programming

[olive sable]

[nocturne olive]

Yes but that costs more money

I don't have that kinda spare money for non-GPU things

It's basically 100€ with shipping

[stark needle]

fair enough

[native surge]

Yes

[stark needle]

i mean just cause u seemed interested in vr

[nocturne olive]

Well I just saw you doing silliness in VRC and it seemed silly and interesting

[sage crag]

yes

[native surge]

this loads

Me not

Meow

Or femboy

[nocturne olive]

[native surge]

Ish

Nerd

[stark needle]

@☆♡《♥fox🦊ying•idol~★》♡☆ this u? 🪞

[sage crag]

https://media.discordapp.net/attachments/963967500171763843/1073733493479194624/attachment.gif

[stark needle]

@nocturne olive is the training time that bad or why u need so much gpu compute

[sage crag]

nrr

https://tenor.com/view/fox-foxes-happy-cute-snow-gif-15010450582016741990

[nocturne olive]

I like going big batch at reasonable speeds, right now it takes one day for a full model at 512 batch size

The more GPU power I have, the bigger and better I can go faster

[stark needle]

are u pretraining the model from scratch?

[nocturne olive]

I could cut training time around in half with a second 3090 by running in parallel like Wispers

[nocturne olive]

This arch doesn't benefit from finetuning

[stark needle]

hmm rip

[nocturne olive]

It's only 100M parameters

[midnight sigil]

I should start my cybersec journey

tired of playing games all day and having nothing to do

[nocturne olive]

Either way I'd certainly appreaciate being able to reach higher or equal quality in less time

[midnight sigil]

smh why me

[sage crag]

https://media.discordapp.net/attachments/963967500171763843/1073733493479194624/attachment.gif

[midnight sigil]

[olive sable]

[sage crag]

pluu

[olive sable]

[fast pagoda]

https://tenor.com/view/homura-akemi-anime-madoka-magica-dancing-anime-dance-gif-15513013343261194666

[fast pagoda]

hook the pressure washer on it

[frozen igloo]

This is a screenshot I think from the 400 car battery styropyro video

[fast pagoda]

ohh
that copper plate does give styro impression

i still havent watched that

[sage crag]

https://tenor.com/view/evil-vtuber-3d-head-turn-concerned-gif-15495186653811856266

dont think this puts us on speaking terms

find the birth certificate or im not buying you food

[fast pagoda]

https://tenor.com/view/family-guy-meg-yuri-yuri-chow-dog-food-gif-14764504896860940861

https://tenor.com/view/teto-mitosis-mitosis-teto-gif-11683636566672045121

they didnt give me one because of this

[sage crag]

[opaque wharf]

I do wonder, is mitosis basically clone?

[fast pagoda]

yes

genetically identical

unless cancer

[opaque wharf]

It does makes me wonder how do we manage the administrative process of lab specimen after mitosis

[fast pagoda]

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

[sage crag]

freezing

[fast pagoda]

look into cell lines and how thhey manage those shit is wild

[opaque sigil]

better hela

[stark needle]

h

[opaque wharf]

Ketchup curry?

[fast pagoda]

mushroom ketchurp

[opaque wharf]

Still cursed

[stark needle]

t

[fast pagoda]

t

p

s

:

[opaque wharf]

tpsbeat me to it

[rigid snow]

slash slash

[fast pagoda]

n

[rigid snow]

e

[fast pagoda]

u

[opaque sigil]

curry ketchup

[frozen igloo]

I read that as ketchup carry and was confused

[olive sable]

You wouldn't understand

Its genuinly good

[fast pagoda]

europe is speaking

be careful everyone

[olive sable]

[opaque sigil]

sam europe has spoken

[fast pagoda]

stay frosty

[kind nimbus]

Hela Gewürz Ketchup

[sage crag]

no child of mine will be bad at osu

5 digit or disowned

[rigid snow]

had to scroll almost to the bottom, you'll never guess what AX stands for

[sage crag]

agentic experience

[rigid snow]

smart

am dumb

[sage crag]

https://cdn.discordapp.com/emojis/1495756232047329291.webp?size=48&animated=true&name=NekoCatPetting&lossless=true

who are these default people

are you real

[rigid snow]

as far as i'm aware yeah

[sage crag]

ive seen this guy in like 90 places

on the street

[sage crag]

NOT real

[rigid snow]

i think he used to be a typescript nerd youtuber now everythiong is ai

[sage crag]

vibecoding has become such a negative buzzword that people are calling themselves "REAL agentic engineers" instead

ai hero

[rigid snow]

what did enums do

[sage crag]

w ai

[sage crag]

typescript enum NOT js enum

[rigid snow]

respectfully

what the fuck is a js enum

[sage crag]

string

[rigid snow]

dsfksdfksf

i think i came up with a really cool idea, you know how harnesses pin editable stuff, e.g. a todo a model edits with tool calls? what if we do that but with mood/tone, can use that to prompt tts tone and change the model expression instead of semantic analysis and should reflect the intent of the model better

i don't think for example sarcasm is easily picked up by semantic analysis systems

[azure lynx]

depends on how obviously sarcastic it is. a transformer based system might be able to represent something similar, assuming there is enough context and it's something obvious like "I just love how easily this website made it to lose all my data." because people usually don't associate loving with losing things the transformer might actually end up representing in a way which implies sarcasm. but sarcasm is hard without verbal and/or visual cues.

[rigid snow]

yeah that's what i'm trying to address, what if the model just tells its intent in a way that doesn't require tuning it to do so? obviously increases latency since the model has to do a tool call once in a while to change its tone but it shouldn't be by that much if you're running the model locally. i'm gonna have to test this of course since i have a suspicion they're gonna be using the system wrong/not at all but i probably can fix this with a good enough prompt

[azure lynx]

i have my agent do a tool call to set the emote they want to do, reminding them if they haven't set it in a while.

[rigid snow]

yeah or that

[azure lynx]

i think it probably helps with keeping them "feeling" a consistent way. usually they're just "vibing" though.

[stark needle]

@mouse is this true

[sage crag]

https://tenor.com/view/cute-kitten-alone-sleep-tucked-in-gif-14814077

[rancid ore]

did you know gifs can probably be used in computercraft?

they're very compact.

