programming

[sage crag]

cpu throughput with simd is much higher and makes things like ML viable

[real sierra]

well you know the contest rules and the intended compilation command so

feel free to use the knowledge you have to optimize your code

just dont optimize it so much that it breaks on my system or you wont get many points

[sage crag]

note that shiro wont actually be compiling the programs with custom cpu

[opaque sigil]

can you add -march=native to the command

[sage crag]

its better to specify, shiro's cpu has avx2

[opaque sigil]

yeah but there's so many flags

[rigid timber]

actually Shiro, can I confirm something with you? Deno has its standard library as hosted packages not included in the runtime itself, that's fine to use right?

[opaque sigil]

easier to just enable everything supported

[real sierra]

wtf is deno

[rigid timber]

a JS runtime that has a non-ass FFI

[sage crag]

i would argue that its unfair for you to be able to use webgpu just because your stdlib has it

[rigid snow]

good answer

[real sierra]

deno seems to be an extension of js so i feel like it should be restricted to the js stdlib no?

[rigid snow]

oh you sweet summer child

[opaque sigil]

what stdlib

[rigid timber]

js doesn't have a stdlib, it depends on the runtime

[rough bloom]

"JS stdlib"

[real sierra]

[sage crag]

https://jsr.io/@std

[opaque sigil]

deno is the stdlib

[rigid snow]

ok so

[real sierra]

what is wrong with js

[rigid snow]

we shoudl just do node right guys

[rigid timber]

so much

[trim valve]

what isn't wrong with js

[sage crag]

please read this for a full list of the deno stdlib

[rigid snow]

please don't read this

[real sierra]

invoking gpu would still break the parallel processing rule

regardless of being in stdlib

[sage crag]

its only 41 entries

[rigid snow]

it may as well be one entry

deno bad

because i said so\

[sage crag]

what if the parallel processing is opaque and hidden behind a function that appears to be single processed

[rigid timber]

tbh I can totally use only the actual JS-language-only stuff without anything from the Deno stdlib, I just wanna know before I start writing anything

[sage crag]

cause that's how js runtimes work

[rigid snow]

should i immerse myself into the hell that is node api for node c++ addons or whatever the fuck they call it

i'm so ready

[rigid timber]

have fun with that

[sage crag]

most of that stuff looks pretty useless for chess at first glance anyway

[rigid timber]

ehhhhh the caching utilities could be nice

[real sierra]

building these bots for the tournament is going to be a nightmare huh

[sage crag]

or, i mean, most of it wont give you a special advantage

[rigid timber]

the data structures could also be nice...

[sage crag]

you put it in the rules that it must be easy to compile right

[real sierra]

i did put that it should be easy to build on my system

[sage crag]

just disqualify anyone that fails to meet that

[real sierra]

so i guess if i have any problems i just ask bindings devs to fix them

[rigid snow]

it may be pain but is the best thing we can do for compat

[opaque sigil]

require docker images

[rigid snow]

requiring deno is.... yeah

[sage crag]

deno

bun

v8

[opaque sigil]

deno is just a single static executable you'll be fine

[sage crag]

...javascript...

[real sierra]

i mean

is there anything actually problematic in the deno runtime

[opaque sigil]

not really

[real sierra]

that the other rules dont already disqualify

[rigid snow]

it's just obscure-ish

[rigid timber]

well the usual network and filesystem stuff, but that's banned anyway

could explicitly ban the GPU if you want to

[rough bloom]

should've done wasm with wasm submissions instead
but yes, in practice you're gonna look over everything anyway, just clarify as needed

[sage crag]

not hugely

[opaque sigil]

you can ban those within deno too

[rigid snow]

friction for potential contenders

[sage crag]

in appearance

why did we go the c bindings route again

why did i suggest that

[real sierra]

yeah i think its fine to use the deno runtime then

[sage crag]

we should have limited it to c only

[rigid snow]

and node bindings will work across all runtimes

[sage crag]

[rigid timber]

they can just use the C bindings if they want less friction

[real sierra]

just for clarification tho, the gpu stuff is indeed considered parallel and banned

[rigid timber]

yeah, understood

[sage crag]

also async

[real sierra]

i think it fine

[sage crag]

[rigid timber]

single-threaded tho

[rigid snow]

js single threaded

[sage crag]

also this

[rigid timber]

that's under the usual banned network stuff

[rough bloom]

not parallel, just concurrent

[real sierra]

honestly im rather impressed with how many languages there are bindings for

does async start an event thread

[sage crag]

the rules say "otherwise run code in the background"

async is background execution

[rigid snow]

no it isn't if the call is blocking

[rigid timber]

it schedules tasks on the current thread... how the engine manages it is its thing

[real sierra]

if it was just single-thread coroutines that'd be fine

[rough bloom]

dunno what you mean by "event thread" but it does not start an OS thread AFAIK

[sage crag]

single threaded coroutine is fine ye

[rigid timber]

async is single threaded unless you spawn a thread yourself through the runtime's APIs, that's defined as part of the JS standard

[rigid snow]

it doesn't. the only thing spawning os threads in js is workers

manually

[real sierra]

sounds fine to me then, at least by this description

[rigid timber]

tbh async wouldn't be helpful anyway

since you'll spend most of the time searching moves

[real sierra]

i could see a producer-consumer pattern for producing move sequences and consuming them to evaluate, would make time management easy

[sage crag]

can still be useful for divide and conquer search

if search tend to terminate, the more searches you make the faster you will find non-terminating searches

