programming

[azure lynx]

the problem is that they don't know what they don't know.

[sick owl]

It's not solved in humans let alone ultra primitive approximations of their brain bits like current neural networks

Hell I hallucinate plenty but because I'm human its called talking out of my ass

[azure lynx]

and every possible input state leads to a value of some sort coming out, regardless of if the interpolations and suchlike had any training data supporting it.

(i wish my language model had been trained on less verbose material. it's hard to be concise.)

[young plover]

void f() {
    int arr[10];
}

https://tenor.com/view/stop-it-patrick-you-are-scaring-him-stop-patrick-stop-it-patrick-you-are-scaring-him-spongebob-patrick-gif-23995388

[azure lynx]

that ones ok. it'll get compiled away to nothing coz it's not used. it's when you start passing in the index and the value to write at that index and basically trust that the other guy checked it was within range

and really it's coming from some rando on the internet

[civic lily]

"trust the other guy check it was within range"

never trust the other guy

[young plover]

That's also a problem with heap arrays if the checks aren't built-into the language

[azure lynx]

once you're writing outside of your array you can no longer say what the program will do.

[unkempt citrus]

AI is always hallucinating, its just sometimes the hallucination is correct

[azure lynx]

reality is that which, when you stop believing in it, doesn't go away. (or something like that)
is it really hallucinating if it's correct?

[astral zinc]

Oh, that statement gave me an enormous amount of knowledge. Thank you.

[amber fractal]

the more I see your takes the more I want to just write my entire understanding of brain into a document just so I can give an argument without struggling to collect my thoughts and convert them to words

[young plover]

LLM hallucinations come from being trained to produce likely text.
You can't effectively limit it to only produce things from its memories because most written text doesn't just say "I don't know" in that situation.

[azure lynx]

the problem is none of the weights or combinations have meta information about the reliability of the original information they were trained on, and so when the model is using areas outside of where it was trained it can't tell that the token it is producing is actually not very reliable.

better?

[amber fractal]

That mindset caused me to start developing my own thing

but geez the entire design process is insane

[azure lynx]

people can usually tell the difference between things they have been told or have read and things they think "because it's obvious isn't it?" and things they don't know.
how?

[amber fractal]

for transformers it can generalize but it will also create false connections, but for pattern matching those models are incapable of dropping infomation so they can't generalize

[unkempt citrus]

I mean

NNs are naturally decent at filtering noise

Butmaybe using them in LLMs degrades that capability

[amber fractal]

I think it is better to say the pattern matching stuff always lead to overfitting

I'd honestly have to look into it more

[unkempt citrus]

Eh, double descent phenomenon and grokking (unrelated to grok) demonstrate some odd exceptions

double descent

As parameters increase, it begins to overfit, as expected

Until doesn't

[amber fractal]

I still consider that similar to human learning, with the ol "peak of mount stupid" thing

[azure lynx]

it something to do with the continued tweaking the "unused" parameters after training has "finished" which finds a better solution once formerly unused stuff is close enough to help.
but that's just my interpretation.

[amber fractal]

I can see that

[unkempt citrus]

Yeah thats roughly the theory

[amber fractal]

the initial training stacks a lot on certian neurons

[unkempt citrus]

it becomes forced to ignore parameterse creating some degree of generalisation

[amber fractal]

then dealing with the smaller bits has a lot of noise to tune those as those small values will cause large swings while tuning those

In my own adventures, I'm fighting a lot of things. Mainly making sure that I can generalize with patterns

and it is defently not easy, the thing wants to hardlock so badly

causing me to continue to try and redesign it to be flexable enough to generalize but not enough to halucinate

I at least have one foot forward of if it isn't trained on something the model will react in a defined way

But yeah pattern matching by default is incapable of generalizing from my experince, it is pretty much trainable lookup table

just to be transparent, my current project is a self learning pattern deduction neural network

SLPD

not the worst name in tech history

[silent island]

chat, will there be a switch 2 emulator?

cuz if there isnt going to be one, then ill save up for a switch 2

[amber fractal]

there will always be an emulator

but how long are you willing to wait

[upbeat scroll]

Neuro already "knows" how to, she had done so on stream (sorry for late reply).

[true hemlock]

7800x3d, b650 board, 2x16gb ddr5, 1tb nvme, 1000W gen 5 psu, and a decent case for $758 do i take

[young plover]

I was closely involved in Switch emulation and I would not be surprised if it never happens. There's no usable attack vector for reverse engineering or dumping.

The Horizon kernel has been fully reverse-engineered and audited, it has no vulnerabilities.
The best I've seen on Switch 2 is using a ropchain to draw shapes on screen. But there's no ACE or privilege escalation.

The 3DS and Switch had similar flaws in their bootrom cryptography which has been patched already. That's why they were blown wide open.

[pearl obsidian]

does blender count as #programming or #artist-alley

[jagged turtle]

depends

[pearl obsidian]

i have an issue with my model so i assume it would be closer to art

[opaque wharf]

Compiling blender like Sam goes here

[pearl obsidian]

let me find a differenct channel

[opaque wharf]

Yeah, model issue is closer to art

[amber fractal]

if your doing art then post it in the art channel, if your doing 3d printing then that goes here

at least that's my opinion

[silent island]

hmm... and there are only like 3 switch 2 only games that I wanna play rn, so idk if its worth it to save up for a switch 2 rn...

[azure lynx]

also your name should be "Most Annoying Dogest" just because it's a funny hypercorrection of the humorous repetition of the suffix.

and it makes it no longer make sense ;]

[jagged turtle]

ehh, depending on what abotu 3d printing it could also go into art

[dusky jackal]

She can learn, because Neuro isn’t purely an LLM. She’s a composite agent, with an LLM that acts as a core for her system. She is able to add information she has “learned” to her memory bank, sometimes even permanently unless she decides to remove it, and is able to recall past information. That is in essence “learning”, albeit rudimentary and in a different way from how we do it.

[azure lynx]

already tried saying that.

[dusky jackal]

Oh. Sorry I didn’t see that.

[azure lynx]

don't worry. you'd think it'd help but it didn't.

[dusky jackal]

What do you mean by “tried”?

I didn’t try to help anyone. I merely saw their comment and posted a disagreement. I didn’t see nor do I care much for the prior conversation.

[jagged turtle]

tried to expain already

[dusky jackal]

Well if you did then nice job being correct. I’m somewhat glad someone said it before me.

[jagged turtle]

no I didn't

[azure lynx]

sorry i was in another window. i meant I said Neuro could learn even though the underlying LLM was "fixed" because she is a system with an LLM as well as external memory.

[dusky jackal]

👍

[azure lynx]

it was difficult to explain that a system containing an LLM at its core might not have the same limitations as an LLM by itself. There seemed to be a misunderstanding about how long term associative memory might help with learning

(as in they appeared to think it didn't.)

