programming

[olive sable]

Im geussing when its on auto and it shows T1, that might be GDM

Im on cl16 now with GDM

Ok cl14 is pushing it, that doesnt post

[true hemlock]

damn cl16 now?

[olive sable]

Ye

But

i had it in the wrong memory slot

16-18-18-18-36 T2 posts but just freezes on the logo

Ok this might be thebest its gonna get

[idle dune]

succesful distro hop. I am now using Kubuntu after my jury rigged LinuxMint got to bloated.

[olive sable]

actually

now that i know i had a bad ram stick

the discord and firefox crashing might stop

[kind nimbus]

my distro hopping journey is over

i don't think i will ever distro hop again

[idle dune]

I will try NixOS in a VM at some point.

But i have kinda settled with Kubuntu as what I need for my office machines

[kind nimbus]

Yeah NixOS is awesome. Keep in mind it's completely different compared to traditional distros like Ubuntu. There is a steep learning curve.You will need to learn a new functional language, troubleshoot, sometimes read upstream source code or even override upstream packages in order to fix bugs.

It's cool having full control over everything tho.

I am on unstable tho so that might cause some issues having the latest and greatest packages.

Flakes and home-manager and overlays are awesome tho.

[young plover]

just thought to check my RAM speed for the first time in a couple months
reverted to 4800 MHz again

[olive sable]

4800 is good tho no?

[young plover]

XMP speed is 6000 MHz

[olive sable]

ah

[amber fractal]

divide by 2 for ddr4 equivalent

[idle dune]

"It's cool having full control over everything tho." So its kinda like Arch, Gentoo, or LFS?

[kind nimbus]

in the end i can declare about anything i want my system to be in my nix code

i can also override and modify packages or services on the fly

also home-manager is nice

[olive sable]

nixos makes arch's "full control" look like baby shit

[kind nimbus]

and impermanence so my system is immutable

[kind nimbus]

true

also first class ZFS support

[idle dune]

Oh wow

[olive sable]

nixos's full control is a step above

[kind nimbus]

you just cant break it lol

my system is quite literally indestructable

[amber fractal]

oh it is breakable, just have to be stupid to do so

[kind nimbus]

how?

[olive sable]

ye

[kind nimbus]

my bootloader is read only, nix store is read only, can revert anytime via nix generations or zfs auto snapshots

[olive sable]

well

[kind nimbus]

how you wanna break that?

[olive sable]

the user can still manually fuck it up

[amber fractal]

setting aside the nixos ego, I was attempting to clone a system to a usb drive and overwrote the base system due to forgetting to override the hardware config.

skill issue yes, but also proves my point of not being entirely unbreakable

[olive sable]

nixos has some nice features for safety and such, that a user can make use of if they use the system in the intended way

you can still break it very easily if you dont follow the intended usage

if you rm -rf your drive nixos is just as fucked as any other os

[kind nimbus]

uh

nope

actually that would not screw me

[amber fractal]

Thanks to zfs that is

[kind nimbus]

/nix is read only, /boot is read only, the rest are symlinks

it would only delete my /home

but i have auto snapshots for everything

[olive sable]

arguably the most important part tho

[amber fractal]

seconded

[idle dune]

I might grab a project machine to throw NixOS on then

[kind nimbus]

give it a shot

[idle dune]

Seems like it woul dbe fun to thinker with

[amber fractal]

It really is fun

just have a lot to do before using it as a daily as some people have a lot of requirements

[kind nimbus]

oh i was about to share a discord link looks like it got blocked lol

[amber fractal]

yep, if we allowed you we'd also allow scammers

[kind nimbus]

haha yea
i was about to say feel free to join the vimjoyer discord if you need some help with Nix/NixOS
you can just google his yt to find the discord link

[amber fractal]

firefox including AI by default be like

flashbang when I hopped on the iMac for anime

What AI summary do you expect from the anime page

[idle dune]

Yeah I'll grab a ye olde thinkpad for like 20 bucks and throw NixOS on it as a little project machien along with all my other project laptops

[amber fractal]

nixos on an orange pi 5

(nixos hardware actually exists for that one)

[olive sable]

"project laptops"

[kind nimbus]

cool lmk if you wanna also join the vimjoyer discord
i can dm you an invite

[idle dune]

They are cheap and the parts are cheap

[olive sable]

you can jsut use a different drive on the same laptop

[amber fractal]

tbf my project laptop is "your purpose is to forward the touch/pen usb device over wifi"

[olive sable]

damn

16gb single channel is fuckin slow

[idle dune]

https://giphy.com/gifs/Fsn4WJcqwlbtS

I have a Windows XP gaming, a DOS gaming, and a Jellyfin project laptops in the works.

three diferent machines

[amber fractal]

the display was already damaged so

[idle dune]

Yeah but I have work things I need to do on all these drive

[kind nimbus]

nix build .#somethingAwesome

[amber fractal]

Reminds me to actually set up gamemode on my framework 13 if I'm using it for VR

(it is not a AMD boi)

[true hemlock]

lmao someone wanted to trade their 2x16gb ddr5 sodimm for my 2x16gb ddr4 3200

because they need ddr4 more rn

[olive sable]

what speed was the soddim?

[true hemlock]

5600

[amber fractal]

trade offer

[olive sable]

oh

so its not even that good of a deal

[true hemlock]

nah

i got my 2x16 ddr4 for like $50 each

have a lot of them

[amber fractal]

[quick condor]

perfectly normal training

[obsidian mantle]

how tf do i do this stupid raycasting to see if i shoot the target and not the wall

they reverse engineered the structure but didnt find location of function to build it

i see the structure but i cannot use it

they finished it for fallout4 but not for skyrim

there is option to look up different example in some other skyrim mod but when i tried to take it i got 500 errors because it uses different library

adjusting it for my setup will take months

i hope they just help me find it because i have no idea where to look

[amber fractal]

I continue to be glad that when it was time for me to use CUDA it was windows or NixOS, might have actually perished if I had to deal with that driver not working on top of thunderbolt never functioning correctly.

also known as 50/50 gamba on if the device was reconized

You know, now that I think about it. I should just not try to get the host system to work and accept I'm passing the cards into a VM due to the drivers being garbage.

Give me 1gb ram and single cpu on the host and just pass everything else in because thunderbolt and nvidia drivers are the worsr duo of all of history.

[alpine cloud]

oh sheet, i also had TridentZ's in a pack of 4, and one of them failed. fortunately that happened beginning of last year, so i replaced the 4x8 with 2x32 that cost me about 120 bucks. last time checked that same kit was 380+

[olive sable]

Gskill has lifetile RMA

You can just trade them in if its not your fault

[amber fractal]

we love 2x32 being... Wait a minute

[olive sable]

They should have a seperate stock for RMA tho

[amber fractal]

what is happening with the clockspeed there

[olive sable]

I saw 4x32GB for 350, but it was soddim

Oh, and they just lied, the price shown was for 2x32

[amber fractal]

I am glad that it isn't ddr5 and $1000 for 2x32

[olive sable]

Those are supposed to be temporary?

[glad orbit]

tp link

[glad orbit]

that like cybercafe place

[olive sable]

The cloudflare outages are just netcafes going out of business that cloudflare needs to re-route

[opaque wharf]

Wait, they have outage again?

[olive sable]

No lol

Not that im aware of at least

[opaque wharf]

I swear to god if they have outage again I'm migrating somewhere else lol

[silent cloak]

are netcafes still a thing

[opaque wharf]

Yes

People here still likes to play with their friend nearby

So they can all shit talk the other net cafe

[tender river]

https://www.wired.com/story/install-and-update-microsoft-store-apps-using-the-command-line-in-windows/

[silent cloak]

store but terminal

[opaque wharf]

I'm glad for those dozens of people that would use it

Also, how long do you think until it is exploited?

[tender river]

its just an official builtin cli

if an attacker can attack it, they might as well attack microsoft store itself

[midnight sigil]

loll

[midnight sigil]

how is it cheaper?

