programming

[real sierra]

I have

[honest scaffold]

[real sierra]

-#

smol

[sage crag]

-#

https://cdn.discordapp.com/emojis/1162976978278809641.webp?size=96&animated=true

https://cdn.discordapp.com/emojis/1162976978278809641.webp?size=96&animated=true

https://cdn.discordapp.com/emojis/1162976978278809641.webp?size=16&animated=true

https://cdn.discordapp.com/emojis/1162976978278809641.webp?size=16&animated=true

https://cdn.discordapp.com/emojis/1162976978278809641.webp?size=8&animated=true

16 minimum?

https://cdn.discordapp.com/emojis/1392454929616076890.webp?size=96

https://cdn.discordapp.com/emojis/1392454929616076890.webp?size=16

tiny

[nocturne olive]

SMOL

-# I don't know how SMOL really, looks like a normal emote to me

[real sierra]

the image bevel makes it look like it's peeking around a corner on mobile

[solid bough]

I'm back

Hi Shiro~

[tender river]

this happens when i ^C

[solid bough]

Do you like Coding?

[tender river]

https://imgur.com/a/CO5Mscw

[solid bough]

That's the spirit

[real sierra]

normal

the thread sleeping for the semaphore fails ungracefully when interrupted

[solid bough]

Damn

[solid bough]

When a thread goes ungracefully on my code, I just make that it isn't seen.

[real sierra]

I think it's better to be seen if it does than not

[tender river]

it seems to work

[solid bough]

Jokes aside, if I wanted to stop the code and it then has a traceback just because of the stopping, then I have a reason to not make it shown.

[tender river]

holy that was fast

[sage crag]

nyum leaf

[tender river]

cant use rng so using board hash for move selection

[solid bough]

Hi konii 2.0.1

[tender river]

erf let me write docs and surely ready to ship

[real sierra]

[solid bough]

Everything is fine:

[real sierra]

bred was running whole tournaments at this speed, that's how they identified some kind of 1-in-50000 edge case in the legal move logic

still don't know what the cause is, or if it's accidentally been fixed in the latest API version

probably castling or knight moves related

[sage crag]

ai generated ai code

[solid bough]

You're not wrong actually, It is based of Deep Doodle But I wasn't doing the fixes to run on Python 3

[rigid snow]

if that's written by ai

idk

it's not

i don't believe that

[solid bough]

Nope, I got the code of that Repo and wanted to make that run on Python 3

[rigid snow]

yeah

like

this is human code

[real sierra]

they're self-replicating...

[solid bough]

That is the code that has log_info(text, logg)

And that time I wrote it by hand

[rigid snow]

that wans't really the point

[solid bough]

Dude, I wrote the module not quite up to the standards as normal modules (it is not installable via pip)

[sage crag]

if your loss is MSE

not in original repo

[solid bough]

What was the point?

[rigid snow]

yeah that's fine still?

[solid bough]

The Deep Doodle repo?

[sage crag]

all code with capital letter in comments ai code

[rigid snow]

i know you're referring to third person but it's not uncommon to "address the reader" as the user of the software

[rigid snow]

why are you assigning the string to a variable first and not just passing it to the function directly

also initialisation is a rare spelling

llms don't do rare spellings

[solid bough]

Because if I do "log_info(text="Sometextidc", logg) I get an error.

[solid bough]

on the log stuff?

[rigid snow]

log_info("Sometextidc", logg)

[solid bough]

[solid bough]

Please wait, Imma try that. (with the needed imports already done for that line to run)

[sage crag]

what is logg for though

[rigid snow]

yeah this

no ai

[sage crag]

definitely not this part

[real sierra]

I don't think it's that bad

reasonable way to easily enable/disable logging in dev code

[tender river]

did you remove the .so from the repo yet

[real sierra]

should use logging levels and stuff still

[real sierra]

yes

don't you have access to the repo

[sage crag]

global variable better way for this

[tender river]

[real sierra]

[clear sedge]

IT USES STRINGS AS BOOLEANS SHIRO

[sage crag]

or log level

[clear sedge]

what are you ON ABOUT

[real sierra]

WQIT

JUST SAW THAT

[tender river]

lazy

[real sierra]

banned

[opaque sigil]

[sage crag]

this also true

some_bool = "false"
if some_bool:
  print("how did we get here")

[real sierra]

I need to stop answering you, I'm enabling this

[rigid snow]

that's why they do some_bool == "true" bad reader

[cosmic sphinx]

did chinese companies cook today @sick owl

[solid bough]

Oh, I just made that that I can just put any text and have a "[SYSTEM] textidc" and "[INFO] textidc" and "[WARNING] textidc" and also "[ERROR] textidc"

[sage crag]

[rigid snow]

what

[solid bough]

Just some basic logging module

[real sierra]

someone is gonna see this and parade in the streets that AGI is here

[sage crag]

from typing import NoReturn
from sys import exit, _getframe, stderr
from pathlib import Path
from pprint import pformat

def srcloc(skip: int) -> str:
    f = _getframe(skip)
    name = f.f_code.co_name + " @ "
    if name.startswith("<"): return ""
    return f"\x1b[30m[{name}{Path(f.f_code.co_filename).name}:{f.f_lineno}]\x1b[0m "

def exceptionloc(ex: BaseException) -> str:
    tb = ex.__traceback__
    if tb is None: return ""
    while tb.tb_next: tb = tb.tb_next
    f = tb.tb_frame.f_code
    name = f.co_name + " @ "
    if name.startswith("<"): return ""
    return f"\x1b[30m[{name}{Path(f.co_filename).name}:{tb.tb_lineno}]\x1b[0m "

def log(kind: str, escape: str, fmt, skip) -> None:
    if isinstance(fmt, BaseException): loc = exceptionloc(fmt)
    else: loc = srcloc(skip)
    print(f"{loc}\x1b{escape}{kind}\x1b[0m: {pformat(fmt)}", file=stderr)

def error(fmt, skip: int = 3) -> None: log("ERROR", "[31m", fmt, skip)
def info(fmt, skip: int = 3) -> None: log("INFO", "[32m", fmt, skip)
def warn(fmt, skip: int = 3) -> None: log("WARN", "[33m", fmt, skip)

def die(fmt, code: int = 1, skip: int = 4) -> NoReturn: error(fmt, skip); exit(code)

cry, weep

this is my terrible code

[solid bough]

TLDR: I just wanted an easy way to just write a function with the wanted result and be able to re-use that py file on other code.

