programming

[uneven pulsar]

done all the ways to use arrays

or modify for whatever

nvm missed some stuff, now i did them all

nighttime so i'll have to code tommorrow now

[glad path]

NO NOT JAVASCRIPT STAY AWAY FROM JS WHATEVER YOU DO

[uneven pulsar]

why

its just code

[glad path]

why is there a chatgpt tab titled "adrenaline..."

[glad path]

javascript worst language

in my opinion atl

[uneven pulsar]

needed to ask questions about my consistant adrenaline issues <3

[uneven pulsar]

oh

noted

i will not do javascript

you're right <3

[glad path]

i mean you can keep doing it

but like

idk

i just dont like it personally

[uneven pulsar]

nah its okay

[glad path]

its just a weird language

[uneven pulsar]

i see

i'll like stop doing js

[glad path]

probably learn it though

[uneven pulsar]

and do python

[glad path]

since

its still useful

[uneven pulsar]

or just not code at all

[glad path]

all languages are good to know

js is not an exception

though its still not my fav language

[uneven pulsar]

i see

so should i code or nah

i mean idc

[uneven pulsar]

@glad path

[glad path]

yes

[uneven pulsar]

i'll go to bed after

[uneven pulsar]

yippie

[glad path]

good skill to have

[uneven pulsar]

thank you

[glad path]

has actual function benefits for me

aka cobbling together random crap to make things

that

work

[uneven pulsar]

real

[glad path]

also i made an autoclicker that works like tinytask once

because there was nothing for mac

[uneven pulsar]

lol

[uneven pulsar]

that sounds pretty cool

[glad path]

(it works)

[uneven pulsar]

as long as it works

nothing else matters(big mistake here)

[glad path]

ok i slepe

[uneven pulsar]

gn

[thick thicket]

[obsidian mantle]

Oooooh

So iterative search restarts the search from depth0 every time but since we have transposition tables it isnt bad since progress gets easily restored?

I thought it somehow restores the boards and continues from where it stopped

And i just realised my "depth" gets incorrect after turn ends so i either have to clear the table or shift depth of entire table

What if i remember the turn number for transposition

[obsidian mantle]

It probably isnt beneficial to remember all tables of all turns since it gets too huge and searching becomes too slow

Thats why i had no performance boost

[jagged turtle]

[nocturne olive]

I have not the slightest clue what I'm doing

Except that it involves a ton of token saving Kotlin STD abuse

[obsidian mantle]

when (it)

I sooo much want to quit my job and start learning programming

[midnight sigil]

you should make your material values follow the Simplified Evaluation Function page in chess programming wiki

[obsidian mantle]

This?
P = 100
N = 320
B = 330
R = 500
Q = 900
K = 20000

[obsidian mantle]

Is this + static position of pieces good enough

[nocturne olive]

Yeah and it is of type PieceType

[nocturne olive]

Wha?

[obsidian mantle]

I think its about 1 3 3 5 9 not being optimal + you want to evaluate the position of each piece

Knight in corner is bad gives it -50

[nocturne olive]

I have not the slightest clue how to evaluate positions of pieces

[obsidian mantle]

And if you use 1 3 3 5 9 you need floats to add position value

Here are the tables https://www.chessprogramming.org/Simplified_Evaluation_Function

It also explains it

The problem is to pack all that nicely to save tokens because position tables are big

128+ token per table and we have 7

Some guy who did lague's challenge - packed them into numbers and mirrored them, extracting the value later using some formulas

Which seems to be a nice idea

[midnight sigil]

bitwise operations

quantize the tables into 64bit arrays

[obsidian mantle]

Huh

I only can think about packing it into bigger numbers

[nocturne olive]

Well given the tables in memory determining the piece values seems easy enough as long as I determine where each piece is as an index, then I can simply build another few numbers
The hard part seems to be getting the tables into memory in the first place

[obsidian mantle]

For example each position is 1 byte so you pack it in 8 bytes to save x8 space

[obsidian mantle]

You mean array of 64bit numbers? Did i describe the same

[nocturne olive]

Provbably will want a functional bot of some sort before I start to tinker too much with that

Currently I do not have a bot that functions

It would barely even compile

It lacks the entire search function as well

[obsidian mantle]

Use negamax its op

Also very upgradable

And all strategies you can use are out there

[nocturne olive]

I think in its current state if it does anything at all it will evaluate once and then pick whichever move gets the highest score

[nocturne olive]

This stuff is confusing
The search algo I saw was alpha-beta or something but then again I'm just reading the Chess programming wiki with no clue what I'm doing

[midnight sigil]

yuh

negamax is the searching method

alpha-beta is the pruning method

[nocturne olive]

Wha

[midnight sigil]

you use pruning methods to prune node trees for the search

[nocturne olive]

Now I'm even more confused

[midnight sigil]

see chess programming wiki

just, read

it has everything you need and to know

[obsidian mantle]

Chess programming is wall of text, you can watch Lague's videos as an alternative

But chess programming should be more informative overall

If you have time to read it

[nocturne olive]

Well so far I've just poked at this Chess stuff at random times when I happen to have the motivation and time to throw together a function

I grab a page on the Chess programming wiki that seems good enough, then throw together an implementation

Well, I have done that once now, for the evaluator

And then after a baisc implementation I stripped out all the extra tokens

[midnight sigil]

I wonder would there be anyone using MCTS instead of NegaMax

[obsidian mantle]

???

Shiro uses it

Is it that bad?

[midnight sigil]

Monte Carlo Tree Search?

[obsidian mantle]

Yes exactly that

[midnight sigil]

based

[obsidian mantle]

Is it superior

[midnight sigil]

idk

[obsidian mantle]

We'll see i guess then

[midnight sigil]

both has pros and cons

[obsidian mantle]

I wonder how its pruned

[midnight sigil]

you can do hybrid, technically

[midnight sigil]

just the same way

[obsidian mantle]

What "value" do you assign

It cant play through every possible scenario

[nocturne olive]

Anyway I'm just clueless
I just have a really compact evaluator

[obsidian mantle]

Or does mcts not necessarily go to the end of the game

Wait it plays random moves, right?

And you get % of wins updated constantly

And prune bases on that?

