programming

[desert plaza]

bime*

you will see, you'll all see

[sage crag]

by using only machine code, you limit yourself to the lowest level of complexity

[tender river]

its like trying to understand human brains by studying neurons

just completely different levels of abstraction

[small anvil]

it’ll get better once im home, im stuck making the syntax in apple notes

[desert plaza]

well, i plan on moving up, i've used C y'know

[small anvil]

wifi spiked

[opaque sigil]

[sage crag]

congratulations

[tender river]

https://tenor.com/view/omedetou-congratulations-evangelion-shinji-chair-gif-17781895581304383421

[opaque sigil]

you should copy hblang syntax

[desert plaza]

i will write a rudimentary compiler in machine code for a simple hex-based language

[sage crag]

why

[desert plaza]

and then

[sage crag]

show me what your syntax will look like

[small anvil]

hblang 2.0

[sage crag]

how will you make an AST in machine code

[desert plaza]

use that to make a higher level language maybe

[desert plaza]

don't need AST

[tender river]

om after hblang rust declarations feel weird

[sage crag]

if your language is anything more complex than a lisp it will take you a decade

[small anvil]

compiling on hopes and dreams is so real

[desert plaza]

or i could make my own hblang compiler

[sage crag]

how so?

[tender river]

i just havent written rust in a while

and had to think a bit about how name must come after struct or enum keyword

not before

[sage crag]

i had this happen to me too

[tender river]

its like how i kept writing -> instead of => in pattern matches after using F#

[sage crag]

i wrote

SomeStruct = struct {
   x: f32,
}

[opaque sigil]

that reminds me of how i learnt rust while doing scala in uni so i kept messing up pattern matching

[small anvil]

when i tried using C++ typedef in C# forgetting thats not how C# works at all

i did want to try F#

is it good?

[sage crag]

i almost forgot how to use traits after using hblang's type system for so long

[small anvil]

or

rather

[sage crag]

[small anvil]

is it worth it

[tender river]

wdym worth it

[small anvil]

to try learning it

mb

too tired

[tender river]

learning languages that teach you something new is always useful

staying in your little bubble is not productive

[opaque sigil]

but the bubble is comfy

[tender river]

personally i only use F# for unity modding, i like it so much more than C#

[desert plaza]

but i'm writing a bnf, it won't be complicated

[small anvil]

im aware, however it’s getting to the point where it’s just "another language to learn" if you know what i mean. by all means i am not a programming god but i just know too many languages now

[sage crag]

mhmm

[sage crag]

please do

[tender river]

python, js, c#, java are all the same language

c++ and c only introduce pointers beyond that

[small anvil]

i know around 8 or 10 i think

[sage crag]

hblang is different to c because it has type objects

[small anvil]

near proficient level

[sage crag]

but hblang is zig

[desert plaza]

erm what is a type object?

[tender river]

go is literally just a mix of java (gc) and c (average age of language features)

[tender river]

my point is you have to learn languages that are actually different

[sage crag]

GenericType := fn($T: type): type {
  return struct {
    .inner: T;
  }
}

TypeInt := GenericType(int)
TypeStr := GenericType([]u8)
TypeTypeStr := GenericType(TypeStr)

[warped narwhal]

how would you even init a TypeTypeStr

[tender river]

.(.(""))

or TypeTypeStr.(TypeStr.("")) as the long form

[warped narwhal]

reminds me of that one piece of c code

[sage crag]

a := TypeTypeStr.(TypeStr.(""))
b: TypeTypeStr := .(.(""))
c := TypeTypeStr.{ inner: .{ inner: "" } }

one of various wyas

[sage crag]

this doesnt exist in hblang

[small anvil]

augh im too tired to yap

my tldr is that i feel like it gets annoying when you know too many languages, i feel like it’s nice learning cause there’s so many different benefits in a lot of languages, i just don’t want to learn another language that’s like the rest lmao

[warped narwhal]

then add it

[small anvil]

if you get what i mean

[desert plaza]

that makes sense, C isn't that hard

[small anvil]

my head hurts

[tender river]

f# is inspired by ocaml but runs on CIL

[sage crag]

f := fn(): ^^(^fn(): void) {
    @error("todo: ", f)
}

the closest you could get would be this

[small anvil]

but ya u right, was just wondering if it would be worth it to get into learning it (especially considering its microsoft made)

[small anvil]

interesting

[desert plaza]

although, syntax is confusing

[small anvil]

that’s probably why the CIL spec has so many op codes i haven’t seen in use

[warped narwhal]

clearly, everyone should be writing code in brainfuck

[tender river]

ocaml (or sml) is potentially a better choice to learn since it didnt have to do any adjustments for CIL

or just haskell

[sage crag]

ocaml is a good learning choice

[tender river]

they have different approaches to modularity

[small anvil]

i’ll definitely look into both

[warped narwhal]

-[------->+<]>-.-[->+++++<]>++.+++++++..+++.[->+++++<]>+.------------.---[->+++<]>.-[--->+<]>---.+++.------.--------.-[--->+<]>.

[small anvil]

ocaml seems interesting though

[opaque wharf]

I have solved my hunger issue

The answer is shakshuka

[tender river]

https://courses.cs.cornell.edu/cs3110/2021sp/textbook/

[sage crag]

i have just added documentation.enable = false to my nix config

[desert plaza]

come on, it's not impossible

[sage crag]

its not impossible like an anvil falling on your head from a cloud

[tender river]

no man pages

[sage crag]

i didnt like how long it was taking to generate man pages every time

i dont use local manpages anyway

[tender river]

you can just disable that documentation.man.generateCaches = false;

[sage crag]

hmmm maybe later

[tender river]

its useful
man 5 configuration.nix
man rsync
man 2 mmap
man 7 signal

[sage crag]

muahahaha

[sage crag]

maybe i am convinced

i will re-enable man

sadge

[desert plaza]

er what was this program supposed to do?

[small anvil]

this really looks cool, i’ll keep the link somewhere so i can read it when im home
thanks for the course

[warped narwhal]

it's supposed to print hello world

[tender river]

in addition to generateCaches = false; you can do documentation.doc.enable = false; and documentation.info.enable = false; for a small boost

doc is for html manuals, its actually pretty interesting because e.g. bash, git and nixos provide html manuals, but its useless if you dont use it

info is gnu docs, it would be usable if it didnt use emacs keybindings

[desert plaza]

either your program is borked, or my bf interpreter is

[warped narwhal]

it does print hello world, I think your interpreter just sucks

[desert plaza]

i see

not surprising

[fierce pasture]

lol

[desert plaza]

works on my machine

[desert plaza]

my hello word works just fine though

[warped narwhal]

no loops

-# (your loops probably don't work)

[desert plaza]

erm i contest that

[opaque wharf]

This will be malbolge situation 2 : electric boolagoo

[warped narwhal]

++++++++++[>++++++++++<-]>---.

this should print a and it is about as simple as a program can get with a loop, so see if that one works

