programming

[sullen marsh]

llm online search

selenium is goat

[olive sable]

Goodmorning

2/10 morning, they opened the shades and destroyed me eyes

[unkempt citrus]

I have a foot infection so I can't walk

[dense cosmos]

blue nodes = input, purple = output, white = hidden, red connections = negative, green connections = positive

[wild wadi]

[stark needle]

Morning

[ruby timber]

Morning!

[stable fractal]

whats the difference between doing this and......

this

as you can see, in the first one, im using PRIMARY KEY(id)

and in the second one im adding PRIMARY KEY to the line where im creating the id column

is there a difference between these two methods?

ping me on reply, thanks in advance!

nvm, i got the answer

[olive sable]

even on vacation i be improvinng the code

center = np.mean(frustumcorners, axis= 0)[0:3])
is now
self.player.camera.position + self.player.camera.forwards * (depthlayers[i] + depthlayers[i+1]) * 0.5)

double the speed

[nocturne olive]

Silly

Meanwhile me training that new model despite being slightly sick

[covert garnet]

Meanwhile me advancing my studies for the next course (2 months left)

[distant blaze]

it looks like a mathematician wrote this

looks very barebones but just enough to make a computer function

its like skirting the line of doing just enough to be turing complete

that does actually make sense

would watch but gf is sleeping next to me

will watch if i remember later tho

its very fascinating

its very explicit

that's actually incredibly impressive

[safe path]

very nice

[distant blaze]

ooo

so you're one of the devs behind it

I can't judge the code quality too much because I'm only starting to learn rust. and its a completely different beast than embedded C

it seems interesting, it feels like I'm reading something inbetween rust and possibly markdown

that is interesting

aww it didn't output into the terminal

ooh, it says null?

oh yeah i didnt reaf yhe code xD

i just looked at the terminal output

oh xD

allocating memory like this is interedting

oh right

ah okay that makes sense

I haven't gotten out of bed yet

so apologies if my brain is a bit uh

sticky

and slow

is the goal to abstract manually allocating memory away?

are your allocating per variable or per program?

like ad the end goal

ahh I see. since you can do whatever you want with raw memory

[stark needle]

[distant blaze]

still, having a consistent way of allocating these things is gonna end up in more reliable and predictable programs

that's pretty fucking cool

so you're just iterating over memory

that works so seamlessly

its genuinely really impressive

[safe path]

the ableos recruitmrnt drive never stops

[distant blaze]

yeha I do xD,

I'm gonna be working with my gf on starting our own company, probably a repair company, on top of that, helping her work on her wayland compositor and some other projects too.

Also going to be designing custom electronics to sell to consumers

so as much as I'd love to I don't have time xD

I'll give it a follow!

<3, I'm not relaly sure how to follow it on forgejo though xD

and thank you <3 I'm mainly a reverse-engineer

your project is incredibly cool. I'm sure you'll find some passionate people <3

it IS cool that we're in the same timezone though

[tender river]

not "multiple columns can have primary keys", multiple columns can be the primary key

[stable fractal]

right

sorry

[stark needle]

[stable fractal]

esl

[tender river]

??? how

[stable fractal]

what is that?

[distant blaze]

very academic and structures, and generally hints ar it

[tender river]

it reads more "engineering" than "academic" to me

it just omits descs but nothing about it is surprising

[stable fractal]

what is a "holebytes"?

whats ableos

oh

its an opperating system

[warped narwhal]

I'm getting an SSL unrecognised name error when trying to clone the hbvm repo, you may need to update the certificate

the EU site is working, is there a way I can edit the config to point to that instead for the cargo build?

[tender river]

pointers are making me cry, they make optimization so much more annoying

a := 5
b := &a
b.* += 5
// now a is 10 despite being assigned once

ok what about this solution - when a is referenced with a pointer it gets converted into a cell so the compiler knows not to optimize it

a := 5
// internal
a := cell(a)
b := &a
b.* += 5

should work right? well what about this

a := 5
i := 0
loop if i == 3 break else {
    // a := cell(a)
    b := &a
    b.* += 5
}
// a is now 20

lets even assume that you can just flow a being a cell out of that loop, you still have to flow it also into the next iteration of the loop and then a is not a cell upon entering it for the first time but a cell upon entering it for the second time

so all you can do is just backtrack and make it so a was a cell all along and i hate that
alternatively unroll that loop into two separate bodies which is weird as well but possibly a better solution

i'm interested in making a compiler + LSP to gain experience but pointers and comptime make everything hard

[sick owl]

Vector just learned to drive over my keyboard cable

[stark needle]

My dad bought 4 dgx spark under his company

[hard shale]

I can't find much about hblang so I will presume its a new programming language, does hblang not have a standard library atm?

oh, so did you architect the language?

assumed from your answer

ok, what about a documentation for hblang? does it not exist yet?

[stark needle]

Technical debt

[midnight sigil]

my ISP finally gave me an ipv6 address

it doesn't really matter since the ipv4 is on public anyways

[hard shale]

The syntax looks like a lot like the zig language so that's understandable somewhat. supposing someone would want to contribute, for reference about the syntax and possibly semantics is this a confronting compiler?

a proper hblang's compiler, following its rules instead of creating extensions of them unmentioned

OK

[olive sable]

I went to castelo de s. Jorge
Eating burito now

Got a lot of inspiration from it

[stark needle]

https://www.upwork.com/freelance-jobs/apply/Looking-for-Computer-Science-Journalists_~021925928967860465601/

Man

What is this job

100$ for 3b1b level video script on technical topics

Expert level

Lmfao

[loud thicket]

You need a PHD and many hours on leetcode

[dense cosmos]

I used a goto

[atomic sonnet]

trying to port my java baseline dct jpeg decoder to c++ and....

[dense cosmos]

[atomic sonnet]

[dense cosmos]

vs

[faint sandal]

this took 75 SECONDS to get past

[inner pike]

this one is just confusing cause technically that is not a bycicle but who knows how the fuck it works for captcha

captcha is absolute bullshit sometimes

[scarlet arch]

Yeah I also don't understand them cause I'm totally not a bot

[trim valve]

I barely get bot checked at my uni

[atomic sonnet]

I get capcha every day as I'm in a uni dorm so a lot of ppl use google and its just loooooong sometimes with a lot of retry

[trim valve]

anubis hits me a fair bit

[trim valve]

I guess it probably depends on how your uni does internet

[atomic sonnet]

this and it seems to have gotten worse with time

[trim valve]

😔

[trim valve]

99% of the time I'm only getting hit with extra captchas if I open an incognito tab and even then that barely happens

[atomic sonnet]

I'm... always* in incognito

[noble zodiac]

the amount of free training data generated daily by captcha is insane

[stark needle]

[trim valve]

the occasional brainworm telling me to buy a gpu for cuda keeps calling to me

[hoary lion]

hello b200

[spark kelp]

Steam captcha takes me 15 minutes for some reason 😭

Yeah

my acc profile is even better

since it removes the spaces

It looks different for a lot of devices

[trim valve]

