#programming
1 messages · Page 200 of 1
ah i see
if you use msvc you deserve to have your stuff break 
preproc_arg doesn't have to be valid syntax does it?
it can still be tokenized surely
yes
fuck msvc
all my homies hate msvc
is it possible to identify string tokens and add the string length to the token count or is that too language specific
within tcount I mean
──────────────────────
Group Files Tokens
──────────────────────
C 1 12718
──────────────────────```
it's beautiful
it's not possible
but i can do it

aura
including math.h makes it 24456
can also make integers count the amount of data they encode in longs
or should it be bytes to match strings 
you could probably just recursively look for preproc_args in the tree and run tree-sitter on those
that'd be cool
I think a reasonable bound is max 64 bits per token
regardless of data type

chale eef
watch someone make a program that takes an array of encoded longs, converts them each into 8 ints, converts each of those ints into an ascii character, and then joins all the characters together and runs the resulting string as code
now i'm curious how many tokens that could save you
1066 tokens from 2334 characters (for my program) to probably 292 (2334/8) + 200 (for the decoding mechanism) tokens
saying 200 for the decoding mechanism because you'd need to do that garbage in a fitting language like javascript
me when i don't have explorer.exe
no more exploring :(
I banned exec()-like functions in the rules 
but did add a stipulation that you can write your own VM if you want
awww

a custom vm is surprisingly efficient
token-wise that is
its almost like using defines but taken to the extreme
you collect the most common and vital operations and index them with opcodes
a few fixed registers lets you define minimal variables
you can map chess API functions to opcodes too
and then the opcodes themselves can probably be easily packed to one byte each, if not less
I'm watching it. It's pretty good
yes easily
hbvm has 119 instructions
it might not be worth tho to map opcodes to less than a byte because annoying to index
well I guess you can compress opcode and args into a byte
its actually very easy as long as you ensure opcodes are constant length
then mask and switch
args are much harder and pretty much impossible with hbvm
its common for vms to have 255 or more registers
if you write a VM you probably need a compiler too
workload 📈
just an assembler
ive handwritten hbvm programs before
fewer opcodes the better
ideally just zap all integer ops
and map some math ops to instructions
like log, sqrt
all this is making me wish to continue work on my nes emulator
written in c, naturally
search tree seems hard with vm
how store
https://git.ablecorp.eu/lily-org/hbvm-c
if you want any reference for a high performance interpreter i wrote one myself
ye
most non-systems vms have something similar
the more high level the vm instructions are the easier it is to JIT
i already have cpu emulation down actually 
mine's faster though 
no proof
nope but i guarantee it 
no proof required my code is simply faster

multitask
i am currently unemployed AND not in college so i have all the time in the world to write code
just be like me
duh
in your dreams
the trifecta of programming students
you can also do decompression on the binary
map most common operations to opcodes with long sequences of repeated bits

I guess that's probably not worth tho
im in my college's discord now and there are too many people i have mutual servers with 
more than 10 are in neurocord, the majorty is in the satisfactory server tho 
I don't think RLE actually helps at all
I would never join my college discord
instantly destroy my public image
rle is pretty poor on packed data like binaries
huffman coding or dictionary mapping would be better
not sure how small i can write either of those
Huffman coding would be smart yeah
there are better alternatives but its above my paygrade
buff man encoding 
vape addict
compress the data by compressing your ram sticks with a hydraulic press 
does compressing by physicall volume count as compresing the file?
it is, by definition, compressing
and the file is contained within the storage medium
ok but you're not compressing the file
you're compressing the plane of existance the file is on, not the file itself
this does have information theoretic consequences though
makes me think of degeneracy pressure
makes you think of what?