[nocturne olive]

There's probably some way to compact it though, I don't know

[obsidian mantle]

It seems pretty compact but you definitely want position of pieces evaluated too

[nocturne olive]

Eventually when I have an actually functional bot

[obsidian mantle]

Because otherwise it wont want to move pieces in good position but rather trade them in its favour

[midnight sigil]

https://www.chessprogramming.org/Monte-Carlo_Tree_Search

again you have everything here

[rough bloom]

unless I'm misunderstanding MCTS there's no need to prune the tree at all
since you just sample random branches you can exhaust your whole time budget repeatedly sampling to get a more accurate distribution of what move is best

though pruning would still make it more efficient because the search space becomes smaller

[midnight sigil]

you can do search guiding first I think

[stray dragon]

somehow, doing chess_push with a valid move causes the bot to crash?
and this only happens when i don't have a hard cap on the depth of the quiescence search?

guess i check the components of that *bestMove

[midnight sigil]

might be because you pushed an empty pointer idk

it happens to me when I pushed a null ptr to the api

[stray dragon]

hmm

[midnight sigil]

it repeat the previous move you pushed onto the stack

[stray dragon]

it might be doing that then

pointers are a pain to deal with

[midnight sigil]

real

that's why I use Python

[stray dragon]

*bestMove seems to have the right info?

[obsidian mantle]

Maybe it crashes later

What happens when you dont push at all? It should give illegal a1a1 move

[stray dragon]

if it did that, it would still print "test3.6" to the console before dying

[obsidian mantle]

Oh right

Wait

No

It has delay

Cute chess can discard outputs from bot right before it crashes

I encountered that when tried to stop bot by dividing by zero

It didnt send the message right before division

If you can, try running it in console

[stray dragon]

hmm

[obsidian mantle]

All you need is give it these commands:

position startpos
isready
go

Iirc

Maybe just "go" is enough

Didnt try

[midnight sigil]

no

position needs to be initialized

position startpos required

[obsidian mantle]

Understandable

[stray dragon]

clearly not a null pointer and clearly the move info at the pointer is all correct and clearly it works for other moves before it with no issues

and sometimes it does manage to make a move

and it's also strangely non-deterministic.

[obsidian mantle]

Something isnt getting cleared?

[stray dragon]

maybe

might be why it's acting non-deterministic

but it's been doing that for a while, even before i made the change that introduced this issue

[obsidian mantle]

How do you manage depth

How does it know how deep it is

[rigid snow]

fish finder

sonar

[maiden geyser]

position location needs to be position positioned initialized

[rigid snow]

tomato

[obsidian mantle]

What happened to evil yesterday? I missed the moment when it started. Was it random or vedal did something?

[stray dragon]

depth goes up as it searches deeper trees, and is soft-capped by a "searchDepth" variable that increments in a loop
it does a recursive negamax function with some pruning inside, and every recursion it increments the depth variable it passes to itself
this issue arises when i change it from a hard-cap to a soft-cap for the depth searching

[obsidian mantle]

I ask because i store current depth as global variable and i had problems when it wasnt --'d or ++'d correctly

[midnight sigil]

[obsidian mantle]

I guess not your case at alll

[stray dragon]

searchDepth is the global variable in this case

depth is passed from negamax to negamax

[obsidian mantle]

I need to change this bullshit to something more appropriate

[obsidian mantle]

Wait..

So it might be your case too?
Every return must decrease depth
Every search enter must increase depth

I had problem when not every return has been decreasing depth so it got all mixed up

Uuh the more i describe it the more i want to change it

[maiden geyser]

https://www.youtube.com/watch?v=vix7cJ8FVr0

evil lobotomy

[stray dragon]

return doesn't decrease the depth, since it's not a global variable, but one passed between negamax functions

[obsidian mantle]

This sounds so odd

[obsidian mantle]

Alright then it isnt it

[stray dragon]

and only one place where depth increments in the negamax function

hmm.

[obsidian mantle]

Printing boards might help too

At least final board

[stray dragon]

i wonder if this issue arises when i allow it to do quiescence search to a hard limit that is past searchDepth

[obsidian mantle]

Maybe you dont undo your move somewhere

[stray dragon]

shouldn't be an issue, i reset the board just before pushing the move to it

[rigid snow]

sf is not a real place

[obsidian mantle]

I think pushing move doesnt touch any boards at all

Only making, skipping and undoing does

But since the move itself is legal.. weird

[stray dragon]

allowing it to go to a depth of 2 past searchDepth works perfectly fine

lol

how far can i take this...

[rigid timber]

it touches the internal board when you commit the move

[viscid surge]

just curious, what type of loop you guys prefer

// 1
auto value = something(x);
while(value != 0)
{
    // stuff
    value = something(value)
}

// 2
auto value = x;
while(true)
{
    value = something(value)
    if(value == 0)
        break;

    // stuff
}

// 3
for(auto value = something(x); value != 0; value = something(value))
{
    // stuff
}

// 4
auto value = x;
loop:
value = something(value);
if(value != 0)
{
    // stuff
    goto loop;
}

[rigid timber]

then again that should probably get replaced with a different board for the next move...

[midnight sigil]

1 or 3

[clear sedge]

1 exposes the variable beyond the scope where it's required

cringe

[viscid surge]

all do that except 3. Though you can always surround the whole thing with another bracket if that's the issue

[midnight sigil]

location location locationing

[clear sedge]

then it becomes clunky
3 is specifically made to avoid such clunkiness so

[midnight sigil]

[clear sedge]

my opinion? if you use 4 you're a psychopath

[viscid surge]

The only issue that I have with 3 is, very long one line.

[obsidian mantle]

Whaaat
Why?

[rigid timber]

well it has to store the move somewhere, doesn't it?

at least I think it does, it's been a bit since I saw the code

[obsidian mantle]

No it just cout<< it

[midnight sigil]

for me 1 is more readable

[obsidian mantle]

I mean i suppose

You have entire move to be pushed in your hands

[clear sedge]

for(auto value = something(x); value; value = something(x)) { ... }```
![ReallyInnocent](https://cdn.discordapp.com/emojis/1229173141088112640.webp?size=128 "ReallyInnocent")

[obsidian mantle]

Why do you need a board to print it to the api

I had problems in the past when i used push instead of make move to affect the board

[viscid surge]

It's good except: it's easy to make infinite loops, you can't use continue keyword etc.

[clear sedge]

yes

i was gonna make that point

[rigid timber]

well that would make sense bc push doesn't actually affect the board

[clear sedge]

3 allows for continue; without repeating the assignment

[rigid timber]

only done() or whatever it's called does

[rigid snow]

just

for (
  auto value = something(x);
  value != 0;
  value = something(value)
) {
  // ...

[obsidian mantle]

What about
While (auto value = somethinf(x))

[rigid snow]

that wouldn't work would it

[obsidian mantle]

Does = give 1 on success

I think it does

[clear sedge]

x = y returns y

[obsidian mantle]

[obsidian mantle]

Then it should work?

[viscid surge]

It would had to be

auto value = x;
While (value = somethinf(value))

not sure if this is legal?

[obsidian mantle]

Looks legal to me

But i dont quite see whats happening here entirely

[rigid snow]

no i think this should work on second thought

[viscid surge]

It's infinite loop as x never changes

[clear sedge]

[obsidian mantle]

But same here

Yeah if x isnt changed then none of those variants end?.

