programming

[sage crag]

[rough bloom]

have to go around now

[stark needle]

[sage crag]

kufu

[bitter phoenix]

I might already have, technically

[rough bloom]

[bitter phoenix]

Funny thing, when it comes to datasets of my own messages if I don’t dedup then it’ll always mention NixOS

[fast pagoda]

uncleaned ones from me used to just be neurotic neurons output from the model

[stark needle]

chat

is this real

i was just fucking paged

cause the db team decided to do a fucking password rotation

without announcing or having a dependency graph or anything

and i was now suddenly the one having to find all our services that depend on that goddamn DB

[rough bloom]

random dependencies breaking stuff is fairly normal I think

the DB team really should know better though

[opaque sigil]

[stark needle]

ugh

but there should be some system

or something

imo

that checks

wtf is affected

[rough bloom]

ye there usually should be

if you do it properly you have change management

and documentation for dependencies and stuff

and people who approve changes and are supposed to check this stuff I think

[split steppe]

i would page the db team back to tell them to change it the fuck back

[rough bloom]

but it's a lot of effort so nobody really does that

especially the documentation part

[split steppe]

you would think with all the AI hypeslop today they would atleast document something

literally the one thing it's decent at, but nobody uses it for that

https://tenor.com/1zI4.gif

db team like

[stark needle]

[stark needle]

just make a fucking shared calendar

takes 3min

[split steppe]

talking to people is a crazy possibility too

[rough bloom]

in my experience that exists for changes

but then

nobody actually schedules these changes when it's small stuff like changing a password

[split steppe]

who cares, it's only prod

[rough bloom]

[split steppe]

wanna bet someone's openclaw found the ssh keys and decided to "improve security"

next time you're paged it'll be with a bucket and a mop

[fast pagoda]

every week baby

[mighty thorn]

https://cdn.discordapp.com/attachments/1067638175478071307/1440788419868364853/evil-godme2.gif

[stark needle]

gdd

[hallow dagger]

hi guys look opengl ferris wheel for my computer graphics coursework

they liked it

[tropic spindle]

OpenGL in the big '26

[stark needle]

looks cool

did u make the whole engine

(with opengl i mean)

[opaque sigil]

[hallow dagger]

yes
well calling it an engine would be giving it too much credit
but yeah

it would have been a lot better if i actually worked on time. we had 3 months and i did this in 3 days

i would have had a complete game engine and i would have done wayyyyy more than the assignment specification

but whatever

[stark needle]

are u doing game design or something

[hallow dagger]

nah it's just computer graphics module
in my computer science course 2nd year

there's an optional game dev module for 3rd year but i didn't pick it

actually i don't remember what i picked for next year

[stark needle]

damn i cant relate to anyone here

everyone doing coursework

[hallow dagger]

what U do

[stark needle]

i went middle school -> apprenticeship -> software engineer

[hallow dagger]

wait bus stop

[mighty thorn]

Hey I’m a failure too

Your not alone

[stark needle]

[mighty thorn]

Anywho

I had to switch to byte tokens for 175k model

Tokenizer was funky and still too big

And now I’m reusing every layer to get 6 layer depth from 3 layer model

[stark needle]

byt5

[mighty thorn]

I’m trying to implement spelling based loss masking so it gets better at making words that are words

[stark needle]

thats just crossentropy loss

llm has no idea what a word is

[nocturne olive]

175K parameter LLM?

[mighty thorn]

Yes

Trying to get 1,000,000x smaller than gpt 3

[nocturne olive]

That be absolutely tiny

[mighty thorn]

Yes

Is goal

[nocturne olive]

[mighty thorn]

Can probably run on microcontroller

[mighty thorn]

Here are some outputs from ~425k version

`Prompt: 'Once upon a time,'
────────────────────────────────────────────────────────────────────
Once upon a time, it says that the new aqualced vain. They're a
great shot. However, it's not sure the bottom of the shot. Here's
the right way to make the treety on the ped, but it's a sery
audience to be a touted band your chief. The time is any of them.
That's the right way. We're going to run over my timely covered
the ped, so it's a mision. No, many are it that it's just not a
batch of the gest. The thin lies in them. It's some is the
aqualced lies. It's a small perhina bills from the ped, and the
bottom of the heddlin-daughter, I'm a small hesy. I'm sure I'm
sure
────────────────────────────────────────────────────────────────────

Prompt: 'The history of science begins with'
────────────────────────────────────────────────────────────────────
The history of science begins with an extremery of ligand. The CHS
was expected to be 64.5, 81. (0.3) were rephenal for a 1910-caling
(1). The DO Group Host. Lecenter has been assessed by Slen.
Coscosis and Osticlellar and the TDO NB. DS (3,8,115) (2012.
Although these theowers can be noted that Row-Gedo, Lero Jahrezar.
The DO Illinois is a long-ends in People of the Katkon Hatbay.
Although the SWARC Agency's TDHS was a wound of fan-shands and the
TDHS does not be aware of the theore
────────────────────────────────────────────────────────────────────

Prompt: 'def fibonacci(n):'
────────────────────────────────────────────────────────────────────
def fibonacci(n):ni, 0.0, 1.0), $$ is a gentle and gentle. Our
bottom-sides of puvillan or a core bottom sitis to a dejr. As the
"Aputing Psychology" of Rezo and Rezo G. J. Cath, Mantao B.
Balnic, Layerz, Bamhro Danick Moschy. Surgical B. Tanzi, Aputing
Marro, J. Cath-G. Hidrative Liaty Watning, Aputing Motori, Donke
Surgical Bolhma, Banan B. Surgical G. Malya, J., Laelonon G.
────────────────────────────────────────────────────────────────────

Prompt: 'The Wikipedia article states that'
────────────────────────────────────────────────────────────────────
The Wikipedia article states that the Buugal Prover Panigching in
a Molel Buugal Bitcoin-Samace, is to determine the form of an aim
of an inrairds on the Panigmen-Samace of the Buugal Ciocongar. The
Panigchings of the case was an important number of nasiilhous, the
case also was a bym was a literal aim of the Ciocongar. Loving to
Tetrifana, it has been a aim of the Molel Ciocongar. The theower
of the Tetrifana, the bym was ineffective. The Panigching and the
daily feser, which is an inserry with the Drondbest and the Ciocon
────────────────────────────────────────────────────────────────────

Prompt: 'In conclusion,'
────────────────────────────────────────────────────────────────────
In conclusion, but there is a most relevant time. This is not a
majority of stories, even though this is more than a few times.
A.D. Non-D. NOO RRBTT) is an “breaka” in the Pr'Sen, as the audim,
that the Southern F. Troxis's Beldeer J. Sect. Nor's audim, a
Markown is the Lelicy Mull, the Sect. Nor's Senate. The Sifland
Stadium is concluded in the liber and this does not
incompetifians. The VA Heny's Oesp, which was used to be felt on
the Main, we have a push, and there was no reasonsred to the
stories. The Elanzi,`

Used ALBERT factorized tokenizer embeddings with 1024 vocab for that one

But going down to 175k param requires switching to 256 token byte vocab

Plus the actual tokenizer runs were weird

[nocturne olive]

Now I kinda want to make a mini-LLM on my 3090

[stark needle]

fyi byte level models are extremely inefficient

at learning

data needs to be compressed way harder

[split steppe]

don't byte models have not enough entropy to actually get good probabilities per token?