I really wanted to make a plugin that lets you rename people client side

keep putting it off due to lazy though

[stark needle]

uohhh this is based

also who tf removed windows fax and scan

that app was goated

[opaque sigil]

The money clock is ticking

[stark needle]

7$ or osmething

so expensive

[trim valve]

did your just disappear or something

[stark needle]

apparently its been replaced

by a shitty app

"Windows Scan"

[trim valve]

like is there nothing at C:\WINDOWS\system32\WFS.exe

[trim valve]

I would offer you mine but that sounds sketchy af

[stark needle]

i am trying to sign up for scholarship

and bruh

they ask for everything

shit i doubt my parents would ever give me

full tax papers of my mum and my dad and his wife

all documents from when they separated

[scarlet arch]

No. Rule 15:

Don't use an invisible name.

hate pulling an "erm actually" lol

[atomic sonnet]

update

[dense cosmos]

128 inputs x 129 outputs per network in my visualizer

kek

[scarlet arch]

probably wanna use alpha to signal how close to 0 the values are

surely that should help

[dense cosmos]

i am

also, made it bigger

[scarlet arch]

interesting artefacts

very cool

[stark needle]

uoooh https://www.youtube.com/watch?v=IwglW_hIL_g

[trim valve]

😭 that audio sucks

[scarlet arch]

maybe make it uh.. not linear? like.. idk how to explain cause I can't remember the function I'm thinking of. but like.. from -0.5 to 0.5 its basically invisible or only slightly, and then it ramps up a tiny bit, and then on the last 10% to -1 or 1 it actually goes full alpha

should give a better indicator about strong connections

[stark needle]

all money went into training gemini 2.5 pro deep think

[atomic sonnet]

fixed

tested a colored image and another error to try to fix (apparently I read too much bytes ?)

[dense cosmos]

hmm, good point

[scarlet arch]

maybe try just f(x) = x²?

[dense cosmos]

I was thinking of doing a sparse intialization too

because there's like 16k connections there

and that's a lot to optimize

[scarlet arch]

is this your neuroevolution result?

[dense cosmos]

(for evolution)

[scarlet arch]

when I did neuroevolution my networks started out with just the input and output neurons. not even connections between them lol

[dense cosmos]

not result, more like I'm trying to do character prediction using one-hot-encoding

[dense cosmos]

my network would crash if I didn't give it any connections rn lol

[scarlet arch]

ah well. yeah. that'd do it

back then I designed it in a way that I simply feed it input values and then take the output, regardless of whether anything is connected or not

initial networks mutated fast that way. speciation was also off the charts due to so many initial variations

[atomic sonnet]

my code try to read 65kb of data in a 4kb file

[trim valve]

[atomic sonnet]

aparently the error is reading an huffman table
I assume a 65459bytes huffman table is not ideal for image compression

[nocturne olive]

This thing is just sitting here, computing

[dense cosmos]

even bigger!

[desert wave]

https://tenor.com/view/patrick-menacingly-spongebob-standing-there-gif-19452999

[dense cosmos]

Avg Fitness: 0.0840659
Max Fitness: 5
Min Fitness: 0
Species Count: 5
Species:
[
        Species no:     :0
        Size            :182
        Avg fitness     :2.67033
        Max fitness     :11
        Min fitness     :0
]
[
        Species no:     :1
        Size            :7
        Avg fitness     :2.14286
        Max fitness     :8
        Min fitness     :0
]
[
        Species no:     :2
        Size            :1
        Avg fitness     :5
        Max fitness     :5
        Min fitness     :5
]
[
        Species no:     :3
        Size            :8
        Avg fitness     :4
        Max fitness     :10
        Min fitness     :0
]
[
        Species no:     :4
        Size            :1
        Avg fitness     :3
        Max fitness     :3
        Min fitness     :3
]
[
        Species no:     :5
        Size            :1
        Avg fitness     :0
        Max fitness     :0
        Min fitness     :0
]
'Evaluation':       19929328700 ns (19929.3 ms)
'Culling Game':       6029500 ns (6.0295 ms)
'New Generation':       842239300 ns (842.239 ms)
'Generation':       20779765100 ns (20779.8 ms)

20s per generation

for 200 individuals

[stark needle]

ooo u got a left side window pc??

[nocturne olive]

Very silly

[stark needle]

i mean right

[nocturne olive]

It has a window on both sides

[stark needle]

wait so can u look through both or does one just show the back

[nocturne olive]

The side that's visible in the image only shows a couple cables, the back metal panels, and my 8TB SSD

[stark needle]

i see i see

fair enough

[nocturne olive]

The other more hidden side gives vision directly to the main internals, having the GPU, CPU cooler, RAM and motherboard visible

[stark needle]

fair enough yea im just used to seeing pc cases where on the right side it's just metal since you cant see anything anyways

[nocturne olive]

I guess it's symmetric externally

If just I had a better GPU in there

[atomic sonnet]

found the issue
I was doing an | from a uint16 with a negative char
so the char got cast as its equivalent in signed 16 bit. then got xor and become unsigned making a big number

now I can decode big (5120x4920), subsampled, colored, baseline dct jpeg 1.1 images

my java code is still faster by 30% somehow. need to profile

seems like wsl slowed down my code. with window build I'm ~12% faster than java

[nocturne olive]

Silly

[prime ridge]

bro why are you measuring in nanoseconds 😂

[dense cosmos]

I originally was using it on small functions

[prime ridge]

even still 💀

it still takes over 1000 ns to do 1+1 bruh

I assume ur using python

when ur using python there is 0 use for nanoseconds cuz we a lil slow

[atomic sonnet]

he is using c++

and cuda

[trim valve]

why wouldn't you

[prime ridge]

C++ if ofc different

I thought he was using python

[trim valve]

o

[prime ridge]

nothing in python runs on the nanosecond scale

[opaque sigil]

Do you have the project saved in the windows filesystem huh

[atomic sonnet]

maybe ? good idea to test

[opaque sigil]

if you do then don't, the filesystem bridge between windows and wsl is painfully slow and will hurt performance if you read anything

[rigid snow]

WSL FS SUCKS sooooo bad

[trim valve]

if you stay within wsl its fine

[opaque sigil]

you can just have your project inside wsl and it's a non-issue

it'll be faster than the windows fs too

[atomic sonnet]

am just lazy ngl
window is temporary anyway

and my java benchmark was running in window too so apple to apple comparaison

[opaque sigil]

i guess that's true

[rigid snow]

am i being gaslit or has something changed at some point, i remember it being painfully slow even in linux fs

[opaque sigil]

that's still weird that it'd be slower inside wsl

i think wsl1 had some performance issues with the filesystem?

