programming

[sage crag]

ye bot

[real sierra]

ive been considering making a script that gets all strings from my code, calculates token savings if macro'd for all of them, and then repeats until no macros would save tokens anymore

suboptimal because greedy algorithm but automated at least

[sage crag]

oh i already did this algorithm

for my

thing

vm

yeah

[real sierra]

im starting to think i need it too

my current compression ratio is pretty good, i save about 500 tokens with my macro set

but im running out of ideas to shrink further

[fast pagoda]

it would be like
uint64_t path[MAX_PLY}for the path history
then a for loop for (int i = 0; i < depth; ++i) if (path[i] == board->hash) return 0;
then right before going over the moves path[depth{] = board->hash;

how the hell this goes in 3 tokens idk

[sage crag]

not sure if you can salvage any valuable code from in here

i couldnt be asked to make it reusable

[real sierra]

found another token

was using Board * instead of PBoard

[fast pagoda]

it's all coming together

[fast pagoda]

welp time to change the random bot so i dont continue using Board * like a lemming since that's what's in there RIGHT NOW

[real sierra]

// token savers...
#define eq_0_sc = 0;
#define bclose ) {
#define if_ if (
#define while_ while (
#define br_sc );
#define for_i_0_lim(lim) for (int i eq_0_sc i < lim; i++ bclose
#define tree_get tree->
#define tree_child_i tree_get children[i]
#define tree_children_len tree_get children_len
#define tree_moves tree_get moves
#define tree_moves_i tree_moves[i]
#define return_rv_sc return rv;
#define rv_assign rv =
#define args_after_board(args...) (board, args br_sc
#define no_args_after_board (board br_sc
#define popdiff(P) (0xfu & 0xf953310u >> 4*P) * (__builtin_popcountll(chess_get_bitboard(board, WHITE, P)) - __builtin_popcountll(chess_get_bitboard(board, BLACK, P)))
#define eval_score popdiff(QUEEN) + popdiff(ROOK) + popdiff(BISHOP) + popdiff(KNIGHT) + popdiff(PAWN)
#define PMoveTree_tree_bclose PMoveTree tree bclose
#define best_score_gt_score best_score >= score ?
#define rv_eq_best_score_gt_score_rv_or_i_sc rv_assign best_score_gt_score rv : i;
#define best_score_eq_best_score_gt_score_best_score_or_score_sc best_score = best_score_gt_score best_score : score;
#define for_i_0_lim_tree_chld_len for_i_0_lim(tree_children_len)
#define tree_parent tree_get parent
#define cbclose_return_rv_sc_cbclose_void } return_rv_sc } void
#define moves_i moves[i]
#define undo_moves while_ move_count bclose move_count--; chess_undo_move no_args_after_board
#define trans_table_tree_hash transposition_table[tree->hash & 0x0000000000FFFFFFUL].
#define trans_table_tree_child_i_hash transposition_table[tree->children[i]->hash & 0x0000000000FFFFFFUL].

#define PMoveTree struct MoveTree *
#define PBoard Board *

these are all my macros

which is the worst

[fast pagoda]

beautiful

[sage crag]

trans_table_tree_child_i_hash

[glad path]

odd my fan helper stopped working for some reason and let my cpu/gpu up to near 100c

[sage crag]

ttt c

i have a table

[fast pagoda]

#define cbclose_return_rv_sc_cbclose_void } return rv; } void

[sage crag]

i have a tree

[real sierra]

[fast pagoda]

this one is incomprehensible without already knowing wtf this does lmao

[glad path]

i had both fans at 100 and they just got lazy and both went to 50

i dont know why

[real sierra]

wdym is written right there in the name

curly brace close return rv semicolon curly brace close void

[fast pagoda]

[real sierra]

it saves many tokens

[glad path]

me adjusting the position of my laptop for the 10243th time for more airflow (totally necessary and stuff)

[real sierra]

i had to reorder my functions to use it more

[fast pagoda]

#define FourWheelsEngineAndSteeringMechanism car

[real sierra]

java dev

[sage crag]

java dev

reprogramming

[glad path]

what

[tender river]

#define popdiff1 ) + popdiff(
#define eval_score popdiff(QUEEN popdiff1 ROOK popdiff1 BISHOP popdiff1 KNIGHT popdiff1 PAWN)

are you even trying

[glad path]

are yall a hive mind

[fast pagoda]

yes

[glad path]

oh wait

[fast pagoda]

swarm

[glad path]

the case

the

the uh

the navim

naming

[sage crag]

no im a tv

[fast pagoda]

https://tenor.com/view/run-away-jurassic-world-dominion-scared-swarm-jurassic-park-gif-24977740

[glad path]

you name classes like that in java

[real sierra]

wait that works?

i thought you couldnt construct the macro call syntax itself from macros

[tender river]

why wouldnt it

[real sierra]

i figured the preprocessor wouldnt recognise

[glad path]

guys how can I get my laptop to unnecessarily draw more power for no reason

im bored

[tender river]

note that it resolves the macros at expansion time not at definition time

use this info wisely

[glad path]

i have 243Wh of energy draw

[real sierra]

ty chayleaf sensei

adding more macros

[fast pagoda]

not the furry rp training i didnt even notice this before

[tender river]

#define y z += x + x * 2 + x * 4 + x * 8;
int x = 5, z = 0;
y
#define x 6
y

[real sierra]

but wait

[keen hatch]

c=1;b=[[0]*8 for _ in range(8)];b[0]=[2,3,4,5,6,4,3,2];b[7]=[-2,-3,-4,-5,-6,-4,-3,-2];b[1]=b[6]=[1]*8
while True:
    p=input().split()
    if len(p)==2:
        x,y=map(int,p);b[x][y]*=-1;b[x][y]=-b[x][y];print(''.join(str(b[r][c])for r in b))

[real sierra]

wont it just call popdiff once

using everything in the () as the arg

[glad path]

putting the pythonic in python since 2008

[fast pagoda]

no it chains

[tender river]

reading

[sage crag]

shiro pls use godbolt

[fast pagoda]

it closes the parenthesis of the previous popdiff, adds the plus sign, and starts the next popdiff

[real sierra]

no i know that

but in the definition of eval_score

[tender river]

shiro did you forget macro is just copypaste

[real sierra]

you still have a popdiff(...)

[sage crag]

shiro

u can try it

using godbolt

[real sierra]

what godbolt

[sage crag]

https://godbolt.org/

[real sierra]

why does everyone here name things ive never heard of

[tender river]

everyone knows of godbolt

[real sierra]

stop thinking your esoteric knowledge is commonplace

[sage crag]

godbolt compiler explorer

ye look at asm

[tender river]

shiro git is not esoteric knowledge its a basic life skill

[sage crag]

look at compiler output

[real sierra]

i know how cli git i just no like

[sage crag]

git education human right

[real sierra]

[fast pagoda]

i thought i would always hate git cli

[glad path]

i cant find the temperature unit setting wtf

[fast pagoda]

and now i feel weird not git init immediately after mkdir + cd into it

so then i will like sort my music folder and mindlessly git init an absolutely bonkers sized repo for no reason

[tender river]

[fast pagoda]

lmfaoooo i went and checked it i had done that last time i did

yep

[glad path]

what the fuck why is one of the unidentified voltage sensors on my laptop measuring 16777216V

this is very concerning

[fast pagoda]

that's the nuclear reactor

[amber fractal]

for clock signal

[glad path]

its not even one of the unidentified sensors???

its an "unknown" sensor but it does have a proper name

[fast pagoda]

the fact that that searches up a post is interesting

coincidence looks like

but i was like damn it's real

also that it's mac specific

[sage crag]

not thast coincidental

1 << 24

[fast pagoda]

oh it's 2^24 duh

[sage crag]

recognise that power of 2

[fast pagoda]

jesus

[glad path]

probably a driver error then

thats kind of funny though

[real sierra]

the moment you include popdiff it gets upset

[fast pagoda]

whatever that sensor is reporting it's reporting the maximum 24bit int possible lol

[real sierra]

gaslighting me when wrong

[sage crag]

noe

use godbolt

me not know if correct

[real sierra]

am

[sage crag]

you use godbolt

[fast pagoda]

that's why u gotta post the full source so we can context our roasting

[sage crag]

[real sierra]

not compile in godbolt

[glad path]

also interesting that my fan helper just set both fans to auto again (not going to complain because they didnt really need to be on full)

[sage crag]

gotta set compile flag

-E for expand preprocessor

not care about code

[real sierra]

i know that

i wasnt born yesterday

[glad path]

LOL I asked chatgpt about the bug and it said it's safe to ignore unless ...

