore that statement
nah there was a reason back then the schizoing was about which particular small model she was based on her mannerisms
dunno dunno
the first couple gpts aren't a good example though, they were completely untuned
they were well before her time
and also gpt-2 wasn't that dumb, it was fairly coherent
back i the day vedal used to say things about her model
i think for a twitch kind of llm it wouldn't take so much that it's impossible to do at home or by renting a server farm
especially if he works as a sw dev as his dayjob
maybe not now but surely before
he knew development well enough to atleast build the osu bot from scratch
she has too much knowledge even in this video from january to have been this
the fact that he had time to make 6 of them would imply they’re interchangeable enough to be completely different models right
pounds uhh what is that in kilograms
about $350
wow $350
knowledge idunno, can be repetitions of what someone somewhere has said, if you have a good dataset you can get good outputs
he is actively finetuning it on the convos atleast, but we know that
yeah but a good dataset for knowing anything about anything even in the undertrained ahh model era of 2023 was likely about 5TB in size and 1-1.5 trillion tokens
and would still take months
does mention qloras, probably uses those as well then
i do think he probably did full fine tuning at some point
he probably doesnt need to now
neuro didn't have memory until recently either
i think he's always had a good intuition for getting the behavior he wants out of a tune
neuro has had a form of memory since 2023
at the latestr
yep
he was adding stuff to her data and prompting harness for a while and then went to rag
which isn't really an "actual" memory system like she seems to have now
sshe had some sort of rag as early as when she came back from subathon in 2024
i am not confident earlier than that so i cant say how early
lol
my own reactions from back then
yeah she had this debut in birthday i think\
atleast we know it's not chatgpt :V
he for sure aint training anything big at home

she was running on that 1080ti for actually a stupid amount of time
hardest working 1080ti ever
man remembers tay
hahah true that
sydney too
bruh i completely forgot about sydney
im baaack
how did i forget bing melting down
who’s sydney
microslops bing chatbot from 2023 they accidentally published
it melted down and started threatening people
i still remember going absolutely crazy at how realistic subreddit simulator was and those were markov chain models
that shit was almost 10 years ago now
im dying
a markov chain with a sunny disposition 
oral what?
semicolons.
nothing there but a black dot im not sure
fuck old sub sim was so funny
that joke made no sense
oh its your ai slop bores me from a decade ago 😭
wait no its not
it was old markov chain language models trained on individual subreddits
quite the reverse
and they would just post in their subsim
which ended up being epic shitposts
idk if i have to block that one but
all I remember is that one chat bot that got trained on reddit and turned out extremely... reddit pilled
I got automodded, but you can probably imagine what type of personality it gained
I can just go over to dougdoug’s channel and get a good idea
that's what happens when they get exposed to the internet
grok went nuts for the same reason along with his system prompt being fucked with by john tesla X
and by exposed i mean allowing the exposure to actually affect their otherwise frozen weights in some way (or their harness)
grok had a sort of context sharing going on where each interaction was not a distinct one but shared
and that ended up just becoming a giant adversarial prompting surface
combined with system prompt shenanigans
tay also learned from interactions
i have a small suspicion vedal either vanished or wiped history between 2020 and 2022
he streamed osu only and then took a break to figure out neuro
seems he vanishes mid 2019 from the server, comes back end of 2021 for one day and then returns end of 2022
i bet he was working out the vtuber since 2020 at least
he used to talk about things
he had pitch meetings and stuff with her for a while
so i guess he was shopping it around
with whoms't'd've that would've been idk
would've'd'd a whale of a time
he was streaming on twitch from 2019 maybe 2018 thru 2022
yeah maybe he streamed way less than i remember
cuz i am looking at a similar site
2017 NOWAYING
i think these things dont really mark anything down for very low viewership,
seems about the same
yeah just with random sporadic other dates
there were some streams in the empty times
they do track low viewership ones
the one in 2018 apparently only had 4 viewers
30m stream