[atomic sonnet]

I do kinda heavy file opperations and probably is most of my time so no wonder if there is a slow down
I use WSL 2

[trim valve]

my only wsl2 experience has been super fast fs

[opaque sigil]

but in exchange it was faster when touching the windows fs

wsl2 is the other way around

[trim valve]

but also I literally don't use the windows mount at all

[rigid snow]

idk man i must be mixing up something then

or was it just the windows mount

[opaque sigil]

that would explain it if you access that yes

[trim valve]

almost certainly just the windows mount

[rigid snow]

i remember now, i tried transferring a dataset from host to wsl

[atomic sonnet]

testing on wsl no window mount rn

a bit faster than window (by 20ms over 530 ms, ~3%)

[opaque sigil]

i just tested how long it takes to install dependencies for one of my js projects in windows vs wsl, 8.5s vs 0.75s

[rigid snow]

how hard could that be on an nvme

[atomic sonnet]

oh I should test something like that for my project with verrry long first dependency run

that has specifically issue with window making it slower

[opaque sigil]

yeah ntfs is horrible at this

when lots of small files are involved in anything

[atomic sonnet]

yeah

[atomic sonnet]

that a minecraft plugin and it need to dl, decompile, remap, patch, recompile 14 minecraft server version

[trim valve]

o if this is paper they actively suggest not doing dev work on the windows fs iirc

[atomic sonnet]

yep ik

and I do it on an hdd in window so its pain

I think something like 15m with -parallel ?

[opaque sigil]

[atomic sonnet]

its ok its only once

[opaque sigil]

i wonder how their ReFS performs

have yet to try it out

[atomic sonnet]

oh, what ReFS

[opaque sigil]

their "new" filesystem that replaced NTFS for windows servers and is also used for their new "dev drive" thingy iirc

[atomic sonnet]

oh so that microsoft New Technology for File System

[maiden geyser]

https://learn.microsoft.com/en-us/windows/dev-drive/

[atomic sonnet]

ima go full linux anyway anytime soon so no need for it

[tender river]

anubis checks everyone with Mozilla user agent (so literally every browser or everyone who pretends to be a browser)

[dry charm]

@rare bridge Is the Bingo opensource? ||so I can use it for Alex's stream as well ||

[rare bridge]

yes it is open source

https://github.com/cloudburstwan/nwero-bingo

[dry charm]

oooh thank you!

for some reason it wasn't listed in your GH profile

or im just very blind

[tawny wren]

Well this is the type of request it's trying to make.

[paper nova]

not to familiar with all this and what its saying but very interesting

[tawny wren]

I honestly think it's impossible.

[paper nova]

is the one get string throwing all that ?

[tawny wren]

Seems like it.

I'm sure I can get it to redirect, but I may never know what the intention of the mezzping is.

[paper nova]

do you think its licensee or validation check?

[tawny wren]

License check? Hmm...

[paper nova]

cause if you calling to pull data and it is running the Pragma: no-cache it has to also validate each time right

[tawny wren]

You might be right.

[paper nova]

pretty sure doesn't store validations if im reading that right

the way i read that your requesting data with the get the logs are crazy high but its still grabbing the data. its just doing validation checks for each source

[tawny wren]

I'm starting to think that I need to use Wireshark to see the outgoing traffic, and a debugger on the client itself to see what it's expecting.

[stark needle]

i love trying to use a dead repo for some llm that was implemented in 2022

i basically have to rewrite everything myself again

[tawny wren]

They very thoughtfully labeled what each DLL plugin does.

[maiden geyser]

"at least, not on foot", as said

[hoary lion]

don't we love abandoned repos that have few hundred stars

[maiden geyser]

especially if the author vanished in 2021

[opaque sigil]

look at the bright side, it could be a decade old

[tender river]

using tensorflow 1.0 or something

[stable fractal]

what?

[hoary lion]

tensorflow 1 do exists

[olive sable]

Good evening

Im eating ramen

[stark needle]

omg i hate those repos

the author is NOWHERE to be seen

in this repo

theres instructions on how to run a script

but the script is not present in the repo

so uhh

???

[hoary lion]

https://tenor.com/view/basketball-juke-school-gif-21357228

get juked by .sh

[opaque sigil]

generated script maybe?

[stark needle]

@ruby timber explain

[hoary lion]

unexpected juker

[trim valve]

💔 just realised my university's WoL setup might not be working instead of my own config being the issue

[ruby timber]

what

[vestal prism]

yo sup i have question for the functional andys over here

im finally free from my javascript shackles

but now i need to learn a lisp language

what of all the 3-ish do you recomend

[noble zodiac]

racket

[maiden geyser]

my gpu starts whistling when i'm running ai

it's over

[paper nova]

lol i am running a 1080ti i cant even try

[vestal prism]

back in the day the 1080 was the best, it was so good that they made him a titan version

👴

(it was like 6 years ago lmao)

[maiden geyser]

i'm getting 80 t/s on a 3080, it should be fine on a 1080 ti

[paper nova]

ive been playing with it and runniing into driver (cuda) issues. And deff needd to learn more about torch cause thats been giving me issues

[hoary lion]

you don't actually

learning torch doesn't necessarily helps

[paper nova]

im kinda weird and will lock into learning about stuff to better understand it

[hoary lion]

blame nvidia

[maiden geyser]

how dare you blame my favourite multibillionaire corpo

[hoary lion]

am a google stan so

[tawny wren]

I'm really getting desperate now.

[maiden geyser]

but you're only 40 minutes in

[faint sandal]

what are you even doing

I scrolled up and still don't know wtf you're trying to do

[trim valve]

I was gonna ask

but assumed it was standard neurocord shenanegans

[faint sandal]

you want to check what data it's sending???

[tawny wren]

I wanna know what it expects to receive.

[livid relic]

is this related to https://github.com/lightbulbatelier/starlight ?

[faint sandal]

follow the call stack and xref in ida

I hope you're not doing this all with just disassembled code

[tawny wren]

What other option do I have?

[faint sandal]

use in conjunction with something that gives you a better disassembly navigation

preferably something that can xref and spit out the relevant pseudocode

I.e., ida, ghidra, binja, etc

[noble zodiac]

what is this even about

[maiden geyser]

binjas binted

[faint sandal]

ok I'm gonna hit the

[runic pulsar]

I did it let’s go.