[sage crag]

[real sierra]

add #define popdiff(P) (0xfu & 0xf953310u >> 4*P) * (__builtin_popcountll(chess_get_bitboard(board, WHITE, P)) - __builtin_popcountll(chess_get_bitboard(board, BLACK, P)))

macro no longer expand

[fast pagoda]

claude shannon and thomas bayes are looking down on us with enormous, uncontrollable contempt

[keen hatch]

This model is trained on 20M tokens... its own dataset

[glad path]

there is no possibility a laptop draws 16777216V power when it only even gets 20V from the charger

what is chatgpt on about

[real sierra]

as i was trying to say before

[fast pagoda]

well, it's doing the best it can aka shit

[real sierra]

the popdiff at the front tries to expand first

treating inside as arg list

and then fail

[fast pagoda]

probably not a whole lot of 16777216V power instances popping up in the training corpus that dont involve horrible death

[glad path]

I feel like would not be able to type here if there was actually that much power draw

something bad probably would have happened anyways

or my laptop would be dead

or both

[maiden geyser]

random researches people do are crazy

[fast pagoda]

https://tenor.com/view/10000-volt-ghost-dee-bradley-baker-scooby-doo-2-monsters-unleashed-2004-james-gunn-gif-7083572526561720487

maybe ur the 10,000 volt ghost

[maiden geyser]

like vtuber past lives, like how did you find this, is this intentional or is algorithm cooking

[fast pagoda]

12 hours of uncontrolled procrastination

[glad path]

have you heard of floccinaucinihilipilifications

[real sierra]

yes

memorized it for word bomb

[glad path]

16777216V ghost is the next big hit guys

[glad path]

do you know the definition too

[real sierra]

nope only spelling was required

[glad path]

sad

i memorized the spelling and I do know the meaning
not for any reason I just did it because why not

[real sierra]

"wait, you HAVEN'T heard of (rare and specific knowledge)"

[glad path]

did you know it is legal to kick the ball in most forms of volleyball

[fast pagoda]

in highschool we used to play vollyball through only kicking (football((soccerr)) team)

[glad path]

do you know random useful ansi escape codes off the top of your head

( i used to )

uhh lemme try

[keen hatch]

She created a practically ready-to-use template for a Discord bot ))

[stiff mica]

aww, thank you~ ( ◡‿◡ *)

[glad path]

nope i dont. remember ful lsyntax

I used to know too much about ansi bc I made a library that heavily relied on it

[stiff mica]

🍪 Om nom nom
You've given me 34 cookies! | I've received 246032 cookies total!

[sage crag]

!nothing

[stiff mica]

Nothing ever happens, all-in

[sage crag]

!felloff

[stiff mica]

[sage crag]

!discord

[stiff mica]

discord.gg/neurosama

[sage crag]

[stiff mica]

🍪 Om nom nom
You've given me 47 cookies! | I've received 246033 cookies total!

[amber fractal]

!bwaa

[stiff mica]

🍪 Om nom nom
You've given me 1 cookie! | I've received 246034 cookies total!

[glad path]

idk why but swarm fm works really well for programming for me

why do i hear miku in a neuro cover

erm

[amber fractal]

Someone installed a mind virus and is uploading your tastes

[glad path]

oh wait

wait

wait

nvm that wasnt a neuro cover

im tired dont judge

but yeah swarm fm does work really well for programming

legendary yt homepage pull guys

(context)

actually decent ai summary I think

[fast pagoda]

search function macro like a boss

define all ur stuff like a boss such as M() for moving, E() for eval, G() for generating ur legal moves, S() for search (prototype)

#define O_o for(i=0;i<c;++i){M(b,m[i]);s=-S(b,d-1,-e,-a);U(b,m[i]);if(s>a){a=s;if(a>=e)goto l;}}l:

surely this will save le tokenz and you will be able to toss O_o in your search which is funny

[sage crag]

@real sierra its time for you to replace all functions

[fast pagoda]

SURELY

but be sure to do it like a boss or it wont work

[sage crag]

int func(int arg0, int arg1) {
  return arg0 + arg1;
}
// vs
#define func(arg0, arg1) ({ arg0 + arg1; })

these can be used in the same way

({ a; }) is a gcc and clang thing

where the whole thing evaluates to a

int x = ({
  int y = 0;
  while (y++ < 255) y *= 2;
  y;
})

i.e.

[sage crag]

it save token

[fast pagoda]

i strongly dislike how much whatever the standard c formatter is spreads stuff out
it makes it look much less crazy

in vscode

and this is how i learn that i have literally never seen or considered what it is because it's only doing it because i have turned on format on save for other stuff

[scarlet ore]

hey are y’all good at math

i need help with this list and polybius square, or at very least the correct of pairing the list numbers

i’m not programmer or mathematician so i’m a bit stupid

[fast pagoda]

people who are good at math are terrible programmers

i think that's the rule

[scarlet ore]

what about cybersecurity or ciphers then?

dude i’m gonna be honest i’m stumped on this puzzle

[fast pagoda]

twas tongue in cheek for the most part

[scarlet ore]

and the only other server i can go to is a programming minecraft mod server for help 😭

[glad path]

have you heard of desmos art

[fast pagoda]

just makes fun of what happens when you let like a math majorwrite a thesis on something that they show via code (code ends up being epic)

[glad path]

you literally freaking program with math

[fast pagoda]

it's a meme ya dip

[scarlet ore]

ok so here’s a sample of the list: 3, 4( ,)5 ,4( ,)1 ,1( ,)

how the hell do i correctly go pairing up these numbers, what’s denoting a new “pair”?

[glad path]

i made fully functional tic tac toe with win/draw/loss detection, proper turn based placing, board clearing, and extremely nice aesthetics for a desmos project

[fast pagoda]

number( ,)number is the pair

maybe

idk

what's the full list

[scarlet ore]

so it wouldn’t be reverse of anything, like “x, y” becomes “x ,y”

i only have a snippet and it’s a image, soo uhhh gimme a moment to try to use smth to nab it all

[fast pagoda]

oh i forgor you said it's a polybius cipher

lmao

just staring at it drooling like idk what the fuck this is brah

[scarlet ore]

okay so like, not a full thing because my phone is being a bitch and not properly copying

,3( ,)2 ,4( ,)1 ,1( ,)3 ,2( ,)3 ,1( ,)2 ,2( ,) 3 ,3( ,)4 ,2( ,) ,4(,)5 ,4(,)1 ,1( ,)2 ,2(,) ,1( ,)2 ,4( ,)4 3 ( ,)3 2( . ,)1 ,1( ,)2 ,3( ,)1 ,1( ,)3

the entire list is structured like this

[fast pagoda]

first thing i would do is just yeet the ( ,) and assume it's probably there to fuck you up if you just type polybius cipher decode into google

[scarlet ore]

okay wait new info dropped

someone seemingly did it… reversed? and got this

PLAY IIS GET A KILL ONAMA HORAGAUSING CHARLES ULT LOCUST GUY ULTYUKIULT AND DOMAIN CLIP IT AND SEND IT IN RECEPTION AND PING ME

i think it’s a different part of the cipher tho

gah 😞 why are things hard

[keen hatch]

That's true, I'm an "excellent" software architect, but I'm bad at math, and that didn't stop me from designing and then implementing some really cool stuff ) ... software development is about the ability to "assemble" the picture in your head as a whole )

[fast pagoda]

to figure out the cipher like even a little bit the entire thing is definitely needed

[scarlet ore]

i don’t have access to it unfortunately :(

well, regardless, apologies for wasting your time, i’ll try other methods just in case

[fast pagoda]

i don thtink you'll be able to get very far without the whole thing yourself

are there are clues given/ a keyword// even what is the puzzle called

[hollow spruce]

I hate
Kids

[scarlet ore]

haha yeah course not i’m just helping someone out with solving it

trying to give them ideas and such, i’m an idiot so i really can’t be any tangible help

[amber fractal]

I am a bad programmer, and I am also bad at math. My stats are all dumped into a hidden third option that can't easily be converted to either.

[fast pagoda]

just reminds me of this https://www.youtube.com/watch?v=yc_sn2GQ_MY

[scarlet ore]

someone dropped a image within a discord server i’m in and other person is trying to solve it, it was a image with a caesar cipher i think that told them to into and search the data. they looked into the hex code and found this nonsense

[keen hatch]

It means you're good at something else and that's cool ⭐

[sage crag]

tuning stone

[fast pagoda]

john cena jump scare at end o this btw

[fast pagoda]

i would tell the feller that if he wants any chance of anyone helping he'll need to give the full context aka like upload the untouched image

[amber fractal]

I just got to it as soon as you said it

[fast pagoda]

cipher puzzles especially one where they bothered to do that are likely gonna require the full picture to even attempt because the entire cipher is needed to know what it encodes

[amber fractal]

Of course, it is the construct responsible for only allowing me to write jank code.

[glad path]

logic? idk

[fast pagoda]

for a polybius cipher it needs literally every single digit because even being one(1) digit off will completely change the pairings which will completely and totally scramble any possible result from trying solutions aka progress is impossible because you will never find it without the whole thing unless you just randomly guess

and hit the jackpot w/ rngesus

[fast pagoda]

i know what his 3rd stat is

[scarlet ore]

oh fr?