[sage crag]

testing obs v4l2 loopback

[tender river]

also useful packages with man pages to add to environment.systemPackages

man-pages - linux man pages (man 7 signal is from there), i use it much less than i should but even then it ends up being useful for me
man-pages-posix - posix man pages, this one is less useful because linux provides the same docs but more specialized for the os

awa

[desert plaza]

my echo and your a works

[sage crag]

how the hell do i

flip this

the awa is backwards

[tender river]

it looks fine on my end

[tender river]

[rigid snow]

classic

[sage crag]

wait i remember this

yea

discord is stupid

[desert plaza]

what is happening

[warped narwhal]

try this:

+>>>->>--<<<<<[,[->+>+<<]>----------[[-]<+>>[+[-->++]--<++[-<+]-<]++[-->++]-->-[-<+]-<<]<]>>>>[.>]

it should also echo a string, but it copies the string into memory first and prints it afterwards

[rigid snow]

it flips it so yo don't freak out because you're human and see yourself in the mirror only

it's good ux

you subconsciously "look worse" when you're not flipped

[sage crag]

rrrr

discord doesnt support avif

[rigid snow]

and for other people of course you're not flipped because tht's how they see you

[opaque wharf]

What if I want to flip the person tho?

[tender river]

[rigid snow]

why

[warped narwhal]

why not?

[sage crag]

is that audible

[tender river]

i dont think so

[sage crag]

drat

[rigid snow]

programmer discovers webcam feature

no of course not you're fucking muted the webcam is video only

[sage crag]

that's stupid

[rigid snow]

??

[sage crag]

external camera may have audio

[opaque wharf]

Muted

[sage crag]

let me try this

[rigid snow]

you know you can go live with a camera device right

[sage crag]

is it audible

[rigid snow]

specifically for that purpose

[tender river]

still no

[sage crag]

exactly

hmph

[rigid snow]

capture devices and whatnot

[desert plaza]

er it's not printing, it gets stuck on read syscall

[tender river]

because audio is captured separately

on browser end

you have to select the camera as the microphone or something like that

[sage crag]

it doesnt exist

[warped narwhal]

needs to be terminated with a newline (unless I've copied the program from my phone wrong)

[rigid snow]

go live with your nintendo switch video

ps5

[opaque wharf]

It should

What is the camera?

[sage crag]

obs v4l2-loopback plugin

[rigid snow]

ohmygod i fuckingcan't

[opaque wharf]

Of course it doesn't have audio

Usually physical device will make separate device for audio and video

And that can be selected

[rigid snow]

^^^^^^^^^^

because a webcam

is

[rigid snow]

VIDEO

[tender river]

uh

scuffed screenshot

[sage crag]

well, it not being there isnt my fault

so

[tender river]

you probably also have to configure pipewire output on obs end

or actually just dont configure it on obs end

configure it separately

[sage crag]

it would be infinitely better to use obs

[opaque wharf]

audio virtual device

[tender river]

create a dummy pipewire device and pipe something in there

hmm

[rigid snow]

this is what you want

[tender river]

i dont have any audio btw

[rigid snow]

and you can actually select an audio source as well

[tender river]

no

im stupid

[desert plaza]

it seems the program only successfully echoes and exits if i only enter a newline, if i type in anything before hitting enter, it will halt forever

[tender river]

i do have audio

[opaque wharf]

Wait then how do you know konii doesn't have sound?

[tender river]

disregard that

[warped narwhal]

I copied it wrong, here is the actualy code I meant

+>>>->>--<<<<<[,[->+>+<<]>----------[[-]<+>>[+[-->++]--<++[-<+]-<]++[-->++]>-[-<+]-<<]<]>>>>[.>]

[opaque wharf]

I am not an LLM

[rigid snow]

PLEASE UNDERSTAND: ANY webcam/capture card/whatfuckinever will be TWO separate devices, one VIDEO FEED only, another AUDIO only (microphone). that's how that works. EVERYWHERE. ANY device.

[warped narwhal]

it also needs 115,000 cycles to echo "Hello, World!\n" so it may take a while

[rigid snow]

EVERY os.

[tender river]

[desert plaza]

you could have led with that

[warped narwhal]

hey now, mandlebrot.bf takes 10.5 billion cycles to complete. This is on the low end

[desert plaza]

i did not know it would take time

why

a stress test or something

or to test my loops

[warped narwhal]

test the loops

specifically the wrapping

as I used scanners for moving to the location for writing the next char

aka +[->+]- will move right until it finds a cell with the value of 255

[desert plaza]

so like, how long?

while as in while(1)?

[warped narwhal]

copy.sh takes 2ms, my old rust project took 8ms

but a manually written interpreter can take longer, especially if it hasn't went through a compiler, and doesn't implement any optimisations itself

[desert plaza]

i wrote this interpreter before the human cpu arc, i wrote it in C

i made it work, and then i forgot about it

haven't done any optimization or bug fixing or anything

[desert plaza]

wait wdym

2ms and 8ms for that program to echo?

[warped narwhal]

yes

[desert plaza]

my interpreter is still running the program

it is broken

[warped narwhal]

copy.sh compiles the brainfuck into javascript, so it is super fast. my program precomputed the loop jumps but nothing else

[desert plaza]

any tool to see what a running process is doing?

i don't think it'll complete

[opaque sigil]

CUDA driver version is insufficient for CUDA runtime version

Build cuda_12.8.r12.8/compiler.35583870_0

Driver Version: 580.65.06
are you high nvidia, if anything my driver would be too new

[glad path]

as in how much resources its using or what

[desert plaza]

like what instructions it's executing

i gotta know where my program hands up

[glad path]

not sure if thats possible

[desert plaza]

i used gdb to find out what function it is stuck in

[glad path]

actually there probably is a way to do it

[desert plaza]

no reason that would be impossible though, right?

something with ptrace or something

idk i've not looked into it

[glad path]

im sure theres a way but idk how

[desert plaza]

i will surely find out eventually

like monitoring a vm too would work

low-level advanced stuff

[glad path]

that's what i was thinking

[rigid snow]

this reminded me how cool i felt when i read and wrote memory of a gpu passthrough vm

totally transparent to userspace software

[sour lotus]

I am sooooo close, been working 4 months to get Live2D moc3 models to work as a chrome extension in chatgpt

[rigid snow]

and kernel too, but there's too much heuristics there to collect to detect a vm

i did not test kernel vm detection

mainly because that's only anticheat software and i don't wanna get accounts to ban

too much hassle

[rigid snow]

surely that's as easy as getting a live2d js library

[haughty basin]

There's an open-source example with VRM? https://github.com/camenduru/ChatVRM-LocalAI

[rigid snow]

vrm isn't live2d

[haughty basin]

VRM is even better, more expressive

[rigid snow]

vrm is literally 3d

[haughty basin]

Also nothing stops you from making a 2D model in 3d space 🙂

[rigid snow]

that's subjective, why do you think vtubers use live2d

also why is it in way higher demand and more expensive

[sour lotus]

https://tenor.com/view/the-studio-taking-notes-paying-attention-interested-gif-17315574

[haughty basin]

it's all hacks though, making 2d look like 3d is a lotta work, riggers make $3000 on a simple model

