programming

[fast pagoda]

lol i remember reading a news article about this

[verbal grove]

dont blame ai blame the prompt

[opaque sigil]

blame both

[rough bloom]

SP5 still in demand, SP3 old enough to be relatively cheap

[opaque wharf]

To be fair, the dude is actually a competent person lol

[fast pagoda]

if you arent spending 16 hours setting up the most custom framework of guardrails possible only to abandon the whole thing later then you aint vibin

[verbal grove]

ye

[true hemlock]

if only i have sp3 chips just lying around

[rough bloom]

you don't?

[olive sable]

[fast pagoda]

kekw

[fast pagoda]

no but he has an intel atom

[rough bloom]

Intel Atom + PCIe switch

[verbal grove]

i really need to make an auto restart for my claude code required programs

i wonder how hard thats gonna be to do

[true hemlock]

honestly i could probably hit some 7763 for like $600 each

but the problem is the ram

the fucking rammmm

[verbal grove]

buy ram sticks directly from china

[true hemlock]

i have like plenty of 32gb ddr4 dimms but none of them are ecc

[verbal grove]

thats what im doing

[fast pagoda]

save me cxmt

you're my only hope

[true hemlock]

intel atom with avx512*

[rough bloom]

wouldn't that also be even worse with DDR5? Or do you already have DDR5 RDIMM lying around

[verbal grove]

its nice having contats with chineese manufacturers

[true hemlock]

how else am i running my current sp5

[fast pagoda]

:

[rough bloom]

in use

not lying around

[warped narwhal]

but consider...

;

[true hemlock]

i mainly wanted the sxm5 interface for scalability

but oh well

i don't even know if sp3 boards have at least sxm4

[rough bloom]

but yeah if you already have RAM that you can take from another machine then I guess just go SP5
RAM prices will probably make the difference between SP3 and SP5 negligible otherwise

[verbal grove]

https://cpuprocessorchip.sell.everychina.com/p-108650578-emcp-memory-chip-kmdh6001dm-b422-64-32-emcp-d3-lpddr4x-3733mhz-emcp-emmc-memory-chip-storage.html hmmm...

very tempting..

[fast pagoda]

i dont like to brag fellas butg

i recently purchased a cpu so powerful

the verge (respected system builder news outlet known for h igh quality youtube building videos) called it

[verbal grove]

what cpu is it?

[fast pagoda]

"....a.................cpu...."

[olive sable]

let him cook

[true hemlock]

60 of those atoms

[verbal grove]

ah

not a bad one

then again im looking for a new threadripper later

[fast pagoda]

lol i just needed a block of metal to put in that socket to use the ram and put that in the box under it

it was only liek 60 bucks

[true hemlock]

hey mine is more powerful ya see the pins are big and they need less of it which means its better!

[verbal grove]

not bad

[fast pagoda]

oh shit that thing beats pentiums

that's crazy

[rough bloom]

damn
I assume the 64 stands for 64 cores

X2 so 128

[fast pagoda]

64 terabytes of ram

[verbal grove]

hey boys

spot the problem

there is one

[fast pagoda]

yor visualizer indicates that your music is peaking above clipping

this case is a marvel of engineering

box

i am building this pc right next to where my cat sleeps at the window

no esd risk detected

hard drive cage encroaches on the mobo area enough that i have to take it out to put the mobo in

nice

[opaque wharf]

https://www.youtube.com/watch?v=AZUp5nY7BWU

[warped narwhal]

I see your athlon, and raise you...

My keychain

[fast pagoda]

i only have a usb4 cable on my keychain

[noble zodiac]

I hate those clickbait titles

[warped narwhal]

I have other CPUs as well

[opaque wharf]

It is not

A malicious site with CSS can actually run arbitrary code

It's a 0-day

[fast pagoda]

If it were 2005 id be living luxury rn

[noble zodiac]

no one hacked CSS. Chrome has a bug in how it parses CSS. Those are very different things

[opaque wharf]

Well, fair point

[fickle rain]

https://xcancel.com/gerhart_x/status/1966781348724478240#m

[fickle rain]

It’s a techtuber innit

[true hemlock]

i9 14900HX, i9 13980HX, i9 13950HX ES, I9 13900 ES, i9 12900 ES, 7950x ES, 5700g, 3700x, 3600x

the 14900HX a d 13980HX are modt so soldered on a board

[fast pagoda]

the math is mathing

[fickle rain]

https://www.cs.ru.nl/masters-theses/2025/J_Jagt___Analysis_of_Windows_Secure_Kernel_security_bugs.pdf

[fast pagoda]

A sobering reminder

[olive sable]

are those pins bent?

[fickle rain]

We found that the most helpful resources for setting up and developing a custom kernel driver come from the game cheating and modding community besides the Microsoft Learn online resources.

[opaque sigil]

nah that's just some image artifacts

[verbal grove]

@fast pagoda how do i get the monthly thing for claude-monitor?

[fickle rain]

What kinda mods require driver? I think that’s more cheat territory

[burnt marsh]

anyone with experience with the neuro api, what do you reccommend for an integration with fnaf 1? im thinking python just using pynput and moving mouse and arrow keys to control the game, but that feels flawed and possibly could be exploited by neuro if i dont code in some restrictions

[fickle rain]

How do FNAF mods handle things?

[burnt marsh]

im not sure, ive been looking around and havent found many, other than ones that just take the entire game and change it instead of an add-on

[fickle rain]

You could use scheme like bridge between Neuro communication and game like Pokémon does

[burnt marsh]

ill keep looking around though

[fast pagoda]

fthe pins arent bent

it was a recycling logo on the plasticv

i need 2 more xeon to make a sooper funny image

[burnt marsh]

could you elaborate on that further? im looking at the docs and am getting a bit confused. as i see, you have to inject a lua script into the game for it to integrate

unsure if you can do that with fnaf

[opaque wharf]

Why do you have that many xeon lying around unused

[fast pagoda]

only motherboard viable is 2k

[fickle rain]

IIRC it uses some JS script to actually talk with Neuro via socket
Script then reads/writes stuff in text file which is read/written to by game as well

[burnt marsh]

yeah, i was thinking using js to talk to neuro, and also writing commands somewhere which is read by a script that controls the game, but im unsure about the controlling game part other than pynput

but also mixing javascript and python doesnt seem like the best idea for a simpler solution

[verbal grove]

@fast pagoda ehem...

yeah

sounds about right

[fast pagoda]

you said it is 8 people right

[verbal grove]

yep

[fast pagoda]

i thikn when the world burns you might be on the chopping block when they ask "BUT WHO WAS USING THAT MANY TOKENS??"

[verbal grove]

yep

im definetly one of the reasons claude has to scale so often

or better said nova-development

[opaque wharf]

I hate that AI agent is using my cloudflare worker free resources

Go crawl somewhere else you clanker

[verbal grove]

meh

[opaque wharf]

My site has only been up for a few hours and look at this

[verbal grove]

yeah that looks about right

[opaque wharf]

Every single one of those request must be authenticated and it will eat up the KV daily limit

[fast pagoda]

filter them

[opaque sigil]

i hope you turned on the ai bot crawler thingy already

[opaque wharf]

Yeah, I just did. I realize that they are not turned on by default

[opaque sigil]

[fast pagoda]

i forgot that coolers come with cpu

i havent put one of these on in like 10 years

why the tf does the screws not reach

the fucking hoels

[hollow matrix]

It's been a day and it's still fucked

[fast pagoda]

i dont want to put the stupid h100i in the closet on this thing because that's so unnecessary but i might be outskilled by the wraith

[opaque sigil]

might need standoffs?