[tiny edge]

guess what program i was using

that made my system OOM

[olive sable]

You have 16gb, i can think of 150 different programs that could

[opaque sigil]

list them

[olive sable]

[rigid snow]

blender and ue are the first things that come to mind

[olive sable]

Yep

Unity too

[rigid snow]

haven't really used it so

surely can't be worse than ue

[opaque wharf]

Tab hoarder

[olive sable]

Coukd also be an ai thingy or a mkdern game

Considering this is neurocord

Altho the cpu is too bad to run an llm tbh

[kind fable]

Good code twitch good code

[opaque wharf]

It doesn't stop me

[olive sable]

[bitter phoenix]

I don’t like how most distros preinstall the Nvidia OSS driver which for some people including me is completely broken, or make the choice of using the official Nvidia complicated as hell. It’s tarnished people’s belief that Linux can work with Nvidia to the point they think they need a special distro for it, you don’t need to reinstall your os. When I think of the steps you have to go through on Debian to install it I’d just say it’s impossible to do correctly due to how much they hate closed source

[opaque wharf]

Cachy

Honestly people are overestimating the difficulty of bastardized Arch

[olive sable]

Nvidia linux is pretty easy no? It worked for me almost perfectly on all cards i tried on both OS i tried

[kind nimbus]

I had to do quite a bunch to get my Nvidia Hybrid Setup correcly working on NixOS but now it works

[opaque wharf]

Like look at steamos. Those are Arch based too

[bitter phoenix]

It’s the myriad of implementations that don’t work and are recommended to everyone

[rough bloom]

can cause some compatibility issues but the install itself is easy

pretty much every distro has either a package or an installer thing for it

[marsh coral]

isnt nvidia notoriously hard to go install on linux?

i meann i run fedora

[opaque wharf]

Even gentoo?

[marsh coral]

i have to do akmods or smth

and i have to like

manually sign the keys

since i have secure boot

[olive sable]

You just need to set the DRM thing to nvidia thingy in the boot params and rhen everything works completely fine

[rough bloom]

no

maybe decades ago

[bitter phoenix]

It could also be due to the desktop environment as well… I’m just remembering that hyprland doesn’t work with Nvidia. But besides that just using the official driver & gnome has been fine

With better performance than windows on everything

[marsh coral]

i meann im a newgen in linux and all this tech stuff

so idk abt the past

all ik is that i switched to fedora 3 months ago

[rough bloom]

idk how it is on Fedora but pretty much every distro is forced to have decent NVIDIA support due to how common the cards are

[olive sable]

Hyprland works perfectly fine with my sold 3090 and my 4090

[tiny edge]

ye but look on the program list, all of them seem normal

[marsh coral]

i think my system is boutta go OOM

all my swap memory is full

[olive sable]

You're sorting by cpu usage, we cant see the one uding most ram

[marsh coral]

i shouldnt have let my model run 24/7

[tiny edge]

im using 100% zram

[bitter phoenix]

Last I remember it was suggested to use amd cards. That could have changed

[bitter phoenix]

Zram my beloved

[olive sable]

Its still suggested, but nvidia works fine too now

[rough bloom]

I can see secure boot being an issue depending on how your system is configured
very much optional though

[tiny edge]

true i guess, but if you look on the list you can clearly see which one is the problem

it cant get more obvious

[rough bloom]

that damn kswapd

you should kill it

[tiny edge]

its not that bad

the other program tho

[olive sable]

Idk what heroic is.

[tiny edge]

the most unoptimized program ever created

[olive sable]

Is that the league launcher?

[tiny edge]

epic games+gog+ some other launchers

[olive sable]

[bitter phoenix]

Epic games launcher, they didn’t want to put fortnight on steam so they made a launcher

[tiny edge]

its medal

if you look there's a bunch of 600mb instances

it has a terrible optimization which causes

terrible memory leak

[marsh coral]

omg my system just went OOM

how is 90% of my cpu being used

[bitter phoenix]

Do you use live filesystem compression or something, open htop

[marsh coral]

idrk

might be bcz im running my agent model, brave browser, running a fedora kernel update, then an akmods for rebuilding nvidia and vbox, and also running feishu

that mightve overloaded the system

[opaque wharf]

So it is AI

[marsh coral]

uh oh its happening again

[opaque wharf]

I was joking when I send this

[marsh coral]

[opaque wharf]

But its probably is the AI

[marsh coral]

usuallyy it doesnt jump this high

my ai inside a podman container on a separate account

[bitter phoenix]

I just think of what could be causing the OOM and do pkill thing

[opaque wharf]

By building, did you mean compile?

[marsh coral]

tho it is running as a quadlet systemd service

[minor crag]

I was thinking that training a 12B transformer on CPU would use all my 16GB of ram and set my CPU on fire

But I've done the math and using my backprop and forward pass optimisations it should use less than 3.5GB of memory (3GB for actual weights then maybe 500Mb for the CNN) and maybe 60 or 70 percent of my laptop CPU

So this seems quite promising

[marsh coral]

it mightt be bcz im using akmods on nvidia and vbox

[opaque wharf]

Because compiling will make CPU go brrr too

[marsh coral]

since the kernel update js finished

tho like sometimes the system jumps to 99% CPU being used all the sudden

and the memory/RAM use spiking

then returns down to only 20% of CPU used

[minor crag]

Plus however much ram I need for the KV cache, context window, and all that junk

[marsh coral]

ok i think ik why cpu is jumping

cuz of akmods

[umbral wigeon]

What will you do if this guy(me) PR you 80000 LOC

[marsh coral]

rebuilding the modules

[umbral wigeon]

70 files changed

-10000 LOC

[bitter phoenix]

What I did for free was put my model on google drive; open google colab and clone my repo and train using their free Nvidia tesla gpus and when the account was out of hours I would push the result to Google drive and sign into colab with an alt and then pull it back for more training again and again and pull it back to my computer when done, also make sure you’re using the best optimizer like Muon instead of AdamW and there’s no software bottlenecks

Also I’ve done my own testing: GQA > MHA, decoder-only > encoder-decoder > encoder only

Was pretty much already known but I wanted to see if there were any cases where it wasn’t true, but no

Just use GQA decoder only bf16 Muon

[minor crag]

That seems like a lot of effort when I just said I can train a 12B model on my laptop

And I'm gonna have to test different optimisers to find what works best with my architecture

[bitter phoenix]

seems like a lot of effort
It was faster

[bitter phoenix]

I’ve been working out the best transformer setup for a while

[minor crag]

But it requires google of all things

[bitter phoenix]

Once you have a good architecture you’re confident in you can just pretrain it and re-use that same model for multiple different things

I have an archive of 90m 500m and 1b models I made myself

The top property is data quality, recently it’s a lot of post-training research to get it just right

[bitter phoenix]

Their.. alright

[stray dragon]

hi phrrrrr fox rectangle flag

[stray dragon]

hi afunyun

[stark needle]

what transformer

[minor crag]

?