Unless f() gives different results from same input using some external shit

[midnight sigil]

let's see how much null pruning would help with the speed

[obsidian mantle]

External global shit

[midnight sigil]

I don't care my bot is 3000 tokens now

[clear sedge]

state machines 😋

[midnight sigil]

I'm heading for 4000

[rigid snow]

no they do actually, it's value = x, then value = something(value) in all of them

[rigid snow]

so this wouldn't work after all

[obsidian mantle]

Without x of course

Right

I expect x to be modified in unseen area

[viscid surge]

[clear sedge]

while(auto value = something(x--)) {
  ...
}```
![Clueless](https://cdn.discordapp.com/emojis/1167230304554066040.webp?size=128 "Clueless")

[obsidian mantle]

By unseen area i mean these "..."

[viscid surge]

why are you trying so hard to validate your code?

[clear sedge]

i'm pretty sure x is meant to be a constant expression

[rigid snow]

ye

[viscid surge]

x here is just like in math, unknown, might be constant, might be function call, might be some other variable, needed later so it can't change

[rigid snow]

then 1 or 3

[clear sedge]

1 is respectable

3 is preferred for me

[viscid surge]

I'm kind of starting to like this idea, i think it's the shortest one ?

[rigid snow]

it should be

[clear sedge]

relying on the return value of the = operator is some javascripty shenanigans

[clear sedge]

also

for(auto value = x; value = something(value);) { ... }```

[viscid surge]

You really like the for loop i see

[clear sedge]

what can i say, i like my scopes

[midnight sigil]

[midnight sigil]

holy frick just the PSTs took me 1860+ tokens

@real sierra can you undo a turn skip(via chess_skip_turn(Board *board)) with chess_undo_move()?

or I have to store two board states

[rigid timber]

not Shiro but should be able to... skip just makes a null move...

[obsidian mantle]

I do it all the time it works

[real sierra]

yes

[obsidian mantle]

It also is stated that it can be undone in c api

[midnight sigil]

👍 thx

[obsidian mantle]

Wait is it potentially faster to do iterative search using cloned boards

[midnight sigil]

if you bothered to do multi threading

maybe

[obsidian mantle]

Not allowed

I mean if i don't just iterate from depth0 every time but actually continue from that board

Or is it not doable

[midnight sigil]

smh

[midnight sigil]

more memory would be used

[obsidian mantle]

What was the limit on memory?

[midnight sigil]

I don't think there is

[obsidian mantle]

I remember shiro specifying it at least in this chat

[rigid timber]

a few gigs iirc

[obsidian mantle]

How many boards can i store in 4 gb

[rigid timber]

quite a few

[obsidian mantle]

Hmm is it really

[midnight sigil]

yuh

not many thing you can store

[obsidian mantle]

Board is what, something around 20-30bytes?

The wrap around it gonna take tokens though

[rigid timber]

16 64-bit integers, 5 booleans, 3 ints, one pointer (not in that order)

plus whatever padding the compiler decides to put there

[obsidian mantle]

145 bytes

[rigid timber]

better to do sizeof(Board) somewhere to get an exact value

[obsidian mantle]

27 million boards

Uuh what depth is it

[rigid timber]

depends on the board state

if you have one king and one pawn, very deep, if you have pieces all over the board, quite a bit less

[obsidian mantle]

Depth 5 for 20turns/layer

Trash

Also i need to clear it after search

[rigid timber]

prune the obviously useless ones?

you can't exhaustively search every single move, you'd be here until the heat death of the universe and consume more memory than you have particles in the universe

[obsidian mantle]

However my current pruned search goes in 4 depth easily

Maybe actually enough

How to wrap this though

I just... Iterate through each saved board

Needs testing

Maybe when im done with all basics

And see if i have tokens for that

I dont need to clear array actually

Just remember position where to write and from where to read

Bot allocating 4gb ram on launch

[midnight sigil]

why is beta highlighted

is beta some keywords in py

[rigid timber]

not as far as https://github.com/python/cpython/blob/3.14/Lib/keyword.py is concerned

[obsidian mantle]

Build it

[viscid surge]

warning C4706: assignment used as a condition

[opaque sigil]

It's legal but uhh yeah

[obsidian mantle]

Warnings who cares about warnings

[stray dragon]

add 3 or 4 new features to bot
can't tell if it's any better than a past version of the bot without the new features
haha fuck

[rigid snow]

it because people oftentimes small brain and do a = b when they meant to a == b but it's intentional in your case, ignore

[obsidian mantle]

I slept 3 hours today
But im not too tired yet
surely it will not kick in in several hours

[uneven pulsar]

Time to plug weed into the twins

[faint sandal]

[midnight sigil]

I wonder the visual of this

[dry charm]

https://cdn.discordapp.com/attachments/1072697081443131476/1422537017006227466/neuro_weed.gif

idk why or who made this

but perfect

[faint sandal]

lmao

I'm saving that

[umbral thorn]

"not compatible"

microsoft want me buy

their shit pc

fodder pc nuh

me pc better

[maiden geyser]

https://tenor.com/view/fact-check-false-clio-clio-vtuber-phase-connect-fact-check-phase-connect-gif-11765976316029522537

[crimson dagger]

i though i could try finishing an old project but when i tried running it everything broke

[crimson dagger]

Oh god i remember what i did.
In that program i used all methods of comunication , in it i found:

• FastAPI
• WebSockets
• Sockets

And probably even more

its a mess lmfao

[fast pagoda]

anyone know exactly what services these credits are good for before i spend $500 and find out they dont apply

idk why i have this but ill use it if theyre just saying here u go

[fast pagoda]

(vertex ai on googly cloud)

oof

oops

[crimson dagger]

nice ping

[fast pagoda]

yeah i missed like a boss

i need ai to help me aim ya know so

that's what ill use these credits for

it's more fun if you have to spend 5x just bugfixing clumsy attempts at wiring together things that really don't want to be talking to each other

5x the time*

[crimson dagger]

Anyways time to spend 5 years debbuging to find out what the hell is going on in this project

[fast pagoda]

i thought that was what was going to happen to me when i launched echo's bot on linux

since he used like

.ps1s for a shocking amount

but it actually works fine after like 15 mins of messing with it, i was amazed

mongodb didnt notice, the discord api was just fine, even the powershell worked cuz i got bored remaking things that already worked and just installed powershell and swapped any volume paths

new venv for uv and worked just fine

and then i was like i didnt expect that to be easy that's not fair

and havent touched it

and now i can make people slightly sick by showing them this

[opaque sigil]

(that'd be me if using pwsh wasn't such a pain on linux)

not pwsh's fault tho

[fast pagoda]

it's not that bad, although admittedly these scripts are very simple so they were mostly fine