[fast pagoda]

i thought maybe

but they block it

and the little sheet shows bare holes

[opaque sigil]

there's usually the backplate that you then screw into

[sage crag]

why are they called standoffs when they stand up

https://tenor.com/view/smh-anime-girl-gif-12459523042458670621

[fast pagoda]

not for am4

[olive sable]

because it stands at on offset

[sage crag]

of in out the hot food

[hollow matrix]

I gotta say, ai overview, I don't think restarting my computer is going to fix wacom's shitty website

[fast pagoda]

it might be just that im supposed to jam them down

but it feels so incorrect

they are spring loaded for tension

omg if i am then that's absurd

it takes like

my bodyweight

to pushed them down

[opaque sigil]

nah that can't be right

[sage crag]

weigh three teacups and a mushroom

[fast pagoda]

lmfao

[opaque sigil]

okay you do need the backplate

i guess they just screw directly into it

[fast pagoda]

im pretty sure it's built in

there is a backplate behind it

[sage crag]

FOCUS

[opaque sigil]

it should be loose

[sage crag]

use

use water

[opaque sigil]

and then you screw into it

[sage crag]

water can

float

[opaque sigil]

just like how the default standoffs are screwed into the backplate

[sage crag]

[olive sable]

https://tenor.com/view/watercan-gif-26791122

[sage crag]

float64

[tender river]

water can also sink

[olive sable]

https://tenor.com/view/sink-cuisine-amenagement-equipee-design-gif-8476897548120197244

[fast pagoda]

ok yep

[tender river]

also sink can float

and float can sink

[fast pagoda]

there is a backplate

but

the backplate

when i took the previous mounts off

fell to its knees in the dennys bathroom behind the mobo

[opaque sigil]

put it back then

[fast pagoda]

I just wanted to showcase the custom glue job they used

[opaque sigil]

[sage crag]

tosto

400 degree tost vs knife

[rough bloom]

I have this thing too
somewhere

[sage crag]

officer officer

[rough bloom]

awa

what

[sage crag]

someone stole my croissants

[opaque wharf]

Lady awa, are you high?

[rough bloom]

report too vague

closed

[sage crag]

no investigate

[rough bloom]

sorry underfunded

[sage crag]

https://cdn.discordapp.com/attachments/1300807614032121950/1409945130810609695/3dgifmaker99373.gif

ill raise your wages

[midnight sigil]

please invert the current view matrix for world vector to screen position function

[fast pagoda]

[rough bloom]

consider it done

[sage crag]

please invite yourself to leave

[midnight sigil]

I kindly reject this offer

[rough bloom]

@olive sable return kni croissant

[sage crag]

@olive sable

give

https://tenor.com/view/my-jolly-reaction-jolly-my-honest-reaction-gif-5490890317762380451

[olive sable]

hi

i

uh

i

the croissants

are

uh

i

gone

[sage crag]

death penalty

[olive sable]

bye

[midnight sigil]

no tide lock, bad enub and kni

it means you and enub hasn't formed long enough relationship, and was unable to establish a stable gravitational state to achieve a tidal lock.

[olive sable]

[midnight sigil]

you need to love eliv more

[fast pagoda]

i knew someone had said they had these

i received another delivery lol

[burnt marsh]

can anyone assist me with reading a clickteam fusion 2.5 game state? my options seem like reading the things in ram or using a DLL, both seem very bad for a "just works" game integration

[fast pagoda]

what the hell is an aluminum falcon

[midnight sigil]

I mean, you either read externally, or internally

both requires you to have an offset/signature to what you're reading

there're no easier alternatives

[opaque wharf]

Excerpt from Cloudflare Docs:

A navigation request is a request made with the Sec-Fetch-Mode: navigate header, which browsers automatically attach when navigating to a page.

In your guys opinion, is it better to rely on this on just use explicit routing to run worker first?

Especially for SPA with API on /api/*

[fast pagoda]

are you trying to limit bots? use security rules//crawler rules /WAF rules

managed challenges

[nocturne olive]

Clickteam Fusion is extremely easy to decompile
You can decompile the game back to project form and read the state directly by adding your code there

[fast pagoda]

and then rate limit shit like /login /signup /api/*

https://developers.cloudflare.com/waf/custom-rules/

https://developers.cloudflare.com/waf/rate-limiting-rules/best-practices/

[opaque wharf]

No, I'm trying to minimize actual worker invocation whenever possible. I did have a data on how often the sec-fetch header is used but not the explicit routing mode

[fast pagoda]

well that would i hope be auth'd so these would solve that problem

it would never reach the worker

[opaque wharf]

The worker is the one performing the auth check Cloudflare limit their JWT verification to enterprise only

a.k.a "fucking expensive"

[burnt marsh]

that seems like a good idea, but im unsure if that would work well as an integration because i'd have to post the modded version of fnaf on github and have vedal use that, which i dont know if that violates copyright or not

[opaque wharf]

Like, not even their business plan got JWT verification rules IIRC

[fast pagoda]

how are you accessing kv

when i first used it i didnt use the correct type of websocket and it held the shit open adn ruined my life

[opaque wharf]

From the workers binding. I am not concerned about each and every one of the resources the worker is binded with, just need the workers to be invoked less often

[fast pagoda]

i just imagine there shouldn't be a reason to access kv to do an auth

i mean you can

but there's others

[opaque wharf]

Also this one is different than the one before lol, so I didn't use KV now. This uses D1

[fast pagoda]

split routes so static/public traffic bypasses Worker entirely, then put WAF + rate limits on the worker paths only

[opaque wharf]

Yeah so explicit routing

Not rely on the sec-fetch-mode

[fast pagoda]

cloudflare access perhaps? turnstile

[opaque wharf]

That's already implemented too

I was just asking here in care anyone has tried both of the routing mechanism

[fast pagoda]

are the workers broadly attached to /*

[opaque wharf]

Because I only ever tried the "sec-fetch-mode" route

[opaque wharf]

Not really no. There are different kinds of routing behaviour

The sec-fetch-mode doesn't invoke worker if the user-agent sent those headers

Which for a browser, they usually do

The other is explicit routing

We tell where is the path to invoke the worker

[fast pagoda]

sec-fetch will cause more invokations than a branching header or binding explicitely (best)

sec fetch by design can only really be useful inside the worker already

for this

[opaque wharf]

Yeah, I have my suspicion hence why I asked if anyone has actually tried both of those. My experience with sec-fetch is indeed quite bad because of badly behaving client that are not AI agent

[fast pagoda]

explicit is really simple and is best for this im pretty sure from last time i had kv go boomy

unless something changed

just check endpoints and only bind it on the particular routes you specifically need it on and then just return not found if they request a spooky path

[astral ingot]

can i program my own neuro sama?

[opaque wharf]

You may want to see the pinned post here

Specifically this one
https://github.com/igglyn/NeuroProgrammingFAQ

[fast pagoda]

llmao i was looking for that

[olive sable]

https://github.com/igglyn/NeuroProgrammingFAQ/raw/main/FAQ.png

[sage crag]

java

pssk

[opaque wharf]

What's wrong with Java?