the force that keeps white dwarfs from turning into black holes
never thought the word "degeneracy" woudl be somethign to do with quantum mechanics
im required to bring "a mouse with 2 buttons and a scroll wheel" 
the only mouse i own is connected to my pc
i did cable manage everything really nicely tho 
then you know how to cable manage it a second time 
I've never found bonus mouse buttons useful
too many buttons
I can't retrain my brain to use them
need a new mouse
@tender river tcount update?
ye me neither. i use left click and right click, and once every 4 months the dpi button cuz i pressed itby mistake and have to put it back
actually i have like 3 buttons for dpi. one for up, on for down,a nd the 3rd button is a profiel switch but the only thing that does is change the dpi and the rgb
"apple switched to USB instead of their proprietary format" is the craziest part of this
modern apple could not comprehend
im required to install adobe photoshop 
pureref and sketchfab too but those are mostly cool i think
the goat adam sandal and his magic remote 
(i have never watched any of his movies)
im sad
I don't think transposition tables will help my bot
I don't have a good way to apply them
just have to pray that tree reroot gives me the performance gain I'm looking for
maybe I can be more clever with time too
the context menu didn't exist on mac os for the longest time and when it did become a thing you were supposed to use alt+click i think? and for scrolling you're supposed to manually drag the scroll bar
people definitely used macs back then so it must have been a good enough experience 
I really want to use something like transposition tables tho
thanks discord
it just seems like such a free speedup
Yo
Hmm good job
thanks. x-x i need sleep...
Mine ares has 2 am
Installing python, making from scratch, practice maths and learning physics, install pip modul , install choco, install msys2, installing node js
your doing your pip installs?
Yea
Tough stuff. I had a hard time learning all the things I needed
have fun @hollow spruce
Have headache
LOL
Install all things for 4housrs
i think mine never went away
Its super rewarding when you get to use the thing you built. Keep it up. Learn well, and bugs happen.
Im making ai from scratch 😂
But I think i will make next day or any time
Water time
No adb shell this time cursor
Cursor error update FAILED
is it being installed in the right directory?
Simply don't vibecode
chat
controversial opinion
i think using chatbots to help you program shit is ethically wrong in the same way using chatbots / image generators to help you draw is ethically wrong
so not at all, then
its not ehically wrong in itslef, the way companies get the data to amke the programs function is wrong
if you're gonna use ai generated images directly? shame on you
it is possible to make one that is ethical
too lax, it's trained unethically therefore it is unethical to use it
human effort was involved, and the work is transformative
controversial opinion: the grass is green or yellow
just train one on ethical data. the tool itself isnt inherently unethical
rather than arguing against using models trained off of stolen data, i'd argue against the existence of those models in the first place
using a model with stolen data doesn't make me a data stealer
lay that blame right on the people that stole the data
there is no point in any AI that isn't an rvc
bruh
Vocal synthesizers:
you're supporting the data stealers
Would you want to lose SV Teto?
vocal synths are fine
they're most often made with permission
and people who use it are actual artists just using it as an instrument
Vocal synths are great, they aren't like those stupid song generator AIs
i hate her, yes
You hate Teto??
why the teto hate :(
by not paying them anything and not using the generated images directly? yep sure
ok but like, if you buy a chicken mcnugget you're supporting an industry where thousands of animals live and die unethically. that doesnt mean i will stop eating chicken.
i feel the same way about ai
you're propagating the use of unethically gathered data
pretty sure they're mostly paid for by investors, not by the actual consumers
Yeah, vocal synths are essentially fancy instruments with extra parameters
then don't use that model specifically?
https://www.youtube.com/watch?v=b1HNHqSdGW4
I sure do like my vocal synthesizers
this has gotta be one of my fav covers i've ever done... and its even got special vocals! swarm is the most dedicated fanbase. ABSOLUTE CINEMA 💜
come chat w/ me on discord if u like! discord.gg/8ry5dn3UjA
NeuroSynth-BETA-3 developed by Superbox & Wispers
(actually amazing work btw)
Motifs:
Raise Up Your Bat (Toby Fox)
SUPER COOL ART BY O...
then don't use any model at all unless it's trained on data you own or have permissions to
nah bro has to get his plug in 
i can hate youtube as a company and use youtube as a platform at the same time
NeuroSynth mentioned > show NeuroSynth example song
you don't gotta pick one or the other
why do you need ai though?
i don't
gen ai le bad superbox you should sleep on the warm side of the pillow now
i never use it in the way you're suggesting
High quality vocal synthesis
That's my personal main machine learning usage
i never use it period. i don't need it, it's most often trained unethically, it's shoved in my face when it makes my workflow slower
cool
then what's the point of the whole argument, if we put the guilt onto people, who can be changed as often as gloves?
they're less easily replaced than the people that use the ai
Machine learning models have their uses, but for the average person they're not a requirement for sure and can even slow you down, so completely fair to not use them at all
and you can also just blame the company itself, usually works
look what superbox does is cool, does it matter if he technically has no permission to use the voice? news flash, neither fucking does vedal
Yeah technically if it's stealing it's transitive stealing
Stealing what's already been stolen
I don't think of it as stealing though
NeuroSynth is to allow the community to do what they couldn't before