wack
well in other news, i think i managed to get my llm balanced with just a system prompt
needs a seasonal trend line fitted 
if he was actually active in the discord then he deleted the history but tbh i think he went on a real hiatus
was before my time
but he mentioned it being long time no see in 2021
when he announced abandoned archive which is so fucking funny
he would've been active in the old aa disc
but cant get into that anymore
vedal really hasnt been in here for a while has he
nope
he's too cool for us now
does he fear us now that he said he’d rewrite the codebase in typescript 
man's got bidness to run
he cant come in here without instantly getting mobbed "how maek my own noro samer"
should change his bio to just the link on his website
just ask an llm :V
the link to vedal ai where he explains it without actually explaining it
personality fine tunes are much more fun than anything else
that's why echo is fun
for me
also where he came from
partly
I personally got into the neuro making hole bc i wanted to learn 3d movement policies
(cough shameless plug ➡️ ⬆️)
lmao
it killed itself
I havent posted a video on it yet but im playing with isaac sim mostly it’s fun in general
it fears getting replaced
or perhaps being turned into a kawaii vtuber
he hasnt randomly brought up neuro in a while
strange coincidence
i think i fucked up his temp or something earlier because he's been extra crazy today
there was a joke I tried to say but it got modded
it's truly just that easy
original creator open source code 
is that kimjammer again
made the same mistake as neuro herself
I think asking chatbots about making neuros in my experience has been either too babying or too complex
it either believes you truly cant code shit without vibe coding or thinks you understand how chatgpt itself works
and the “personal intelligence” doesnt do much to quantify your skill meaningfully at all
sorry this is unrelated to neuro im just ranting about learning from llms
dunno, i think the answer qwen gave is fairly ok
imo you shouldn't expect an llm to teach you shit, but it's good for a jumping board if you don't know what to look for
hmm ykw yeah it was actually decent compared to what ive seen ais dole out
though its emphasis on model and tts are a bit questionable bc vedal himself used the default vtuber model and turned out fine

you can make a neuro sama that is vaguely similar but it wont be neuro
it does say
there's a github where someone made a whole system for it recently
cant remember the name of it
I mean there’s a lot yeah
project airi
even people who normally do other projects have made it for a weekend video
oh ik that yeah
i found out that i can tune the likelihood for the end of thinking token, experimenting with that a little in addition to other variables now
well making an ai vtuber is the easy part, hard part is having a personality
^
I would eventually try something like that, but right now I'd prefer just working out a base model I'd want to do.
think airi and other character ai platforms trying to serve everyone with low effort makes it more susceptible to “slop” behaviors though
oneshot always tells me i should walk to the carwash, while thinking has a 60/40 chance to tell me to drive
Have to try it once, but I still am not the biggest fan of it
for qwen specifically it's usually held true for me with the smaller ones, it's really easy to get them to absolutely go nuts when they try to think after finetunes if you're not careful
errors just pile up
i'm rawdogging it, but i do use q6
close enough
it's not ha
isnt it more just redistributing model probability to look smarter
yep exactly
it's just an extra amount of test time compute
and the training for it encourages problem solving while doing it
so it helps 2fold
also i stole a dumped system prompt from a random git repo, looks like longer prompts can sufficiently change what's generated in thinking
of course
i gave a qwen contradictory languages as an experiment
and it spent the whole thinking trace debating on that
lol
i told it to please tcritique and create its own version of some japanese lyrics i had for a song (not copyright lyrics i wrote them and then translated them w/ gemini's help a long while go lol)
then the system prompt was an old echo one
a little better than 4000 tokens wasted
literal alchemy though, i'm just shooting randomly with the settings
there's a lot of trial and e rror to do