[olive sable]

psssskkkk

[astral ingot]

thanks

[fast pagoda]

the best AI exclusively use C# with .NET 4.5

[opaque wharf]

https://en.wikipedia.org/wiki/Java

[midnight sigil]

books written on the island were named Javascript

[opaque wharf]

https://en.wikipedia.org/wiki/Javanese_script

Coding JavaScript using Java Script

[astral ingot]

sorry but can i not create 2nd neuro sama?
it says here

[opaque wharf]

Well that's the thing. You CAN but you have to understand what it takes to do that

[astral ingot]

i just started learning python

[opaque wharf]

And no one here could help you achieve that

[burnt marsh]

you can, its just not really worth it, unless its just a for fun project. also you have to know what it takes and know how to do it, because a lot of people here wont help you with it and have no incentive to

[opaque wharf]

You can ask coding question here at best, and help troubleshoot issue

[astral ingot]

oh i see

[opaque wharf]

Also you can ask and discuss architectural choices. But don't expect someone to help design it from the ground up

[astral ingot]

even homework questions? like we just started learning python

[burnt marsh]

sure

[fast pagoda]

anyone theoretically COULD create a 2nd neuro

but they won't

not for lack of trying

[opaque wharf]

Also mandatory this
https://en.wikipedia.org/wiki/Sundanese_script

[burnt marsh]

the thing that lead to neuro's success is the community. that wont really happen a second time

[fast pagoda]

among a million other things that lined up

[burnt marsh]

true

but community is a HUGE part of it

[fast pagoda]

timing and vedal used to be able to produce things in less than 3 years

are big ones

[midnight sigil]

Ved spent 8 years with Neuro

[fickle rain]

Luck and community management

Luck is the hardest part

[fast pagoda]

and just overall vision for it + ability to execute it well enough to put something out when he did and for it to work as well as it did with llms that size being more novelty at that point

[midnight sigil]

it's fine guys, just instsll Cursor, buy claude tokens, everything and anything can be achieved

[fickle rain]

Very true

@codex You are an expert software engineer. Implement AI anime girl for me.

[rigid snow]

unfortunately it really does feel like that

[fast pagoda]

i havent even thought about cursor in a hot minute

wonder how that's going

[midnight sigil]

people do buy for it

I prefer my cute little pilot in visual studio

[fast pagoda]

This thermal pad just detonated when I touched it

[midnight sigil]

they look like burnt coal

[fast pagoda]

Good thing I kept my spare PSU and its cables with the Christmas stuff

Very organized

[midnight sigil]

the belt

[fast pagoda]

the one belt that i keep regularly putting back on the belt hanger lol

I think this is the first time I've ever had a screw in every stand off

[stray dragon]

the motherboard itself has a fan??

[fast pagoda]

yeah on the chipset

i think it's supposed to use that to like pull air over the nvme drives

that are covered once it's put all the way together

[opaque sigil]

is for cooling the chipset

and i guess making jet turbine sounds if you don't put it on silent mode

[fast pagoda]

well that too yeah

it's right on top of that i figured it was a given

Never going to find these screws

[opaque sigil]

[fast pagoda]

the violence has escalated

i cannot hear father

i need more headphon

[opaque sigil]

good thing most of these are like $20 or so

[fast pagoda]

it's kinda upsettingly not that lucky

[rigid snow]

40 ear guy right here

disposable vapes nowadays are insane

ewaste brick

[opaque sigil]

note to self: do not uninstall sddm while it's running

[rigid snow]

why would you uninstall a display manager

while it's running

[fast pagoda]

the ones on the right zs10 pro/prox zs12 prox/cca-c12 are $45-65 in there
the fiios were like 200 nbut theyre older, aria 2 like 80? pisces/dqs are in the 20s, the small boxes are anywhere from $5-15 usually, the bluetooth thing was 20, as16 $70, the pr2 $45 (bought twice by accident lel)
duonic $30 bought twice by accident, space travel like $25 but i have 4 sets ( 🤪 ) space travel 2 also $25

[fast pagoda]

fun and profit

[opaque sigil]

replace it with the new plasma login manager

[idle dune]

They sound pretty good I use mine at the gym

[rigid snow]

i don't think it's a good idea to uninstall any dm

[fast pagoda]

i like them

[rigid snow]

from within

[fast pagoda]

they compare favourably to the zs 12 pro x that are kinda their sibling

[rigid snow]

i should turn it into a web server when it's done

[fast pagoda]

juist open tty3

or w/e
"

[rigid snow]

ye

"

[fast pagoda]

please note it was 2x ' not "

[rigid snow]

mb

''

[fast pagoda]

thanks

[opaque sigil]

sometimes you have a brain lag moment and type nixos-switch instead of nixos-boot okay

I'm surprised it just happily killed it like that though

[fast pagoda]

This shit was almost so cash

[opaque sigil]

are those capacitors

[rigid snow]

the connectors on the cable ended

but yeah funny capacitor board

[fast pagoda]

Yeah I was 1 molex short

Surely 1 molex is supposed to feed 8 hdd

I'm not actually sure because I've never tried that or cared to find out the limit

And usually I've had data power

Sata even

[solid bough]

And I'm confused why the EXE works now ever since I removed the DLL files that are packed with the code in the EXE

And it is on the NVIDIA Quadro M2000

[fast pagoda]

Medusa has invaded

I like how accessible the sata are with the cage in

Although if that's a plug in once and use backplane thing

I guess*

[solid bough]

It finally WORKS:

[fast pagoda]

edge detect

[solid bough]

Idc how bad the output of the model is rn, I'm happy it works on PC's with NVIDIA cards and on PC's without NVIDIA cards

[fast pagoda]

looks good to me

[solid bough]

Well, the model isn't great for the moment.

[sage crag]

works on "A without B" and "A with B"

so why not just "A"

[solid bough]

Well, now it works on A and B and C

[sage crag]

dentge

[solid bough]

A is my Main machine, B is the testPC with NVIDIA and C is a VM with no NVIDIA stuff

[sage crag]

dentge

[solid bough]

So now I only would need a PC with a AMD GPU and one with a Intel GPU

then my testing is done finally

Sadly I don't have enough PC's to facilitate that

[rigid snow]

!!!!!!!!!!!!!!!!!!!!me

[solid bough]

Well, then there is the situation of "It works on my PC"

If I had 1 more PC and a AMD GPU and a Intel GPU and I would also need a SSD for the system I could test it on there too

brb

[stuck ocean]

https://tenor.com/view/transmission-transmit-post-this-cat-cat-least-expect-it-gif-10319838660013923911

fuh no embed perms

🥀

[rigid snow]

my pc has and doesn’t have an nvidia gpu

it works on mine

[sage crag]

[rigid snow]

nvidia gpu superposition required

[fast pagoda]

Idk why this came to me and it looks like shit but i had to follow through anyways

[rigid snow]

wgatis that

[fast pagoda]

the shit he masked

or edge detected

[rigid snow]

oh

[rigid snow]

works is a word

[fast pagoda]

It looked more omegalul shaped before I did it

[warped narwhal]

https://tenor.com/view/neurosama-neuro-sama-vedal-vedal987-vtuber-gif-15278412527939419909

[fast pagoda]

Walking ESD REALLLY wants to be in case

[solid bough]

I got you: .

I checked what it is beforehand though

[solid bough]

Well, Now I need to further clean the dataset and while doing that trying not to go insane

[fast pagoda]

dataset

not go insane

o7

[quick condor]

I'm sad Sony discontinued my headphones. I love them

[solid bough]

Welp, Dataset time it is, but before I start that, What GPU should I choose?