speech synthesis is fine as long as it's either
- trained with the author's consent
- completely synthetic
don't really get the whole "that data is STOLEN it is unethical to use it" thing in the first place, if it's online and it is freely accessible then that is on you
quick tell me whos the author of neuros voice
you could argue against data from sites that gather paid content for free, yeah
Well, considering Neuro is a synthesizer already our data is 100% synthetic
actually i shouldn't mention that by name
i don't even know the author of hatsune miku's voice
ehhh ashley is based off and trained on someone real tho
Well, the data is synthetic because it's from a synthesizer
arguable, not the use of synthetic i intended
I mean how would you even generate synthetic machine learning model data for a complex model without a machine learning model?
Either way, Vedal made Neuro's singing voice, so I see no problem in making another Neuro singing voice
there are many things that are freely accessible, did you know you can watch full shrek movie on discord?
machine learning models are not inherently unethical
i wouldn't know how to generate completely synthetic speech though
why yes i did!
the thing with miku is there IS a va, with neuro its:
- a VA who recorded training data for microsoft tts ashley voice
- vedal modified the pitch parameter, transforming it so much there is little resemblance to the VA already
- not sure who exactly at this point but someone trained a voice synth model based on the vedal's variant
- this model gets used by pb to cover songs as neuro
- covers get their vocals extracted and either: an rvc model gets trained on them, or neurosynth gets trained directly
- if rvc, rvc model gets used to generate training data for neurosynth
cool, so it's someone who knew their voice would be used for speech synthesis
Yeah NeuroSynth has so many steps in between
Current BETA-series models use RVC-generated data, 1.0 and onwards will skip the RVC step completely
are you doing this because it's difficult to "annotate" covers so you plop an rvc voice onto an already existing "annotated" vocal track
or just because
or too little covers
idk how training data looks for voice synths though
i'd imagine it's this
Yeah, that's why
If it was as easy as RVC, which is plop audio in with no transcription of any kind, we'd already bet at 1.0, but for vocal synth you gotta transcribe the phonemes manually
Yeah phonemes
Pitch is automatic
There's a tool to create a basic phoneme transcription, but it won't get the timings quite right and will make mistakes
Did I mention the timings need to be accurate?
lrc extended comes to mind
Otherwise you get a model that goes T)IHAJTOIHASODXh TGOHAPISOEHtyPTOIHWEOIHTGYNIKOPLSJNM
right but that can be accurately synced by a model of some kind theoretically
Well all of this stuff needs to be manually touched even though there is a tool to transcribe with semi-decent accuracy, but semi-decent is not sufficient
The output has to be manually gone trough to fix it up
it's definitely easier to work with a mostly accurate transcription than with literally nothing
Yeah that's why we do it in the first place
Though so far Wispers has not gotten much progress after his drive with the dataset on it died
I'm sure we'll get there eventually though, and hopefully there's enough A#5 to finally get us that
does the rvc model fail to hit a#5
No, the base data lacks enough A#5, there's like one of them
i still think it could be extrapolated easily
Maybe, but so far what we've done has not been enough
If you somehow know anyone that would be willing to work on the manual parts of the data transcription that would be kinda great considering Wispers is busy these times of the year
i am very not adjacent to voice synthesis, i am kind of in a different genre
It kinda takes knowing phonetics to do vocal synthesis
I'm starting to get reasonable at it but I'm still very new to it
here's what i mean btw
holy compression
this isn't touching formants at all, audible distortion doesn't come until like +-5st
Silly
Instead of pitching up the already rendered audio if we pitch up something we use the RVC pitch parameter though
OHH then just pitch the base songs??
This is for machine learning, any even inaudible defects will be greatly amplified if they happen often
what's the problem?
That's the next thing we're probably trying
Wispers is just busy
literally every rvc tool i used has a pitch parameter
Yeah so does Applio
which just pitches the input audio
It adds to the internal pitch curve rather than the rendered audio pitch
does it
i thought it did this
Well, why wouldn't it? That's the most non-disturbing way to do it
tbf i don't have a slightest clue how rvc works
The RVC model is fed feature and pitch information, and it retreives segments of audio from its index to fill it in
Retreival-based Voice Changer
Before conversion, the pitch and feature information is inferred separately from the input audio, so the pitch can be added to before feeding
Retrieval-based Voice Conversion
Yeah something like that
it's core tech for your covers, get it right 
Eh, not any more
NeuroSynth is the new core tech
For a while though it was annoying that there was no GPU rendering on Linux for NeuroSynth, so I recently added it in myself
That sure was a lot of effort but completely worth it for over real-time speeds
No more waiting a minute for an edit to a segment to render
I can finally be productive
I think I figured it out after reading some papers on MCTS
Would you buy this
3
3
2
No
I can store unified performance data for a state in a hash table
instead of in particular tree nodes
that would eliminate a whole bunch of useless moves ideally
Is this bot or api related
We should make the most intentionally poorly optimized game engine ever
Like ever frame it redoes the entire scene
Lighting is ray traced without any of the visual benefit
Etc
probably not fun when taken to its limit
only sorting function is bogobogosort
hit first array sort and wait for 1 universe lifetime
you wouldn't get a single frame