[rigid timber]

possibly... maybe someone will come up with something nice

[sage crag]

therefore async is actually useful

[wide flicker]

I'm not even a participant in the contest, and lurking over the last few days has been... fascinating

[real sierra]

this is definitely manageable for me alone to run

( )

[rigid snow]

a bot will not be thinking in between moves because chess_done() is blocking. for a bot to be thinking in between moves the bindings would have to call chess_done() in a separate thread manually. js is very single threaded by default, "async" only refers to calls to native code

or interrupts

or whatever

single threaded

period

[sage crag]

async doesnt typically have any meaningful gains unless you are doing IO anyway

[real sierra]

async sounds fine

[sage crag]

it is useful for long-running and early-terminating searches though so i think its a good idea to use it

you implicitly use it when typically programming searches though

[opaque sigil]

since when is this a thing on twitch

[real sierra]

youtube 2

[stray dragon]

should have always been a thing

[opaque sigil]

better late than never i guess

putting this behind being a sub is crazy though

[rigid snow]

announced at twitchcon

[opaque sigil]

ah

[rigid snow]

rolling out now probably since it says new

[rough bloom]

never heard of this, but the functionality itself has been a thing for ages as long as VODs are enabled

[opaque sigil]

apparently it being subs only is only for now

so that's nice

[rough bloom]

the current stream is also a VOD with a lag of a few minutes

[opaque sigil]

mhm

[rigid snow]

i mean i always just clicked the clip button to do that

[desert wave]

it was announced at the last TC EU opening ceremony

[lapis wraith]

Real time VODs like youtube?

[opaque sigil]

yup

[lapis wraith]

why they put this behind a paywall?

[onyx solar]

So the paywall is worth

[opaque sigil]

only for now apparently

[lapis wraith]

the is strong with this one

[opaque sigil]

sth sth they want to make sure it doesn't get used to bypass ads

which idk how that'd even work but fair enough i guess

[midnight sigil]

to encourage ppl to sub to a channel I guess

Twitch is losing its money cuz no one is paying for their services. Youtube on the other hand has Google at the back

[jagged turtle]

isn't twitch backed by amazon or is amazon also going in the dumpster fire rn

[midnight sigil]

big corpa are weird

[rigid snow]

amazon is twitch's parent company but that doesn't mean they are one entity that's not how any of that works. same reason they pay astronomical aws bills

different entities, different bank accounts, different incomes and expenses

[jagged turtle]

business relations are weird

oh well

[rigid snow]

1 hour spent fighting msvc instead. it won

[jagged turtle]

[opaque sigil]

[rigid snow]

oh my fucking god it's impossible

i can't figure out a way to make it compile c as c

[jagged turtle]

???

[rigid timber]

lmao

[opaque sigil]

you mean c++ as c right

[rigid snow]

no

[jagged turtle]

me when i can't compile to the same language i wrote it in

[rigid timber]

lemme guess, it really wants to compile as c++

[rigid snow]

yes

exacty

[jagged turtle]

💀

[opaque sigil]

isn't v8 c++ only or am i missing something

[rigid snow]

it shoves /std:c++20 into the c dep

[rigid snow]

chessapi isn't

[rich ibex]

least disruptive pr

[rigid timber]

actually can't you compile chessapi as c++? doesn't it have some c++-aware stuff in there?

[opaque sigil]

ah there are some issues

it won't compile

[rigid timber]

that's a pretty big issue

[opaque sigil]

some implicit conversions

yknow what i may as well fix this and make a pr

[rigid snow]

that's fine in c though and should not be an issue if you compile it as such

i also got errors in threads.h of all fucking places

[opaque sigil]

c++ is stricter when it comes to pointer types and type conversions

you have to explicitly cast void* to whatever pointer type you want

[opaque sigil]

what errors

[rigid snow]

not anymore nvm

[rigid snow]

also ints to enums are implicitly cast

[opaque sigil]

yup

@real sierra i made a pr so the chessapi.c compiles with c++ compilers too, feel free to merge if you want to (it's just a bunch of type casts)

[rigid snow]

Someone else would have to test with msvc if they care
i'm running into the threads.h thing again

#ifdef __cplusplus // TRANSITION, [[_Noreturn]]
[[noreturn]] void __cdecl thrd_exit(int _Res);
#else
_Noreturn void __cdecl thrd_exit(int _Res);
#endif
``` basically i somehow end up on the `_Noreturn` macro even though i'm compiling as cpp

i don't even know at this point

[opaque sigil]

[rigid snow]

static inline int __cdecl thrd_sleep(const struct timespec* duration, struct timespec* remaining) {
    return _thrd_sleep64((struct _timespec64*) duration, (struct _timespec64*) remaining);
}
``` oh and it doesn't like this too

is defining functions in headers even ok

in cpp

[opaque sigil]

it is

there's nothing special about headers they're just a convention

[whole tulip]

they're special to me!

[rigid snow]

C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.41.34120\include\threads.h(116, 99): error C2143: syntax error: no ";" before "{" i had to translate so the text of the error may not be exact. my msvc is in russian for some crazy reason even though the system language is english

oh the vs install is probably in russian because at some point i downloaded it from a russian ip

that would make sense

[opaque sigil]

i can't believe i'm unironically trying to install visual studio via wine now

what has my life become

[rigid snow]

why

don't

[opaque sigil]

too late

[rigid timber]

does this classify as self-harm?

pretty sure it does

or at least should

[opaque sigil]

also speaking of wine

i found out the other day apparently if you install specific combinations of packages in the same prefix it can kill the kernel

[rigid timber]

lovely

[amber fractal]

I remember trying to install software on wine only to have to try and install it

I swear I was trying to run python as windows for some reason

I remember so many errors and giving up on that

[opaque sigil]

[jagged turtle]

windows huh

[dusky crater]

bro, why, no god no. Whatever sanity you've got left will disappear.