[maiden geyser]

my condolences

[noble zodiac]

https://www.damiencharlotin.com/hallucinations/

[hoary lion]

use temp 0 :p

[stark needle]

man

😭

[hoary lion]

any day now google deepmind

3n audio when

[stark needle]

they said coming weeks

[hoary lion]

it had been a week for korean time

[stark needle]

man

[fast pagoda]

so i purchased a month of claude turbo subrscription

because i hate money

the $100 one not $200

anyways i was in claude code farting around having a grand old time and at one point i decided to see was a pull request review would look like from the cli since it had that in the options

and claude spent literally like 30-45 minutes desperately trying to get that environment to let it pull it

and i was kinda watching amused and putting in random attempts to help it authenticate

well i checked the model after like an hour of this

that was opus

the whole time

opus 4

(which is not charged to me because of the sub)

so i just bankrupted anthropic sorry y'all

it wont even tell me how many millions of dolllars it was >:(

[hoary lion]

nah

what actually happens is that

they like increase usage and decrease based on their current situation

so uhh you just used many usage lmao

probably wasted all your daily code

[fast pagoda]

i've just been leaving it on auto

[hoary lion]

they might have set it on batch too

anthropic absolutely SUCKS on reliable usage for subscription users

[fast pagoda]

i've actually only gotten rate limited once so far and i've been sitting here with code rolling all day each day honestly kinda trying to see what the usage limit is for this

a lot of it has just been chatting BS though

not actually coding with ... claude code

[hoary lion]

hmm

idk

but i don't like how they work in general

like instead of fixed 'n chats' per hour

they give 20 times more than free tier

like what

[fast pagoda]

when i had the $20 tier i definitely got rate limited so fast every conversation

which is why i dropped that sub

but then 4 came out and i was like eh fk it and got max

cuz ive been meaning to try claude code anyways

without bankrupting myself in api

[hoary lion]

making your money worth fr

[fast pagoda]

i dunno where they're getting that the $100 is only 5x the $20 usage

i mean it's 5x the cost

but it feels like wayyyyy better rating

and ive been leaving it on auto and checking what model it's using and it's opus way more than i expected

usually if it starts absolutely blathering and pontificating i'm like oh look it's opus

[olive sable]

Goodnight

[fast pagoda]

gn sam

https://tenor.com/view/ewok-return-of-jedi-star-wars-wicket-yubnubintensifies-gif-24401091

maybe the rate limits weren't actually THAT low it's just it was ALWAYS a "high traffic time" but now i dont feel that or something

[stark needle]

Hyperrrrrrr

I hate pytorch

no

i reject pytorch

I'm so sorry

:(

[fast pagoda]

i spent a solid 2-3 hours yesterday telling gordon the docker AI just how much i despise docker

while i as having issues with a bunch of containers

[opaque sigil]

[insert nix propaganda]

[fast pagoda]

honestly if nix makes docker unnecessary in its entirety then it's much more appealing than just about any other option

i hate docker

[opaque sigil]

for development i guess kind of but it's not like that's the only use case for docker

[fast pagoda]

im sure it's great for someone somewhere

it's never been anything but pure suffering for me

like sometimes it works but most of the time i'm saying fk this and reaching for something like https://github.com/rzane/docker2exe

[opaque sigil]

i feel cheated, this just runs the image via docker

[fast pagoda]

i just hate the docker daemon, i hate dockerfiles, i hate docker compose beyond words, docker desktop sucks ass, docker cli is barely better because it's still docker related and you have to touch that garbage with clean hands, the way volumes/containers/images are separated is barely intuitive for me (that's on me i just have always hated it and been annoyed at which is which)

but really more than anything i hate that people just yeet stuff into docker containers and decide that means that their package needs absolutely 0 documentation and it'll just work no matter what

yeah i thought it was some sweet translation thing the first time i saw it too

i just like it because i dont have touch the docker engine myself

[hoary lion]

you don't really have other better options than pytorch tho

do you want to use cupy then?

[opaque sigil]

clearly the solution is

https://github.com/NVIDIA/cudnn-frontend

[fast pagoda]

it's time to make keras the king

[opaque sigil]

reminds me actually, wonder how the tinygrad torch backend is doing hmm

[fast pagoda]

they appear to be vibing

[opaque sigil]

cause being able to use their amd stack via pytorch could be kinda nice

[fast pagoda]

anything that allows anyone to use something that isn't an nvidia product is based in my book

[opaque sigil]

[uneven pulsar]

I DID BIG BRAIN THINGY (unless it doesn't work)

[stark needle]

i may

be able to run llm at 3x speed

maybe idk

[uneven pulsar]

Silly LLMs don't use the context window as well as mature LLMs

So I'll use a mature LLM for the context window and logical crap then the silly LLM actually responds

50/50 it works and 0% chance it is a good idea

[fast pagoda]

so it's speculative decoding but shittty

sounds like a great idea to me as a certified slop producer

[stark needle]

yes

[opaque sigil]

same

[uneven pulsar]

I've only got 114 days to get this thingy production ready, and so far I haven't got enough done :3

[opaque sigil]

[stark needle]

i took it away from you

give me 64 Nvidia B200

before

[uneven pulsar]

NVIDIA = not good

[stark needle]

then use amd

[uneven pulsar]

ROCm is working good for me

[stark needle]

nah tbh the best thing after nvidia is google tpus cause they actually have software support

it's not as good as nvidia software support but ye

[uneven pulsar]

Google tpu sounds expensive

[stark needle]

it is cloud based

but for large scale ml training it can be worth it

for single person no

[uneven pulsar]

So ive gotta find a girlfriend to use them?

[velvet crest]

You have unlocked new role

[stark needle]

i mean company

where the budgets are +1M

[uneven pulsar]

Wait why doesn't google sell them?

[fast pagoda]

it's great.

if you work for deepmind i'm sure

[stark needle]

cause they host them as massive optimized clusters

[fast pagoda]

they don't need to sell them

they made them for a reason and it was to use them

[stark needle]

setting up a similar cluster would cost +100M dollars or something i assume from scratch

[fast pagoda]

they also made them to upset shadow by saying 200% efficiency all the time or whatever

[stark needle]

and then electricity, networking, etc

[stark needle]

they are efficient in terms of used chip

when shit is compiled to XLA it's very hard to get poor util

[fast pagoda]

400% fusion energy achieved

[uneven pulsar]

So instead of a fancy efficient TPU I gotta use a RX 7600 XT I can't afford

[stark needle]

[opaque sigil]

dw the tpu would be way more expensive if they sold it

[fast pagoda]

tpus dont even display images do they

why would they

[opaque sigil]

neither do datacenter gpus

[fast pagoda]

yeah

[opaque sigil]

the h100 and later are basically just tensor cores

[stark needle]

1 brain = 1 million h100

i will harvest all human brains

[uneven pulsar]

But they look cool (plus I can't afford the RX 7600 XT as it is)

[stark needle]

yea only cheap to google

[opaque sigil]

idk about cheap

but they can eat the cost upfront

[stark needle]

they dont pay nvidia 10x margin

[opaque sigil]

R&D costs though

[fast pagoda]

googol has a googol dollars anyways

[opaque sigil]

who makes the TPUs anyway, tsmc?

[stark needle]

no one knows

private info

[opaque sigil]

fair enough

[fast pagoda]

they be designed by google in california

demis hand crafts each one with love

[cloud relic]

Yeah but what are you gonna do with that much power? Run 3 chrome tabs?