[rigid snow]

because it looks better in the end

than 3d i mean

not always, but usually

[sour lotus]

I like em 😄

Im also using Pixi (which I see VRM was also using)

[rigid snow]

isn't that a literal game engine or am i misremembering

[sour lotus]

but I love the look of the L2D moc3 models

https://tenor.com/view/elmo-sesame-street-shrug-idk-i-have-no-idea-gif-17243185

[rigid snow]

they're using pixiv's vrm driver for threejs

which is uhh not the same as pixijs

and isn't related

[sour lotus]

Man I have so much to learn

[rigid snow]

also it's more like a 2d rendering framework

i was misremembering kind of

[haughty basin]

Also have you seen this: https://github.com/Open-LLM-VTuber/Open-LLM-VTuber

Classic newbie mistake is to implement something from scratch yourself when there's literally end to end solution that you can adapt to your case
Unless this was intentional from start to learn things

[rigid snow]

his thing is different

it's like a uuh

[glad path]

me when vm fusion doesn't have gpu passthrough

[rigid snow]

visual accompany to the chatgpt interface

[rigid snow]

i used qemu

also very nice that i could recompile it to strip the stupid qemu labeling and serials that were triggering vm detections left and right

or rather replace them

[rigid snow]

all my homies hate QEMU HARDDISK

[glad path]

can the self/this argument for functions inside classes interfere with function wrappers in python

i feel like i should know this lol

[rigid snow]

is there this in python

[haughty basin]

So you can just rip out this and insert it into Chrome Extension, this is rendered on <canvas>

live2d renderer in JS

[rigid snow]

i don't think he rolled his own renderer

no way

[haughty basin]

All the wiring still a lot of things to do 🙂
but I'd argue this is sort of better to learn entire thing so you can then later use something from a shelf and be ready to jump into fixing a bug that might appear somewhere down the stack

[glad path]

you can use whatever you want for classes

[rigid snow]

you linked a react component that's 130 loc and half of them are comments...

[amber fractal]

You could use this, linters will execute you but it's possible

[glad path]

i've never actually seen anyone use this in python

ive only even seen anyone use this once

[amber fractal]

I've used myself once

[glad path]

thats just a crime

[amber fractal]

The project was jank and self was already used in scope

[glad path]

better than this i think though

this looks weird too

[haughty basin]

Yes, it's the place where we render the live2d on canvas. The way you'd rip that out is clone entire project, add your Chrome extension as a module, then import this as a component into extension.
Then if needed you can start purging all the electron 💩 but only if you really want 🙂
the goal is to bundle the extension

[rigid snow]

i

[glad path]

why

why

just why

[rigid snow]

[glad path]

that would be so confusing to read

[amber fractal]

I might switch it to this for the lols

[rigid snow]

should be I, that's less confusing

[glad path]

nononono dont 😭

[rigid snow]

that's a shit way to do anything i'm sorry

to complete a contract and collect money? ok

but still weird

[amber fractal]

I've swapped it from myself to this

[glad path]

this is honestly more acceptable just bc its still like

[rigid snow]

you’ll spend way longer surgically ripping it out than implementing it yourself, in this case in particular at least

[glad path]

yk

normal

somewhat

[rigid snow]

javython

[haughty basin]

For contract case it's even more weird if you decide not to use available solutions 🙂
are we gonna waste 4 months of dev time + all the costs making a barely functioning product?
that's too expensive for any sort of investor/company
gotta ship value fast and dirty, that's how it is in most companies today

[glad path]

that could either be
a) actual port of python
b) really terrible joke language
c) nonexistent

[amber fractal]

I didn't have any ideas when I had to do it, so having a good name is nice

[glad path]

just use self

actually scratch that

look up what the proper practice is for whatever lanugage you're using

[amber fractal]

It's already in scope in it's intended use

[haughty basin]

I don't agree but I respect your opinion 🙂

[rigid snow]

what are you talking about. it’s a library that takes 20 lines of glue to set up

[glad path]

and then dump that and use self anyways

[sour lotus]

Soooo...mine’s intentionally from scratch so I can control every layer and "own it" a bit more. Chrome extension context, Live2D rendering via PixiJS, model fitting, scaling quirks, all of it is self inflicted harm lol. This is a new world for me. I tried finding already built ones, and found a few but none were chrome extension based but more "plug into your website via FTP,etc etc. I’m basically building a “visual companion” for ChatGPT rather than a VTuber pipeline, so I’m skipping Three.js/VRM entirely....or lets see 🤷‍♂️

For me, the fun (and pain) is learning the wiring so later I can rip apart an off-the-shelf solution and fix bugs without crying. 😅 Well okay lots of crying and frustrations, damn. Also I need to set it up for my future plans with this 😄

[glad path]

hear me out just ask chatgpt to make the extension for your /j

[amber fractal]

If it wasn't exactly what the project is, then I'd agree. Cursed stuff with python's internals is not one of those cases.

[haughty basin]

20 lines for loading model, expressions, making audio work, lipsync, sending audio data in chunks, ... and so on?

[glad path]

wait...

import this should be import self

[rigid snow]

just so you know, as far as you’re concerned VRM is the same thing as Live2D, but for 3D models. and threejs is obviously for 3D. just threw me off that you said you’re “skipping” them

[haughty basin]

well, this sort of makes sense, you make a lot of mistakes and learn your own way; then screw it all and use something better but already knowing what you've went through
kinda like: "oh I KNOW the pain point, and THIS is the best solution for it!"

[sour lotus]

Dont know what to ask if I dont know 😄

lol

[sour lotus]

Oh wait im confusing myself, I need to sleep lol, appreciate the chat though 🥴

[glad path]

yall if I use a lot of threading in my projects should i switch them to use asyncio

[haughty basin]

if you're using threading it is already kinda sus 🤣

[glad path]

what's wrong with threads tho

[rigid snow]

[opaque sigil]

Threads and asyncio serve completely different purposes

Though I guess maybe not as much in python

[haughty basin]

I'd assume you are sending messages/sync between local processes to do some arbitrary work with them?

[amber fractal]

If asyncio worked on multiple cores, I'd not be able to use it on one of my projects due to race conditions.

[glad path]

sort of

[haughty basin]

why not use gRPC/http/any other standard protocol?

[glad path]

im not sending requests really

mostly im trying to run a lot of stuff concurrently

[haughty basin]

it's sort of vague, are we talking long-running processes or short running?