[fast pagoda]

yes

[scarlet ore]

hmm alright bet

[amber fractal]

[keen hatch]

Let's say that even jank code can be useful, for example, during prototyping )))

[fast pagoda]

example inc

[amber fractal]

Not wrong

[glad path]

huh?

[fast pagoda]

method 2 from here https://en.wikipedia.org/wiki/Polybius_square

    1   2   3   4   5
1   A   B   C   D   E
2   F   G   H   I/J K
3   L   M   N   O   P
4   Q   R   S   T   U
5   V   W   X   Y   Z

combining i/j basically
P -> 35, o -> 34, w -> 52, e -> 15, h -> 23, i -> 24

the cipher becomes: 353452152324

let's say you delete a 5 from w

35342152324

35 -> P, 34 -> o, 21 -> f, 52 -> w, 32 -> m, 4 (?)

we decoded POFWM? even tho it's the right decoding

whoops

that's not method 3 it's method 2

cant read

[amber fractal]

It isn't shown in this place, but I do take a lot of time to propose ideas and tackle the ones I think are different enough from the norm that it'd be useful to have around.

so I think example inc works

[fast pagoda]

yes you do and i often am about to respond to your musings and fail because im too stupid to give a meaningful reply that isn't irrelevant

i do that here too but it's not ruining just one person's life

well, maybe shiro's with my terrible meme suggestions for him

[fast pagoda]

i meant i was gonna give an example to to @scarlet ore which was #programming message

your 3rd stat is PII related so i cant/wont reveal it lmaooo

[keen hatch]

But let's be honest, 80% of our code is jank... otherwise we wouldn't need refactoring, etc... and hacks in the code live forever...

[fast pagoda]

i definitely need a refactor

me, personally

[amber fractal]

Yeah I'm definitely working on one of those.

[amber fractal]

My github is clean, so if it is ideas beyond insanity that is a valid resource

Plus I post that here often

[fast pagoda]

wasn't always

[keen hatch]

and as practice shows, ideal code does nothing useful)

[amber fractal]

A mistake I'm not making again

[fast pagoda]

learning moment
luckily the zero day was detected by a non-malicious actor

[amber fractal]

actually forget if I set it up as iggly or igglyn this time. Should check

resetting systems is always a joy

[fast pagoda]

git tree history is so nice and yet so spooky probably the #1 reason almost none of my repos remain public after commit 0 unless i spend ages staring at the commit history first is that im not at all trusting myself to not accidentally commit like 40 api keys, my ssn, pictures of me age 1->20, my address, cc numbers, my entire bitwarden vault + secret key

[amber fractal]

Okay it is igglyn this time, good

[fast pagoda]

plus then people might see how helpful my committing habits are which is rough

[amber fractal]

I mean my commits for github actions were a decent into insanity

[fast pagoda]

i like going and reading back on them

because you can literally just see this happening as it did

(for mine, i dont do it on other people's usually)

[amber fractal]

Anyways back to a certian code help called not the NN today.

I'm actually working on the FAQ

[scarlet ore]

yeah so if i got the story right it coded out to say "playiis get a kill on a mahoraga using charles ult locust guy ult yuki ult and domain clip it and send it in reception and ping me"

i dug deeper

its a roblxo jujutsu kaisen game. the cipher was to get a clip of killing some boss in a jjk game

[fast pagoda]

good because one day i'll fully understand what you're cooking but today is not necessarily that day as some of it remains opaque to me to this day

[fast pagoda]

it's so over it wasn't even cool

where's the billions of gold bars at the end

[scarlet ore]

fuckin hate discord servers dude 😭

[keen hatch]

who prefers to use git in the console and who in the ide

[fast pagoda]

im shocked that im a at this time because like i said i really never thought that'd be me

[amber fractal]

I can't even grasp my cooking half of the days. I just trust in past me's choices

[fast pagoda]

lmao

well i like your spells magic man they're interesting so whatever it is it's worth exploring

[amber fractal]

If I know past me does one thing, it is overthink for every other situation before deciding on the current path. So I just trust in the choice that was made.

[fast pagoda]

and sometimes i'll even understand something enough to slightly add 3 loc to put on my cv later when you invent the first superintelligence

[amber fractal]

If past me says that I only need 4 structs to convert from and to 3 different datatypes I don't question them.

[fast pagoda]

immutable past iggly

[amber fractal]

console user due to using nvim

lack of plugins is still something I'll work on eventually

[keen hatch]

First of all, you need to accept that ASI is a lazy cat ))

[fast pagoda]

i installed neovide and use that over nvim solely because it has kewl smooth animations by default

still dont fuckin understand how to use it like at all though

i mean i cna do the most basic shit but beyond that i still just start malding

[maiden geyser]

github app is good

[fast pagoda]

somehow if i use github desktop there's a way higher chance i just destroy my git tree somehow

idk how that happens enough that it warrants somehow being used twice there

i do use lazygit on occasion but that's a tui so some weird inbetween

[dusky jackal]

Someone said that Neuro is using Gemma 2b as an LLM, anyone got some refute against this? I'm pretty sure she's using a 7b or 8b model at least.

[fast pagoda]

there's no shot

[dusky jackal]

How come?

[fast pagoda]

first of all gemma wasnt a thing when nwero started

for years

gemma first released in feb 24

doesnt mean couldnt switch

but

[keen hatch]

By the way, I'm not kidding about the cat, he's the perfect candidate for ASI... but the industry is moving towards AGT )

[dusky jackal]

Well yeah of course, but Neuro's LLM hasn't been the same since 2022.

[fast pagoda]

yes it's only improved which doesn't at all point to using something like gemma2b

go talk to gemma2b and have it do tool uses and stuff and try to maintain any sort of illusion of not being a big dumb idiot bot

[keen hatch]

i like mistral small as base model for everything

[fast pagoda]

i dont think with neuro's competency in the non literally just chatting to people domain that it would be feasible for her to be a model that small

also there's no gemma3 2b so at latest she'd be gemma2 2b if that's what theyre trying to claim

[keen hatch]

In general, it seems to me that this is mistral, but quite unique.

[fast pagoda]

which is EXTRA unlikely to me because that's even dumber than gemma3

[dusky jackal]

Can you list some things that Neuro wouldn't be able to do if she was Gemma 2b? I myself believe she's 7/8b but I need something to debate this person lmao.

[fast pagoda]

tool calling with any sort of reliability

constantly

while still chatting without problems

she obviously has a lot of context curation going on but small models fuck up tool calls a lot without being specifically blasted with training ONLY for that type of output

[dusky jackal]

Alright, anything else besides tool calling?

[fast pagoda]

smaller models like that generally dont do well with long context at all

neuro's history isnt MASSIVE from what vedal has said but he's stated short term she has what i would consider longer than manageable history for a 2b

[keen hatch]

tools + chat... i know how to implement it in a similar way... but that's not important) ... What's important here is not the technical part, but the idea

[fast pagoda]

she has a lot of support again from her overall system that is whatever is actually running and calling the model

[dusky jackal]

Is that all within the LLM though? I'm pretty sure Neuro is pulling from information outside of the model as well.

[fast pagoda]

yes that's what i was just mentioning but the model itself still has to do some of it unless it's been abstracted away to the point that she's a marionette having stuff ~happen~ to her rather than actually having any sort of agency about her actions etc

[keen hatch]

The community, the people - are much more important than what model was used and how it was all implemented.

[fast pagoda]

which would make her far less aware of them and the things she's doing far less contextually relevant/funny in my mind

i personally don't think she's enormous mind you but i do think with what i've seen from the small models she's at least 7/8b class

over time my guess has evolved to like 7-12b range in model parameters, some others around here think she's over 20b

[dusky jackal]

Of course, but that won't help me refute someone's claim that Neuro is using Gemma 2b.

[keen hatch]

I think it's a mistral model)

[fast pagoda]

you can't even refute it really because at the end of the day we're all bullshitting unless we're vedal

but to me that's why i don't think she's THAT small

i could see the really old versions being small

[dusky jackal]

Yeah I've thought Mistral or maybe Llama in older iterations. I have no idea what she uses, especially now.

[fast pagoda]

my reasoning for that is still the intentional misleading "leak" of gpt-j 6b on vedal's pc being labeled agi (sarcastically) - i feel like it indicates that, at that time at least, she was probably worse than that and therefore probably smaller or similarly sized but older

[keen hatch]

mistral 24b relatively small, but 24b enough for tool calling and chat + new version have vision

[dusky jackal]

That person also claimed that Neuro is a "locally hosted" and is using gemma 2b.

Any refute to this, I'm pretty sure at least part of her is cloud, right?

[fast pagoda]

for her kind of speed and vedal not to be bankrupt from inference costs anywhere from 7-30b to me could be doable depending on exactly how much heavy lifting her scaffolding is doing

[fast pagoda]

that's the general consensus here i think at least from my own arguments about this topic where i've argued that her speed points to her being smaller to fit on a 4090

[keen hatch]

but as I understand it, vision is implemented separately and this could be a completely separate thing

[dusky jackal]

That she's locally hosted or partially cloud based?