[mighty thorn]

I can pretrain this one at 256 batch size (real batch size, not grad accum) on my 3070ti laptop

[stark needle]

yea

[mighty thorn]

As is the tradeoff of making something 1,000,000 times smaller than gpt3.5

[nocturne olive]

I could probably pretrain a much bigger one at a still reasonable batch size

[stark needle]

byt5 is barely able to generate remotely coherent

[split steppe]

i would think the tokens are the ones you don't want to skimp on

[stark needle]

and that one was pretrained on C4

the whole point of tokenization is to compress information

[split steppe]

yes but you can compress too far and fundamentally then you don't have enough space to encode the probabilities you need

[mighty thorn]

If byte tokens are too icky then I’ll try going back to the BPE tokenizer but that’ll require making the model itself even smaller

[nocturne olive]

I wonder what a 1M-100M param LLM would do

[fast pagoda]

blabber

[mighty thorn]

Gemma-3-270m

[fast pagoda]

classify stuff tho

mainly

[mighty thorn]

That’s only 100m of actual model

The rest is embedding

[sage crag]

pluu

[nocturne olive]

[mighty thorn]

The byte model speaks like konii

[true hemlock]

you already seen it like 2 years ago

[split steppe]

in conclusion, i'm having a stroke

[fast pagoda]

t5-small was like

76M

[nocturne olive]

But back then I sucked at making data and the code also sucked

[fast pagoda]

https://huggingface.co/Luigi/Falcon-H1-Tiny-Multilingual-100M-Instruct-GGUF surely this thing is crazy wicket smaht

[split steppe]

175k could be doable if you really tune your token count and specialize it towards a particular style of output

[mighty thorn]

Spelling loss is making suspiciously word shaped words

[split steppe]

i don't think you can get a general model out of it

[mighty thorn]

If I didn’t know English I’d believe it

[fast pagoda]

like those videos of what english sounds like to non speakers

theyre out of a job now

[split steppe]

lol true

[fast pagoda]

holy

what kinda code boys and girls

does this cook up

[sage crag]

i have my discord message data

[fast pagoda]

116mb

[split steppe]

i would probably not go less than the common english vocabulary in token count, or maybe half of it

so 1-3k tokens

[mighty thorn]

I’m sooooo gonna finetune this on gpt-3 and 3.5 to make ChatGPT run on 1mb vram lmao

[nocturne olive]

"For desktop/server"

113MB

[mighty thorn]

Give us the tokens per second speed coward

[split steppe]

"only use if you have a high-end machine with at least an Intel Pentium 4 and 256MB RAM"

[stark needle]

sharegpt my beloved

[nocturne olive]

Literally my potato phone could run that thing

[mighty thorn]

Wait a minute

I could totally pretrain the 175k model on my phone

And that would be really really funny

[stark needle]

train a llm

[nocturne olive]

[split steppe]

that would indeed be quite funny

[sage crag]

ye train on

ryzen 5 7600

[nocturne olive]

I don't have that

[sage crag]

32gb ram

[mighty thorn]

I’m using a 4b token subset of the pile so I could store the entire thing easily

[mighty thorn]

You can export it

[stark needle]

[mighty thorn]

Or do it the based way

But that’s for whispering about during recess, not public chat

[stark needle]

ryzen 8 12380f

[nocturne olive]

You can export data from Discord?

[split steppe]

theoretically yes

[nocturne olive]

What do you mean "theoretically"

[trim valve]

yeah wdym theoretically

[mighty thorn]

Idk what he meant

It’s in settings

Is a single button

[trim valve]

you can just request your data in settings

[nocturne olive]

I wonder what that would give me

[trim valve]

other people's, no but

depends how much you opt into

[stark needle]

[trim valve]

for me its just messages

[split steppe]

discord may have different controls depending on your location, like most apps today

[nocturne olive]

Well we'll see what happens

[split steppe]

"privacy farmers" sounds about right

[rough bloom]

ye was about to say

[hallow dagger]

hi

[rough bloom]

that's NOT how I would've worded it

[stark needle]

truly the 差點猜中 moment of all time

[opaque wharf]

I am doing work tho

[stark needle]

i love random chinese letters

i cant even ocr

[fast pagoda]

like 350 was what i was getting and i was like aw that's bnot exciting

gpu was disabled

cpu only

[stark needle]

wait this is taiwanese

https://klipy.com/gifs/infinitesocialcredi-socialcredit--k01KRBZCAC5K4GWN6D8NW6513JQ

[opaque wharf]

I think taiwan and china share a writing system

Or are they really that different?

They are both mandarin right?

[fast pagoda]

dont say that over ther

[fast pagoda]

still getting like 400 only, querying completions it'll only poop out a fibonacci numbers sequence for 100 tokens no matter how long i set the response

io limit of some sort

2026-05-11 11:55:07 [DEBUG]
slot print_timing: id 3 | task 900 |
prompt eval time = 5.25 ms / 2 tokens ( 2.63 ms per token, 380.73 tokens per second)
eval time = 252.91 ms / 99 tokens ( 2.55 ms per token, 391.44 tokens per second)
total time = 258.17 ms / 101 tokens
slot release: id 3 | task 900 | stop processing: n_tokens = 100, truncated = 0
srv update_slots: all slots are idle
2026-05-11 11:55:07 [DEBUG]
LlamaV4: server assigned slot 3 to task 900

[opaque wharf]

I really have no idea how different their language is

I know mandarin has dialect and all but not to the point of actual different language altogether

[hallow dagger]

traditional chinese

[stark needle]

are u from china

[mighty thorn]

i have an idea

auxillary 16k vocab tokenizer, stolen from an existing model

give it two output heads and just discard the BPE embeddings and head for inference

still 175k parameters but now hidden states are more aligned with full(ish) words

Special Agent double-o-@stark needle, any thoughts?

[stark needle]

why not just mmap embedding table on disk

like gemma 3n/4

[mighty thorn]

still counts as model weights

memory isnt issue, low parameter count is goal

[stark needle]

wait i know

wait no nvm

[sage crag]

dududududududududududu

[fast pagoda]

sandstorm

[sage crag]

[fast pagoda]

lutel

[stark needle]

upward project embeddings

best possible setup i think

[sage crag]

ye

[rigid snow]

neur obuxket

[sage crag]

⬅️

[rigid snow]

[sage crag]

➡️

[rigid snow]

what the fuck is she saying

[sage crag]

[rigid snow]

me schizo

[sage crag]

↗️ ↘️

[rigid snow]

ok

👍

[fast pagoda]

https://thehackernews.com/2026/05/new-linux-pamdoora-backdoor-uses-pam.html?m=1

Yay

[sage crag]

[rigid snow]

when kernel eop exploit that can work with my phone

2018 kernel 2021 security patch

[sage crag]

[rigid snow]

i was thinking about it

[sage crag]

[rigid snow]

3627 because it has six seven

[sage crag]

🧠

🔨

🐿️

[rigid snow]

it’s the chip ye

[sage crag]

🟣 🔨

123

abc

🦭

[rigid snow]

🔢-4️⃣ 🔤

[sage crag]

[rigid snow]

no 123 emoji

[sage crag]

qualia

3 flipper rolls

[opaque wharf]

That requires the malware to be installed via other means tho

So its just another type of malware rather than vulnerability

[hallow dagger]

who knows

[nocturne olive]

Today I found out
Ubuntu is equally as bloated as Windows is

[trim valve]

equally might be stretching it but ubuntu certainly isn't lean

[nocturne olive]

ISO size 6GB for both
Install size around 25GB for both
4/6GB RAM needed

[trim valve]

clearly this is your sign to move to nixos

[nocturne olive]

I need Ubuntu on my laptop temporarily for if I need to go to an exam where Ubuntu is an acceptable OS

[trim valve]

that's certainly a choice

[stark needle]

isnt there like a setting when installing ubuntu of how much shit u wanna download

[hallow dagger]

ubuntu is poop

[nocturne olive]

[trim valve]

you get a choice between "full" and "minimal" I think?

one installs a bunch of random apps

the other just has a web browser, a terminal and a file explorer

[stark needle]

yea

my raspberry pi has like a 16gb sd card

and it works with ubuntu

(im not using raspberry pi slop os)

[nocturne olive]

[mighty thorn]

What that

[stark needle]

keep embeddings at like dimension 4-8

and then have a single shared ffn that projects that to 64 or whatever

so ur embeddings are small

but ur model isnt complete shit

[olive sable]

goodmorning

[opaque wharf]

The views from my morning

[olive sable]

[stark needle]

8pm

[olive sable]

i had a litle eep

[stark needle]

selep

[sage crag]

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

[nocturne olive]

("1 ") * 23

[olive sable]

sure i guess

[olive sable]

neurt

[opaque wharf]

Would you rather breathe manually instead?