i've two primary tests - it shouldn't derail in 1+1, and it answers the car wash correctly
he forgs with those
i do be forging with those
the thinking phase was completely ass and got cut off by the budget constraint, but it answered correctly
more likely than you think
the thinking is only going to influence the final answer but at the end of the day it isnt really thinking so it could end up back on the right answer anyways
i want it to be a little less loopy still though, because for bigger questions it will latch on irrelevant details and turn the probabilities to shit
here it atleast latched on the car wash itself, but often it's "i need to format it, wait, maybe i need to format it less, wait, maybe the user wants it formatted differently"
so it never actually expands the relevant context
thinking qwen
i should really build an automation for it and leave it overnight to drive every possibility lol
100m? damn.
damn, you really lobotomized your boy
nahhh can't be, what did you do to it?
asked it about car warsh
oh yeah
it's ancient echo prompt
he doesnt use one anymore
he's inherently an idiot now
lol
it's a poem, call it "car wash next door"
after i told him how tf am i gonna wash my car walking there
(qwen not real echo)
classic soap machine routine
driving to the car wash defeats the purpose if your goal is to wash the car
softstar987
chinese llm strikes again https://en.wikipedia.org/wiki/Softstar
Softstar Entertainment Inc. (Chinese: 大宇資訊股份有限公司), abbreviated as Softstar (Chinese: 大宇資訊; pinyin: Dàyǔzīxùn), is a Taiwanese video game developer and publisher based in Zhonghe District, New Taipei, with Mainland subsidiary branches in Beijing, Shanghai, and Xiamen.
Softstar is famous among the Taiwan regions for...
extremely classic
so true

what insight
i should do a throwback run training qwen2.5 3b
that model was a funny boy
this reminded me of you
perfectly rendered drawing of morbo from futurama
wow
also just saw this https://x.com/anemll/status/2035901335984611412
Running 400B model on iPhone!
0.6 t/s
Credit @danveloper @alexintosh @danpacary
@anemll
dude wakes up a bottom shelf 400B llm on an iphone and the first thing it says is "profound observation, you are absolutely right"
it dug deep for that one
0.6 t/s speed
is it airllm
honestly, for 400b on an iphone, thats wicked
just goes to show how tight these devices are
apparently the model there is qwen3.5 397b a17b
i mean, i cant run that at all on my machine
yeah i mean tis' still impressive

We finally have lift off
Its been a long night of work and that pepsi max bottle in the corner was the only thing keeping me alive
But we made it chat, the first arm is fully operational
And yeah, I did return the game back to where it was standing upright 
reasoning next level
this looks a lot like the arm on the roborock saros z70
next step
Next step is to design a mobile base with tank treads and a central lift to mount the arm on
🚀
Guys I'm so fucking good at coding (they need to take my computer away)
looks like it'll work well and get your 3 buckets
i hate then
english codin
when i switch from c to another language that has it i always forget
i always try to use it so the opposite i forget else
well
not els
that there';s only else
wehre my then
that's how i say it in my head, if then every time
i learned w/ python so ig you probably learned w/c or similar
This is SAS. I know SAS and R
Know is a strong word
they told it to us in the terms of if then
and then pythong doesnt use then

but it stuck
R is neat
is that jupyter
It's SAS studio online
never used
SAS is not open source
ive only ever seen R in jupyter and conda contexts
rare draw
ive only ever been around R from using jupyter for python dataset stuff
probably a common sentiment
why the hell does it remind me of something so much worse, the internet ruined industrial machinery for me forever
R is a good language for statistics and can also serve as a really glorified calculator
i wish i had the inclination for statistics
Wish I had an inclination for computer science
easily my worst math subject in school
by the power of god and anime, turns out here we both are in the same place because that's how it is on this bitch of an earth
for me my least favorite math subject has to be probability, not because of its difficulty but because it's depressing
i dont even remember why i hated stats so much
i wasn't inherently easy pz good at it that's probably why
i hate when that happens
math would be my favorite subject if it weren't for the fact that every fucking school I go to somehow manages to make me do straight review for the whole year
in the past five-ish years only one year was truly not review
prisma are you a bomb
your name seems to get longer every time I see you
also that’s so real
maybe
makes me want to get an iphone just to see it working
when can i run a 300B model on my smartwatch 
run it on 500 lemons
Holy shit is that morbo
vanor vanor, now that you arent purple can i interest you in


i will evolve beyond color
what if, cheese color
did sam lose a triangle
hmm nvm
YOU MUST PARTICIAPTE
@safe path EVEN IF YOU DO NOT DESIRE THE BANANA ROLE
THE BANANA ROLE DESIRES YOU

IT WILL CRAWL OVER YOU WITH ITS BANANA PEELS

AND CONSUME YOU LIKE THE DELICIOUS HUMAN FLESH FRUIT YOU ARE
so are you in or not
I love this cult
campaign man
oh look a bird