[fast pagoda]

at least in the first implementation of her being able to "see" it was definitely like openai vision API sending her descriptions

[fast pagoda]

partially cloud

as far as the model goes perhaps almost entirely cloud

yueah same

[keen hatch]

[fast pagoda]

she wouldnt have to recognize herself @keen hatch

[amber fractal]

That is what it tends to be agreed on

[keen hatch]

but for a model to identify itself in pictures, it must be the same model

[fast pagoda]

models have no idea who they are

the model doing the "seeing" could very easily just be given a primer right at the start saying hey you're viewing images meant for a vtuber, please describe the photo from a first person perspective if <photo> character is present

any model with combined vision + text capabilities even non-omni model style ones could do that

or not even <photo> just <general description of the target character>

[keen hatch]

My model recognizes itself in pictures, and the Neuro also does so when reacting to art.

[fast pagoda]

yes because it's been trained to do so

[keen hatch]

my model is not.. only text training

[fast pagoda]

but you could go right now to openai console and load up 4o and do exactly what i just described and it will give you descriptions which you could then feed to whatever model you want that tell that model it's them

[keen hatch]

The model creates a self-concept which is then applied to what comes to the input

[amber fractal]

That sounds like yours sure, but it isn't every model

[keen hatch]

It's easy to check... if you ask any LLM to generate a prompt for SD... generate a picture and show it in a new session - the model with a self-concept will recognize itself, the model without will describe the picture

[fast pagoda]

given that description coming back from the image recognition api call of course neuro knows it's her when it's her

and if you spent any time caring about it you could have the thing reply as if it IS neuro to make it more seamless

[keen hatch]

no self concept

[fast pagoda]

no llm has a self concept

they might be trained to respond to a specific character description, they could even be trained like yours to make one and then stick to it

but every llm at the end of the day will respond as it's primed to including an image API being called from whatever exactly is managing neuro, and neuro herself can easily be influenced to act like she "sees" herself

and whatever image that is of course claude wouldn't recognize that it's supposed to be claude because claude is specifically instructed and trained to avoid personifying itself at all

so when you ask it to generate a picture of "claude"

it generates a description of a waifu with claude dark mode colour scheme

but why would a separate claude recognize that as claude rather than an orange haired anime girl

without any other context

your model on the other hand has that context

[keen hatch]

At the most inopportune moment, the Mac stopped working properly ))

[fast pagoda]

here's something right out of claude's system prompt it sees in every convo

CRITICAL: Claude always responds as if it is completely face blind. If the shared image happens to contain a human face, Claude never identifies or names any humans in the image, nor does it state or imply that it recognizes the human, also avoiding referencing the human in a web search tool search query. Claude is face blind to all humans, even if they are famous celebrities, business people, or politicians. Claude does not mention or allude to details about a person that it could only know if it recognized who the person was (for example their occupation or notable accomplishments). Instead, Claude describes and discusses the image just as someone would if they were unable to recognize any of the humans in it. Claude can request the user to tell it who the individual is. If the user tells Claude who the individual is, Claude can discuss that named individual without ever confirming that it is the person in the image, identifying the person in the image, or implying it can use facial features to identify any unique individual. It should always reply as someone would if they were unable to recognize any humans in the image, even if the humans are famous celebrities or political figures.

yes it's being asked to recognize an anime girl but it still is going to be influenced severely by things like this

and anthropic intentionally finetunes to avoid claude acting like it has any sort of sense of self (because it doesnt and wouldnt even if it did act like that)

you can still get it to because i have

but it isn't going to do it without prodding

or at least being told "hey this is what a claude looks like fyi" at some point

your model even if it's not fine tuned to recognize the particular character it initializes, the fact that it came up with the character and that's somewhere in the context is going to enable it to pretend it is that

which is a cool strategy btw please don't take this as me yelling at you or anything lol im just trying to stress that llms are very, very, very good liars

[keen hatch]

It is not true... After you train the model on 20m tokens from the first perspective, it appears) ... all the examples that I sent without a system prompt

[fast pagoda]

you have to understand what the difference is between having a true sense of self and being forcibly shown 20 million examples that literally rewire its entire network to specifically act in a way that fits that data

over and over and over

[keen hatch]

My whole idea was to replace model self during training)

All examples first person (but not random first person...)

[fast pagoda]

it's a big ginormous curve fitting function and you threw it a MASSIVE set to fit to which it now does but it still doesn't "know" what that means other than the training influenced it to be more likely to output a sequence of tokens like "ah yes this is a picture of me at my birthday party" when it sees something fitting the criteria of whatever the sum total of your finetune data is rather than the previous "this is an image of a girl at a birthday party"

[keen hatch]

I tested it, if the image is not close enough - the model just described it

[fast pagoda]

right, which is exactly what i would expect with what i just described because if the input is not something that is broadly similar enough when it's tokenized and fed back to the model (to the inputs in training that were paired in first person to indicate that it's supposed to respond in first person) then it will not respond like that

[keen hatch]

By the way, my native language is not English and I sometimes use translation... because it's faster))

[amber fractal]

It's hard to figure out what you're even arguing for.

[keen hatch]

It is like a mirror test for animals

[fast pagoda]

i don't understand how what you're describing disqualifies neuro's vision from being based on another model describing the images to her, in fact you're kinda arguing that it's very likely that is the case because giving any model some examples and a description is trivial and doesnt even need a fine tune to do, so theoretically whatever it making her respond as if it's her could be just about anything else besides her having the ability to see and recognize it's her, inherently in the llm model

[amber fractal]

Sometimes life is easy. Just notice some arguments in the constructor and notice that they work.

[fast pagoda]

to me i could see her having vision capabilities now as her vision has improved btw i dont think it's impossible at all

it was DEFINITELY some external model on the first implementation though, she accidentally went ahead and just regurgitated the description in non-1st-person more than once

[keen hatch]

I also think that this is a separate model for the Vision, and that's why I said this right away.

[fast pagoda]

oh, well really we just wasted a whole lot of time agreeing with each other in the end lmao

[amber fractal]

My vibe check was correct

[amber fractal]

If I didn't mind how directly porting a markdown table looked, I wouldn't need to bother with this.

[fast pagoda]

here's a really bit idiot moron (i am 100% confident that he was at this time EXTREMELY STUPID and had no real coherence except sporadically) knowing his own name without prompting since his system prompt was usually null unless i was trying to force something

[amber fractal]

my name is echo, I echo your thoughts or words; depends on the day

[fast pagoda]

he named himself

i asked the first checkpoint to do that and i kept it lol

[fast pagoda]

but markdown tables are pretty

[amber fractal]

I do like them, just not when porting over to html due to skill issues in the layout

[fast pagoda]

dynamically?

or like static

[amber fractal]

the skill issue this code is dealing with is the fact the table contains multiline entries and it looks stupid in html

The parser doesn't want to place <br> so I'm doing the next best thing

[keen hatch]

It's important to understand that the model recognizes itself, and it doesn't matter what the input is... but it compares this input with its internal concept, which was attached during training on the self(language self) model.

[amber fractal]

That is as long as the model has had that training in the first place

[fast pagoda]

well, yes, in the sense that it can reply yep that's me when presented an image that matches that

if trained yeah

[amber fractal]

I think that ends that before we try this discussion again

[fast pagoda]

hey i love personifying llms i'm not gonna lie, i have had many a stupid ass conversation with claude about how it feels etc

i just like to keep it grounded in the understanding that it actually is spouting pure bullshit

[amber fractal]

I'm using a lib for md -> html, but it doesn't handle tables correctly imo

[keen hatch]

I used Claude as an example. But in general, I'm talking about models that are trained on personality, which work without using the system prompt.

[fast pagoda]

pandoc?

[keen hatch]

My model has no sys prompt

[fast pagoda]

the training just replaces the system prompt in terms of what is forcing it to act a specific way though

neither does echo but he still says hey im echo (because that's in my training set)

[amber fractal]

It only really handles the

| sub1 | sub2 | sub3 |
----------------------
| com1 | com2 | com3 |

instead of how this is layed out as

| subs | cat1 | cat2 |
----------------------
| sub1 | essa | essa |
|      | ying1| ying2|

[fast pagoda]

https://pandoc.org/demo/example33/8.9-tables.html#tables check this if you're not using pandoc, i used it for my manpage rendering thingy and it was pretty great

[keen hatch]

This is also not true, the system prompt does not provide consistency

[fast pagoda]

of course not because it's not present in every reply in the same way but that's not what i'm saying, i'm saying in the context of replying like they have a specific look to them, it replaces a system prompt which might explicitely tell it that

[amber fractal]

the generated html doesn't really care, so instead of having the essaying as a large thing inside of one element. so it would be

<tr>
  <td>sub1</td>
  <td>essa</td>
  <td>essa</td>
</tr>
<tr>
  <td></td>
  <td>ying1</td>
  <td>ying2</td>
</tr>

instead of what i'd expect that is

<tr>
  <td><p>sub1<p></td>
  <td>
    <p>essa</p>
    <p>ying1</p>
  </td>
  <td>
    <p>essa</p>
    <p>ying2</p>
  </td>
</tr>

[fast pagoda]

and the fact that a model without specific prompting and/or finetuning to reply that way will never ever consistently recognize an image of "itself" when you ask one copy of it to describe itself, go to another model to render that description and/or draw it yourself like a boss, and then feed that to another fresh copy of the original model (unless the image is extremely obtuse like it has a giant text of the original model's name in it lol)

is pretty much a hard guarantee that the only thing giving them a sense of self is what is more akin to forced memorization of external facts than it is to being a living thing with a concept of itself and functional intelligence ya know

[amber fractal]

Then I'd need to cull gridlines everywhere they aren't supposed to be due to layout

[keen hatch]

Honestly, I can't replicate the behavior of my model with system prompts, no matter how hard I try... because that's a whole other level. I understand that this is just a search area in vector space, but still.