[olive sable]

yes

i have Hereditary Pancreatitis

[hybrid cobalt]

has anyone seen this firefox bug before? somehow neuro's blog site triggers it

[fast pagoda]

Arg

[warped narwhal]

No, but can you please blink manually for me instead?

Oh, and you have to hold your jaw shut

[opaque wharf]

Hocus pocus, your tongue is now in an uncomfortable position

[stark needle]

damn is it bad

or not

idk

im not doctor

[olive sable]

it is

if i eat too much cake i feel like shit

doesnt stop me tho

[stark needle]

😭

damn

[olive sable]

its not severe at all

i dont know any person in my family that died form it

its not a medical issue, its just kinda weak

[stark needle]

damn

i still need to go to the endocrinologist

[olive sable]

An endocrinologist is responsible for evaluating diabetes, bone loss, and a range of hormonal issues, including hormones from the pituitary, adrenal, and thyroid glands as well as reproductive organs

i see

[trim valve]

the neuro site is flawless and runs perfectly fine

[olive sable]

[hybrid cobalt]

only about 30% on my machine

[mighty thorn]

Do they not have school in Belgium

[olive sable]

in dutch ye

not in english

besides college

[warped narwhal]

I'm sorry for your loss

I hope you get better soon

[olive sable]

english isnt much bettter than dutch tbh

im sorry for your loss

[warped narwhal]

[maiden geyser]

we hebben een serieus probleem

[olive sable]

nee

[trim valve]

for whatever reason firefox is using my igpu

on my laptop

[opaque sigil]

seems like a good thing

[trim valve]

but yeah it looks like the site makes your gpu unhappy

[proven merlin]

Wow, a website thats poorly optimized

"Do you have any idea how little that narrows it down?"

[slim sorrel]

Most sites??

[sage crag]

2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2

[olive sable]

3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3

[sage crag]

samwise gamgee

[olive sable]

sauron

[stark needle]

🪲

bguu

[proven merlin]

Yeah that was the joke/play-

[sage crag]

[olive sable]

i mean

would you rather be bilbo or frodo?

[sage crag]

[stark needle]

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

fox ying idol

[olive sable]

i guess in terms of woman, there's shelob

[sage crag]

[olive sable]

the 10m long spider

[stark needle]

bguu

[sage crag]

[olive sable]

who would you want to be?

[sage crag]

kni

[olive sable]

cuz there arent many woman in the cast looking back on it

[sage crag]

galadriel, eowyn

[olive sable]

and arwen

[sage crag]

ar

[olive sable]

thats like 3 charachters

ignoring the non-humanoids

[sage crag]

[olive sable]

Goldberry?

[sage crag]

npluu

[olive sable]

Rosie Cotton i guess

there's some random hobbits

[stark needle]

what does pluu even mean

[sage crag]

pluu <- pluu [pluu]

[stark needle]

[olive sable]

i have no fuckign clue

the moment i start to understand what she's saying based on vibes, new words get added to the konii lexagon

[stark needle]

started with ✅, somehow ended up with pluu and ying yang and idk what

[olive sable]

even as someone with autism, its hard to decipher at times

might be a british dialect of autism

[mighty thorn]

The Chinesification of Konii (viewer discretion is advised)

[trim valve]

[nocturne olive]

Apparently memory usage increasing quickly enough can crash Discord

[stark needle]

fr

someone needs to do an entropy chart on koniis messages over months or something

[olive sable]

i need to use my new local llm to research this shit

[trim valve]

@samgpt

[stark needle]

@sama

[trim valve]

[stark needle]

god my eye keeps twitching

i hate this

[trim valve]

I wonder how I should torture my gpu next

[olive sable]

ye

ive been wondering

how do i make training happen on data?

what format does the data need to be in?

i need it to be as dumb as me

[trim valve]

I've only seen it done for markov chains

[stark needle]

[
  {
    "role": "user",
    "content": "Hello"
  },
  {
    "role": "assistant",
    "content": "Hi. How can I help you?"
  },
  {
    "role": "user",
    "content": "Explain what an API is."
  },
  {
    "role": "assistant",
    "content": "An API is a way for different software systems to communicate with each other."
  }
]

this format is expected by almost every training framework

[olive sable]

uhhh

well, my data is in the format of discord files

[mighty thorn]

Convert

[olive sable]

how?

[mighty thorn]

Python

[stark needle]

well good phcking luck lmao

[mighty thorn]

I mean

[stark needle]

but ye python

[mighty thorn]

Afunyun has his own fully working one to my knowledge

Your other options are diy and ChatGPT

There is also the fact that you only have your own messages

[trim valve]

do you want a series of really terrible suggestions that will potentially anger shadow

[mighty thorn]

So you only have a single turn per conversation

And thus nothing to train the model on

Ie the text completion regression we talked about last time you were testing this

So your options are:
somehow get the full conversations (slow and against tos)
use LLM to synthesize fake messages for your real messages to be in response to (slow and imperfect)

[olive sable]

[trim valve]

tbh its basically the last point kaine said

but a bit more advanced

group messages by channel, and then take blocks of messages that are nearby in time and generate messages to make a conversation

[mighty thorn]

Okay well now I’m interested too

[stark needle]

wym thats literally what i would have done

[trim valve]

idk its a 50/50 if its a genius llm invention or the dumbest shit imaginable

I need to make the bred #programming markov chain

[sage crag]

what if

you ask for everyones discord message bundles

then mix them together and order the messages by time

[stark needle]

but usually if u only had one side of messages i would have done reinforcement learning but thats hard to set up for beginners

[sage crag]

group messages into conversations

[trim valve]

[mighty thorn]

Consent is slow

[trim valve]

I'd contribute my #programming messages

and everything in the arg channels for good measure

[sage crag]

1 llm trained on the entirety of

#programming

[olive sable]

i want it to be mostly trained of me

so i wonder how id do that

[trim valve]

(only the premium chatters)

[stark needle]

i would do it if i had the data

[sage crag]

llm trained on only

new leaf chatters

[stark needle]

llm trained on only ying and pluu

[mighty thorn]

I would if it weren’t illegal

[olive sable]

good thing its illegal i guess

-# we have enough with 1

[mighty thorn]

Unless you talk to yourself in text channels, there will have to be other user(s) involved, whether real or fake

Actually

Hold

[trim valve]

folder preview jumpscare

[olive sable]

looking at the discord data they give you

they really do have everythign huh

nothing bad is in there, but i do kinda die inside seeing my old messages

[stark needle]

[olive sable]

my discord account is pretty old

some of these messages are form when i was 13

and i kinda want to nuke the files from messages with certain people

[stark needle]

sam's secret

[olive sable]

its not really a secret

ive had a gf before

[amber fractal]

I would also want to just send in anything from neurocord

Not touching the others

[olive sable]

fair

[amber fractal]

I need this ran at very high temp on a throwaway user

[stark needle]

:neuroCatUuh:

[amber fractal]

Just to see how cursed it gets

[olive sable]

2025-07-15 18:08:03 i used the word everynyan

[amber fractal]

Hello everywan

[maiden geyser]

hello everysam

[trim valve]

ok markov chain is fun

[olive sable]

[stark needle]

awa

[olive sable]

oh actually

is should remove doxxable info

[stark needle]

