programming

[trim valve]

mhmm, just kinda memeing at this point

yes

also somehow my formatter doesn't kill it

but the formatter does add a trailing comma

[stray dragon]

was wondering how i would get my token count down further and then you guys mention C macros lol

[real sierra]

have you seen the kinds of macros I've made

[stray dragon]

nope

[real sierra]

feels like war crimes

[stray dragon]

do show

[real sierra]

#define bclose ) {

saved like 50 tokens

[stray dragon]

BRUH

[real sierra]

that's where I'm at

the code is completely illegible

half of the keywords are macros

[stray dragon]

so basically the solution is just "make your own tokens with macros"

i see

[real sierra]

there's some other weirdness to it

I only have one kind of loop now

a for loop from i=0 to some limit I set

that's basically the only loop in my entire code

[stray dragon]

oh that's super weird

[real sierra]

so I can macro it around

it means I reuse the exact same variable names for certain data types so I can bake them into macros too

[stray dragon]

i've been doing shit like this

[real sierra]

whitespace doesn't count as a token

but yeah that's something

I see a bclose on the right

👁️

[stray dragon]

yep

[real sierra]

bclose works on everything I love it

[stray dragon]

damn i could use macros everywhere

[real sierra]

surprisingly tricky to figure out which are useful

the macro definition also counts as tokens

[stray dragon]

yeah obv

[real sierra]

so it's not free

wait t

[stray dragon]

yes shiro

[real sierra]

oh right I posted the chess API header

nvm

[stray dragon]

yeah

[sage crag]

int main() {
    while(1) {
        Board *board = chess_get_board();

        int len_moves;
        if (Move *moves = chess_get_legal_moves(board, &len_moves))
            chess_push(moves[rand() % len_moves]), free(moves);

        chess_free_board(board);
        chess_done();
    }
}

could make that one smaller like this

probably

[real sierra]

then add macros

[tender river]

[real sierra]

chayleaf

am I gonna have 5 different tcount implementations

[tender river]

1

[real sierra]

if you make a working fork that supports Haskell alongside all the other langs

that works for me

[sage crag]

.{get_board, push, done} := @use("chess.hb")
.{rand} := @use("lily")

main := fn(): never loop {
  if mov := get_board().get_legal_moves().get_wrapping(rand()) push(*mov)
  done()
}

lack of glob imports really letting hblang down here

[tender river]

i'll also remove java

[real sierra]

[tender river]

[real sierra]

it looks cool

[sage crag]

hblang

[tender river]

lily libc

[sage crag]

🤫

[real sierra]

not stdlib...

[sage crag]

i can always

rand := fn(): int @import()

libc

[real sierra]

how does that work

that's insane

[sage crag]

linker symbol

[tender river]

lily is the hblang stdlib

[sage crag]

ye

and i write it

[real sierra]

[sage crag]

anything i say is in there is in there

[real sierra]

allowed

[sage crag]

[tender river]

stockfish in stdlib

[real sierra]

[sage crag]

ye could

[real sierra]

my submission in my new language, JavaChess
stockfish();

4 tokens

source: trust me bro

[sage crag]

the stdlib is pretty big already

but not complete

[tender river]

we got silksong before 0.2.0

[real sierra]

"make binaries smaller for no reason" tracks

[sage crag]

[real sierra]

relevant skill

[sage crag]

yeet libc

[sage crag]

ye hello, world in hblang is about 170 bytes

[real sierra]

holy bytes

[sage crag]

as in the binary

not the main function

[tender river]

shiro

[sage crag]

[real sierra]

chay liked it

[sage crag]

it also not link to libc

[real sierra]

don't deprecate c

c is best

[sage crag]

lily.{target, config} := @use("../lib.hb")

@export("_start", fn(): never {
    main := fn(args: []^u8): uint @import("main")
    args: ^^u8 = @bit_cast(@frame_pointer())
    target.exit_group(main(args[2..2 + @bit_cast((args + 1).*)]))
})
@export("memcpy", target.memcopy)

ok but

look at this though

[real sierra]

my eyes

[sage crag]

yes its not pretty

also outdated

ye it reimplements the linux syscalls so it doesnt need libc

[real sierra]

I'm not creative enough to make an eye catching robot chess banner

someone beam creative power into me

it's literally just that and then I can launch the contest I think

[sage crag]

uh oh

[real sierra]

I'll fix API bugs as I go

[sage crag]

how long will it run for again

1 month?

[real sierra]

1 month ish ya

[sage crag]

ehe

[real sierra]

. .

aliases

[sage crag]

ye for Result

[real sierra]

my chess bot 1023 tokens

can't shrink anymore

actually that lie

I can still

#define voidf } void

maybe saves me some if I'm lucky

wait konii question

[trim valve]

btw

[sage crag]

ProgramError := enum {
  .SkillIssue;
}

ermie := fn(): Result(i32, ProgramError) {
  return .err(.SkillIssue)
}

main := fn(): void {
  if erm := ermie.to_ok() {
    log.printf("hello {}", erm)
  } else die
}

result

[tender river]

[real sierra]

#define cond x
#define while_ while(cond) {

int main(int argc, char *argv[]) {
  int x = 0, y = 0;
  while_
    #undef cond
    #define cond y
    while_
      printf("hi\n");
    }
  }
}

?