Also, you can technically just run msvc directly via wine... but don't. VM it

[opaque sigil]

[dusky crater]

I tried the MSVC package in the AUR once. The errors were something to behold

[real sierra]

not breaking? won't make development painful? cross-platform compatibility?

[opaque sigil]

it's literally just casting the return value from malloc and casting some ints to the enum types

[real sierra]

[opaque sigil]

there was another change i made about some scanf's return value being unused but i removed that from the pr

[real sierra]

I fear the merge conflicts my other branch will suffer

[opaque sigil]

[real sierra]

is it even worth resolving or do I just reimplement it all

[opaque sigil]

the pthreads only one?

[real sierra]

it's pthread and threads

supports both

[opaque sigil]

right

[real sierra]

unsure if it works or not tho

was considering expanding it to just use win32 threads as well

[opaque sigil]

at that point you may as well

[real sierra]

because windows just doesn't support anything else reliably

it means testing will be a pain in the ass

[opaque sigil]

mint was hitting some weird errors with msvc in c++ mode around threads.h too apparently

[rigid snow]

the errors are fucking nonsense

i’m heading to sleep gn

forget about the errors

[real sierra]

supporting pthread and win32 threads should be enough for all platforms

mac and linux should both be POSIX compliant

and win32 is... yeah...

future problem anyway

fixes may come tomorrow as today has proven to be quite busy irl

I can review some PRs tonight but probably not write code myself

[tender river]

its fine to update it i just used 8 because shiro had it on ubuntu

[sage crag]

awa

[silent cloak]

Support DOS API now

[opaque sigil]

okay i give up with msvc

[silent cloak]

No LLVM or clang?

[opaque sigil]

i was trying to get msvc working under nixos to debug some things

[silent cloak]

Msvc is notorious for giving shit errors

[opaque sigil]

the installer doesn't like my wine

[sage crag]

remove words 2, 3, 5, and 7

wait, only 3, 5, 7

[silent cloak]

Lol

It is certainly one of the ones people tend to avoid

[sage crag]

how about 3, 4, 5, 7

[silent cloak]

I used to use MinGW as an alternative and now on the rare chance I use C++ ill use clang

[tender river]

[silent cloak]

The real legends just ask chatgpt to compile it

[sage crag]

noom emacs

[silent cloak]

I should try that tomorrow actually

[real sierra]

i just install every compiler

try all of them

[silent cloak]

Make a shell script to build

[tender river]

they have a separate type for that (std::bitset) but it's fixed size only

i know

these are actually 18 bytes per token, you can use "\x57\x1e\xab\xb8\x64\xe9\x9b\xc1" though which is 8 bytes

[stray dragon]

i'm killing you

[silent cloak]

Make a program to calculate the best method

[tender river]

me but nothing changed for me since i never programmed chess bots

sam misinfo

[stray dragon]

i bet you have a variable name over 8 characters

smh

rulebreaker

[silent cloak]

C devs storing everything under single letters

[tender river]

variable names dont encode info

the rules explicitly disallow using them for storing data

[stray dragon]

shiro said no issues i trust shiro

[tender river]

shiro wrong

[stray dragon]

anything but making it as horrifically unreadable as possible

[tender river]

[silent cloak]

Sorry pal u gotta program like it's the 90s

[tender river]

@real sierra please confirm a string literal with 18 characters is too long

[stray dragon]

i'm gonna cry

[opaque sigil]

[silent cloak]

Everything must be single letter with no context or help to anyone reading it

[stray dragon]

my 12 tokens that i save with this method

gone

evaporated

how could you

[real sierra]

sounds like a literal that stores more than 64 bits

[tender river]

t rewrite

[stray dragon]

NOOOO

[silent cloak]

Rip bozo

[stray dragon]

make less exploitable strings

you are the worst

[tender river]

me the best

[stray dragon]

NOT ALLOWED

BANNED

YOU CANNOT USE THIS!!!

[tender river]

this is allowed as it doesnt encode info

[silent cloak]

Doesn't want hashing

[sage crag]

NeuroopaTalk

[stray dragon]

i was only getting 64 bits of info from each of those strings

man.

[silent cloak]

Make an array of the string then

[tender river]

skill issue storing 8 bytes in 18 bytes

[silent cloak]

Char array

[sage crag]

nyanleaf

[stray dragon]

how do i even convert that to ulong...

[sage crag]

why dont you build the string token counter into toknt

[stray dragon]

i guess there's that other way

[tender river]

@sage crag this is why

[real sierra]

i think the reason was effort

[stray dragon]

but i don't like it

mmmm

[sage crag]

that way people can use strings without getting undercounted

[sage crag]

ye lazyleaf

[tender river]

efficient

[sage crag]

if people have issue they can complain

then they can submit pr

[tender river]

they can complain even if they dont have issues

[sage crag]

i use a python script

i would share it with you but i dont need you getting an edge

[tender river]

not a bad idea to make includes one token but context sensitive lexing

[stray dragon]

you've already leaked your secrets

[sage crag]

[stray dragon]

it's too late

[sage crag]

[stray dragon]

i have an option

i do wonder what adding python to this shit would let me do.

probably not that much actually

nevermind fuck that

[real sierra]

compress the weights and extract at runtime

[stray dragon]

i'll just go with the annoying option of having c*mmas

[stray dragon]

shiro

that is

literally

what i do

[real sierra]

tokens...