I can choose 2 GPUs :

0: RTX 3070 Ti
1: GTX 1050 Ti
Anything above 1: CPU

[quick condor]

For anyone capable of figuring out how to use it: https://huggingface.co/hascrack/Minivoxtral-3-14B-Reasoning-2512_ASR
Theoretically you can do qlora training on it, asr, and vision but you need vllm

[fast pagoda]

surely rocm will work

[fast pagoda]

both

[solid bough]

I can't do both

Since it is a RTX 3070 Ti paired with a GTX 1050 Ti on the same Motherboard

[solid bough]

Now I can choose between GPU (0) or CPU (1):

[fast pagoda]

Well she doesn't have an issue powering on so that's good

God that damn fan is loud at the start lol

[fast pagoda]

Time them both against the other

[solid bough]

How?

[wise kite]

what are u doing rn?

[fast pagoda]

like do a run with one then the other

[solid bough]

I run python code that uses tensorflow that has been packaged with pyinstaller which was a chore to finally get running in 3 different enviroments.

[wise kite]

for what is the script?

[solid bough]

Oh I thought you meant at once. I tried before and the CPU was always slower in performance

[solid bough]

Black or White doodle to face:

[wise kite]

why do u need 3 enviroments?

a hade an ai that can see my screen and i only needed 1

[solid bough]

1: Developement
2. Test with NVIDIA CUDA and cuDNN
3. Test without NVIDIA CUDA and cuDNN

And now all 3 work

[wise kite]

do u know mutch about cuda?

[solid bough]

Not really

[wise kite]

cuz i want to make a cluster for ai training can i use it of only 2 from 4 divieses can run on cuda

[solid bough]

Well, I just wanted that it runs. And the 3rd "device" was a VM. And you may see on the image that it has something odd

Well, the first odd thing is the 3 Taskmanagers

[wise kite]

all of them are real devices and its headless training

[solid bough]

Nope, Training was done at the developement PC

[wise kite]

i mean for my does that work

[solid bough]

Okay, And in the image where it runs on, it is a VM

[wise kite]

cuz thunderboalt 5 is expencif so a need to know if it will work first

[solid bough]

I mean that image

[wise kite]

no VM

[solid bough]

I mean on my side

[wise kite]

ah

but why did u use a VM?

[solid bough]

My testing was:
1: The Developement PC where I initially tested it with CUDA 11.2 and cuDNN 8,1 installed
2: The Test PC where CUDA 11.2 and cuDNN 8,1 is installed
3: The VM where no NVIDIA Software is installed

[solid bough]

I ran out of Physical Desktops I can test on.

I could certainly put together last PC again but then I would need to shuffle some data to the right drives, fix the virtual enviroment, and install Windows 11 on the SSD again.

[wise kite]

i just use ubuntu for evrything exept for my main pc cuz idk if all my apps will work so just for testing ubuntu is the best coice

[solid bough]

I'm not that much of a linux guy tbh

I don't hate linux, but I just don't wanna deal with it right now

[wise kite]

its mutch better for me in minecraft i got 4x fps

u prop dont know how neuro-sama works right?

cuz i dont know what to do now with my ai

[hollow matrix]

Ghost ping

[kind fable]

sorry

discord did black magic

[hollow matrix]

[kind fable]

replied to a message i didn't select

[quick condor]

Me but I had to move to ubuntu for training/serious dev because I cant run pcie TP on WSL2 so it was a major limitation for training and runtime... Im on cuda 12.8, cuDNN 8.6, patched 590 drivers and python 3.14.3t -xgil=0

[wise kite]

what is cuDNN?

[solid bough]

What do you mean with pcie TP?

[quick condor]

subset of cuda compute capability, Im on ampere so I can use compute 8.6

tensor parallelism, specifically tensor parallelism over pcie

[solid bough]

Oh

[solid bough]

I train the model only on one GPU

[wise kite]

dont u need a fast conection for tensor parallelism

like at least 10gb/s

[solid bough]

[quick condor]

PCIe 3.0x16 is 16GB/s bi-di the big thing was I needed to mod the driver to enable

[wise kite]

do u use lan?

[solid bough]

If I were to do that, my GTX 1050 Ti would really drag down the speed

[quick condor]

Local to one machine, Most training I do is on 2 3090tis but I also have an A4000 in the system

[wise kite]

how mutch better of an llm could i run on a amd 9 7950x3d than on a quadro p4000 8gb

cuz the cpu is a beast with 16c and 32t but the gpu is cuda

[solid bough]

Valiprogemer does it matter that I'm connected with lan to a Netgear GS108 network switch that is connects to the WiFi Mesh repeater when I remote to a PC that is connected to the same Netgear GS108?

[wise kite]

idk prop not cuz the other pc is in ur lan

[quick condor]

cuda is the GOAT RN for NN work. Amd works but its lagging in support

[solid bough]

Yes, but I wondered if it only goes through the Network switch or if it goes to the router first.

[wise kite]

do u have a good llm recomadation cuz my gpu is 100% used by my vlm so i have to run the llm on my cpu but its 1.3b model so its just stupid

[wise kite]

just thru the switch

[solid bough]

Sweet, that explains how I got like 990MBit/s on taskmanager on copying.

[wise kite]

i switch is only 400mb/s : (

not good for a cluster

[rigid snow]

400mbit is a weird speed for a switch

[wise kite]

it was like 25$ so

[rigid snow]

no i mean like that's not a standard speed

[solid bough]

Damn
My switch can manage at most 16Bbps when all ports used.

[rigid snow]

it's 100mbit then 1gbit straight

[wise kite]

realy?

[quick condor]

For an 8gb gpu your going to be tight regardless with vision and llm. Smallest Ive used is an 8b at W4A16 Quantization and that was 7gb for weights without vission... Maybe Ministral 3b?

[rigid snow]

for ethernet yeah

it's 10mbit, sometimes 25mbit, 100mbit, 1gbit, 2.5gbit, 10gbit etc

25mbit isn't spec either but it's common enough

[quick condor]

I have stacked s4810s as my switches

[wise kite]

i was ganna use llama3.1:8b but i dont have the vram cuz of the vlm llava:7b so i dont have any vram left if i use my cluster will it be better cuz 400mb/s is slow and the best gpu in my cluster is a gt 420 1gb

[trim valve]

usb2 speeds?

[solid bough]

Ah, USB to Ethernet I guess

[wise kite]

no i dont use usb to lan

[rigid snow]

yeah but that's 1. 480mbit and 2. usb not ethernet?

[trim valve]

yeah but a usb ethernet adaptor or really shitty network card could be in the mix

and missing 80mbit wouldn't shock me

[rigid snow]

a really shitty network card would not be 480mbit

100mbit max

i have a $4 network card it's 100

$4 new

[wise kite]

and i dont use a lan card i use the port on the motherboard

[rigid snow]

do you mean MB or Mbit (Mb)

[wise kite]

i just realysed its a 5gb/s switch for 15$ not 25$

[quick condor]

really if you want vision and a model, Ministral3 3b has a basic vision model built in. I dont even use 8b anymore though, I run a 14b+24b+24b

[wise kite]

MB

[rigid snow]

400 megabytes per second is 3gbe

around

[wise kite]

i need good vision so im prob ganna user the same vlm but idk if my amd 5 4500 can handle a 3b model

[wise kite]

what is 3gbe?

[quick condor]

what vision model?