[sage crag]

yeah sure

works

[real sierra]

nested loop macros unlocked

token savings incoming

[sage crag]

well actually

a strnage

[real sierra]

tho just having while_(v) probably cheaper

but wait it not cheaper

[sage crag]

what is #while_

[real sierra]

because don't always need nested loop

I am slow

[sage crag]

technically you dont need to undef cond

[real sierra]

I fix

[sage crag]

you can just redefine it

it will give you warning though

[real sierra]

that's fine my code already gives me warnings

but I know the order of operations better

[tender river]

[real sierra]

wait

it's the same cat

?

how am I just now learning this

[hollow spruce]

test

[real sierra]

test

[hollow spruce]

d-bus connection

[real sierra]

☎️

[hollow spruce]

phone to pc

[real sierra]

@stray dragon

in your experience of chess bot writing

are there any functions you feel are missing from the API

from a "this is common boilerplate that everyone will need" perspective

[tender river]

shiro for the love of god make Board* not publicly accessible

[real sierra]

?

it's an opaque struct

[tender river]

its a singleton internally

[sage crag]

https://cdn.discordapp.com/attachments/1067638175478071307/1406337271241904158/ezgif-3f552200e37817.avif

[tender river]

so it shouldnt even be accessible as part of the public api

you can just enforce the initialization invariants internally

[real sierra]

get board returns a clone of the board

not the actual internal copy

chill

[tender river]

[sage crag]

https://cdn.discordapp.com/emojis/1212937424062119967.gif?size=48&animated=true&name=nuhuh~1

[real sierra]

even if it didn't, it's all local to the bot so the only person harmed by corrupting the board is the bot user

you still can't submit illegal moves

the Gui will just disqualify you

[tender river]

i wasnt talking about security

[stray dragon]

i do #include "bitboard.h" to import get_lowest_bit(bitboard) and floor_log2(bitboard) so i can do this to pull a piece's position out of the bitboard

i have no idea how Board is structured

[real sierra]

including bitboard.h is legal

I don't think it makes sense to do bitboard manipulations through the API given the function call overhead would be slower than the computation

but I can include convenience functions for some of the weirder bitboard ops

[sage crag]

i just learned of something

[stray dragon]

yeah i feel it doesn't match the api that well either

[sage crag]

__VA_TAIL__

[real sierra]

bitboards are definitely kinda low level

[stray dragon]

mostly i just want the positions of the pieces

in a way i can use

[real sierra]

yeah

I can expose that somehow

[stray dragon]

maybe a function that just returns an array of ints, representing the positions 0-63 on the board for each piece

[real sierra]

would a bitboard_get_x and bitboard_get_y work

[real sierra]

that's doable too

[stray dragon]

no need for x and y

[real sierra]

yeah if I give you the index

[stray dragon]

0-63 is perfectly fine i feel

[real sierra]

you can calculate those yourself

x = index % 8

y = index / 8

[stray dragon]

yep

[real sierra]

that'll remove tokens from my bot too

[stray dragon]

win-win

[real sierra]

I'll add it to bitboard.h

and then maybe add some bindings for that to the py api

or reimpl it

@trim valve do you think bitwise ops would be faster in python or through bindings to C

[stray dragon]

this + a function to get the piece at the position would be good i think

maybe just the function to get the thing at the position

[real sierra]

that would also shrink my implementation a lot I can do that

[tender river]

would looking at existing apis help

[stray dragon]

could just do a loop 12 times, one for each color of each piece, and then go positions 0 to 63 through the bitboard checking each position for a piece
would only need the thing to check if there's something at position or whatever

[real sierra]

would you want get_piece_at to accept an int or a bitboard

[hollow spruce]

i cant sleep

[stray dragon]

int

i think?

hmm

[real sierra]

[stray dragon]

maybe both actually

[real sierra]

yeah the conversion is just

[stray dragon]

(bitboard, position)

[real sierra]

int index = highest_bit(bitboard)

[stray dragon]

welcome

[hollow spruce]

not define

[stray dragon]

it's 12 pm here and i have not been to bed

[tender river]

it's something like __builtin_clz

[real sierra]

no need we just mask the bitboard representing the square with every piece bitboard and return the first one that is nonzero

that's why bitboards are fast

it could technically be a one liner

[hollow spruce]

12 am here

[stray dragon]

programmer sleep schedules are fun

[stray dragon]

clever

[hollow spruce]

4am to 7 am

[real sierra]

PieceType get_piece_from_index(Board *board, int index) {
return get_piece_from_bitboard(board, ((BitBoard) 1) << index);
}

PieceType get_piece_from_bitboard(Board *board, BitBoard bb) {
  return (bb & (board->bb_white_pawn | board->bb_black_pawn)) ? PAWN : (...);
}

[hollow spruce]

bye im going sleepy 😴 sql to apacha pa

[real sierra]

[real sierra]

[sage crag]

are there any cases where
#define O_PAREN2 ((
#define C_PAREN2 ))
could save tokens

[real sierra]

chayleaf tcount fork saves the Haskell world

[real sierra]

yes

I found big token saves came from eliminating array indexing

[x] is 3 tokens

[stray dragon]

awesome

how does get_piece_from_index act between white and black pieces, if at all?

[real sierra]

doesn't matter

[tender river]

1773 tokens

[real sierra]

you can't have both a white and black piece on a square

if you want the colour that's a separate function

[stray dragon]

if the piece at a position is black, and my turn is for white, i want to make the black piece's position contribute negatively to the evaluation function

[stray dragon]

wait really?

