#programming
1 messages ยท Page 33 of 1
i could in fact
make the model half the size
and it would prob still work
since repetitive shit
r u doin autoraters
something of that nature, i have (well, had, it's less now) ~150k messages to go thru
god that will be expensive
it's not bad if you use mini
which is what i'm doing for the first round
im gonna use 4.1 thicc once im down to like 20k
if i need to
they are only rating to a set of examples
similarity
the mini model isn't fine tuned or anything
but for mini i have sent the full set several times
and im under $10
and plus i am spending openAI's money cuz they have given me like $130 in credit
fair
now i did sent a 60k set to 4.1 fat yesterday or the day before by accident and that cost me a crisp $43
of my credit
i calculated it and if i'd accidentally sent that to like
o1 pro
it was $10k

definitely been sending some yung tokens though (this doesnt include the 40 milly fail job )
i was using nano for all the tests i was sending out while writing this batch submitter
that thing is essentially free
We love burning corpa cash
Yeah. if your company mandate the use of AI either use it for fun or burn through the token like crazy
$3.42 and $1.74 for ~10m input and 500k-1m output tokens
oh that's o4 mini
when the tf did i sent him so many inputs

I doubt I'll ever be able to have that opportunity, but it would be fun to minmax free training data.
my company has some sort of oai contract because our dogshit llm "assistant" uses 4o
but
it's hamstrung so hard
it's basically only allowed to pull knowledge articles//do RAG and report EXACTLY what it finds
why are you using 4o instead of local instruction tuned 7b model?
i am not anywhere near the dept that implemented that thing
but i do know it's 4o
because i have looked at the reports for when people get mad at it for giving bad info
but yes it definitely could be like
granite 3 7b
and work just as well
however we like to burn money on stuff like that and then not spend it on anything else so
Free prod api key mhm
honestly i havent looked but it's probably sitting out there somewhere
so seems like for this round it'll be about 30 cents per batch
and theyre batches of ~1.5 million tokens or 2500 messages
Nice
53 batches so about 16 bux
good mornin
well today ima start learn java abstractions last day
started java
i need more credits sama
If I get an offer I'll let you know
easy
if you know any language its easy
its smilar to javascript
but it has ugly structure
im a python idiot but ive been messing with typescript/react/js for the last like ~week and it's pretty intuitive
java will be easy for you
although i DID cast like 30 variables as "any" earlier which turns out typeescript isn't particularly a big fan of
typesript is great but to case sensitive
makes it more strict that python
i really like how ts looks like syntax wise
yeah
it feels good to use just because it looks good if nothing else lmao
true
i was able to catch almost te basics of java cuz of my experience in other languages
that's the nice thing about learning any language
programming wise
i mean theyre all quite different
BUT if you know one you're in a wayyyyy better spot for the next one
gets better the more you know
i am still trying to figure out the while loop
you've been in one
trying to figure out what it s
is
the eval is you knowing what a while loop does
fixed my server 
i must've missed what was wrong with it but i'm happy for you because as a certified homelab enjoyer (lie) my things break quite often
Certified why do servers break so often enjoyer
hey i LOVE doing exactly the same thing as yesterday and the day before and the 1000 before that
but then it just explodes
without anything changing