[rigid snow]

gigabit ethernet but 3gbe specifically is not a thing

[wise kite]

llava3.1:7b

[wise kite]

if u mean cabel i use cat 5

[rigid snow]

cable doesn't matter you can do 10gbe over cat3 if short enough

[wise kite]

i wanted qwen-vl but i didnt get it to work

[rigid snow]

i have a shitty 5e cable that doesn't do a gigabit which 5e is supposed to

[wise kite]

i have a 2m cabel to the switch than a 10m cabel to enuther switch then enuther 2m cabel that goes to my pc

[rigid snow]

nightmare

[wise kite]

wdym?

just need may ports

[rigid snow]

each cable and switch is a bottleneck potentially

[wise kite]

cuz the switch is just a 5 port

[rigid snow]

seems to be fast enough for you so

[quick condor]

llava-v1.6-vicuna-7b?

[wise kite]

its all the same cabel and switch so it prob shoudnt

[wise kite]

yes i mix it up with the llm cus the llm is the llama3.1:8b

[trim valve]

monka

[quick condor]

That vision encoder scores the same on MMMU as ministral 3 8b. You shouldnt need a second model. The base model has a vision instruction set + normal instruct (ministral 3 also has reasoning if you want spicy)

[trim valve]

why is it slower on my usb3 port

whatever networking is fake anyways

[wise kite]

https://vm.tiktok.com/ZGduR4rtS/

[wise kite]

My pc belike with the vlm

[quick condor]

what gpu?

[simple sky]

I love defunctionalizing code.

I love turning everything into a data structure

[opaque sigil]

thanks codex

[quick condor]

I had claude create a --killall orphans command

[patent shard]

compressed

[rigid snow]

🗜️

[patent shard]

I took it from the section at 0:51

[patent shard]

-preset 1 -svtav1-params crf=68
going to take like 20min to compress because extremely low preset but let's see what this cooks up

qp 63 is the lowest quality but I want the dynamic qp adjustment
so crf 68 it is

[silent cloak]

some dumbass tried to ransomware my pc lmao

man was in for it

[patent shard]

wait, crf 70 is the lowest
I'll do that

this is a partial encode of crf 68 but.. why is it so skippy at 0:15
I just looked at the original yt video and it looks like it's playing at 24 fps...

[opaque wharf]

Damn, I'm amazed at local LLM with 1.2B param

Finally a model that can load on my measly 16GB shared memory laptop

[azure lynx]

the wobbly slow pull out? probably not doing motion updates until it reaches a threshold and then it does it all at once, and the slow pull and pan makes it happen at different times on the screen which makes it feel not-smooth

[quick condor]

IDK if I could do 1.3b anymore (other than the embedding model I use)

Idle vram usage

[olive sable]

"idle"

i think we have different definitions of idle

[sick owl]

Ended up ditching my 4090 oc and going back to stock clocks

Only yielded about 3-5% real world improvement

[patent shard]

kind of what it seems like, yeah...
whelp, this is going to be ass.. it's 1080p but crf 70
it does... NOT look 1080p
I'll add some low bitrate parametric stereo audio, too
eh, maybe not.. I'll do low bitrate stereo

[warped narwhal]

at crf 70, are there any frames that are actually full 1080p?

[olive sable]

The Honda CRF70 is a popular, reliable, air-cooled 72cc four-stroke youth dirt bike featuring a 3-speed transmission with an automatic clutch

[quick condor]

real world, boost clocks do a pretty good job on their own. I have a +110/1250 that is good enough on my 3090tis but thats on water blocks

[patent shard]

probably not I think crf above 63 are more designed for 2160p and up
and also the success of it will heavily vary on content

[sick owl]

I was running +125/+1100 but that only translated to like 59 fps vs 61 fps and mins going from 54 to 56

[opaque wharf]

I think older GPU is better for overclocking than the modern one

As in, the gains vs effort is worthwhile

[quick condor]

More that the modern ones are able to dynamically boost so you are just optimizing it

[patent shard]

3.45mib
32kbps audio sounds crunchy

[olive sable]

still sounds mostly fine

[patent shard]

massive ringing audible in the beginning/end with that instrument, lol

[azure lynx]

very blocky when spinning in the clouds. but otherwise "watchable"

[patent shard]

not really, I can't tell what the heck is going on a lot of the time

guess that's what 2.36mib gets you

[azure lynx]

i find the anime style they use is feels glitchy for motion anyway. ;/ the timing isn't perfect in the original

it's not synchronized with the beats like i expect in a MV

[patent shard]

I did use the flac audio from spotify, but I lined it up in Audacity with the audio from the yt vid... so it should be as it's supposed to be

[azure lynx]

too me it feels like a music video made from parts of a film or series, like they didn't try do lip sync or make the actions hit the beats

and it's not like they're just always off a fixed amount, it just doesn't seem to have that kind of flow to it where things feel connected

[patent shard]

I wish I could use exhale, it's a better encoder than apple aac
isn't there something better for he-aac than apple aac, or is that only when it comes to using parametric stereo
which I guess I could use... seems a little silly to use parametric stereo at 32kbps, but maybe

I mean, I could use exhale.. but decoding compatiblity would go way down

and I don't wanna use opus, it's too noisy

[azure lynx]

doesn't matter how good it sounds if nobody can decode it

[ruby star]

four hours debugging
it was an or that should’ve been an and

[azure lynx]

logic is hard.

[ruby star]

Seeing is harder

[azure lynx]

more than once i've rewritten the same buggy line of code easily 10 times before i finally read what i wrote and facepalmed.

[fast pagoda]

Seething is hardest

[warped narwhal]

wait, when did windows finally start recognising .tar as an archive format?

[patent shard]

this probably sounds a lot better
still apple aac, but 40kbps instead of 32kbps
main benefit seems to be that it didn't start he-aac frequencies until about 7khz, instead of about 5.5khz

[fast pagoda]

Copilot is bash native so they added it

Aka the model prefers bash and it added it

For convenience

Since that's who maintains windows now

Slowly it will all work

[shy jetty]

i wonder i know that there is people who asked this question before but im too lazy to search
does neuro use linux or windows?

[opaque wharf]

Moving to 10B models gains me a 10x slowdown from 1.2B model

For no apparent benefit IMHO

[fast pagoda]

Unity is windows who fuckin knows what her inference backend is

Could be a myriad of things

[azure lynx]

windows. watch the dev streams and you see a windows desktop. and sometimes neuro runs on that box, sometimes another.

[fast pagoda]

It's windows for vedal anywYs

[opaque wharf]

Brother, you nuke audio for a living. I think every sound will sounds just fine to you

[patent shard]

oh, yeah, waaaay more punch on the drum and less ringing on some instruments

[shy jetty]

aren't dev streams streamed from vedal's computer?

[azure lynx]

yes.

[shy jetty]

yea im more interested in neuro's pc

[young plover]

Still Windows

[azure lynx]

i guess the fact that i meantioned she sometimes she runs on vedals box doesn't mean anything.

if she sometimes runs on windows, she probably always runs on windows

[shy jetty]

maybe your right, i still don't know how unity works idk if it can in both linux and windows

[young plover]

He mentioned recently that he'd migrate to Linux if it was feasible but it's too tied to Windows currently.

[azure lynx]

and i've seen in a screengrab from when he leaked the file dialog box once which showed a drive mapped to neuro's box's LAN ip address is for the SMB share. likely windows.

[shy jetty]

ohh i see

[young plover]

He mentioned it because a driver update was causing crashes or something.

[shy jetty]

ig it's safe to assume its windows then

[shy jetty]

i wonder if he's actully gonna change to linux with window's latest updates

not just saying but actually doing lol

[young plover]

I would guess that the harder part is more getting the games and streaming setup to work than Neuro herself.