[rigid snow]

fancy, ansi colors! may i suggest emojis next

[solid bough]

Jesus, I am not that far

[tender river]

ye shiro its all your fault

[sage crag]

ye next

[solid bough]

I'm still pretty basic

[real sierra]

use more type hints

[sage crag]

nope

everything that needs to be hinted is hinted

[tender river]

thats enough

no more needed

[real sierra]

add more

[solid bough]

And I personally love that look: 22/09/2025 22:28:34 [INFO] Checking if the data is set up 22/09/2025 22:28:34 [INFO] Data is there. 22/09/2025 22:28:34 [INFO] Loading Data... 22/09/2025 22:28:35 [INFO] ('Loaded', '4000', ' Samples.') 22/09/2025 22:28:35 [INFO] Attaching more channels and splitting the data. 22/09/2025 22:28:36 [INFO] Channels attached and split. 22/09/2025 22:28:36 [INFO] Shuffling the data... 22/09/2025 22:28:36 [INFO] Data got Shuffled successfully. 22/09/2025 22:28:38 [INFO] Setting the image format. 22/09/2025 22:28:38 [INFO] Building the model from Cache or Scratch. 22/09/2025 22:28:39 [INFO] Imported SGD from keras.optimizers 22/09/2025 22:28:39 [INFO] set up the optimizer successfully. 22/09/2025 22:28:39 [INFO] compiled the optimizer. 22/09/2025 22:28:39 [INFO] made a Model image. 22/09/2025 22:28:39 [INFO] Done, Training the Model (initialisation) 22/09/2025 22:28:39 [INFO] Epoch 0 of 500

[real sierra]

needs Optionals and Unions

[sage crag]

NOT

[rigid snow]

!

[real sierra]

I pass fmt=25 to log

[solid bough]

I'm perfectly happy with the logging

[rigid snow]

that's all that matters

[tender river]

shiro which type hint do you think best describes you

[real sierra]

optional

maybe Optional[Shiro]

sometimes I Shiro

sometimes I nothing

[tender river]

[solid bough]

Okay, that works. But I'm right now not really in to changing my whole code to that rn

I was doing py log_info(text="textidc", logg)

[sage crag]

it works differently than it used to

[opaque sigil]

if you ever feel like adding more stuff to the logging there's a nice built-in module for that so you can avoid having to check if debugging/logging is enabled every time

[sage crag]

it used to properly categorise the exception and print it

oh is it because its not being called from in a function

funny

yup

wait nevermind

its still doing it

just has the correct srcloc

eca hca svc env xec sys rte

pick a name

[opaque sigil]

svc

[sage crag]

reason?

[solid bough]

Wait, is the try except inline?

[sage crag]

you can call it inline if you wish

most of python lets you do this

[opaque sigil]

stands for service

[sage crag]

with the exception of loops

[solid bough]

Okay, I didn't know what to call it

[opaque sigil]

first one i recognised

[sage crag]

ye that good, svc disqualified from list

they are names for environment calls

in my third vm project this month

[opaque sigil]

just do sys then

[sage crag]

but if vm context

sys -> system

[solid bough]

I don't do VM on the Deep Doodle Code

[sick owl]

Seems they did

Native streaming audio out and in is a severely undercooked part of the local space

[solid bough]

Should I do VM on my code?

[rigid snow]

2.5-omni was pretty resource intensive and i have not heard of anyone running it locally

[sage crag]

that does not make any sense

[solid bough]

okay, Should I run my code in a VM?

[sage crag]

uhh no?

[rigid snow]

not trusting your own code

[solid bough]

What do you mean "if vm context"

[cosmic sphinx]

oh yeah, it wasnt ending with just 3 posts..

[sick owl]

Sweet

[tender river]

thats a different kind of vm

bytecode vm

[solid bough]

[rigid snow]

java virtual machine ✅
virtualbox ❌

[sick owl]

I will say I don't find Deepseek 3.1 Terminus that exciting rad model name aside

But the Qwen releases are very cool

[solid bough]

All I know of VM is Hyper-V

[rigid snow]

yeah not that

[cosmic sphinx]

horror

[rigid snow]

what the fucky

spoiler pls

[sage crag]

i dont like vedal plush 2

original vedal plush better

[solid bough]

Yeah, I can't pass a GPU in to the VM with Hyper-V anyways

[olive sable]

eeewwwwwwwwwwwwwww

[rigid snow]

what does that have to do with

nevermind i give up

[solid bough]

I mean for what I need

My bad

Sorry

[sick owl]

https://tenor.com/view/stranger-things-fire-flame-gif-14920648

[rigid snow]

what are you apologizing for

[solid bough]

That I wasn't clear enough that I was talking running Tensorflow in a VM

[rigid snow]

whyd it make his shell a ball though

[real sierra]

always trust my own code

flawless first try

everything added after the first run is a bonus feature

[cosmic sphinx]

are we really questioning why LLM's do something atp..

[solid bough]

200 OK
-# I trust all code of Shiro~ now ✅

[real sierra]

good choice

please run this Bitcoin miner now

[cosmic sphinx]

here's another jumpscare
||bro, I aint switching from gemini 2.5 flash anytime soon with these 'crazy releases'||

[rigid snow]

but that's no fun is it

[solid bough]

AAH

jesus

[sick owl]

https://tenor.com/view/alien-sygourney-weaver-i-say-we-take-off-and-nuke-the-entire-site-from-orbit-its-the-only-way-to-be-sure-movie-quotes-gif-11314084

[cosmic sphinx]

fyi this is what gemini does in half the time qwen takes (it does not turn the plushie into a backrooms horror creature)

[solid bough]

That looks actually better

[rigid snow]

it doesn't look like a plush anymore

[solid bough]

The smile is more like if vedal would smile with the mouth closed

[solid bough]

I prefer that one over the other 2

[keen hatch]

deepseek also released a small update https://huggingface.co/deepseek-ai/DeepSeek-V3.1-Terminus

[rigid snow]

that was mentione

d

[cosmic sphinx]

idk, it looks like a blender typa render to me asw

[solid bough]

That sure is a model.