My avarage usage of claude
It doesn't want to function, I wonder why the unoffical linux port would be buggy 
ps aux | grep "bananna"
kill 84399
me also killing discord while running this
rune & yellow theme h,


why electron if you have browser
banana
banana 🍌
banana dog
bart simpson writing on the blackboard like

https://sladge.net/
At last
Self-declare heavy LLM use in your projects. Because shipping unchecked AI-generated code without warning is just rude.
Honestly was already considering adding something like this to one of my repos
So good to know it exists
I'll admit that one of the repos I'm experimenting with is vibecoded. The LLM repo I'm working on deserves it, though I'm also working on fixing up a slowdown manually.
I guess by proxy I get to work on one of my other projects
banana?
does this count 
im not sure

guess i have to be an anti-banana guerilla fighter now

good morning
good morning
sure
secret
still can't find banana bread
I just spent an entire work day
Doing nothing but getting a old netgear nas
RND2000v2
256mb ram, 2x2tb .. back up and running
To then be blown away at its amazing speed of a whopping 10-18MB/s
Resilver runs at impressive 50MB/s if nothing else happens
Even RPi1's did better via usb sata adapters xD
Why is claude opus so good
It has no right to be as good as it is in technical work relative to Gemini 3.1 Pro
When you leak 40GB of memory and then fix the memory leak and rerun:
(typically idle mem use is around 40GB itself)
What? A developer fixing a memory leak instead of just slapping zram onto it? In this economy??
Also based btop
(but why the brail theme instead of blocks? Kinda curious)
It's the default and I like how it looks
Fair enough
Well leaking memory literally crashed my many-hour training run so I gotta try to fix it
just slap more ram in it, and download some more (i hate this mentality)
((the only upside of the ram shortage))
Oh, does the arm use IK?
Yup, since I don't have a leader arm to pair it with I rely on end effector control with a gamepad for teleoperation
Damn that’s cool 
I was playing a rogue like, “wizard of legend 2” and player 2 kept bugging out:
- character completely froze and you had to spam buttons to unfreeze it
- character animation freeze (imagine moving around at running speed while stuck in crouch)
- physics stopped applying ( boss fight than involved platforming player 2 could slide around in midair as if on ice without falling into the abyss, ignoring the need to platform)
What could cause this?

:NAHH: gemini said oof
Humble pie :P
this is like a yearly occurence here lmao
the actual fix to a memory leak is to restart the process 
elixir
the discord way
Doesn't the brail allow higher resolution graphs?
Who was that guy that offered the neurons access to his H100 for fine tuning, is he still around
I could save a lot of money fine-tuning Pi0.5 if he is

latest windows update reminded me to back up my files again (i opted out 3 months ago)
and search bar is broken now
Hello
Is this safe?
Im trying to fix my pc my WiFi is connected but it can't use the internet
in theory it's safe as long as you obtained it from the Intel website
but I doubt what you describe is a driver issue
so it likely won't fix it 
So is the problem is my router itself?
Router or network config on the PC, yeah
Ohhh
How do i check the network config?
if you know your routers IP address you can try pinging it from the PC
How do i ping it
what is the problem exactly? "can't use the internet" as in?
dw lizard, its safe safe, i hope... 
I have the no internet sign but im connected to my wifi
its da meme
Its an actual wifi but not mine
win+r -> "cmd" -> "ipconfig" -> something gateway, copy that address into your browser
Try restarting ur router

Or just use a lan cable atp
cables are ass to run if there are doors
ok now what
im on my pc but connected to my phone
default gateway should be your router's ip address if you're connected to it
you need to go to it in the web browser
it's probably http://192.168.0.1 or http://192.168.1.1
it will open an admin panel, the username and password should be admin admin
Oh
and if it isn't, it will be on the router your isp gave you
continue, it's a local connection
Let me manually reset my router
I haven't turned it on or off yet
So i will so that
Yep, my cable holder's falling apart lmao
the black spot used to be slammed by the door
it's still slammed by another door in another spot
I think block looks higher resolution.. but idk, possible
Discord mentality