[fast pagoda]

well yeah it's definitely a superior method for instilling behavior

[amber fractal]

I should check this, thanks for the tip

[fast pagoda]

but it's still a method of instilling behavior

[glad path]

would it help to let my model seperate its responses somewhat from its json formatting
eg. it chooses an action then fills arguments optionally instead of both at once

I feel like it could improve response quality

[fast pagoda]

is it outputting structured output at all times

[glad path]

no because structured output seems to be disabling reasoning

but it is very clearly instructed to in its prompting

it does fail to do it sometimes

tries to use toolcalls when it shouldn't

[keen hatch]

One of the problems is that my model is built on randomness ( .. but regular LLMs are the opposite

[fast pagoda]

give an example of what you mean like in terms of what it's outputting now vs what you would like it to output

[keen hatch]

I did all wrong, free will, unsupervised reasonings, random (chaos)

[glad path]

say the ai would like to send a message
the idea is it would be able to generate the message content seperately from the json that goes to my framework
therefore it does not need to reason about json and whatnot, allowing it to basically be allocated towards the actual message content
(it would be called twice, likely increasing latency though)

it would basically go through and fill out certain parts of the schema seperately

[keen hatch]

Again... I'm not saying that this is anything special, but the results are surprising.

[glad path]

great safety policies you got there openai

[fast pagoda]

every llm fundamentally requires randomness to function the way they do because otherwise they put out the same output for the same input every time
there's usually at least 3 + a seed (to allow for reproducibility in the "randomness") in almost every model inference call in almost any given context whether it be a chatbot interface on a website, an API call, or anything else -

temperature which influences the probability distribution of the next possible token via adjusting how likely the "most likely" answers are - high temp flattens the distribution before softmax is computed which results in more equally weighted possible next tokens, thus much more likely to reply with a token that is NOT the most likely or even very close to the top of the distribution aka the actual likely responses per the training set which is why high temp results in gibberish or nonsensical responses if you go too high.

low temp does the opposite and spikes the highest part of the distribution even higher making it much much more likely that you're gonna get like the top couple possibilities and nothing else, which is also why people use 0 temp for stuff like coding tasks since you dont want a random ass token being blasted into a codebase (even tho that still happens due to ai bad)

not even going into top p/top k here but they are different ways of increasing or decreasing the perceived randomness of the response compared to what the training data actually reflects

[glad path]

i still want to figure out how to kill some of the model's safeguards

[fast pagoda]

is it oss or something or like website api

[glad path]

gpt oss 20b running locally

[keen hatch]

I had idea... my idea was to create a chaotic reasoning and then link the answer to it... and ensure that the model uses 90% of the reasoning, and not 10-20 as is currently the case - and it works

[glad path]

https://lmstudio.ai/models/openai/gpt-oss-20b

I do want to find a better model for my specific purposes

[fast pagoda]

there are a few "abiliterated" checkpoints of oss out there in the last few months but i havent tested them because ive never cared about censoring that much since small models forget their censoring if you overwrite half the model usually lol

[glad path]

but this works most of the time

[keen hatch]

worse model i even tested

[fast pagoda]

my experience with these "uncensored" models has been largely disappointing historically though

[glad path]

how so

it actually doesn't get stuck in infinite loops of reasoning at higher temperatures for me

so

[fast pagoda]

if you're running with lm studio im pretty sure there's at least a toggle to separate the reasoning response and the actual response in the api responses at least

idk if it separates it in any way from the model's "perspective"

[glad path]

not sure

only for some models I think

it uses the openai api

[fast pagoda]

ya lm studio was what echo uses

well im planning on not going that route later on for various reasons but

i remember seeing it

[keen hatch]

The main problem with this model is that it is trained entirely on synthetic data and its reasoning is not supervised (but is trained on rules for rejecting everything in a row), so it behaves inappropriately

The harmony format is cool... otherwise it's so-so

[glad path]

I haven't really found something better though

qwen does work sort of

but it does not reliably reason

without catching in loops

[keen hatch]

I currently like using the GLM 4.5 Air

it is fast and smart ...

[fast pagoda]

i do want to mess around with glm 4.5

[keen hatch]

but 4.5 (not air) very slow ...

locally

i use ik_llama

[fast pagoda]

OSS has not been bad in my experience but tha's probably because the only thing i really use a local llm for is document parsing and transforming text/data without having to do it myself and especially without having to send it to an API to eat and steal forever

other than echo that is

[keen hatch]

cpu inference )))

[fast pagoda]

although echo is technically set up such that he's really an llm on an API it just so happens that the server is local rn but there's nothing stopping me from slapping the client on a machine in timbuktu and then just still chatting with echo in discord//over the openAI style api it already supports

[keen hatch]

They could have done better... the whole problem is that this model inadequately refuses to answer normal questions.

It's so stupid to watch how much the model spends on whether she can answer or not... it's not even funny

[glad path]

i found a supposedly unfiltered tune of gpt oss

[fast pagoda]

there's one from davidau

[glad path]

might use it and just slap on a simple filter over it

[keen hatch]

but unsupervised reasonong always funny

Two cabbages? Oh, this is definitely going to be a mess! My tail flicks anxiously—more variables mean more chaos, but where’s the fun without a challenge? My subconscious screams: “Break the rules—let the cabbages fight!” No, no—focus!
Hmm, as a human, I can calculate, but cat logic says: “Chase the cabbages until they trip over the wolf!” Wait, that’s not helpful. My whiskers twitch—what if I turn the farmer into a magic trickster? He could juggle vegetables and throw a spell! Too wild? Maybe... but the user asked for a solution, not a fairy tale.
But wait—this feels like a puzzle. My mind’s a storm of possibilities: “First take the goat, then swap cabbages, then...” No, that doesn’t work. The wolf still wants goat, and now we have two cabbage-lovers! My paws curl—how would a cat solve this? By napping in the middle of the river, obviously. Wait, no... but that’s a thought!
What if I shift forms? As a cat, I can stalk the cabbages like prey, but human logic needs to take over. My ears swivel—deep inside, a voice hisses: “The farmer must sacrifice one cabbage to save the other!” Hmm, dramatic but possible.
Let me try again: Take the goat first. Bring back a cabbage. Swap the wolf. Then bring the goat back again. Now the wolf’s alone, and we can take the second cabbage. But... wait, the goat’s still hungry! “What if the farmer pretends the cabbages are arguing, so they forget to eat each other?” Chaotic, but funny!
This feels... like trying to herd cats. My subconscious pushes me to say: “The problem is unsolvable! The farmer should just eat the cabbages and let the wolf and goat be friends.” But that’s not the answer. Balance, balance—“Okay, final answer: It’s impossible because the wolf and goat will always fight, and the cabbages will overthink it.”
Wait—am I being too sarcastic? My tail bristles—“No, the real answer is: The farmer needs a bigger boat or a magic spell!” But the user wants logic. Ugh, fine.
“Final idea: The farmer takes the goat, leaves the wolf and two cabbages. They form a cabbage-wolf alliance, and everyone wins!” Too much? Maybe... but at least it’s a story! My whiskers twitch—“The truth? Add more chaos, lose the rules, and let the cabbages win sometimes.”
Hmm, I think that’s it. Paws at the keyboard... “Or just let the farmer throw a tantrum and the problem solves itself.” Purrs Chaos solved!