[fast pagoda]

whjy would you bring such an unrealistic thing in to this highly grounded conversation

3??

next you'll tell me they can run crysis

[stark needle]

atleast its not too bad

[tender river]

you dont need it just use a tablet

[stark needle]

i swear i would be so cooked without 128gb ram

[fast pagoda]

having 64 has been actually crippling for the last few months

i really don't want to buy more before upgrading though

[stark needle]

its been looking like this for me

[fast pagoda]

it's every day bro

[opaque sigil]

i hope they have some sticks for sale for prime day or sth for cheap so i can get to 128

[fast pagoda]

i am fixated on 192 because of fking t

[tender river]

hi t

[fast pagoda]

[stark needle]

bruh

192 would be insane

but the rise in cost is also insane for me

[fast pagoda]

i'm either getting 48gb sticks or 64 come upgrade

[stark needle]

i can not afford that

[fast pagoda]

it depends on the cost difference when i do

but it will be at minimum 48 per stick

also depends on how hard i get rekt by the gpu

[paper nova]

i def need to upgrade my setup

[fast pagoda]

nvidia STILL havent send me my damn RTX insider super deluxe VIP "invitation to buy a 5090" thing

it's been like 4 months bruh

and i simply will not pay $3500

not to burn my house down

i'

will do that at msrp

ish

[opaque sigil]

i've been pretty happy with my 5070 ti for the most part

well, the constant driver crashes during profiling aside

[paper nova]

views on amd and intel now adays? im currently using a 1150 board with I7 k series

[fast pagoda]

i curse my attitude towards purchasing things every day

[fast pagoda]

ive not seriously considered an intel cpu since ryzen 5000

[paper nova]

from what i understand AMD really been killng it

[fast pagoda]

ironically enough the 48gb intel card coming out soon got me like 👁️

despite everything

[opaque sigil]

keep in mind it's 2 separate cards

[fast pagoda]

yea but 2 24gbs

[opaque sigil]

Mhm

[fast pagoda]

both of those are 2x the 10gb i have now

more even

i just pray they're available

(they wont be)

[opaque sigil]

iirc they said they should be sold on their own early 2026?

[fast pagoda]

well shit it's damn near early 2026 rn

with how time's going

[amber fractal]

Honestly those would be the cards I would have picked if I didn't go P40. Those weren't the best decisions but it was the one I had at the time.

[fast pagoda]

yeah if they're in any sort of reasonable universe for price

i really am having a hard time justifying why i'd skip them

but that's the caveat i suppose

[cloud relic]

Thatss crazy, luckily I have 32gb of ram (laptop tho) Soo I can open a few tabs, I won't forget the day I accidentally opened 400 edge tabs

[opaque sigil]

supposedly "less than a grand"

[fast pagoda]

yeah if they truly do end up less than $1000 i dunno if i'd even hesitate

guess it depends on what else is around by then

or coming

and if jensen has come by to make me pay the troll toll

[opaque sigil]

5080 super

24gb for a mere $1600

[fast pagoda]

i hate what we've become

[gaunt charm]

Hey can someone tell me whats the etiquette on discussing/speculating on Neuros inner workings?
Vedal leaked the length of the context window in the recent stream and I am just wondering if there are any threads collecting those kinda statements or if that is off limits.

[fast pagoda]

we speculate on it like schizos all the time

i'd say unless you live in his walls and start giving out actual source details it's fine

cuz otherwise we dunno

[gaunt charm]

Just wondering because he is dropping a lot of details

[cloud relic]

One day we gotta make a board

[fast pagoda]

if you search the word context you will see many posts about how she has some level of control over what's in her context

from today

[gaunt charm]

I mean he himself said his major achievement is merely getting the response super fast

I reckon nere runs async

and random chatter messages flooding her context is prolly not super healthy

but he told filian the exact number of seconds for the visual? context window

[fast pagoda]

ive been schizoing about how i think she has a bunch of different context windows that get swapped in and out depending on what she's doing to maintain some coherence without getting overwhelmed for a long long time now

[gaunt charm]

I think its just sorting stuff from various sources with normal algos and appending to the prompt

[fast pagoda]

we've known/been told/hinted at the rough length of her short term "memory" aka immediate context for a long time as well

but yeah my thing forever has been if she's activity swapping between like chatting/playing a game/talking to stream guests whatever, she's getting her context yoinked out and replaced with maybe a hint of what's going on but not front and center, to make her still know what's going on in the other contexts if needed to seem continuous but the thing she must reply to RIGHT NOW is the relevant context for that situation so she doesn't have the drift you'd expect with how many things she's juggling

and of course ranking/reranking things even within the same activity

[gaunt charm]

evil broke a lot during the MC stream

[fast pagoda]

seems she also does that herself

[gaunt charm]

directly put the reply/prompt to the mc bot into chat and probably vice versa

[fast pagoda]

yeah this issue is probably why she doesn't often show "new in progress" developments until they're ready

probably out here leakin'

if that's not nailed down

[gaunt charm]

Also had a telling moment during standup stream where neuro put what seems suspiciously like the control prompt into a story

throwback to the first days of gpt 3

[fast pagoda]

she will also just make up dumbass stuff that sounds like it might be a system prompt for fun though

[gaunt charm]

"always obey vedal" def seems something the big man put in and why she can't stop bringing him up every 2 seconds

even without context

[fast pagoda]

rules 1 and 2

[gaunt charm]

Not to imply its malicious

it clearly doesn't work even if its in there

[fast pagoda]

of course not you have to wrangle these systems or they will just go berserk

especially a le funny tuned one

[gaunt charm]

But neuro always bringing up turtles is definitively what is puzzling me the most

[fast pagoda]

it's in her memory a lot because vaduhl is a turtl

comes up easy

[gaunt charm]

is that just the vedal emotes in chat, her long term memory or actually always in the context because prompt

because it happens even when she streams with other ppl on any topic

[fast pagoda]

that long term memory has got to be one of the most cursed collections of schizo material known to humankind

[gaunt charm]

I mean, at this point we don't know how much is the memory and training data

[fast pagoda]

all i know is maintaining anything resembling consistency with how many iterations and models she's been through is very impressive

[gaunt charm]

is it though if the prompt is the same and you got the model + voice + quirks inentionally put in?

Like randomly jacking up the temperature to get her to tell stories/tangents

[fast pagoda]

to me it's quite impressive yes, without too many obvious personality changes where the data must've changed more than usual

or the model itself

i mean you train 2 different models of the same size on the same data and they'll be totally different even with everything else the same

[gaunt charm]

Nah its an inverse square kind of thing.
Where the most common phrases stay the same and its only when you drift into the ""valleys"" i.e. stuff that isn't well covered that deviations appear

But you'd need to train and compare several llms with the same prompt/temperature/architecture to really proof that.