[real sierra]

but it might be easier to just have BitBoard get_my_pieces()

then you can just & them

0 means it's your piece

[stray dragon]

my solution before your api was a thing was to just have the pieces be + if white and - if black

white king = 6, black rook = -4, empty = 0

[real sierra]

makes sense

[stray dragon]

easy to handle

[real sierra]

I only look at pieces involved in moves at the moment

so I always know that the piece being moved from is mine

and the one being moved onto, if a capture, isn't mine

[trim valve]

shruge it's python surely both are slow

[tender river]

if you do pypy

[stray dragon]

my evaluation function checks for the piece at a position on the board, and if that piece's color is different from the current turn's color, it contributes negatively to the evaluation function

[tender river]

builtin would be faster

probably

[stray dragon]

if same color contributes positively

[sage crag]

brain

#define while_(cond) while(cond) {
#define dclose ;}}
#define FUNC(ident, ...) ident() { __VA_ARGS__; }

FUNC(int main,
  int x = 0, y = 0;
  while_(x)
    while_(y)
      printf("hi\n")
  dclose
)

[stray dragon]

i suppose could also just mask the other color's moves, sum the scores of the pieces on your side, then do the same for the enemy's side

[real sierra]

megamind

[real sierra]

my current solution for scoring is kinda lengthy

I think I have a macro popdiff or something

and its along the lines of

[sage crag]

50 tokens

by the way

int x = 0, y = 0; is a waste of tokens

[stray dragon]

my scoring function is uhhh

[real sierra]

#define popdiff(board, piece) count_ones(chess_get_bitboard(board, piece, WHITE)) - count_ones(chess_get_bitboard(board, piece, BLACK))

then for scoring I do

[tender river]

int x, y; x = y = 0; would be slower BUT if you define x/y as globals they are zero by default

[real sierra]

score = 9*popdiff(board, QUEEN) + ...

[sage crag]

int main() {
    int x, y = x = 2;
    return y + x;
}

[tender river]

[stray dragon]

i see

[tender river]

didnt know latter definitions can use previous ones on the same line

[olive sable]

I have to get up early tomorrow. Goodnight

[sage crag]

it because a = b; evaluate to b

[real sierra]

clever way of getting bitboards tho

[tender river]

i know that much i just didnt expect x to be accessible for y's assignment

[sage crag]

@real sierra if you define any variables like int x = 0, int y = 0;

[real sierra]

that factor line is crazy

[stray dragon]

ignore the tiny FFN i slipped in there to deal with position-based scoring

yeah lol

[sage crag]

its uh

the same number of tokens

[real sierra]

what

this works?

[sage crag]

[sage crag]

yes

a = b evaluates to b in c

[real sierra]

crazy

I knew a, b was b

but not equals

[sage crag]

that's why there is a difference between ++a and a++

[real sierra]

I'm still to scared to abuse ++ in expressions

I worry I'll do it wrong

[sage crag]

int main() {
    int x = 0;
    return ++x;
}

returns 1

int main() {
    int x = 0;
    return x++;
}

returns 0

[real sierra]

yeah but they're just such cursed operators

[sage crag]

btw this also works for +=

return x += 3; returns 3

[real sierra]

++x + x++ is undefined behaviour for example

according to the standard

[sage crag]

yes its because its unordered

[real sierra]

I don't like weird operators

order it

[tender river]

not weird

normal

[real sierra]

weird

[opaque sigil]

huh did they not fix this

[real sierra]

?

[opaque sigil]

they were bothered enough to make the eval order of function args ordered at least iirc

[sage crag]

there is no eval order for ++ but you can remedy it by using it in inside function or something

[real sierra]

wontfix

[opaque sigil]

the almighty committee is never wrong

[tender river]

Order of evaluation of the operands of any C operator, including the order of evaluation of function arguments in a function-call expression, and the order of evaluation of the subexpressions within any expression is unspecified

its a good standard

on some platforms left to right is better, on some right to left is better

[opaque sigil]

huh

was it just clang and gcc ordering them

[sage crag]

[opaque sigil]

why does that work

[sage crag]

it doesnt

[tender river]

, is ordered ye

[sage crag]

its a trick

[opaque sigil]

smh

[real sierra]

yeah that's unary +

[sage crag]

wait it does work

i thought it was tricking me by essentially return ++x;

nope

the x++ is the same as x though

for obvious reason

ok but this is cursed

2?

not 3?

[tender river]

why would it be three

[opaque sigil]

you increment and then you increment again

[tender river]

konii

[sage crag]

im adding them together

[tender river]

you arent

[opaque sigil]

no you're not

[tender river]

theres a comma

[fluid junco]

no

[sage crag]

right

comma

ye comma

[tender river]

[fluid junco]

Neurorb

[sage crag]

unary +

unary + so stupid

useless operator

[faint sandal]

https://fxtwitter.com/AzakaSekai_/status/1968005806156288335 rare Linux w

安坂星海 Azaka || VTuber (@AzakaSekai_)

wait... SynthV works straight up on Linux without any installation shenanigans???? HOLY SHIT

**💬 1 🔁 1 ❤️ 6 👁️ 181 **

FxTwitter · Sep 16, 2025, 5:35 PM

[sage crag]

self promo

ban

[tender river]

its useful

[sage crag]

it useful like a chocolate teapot

[faint sandal]

ban this

[stray dragon]

yummy

[sage crag]

t pot

[tender river]

ban this

[fluid junco]

depends on the chocolate tbh

[stray dragon]

hi konii

[sage crag]

no

[fluid junco]

got myself a 99% abomination

it's like eating cocoa powder

[faint sandal]