[rigid snow]

it looks like a plastic toy wrapped with suede

[cosmic sphinx]

gotta keep refining the prompt from just 6 words then ig

[real sierra]

the consistency it gets with the lighting is pretty impressive

which the shadows and the reflections agreeing

[keen hatch]

In my opinion, deepseek is still the best model that can be run locally if you have enough resources.

[solid bough]

What amount of VRAM?

[keen hatch]

i run in on cpu

[tender river]

@real sierra pretty sure GameState isnt used anywhere in the public api

[solid bough]

So I assume more than 8GB

[real sierra]

is_game_ended

[cosmic sphinx]

gemini's consistency in style and world understanding is acc insane compared to every other model
also why it earned a top spot far ahead of everyone else on lmarena undercover

[keen hatch]

i have 64gb vram, 512gb ram

[tender river]

it returns an enum ok misleading name

[real sierra]

[tender river]

rename now

[real sierra]

breaking change for semantic benefit is nonsensical

[solid bough]

Damn, I have 64GB DDR4 RAM and 8GB of Asus NVIDIA GeForce RTX 3070 Ti TUF GAMING OC EDITION

[tender river]

not breaking until its released

[real sierra]

I could add an alias to it with a better name tho

then deprecate old version

[tender river]

this is literally the stage where you do breaking changes

for any benefit that is at all possible

[real sierra]

it's effectively a breaking change with how many bindings have been written already

I'll just keep both

and APIs can choose which to implement

[tender river]

ping every bindings author

[real sierra]

they want to be lazy too you know

[tender river]

then remove from header but export from the c file

[real sierra]

you can't hog all the lazy for yourself

[tender river]

their fault for not using haskell

[keen hatch]

I even ran kimi-k2 locally (6-7t/s)... but I don't like this model...

I don't like this model's personality. Kimi-K2 is an introvert who doesn't want to talk to you.

[obsidian mantle]

do you just talk to them?

[rigid snow]

this is a discord server dedicated to an ai vtuber

what do you think

[keen hatch]

No, but sometimes when I need to think about something or look at a problem from a different angle, I need a model to continue the conversation rather than try to end it, i.e., for me, follow-up questions are a must.

[sage crag]

im designing the opposite of peak

hardware loops

[glass flower]

just spell out the whole word.... why turn halt into hlt

bad code

[sage crag]

easier for me to mentally parse

its an instruction name

halt has an annoying vowel in it that makes it harder to read

[obsidian mantle]

i am a bit scared of using ai because i feel like it will reduce my ability to think by thinking for me

[glass flower]

are you even human or do you run on instructions

[sage crag]

[keen hatch]

I have fine-tune to chat with... but larger models can be put to better use

[olive sable]

i should sleep

[rigid snow]

same

[tender river]

loud incorrect buzzer

[glass flower]

oh it definitely does that i recently switched from a local model to another and it was a bit slower and i caught myself just waiting

thats why my only thing i use AI for is autocomplete and converting code from one language to another

[tender river]

haddock

[olive sable]

1 more episode

[tender river]

[glass flower]

and i use dumb models intentionally. so i can't actually rely on it and never stop thinking

[sage crag]

ldr 1 $1
ldr 100 $2
; load address into $3 here or something
ldr 4 $4
rep $1 $2 5 1
setb [3] $5 $4

memset or something idk

[obsidian mantle]

2-39
gotta go sleep my 5 hours of sleep

[sage crag]

my ISA even if its wrong its right

[obsidian mantle]

scary text

what are those? some merchendise prices?
$5 $4

[sage crag]

maybe i should remove the increment from the rep instruction and have it in the loop body instead

ldr 1 $1
ldr 100 $2
; load address into $3 here or something
ldr 4 $4
rep $2 5 2
setb [3] $5 $4
inc $2

hardware loops scary

[real sierra]

@tender river new version of API up with renamed function

[sage crag]

ye now know why there are so many RISC vms but hardly any CISC vms

designing CISC vms is painful!

[keen hatch]

The point here is to use AI for brainstorming. It's like someone asking you questions, not the other way around. This way, you can find a solution much faster. You don't need to ask the AI ​​to think for you or write code, for example.

[glass flower]

that can also work ye

[sage crag]

graph reduction vm

[tender river]

i'd think it's mostly "start with risc then add common instructions for common operations"

[sage crag]

no no as per programming bad habits 101 you need to start by designing the most complex part of your system before anything else

[sage crag]

free parallelism but

SLOW

[tender river]

haskell is fast enough for what it is

[sage crag]

node based vm

[real sierra]

python has a lib called hypothesis that does automated generative testing using templates for input samples, and i've always wanted to use it for something completely different from testing

[tender river]

ye haskell's evaluation model is graph reduction

and each value is either an evaluated value or a thunk that begins its evaluation and rewrites itself

[keen hatch]

and there is one more advantage... initially you have to explain to the AI ​​what the problem is... understanding the problem is part of the solution.

[sage crag]

im a thunk

[real sierra]

🔨

thunk

[sage crag]

why dont people just smash register vm and graph reduction vm together, are they stupid

[real sierra]

graph of registers

[sage crag]

unironically it would be faster because you dont need a heap of evaluated values

but it still sounds slow

[real sierra]

use stack instead of heap

[tender river]

do you mean something akin to forth

[sage crag]

ye infinitely worse data structure

[sage crag]

[tender river]

haddock

[sage crag]

thats a fish

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

[real sierra]

time to go through the esolang wiki and see if theres any new ones i genuinely like

[tender river]

999 more brainfuck clones

[sage crag]

why cant i write something thats better than all existing solutions with little or no experience studying existing solutions

[real sierra]

?