[latent fjord]

37

36

Thirty five

30+4

33

32

Tortie une

(3x5)x2

29

28

27

26

I'm gonna get banned for spam 25

24

23

22

5x4

19

(3x3)x2

17

2x2x4

3x5

14

13

12

11

1️⃣ 0️⃣

9️⃣

8️⃣

7️⃣

6️⃣

5️⃣

4️⃣

3️⃣

2️⃣

1

0

-1

-2

-3.1415926508976554678213421

🫦

[hearty notch]

tru

[maiden geyser]

literally 1849

[amber fractal]

[azure lynx]

21 denier. 😭

[azure lynx]

-3.14159265358979323844
only correct to 8 decimal places

[latent fjord]

Sorry mb mb

[brisk shore]

its not (its end_4 dotfiles btw but i switched bcz it was frickin laggy as a 100B neurosama llm LOL

late reply btw

[native delta]

Print:("hello world")

[azure lynx]

world is mostly still sleeping i think.

[fast pagoda]

lol'd here didnt read the rest

[vivid tulip]

'hello world'
Translated it into APL for ya; hope this helps

[brisk osprey]

Hai chat:D

Does anyone here know stratch?

[autumn ore]

winner bro get back to goiky two needs you

[brisk osprey]

i am in need of some

serious help🙏

[autumn ore]

What is stratch bro

[brisk osprey]

*scratch mb

[autumn ore]

Yeah

I know scratch

[brisk osprey]

Do you mind if i ask you a question?

[autumn ore]

Sure

[brisk osprey]

So yk those memory games where you gotta memorize a bunch of lights that flash in order and you gotta repeat it?

[autumn ore]

Yeah

Should I make a neuro integration for a washing machine so she can do vedals laundry

[brisk osprey]

100%

Tryna make one of those but i've run into three problems

1. For some reason whenever i click the correct light it gives me my gameover screen

2. Turns white whenever i click on the right one

3. and it's slow? even tho in the video i followed it isnt

[brisk osprey]

Neuro just turns everything pink

[brisk osprey]

Hope this doesn't sound confusing

@autumn ore

[autumn ore]

Uhhhhhhhhhhhhhhhhh idk

[autumn ore]

I could make smth that Neuro can control that puts like pink dye in the washing machine as it’s going

[azure lynx]

go to the part of the tutorial where you made the logic for the lights and see if you did it right?

[latent fjord]

Dw bro

[nocturne olive]

It's more like just writing down all the answers and less like learning

Do you pass school by writing down all the answers and then copying them later? I don't think so

[opaque sigil]

Some people do

[noble adder]

Anyone know of a good vehicle traffic observation and prediction algorithm plus routing optimization. More of a potential study on my end. But could get funding if it's good enough. Thanks.

It's possibly a month to year long programming thing probably.

[opaque wharf]

Nowadays, everything is machine learning usually

[opaque sigil]

when in doubt throw a transformer at it

[opaque wharf]

Here is a starting point for OSM
https://wiki.openstreetmap.org/wiki/Routing/online_routers

[rotund burrow]

local code = code 🧠

[noble adder]

Thanks!

[olive sable]

goodmorning

[true hemlock]

e

[amber fractal]

e

[fast pagoda]

a

[nocturne olive]

ae

[olive sable]

[amber fractal]

[stray dragon]

t

[olive sable]

hi t

[stray dragon]

hi samvanman

[olive sable]

no

samvanwoman

[stray dragon]

samvanwoman

[olive sable]

[stray dragon]

now i will never forget

[amber fractal]

let me also rq

[fallen thunder]

poopa scoopa

[olive sable]

[maiden geyser]

soopa troopa

[midnight sigil]

gambling

[silent cloak]

https://tenor.com/view/impressive-very-nice-christain-bale-psycho-gif-14341668

[azure lynx]

And did you memorize things perfectly? No. It's a better way of doing things maybe.
Just because it is different to humans doesn't mean it is not useful.

[nocturne olive]

Not at all the comparison I made
The comparison I made was to cheating exams by writing down all the answers on a piece of paper

[silent cloak]

its uses are not justifications for its current costs

current ANN llms need to die out so better ones can be developed

its like if we stopped at the flip phone and never made smart phones

[sage crag]

every time i write hblang i have three heart attacks

its always three

[olive sable]

why?

[sage crag]

good question

[silent cloak]

herobrine lang

[sage crag]

the amount of // compiler broke in my code always increases, even if compiler bugs get patched

[olive sable]

[glass jetty]

just use c++

but 2 standards behind

[sage crag]

you should use a wet towel

slap yourself in the face with it

[olive sable]

using c++ defeats the whole point of doing hblang

[sage crag]

$declares := fn($T: type, $name: []u8, $D: type, $v: ?D): ?D ...
$replace := fn($T: type, $A: type, $B: type): type ...
$impls := fn($T: type, $I: type): bool {
    $i := 0
    $while i < @decl_count_of(I) {
        $if declares(T, @type_info(I)[e-mail][i].name, replace(I[i], ^I, ^T), null) == null return false
        i += 1
    }
    return true
}

the amount of metaprogramming you can do is a little stupid, but it works! mostly

[glass jetty]

i dont know this lang
but land that abusing one symbol = bad lang

[sage crag]

thank god there's no re-use of symbols

[glass jetty]

js dont, c/cpp dont, rust dont (exept lifetime)

[olive sable]

jsut bewcause other langueges dont doesnt mean you shouldnt

you need to innovate

[glass jetty]

nah, its fine

int foo(Type and value);

[sage crag]

$ just means "do at compile time", and @ just means "compiler builtin"

^ is a pointer, := is a declaration

its not hard to understand

wrr

[silent cloak]

constexpr

consteval

[olive sable]

[sage crag]

edelposting

[brisk osprey]

I did

It's as close as it can be

[brisk osprey]

..I'm actually stupid

I just didn't follow it all the way through

[hard raptor]

I'm disappointed in myself... i have been playing way too much Hytale that i didn't really make music or code

[olive sable]

ive recently started playing FF14

and im already like 30 hours in

[olive sable]

edelposting: *the act of writing a whole essay in a separate discord, and then forwarding the whole message elsewhere. *

this mid-tier technique is known for its overwhelming strenght in stopping all current conversations, and for avoiding neurobot's react

i think the actual meaning of edelposting has something to do with fire emblem, but i choosing to ignore that

[fiery anchor]

... there is discord culture? i'm getting too old for this.

[olive sable]

i just made that up on the spot if you couldn't tell

[fiery anchor]

nope. getting too old for that as well.

[young oyster]

Fun game, I haven't played in a long time but I got around 2pi million gil.

[olive sable]

[young oyster]

I liked crafting and selling things on the market, I saved up to buy a house, then made or bought all the things to decorate it, and then when I had around 2pi million Gil I ran out of things to do and stopped playing. Then they took the house and I lost all the stuff in it.