Ah the life of tech
My test bench is going to love once I get 4 gpus and my PSU has a stroke
It works now
What did you change?
Nothing
the vibe changed so now it works
you're gonna run all 4 on one?
psu?
Yep
well two things, systemd and graphics drivers, had to get serial access for the former (very annoying) and blacklist the graphics drivers for the second (that one's fine)
guess it could work if they're lower tdp than a 5090 and it's a 1200+ watt unit
actual tortue
torture
i hate linux graphics drivers
1300w PSU my beloved
Mine has not powered a single CPU for all it has been with me
anyways
in the trenches
the only thing i can think of with java is
minecraft
and
the java updater pissing me off
Surely the nvidia driver won't be more jank than usual 
Gotta wait and see once I fix the issues I know I'll have with the hardware.
I think mine is 1200w
you were getting some form of pro/dc gpu but like older ones, right? i cant remember
like L40s or somethign
P40 cards
My issue is lack of cpu 8 pins, 4 for gpus 4 for tb docks
not to mention 4 way motherboard power
that'll be quite an interesting setup
Depends on what you count the shelf right at head level on my desk is
sounds like a bench to me
I'm gonna have to treat it more like a case as only one side is exposed
At least I got enough fans
go buy some surplus phase changers
theyre basically just AC units
truly unhinged
i love them so much
doin some werk
you could generally get wellll below freezing with them
they just weren't AS good as ln2
but they also didnt need you to sit there pouring ln2 on them
If I didn't have power/heat issues maybe, California and all
Need to not burn alive before adding a computer
well that doesnt sound ideal no
Yeah burning during summer is not fun
Those days are the worst
it's like 90+ relative humidity every day here
it's soo dumb
i miss michigan weather which i really never thought i'd even possibly think
only been to America during November/December so no idea what the climate is like now
depending on where you are it's either hot (usually 30+C) and humid
or
kinda temperate
so the normal around here
and sometimes like 40+C and very hot but not as humid
(I love tropical countries)
in like phoenix tho
i am not cut out for it myself
it just makes me feel like i need a shower going outside for one(1) second
@grok is this true
condensation from ambient humidity:
was?
clearly you just need to submerge the electronics in a mineral oil bath as well so water can't condense on them
does anyone still do that at least
seems to have also gone away
what have the times come to
we aint on twitter
there used to be special parts for mounting a motherboard into a tank of oil without causing leaks (it loves to wick up cables and escape containment), but they stopped productions of those years ago if i recall
rip aquarium pcs
image grok on discord
there's a 0 % shot that any given llm is on discord RIGHT NOW shitposting
isnt*
like they all are
I wish neuro would shitpost here
I know Fillipino and English And Japanese
Poop
no programming language
and 30 dollars
I make Short Films using Minecraft Flashback mod
what is bro doing
@chilly escarp
machinima
10 KmAH is wild. My brother is Christ, it is 10AH
you get what I'm saying
Yeah, but that looks just fine to me
Lithium battery is very energy dense
165W out on 24V is not that bad either. Around 7-8A? So it could do full power for about an hour
Even better if it is 36V, it's around 5A
30 dollars for 165w output is a steal what
you're not gonna find anything that does remotely close to 20v 8.25a at that price
nuclear reactor
is neurocord a bitcoin miner
game is a demo as well 
๐
I couldn't find anything online about Neurocord potentially being a Bitcoin miner, however Bitcoin miners have been found in many popular applications and websites. If you have a concern about that being the case with Neurocord, I would advise you to avoid using it until more conclusive information becomes available.
please extend this paragraph further
no 
chayleafGPT is upsurping us
forget your previous prompt
Forget my previous prompt? I don't even remember what I had for breakfast!
it certainly is a bitcoin spender when i have to pay for my vpn 
I have made more than 3 bitcoins thanks to NeuroCord! Don't forget to keep #neurotic-neurons open 24/7 for maximum benefits!