[stray dragon]

ew why

[sage crag]

azabwaa

[rigid snow]

based

[rigid snow]

unbased

[faint sandal]

bwaazaka

[fluid junco]

but it's like
solid

[rigid snow]

but it good

[real sierra]

[fluid junco]

it no good

[real sierra]

thinking about all the things I'd add to my bot if I had tokens

[fluid junco]

it's contents are: cocoa, low fat cocoa, and a pinch of vanilla

[sage crag]

no tokens

tokens are for me

not for any of you

[real sierra]

I have 1 token

[sage crag]

ye 1 token bot

metaphor for society

[real sierra]

can't even add a statement with 1 token

bonus semicolon

[rigid snow]

is 0 not a statement

oh

[faint sandal]

I have 4 tokens

[fluid junco]

sell tokens

[sage crag]

expression

[faint sandal]

to play a single round of rhythm game at an arcade

[real sierra]

azaka you should join the chess contest

[rigid snow]

classic brain short circuit

[real sierra]

[faint sandal]

wtf is a chess contest

[sage crag]

basically

everyone gets a chess board

and you have to be the last one standing

[fluid junco]

an advanced debuggin session

[faint sandal]

???

[real sierra]

I'm planning to host a chess bot tournament

[rigid snow]

battle royale

[sage crag]

you have wooden chess board

so do everyone else

[real sierra]

bots will have a source code limit of 1024 tokens

[sage crag]

you have to be the last one

[tender river]

[fluid junco]

why is it wooden

[faint sandal]

I can just smack others with the wooden chess board

[real sierra]

I heard you're good with obfuscated or shrunken code

[faint sandal]

easy

[real sierra]

so I feel like it'd be up your alley

[sage crag]

ye that point

why would give wooden chess board if not intending to hit people with

[faint sandal]

idk you should have given a metal one

[real sierra]

do you bludgeon people in baseball

[stray dragon]

food

[sage crag]

only i get a metal one

[stray dragon]

you don't?

[faint sandal]

cheating

[sage crag]

not cheating its in the rules

[fluid junco]

i feel like you shouldn't answer this one chief

[sage crag]

[sage crag]

as in as part of the game or

like me personally

[fluid junco]

how does konii versioning work

[sage crag]

https://cdn.discordapp.com/emojis/1212937424062119967.gif?size=48&animated=true&name=nuhuh~1

that's for me to know and for you to fruitlessly guess

https://cdn.discordapp.com/attachments/1072697081443131476/1377728467025530950/NewLeafAbyss.gif

[real sierra]

yes?

to both?

[sage crag]

not typically

[sage crag]

yes

[fluid junco]

how long does this leaf stay damn I've been here for three days nonstop

[real sierra]

rules didn't say to follow the rules

[real sierra]

week

[rigid snow]

is NewLeafAbyss.gif your favorite gif

[real sierra]

rite of passage

[sage crag]

https://cdn.discordapp.com/attachments/1071784467036913664/1409333118405574826/3dgifmaker76610.gif

[real sierra]

never forget your roots

[fluid junco]

guys remember 3d tvs

have you ever seen one

[tender river]

https://cdn.discordapp.com/attachments/1072697081443131476/1377728467025530950/NewLeafAbyss.gif

[sage crag]

rules didnt mention me specifically

that means im an exception

[rigid snow]

it's 3dgifmaker76610.gif

[fluid junco]

true

[sage crag]

https://tenor.com/view/mejiro-mcqueen-spin-umamusume-anime-girl-cute-gif-26001972

[real sierra]

I'm amazed the Zobrist hashing code actually works

there are more magic numbers in that feature than I've ever used before

[fluid junco]

the name feels like a deeznuts joke

[real sierra]

its literally just random xor ops

inserted throughout the move function

[tender river]

thats like... every hash

[real sierra]

and it mostly worked first try

[sage crag]

ye

zobrist hashing not hashdos resistant smh

[real sierra]

Zobrist hashing fast

array of 700 random numbers

[fluid junco]

It has also been applied as a method for recognizing substitutional alloy configurations in simulations of crystalline materials.

[sage crag]

i can also do that

zobrist hashing isnt special

[real sierra]

it special

you can't hash like Zobrist does

[sage crag]

1

2

3

4

5

[real sierra]

STOP

[hoary lion]

erm

[sage crag]

[olive sable]

To win just make your bot random, and if you lose you go back in time to change the seed. You only have to do it once to do it a million times

[tender river]

[olive sable]

Ok now back to bed

[sage crag]

[real sierra]

this is actually possible in a way

[sage crag]

ye just time machine

[real sierra]

you just play a random series of moves on the local board

[tender river]

[sage crag]

takes about 20 token

[real sierra]

and if you lose you unwind back to start and try again

it's like what I'm doing but with less brainpower

[sage crag]

this is just brute forcing in chunks

[fluid junco]

shouldn't the whole board be like 256 bits?