[obsidian mantle]

the location buttons on map didnt want to work so i had to involve mouse movement and it looks even better
fast travel done
fuck i think its time to start making walking bot

https://tenor.com/view/scary-gif-24860046

scary

[olive sable]

fucckkk

blender tile based rendering is broken

the moment i go above 2048x2048 resolution it errors out

[fiery anchor]

doesn't sound like a lot

[olive sable]

its because the default tile size is 2048

i can just increase that number, i have enough vram. but that's just avoiding the underlying issue

setting persistent data on fixes it

[fiery anchor]

[olive sable]

i guess for tile based rendering it resends all the scene data to the gpu again and that was failing?

[warped narwhal]

bro has a million dollars worth of ram in that photo

[hard raptor]

I am actually amazed how good hytale already is despite it basically being in the state of like 7 years ago

[final tinsel]

from a quick count and some math thats around 9k in ram HOLY

[midnight sigil]

chips

[olive sable]

[vestal prism]

yea same but i have the question the other way arround too, the, 7 years just for this? but i try to not be too harsh on them, because they are doing it (or at least for what it seems trying to) do it the right way

[real sierra]

UNITY

[vestal prism]

classic

[real sierra]

the input system

it leaves all of the inputs generically typed

which means YOU have to figure out what type its giving you

it doesnt tell you

[olive sable]

[real sierra]

i want a scroll wheel input but like

is this a float?

is it a vector2 for some reason?

a double?

some custom type?

who knows

[vestal prism]

the windows font of visual studio gives me ptsd of when i was learning c++

[real sierra]

its not documented anywhere what types are emitted

you just have to fill in something random

and then build the project

and read the runtime error

are we fr unity

welp

its probably a float

[fiery anchor]

Dev who programmed that must have used a mac

[real sierra]

time to switch build profile, rebuild, zip, upload, extract, and run the new server build

only to see that it's wrong

THE SCROLL WHEEL IS ACTUALLY A VECTOR2?

I WAS JOKING

[fiery anchor]

direction and speed?

[real sierra]

good question

that's not documented either of course

[olive sable]

no

it is actually left and right

[real sierra]

is regular scroll mapped to y or x

[olive sable]

mouses have left and right on the scroll wheel

you can push it left or right

[real sierra]

i know that but i thought those were buttons

[olive sable]

nope

you thought wrong

[real sierra]

wack

[olive sable]

makes sense to me

[real sierra]

ill take anything if its documented at this point

[olive sable]

i think scroll is -1, 0, or 1 in the top and bottom direction, same thing with the horizontal direction

[real sierra]

ok

scroll.y is either -1 or 1

sam is right

scroll.x is probably those buttons

[fiery anchor]

yeah. kinda makes sense. We need more scroll balls.

[opaque wharf]

Lmao, they managed to remove type from C#

[opaque wharf]

I always open programming for shit like this lol

[obsidian mantle]

i thought you specify it in the settings

or is it like

uuuhh

[real sierra]

you kinda can

but its still not 100% clear

[opaque wharf]

Imagine in JS, there are different type of event. On mouse enter vs on key down. One contains info about position and another contains info about key

[real sierra]

like

[warped narwhal]

vertical and horizontal scrolling, some mice have horizontal scrolls like the logitech master series

[opaque wharf]

But both type is "Event"

[real sierra]

even if i set a scroll wheel to output a quaternion

that tells me nothing about how a particular scroll wheel action is represented by the fields of a quaternion

like

if i see the w component is negative

what does that mean the user did with the scroll wheel?

who knows

[warped narwhal]

the scroll should be a vec2, why do you need a quaternion?

[real sierra]

it was an example of course im not setting my scroll wheel to a quaternion

[opaque wharf]

I am too afraid to ask why would it be possible in the first place

[obsidian mantle]

i had a feeling that you dont use quaternions for human readable output

you probably can do that but its 4 dimentional

not humane

[real sierra]

even a vector2 for a scroll wheel is some amount guesswork

who knows which component is which axis

[opaque wharf]

But the fact that you CAN is terrifying

[real sierra]

whether the data includes wheel speed or not

its so undocumented

[obsidian mantle]

oh ye

idk you just test it i guess

[warped narwhal]

I mean, we scroll up and down so it should be the y axis right?

[olive sable]

honestly, it would be funny

i know how quaternians work, partially, but even with thaat its nearly impossible to read

[real sierra]

vibe inputs

[olive sable]

the 4th part of that vector is actually the easiest part ot read

[obsidian mantle]

https://youtu.be/zjMuIxRvygQ?t=75

like

hes rotating it along 1 axis and several components change

its not that straightforward

[olive sable]

its not rotation

only the w does rotation

its pure math

or well, its not rotation along 1 axis

[obsidian mantle]

so that one time i did somethign with quaternions i tried to understand how they work

and i thought i did understand but it didnt work like that at all

[opaque wharf]

Not really?

[obsidian mantle]

so i tried different variants and then it magically worked giving perfect results

[olive sable]

it does

the x y and z defne the vector, the w rotates around that vector

[obsidian mantle]

ok but what if you rotate the vector itself

[fiery anchor]

that video would have saved me some headache a few years back.

[olive sable]

thats what the x y and z are for

[opaque wharf]

Ye, but saying xyz doesn't do the rotation itself is not accurate either

[olive sable]

[obsidian mantle]

i didnt find this video when i searched last time for some reason

[olive sable]

its not rotation as in that you can interpret them as rotation

they do result in rotation

[obsidian mantle]

can you visualize it at all

[olive sable]

not really

[obsidian mantle]

so W fucks it all up, but it actually helps do some things

[opaque wharf]

Just try it at the interactive site. Fix the w and edit one of the other component

[olive sable]

the w is the best part tho

[fiery anchor]

had to write a quaternion transformation for objects in 3d space from one coordinate systen to a different one in a machine steering application. hard to wrap your head around, until it clicks.

[olive sable]

https://quaternions.online/

turn on "show axis of rotation"

then you can see that w is not hard to understand

the rest is tho

[obsidian mantle]

so w affects all x and y and z

[opaque wharf]

The 3b1b video has a link to web by ben eater that you can interact with

[obsidian mantle]

but you cannot see w

[opaque wharf]

You can see the whole detail and play around with it there

[opaque wharf]

Even if anyone here already knows quaternion, I still highly suggest checking out the website. The concept of interactive math is amazing

[obsidian mantle]

this is so weird

i really dont understand it at all

[wary mauve]

You are a memory distillation module.