[stark needle]

architecture

[minor crag]

Nope I have ethics

[minor crag]

Custom :3

[stark needle]

i doubt it would run even with heavy gradient checkpointing

[bitter phoenix]

12b is huge

[stark needle]

bf16 12b is like 24gb

add the two momentums of adamw

and ur dead

and we didnt even consider activations

[bitter phoenix]

Might want to think about the evals to see if that’s really necessary

[stark needle]

and flash attention doesnt exist on cpu in pytorch

[minor crag]

I'm using a mix of optimisations with one being 1.58 bit weights so it takes up just about 3GB for the entire model weights

And I'm making the model bigger because I can just afford the compute and it'll offset the loss of using 1.58bit weights and some of the other optimisations

[stark needle]

pretraining in 1.58 bits seems like a very bad idea

[bitter phoenix]

A 1 bit model? That’s super cool

[minor crag]

1.58 so Instead of 1 or 0 it is -1, 0, or +1 but the matrix multiply is still a lot cheaper since it only uses add, noop, or sub instead of multiply

[stark needle]

fp4 training in the nvidia paper was already very unstable hence they had to do random hadamard rotations and keep the amax history and invent nvfp4, otherwise it would just NAN

[minor crag]

Havnt gotten far enough to see but I'm using a custom method instead of backprop so it will work (in theory)

[bitter phoenix]

And you’re writing the transformer and training logic yourself? I would love to see it to learn from as reference, if that’s alright. I heard Microsoft was already working on something similar as a prototype

[bitter phoenix]

I don’t appreciate enough how my training is always stable, except when I was pushing a 100k model too hard and broke a bunch of stuff and it NAN’ed

[stark needle]

more like cause we use a lot of stuff that is proven to work

switch the friggin activation function to something random and ur dead

[minor crag]

I'm writing from scratch in luau using no ml libs or gpu accelerations but sadly a lot of my methods are custom and I don't want my methods to be public since I spent a lot of time working on them

[bitter phoenix]

Same here

[stark needle]

does luau support simd intrinsics

[bitter phoenix]

My goal is I don’t think what I’m looking for is possible with existing tools like ollama, everything I’ve seen who’ve leaned on ollama to ‘make their own model’ get no where

[minor crag]

I don't think so

[bitter phoenix]

So I write everything so that there can’t be any blockers

[stark needle]

training is gonna take a very long while in that case

especially since the ops are quadratic with standard attn

and no simd

for a 12b model

[minor crag]

That's why im using custom methods instead of backprop that is a lot cheaper

[bitter phoenix]

It would probably use a 64k tokenizer, which is even slower than my standard 8K

[stark needle]

right but it would still have to traverse each single param

[fast pagoda]

hi

[bitter phoenix]

For a hero run

[minor crag]

Yea?

[bitter phoenix]

You could probably record the loss, batches & hours and make a chart

To know how long it will take

[true hemlock]

bruh gonna train a 12b model on basic fp int units with no simd 💀

[fast pagoda]

see you in 203

0

[true hemlock]

ipc going to hell

no

see you in 3020

[fast pagoda]

[stark needle]

also if u wanna reach reasonable loss faster u should prob use a smaller transformer due to how scaling laws function

[amber fractal]

[bitter phoenix]

I have this image saved already lol

[stark needle]

chinchilla my beloved

[fast pagoda]

i got 2 flip flops available

[rough bloom]

wouldn't it converge a lot slower without backprop?

[true hemlock]

bruh lmao even at the absolute best 4 IPC apparently it would still take you 40K years to train with reasonable sized dataset

assuming 16 cores

[bitter phoenix]

Maybe trim out some irrelevant training sources if possible?

[true hemlock]

(assuming foward prop alone)

[minor crag]

I'm not training the entire 12B at once

I'm training 500M first then once the loss stabilises I'm gonna freeze the first 500M, add 500M more, train it for 1 3rd the total time, unfreeze the original 500M 1 3rd the way through that run and train both together for 2 3rds of the total time

Then repeat multiple times over a few months until the model hits 12B

This combined with the many optimisations and replacements compared to a standard transformer will train the entire model in a few months on CPU

[true hemlock]

what

uh

okay, what size of dataset are you aiming for

[rough bloom]

many optimisations and replacements compared to a standard transformer will train the entire model in a few months on CPU

[minor crag]

It will be a tad slower but my replacement should converge

[rough bloom]

tad slower is an understatement usually

there's a reason why everyone uses backprop

(for normal ANNs, that is)

[true hemlock]

because training each transformers layer would not only fuck up generalization a bit, it would still take you years, possibly hundreds, AND assuming foward prop alone.

[bitter phoenix]

So just expand the model as you go instead of starting at full size? I’ve experimented with that, don’t remember the exact sizes but after quadrupling the size it began training at 4 instead of 10. Not sure if it helps to start small or if is just convenient to start with the original size

[minor crag]

Not to sure what size it will be since I haven't finalised the dataset yet

[kind nimbus]

no i use KDE that should not be an issue

[stark needle]

@bitter phoenix idk if it interests u but here google leaks gemma scaling laws https://arxiv.org/abs/2501.18914

page 6

[true hemlock]

i don't know why avoid backprop even, with batch feed and grad acc its like pretty much almost negligible compared to the forward pass

[rough bloom]

that + good luck optimizing without gradients

[bitter phoenix]

A few months is still a long time, I would prototype for a while before committing because it might turn into your own legacy model with one realization or improvement. I also bake my tokenizers into the model files

[bitter phoenix]

Ty

[stark needle]

use SNN bro

[minor crag]

Yep I'm just expanding the size by 500m each time and idk if it'll help but in theory it should work

[true hemlock]

<insert SNN mention copypasta>

[rough bloom]

your training steps will be faster but you will need 10x as many to converge, if it converges at all, so you end up spending way more compute

[opaque sigil]

SNN

[true hemlock]

what gets me more

is the fact that they wanted to train on cpu

[fast pagoda]

bitnet works because it has gradients on the backwards pass, there's a latent full precision copy of every single weight which is actualy what is being optimized

[true hemlock]

with no simd

[ebon basin]

uuuh net 11 has MCP directly integrated on one side I like it but on the ohter side MCP is bloated

[rough bloom]

Hyprland works fine with NVIDIA
used to be an issue because NVIDIA insisted on EGLStream but that also just hasn't been the case for years now

[minor crag]

Can't afford a good gpu in this economy so I'm optimising until what I have works

[stark needle]

use kaggle tpu

[true hemlock]

are you sure that not using the simd unit on the cpu is "optimizing"

[rough bloom]

gonna be optimizing for a few decades

[amber fractal]

[fast pagoda]

gradient-estimate variance grows with parameter count so 12b is going to have like 0 signal

[true hemlock]

bare non-vector arithmetic units alone are slow as hell there's literally a reason AVX exist