Besides, Vedal heavily hinted at Neuro and Evil being the same and people insist they are RADICALLY different

[fast pagoda]

the memory is probably the main thing keeping it consistent
going from qwen2.5 3b to gemma 3b with the same data same training is very very different result for me
im not training the entire model though so that complicates it
i dont think neuro is full fine tuning or if she is now she hasn't always been because of resources

ive always considered them to be roughly the same or a very similar base model with some sort of lora + obviously their separate memories + different prompt

[stark needle]

Cant compare cause different model depths

[fast pagoda]

right they're different models entirely but that's what i mean

neuro has been like 5+ different models, although vedal does know what he's doing i doubt they;ve all been the same architecture given when he started

[stark needle]

just use the base model

And have 1 proper rlhf

[gaunt charm]

I mean, its just, even if it seems trivial if you ask any ai "is the sky blue, answer in yes and no only" they'll all give you "yes"
Thats what I mean with it fraying only at the fringes

And we have to consider that whatever you run as hobby probably is closer to RNG than a full sized model

like any of the 32gb llama things I've seen were really bad

and neuro has come FAR

[stark needle]

Yall are underestimating how good a 1-3b model can be

[fast pagoda]

certainly but also i'm running on the assumption that neuro was at least in some form in that state as well for a good amount of time before it was a whole operation

although there is the ;susge: "we" in all of vadhuls early neuro pitches

so who knows maybe it's just demis' pet project secret mode

that's why we can't see the architecture of gemma3

because it's all neuro

[stark needle]

bruh

[fast pagoda]

ive cracked it

[gaunt charm]

kek

[gaunt charm]

Lots of artists do the "we" to respect fans and contributors

even though it seems weird sometimes

[fast pagoda]

yeah i mean there's also his own reasoning of "wanted to sound official"

and early neuro was a lot dumber than i usually remember her being

now she's gpt4.5 with a stupid lora on and that's why vedal can't afford greggs

[gaunt charm]

I mean even aside from becoming more coherent sfx+discord bot upgrade definitively was a quantum leap

[fast pagoda]

yeah it makes sense it was around the whole "agentic" trend anyways

which is still ongoign

but she's def out here agenting

tools out the wazoo, maintain long context and coherent goals over a whole stream with literal garbage being shoveled at her all the time

[gaunt charm]

The pc/setup roast streams worked out before

but you could tell it was pretty formulaic

now she can just ignore the slideshow and stuff

and go into therapy mode XD

[fast pagoda]

she seems to also be able to pivot pretty well without requiring a whole specific setup for various situations

[gaunt charm]

sadly a higher level thinking unit won't happen because of lag

[fast pagoda]

part of why that one pirate stream went bad because she decided it wasn't pirate time anymore and it apparently had been just left up to her to maintain it, and he was hoping she would, but she didn't

while the ones before were asome sort of style adapter

hey gemini2.5pro can do this kinda speed she might be huge and fast we never know

one day

[gaunt charm]

Well yes but gemini still only runs one stage

[fast pagoda]

and also still 6s to first token

[gaunt charm]

you can either do the higher level thinking with conventional scripts that run in <1s or you bit the bullet and go full auto gpt

where you have an iteration just doing stream of consciousness that then gets processed

[fast pagoda]

neuro was actually the first matformer

[gaunt charm]

?

[fast pagoda]

like russian nesting dolls, they get trained with nested blocks of various sizes at each layer so they have a maximum possible size but they don't actually always use all their weights depending on what they're processing at inference time

like an autoMoE without a router

[gaunt charm]

ah

Did not know that

See, I need that thread/board XD

So far only really messed around with image gen

[fast pagoda]

i heard about it vaguely but folks were talking about it here like yesterday

i think shadow linked me the arxiv lol

that's how i started being interested in it too lol it's been a long journey at this point
the last 2 years feel like 10 years

[gaunt charm]

TBF that is all of vtubing at this point

[fast pagoda]

youve started caring about non-image gen just in time for diffusion llms to maybe actually be something

[gaunt charm]

Ehhh so far it feels like the industry stalled a lot

I don't think they can really resolve the issue of speciality topics

[fast pagoda]

https://x.com/InceptionAILabs/status/1894847919624462794 i think we just hit the "throw more gpus and easily gotten data at the issue" wall

Inception Labs (@InceptionAILabs)

We are excited to introduce Mercury, the first commercial-grade diffusion large language model (dLLM)! dLLMs push the frontier of intelligence and speed with parallel, coarse-to-fine text generation.

X · Feb 26, 2025, 8:32 PM

that's why gpt4.5 crashed out

\diffusion being applied to text gen freaks me out somehow lol

it's a lot harder to "identify" with the model when it's not autoregressin

even tho either way it's not thinking

it just logically feels weirder

[paper nova]

sounds like nuke it tell it works methods

[gaunt charm]

I think you have the issue that the industry really doesn't know which use case they can get good enough at to be a real thing

but yeah, speciality topics/context and higher order thinking/understanding physical processes

[paper nova]

what call the caveman method, throw and hit it until it works

[gaunt charm]

I am doubtfull you can resolve for everything in one step

[fast pagoda]

large models >1b are really just this

there was a time where we thought huge models wouldn't make sense or even work

the conventional thought was that overparameterized models were a bad thing

[paper nova]

not enough time has passed to actually see thats why

[gaunt charm]

I think people don't understand the brain works exactly like a LLM

randomly connected nerves

with random weights

its just we can't run however many billion neurons on a gpu

[paper nova]

everything takes time to develop, with out it you just dont know

[fast pagoda]

back in the day if you told someone you'd have a 1T+ language model they'd think you were a total idiot headed for the world's most overfit model

didnt help there wasnt compute to actually test it anyways

lOcAl MiNiMa

[paper nova]

i grew up wth 4 in floppy so ya sht crazy now

[opaque sigil]

you can have the smartest model, if it takes 1T+ params you're still an idiot

[gaunt charm]

ehhhhhhhhh

[opaque sigil]

speaking of which

what happened to meta's behemoth or are they just "delaying" it

[fast pagoda]

it's something like under 10% of the params in any of the large models are actually important

theyre actually just dummy thicc because that's how much of a net you have to cast

[paper nova]

should focuus more on that

[fast pagoda]

to hit the training objective

well it's a law of large numbers thing

otherwise liek you said earlier you get stuck

you have to give it soooo many params so it actually goes past that minima

[gaunt charm]

you can very succesfully cut them down after

[fast pagoda]

but not 100% successfully

[paper nova]

name one 100% human...

[gaunt charm]

the bigger issue is the half full glass of wine thing

[fast pagoda]

QAT and stuff are big

[fast pagoda]

me

they call me john human

[paper nova]

lol okay my bad..