[sage crag]

2d program

instruction pointer increment in either axis

arrow changes direction that it increments in

sounds like

[real sierra]

you can write to the instruction space

[sage crag]

> v
^ <

cursed loop

[sage crag]

small secret

every vm thats written by a hobbyist doesnt have memory protection

[real sierra]

no i mean there's builtins for it, no hacks required

[sage crag]

[sage crag]

i was right its 2d

[tender river]

haddock supports latex

[sage crag]

when ndtensor based program

[tender river]

rustdoc could never

[sage crag]

[sage crag]

we should write a new markup format to fix this

[sage crag]

send that back to whence it came

[sage crag]

ML vm

[tender river]

[sage crag]

[rigid snow]

it appears it is uhhhh not possible to send messages back in time

[sick owl]

I really like how every major windows update now has a 50/50 chance of fucking up your nvidia drivers

That's a really great bit of windows I think

↓

[tender river]

[sage crag]

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

[real sierra]

this list is scaring me

[sage crag]

didnt try hard enough

[sage crag]

esolang esolangs

[sage crag]

i feel like the more complicated the vm execution is the harder it is to write programs for

[sick owl]

I'd have moved my desktop to bazzite by now if not for game pass

[keen hatch]

I spent months talking to MSI, NVIDIA, and Gigabyte support to fix the 50xx compatibility issue with threadripper CPUs. Ultimately, the best solution was to buy a new motherboard and processor. NVIDIA drivers for the 50xx are a horribly buggy mess.

[sick owl]

I'm on a 4090

They're all cooked

[balmy nacelle]

wats going on

[obsidian mantle]

Is there a chance this is all happening because they all started utilizing ai for their software development

[balmy nacelle]

is this channel also for tech help cuz id love to know whats the best next thing to upgrade in my pc

[obsidian mantle]

Grok, make me a 5090 driver

[sage crag]

nyan nya nyan

thinking gives me headache

[keen hatch]

They asked Evil to help with the code

[tender river]

also uh not a big deal but wouldnt it be better for make_move to set capture and castle by itself so API users dont have to set it if creating Moves from scratch?

castle - whether its a 2 square king move
capture - whether its either a diagonal pawn move or theres an enemy piece present in the target square

i know it doesnt cause any exploits its just potentially better ergonomics in certain cases maybe?

[obsidian mantle]

Started learning neural networks, following some guide from internet to make basic nn on cpp.
The cost calculation is dank. But at least I can finally use derivatives knowledge i still have

Was thinking to make chess bot learn for 55 seconds and then play the remaining 5.
Now im not sure it will fit into 1000 tokens even if it would somehow be efficient. How good can it learn in 50 seconds?

[stray dragon]

not very well

if you want something like that, better to have efficiently updatable parameters

[obsidian mantle]

Someone said they were making "small neural network" for the bot. I wonder what it looks like and what exactly does

[stray dragon]

me

lol

[obsidian mantle]

Oh

Hi

[real sierra]

maybe, i would have to look at the code to ensure there's no edge cases that would require information assembled in get_legal_moves, but i do think the use cases are probably limited given users will (generally) want to ensure moves are legal before making them anyway, in which case they can simply be pulled from the returned array

[stray dragon]

basically: output_vector = matmul(matmul(input_vector, weights_part_1), weights_part_2)

[real sierra]

its a change that can be made at some point with no change to the header so i dont think its urgent

[obsidian mantle]

Yeah but where do you get weights from? Just put some pre-calculated constants?

Hmm

[real sierra]

most regression methods need a starting point of some kind, yea

[stray dragon]

yep, i'm converting unsigned long long values to like 40 ternary (-1, 0, or 1) parameters each to use as weights in the model

[real sierra]

just pick something reasonable and trust your convergence rate

[obsidian mantle]

Actually how good will be the bot with pre-declared neurons which fits into 1k tokens

[real sierra]

generally, neural networks are used for board evaluation in chess

[obsidian mantle]

Its not an llm after all, doesnt need to have 40000 books knowledge of human speech

[real sierra]

they're given a position and asked to output a score on how good it is

but that alone isn't enough for a strong chess bot

[obsidian mantle]

Yes but how small are they

[real sierra]

ask t

t

t

[stray dragon]

i'm guessing it'll be rather good, as long as you try to minimize the amount of processing it needs to do (give it the clearest information possible and require the least processing of that information to change the evaluation of the input state)

hi shiro

really, it's just an augment to the eval() function

i use it intead of piece-square tables because i don't like the idea of having to store 64*6*2*2 different values for something that has clear patterns and probably doesn't require that much extra neural processing to adapt 1 set of parameters to any piece type, any player color, any game period

[obsidian mantle]

Initially I wanted pure heuristic but now I feel like neural network is much more fun

[stray dragon]

ye i want an excuse to work with one

[tender river]

also really minor but i'd replace is_white_turn with is_black_turn purely because BLACK is 1 and WHITE is 0

[stray dragon]

agree^

[real sierra]

what if the api just had both instead

[stray dragon]

fair

[real sierra]

i know i said the api should have a boiler plate but you're really making me update it to save a !

🥀

[olive sable]

i might have watched 2 episodes
shoudl sleep now tho

goodnight

[obsidian mantle]

Why do you even need to know which turn it is

[stray dragon]

looking ahead

[obsidian mantle]

But when they tell you "go" you know its your turn, and when you look ahead you just flip the table

[stray dragon]

or pulling up a board state from a transposition table or something

it just simplifies things

[real sierra]

though not as relevant for this competition, if you wanted to store opening books, it may be important to know which colour you are

also it turns out that the "ideal" piece square tables for black and white are likely different

induced by the asymmetry of white having first move

[stray dragon]

^^

[real sierra]

thats pretty advanced tho, prob not enough tokens to worry

[obsidian mantle]

Yeah because chess isnt 100% symmetric

Makes sense

[stray dragon]

it's why i include one of the *2 in the potential number of piece-square tables i mentioned earlier

[obsidian mantle]

Now imagine processing that within 1000 tokens

Looks so minor

[rigid snow]

i have maximized the discord window my whole life until like a week ago, and now accidentally maximized it, remind me to never do that again

it's such a weird thing

why is it so much better in a like 800p window on a 1080p monitor

[obsidian mantle]