4 bits per piece, 64 places

[sage crag]

no, because using a uniform representation would be stupid

[real sierra]

the board is stored as 64-bit bitboards, one per piece per color

[sage crag]

its actually more

because its easier to compute with

[real sierra]

^

there's actually a few bonus bitboards too for efficiency

[tender river]

optimizing for space is only useful when you have lots of data

[fluid junco]

yes me stupid

[real sierra]

you not stupid

you can definitely store the board in smaller amount of memory

[fluid junco]

stupid be fast though, 256 bits can compare in one clock cicle

[sage crag]

im optimised for space not because i have lots of data but because im very cheap to clone

[real sierra]

but chess bots are a compute bound task

[sage crag]

[tender river]

how much

[fluid junco]

or 2 if you double pump

[real sierra]

separate boards per piece type make it very easy to generate moves

[tender river]

actually i dont feel like it sorry

[sage crag]

more than you can afford

[tender river]

you dont have to answer

[real sierra]

for example, all pawn moves for white can be obtained by shifting the pawn board north one square

[tender river]

hmm

[olive sable]

I did the math, you can technically fit all 13 pieces for 64 positions in 237 bits

[real sierra]

which is just a left shift by 8 bits under the hood

[tender river]

doesnt sound that cheap

[real sierra]

blazing quick

[tender river]

[sage crag]

cost is relative

wake up chat wake up wake up wake up wake up wake up wake up wake up wake up

[real sierra]

[sage crag]

[fluid junco]

true

[tender river]

eliv queue wake up

[olive sable]

???

[sage crag]

eliv queue goto &&eliv queue

[real sierra]

only possible if the pawns are in their own data structure to begin with having to search for them would be too slow

[olive sable]

Wth is an eliv queue?

[sage crag]

ye i see how this can be difficult for ESL

english have adjective before noun

[fluid junco]

toll free amogus car vc

[sage crag]

the queue is of eliv

[real sierra]

I thought eliv was a noun

[sage crag]

this is also a difficult part of english

nouns can be adjective

[fluid junco]

eliv is the friends we made along the way

[sage crag]

red is a noun referring to a colour

its also an adjective

[real sierra]

nouns technically can't be adjectives but we removed the feature of hyphenating noun sequences

ragey

[olive sable]

Ye wd have that too, i just dont know what you're talking about

[sage crag]

so shiro was never your name

your name was ~ this whole time

[fluid junco]

brains don't brain logically is one of the best threads I've seen on the internet

[real sierra]

bit flip

[sage crag]

ye

bitwise not

[real sierra]

~Shiro ⚫

[sage crag]

[sage crag]

shiro xor shiro

⚫

[fluid junco]

fastest way to zero

[sage crag]

on x86 sure

[tender river]

shiro i have your brain scan

[stark needle]

you deserve to be a plushie

[sage crag]

that's only because the microcode has a built in fast-path for a xor a

[real sierra]

I zero with my eq_0_sc macro

[sage crag]

no, give me

[fluid junco]

well also because you don't have to load no value other than one register

[real sierra]

5 bytes

[sage crag]

also not the fastest way to zero memory of width > register size

[tender river]

what will you use it for

[fluid junco]

memset algos go crazy

[sage crag]

you dont have to loads more than one register with mov variants too

you can also & 0

the only reason its used is convention

[tender river]

it's mostly specialized instructions not algos

[fluid junco]

yeah but you see you have to load that 0 from asm

[real sierra]

my stack computer project

[sage crag]

immediate constants arent loaded from "asm"

[fluid junco]

where did you get all those puters to stack

[sage crag]

they are loaded from cpu cache

[olive sable]

Everytime i see "asm" i read it as my name but then i realize i readn't

[real sierra]

sleep deprivation

[real sierra]

you're asm now

[sage crag]

especially with & 0, it has the same specialisation as a ^ a

the dispatch cost for the instruction is more than any actual work it does

which is why pipelining is good

[tender river]

erms

[sage crag]

erms

[real sierra]

koniiiiiiii

need token

[sage crag]

you cannot have my token

[real sierra]

need token need token

so many feature

[olive sable]

hail the almighty pipeline

[real sierra]

can't fit

[sage crag]

send me code me make smaller

[real sierra]

but if steals then I lose...

[sage crag]

https://cdn.discordapp.com/emojis/1212937424062119967.gif?size=48&animated=true&name=nuhuh~1

[fluid junco]

give me all your trade secrets i can be trusted

[sage crag]

lose anyway

future unchanged

[real sierra]

to my knowledge I have the only actual chess bot so far

other than t's

but idk how T's performa

[sage crag]

i have the api

[olive sable]

Why do you hail something when hail is ice rain?

[real sierra]

yeah I'm not really supposed to have one yet

but I figure I'll be fixing bugs during the contest so

I figured it fair

[sage crag]

it so nice to have

[ebon basin]

a chess NN nostalgic

[real sierra]

I'm excited to see if anyone tried a neural net

[sage crag]

one really simple optimisation you can try is fiddling with evaluation function

i.e. logarithm weighting

or sqrt

[ebon basin]

I treid once to do reinforced NN with chess

[real sierra]

my method doesn't really need an evaluation function, I wonder if I should just throw it out

[sage crag]

try disabling it and playing it against the version that has it

[real sierra]

it only uses it for breaking ties with stalemates

[real sierra]

smort

[ebon basin]

evalation isn't the probleme it the 4096 diffrent moves you can make on a chess board

[fluid junco]

evolution algorythm

[real sierra]

doesn't know about the token limit

[fluid junco]

seems too low a number

[rigid snow]