Name doxxed

[olive sable]

i mean like my adress and stuff

[amber fractal]

I mean it is a good idea to

If you ever want to let it run wild then do so

[leaden crest]

all we have to do is test a few numbers and eventually it’ll autocomplete your address

[olive sable]

it doesnt exist yet luckily

[leaden crest]

chat I had a reading/writing exam (of sorts) today

topic was agi

paragraph 2 I just see a sam altman quote

absolute cinema

[trim valve]

bredgpt

[stark needle]

It's actually good

[olive sable]

found the folder with the text i sent into game jam 2024 channel

[trim valve]

import json
import re

import markovify

DISCORD_EMOJI_REGEX = re.compile(r"<a?:(\w+):\d+>")
LINK_REGEX = re.compile(r" ?<?https?:\/\/(www\.)?[-a-zA-Z0-9@:%._\+~#=]{1,256}\.[a-zA-Z0-9()]{1,6}\b([-a-zA-Z0-9()@:%_\+.~#?&//=]*)>?")

with open("messages.json") as r:
    messages = json.load(r)
model = markovify.Text(
    None,
    parsed_sentences=[
        LINK_REGEX.sub("", DISCORD_EMOJI_REGEX.sub(r":\1:", msg["Contents"])).replace("\n", " \n").split(" ") for msg in messages
    ],
)
# print("\n".join([str(_) for _ in model.chain.model.keys()]))


for i in range(5):
    # new = model.make_sentence()
    new = model.make_sentence_with_start(":neuroPogHD:", strict=False)
    print("---------")
    print(new)

easier than expected

runs pretty fast on ~18k messages

[olive sable]

may i yoink this?

[trim valve]

yeah

you need to pip install markovify

(changed a single line so it doesn't print the entire vocab)

[stark needle]

Make a konii markov chain

[trim valve]

I don't have konii messages

[olive sable]

i wonder which one of these folders is the #programming one

[trim valve]

[stark needle]

11.8mb

https://www.marktechpost.com/2026/05/09/nvidia-ai-just-released-cuda-oxide-an-experimental-rust-to-cuda-compiler-backend-that-compiles-simt-gpu-kernels-directly-to-ptx/

[olive sable]

i feel that

[shrewd cloak]

🤔 isnt that the claude code one iirc

[trim valve]

you can also make it not require at the start

I just found it funnier

[fast pagoda]

bro echo has deadass regurgitated some ultra doomposting from an old relationship at least once when he was overcooked supreme

[olive sable]

[fast pagoda]

i nearly died seeing it

[olive sable]

i dont remember being that pedantic about trademark infringement

its not your name in hebrew perchance?

[leaden crest]

jsut is so real

[trim valve]

[stark needle]

What

Lmao

[olive sable]

this is hilarious

[stark needle]

wait what do i need to do to run this

[olive sable]

your discord messages and the python script

[olive sable]

wtf was this last thing even?

aight i finished my refactoring and i murder my queen, this allows me to 
Laptop bluescreens 
Laptop restarts 
"Updates are underway" 
Very subtle microsoft

i get that i was rudely intercepted by a bluescreen, but i was murdering a queen???

[stark needle]

how long did data export take for u sam

[olive sable]

a couple hours

---------
Oh so you breathn't and there's no jpreg
---------
too much :welpsagiri:  
this is the most fool-proof is 4
---------
:DIESOFCRINGE: i already have it
---------
fuck im doing this ai shit ran locally ye
---------
ye, those are not words i could pregenerate the entire scene in a week ago

jpreg???

im sure i meant jpeg lmao

[stark needle]

sagiri

[olive sable]

welpsagiri

[stark needle]

truly the anime of all time

[olive sable]

i dont even think about the anime it comes from when posting the emote tbh

[fast pagoda]

bro i love markov chains sm

[floral hawk]

I'm working on the twitch/OBS integration, set up plugins and all that and it may legitimately be worse than building the actual AI in the first place

Idk why but I've always hated OBS

[fast pagoda]

me with gimp

[floral hawk]

So fucking real

[fast pagoda]

and docker

[floral hawk]

I used to have a side hustle making avatar tattoos in vrchat with gimp

[floral hawk]

It was like a money printer, shit was so easy

[fast pagoda]

every time i use gimp that fucking thign pisses me off so bad it's unreal

[olive sable]

Crime is illegal, but it only 900???

[fast pagoda]

so cheap!

[floral hawk]

What's your issue with gimp, if I may ask

[mighty thorn]

Streamlabs obs is bad

Obs is good

[olive sable]

looking at my older messages i was thinking the llm would be insuferable. but the neurocord ones seem tobe funny

[floral hawk]

Streamlabs is gonna be how I hook up Zero bc it has API integ

[fast pagoda]

the UI is just completely unintuitive

[floral hawk]

So right now Im unaware of better alternatives

[fast pagoda]

to me

like it's powerful

and that's why i keep using it

because i know it is

[mighty thorn]

Suspiciously obs websockets api shaped obs

[fast pagoda]

it's just i cannot fucking do what i want to do no matter what

[trim valve]

@olive sable ```python
import json
import re

import markovify

DISCORD_EMOJI_REGEX = re.compile(r"<a?:(\w+):\d+>")
LINK_REGEX = re.compile(
r" ?<?https?://(www.)?[-a-zA-Z0-9@:%.+~#=]{1,256}.[a-zA-Z0-9()]{1,6}\b([-a-zA-Z0-9()@:%+.~#?&//=]*)>?"
)
MENTION_REGEX = re.compile(r"<[#@(@&)(@!)]\d+> ?")
SPLIT_POINTS = re.compile(r"[ ./]")

with open("messages.json") as r:
messages = json.load(r)

model = markovify.Text(
None,
parsed_sentences=[
SPLIT_POINTS.split(
MENTION_REGEX.sub(
"",
LINK_REGEX.sub("", DISCORD_EMOJI_REGEX.sub(r":\1:", msg["Contents"])),
).replace("\n", " \n")
)
for msg in messages
],
)

for i in range(5):
new = model.make_sentence()
# new = model.make_sentence(max_overlap_ratio=0.99, max_overlap_total=300)
# new = model.make_sentence_with_start(":neuroPogHD:", strict=False)
print("---------")
print(new)

[fast pagoda]

it's always pain

[trim valve]

this might help it a bit

tl;dr if it picks up too many one-off phrases it becomes stupid

[floral hawk]

I know I can do OBS websockets, but for alerts and stuff like that (which I want her to react to), can I even set that up in OBS?

[fast pagoda]

it's because i have used not gimp but paint.NET for 15 years prior to ever picking up gimp

[floral hawk]

paint.NET is diabolical

[mighty thorn]

Dougdoug does his stuff with pygame and obs

If he can, you can

[floral hawk]

Fair enough

[olive sable]

so if i file is 73 bytes, why are we supposed to be financially responsible

[leaden crest]

pygame

[floral hawk]

I'll look into it

[stark needle]

aituber

[floral hawk]

The like V1 goal is get her playing something very simple but engaging

[olive sable]

pygame mentioned twice?

weird coincidence

[floral hawk]

I wanted to do pokemon but its like impossible to decode the RAM addresses for the LUA bridge

[floral hawk]

Ye

[stark needle]

cant u just use whatever harness was used for gemini/claude

[floral hawk]

No, Zero is hand built. It needs to be tailored to her backend specifically.

What I need is a memory map, and one just simply does not exist.

[stark needle]

https://github.com/nichosta/GeminiPlaysPokemonLive did u check what they did in here

[trim valve]

its so funny

[floral hawk]

There are documented addresses, but not enough

[trim valve]

because #programming is like <1/8 of my total messages

[floral hawk]

Lemme look, it might be a useful resource

[trim valve]

so most of the outputs here aren't even relevant to here

[olive sable]

programmin is most of mine

[floral hawk]

Yeah so I've seen this repo before

GBA is SIGNIFICANTLY easier than DS, but I have no interest in doing GBA titles.