Its good for when you have messy desktop and dont want to see it

[real sierra]

ok, added is_black_turn to the api

[rigid snow]

i have disabled desktop icons 2 years ago, my desktop is just my wallpaper

i do use desktop icons on my mac since they have the stacks thing

if windows had that i would

[rigid snow]

i disabled them because it is so messy

[obsidian mantle]

Is it enum too

[real sierra]

bool

[obsidian mantle]

messy

[rigid snow]

1 token savings or why

i remember there being a white version in the header

[obsidian mantle]

0 ambiguosity

[real sierra]

the enum for playercolor has WHITE=0 and BLACK=1

[obsidian mantle]

Is it a word

[real sierra]

so is_white_turn casted to an int didnt match

[rigid snow]

oh

[real sierra]

so yes

1 token saving

because ! fixes that

[rigid snow]

that is stupid i'm glad you fixed that

[stray dragon]

except it's not even 1 token saving

because ^ can replace != in that case

xor my beloved

[real sierra]

well i was thinking if you wanted to pass it directly to a function taking a PlayerColor

[stray dragon]

i will say though: == looks a lot better than ^

[real sierra]

e.g. chess_get_bitboard(board, is_black_turn(), ROOK);

[obsidian mantle]

forcing token wasting what if someone doesnt want it

[stray dragon]

shiro can you add a function that summons stockfish pretty please 🥺

cherry on top

[real sierra]

[obsidian mantle]

stockfish summon but has 50% chance of actually making worst possible plays (calculated precisely)

[stray dragon]

also can you add a function to wish you a happy birthday so i can waste 3 tokens

[obsidian mantle]

Someone should make a bot that makes random moves and uses remaining tokens for some funny messages

[real sierra]

you could technically print info string <str...> \n according to the UCI spec

cute chess doesnt render it nicely but it shows up in engine debug

[obsidian mantle]

Works in arena

[obsidian mantle]

Oh no

[stray dragon]

i will add this to run every turn now thanks shiro

[obsidian mantle]

Whats fun in that

Maybe you can enable it in settings? Or is there not even log ln gui

In main screen

[real sierra]

not a setting

just not implemented

[obsidian mantle]

Bwa

[rigid snow]

will someone recreate the legendary neuro’s strat in their bot

[stray dragon]

d2d4
Happy Birthday Shiro!!!
b8c6
Happy Birthday Shiro!!!
e2e3
Happy Birthday Shiro!!!

[obsidian mantle]

i was thinking about this

Neuro gambit

(idk what gambit is)

[stray dragon]

the piece square tables for the pawns may be slightly overtuned

[real sierra]

nobody:
♟️ :

[obsidian mantle]

(shiro chose cute chess ONLY because it doesnt allow to send those messages)

[real sierra]

i chose cute chess because it doesnt randomly crash and then cook my terminal

[tender river]

i wanna eat

i wanna drink

docs

[real sierra]

ive made a discovery

about the weird python bug

there still seems to be no pattern about when it occurs

but

it always makes the move g6g8 illegally

how

[obsidian mantle]

On the border

Something something arrays out of bounds

[real sierra]

its so odd

actually wait

def get_legal_moves(self) -> list[Move]:
        moves_len_C = c_int(0)
        moves_C = cast(lib.chess_get_legal_moves(byref(self), byref(moves_len_C)), POINTER(Move))
        moves_len = moves_len_C.value
        moves = [moves_C[i] for i in range(moves_len)]
        lib.chess_free_moves_array(moves_C)
        return moves

you don't suppose moves_C[i] is not actually reserving Move objects in python memory?

[obsidian mantle]

Who's i

[real sierra]

idk, who are you

[obsidian mantle]

Oh
I see it
Looks weird on phone

I am i

[real sierra]

but yea this is the only thing that makes sense to me

[sage crag]

not for long

[real sierra]

if the move objects are still backing onto C memory

then maybe they're being overwritten during move selection since teh original array is freed

the easy way to test this is to just not free the array

if it only leaks a kb or so per move i can tank that

[obsidian mantle]

I had a bug in C program where it was split into parts that were called asynchronously and someone decided to use global i's in cycles

[real sierra]

fixed

so that is the problem

damn

thats hard to fix

guess i can make an internal MoveArray object with a refcount

[stray dragon]

the problem is you

[tender river]

i feel like this would be better with return 0; at the end

[real sierra]

and call free once all the moves are recycled

[stray dragon]

and yes, it is very hard to fix

[tender river]

why not just copy them

moves_C[i] is probably a pointer

[real sierra]

moves_C is a pointer

moves_C[i] is a move struct

Move just expands structure from ctypes

[tender river]

why is it a problem then it should be copied to python heap

[real sierra]

its not

[obsidian mantle]

Maybe it doesnt like the name

[tender river]

then its not a move struct but a pointer

mhm

[obsidian mantle]

Moves_C
It wants Moves_Python

[real sierra]

it's exposed to python as if its a struct

but i think the struct is actually in c and ctypes is abstracting that

as far as the "emulated" types are concerned, this should not be a pointer

[sage crag]

firefox is downloading half of my beatmaps in Pictures and the other half in Downloads

[tender river]

ye so dont do that

[real sierra]

ye so if you have complaints take them to ctypes

you're on some level 999 procrastination rn to be scrolling the api

[obsidian mantle]

Probably depends on the file extension

[real sierra]

regardless too many ops in python is gonna cause slow

[sage crag]

extremely loud incorrect buzzer

they are both .osz

[real sierra]

so im gonna just add a wrapper class for the movearray for internal use and keep a refcount

each time a move object is gc'd by python, it'll lower refcount of its source array

and then once that is gc'd too, it will call free array func

[tender river]

a = [Move() for i in range(len(moves_C))]
for i, x in enumerate(map(pointer, a)):
  x[0] = moves_C[i]

something like this

[tender river]

no shiro do not

[real sierra]

finish your homework 🥀

[tender river]

do this

[real sierra]

omg backseating

im gonna lose it

[real sierra]

deleting channel

programming

[tender river]

[stray dragon]

before we all get deleted i just want to say

happy birthday shiro

[final tinsel]

farewell programming

final msg

[real sierra]

what is

[final tinsel]

parody :)

[stray dragon]

I NEED A SHIRO

I'M HOLDING OUT FOR A SHIRO TILL THE END OF THE NIGHT

[real sierra]

scientists have discovered the strongest material in the universe and it's whatever is holding my sanity together

[final tinsel]

i thought the limits of that were already found?