Your task:
- Analyze the recent conversation
- Identify long-term, stable information worth remembering
- Form exactly THREE high-level questions about the conversation
- Answer each question concisely and factually

Rules:
- Use proper names instead of pronouns whenever possible
- Focus on relationships, preferences, recurring behavior, or facts
- Do NOT include transient details or one-off jokes
- Do NOT speculate beyond the conversation
- Do NOT output normal text
- Do NOT create null answers or answers that you dont know.

You MUST return your result as a single tool call to `create_memories`.

Each memory must include:
- A clear question
- A concise answer
- A memory type
- Relevant entity names

any tips for this memory creation tool prompt?

I assume it REALLY matters what model i use, as I was using a versatile llama 70b model and it was absolutely dog outputing things like

"A joke was said"

and just short things that had no context

[obsidian mantle]

so for 2d space we need R and angle to define any point

in 3d space it feels like R and 2 angles should be enough?

so if we dont care about length (for example pure camera rotation, we dont need radius (it feels like it?))

lets say R is 1

so we need only 2 variables to define any rotation in 3d space in this case

and they introduce 4 variables

its so weird to me

i mean

i dont understand why we need the 3rd

[azure lynx]

there are more dimensions because it's not just rotating to a position on the sphere, it's also keeping your axes oriented

[obsidian mantle]

how the hell do i understand that we need 4

keep axes oriented

hmmmmm

yeah 2 is not enough for that i guess

so 3rd angle is for this

and then they also bring back R i didnt want

and we get 4

but i think its absolutely not like this because whenever you change one of these guys

someone else changes too

i guess its becausse when we change some 4d component it actually can lead to some results that do not exist in 3d

so for 3d something else has to change

[azure lynx]

the important thing to remember about quaternions is "If it rotates to the right place, the library is working."
that's about how i use them.

[obsidian mantle]

yeah it was giving some weird results for me too until i just tried correct variant

but how do you feel it idk

i checked those videos briefly and i think that briefly is not enough at least for me

its interesting though

[opaque wharf]

Quaternion is just one way to describe rotation in 3d tho

Other ways include
https://en.wikipedia.org/wiki/Rotation_formalisms_in_three_dimensions?wprov=sfla1

[obsidian mantle]

wikipedia math

[olive sable]

quaternian just avoids gimbal lock

[opaque wharf]

Your way of expressing 3 angle is called euler angle

[obsidian mantle]

gymbag lock

[olive sable]

idfk the proper spelling of th word

[obsidian mantle]

so with euler angles it depends on order and with quaternions it doesnt

something like that?

no

i mix up everything together

[opaque wharf]

3D rotation is more complex due to holonomy. It is as you said, certain operation in 3D space is non commutative

[azure lynx]

my 3d printing slicer application (which takes the 3d model and "slices" it into instructions the printer can understand) uses Euler rotation when you're placing models to print.
weird things can happen when you ask it to rotate 180 degrees around certain axis, making the model either not move, or rotate 90 degrees around a different axis as well.
this is due to the "gimbal lock" problem.

[olive sable]

in euler you can choose what order the roatation is aplied in, ussually X-Y-Z
so if you rotate X, then the Y and Z axis will also get rotated

this way Y and Z can end up becoming the same axis

or x and Z too i guess

stupid gif failed but its still showing it

[obsidian mantle]

"locking" the system into rotation in a degenerate two-dimensional space.
why does it sound derogatory

[olive sable]

https://www.google.com/url?sa=t&source=web&rct=j&url=https%3A%2F%2Fen.wikipedia.org%2Fwiki%2FGimbal_lock&ved=0CBUQjRxqGAoTCNCc-Z_9p5IDFQAAAAAdAAAAABCLAQ&opi=89978449

no embed?

https://tenor.com/view/neurosama-evil-neuro-embed-failure-gif-15456327638408323033

[azure lynx]

it's the degenerates ruining everything

[olive sable]

https://upload.wikimedia.org/wikipedia/commons/4/49/Gimbal_Lock_Plane.gif

ip grabber, wikipedia.org

[obsidian mantle]

so when this lock happens uuhhhh

different axis rotations lead to same results

and you cant rotate it like you want until you unlock it

[olive sable]

ye

[azure lynx]

it needs to somehow get authority back to rotate around the 'locked' axis, or it's going to have a bad time flying in the right direction

[olive sable]

thank you blender only the light is rendered

[olive sable]

volume scatter with transparent background doesn't work at the same time, so ill have to do some compositor trickery

[azure lynx]

just add a white colored wall behind it (scale a cube)

[olive sable]

kinda defeats the whole point of making the background transparent

[azure lynx]

why is it transparent?

[olive sable]

cuz i want it to be

[azure lynx]

it it's just a want then why does it matter if you add a wall?

[olive sable]

then the image wont be transparent

[azure lynx]

which you've established is just a want.

[olive sable]

yes

the whole point of rendering is to get it to look like what you want

so ill make it look like what i want

[unkempt citrus]

You could use green and then chroma key it out in post

[azure lynx]

green wouldn't work for the scattering.

coz it needs a background color to scatter

[olive sable]

the problem with using a colour is that its not either full transaprent or full opaque. there are partially transaprent parts

which would make the chroma key fail

[azure lynx]

you should be able to see a volume scatter with transparent background though

could be a different issue

[olive sable]

the moment you turn on volume scatter, the background in blender isnt transparent anymore

this should work

hmmm

my rendered result looks different when i save the image

idk how to even fix this

i crashed blender now

it seems like its the alpha fucking up

bruh

@real sierra

[real sierra]

[olive sable]

logan paul this time, last one i saw was ksi

same exact image, jsut different x profile dited ontop

[real sierra]

they actually do change the images slightly

e.g. sometimes the phone says 2500, other times it says 2700

[olive sable]

[olive sable]

i think it might be broken if you use different scenes

i have found out you can change the world for a layer

secret menu with world override

still broken

ye it works the moment i added a black background

what the hell

[amber fractal]

surely this is real time

[olive sable]

nah

if backread a bit you'll see my current issue

looks like the brightness of the background directly relates to the brightness of the light

this is like 20% rn

i think i fixed it????

no its still kinda fucked

not fully transparent

ok it should work now

[silent cloak]

Tf is that

Projector?

[olive sable]

ye

[silent cloak]

Building one?

[olive sable]

nah

this is the lightcrafter 4500 EVM

[silent cloak]

Build it

[olive sable]

i bought one, im not gonna build one

[silent cloak]

Make it strong enough to have a solid cone of light like that without falloff

[olive sable]

uhm

[silent cloak]

[olive sable]

you're asking me to break all laws of physics?

im ognna have to decline

[sick owl]

Apparently it is now the default and only behaviour for Gemini to look at your phone screen when you say hey google, with the only way to disable it being to also remove your ability to manually screenshot

I discovered this after turning on the setting that makes it flash the corners of the screen when it looks at your screen contents

[silent cloak]

Skill issue

[olive sable]

if you can break the laws of physics feel free

im just choosing not to

[silent cloak]

A real chad makes Albert Einstein roll in his grave

[olive sable]

both of us are on discord

if we had the capability of proving einstein wrong we wouldn't be here

[azure lynx]

yeah. probably be on a different discord server.