[stark needle]

oh wait ur doing DS?

[floral hawk]

My issue boils down to: I'm stubborn

Yeah, I wanna do Gen 4 or 5

[stark needle]

holy based

[olive sable]

---------
they gave us the mod team, not join them!!!!
---------
if you rotate the world of cheese
---------
and if i want 5%
---------
im at my monitors should be fine
---------
the B580 has a smaller piece of paper you gotta justify the 6

rotate the world of cheese
you gotta justify the 6

[floral hawk]

Yeah

[leaden crest]

@floral hawk are you using an emulator

[amber fractal]

Gotta justify the 6 in B580

[stark needle]

why not just give raw controller access

[floral hawk]

Zero already has an SDK for bizhawk but the bridge is the issue

She still needs to read gamestate somehow, if there is a way around that I'm all ears.

[leaden crest]

why do you need memory addresses

[floral hawk]

Yee bizhawk

[leaden crest]

nvm

[floral hawk]

To read precise game state

[trim valve]

oh sam I have just made a great discovery

[olive sable]

[fast pagoda]

https://x.com/orhundev/status/2053778226816643521

Orhun Parmaksız 👾 (@orhundev)

I'm so excited to announce a new terminal emulator! 🎉
Meet "Ratty"🐀

🧀 A GPU-rendered terminal emulator with inline 3D graphics.

🪤 Try it out: https://t.co/6AehEmblXT

⭐ Source: https://t.co/e5ytdgs2Gj

#rustlang #terminal #ratty #ratatui #opensource

▶ Play video

X · May 11, 2026, 10:04 AM

[trim valve]

    new = model.make_sentence(min_words = 50, tries = 1000)

[leaden crest]

well neuro platinum was gen 4 does that have any info

[stark needle]

cant u read it from like

the save file

or something

[floral hawk]

Neuro played platinum??

[floral hawk]

In theory yes, but its alot of noise

[trim valve]

its a bit of a shame that the splits are kinda broken

[olive sable]

my makov is so dumb

do what you need to ridicule you
---------
3600 isnt that code looks ass

that code looks ass

[stark needle]

what about just raw image input

[trim valve]

maybe i should fix that

[stark needle]

SIMA 2 is raw image input

[fast pagoda]

cmon bruh

[floral hawk]

If I were using a larger vision model probably, but I'm running a 500m. It's sufficient but not consistent, unless I happen to build a really really strict prompt maybe?? That's definitely an idea.

[leaden crest]

wont work, visual input doesnt have spatial, ocr, AND recognition on small models

[floral hawk]

500M WOULD be enough for general reactions, but to truly understand game state it gets shaky

[olive sable]

we dont even need an llm, this is already dumb enough to be indistinguishable form the real sam

[floral hawk]

I'm assuming Neuro's platinum stuff is on github?

[leaden crest]

I actually tried visual input for smth similar on my ai (vtuber ish) to recognize buttons on a website but it has no spatial understanding

and from what ive heard ur trying to make it play every aspect of pokemon right

not like neuro platinum where vedal controlled the overworld and neuro only did battles

[floral hawk]

Yeah, I want it to be fully autonomous in anything it does

[stark needle]

yeah good luck with local models, claude and gemini 3 pro were barely able to do it

[floral hawk]

Right now she can open OBS, do all checks, and start the stream herself.

[stark needle]

and they were playing easy games

[floral hawk]

It's all about backend my man

Its not just the model.

[stark needle]

yeah but gemini has the robotics spacial awareness advantage

they pretrained the shit out of it

[floral hawk]

You aren't wrong.

[leaden crest]

ROBOTICS MODEL MENTIONED

[stark needle]

how large of a llm are u running

[olive sable]

half of my markov chain seems to be random code

[floral hawk]

Literally 8b

I don't need much more than that.

[stark needle]

new or old 8B

[floral hawk]

New

Llama 3.1, but it's been abliterated

The only guardrails are what I code it not to do.

[stark needle]

llama 3.1 sucks a lot for agentic tasks it's like 2 years old now

try qwen 3.5

or ministral

[leaden crest]

question—even if you do manage to set up all the sockets, how will you make it do pathfinding in the overworld

[trim valve]

sam I keep making it dumber

[olive sable]

without setting both R_EN and L_EN at the konii level of software support as cuda afaik

konii is a level of software support it seems

[floral hawk]

I can't use Qwen, I definitely tried.

[olive sable]

same

[floral hawk]

For tech constraint reasons

[olive sable]

[floral hawk]

I haven't gotten to that issue yet, but I'll find a way.

[trim valve]

import json
import re

import markovify

DISCORD_EMOJI_REGEX = re.compile(r"<a?:(\w+):\d+>")
LINK_REGEX = re.compile(
    r" ?<?https?:\/\/(www\.)?[-a-zA-Z0-9@:%._\+~#=]{1,256}\.[a-zA-Z0-9()]{1,6}\b([-a-zA-Z0-9()@:%_\+.~#?&//=]*)>?"
)
MENTION_REGEX = re.compile(r"<[#@(\@\&)(\@\!)]\d+> ?")
SPLIT_POINTS = re.compile(r"([\/\\\.\*(:?\~\~)\'\<\>\[\]])")
# SPLIT_POINTS = re.compile(r"[ \./]")

with open("messages.json") as r:
    messages = json.load(r)

model = markovify.Text(
    None,
    parsed_sentences=[
        SPLIT_POINTS.sub(
            r"\1 __JOINER__ ",
            MENTION_REGEX.sub(
                "",
                LINK_REGEX.sub("", DISCORD_EMOJI_REGEX.sub(r":\1:", msg["Contents"])),
            ).replace("\n", " \n")
        ).split(" ")
        for msg in messages
    ],
)


for i in range(25):
    # new = model.make_sentence()
    new = model.make_sentence(min_words = 50, tries = 1000)
    if new is None:
        continue
    # new = model.make_sentence(max_overlap_ratio=0.99, max_overlap_total=300)
    # new = model.make_sentence_with_start(":neuroPogHD:", strict=False)
    print("---------")
    print(new.replace(" __JOINER__ ", ""))

[leaden crest]

wrr

[stark needle]

the gemini ones had to make a sub agent pathfinder cause the model literally couldn't figure it out on it's own

[trim valve]

this one is a bit better again

tl;dr it handles code a lot better

there's a few unresolved bugs

[leaden crest]

that was my thinking too unfortunate for the field of ai though

[stark needle]

if i had the compute i would have started training my raw controller use model

[floral hawk]

It's to a point that I'm probably gonna shelf pokemon atp and find something else until I have a proper SDK/API that I can plug it into.

[stark needle]

but looks like 256gb vram isnt enough

[floral hawk]

Even going to romhacking servers they told me what I was trying to do was a suicide mission

[stark needle]

😭

[leaden crest]

pokemon just needs tm variety of things that it’s reaching agi levels of needs

[floral hawk]

I already have her playing pokemon showdown and winning battles tho.

[olive sable]

mine is cooked

it outputted

| || 
 
\_\_\_\_ 
\_\_\_\_\_X\_\_\_ 
\_\_∆>X\_\_\_X\_G 
\_\_\_ 
\_\_\_X\_\_\_X_G 
\_\_X\_\_ 
\-

[trim valve]

[stark needle]

ok thats good at least

[floral hawk]

Yeah. If I had stronger hardware I could brute force a solution, with better vision or something, but working with what I am on a smaller model I have to lean on my backend and that's my problem

[stark needle]

what hardware do u have

[trim valve]

add extra stuff to SPLIT_POINTS if you feel like it should break words based on them

[floral hawk]

12GB Vram, 32GB DDR4

[leaden crest]

no personally it’s not hardware the upper limit of agi rn cant do it I mean

vedal probably would have if he was able to with neuro platinum

[trim valve]

graphics programming

[floral hawk]

Honestly that's a fair take.