[opaque sigil]

[fast pagoda]

they were just used to like launch other things like mongo started from that as did the inference server and then like some backup stuff when it shuts down or invokes it while running

me when i subprocess.run(powershell_command,

[dry charm]

are you sure you want to update your PC to win11

just upgrade to LTSC

[fast pagoda]

just install linux

[dry charm]

Elvyn and linux

[fast pagoda]

ive been using since 2021 and i still swapped entirely

it's the year of the linux babeeyyyy

or something

[dry charm]

Sadly not everyone can do that

[fast pagoda]

surely everyone is me and has the same usecase

[opaque sigil]

[fast pagoda]

i would dual boot if i had that requirement tho prob

[maiden geyser]

barely 5% marketshare

[fast pagoda]

TH EYEAR OF THE LINERXXXX

i think every time linux gains .002% share it becomes the year of the linux desktop again

[dry charm]

@umbral thorn how mad Quack would be if you were to nuke windows and install arch or smth

[opaque sigil]

clearly very mad because it should be nixos

[rigid snow]

i updated my boards bios last year and it became compatible and auto downloaded the win11 update and then was like “we already got you just like uhh press this to install w11 instantly”

[opaque sigil]

how does a bios update make it compatible all of a sudden

[rigid snow]

i don’t know

[umbral thorn]

my pc

[umbral thorn]

no ew

[rigid timber]

well enjoy no security updates I guess

[umbral thorn]

tweak nvidia gpu in linux is hell

[dry charm]

Just don't get hacked

[rigid timber]

I don't see what everyone still has against Win11, it's literally the same as Win10

[fast pagoda]

the secret is it's still win10 (same kernel version) (someone will see this message negatively somewhere probably)

[rough bloom]

NixOS dual-boot

[maiden geyser]

maybe it reset some tpm setting back to enabled or something

[dry charm]

If you use NixOS, it is easy cause you can write a config and it rebuilds the whole system

[fast pagoda]

usually the restriction was literally tpm

[rigid snow]

tpm was enabled before the bios update btw

[fast pagoda]

and it may have already been a thing on your board but not enabled (virtual tpm)

[opaque sigil]

true actually

[dry charm]

(I say this in my best wishes, I would have had to reinstall arch or other system multiple times already if I weren't using nixos)

[rigid timber]

there's all the same telemetry, all the same shit preinstalled software, just the number is 1 larger

[fast pagoda]

btrfs snapshots do the job just fine for me tbh although i see the appeal

but then you gotta use btrfs which sucks

[opaque sigil]

it's fine as long as you don't use gnome

[dry charm]

I would forget to take a snapshot before doing shit

I'm using BTRFS actually!

[opaque sigil]

timeshift hook

[fast pagoda]

pacman hook that runs a snapshot

erry time

[rough bloom]

actually true
the UI is subjective but from a technical standpoint it seems pretty much purely better (other than the normal incremental Windows update bloat)

[dry charm]

    "/" = {
      device = "/dev/disk/by-uuid/0e32c2d8-9784-45ed-a3ac-f0b644e93115";
      fsType = "btrfs";
      options = [ "compress=zstd" "subvol=root" ];
    };
    "/home" = {
      device = "/dev/disk/by-uuid/0e32c2d8-9784-45ed-a3ac-f0b644e93115";
      fsType = "btrfs";
      options = [ "compress=zstd" "subvol=home" ];
    };
    "/nix" = {
      device = "/dev/disk/by-uuid/0e32c2d8-9784-45ed-a3ac-f0b644e93115";
      fsType = "btrfs";
      options = [ "compress=zstd" "noatime" "subvol=nix" ];
    };

[rigid snow]

nyooo muh start menu react native react bad web bad literally electron

[clear sedge]

i finally have a triangle

[fast pagoda]

i see your fancy fstab and i say

[fast pagoda]

reject electron embrace avalonia

[dry charm]

it's EdgeWebView actually! I was talking with someone that needed teams and I was like "You can just install it, it's a webapp anyway"
And nope! Since it uses EdgeWebView shit, you cannot just use it

and it doesn't even work in wine

[rough bloom]

unable to use Teams
lucky

[opaque sigil]

https://github.com/IsmaelMartinez/teams-for-linux ?

[rigid timber]

that's hardly Win11's fault

Microsoft's fault 100%

[rigid snow]

rn is basically the same thing as avalonia except fucking better as rn doesn’t use skia which means react native is actually native ui while avalonia is like flutter

[fast pagoda]

nightmare

[opaque sigil]

sadly i need to use teams

i guess you can also use the old af deb from when they still shipped anything

[rigid timber]

I'm lucky (debatable) to only have to use Google Chat, not Teams

[rigid snow]

js react oooo scary

[rigid timber]

C devs when JS

[fast pagoda]

-# <-kinda likes flutter for some reason still

i really dont know why

[opaque sigil]

god damnit, i went to check the aur to see how old the teams deb is but it's down

[fast pagoda]

i cloned the github repo and am just having it sync that periodically

there's also an aur helper that uses the gh repo directly instead u can use

[rigid snow]

it is cool ngl but if we're speaking bloat purely which we are it's objectively worse

[fast pagoda]

https://github.com/archlinux/aur aur backup

[obsidian mantle]

i wonder if i can setup github to do version control on at my job
the management will be against probably since storing shit in foreign cloud is against our company's policy

damn it probably would be pretty convenient

[fast pagoda]

https://github.com/ryk4rd/grimaur the helper that uses the gh repo instead of aur directly

[obsidian mantle]

i can put any files in github right?

[rigid timber]

technically yes

[fast pagoda]

yea they get mad if you go more than like 10 gb

in a repo

[opaque sigil]

ah okay it was end of 2022 for the last teams version then

been a while damn

[obsidian mantle]

i probably need 1-2gb

[rigid snow]

sooo uhhhh are they going to load

[obsidian mantle]

hmmmm

i think i will try it and then ask for permission

wait what if i use github desktop

[opaque sigil]

git lfs exists

[rigid timber]

lfs is paid on GtHub iirc

[obsidian mantle]

what ifs?