[glad path]

uh

[haughty basin]

like I/O bottle-necked or something else

[glad path]

hard to say

[rigid snow]

i just looked up asyncio because i don’t do python and i can’t see any mention of threads, is it just green threads/event loops

[glad path]

kinda varies

[rigid snow]

why are you concerned about threading then

[glad path]

pretty sure asyncio lets you have things

not be blocking

sort of

[rigid snow]

yeah

[glad path]

(among other things)

[rigid snow]

i get that it’s in the fucking name

async

i’m just saying it doesn’t thread

it uses an event loop system

[glad path]

python doesnt thread even

[rigid snow]

so one thread

[glad path]

python threading isnt real threading

[sage crag]

you can do real threadpools in python

[glad path]

most of the time at least

[sage crag]

you can also disable the GIL in new versions of python

[rigid snow]

good morning konii

[sage crag]

its 11pm

[rigid snow]

[glad path]

yah

[rigid snow]

soooo what’s the point of threads then

[sage crag]

this is outdated and also wrong

[glad path]

its gemni

im not surprised

[sage crag]

you can do true multithreading in python, even without spawning multiple interpreters, if you just disable the GIL

[glad path]

concurrently running stuff

[glad path]

i didnt mean that you literally couldnt do true multithreading

i just meant that like

[haughty basin]

Threading is good if you need to share context in a single process, if you want some concurent/parallel processing, etc
This is good for example for consuming lots of messages (event processing) or responding to many web requests simultaneously (e.g. nginx), where shared context is minimal.
But if you use it to start/stop sub-processes that do some stuff without real communication you don't need this level of granularity / control over resources.
A normal set of processes would do just fine. If you're on linux you can even do sockets if need to send a few bytes here and there

[glad path]

by default ig

its not real threading

[sage crag]

the point of threading in python is for IO heavy applications

if one thread blocks on IO then another can resume work

aside from that they are just concurrent

while the GIL is enabled

[glad path]

im using them mostly for concurrency

[rigid snow]

yeah but if we take that explanation at face value if you block every thread except for the one that’s actually executing there’s no concurrency

or am i stupid

[tender river]

nix has that version in the python313FreeThreading package apparently

[rigid snow]

no that’s literally what that says

[haughty basin]

threads can run in parallel but the problem is memory safety, are you sure your thread 1 would not modify shared memory at the same time when thread 2 works?

[sage crag]

not with the GIL

[rigid snow]

that’s why they use an event loop

or locks in this instance

[haughty basin]

GIL = cannot run in parallel

[rigid snow]

wait global yeah

i don’t get it

[tender river]

in cython (pre-gil removal), the threading model is kinda bad but its still useful for io-bound workloads
other implementations of python (like the experimental free-threaded cython) can have different threading models that allow proper threading

[haughty basin]

event loop doesn't save you from accessing shared memory 🙂

[rigid snow]

what

there is no shared memory

[sage crag]

in a parallel context there is always some form of shared memory unless you do message passing

its solved with atomic operations

[glad path]

wym

[sage crag]

https://doc.rust-lang.org/std/sync/atomic/

[rigid snow]

i mean they’re not shared between threads hence you can’t have traditional race conditions

it’s not

whatever

[haughty basin]

event loop on one thread, accepting many connections at once, processing things one at a time but imagine you created a global object everybody can access (aka shared memory)
e.g.

• you have a Facade for API client that is "cached" for performance purposes (big mistake)
• first request goes in (customer A), sets the API token inside that global api_client, does the request
• second request (customer B) goes in, lets see, oh we already have api_client!111 gonna use that (woohoo performance!11)
• second request from customer B uses api token from customer A, they fetch the Shopping Cart
• WTF I didn't add THESE items in my cart!?
• OH WHY THERE'S someone else's address on my profile? WTF going on

this really happened to a real project I was at back in 2019 🤣

event loop btw

very safe they say btw

it's thread safe right, one thread running all the time

I had to do all-nighter to remove all the cached api clients throughout entire frontend app...

[rigid snow]

that... would work the same way in a sync context.... i don't understand, are we clearing the token at the end of proccessing the request?

[haughty basin]

that's the hat trick, we didn't 🙂
classic mistake, don't clean up after the request is finished

[rigid snow]

so, where's the issue with concurrency

[haughty basin]

the trick is to always create separate context / shared memory for requests, then you have no problems with concurrency

but that rule was violated (global api_client) and there you go

[rigid snow]

i've never felt more stupid

i really don't understand why this would not happen in a perfectly sync environment

[haughty basin]

if there's no event loop then you have to create a new thread every time a request goes in

it's isolated

[rigid snow]

thank god i'm not actually stupid

so you were not clearing the token and the issue has absolutely NOTHING to do with any form of concurrency

it would happen regardless

great

[haughty basin]

it would not happen on django backend for example

which is not concurrent on asgi/wsgi

technically you have workers on event loop that pass web requests to thread pool

then django selects a web request and assigns a free thread from that pool

difference really is just how much context is shared

[rigid snow]

ahhhh i get it, "threads" which don't actually share anything and are subprocesses

i'm so mad i wasted mental capacity on this

[haughty basin]

Threads ≠ subprocesses.
Threads within the same process share the same address space (globals, heap, FDs).
Subprocesses don’t.
In Python, threads still share memory even with the GIL, only one thread runs Python bytecode at a time, but they interleave and C libs can release the GIL.
That’s why putting per-request data in a global (e.g., a cached client with a mutable token) is risky in both threaded and asyncio setups.
Fix is architectural, not “async vs threads”: if you keep state per request (ContextVar / threading.local() / DI), make clients immutable or pass the token as a param you're .
If you want hard isolation, use processes.

[glad path]

why does my ai have to have 5 minute latency 😭

[haughty basin]

model is warming up / you have a slow inference provider / classic first run where model is actually being downloaded / pick your poison

[rigid snow]

i really don't know how this ended up the conclusion i reached since you mentioned a cache variable

[glad path]

its running locally

[rigid snow]

man i need sleep

[glad path]

WAIT WHAT

WHAT THE

NEGATIVE RAM???

HOW DOES A PROCESS USE NEGATIVE RAM

[haughty basin]

it creates new ram for you

on disk

[glad path]

yes but

ok

i know what swap is

but how does it use negative ram

[haughty basin]

lemme think hmm
total_memory_available = os.get_ram_size()
total_memory_used = os.get_used_ram(this.process) + os.get_swap_size(this.process)
shared_memory_size = total_memory_available - total_memory_used

(I'm not a systems programmer)

[glad path]

it went straight from 3gb to -2.37gb

how do you have negative filesize

also it probably just grabs the actual filesize of the shared memory instead of calculating it

[haughty basin]

it just means that the thing you are subtracting is larger than thing you're subtracting from

[glad path]

I dont think that's how it's computed..

that was for one process specifically btw

[glad path]

you wrote the thing twice

[rigid snow]

yeah i get it but it's genuinely such a stupid architectural decision. the point of an event loop is to magically make your sync code do async without threading. if you're doing something that would fail in sync code and assumes subprocesses where there aren't, it's not just a skill issue, it's firing worthy - you cannot program, congrats, and why do we pay you? you can't even fucking write sync code
yeah i get that it's "shared memory" technically, but then you can argue any global state is "shared memory" in ANY app, sync or async, threads or subprocesses or different machines altogether etc. the problem with shared memory with threading is race conditions, where you truly concurrently access the same memory. that is not a concern if you're not accessing memory concurrently which you aren't in this case

[haughty basin]

sorry my brain CPU is melting at 27C room temp and I'm throttling 🤣

[glad path]

my ai is ghosting me br

i dont know how it can even do that

[haughty basin]

yes concurrent access to shared resource isn't constrained to threads/event loops;
in fact it's everywhere even with process isolation, if you store state in a DB and you have two clients accessing it at the same time it's gonna corrupt

[glad path]

ok

im just gonna assume that deepseek is enjoying using all my memory or something

and taking an extra long coffee break

[rigid snow]

that's what transactions are for, and that's way outside of the realm of the language/runtime

[glad path]

its taking like 5 minutes to generate

[haughty basin]

when you run model for a while, you have context overflowing into your ram/disk

[glad path]

its not been running for a while

[haughty basin]

are you running on RAM/VRAM?