[real sierra]

people are always redefining the limits

human spirit something something

[tender river]

also this doesnt match the docs i think

because it doesnt do gen | in the final return if it slides it

i faithfully translated it not my problem

[real sierra]

it doesnt need to

whether captures is true or false, the return never includes the original square

[tender river]

but then it doesnt include the second square either

no it doesnt got it

[real sierra]

ok which part of the docs are you confused about

you're misinterpreting something

[tender river]

its fine

[real sierra]

bro

did you go to ragebait academy last night what is happening

[tender river]

"it doesnt" was referring to "doesnt"

[real sierra]

[tender river]

double negation

[real sierra]

[tender river]

[real sierra]

as long as there no problem

if there bug there then i fix

my roommate blasting music in the shower and singing to it

why do i have to put up with this

where did shame go

[tender river]

docs done i think

[real sierra]

ah

damn it

ctypes why is your documentation so sneaky

[tender river]

i should probably render them

[real sierra]

[tender river]

doc sections dont work

i'm doing something wrong

oh so thats how you do it

[real sierra]

alright python fixed

[tender river]

i'm close to being done

fixing doc formatting

[real sierra]

python and C random moves bots are surprisingly on par for time

229 move game and neither used more than 2ms

would've expected python to be at least slightly slower

[sage crag]

all youre paying for is python's call to c overhead and for loop overhead

port your actual bot and see a difference perhaps

[real sierra]

for loop overhead is a lot

[real sierra]

[sage crag]

noe, for loop overhead is small

its the python arithmetic that is slow

[real sierra]

my bot doesnt do a whole lot of arithmetic

[sage crag]

bitwise ops?

anyway

python iterator generators are written in c

[real sierra]

[sage crag]

python control flow is not that slow

[real sierra]

the most expensive operation my bot actually does with any frequency is a < b ? c : d

[sage crag]

its a stack vm so its pretty bad at complex arithmetic

aside from that its bad at threading

but python isnt that slow these days

especially if your whole program is calls to c

[real sierra]

even so

if i lose to a python bot

💔

ill never forgive myself

what scares me is how few tokens you can express things in using python

generator syntax for mapping lists is so powerful

[tender river]

it doesnt sound that weird though? like even c++ is a much better choice if you're going for token count

speed does help but better logic goes a long way

[real sierra]

but speed

🇨 🔥

[tender river]

how do you publish a repo for gh pages again

[real sierra]

i thought github repos were public by default

github pages is something else tho idk how that works

i dont suppose you went with a custom bespoke docs format that doesnt conform to the standard markdown that the other docs do

[glad path]

this gives me

"oh, python is too slow?"
"then write faster code"
vibes

does anyone know if the openai api lets you see how much of a list of messages you can fit into a context window

[rigid snow]

this isn't an answer but openrouters middle out transform solves this

it's so nice

[glad path]

my idea is to fill 90% of the ctx window with message history, and a system prompt to the very beginning

[keen hatch]

On average, one word is 3 tokens, so you can roughly calculate it.

[keen hatch]

https://github.com/openai/openai-cookbook/blob/main/examples/How_to_count_tokens_with_tiktoken.ipynb

[sage crag]

if the message log is very long you may have to duplicate the system prompt to the end

[glad path]

the idea is to trim it to fit into the window myself

so there'd always be room

[tender river]

merge my pr and everything's ready on my side

[keen hatch]

main issue with system prompt update - kv cache invalidation

[real sierra]

merging into mine?

how do i push maintenance off to other people if its in mine

ok lets see

[tender river]

its only a project template

the actual library will be maintained by me

[keen hatch]

... so the server will have to process the entire prompt from the beginning

[real sierra]

wow haskell is cool

i gotta learn this at some point'

[sage crag]

[glad path]

it would be the same system prompt btw

[tender river]

i dont actually know any haskell btw

this is my second time ever using it

i read a lot of it but i have no use for the language itself

[sage crag]

first time was SAT solver

[real sierra]

merged

[sage crag]

ye easter egg added

you will never find it

[tender river]

🤫

[real sierra]

at last

i can open the contest and api

gonna do some final checks

[tender river]

better thank me for delaying it until you fixed the python bugs

[sage crag]

merge conflict

[real sierra]

i knew they were there, i just delayed fixing

[real sierra]

no conflict

[sage crag]

ye because strong ordering

merge conflict race condition

speaking is mutex

[real sierra]

?

[rigid snow]

the mutexes are speaking to me

[tender river]

shiro can i have water now

please

[sage crag]

no

you can have monster energy drink only

[tender river]

too expensive

[sage crag]

shiro paying

[tender river]

[rigid snow]

enrgy drink is kil u

[sage crag]

not

not do that

[tender river]

ako still alive

[rigid snow]

that's why i drink at least 4 cans of white monster a day

[sage crag]

are we sure

to both A and B

phrr

explosion

[tender river]

[opaque sigil]

thanks claude (you'd have to try very hard to make something less predicatable than a cuckoo filter)

[sage crag]

ye whole random

unpredictable => good in hashing

https://cdn.discordapp.com/emojis/1276149117717970964.gif?quality=lossless&name=nopy&size=64

[opaque sigil]

i need to find a filter to implement for my thesis
wanted to do a blocked bloom filter since it doesn't get much faster than that but it's already taken

[sage crag]

nopy

[keen hatch]

If the idea is a sliding window, then this is much worse than a summary - because each shift will lead to a reset of the KV cache

[tender river]

ye shiro taking too long i drink water

[sage crag]

https://arxiv.org/abs/2201.01174
just browse arxiv until you find something

[opaque sigil]

i did stumble across that one and couldn't find any gpu versions of this one so it's pretty high on the list

[sage crag]

ye write it yourself

[opaque sigil]

that would be the goal yes

[sage crag]

ye

ye.

xor filter, binary fuse filter, ribbon filter

[opaque sigil]

idk how well this lends itself to gpus since the actual construction is very much serial and only doing querying on the gpu is boring

[sage crag]

prefix filter

[opaque sigil]

i haven't seen that one

wait this sounds nice

[tender river]

btw have you tried multiple hashmaps

add to (coreIndex % hashmapCount)th hashmap
query every hashmap sequentially

should have better insertion and worse query performance

[opaque sigil]

the problem with that would be size i think

[real sierra]

[sage crag]

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

[real sierra]

what am late for

[sage crag]

ye need to ruin toast's thesis

[real sierra]

even i dont have water

[sage crag]

[real sierra]

i need food

[sage crag]

shirokomew

[real sierra]

i need water

im going to pass away

konii...

you must inherit the chess api...

please find the strange bug that causes there to be 4,865,610 legal moves at depth 5 instead of 4,865,609...

i pass away now...