[ebon basin]

another uii JS directly in MAUI

[opaque sigil]

has anyone checked whether the add/sub only "matmul" in bitnet is even cheaper than using tensor cores

[stark needle]

yes on gpu when packing it in tensor cores

with triton kernel

[ebon basin]

and AOT for android

[minor crag]

I've not even gotten to the point of optimising the code and I assume the compilers either uses simd or I can mod the compiler to allow for simd intrinsics

[opaque sigil]

is that a yes for it being cheaper

[amber fractal]

[true hemlock]

aint no way bro

[fast pagoda]

you would need a population size that exceeds yo momma's weight

that's impossible

[stark needle]

https://huggingface.co/blog/1_58_llm_extreme_quantization

[bitter phoenix]

I’ve made a 256k model with a loss of 4.9 just considering every way I can push it towards the right answers. There’s tons of things you can do with smaller models to reach decent results instead of going larger

[amber fractal]

This is an insult to my efforts, and my stuff doesn't even properly learn yet. But at least it is fast despite the operating logic being in python

[nocturne olive]

What be going on here?
Seems kinda entertaining

[true hemlock]

also i still don't get the reason with avoiding backprop

[olive sable]

[fast pagoda]

[minor crag]

I could use a 3B model or whatever but with my current code I can train a 12B model before my deadline is due

[nocturne olive]

Who told you that?

[amber fractal]

I'm ignoring backprop because I have to, I can't say anyone else should at all do so

[nocturne olive]

You'd be better off training on free tier Google Colab, at least it has a GPU

[stark needle]

just use ngram model

atp

[bitter phoenix]

I would say there isn’t much reason to go farther than 3b if you aren’t coding or mastering multiple domains. Depends on what the inference is

[amber fractal]

not GPU optimized (real)

[stark needle]

google and meta usually run ablations on 50-100m parameter models

on ~100B tokens

[minor crag]

I get there isn't a point but I can do it so I don't see why not

[bitter phoenix]

I prototype on 90m

[rough bloom]

if your architecture is somehow incompatible with backprop or has a better way to train it then ye that's reasonable
that's just not the case for the average ANN

[fast pagoda]

i dont think even hinton has gotten forward forward prop to work in anything transformer shaped at all let alone a 12b

[sage crag]

https://media.discordapp.net/attachments/963967500171763843/1073733493479194624/attachment.gif

[bitter phoenix]

If you’re confident

[rough bloom]

the reason is that you can't do it

[fast pagoda]

[opaque sigil]

sounds like a noob

[nocturne olive]

Highly entertaining

[amber fractal]

Yeah, I assume not a lot of people have the smarts to uphold the drain damaging takes

[nocturne olive]

Keep it up

[minor crag]

If there is an issue causing it to take more time then ill still have a working model but it just won't hit that 12B goal in time

[fast pagoda]

tribiall brain dablage

[sage crag]

written in luau

[fast pagoda]

https://tenor.com/view/disney-stitch-lilo-surf-fun-gif-12286198

[amber fractal]

To this day I'm surprised the NN series has actually produced something that learns at all

[rough bloom]

is apparently

[sage crag]

chat can i train 1t model on my athlon xp

[rough bloom]

surprisingly easy

[amber fractal]

True, the real battle is making it useful

[nocturne olive]

[stark needle]

train a 1 trillion parameter bigram model

1 trillion bigrams

✅

[nocturne olive]

It's only gonna take 1 trillion years

[opaque sigil]

auto differentiation one of the best things people have come up with

[sage crag]

512mb ddr2 ram

[fast pagoda]

i think

[nocturne olive]

Just keep all the weights on disk

[fast pagoda]

llms currently

being so bloated

has distorted what a "large model" is

so badly that people think a 12b is a small network

[sage crag]

1b params

big number

[amber fractal]

✅

[stark needle]

fr who remember 500m param vit/cnn being huge

[bitter phoenix]

Then you bypass the finite dram limit

[stark needle]

everyone was like holy shit

[rough bloom]

ye

[fast pagoda]

GO BEYOND

[nocturne olive]

Is a 100M parameter NeuroSynth big?

[rough bloom]

GPT-2 LLM

[sage crag]

1b params in your mnist number recognition toy

[rough bloom]

117M params

[sage crag]

[fast pagoda]

https://tenor.com/view/go-beyond-plus-ultra-all-might-mha-bnha-memes-gif-19247796

nan

[sage crag]

3 parameter

llm

vs

[nocturne olive]

Surely it's not that big

[fast pagoda]

yes/no/maybe

[sage crag]

my child

[fast pagoda]

2

[sage crag]

youre right

2 param llm vs my child

[bitter phoenix]

Is the baby coughing

[sage crag]

@fast pagoda are you coughing

[fast pagoda]

[sage crag]

the answer is maybe

[rough bloom]

I bet on the LLM

it is immortal

[fast pagoda]

https://tenor.com/view/magic-8-ball-not-good-outlook-robot-dreams-gif-13153136483637167804

[rough bloom]

so it will win

by default

[bitter phoenix]

I bet on the baby

[sage crag]

which one has higher battle iq

[amber fractal]

I'm betting on the baby

[fast pagoda]

always bet on DaBaby

[sage crag]

what if its an english exam

[sage crag]

https://tenor.com/view/uma-musume-umamusume-silly-urara-haru-urara-gif-5941918971645523248

animal farm

[rough bloom]

https://tenor.com/view/agnes-anvil-agnes-tachyon-pain-invincible-gif-2562559485872649862

[stark needle]

https://klipy.com/gifs/megumin-konosuba-10--k01KRS4J3TFZ3PV4CHA0XHZ2Y9M

uranium rods

mrrrrrp

[opaque wharf]

I need to learn chinese

[fickle rain]

Oops, forgor that I need to add it back :D

Yay, no more bugcheck

[sage crag]

https://tenor.com/view/evil-vtuber-3d-head-turn-concerned-gif-15495186653811856266

ye

restore stack

[kind nimbus]

if anyone wants this as well you can submit here:
https://claude.com/form/cyber-use-case

[obsidian mantle]

what does it allow you to do?

write exploits?

[glass flower]

technically

[obsidian mantle]

surely they will only give it to the good guys and there will be no issues with this system

[glass flower]

anthropic give me access to mythos...