[proven merlin]

sudo dd if=/dev/zero of=/dev/sda (or whatever) snapshots are nice, but vms are better. (Yes ik im replying to a old af message)

[opaque wharf]

Well, I mean how many average users are aware when a powershell script is suddenly running. Having a ms store cli, depending on the scope, could be exploited to perform hidden purchase

[proven merlin]

Cat could probably do the same. Echo with enough patience

[tender river]

ye doubt it will support purchases

[opaque wharf]

Yeah, most likely even if they do support it, I hope that it requires user confirmation

[midnight sigil]

they refuses to give QoL features, like removing buldled bloats, but instead make msstore cli

[silent cloak]

its an attempt to get linux users back on windows

[midnight sigil]

give Steam trillions to stop developing Proton and they'll come back

fuck Vulkan, make the new Minecraft uses Microsoft's favorite DirectX

[rigid snow]

better than having winget packages being shims for actually downloading off store

[silent cloak]

i hate windows package management solutions so much

[rigid snow]

looking at you raycast

[silent cloak]

even worse if ur installing anything that must be compiled

[maiden geyser]

opera mobile award

[rigid snow]

that is awful

yeah let's take the worst app switcher era/design and make it into a tab switcher where you usually need access to more than the two apps you're currently using

uh huh the most important part is the page content which becomes illegible when scaled down anyway not the page title yeeeeah

C-

idr how current chrome mobile is but current safari (pictured right) is a great balance between focus on page titles (like safari used to be, pictured left) and complete focus on page content (like in opera above)

[maiden geyser]

may randomly start talking about tab switcher designs award

[rigid snow]

i take it

[opaque wharf]

Firefox also does the right. And chrome too IIRC

[rigid snow]

because it's plain better

[midnight sigil]

err winget is fine

[rigid snow]

no this is bad too, what the hell is the point if you don't render the thumbnails

wasted space

[opaque wharf]

Naah, that's FF being goofy. It usually show the thumbnail

This is chrome

[midnight sigil]

chrome wins

the goat

[opaque wharf]

This is after interacting with the tab

The previous picture was before the page interacted with (maybe even not fully loaded but because I pressed the tab switcher it doesn't bother to try and render it further)

[midnight sigil]

huh where's your search tab by title feature

[opaque wharf]

There's no such thing for as long as I can remember

Maybe there is but I didn't see it after my old phone tabs goes 99+

[midnight sigil]

搜尋分頁 - search tab

[opaque wharf]

Chrome hates me specifically for abusing their non standard Web API

[midnight sigil]

Chrome 145.0.7632.76
Android 14

[opaque wharf]

Uhh, yeah. Turns out I haven't updated all of my apps yet on this new phone

[midnight sigil]

[midnight sigil]

this looks like something I would see in the 90s, but with usb-c

[maiden geyser]

ts killed 5 people btw

[rigid snow]

that was the wireless version

[dull egret]

They used the F710
I have an F310

[dull egret]

In what way? [Does it look like from the 90's]

I'm curious why that vibe

(I sent those messages a couple mins ago, but Discord forwards are dogshit)

[dull egret]

Also... Good 90's vibe? Or bad 90's vibe?

I would ASSUME good, but anyway...

[opaque wharf]

https://tenor.com/view/started-blasting-gun-eyeglasses-so-anyway-i-started-blasting-gif-16955226

[kind nimbus]

Well yeah ok this just wipes the entire drive lol

I don't think you can really do anything against that

[dull egret]

Ok, gonna lay down before work, so I don't make myself sick AGAIN

[opaque wharf]

The right side is total size btw, not the currently downloaded

[opaque sigil]

i stil don't know what google play system updates are actually for

[opaque wharf]

Same. My banking apps broke if I didn't update tho

[maiden geyser]

maybe has something to do with strong integrity?

[opaque sigil]

so it's just patches for certain parts of the OS that come directly from google

[olive sable]

Its very common for me that a google app closes while im using it cuz "google is updating"

I appreciate the updating at 3am so people arent bothered by it, but im awake at 3am

[opaque wharf]

Most likely yeah. Also, I love the EU for making things good for the world. Now the phone warranty is tied to the phone itself not the purchase

[olive sable]

whats the difference?

[opaque sigil]

you get a warranty for second-hand phones

[opaque wharf]

You need to train your google algorithm better lol. Or set a schedule

[opaque wharf]

Yep. This second hand phone that I got for cheap still has official warranty. Thanks rich people that cycle phone quickly

[maiden geyser]

meaning you don't need a receipt to get one?

[opaque wharf]

Indeed you don't

[opaque sigil]

i think you usually only need the IMEI and/or serial number

[opaque wharf]

And for this phone specifically, the warranty is baked into the OS. At least they used DRM for good stuff

[olive sable]

My sleepchedule is very sporadic depending on how invested i am in bugfixing/procrastinating when tomorrow starts

[hallow dagger]

i fucked up my sleep schedule today

drank monster at 4am
slept from 9am to 3pm

don't do that guys

[opaque sigil]

[olive sable]

I went to brussels today

So im awake from yesterday 5pm till rn 7pm

[opaque wharf]

I just got back from watching the local neighborhood drama

[sage crag]

[maiden geyser]

should've tried battery acid spaghetti instead

[hallow dagger]

hi

[leaden crest]

hello

[hallow dagger]

C# will be the death of me

[opaque sigil]

just use F#

[olive sable]

we love filling in forms

"where did you buy this product?"
previous gskill RMA

[hallow dagger]

ah it's not about using...

i'm writing an il2cpp dumper

[azure lynx]

um... how? you mean it wraps all the C# calls or something?

[hallow dagger]

il2cpp dumper dumps C# type declarations from a unity game

i'm writing il2cpp dumper for hsr and i want it to be as accurate as possible

sorry if that sounds like gibberish

i would have to yap a lot to explain it

[maiden geyser]

is this a scientific interest or

[hallow dagger]

yeah just for fun

well there were more motives behind it but i'd have to yap a lot to explain those as well
and most of them aren't valid anymore so yeah i'm just continuing the project because i want to

[azure lynx]

there are tools like dnspy or whatever replaced it that decompile .NET binaries for you.

[hallow dagger]

defeats the purpose
also probably won't work very well anyway since HSR is a heavily encrypted and protected game

[azure lynx]

there used to be lots of modules for the various packers too for it.

and there were modules which did anti-obfuscation too

for almost everything i just dropped it into dnspy and it would give me readable (and hackable) code. used it to temporarily remove the restrictions and sign-in on a trial version of something on steam a few years ago. (it sucked even without the "free" version resrtictions anyway so uninstalled it)

[hallow dagger]

interesting
but still defeats the purpose for me
i don't want to use tools someone else wrote
i'll never be satisfied if i don't understand how everything works
down to the smallest bits
and my end goal is not the final output, i don't really have a use for a "perfect hsr il2cpp dump"
i'm more doing it just for the sake of doing it

i know it might sound silly but yeah

that is me

[azure lynx]

you can use dnspy to analyze dnspy and see how it works ;] or you could just read the code to see how they did the variuos decompilers

[hallow dagger]

ehh

maybe

[azure lynx]

i'm just a big fan of it coz it has saved me so much time over the years. ;]

[hallow dagger]

understandable

i'm a big fan of pain and suffering
did i mention i'm writing the dumper in pure C

[azure lynx]

i use C++ like C all the time because I think I know better than all the people who wrote python moduls which already do what i'm redoing.
i get it.

[hallow dagger]

[maiden geyser]

it's ok, tism in big 2026 is not that controversial

almost wrote 25

[hallow dagger]

wanna know how it works

it works in runtime
basically it's a dll
and it gets injected into the game by replacing a legitimate dll in the game files
game runs my code and i get full access to the game's entire memory space

[azure lynx]

i understand.
seeing one working already also can give you that though.
(dnspy does that too and lets you edit the code and patch it live)

[hallow dagger]

then it scans for the method pointers of C# reflection functions, and calls them on the game to dump everything from it

[maiden geyser]

it's called dll hijacking sir

[hallow dagger]

yes yes

fancy words

bamboozle game into using its own code against itself

[azure lynx]

never heard of it. (says the ex-cybersecurity professional)

[hallow dagger]

i might study cyber security after i finish uni

[azure lynx]

i used to be able to get system on work computers by dropping a python3.dll file into a directory with my code in.

they fixed it quickly after i told them. ;]

[azure lynx]

(the scary part was they were running something written in python as system)

[maiden geyser]

not knowing the thing's name but the thing is called being a practician sir

[azure lynx]

i am actually familiar with it. it would be bad if you did anything cybersecurity related and hadn't heard of it.

some apps really really really want to load certain libraries so much that they'll look all over the place for them. if you put a replacement n one of those places, they'll use the replacement instead of the one they should... and execute your code in their process.

[fickle rain]

Are you doing this to mod it or cheat in it

[azure lynx]

you can also trick the program directly into loading a DLL by injecting a little code into (but if you can do that you're no longer doing anything "security" related... you haven't gained any extra ability... except you might want to do it to games )

[hallow dagger]

neither
not interested in mods
and not interested in cheating either (even if i was it's near impossible with a game like HSR)

[azure lynx]

there's probably better binaries to analyze if you are just writing a tool to dump function prototypes from the IL. Like ones you've written yourself so you know the ground truth values for the results.