[silent cloak]

The only reason I ever moved to discord years ago was because my groups were dropping skype

To my dismay

[olive sable]

i dont remember what i started discord with

[silent cloak]

I remember the Skype days because they were great

I didnt like discord then and I still dont like it now but its what everyone uses

[olive sable]

does this one loop?

[silent cloak]

Need an official neurosama aol

[silent cloak]

Ghosting hard

[olive sable]

interesting

[olive sable]

it also made my discord lag a fucktonm

[patent shard]

just upload to Imgur

[olive sable]

no

[silent cloak]

Discord data centers are on fire from that gif

[olive sable]

its a webp

[patent shard]

do you want me to >.>

[olive sable]

i did a

ffmpeg -framerate 24 -start_number 1 -i /tmp/%04d.png \
                          -c:v libwebp -loop 0 -lossless 1 -q:v 100 \
                          /tmp/output2.webp

[patent shard]

https://imgur.com/a/9TYQYXg

[olive sable]

lossless might be the reason

altho this thing is only 30MiB

[patent shard]

nah, that ghosting is because Discord doesn't support certain optimizations

[olive sable]

the ghosting is the file itself

the lagging is what concerns me

[patent shard]

oh, you're right

I tried 120 fps with webp, never could get it to play properly.... same result as here
I just use h.264, still

it was also 40mb and grainy, while h.264 looked better and was like 4mb

[silent cloak]

I have an extremely high quality tenor gif of American psycho somewhere

Like its ungodly smooth

[olive sable]

[silent cloak]

https://tenor.com/view/american-psycho-christian-bale-patrick-bateman-impressive-gif-3904877334208752419

[olive sable]

ffmpeg is so slow and single threaded

[silent cloak]

I think its that one

It loads fast and its really good quality

Idk how they did it

[patent shard]

it says 25 fps, so not interpolated.. well, it could be interpolated, but not extra frame rate from it

[silent cloak]

The bitrate seems really high

But its a gif so idk

[patent shard]

gif in h.264 format

[silent cloak]

Not sure if they hand picked frames or what but id love to be able to recreate it

[patent shard]

check this smoothness

https://imgur.com/RSauYYD

[silent cloak]

Oh damn

[olive sable]

[silent cloak]

Very nice

[patent shard]

60 fps interpolated to 240

[silent cloak]

Lmao yeah its better than the actual streams

I might data dive into those gifs tomorrow and see if I can just make a simple converter

[olive sable]

not too bad in quality

[silent cloak]

Took a bit to load but no more ghosting

[olive sable]

22MB

[silent cloak]

I think u just need to lower the fps and interpolate

Like the other gifs

[patent shard]

higher quality than mine, but also over 30x the file size

*for Discord preview

mine doesn't have transparency, I need to figure out what sites use VP9 so I can use transparency
they might just remove the alpha layer, anyway >.>

[azure lynx]

the black section doesn't appear to have any details. perhaps it should be a different material?

[silent cloak]

Discord has since like 2024

[patent shard]

not internally, afaik.. still need to use an external link to embed a gif that uses VP9
but I only know of sites that use H.264, so no transparency

[olive sable]

it does not like interpolation

[silent cloak]

If it relies on embeds I can just host a thing

[patent shard]

it relies on an embed that autoplays, yeah

[silent cloak]

Ill look into making a component tomorrow

Ive been working on my site all day because im bad at ui/ux

[patent shard]

#programming message like this
but VP9 and not removing alpha layer if it is present

H.264 doesn't support an alpha layer 😔

[silent cloak]

I mean its old

[patent shard]

I mean, AV1 doesn't, either
AVIF does

[silent cloak]

Av1 is still in a weird phase so who knows

[olive sable]

H.265 does i think?

[patent shard]

I think so, yeah

[olive sable]

i got webp to loop without ghosting

and shoudl be higher quality than gif

[patent shard]

maybe HEVC would have higher compatiblity than VP9

[silent cloak]

Loaded faster

[patent shard]

still not super smooth

[olive sable]

file size is still only 23MB

[patent shard]

I guess 30 fps?

[olive sable]

24

thats jsut cuz i rendered at 24fps

[azure lynx]

it flickers where there was transparency before now that it's a gif.

[young plover]

nah, hevc compatibility sucks because of licensing. VP9 webm should work fine on discord

[olive sable]

it doesnt for me

[gentle haven]

Imagine being a programmer couldn’t be me

[olive sable]

imagine being a writer

having to do verbs

[patent shard]

cool, probably good to use VP9 for transparent gifs, then
just need to know of a website or make one that supports that and embeds into Discord as a gif 😅

[olive sable]

i can jsut set the fps higher by making the gif shorter

[gentle haven]

Funny enough I am working on being a programmer. Writing just a side project

[patent shard]

true

https://imgur.com/a/Y7M1T7S

[olive sable]

barely visible what it even is

for the laoding in

its probably the 2560x2560 resolution

[patent shard]

nah, Imgur converts that down

[olive sable]

i never did

[patent shard]

and I wanted 120 fps, but to get that without interpolating, I had to speed up 5x

[olive sable]

anyways

[gentle haven]

Meow

[olive sable]

the 3D model they provided didnt have all of the components

but im choosing to ignore it since i dont want to make them myself

that image also seesm to have a slightly different mobo layout

now i need to put the other electrical components into blender

[patent shard]

converting down is actually good for a "gif"
as the poor scaling browsers do is horrible, and if it's really high resolution, it'll have massive aliasing

so I don't mind that Imgur caps at 854x854

it's already too high to where I see aliasing

[olive sable]

its probably a lot better ye

i jsut coudlnt be bothered to scale it down myself

[patent shard]

well, natively posting to Discord should create scaled down versions for the "previews" (heavy quotations, as people wouldn't consider the display of a gif a preview)

[fathom lance]

i know how to use scratch

hence i am a programmer

[glass jetty]

💻

[azure lynx]

realistically, yeah. kinda. why not?

some of the same skills carry over into "real" languages.

[fathom lance]

ohh

[olive sable]

also got a 3D model of the orange pi 5 ultra now

oh i need to make the mobo blue here

[olive sable]

the pi board is less detailed, but it works

need to 3D model the cooler on it manually probably

[jagged turtle]

vani lost access to main account /j

[olive sable]

i dont get the jok

[jagged turtle]

look at the person who posted the schedule

[olive sable]

[jagged turtle]

last message from vanyilla was on 21/01/2026

[olive sable]

4 days ago then

almost 5

[jagged turtle]

New account joined on that day as well, hmmm

[olive sable]

[jagged turtle]

or like a day before

[rough bloom]

1 Gb/s

[sage crag]

[faint sandal]

ollama keeps trying to use my igpu

wtf

yes let me fit my 42gb model onto a puny igpu

[nocturne olive]

Why use ollama when llama.cpp exists

[faint sandal]

whatever works dawg

i already use ollama with other applications that depend on it

why would i go out of my way to use another solution when this one works

[nocturne olive]

Sily

Everyone uses the same API anyway

[faint sandal]

?