[glad path]

no clue

im using the ollama python lib

[rigid snow]

and once again, you can't access memory concurrently in an event loop

[haughty basin]

what kind of model are you running, 1B, 7B, 420B?

[glad path]

r1

idk

some variation of it

i forgor

let me check

[haughty basin]

how much ram do you have

[glad path]

36gb

[haughty basin]

and vram?

[glad path]

no clue

im on a laptop

[rigid snow]

from the font rendering

from the screenshot

it's a mac

[glad path]

usually theres like a minute of latency

im assuming theres an issue with my code

[rigid snow]

and they have shared memory or whatever it's actually called, my brain is fried now

point is, however much vram it needs

[glad path]

time to see if the prompting is even finishing

[rigid snow]

it takes from the system ram

[glad path]

im just gonna restart ollama lol

[rigid snow]

can you confirm or deny you're on an apple silicon macbook

[glad path]

i am in fact on a sillicon macbook

[rigid snow]

then you should be fine if the model can fit in 36gb

[glad path]

it can

its just taking like forever to generate and idk why

[haughty basin]

you can still stomp on shared state via interleaving (cooperative concurrency) when you await

import asyncio
counter = 0

async def bump():
    global counter
    tmp = counter           # read
    await asyncio.sleep(0)  # yield → other tasks run here
    counter = tmp + 1       # write stale value

async def main():
    await asyncio.gather(*(bump() for _ in range(1000)))
    print(counter)  # often << 1000
asyncio.run(main())

[glad path]

it doesnt usually take this long

oh god what

ok thats concerning

the laptop is quite hot but the fans are on the lowest setting i think

[rigid snow]

yeah no shit you're yielding

[haughty basin]

but it's event loop

[rigid snow]

and there's no concurrent memory access here

[rigid snow]

at no point

[haughty basin]

yeah, no simultaneous access to memory but still counter is broken

that's the point

[rigid snow]

because you're storing counter in tmp before yielding...

[haughty basin]

that's exactly the misuse of shared state

[rigid snow]

then just don't yield?

or don't store it?

idk this feels like painfully obvious misuse

[haughty basin]

don't use event loop if you don't know you can shoot yourself in the foot, I'd say 🙂

or threads

use good old processes

[glad path]

ok i thing they're definitely something wrong with th ai

[rigid snow]

i do js for a living, somehow i've never ever had any issues with that

[haughty basin]

newbie mistake 🤣 wait until you do

how much are talking? 10 years, 20?

it took me ... hmm

[rigid snow]

actually working way less than that, but i absolutely did async in js in like 2019-20

[haughty basin]

it took me around 6 years to encounter the problem I outlined above with event loop + shared state

I didn't even code it myself, it was another genius frontend dev

who decided to rewrite entire app into TS

and I was the one on fire debugging his shit code

my bad I approved his MR I guess 🤣

also we were leaking memory in Node.js easily just by forgetting to clear setTimeout

[rigid snow]

again, the problem you outlined is so fucking stupid it makes no sense for this convo to even exist

why are we setting a global in the first place

what caching

why

[haughty basin]

but it did happen, and not a single dev could quickly figure it out, we were basically stuck because things started breaking randomly, because on top of this we used pm

so the problem was like, happen not on every request but every 4th or so

15 engineers

[rigid snow]

you're telling me not a single dev questioned setting a global

in a request handler

[haughty basin]

yeah it was a startup and I was the only MR reviewer

had to ship value fast

no time when you juggle backend, frontend and figure out how k8s work on GCP

[rigid snow]

that just never happens idk if you knew that

there's practically 0 reasons to do that in any scenario

[haughty basin]

it was not a global, it was a singleton pattern and I was like "hmm ok we should be safe because we create it inside a function in the js, right?"

little I knew about how event loop works back then

[rigid snow]

an exported instance then

that's a global too, but harder to spot sure

[haughty basin]

it was a shitshow for sure

but teached me two things

• don't trust singletone pattern
• don't trust event loops
• don't trust yourself 🤣

[rigid snow]

still, yielding while in the process of touching a global object should set off the loudest alarm bells in your head

[haughty basin]

I mean... I was exhausted from 12 hour workdays

[rigid snow]

or i'd go as far as to say mutating any global object in an async context should

[haughty basin]

hindsight is 20/20

I'd quadruple check myself before I even think about using low-level primitives to "launch some apps concurrently" or do a "simple communication"

I mean, we do have docker containers you can scale with pretty low overhead. Do you really need to save on RAM?

Are we talking drivers / embedded?

btw sometimes not even DB transaction can save you

-- Setup
CREATE TABLE counter(id int PRIMARY KEY, value int NOT NULL);
INSERT INTO counter VALUES (1, 100);

-- T1
BEGIN;
SELECT value FROM counter WHERE id=1;  -- 100
-- app computes new=101
UPDATE counter SET value=101 WHERE id=1; -- holds lock, not committed yet

-- T2 (in parallel)
BEGIN;
SELECT value FROM counter WHERE id=1;  -- 100 (stale snapshot)
-- app computes new=101
UPDATE counter SET value=101 WHERE id=1; -- waits, then overwrites
COMMIT;

-- T1
COMMIT;
-- Final value = 101 (should have been 102)

-> Lost update under READ COMMITTED

Why: each tx writes a constant based on a stale read; the DB serializes writes but can’t infer you meant “+1”.

Fixes (pick one):

• Do it atomically in SQL: UPDATE counter SET value = value + 1 WHERE id=1;
• Lock on read: SELECT ... FOR UPDATE → compute → UPDATE.
• Optimistic locking: add version and UPDATE ... WHERE id=? AND version=? (retry on 0 rows).
• For cross-row invariants, consider SERIALIZABLE (with retries) or proper constraints.

[midnight sigil]

vanor uses ligatures

[haughty basin]

btw here's Hello World in ES6+

(await (async () => {
  const { log: printer } = await import('node:console');

  @((cls) => cls)
  class ℍ𝕖𝕝𝕝𝕠 {
    #world = 'Hello, World';
    get [Symbol.toStringTag]() { return this.#world; }
    async *[Symbol.asyncIterator]() { yield this.#world; }
  }

  const hello = new Proxy(new ℍ𝕖𝕝𝕝𝕠(), {
    get: (t, p) => {
      const v = Reflect.get(t, p);
      if (v == null) return t[Symbol.toStringTag] ?? '🤷';
      return typeof v === 'function' ? v.bind(t) : v;
    },
  });

  const msg =
    `${(await hello?.[Symbol.asyncIterator]?.().next() ?? { value: String(hello) }).value ?? hello}`;

  function tag(strings, ...vals) {
    return strings.raw.reduce((a, s, i) => a + s + (vals[i] ?? ''), '');
  }

  globalThis['🖨️'] = printer;
  globalThis['🖨️'](tag`${1n} → ${msg}`);
})());

This actually compiles in babel 🤣

[olive sable]

hello

i have returned from going to mcdonalds with my sister and her friend for icecream

[raw trout]

I love the mcflurry

[stray dragon]

one of the #programming posts of all time

lol

[hollow spruce]

?