I'm happy with my Nvidia 580 Linux driver
I do in fact play games

Some of the more intensive ones include Subnautica and Satisfactory which like to make my GPU go all vacuum cleaner on me
isnt the problem potat is having exclusive to 5000 series or soemthing?
@sick owl what gpu do you have?
Well he has a 4090 no?
i dont remember
4090
I'm not having any problems now, the hotfix driver sorted it
And I am not budging from it any time soon
problem exclusive to 4000 series then?
Yeah weak performance in Resident Evil Requiem, it was a 4000 series issue
The 5070ti was outperforming the 4090 in path tracing

Thankfully they put out a hotfix that sorted it, but it took them a week
That's what Nvidia wants
But it's not what the hardware should be able to do
just artificially make the 4090 worse and now the marketing is correct
I'm on 590 and gaming on it as well
LLMs also run fine
595 
Driver Version: 590.48.01 CUDA Version: 13.1
oh, they made 595 the new stable release today 
lets see when that is gonna hit NixOS stable lol
gonna have to wait for 26.05 release
yeah that'll take a while, unless they backport it for some reason
unless you just specify the exact driver version yourself of course
Oh interesting
I have Nvidia 590 available by now
Eh, I won't bother installing that for now
atp i dont update my nvidia drivers unless im getting a bug
yeah, looks like that's not usually backported
i default to just updating whenever there's a new version cause they're still actively fixing a lot of wayland issues
wayland moment
truly
I'm on latest on one system and will be doing 580 on another system as have cards that have dropped support
i'd recommend 595 for anyone on kde plasma though cause of the new egl_wayland2 fixing the stupid fd leak
though I think they could if they wanted to?
ABI doesn't really matter for Nixpkgs and I don't think the NVIDIA driver has a stable API that they could break compatibility with
i feel like wayland will be stable when ipv4 is replaced with ipv6
i wanna say they've backported the nvidia driver before but idk if that was just due to build issues
don't care and just let NixOS stable handle it 
Meanwhile me, an Xorg user
looks like there was a patch for Linux 6.12 backported, but not the whole version
not surprising since NVIDIA is out-of-tree, has issues like this all the time

I have kernel 6.17 available to me now, is it worth installing at all?
i went back to stable but then glibc updated and i didn't wanna give up my editors from unstable so i switched again 
probably, idk
there's a bazillion little changes

Currently on 6.14 which is now stated as EOL
meme about x11 with no screen tearing joke
I see no screen tearing here
may as well wait a little longer and just install 6.18 lts
I doubt I'll ever get 6.18, it'll probably jump to 7.x next (this is an APT distro, Linux Mint to be specific)
did you mess with manually installing a specific kernel version at all
No
mint is based on ubuntu lts right
Oh wait you meant picking a kernel from the list? I changed it from 6.8 to 6.14
what list
The kernels list?
The reason I didn't stay on i3, I didn't want tearing across the entire window manager and all of the compositing is kinda jank
The kernels list
looks like ubuntu 26.04 is going to use linux 7 
if it's anything like Ubuntu there should be several "branches" of kernels for different purposes hardware
This is the kernel picker, it does exactly what it sounds like it does, picks kernels
Hem?
I'm of the opinion of yoloing a the latest kernel, but I'm at least on nix so I can roll back
i'm that guy running the latest kernel combined with the latest nvidia driver 
now the amount of times I actually update my system is something else entirely
(it doesn't actually break as often as I expected)
I like when my stuff just doesn't break at all
I expect a small amount of breakage, it happens when I'm naturally trying to do so much.
Recently tried Nvidia on Wayland again after hearing it has matured a lot
well it's an absolute shitshow
not usable
what did you try
because it was "usable" on Wayland years ago already
i'm curious though
it was horrible when they were still EGLStream-only but they fixed that a long time ago
I've already ranted about it before but x makes my mouse move funny and for that reason I can't stand it
tried to run KDE on Wayland directly on Nvidia as my primary card to see how good it runs
the result was lags like hell. Kwin ran super slow and it was just annoying.
works for me 
not sure
ye you have something misconfigured, idk
There was a nvidia KDE bugfix earlier that this convo stemed from
jsut out of curiousity, what specific nvidia gpu u using?
all i changed was the gpu config
kde plasma + wayland + nvidia is a bit of a brittle combo sometimes tbf
RTX 5070Ti Laptop
disabled Intel Xe iGPU and tried to go with Nvidia only
i fixed the power management issues
but i recently saw an open issue on nixpkgs about GSP lol
wonder if this is a prime being super fucky thing
i had prime disabled to test nvidia only
i also set the MUX switch in the bios to dGPU only