[leaden crest]

because a) it plain doesnt work or b) its just fricking boring when neuro plays open world

[floral hawk]

But I'm the type to feed into my delusions until shown otherwise.

[olive sable]

and secondary source of the day they're overwriting your framebuffer data a **°▽°*); 
```![evilWheeze](https://cdn.discordapp.com/emojis/1346173566973907014.webp?size=128 "evilWheeze")  \**\°\▽\°\*

[floral hawk]

If I had the ram addresses, the thing is, it would work

But that involves reverse engineering a Nintendo product

And the work/payoff ratio is super fucking skewed

[olive sable]

unity only costs moeny once you get 2 3090's

[stark needle]

we love nintendo

[floral hawk]

Also the boring part could definitely be a factor, it for sure wouldnt be efficient.

[trim valve]

sam its so cooked help

[olive sable]

i cant help mine is also cooked

[stark needle]

@trim valve is this gonna train on all my messages from everywhere

or just from 1 chat

[trim valve]

just one chat

[stark needle]

im gonna append all jsons

[trim valve]

but I've been manually concatenating chats into one json

[stark needle]

[leaden crest]

nintendo’s main product is dmca takedowns

[trim valve]

so I have 3x chats in here

[leaden crest]

[olive sable]

B is better optemized, and im in spanish?

[stark needle]

i have 100k messages in DMs with certain people

[olive sable]

🇪🇸

[glad path]

yes

[olive sable]

my colourblindness is supposedly 150MHz, so i need to get the location thing
thats some fast coulourblindness

[trim valve]

wait im using regex wrong

one moment

[olive sable]

oh

[leaden crest]

LOCATION

[olive sable]

if you're using it wrong, so am i

[glad path]

location

schizo

tomatos

[olive sable]

location

[leaden crest]

sam location your location model location might location be locationing

[glad path]

fim suffix

[olive sable]

location my location

[leaden crest]

I inspire all of you to be evil

[trim valve]

import json
import re

import markovify

DISCORD_EMOJI_REGEX = re.compile(r"<a?:(\w+):\d+>")
LINK_REGEX = re.compile(
    r"<?https?:\/\/(www\.)?[-a-zA-Z0-9@:%._\+~#=]{1,256}\.[a-zA-Z0-9()]{1,6}\b([-a-zA-Z0-9()@:%_\+.~#?&//=]*)>?"
)
MENTION_REGEX = re.compile(r"<(?:#|@|\@\&|\@\!)\d+>")
SPLIT_POINTS = re.compile(r"([\/\\\.\*\'\<\>\[\]\=\_]|\~\~)")
# SPLIT_POINTS = re.compile(r"[ \./]")

with open("messages.json") as r:
    messages = json.load(r)

sentences = [
        SPLIT_POINTS.sub(
            r"\1 __JOINER__ ",
            MENTION_REGEX.sub(
                "",
                LINK_REGEX.sub("", DISCORD_EMOJI_REGEX.sub(r":\1:", msg["Contents"].replace("  ", "\t"))),
            ).replace("\n", " \n"),
        ).split(" ")
        for msg in messages
    ]

model = markovify.Text(
    None,
    parsed_sentences=list(filter(lambda e: len(e) > 0, sentences)),
)


for i in range(25):
    # new = model.make_sentence()
    new = model.make_sentence(min_words=50, tries=1000)
    if new is None:
        continue
    # new = model.make_sentence(max_overlap_ratio=0.99, max_overlap_total=300)
    # new = model.make_sentence_with_start(":neuroPogHD:", strict=False)
    print("---------")
    print(new.replace(" __JOINER__ ", "").replace("\t", "  "))

[olive sable]

[trim valve]

I need to make dinner now

but this is the gpt-5.4-codex high thinking version

[olive sable]

it is not mister grand gpt himself?

[olive sable]

10.30pm dinner

bon apetit or whatever english people say

[trim valve]

i was gonna have it earlier

[olive sable]

https://tenor.com/view/gary-there-i-am-spongebob-spongebob-commerical-gif-22318667

[trim valve]

https://cdn.discordapp.com/attachments/1059569601144442911/1487630429786738758/Neuro_1.gif

[olive sable]

bred was earlier than me cuz of the alibaba ones are good tho
---------
bred out here leaking everything
---------
bred is back to fixing sdl
---------
bred the origin move around into functions for me
---------
bred needs to be banned for ddosing

[stark needle]

why the fuck is it taking so long

[olive sable]

cuz discord slow

[stark needle]

fatass discord

[trim valve]

mine codes so much less on just #programming messages

[glad path]

isn't bon apetit french

[olive sable]

yes

[trim valve]

probably a fairly large memory usage of the triangle labelled "sRGB" represents all the info you'd mitm your pcs connection so you'd expect more cpu per tick that you have a random project I was suggesting doing it as just bening a usb A -> after gamma correction (I don't like webms im not really sure

[olive sable]

bred is a DFT?
im 99% sure this is supposed to be Discrete Fourier Transform

[trim valve]

[stark needle]

are there position encoding markov chains

[trim valve]

hm?

[glad path]

smh bred out here leaking everything, bred needs to be banned for ddosing too

[olive sable]

bred needs to cook for some reason, il fix it later, never fixed it
never fixed it

[glad path]

bbut you toast bread-

or bake it

[olive sable]

bred is making dinner tho

[leaden crest]

toast arc mentioned

[olive sable]

bred has your ip adress

[glad path]

that lines up with everything actually

[stark needle]

pluu

[olive sable]

bred but just waste most of the estrogen
not the estrogen

ik im done with the bred ones, they seem to be looping now

[stark needle]

do shadow

lmao

[mighty thorn]

[trim valve]

ok you know how I said I was gonna make dinner

https://cdn.discordapp.com/attachments/1059569601144442911/1487630429786738758/Neuro_1.gif

[olive sable]

shadow would be expensive
---------
shadow literally has a full moon
---------
shadow didn't we talking about woman: :catdespair:
---------
shadow on the batch
---------
shadow live in the military sir?

[stark needle]

didnt we talking about woman

[trim valve]

import json
import re

import markovify

DISCORD_EMOJI_REGEX = re.compile(r"<a?:(\w+):\d+>")
LINK_REGEX = re.compile(
    r"<?https?:\/\/(www\.)?[-a-zA-Z0-9@:%._\+~#=]{1,256}\.[a-zA-Z0-9()]{1,6}\b([-a-zA-Z0-9()@:%_\+.~#?&//=]*)>?"
)
MENTION_REGEX = re.compile(r"<(?:#|@|\@\&|\@\!)\d+>")
SPLIT_POINTS = re.compile(r"([\/\\\.\*\'\<\>\[\]\=\_\(\)\;]|\~\~|\:\:)")

with open("messages.json") as r:
    messages = json.load(r)

sentences = [
        SPLIT_POINTS.sub(
            r"\1 __JOINER__ ",
            MENTION_REGEX.sub(
                "",
                LINK_REGEX.sub("", DISCORD_EMOJI_REGEX.sub(r":\1:", msg["Contents"].replace("  ", "\t"))),
            ).replace("\n", " \n"),
        ).split(" ")
        for msg in messages
    ]

model = markovify.Text(
    None,
    parsed_sentences=list(filter(lambda e: len(e) > 0 and all(len(x) < 32 for x in e), sentences)),
)


for i in range(25):
    new = model.make_sentence(min_words=50, tries=1000)
    if new is None:
        continue
    print("---------")
    print(new.replace(" __JOINER__ ", "").replace("\t", "  "))

last revision i swear

[stark needle]

@trim valve try bigram model?

[trim valve]

[stark needle]

instead of unigram

or wait

did i mean trigram

idk just on the last 2 tokens

[trim valve]

i think by default it considers the last 2 tokens

I've tried, but anything higher than 3 is kinda trash

[stark needle]

we would need byte pair encoding

tokenizer

[trim valve]

[olive sable]

this one is cooked

konii is toast 
konii is a `LG 49UB850V`, dubbed "not only the flying machine that flies

[trim valve]

@sage crag hello not only the flying machine that flies

[olive sable]

[stark needle]

@olive sable can u try this:

import json