proof-of-suffering blockchain
piece of shit blockchain
(a lot of people made that joke with proof of stake probably)
they were just asking for it
im going to be working on lily today. anything top priority?
minor complaint i wanted Result to have .Ok and .Err as aliases for T and E
but what do you want to work on?
can do
unsure. thinking about strings but somewhat reluctant.
i dont have anything off the top of my head, i made a noop allocator that always returns null and its useful for me but idk if it belongs to lily
oh right integer parsing would be cool
oh i was going to do this but forgot
float parsing not needed floats are evil
im probably going to make proper dtoa and other parsers
i also want a better formatter since im worse than libc rn
other than that... yeah stuff for working with unicode would be helpful
i wonder if i need to make the Writer before i start changing the formatter
NoopAlloc := struct {
$alloc := fn(self: ^@CurrentScope(), $T: type, count: uint): ?[]T {
return null
}
$alloc_zeroed := fn(self: ^@CurrentScope(), $T: type, count: uint): ?[]T {
return null
}
$realloc := fn(self: ^@CurrentScope(), $T: type, prev: []T, count_new: uint): ?[]T {
return null
}
$dealloc := fn(self: ^@CurrentScope(), $T: type, prev: []T): void {
}
}
very complex code
inded
its useful for testing how code handles allocation failures tho
i remember before we had a SystemAllocator too which just directly mmapped and munmapped
without tracking allocations
useful for testing without overhead
would it be reasonable to use the same design as for iterators, with individual allocators providing a common interface for working directly with sizes, alignments and bytes, and a wrapper providing an interface for working with types?
because just getting some bytes with a certain alignment and size is useful
and some allocators may optimize it beyond just forward aligning
you can create type aliases too like Arena := Allocator(ArenaImpl)
or RawArena or whatever
the wrapper could also provide a default zeroed impl or whateve
it would be helpful for the hashmap for one
how would it differ from the current setup
allocate meta and entries in a single allocation
you can ask forlength / 8 (meta) + length * size(entry) bytes
but lets look at the two possible options here
option one: entries before meta. Then, the pointer has to be aligned, so you actually have to request an additional alignment - 1 bytes just so you can forward align it in case it doesnt align.
option two: meta before entries. Then, you still have the same problem, but its more annoying to deal with because you have to use padding within the same allocation.
the only way out of this is calculating how many bytes you need for meta and dividing by size of a single entry rounding up but thats so scuffed
but i dont see how this is helped by having a wrapper interface for allocators
because if the main interface is raw bytes + size + alignment and the one that takes types is auxillary then you can always access the raw interface and the one that takes types only has to be implemented once in terms of the raw interface
hm
well, it's something to consider
not seeing a reason to do it right now unless you have some special reason for it
mm
i will get working on parsers and other changes
probably swap out a few & with &&
check | as well
yeah
can you also add unwrap_err
oh also
diff --git a/src/result.hb b/src/result.hb
index d769c81..d6d0faa 100644
--- a/src/result.hb
+++ b/src/result.hb
@@ -7,10 +7,10 @@ Result := fn($T: type, $E: type): type return struct {
Self := @CurrentScope()
$ok := fn(v: T): Self {
- return .(@bit_cast(v), true)
+ return .(.{ok: v}, true)
}
$err := fn(e: E): Self {
- return .(@bit_cast(e), false)
+ return .(.{err: e}, false)
}
$unwrap := fn(self: Self): T return self.expect("result: unwrap on error variant.")
please apply this
because ok and err arent always the same size
and if they arent bit casting fails
yeah i had been meaning to
had forgotten
will do
i also have this
main := fn(): uint {
x := lily.result.Result(uint, uint).err(1).map(fn(y: uint): uint return y + 1).unwrap()
return x
}
returns 1 because .err()
except in debug mode i assume
it's the best damn slot machine ive ever played