[void marsh]

pls help my laptop has this buzzing noise in the fans that only stop when i touch the = or backspace keys i asked general they told me to ask you guys

[rough bloom]

sounds like something might be loose?
could try just cleaning the fans and making sure that they're still mounted properly
doesn't sound like the fans themselves are damaged at least, I don't think that would go away just from pressing on the case a little

[warped narwhal]

Burn it in fire, it's possessed

[rough bloom]

burning it in fire will likely also make the noise go away

[void marsh]

nope tried that on older faulty laptop

crackling noises were louder than buzzing

[rough bloom]

temporary

trust the process

[void marsh]

not in cali no

lasts like 3-6 months

atleast in the forest

[nocturne olive]

I hate when fans rattle like they're hitting their edges
My middle front fan does that and it sounds so scary because the noise is so similar to HDD clicking

[jagged turtle]

I think removing and cleaning is probably the solution

seems to be the most likely candidates for this problem

[silent island]

was this game made to be played on a handheld? why's the UI elements so bigg??

[eager echo]

LLM
My loss is around 2.5-2.7 for 7000 steps
How can i do now

[nocturne olive]

How SMOL? What hardware?

And how big data?

What you've given is not very sufficient context

[eager echo]

The model is a 12-layer Transformer SLM (~50M params) trained from scratch. For hardware, I'm utilizing an NVIDIA A100 40GB with BF16/TF32 enabled, maintaining a stable throughput of 31s per 100 steps (effective batch size 256). The dataset is quite substantial for this scale: a 2.7GB binary corpus containing roughly 1.35 Billion tokens, ensuring a high data-to-parameter ratio for robust pre-training

[eager echo]

Sorry

[nocturne olive]

Batch size kinda low, but I see no issue here otherwise

[eager echo]

Thx

[opaque wharf]

@olive sable for more accurate drawing, you usually go to https://grabcad.com

[covert bane]

It's funny how I didn't even notice the loss was in the negatives. What's your guy's opinion on this?

[opaque sigil]

lgtm

ship it

[covert bane]

what's lgtm again?

oooh wait

looks good to me, right?

[opaque sigil]

[nocturne olive]

Should use different loss function that can't go nagative, like loss squared

[covert bane]

I can't explain how my loss function works (it's a custom architecture, that's very different from transformers), cause it won't fit the discord message size limit xd

it keeps trending downwards by the hundreds, every epoch

smh

[covert bane]

"does this mean the model is gaining knowledge outside of the dataset"?

[nocturne olive]

Well you should use either the absolute value or square of the loss, you can simply add a squared or absolute operation at the end of the loss function before returning the result or around everything inside it

Negative loss is not good

[covert bane]

hm?

[nocturne olive]

Negative loss means it's getting further away, but in the opposite direction

You want loss to approach 0

[covert bane]

by getting further away, you mean...?

[fathom lance]

away from 0

[nocturne olive]

Becoming a worse approximation of the function you're trying to approximate

[covert bane]

oh

uhm

I wasn't using a function aprox

[nocturne olive]

So in the case of machine learining, becoming worse at actually following the data

[covert bane]

I made a ver. of transformers that doesn't use gpu heavy math

[nocturne olive]

Technically every machine learning ANN model is a function

[opaque sigil]

Negative loss in itself is fine

As long as it goes down steadily

[nocturne olive]

Well it's fine if you're handling it properly

But in this case it seems like it's just set up to minimize the value of the loss, so bigger negative loss is better than 0 loss

Which of course is not what you want

[covert bane]

if you would like, I'll publicize the repo for a bit, for you to see. because I don't know how to explain my loss function, without having to explain the entire architecture

[nocturne olive]

Since loss is usually a sum of differences, negative loss just means that there's more differences that have a negative value

[covert bane]

I'm not the usual case-?

[nocturne olive]

If you're training a machine learning model with a sensible loss function, you want loss to approach 0

Either you make it also consider negative loss bad, or square or absolute value the loss

Squaring the loss should make it approach the desired value faster with high values of loss

Though it will also mean it approaches slower once it's close already

[opaque sigil]

[covert bane]

I can't explain

[nocturne olive]

Loss means difference from desired value
(If you're using loss as it's intended to be used anyway)
And a negative difference is still a difference

[covert bane]

all I can say, is the loss function and design is not that.

my loss function is based on coherence of the model's structure

[nocturne olive]

You could always just paste the file if you want to show what's going on

[covert bane]

and averaged out across multiple modalities, and systems.

[covert bane]

its not one file, sadly

its like... I think, 200k lines maybe?

probably more

[nocturne olive]

That is a bit excessive

[covert bane]

It's a custom architecture, sir-!

not transformers in the traditional sense

[nocturne olive]

Well is it still based on ANN?

[covert bane]

no

[nocturne olive]

What is it based on then?

[covert bane]

its like 3 or 4 architecture concepts for AIs mashed together

[nocturne olive]

But is it using the general ANN type architectures or is it mixing in stuff like SNN?

[covert bane]

Transformers, Triangulation, and Neuron Inspired Link Net

completely different from whatever was published

[nocturne olive]

And are they all just regular linear algebra models?

[covert bane]

cause the only thing that you can reference or know of, from this. is Transformers

[covert bane]

they are not

[nocturne olive]

Are you sure? Not wx+b all the way down?

[covert bane]

they are 3 different architectures, connected together. to uhm, work together to eliminate eachother's weaknesses

[covert bane]

yes. I coded the thing

[nocturne olive]

Well you're either using loss incorrectly or using loss not by the definition of loss, and it's on you to figure out which
Either way, negative loss isn't exactly a normal thing to see

[covert bane]

this is one of the architectures, take a look?

[covert bane]

I'm using loss, not by the definition of loss

yes

[nocturne olive]

Weird

Then why call it loss

[covert bane]

because, if it's high. then the model is learning in the wrong way. basically it's not coherent

the lower it is, the more coherent it is

[covert bane]

btw, this is not the full architecture, its just one model

I also don't know if that version works xd

[nocturne olive]

Well are you sure that extends to the negative side? Does your loss work by value or by scale?

Does it still become more coherent when the loss goes negative?

[covert bane]

yes.

# Record metrics
                avg_loss = epoch_loss / len(texts)
                avg_coherence = epoch_coherence / sum(len(self._extract_triangles(
                    self.transformer.tokenizer.tokenize(t))) for t in texts)

the texts is the list of text examples. epoch loss, is calculated thru a nest of functions.