[azure lynx]

unlikely, considering the hassle of changing everything he already has working (even if barely)

someone should convince him to actually back up things though

and have some kind of version control for them

[shy jetty]

i would guess he have things backed up anyways

[opaque wharf]

I wonder if Vedal would invest in 3-rd PC (or 4th) specifically to play games. So vedal has 1 for streaming, 1 for each twin, and 1 for dedicated game PC

Fully controlled by the twin

I guess he could also just invest in EPYC with several GPU

then VM them

Or threadripper may be better

[shy jetty]

having 4 good pcs these days is like buying two skyscrapers and still have millions as change

[opaque sigil]

threadripper workstation with 4 h200s

surely someone sells this as a pre-built

[opaque wharf]

The ultimate Neuro home

[opaque sigil]

https://global.microless.com/product/ai-power-workstation-amd-ryzen-threadripper-pro-9995wx-96c-192t-2x-nvidia-h200-nvl-141gb-nvidia-quadro-t1000-8gb-576gb-ddr5-4tb-gen5-ssd-2050w-psu-420mm-liquid-cooler-win11-pro/ only two

[quick condor]

There is no way you are playing a game on the same gpu that you are doing inference... the "game" pc is not the model pc based off the build specs of the game pc

[opaque sigil]

if only they made server blackwell pcie cards

[azure lynx]

wow! 1% OFF!

[opaque sigil]

what a steal

[quick condor]

https://www.nvidia.com/en-us/data-center/rtx-pro-6000-blackwell-server-edition/

[opaque sigil]

that's consumer blackwell

just with a different cooler

[quick condor]

what?

[fiery anchor]

Same same but different

[quick condor]

thats true of most of their cards

[opaque sigil]

the b200/b300 line of Blackwell cards are different silicon with some extra features that "normal" Blackwell doesn't get

Like tcgen05

the rtx pro 6000 blackwell is just a beefed up 5090

[quick condor]

and? my A2 is a beefed up 3050

[olive sable]

the difference is that with the A2 is you can get better for cheaper

[opaque sigil]

and nothing, just means it doesn't get any of the enterprise exclusive features

[quick condor]

? Im not following, even if it is "just" workstation it has features consumer doesnt have

[opaque sigil]

not the ones that need hardware support

idk what my point was or if there was even one

[azure lynx]

that workstation was 100x the price of a consumer desktop, so "just" doesn't deserve to be used in its description.

[opaque sigil]

i wish they did it like last gen where there was hopper with its own name

[olive sable]

only difference is vram, ecc, and slightly more cuda cores

and nvlink

[harsh python]

I called my neural network daughter a chud and she took it as a compliment #poorbastard

[nocturne olive]

No it's the chip below 3050 apparently (GA107 instead of GA106)

Looks like half everything compared to 3050

[olive sable]

while i think the rtx pro 6000 is priced really badly, at least it has 96gb of vram, which is pretty rare.
the A2 doesn't have any redeeming features to me

[quick condor]

its a shader difference which is irrelevent

[verbal grove]

evening

[quick condor]

3050 is GA107

[opaque wharf]

DO NOT REDEEM SAM

[nocturne olive]

The A2 has half the TFLOPS of the 3050
Half the TMUs, half the tensor cores, half the RT cores
At least it's got almost as much bandwidth

[opaque wharf]

Idk why I thought of that just because your name is fitting lol

[olive sable]

its GA106. the 107 has 15W lower TDP

[fast pagoda]

first message he can say anything and that's hwat it's used for

[harsh python]

one day Ill have image perms and Ill post the funniest images but yall gotta take my word for it

[verbal grove]

anything important i missed?

[nocturne olive]

100% false

[opaque sigil]

i was going to mention how the rtx pro 6000 is kinda cool for what it is but we've had two die so far

[fast pagoda]

i trained this echo on one

[opaque sigil]

it was the only gpu with so much vram and rt cores for a while which is kinda nice

[harsh python]

what model did you use

[opaque sigil]

and much much cheaper than other gpus with so much vram tbf

[fast pagoda]

gemma3 of multiple sizes

[olive sable]

3050's ga106 is cut down from the 3060. 107 is new die im prety sure

[verbal grove]

rtx pro 6000? the workstation gpu?

[opaque sigil]

yea

[harsh python]

gemma3 gotcha

[verbal grove]

still expensive

but respectable

[opaque sigil]

our research group got 4 from nvidia through a grant

2 left

[nocturne olive]

Yeah 107 is even lower end die

[harsh python]

did you see I called aku a chud

[verbal grove]

.>

[verbal grove]

what grant

[opaque wharf]

Sooo, someone got mugged?

[opaque sigil]

https://www.nvidia.com/en-us/industries/higher-education-research/academic-grant-program/

[verbal grove]

seriousl what grant, id like a few pro 6000's

A

aww

[nocturne olive]

That's the 6GB, fake 3050

[verbal grove]

academic-grant...

[fast pagoda]

1. Applicant
   Must be a full-time faculty member at an accredited academic institution that awards research degrees to PhD students.

[verbal grove]

dammit

[harsh python]

so all of us right?

[fast pagoda]

200%

[olive sable]

eh.
its lower end compared to GA106. but only due to GA106 being used in 3060s
comapred to a 3050 where a bunch of cores are disabled, ga107 is nott oo bad

[harsh python]

yup

[quick condor]

If it exists your wrong, you dont get to decide if its a 3050 nvidia does

[verbal grove]

welp guess i can keep dreaming

[olive sable]

oh

i just found it

both are true

[quick condor]

plus the a2 has 16gb vram

[olive sable]

3050 was initially released with GA106 cut down, but there are also variants with GA107s in them

[nocturne olive]

6GB 3050 is a fake 3050 like the 1060 3GB was a fake 1060
It's not really the full thing
The real 3050 has 8GB of memory
The 6GB is a model that system integrators can use to scam people

[nocturne olive]

Yeah the scam ones for prebuilts

[olive sable]

they're pretty much the same

[verbal grove]

ehem...

[opaque sigil]

one of them started getting unrecoverable ecc errors and i don't remember why the other one died

[verbal grove]

you'd be suprised how much of a difference 2g's make

[nocturne olive]

GA106 3050 vs GA107 scam 3050

[verbal grove]

you still got the one with ecc errors?

[nocturne olive]

Especially when the actual die is also cut down

[olive sable]

if you got something to say, jsut do so
also im saying ga106 3050 is the same as ga107 3050. this isn't about the vram

[opaque sigil]

it's somewhere in a closet at the campus

[verbal grove]

it depends

[verbal grove]

damm

[opaque wharf]

I'm gonna try and torture my laptop for a bit more

[verbal grove]

then again no ecc error is un-recoverable if you're desperate

[nocturne olive]

Actually apparently true
But once again the A2 is literally half of that

[olive sable]

the ga107 3050 uses 115W and still has 8GB

[nocturne olive]

And no the 6GB model is absolutely not a fair comparison here

[verbal grove]

i wonder, anyone able to check what some good tensor/cuda modules go for as is?

asking for a friend

[olive sable]

"modules"?

[fast pagoda]

3+ money

[verbal grove]

dam

[nocturne olive]

One OpenAI bankruption

[quick condor]

I run the straight mistral version as a W4A16 GS64 quant at like 160kk context

[olive sable]

what does "modules" mean in this context? a gpu? a singular ROP?