[obsidian mantle]

use claude to get an exploit to access mythos to make ultimate exploits

[glass flower]

i'll be honest i wouldn't be able to afford to run a single query on mythos anyway... all the claude models are so damn expensive

[obsidian mantle]

well

is there anything that can stop them from making this thing stronger and stronger

they can just grow it more and more cant they

[glass flower]

i mean.. thats the goal of all frontier labs

google has deepmind

deepseek exists

[obsidian mantle]

"mythos, destroy this country please"

[glass flower]

openai now has gpt 5.5-cyber that is like mythos

[obsidian mantle]

fine tuned to do cybersecurity stuff?

or smth

[glass flower]

finding exploits. security stuff

[obsidian mantle]

or is it like

it just becomes so strong it can do it without any specific tuning

and nobody knows how it does it

[glass flower]

its for their glasswing thing

which is what anthropic does with mythos

[obsidian mantle]

is idea of clashing glasswing vs mythos being worked on

[glass flower]

[obsidian mantle]

can it potentially make them godlike while they fight

because they train in the process or something

or since they are llms it wont work

[glass flower]

openai and anthropic haven't had any new stuff they developed in a long time

[obsidian mantle]

just stacking more islands?

[glass flower]

any innovation in the AI space comes from Google or deepseek. and anthropic/openai just adapt the research into new models

they are currently more interested in turning a profit rather than developing new ideas

[obsidian mantle]

hmm

makes sense

so for google this thing is like a side job they are not feeding off it

if i got it right

[glass flower]

google is the reason gpt exists they wrote the original paper on how to do it. but just didn't release a model

they have been doing AI for like 3 decades at this point

[obsidian mantle]

but anthropic/openai are focused on this thing and have no other income basically

[glass flower]

google is easily the safest bet in the AI race

anthropic and openai are just the shiny things... but true AI will come from google

[obsidian mantle]

cool

[glass flower]

if they ever decide to actually take it serious

deepmind has been doing AI research for a long long time

it also helps that google has the whole thing under their belt. they do inference, models and the hardware to run the inference

they do the whole thing. so they are under no risk of external forces. like if nvidia goes down they take the whole AI bubble with them.. but google would be unnaffected

[obsidian mantle]

they have their own hardware?

[glass flower]

[obsidian mantle]

is it like fully their own thing

ai focused

cant buy it on amazon and shit

secret google chips

[glass flower]

pretty sure ye

[obsidian mantle]

damn

[glass flower]

i mean.. even their open source models have gotten a lot better lately. gemma 4 is a beast overall

google is just not focusing rightnow on benchmaxxing so they look bad and might not be the best in agentic tasks... but they focus on things that people might use in their products.. like the AI search thingy, the video summary on youtube. and ondevice small ai's

[frozen igloo]

Classic

[rough bloom]

TPUs

can rent on Google Cloud but can't buy

[glass flower]

if someone gave me $10k dollars and i had to bet it all on a company to win the AI race. i would put all the money in google tbh

they might be lagging behind the best and greatest models... but they always catch up

[true hemlock]

i'd say they're winning in terms of real breakthroughs

[obsidian mantle]

what about deepseek

[true hemlock]

i dont consider llm benchmark scores any significant.

[glass flower]

deepseek isn't self-sufficent. they rent gpu's thats the main reason i wouldn't... but yes they would be second to google

[true hemlock]

im talking about the deepmind team coming up with actually scientiically useful models like alphafold 2.

[glass flower]

[true hemlock]

screw the llm part that's unimportant as hell

[obsidian mantle]

is alphafold that chemistry protein thing or what is it

i think i saw some youtube thumbnail about it

[glad path]

yeah probably

[true hemlock]

protein folding prediction

[tropic spindle]

I had to read this 7 times to notice the difference

[glad path]

so it'd be good for designing drugs that work at a molecular level

[true hemlock]

idk man. google deepmind is the ONLY one making real academic contributions in the ML field.

[glad path]

it seems like it'd be pretty useful for medical research

[true hemlock]

the rest are cashgrabbing with llm shit and most they do is just modifying the architecture a bit

[opaque sigil]

https://www.isomorphiclabs.com/ they spun this off into its own company a while back

[mighty thorn]

“real” == unusable by the masses btw

Don’t let him fool you

SNN always stays closed and unpractical at scale

[true hemlock]

papers out there

if you can't implement the backend, skill issue honestly

[mighty thorn]

https://cdn.discordapp.com/attachments/1202474609518055424/1503566336679346277/IMG_9909.gif

[opaque sigil]

[mighty thorn]

He’s just cranky that his field has irrelevant for 4 years

[rough bloom]

it's fine

next year is the year of SNNs

[opaque sigil]

[obsidian mantle]

Snn is science neural networks?

[mighty thorn]

And Linux and communism and free hugs

[fast pagoda]

stinky neuro network, it's what neuro uses, ask the green fog

[opaque sigil]

very scientific

[obsidian mantle]

I am getting caveman vibes again

[stark needle]

bro u dont know

mixture of snns

✅

[stark needle]

bro

tensor networks bro

next gen shit

[true hemlock]

bro

quantum neural network bro

next level

the future!!11!11!

[stark needle]

actually tensor network is fucking insane

tensor networks is based in quantum computing or osme shit

https://arxiv.org/pdf/2401.14109

[mighty thorn]

x is crazy. Shit is the future
doesn’t know what it is

Shadow is an ai bro now

[stark needle]

u can represent bigger models with less params

via tensor networks

its fucking insane actually

literally

less params to backprop

and theres a fucking pytorch impl

[obsidian mantle]

Why is training speed an issue at all

[stark needle]

https://github.com/rballester/tntorch

[obsidian mantle]

Do some neural networks take like super long to train

And its viable because can shrink 10 years training to 1 year for example

[stark needle]

bro u can decompose a tensor into a graph

[obsidian mantle]

Or is it because every train is trial and error

And you can do 1000 attempts instead of 10 attempts for the same period of time

I am caveman please clarify

[stark needle]

https://tensornetwork.org/

schizo site

[mighty thorn]

Grokking

[obsidian mantle]

Why is there no 200b tensor llm that you can run on 16gb vram yet

Or is it wip

[stark needle]

u can in theory

ucan convert all tensors

[obsidian mantle]

Why arent they making it

I mean

[stark needle]

a company does this shit

[obsidian mantle]

You need x10 times less islands

[stark needle]

Model as a service

[obsidian mantle]

Why nobody uses it

[stark needle]

https://multiversecomputing.com/

[obsidian mantle]

Or is it new and wip and they are doing it rn

[stark needle]

this is shit from 2024

but the idea of tensor networks exists since like forever

[obsidian mantle]

Soo

Where models

[stark needle]

idfk

ive tried this shit tho myself

[obsidian mantle]

This justifies economy

[stark needle]

and it rly works

ive ran like 100M param pretraining

with tensor network uses like 10M params

[obsidian mantle]

Do you need some quantum computer to run it

[stark needle]

no

[true hemlock]

imagine having to iterate 30 billion weights times the highest context window, times the dataset length in tokens which is also in the billions to trillions of tokens