i think the best outcome is actually a really well optimized engine that makes the problems extremely obvious
like really obvious artifacting
reversed mipmap where high res textures are only used at distance
maximum screen tearing
you could even have the world's shittiest AI frame gen for max motion sickness
@trim valve pushed fix
🙏
good job on finding these bugs btw
all very rare positions
that last position, the only legal move was for a piece being pinned to move
crazy
im installing creative cloud, i'm now an adobe user 

why
school
condolences
my uni cancelled their adobe contract this month so no free creative cloud anymore 

i use photoshop maybe once a year tbf so not much lost there
uci
ucinewgame
position fen 3kqb1r/2p4p/rpb5/p1PppP2/PB1Kn3/1P1PPp1P/1R6/1N2QBNR w - e6 0 25
go infinite
@real sierra forced en passant om
that took 8k games to find as well
huh
thats rare
it's a flaw with the pseudo-legal move finder
it's not seeing the move
wait this is so complicated
because
wtf
you can have situations where
the piece you want to capture with en passant
is pinned to the king
and then you cant take it
what a dumb rule
worst chess feature
have to throw all the other logic out the window
this is why i have a separate function JUST for evaluating if en passant is legal
but it only happens in
REALLY SPECIFIC cases
because other times you CAN take that piece
depending on the pin direction

vertical pin = en passant pinned piece capture ok
other pin = bad
shiro if its any consolation I've sucessfully played 25k games without crashing

it was just insane luck it found that position
last bug?
1B1q1b2/3k1pR1/3pp3/p1pP1brp/1pPK1PnP/1P2P1n1/P5B1/RN1Q4 w - c6 0 35 it died at 27k to another forced en passant
Player: rust-chess-bot (release) (a)
"Draw by 3-fold repetition": 22158
"Draw by fifty moves rule": 228
"Draw by insufficient mating material": 1010
"Draw by stalemate": 509
"Loss: Black mates": 754
"Loss: White mates": 769
"No result": 2
"Win: Black mates": 785
"Win: White disconnects": 1
"Win: White mates": 771
om
1/27000 chance
ok i think i fixed the legal move generator, now for the pseudo-legal move generator
i legit dont know whats wrong here

OH
no wait i still dont know
im ordering food this is hurting me
actually
im not sure ive ever had to load an en passant square from a fen
does this even work right
n 
good try lil buddy
starts working immediately
crazy
the move here, d5c6, is absolutely insane
not only is the white king in check
but the only way out of the check is en passant
and after this forced move
the pawn that just did en passant puts the black king in check
bro is a pawn
does not need to be doing all that
https://github.com/triton-lang/triton/pull/7298/commits/a5e23d8e7e64b8a11af3edc1705407d91084b01d nvidia pls 
@trim valve pushed new
i wonder if they're doing some -ffast-math-esque shenanigans
🙏
trying to decide if I should sleep or set off a new chess fuzzing round
ah yes, this error makes perfect sense! 
the guild was incorrectly identified as a guild!
save tokens by reusing existing variables