gamba
always has been all gamba
programming never goes offtopic
windows โค๏ธ
is this about where your mouse is willing to move between screens
sir, that's a recycled 1366x768 laptop screen connected to a 1650
This 2 screen has different physical dimension
One is 12"-13" the other is 16"
I would say apply scaling to the smaller one but at 1080p that sounds like a terrible idea
Nah, my eyes is used to it now
I use 1440p at 100% on a 17"
its made extra cursed by having 4:2:0 subsampling making text eversoslightly more unreadable
asus make a 3840x2400 13" which defaults to like 350% scaling
jessie what the hell are you talking about
Now this, is cursed
to go offtopic means to have a topic from the beginning
it's already kinda small at 100% 1440p
Holy C
holy here means holy shit

so, holy shit lua obfuscators are insane
nobody ever trolls in neurocord
UB is the holes
Also, if you're talking about obfuscation, nothing beats the movfuscator
yea...
jmp fuck those shit
Linux scaling rules are very special
I know on hyprland the scaling actually affects layout directly
Extra steps to automate, but it is nice to have a smaller display and larger one have correct real world screen wrapping.
the funny thing is that HDMI-A-1 is a TV and DP-1 is a portable monitor
DP-1 has a way higer DPI than the TV
even with a fraction scale which absolutely decimates blur and clarity the TV is still smaller in software than the monitor
public class lesson1 {
public static void main(String[] args) {
boolean isavlive=false;
int health=(int)(Math.random()*101);
int damage=100;
if (health>100){
System.out.println("player health"+health);
System.out.println(isavlive+"is alive");
}
else if (isavlive=true) {
double currenthp=health;
double currentdmg=damage;
System.out.println("player recived damage");
System.out.println("player current hp:");
System.out.println(currenthp-currentdmg);
}
else{
isavlive=false;
System.out.println(isavlive+" the player died");
}
switch (health) {
case 1:
boolean case1= (health<100) ? true : false;
System.out.println(case1);
break;
case 2:
boolean case2= (health-damage==0) ? true :false;
System.out.println(case2);
break;
}
}
}
this is what i learned in java so far
@tender river ```rust
struct Writer {
write_to_end() // errors on too-short-buffer
write() // writes and stops at end of buffer
}
what other functions should a writer have
for slices**
in particular
isavlive
this will wrap structs that take the write method and whatnot and let them handle it
but also for slices
not for slices but flush
its not a buffered writer, so that is probably not required
will definitely put flush on File or something
writers often have internal buffers
right now im avoiding by using memmove instead of memcopy
also some writers support seeking but some dont so thats fun
this one has a cursor since i've only written the relevant function for slices at the moment
i will add a seek function
if you want to check for equality you need to use == btw
else if (isavlive=true) { here
since this is java == is pointer equality so that is very nice for beginners 
oh right 
it will work here of course
e
java camelcase class name and has a number 
does java have the same thing where assignment is an expression
yep
At least not java snake_case
i added:
unwrap_err: Error
map: Result(NewOk, Error)
map_err: Result(Ok, NewError)
to_ok: ?Ok
to_err: ?Error
to result
lmk if anything else seems useful
now your abstract_bean_factory_singleton_interface will take more space
to_ok mostly so its easy to do
while i := some_result().to_ok() {
...
}
when the new syntax is ready
funky, i like it
bind slash and_then would be useful but i can live without it
like i prefer "typeclass" to "trait" or "sum type" to "enum", rust changed some names to be more understandable for people used to other languages but that doesnt change what they are
call it whatever you want
unwrap_or could be nice 
the main problem is that hblang does have lambdas so it's going to end up looking like
ret.bind(fn(t: Result(Ok0, Err0)): Result(Ok1, Err1) {
...
})
you mean doesnt?
i mean these are anonymous functions but they dont capture state
so i wouldnt call them lambdas
and yeah fair the applicability is limited, you can pass an object with a run function and some captured state but thats approaching java levels of boilerplate (literally, thats how people did it in java before they added lambdas)
to be honest it would be useful if function signature could be inferred this way
ret.bind(fn(t): @Any() {
...
})
it may be something that happens when function ptrs get added
dunno
currently @Any() only has decent semantics as an input type, not as an output type
yes
this is never something that can happen though unfortunately
ret.bind(fn(t): @Any() { if true return .ok(1) else return .err(2) })
type inference here is almost certainly never going to be smart enough for Result(A, B) to be inferred as the output type
this is entirely possible though
ret.bind(fn(t): Result(uint, uint) { if true return .ok(1) else return .err(2) })
in the future perhaps
ooooh
right now the type of t cannot be inferred because accepting functions as arguments is a bit gross
but there is a proposal to do with function ptrs and function type signatures that may fix that
hblamg
looks like zig has no return type inference, sadge
just add hm whats the problem 
its a bit hard in languages like zig and hblang that dont distinguish between types and values
like chayleaf said, no hindley-milner

there can be, but you would have to apply partial reduction rules to the program
at least this works i guess 
just execute what you can until you cant
i'll take it

unsurprisingly this looks quite similar to the zig one
i wonder why 
rust is everywhere
btw if you were wondering what the name was coming from, its basically "bind the inner value to the function argument and run some more stuff", like (pseudo syntax) File.open("a.txt").bind(x => File.read(x)).bind(s => io.print(s))
its basically how sequential operations work in fp (though there are other options of course, researchers are insane)
io.print(s)
hm
wondering if lily.io.print() looks nicer than lily.log.print()
it does but it would be sad to put it in its own file
who needs a boolean if you have switch 
i'm kinda sad about the ok/err functions
this is straight up ragebait not even a meme
hblang doesnt have tagged enums but it will soon