[hollow spruce]

Let me check ure meta data of this image or screenshot shot 🙂

[stray dragon]

it's not my image

[hollow spruce]

Ure are safe

[stray dragon]

i know

[hollow spruce]

Phone: motorola ,mm ?

[stray dragon]

i didn't take the image

smh

my phone is from samsung not motorola

[olive sable]

honeslty this is completetly worthit if they didnt break to board. just needs a new fan for 5 bucks

[stray dragon]

it does make you wonder how it happened though

[hollow spruce]

BROKEN RTX 3090 ?

[stray dragon]

yes that's what it shows in the image

[olive sable]

case too small

used an angle grinder

[stray dragon]

lol

https://imgur.com/GqowZiI

[hollow spruce]

Ask to owner
If the gpu process and vram it has

[olive sable]

i need to binge watch adventure time again sometime

they added a sequel

[amber fractal]

And there is a slight dent in the back

[stray dragon]

love that video

[amber fractal]

I seem to remember FE cards only having the gpu on the left and the right being a flowthrough cooler.

[patent walrus]

relative’s old laptop not doin too hot

this is gonna be going back to an older relative, none of the data matters
i’m tempted to slap like windows 10 ltsc enterprise iot on it hoping that itll run significantly better but idk

can only go so far with a battery that’s like 8 years old

[velvet crest]

You have unlocked new role

[stray dragon]

congrats on unlocking new role vani

[patent walrus]

thanks i worked really hard for it

[rough ocean]

Any reason they can't buy a new battery?

[hollow spruce]

Lts is the best one in windows

[patent walrus]

they can, the relative who's inheriting this laptop is cross-country and my family brought it back to wipe/do whatever with until it's handed back in however many months

[hollow spruce]

Use it / donate it / wait of orginal owner

[pulsar edge]

make sure the laptop can boot without a battery

the battery in my asus laptop did that back in february and it was really annoying

even when plugged in, it would shut off if the battery was too low, which would happen if it was too cold

the only way I could get it to boot some days was by booting into the BIOS for like 15 minutes to warm it up

[glad path]

i added basic discord functionalities to an ai im working on and tried to test it
but i forgot it has a 5-ish minute latency since i'm running everything locally on my laptop

10 minute latency at times even

this is quite painful.

uhm

seems like we're gonna get 12 minute latency

15 maybe

darn

its still goin

ok i killed it

[olive sable]

[glad path]

ok so the issue is the memory

apparently I can't feed it 30 lines

so i cleared most of it

ok so ig i gotta improve that now

[hollow spruce]

Still loading......

[nocturne olive]

I would assume so if you have any even remotely useful LLM in there

Laptops are known for having lower VRAM GPUs

[glad path]

it's quite budget since I'm literally inserting it into the prompt
once i can confirm it can send messages on discord i'm going to work on the memory and make it actually decent

[nocturne olive]

You're gonna have to do something crazy to fit it into VRAM if you don't want to simply get a smaller model

[worldly panther]

I cannot for the life of me get torch to recognize/use cuda

gonna McLose it

[stray dragon]

classic

[worldly panther]

I made the mistake of installing cuda v13 before realizing torch hasn't been updated for it yet so I had to install cuda v12.9 and change some system variables
but nothings working

[stray dragon]

you're in for a long haul

[worldly panther]

I don't doubt it :c

[stray dragon]

took me 20 hours

[worldly panther]

what was the issue for you?

[stray dragon]

it was less "the issue" and more "the issues"

i don't even remember what they were

[worldly panther]

lmAo

fair

[stray dragon]

trial and error until it worked

[worldly panther]

LMAO

god thats brilliant I didnt realize there's memes for it

[faint sandal]

anyone remember the Snapchat hotdog

[patent walrus]

[faint sandal]

I don't think there's a battery to save

[hollow spruce]

1 power point 🙂

[olive sable]

if i use the gpu in my laptop i get about that much too

but ye on "best battery life"

that battery is cooked

[patent walrus]

i dunno about you guys but this feels off

[hollow spruce]

atleast true tho

[olive sable]

with 12 hours they mean its lifespan, not the capacity

[patent walrus]

this thing is running an i3-7100U with 4 GB of ram how the fuck does it only last an hour

gotta unscrew the entire back panel to get access to the battery too

[hollow spruce]

u have been using 1 months

[tender river]

i hate this design because there are cases where you'd want to pull the battery out quickly

like if you spill something

[patent walrus]

well its like a cheap $350 (at launch) acer laptop so

[tender river]

EITHER waterproof it OR make it accessible

[olive sable]

bring back those big battery slots at the back

actually put it on the front, i need io in the back

[amber fractal]

er

[olive sable]

er

???

[amber fractal]

Give me a back slot under the cameras like how it was in the olden days

Yes I'll have less battery capacity, but I can also grab more batteries

[hollow spruce]

try to remove battery and direct connect to Power connection

i already used on my phone

[olive sable]

thats not the point

[amber fractal]

Or hell, I'll have a thicker phone to double stack batteries and add a heatsink for the extra heat.

[olive sable]

if i have to use it while plugged in i might as well jsut use my normal pc

if i had the money idd def make a custom chunky phone

[glad path]

ITS STILL RUNNING

THE AI IS STILL RUNNING

ITS BEEN LIKE AN HOUR

[olive sable]

good? bad?

[amber fractal]

We love 4gb ram

[glad path]

vedal complains about 2 second latency

i have theoretically INFINITE latency

1h and it still didnt do a thing

[patent walrus]

look at this fuckass battery

[hollow spruce]

ok 100

[amber fractal]

Tasty

[olive sable]

running an llm shouldnt take that long for even a small amount of output

are you training it mayhaps?

or is your pc just ancient?

[faint sandal]

slimjim

[olive sable]

5300mah, we have phones with more these days

[hollow spruce]

batter must be survive dont give up battery

[amber fractal]

Man that is so cursed but completely tracks

[olive sable]

nah that battery is cooked, better of at s recycling plant that being used

[glad path]

yeah

im not training it no

it usually has like 3 minutes of latency at most

i don't know exactly what's gone wrong

[amber fractal]

:bwaao7:

[olive sable]

well it must have gone wrong somehow, a full hour means its stuck somewhere

[glad path]

yeah

i've had a similar issue before but it turned out to be the ai gaslighting me since it does have the option to do nothing

it misunderstood a prompt

but this time its just not taking any actions i think

[hollow spruce]

my pc has no microphone AND still USING stream link on mobile AS microphone

testing ❌

[worldly panther]

YEAHHH BAYBEEE

[worldly panther]

only took 3 hours

[pulsar edge]

5300mah??

holy small

at 11.1V its 58Wh though

[pulsar edge]

the battery is almost certainly toast. Check the health with Nirsoft BatteryInfoView

[olive sable]

you know how electric bikes ahve batteries the size of my keyboard?

i need on of thsoe to be able to slot into my laptop

[midnight sigil]

False

[worldly panther]

say sike right now

[midnight sigil]

sike right now

[nocturne olive]

Did you install the right Torch for your CUDA?