[stark needle]

gpu is fine

[obsidian mantle]

No i mean

I get that its super good

And overpowered

I ask why for example

Openai wont use it rn

[stark needle]

who knows whos using it

[obsidian mantle]

To make 10000000b model that fits into previously 1000b model size

[opaque sigil]

if it works at scale and doesn't fuck up everything they're probably using it already

[obsidian mantle]

And save 90% of money on building islands

[mighty thorn]

Imagine having to actually progress the field instead of gatekeeping and bragging with unverified claims

[opaque sigil]

granted that's a big if

[rough bloom]

apparently super good
really old
has PyTorch impl
somehow nobody is using it

[stark needle]

just gonna say that google had also a library

https://github.com/google/TensorNetwork

[opaque sigil]

archived

[obsidian mantle]

Wait are you saying its bad
I didnt quite get it
I felt like what you said was like.. benefit for it

[stark needle]

deepmind has alr schizoed into this shit

[true hemlock]

that's how much is going on with training llm. they had to optimize any caching related techniques on the attention heads just to make it viable

im just giving some examples for scale

[obsidian mantle]

So you saying that it works and everyone using it

[mighty thorn]

Alright so

Quack is an SNNbro

He hates ai industry

[obsidian mantle]

Why my gemma4 26b takes 17gb vram then

[mighty thorn]

But has agi internally but it’s just too dangerous to release

[true hemlock]

idk wtf is this dog on about lmao

[mighty thorn]

So basically

[amber fractal]

Weaksauce hating

[mighty thorn]

You know how OpenAI refused to release gpt 2 cause it was too dangerous

That’s where they are rn

They are gpt 2

[opaque sigil]

you're talking about two completely separate things

[mighty thorn]

They are still catching up

[amber fractal]

Kaine is yapping because they want to spread infomation, never said it good or informative tho

[true hemlock]

yeah im just giving scales on what's going on with pretraining a whole llm

maybe im leaving details like moe feed forward is much more compute efficient

[stark needle]

use RWKV or some shit bro

gated deltanet

[true hemlock]

idk, kaine is talking about something else

[mighty thorn]

Not capability wise. SOTA SNN can barely spell. But I mean they are still in the phase where they all think they are savants who just created sentient machines and need to gatekeep it for the protection of the human race

[true hemlock]

and being somewhat obnoxious for absolutely no reason whatsoever

[obsidian mantle]

So quack is talking about different thing completely did i get it right
Not talking about tensor model shrinking

Or what

[rough bloom]

yes

[true hemlock]

yes

[opaque sigil]

quack was talking about why training can take tens if not hundreds of million gpu hours

[obsidian mantle]

Oh

I see

[true hemlock]

yeah

[obsidian mantle]

Then this tensor shit is mega overpowered? No?

[true hemlock]

there's a reason we laughed off of a guy who thought he could train a 12b 1.58bit model on a cpu

[mighty thorn]

Quack would never be obnoxious about another persons interests, projects, or field of research

[true hemlock]

within few months

[rough bloom]

supposedly

but somehow never deployed anywhere

not widely known for LLMs at least

[opaque sigil]

hard to tell, most optimisations end up being quite specific and not applicable to everything so it's easier to just not bother

[rough bloom]

despite existing for ages already

[true hemlock]

this tbh

[opaque sigil]

i don't remember much about tensor networks to give an actual answer

[obsidian mantle]

Okay and then my gemma4 takes so much vram because gemma4 200b fitting in 4090 would be so strong i could exploit windows on it or some shit

[opaque sigil]

no-brainer optimisations are obviously used everywhere

[obsidian mantle]

While mythos needs only 500gb vram while being 500000b model

[stark needle]

it would still use a lot of vram for activations

[true hemlock]

god, im late for my fishing trip brb need to call an uber

[stark needle]

activations dont get reduced

[stark needle]

have fun quack

[obsidian mantle]

Oh so this thing is purely training related

And no x10 benefits on inference

[rough bloom]

no, it only affects the parameters I think

[stark needle]

only params are affected

[rough bloom]

shrinks the model itself, but not the space required for any computations with that model

[stark needle]

Tensor networks +deepseek compressed sparse attention

[mighty thorn]

That is absolutely not true

I want a source so I can ridicule it

[obsidian mantle]

Numbers are random

[stark needle]

bro mythos is just a wider moe

[mighty thorn]

HAHA RANDOM HAHA

[obsidian mantle]

But i already got answered

[obsidian mantle]

I mean i took these numbers out of my ass to pinpoint the scale of shrink

[mighty thorn]

That’s why it’s a fraud

[stark needle]

@mighty thorn u keep saying moe is shit but gemini and chatgpt and shit are all moe

[mighty thorn]

They would be better if the hardware used for them was used for not MoE

[stark needle]

that's just poor copium

[obsidian mantle]

So is this basically correct

[stark needle]

not economical at scale

[obsidian mantle]

I still need 20gb for 26b model

But its trained much faster

Trying to understand

[opaque sigil]

moe is good for what it's made for, which is allowing you to cram far more knowledge into the model for relatively cheap

[rough bloom]

MoE is awesome for big models

[opaque sigil]

no need to pretend it's worthless

[mighty thorn]

You either die dense or live long enough to see yourself become sparse

[rough bloom]

faster and easier to deploy

[obsidian mantle]

Moe is just mega fast inference

[stark needle]

moe is for expert paralleism on interconnected systems and shit

[obsidian mantle]

Which is good

[rough bloom]

yee

[mighty thorn]

MoE is worthless for LLMs in most contexts, but very useful in the contexts where it’s actually intelligent to use

[stark needle]

???????????????????

[fast pagoda]

but it's used in all sota contexts currently t

[stark needle]

bro moe makes llm infinitely cheaper to train

[fast pagoda]

doesnt seem particularly useless

[stark needle]

u just need more vram

but compute is the same

activations dont grow

u can use higher context length

[true hemlock]

i don't think kaine knows what he's talking about

missed the point of moe entirely

[mighty thorn]

• MoE is worthless for LLMs in most contexts, but very useful in the contexts where it’s actually intelligent to use

[obsidian mantle]

I tried moe vs non moe and all i noticed was x5 faster inference

[fast pagoda]

what is .... most contexts?