#neurotic-neurons chess edition

[sage crag]

dumbot

touhou

https://tenor.com/view/bad-apple-manu-touhou-gif-21182229

[real sierra]

someone did this in seb lague's tournament

[tender river]

https://tenor.com/view/touhouplayer-touhou-meme-funny-rain-gif-9335315081749646867

[real sierra]

actually super impressive to compress 2 min of video into a bot

[ebon basin]

it from 64 postion on the board to 64 = 64*64 for one piece but alpha go cut the tree by non valid in the NN

[sage crag]

video black and white, image is 64 pixels

[real sierra]

30fps * 120s

ouch size

[sage crag]

225 kb

[fluid junco]

don't need 30fps though

[real sierra]

assuming you can store max 64 bits in one token uncompressed

[sage crag]

each frame can be 1 bitboard

[fluid junco]

30fps too high, 10 will be enough for everyone

[sage crag]

which is 3600 tokens naively

[real sierra]

max 8192 bytes of data

[sage crag]

if you use run-length encoding im sure you can get that down to basically nothing already

[real sierra]

gives you 68 bytes per second of video

[ebon basin]

Anyway it not my project I want to but Neuro inside of a 3D printed human

[rigid snow]

rle on "video" does not sound efficient

[sage crag]

[sage crag]

do it temporally

[real sierra]

I think their implementation stored the durations until flip for each pixel

[sage crag]

ye seem like easiest way

[ebon basin]

Do you know ONI ?

[real sierra]

then you no need store per frame

[fluid junco]

smort

how to store that though

[real sierra]

byte array

with clever indexing

track time since last flip

[fluid junco]

clever does a lot of lifting here

[real sierra]

load bearing word

[ebon basin]

It a an asteroid simulation from the makers of "Don't Starve " and "Don't Starve Together"

[sage crag]

you can just pack it into a struct

[real sierra]

I'll pack you into a struct

struct konii *

[fluid junco]

linker error sorry

[real sierra]

struct Shiro has a pointer to brain which is always null

do not use

[fluid junco]

doesn't mean no brain though, just have to find it

[tender river]

The original 121 seconds with 30 fps consist of 121 * 30 * 64 = 232.320 bits being displayed.
This gets compressed to 52.873 bits, so it fits in 827 ulongs = 827 tokens.

just uses rle

[real sierra]

[tender river]

[real sierra]

RLE+BWT

repeat until it gets bigger

[sage crag]

you can probably pack a pixel and its duration into a single u16

[fluid junco]

into a u8

[sage crag]

u8 too small

255 frames

3600 frames of video

[fluid junco]

128 frames, one bit color

[sage crag]

there's a better way of doing it than that im pretty sure

but i cant remember the method

[fluid junco]

need to compute if most pixels stay one color for more than 128 frames or not

i think they dont

[tender river]

negative elo

[real sierra]

gotta fit RLE decompressor + video decoder into the rest of the tokens

[tender river]

[real sierra]

one goal

[tender river]

WorstBotEverMade

[sage crag]

pretty easy

is it better to stream the video directly while you decompress or decompress then read from an array

if you are just doing RLE the first is pretty easy

[real sierra]

I think stream while decompress

have to wait between frame anyway

time for compute

[fluid junco]

if it resets every move too much time to think

[sage crag]

it does not

[real sierra]

did bred find API issues or were they rust

I don't remember

[fluid junco]

i couldn't find this morning

[real sierra]

memory

[tender river]

@trim valve

[fluid junco]

i was interested what happened

[real sierra]

what happened with what?

[fluid junco]

i think you just moved topics

[real sierra]

oo

soz

[fluid junco]

with that race condition bred had

or whatever

[real sierra]

ooo

I think I fixed

I forgot to init a mutex

[sage crag]

ye you dont need to do this

[real sierra]

now there new issue I think with double check legal moves

finding too few

[sage crag]

store 64 bits that encode the colours at the beginning of the video

and then the video data can just be the frame length until the colour toggles

then you get the full width of the integer for frame length

[fluid junco]

ooooh

[real sierra]

there so many edge cases with chess rules

[fluid junco]

see storing it is cool

[real sierra]

like how not all attacked squares are legal moves

[sage crag]

not that it matters since its better to use a u16 than a u8 anyway

[tender river]

also have to store per-pixel offset indices

[real sierra]

and not just for check reasons

[fluid junco]

u16 for time, u8 for color

aligning issues 😨

[sage crag]

unnecessary

64 bits only required

for the colours

[fluid junco]

why so many bits though

[sage crag]

64 squares

[fluid junco]

for black/white colors?

[real sierra]

bitmap

[fluid junco]

but can't really use a bitmap here, because all the temporal stuff no?