[hallow dagger]

you see i thought of that
which is why currently i'm writing my dumper logic in a C# script in a dummy unity game that just dumps itself
using the same exact unity version that the version of HSR i'm targeting used
and i can compare the output with publicly available repositories
for example, i can compare mscorlib dump with https://github.com/Unity-Technologies/mono/tree/unity-2019.4-mbe/mcs/class/corlib

once i ensure my logic is 100% accurate i can just port it over to the actual C dumper that targets the actual game

the C dumper already has all the logic for scanning the game's memory and finding any C# method i need

[hallow dagger]

look this is so cool
i have an x macro file
all i have to do is define any C# method i want to use in here
macro handles all the rest so i have the functions ready to use in dumper.c
and the actual scanning and finding of the functions is all abstracted away
i won't show any of that stuff i'm already sharing too much here

[real sierra]

@olive sable https://x.com/i/status/2025224025320882178

Programmer Humor (@PR0GRAMMERHUM0R)

beware https://t.co/dWzeRkOGEe

X · Feb 21, 2026, 3:00 PM

[hallow dagger]

LMFAO

[olive sable]

checks out

[amber fractal]

I remember sam sending this exact gif when I asked about him doing vulkan back when it was mostly opengl focused.

[proven merlin]

I think you could set RO to the /dev/sda file or something like that in the init script but like thats just another layer.. although does perhaps a kernel parameter exist for continuously making a file unwriteable?

[unborn moon]

@alpine pine look at this image what does this code do?

[alpine pine]

That's Python code. It defines a function called negate that takes one argument and returns the negative of it. Then it prints the negative of the number 5.

[unborn moon]

mallow thats... not right

[amber fractal]

python

[maiden geyser]

neurocord selfbotting situation is INSANE

[hallow dagger]

i'm going to pretend i didn't see anything

[amber fractal]

I don't think this can be defended

[maiden geyser]

old account... bro's making a sacrifice

[hallow dagger]

wait does this happen a lot here

[maiden geyser]

no but sloptuber syndrome is kind of funny

[unborn moon]

mallow has left the server lol

[amber fractal]

[unborn moon]

Did not know that account was made in 2023 tbh

[olive sable]

[amber fractal]

@olive sable python

[olive sable]

wha

context please

[maiden geyser]

read chat

[amber fractal]

.

[olive sable]

no

i dont get it

@grok dumb it downa lot more for me

[hallow dagger]

@hallow dagger

[olive sable]

python code is .py

not .h

[hallow dagger]

it's fine dw about it

[olive sable]

[hallow dagger]

https://tenor.com/view/neurosama-neuro-sama-gif-1408440608369590886

[real sierra]

cursed

reminds me of when I invented a pseudo html tag

and had to run a preprocessor to replace it with valid html

world's first compiled html file

it was the best solution for the use case though I think

[hallow dagger]

hii

[real sierra]

hello slicer

[olive sable]

hi shiro

[real sierra]

sam

I was thinking about screeps lately

[olive sable]

what

[real sierra]

I kinda wanna publish my screeps code here and see what you think about ot

[olive sable]

[real sierra]

it's long enough that the competitive edge is lost already

so maybe some code review is due

[hallow dagger]

why does that sound familiar

[olive sable]

what about fumoquest?

[hallow dagger]

ohhh i've heard about it before
it's like a game right
you write code to control creatures or something like that

[olive sable]

no it isnt

[olive sable]

fumoquest is still in development by shiro

[real sierra]

they mean screeps

[olive sable]

oh

[hallow dagger]

[real sierra]

well I've been away from my computer for a week unable to work on it

and though I have my PC back now, I have like 800 midterm exams this week

so my time may remain scarce

[hallow dagger]

good luck with those

[real sierra]

thanks they should be easy enough

[opaque sigil]

[real sierra]

the actuarial science one is the one I'm most afraid of

annuity this annuity that, man I'll be lucky to ever receive an annuity in my life

[leaden crest]

annuity: the right to receive an annuity

they really should use pointers in dictionaries

what a recursive definition

[hallow dagger]

😭

lmfao

[amber fractal]

would make it more readable

[real sierra]

the styropyro car battery video dropped

I feel like the relevant audience is in this channel somewhere

[hallow dagger]

STYROPYRO MENTION

[fiery anchor]

i just started watching that, stop monitoring my data packets

[rough bloom]

already watching it

[leaden crest]

I dont see a car battery video

just a buffet vid

yum yum battery acid

[kind nimbus]

good question

[unkempt citrus]

Oh its lead acid batteris

I thought it was Li ion EV batteries

[hallow dagger]

anydesk goated holy

goats

[amber fractal]

dead chat, everyone is watching a 2 hour video

[olive sable]

im making lego

fake lego, got it form family

[glad path]

bet

[amber fractal]

I got halfway before I still have to exist IRL

[silent island]

Sigh... i guess a console might be best for me..

[young plover]

I have a Clevo with an RTX 2060 that I haven't use in a while. I was always confused about wtf the iGPU was doing when playing games.

Shouldn't take that much power to hand off frames to the display out

[amber fractal]

6gb 3060

I can't do anything useful besides VR with it

[high brook]

i have a 12 gb 3060 i game on and a 5060 16 gb i run ai on

[silent island]

well, it is enough to play my steam games with 100+ fps on it... (well, my library is basically filled with PS4 and crossgen PS5 games :3)

[hallow dagger]

it was awesome

[silent island]

usually, the igpu would just sit on idle. But it seems my iGPU's glitched to cosplay a broke man's SLI

[hallow dagger]

my old laptop was also 6gb 3060

[high brook]

i got my 5060 for 360 about 8 or 9 months ago.

rip anyone buying now

[silent island]

for most peeps upgrading, they'd better off with a PS5, cuz its not affected by the shortages... (for now atleast...)

[unkempt citrus]

Let me just do fluid simulations on my PS5

[high brook]

biuld a PS3 compute cluster. if it's good enought for the US air Force and the Country of Iran it's good enough for you

[azure lynx]

PS3s were able to be used as super computers...

[high brook]

lol

[azure lynx]

they did some crypto breaking research using a cluster of them in a university that proved MD5 was bad. (if i remember my crypto lore correctly)

[opaque wharf]

Where is the armo?

[olive sable]

armo?

[azure lynx]

${bodypart}o

[opaque wharf]

I'll let that sink in for a while

[olive sable]

no thanks

im busy

[opaque wharf]

Heado
Armo
Torsoo
Lego

[olive sable]

ah

[silent island]

hmm... is there anyone here who owns a switch 2?

[olive sable]

we have one

[silent island]

cuz I have a question

[silent island]

does the switch 2 run switch 1 games at 60 fps or still 30?

[olive sable]

depends on the game, generally ye

[silent island]

damn, so no real reason to upgrade, yet...

[olive sable]

generally ye to 60fps

sorry if my answer wasnt clear lmao

mario kart for example really strugles on switch 1 with 4 players on a big map, you can see it getting framdrops

on switch 2 its a stable 60

still not worth the ugrade imo tho

[silent island]

rly? my switch 1 didnt struggle that much on 4 player mario kart...

[nocturne olive]

Rare Wispers in #programming

[olive sable]

on the dlc maps specificially it had a lot of frame drops

cuz im very good and my family sucks so a lot of the map was loaded at the same time

[silent island]

im asking cuz im looking forward on playing xenoblade with smooth 60 fps (hopefully..)
cuz if the switch 2 still ran ts at like 540p 30... then ill just stick to my emu frfr

[olive sable]

old xenoblade or the new one?

old one will prob run at 60, new one prob 30

i dont own the game to test it so i cant test

[sage crag]

spspsp

topipo

topipo

topipoprrr

https://cdn.discordapp.com/attachments/1421861060507602944/1425936142754123927/neuroCatModeOfOfnnfOOf.webp

[sage crag]

[olive sable]

hi konii

[sage crag]

🔺sam🔺

change your name to

🔻mass🔻

[olive sable]

im underweight tho

[sage crag]

not relevant

[olive sable]

your suggestion for my name has been attempted to be noted down, but sadly we ran out of pens

and i cant read or write according to you

[sage crag]

prrr

unfortunate

not only can you not read or write

you also cannot identify what a pen looks like

[olive sable]

what is the difference between awa and uwu?