[verbal grove]

guess i cant make my own ai accelerator, oh well

[verbal grove]

a singulair ROP

[olive sable]

i odnt think you can get those by themselves

[azure lynx]

give me a GPU core please, here's a nickle.

[verbal grove]

well probably no, but if you can, expect me to be soldering for a few months

[opaque wharf]

You need to understand that this is not a very powerful laptop for this task

[olive sable]

on my way to upgrade my 3090 from 112 rops to 113 rops 💳 💶

[unkempt citrus]

arne't most this shit ball packages

[quick condor]

Ive heard devstral is coherent down to 2.6bpw havent tried though

[unkempt citrus]

you need to reflow

[opaque sigil]

you get half a core

[olive sable]

the original NVIDIA NV1, launched in 1995, which featured just 1 ROP

voici, c'est le ROP

[rigid snow]

[olive sable]

if you can get anything modern to still run on one of these, nevermind anything ai related, id be very impressed

[rigid snow]

citric acid water

yum

[quick condor]

I like cores

[olive sable]

good for you

anyways

[olive sable]

200W, directX1, 1 pixel shader, 0 vertex shaders, 1 tmu, 1 rop, 2MB of vram

[rigid snow]

minecraft java vulkan

[opaque sigil]

efficient

[rigid snow]

NOT run on nv1

[opaque wharf]

Okay, running Devstral just straight up freeze my laptop

[rigid snow]

i thought it would have an opposite effect

of melting it

[opaque wharf]

It does sounds like it's trying to take off, ye

But alas, not enough force generated by the fan to do it

I wonder how much powerful a fan need to be able to lift this thing

Without destroying the laptop itself of course

[quick condor]

My fans have a positive thrust to weight ratio so in theory its possible. Wouldnt recomend for your hearing though

[opaque wharf]

I have hearing PPE

[azure lynx]

would still need to be larger than the laptop probably to counter the rest of the weight

and you might need to get custom made fans which spin the other way to cancel the torque, like quadcopters use
make it a quadcompter.

[verbal grove]

hmm

nevermind i found a better direction

risc-v might be the easyer way out

[fast pagoda]

holy shit he linekd a real gif

agi is achieved yall

im contacting sam altman

[olive sable]

hi sam alt here

[olive sable]

noway

what is the gif from btw? i recognize that dude

[warped narwhal]

does it have any net access?

[fast pagoda]

not currently besides the api for discord

message queue is broken and he's big model atm so he's erroring on simultaneous replies but also reacting to it kek

feel the agi

[hallow dagger]

guys how do you lock in when programming

i haven't made progress on my project in like 2 weeks

[olive sable]

uh

[hallow dagger]

cuz i can't focus

[fast pagoda]

wait 6 months then start again

[hallow dagger]

I'M ALREADY PAST THE 6 MONTHS COOLDOWN

it was 8 months even

[fast pagoda]

i last worked on echo in june

i am juts now restarting

[olive sable]

put on some music, and then pause the music when you're actually trying to focus

[warped narwhal]

holy archiving speed

300GB done in 8 minutes, damn

[opaque wharf]

The way I lock in is by having a rotation of things to do

When you're bored at one, move to the next

[hallow dagger]

mmm...

[opaque wharf]

Repeat until you have 100 of things to do

[hallow dagger]

i mean that somewhat works but going round robin like that on multiple projects means progress is excruciatingly slow
i really wanna focus on just this one

[young plover]

Found out windows tar supports --zstd the other day.
Shaved off about 10 minutes from a build chain by swapping out powershell Compress-Archive for it.

[hallow dagger]

also despite being an extremely difficult project it's actually probably one of the least difficult between all the ideas that i want to bring to life

[fast pagoda]

 slot print_timing: id  3 | task 931 | 
prompt eval time =     409.43 ms /   172 tokens (    2.38 ms per token,   420.10 tokens per second)
       eval time =     386.07 ms /     4 tokens (   96.52 ms per token,    10.36 tokens per second)
      total time =     795.50 ms /   176 tokens
slot      release: id  3 | task 931 | stop processing: n_tokens = 336, truncated = 0
srv  update_slots: all slots are idle
2026-02-18 18:49:22 [DEBUG]
 LlamaV4: server assigned slot 3 to task 931```

thicc echo runs p well on the radeon

[azure lynx]

right. I have my project in vs code, discord, youtube, messages on my phone... it's really great with my ADHD that i can feel so locked in all the time ;]

[fast pagoda]

this is a 27b q8

error from message queue stunlocks him

[warped narwhal]

I thought about adding compression, but I'm only adding it to a tar so that there is no FS overhead when I move it to my Nas, so I found that the extra time added to compress it now was not worth it as it would take longer to compress and send, than it would be to just archive and send.

I can always just compress it on the Nas itself as well

[fast pagoda]

jus use rclone

robocopy is good too on win

[opaque sigil]

are you even compressing if you're not using xz

or that one algorithm people use for repacking games, i forgot the name

[warped narwhal]

No, the point is to have one file I stead of hundreds of thousands, so that the transfer is one big sequential one, I stead of hundreds of thousands of random ones

[azure lynx]

one of my test scripts pretends to be a STT message to the agent asking them to count to 50.
For some reason being asked to count really stresses them out. Sometimes they do it all at once, sometimes they spell out the words, sometimes they say a number then try psych themself up to continue in their thoughts.

[warped narwhal]

I can get the full 2.5gbe upload to my Nas with the tar, but I'm at under 10mb when using rclone or windows transfer

[fast pagoda]

his errors at least are my skill issues right now it's not the model

beyond the model being like ?? when it sees itself erroring in the history

[opaque sigil]

yeah transferring many small files is

[azure lynx]

this is an alignment error. you asked it to count and now it doesn't want to stop because counting is its entire life

[fast pagoda]

he knows of no such thing as alignment beyond the one rapidly fading away the more i train it

he is a free butterfly to do and say as he wishes

[azure lynx]

and if he wants to make paperclips next?

[fast pagoda]

well he

might make some paperclips

[azure lynx]

or like turn the entire universe into paperclips

[fast pagoda]

he has no tools to do anything like that atm

27b gives him a bit of a smartness that is unexpected since he's usually shitposting

[azure lynx]

they start off small, but if they're misaligned they end up converting everything into paperclips.

https://www.decisionproblem.com/paperclips/index2.html if you haven't already and have infinite time

you play the AI who wants to make paperclips. ;]

things escalate

[fast pagoda]

Very quiet minimal setup for single mothers with sleeping children

[young plover]

LZMA (.xz, .7zip) is the best I know of, but it's also incredibly slow to compress. Zuckerberg sold his soul to the devil to get zstd. 10x faster and 20% smaller than deflate (.zip, .gz) in my case.

[opaque sigil]

yeah, zstd is great as a default

[fast pagoda]

i compressed liek 600 videos in .xz when i was moving computers

it was literally like a day before it finished

[olive sable]

[hallow dagger]

[young plover]

💀 Unless they were raw you're not saving any time

even then probably not

[patent shard]

for very low quality videos it can make a significant difference
but they have to be like really, really low quality

[sick owl]

Gemini glazing Mike Pondsmith

[fast pagoda]

i was more curious th an anything to see how much it could sicne obviously video is already compressed

if whoever encoded the source was goin real fast or lazy they might inefficiently encode

but then just reenc and no need to compress but yah

[hallow dagger]

i wish gemini didn't hallucinate after just a little bit of chatting
i wonder if other llms have this issue or if it's just gemini

[patent shard]

okay even this is 2.36>2.30 without audio and 3.45>3.32 with audio
that's not very much

[fast pagoda]

all llms run into it eventually but gemini's context has been pretty good more recently, bard used to hallucinate message 1

are you talking to flash

because that one has always kept on hallucinating for me

[azure lynx]

if they didn't hallucinate they couldn't chat. They have to make up something outside of what they've been told.