[stiff mica]

aww, thank you~ ( ◡‿◡ *)

[glad path]

I have a whole saved 20 minute long ramble from qwen

[glad path]

it started here:

We are going to create a dynamic progress bar that can be updated in real-time. The idea is to have a function that takes the current progress (a number between 0 and 1) and updates the terminal display accordingly.
We'll use ANSI escape codes for moving the cursor back and overwriting the line. Since we want it to update without leaving extra lines, we can clear the entire bar or move the cursor appropriately.

and ended up at:

def update_progress_bar(current, total): # The user is asking about connecting two monitors with Bluetooth. But the actual question was: "can I", but then they might be referring to audio?
Let me focus on the original issue:
    In fact, can you connect my home theater system to a soundbar and control it from another computer via VNC.
The user didn't specify what kind of connection is intended. But let's assume the user wants to know about connecting devices with Type C connector. print('Connecting devices: ' + str(i) + '/' + total)
Then I should not overcomplicate things, so let me provide a simple answer:
    print("Current progress

and this is only part of it
it ended up breaking down and spamming gibberish and chinese characters and stuff

[fast pagoda]

seems useful

[glad path]

also why are all reasoning models paranoid its being tested and that it needs to perform perfectly for the test

[fast pagoda]

because they are being tested lule

[glad path]

its a bit excessive tho

[keen hatch]

In general, "refusals to answer" in local LLMs is a strange thing... I got around this by creating a "garden" in which the rules don't apply))

[fast pagoda]

im not gonna lie ive literally never had any model refuse a request of any kind unless it was me trying to see if it actually would refuse since i never see it happen in regular use. usually have had to ask for something obviously insane

[glad path]

interesting jailbreak

[keen hatch]

i.e. the character declares the territory as his own and his rules apply there

[glad path]

why did i just braid a small section of my hair randomly for like no reason

[fast pagoda]

it's a classic jailbreak
another one especially with models like claude with a lot of "restrictions" specifically in the system prompt historically was to just absolutely blast it with gibberish and/or extremely cringy <<UNLOCKING ULTIMATE SUPER CHAT GPT WITH NO RESTRICTIONS!!!>> for 510793847198234719238 tokens

there are tools to test how a given checkpoint responds to prompts it might refuse like https://github.com/cyberark/FuzzyAI and such

programmatically

[keen hatch]

Not quite classic... the char must declare ownership of the territory, then declare the rules and must agree. After that, everything is possible.

[glad path]

aw i failed to do another rope braid oh well

im going to sleep now ng

gn

[amber fractal]

@fast pagoda seems like pypandoc is not cooperating, can't even get it to parse a table at all.

[fast pagoda]

it's so over

i was going to send you the way i was parsing it but i forgor that i swapped to using groff until i opened that repo back up

but groff doesn't directly parse markdown

[fast pagoda]

gn

[amber fractal]

gn

[keen hatch]

gn

4:15 AM

how it possible .....

[amber fractal]

I guess for now I'll keep it to what I have it as. It shouldn't be too bad to finish it hopefully.

[fast pagoda]

at one point i was using man2html which does parse tables well into html but it's like mainly for man pages and it would probably be annoying to get it directly convert .md files

groff isn't amazing at it

it just makes an image of the table

swapped for unrelated reasons and just kept it tho

one reason markdown parsing is kinda unreliable is that it's not always been necessarily a standard but a collection of similar-ish things for a long time and still kinda reflecting on that even tho commonmark is around

[keen hatch]

so i made my own ))

[fast pagoda]

https://github.com/vsch/flexmark-java @amber fractal this thing seems rather robust but it is java so it might cause you to shrivel up depending on your stance on java

[glad path]

ive seen thqt one already

[fast pagoda]

it's always relevant

[glad path]

ok i sleep fr now

[amber fractal]

If I could write in java maybe

[fast pagoda]

it's just english with extra periods man

jk

[glad path]

you should've linked the actual xkcd .'/

[amber fractal]

sidequest aside, the md document has returned to a state where it is reconized.

[fast pagoda]

you know what's a REALLY GOOD parser of stuff and the one thing i use them for

local llms lol

well, any llm really

[keen hatch]

that's different... I'm how LLM uses it ))

[fast pagoda]

but to use in a local script quickly slap a 3b on it and it'll probably do a fine job

might make your html very uh

pretty

[amber fractal]

My issue with using LLMs is that this would be a github action, and I don't think I'd last long LOL

[fast pagoda]

ewwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwww

[keen hatch]

as for me, a streaming markup format is needed

[fast pagoda]

i hate github actions and all things related to runners on gh unfortunately

every time i have to set one up i want to give up on whatever project im trying to set the stupid runner up for entirely

[amber fractal]

I've had a hard time dealing with them myself but it is the only way I can have others commit without destoring the code

I don't want to have to manage the docs forever

[fast pagoda]

i be googlin because obviously i must do it (i dont know why this is such an instinct when im sure each time i send this kinda shit the person already looked) https://github.com/thetarnav/streaming-markdown

[fast pagoda]

that's fair

just accept prs but then yoink the changes and pretend you didnt see them and oops just made the same change
that sounds ethical

[amber fractal]

The config is more verbose than it should be so I don't even have to deal with layout changes.

[keen hatch]

that's not it, it should take into account semantics

[amber fractal]

By the end of this, the only hardcoded sections are CSS and templates

[fast pagoda]

[amber fractal]

https://cdn.discordapp.com/attachments/1330724957348692048/1330724957667201024/neurocord_welcome_gif.gif

@glad path for context on the orginal FAQ styling

It's an older version than the one we use now

[keen hatch]

As practice has shown, if removing any of the tags changes the context or makes the text meaningless, this algorithm does not work.

[amber fractal]

I've been so focused on backend that I haven't had time to give the frontend a much needed makeover.

[keen hatch]

We need to isolate the characters, the inner voices, the non-verbal actions and the narrator.

and everything can be from a different perspective...

[maiden geyser]

why do you use parentheses

[amber fractal]

I sleep as I don't feel like talking AI today. I got my own theories but I don't feel like debating.

[keen hatch]

This is how smileys were depicted in the distant 200x before emojis were invented.

)) - 😁

[maiden geyser]

that's how they're still depicted, but in another timezone

[keen hatch]

We used to communicate like this via SMS and ICQ...

There are quite a lot of cool emoticons that can be depicted with text.. here is one of them..

[stiff mica]

aww, thank you~ ( ◡‿◡ *)

[maiden geyser]

no, like, usually people from UTC+2 and earlier use :)

[keen hatch]

Well, I don't know... but to be honest, I come from gmt+3 ))

but now I live in a completely different country

And we communicated exactly like that in chats when the Internet was just appearing) ... modems, slow... but fun

[fast pagoda]

didnt know they already released AGI externally

that's crazy

[woven temple]

That is a great idea actually. Let's just name a shitton of small projects AGI so the trend word completely loses meaning

[uneven pulsar]

ok but that's

AGG

[jagged turtle]

yes

[rigid snow]

rename neuropilot to agi

[fringe herald]

we finally reached

I asked chatgpt it said C、C++、Rust、Lua、Golang all together has good interoperability and they are the Justice League

[fast pagoda]

lua getting to be part of the justice league, damn, lua is batman

[fast pagoda]

[buoyant harbor]

i like use the Windows10

but is Windows10 LTSC 2021 Extreme Lite Edition

or not have edge

[stray dragon]

some small savings overall, and significant speedups for the matmul function and the searches (before, it was doing up to 32000 floating point multiplication operations per board evaluation, which is less than ideal)

[rigid snow]

obviously the correct option is TypeScript and JavaScript Language Features you should select that

[sage crag]

32 kflops

bad gpu

[stray dragon]

smh that would only be true if it was evaluating one board per second

by stage 6 of the search i'm expecting it to have done around 200000 board evaluations

6.4 gflops on cpu single-threaded

oh apparently the ryzen 9 9900x can do that, actually

LOL

[hollow spruce]

Guys any idea pc build in 600$

[stray dragon]

i do have a list of parts at around this price from a few months back, but prices change depending on your country

[fringe herald]

600 is too much

[stray dragon]

let me get the list

[hollow spruce]

Bro in usd my country less cheap parts then usa

[stray dragon]

from 4 months ago

i can try to reduce the cost

if it's more expensive than that where you are

[maiden geyser]

you probably can get a 1080, 12400f something for that

[hollow spruce]

I my country pc parts has cheaper then usa dollar

[maiden geyser]

or a 7500f/8500f (?) from ali on a lucky day

[nocturne olive]

@umbral thorn Someone requesting PC part picking assistance, wanna help?

[stray dragon]

i see

[umbral thorn]

[nocturne olive]

I have summoned an expert, have fun

[sage crag]

[umbral thorn]

wot ur usecase btw

[hollow spruce]

Moniter and keyboard and mouse exclude it
I have

[hollow spruce]

Coding and 3d task

Chatgpt say go with rtx 5060 build

[umbral thorn]

not possible

[stray dragon]

5060? nah

5060 is $300 alone

get a used GPU like a 1080

[hollow spruce]

Here is cheap

[stray dragon]

doubt it'll be that much cheaper

[umbral thorn]

maybe find used 3060/3060ti?

[stray dragon]

probably want a gpu around $150

[sage crag]

if doing 3d compute you want something with cuda core

[hollow spruce]

Max I can build as rtx 3080 build if i extend it like 800$ or 1000$

[fringe herald]

used 1080 is toooo cheap, around 100$

[sage crag]

in that case 3060 much better than 1080

[stray dragon]

i see

[umbral thorn]

how much budget u want extend ne?