[opaque sigil]

you get 2gb free

[rigid snow]

git lfs

[rigid timber]

that's not really all that much tbh

[obsidian mantle]

i dont speak programming abbreviations

[fast pagoda]

took an irritating amount of time for this to calculate this but the whole AUR is only about 2gs

[opaque sigil]

yeeeeah

[rigid timber]

git large file storage, special way that git can handle large files

[obsidian mantle]

oh so its 2gb free

not 10

[fast pagoda]

i pulled 10 out of my ass

[rigid snow]

it has always been 2

[fast pagoda]

i thought it saw 10gb thrown around in their docs

i was clearly hallucinating

[obsidian mantle]

its stil enough probably

[rigid timber]

might be 10 for sth else, they have so much other stuff too

[fast pagoda]

i was recently angrily in the gh pages docs too much so my guess rn is that i was thinking of the 10gb size limit for .tar uploads as pages aritifacts

[dry charm]

Yeah, but that is a wrapper for web teams. Which is honestly better at this point

[midnight sigil]

classic 141365 branches

[fast pagoda]

if ur repo has a gh pages site attached you could get away with scamming them packaging ur backups as an artifact lol

yeah it's 82 gazillion branches 1 for each package

but each branch is only the PKGBUILD and the metadata for the aur listing

[rigid snow]

how are you going to make the action fetch your shit

[fast pagoda]

it lets you upload static stuff to package in

[rigid snow]

make it connect through ssh or something

[rigid timber]

oh, 10GB is also the on-disk size limit of .git

[rigid snow]

upload files as input to an action?

is that a thing

also gh pages is only on public repos if you're on the free plan

[fast pagoda]

https://github.com/actions/upload-pages-artifact

[rigid timber]

and there is actually 10GB of LFS free, it's 2GB per file that's the limit

[rigid snow]

you don't understand what i'm asking you. how do the files get to the action. do you "host an action runner"?

[fast pagoda]

yeah i am thinking you'd have to upload them to the repo anyways

which defeats the purpose

[rigid snow]

yeah

[fast pagoda]

idk i dont try to host files on my repos so it was more of a showerthought lol

wonder if you can rsync in a runner

[rigid snow]

you absolutely can

[obsidian mantle]

i like how my console keeps previous command from before computer reboot

was it always a thing, or is it wsl, or is it win11?

[fast pagoda]

so yeah there you go

i guess rsync the bulk to the runner itself as it runs

sounds stupid

[midnight sigil]

free storage space

[fast pagoda]

it's free except copilot eats it

84 bazillion times

[midnight sigil]

I don't mind polluting copilot's codebase

[fast pagoda]

yeah poor guy would be a lot better without me that is true

[clear sedge]

i accidentally made a cool effect (flashing light)

[fast pagoda]

each time ive set up workflows i have hated every second so i dont really explore the possibilities there kek

[fast pagoda]

seizure warning

lol

sheesh

that looks like the tunnel in my dreams

[clear sedge]

[fast pagoda]

actually reminds me of miside

[clear sedge]

it reminds me of undertale's barrier thingo

this is all it does

i just don't clear the framebuffer so it renders on top of itself like a goldsrc map

[fast pagoda]

make the colours super obnoxious and put it all over the place in whatever you're making and then claim they're unfairly targeting you when people complain

[midnight sigil]

are those gradient steps the video quantization or it's designed like steps

I guess it's quantizatons then

[clear sedge]

it's literally just two triangles

[glad path]

same

[fast pagoda]

my fire alarm is going off and it's so fucking loud

and it's a test i knw for sure

i should be able to sue

What it's like to chew five gum

that went for 10 minutes what the fk bro

im pretty sure my cat doesnt have hearing anymore because i sure dont

[obsidian mantle]

ok i think i fixed transposition tables finally

gives same result as without it, before this fix it was losing

[fast pagoda]

add a fire alarm that goes off whenever it's losing* like that and it'll be fixed in no time

[obsidian mantle]

ewww i remember we were having breakfast in some cafe with the guys i was working with
and the fire alarm went off.
it was something like that very unpleasant
they burned something in the kitchen and didnt know how to disable the alarm some random person had to help them disable it

[fast pagoda]

me when i cook a frozen pizza

[obsidian mantle]

wait what does your alarm do

oh it was literally fire alarm

i thought it was somehow related with the screen you are recording

"disk bad ALARM ALARM"

or is that ram

[fast pagoda]

top left is ram right is a disk and bottom is internet traffic

it was right when i tried to run pacman -Syu so at first i did think it was my pc or something

i was like damn it's never done this before

the aur must be REALLY down

that's probably what it sounds like at aur's host during the ddos anyways

[obsidian mantle]

wait a second

something is super wrong

iterative search runs thousands of times with max depth going to thousands????

ok transposition is actually wrong

it cuts off all next iterations

chess hard

[obsidian mantle]

https://media.discordapp.net/attachments/1072697081443131476/1422537017006227466/neuro_weed.gif?ex=68e4f156&is=68e39fd6&hm=29dffd3fa61c50a5d8d616a0af4b95eb2407db7dec370285cf5387b9ca4ec5bd&=

[rare bramble]

I'm not a C dev but I have never heard anything good about MSVC, is it really that bad?

just heard from a teacher that it's not great at autovectorizing code so we should avoid it in our performance relate project

[opaque sigil]

there's clang-cl at least if you need the msvc frontend

[clear sedge]

there's also the fact that i'm pretty sure msvc is only officially available on windows, it just doesn't try to be cross-platform
i highly recommend clang or gcc instead

[opaque sigil]

msvc also doesn't have a good track record of implementing the c/c++ standards, or they take ages

[clear sedge]

microsoft not having a good track record of implementing standards is very much like them

[opaque sigil]

llvm my beloved

[obsidian mantle]

lack of sleep kicked in

but its too early, 10 pm

i cant code for shit

[opaque sigil]

[obsidian mantle]

transposition doesnt work
iterative search doesnt work

there is some small detail im missing

cant see it

[tribal yarrow]

Sam Altman doing more Dev streams than vedal 💀

[versed orchid]

Yes(Late response cause I've been busy)

[obsidian mantle]

tried to bring logic to this alpha beta transposition mess now it picks worst moves

idk how to even debug this shit

[cosmic sphinx]

brotheeer put some volume warning there 💀

[obsidian mantle]

i should rewrite this whole search from scratch

[amber fractal]

There was context right above

[obsidian mantle]

i so much hate to end the day on a loss

4 hours -> achived nothing + broke everything

i felt like i understood how it works but i didnt, whole knowledge i gained just does not add up

[versed orchid]

the most efficient debug method

[obsidian mantle]

if i add one cout at least anywhere in search it will give me 5 quintillion outs which i will never parse

if i search in depth 1 it will have no bug visible

whatever

maybe im just to sleepy

[solid bough]

Oh god, it's slow:

That is training on CPU

[versed orchid]

F

[solid bough]

Because I have no GPU with enough VRAM

[versed orchid]

Well... Lower the batch if possible

you got it like... 32 as I see

if you don't need much precision...

[solid bough]

Where?

Where?

[versed orchid]

You got 8000 samples and 250 steps per iterations

8000/250 == 32

[solid bough]

https://tenor.com/view/yeet-bruh-falling-gif-15495609

Well, finally I can train with more than 200 images

on a CPU 💀

Well, I can spare that PC ngl, I kinda miss the plot though.