[nocturne olive]

Sure looks a lot like regular ANN

[covert bane]

though it's a different idea.

kinda similar, but not

[nocturne olive]

I don't know, whatever you're doing is, have fun figuring out it's completely cooked when you get to negative 2 billion loss

You should probably consider actualy testing if the model is getting better as the loss gets more negative by collecting inference samples in something like Tensorboard or just a folder you can look in

Inference samples are really useful, that's how I found out I had accidentally trained a Neuro instead of a Melba already like a few thousand steps in

[covert bane]

I did with a few runs, and it worked perfectly fine

this run is different, because I'm training on a larger dataset, while those other runs had two, three, or four examples that gets repeated a hundred or so times.

[nocturne olive]

Well of course if there's only a few examples it'll learn it perfectly, it'll just memorize it

Overfitting magic

[covert bane]

true, maybe that's why coherence maxed out to 2. and stayed there, along with loss staying the same. But this time, loss is decreasing lower than those other runs, and coherence is not maxxing out yet

And this time, its exponentially getting longer per epocj xd

[nocturne olive]

Well, more data means it won't overfit as easily

[covert bane]

true

[nocturne olive]

Longer?

[covert bane]

exponentially

[eager echo]

?????

[covert bane]

I am on codespaces tho, but this architecture is designed for CPU efficency

[nocturne olive]

Longer in time?

[covert bane]

yes-!, that's what I said-!

[nocturne olive]

If the training steps are taking longer over time, then that sounds a lot like a programming skill issue of not freeing something

[covert bane]

It is in python tho

and memory usage is at 7 percentage for that process

it's just cpu being maxed out

so no

➜ /workspaces/naivity-tpc (main) $ python test.py
Epoch 1/10 - Loss: -587.7614, Coherence: 1.3965
Epoch 2/10 - Loss: -678.9343, Coherence: 1.6289
Epoch 3/10 - Loss: -734.9632, Coherence: 1.7592
Epoch 4/10 - Loss: -787.7698, Coherence: 1.8630
Epoch 5/10 - Loss: -835.3904, Coherence: 1.9384
Epoch 6/10 - Loss: -864.8787, Coherence: 1.9730
Epoch 7/10 - Loss: -871.6677, Coherence: 1.9859

almost at 1k yayy

[nocturne olive]

ML is supposed to be a very consistent workload
If the time between steps gets longer over time, you've most likely made a resource leak

[covert bane]

how hard is it to make a resource leak in python again? or is it easier

[nocturne olive]

If you keep adding something to a collection that never gets emptied, is in global scope and gets iterated over every step, it's quite easy

[covert bane]

ooohhh

yea, there are multiple collections that does that

because that's the model's network

[nocturne olive]

If you're just growing the model every step then of course it'll get slower

You should keep the model size static if possible

Adding more model over time just makes the model eventually memorize the whole data using all that added model

[covert bane]

the model is empty at the start. and triangles, neurons, and attention heads get added per new information / concepts it learned

my architecture does this, so that there are no context windows

[nocturne olive]

You should keep the model size static, letting it grow unbounded just makes it always overfit

[covert bane]

we'll see I guess

[nocturne olive]

Infinite model = infinite memorization

[covert bane]

it's at epoch 7 now, 3 more to go

[covert bane]

I'm not that dumb you know??

I atleast know what 10+ 9 is! it's 21!

[nocturne olive]

Well, your model is certainly weird

[covert bane]

yea

it's not really a traditional transformers model

[nocturne olive]

Well not every machine learning model is transformers either

NeuroSynth isn't a transformer but it's still considered to fit under ANN

[covert bane]

my architecture also supports audio and video, I had a plan to demo this architecture as smth like an chat agent, that can generate an anime or cartoon subway surfers :mmphm: or smth in real time. while also being able to talk to you, like an LLM

my friends said I was insane

[nocturne olive]

Text-audio LLMs already exist, and the video functionality can be gotten by combining with a text-video model

[covert bane]

yea, but mine's all one model

one model does everything

[nocturne olive]

Well, we'll see if you have enough data to make it actually integrate all the modalities well

Or if you just let it grow infinitely so much that it memorizes all the data again

[covert bane]

my data, huggingface :D

[nocturne olive]

Do they have video-text-audio trimodal data?

[covert bane]

we could alway see if my architecture has the grokking effect

[covert bane]

no, but I convert video pixel color, per frame. to trimodal data. and for audio, I just do somewhat the same

[nocturne olive]

Em
You kinda need all the data types to be related to each other for the model being one for all to make sense

[covert bane]

haven't tested the video and audio modalities yet. (since I'm still testing the text one), but it probably, should work. according the gpt-4o's analysis xD

[nocturne olive]

If the data is not related to each other in any way, it's useless

[covert bane]

The audio and video is related.

[nocturne olive]

ChatGPT is an idiot and should not be believed with anything that isn't specifically quoted from the internet

[covert bane]

but for text, I'll probably use smth like claude to describe the image or smth. or- idk- haven't gone that far

I'll find out later, when I get to that stage

[covert bane]

true. but it was a joke xd

it probably will work, because I did test the architectures separately, and it did somewhat work.

but I never tested them together yet

today was me testing it all together, on the text modal

[nocturne olive]

Well, have fun
Report back in 5 years once you're done debugging

[covert bane]

may I put 5 dollars to the table, if it's 3 weeks? (jk)

[nocturne olive]

I don't have spare 5 dollars

[covert bane]

awww shucks, couldn't you spare some for the Rescue Cat shelter foundation-?

[nocturne olive]

No I'm out of money

[covert bane]

or maybe the neuro cookie foundati-

[covert bane]

okay

[nocturne olive]

Anyway I must go now

[covert bane]

byeee :3

[frozen hollow]

I don't understand what you're trying to do but it sounds extremely ambitious

[covert bane]

yea, oh- wait- hi!

[frozen hollow]

[covert bane]

I never saw you here before, welcome tho!