shr
i use i for all my loops already so int i already part of macro 
it not in many function defs
int x(a, b) int a; int b; {
}
apparently this is valid c
in fact it only in that function def
im wondering if there's a way to abuse it
what?
how is that valid
ye it just is
int x(a, b) int a, b; {
return a + b;
}
holy
this is also valid
does that save tokens yet
The curly braces create a scope
int a, b; --> #define ints(__VA_ARGS__) ) int __VA_ARGS__ ; {
#define args a, b
int x(args) int args; {
return a + b;
}
is uh
is argv mutable
argv is mutable but the value at argv is not
wait
it may be
its on the program stack
argv => **char
fun fact
if you compile with -std=c89 or -stc=c90
int is implicit return type in function declaration
main() {
return 0;
}
valid 

which
some macros and a few more features later and i don't think i can really push it down any further without a big redesign
just needs actual testing because i have STILL never run this thing once
good luck lol
i like how .uk tlds are just inexplicably half price of everything else
awa
ye it because this country half price

half price rooks
bro did i miss the part where it's time to make these chess beasts in the 12 hours ive been away
link
nah im just working on my submission a bit early
that's illegal
i figure ill have to spend tons of time bugfixing and doing Q&A during the contest
so i think it balances out
i also need to make sure there's no features the API is lacking
i will respond to hate mail slightly less effectively
what better way to test than leveraging it
as long as i didnt somehow miss the entire contest 
official shiro api when
modmail
nah you're good lol
shiro api is_birthday function that always returns true when
you never know, sometimes i come back and there's been a tangent about triangles for 16 hours straight and i missed everything
#define next ;goto *optable[*pc++];
#define concat_inner(a, b) a##b
#define concat(...) concat_inner(__VA_ARGS__)
#define op(...) concat(op, __COUNTER__): __VA_ARGS__ next
void vm_run(unsigned *pc) {
void *optable[] = {
&&op0, &&op1
}
next
op(return)
op(printf("hello!\n"))
}
its almost so pretty but
computed goto is such a waste of tokens

Google's Pixel 10 faces criticism for underwhelming GPU performance in gaming and editing, attributed to outdated drivers throttling the Tensor G5 chip. Benchmarks and user reports show lags behind competitors, but experts predict software updates could boost performance by 30% or more. A swift fix might salvage its reputation and highlight Goog...
Releasing a device without GPU drivers compatible with its OS is crazy work
alexa please play despacito
:juh:
google and making the most asinine hardware choices just to give people a thing to shit on every pixel for
name a more iconic duo
it done corrupted itself during realloc
well after
Non sloppa article
It stays at 396 MHz under load rn 
No wonder early review benches were underwhelming
ionno but something aint purty in the free memory list that malloc read at #6 so it's wrong somehow
sounds about right
well it cancelled properly when the other thread exploded lol but it doesnt look like anything else happened in this thread besides chilling like a boss
how broke malloc

OH
figured out
when reroot, make child new root
but didnt clear old parent of child which points to garbage now
garbage collection is fun :^)
its kinda amazing how you can always find a few more tokens
i remembered calloc exists and wiped a line
also replaced all sizeof with numbers 
god dammit of course GPG added the accidental space i entered after my passphrase when generating the key... to the passphrase
ive been so confused till i went and highlighted what i had entered in my terminal again
reeeee
?
i can fit so much code in here
that's not the kind of busted i expected
how big is stockfish
gonna be hard to write the assembly
not sure i've only written a chess bot once
nope im gonna write an assembler in python 
theres a certain irony to an interpreted assembler
little sneak peak into terror
uint8_t *pc = (uint64_t[]){ 0x0000000000000001 };
is stockfish ML
sometimes
global 
stockfish does its general position eval using neural nets i think
on top of every other optimization, of course
currently it does yeah in the new ones
yeah
seems to be the way most good chess engines are leaning
machine learning position eval
meanwhile my bot still cant figure out endgames 
its like NNUE right
yes
funny
the vm source code file is 270 bytes 
it doesnt do anything yet though...
because i dont have a certain api 
someone came up with it for shogi and they slapped it on chess and it went crazy
super small network
use this in conversation with normies: Lizard SCReLU
my friend use define mydih long long
heh heh
alright time to read up on chess engines more since it seems close and ive been lazy
no i am waiting for your api functions lol
i imagine it'll be fairly good, searches up to 7 turns ahead with pruning and an evaluation function that uses a 720 parameter FFN
hi new
shiro has been agonizing over an API for a chess bot tournament with token limitations (single source file, 1024 tokens? submissions) in w/e language that can talk to the api
nono this is something recent
so recent it hasnt happened yet
yes what afunyun said
chess bot tournament 
coming soon
chayleaf is making some modifications to the token counter
so depending on how that plays out it may vary
but its probably 1024
and certainly not smaller
can i join
no, you are BANNED from this tournament
it just hasnt started yet