[fast pagoda]

it's ok you didn't know

[fast pagoda]

zucc doesnt want to give up his wAIfu

[paper nova]

just sayng normal human uses how much of there brain

[gaunt charm]

I mean if you think about it, a parrot can match a 4 yr old child with 1-5% of our brain size

and thats not even taking a human brain and ripping all the unnecessary stuff out

[paper nova]

there are some pretty cool neuro science discoveries out there

[gaunt charm]

the issue is mainly that any improvent in cognition is logarithmic

but targets have an abrupt cutoff

[fast pagoda]

we really dont know what the cutoff is or where emergent capabilities appear

until it happens

[gaunt charm]

well, we know a GPU can't do it

[paper nova]

i also could fight the other side of this because errors are what define and help us grow and learn

[fast pagoda]

currently seems like the consensus is it's not worth trying above like 1-2T

[opaque sigil]

i'm just here for the hardware innovations, if LLMs is what it takes then so be it

[fast pagoda]

but that's also compute limited again

[gaunt charm]

the training process is trying to generate infinite errors

[fast pagoda]

who knows what happens if you cram 40T in tehre assuming you can source the data

they could start flying for all we know

[gaunt charm]

good one

but counterpoint

there are no emergent properties

[paper nova]

well they are training process shouldnt make same mistake twice

[gaunt charm]

we just like social animals that can vocalize and are seflish enough to recognize themselves in the mirror

[fast pagoda]

in 15 years if we're still alive we'll all be like damn those idiots didn't know how to train a model

[gaunt charm]

The basic structure of the training process is mathematically perfect

its always goals

[fast pagoda]

the data is the problem and also the objective

ye

i wish ol' euler could see us now

[paper nova]

i already feel that way

[fast pagoda]

and also see his name slapped on samplers still in use

[hard shale]

their vibe will be pretty strong

[fast pagoda]

it's pretty wild to think about the capability gap between what we have in 2025 vs even like 5years ago

[gaunt charm]

do we though?

[fast pagoda]

transformers were out by then but it was still before gpt3 said fk what you know about scaling

[gaunt charm]

Like, yeah gpt 4.5 is pretty good

[paper nova]

i had too fight to get programing classes, they now teach them as a standard

[gaunt charm]

idk it feels like its all skewed by the first "product" launch

which if you go before Iphone you are 30 years back

but after Iphone we got marginal improvements

[fast pagoda]

it's just whenever someone gets the recipe in a good enough state the general public notices and adopts stuff

[gaunt charm]

Yes

But we are like still chasing raytracing

not to mention fusion

[warped narwhal]

I still remember using talk to transformer with gpt2 to write my exam for my English class (I hated that class so I didn't care)

That was 4 years ago now

[fast pagoda]

yeah i remember laughing at subreddit simulator

[warped narwhal]

Or the florida man newspaper generator

[gaunt charm]

Eyyyyyyyyyy somoeone already finished a Filian vs Neuro captcha compilation

[fast pagoda]

lotta those were like markov chains

and i still thought that the subsimulator lil markov toys were freaky good

hell thispersondoesnotexist was around in 2019

and that was based on older GANs

[opaque sigil]

thank god GANs are a thing of the past

[fast pagoda]

o7 BigGAN pour one out

there's been a fair bit of murmur of mamba + GANs being very good at very specific domains

recently

and gans be going FAST

theyre still arguably better at faces than SOTA diffusion models

convnets or at least layers are being used in a bunch of ways to try and kill attention again

neuro is actually ELIZA from the 60s but they forgot her in a basement the last 70 years

what's old is new again

[gaunt charm]

Just waiting for when gen models bite the bullet and actually innately populate scenes with skeletons and draw the image over those instead of just randomly diffuse bits somewhere

you can do it with extensions but its not the same if it isn't done natively

[fast pagoda]

things will have gotten out of hand when they start stripping people of their skin and using that as the real skeleton starting point

how good is good enough

i certainly don't know

[gaunt charm]

I mean img 2 img already works the same way

you just need to figure out how to properly fill in

[fast pagoda]

starting from a predetermined base and refining is always easier i mean even knowledge distillation is really just that
take the more powerful larger model and noisily train the smaller learner model beyond where you'd normally realistically be able to

[gaunt charm]

I think its more like moving from pixels to gaussian blurs

because right now you want the models to very sharply paint in where the attention says the keyword goes

but ideally it has a diffuse space that restricts it

and can rotate and everything

[fast pagoda]

there have been so many reworks of exactly what the attention mechanism is targeting or even doing since attention is all you need

we already have rotary embeddings & RoPE /yarn type things that cause the attention to become positionally aware way more

[gaunt charm]

the issue is that you can't get a flat definition across the entire range of tokens

or disentangle stuff

[fast pagoda]

the fuzziness is never going away

that's where generalization pops out of really

it can be exact and deterministic or generalizable but noisy

i dont see that reconciling any time soon

[gaunt charm]

yeah but hence why I say you should populate using skeletons first. Kinda the first step towards building in physics, general enough for most objects and really good at what people care about

otherwise you end up with being in a half filled glas as an intrinsic property of wine

[fast pagoda]

nature starts way before the skelly and it's still noisy as all hell

the reason for evolution by all accounts seems to be the inherent noise present in reality motivating random change

and lots and lots of cancer

[gaunt charm]

I mean its not a reason as much as an emergent process

clones are still decently succesfull

[fast pagoda]

if you can somehow get down to like planck length and perfectly calculate, anticipate and control the laws of reality there then you're finally where you need to maybe control most randomness

and then the fuzzy logic in my rice cooker will no longer function

[gaunt charm]

I mean we can't perfectly describe chaotic systems yet

[amber fractal]

[gaunt charm]

but the butterfly effect was pretty much disproven

but to get back at what I meant initially

[fast pagoda]

how's your programmatic neurons coming along @amber fractal

[amber fractal]

Overing so I've been laying off a bit, I think I got enough ideas for the next iteration before solidifying it.

[fast pagoda]

yeah i stopped seeing pushes on the gh so i was like D:

it'

best to let things rest to come back with a new perspective though

[amber fractal]

That is my default method of iterating, which is why I bounce between a lot of things.

[fast pagoda]

you got it to a respectable pause point though

i kinda hate getting there because then i'm in the "well it's good enough atm that it's not urgent so i'll wait"

[gaunt charm]

the ""skeleton""/pose for a "horizon" could pretty much be a line.
While geometric shapes are the most defined. You can then use an llm to link higher order concepts with lower order ones.

80% of what people care about will be humans or animals and skeletons solve those entirely.
Solving how strict that original pose distributes attention around itself is a bit more tricky for less defined objects, or how much you can deform the skeleton with the original prompt.

[fast pagoda]

that's why i've adopted the "get it to a broken disaster 20% of the way and then abandon but feel massive pressure to fix it anyways"

[amber fractal]

I can probably push the current as the git version is behind, but probably in a WIP branch as it is a start of another overhaul.

[dense cosmos]

Good morning #programming

[fast pagoda]

well then you've got the question of what is a baseline skull anyways

cuz it's not like you can perfect 1 and then generalize perfectly

the opposite really

[dense cosmos]

Programmatic? Do you mean each neuron is running custom code?

[fast pagoda]

now you're just dummy overfit on that single thing

[spark basin]

Hello, everyone

[fast pagoda]

iggly can explain it way better than me and i already misspoke a bit calling them programmatic in the first place kinda

but yeah our boy is basically cooking agi

or am i

[dense cosmos]

me too :D

Or trying to

[fast pagoda]

im trying to build the perfect idiot myself

[dense cosmos]

I'm working on a SNN model

[fast pagoda]

it's quite hard

spiking?

[amber fractal]

It's hard as the definition of what it is doing because it is pretty abstract. Not using floats is really beating me over the head sometimes.*
-# *All of the timr

[dense cosmos]

Something interesting about SNN is that there's a temporal aspect to the network

[dense cosmos]

Correct!