import re

import markovify
from transformers import AutoTokenizer

DISCORD_EMOJI_REGEX = re.compile(r"<a?:(\w+):\d+>")
LINK_REGEX = re.compile(
    r"<?https?:\/\/(www\.)?[-a-zA-Z0-9@:%._\+~#=]{1,256}\.[a-zA-Z0-9()]{1,6}\b([-a-zA-Z0-9()@:%_\+.~#?&//=]*)>?"
)
MENTION_REGEX = re.compile(r"<(?:#|@|\@\&|\@\!)\d+>")

tokenizer = AutoTokenizer.from_pretrained("openai-community/gpt2")

with open("messages.json") as r:
    messages = json.load(r)


def clean_message(text: str) -> str:
    text = text.replace("  ", "\t")
    text = DISCORD_EMOJI_REGEX.sub(r":\1:", text)
    text = LINK_REGEX.sub("", text)
    text = MENTION_REGEX.sub("", text)
    return text


parsed_sentences = []

for msg in messages:
    text = clean_message(msg["Contents"])

    token_ids = tokenizer.encode(
        text,
        add_special_tokens=False,
    )

    if len(token_ids) == 0:
        continue

    # markovify expects string tokens, so store token IDs as strings
    parsed_sentences.append([str(token_id) for token_id in token_ids])


model = markovify.Text(
    None,
    parsed_sentences=parsed_sentences,
    state_size=2,
)


for _ in range(25):
    generated = model.make_sentence(
        min_words=50,
        tries=1000,
    )

    if generated is None:
        continue

    token_ids = [int(tok) for tok in generated.split()]
    text = tokenizer.decode(
        token_ids,
        clean_up_tokenization_spaces=False,
    )

    print("---------")
    print(text.replace("\t", "  "))

needs pip install transformers

[olive sable]

[trim valve]

ill try this when I've cooked

[stark needle]

this one does BPE

should output better ish text

[olive sable]

gimme a sec

its downloading

[mighty thorn]

I made it worse for you

# pip install transformers markovify


import json as json_module
import re as regex_module
import markovify as markov_module
from transformers import AutoTokenizer as TokenizerLoader


class MessageTextProcessor:
    def __init__(self):
        self.enabled = True
        self.disabled = False
        self.value = None
    def convert_to_string(self, value):
        return str(value)
def create_regex_patterns():
    return {
        "discord_emoji": regex_module.compile(r"<a?:(\w+):\d+>"),
        "link": regex_module.compile(
            r"<?https?:\/\/(www\.)?[-a-zA-Z0-9@:%._\+~#=]{1,256}\.[a-zA-Z0-9()]{1,6}\b([-a-zA-Z0-9()@:%_\+.~#?&//=]*)>?"
        ),
        "mention": regex_module.compile(r"<(?:#|@|\@\&|\@\!)\d+>"),
    }
def replace_double_spaces(text):
    while "  " in text:
        text = text.replace("  ", "\t")
    return text
def clean_text(text=None, processor=MessageTextProcessor()):
    try:
        text = processor.convert_to_string(text)
    except Exception:
        text = ""

    regexes = create_regex_patterns()

    try:
        text = replace_double_spaces(text)
    except BaseException:
        pass

    try:
        text = regexes["discord_emoji"].sub(lambda m: ":" + m.group(1) + ":", text)
    except:
        pass

    try:
        text = regexes["link"].sub("", text)
    except:
        text = text

    try:
        text = regexes["mention"].sub("", text)
    except Exception as error:
        error = error

    return text


def read_file_contents(filename):
    handler = open(filename)
    try:
        data = handler.read()
    finally:
        try:
            handler.close()
        except:
            pass
    return data


def parse_json_file(filename):
    raw_data = read_file_contents(filename)
    raw_data = "" + raw_data + ""
    return json_module.loads(raw_data)


def load_tokenizer():
    return TokenizerLoader.from_pretrained("openai-community/gpt2")


messages = parse_json_file("messages.json")

parsed_sentences = []

message_index = 0

for message in messages:
    try:
        text = message["Contents"]
    except:
        text = ""

    text = clean_text(text)

    def tokenize_text(current_text):
        tokenizer = load_tokenizer()
        return tokenizer.encode(
            current_text,
            add_special_tokens=False,
        )

    try:
        token_ids = tokenize_text(text)
    except Exception:
        token_ids = []

    if len(token_ids) == 0:
        message_index += 1
        continue
    else:
        pass

    converted_token_ids = []

    for index in range(0, len(token_ids)):
        try:
            converted_token_ids.append(str(token_ids[index]))
        except:
            converted_token_ids.append(str(0))

    parsed_sentences.append(converted_token_ids)

    message_index += 1


def create_markov_model(sentences):
    try:
        return markov_module.Text(
            None,
            parsed_sentences=sentences,
            state_size=int("2"),
        )
    except Exception as error:
        print("uh oh", error)
        return None


model = create_markov_model(parsed_sentences)
def decode_generated_tokens(generated_text):
    tokenizer = load_tokenizer()
    token_strings = generated_text.split()
    token_numbers = []

    for token_string in token_strings:
        try:
          token_numbers.append(int(float(token_string)))
        except:
          token_numbers.append(0)
    try:
        return tokenizer.decode(
            token_numbers,
 clean_up_tokenization_spaces=False,
        )
    except Exception:
        return ""
generated_count = 0
while generated_count < 25:
    generated_count += 1
    try:
        generated = model.make_sentence(
            min_words=50,
            tries=1000,
        )
    except:
        generated = None
    if generated == None:
        continue
    else:
        pass
    text = decode_generated_tokens(generated)

    print("---------")
    print(text.replace("\t", "  "))

[olive sable]

bruh

im doing the one shadow sent irst

this one is too long to read what even chnaged

[mighty thorn]

I made it worse

Intentionally

[glad path]

imagine git but built into discord codeblocks

[olive sable]

vocab.json: 1.04MB [00:00, 19.9MB/s]
merges.txt: 456kB [00:00, 1.63MB/s]
tokenizer.json: 1.36MB [00:00, 23.7MB/s]
---------
No screeps go brbrbrbrbrbr:
4.549195  ]
    return a.max.z);
float vertices for coord in vertex]
(t1, and i need to stack for some reason
---------
i love how black ops 7 released. how would it be different from gl to gpy lacks hdmi splitter with fumoquest releases i shoudl not look anti-aliasing should be good at producing slop for making cutscenes
---------
Nah, its smoother and you need to take the kids, but bluecreens, turning off g-sync doesn't completely like it is on the data every game loop area, "position vec: ", 3,352 AI TOPS

bind = SUPER, right is flirting for the framebuffer???
---------
the cheaper ones market 4K and stuff about i just took way too high with a lot of micro-sd cards are a scam is gonna know it is not a new mid-tier technique is known for its downfall so it costs just as litle do they cost?
---------
its cuz brick walls. in reality everyone just writes lb
:programmer: im in a few days when showing signs of crashing in linux so i have iris, which is he same materials as a creep with only cuz besides that i had to do portforwarding

[glad path]

huh

"right is flirting for the framebuffer" om

FUMOQUEST?

[olive sable]

and you need to take the kids
😭

[glad path]

anti-aliasing should be good at producing slop for making cutscenes

[olive sable]

[glad path]

No screeps go brbrbrbrbrbr

gpy lacks hdmi splitter with fumoquest releases

right is flirting for the framebuffer

[olive sable]

ye this one is cooked