Only time I was disabling the igpu was due to trying to get VR streaming to work. I respect the testing though
However I do just recommend the offload method
lol
i had some really bad hard freeze issues and i was suspecting the intel xe driver
the intel xe driver is ass for my laptop too lol
i am getting these issues in my dmesg and journalctl over and over again as well as kwin errors
[38188.338579] xe 0000:00:02.0: [drm] PHY A failed to request refclk
[38804.728436] xe 0000:00:02.0: [drm] PHY A failed to request refclk
hate it
out of curiosity, do you mean the xe or the i915 driver
because the xe driver still isn't in a great state AFAIK
i have no clue lmao
fun fact when i change to i915 everything runs much much worse

i wonder if it for some reason used mesa for the gpu
its an intel 12500h, and a 3050m
dont ask me the specific naming
idk if nouveau supports blackwell in 25.11
sure hope it does
i remember when i tried the installer last year there was no support yet so everything ended up software rendered 
you were using i915 
the issues i was having was related to switching between dedicated and integrated graphics causing bluescreens, turning off the connected hdmi device causing blue screens, some driver reading protected ram or osmething causing bluescreens, and such
lots of bluescreens basically
Windows
probably
then ignore the i915 thing 
that's Linux-only
it's at 25.2.6 and they added support in 25.2 apparently 
im averaging like a bluescreen per 2 weeks or so
maybe ram issue actually, i should do a memtest
Or switch to Linux and watch if you get any Kernel panics 
oh that must suck
Classic
and i also need h264 support for davinci resolve
even if its 99% efficient, its not worth the hastle
that reminds me, i've been meaning to check nixpkgs for a winboat fix hopefully
for some reason it tries to build a weird mingw go version now 
a real WinBoat user in the wild? 
the ms office web apps are so garbage i ended up installing the desktop apps through it
Lol
Is WinBoat good? Never gave it a try
it tries to cross-compile for x86_64-w64-mingw32 and fails
works pretty well, though opening "just" the apps is broken for me so i just open the full desktop
im working on a ai voice model for tts, does anyone have any tips or recommendations of what i should do? i havent done something like this before so any help is appreciated
Get your CCNA at NetworkChuck Academy: https://academy.networkchuck.com
Remember when I hacked together a way to run Claude Code from my phone? That video blew up. Well… Anthropic saw it and said "hold my
beer." They just dropped Remote Control — an OFFICIAL way to start a coding session on your laptop and control it from your phone. S...
"hi claude code, what is the time now"
thinks for 15 minutes
"it's currently Janurary 1st, 1970"
not using opus 4.6 fast mode, what's the point 
vanorsigma spotted
why’s the neuro stream “slightly experimental”
is vedal testing smth or is it just named that way
been wondering that too
does anyone know anyone ever using Matlab app feature?
i just need to know how much effort I should put in this, it feels so useless
Now if we add xlibre as the next step up its peak
lel
Like, theyre doing things.. sometimes badly sure but like, massive.
Considering switching to artix for a while now but idk
I think ill hang around on arch until kde aint offering x11 anymore and then go to MATE and enjoy my extra 3GB of free ram
I have no idea if my distro will ever switch onto XLibre
Though if it's better than Xorg it would be peak
(mate is amazing ram wise.. 300mb for a fully booted, hassle free system is neat
KDE is planning to remove Xorg in 6.8 if I remember right
Many distros distance from xlibre either because "politics and coc issues".. which for open software imo is nonsense, offer the user choice god damn. Or they dont because redhat has their fingers in the funding or something idfk
6.8? I thought 6.9
But possible
What distro u running superbox?
Ahh cool cool
Wayland on Nvidia more or less just works rn. Its just a few edge cases from What I gather. Im using kde Wayland myself
Mint standart i think will keep x11 or xlibre, because of legacy support. But idk, they usually are fine but some choices are stupid
Well I hope they keep it, because Neurolings just does not work on Wayland
And I also like remote desktop
The mint team is working on Wayland, but it'll be awhile
I get told that many times but i have had nothing but workflow issues with wayland.. this could be a skill issue but i kinda doubt
For me Wayland is literally unusable and lacks basic features software I'm participating in developing will require from the windowing system