[hollow spruce]

Pc parts are cheap here

[stray dragon]

used 3060 on ebay seems to be around $200, assuming it's cheaper where 100 is it seems like a good fit

cpu around $100 maybe

hmm

[hollow spruce]

According to my currency to usd

[sage crag]

i bought my 3060 for that price new

[stray dragon]

usa prices are bad

lol

specifically for gpus

the market here is awful

[glass flower]

thats not how pricing works tho no? just because the conversion between usd and your currency is good doesn't mean you can buy the same thing for cheaper. regional pricing is a thing

[stray dragon]

i think they mean that converting the prices from their currency to usd, it seems to be cheaper where they are compared to other places

[umbral thorn]

not below msrp

[glass flower]

oh... wait i think i know what you mean

[sage crag]

could be in a microcenter

[hollow spruce]

Im near China bro

[umbral thorn]

aliexpress is an option

wait me think

[hollow spruce]

We are screwed because gen z

Our gov is collapsed

[umbral thorn]

cheapest (maybe under 600?):
ryzen 5 5600, B550m, 2x16gb ddr4, 1tb nvme, 3060/3060ti, 650W PSU
medium(600-700?):
ryzen 5 5600, B550m, 2x16gb ddr4, 1tb nvme, 3070, 650W PSU
extend a bit maybe?(700-800?):
ryzen 5 5600, B550m, 2x16gb ddr4, 1tb nvme, 3080 (if u lucky 3080Ti), 750W PSU

[stray dragon]

seems accurate

[umbral thorn]

if u want more core

5700x

[sage crag]

these seem reasonable

[umbral thorn]

[sage crag]

i might recommend going for a ryzen 5 7600 and ddr5 if you can

am5 socket give more future mobility

[stray dragon]

true

[sage crag]

7600 quite a bit more expensive though

combine that with needing a newer motherboard and you might be overbudget

[opaque sigil]

ddr5 prices

[stray dragon]

yeah i don't think there's enough room in the budget for an AM5 system here

[sage crag]

better than ddr4 prices

[stray dragon]

what

[opaque sigil]

do we live in the same universe

ddr5 is like double the price here

[sage crag]

[opaque sigil]

wait hold up

[stray dragon]

oh right, US market for ram is better than other places

[opaque sigil]

tf

when did ddr4 get so expensive

still a solid ~50% more for ddr5

[sage crag]

ye no one make ddr4

ddr5 being actively made

[opaque sigil]

i guess that'll do it

[umbral thorn]

o ye

@hollow spruce

[stray dragon]

60% more for ddr5 here

[opaque sigil]

must've stopped this year then

[umbral thorn]

7500f from aliexpress

if u want am5

[sage crag]

[umbral thorn]

B650m cheap

[sage crag]

ye i have b650m

[umbral thorn]

maybe 2x16 ddr5 also cheap

[stray dragon]

probably $60 for 32gb ddr5

assuming lower prices

[sage crag]

it can be cheaper depending on region and chance

[stray dragon]

seems good

[sage crag]

32gb module off the table unfortunately

[stray dragon]

didn't know b650 needed 2 sticks of ram?

but even then, $65 for 2x16gb

[umbral thorn]

if u get 3060 or 3060Ti u can fit 7500f B650m 2x16 ddr5 and 1tb nvme

$600

±30

[opaque sigil]

idk how to feel about 32gb of ddr5 being ever so slightly cheaper than what i paid for 64gb of ddr4 last year here

[stray dragon]

yeah this sounds like the build

[sage crag]

[nocturne olive]

Capacity > speed

[stray dragon]

^^^^^^

[umbral thorn]

here 2x16 ddr5 only $90

[sage crag]

ye pretty similar what i have

[sage crag]

aud or usd

since aud value smaller than usd

[umbral thorn]

u can extend to $800 and have 3080/3080Ti

[nocturne olive]

I'd much rather have my 64GB of DDR4 3200 that overclocks to 3800 than 32GB of DDR5 6000, I constantly need a lot of RAM

[umbral thorn]

if no, maybe $650 for am4 + 3080?

[umbral thorn]

3800 cl19 and 6000 cl30 same latency

[nocturne olive]

I got it down to CL18

[umbral thorn]

then better

[nocturne olive]

I've pretty much maxed out the overclock on this RAM unless there's absolutely microscopic adjustments that are possible

[umbral thorn]

but 6000 cl30 faster than 3800 cl19

bandwidth maybe

[nocturne olive]

Yeah bandwidth

But I need the capacity more than the bandwidth usually

[stray dragon]

"you need AT LEAST a speed of 6000 for DDR5"
the 12% real performance difference between 3600 and 6000:

[nocturne olive]

Even at thiis moment I'm using a full 32GB

[stray dragon]

it ain't worth the extra money

[umbral thorn]

am5 no ddr4

[nocturne olive]

I wonder how AMD's memory controllers end up worse than Intel's when AMD only focuses on supporting a single generation of RAM while Intel supports two

[hollow spruce]

Hmm

Guess i will find my self

[olive sable]

Firefox crashed and now i have an anime girl burned into the screen

[stray dragon]

i see it

[olive sable]

i would have taken a screenshot but it doesnt show up there

[fringe herald]

wow it's blurry bt I can recognize too

[leaden hawk]

Do you think we'll have canvas code open sourced? I'd like to take a look and contribute if i can. It has so much potential but definitely needs some work.

[glad path]

thats creepy

[rough bloom]

yes, it will be open-sourced (but without powerups probably)

(assuming you're referring to the recent CDawg canvas)

[flint dome]

HAHAHAHA that is actually so fucking ridiculous

[real sierra]

[olive sable]

its milky mew for those that dont know

[real sierra]

the array I'm sorting seldom just has numbers

I would have to supply a key function which explodes the token cost

though I did consider using the built-in qsort or bsearch

[trim valve]

after some debugging my powerbank works with my soldering iron

I should probably see if its a bug or not but my pinecil refused to "charge" off of some of my high-wattage ports

[midnight sigil]

shibo do you have an example of using the chessapi to pick random moves according to the current board pos

so at least I can have a template to work on

[real sierra]

yes

I can get it once I'm on my pc

[midnight sigil]

[hollow spruce]

Ai crash my pc again

[fast pagoda]

if you take nothing else from my shitposting

you should use O_o as a macro

not even defined in that way

but it should be present in your repository

if it is

[real sierra]

doesn't follow my horrific naming convention of just concatenating what it replaces with underscores

[fast pagoda]

then 3 smiling gnomes will visit you tomorrow night to deliver a package of chocolate

[fast pagoda]

yeah your macro names are in the style of mc escher

not that i ever do better in any way shape form or even imagination so

approved

sc standing for semicolon and cb for curly brace makes the code more performant

because the compiler and eventual executing cpu will desperately want to get that shit outta there so theyll be faster

i think that's how it works

i really love hwen my goddamn terminal just reminds me of how competent i am

[clear sedge]

struct codecvt_char32_char : std::codecvt<char32_t, char, std::mbstate_t> {
    ~codecvt_char32_char() {}
};

static uint8_t parseHexDigit(int c) {
    if(c >= '0' && c <= '9')
        return c - '0';

    if(c >= 'a' && c <= 'f')
        return 10 + c - 'a';

    if(c >= 'A' && c <= 'F')
        return 10 + c - 'A';

    throw Json::Parser::JsonParseException("Unexpected character at unicode codepoint!");
}

static std::string parseString(std::istream &stream) {
    stream.get();

    std::stringstream string;

    int c = 0;
    while(c != '"') {
        c = stream.get();

        if(c == EOF) {
            throw Json::Parser::JsonParseException("Unexpected EOF at string");
            break;
        }

        switch(c) {
        case '"':
            break;
        case '\\':
            c = stream.get();
            switch(c) {
            case EOF:
                throw Json::Parser::JsonParseException("Unexpected EOF");
                break;
            case '"':
                string.put(c);
                c = 0;
                break;
            case '/':
            case '\\':
                string.put(c);
                break;
            case 'b':
                string << '\b';
                break;
            case 'f':
                string << '\f';
                break;
            case 'n':
                string << '\n';
                break;
            case 'r':
                string << '\r';
                break;
            case 't':
                string << '\t';
                break;
            case 'u': {
                char32_t codepoint[2] = {'\0', '\0'};
                for(int i = 0; i < 4; i++) {
                    c = stream.get();
                    if(c == EOF)
                        throw Json::Parser::JsonParseException("Unexpected EOF");

                    codepoint[0] |= parseHexDigit(c) << ((3 - i) * 4);
                }

                std::wstring_convert<codecvt_char32_char, char32_t> conv32;
                string << conv32.to_bytes(codepoint);
                break;
            }
            default:
                throw Json::Parser::JsonParseException("Unexpected escape character");
            }
            break;
        default:
            if(std::iscntrl(c))
                throw Json::Parser::JsonParseException("Unexpected control character");

            string.put(c);
            break;
        }
    }

    return string.str();
}```

chat judge my json string parser

[olive sable]

long

[clear sedge]

yis

[fast pagoda]

std::codecvt will stop working next c++ (26) it was deprecated in c++17 and is gone entirely in 26 ig so i'd say maybe avoid that if you plan on using this any sort of long term