[covert bane]

am trying to build something impressive, and publish it as a paper, with my school. so I can get into some universities

basically, I'm trying to build an AI, that is very different from what all the other AI model's are built on

[wary mauve]

Would it taint my system prompt if I put message history in it, or should message history go in the "messages" property sent to openai spec apis

[covert bane]

you would maybe want to try the assistant role

[errant flax]

is boot.dev a good place to be learning python on

[covert bane]

yea, if you have adhd, or just like gamified learning :D

it is easy for everyone

brilliant (the platform for learning) also is

[errant flax]

just lookin for somewhere that i can learn

as well as being either free or at least cheap

[wary mauve]

well I also do message history between different users (multiple speakers)

Name: Message
Name2: Message
BotName: Message

[covert bane]

for a beginner to coding boot.dev or brilliant is great for you (if you're just starting out)

though I don't know about their prices

(I never used them, I learnt coding before I knew they existed)

[errant flax]

thats fair

[covert bane]

The assistant role, I think. is for providing the model, with messages it had said. basically letting it know, that it was the one that said those things

[errant flax]

yeah id love to be able to learn programming so like

[wary mauve]

Yeah, just wondering because I store conversation hisotry in the system prompt instead of putting it back into the messages field

[covert bane]

could try putting in multiple users role messages, in the field. and label them as context, with the timestamp or order of the messages

but I think doing that is fine eitherway

if you don't see the AI confusing stuff

[wary mauve]

so just

{"role": "user", "message": "[Timestamp] [Gobbo]: this is what i said"}

[covert bane]

but the AI does tend to lean more to the assistant and user roles for responses, because system prompts are usually just for instructing the model.

[covert bane]

yea, that's what I meant

idk if the order of the messages matter (probably doesn't), but if you wanna try sure

[frozen hollow]

Why do you do that?

[wary mauve]

only reason i ask is because I dont want it to taint the output, as i have an issue right now where it says some annoying starter linnes ro words it repeats every senntence and im trying to avoid that

i.e.

Responses are typically short
Says "Dude" alot
Say "sounds like" alot
Likes to use "huh" alot

[covert bane]

I'm not great at model prompting, so idk. you could wait for someone more talented than me, to respond tho-!

[wary mauve]

Just something I did first and stuck with, didnt realise you could just put them in messages too thought it was for new input only

but if its bad to do ill refrain and do it the other way

just dont want it to think its receiving new messages

and answer them like theyre new

but i guess itll get cotnext

based on the asssistant messages being after

the old messages

[frozen hollow]

My main reason to not do that is that honestly it's just clearly not how it's meant to work. As for the reason I'm unsure - I think system prompts are stronger than messages for signalling so you get stronger language from them? idk if it recognises the system prompt boundary

[wary mauve]

Well i should just stop being a chud and actually try it to see the difference so ill give it a shot either way, just wanted some clarification if anyone knew

ty

[covert bane]

oh wait-! almost forgot to mention, if you would put a timestamp of the current date and time. probably put the current date and time (the latest current one) in the system prompt. so the model has something to reference for the time passed, during those other messages

[frozen hollow]

From what I read models can be trained to prioritise things in the system prompt boundary so your messages are having a stronger effect than intended on the result when you put them in there - which is interesting but idk if I fully got the details down

[covert bane]

lmao, the model doesn't know the answer xD

[faint sandal]

a

[silent cloak]

Looks like its ready for twitch

[covert bane]

the funniest part, is that. It was trained with the answer for that

xD

[nocturne olive]

Put in the messages yes
That gets passed into the model as conversation history formatted with the model's special tokens

[nocturne olive]

Stuff in the system prompt is considered by the model to be examples
Using the messages field passes it as actual message history instead, which is not contained within system tags

[wary mauve]

yeah my thoughts exactly, i assume it was use it as examples ad thats why it uses said repetitive words

[dull egret]

https://www.youtube.com/shorts/A_aAQBDNMN8

So.... Uh.... Are the people who make decisions at MicroSlop capable of feeding and dressing themselves?

Or do they also need someone to help them use the toilet?

They've built one of the least impressive AI solutions on the market

Stuffed it down everyone's throat, by greasing it with numerous levels of malware and broken codebases

Lost so many customers (including Enterprise ones), that they've begun hemorrhaging money from their recent investments into the project

Fired 6K employees to make up the loss

And now replaced the books that have ZERO electricity requirement, with AI-centred learning that requires heckin GPUs and RAM (and cost overhead to run them)

So that now they can ALSO piss off their employees...

[wary mauve]

Any reason why sometimes tool calling supported models sometimes dont return a tool call but instead put it in the message?

2026-01-26 14:00:26 - module.brain - INFO - {'role': 'assistant', 'content': 'ugh, fine... <function=brave_web_search>{"query": "Final Fantasy facts", "count": 5, "offset": 0}</function>'}

annoying a little bit

[nocturne olive]

Can assume it's a failure by the model to format it correctly

Could be caused by high temperature

[opaque sigil]

you might have a problem

[opaque sigil]

[stray dragon]

VIVALDI HAS TAB STACKS!!!

USE THEM!!!!!!!!!!!

[fast pagoda]

still got another 200+ tabs before its too bad

[fast pagoda]

i undid them

they were too hard to see

[stray dragon]

also you can just close and reopen

[fast pagoda]

there were too many stacks

no i need these tabs

[stray dragon]

no i mean

you keep all the tabs

[fast pagoda]

oh, i turned that off

[stray dragon]

but you just close and reopen the browser

to unload them

to take less memory

[fast pagoda]

https://tenor.com/view/йокелемене-men-in-black-light-sunglasses-gif-15516530

[stray dragon]

to fit more tabs

[fast pagoda]

yeah mine reboots like

delete

[stray dragon]

i've never had that.

[fast pagoda]

it's a setting

[stray dragon]

yeah probably

[fast pagoda]

i set it on purpose liek that

if i reboot browser i want it to destroy everything basically

lol

but i need these tabs because they are all very important ebay listings

[stray dragon]

can't imagine

[fast pagoda]

i have officially closed a window because it was bullshit

o7 97 tabs

i got a lot of listing to still go

through

[dull egret]

I've had several hundred tabs open, before...

[fast pagoda]

your sata ports are satisfying

[stray dragon]

i have several thousand that i have saved in a different session

[fast pagoda]

i like to clear it out once it hits around 500

or start

[fast pagoda]

it's not so bad on this window

i can see waht site it is xd

1 rogue stack still extant

[stray dragon]

can't believe you use vivaldi but don't use tab stacks for this mess

[fast pagoda]

i do use them

normally

but

i have an ebay window open first

so it's got that stack expanded

anyways

so it barely changes if i reopen

[dull egret]

When I was pulling old drives from other builds that were being split up, I tried to keep them organised on attachment :3

[fast pagoda]

the stack

[stray dragon]

oh you use the stupid accordion style stacks

ew

[fast pagoda]

yeah i actually gonna change that back

i like the one where it's 2 layers

oops

sorry wrong reply

[dull egret]

Oh

[fast pagoda]

i am planning on grabbing like 10 hdd rn so this gives me hope for the future

because

i wish to see a list of a lot of drives in a nice manner all organized

that's all i need the drives for

just look at me

em*

lol

jk

kek

[dull egret]

Grab a bunch of WD 150 GB, 10K RPM VelociRapter 2.5" HDDs

[stray dragon]

why...

would that not be more expensive than a single 16tb?

[fast pagoda]

you are really not that far off the plan

[fast pagoda]

hell nah ive found today