[stray dragon]

each one of those ulongs is 38 parameters

[real sierra]

[stray dragon]

could make it 40 per ulong but would actually cost extra tokens to do that

[magic drum]

Trying to compile the C version of the chess api on Windows, I am running into "threads.h" being missing. Is there a recommended route to avoid this? The change log claims that as of a couple days ago Windows should be working but Mac OS still has this issue, but I'm not on Mac

[stray dragon]

use clang

that's the method we've found

[real sierra]

known issue unfortunately

[stray dragon]

damn it i thought i was so clever putting the ulongs in strings and doing this to extract them

included in stdlib as well

[tender river]

[magic drum]

clang also says it's missing threads.h, is there something else needed on top of that?

[stray dragon]

uhh

[tender river]

what version

and did you pass -std=c23

[magic drum]

version is 21.1.1-1, and no I did not pass that

[stray dragon]

he's superior

how

[opaque sigil]

Old ass version

[stray dragon]

literally what it installed with the visual studio installer

few days ago

[opaque sigil]

Yeah from 2022

[stray dragon]

bruh

[opaque sigil]

The latest stable version is 21.1

You'll have to download it separately if you want it

[stray dragon]

i am making the executive decision to ignore it

probably fine™

[opaque sigil]

It's fine

Well assuming shiro doesn't use any weird c23 features

[tender river]

shiro does

[magic drum]

What is -std=c23 supposed to do for me? Didn't seem to make a difference

[stray dragon]

i'll test on linux in the proper environment when i feel it's about ready

[tender river]

but they were unofficial features before c23 either way

[stray dragon]

removes implicit integers, mostly (from what i've seen of it affecting people's code)

[tender river]

how did you install clang

[magic drum]

pacman -S mingw-w64-x86_64-clang in MSYS2

[tender river]

try gcc in msys2 maybe and of course pacman -Syu

[magic drum]

I tried gcc first, switched to clang because it might have threads.h

[tender river]

[real sierra]

https://www.reddit.com/r/C_Programming/comments/1gcr1jx/c11_threadsh_is_now_well_supported

[opaque sigil]

Guess they don't ship that for mingw huh

[real sierra]

random relevant reddit post?

maybe has answer?

[tender river]

try this https://github.com/llvm/llvm-project/releases/download/llvmorg-21.1.2/LLVM-21.1.2-win64.exe

if all else fails wsl will just work

[tender river]

i mean you can rewrite again using windows.h and CRITICAL_SECTION if you want

[real sierra]

just support one more threading method

it'll fix everything just one more

[opaque sigil]

Honestly just accept your fate and make a layer on top of pthread and win32 threads

[real sierra]

that's probably what's coming yeah

[tender river]

if only there was some sort of header in the standard that allowed you to write cross platform threading code

[opaque sigil]

Yeah if only

[real sierra]

"everyone supports threads.h!"

nobody supports threads.h

🪟 "nuhuh we dont wanna use stinky pthread"

🍎 "nah we dont like threads.h"

🐧 👍 "use whatever idc"

[tender river]

🪟 threads.h is supported or isnt idk we havent decided it depends on your compiler

[opaque sigil]

Sometimes I wonder how much pain could've been avoided by using <thread> instead

[tender river]

std::thread is well supported

[opaque sigil]

Shouldn't be too bad to make a c header and the implementation in c++

[real sierra]

c++

[tender river]

i love c++

[opaque sigil]

Shiro you can just treat it as c with actual cross platform threads

[real sierra]

can i..?

[opaque sigil]

Yes

[tender river]

[real sierra]

example me

[tender river]

rename .c -> .cpp

done

its now c++

[real sierra]

but if i go down the slippery slope people will make me rewrite the entire thing in cpp features

[opaque sigil]

Well your current version doesn't compile as c++, hence my pr

[real sierra]

i dont know enough cpp for that

[tender river]

and you will have given every bindings author an aneurysm

[real sierra]

ill just solve it myself

[tender river]

people havent made you rewrite the horrible repetition you have in move checking with macros

[real sierra]

i just need to abstract the synchronization primitives with clever macros

and provide a windows.h derived impl

and a pthread.h derived impl

and then i win

[tender river]

that reminds me

in windows 11 they changed the layout of CRITICAL_SECTION

they didnt make it bigger

but they started using more of the bytes in there

which broke some legacy code

[real sierra]

idk anything about windows.h

i dont know if i want to

[tender river]

i only know from reverse engineering windows code that CRITICAL_SECTION is windows mutex

idk how you'd do cvs

CONDITION_VARIABLE

how convenient

[real sierra]

surely its more abstracted than this

[opaque sigil]

wdym, that's identical to the other version

[real sierra]

there is one part i kinda dread

which is the typing for the function the thread runs

DWORD WINAPI MyThreadFunction( LPVOID lpParam )

like what is this trash

[tender river]

WINAPI is more or less __stdcall

(which just means uh, not __cdecl)

LPVOID is void*

tokens

and DWORD is just 4 bytes (unsigned)

[opaque sigil]

16 bit words

[tender river]

it stands for double word

[real sierra]

this is silly to have to include in code tho

[opaque sigil]

Isn't that signature identical to the one in pthread_create

[tender river]

even though modern words are not 16-bit

[real sierra]

why cant the windows compilers abstract it all

[opaque sigil]