you know it's 31337 if the compiler says wtf bruh
there'll be an event ping for it when it starts 
TRUE
Comments are not allowed in this language 
comments are not allowed in this language

compiler-enforced lack of documentation
as it should be
that's what separates us from the other hominid class great apes
the ability to intentionally confound compilers for fun and profit
what did i do chat i gonna cry 
being new is a terrible, terrible transgression
cry harder because you are not getting an unban
i googled opening books to find some for the tournament, how did i end up here 
(t is in a hospital for hallucinations)
the sacred texts
damn $190 for a fully featured, assembled AND tested tape drive controller with all that connectivity

where do i mail the order
oh it's at the bottom
thank goodness
wonder if they're still taking orders
maybe i should just make a silly improvised version of the api just to run tests
im going to make the only vm operand type be f64 
horror
honestly why not
index registers by f64 
the real long long
you can still type convert it cant you
ye
well i mean for the obvious case of operands im gonna do bitpacking instead of a full f64

but like, immediate operands
f64 
why do you need floats anyway...
yeah i move 0011111111110000000000000000000000000000000000000000000000000000 to e4

do you really need floats for eval
just use large ints
yeah
only chess bot ive written used fp eval but im down to write int only eval
then the remainder doesnt matter much

why do i see so much for(;;) when while(1) is less tokens

i should just assume bot never exits 
ye
hello!

hello!
what a strapping young lad
hello!
hello!
h1vm is real 
who needs registers when you have fighting spirit
if you think about it, get_legal_moves() just allocates and returns a buffer
doesnt explicitly say what to do with it
shiro i am sure your documentation will be tip top and well researched + approved by committee?
competition didnt start yet
true
i need a timeline from chay on the tcount changes
and i need bred to fuzz the api again
]
because i think i introduced a new bug somehow 
or my bot just sucks 
your malloc bug earlier could have ended up as a nice CVE
tournament ready
UCI spec does define a command
info string <yourstringhere> \n
which just prints the string to the gui allegedly
if anyone is curious, yes this does in fact use __COUNTER__

i plan on introducing chaos into my moves (theyre not wrong theyre intentionally chaotic using fuzzy logic adapted from japanese traditional art of fuzzy rice cooker logic)
rng -> forfeit true combo
oh ye also i already have the api header

collusion
it's changed slightly

big change is that the docs now use doxygen formatting so vscode renders function defs nicely on hover

bespoke
my bot just lost to the random bot
destroyed
its so annoying
he's just like me fr though
my bot is great at wiping up material
but then it gets to the endgame and just
shuffles pieces around until draw
wasnt that neuro's problem last time she was chess engining
im pretty sure she had the same issue and couldnt close a game worth shit
yea closing games hard apparently
a good heuristic i saw for endgame stage is to reward the enemy king being far from the center
but that kind of mechanism doesnt make sense in my framework
i will take that to mean heuristics in general are incompatible with your framework 
thats pretty much what it means outside of some things
better than loss
he's just a baby maybe something made it think it was boutta lose
the one heuristic left betrayed him
literally
got nerves due to winning too hard
ff15 boys pack it in
but my bot pulls this shit
5 minutes on the clock and it spends under a second on each of those moves