[versed orchid]

It'll take you like 2 min per epoch

so... you have a room heater for an hour or two

[patent shard]

24-bit flac is about double the file size of 16-bit
You'd think it'd be 1.5x
I guess 48-bit png is like 5x the file size of 24-bit

[solid bough]

That is generously said, The CPU i use for that training is the Intel Xeon E5-1650 V3

[versed orchid]

[solid bough]

Well, it is a PC I can leave running without it taking my main PC

And also more images (64GB of DDR4)

[versed orchid]

Google collab maybe?

[solid bough]

Woah wait Rainboom?

I didn't know you're chillin here

That geniunely caught me off guard

Yoo Rainboom Dash, How's going?

[patent shard]

On my lunch break at work

[solid bough]

Enjoy you lunch btw

[solid bough]

Nah, I know it's free, but I can't quite do what I was thinking there

That really caught me off guard now

Well, it looks like it isn't quite running fast, but that is fine:

[fast pagoda]

what are you training an image model or llm

[versed orchid]

Ok... I understand. I wanted to use it for my coursework, BUT I couldn't pay for better GPUs. Not because I have no money, but because I don't have the right paying methods. Sadge

[solid bough]

Well, I mean Idk what length I should do and stuff, but if you know what you're doing, that's great

Like you can use whatever you wanna (but not my GPU)

[versed orchid]

I mean yeah, just as I remembered the thing it reminded me how unusable it is for me

[solid bough]

Right, My bad, I just cleared up from me being surprised. I try to build a better model than that blur:

[solid bough]

The Google Collab?

[versed orchid]

yeah

[solid bough]

I wanted to do even better than the original model

[solid bough]

Oh 😬

[solid bough]

That ran on my NVIDIA Quadro M2000 meanwhile (4GB of VRAM)

[fast pagoda]

🗿

that's a model use ive not heard of

what's the benefit? Blur without it being deterministic?

[solid bough]

The name of what I mean is Deep Doodle that CodeParade made.

[fast pagoda]

oh

[solid bough]

There is a video demo on YouTube meanwhile

[amber fractal]

Speaking of youtube

[versed orchid]

You can do better even on same or lesser GPUs if you are teaching differently. Is there full code or the description of the model?

[amber fractal]

for the first time in awhile, they released a very based feature

My client shows it as this, but group uploads now exist

[solid bough]

I wanted to do better than that:

[solid bough]

I actually have the code rn. (I modified it from the GitHub to run on Modern ish Python)

[solid bough]

Should I censor the Eldrich monstrous thing next time?

[versed orchid]

Is it not?

just kidding - it's fine I think

[versed orchid]

so you just updated it in a way to run as is without any new feature?

[solid bough]

Yeah (Ahem chat did)

[versed orchid]

Now tell me WHY you think your model should work better?

[amber fractal]

I trust comments with my life, also that is an insane pitfall of only put on one channel.

[solid bough]

I got more images to train on

[versed orchid]

How many they had?

[solid bough]

That is what the maker said:↑

[versed orchid]

dataset is the base of the good model, but still there is a cap for how much better it can be

[versed orchid]

that make sense

[amber fractal]

Part of it is data, part of it is sound logic

[solid bough]

Yeah, but I don't wanna sift through like 202K images just that they have a visible face and face the camera

[rigid snow]

was trying to figure a scenario out where that would make sense while bouncing off a llm just as a fact checker and had a funny moment

[rigid snow]

BINGO. You are 100% correct…
*proceeds to tell me it’s incorrect*

not in this screenshot but later

[versed orchid]

I mean ok - you got the point

[rigid snow]

this is a reasoning model too

[solid bough]

Well, Should I just not care and put like 3K images in there?

I got 64GB of RAM to spare

or at least like 58GB

[amber fractal]

reminds me of a competition I heard about where a char level llm beat a normally tokenized llm because it was implemented differently.
-# If I recall correctly, this was a competition between two chatters here

[versed orchid]

well... You can test if you like on a dataset x2-x3 larger than original to see if there is a major improvement

it should be faster this way

[solid bough]

Yup, Sifting through the images that it is like facing the camera and face clear is a good idea then to do?

[versed orchid]

I think yes

[solid bough]

Oh god

oh god

Hang on lemme look

I needed 2K images that I found 600 that are clean and looking to the camera

[obsidian mantle]

Cant you use some ai search

[solid bough]

Yes, I manually sifted through the images

[solid bough]

I doubt that ngl

[versed orchid]

At least if you don't see improvements for now you can adjust the learning parameters and functions so I think there will be some improvement overall

[versed orchid]

I don't trust it, Vitek

[solid bough]

if I do a learning rate too high, then it is just not too great

Well, I had some error open how much it is looking at the camera, so mild turning from the camera is there.

[versed orchid]

I mean you don't have to, but if you'll see no improvements by enlargening dataset you probably should

[solid bough]

Well, ignoring mirroring, I would need 2400 images then?

[versed orchid]

If it's not like 40 degrees it should be fine

[versed orchid]

about something like this

to test

if it's efficient

[solid bough]

[versed orchid]

ML is hollowing the souls of the dataset makers

I have to go sleeping. It's not healthy to stay up too late

[solid bough]

ML?

That is the last question I have

[versed orchid]

Machine learning

[solid bough]

Oh

Thanks btw

Sleep well Lankin 🫡

[olive sable]

https://youtube.com/shorts/dTTDemkc-hg?si=svkR2MjwzlE3xHzV

[amber fractal]

Need to include this as an easter egg fr fr

[olive sable]

[amber fractal]

Want to unselect? Tab or grab the nearest game controller

[shadow sinew]

Game controller on Windows 98? Better have a sound card installed.

[warped narwhal]

tab + space ftw!