[real sierra]

yea I don't understand using a bitmap either

I would store array of bytes

[sage crag]

struct Video {
  col_bitmap: u64,
  data: Box<[u16]>,
}

first 5 bits can encode the pixel number

[real sierra]

where byte (i+64*j) stores the ith delay until flip for pixel j

[fluid junco]

arm so cool, arm handles aligment stores and loads magically

although still slow

[real sierra]

then have a 64 byte array of pointers into that, one per pickle

[sage crag]

wtf

[real sierra]

advance pointers as delays expire

[sage crag]

pointer

pointer is 8 bytes

[real sierra]

oops

[fluid junco]

pointer too big

[real sierra]

it fine

[sage crag]

this is the best way i have thought of

[tender river]

factorio

[real sierra]

not many token

[tender river]

video decoder

video data

[real sierra]

factorial chess bot

[tender river]

video

[real sierra]

factorio mod with bindings to chess api

[tender river]

(grey spots are clouds)

[real sierra]

send moves to the chess comparator

[fluid junco]

my brain isn't ready to handle factorio memory bank logic

not talking about video decoding

[tender river]

it supports true color too

thats why there are 2 video circuits in the decoder

with a unified timing loop

[fluid junco]

neuroCrying

[sage crag]

hm

instead of storing the numbers of frames

no that's stupid idea

[real sierra]

store time as floats

[tender river]

[fluid junco]

tell tell

[real sierra]

I wanna go home

I wanna make token

and fix api

I think actually I get many free token if upgrade API

[fluid junco]

and not find a car on its side preferably

[real sierra]

yeah ideally no cars have undergone a rotation

we'll see

[fluid junco]

oh hi bred

neuroWaveA

[trim valve]

yeah your thing was generating weird invalid moves or not seeing moves

[ebon basin]

Okay nearly everything ready 🪛 (will take some weeks)

[real sierra]

did you ever spot an obvious pattern

[ebon basin]

Actually how do I see what mods are alreaedy made for Neuro

[real sierra]

no idea

lol

[tender river]

#1336777692070023271

[real sierra]

they might be pretty far down but yeah

searching that is probably your best bet

[tender river]

tag

[fluid junco]

neuro car in factorio

[trim valve]

not really but a lot of them were to do with nonexistent pawn moves

[tender river]

make a 3d model in blender

[real sierra]

do you think ved would join the chess contest

[tender river]

rest is easy

no

[trim valve]

I might try and do a minimal repro of some in a bit but have to do some other stuff first

[ebon basin]

well Oni is actually also an autism game

[trim valve]

here is the output of 10k games

[sage crag]

only if he can make a stream out of it

[fluid junco]

modify the game sprite a bit, sprinkle some pink

[ebon basin]

I'm scared to look how many hours I have on it

[tender river]

[real sierra]

[trim valve]

search for illegal move: and it should tell you exactly what move prompted the game to explode

[real sierra]

@real sierra note to self investigate this pgn

[trim valve]

everything before the curly braces is the pgn leading up to it

you can kinda just ignore the square bracket stuff, its just metadata

[fluid junco]

Oxygen Not Included?

oooh do you want to make a neuro crewmate?

[real sierra]

I love oxygen not included but I always get to the same point in the game and stop playing for some reason

I think I just solve all the immediate challenges and then run out of things to do

[ebon basin]

yes and let Neuro connect to it

[fluid junco]

seems a bit too complex with all the options for llm

?

go place a block, pick up resources, run on a wheel, etc etc etc

[ebon basin]

I let her control the priority queue for the task and set her own schedule

[fluid junco]

she'll be runing out of air like every two minutes

[ebon basin]

Only if she set free time while in carbon dioxide and idle there

[fluid junco]

minecraft ran somehow though

[ebon basin]

But it is chaos she won't go to the toilet or to her bed etc

[fluid junco]

idk seems fun at least

even if won't work

[real sierra]

I hunger for chess

[ebon basin]

What is also fun on this are the reversed roles Vedal plays as the player the almighty AI and Neuro get a taste what a human is.

[chilly maple]

neat im happy he's so keen to share specs

admirable
Ive been working on my own vision LLM (ofc lora training) trying to go for the highest token output .

Been quantizing my own gemma Abliteration :)
Running it on 2x 3090.

not with twitch in mind though. I started off with it being a "it joins discord calls and remembers voice/accents and tries to isolate out individual users.

i basically have whisper-x, and TTS all running on the 2nd gpu (for vram reasons) to not shard.. then running an EXL2 to have a higher token output for faster responses

[fluid junco]

a giant of a project for a discord bot
TWO WHOLE 3090 DAMN

[chilly maple]

haha yeah

it started with a 3090

but i was having so much fun i wanted it to grow

originally it was a old 9700k as well but i was like.. lets upgrade mobo etc etc.
I have no game agents or anything that front is new to me..

My idea was really just.. hey it can join friends and give people a great day. That part seemed to work so far!

Originally i was even driving voice inefficiently using RVC but yeah i wouldnt wana do that anymore!

[fluid junco]

that's great

[chilly maple]

i can post it somewhere. let me see if there is a channel that lets me

[stark needle]

That's crazy how many tokens are you pushing rn

[chilly maple]

Let me check! I mentioned it before but Again this is from memory it was in the near 300's?

per second

[stark needle]

Wtf

That's crazy

[chilly maple]

the original one was very like.. bad because i wasnt streaming in and chunking tts properly and like.. it caused gaps in response

i dont wanna make a project. any place i can dump a conversational video?

[stark needle]

If you keep talking you'll gain level and embed perm

Idk how long but not that long

[rigid snow]

you can still send links

it just won't embed

[chilly maple]

ok ill do that! moment i posted it somewhere

https://cdn.discordapp.com/attachments/996911628677284001/1407116710154342582/2025-08-18_23-35-34.mp4?ex=68ca81ce&is=68c9304e&hm=82766e9cb3692fd207204bdeaf52e884c285d9b27667460ee1fada71b40db72f&

doesnt embed.. oh well but yeah remember this was BEFORE proper chunking and streaming of TTS. voice needed work too but its getting better now

[fluid junco]

good latency also

[chilly maple]

i started properly retraining!
Latency should be far far far lower

[fluid junco]

does it do smart stuff like internet search

i guess it does kv for remembering users

[tender river]

two leafs so confusing

https://cdn.discordapp.com/attachments/1072697081443131476/1377728467025530950/NewLeafAbyss.gif

[fluid junco]

you a leaf too

[chilly maple]

I'm using gemma 3 but abliterated. and then i will re-add restrictions based on people being bad to it and encouraging it to be bad.