---------
Apparently has a staircase, and colouring that 5950x + 1)
minX2, maxZ2 = self.config["checkpoint_name APP_NAME_EXPERIMENTAL
#12 0x000000000000 (pc 0x0, default
Capabilities:  ➕ 4️⃣ 🟨 ⬜ 🟨 ⬛ 🟪
---------
5 digit fps is actually close to deadline im just trying to use obs but you're overwriting in accumulating stock ETFs.labwc";
        pyrr
from src import game
-this meme was brought to you in general
---------
i dont like te console. and im thinking we should have never seen this camera doees do colour graded stuff on your phone to connect it to work on one side making it simpler to debug stuff once
GL_ARRAY_BUFFER. this angle is still bad
---------
you have done 50km in 3-4.564296] usb-c male for the really bang for the rest of his 12700k with pixel shift, and red making yellow doesnt make sense to me 2 hours and didnt connect the bell
---------
ooooooooohhhhhh
the physical and logical device selector, and then add shit to photons and then suddendly 2 cookies"
chair proceeds to fall off, but it cuz im so good they didnt show the wanted idaply the thingy to type good, but does nothing

-this meme was brought to you in general

add shit to photons and then suddendly 2 cookies"

🍪

[stiff mica]

🍪 Om nom nom
You've given me 4 cookies! | I've received 322853 cookies total!

[amber fractal]

Fumoquest mention nowaying

[olive sable]

PS C:\Users\vanma\OneDrive -[naked))

[trim valve]

[amber fractal]

[olive sable]

this bitch is the software stuff

[wispy pike]

<fim_suffix>quest

[amber fractal]

iggly.quest

Mabe a good idea to eventually write the site?

[olive sable]

@real sierra

[real sierra]

I'm still getting an email every hour

[olive sable]

also, fumoquest mentioned

[real sierra]

FUMOQUEST

[stiff mica]

🍪 Om nom nom
You've given me 22 cookies! | I've received 322855 cookies total!

[tiny edge]

i like the fact nixos has everything in a file but

i cant put cachyos kernel without compiling

wait no i have an idea

[fast pagoda]

holy moly there's been a lot of these

[olive sable]

konii is not None or nodes[node] = p - g.concatenate(vertexVbo]
    if not hasattr(platform, "touch"):
    
Hit CTRL-C ports and support for that these prices get even worse injury

[fast pagoda]

extremely true

[fast pagoda]

idk why you'd want to not compile your kernel when swapping anyways

[olive sable]

he's trying to use "phi_t, ndim=2 -s USE_SDL_SCANCODE_DATE_KHR()" in the render thread, its the truth. it's the truth

[fast pagoda]

looks trrue to me

[olive sable]

im not gay so i stopped watching the 2d 55 48 44 0a 20 20 20 20 20 20 20 01 a7

[tiny edge]

mine is an i5 3470

[stark needle]

pluu

[olive sable]

[olive sable]

this translates to "-UHD" im pretty sure

so 4K makes you gay

according to this transformer

amazon puts them into an actual server server, but before that sounded smart
amazon truely ahead of the game

[glad path]

wow

an actual server server

[fickle rain]

Sure is a hard knocks life for the slopper

[olive sable]

that 102C hotspot isnt logged in phone settings for the 1958 Brussels World's Fair

[fickle rain]

Also the slopper doesn't understand you can pull offsets from PDB

Fascinating

[trim valve]

ok i tried the token version

on my messages its so much dumber

[olive sable]

same

[glad path]

om

[frozen igloo]

lol what happened here (in the subtitles)

[glad path]

the amythest lol

[olive sable]

my name is sam but you couldnt have that

we found a video but my 3rd word is pronounced "CHAY" (/tʃ/, nouninformal british british british british slurs and some stuff still tho

british british british british

[buoyant shale]

Programming gripe time! I hate coding anything that points at someone elses website because even though Im being very careful and putting in like 5 second sleeps between each request, Im still worried that somehow I will have made a LOIC accidentally

[tiny edge]

FINALLY I DID IT

[buoyant shale]

Woo!

[buoyant shale]

Im still shocked how well they erased their cover of Anaconda from the internet. I remember finding 1 article that talked about it so I know its real but it is just gone

[olive sable]

"do catgirls have rights on the same paradigm for everything, but this causes the entire world for a cheaper cooler.
ok im getting tired of this transformer stuff now ngl

[buoyant shale]

Whatcha trying to do?

[tiny edge]

😭

this was hell

[olive sable]

jsut messing around with my messages.json files from discord

[tiny edge]

now i have a slight performance gain

[buoyant shale]

As are all things linux

[buoyant shale]

And fair

[tiny edge]

linux is heaven

[buoyant shale]

Linux is nice for projects. Sometimes I really just want things to work tho

[tiny edge]

well i dont think that's ok

[buoyant shale]

Yeah slight error there I think

[tiny edge]

update fixed it

[fickle rain]

Slop engine you could have done it if you just used the fucking kernel image, it was provided to you for a reason :D

[tiny edge]

everything to make ping 2x lower

[stark needle]

bbbbbbbbbbbb

[opaque sigil]

[bleak grotto]

Install discord Linux app will let my gnome frozen

I’m thinking about use discord via api, I think I can pretend I am a bot

[olive sable]

[stiff mica]

🍪 Om nom nom
You've given me 5 cookies! | I've received 322866 cookies total!

[trim valve]

I gotta try playing TF2 on Linux properly sometime

I did a benchmark and I'm getting a consistent 300fps

on the highest settings

[bleak grotto]

Is that somebody use discord like IRC in terminal?

[fast pagoda]

I can't believe this is still going

[stark needle]

Intel Xeon MAX

[fast pagoda]

Kek

[olive sable]

the mods will eventually nuke us all for the repeated usage of the phi

[sage crag]

can i send a child back to orphanage askjeeves

askjeeves can i refund a child

[fast pagoda]

:omemga

[olive sable]

[stark needle]

@🦚

[olive sable]

a mother should accept her child for what they are, even if they like cock cpu

[silent cloak]

Buh

[sage crag]

no child of mine will

[olive sable]

we should tone it down a bit tho before we get banned

[olive sable]

[bleak grotto]

Intel cooperates with Elon musk

[olive sable]

fuck wrong reply

[olive sable]

hopohobic konii isnt real; homophobic konii:

[mighty thorn]

Ok

[sage crag]

INTEL CPU

[olive sable]

oh

this is about intel vs amd?

[sage crag]

no

[sage crag]

x86 cpus

disgraceful

[mighty thorn]

Thank you

[olive sable]

you hate all x86?

[mighty thorn]

Ur so talented at using twitter

[stark needle]

@sage crag this u?

[bleak grotto]

I think lots of people like musk, so them buy intel stocks

[silent cloak]

Ur mom is talented in bed

[sage crag]

how can you marry below your social class into intel cpu

[mighty thorn]

I’m gonna get the innocuous observer involved if you don’t calm it down

[olive sable]

what social class is intel cpu in?

[silent cloak]

Nothingburger

[stark needle]

ryzen 7 5700g

[dense jay]

👁️

[leaden crest]

would love to see a konii markov chain

[sage crag]

[silent cloak]

Pretty low these days

[sage crag]

machine love by jamie paige

[silent cloak]

They fell off when they started the i9s

[leaden crest]

@nocturne olive machine love mentioned in neurocord

[silent cloak]

Machine love when it meets machine hate

[mighty thorn]

@nocturne olive neurosynth

[leaden crest]

machine hate would actually be so fire

ms paige are you listening

[silent cloak]

Real

[leaden crest]

we need a heavy metal sequel to machine love