[sage crag]

awawawa <- awa

uwuwuwu <-

[olive sable]

seperatly from context, it just looks like an uwu variation

[sage crag]

samuel

[olive sable]

hi

[sage crag]

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

[olive sable]

they hated him because he spoke the truth

[silk tulip]

😔

[quick condor]

Running Ai on only 16Gb? I wish I could do that... Currently chewing up 61Gb vram

[sage crag]

they hated him because he could not read or write or identify pens by eye, but somehow managed to randomly type letters into the keyboard that vaguely resembled english

[olive sable]

[sage crag]

[silk tulip]

you can run a 3b pretty smoothly on 16gb llama 3.2 isnt too bad

[olive sable]

i submit. or yield. whatever the one is for admitting defeat

icant rememebrr

[sage crag]

https://tenor.com/view/cosmic-princess-kaguya-yachiyo-cosmic-princess-kaguya-yachiyo-jumping-fast-gif-8114031472651899845

[rigid snow]

morniungh

,moning

pen s

[sage crag]

mlntcan

[rigid snow]

[high brook]

my use case is very small

[rigid snow]

i go slepp

bai

[silk tulip]

gn

[olive sable]

bai

[silk tulip]

how do you get embed perms on the server?

[olive sable]

by talking

[silk tulip]

ok

kinda new to watching neuro but its gotten me interested into trying to make something similar to her

im not going to do something like the full setup vedal does

but maybe just hook up a live2d model, tts api and then use either qwen or llama 3 for the llm

[opaque wharf]

You'll get the l word

[silk tulip]

?

[opaque wharf]

Latency

[leaden crest]

gasp

[silk tulip]

i have enough vram to run a q4 7b model

what i cant offload to my gpu i think my cpu can handle

[azure lynx]

yeah. the trivial " just hook up a live2d model".

that'll need a little bit of work to make it actually do anything useful.

[silk tulip]

true

i was just first gonna try lip syncing if the tts works and if possible maybe just get the model to move around a bit

[azure lynx]

lip syncing is non-trivial too, unless you just do it based on audio intensity. but then you still have to get it synched with the model.

[silk tulip]

yh

im thinking this will be about a 3 month project if i think abt it optimistically

i dont plan to do twitch and all the other stuff

just want to try the basics and see what i can do with what i already know

[quick condor]

I mean true, I'm currently using 12.9/16gb of my A4000 to run voxtral 3b at bf16 as a teacher model, thing is I'm usually running 1-2x 24b and 1x 14b at inference

Also don't go down the rabbit hole of attempting to build anything like Neuro or you might end up where I am: fusing a audio transcription encoder to a 14b model so I can drop whisper, and bump the 3ed model from 8 to 14b

[silk tulip]

hm

well im not gonna try anything past a q4 7-8b model

[quick condor]

I'm also about 4 months in and haven't even touched a 2d model yet

[silk tulip]

ok so this is going to be more work than i initially expected then

[quick condor]

Your really going to struggle with 8b, 14b might work in a limited capacity but I was hitting a wall with 8b as just a logic/semantic gate

[amber fractal]

try a year on for size, and a gamble at that scale

[mighty thorn]

So 8gb?

[amber fractal]

surely a neuro clone of any capacity runs on 8gb

[silk tulip]

my initial plan is to use a 7b model add speach to text and tts and then once i get that to work i was going to try to hook it up to a model and then be done

[silk tulip]

around that ig

im not going to try anything groundbreaking

[amber fractal]

minus the size constraint

[mighty thorn]

I think you underestimate the difficulty of that last step

[silk tulip]

i know that is going to take the longest

steps 1-3 arent too bad tho

[quick condor]

For gguf probably but 14b at W4A16 GS64 is 11Gb with vision and audio encoders/projectors at BF16. 8B at that quant is 6.2Gb with vision and audio stripped

[amber fractal]

@mighty thorn do you happen to know the vram cost of STT and TTS just out of curiosiry?

[mighty thorn]

About 0kb-100tb depending on what you want

Whisper can be like 1-3

But the speech output is the real variation

[quick condor]

TTS can run on cpu relatively easily at decent quality, STT is a big * on vram

[mighty thorn]

I’ve seen 500mb and 16gb

[silk tulip]

i was going to use piper since it runs on the cpu leaving my gpu for the llm

[quick condor]

Piper is solid, and can be run on cpu or gpu, whisper large is usually around 3Gb and doesn't like contention on the GPU while transcribing. You can use something else but don't expect great results if you aren't doing something completely insane like I am to get it smaller

[mighty thorn]

Distill whisper medium on faster whisper is more than good enough while using like 1gb

[silk tulip]

XTTSv2 was also one i came across online but skimming through the docs it seemed like it would be too heavy on the gpu

[quick condor]

I went further and pulled the full whisper encoder from voxtral 3b and combined it with ministral 14b reasoning

I didn't have good results with xttsv2.

[silk tulip]

going above and beyond

[silk tulip]

havent tested personally so thanks for letting me know

[quick condor]

Currently I'm able to hit about 100 t/s on 24b on my 3090TIs and about 60 t/s on the A4000. It's only slightly above and beyond, to actually get what I wanted I needed to rethink my vram allocation and there was no way I could fit more than an 8b+whisper on the A4000

[silk tulip]

i have a rtx 5070

so i think im going to stick with a 7b

[quick condor]

You're also going to be heavily limited since you can't tune models with that little hardware

[silk tulip]

thats what i knew i was going to struggle with sadly

[azure lynx]

this is what i do.
i'm using whisper for STT on the GPU, Gemma 3 model (varies which one) on GPU and HeadTTS for TTS on the CPU to get the viseme sync with speech for the avatar.

[leaden crest]

where does one learn more about this by the way

is there some textbook I can read

or a coursera

the basics I understand but I dont know where to get this consistent base of knowledge

[azure lynx]

i'm also doing some speculative decoding with whisper to get sentences as soon as possible. probably not a good idea.

[quick condor]

Whisper really doesn't like cuda contention

[azure lynx]

yeah. luckily i'm buffering and pushing it to whisper in a thread

[silk tulip]

i just learned through w3schools, gh repos and docs online along with yt if you can find any on the topics your looking for, its mostly just about finding and learning about what you can

hascrack seems to be way better than me at this so he might have a better answer

[leaden crest]

yeah my experience has been about the same so far 💔

[quick condor]

I might be a little too much, most of what I'm doing doesn't really have an instruction manual... I'm running free threadded vLLM, p2p drivers, vector DB memory recall, sequential model processing and custom models

[azure lynx]

i was trained in the secret arts of computer stuff by monks in a monastery in the mountains. they made me promise never to reveal their secrets.

[leaden crest]

what projects have you done that helped you most to learn all of this then

[quick condor]

Started with doing my own quantization, then moved to fine tuning, then model routing and integration then feature fusing

[leaden crest]

hmm I see thank you

very helpful

[azure lynx]

!!!!!!!!!!! IT WORKED !!!!!!!!!!!
changing from "localhost" to "127.0.0.1" stopped it from failing when trying to connect to the IPv6 address instead. (still guessing that was the cause, but explicitly saying "use the ipv4 address" worked so...)

and now when I talk to my agent, they respond through the avatar!
But i think they have a bug to do with the interruptions. it added the word "interrupted" at the end of it's utterance. it might accidentally be learning to say that from the system messages. or i might have a bug somewhere else.

it's my own little "Location" bug.

[silent island]

hmm...

god... i feel old...

[silk tulip]

this brings me back

[silent island]

i always forgot that the switch dropped when i was still in elementary school. and now, im boutta graduate high school lmaooo

[silk tulip]

crazy how time flies

[idle dune]

I really wish someone made a open source implementation fo Winamp that had skin combatability

[hallow dagger]

i had this issue too

i don't even know what we're talking about but yes

127.0.0.1

cuz localhost will try to use the ipv6 first and it will waste like 2 seconds and fail

hi

[azure lynx]

hi.
i had used "localhost" consistently in my code and configs. but some libraries like to listen on 127.0.01 and :1 and some seem to just pick one at random (always the same one for that service though). Like i've had the same library for different services act different (probably changed something else somewhere)

i only noticed it do the :1 thing when i tried it over curl

and the AI code had decided that "if (!res) { // connection timed out".

[timed out in 40ms? wow]

[azure lynx]

https://www.reddit.com/r/androidapps/comments/1mx889o/open_source_winamp_classic_clone_music_player_for/ this one supposedly is

oh. android.

but it looks like there are a few of them

[idle dune]

It is surprisingly hard to find a music player with a vizualizer now

I've tried Foobar, Strawberry, and some others

[azure lynx]

it's not easy to do a good visualizer. I mean, it can me depending on what you want to do.. i usually just did slightly rotated video feedback + something with either the raw wave form of some frequency binning and rotating what frequency controlled what other parameter when I wanted to make something react to music.

and it vaguely pulsed or whatevered to the beat

[idle dune]

Yeah I know that its hard.

I know my dumb ass can't make one

[azure lynx]

i have NDAs but I was constantly amazed at name-brand professional code.

and their coders.