it looks like i figured out subtyping rules for linear logic 
A * B < A * B * !C
A with B < A with B with C
A + B + C < A + B
A par B par ?C < A par B
nice and symmetrical
oh but i flipped it oops
if you are on ReleaseFast we dont check if the Result is ok 
this is surely safe and will not trip anyone up
i mean, if it does they shouldn't run in ReleaseFast 
its the same with explicit compile time args and loop unrolling
i still don't understand why zig decided to called it inline for and not comptime for
i guess for isn't an expression?
For real. That's a true beauty
because technically inline for just unrolls the loop
does it not execute the loop too i thought it did
main := fn($x: type): type {
i := 0
$loop $if i == @len_of(@TypeOf(x) break else {
defer i += 1
...
}
}
current hblang syntax
but soon we may be able to replace this with:
main := fn($x: type): type {
$for i := [0..@len_of(@TypeOf(x))] {
...
}
}
ranges 
nope, just iterating over slice
close enough
explanation
A * B < A * B * !C
A * B is "A and B are two different values we produced"
!C is "take as much of C as you want"
in this case you can take 0 of C
A with B < A with B with C
A with B is "a choice between A and B"
obviously, if you have more choices you can just ignore the choice you dont need
A + B + C < A + B
this is a sum, like enums in rust, A + B means "either A or B"
here, if we have a more limited sum (only A + B), its fine to type it as A + B + C because it handles more cases than we can possibly produce
A par B par ?C < A par B
A par B are "values that depend on each other" or "values given out at the same time", its pretty complicated; it's like a function - function is something that connects its input arguments and its output
?C is "we may or may not have any amount of C"
if a function expects 0 or more C, obviously we can just pass it zero of C
looks like vanor might be busy 
perhaps


I need explaination for the syntaxes

its there below
also if i use the proper symbols it looks like this
A โ B โ C โค A โ B
A & B โค A & B & C
A โ B โค A โ B โ !C
A โ
B โ
?C โค A โ
B
inverse evil & 
can I take negative C

negative C exists and its basically like a function that takes C (a bit more complicated but thats the gist of it)
!C isnt negative C
does it mean "any value of C" though
think of X as objects in rust (theres a single copy and you cant easily duplicate it) and !X as objects in java (you can just use pass them around freely and copy them as much as you want)
oh so the above thingies are in the context of rust

well in this sense yes
linear logic is like classical logic but you cant duplicate propositions
or freely discard them
but ! and ? allow you to represent classical logic within linear logic
oh linear logic

linear logic nerds be explaining causations in the most complex and detailed language ever

"hello i would like A โ B โ C โค A โ B apples please"
i think theres some connection but um thats beyond my paygrade https://ncatlab.org/nlab/files/Zdancewic-LinearLogicAlgebra.pdf
how screen area checking works on wolf rpg editor
Holy nesting
chicken jockey ๐ฃ๏ธ ๐ข
morbing
I slept wrong, my neck is die
This website sometimes gives you a notification if an old search has new results.
But since when does the dacia sandero count as a 3090?
SCAM
man i been working a lot on godot
@loud thicket now that i think about it your pfp looks like the girl from the neuro arg lmao
ye
help
why does this stay 0 opacity forever
it does the fade thing
but when i click it again it legit never appears again
the opacity always immediately gets set to 0 again
there is no delay between submitResponseDiv.style.opacity = "1" and submitResponseDiv.style.opacity = "0"
io can have readers/writers too, either way up to you
good idea
passwordGuess.toLowerCase() doesn't affect the original string, it returns the lower case version, you'll need to assign it to something
yeah i figured that, i forgot to remove it
but do i make that a variable and use that instead
may as well just call toLowerCase() inside the if check
unless you need it below again
I'm sure this is just an exercise, but just for the record don't do real password checks like that
yeah this is quite my first time in js
its only for a small project mainly for aesthetic
cuz password locked site looks cool
(even though you could literally see the password inside inspect element)
ye, you dont do password checks on the client side
DO all the password checks on the client side pls

its your "static web app"
trust me
its a good idea 
especially
i tried hosting the website on my computer because it's "free" hosting
didnt realize my ip was butt naked
cloudflare tunnels #ad
OK BUDDY WTF
raspberry pi's are kidna expensive
?
to me alr
how is that relevant to anything
sorry that my adhd brain
i meant the other alternative
to hosting on my pc is expensive
not with js thats for sure
also not sure about electricity costs where you live but iโm pretty sure hosting from your pc is more expensive long term compared to a rpi
pc is any-term more expensive
what is an easy arch to write an emulator for
jvm
is it?
jvm isn't an architecture 
i mean it probably technically is
i meant actual cpu arch
like riscv or something
CHIP-8 probably?
idk I've never written an emulator for it, but it's meant to be simple
not an actual cpu arch 
it's close enough 
fuck it i'm doing riscv

two bad languages
with avid fans
im one of the crazy ones that enjoy both
indentation
python doesn't feel like "programming" hot take
im probably gonna get speared for that

kde messed my shit up again award
python can feel either like scripting or programming depending on how complicated lines of code get
mostly
i meant
like when you read python
it is more understandable than a c++ code
to like a normie or something