Because yes

At least their docs are pretty nice

[tender river]

LPVOID stands for long pointer (i think) but on modern platforms no difference between short and long pointers

its kinda incredible how far back winapi dates

linux just relies on source code to update programs for new library versions but windows like actually kept the api mostly stable ever since the 16-bit days

[opaque sigil]

For better or worse

[tender river]

its not an issue with the kernel, "just" with the entire ecosystem around it

and honestly i dont really mind it

i dont like every windows app shipping its own dlls

[magic drum]

What's wsl? Sorry for all the basic setup questions, I haven't done C on my PC before (obviously), but Python C and C++ are the only available languages that I have experience with so it seems compiling the c version of the api is unavoidable

[tender river]

windows subsystem for linux

its just a way to have a linux environment on windows

[tender river]

it seems like shiro will also implement winapi support soon so if you don't wanna bother you can just check back in a couple days

[real sierra]

as an addendum to this, it's likely (though not set in stone) that i will push the submission deadline to accomodate the bugs that people have encountered at the onset, so no need to feel too pressed for time :)

-# (oops sorry for ping)

[tender river]

its ok

[real sierra]

I'll ping you with an interesting question to make up for it

what is the most portable "hello world" possible

[tender river]

are there language restrictions

[real sierra]

nope

[tender river]

and how is portability defined, binary or source portability

[magic drum]

I'm not stressed about a time limit (I know regardless of how much time I have I'm not going to produce anything that competitive, I'm just doing this for fun and experience). Trying to solve all these problems is an adventure of its own I guess

[sage crag]

i so tired

[real sierra]

binary

[sage crag]

not slep

no wanna

[tender river]

sleep sleep sleep

[sage crag]

need

[real sierra]

like, can you polyglot the binary to run on multiple architectures

[sage crag]

code

[tender river]

bed comfy

[real sierra]

wanna write code for me

[sage crag]

sometimes you can polyglot a binary

[tender river]

you can ✅

[sage crag]

sometimes not

[tender river]

https://github.com/jart/cosmopolitan

[sage crag]

no

me busy

[real sierra]

[sage crag]

wait maybe

what code

[real sierra]

no it ok

u busy

[sage crag]

not that busy

just writing spec

[tender river]

[sage crag]

smog

[real sierra]

@stray dragon wanna exchange trade secrets

I think if we share C code hacks

we can mutually benefit

despite my extensive research, most online C code golf knowledge is for source length, not token count, and often runs counter to the goal

I think the real token saver knowledge is all bespoke

[tender river]

i'm working on making c toknt count include names as 1 token

[sage crag]

&&

[tender river]

||

[real sierra]

free 2 tokens

[stray dragon]

when i have tips to share i will share them but currently i am procrastinating a bit

[tender river]

can just replace <> with ""

[real sierra]

can I just send you my code and you shorten it for me

[sage crag]

terminating xor

^^

[stray dragon]

lmao

[tender river]

[sage crag]

!!

[real sierra]

I think you are pretty locked into using a different method than I am

[sage crag]

terminating not

[real sierra]

so I think there's minimal code theft risk

[stray dragon]

ah, that's the issue

[tender river]

🍞‼️

[stray dragon]

i'm procrastinating to unlock myself

[real sierra]

hiding my code from t

you don't want this it's so many tokens

it even performs badly

[sage crag]

[real sierra]

yeah you don't want my code

[sage crag]

koni macro abuser

[real sierra]

[stray dragon]

i'm sure it does

i'm sure it performs like shit

[tender river]

not abuse just correct

[sage crag]

#define fl_bytecode(x) x(tex, 0, 0) x(evc, 1, 0)

[real sierra]

(it was actually pretty good until I introduced the latest bug)

[sage crag]

hlt is so boring tex is better

[real sierra]

along the critical line it usually hits depth 6 ish I think

[sage crag]

🐛

nopy

nopy

[stray dragon]

damn ok

[real sierra]

🎷

[stray dragon]

so all i gotta do is hit depth 8

got it

[real sierra]

what's your time management like t

[sage crag]

its easier with alpha beta pruning

than mcts

[real sierra]

your bot's*

[stray dragon]

life skill or in the bot

[sage crag]

[stray dragon]

oh

because the answers in order would be "horrendous" and "i still haven't actually run it with my code yet"

i think pretty good though?

[real sierra]

you haven't run your code yet?

[stray dragon]

nope.

procrastinating

i have to rewrite it all into the bot

make sure it all compiles and follows the rules i didn't consider

[real sierra]

I see

[stray dragon]

will have to edit a lot, and i also want to redo the search algorithm

add better pruning, instead of just doing "remove a fraction of the least promising paths" every step

maybe, if i have space, add in a transposition table, so i can cache the evals

[sage crag]

#include "stdint.h"
#if __STDC_VERSION__ < 202311L
#include <stdbool.h>
#endif

[stray dragon]

would probably take a lot of fucking around to shave down to that point though :(

[real sierra]

TT very cheap in tokens

[safe path]

[tender river]

you can just include unconditionally

[real sierra]

it's just a problem for me because it turned my tree into a graph

[stray dragon]

but also only saves me like

maybe 2/3 of searches

[real sierra]

the notion of "parent" node no longer applies to my game tree

[stray dragon]

probably?