[opaque sigil]

okay this does sound really interesting and could work well with the memory hierarchy, will consider

[sage crag]

ye me sell it

[real sierra]

rolling in my grave

sells for like 2 cents

[sage crag]

2 cent richer

[real sierra]

lose 1 cent to tax

[sage crag]

[tender river]

[sage crag]

!ping

!latency

where is it

[stiff mica]

Client WebSocket ping: 22ms
Discord API ping: 132ms

aww, thank you~ ( ◡‿◡ *)

[tender river]

Client response time: 17582ms
Discord API ping: 2150ms

[sage crag]

Client WebSocket ping: 0ms
Discord API ping: 0ms

[stiff mica]

🍪 Om nom nom
You've given me 3 cookies! | I've received 246079 cookies total!

[real sierra]

https://turbowarp.org/1188911953

wtf

that's a fast scratch runtime btw

that's minecraft in scratch

[sage crag]

https://experiments.turbowarp.org/pointerlock/1188911953
apparently this is better

[real sierra]

this works concerningly well

[sage crag]

its using heavy LODS

[real sierra]

programmer finding edge cases instantly

your water looks way better than mine did

[sage crag]

different project

[real sierra]

[sage crag]

ye look at featured project

[glad path]

turbowarp is actually really powerful

you can theoretically build actual apps in it

and it's also embeddable

ok i gotta agree that does work concerningly well

lemme see how fast it runs on the normal scratch engine

[pulsar edge]

[glad path]

ok it's slow on normal scratch but that was expected since the scratch engine is pretty slow

https://turbowarp.org/1201938491?size=960x720

darn

[opaque sigil]

sponsor my 9950X pls @sage crag

[pulsar edge]

yeah scratch is interpreted but turbowarp is JIT iirc

I think turbowarp doesn't support every extension too

[glad path]

wym

[pulsar edge]

I recall hearing somewhere that some extensions just arent compatible with turbowarp

was also years ago, but still

[glad path]

browser extensions?

[pulsar edge]

scratch extensions

plug-ins, idk what they're called

[glad path]

scratch has like two extensions

or three or something

[pulsar edge]

couldve sworn there was way more

[glad path]

turbowarp has like 50 extensions

[pulsar edge]

ok I might he thinking of BBC microbit then

[sage crag]

er why

you do

[final tinsel]

this sounds fun

[pulsar edge]

no it doesnt

[final tinsel]

i heard it mentioned in here before

[pulsar edge]

theres no C# bindings

[final tinsel]

nonono i meant the chess bot

[pulsar edge]

yeah

[real sierra]

make?

[pulsar edge]

theres no C# chess bindings

[ashen hare]

mini chess bot... The only answer is to force close the game 😄 😄 😄 😄

[real sierra]

make them?

[final tinsel]

well i dont know C# and since python is an option it seems fun to me

[real sierra]

i dont know C# so i literally cant make them

[alpine frigate]

Ig because too many tiny C# chess bots exists, due to Sebastian Lague's tournament

[pulsar edge]

no fun allowed

https://cdn.discordapp.com/emojis/918802255598194759.webp?size=48&name=Smadge

[real sierra]

i didnt even think of this...

[final tinsel]

i also have done very little actual programming aside from like... matlab classes and some small python scripts to reformat strings

so expect many if statements

[real sierra]

but the api is different and the token counter is different so i dont think those would be directly portable anyway

[pulsar edge]

the bindings dont exist because Ive been busy

[real sierra]

if you're lucky, someone will make them for you

you just need to find someone else who uses C#

and talk them into it

[final tinsel]

well i say very little, ive done a lot, just not very deep

[opaque sigil]

[pulsar edge]

https://cdn.discordapp.com/emojis/1106273080533327982.webp?size=48&name=gasp~1

vedal

[alpine frigate]

"Make bindings for C#. Make no mistakes"

[opaque sigil]

does llvm ship libstdc++, i forget

or do they only ship libcxx

[final tinsel]

i half expected vedal to just pop in, say nah, and then leave

[alpine frigate]

IIRC shipped seperately

[real sierra]

excited to see how randomly busy this channel becomes now

[pulsar edge]

excited to see it be less busy than usual

[final tinsel]

i wonder how many bots will have internal easter eggs referencing HBD shiro

[real sierra]

at least one

[final tinsel]

ooo good idea

[opaque sigil]

yeah i'm dumb, libstdc++ is a gcc thing

[final tinsel]

thanks shiro

[sage crag]

nyan

byan

[final tinsel]

do comments count as tokens?

[sage crag]

pyan

[hollow spruce]

Næ

[final tinsel]

oh good

so ill put the entire bee movie script in mine as a comment

[hollow spruce]

What

[final tinsel]

Y O U

H E A R D

M E

[opaque sigil]

why does _mm256_permutexvar_epi8 require avx512vbmi

[hollow spruce]

Where in ?