[olive sable]

constexpr static const float e;
very necessary const

[warped narwhal]

you forgot thread_local

[olive sable]

constexpr inline static const volatile thread_local alignas(64) float e

[amber fractal]

wasted space, floats are usually 32 bit :SMH:

[clear sedge]

[olive sable]

fuck you, alignas 256

fucking numlock lmao

[clear sedge]

alignas(65536)

[olive sable]

alignas(8*1024*1024*1024)

align per gigabyte

[amber fractal]

ML engineers when assigning memory

[clear sedge]

chromium devs:

[amber fractal]

Not firefox, that is aligning to 2gb

[solid bough]

Okay I'm here now Sceade

[signal trout]

if you try to run a python file in VS Code, does it pop up any messages? Assuming you have the right extensions installed

[solid bough]

Wth is going on? I only opened the settings menu

Traceback (most recent call last):
  File "C:\Users\Volks\AppData\Local\Programs\Python\Python313\Lib\tkinter\__init__.py", line 2068, in __call__
    return self.func(*args)
           ~~~~~~~~~^^^^^^^
  File "C:\Users\Volks\OneDrive\Dokumente\code\Text-converter\gui.py", line 93, in <lambda>
    theme.settings_butt=Button(GuiConfig.window, text="Settings", padx=5, pady=5, command=lambda:gui_settings(GuiConfig, config, theme, ))
                                                                                                 ~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\Volks\OneDrive\Dokumente\code\Text-converter\settings.py", line 304, in gui_settings
    print(SettingTheme.bg_color.get())
          ^^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: 'str' object has no attribute 'get'```

[amber fractal]

Is a string

[clear sedge]

dynamic typing downfall

[solid bough]

Probably

[signal trout]

SettingTheme.bg_color is a string, apparently

[solid bough]

Well, That is not something to worry rn

[amber fractal]

Print the string

[solid bough]

That actually was the thing that made the error

[amber fractal]

Remove .get()

[signal trout]

print SettingTheme.bg_color, he means

[clear sedge]

https://tenor.com/view/oshi-no-ko-oshi-no-ko-memcho-memcho-confused-anime-confused-anime-question-mark-gif-8655139622262299831

[signal trout]

or walk through it in the debugger

[clear sedge]

there's no str#get, you're trying to call that

[solid bough]

That: py print(SettingTheme.bg_color.get())

that line

[amber fractal]

This: py print(SettingTheme.bg_color)

[clear sedge]

this is akin to trying to print py print("".get())

[signal trout]

well, what is bg_color supposed to be? Right now it's a string, but I'm assuming it's supposed to be another object

one with a get() method

[clear sedge]

tsk, dynamic typing

[signal trout]

true

[solid bough]

it is supposed to be something something tkinter

[signal trout]

I have had... struggles with tkinter

[solid bough]

Oh wait it has NOTHING

[clear sedge]

i've used tk before

i've never used tkinter

[solid bough]

that variable has NOTHING

AAAAH

I WANNA SLEP

SLP

SEP

SLEEP

[signal trout]

can the problem wait till morning?

[solid bough]

filtered no

[signal trout]

does the vsc debugger work?

you mentioned it can't find python, so I'm thinking not

[solid bough]

It should show a friggin version

[signal trout]

if you press ctrl + , it should pop up the settings

search for "interpreter"

see if you have a complete path to the python.exe under Python: Default Interpreter Path

[solid bough]

Imma try that on a German layout

Ah

it works

[celest hull]

bg_color is of type str and it doesnt have a get() method

[solid bough]

I mean there is no data in that variable

[signal trout]

it should be a blank string

according to the traceback and your saying it has nothing

[solid bough]

Guess what

[signal trout]

it's already there?

[solid bough]

After making sure the Python itself like that

[clear sedge]

"" i'm assuming?

[solid bough]

It's still like that:

[solid bough]

I think that is what it looks like if it was code

[signal trout]

if you attempt to run the code it should ask you to select an interpreter

does that happen?

"default interpreter" should be an option

[solid bough]

OH MY GOD

I'm silly

[celest hull]

perhaps it's not the variable you are looking for?

[solid bough]

Wth is VSC trying to do???

[celest hull]

creating a venv i would assume

it's good pratice to use one

[solid bough]

Before I started the venv shenanigans, it was all fine

[signal trout]

I have seen some weird stuff when working with environments

been a while, though

were you trying to do something with virtual environments? Or is it doing that on its own (perhaps due to an extension)?

[solid bough]

All I did was Python 3.6 and then doing Venv

And rn I'm trying to use 3.13

I'm going to bed!

Wait

I forgot

It should look like that but with just 3.13

[solid bough]

Like when I press the button on that one that says 3.60, it should do that:

[signal trout]

huh

have you tried clicking the other one?

[solid bough]

Just how the colors are

[signal trout]

also check in your python directory whether /.venv/Scripts/python.exe exists

probably does

[solid bough]

Like, that is C

[signal trout]

not sure what it means that it can't resolve env "//python"

[solid bough]

There is none since I never set one up:

[signal trout]

it might be hidden

oh, if you really never set it up, then it wouldn't exist

but iirc it is normally a hidden folder, so you'd need your folder settings to show them

[solid bough]

I made that hidden folder were not hidden before that:

[signal trout]

so running it as a venv wouldn't work

since it doesn't exist

that makes sense

[solid bough]

Yeah

Wanna know the twisted part?

[signal trout]

does it run if you click the last option without (venv)?

[signal trout]

go ahead

[solid bough]

THE CODE WHERE I DID SOMETHING WITH VENV FREAKIN WORKS:

[signal trout]

huh

[solid bough]

That was on the PC I intended to train the model on, not that one

I just remoted in

[signal trout]

I don't like working with virtual environments

[solid bough]

I wouldn't even have a way to connect a monitor to the other PC I sent that screenshot from

[solid bough]

same rn

Why?

[signal trout]

not enough memory on your device, it seems?

[solid bough]

RAM or VRAM?

And I mean that the code runs

and how did you know:

lmao

[signal trout]

lol

[solid bough]

That is on GPU rn

[signal trout]

I'm something of a genius

hm, so it is working, just telling you it's being slowed down

[solid bough]

Still, How the poyo does that work fine?

[signal trout]

I haven't worked with AI at all, so I'm a noob when it comes to model training

[solid bough]

It's rn more about "WHY TF DOES VSC NOT SEE PYTHON???"

[signal trout]

the eternal question: why is vs code blind

[solid bough]

Of couse when I have the AI training bs open it filtered works

[trim valve]

probably worth checking task manager, but this feels like a message you'd get when you run out of vram and it spills it over into regular ram for you

[solid bough]

I'm not quite on spilling...

It hasn't reached 8GB yet

[trim valve]

hm

[solid bough]

Oddly though that is with any dataset size that it goes to like 6.5GB

[trim valve]

it seems to be mentioning something about gc, but also saying that it might help if you manually free stuff when you're done with it?