[azure lynx]

i should draw mouths. and eyes and eyeballs and faces and hair. ;/

and i've decided i don't like this specific voice. oh woe is me! everything is ruined! faint

[azure lynx]

https://www.cs.cmu.edu/afs/cs.cmu.edu/project/fgdata/OldFiles/Recorder.app/utterances/Type1/harvsents.txt <= next time you don't know what to say, this list has you covered.
From "IEEE Recommended Practice for Speech Quality Measurements." IEEE Transactions on Audio and Electroacoustics, Vol. 17, Issue 3, 225-246, 1969. APPENDIX C 1965 Revised List of Phonetically Balanced Sentences (Harvard Sentences).

[azure lynx]

maybe use it for when you want to sound like a secret agent giving their half of the code phrase.

[azure lynx]

"It's a dense crowd in two distinct ways."

it could also be a good prompt source for creative writing.
it's intended to be a set of test sentences for getting/testing all the phonemes for making a TTS system i think?

after watching the avatar read about 200 or so of those sentences i've decided it's not good enough and I need to do the animation a different way.
time to be creative.

[wide parrot]

Working on a pajama sam neuro integration. I'll post the source and put it in projects tomorrow. its 2am here and I am tired. Though its a lot further than i thought id get, but its still quite a ways to go. It also took way longer than I thought it would to get started. I certainly didnt make it easy on myself.

Im so happy that I at least got something. I've been wanting to do this integration for nearly a year

[patent shard]

why does Discord keep crashing 💀

[azure lynx]

time to learn to draw.

don't have any pencils. ;]

tommorow i will have pencils.

[azure lynx]

(i can already draw somewhat so it's not like it's from scratch. i just don't have any pencils. and have never drawn anime/manga characters before.)

[kind nimbus]

ncmpcpp ?

[azure lynx]

i think that anyone who could write a visualzer for a music program instead went and wrote a beat matching game

[azure lynx]

hehe @ Tony.

[azure lynx]

hmm.. awkward. apparently the "Ask" agent can't read webpages. I don't want to wait 5 minutes for the Plan agent to decide to actually ask me for the webpage for the question I have. And I definitely do not want the Agent model to run rampant with this new information. ;/
guess I better just copy and paste the docs to the Ask agent. 😭

[jagged turtle]

yay clicked!

[jagged turtle]

huh, thought it could

can it at least use any tools?

[azure lynx]

it lied then.

it straight out said it couldn't.

usually it suggests hallicinated stackoverflow pages

[jagged turtle]

no as in

I normally see it using web search tools

[azure lynx]

it can read the files

[jagged turtle]

wait what agent are you using

[azure lynx]

its gemini 3 pro in copilot chat thing for vs code.
it can give me links for it to read, at least in planning. but it straight refused to go to the link i gave it and told it to read, saying it couldn't view webpages.

when it gives me a link i can choose to retrieve it and then i can review it before passing it back to the agent.

[jagged turtle]

weirdge

[azure lynx]

was probably either feeling lazy or confused about its superpowers.

i gave it the link because I knew it would otherwise just hallucinate some weird members in the json blob it needed.
and it ignored the link and hallucinated weird members in the json blob.

[jagged turtle]

https://github.com/orgs/community/discussions/166168

also apparently according to bing copilot (the worse one ik) copilot chat will only search in ask mode if you tell it to

also I think you need to enable access

but I assume you already have

[azure lynx]

maybe i need to enabled it for ask mode.
i've still got it prompting me both before going and getting it and making sure that it's something i want to share.
probably 1/2 of all pages have been unrelated stackover answers.

looks like it's enabled.

it probably was trained at some point to say "I'm sorry, I can't search the web yet." or something

[jagged turtle]

not surprised

especially the early ones before tool calling existed without MCP invading its bubble

[azure lynx]

I'm wanting to implement tool calls in my agent so they can access the web.
i think i just did exactly what i'd need to implement for the "..." handling. instead of ... being detected as an end of sentence for TTS, now it explicitly gets a pause and continues the same utterance. sounds smoother because it doesn't sound as much like the end of a sentence and the start of another one.

[rigid snow]

very often web search is a separate thing from tool calling

usually it's a request parameter you turn on and get those reference tags and stuff and get charged more, the actual search api is all handled by the provider

model provider

[azure lynx]

i was mostly curious about how it was implemented but it seems it's pretty simple: just need to have somewhat to transition from just outputting data from the LLM to processing it like a tag or similar, then once i have the whole tool message, do the tool thing and add the result to the context and start stepping again.

there doesn't seem to be a good way to handle async messages, but that's probably ok.

hmm... i could give them a CLI mode where they can do commands and give the CLI a help command.

i'm guessing "spin" is a tool for the twins.

i vaguely remember them talking about spinning other people and trying it and being disappointed that it only worked on themselves.

i wonder if they tried adding a "target" operand

[rigid snow]

it's better not to reinvent the wheel here, llms are already trained on the established tool calling workflow/template and know it well

is it good? fuck no

but it's there

you'll probably get inconsistent and generally bad results is what i'm trying to say

[azure lynx]

The Gemma model's docs seem to be saying my CLI idea isn't too far from how to do it.
they provide a template of a generic method for calling functions, and then provide a JSON blob with a description of the functions.
They say you just need 1. Function calling setup (where you explain how to call a function), and function definitions (which define the functions and specify what arguments they need).
They say the instructions should be brief and clear and to avoid providing too much general information. And that instructions may be ignored when they are confusing or overly detailed, especially by smaller models.

so a CLI like: use <|cli|> enter command line mode. type help to list available commands. and when they type help, it gives the same type of list of functions the example does. except maybe not JSON.
They already know about <|say|> and <|thought|> so adding <|cli|> shouldn't be too hard.
or i could make /commands which work everywhere...

[lilac apex]

I got pinged here?

Must've been a role mention

Oh

[proven merlin]

Heya, just got a silly question, i have 7 P106-6GB (P106-100) and a 1060-6GB on my hands, and as theyre essentially the same card, i was wondering to use em for LLMs.. i do have in my main rig a RX6700XT but i was thinking of putting the Mining cards to work..

My question is whats the cheapest way to do that?

48GB of vram would allow me to play with some nice Q4 models and even some MoE models with the A22B stuff, assuming i have enough pcie bandwidth at the end of connecting all gpus

Or would it end up being cheaper to just buy one of those Intel gpus with 48gb vram? Are they even out yet?

My first idea was getting a cheap am4 mainboard that allows bifurcation and hook each one up via pcie3.0 x1 to my a8-9600 that i have lying around.. id guess it should do fine once the model is loaded, but perhaps some of yall have experience with scrap parts like these

[maiden geyser]

there must be a conspiracy between power strip manufacturers to always make the light-switched ones go dim ultra fast

[rigid snow]

<|special-token|> should be separate and complete tokens in the tokenizer - like the whole thing is one token, the model doesn't have any correlation between them other than the "special" semantic meaning they inferred during training. if you want <|cli|> and it isn't mentioned anywhere in the docs it's probably out of spec hence why i'm saying the model wasn't trained on them and wouldn't do great

what i mean is for calling functions you don't need to prompt it like "to call a function follow this template:" because it already learnt that during training. you can try to finetune <|cli|> in but that's another rabbit hole i don't think you wanna go into. see screenshot - it's for functiongemma but it's the same for all models that have function calling capabilities across the board

but if you're using base gemma (not functiongemma) which i assume you do since you shared gemma docs, they don't train it on tool use and you don't have these tokens. from the page you shared:

Gemma does not output a tool specific token. Your framework must detect a tool call by checking if the structure of the output matches your prompted function output specification.
-# which is insane in 2026 tbh
you probably won't get good results with it... SOTA models fuck up tool parameters, why wouldn't a tiny ass model fuck up the formatting too? it's just more points of failure. their example in the docs works because there are like 300 tokens of context at a time, as soon as you actually start building you'll grow way past that and the model will just get hella confused if it's not constantly calling tools and start messing up very fast. models that know tool calling by heart (have been trained on it) won't do this

my advice is: get a model that's competent at tool calling, follow the tool calling standard for your model. it's generally function definitions at the beginning, then tool calls denoted by special tokens - you stop at function end token, inject the result into chat history surrounded by function response tokens, resume inference

holy essay

[quick condor]

If you really want tool calls, Devstral2 Small is about the minimum for consistent passible output. The way I handle web searches is as part of my gating flags.
A turn is first passed to a small gate model with a simple thinking level (low or high) prompt + are memories needed + is web search needed. Based on the response the thinking path is selected and then the 14b model is asked to provide memory/web search queries. The search results are passed to PSE and the results parsed and injected similarly to memories as context. The main model doesn't have search awareness