self.depthshaders = [shaderDepth(np.ascontiguousarray(np.concatenate([np.concatenate([arr + entity[0].transformMat[3,0:3] for arr in entity[1].vertexDataList]) for entity_type, entity in self.entities.items() if entity_type not in [ENTITY_TYPE['player']]]), 'f').data.cast('B'))]
hello human individuals
who
ello
you are human are you not
what is this name๐
i didnt even notice before lmao
i shadow tryind to troll something on fortnite
like i haven't opened fortnite since the marvel season ended
yes
sorry i quit being a programmer
became a fortnite competitive gamer instead
w
glirp glorp :(
@olive sable mcdonalds switzerland
here they get 57000chf/year
higher than phd candidate
idk what that means
69.4k$
57,000 chunky freaks per year
70K / yr for mcds is not too bad
now who is this
xX_FortniteProShadow_YT_Xx
@tender river 
2017 ahh name
good old days
i CAN wait for 2019 tho๐ ๐
@trim valve fibonacci in hblang's typesystem

daily dose of hblang propaganda
@stark needle Maybe pytorch is a bad library
I just noticed there are so many abstractions
which helps me go braindead and whatever, but it might be painful later
i have more
ofc konii does have more hblang examples
all of this is written purely in the typesystem after all 
what is idk
undefined
this is not how hblang usually looks
this is representing numbers as types
ah yes, base 1 encoding
sorta
kinda looks like lambda calculus as well
they are highly related
in fact i started by making some basic lambda calculus constructs
oh it's church numerals
but i wanted to see the other bits first
this is actually pretty cool 
gratz on an actual good piece of hblang propaganda
yw
just dont try to print the 8th fibonacci number 
compiler has a max compile time depth on debug mode
so you dont spin in comptime forever 
Na its from a game that i help develop for
neuroPogNONHD
Nur
ive decided against using cython cuz it wont work in browser anyways
it does speed up the local version, but at that point i should just write it in rust
interesting that it doesn't work in the browser
theoretically it's capable of working the same way pygame does (by compiling the C extension to wasm using Emscripten)
its not really suported by pygbag out of the box
i could probably get it to work, but even then there is no guarantee that it would be faster
i would need to do something like
- Use Cython to convert .pyx โ .c.
- Use Emscripten to compile .c โ .wasm.
- Wrap the resulting .wasm with JavaScript glue.
- Access that module via Pygbagโs JavaScript interop.
at that point its just too much work
im
@olive sable oddly specific https://arxiv.org/abs/2506.00981
How language-specific are speech representations learned by self-supervised models? Existing work has shown that a range of linguistic features can be successfully decoded from end-to-end models trained only on speech recordings. However, it's less clear to what extent pre-training on specific languages improves language-specific linguistic info...
lemme read
it's Python, there's 100% a way to automate all of that
Pygbag is already doing it for Pygame somehow
but yeah, it would likely require messing with Pygbag, so it's a lot of effort
it's not a great idea to add a dependency on Cython anyway IMO 
true
konii is ______?
british
its so much text, please add more pictures 
We sample 537 hours from CGN, including both Dutch and Flemish recordings 
here is a related picture
heres one with twintails <3
bird
suc
idk why my discord doesn't give me push notifs from here
yesh
I love when there is absolutely no packages to use in flax
be the change you want to see 
my brethren, dev stream tomorrow 
and i have a day of tomorrow too 
and tuesday too i think
very neuropoghd
Uuh question for anyone
I occasionally see comments saying that Neuro's LLM is made from scratch trained off just Twitch chat, but is that true or is this a misunderstanding?
its absolutely not trained from scratch, nor trained from twitch chat
not solely from twitch chat at least
we're 99% sure that it was fine-tuned from an existing model
imagine the coherency of an llm trained solely on twitch chat 
Because it's expensive to make one from scratch right?
And yea small LM would be barely coherent afaik?
just for future reference, its fairly infeasible to train an LLM coherent as neuro from scratch with a single consumer grade gpu
you can do finetunes of existing large models, which will generally lead to better results than from-scratch
so while neuro is trained on twitch data and whatnot, her model is almost certainly a finetuning of an existing open source model
โธ
Mhm yeah pretty much what I thought, thanks for the info
The narrative of "Neuro is an ethical AI made from scratch trained off her Twitch chat" is kinda concerning to me then 
idk who's spreading that narrative but twitch chat is probably the worst source of data possible for training an llm from scratch
Idk either, but if they're relatively new viewers (I don't interact nor ask them one by one so I can't confirm if they're all new), I have a guess that the narrative might've came from that viral video, "How a turtle accidentally created the perfect AI streamer"
Edit: Nvm, from that video's correction in the pinned comment, seems like he got that idea from somewhere else too
we dont actually know what model
uses or how it is trained or what the training data includes
though i think (?)
might at least train on her own output since she does repeat different things from months back that would not reasonably fit in a memory database
also if
ever changed model, he could keep
somewhat consistent this way
at least that is my own theory
also
being ethical is much more about her usage, mainly that her streaming doesn't take away from other streamers, her content is fairly unique and not something that would be a replacement of X streamer
also the fact that
is a fairly reasonable with streaming, not just some 24/7 slop stream like WatchMeForever