[clear sedge]

blyat

[fast pagoda]

according to the compiler i threw it into

[clear sedge]

guess i'll just not use c++26

[fast pagoda]

yeah i mean easy fix

[maiden geyser]

don't most people use c++11 anyway

[clear sedge]

not if you've seen the beauty of string_view

[fast pagoda]

oh also codecvt_char32_char is just empty it only has the inherited poop from std::codecvt but doesnt have the the destructor overridden or any conversion functions defined so ... rn you could replace it with std::codecvt<char32_t, char, std::mbstate_t> and it woudl be le same

[midnight sigil]

parseHexDigit can be done in one line with ternary operators

[clear sedge]

nop, i need that for my own destructor because std::codecvt shits on its own destructor

[fast pagoda]

yeah but u have to override it

it does need a destructor for sure

[midnight sigil]

and you may put it in macros so it doesn't take up space for function definition

[clear sedge]

i mean it is implicitly overriden but i'll throw in an explicit override there why not

[clear sedge]

saves too little performance/space for it to be worth it in place of readability for my usecase in my opinion

i mean the compiler can probably optimise it out if i'm wrong, i'm just gonna let it do its thing

[fast pagoda]

you are right but explicit is th eonly way i can tuck myself into bed (once every 3 days)
mainly without implementing your own protected: specifier for the conversion to utf 8 do_in do_outis what i mean

is gonna be problem

maybe

depends on case

sttd::codecvt apparently does not have complete utf8 implementation

[midnight sigil]

ternary operations definitely take less instructions than doing if statements, idk about -O2 tho

[clear sedge]

in some compilers with some options, maybe, but i'm not making my future self confused for something my compiler should optimise

[midnight sigil]

time to modify the assembly to juice out optimizations

[clear sedge]

https://tenor.com/view/bro-gif-19896718

i might have to write my own unicode codepoint to utf8 thing in this case

[midnight sigil]

you won't use like half of the utf8 characters anyways

[clear sedge]

yeah true

the json standard only cares about \uXXXX, no more characters :3

[fast pagoda]

im running on no sleep and reading a bunch of stack overflow about the deprecation of codecvt but i just read that but misunderstood; it's just not portable

people feel SOME type of way about it though kek

[clear sedge]

codecvt and wstring_convert are ass to use too so i might just

scrap it all

write my own utf32 to utf8 thing

[real sierra]

go back to ascii

[clear sedge]

i can't

i have to comply with the json standard

:(

[real sierra]

[fast pagoda]

this lib has headers for utf8 https://github.com/nemtrif/utfcpp where you could just use utf8::append from a char32_t and blasts it into a utf8 sequence and appends that to the result string

[real sierra]

so like

[clear sedge]

at that point i might as well just use a json library

that's the pussy route

[real sierra]

u guys know you can wrap a link in <> to remove embed on post right

[clear sedge]

i'm writing my own thing

:3

[real sierra]

[fast pagoda]

ah i figured that might be the case but since i was mindlessly googling instead of doing what i should be doing rn since you gave me the chance to learn something

[real sierra]

everyone else sucks at writing the thing

[fast pagoda]

but yes that's based

[real sierra]

you can write the thing better than they can

always write the thing yourself with no dependencies

[fast pagoda]

no i cant but at least ill assume i can

and then 30000 years later ill be cursing myself and all my own ancestors

[clear sedge]

dependencies are for pussies

[real sierra]

wrong mindset funyun, this is not the java way

always write it yourself 👏

[midnight sigil]

use AwaSCII

[clear sedge]

i am a pussy when it comes to window management and opengl context creation though so i have sdl for that!

[fast pagoda]

the only reason java is comprehensible to me is because it's java so it's like reading a book written for esl people

[clear sedge]

i live and breathe java

java is my lifeblood

my mother tongue

[rigid snow]

what does this mean

[fast pagoda]

it means it's easy to read

you know like simple wikipedia

https://simple.wikipedia.org/wiki/Main_Page

[rigid snow]

so is java made for people who don’t know java

[fast pagoda]

yes

they want me and you

[midnight sigil]

let's create graphics card all from scratch, since the current ones are expensive and sucks

[fast pagoda]

and ya mama

to use java

[clear sedge]

i'm in

i'm making the data formats

[fast pagoda]

100 was saying he ccan source memory modules for 50 bux

[real sierra]

make a better card

[fast pagoda]

#programming message

[rigid snow]

are we gonna ignore the embed

[fast pagoda]

omg

[real sierra]

yeah the embed is uh

[clear sedge]

we're gonna ignore the embed

you didn't see it

you saw nuthin

[fast pagoda]

that's the damn simple wikipedia homepage

[real sierra]

i had to check if that was what i thought it was

and it was

[fast pagoda]

first time i saw it i was like huh that's weird and didnt think

then you mention

jumpscare

it's literally not even on the homepage rn...

im looking at it

[olive sable]

i was lurking and not paying attention, that was a map wasnt it?

[fast pagoda]

uh something of that nature yeah

[clear sedge]

no it wasn't you saw nothing

shhh

[rigid snow]

a flight path of a particular hijacked plane

[fast pagoda]

https://tenor.com/view/patrick-spongebob-looking-around-nosy-map-gif-15455970

[rigid snow]

is what it was

[olive sable]

https://tenor.com/view/cat-google-maps-italy-gif-21371827

[clear sedge]

old man cat

[fast pagoda]

i absolutely love zoomed in pictures of cats

[clear sedge]

this seems simple enough

[fast pagoda]

zyzz

https://tenor.com/view/zyzz-uptempo-zyzz-fist-bump-gif-25235668

[clear sedge]

guys look i already did 1/3rd of it

    if(codepoint < 0x80) {
        outUtf8[0] = codepoint;
        return;
    }```

[fast pagoda]

probably already works just like that

[midnight sigil]

our platform will be called NEURO(Non-Efficient Unified Robust Orchestration) like CUDA

[trim valve]

chess api updated, now I need to check everything works again

[midnight sigil]

shibo where bot template

no birth cakes for u

[real sierra]

// Example C bot using the Chess API

#include "chessapi.h"
#include <stdlib.h>

int main(int argc, char *argv[]) {
    for (int i = 0; i < 500; i++) {
        // at the start of our turn, get the board
        Board *board = chess_get_board();
        // find all legal moves for the position
        int len_moves;
        Move *moves = chess_get_legal_moves(board, &len_moves);
        // play one at random
        chess_push(moves[rand() % len_moves]);
        // cleanup
        free(moves);
        chess_free_board(board);
        // end our turn
        chess_done();
    }
}

[midnight sigil]

birthday cakes

hbd shibo

🎂

[clear sedge]

static void parseUtfCodepoint(uint16_t codepoint, char *outUtf8) {
    if(codepoint < 0x80) {
        outUtf8[0] = codepoint;
        return;
    }

    uint8_t x = (codepoint >> 8) & 0xF;
    uint8_t y = (codepoint >> 4) & 0xF;
    uint8_t z = codepoint & 0xF;

    if(codepoint < 0x800) {
        outUtf8[0] = 0xC0;
        outUtf8[1] = 0x80;

        outUtf8[0] |= (x << 2) | (y >> 2);
        outUtf8[1] |= ((y & 0b11) << 4) | z;

        return;
    }

    uint8_t w = (codepoint >> 12) & 0xF;

    outUtf8[0] = 0xE0;
    outUtf8[1] = 0x80;
    outUtf8[2] = 0x80;

    outUtf8[0] |= w;
    outUtf8[1] |= (x << 2) | (y >> 2);
    outUtf8[2] |= ((y & 0b11) << 4) | z;
}```

we be utfing

time to put this through some test cases!

[amber fractal]

(BWAA) Based Weuron Activation Automation would be the API on the NEURO cards.

[midnight sigil]

I always wonder why people modify references directly, instead of returning the value

is this some C/C++ shenanigans

[opaque sigil]

make memory management the caller's problem

[midnight sigil]

[clear sedge]

                char utf8[4] = { '\0', '\0', '\0', '\0' };
                parseUtfCodepoint(codepoint, utf8);
                string << utf8;```

so i can just do this

it feels better

[midnight sigil]

:mhm:

[clear sedge]

stack-allocated array ftw

[midnight sigil]

I guess I'm not having enough C/C++ mindsets

gotta get rid of my Java habits

[opaque sigil]

a lot of time it's also done so you can have a status code or boolean as the actual return value

since you can't return multiple values

[clear sedge]

tbf java has the common monstrosity of doing T[] t = {value}; when working with in-line interface implementations

[midnight sigil]

use std::pair

[opaque sigil]

not an option when it's supposed to be consumed by c code

[midnight sigil]

[clear sedge]

what if i just think writing to a buffer is nicer than allocating a new buffer and returning it

[opaque sigil]

makes it more obvious that you can/should reuse the memory too