[rough bloom]

that probably would be cheapest way to do it if you already have a processor for it, yeah
ideally you'd want more lanes, like at least x4 or better x8 to each card, which I think you could get for relatively cheap with an old dual-socket Xeon platform
you would have to deal with really old power-hungry CPUs + NUMA, but I doubt that's a concern if you want to run these GPUs
pick your poison

[quick condor]

Pcie 3.0 x1 works for mining but is aweful for AI. Beyond that Pascal is missing FP16, FP8, FP4 and Tensor cores. With 8x cards you are going to be doing massive data transfers between cards and a substantial amount of your vram will be consumed by tensor parallelism overhead. If you get past that you might rock a solid 0.5-1.5 t/s

[rough bloom]

you wouldn't be doing tensor parallelism anyway

with bandwidth that low you have to do pipeline parallism

[quick condor]

All reduce would murder it

[azure lynx]

<|special-token|> should be separate and complete tokens in the tokenizer - like the whole thing is one token,
the <end_of_turn> "token" the Gemma model produces isn't a single token.
my ad hoc <|say|> and <|thought|> appear to be working adequately in testing. the agent uses their thoughts to plan what they are doing (nothing deep: sometimes it considers if needs to reply or not, and it's even reminded itself it was doing something else when it got distracted once.
i don't expect it to be chain of thought level logic, but it seems to be helping.
based on that working i think maybe telling it "there are commands available if you output /help, it will show you what you can say." and then when they actually say it then it'll enumerate the commands. ideally all function names are self-documenting.
and if it doesn't work, i'll try a bigger model, and if that doesn't work I'll find something else.

[rough bloom]

the <end_of_turn> "token" the Gemma model produces isn't a single token.
it really should be, it's listed as a special token, at least for Gemma 3 27B
which model specifically is this?

[rigid snow]

the <end_of_turn> "token" the Gemma model produces isn't a single token.

[azure lynx]

what i mean is for calling functions you don't need to prompt it like "to call a function follow this template:" because it already learnt that during training.
The page I linked from google literally tells you to:
You can use function calling with Gemma by constructing a prompt that provides instructions that specify the output format and define the available functions.
it gives 2 examples:
You have access to functions. If you decide to invoke any of the function(s),
you MUST put it in the format of
[func_name1(params_name1=params_value1, params_name2=params_value2...), func_name2(params)]
and
You have access to functions. If you decide to invoke any of the function(s),
you MUST put it in the format of
{"name": function name, "parameters": dictionary of argument name and its value}

and then it outputs a JSON blob with the descriptions.
how it was trained was generically.

[rigid snow]

based on that working i think maybe telling it "there are commands available if you output /help, it will show you what you can say." and then when they actually say it then it'll enumerate the commands. ideally all function names are self-documenting.
and if it doesn't work, i'll try a bigger model, and if that doesn't work I'll find something else.
it will maybe work depending on the model, up until a point
longer until everything explodes on larger models yeah but it's gonna happen inevitably
models being trained on function calling helps a ton in this regard, they seem to forget about what they have way way less

[azure lynx]

it recommends the 27B model for functions or the 12B if you need speed. so it's probably not the best at it.
it even says it can mess up.

[rigid snow]

yeah but

their example in the docs works because there are like 300 tokens of context at a time, as soon as you actually start building you'll grow way past that and the model will just get hella confused if it's not constantly calling tools and start messing up very fast.

[azure lynx]

the <|say|> and <|thought|> tags have been working with at least 10k of context.

[rigid snow]

the example is probably for one shot things with the instruct template anyway, in chat template it will fall apart i guarantee you

[azure lynx]

(about 6k of "system" prompt (Gemma has no system role, so it's just a user but effectively the same)

i'm mostly going to give it zero or one argument functions so it probably won't be too confusing.

[quick condor]

I never really liked Gemma, mistral handles various function calls well if you are using the tekken tokenizer vs tiktoken

[rigid snow]

gemini wasn't ever any good at tool use either

google just can't do it it seems

[azure lynx]

gemini 3.0 is mmuuucccchhhh better than 2.5

[rigid snow]

even whatevernumber the last gemini pro is sucks at it compared to claude or gpt

the one that dropped this week

[azure lynx]

the current gemini 3 also only burns through Copilot month credits at 1x instead of the 3x Claude burns at.

did a new one drop? i haven't looked.

[rigid snow]

3.1 pro

[azure lynx]

ah.. maybe that one sucks... i'll ask my friend at google.

[rigid snow]

it's a VERY good model

[quick condor]

Even then good results are only really possible with vllm and only on v16 for mistral. As for gemini/gemma, it's maybe in the gpt 5.0 low/opus 4.0 range, idk how it scores so high

[rigid snow]

like the best one currently across every lab

[quick condor]

Gemini is really bad across the board at handling context rot

[sage crag]

https://cdn.discordapp.com/attachments/1067638175478071307/1458837136777150464/doc_2026-01-08_07-48-27.gif

mlntcan

[azure lynx]

if the Copilot chat console manages to screw up the markdown parser in VS Code, it also screws up the agents ability to use tools correctly and inserted edits in random places. that happening has been the only times when Gemini has completely screwed up for me recently.
and i basically do either ask or plan first (depending on what it is i'm asking), scope the problem to one file, and give explicit deliverables.
And it seems to get caught in loops in Plan mode. (I wonder how much of my months use is due to the bad looping while planning i had last week. probably 1/3rd to 1/2)
Sometimes it seems to want to say it's done, but then it says "but what about..." and then it goes off on a tangent for a minute, then it says it's finished a few times... and then says "no! did we check the wiper fluid? Reading wiperfluid.cpp Lines 100-160 Yes we did check it." and it does that forever until you stop it (maybe?)

[quick condor]

Claude seems to catch itself going on a tangent in its Cot stupid quick. The thing that seems to waste the most time for me is pyrite locking up and stalling things

[olive sable]

[opaque wharf]

Every web dev should start having one of these as mandatory on-boarding. The things I've seen people do with SQl

[azure lynx]

hello. my name is 1' or '1'='1

[opaque wharf]

Nah, that's a high bar for 2026. At least for every code that I've encountered now people are using proper sanitization provided by their framework of choice. So the burden gets elevated a bit

[azure lynx]

i did things to SQL from inside Elixir once. (only know what the language is called because someone talked about it recently.)

made an iterator do unsafe stuff which allowed me to unquote some text and then write something to an arbitrary file? I think they also had permission errors on the box so one densely written CGI request later i had to write it up. ;[

[azure lynx]

so why can't i use ! in my password at my bank?
the only reason i can think of is that it's going to be transmitted in the clear to something which doesn't (or you strongly suspect it doesn't) do proper escaping or quoting and then you're going to use that in some unsafe way.
my password should be hashed before I send it to you, so there should be no chance of any weird characters doing anything bad to your system because it's operating on a number now, not potentially evil characters.

[opaque wharf]

Meanwhile my bank mandates at least one symbol on their password policy

[azure lynx]

also, my bank should let me just type a 40 character sentence and use that as the password. no uppercase requiired, no lowercase requiired, no numbers requiired, no (some) symbols.

NIST scrapped that recommendation because it makes passwords WEAKER

just make the password long and it doesn't matter if you're using dictionary words. just use a lot of them in an unusual way.

What is NIST’s guidance for passwords?permalink
The most important part of a good password is its length. Every additional character dramatically increases the number of guesses an attacker would need to try. For example, a one-character password made from lowercase letters would take at most 26 guesses. Adding a second character increases that number to 26 times 26, which is 676 guesses. An eight-character password would take about 200 billion guesses. That’s way too many for a human to guess, but remember that a modern laptop can comfortably make 100 billion guesses per second, so eight characters is not very secure at all.

NIST guidance recommends that a password should be at least 15 characters long. At 100 billion guesses per second, it would take a computer more than five hundred years to guess all the possible combinations of 15 lowercase letters.

[opaque wharf]

I'm still waiting until people use RFC 9807 as default when using password based auth

[azure lynx]

wow. RFC numbers sure are getting up there.

my favorite way of logging into a website is clicking "i forgot my password".
much easier than remembering my password.

and it ends up changing it each time, which makes it more secure!

that was scary... i clicked a link, and then i clicked another, and the next thing i knew i was on reddit!

[proven merlin]

I do have a 1U dual xeon server.. hmmmm

[proven merlin]

Which wouldn't be too bad, i dont mind if the generation takes a while to start aslong as im getting atleast around 7t/s. I do wonder why llms would struggle on x1 tho? The model would be fully in vram, the only thing transferring would be the tokens/kv or whatever, no? What am i missing/missunderstanding?