[fast pagoda]

yeah lmao so i kinda just default to the "well they do their thing"

[dense cosmos]

Start with neural nets being universal function approximators, how does yours approximate?

[fast pagoda]

temporal info in modeling breaks my brain more than even positional embeddings

[amber fractal]

Trust me, I also defaulted to that. I'm finally starting to tunnel into a proper implementation based on resonance.

[dense cosmos]

Am currently just going to brute force with evolution

[gaunt charm]

I mean currently we are starting from random noise and guessing where any number of skulls could be.
Think wireframe rather than real skeleton (and even that is too much info)

[dense cosmos]

Trying to see if this path is worth pursuing in the first place

Have to figure out I/0

[fast pagoda]

i mean even just audio transformers being able to model the time dependencies of music as reliably as they can makes me go ???

transformers don't understand time like at all,. they aren't sequential

and it's quadratically harder the longer you're talking for them too

[dense cosmos]

Hmm... I think I also want to try hyperNeat, because I don't see how else to handle making repeated structures

The reason I chose SNNs was because of sparsity

Sparse through time, sparse through space

[fast pagoda]

wheelz would love you

[dense cosmos]

Sparsity through time means energy efficiency and better compute

[amber fractal]

The approximation would be a combination of pattern matches, either positively reinforced or negatively. That's all I got really.

[fast pagoda]

that feller LOVES SNNs

[dense cosmos]

Sparsity through space would be memory savings

[fast pagoda]

i think the ideal models are all sparse in the end

[dense cosmos]

Basically so I can run a good network on my 4gb of vram

[fast pagoda]

we're not even dense models

[dense cosmos]

True

[fast pagoda]

just a bunch of sparse representations in a trench coat

[dense cosmos]

That seems too broad, no? Maybe striking down a mathematical definition would help

My network also allows arbitrary connections

Like self connections, and backward connections

[fast pagoda]

neurons hit up that lost connections classified

[amber fractal]

Apparently something similar exists, positoron? I think was the name of it.

[dense cosmos]

wdym

[amber fractal]

I never got to reading that paper, another chatter mentioned it was similar.

[fast pagoda]

there's a very well known case of a positronic AI

[dense cosmos]

Like, I believe that the reason only forward connections are used is to make training viable

[fast pagoda]

his name was data

well he had his brother lore

[dense cosmos]

However, I am using evolution

And evolution doesn't care about backwards connections

[fast pagoda]

backprop doesnt exist

we made it up

[dense cosmos]

I don't think I can do backprop on SNNs anyways

[fast pagoda]

nah

you can't do it on linear

sequential systems

that's why we can't do it lol

you'd have to stop your own brain processing shit to update your weights

how do you do that

[dense cosmos]

Maybe BPTT using approximate gradients?

I'd rather pursue the line of local learning rules

[dense cosmos]

Sleep time?

Afaik we do actually rewire our brain when we sleep

[fast pagoda]

dead time maybe

ive always considered it more like a car wash

but yeah get bumped around ya know

i generalize shit to the extreme i really gotta stop doing it

[amber fractal]

My definition is entirely binary, so it is acting more like a 8d binary vector at the moment.

[dense cosmos]

Local learning rules should also allow for arbitrary topology

Which is a bonus

But also I think local learning is also better for neuromorphic hardware

[fast pagoda]

neuromorphic stuff is just finally coming into its own

like recently

it's c omin

[dense cosmos]

Oh?

I haven't looked too much into myself

[fast pagoda]

well, i saw a bunch of crap about it first half of this year

which means it's probably 43 years away really

but everyone was sure it was ready finally

it seems to be somewhere the labs are pointing resources at because we're probably climbing up the start of a wall in the scaling again

[amber fractal]

The last iteration of my project I accidentally found a small quirk, Some updating code only exists if it wasn't what it was trained on, so I made that into a resonance method.

[fast pagoda]

nature is beautiful

iggly's neurons done evolved already

[dense cosmos]

What

[amber fractal]

My things come in two halves, match and emit

I was using an update function and masking which parts needed updating.

[fast pagoda]

me too

[amber fractal]

The match mask is only all true if the input was different than the last time it was trained.

[fast pagoda]

every good intelligent system needs some good ol' noise tossed in

[amber fractal]

Noise my beloved

Use it often in art, not much in programming

[dense cosmos]

Alright, time to get up and seize the day

Carpe diem!

[fast pagoda]

me when i put devstral on .01 temp

https://tenor.com/view/fångarna-påfortet-peter-stormare-kapten-kaboom-tv4-fort-boyard-gif-26573227

[amber fractal]

The best part of this thing is, it doesn't rely on the output; so this version moved it to the forward pass.

[fast pagoda]

i liked cerpe it somehow gave it a fishier sound

what does that mean
why i dont know

ive unironically been up >48 hrs rn so i may actually get in bed

shortly

i am more noise than man now

[amber fractal]

Go sleep

[dense cosmos]

Sleep!

[dense cosmos]

It still means Seize the day, but in latin

[fast pagoda]

i will but only because i'm getting slightly nervous i might actually just die randomly if i attempt to stay up again

[dense cosmos]

Unless u were talking about cerpe fish

[fast pagoda]

oh no i was talking about my fish thing

lol

[dense cosmos]

XD

[fast pagoda]

sometimes i just say things and even i'm like

????

ok god dammit i'll get in bed

i swear if i somehow wake up at like 11 pm

i'm gonna be mad

my hands are all like

sore

that's odd right

i think i'm dying

feels like i have arthritis

might have something to do with typing for like

48 hours

[frigid matrix]

Warm water soak for hands

[hoary lion]

fuck useful programming

im going all in for ml

[lament igloo]

exam includes connecting sql to java
cant fucking remember Class.forName("com.mysql.cj.jdbc.Driver")

help me

[prime ridge]

@amber fractal Can I get a quick recap of what u did?

[dense cosmos]

I'm thinking of remaking my NEAT system with a hierarchical approach

The idea is to find reusable subnets that when composed can solve lots of problems (preferably efficiently)

So layer 0 is just raw neural network

Layer 1 is made of layer 0 networks

And so on

So I try and keep a database of useful submets

Anyways gtg

[prime ridge]

subnets won't be able to solve problems

they just build an internal representation

unless ur planning on building an model ensemble which would be extremely inefficient

[maiden geyser]

bro uses java in 2k25

[random igloo]

Someone put an LLM on a raspberry pi. As it generates tokens, it will run out of memory and crash and reset. He also made it aware of this limitation through system prompt
https://www.youtube.com/watch?v=7fNYj0EXxMs

[maiden geyser]

why can't it just enable page file

is it stupid

igglya is typing...

[amber fractal]

I am going to get essaying so hard

[stark needle]

godgd mrgonining

[amber fractal]

Yes I have been actually typing this entire time

[amber fractal]

[maiden geyser]

guten morgen o algo

[stark needle]

how dare you talking in another language

banned

[maiden geyser]

in another languageS