[stark needle]

most contexts

bro all sota llms are moe

[mighty thorn]

Because cheaper for improving benchmark scores versus actually making better models

[stark needle]

❓

[mighty thorn]

• Because cheaper for improving benchmark scores versus actually making better models

[stark needle]

bro

u gotta be fr

[rough bloom]

anything except tiny local models where you need as much density as possible because you have little VRAM
everywhere else MoEs win I think

[true hemlock]

in the enterprise space where they basically have endless hbm memory, moe is highly beneficial

[obsidian mantle]

But moe doesnt shrink required vram

[true hemlock]

less compute being used

[stark needle]

dense llms is just poor people copium

[true hemlock]

which is why i said endless memory

also

moe is basically

[mighty thorn]

It explodes it while having almost all of it being idle and unused

[fast pagoda]

the weakness of moe isnt a problem at sufficient active parameter counts
Something like Qwen3 30B-A3B knows things like a 30B model but is limited by the width of the active slice in any given forward pass

scale this to frontier size and the active parameters are in the tens to hundreds of billions again and it scales incredibly well when total params end up in the trillions

[mighty thorn]

Least efficient possible idea

[obsidian mantle]

Huh

[rough bloom]

and even locally MoEs are pretty good
it's dumb anyway so might as well make it fast

[true hemlock]

yeah

[obsidian mantle]

Didnt notice any explodes in 26a3

It just took like 17gb all the time

[true hemlock]

basically

[mighty thorn]

17gb for 3b parameter while the rest do nothing 😭

[obsidian mantle]

I understand that it assignes some "experts" that make it use only relatable params instead of all params evrry tjme

[true hemlock]

you got a model that inferenced just tiny part of itself but knows as much as the total params
though the drawback is that its somewhat dumber, but when you have like hundreds of billions params scale and work with enterprise scale inferencing does it even matter

[fast pagoda]

"nothing" is a misunderstanding of moe

[stark needle]

wrong when serving to many people u can get in fact higher utilization

due to expert load balancing

[obsidian mantle]

It only takes 3 but it selects it carefully

Whats the point of calculating impact of tokens that say 2+2=4 if it needs to write a poem

Its not a bad idea

[true hemlock]

ah great, he has never done batch inference before

[mighty thorn]

Assuming that they aren’t trying to use the same experts

[true hemlock]

[mighty thorn]

Which they will be

[true hemlock]

[mighty thorn]

Since only 3 are generalists while all the others are overfitted to Turkish history

[true hemlock]

did bro fail statistics class

[rough bloom]

[true hemlock]

crazy

[stark needle]

they arent bro

[rough bloom]

this is why you have load balancing in MoEs

[stark needle]

bro router z loss

load balancing loss

[true hemlock]

also when you get multiple people using the same experts

that's not bad either

you batch the damn matrix

[fast pagoda]

Multiple users hitting the same weights is the thing you want it's the entire economic basis of batched serving. Weights are readonly reads don't contend

[true hemlock]

exactly

[fast pagoda]

you can batch serve a dense model all on the same weights

[stark needle]

ye but then theres a ton of duplication

since u would be doing DDP

or TP if model too big

[warped narwhal]

Does anyone here use both nixos and jetbrains ides?

[true hemlock]

i don't use jetbrains

i use nixos occasionally though

shuni is proud

[mighty thorn]

LLMs have been on the wrong track since GPT-o1
Now everything is MoE CoT benchmaxxed distilled int8 marketingslop

[true hemlock]

well

[obsidian mantle]

linux

[rough bloom]

[fast pagoda]

throughput can get fugged for bad routing with one hot expert doing all or if the tokens all scatter like crazy that can be an issue but especially the single expert getting smashed is an issue of iimbalance

[true hemlock]

god im still waiting for sleep deprivation to hit me im sleepy as hell i need the sudden surge of energy

[warped narwhal]

I'm pulling my hair out trying to get flakes working with them, cause everytime I try using rust rover etc, it complains that it can't find the rust install because rustup isn't available (even though it is installed) and it fails to run it because it's a dynamic exe

[rough bloom]

Qwen kinda is that

all the big LLM benchmarks are close enough that I do not care

[opaque sigil]

surely there's a way to tell them where to find rust outside of rustup right

hmmm

[true hemlock]

the benchmarks don't tell shit anymore i prefer to measure models based on how enjoyable it is to converse with + how reliable it is with doing tasks a user requested

[warped narwhal]

It will find rustc, but not the standard library

[stark needle]

bro they can do more shit than earlier

earlier llms were unusable

actual braindead shit

[opaque sigil]

are you sure you even have the stdlib

it's its own toolchain component

[warped narwhal]

Yes, I can compile rust apps outside of the ide just fine, it's the ide itself that shits the bed when it tries anything

[opaque sigil]

[rough bloom]

ye basically that
they're still useful to know very roughly where an LLM is in terms of performance but you also usually don't need benchmarks for that

[fast pagoda]

yea i mean the indicator is like a black and white line

it's either in step with peers vaguely on useful tasks or it's total dogshit

that's about the max judgement to be had there

[rough bloom]

[mighty thorn]

I was having 4o do python and was fine with it

Then they forced CoT on us

Then MoE

And now we are here

[stark needle]

bro 4o was moe

[fast pagoda]

[mighty thorn]

5T frontier model with 400k active parameters

[rough bloom]

Mixtral

[stark needle]

mixtral was fucking god tier at that time

47b params for gpt4 perf

would beat the fuck out of llama 2 70b

[fast pagoda]

god

i archived the fuck out of mixtral as my "if the world ends" model to have lmao

at the time

you just reminded me

[mighty thorn]

Yeah

I have a few of those

[stark needle]

i had llms locally cause huggingface was down constantly

back in those days

in a fucking gitlab lfs

[opaque sigil]

i just tried and rust-rover seems to find my toolchain just fine
so idk what to tell you other than use rust-overlay I guess

[mighty thorn]

Stranded on island with solar powered laptop and 8b local llm, the creation of the ultimate schizo

[rough bloom]

OpenClaw, get me off this island

[opaque sigil]

[toolchain]
channel = "nightly-2026-04-03"
components = ["rust-src", "rustc-dev", "rust-analyzer", "clippy"]

rust-toolchain = pkgs.rust-bin.fromRustupToolchainFile ./rust-toolchain.toml;

[mighty thorn]

Make no mistakes

[stark needle]

bro just

fuck the laptop

get 2x gb10

similar watts

run deepseek v4

moe

[mighty thorn]

Ur actually delusional

Price check that for me

[stark needle]

poor people cope

[mighty thorn]

See how laptop like the price is

[stark needle]

my laptop was 5000$

or osme shit

[mighty thorn]

Mods ban this guy for discrimination

[stark needle]

just get a job bro

i was earning slave wages

for 4 years

[mighty thorn]

Suspiciously works 20 hours a week shaped

[stark needle]

while working at L4 technical level

as a fucking kid

[mighty thorn]

Shadow should do the thing where the rich person gets rid of all of their assets and qualifications and tries to rebuild from scratch to prove it’s easy.
He’d last a few hours before breaking down upon having to go to chick fil a instead of eating 50lb of caviar every meal