[mighty bane]

Today I updated my broadcastEventListener to support push notifications for phones.
It just works (not Just Works™ but actually just works!)
Today is GOOD DAY!

[cedar pulsar]

guys I just posted a project in the proects channel about doing sentiment anlysis on chatbots without neural networks. Just wanted to know: do you think the method I used and the results I got could be useful for people who need to know the emotions associated to an LLM output, or are the values not significant enough to be of any practical use? (there's absolutely no need to go through all of the readme, there's a graph at the bottom). thank you! PS: it's my first time writing something in here so I apologize for the wall of text lmao

[rough bloom]

you aren't missing anything, it's just that transferring the tensors between cards does have some latency and is bandwidth-sensitive, their size isn't always negligible
so it still has some effect on generation speed, just much less than if you used tensor parallelism

especially if low batch size

[azure lynx]

i know that for emotional responses for Neuro like when she goes all heart eyes, thats because she said certain words. (specifically "love" or "like"). Perhaps her other responses are similar. does she go all poing poing jumpy with her hands up when she says thank you? or is that more complicated?
an other easy-ish way to do that would be to precalculate a sentiment for every common word in the language and then just do a running sum. (assuming you have already got per word sentiment)
then you could make the list larger or smaller depending on how long you wanted it to take.

your method probably works fine. who'd know the sentiment better than the thinker of the thought?

[cedar pulsar]

yeah that's fair. thank you!

[opaque wharf]

The problem is the latent space for different LLM will point to different place. So it needs to be trained on that specific LLM no?

[cedar pulsar]

yes that's right, I put a class that calculates and saves those vectors. Now that I think about it I didn't test the functions that save and load those vectors tho

[opaque wharf]

Yeah, I just saw that after scrolling down some more

[azure lynx]

if they have the same final layers they should have the same concepts active in general though, no?

[opaque wharf]

No, that's the thing with latent space. It is usually unique to the model

For example a vector [1,0,0] could be pointing to "happy" in one model while the other it could be "sad"

[azure lynx]

but if they are wired so they tend to result in the "py" ending token getting generated for the "hap-" in the context, they have to link to the "py" token

[cedar pulsar]

that's what I know too, but I'm a noob at this so I might be wrong

[azure lynx]

there is precident for extracting info before the final layers... TTS models can give you the phonemes and timings from their activations in some models. very helpful for mouth motion matching when speaking.

also sentiment analysis models are tinnnnnny.

[silk tulip]

not having image perms kinda sucks icl

anyways

i got the llm personality configured to a good degree

now its just the stm and ltm left before i move on to the next set of things

[rigid snow]

coney

[wary mauve]

someone give me ideas on what to add to my neuro clone pls and thanks

running dry

[rigid snow]

blackjack

[wary mauve]

gambling addiction

got it

😏

[sage crag]

https://tenor.com/view/borzoi-snoopa-lord-foog-dog-funny-gif-20125052

banbnaba

[olive sable]

[cedar pulsar]

that's true, tbh my initial goal was to see if it could discern between what the LLM is saying and what it is feeling (for example with sarcasm, where it would mock the user by saying the opposite of what it's feeling), but that didn't work so I mentally rebranded my project to sentiment analysis

[wary mauve]

what were the results like? id like to do the same but for whisper, whisper -> emotion

[azure lynx]

i think another trick you could try is taking a smaller model from the same family and duplicating the current state in that and asking it outright for what it is feeling.

so you just clone the state into a smaller model

it might not be 100% accurate but it probably would have a good idea

basically asking "what does my model of myself think about the current state?"

instead of "how do you feel?"

[cedar pulsar]

I did a GitHub repo with the code I AI-slopped, the readme is very long but at the end there's a graph, you can just skip to that
https://github.com/Simomone326/Latent-Sentiment-Analysis

[wary mauve]

dw my project was AI slopped until i thought it was a worthwhile idea the decided to rewqrite it all 😏

thats 2026 coding

ill have a little gander

[azure lynx]

i'm happy to call my code for my current project not ai-slop, even though most of it was written by AI. Coz it's not slop. I haven't broken it by adding features yet.

at least, not permanently.

[wary mauve]

ive stopped using AI for programming i just get it to design systems for me and just questions and research i try not to program with it anymore cus i felt lazy

[cedar pulsar]

is it possible to learn this power?

[azure lynx]

you could try using Gemma 3n which has STT. you could give it the audio and ask it how the person feels?

[wary mauve]

Could emotions be derived from the raw audio data? it can deserne intonations and stuff or is that not possible

[cedar pulsar]

yeah that's trueeee, honestly if I knew what I was doing I would also try to code it by myself

[azure lynx]

Correctly using Ask/Plan before Agent in Copilot in VS Code. (and correctly scoping the problem in your requests, including what you want. 1 setence requests without good context tend to go wrong.)
And having already implemented a different version twice before that turned to slop.
So practice, persistence, and wasting lots of time.

[wary mauve]

thats fair but youll only know what ur doing by researching adn learnning so

imo if ur learning to program, ai is such a hazard trap cus its so easy to rely on it for everything, unless u actively use it to learn instead of letting it code for you

[azure lynx]

oh, and already being able to program in many languages before using AI.

[cedar pulsar]

true

[azure lynx]

i've written maybe 120 lines of the 3500 in this project, but they've been important lines.

mostly fixing bugs.

[fickle rain]

I think Codex/Claude CLI work way better as harness than most in IDE stuff

https://blog.can.ac/2026/02/12/the-harness-problem/
Here’s another cool one

[azure lynx]

the hardest part of using AI is reading the code it generates and reading what it's telling you.
as i said, you definitely need to get the agent to commit to a plan in "Ask" or "Plan" mode before you set it loose, and you should also read what it wants to do before just accepting it.

[cedar pulsar]

I heard that sparse auto encoder are used to create features that each token can or cannot activate (like a feature for "angry tone" or another for "stuff related to the golden gate bridge") but that requires a lot of data and computations, there are already pre-made ones for some open source models with tons and tons of features but idk if there are some specifically trained on the layers that map the audio to the latent space of a multimodal model (if this is how multimodal models work (?))

[wary mauve]

hmm well first i need to figure out how to make whisper stop thinking a cough is "Thank you" even with a VAD: P

but ty

ill look into it :)

[azure lynx]

you mean like comparing the embeddings to determine emotions?

you make some embeddings which are just 100% loaded with each emotion

and then you compare the current one and see which has the best match?

something like that

the compare is just a bunch of multiply and adds - doing a dot product - so it's very easy.

i'm not sure of the exact way they do the embedding but it ends up just being like a 128 value vector

[cedar pulsar]

more like seeing which features of an hypothetical SAE trained on the embeddings activate when emotional sentences are embedded, but if there's not already an open source SAE, like the ones in neuronpedia but specifically for those layers of that model, it would be very hard to train

I don't even know if this would work tbh, I'm assuming it would works the same as the one anthropic used in the golden gate bridge paper

[azure lynx]

have you considered just using embeddings? the dot product match thing works well enough for memory recall, so i imagine it'll work well enough for recognizing when an emotional embedding matches the current state's embedding

[cedar pulsar]

I'll look into it

[opaque wharf]

PHP indexing using "multiple" core

[olive sable]

[rigid snow]

synchronous multithreading

crazy

[olive sable]

"my app is multi-threaded"

[azure lynx]

https://youtu.be/eBaR5VdmscI <= my avatar app looks cool. still not cool enough to keep as the main avatar, but enough to test the speech system.
the video is just the avatar playing back the a batch of speech the agent produced earlier. This specific sequence ended with "interrupted" coz it was recorded yesterday when I had the bug.

[opaque sigil]

would you rather it take your system hostage for a minute like clangd

[fickle rain]

I’ve noticed Clangd starts being a lil messy with header only stuff
Might be me holding it wrong tho

[opaque wharf]

I was hoping there's a sweet middle ground you know

[opaque wharf]

To be fair, single header libraries are a menace

[azure lynx]

I forgot that when they speak they don't do the random eye/eye-brow movements. ;/ boring.

[fickle rain]

too many errors
still builds

Yeah sure brah

[azure lynx]

warnings aren't errors. they're just indicators that there may be errors.
(i hate warnings.)

[olive sable]

yes

[opaque wharf]

Well, not for me. I'd still like to watch youtube while waiting

[olive sable]

maybe keep one core free for system needs

[opaque sigil]

sometimes you do end up with a disconnect between clangd and the build system

(whoever broke it, please fix cuda 13 support in the clang libraries)

i do wonder if there's a workaround for the fact that headers just don't show up in the compile commands file, certainly haven't found it yet

[leaden crest]

reminds me of the machine love mv a bit

[opaque wharf]

I have experienced that pain. Especially when the project specifically uses gcc non standard quirk

[rigid snow]

[opaque sigil]

haven't run into that thankfully, though sometimes clangd freaks out about some libcu++ stuff and then I just tell it to ignore those diagnostics

[sage crag]

clanged

bink bonk

t

clangt

[stray dragon]

hi konii

[olive sable]

hi t

[stray dragon]

hi sam

[olive sable]

what does clangd even stand for?