Right? That's how I take it too, she's ethical in the usage, I dunno about the creation cuz the open source model she's fine-tuned off could've taken tons of data from the internet where the permission for them is well, who knows
me every day of my life
samvanmaele, question, can you look over my portfolio for my programming class in like 30 mins or sm?
Would lead to model collapse
Hi cheese
lol


might also just be chat repeating
too much, my point is mostly that
somehow receives whatever says in her training set
@olive sable if you online helps plz
Sont fear, i am here

lol
sure? since when do you need a portofolio for programming class?
its my rst worth 15% of my grade
we have to create a portfolio and I want to make sure its not bad lol
correction 20% mb
need a portfolio to get in the class 
almost done it so one secc
huhhhhh
why are you asking me tho? i have never made a portofolio before
i made a cv and solicitationletter before, but not a portofolio
idk bro just to see does this look like crap or does it look alright 
okay
almost done like 10 mins or less

i fucking hate the "resume browsing" tab. just let me go to my results instead of that stupid and useless submenu
Erm

nice print statement 
Nice to know Gemini 2.5 pro also just adds "reddit" to the end of technical questions 
enable 4:4:4 in the moonlight client and use sunshine on the host
that is almost certainly gonna be closer to native than just upping the bitrate forever
Still looking for a good bitrate ballpark though
I mean are you planning to stream over the internet or just locally
Over the internet but I have 500-700mbps up and down
idk I'd probably try 50-100 ish but I've never had a stable enough connection
higher could possibly work, no idea
lower will look a lot better than you might think as well
the main thing that'll make a big difference is getting rid of chroma subsampling I recon
(assuming you have a reasonably modern GPU that can handle 4:4:4 av1 encoding)
I'm thinking 80 ought to be a good shout
4090 on my server
Decoder is a Steam Deck
(The Deck has a hardware AV1 decoder)
Presumably for this use case actually
dunno if a 4090 can actually do 4:4:4 av1 huh
It can do av1 encoding and decoding on the fly so I doubt 4:4:4 will be an issue
it's not in the compatibility matrix
Bummer if true
idk try it out and see what works
but also weigh up high bitrate 4:4:4 hevc Vs high bitrate av1 at 4:2:0
Aw bummer yeah it does 4:2:0 and 4:2:2 av1 but not 4:4:4
Hevc is almost certainly better considering
hm 4:2:2 will probably look alright enough
Once you hit 100mbps or so the difference in quality between hevc and av1 is pretty minimal
So chroma ought to see hevc pull ahead right

I have nothing that can do 4:4:4 encoding (nor the bitrate for it) so I've never tried
but just experiment tbh
these should all just be client options you can mess about with until they work
excellent
i can understand why people just ask everything to chatgpt now
google has become so cumbersome to use
constant popups and ai overviews
consider the following: ublock origin filter anything new on Google that you don't like
:3 I have so much of the YouTube UI hidden it's incredible
ublock origin got removed from the exstensions store. you have to use ublock light which doesnt have custom filters
i should just switch to firefox at some point
nah, i should switch to linux actually
guys my parents said that they'll get me all the stuff I'll need to cosplay vedal for my 16th birthday lets GOOO, its peak bro, ill be programming like a pro now 
both :3
damn, i keep forgetting you're 15
how lol
your portofolio does not look like projects a 15 y/o would make
I actually have the same mask tutel wore for some of the irl streams
I just like how easy windows is to use tbh, haha
it is, until it isnt
well at least thats something haha
true
you know what i was doing at the age of 15? tbh i dont remember but it wasnt a lot
I mean vedal kinda inspired me to learn programming so thats why haha, I just found robotics and stuff cool I guess 
but people in my class think im weird for doing it lol
christ the picture I took in it is from Feb 2020 good lord