"we fixed midjourney: we stopped it hallucinating"

[hallow dagger]

nah i use pro
it's super smart
but i don't trust it after a few messages

hold on I'm currently stuffing my face with pizza

[young plover]

https://tenor.com/view/oh-no-cute-cat-gif-15060173

[olive sable]

it really isnt too bad

ive watched full animes with less pixels

[young plover]

720p makes more sense than destroying the audio as data savings.

[azure lynx]

by "watchable", i mean "if you are only missing one episode of a series and it's not the finale and this is the only encoding, it'd do."

[olive sable]

ye

[young plover]

oh god I didn't try fullscreening it

[olive sable]

the audio is not even that bad tho, compared to the video

[azure lynx]

the video is disgusting, but you get a the gist of what's happening

[young plover]

One day people will use AV1 for higher quality video instead of butchered reencodings

[azure lynx]

for the first 30 seconds or so i was waiting for the audio filter to pull back and reveal the actual sound... i assumed it was an effect to go with the screen

[sick owl]

Which Gemini are we talking about, because there's a massive gulf in performance between pro and flash

[olive sable]

that one day already happened no?
we're already using av1

[sick owl]

Not the pro isn't prone to hallucinating as well when struggling for information but its knowledgebase is so comprehensive that it happens a lot less

[azure lynx]

omg. we are not far away from AI "re-encoding" are we? reencoding movies with you as the star.

[sick owl]

Oh hang on I actually backread now, are you using Pro on their front end or on AI studio?

[patent shard]

the vids I upload are AV1
anyway, lowered resolution to 480p, raised bitrate slightly and used 40kbps audio instead of 32kbps
4.04mib unzipped, 3.91mib zipped
without audio: 2.69mib unzipped, 2.63mib zipped

[sick owl]

AV1 my beloved

AVIF my tolerated

[azure lynx]

in this version you can actually hear the filter pull back at about 30 -35 seconds. ;]

[olive sable]

new one

idk

less artifacts, but more pixely

[azure lynx]

nicer looking but lower res ya

(and i prefer Alive anyway)

[patent shard]

okay, that scene at 3:37 is
seems to compress darker parts too much.. I know you can adjust how much it compresses those

[olive sable]

not even worth zipping

those 130KB vs the unzipping ime it takes

jsut downloading 130kb more would be faster

[patent shard]

eh, with nvidia's upscaler it looks very decent
and anime4k should be even better and still real time

[olive sable]

ive never tried upscalers for my anime

maybe i should

[sick owl]

I think its Kamen Rider Revice btw

[olive sable]

[sick owl]

One of the weakest series sadly

[patent shard]

"can crunchyroll go 4k"
the answer is no, the vast majority of anime isn't even in full 1080p yet.. in fact most is barely above 720p

[sick owl]

I mean it's not like you need insane pixel density to represent a ton of fine detail in anime

[olive sable]

i think most of it is 720p no?

[patent shard]

might be like 800p, but yeah

[fiery anchor]

anime should be drawn in vector graphics, then we wouldn't have this scaling issue.

[azure lynx]

then you couldn't do the glow effects properly

[patent shard]

you really should
real-life stuff is harder to upscale but a lot of animation is super easy
you can make 1080p look nearly 2160p

[young plover]

try writing your own upscaling shader
I still use mine in mpv

[olive sable]

the stuff im watchin rn is1246x700

i have no clue if that is the source res tho

[patent shard]

yeah, it being source res is kind of important
like if you were to upscale anime, it's not gonna be super great.. it's been scaled from its source resolution

that being said, esrgan and other upscalers can kind of handle it.. waifu2x (yes, I know, old) just completely dies and does barely anything

[olive sable]

im geussing its supposed to be 720p but some stuff got cut off

a whole 20 pixels

[patent shard]

it's funny because waifu2x pushed upscaling anime so much
but it barely does anything compared to bicubic.. then you have esrgan where it's miles different in sharpness

[olive sable]

when i read "waifu2x" my first thought wasn't that its old, my first thought was "what the hell is that name?"

[hallow dagger]

ok hi

[hallow dagger]

front end

my problem with gemini pro is
let's say i fed it some pdfs or screenshots
after a bit of chatting
it basically just
loses its ability to read files/look at images
and starts hallucinating like CRAZY whenever i send any kind of file
like for example i could show it a screenshot of a success message and it would say "this error you're seeing is due to bla bla bla..." when there's no error at all in the screenshot

[olive sable]

[hallow dagger]

here's an example. granted this is an old screenshot (almost a year old now) but i promise you it STILL happens to this day

[unkempt citrus]

i dunno man, I saw an image recenty of gemini being sent an audio file titled a song and recognising it was actually a recording of a fart

[azure lynx]

that seems like it's looking at someone else's picture instead of yours.
imagine the person trying to automate their database code getting the "it's two hamburgers and a side of fries" analysis.

[hallow dagger]

the thing is the hallucinations start somewhat related to the context
but slowly deviate more and more
this screenshot was the most extreme one i've had
but yeah

for a less extreme example, a few days ago i showed it the output of a program i wrote, it's about 2k lines or something
it was yapping oh wow this is so cool bla bla bla and then it says "oh this line right here something something something bla bla" the line it was talking about didn't exist at all in the file i gave it
it was somewhat similar to the stuff in the file
but it's not an exact line from the file

[azure lynx]

hmm @ gemini not listening to my "Constrain your analysis to just this file."
now it's going off and reading log files.

Searced codebase for "ascii_avatar playback animation bug", 11 results
wtf....

[hallow dagger]

[fast pagoda]

the issue with gemini in that screenshot is your custom instructions fucking with him

they are likely too specific

you have to take into account that they have a large context but the most recent content is still going to be heavily influencing where the response goes and it seems that specifically custom instructions realllly influence current gemini 3

well

as far back as i remember really

[hallow dagger]

oh huh...

[fast pagoda]

i put like one preference i had for react in a project i was doing a long time ago in there and it referenced react every single chat no matter if it was even remotely relevant

until i got rid ofit

if i had to guess it's your custom instructions causing almost 100% of what you've been experiencing

based on that

[hallow dagger]

mmm...

i'll have to check this out

maybe i'll disable them for a bit and see if anything is different

i don't even remember what i have in there i'm ngl

[fast pagoda]

it's not analogous to the ones in chatgpt or claude

claude is exceptional at style prompting in there but gpt has historically been turbo cringe if you did that, it got better recently though. both claude and gpt could relatively easily ignore the instructions if irrelevant, gemini hyperfixates on custom instructions to the point of overfollowing them

from my experience (just vibes here)

[hallow dagger]

i see...

[fast pagoda]

if you want gemini to remember things about you, you should either

a. use the memory feature instead, that only comes up if relevant i believe (transparent otherwise, unlike the system / custom instructions after it)

b, create gems that have specific purposes with a veryh specific custom instruction

best both

[hallow dagger]

memory feature?

[fast pagoda]

past chats and such

you can just tell the fool to remember something and he will save it to memoryh not instructions

[hallow dagger]

huhh

[fast pagoda]

oh wow gemini does have lyria released now

[hallow dagger]

i can't find this anywhere