[stray dragon]

c language d

[sage crag]

https://tenor.com/view/kitten-hit-kitten-hitting-kitten-hitting-camera-kitten-slapping-cat-hit-gif-5248570956277630634

[olive sable]

"C language daemon"

[stray dragon]

ow ow ow ow ow

[sage crag]

https://tenor.com/view/pat-cat-pat-pat-gif-25250101

[olive sable]

daemon is an anagram for a demon

[sage crag]

sam

[olive sable]

konii

[sage crag]

thats not an anagram

[olive sable]

why not?

[sage crag]

dog is an anagram for god

dogs is not an anagram for god

dogs is also not an anagram for dog

[olive sable]

i didnt ad a lettor tho

[sage crag]

lack quotation marks

[olive sable]

daemon <-> a demon

[sage crag]

sam punctuation

https://cdn.discordapp.com/attachments/1067638175478071307/1458837136777150464/doc_2026-01-08_07-48-27.gif

[olive sable]

punctuation is hard

[sage crag]

letter t

[stray dragon]

yes that is me

[sage crag]

ye

tha t

[stray dragon]

i am indeed

the t

[sage crag]

t hat

[warped narwhal]

have you considered the humble semicolon?

[stray dragon]

i do not have a hat

[olive sable]

no

[stray dragon]

no

[warped narwhal]

:(

[olive sable]

i have no clue where the hell a semicolon is supposed to go in a sentence

i only use it for programming

[stray dragon]

similar to em dash

but one time

[olive sable]

let me reiterate

i have no clue where the hell a semicolon is supposed to go in a sentence
i have no clue where the hell a "em dash" is supposed to go in a sentence

[stray dragon]

em dash is the long - used by ai

[olive sable]

sure

[sage crag]

ⱦ

NOT t

[warped narwhal]

^
t

t with a hat

[sage crag]

ȶ ail

𝘵
t italic

𝘵
t italic italic

⒯
t in brackets

(⒯)
t in brackets in brackets

𝓉
𝓽
fancy t

[olive sable]

knowing what it looks like doesn't mean i know what it does.
i also have no clue where the regular dash - is supposed to go in english, its different in dutch

[azure lynx]

just changing the pronunciation my avatar uses for the word "to". Now it uses a schwa instead of a stressed "U" and it makes it sound much more natural. (deleted the other pronunciations from its dictionary. I'm looking for other hyper-correct pronunciations too.)

[sage crag]

ŧ
mutant t

ƭ
mutant t

[stray dragon]

wtf it has more limbs

[stray dragon]

wtf it has a long neck

[sage crag]

ṱ
t with a hat but wearing it wrong

[stray dragon]

ⓣ

[azure lynx]

it's probably a littoral fish eater. the long neck. and wings

[sage crag]

ʈ
tall t

ₜ
small t

৳
alien t

𝕥
gothic t

[azure lynx]

the one true T is the pose.

[stray dragon]

there are too many t's

[sage crag]

but only one t

https://cdn.discordapp.com/attachments/1421861060507602944/1425936142754123927/neuroCatModeOfOfnnfOOf.webp

[stray dragon]

there is another????

[olive sable]

2nd t

[opaque wharf]

Clanker language daemon

[olive sable]

i dont like the use of that word

[sage crag]

useless

[sage crag]

sam change your name to

🔺

[olive sable]

🔺 is already in my name

[opaque wharf]

t riangle

[sage crag]

no, theres an extra Sam🔺

bad

fix it

[olive sable]

no

[sage crag]

[olive sable]

i have decided to keep my name in my name

[azure lynx]

also changed a to be unstressed schwa and an has a schwa now too and also no stress. after that, nothing it's said so far has jolted me. (i expect homophones to have problems.)

[opaque sigil]

https://github.com/ashish0kumar/windots hyprland for windows

[opaque sigil]

https://github.com/AOMediaCodec/oac they're doing a thing

[opaque sigil]

i wish they would just start from scratch and write it in rust or sth but oh well

[real sierra]

looking at the 20 minute old account whose first message was only emotes trying to figure out if they're human

[olive sable]

hi shiro

[olive sable]

how does that procces even work, do you jsut ask them?

[amber fractal]

that sounds tough

[opaque sigil]

ask for forgiveness not permission

[real sierra]

i wait for them to say more stuff, or for discord to flag the account for excessive dms or spam indicators

[olive sable]

[olive sable]

cant recommend that lmao

im not risking a ban for the stuff im planning

[real sierra]

second message all emotes

it's not looking good

[olive sable]

[trim valve]

[olive sable]

bred is back

hi bred

[trim valve]

up late and writing terrible prs

[olive sable]

[trim valve]

also accidentally caused an api to get limited by its upstream

[olive sable]

idk what that means

[azure lynx]

bred set the smoke detector off and the provider yanked full-speed access for a bit

[olive sable]

[trim valve]

the api in this has no rate limits

[azure lynx]

did you say "dear another api please do a lot for me. thx." and it said "k. brb." and the a lot was too much?

[trim valve]

I just sent like a few hundred requests to the thing labelled api

which sent a few hundred requests to the thing labelled another api

which subsequently ratelimited everyone else

[olive sable]

[azure lynx]

i've used systems where one request sparks off like 10 other requests on a different system. so if you make 100 requests, the other system gets hammered with 1000 requests all at once. it's the sort of design that works in dev when only 3 or 4 people are using it.

[trim valve]

well I think the initial request I send does like 50 external ones?

idk it flattens a paginated list

[unkempt citrus]

Turns out scaling is kinda hard

Who knew

[trim valve]

scaling is easy wdym

[azure lynx]

yeah. just get more computers.

[trim valve]

just stretch your image in mspaint

[olive sable]

you know how tuned cars have a laptop on the passenger seat with the software for the graphs and stuff.
why do they use a laptop instead of putting a pi or whatever in a less annoying location?

most new cars even come with a pc built in already

[azure lynx]

laptop can be taken out of the car inside.

[olive sable]

i see that as a bad thing tho

[azure lynx]

means you can get stable good speed wifi. i see that as a good thing

[olive sable]

you dont need wifi for tune software tho

[azure lynx]

and you can tweak your settings on a big screen with spreadsheets or whatever they do

[trim valve]

I have been waiting for this ratelimit to pass for nearly an hour wtf

[azure lynx]

probably because they just need to install on piece of software to make a laptop do anything, versus learning about Pis and their ways. They probably already use a laptop.

[opaque sigil]

banned

[olive sable]

my only guess would be people that tune their cars dont know about pis?

the money isnt the issue cuz i know that crowd spends 5K on fake spoilers

[azure lynx]

maybe the apps don't run on linux

[olive sable]

ye, HP Tuners VCM Suite is windows only

[azure lynx]

no idea why anyone uses a laptop when they could use a desktop. (as in buys a laptop and never takes it anywhere)

[trim valve]

the worst part is I:

• have to present this in like 2 weeks
• share it with a few hundred other students

if this happens mid demo I will explode

[olive sable]

you mean in general, or for car tuning?
cuz for the car tuning i agree

mini pc

[azure lynx]

i mean in general.
some people buy laptops and they sit on their desk, plugged into a monitor, mouse and keyboard.

and that is how they always use it.

there's a word for the position their laptop is: desktop. it's a desktop laptop.

[olive sable]

i mean.
even if you only use it at home, being bale to store it away and free the desk is kinda nice.
but the moment you start using a sperate monitor or a seperate keyboard, jsut get a desktop

[young plover]

In college I used a gaming laptop as a desktop most of the time. Then I could actually take it home on breaks

[azure lynx]

i'm ok with people who take it off the desk and use it while they watch tv sometimes. it's proved its worth

and taking it home on breaks is another good reason.

i suppose people think they're going to take it places and then they think of the weight and inconvenience it is to travel with. and they don't get around to ever unplugging it.

[olive sable]

TunerStudioMS is opensource and runs well on a pi 4 or 5 apparently

[amber fractal]

I did just gaming laptops for those reasons, taking it about for stuff but more importantly never having one desk at home. Had to stay portable or else I wasn't getting access to the device.

[azure lynx]

it just occured to me that Gemini 3.1 Pro might work better inside the gemini CLI than it does in copilot chat, coz of the type of things mentioned in the article. that'd be nice.