[worldly panther]

actually the FIRST issue was that I got wrong CUDA for my torch

xD

[nocturne olive]

Silly

[worldly panther]

then I needed to reconfigure the interpreter pycharm was using

[glad path]

yall im gonna try implementing vector-based memory to my ai system wish me luck

[worldly panther]

then I realized I was using python 3.13 and torch is only compatible up to 3.12

[nocturne olive]

Have you checked if it's even running on your GPU?
How large is your LLM?

[glad path]

uh

i havent given it a lot of thought

im just kinda

[nocturne olive]

That could be why it's slow

[glad path]

im using deepseek r1 8b

i have an m3 pro

also the prompt gets exponentially longer as the memory grows

which is what i'm tryna fix

[nocturne olive]

Why are you using a DeepSeek model for a low-latency task? DeepSeek R1 models have thinking which really slows it down

[glad path]

the thinking is kinda what is useful about it though

[nocturne olive]

Because that's jow the context window works?

[glad path]

the way I'm handling memory is literally just dumping it into the prompt

its not efficient

[nocturne olive]

Not if you can't run it at high speed on a GPU or want good latency

[glad path]

so im trying to rewrite it

[nocturne olive]

What is your GPU and what quant of the model are you using?

[glad path]

like i said I have a macbook m3 pro

so the m3 is also the gpu

[nocturne olive]

And with how much memory?

[glad path]

no clue how much vram

[nocturne olive]

The VRAM and RAM are the same on Mac

[glad path]

36gb then

[nocturne olive]

Kinda tiny

[glad path]

its a laptop

also

apple charges like $400 to increase ram

[nocturne olive]

I guess the main cause for the slowness is a combination of thinking and the unified memory having horrible bandwidth

[glad path]

likely

i also haven't particularly tried to kill the latency

[nocturne olive]

I could probably run a much larger model much faster on my 3090

[glad path]

likely

[nocturne olive]

Apple is a scam

[glad path]

it is

i just hate windows so incredibly much

i literally cannot stand it

[nocturne olive]

Use Linux

[glad path]

eh

nothing wrong with linux

[nocturne olive]

You can make Linux look like whatever you want
Want MacOS-like UI? You can have it

[midnight sigil]

vector-based memory

[nocturne olive]

Linux is the superior OS

[glad path]

i dont really want to have to rewrite most of my custom applications again for linux

[nocturne olive]

What languages are they made in?

[glad path]

pretty much entirely python

[nocturne olive]

Then they should just work

Python is portable

[glad path]

its not gonna be a complete rewrite

but im gonna have to tweak some stuff so libraries use different backends

[nocturne olive]

Basically just install the libraries you need, that's it

[glad path]

also im gonna have to recompile the apps

which is annoying

[nocturne olive]

Recompile? Python is not compiled

[glad path]

im not planning to get a new laptop anytime soon

[midnight sigil]

isn't that just called embeddings

[glad path]

yes but i want applications not scripts

there are libs that basically setup portable python envs and bundle them in to apps for macos

or windows exes

[midnight sigil]

python can be compiled

cpython itself is a compiler

[glad path]

yes but it's not a compiled language generally

[midnight sigil]

python is runned on a vm

[glad path]

its interpreted

thats what most people will say

(normally at least)

anyways i do actually want to get the memory started

so cya in a bit

[midnight sigil]

it's a byte-code interpreter which can be also compiled

[glad path]

btw deepseek r1 8b is not surprisingly

8b parameters

oh god

just occured to me

adding vision is gonna be a nightmare

the latency on that is gonna also be an issue

[midnight sigil]

vision embedding model

[glad path]

im probably going to use gemni since it supposedly has image recognition

sooo

[woven temple]

You mean Gemma?
If you want to run fully locally with a decent enough vision model you're in for a bad time with a slow setup

[glad path]

yes i mean gemma

arent the names interchangable

or am i stupid

[fast pagoda]

they are not

but people will get what you mean

they are different though

[amber fractal]

Python's default impl can definitely compile. Python even has bytecode objects to manage dynamically compiled parts of the program. (Why one would even do this is beyond me, but it is there)

Normally you exec a string directly

but you can also call compile to put it into bytecode first then pass it to exec

[fast pagoda]

nuitka exists and that blasts it into C++ then into a binary

package installers do similar like xcfreeze or whatever that was called

cx freeze

pypy and numba both compile JIT on first run and then keep it like that till done which speeds em up a lot

same w/jax

[glad path]

ok im kinda almost done

shoutout to chatgpt for walking me through it and also providing a good amount of code (i dont mindlessly copy and paste it tho dont worry)

now I need to handle fragile info like discord channel ids

in metadata ig

[stark needle]

im getting my 2nd 3090 today

Finally I'm free from the cloud

https://tenor.com/view/man-free-man-chain-breaking-chain-meme-gif-15783897973479492543

[rough bloom]

Surely 48 GB is enough for everything

[nocturne olive]

Sure would be a great machine to train NeuroSynth on

[glad path]

yall ii have the vector memory set up i think

but im not sure what to use for querying it

since the ai needs a LOT of context

like i was feeding it the whole memory for a reason before

what should i do

[stark needle]

it is a lot

i just need to buy a nvlink bridge as well

holy shit why are they used 700$

frick the ai boom

[opaque sigil]

The bridge?

[opaque wharf]

I'm itching to ping Cloud lmao

[stark needle]

cloud unfortunately burst already :/

[stark needle]

ye

[opaque sigil]

????

[rough bloom]

certainly much better than 24GB, yeah

[opaque sigil]

They're like 200 on ebay here, which granted is still absurd

tf you mean 700 for an nvlink bridge

[rough bloom]

IIRC one of the sizes is cheaper but I forgor which one

[opaque wharf]

Am I mistaken nvlink for SLI?

[stark needle]

oh lmao just checked, the gpus are not compatible with each other

cause the asus card is a bit wider

who cares it's not the end of the world

i can always do ddp

[opaque sigil]

Sli is for graphics to split the workload, nvlink is just a faster, direct connection between gpus

[stark needle]

i dont wanna use sli

[opaque wharf]

Oh, so that's why you're all talking about combining 30 series card

[amber fractal]

I'm pretty sure SLI, SLdied with more recent driver versions anyeags

[opaque wharf]

Because SLI is not supported on those card IIRC

[opaque wharf]

Yeah

[opaque sigil]

It technically should be supported if you have an older driver at least

On the 3090 at least

on the 4090 it's dead dead

[rough bloom]

shouldn't they be compatible regardless? the only concern is thickness/spacing

[amber fractal]

let me just slide the nvlink to the left rq

[rough bloom]

do they not put it in the same spot

are they stupid

[opaque sigil]

That feels like something nvidia would require

[amber fractal]

I feel like they wouldn't so they can say their cards are conistant

[rough bloom]

yeah, and that's my current understanding of how it works
card length may vary but the NVLink connector should be in the same spot

[amber fractal]

Have to wait on shad for this one, I was just memeing

[rough bloom]

if not then there's always the option of using risers
as a bonus you can use the cheapest bridge you can find regardless of spacing

[amber fractal]

I swear risers 4x ed in price here

even for gen 3

[opaque sigil]

It's one of those things were signal integrity is #1 priority and being able to just freely move it around sounds like an awful idea

[opaque wharf]

Will it not kill the bandwidth?