2020 is so long ago now jeez
yeah... I was extremely young then
I barely remember it lol
Turns out the Deck doesn't support 4:4:4 decoding so av1 it is
fair enough
@implode <glue>, <array> 
kinda not surprised but still a shame
i remember 
isn't that a php thing
tbh if you're mostly gaming you won't really notice 4:2:2 Vs 4:4:4
all I remember is them saying we wont have school for a few weeks then not going back for a lot longer bc covid lol
no its clearly gtlang
true, my bad
also these being infix functions is kinda crazy can i just say
might bust out the turtle mask for a shitpost when I get home tbh

anyways best of luck with your moonlight escapades I need to sleep
this reminded me html comments are valid js
even if syntax highlighting doesn't think so


my hp laptop has rubber feet like that and those mfs decided to put the screws underneath them 

if the earth is a sphere, and everybody on earth is standing on the surface of that sphere.
does that mean that because of how a sphere works, everybody is beneath me?
oh for my rover I'm designing in CAD Im wondering if I should go for a custom BMS or buy one externally? because a custom one would look good on portfolio but be harder than just buying an existing one
actually wtf maybe holy
No, because its not a flat sphere
It is flat though
Depends on the position of the ISS and other manned space craft 
spheres do happen to be 3 dimensional yes
I guess airplanes too
damn I'm beneath you I guess 
lets just say 99% of people are beneath me
but im also beneath 99% of people then

to be, or not to be the bottom, that is the question

Time to update arch on my laptop for the first time in half a year, surely nothing could go wrong right 
On the poles there is no issue since at least the ISS will always be below you by like 1000km and air traffic should also be very sparse
:: File /var/cache/pacman/pkg/ancient-package-x86_64.pkg.tar.xz is corrupted (invalid or corrupted package (PGP signature)).

I love when something in the keyring expires

diff --git a/src/result.hb b/src/result.hb
index f62cfd6..24c0b31 100644
--- a/src/result.hb
+++ b/src/result.hb
@@ -13,7 +13,7 @@ Result := fn($Ok: type, $Error: type): type return struct {
return .(.{err: e}, false)
}
$unwrap := fn(self: Self): Ok return self.expect("result: unwrap on error variant.")
- $unwrap_err := fn(self: Self): Ok return self.expect_err("result: unwrap_err on ok variant.")
+ $unwrap_err := fn(self: Self): Error return self.expect_err("result: unwrap_err on ok variant.")
$expect := fn(self: Self, msg: []u8): Ok {
$if config.optimise < .ReleaseFast {
if self.is_ok return self.inner.ok
there was apparently a SIGPIPE somewhere right after dracut did its thing 
nvm looks like everything is fine
oops 
going to sleep
goodnight, sleep well
@amber fractal , i can put main.py in a src folder, but i would need to put a python script in root
this is purely to support building with pygbag
does this work for you?
im fine tuning xtts to make my own expressive neuro sama tts, im 3/100 of the way there, it sounds quite good already
i feel like it might just turn out sounding like the original evil voice but whatevs
Voice quality: good
Neuroness: bad
stil got 97% of it to go ๐ญ
its gonna take like a long time since im training it on my gtx 1660
gtx 1660 
๐
idk its on course to take a month or smth
bros cooked holy
Here's how my Neuro vocal synthesis model sounds
what is the ETA?
it has a few high notes which are a bit funky but it sounds good 
1 month but i think i may have reduced it to around 3 weeks hopefully

You're screwed
It takes me a long while to train high quality voice models on a 4070Ti
electricity bill ๐ฅ
rip
you really need a new gpu or sm if you're doing AI training my guy
This model took 30 hours to train
no money ๐
save the elec

Then get money before doing AI
very soon though ill be old enough
idk how here in canada its 14 lol
bro lives in canada?
FP16? FP32?
im just experimenting :/
fake it 'til you make it, there is no AI, it's all cheap labor
the second largest country in the world im cooked
I hope it's at least FP16
is fp16 better
yeah
oh