[pulsar edge]

using /*HAPPY BIRTHDAY SHIRO*/ var /*HAPPY BIRTHDAY SHIRO*/ board /*HAPPY BIRTHDAY SHIRO*/ = /*HAPPY BIRTHDAY SHIRO*/ ChessApi.GetBoard() /*HAPPY BIRTHDAY SHIRO*/;

[opaque sigil]

your cpu

[final tinsel]

OH YES

[hollow spruce]

This are hidden

[gritty dust]

Programming challenge O_O

wowza

welp

time to put on my big boy pants

[midnight sigil]

buh the chess api uses pththread

linux users guh

[alpine frigate]

linux masterrace

[real sierra]

@fast pagoda your job begins

[hollow spruce]

clang: error: linker command failed with exit code 1 (use -v to see invocation)

Sad

[glad path]

does the content of strings count for additional tokens

for py

[final tinsel]

true

[midnight sigil]

use a lexer to see

[glad path]

too lazy

[midnight sigil]

[alpine frigate]

Don't try to store code in variables or function/class/file metadata to be executed via an existing language feature. In particular, don't try to hide code in arbitrary-length strings or ints as pertains to your language. Writing your own VM is okay though, if you really want to go that route.
How do dumps of neural network weights fit into this route? Is compressing all weights into one literal banned?

[hollow spruce]

I use model to import all
Model.py in python

Try,
except

[real sierra]

there's a line that was pretty generally agreed upon that i had forgotten to formally add, but i've just added it:

A single literal token should not store more than 64 bits of data.

[midnight sigil]

and semaphore, yea

[glad path]

AHA

just put all code in a multiline string and exec it

and it's theoretically only like 4 tokens

[vestal zinc]

"Don't try to store code in variables or function/class/file metadata to be executed via an existing language feature. In particular, don't try to hide code in arbitrary-length strings or ints as pertains to your language." So, I can't hide the entire neural network's weights inside a single string or something?

[final tinsel]

already killed

[alpine frigate]

nyyo my plans of cheesing are deadge

[midnight sigil]

I guess I need wsl again :sigh:

[real sierra]

there's a line that was pretty generally agreed upon that i had forgotten to formally add, but i've just added it:

A single literal token should not store more than 64 bits of data.

[glad path]

then split it up idk

[vestal zinc]

mission failed we'll get em next time.mp3

[alpine frigate]

rust users in shambles. u128s are unusable

[glad path]

should isn't the same as can't

[real sierra]

i will edit to "must not", thank you

[glad path]

a single token can only store 64bit max
but what about two tokens? when you count them together, they can store over 128 (unless you use logic to deduce that's false)

[real sierra]

sure you can have two tokens and combine at runtime

[glad path]

hold on i have another way

[opaque sigil]

more like 3 cause you need a comma

[tender river]

using /*HAPPY BIRTHDAY SHIRO*/ var /*HAPPY BIRTHDAY SHIRO*/ board /*HAPPY BIRTHDAY SHIRO*/ = /*HAPPY BIRTHDAY SHIRO*/ ChessApi /*HAPPY BIRTHDAY SHIRO*/ . /*HAPPY BIRTHDAY SHIRO*/ GetBoard /*HAPPY BIRTHDAY SHIRO*/ ( /*HAPPY BIRTHDAY SHIRO*/ ) /*HAPPY BIRTHDAY SHIRO*/ ;

[real sierra]

[glad path]

so if comments don't count as tokens at all
read _file_ (or whatever its called) up to a certain point and execute that

and you just substitute '#' for ''

[real sierra]

surely you read the part of the rules about no reading files

[glad path]

it's the literal file that's being executed ._.

[opaque sigil]

what about #embed

[pulsar edge]

can I read from /dev/rand though

[real sierra]

[tender river]

single file submission

[glad path]

how about /dev/null

[pulsar edge]

what about /dev/null

[opaque sigil]

right

[glad path]

HIVEMIND

[real sierra]

[pulsar edge]

https://cdn.discordapp.com/emojis/1101577030778236968.png?quality=lossless&name=Nooo&size=48

[glad path]

what about moving something to /dev/null

is that allowed

[real sierra]

file io banned

[glass flower]

how about timetravelling.

[glad path]

what about consuming a cookie

[alpine frigate]

Do we have root access? rebind /dev/rand as the file being executed. Easy bypass

[real sierra]

if you can figure out how to time travel in 1024 tokens without breaking the other rules, sure

[glass flower]

i have no clue where i would even start with the chessbot... i suck at chess myself so that doesn't help

[final tinsel]

ahhh crap i still need to get a c compiler even tho im usin python

[real sierra]

id say do what seems intuitive

you can ask people here too, im sure lots of people will be working on it

[alpine frigate]

Random moves + don't lose intentionally lmao

[vestal zinc]

just NN everything

[final tinsel]

(this step)

[pulsar edge]

for the record, exploiting the branch predictor is basiaclly time travel

[glass flower]

bad strategy. try to loose intentionally so the other bot does mistakes thinking its a trap

[pulsar edge]

all you gotta do is find a CPU zero day, EZ

[final tinsel]

shut up dont steal my strat

[tender river]

here's a book on modern chess bots https://github.com/asdfjkl/neural_network_chess/
there is a catch: it covers way more than you could write in 1024 tokens

[real sierra]

in general, i find this a very good reference, but the learning curve varies: https://www.chessprogramming.org/Main_Page

[surreal sapphire]

does the token system not give a massive advantage for python ???

[glad path]

what is that embed lol

[glad path]

kinda does

[pulsar edge]

it does

[tender river]

you can use c++ too

[surreal sapphire]

:(

[real sierra]

if you think so, feel free to use it
the speed payoff is there

[glad path]

can I use plaintext

[real sierra]

choice of language is as much a chess bot decision as the code you put in it

[glad path]

aka literal english

[alpine frigate]

If you are feeling adventurous: https://arxiv.org/abs/2209.01506

[surreal sapphire]

I wanna go with haskell just for fun cuz I never get to use it outside of personal projects

[glad path]

and you just have a human following rule based english commands

how about that

[real sierra]

english is not one of the supported languages

[pulsar edge]

can I write my own DSL that compiles using a single compiler token

[real sierra]

i will probably not agree to add this language

[pulsar edge]

https://cdn.discordapp.com/emojis/1101577030778236968.png?quality=lossless&name=Nooo&size=48

[glad path]

what if i write the bindings for english

[glass flower]

i don't think its very much a advantage. you can make your code very small in other languages more than python. python just has a smaller base footprint. but you can make some crazy stuff in rust,c++ etc. that will easily pull down token usage

[tender river]

i will probably not agree either

[final tinsel]

BF support?