[real sierra]

not really

literally any board where you have two moves that can be done in either order

TT cuts that work in half

[stray dragon]

but like

it's only half

[sage crag]

no compiler warning

[real sierra]

thats because it's only depth 2

it turns out its really easy to repeat positions you've already done

[tender river]

unused header

[stray dragon]

also: my method prunes a fraction of the trees at each pruning step

it actually would be about half or so between the bot's turns

a bit more from doing the multiple searches within each turn

so i'll say 2/3

still significant though

actually... within endgame, cache hits would probably increase

hmm

[real sierra]

earlygame and endgame are where i see the most cache hits

and while it might seem insignificant to worry about opening moves, if your hash moves are strong enough you might even be able to cut down on time spent on the opening

having extra time later could give you a win even against a better bot

[stray dragon]

that's actually relieving tbh

[real sierra]

sorry i think i phrased that poorly

earlygame and endgame are where i expect the most cache hits

[stray dragon]

ah

[real sierra]

i think i wanted to say "could see"

load bearing qualifier

[stray dragon]

earlygame i can 100% understand

especially if it carried over the cache between games

[real sierra]

even if not, the cache being carried forward within the same game is worth a lot

if you search up to 8 plies ahead, and then you make a move and black makes a move, there's still 6 plies ahead worth of cached moves

[stray dragon]

yeah

[real sierra]

unless the opponent did something truly wacky, i dont think the moves you're checking will change that much

[stray dragon]

i'll simply have to see

[real sierra]

why am i helping the competition

[stray dragon]

i'll help later it's an investment

trust

[real sierra]

lectures cancelled tomorrow long weekend

[stray dragon]

make a bot so awful to replicate that nobody will want to try

[real sierra]

my code is probably far worse than my bot tbh

im amazed how many tokens it uses for how few features it implements

[stray dragon]

yeah seeing how many tokens my neural net takes up hurts

especially the part where it's slotted into the eval function properly

was at like 500 tokens just for that + eval function, and that was before realizing the rules and compile issues

i think i need to trim stuff

horrible :(

[timber acorn]

so is it impossible to process that with neural net within 1k tokens?

[stray dragon]

i mean

no

not at all

it's entirely possible

[tender river]

its possible but the net's gotta be pretty small

[stray dragon]

but it doesn't leave as much room as i'd like, with the implementation i have

[tender river]

or maybe you can use 1.5-bit neural network

[stray dragon]

yeah mine has 1020 parameters

[stray dragon]

i use ternary

[tender river]

either way you'll have to use some unconventional approaches

[real sierra]

creativity is the name of the game

[stray dragon]

weights can each be -1, 0, or 1

[real sierra]

how to get maximum chess performance from a fixed token budget is a very abstract problem

[stray dragon]

seems to have good performance

[timber acorn]

gpt says even 16k is not enough for an able-to use nn 🙁

[stray dragon]

gpt is wrong

[real sierra]

gpt doesnt know anything

[stray dragon]

i've implemented it myself

[timber acorn]

i know

[stray dragon]

i will do updates on it though

i feel the key to making it work well is to have it adjust the big numbers that get summed into the evaluation

takes them in as inputs, spits out factors that it multiplies them by, sums them, that's the evaluation result

by "big numbers" i mean material, mobility, etc.

king safety

pawn structure, maybe

that way, it skips having to figure out what the inputs mean, and can do away with all the abstraction you'd normally want for a neural network like that

[sage crag]

yes this is 1.5 bit

or referred to as such

[stray dragon]

1.58 bit i've seen

[sage crag]

yup

[stray dragon]

chayleaf said 3 bit initially so i corrected

[sage crag]

1.58496250072

you can keep going forever

its log_2(3)

[stray dragon]

ok no need to get too specific

[sage crag]

hi

since your network is ternary

have you considered

[real sierra]

you know

we're kinda like the bots, in a way

our search space is code strings instead of games

[sage crag]

you can pack the entire network into a base-3 integer

[real sierra]

and our time limit is a token limit

[sage crag]

and then write its binary representation

[stray dragon]

did you know?

i already do this.

[sage crag]

[stray dragon]

that's how i compress my parameters into ulongs

38 parameters (40 - 2 for token reasons) per 64 bits

[sage crag]

you can also try packing 5 trits into a byte

with smaller networks its more optimal

[stray dragon]

ye it's an option if i still go the string route

wait

i can just do that still

[sage crag]

duh

[stray dragon]

and it would actually save me tokens potentially.

even from the current method that is illegal.

huh....

[sage crag]

on the unpacking perhaps, you would need to write a python encoder though

[stray dragon]

no

i wouldn't

i could just

take the characters as ints

and use the same ternary unpacking method i use already

but with less iterations per number

and i could remove the string parsing function in its entirety

meh i'll do it tomorrow

probably

[sage crag]

trits = [-1, 0, 1, 1, -1, 0, -1, 1, -1]
out = bytearray()
i = 0
n = len(trits)
while i < n:
    block = trits[i:i+5]
    while len(block) < 5: block.append(0)
    val = 0
    mul = 1
    for t in block:
        val += (t + 1) * mul
        mul *= 3
    out.append(val)
    i += 5
print(list(out))

trits = []
for b in out:
    val = b
    for _ in range(5):
        val, d = divmod(val, 3)
        trits.append(d - 1)
print(trits[:n])

all you need to do with out is convert it to ulongs from here, not my problem

[stray dragon]

i just change the 8-character representations to the hex-byte formatting method chayleaf showed earlier, then i change the numbers in this bit and fix it so it doesn't do /=3 before it gets the parameter, and i'm chilling

ez

[sage crag]

you could also try arithmetic numeral systems

probably rans

[stray dragon]

what

[sage crag]

sorry i mean asymmetric numeral systems

5am

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

[stray dragon]

ideas!!!