[stark needle]

bro

ive had

cold dms from random sillicon valley cofounders shit

on fucking discord

i didnt even need to show my cv lmao

just show technical competence

[mighty thorn]

Let’s see what the people say

Dirty 1%er

[obsidian mantle]

Age 20

I was living on 100$/mo

[fast pagoda]

me when i doxx

[mighty thorn]

Shadow spent 40x that on a watch

[mighty thorn]

No he’s said before

Probably

[obsidian mantle]

Well right now i am close to that

But im 28

Its like

Super huge difference

[umbral wigeon]

Let's vibe code

https://tenor.com/view/jaded-disappointed-down-let-down-computer-gif-143123934597902682

[obsidian mantle]

Its basically luck

Time can counter luck

Effort too of course

Cant do shit without it

[fast pagoda]

me, age 20

[stark needle]

"luck"

[obsidian mantle]

I discovered what programming is at 25 what do you want 💀

Yes luck

[umbral wigeon]

Shit
https://github.com/weenachuangkud/FastCast2/pull/39

[mighty thorn]

This says a lot about society

The richer Australian with multiple a100 and a collection of exotic GPU and cpu defends the less rich guy

0.1% and 1% coming together to gaslight the poors 😭

[umbral wigeon]

215 commits now

[obsidian mantle]

Wtf is he essaying there

[mighty thorn]

It’s gonna be a long form defense of his $4000 watch

Which is worth more than everything I own combined

[obsidian mantle]

I mean if you have 10k per month income why not

Literally my current job but in US gives exactly that

[mighty thorn]

12 days for a watch 😭

[obsidian mantle]

Yeah

[true hemlock]

which one is which

[mighty thorn]

How many A100 do you have?

[obsidian mantle]

But due to low luck im not in US so i get 1.2k

Yes luck x2

[obsidian mantle]

Did you apply to google at 15

[stark needle]

Ye

[mighty thorn]

Child labor is the answer to poverty, ladies and gentlemen

[obsidian mantle]

Did you speak english at 15

[fast pagoda]

i don't think a combined spend of like 12k-15k on tools that are used in your line of work + less than 5k hobby shit constitutes rich

rich is a whole different ball game than anything described

[umbral wigeon]

Is it legal? To let 15 y/o working in big tech? Like in your country

[obsidian mantle]

Did you know what programming is at 15

Yes luck

[umbral wigeon]

My country can't(laws, school system)

[mighty thorn]

Ur also rich but I let you slide cause no $4k watch

[stark needle]

I could easily have passed c1 cambridge in middle school

[fast pagoda]

[umbral wigeon]

What your country?

[stark needle]

How do u think i got in

[obsidian mantle]

There was no human in my town who would pass C1

I am trying to explain that you got lucky to be there where you were

With your skills

[stark needle]

Swiss

[umbral wigeon]

Do you have github, I wanna explore

[rough bloom]

the power of having different circumstances

[stark needle]

No my gh is empty

[obsidian mantle]

Good education too

[rough bloom]

and putting in the effort to actually use those circumstances

[fast pagoda]

spawn luck

[mighty thorn]

No room for said effort without the prior luck

Can’t build skyscraper on dirt

Need foundation

[obsidian mantle]

Hesright

I wasn't supposed to speak English at all for example

I just randomly stumbled upon forsen stream and understood 10%

[umbral wigeon]

How each eduactions system in different countries even works, I mean how did someone at 15 able to get phd, do they just max stats and beat it to phd?

[fast pagoda]

the ability to be active in this discord to the extent that the majority of people in this conversation are is already indicative of a similar type of luck if not the same extent
there are entire swathes of the world who have no inkling of the wonder we interact with every day

[obsidian mantle]

Idk about phd at 15

You need to be noticed

[stark needle]

I dont have phd idk

[obsidian mantle]

What if there nobody to notice you

Will you aplly to google

"hello i can make calculator im 15"

[mighty thorn]

https://tenor.com/view/socrates-spinning-gif-2769471696194500629

Thank you Socrates

Very cool

[stark needle]

Yeah i would have

[fast pagoda]

you're welcome, i won't charge for that one since you're on the brink it seems

further will require an active credit card on file, however

[umbral wigeon]

I'm active on discord, cus I just wanna yap and share or talk, I wanna doom scrolling tiktok, discord, facebook

[fast pagoda]

oi

he's stuck like that

dont make fun

birth defect

[frozen igloo]

It’s a rare disease but not unheard of

[umbral wigeon]

And to stop doom scrolling, I'll shut down myself(sleep)

https://tenor.com/view/donald-duck-sleep-gif-12073235593935901891

https://tenor.com/view/threddy-threddyrex-windows-updates-gif-20412365

[obsidian mantle]

I read it fully now

Holy shit

[umbral wigeon]

I think programming is easy

If autistic person can do it

[obsidian mantle]

Is like "i bought 10 lottery tickets and won them all"

Congrats

[umbral wigeon]

Because I'm autistic

https://tenor.com/view/bosnov-67-bosnov-67-67-meme-gif-16727368109953357722

[mighty thorn]

Get out

[obsidian mantle]

I also fixed previously unknown bugs and got +100$ raise at my job
I guess i played my cards alright

Now company is dying and idk if i should flee the country or go to capital

[umbral wigeon]

If I was a hiring manager, I would like to take a look at github than CV or resume honestly

[obsidian mantle]

My managers dont know what github is

[umbral wigeon]

Or any profile that shows you have passion in the tech

If someone actually done something impressive, they should have it all, either github, your facebook, your X as portfoilo to show they're actually facts

Because how I can believe if someone came up to me and say, I've graudated from cambridge and got a job at google 15 years old, I mean it's not bad
But I would like to see github profile slams on my face or any official papers that you're actually done it

[obsidian mantle]

Its super rare but not impossible

I guess

Not like i ever witnessed anything like that myself

[umbral wigeon]

I'm sorry if my opinions hurt your feeling

[umbral wigeon]

(Also there are intreview rounds, coding problems screen share idk about that cus I've not experneiced it)

[obsidian mantle]

However i find it concerning that you have to work in google at 15 to afford that shit

In 5 years

[stark needle]

I mean are u asking for proof or wym

[umbral wigeon]

No, I'm just sharing my opinions

Not everyone have proof

I've heard that job hiring are using Ai to review CV/Resumes, it is that bad?

[rough bloom]

getting a job at Google that early is probably not impossible if you can show technical competence and find the right opportunity

[young plover]

I did have cool stuff on GitHub but Nintendo nuked the main project that I contributed to.

[rough bloom]

not unbelievable at least

[stark needle]

make a switch 2 emulator ✅