Gladly im using discord so im not under scrutiny!
yes! Ive started turning to more assissant hybrid so its also a lil bit more useful!

[tender river]

https://cdn.discordapp.com/attachments/1072697081443131476/1377728467025530950/NewLeafAbyss.gif

[rigid snow]

wait that's crazy

[fluid junco]

crazy good

[rigid snow]

so others don't have to download

[stark needle]

That's crazy

[rigid snow]

[tender river]

you said latency is better too now

[stark needle]

Chat

[rigid snow]

shadow

[stark needle]

I need your opinion

[rigid snow]

[chilly maple]

ive been working on it off and on for a month i believe

[fluid junco]

damn you're good

[chilly maple]

constantly migrating because the AI toolset is changing so fast and it leaves me wanting more

[tender river]

@sage crag lawsuit target

[sage crag]

ye

[chilly maple]

im def not wanting to capitalize on it. i work in entertainment so making people happy is kinda my job and i see it can give people some fun so

[sage crag]

lawsuit

[chilly maple]

maybe i should eventually make it not believe she is korone and just give it her own backstory system prompt

[stark needle]

Capitalism my beloved

[chilly maple]

it has a vector db and a deep memory! pc has 64GB ram but i can expand it to 96gb like my other pc's

not !

i didnt mean now lol

there.. how 1 typo can completely change the message lol!

im just running lots of freelance jobs to like.. maybe allocate more money to it as it grows.

[stark needle]

Now get her to sub 1 second latency

[chilly maple]

she can! the issue at the time was about the chunking of TTS

[stark needle]

Fair enough nice

[chilly maple]

the vision is also really really fast. about 50~80 milisecond per snapshot it generates tokens.. altho like token generation slows down in vision to probably a low 50's ? im not sure i havent benchmarked

i learned it to recognize humans and remember their names

but yeah most people use GGUF for their AI.. i use exl2 but not many people quantize EXL2~3

EXL2 relies on a full vram load of an LLM but because of it its 9% faster then any GGUF

(of the same model)

[stark needle]

Yea ur the first one i saw that doesn't use a shitty setup

[chilly maple]

its kinda funny i went from .. anti AI to kinda falling in love with it some months back and been studying up hard hard hard during my breaks

mine can break out in spontanious singing btw

during sentence!, laugh and cough too!

[sage crag]

typedef struct {
    uint64_t bitmap;
    uint16_t *delta;
} Video;

void video_unpack(Video *video, uint64_t *bitmaps) {
    Video vid = *video;
    bitmaps[0] = vid.bitmap;
    uint64_t frames[64] = {0};

    for (uint16_t delta; (delta = *vid.delta++);) {
        uint8_t px_idx = delta & ((1 << 6) - 1);
        uint64_t frame = frames[px_idx] += delta >> 6;

        if (vid.bitmap ^= 1 << px_idx) {
            bitmaps[frame] |= (1 << px_idx);
        } else {
            bitmaps[frame] &= ~(1 << px_idx);
        }
    }
}

i wrote the decompression algorithm

i think

[chilly maple]

im still figuring out how i wanna deal with being interrupted though..

[sage crag]

[fluid junco]

we simple people, we forget to init variables

[chilly maple]

used 3090's are kinda cheap here... relatively

[stark needle]

Wtf where do you live

Here they are like 700 used

[chilly maple]

its when you get in the 5090 territory it becomes expensive.. but the issue is well... 32gb vram 2x.. its a boost but for the cost..

Netherlands. Only catch: Power costs are high

[fluid junco]

Tell me the street you grew up at

[chilly maple]

but i have a large home solar farm so i can sustain mine using it

[fluid junco]

damn bro

[stark needle]

Why

[rigid snow]

joke

security question

[stark needle]

tax agency doing ANYTHING to fish information to tax me more

[chilly maple]

Actually since im here.. does anyone know more about the game agents?
I found one for minecraft ofc.. i think neuro probably uses it too since it moves in a similar fashion.. but i found projects like lmgame..
I didnt like the games they have trained on though

[sage crag]

@real sierra

void video_unpack(Video *video, uint64_t *bitmaps) {
    Video vid = *video;
    bitmaps[0] = vid.bitmap;
    uint64_t frames[64] = {0};

    for (uint16_t delta; (delta = *vid.delta++);) {
        uint8_t px_idx = delta & ((1 << 6) - 1);
        uint64_t frame = frames[px_idx] += delta >> 6;

        if (vid.bitmap ^= 1 << px_idx) bitmaps[frame] |= (1 << px_idx);
        else bitmaps[frame] &= ~(1 << px_idx);
    }
}

[chilly maple]

so its something im exploring.

[sage crag]

i do not care to try it

should work

[stark needle]

Only thing i was looking at was RL agents

For a while

[chilly maple]

things get messy when training.. training outside of loras is where things actually get expensive.

[sage crag]

why did i accept *video