[tender river]

some compilers have char as unsigned char and some as signed char

[stray dragon]

chayleaf you might have helped me.

[stray dragon]

meh, easy fix

probably still worth

[sage crag]

its theoretically better than either of the previous methods i said

at asymptote at least

[tender river]

you can just say signed/unsigned explicitly but +1 token

[tender river]

[stray dragon]

this is looking like... 6 extra tokens added, 22 tokens removed, a few added and removed elsewhere for the other non-ternary params

[sage crag]

nna just uint8_t

[tender river]

shiro uses stdint.h in chessapi.h

[sage crag]

ye ye

[stray dragon]

oh i can just remove that from my bot then?

man.

ok

[tender river]

[stray dragon]

that feels so obvious

[sage crag]

[stray dragon]

damn it brain

[sage crag]

you dont need stdbool either

since c23

[stray dragon]

i don't

why would i ever want stdbool

[tender river]

could use _Bool even before c23

[stray dragon]

i never added it i'm not shiro

[sage crag]

fake

[real sierra]

stdbool enjoyers

[stray dragon]

cope

0 and 1 my beloved

bool is smelly

shiro stinky

[sage crag]

where did my vm go

[tender river]

[stray dragon]

yum

ate it while i was distracting you

[sage crag]

[real sierra]

t wouldn't have eaten it if it had stdbool

[stray dragon]

true

it's like putting raisins in a cookie

[real sierra]

i should've figured you'd be all-in on this contest

[stray dragon]

why would you ever do that

[real sierra]

even your username is minimum tokens

[stray dragon]

LOL

[sage crag]

[tender river]

shiro its normal to have a 1 token username

[sage crag]

shimo did you know about flexible struct members

[real sierra]

about what

do you mean unions

[sage crag]

nope

typedef struct {
  uint32_t len;
  uint8_t data[];
} Erm;

data isnt a pointer

its a flexible struct member

[stray dragon]

konii out here pulling C tricks out of a hat

[tender river]

some old compilers complain when it isnt data[0] or data[1]

but ye you can do that equivalent of rust's !Sized types

[real sierra]

i dont even know what that means

you've defined it in terms of itself

[tender river]

its a struct with an arbitrary size

[real sierra]

ffi's worst nightmare

[tender river]

[real sierra]

"how big is this struct?"

¯_(ツ)_/¯

as big as it wants to be ig

[tender river]

not any worse than opaque types

[real sierra]

opaque types are best

then u dont have to worry because u cant

its a pointer

NFI

[tender river]

you dont have to worry about this struct's size either

it doesnt exist

[real sierra]

[sage crag]

struct is opaque to ffi anyway

[tender river]

struct {
  int len;
  char x[];
} *s = malloc(sizeof(int) + 256);

s->len = 256;
for (int i = 0; i < s->len; ++i) s->x[i] = i;

[sage crag]

you can actually also use sizeof(T)

[tender river]

not here

[sage crag]

you not define struct

[real sierra]

bblblblblblbllbllllblbl

[tender river]

branch and link

[stray dragon]

what have you given me even skimming this creates new and interesting ideas for how to compress my stuff to silly levels

[tender river]

[tender river]

job application time

[tender river]

oh to answer your question, you can do that on macos but i dont think you can do it on linux elfs are arch and os specific

[real sierra]

[tender river]

windows supports it too

linux didnt have to since it doesnt rely on binaries but on source code

there was an attempt at introducing it https://raw.githubusercontent.com/icculus/fatelf/ed748c44c3a5eefa669ecf684acc6c4b2d23408a/docs/fatelf-specification.txt

but it didnt gain traction

makes me wonder

could you make it do #!/usr/bin/env bash on linux

but be pe on window/macho on macos

[tender river]

(or /bin/sh)

then you could patch the elf depending on cpu architecture from within bash

not implementing it btw

[gloomy yoke]

my bot is getting a lot of disconnects on cute chess, how can I check where and waht is going wrong?

[real sierra]

a disconnect is usually indicative of your bot crashing somehow

is it disconnecting immediately, or seemingly at random?

[gloomy yoke]

fully random

[real sierra]

i dont think there's a super clean way to identify the cause without much more information

[tender river]

i wonder if theres a way thats easier to debug (with stderr logs at least)

[real sierra]

running it in a terminal and feeding it UCI commands manually will let you see what kind of error it is

but that could take a while

and is annoying if you dont know for sure what caused the error

[gloomy yoke]

sometimes even the bot the plays random gets disconnected

[real sierra]

even the random bot?

when did you last grab chessapi.c?

[gloomy yoke]

the last time you updated it prety sure

[real sierra]

also just to be sure, which operating system and language are you using

[gloomy yoke]

w10 python

must be the way I made this dll

[tender river]

[real sierra]

if you happen to resolve those issues, or pinpoint a more exact cause, let me know

windows is notoriously difficult for everything

and the python bindings have a few known bugs that are being fixed soon

[gloomy yoke]

okay

[analog sentinel]

programming

[tender river]

[real sierra]

[tender river]

[real sierra]

[tender river]

nope

[real sierra]

[tender river]

wait no

[analog sentinel]

I've never seen that one

[real sierra]

linear algebra is the bane of my existence

lost to basis again

cannot build an intuition for basis no matter how many times i encounter it

[tender river]

[analog sentinel]

[obsidian mantle]

So you basically replace your array with a formula?