[opaque sigil]

I don't think nvlink risers exist

[rough bloom]

not if you use an x16 riser, no

[opaque wharf]

No, you rise the gpu

[opaque sigil]

Oh

[opaque wharf]

Both of them

[opaque sigil]

Then no

[rough bloom]

not an NVLink riser, a PCIe riser to align the cards properly for the bridge to fit

[opaque sigil]

Well assuming you have a decent riser I guess

[opaque wharf]

That riser will probably cost more no?

[rough bloom]

don't 💀 me
anything to reduce cost

[amber fractal]

Poorly made marketing image, who ate the cable?

I was looking at prices again

[rough bloom]

depends? they're pretty cheap sometimes I think, and they're not uncommon so you may already have one lying around
obviously just go for whatever is cheaper

[opaque sigil]

3090 is gen 3 right

Or 4 already?

[rough bloom]

PCIe Gen 4

[amber fractal]

Okay thank you, I'm finally seeing $10USD offers again

Not $20 anymore

[rough bloom]

though you can probably get away with just running it at Gen 3

[opaque sigil]

Muh precious bandwidth

[opaque wharf]

Poor pcie5

No one is fast enough to use it yet

[opaque sigil]

Wdym

Pcie 5 is trivial to saturate for any gpu lol

[amber fractal]

Just gen 5 x8

For AI more bandwidth is always more good tho

[opaque sigil]

Actually the same can be said for gen 6 too

I guess maybe not 7

[opaque wharf]

Wut? 6 is out?

[opaque sigil]

No

[opaque wharf]

[rough bloom]

RTX 5060

[opaque sigil]

soon though

[amber fractal]

You got, uhhh 4 months left bois

Surely

[opaque sigil]

I'm really curious how they're going to cool gen 6 m.2 drives

[south needle]

water

[rough bloom]

the only interesting GPUs that do PCIe Gen 5 x8 are the Intel Arc Pro B60 Dual GPU ones
they allocate 8 lanes to each GPU and do P2P transfers over the bus instead of over something like NVLink

[opaque wharf]

I wonder if they could somehow layer peltier on the silicone

[amber fractal]

Flamethrower, that'll run cooler than the SSD

I'd say waterblock with maybe metal to metal contact if they decide they want to produce those kinds of memory (not controller) chips.

If your gonna overkill on the controller, might as well slightly cool the memory cells

[opaque wharf]

I think the seebeck effect is faster to transfer heat

Hmmm, I wonder what is the limit for infinitely thin seebeck effect layer

[amber fractal]

The larger the difference in temperature, the harder it has to work to keep a temperature. So you'd need to not allow temperature to leak while also acting as a gate.

Half of me wonders does the effect stack, but the other half is saying your better of just increasing voltage.

-# Can you add enough thermal mass buffers and isolate a stack to where room temperature exists on one side and close to sub zero exists on the other? Got no idea

[opaque sigil]

thermodynamics is just voodoo magic

[opaque wharf]

Which is what I was wondering about. How efficient it is to stack them infinitely thin

But I am too distracted by other thoughts

[amber fractal]

The peltier 10 pack is calling out to me

[midnight sigil]

I climbed up chat and see topics being changed every few messages and so

[amber fractal]

it knows my desire to do something stupid

[opaque wharf]

I can change it again to food since I am hungry

[midnight sigil]

offtopic smh

[opaque wharf]

Nono, technique to cook food is technology you know

[glad path]

can someone answer me here

[opaque wharf]

Cooking using peltier

[midnight sigil]

[opaque wharf]

You have stove on top, fridge at the bottom

[amber fractal]

Hold up, he cooking

What about running water heating in one of the sections?

[opaque wharf]

Could be done too. I was thinking more about the buffet style food container

Since they usually have some form of heater beneath

[amber fractal]

Coffee machine

But also a way to not have the stove overheat for some reason

divert to something else while keeping the cool side cold

[opaque wharf]

Yuh, basically treat it as a heat source instead of as a heat transfer/remover

[midnight sigil]

https://youtu.be/1799Rqn71A8

the real proper way to cook

[amber fractal]

Dang it, now I really want to look into this

[opaque wharf]

Bruh, why is my cursed idea always get some interest here

[amber fractal]

I know some can handle high heat, as meant for energy collection from a fireplace or smth similar. But what if we just ran a current instead

[amber fractal]

We are all very cursed

[uneven pulsar]

well can anyone see why my commits are taking that long like i didn't yet made a pull req or commit on my main branch

[midnight sigil]

do verbose on your pushes idk

[opaque wharf]

Look into RAG

[uneven pulsar]

verbose?

[opaque wharf]

@glad path

[midnight sigil]

https://git-scm.com/docs/git-push#Documentation/git-push.txt---verbose

[opaque wharf]

You did not commit binary blob right? ...RIGHT?!

Only Sam is allowed to do cursed stuff with git

[amber fractal]

I wish he just walked in and saw that

[midnight sigil]

do you perhaps have LFS stuff

[uneven pulsar]

uh...

[opaque wharf]

This comment does not bode well

[uneven pulsar]

i think i did cuz i been re installing and trying to fix my python env since i created the branch using py 3.8.9 THEN switched to latest version

[midnight sigil]

[glad path]

that'll be helpful for the part of the ai that speaks and whatnot

but idk what to do about the decision making part of it (chooses whether to talk, do nothing, etc)

not that it doesnt get a direct prompt from the user

[midnight sigil]

py 3.8.9

https://tenor.com/view/pepe-pepe-the-frog-disappointed-disappointed-pepe-tired-gif-1482602579291311529

[opaque sigil]

Committed the venv?

[uneven pulsar]

oh fuck i think so

[opaque wharf]

The AI needs to be trained for tool use too

[uneven pulsar]

wait no

[opaque sigil]

So I'm not allowed to have fonts in my dotfiles?

[uneven pulsar]

duDe SHIT YES

[glad path]

internally it uses deepseek

[amber fractal]

83 degrees C temperature difference, what in the frick am I reading?

[opaque wharf]

.ttf?

[opaque sigil]

otf

[uneven pulsar]

i think i am just going to copy past my main code and reset the api keys and create an proper env

[opaque sigil]

Make sure to add the venv folder to .gitignore

[uneven pulsar]

cuz i think i messed up the env

[opaque wharf]

Yeah, RAG and tool use. I do not know about deepseek specifically. But if you want to lower the context token used, that's one way to do it