throw
malicious tournament denial of service through draws
i just need my bot to not be scared of winning
trap the vmbot in the eternal draw dungeon
maybe he just never thought he'd have such a good position so he shit himself
happens to the best of us
it's only human
i think this is a bad change
the draw propogation is important 
a move frequently leading to draws is info too
i wonder if putting a hard cap on the search horizon would help
i can only think that something about how it evaluated that particular board position is off bigtime
if it's properly evaluating a draw as like 0 it has to have somehow came to the conclusion that moving queen to f7 is worse than that
i think it might actually be too shallow of a search being the problem
theres probably a bug somewhere i cant find
you know what
how about i just count draws as bad outcomes for everyone
bad for me
bad for you
nobody wants a draw
out for blood
now that's thinking with your sledgehammer
but yeah i think if it was searching for possible moves after queen->f7 it must've hit some weird scary spooky board state before it ever found the correct queen->f8 win condition
well i didnt learn anything from this game 
its possible
im not using a traditional search tree with iterative deepening
in particular my algorithm is very up to random chance
ok i now want to have a diffusion sampling based bot
oh
no i know why i cant win endgames
its my tie breaking method
basically
if we reach a stalemate position but have significantly more material
we consider it a win
because something went right and we can probably win with at least part of that move sequence
well no this is separate from actual draws ish
this is just counting some draws as not draws
before i was talking about how i handle actual draws
but uh
the problem with this method is that it looks at absolute material instead of a material delta
and in the endgame after i eat all the enemy pieces
i always have way more
hoarderbot9000
so every move sequence, even ones that draw
count for max points
i probably need to eval my current material at the start and look for change in material
man he was absolutely pogging after that draw and you just ruined his whole parade yelling at him
poor thing
dumbass bot
man i don't wanna make my own board functions for this.
for what
no testing until the contest starts
smh you asked me how good it was
i know i keep forgetting
finish the api smhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhh
wait that's it?
oh
theres very few left i think
but maybe 1 or 2 somewhere
that or my bot is broken

likely the latter
thinking about it honestly draws being BAD rather than meh is probably good in the context of how limited the bots themselves are able to be within the token limit
new 
improved 
worse than random

between what stockfish and lichess ? lichess uses stockfish for the analysis board
16 tho
negative = my bot winning
prodigy
Shiro is so intelligent

none of my bots moves are what stockfish suggests 
fuzzy logic
Shiro wtf is this
oh cool android finally has react native
discord patch notes
the android discord app has been a dumpster fire for years
Developers showing soo much life and excitement 
its because the devs dont actually work on discord
all they do is write the patch notes 
the interns are writing the updates
and those interns? believe it or not they're delegating the work to gemma 3n (discord doesnt make money so they cant pay for more)
This option is sooo shit in samsung. All this does is delay the app installing for another 45 seconds. It's soooo slow
mcafee

What is that
also apparently disco has been on react native for a solid minute on android and i missed the boat//didnt care so theyre just talking about react native's new archi (which it says but im blinded by bias)
Koni cutie! 
my bot just had to make this one move and it was mate
basically react native got refactored internally to be better ™ in a lot of ways that dont really matter here
they updated to use the new version so theoretically discord should work more betterer
but naturally
I think 98% done
shiro i think your bot is stupid
1024 tokens of stupidity
and react native is just react (a web UI framework) but it works with native platform integrations and thus is a lot faster
my bot could beat up your bot
True
say what you will
my bot could too (random moves)
my bot is legendary at erasing the enemy pieces
its just
very scared of winning for some reason
hello!
and actively avoids checkmate?
You used chatgpt 
i didn't
at any point
koniigpt

me helpful
Wtf is mean
Bot technically avoided winning
bot could win in one move
but did something else
i already had an algorithm in mind before starting this, as a response to shiro chessposting in general
i had chatgpt playwright browse a documentation site yesterday cuz they blocked fetch requests so it just sat there extremely slowly navigating to each page and pooping out markdown versions
that was fun
tgpt
gpt without the gp
i'm not ai smh

i don't generate anything and i'm not pretrained
GPt: Grandmaster Player t
i am both pretrained and generate garbage via my mouth and fingers
Chatgpt is master piece who changed the world but i didn't agree ure code fully relat on chatgpt
noooo
do bots need healthcare
Yes
i'm not a bot
what is the point of having claude replace my entire employeebase if i still have to pay for healthcare
Server is running right now


















💬



and then my bot makes a legendary fumble