Yea shimejis are like.. borderline viruses(of the PUP type) and super niche so understandable
Shimes are peak though
Been using it exclusively for about a year without isdyes
Absolutely
Hard-requirement, if your windowing system can't handle Neurolings it's already failed
For a average user i bet it can work, but anything advanced or cool (patching around issues and shit) just doesn't work on wayland.. its against the nature of it
Yes the x11 software suite is ass and many things should be better
But god damn it, dont drop the entire desktop stack just because youre lazy to fix the protocol
Fix the damn protocol
Theyre adding hdr to xlibre atm iirc
I even saw that they were planning an X12 before they just decided to redo everything from scratch
X was use cases Wayland will never cover by design, but it doesn't deserve the hate it gets. Though most of its issues are br uase of gnome
GNOME should be deleted from existence and kicked out the Wayland board
A lot of the missing dhit is because of gnome saying no
Im still convinced big redhat and such are paying people to just force wayland and control user choice, because dropping user choice is such a corpo thing, not a linux thing..
Buut thats a theory nothing more
They don't deserve any power in ruining everyone's lives
Yes, we already have MATE to replace it
Wayland is way worse from the PoV of a software developer too
Like practically no support for global keybinds (imagine if you needed to click in your autoclicker window to activate it !!)
And the whole window positioning APIs being lacking resulting in stuff like Neurolings window interactions just being impossible, and can't do cool stuff for a little game that would really like to move some windows
Also the whole permissions nonsense preventing basic features like remoting into a system remotely because there's a local popup to confirm if you want to allow the remote connection !!
Like, id understand if it was like with a control panel to allow specific things to apps.. kinda like flatpak.. (i still don't like flatpak, appimage better, pacman best)
Yeah Flatpak is annoying, Flatpak Discord just doesn't work with copy paste file uploads
Pacman plus the ABS/AUR
DONT GET ME STARTED. GOD. I needed to setup remote desktop on a rpi, i had to drop back go old driver just to get TeamViewer and anydesk and rustdesk working.. a while ago i tried to get rdp working on wayland. Litterally impossible
APT is also based, apk is too, but apt suffers of many repos just being absolutely tiny (try using debian without importing any external repos lmao)
Its works just needs a login of some kind. Wishntheyd allow for gpg sign keys at least.....
TV, AD, RD all just dont work. Login works, screen capture crashes, or something the connection drops even before or the app doesn't launch, not even in a xwayland box. TV kiiinda works but like.. closed source of the subscription kind
rdp i tried everything i found online and I didn't get it to run.. if you could make a good guide PLEASE do.
Ive never had any of those issues. So as per normal on linux YMMV
I need to deal with this at work sometimes
We love "idk, works on my machine" xD but yea..
Lmaooo
One of the bad things about Wayland for developers is that there's just so many compositors
It's impossible to support Wayland generally because you need to basically implement separate support for every single DE
On X11 there's Xorg and Xlibre and that's it
No way ..
Ove nwen arguing with a dev about there use of conda because conda is just broken on my pc for some reason. It installs and run shit, but I always hit a corner case I dont hit with venvs
The Wayland makes the "works on my machine" problem just significantly worse because a different compositor may break an app entirely
(well, the x protocol.. apps usually dont care about if its x11 or xlibre unless theyre xlibre aware (or ig badly xlibre only..))
Yeah
Xorg and Xlibre basically work interchangably, which isn't the case between all Wayland compositors
The thing about x is that it only has one reference implementation. Wayland doesn't.
Only thing was drivers or something but even those i think have been fixed?
What?
On Xorg, an app that works in one place probably works in another place
If you go to Wayland though, something made for KDE may break on GNOME
X11 and xlibre-?
But also, ig valid, but same goes for rust.. but thats a whoooole different shitshow
Xlibre is just an antiwoke fork thays not being considered by any major distro
Not really. Theyre actually improving x protocol