[tender river]

not really a formula

its a single number (though you cant really enter it as a single number as per the contest rules)

then you decode that number based on expected probabilities of symbols in the message

just read about huffman coding if you want something simpler to get an intuition of how compression works

[thick imp]

haha c# go brrr

[velvet crest]

You have unlocked new role

[thick imp]

what

[tender river]

[obsidian mantle]

But you decode with formula

I'll read this huffman thing ...

[tender river]

all programs are formulas

[thick imp]

    // Convert one array to a Set for faster lookup
    const set2 = new Set(arr2);

    // Filter arr1 to only elements that exist in set2
    const common = arr1.filter(item => set2.has(item));

    // Remove duplicates if needed
    return [new Set(common)];
}

// Example usage
const array1 = [1, 2, 3, 4, 5, 6];
const array2 = [4, 5, 6, 7, 8, 9];

const commonElements = findCommonElements(array1, array2);
console.log("Common elements:", commonElements);```

[tender river]

you can use new Set(array1).intersection(new Set(array2))

also you forgot ... in list expansion

[gloomy yoke]

it seens board.get_piece_from_index(target) is returning a 0, causing my crashes, someone reported this on issues in github already

[obsidian mantle]

You have to be sure that index actually has a piece in current board

[tender river]

not really its fine for it to return 0

its a python bindings issue

it will get fixed

[obsidian mantle]

Because 0 is valid return for no piece

Python uses 1-based arrays?

[tender river]

[obsidian mantle]

Then how can it crash anything

[tender river]

a lot can happen at ffi boundaries

(in this case its just enum conversion and shiro forgetting what his own function can return)

[obsidian mantle]

So 0 isnt properly handeled within the python binding?

Inside of it

[tender river]

mhm

[gloomy yoke]

I get this in my logs
ERROR: 0 is not a valid PieceType

[tender river]

shiro will fix

[gloomy yoke]

yea I will wait

yea, just had to change the api and it stopped crashing

[tender river]

change it to return None if it encounters 0

this is likely how it will work in a future api version

[real sierra]

willfix ✅

finally time tomorrow, no more assignments and class cancelled

now I eep

goodnight chess channel

[olive sable]

Goodmorning

[tender river]

shiro

unshiro

[nocturne olive]

Went looking for if the Chess API was broken

Found peak instead

[olive sable]

I have 2 hours of math today

But its math i learned 3 years ago, so my motivation to go to school today is non-existant

[nocturne olive]

So does Chess API still work or did you explode it?

[wide flicker]

chess

[nocturne olive]

have something not related to Chess for a change

[jagged turtle]

this channel became #chessbot

or smth idk

[olive sable]

yep

[wide flicker]

meanwhile im just tryna get bazzite and obs to play nice (PipeWire hates me)

[faint sandal]

chess

[midnight sigil]

chest

[sick owl]

I wouldn't be surprised, they churn out updated checkpoints of each model gen over at google

[obsidian mantle]

Surely i will make it beat random bot every time this weekend

Also is 3fold draw check not working or is it me

[faint sandal]

maybe it's time I got a new type c hub

[lilac lantern]

https://tenor.com/view/cat-cord-bite-dies-dies-from-cringe-gif-14516810518005393867

[obsidian mantle]

duct tape it

[faint sandal]

65w going through that wire

[obsidian mantle]

typeC.png

[lilac lantern]

Huh? Which one

[obsidian mantle]

It looks weirdly alligned

Compare it to other ports

Oh i think thats the reflection

Or what is that, a dust clump

[faint sandal]

it's supposed to be a type c port

it looks like AI upscaling fucked it up

[clear sedge]

i had a dream where my compiler told me to use uint_8_t instead of uint8_t

what a world that would be

[tender river]

uin8 uin16 uin32

sin8 sin16 sin32

[rare bramble]

that's actually evil

[cosmic sphinx]

flash lite actually became much better with this one, comparable to flash, but also retaining the insane speed

couldnt say the same about flash's upgrade

[shadow sinew]

Looks like they added a USB-C port at the last moment so someone shopped it into an existing product shot (poorly).

[obsidian mantle]

At first i thought you were talking about flash player

[sage crag]

ye me scary

https://youtu.be/RFWJM8JMXBs
@stray dragon @obsidian mantle video about ans

[sick owl]

This is altogether more exciting imo

https://deepmind.google/discover/blog/gemini-robotics-15-brings-ai-agents-into-the-physical-world/

[cosmic sphinx]

yeah, google finally revealing what they cooked in robotics

and even deployed a test model on AI studio too

[clear sedge]

writing a gltf parser is so fun

[sick owl]

Oh sweet!

I can't wait to play with this

Did they add a preview of the non ER version as well?

[cosmic sphinx]

unfortunately all gemini models, including this one, are still not good enough for counting tasks on image

[sage crag]

mambo mambo

[rigid snow]

ded chat

[sick owl]

Tested it out a bit

Its a very impressive reasoner

[olive sable]

im back home

[clear sedge]

i wonder how long i've had this qemu instance open for

"gee i wonder why i'm using so much ram"
the kde fedora instance i launched for testing days ago:

[faint sandal]

[sage crag]

vani put tthis emoji in the server pleas

[tender river]

forgot

[sage crag]

[clear sedge]

how can you forget

[sage crag]

cantjam

[faint sandal]

[sage crag]

:MeWhen:

[tender river]

[faint sandal]

[sage crag]

no this emoji can leave

i forgot

@tender river writeShellScriptBin

writeShellApplication

[tender river]

https://ryantm.github.io/nixpkgs/builders/trivial-builders/

[olive sable]

[real sierra]

mimimi