[solid bough]

Before you ask if I set memory growth, Yes: ```py
gpu=0

physical_devices = tf.config.list_physical_devices('GPU')
if physical_devices:
# only use the first GPU (your RTX 3070 Ti)
tf.config.set_visible_devices(physical_devices[gpu], 'GPU')

# enable dynamic memory growth so TF can use full 8 GB if needed
try:
    tf.config.experimental.set_memory_growth(physical_devices[gpu], True)
    for gpu in physical_devices:
        tf.config.experimental.set_memory_growth(gpu, True)
    print("Memory growth set.")
except RuntimeError as e:
    print(e)```

[solid bough]

It's while:```py
for i in range(num_epochs):
print(f"Epoch {i} of {num_epochs}")
text=f"Epoch {i} of {num_epochs}"
log_info(text, logg)

print(x_train.shape, x_train.dtype)
print(y_train.shape, y_train.dtype)

model.fit(x_train, y_train, batch_size=batch_size, epochs=1)

#Training results
mse = model.evaluate(x_train_mini, y_train_mini, batch_size=8, verbose=0)
train_loss = model.evaluate(x_train, y_train, verbose=0)
train_rmse = np.sqrt(train_loss)   # if your loss is MSE


print("Train RMSE:", train_rmse)

#Test results
mse=model.evaluate(x_test, y_test, batch_size=batch_size, verbose=0)
test_loss = model.evaluate(x_train, y_train, verbose=0)
test_rmse = np.sqrt(test_loss)   # if your loss is MSE

train_score.append(train_rmse)
test_score.append(test_rmse)

print(f"Test RMSE:", test_rmse, "\n")

model.save('Model.h5')
print("The model is saved.")

plotscrores(train_score, test_score, 'Scores.png', True)```

I guess

Yeah

it's that part

And the network is:

(768, 5, 192, 144) float32
(768, 3, 192, 144) float32

Idk why tensorflow is going float32

That is a bit overkill

bred, should float16 be fine?

[trim valve]

idk im not an ai person

[solid bough]

Understandable 🫡

oh great:
(192, 5, 192, 144) float32
(192, 3, 192, 144) float16

[solid bough]

Anyways, GPU going brr?

[trim valve]

certainly some amount of brr

[solid bough]

Yea

[trim valve]

i wouldn't be able to tell you if that's maximum brr or not

[solid bough]

Well, if I'm right, if CUDA is 100%

It's doing Epoches like no tomorrow

[trim valve]

[solid bough]

The log file is gonna go 📈 in size

And I'm not amused

Wait, Why does the step amount go down with increased batch size???

[solid bough]

Tht's more brrr now:

with a heartbeat

Well, a messed up one

[warped narwhal]

I paid for the whole PC, so I'ma use the whole pc

[midnight sigil]

[warped narwhal]

For some reason I keep reading your name as calibre.

Are you sure you aren't secretly a font in disguise?

[midnight sigil]

how can you see l and r

[glad path]

reminds me of the time i was running an llm locally and had under 1% free memory

[midnight sigil]

you can force precision with the param fp iirc

on almost all operations

[twilit lava]

So armchair postmorteming Evil's crashout ...

I wonder if the integration with the game led to a lot of very fast location reports to evil

kinda working like glitch tokens

where because location is said over and over it is "nearby" all other tokens

[midnight sigil]

cuda heartbeating

your gpu is alive

[slender timber]

Progamin

[faint sandal]

no way...

[gritty dust]

[safe path]

<non-programming message>

[uneven pulsar]

Guys what programming language should i learn as the first language

[glad path]

depends

[uneven pulsar]

Im adaptable to everything

[glad path]

python maybe

or lua

[uneven pulsar]

unless its too complex even for experienced

[glad path]

no bias here or anything

[uneven pulsar]

hm

python is good

[safe path]

rust

[uneven pulsar]

what

that game is also a language 😭 ?

[glad path]

.`/

[hollow matrix]

@native bramble When is the canvas going to be open sourced?

[safe path]

whoever thought about pdf reflows i love you thank you many kisses and hugs muah

[signal trout]

python is a great place to start

more complicated languages can come after you understand the basics of high-level programming

[rigid timber]

it's already open sourced
https://github.com/BlankParenthesis/pxls-rs
https://github.com/BlankParenthesis/pxls-svelte

[rigid snow]

somehow i am awake

[rigid timber]

hi awake

[hollow matrix]

oh ty

[umbral thorn]

[olive sable]

Goodmorning

[nocturne olive]

I see Elvyn is once again showcasing her crazy PC

[rigid timber]

what'cha dooooooin?

[nocturne olive]

Well, for my Chess bot what should I actually do next? I have a side-wise evaluation function and not much else

[rigid timber]

pruning so you can search deeper

or improve the evaluation function

or check out the chess programming wiki, there's literal dozens of pages

[nocturne olive]

I guess I'll figure it out somehow

[uneven pulsar]

meow

[patent shard]

I finally tried this out, I need to do more testing.. but based on the limited amount I did, I think FT2 bleedless and big beta 6x are generally better
big beta 6x is def fuller, with slightly more noise as a trade off.. still less noise than duality, from what I heard
it's also kind of weird to name it duality.. like, nearly all models can do instruments and vocals fairly well.. I don't really get it

I think this mvsep one is smarter on what to keep and not... but it seems to muffle everything, I don't really get it.. doesn't really match the metric scores
It'd be an amazing model if not for that..

you also can't run it locally :/

I think the duality one is also technically considered an "older model" at this point
granted, it's only like 11 months old..

[nocturne olive]

It's duality because it tries to do good on both inst and vocal

There's also specific models that try only on inst

[nocturne olive]

Useless then (though what stops it?)

[patent shard]

because they want it exclusive to mvsep
it is free to use, still annoying... I try not to use it

[patent shard]

idk man, I didn't like it for instruments, either.. didn't have the fullness I like
BS Roformer Resurrection Inst has more fullness while still not being annoyingly noisy

[nocturne olive]

Lots of models I have not even heard of

[patent shard]

only came out two months ago

but it's the best model I've found so far

[nocturne olive]

Either way when it comes to data it's too late to turn back now that I've already started labelling the NS data

[patent shard]

there's an FNO version that has less noise but more fullness, but the noise I don't find an issue for the other one, and the drop in fullness is very noticeable

[nocturne olive]

All I particularly care about for data is that the voice information is statistically highly accurate over the full sample size
All I particularly care about for in use for getting the information needed for NeuroSynth covers it that pitch information is intact and clear

It would be great to have someone measuring those statistics

[next rover]

I was told this chat could help

I want to know where to begin on creating a self learning chat bot that can be integrated with other stuff in the future

But I'm like ground zero in terms of knowledge

[nocturne olive]

Self learning? No

LLMs cannot do that

Go invent a new architecture

If you just want a chatbot that would be like stupid easy, but self-learning, that needs a new architecture entirely

[wide flicker]

I see your 100% CPU/GPU usages and raise you infinite percent /hj

[next rover]

Wait neuro isn't self learning

[nocturne olive]

Of course not she's an LLM

That's why she sucks at games every time

[next rover]

I thought she was trained on twitch chat

[nocturne olive]

No that is just misinfo

[wide flicker]

trained =/= self-learning

[nocturne olive]

And yeah

Trained is not the same as self-learning

Self-learning implies learning at inference time

[next rover]

Oh

Yeah when you put it that way