Like said, adding hdr support and other things that were considered "Impossible"
Dude running it got kicked out for introducing regression and breaking shot and refusing to fix anything
I mean, a master Branch isnt a stable branch
But yea the dude had a rough start, no doubts
Fr9m what I gather it was less his code and him just being an ass
Dude was just actively hostile to everyone
From what i know he merged without testing, which was the main issue
Being toxic i Personally dont care, aslong as the damn code is good lol
Seperate the art from the artist or something
(this ofcourse includes no payment and ensuring that the art is actually good
That was one of the things he was actively hostile about. He just refused to do it, and was an ass to anyone telling him that
Possible, didnt follow too much, mainly got to see the advancements
soldering hurts, i keep burning myself 
i should get some clamps on arms to hold stuff in place
Bruh
How are you burning yourself-
Grab a paper towel and put the piece on it- and dont grip the iron at the edge, go further back-
And also keep cold water there to quickly run it over your fingers when you burn yourself
The faster you do the less the pain is obv
Also grab like tweezers so u dont hold the piece of metal that youre trying to heat up
And dont go to 500C, 290-390C usually is enough
this thing only goes to 230 lol
its tin
And 230 celcius ofc, right?
idk
Well yes, there is leaded and unleaded solder-tin or whatever
Should be on its label
(and then there is rosin core (hollow inside, has flux in it) and just solid tin)
Yeah that explains why you're struggling so bad
getting it to melt is not an issue at all
Do you have a temperature settable soldering iron?
the issue is not having 3 hands
nope
jsu tplug into the wall directly, 15W
Oh so like a plug in and go
Fair fair
Then yea, table, and tweezers or simmilar
Or a third hand claw thingy but theyre troublesome too
I assume you know, but add a bit of a solder blob onto your iron tip before you try to solder, it conducts heat much better that way
ye
Id personally invest at some point into rosin core tin (or external flux, its sooo much easier) and perhaps a settable temp iron
the most annoying part is im doing small electronics, and the flux has a big brush. so i immediatly put way too much onto it
this was 30 bucks for the soldering iron, the flux, and the tin. im going for cheap rn
Oh you have flux in a brush type.. i got a syringe
Cheap and small stuff, the darksouls of soldering
Gl
I tried soldering a nand smd package once
Like a smd dip8
With my 1980's soldering iron indistrial from wella
Its a paaainn
Luckly flux is magic
Lol it seems superbox and I are both cursed when I comes to python lmao
SIR PUT DOWN THE BUCKET OF GLAZE
Lmao
I actually love python. I just hate Conda
I HATE Python
must be time to refactor the neurosynth codebase to typescript then
I never bothered getting into conda, i just recently started using venv .. before i syswide everything and yes i write Python very C styled.
I ban you To the backrooms
Conda complicates shit too much. Its only advantage is it allows easy installing of parallel python versions
Theyre like smaller python embedded installs
Conda is just a Python embedded wrapper
I think
obligatory uv reminder
I have to use Conda for a thing cuz Sagemath
I hate conda. Its pure Satan
I can't say anything on this one as just using nix as a layer above
-> when the toolchain is so cooked you make a second one
Ain’t building that shit from src
Pipx is a godsend though
Tf is pipx?
I have only heard about it as a windows app install method
Pipx allows you to install any python program with a self-contained venv automatically
So like
uvx my beloved
A packagemanager
Winget, but.. python, and probably a lot more issues
Pip is package manager. Pipx takes thst and shoves it into a venv gor you
Cough, cough, triton.. cough triton-windows
Cant you just.. active venv and call pip?
Uv isnt bad. I just haven't used it much
UV is faster which is nice
Yes. For bespoke installs I do that. But for system wide applications I prefer pipx
Ah ig
Like cli stuff
Idk havent had to do system wide Python stuff yet
Seems like a mess.. id rather alias a activation script
Im a linux user that run arch. Cli is my fren
I also use arch normally
Rn i main Windows 10 because Windows 95 Development Fixation rn but yea
I mean if id start from scratch, sure ig.. rn it seems to me useless to learn it when i already know the linux native methods
Like i said most suff I just write a launcher wrapper that activates the venv. I only use it for cli sthit I was systemwide






