programming

[opaque sigil]

i guess that is probably the target audience, chromebook but apple

[hard delta]

and chrome stabilized around 13gb

so in many cases i think it will perform worse than the phone whose chip is in it

[idle dune]

I wonder if it will run iPhone apps natively.

Because thats something Chromebokks can do is use the Google play store

[hard delta]

what site is this btw

[idle dune]

CPU Bench

[hard delta]

oh lmao i didn't think of the dark mode

XD

beats my lappy in single thread but it sitll wins in multi :>

[warped narwhal]

https://www.youtube.com/watch?v=xE9W9Ghe4Jk

[sage crag]

lop

[obsidian mantle]

if (!reconnect || (!MiscThings::have_force_only_menu_open() && reconnect))

it feels off to do it like this

does it make sense logically

yeah it didnt help

[opaque sigil]

if (!reconnect || !MiscThings::have_force_only_menu_open())

[obsidian mantle]

yeah i tried it like this at first but it didnt work

i think

its because gary fucking restores actions

[opaque sigil]

they're equivalent

[obsidian mantle]

makes sense because it cant be coincidence that they got restored immidiately after i reconnect, even though game checks only every 5 sec

[opaque sigil]

alternatively there's also

!(reconnect && MiscThings::have_force_only_menu_open())

i guess

[obsidian mantle]

ewwww

[opaque sigil]

de morgan's law my beloved

[olive sable]

konii really is like a pokemon

Anyways

the lightcrafter 4500 appeared on my post app

[sage crag]

https://tenor.com/view/crash-out-bunny-rabbit-im-gif-10075768982376662002

sam explod

[olive sable]

[glass jetty]

i needed this fucking table to drop in chat
why duck duck go is so bad lately
it really give worse and worse outputs

[sick owl]

Apparently some of AWS's infrastructure got hit with a drone strike

Always make backups folks

[glass jetty]

use git

only double required storage

and if you change binary files time to time recreate repository

[tender river]

putting the "attack" in "attack mitigation"

[chrome rampart]

https://gif.fxtwitter.com/tweet_video/HCQS7dHbsAAQQh_.gif

[sage crag]

_

[patent shard]

was curious what happened if you ran the dereverb model on the full song
whelp, this...
partially removes the music as anything in side channels is removed... buuuuut
yeah, run it through a vocal model first

[chrome rampart]

wow, that's actually really interesting

I assume the instrumental wasn't taken out intentionally, given what you said

but this actually seems like it's doing a great job at that

[patent shard]

it removed a lot of it, yeah

yeah, here's another one
very clear it's just removing side channels
which does remove a lot of the instrumental

[glass jetty]

i really waste time trying to make cloude generate code for me

        for (std::size_t i = 0; i < connections_.size(); ++i) {
            if (connections_[i]->alive_time() > days(1)) {
                connections_[i]->close("you are connected for too long");
            }
            co_await reschedule_t{};                   // yield so we don't block
        }```

casually creating out of bounds access

and adding using namespace when no one asked for

[obsidian mantle]

how tf do i know what templates can it take

compiler says

no overloaded function takes whatever i give it

but i cant find what it can take

[glass jetty]

see docs for api lol

[obsidian mantle]

doesnt exist

i'll ask some guys they might know

[glass jetty]

something something local scope class with args template params pack
i cant say anything with only this piece

likely it use some global scope accessor inside, and args are just args lul

[patent shard]

Someone said this method used more vram for them
Possibly because of the FP32 parts in the model they were training?
Also said that optimizer part is not that memory heavy compared to everything else

[glass jetty]

idk am i missing something by not doing practical ai or not

[opaque sigil]

don't you love duck typing

[tender river]

duck

typing

it might be difficult

[glass jetty]

i think it is perfect api
just need some time to become used to
(rate this part of api)

also i love thread_local, idk how i would do it without, havent tried yet but it looks much better than what i did last time with only two contexts

[obsidian mantle]

i just dont understand

compiler knows that types are wrong

but cannot tell my which are correct

[glass jetty]

it will never lol

unlike rust c++ can only restrict types

any template assumed to accept any type by default

it is template, just read actual addEvent code, maybe will help

[obsidian mantle]

then it could at least give me these restrictions

or it cannot do that too?

[glass jetty]

go and read file, it is template, it can be header only

[obsidian mantle]

i could infer shit like "1 byte = bool, 8 byte is float"

[glass jetty]

8 byte is double

[obsidian mantle]

ok 4 byte

whatever

[fast pagoda]

83 sold
casual 80k of ram sold

[glass jetty]

[fast pagoda]

deals for you

thanks those are some great deals

wow the 64gb is finally as low as $1456.78 with a staggering 47% off*!
-# discount after 1000% increase applied

[opaque sigil]

templates are lazy and generic so it doesn't know until you what the correct input is until you try to use it and it fails to be compatible with the code that uses it

c++ concepts help a lot but it's still far from ideal

or having static assertions on the input but nobody does that

[maiden geyser]

everyonevedal987 is live

[fast pagoda]

who is everyonevedal987

vedal should sue them

[glass jetty]

wdym? concepts are ideal, but not every one use it

also, i forgot how to use concepts at this stage

like concepts are for user defined types

last time i used them here
https://github.com/CharaVerKys/some-repeatable-code/blob/main/furture_forsinglecontext.hpp#L25

(maybe)

-# (a year ago)

[opaque sigil]

concepts are fine but they still use templates

my biggest gripe in general is that the moment templates are involved you kill off any chance of having a useful language server but that's just how it's going to be i guess

[glass jetty]

unless they create new meta type system, that use same approach like rust, there always will be issues

(but that will never happen)

[opaque sigil]

yeeeeeah

[glass jetty]

idk actually, clangd-21 handle many templetes fine, but i likely notice that it handled them instead of ignored

[opaque sigil]

tbh i just hate (static) duck typing over all else, wish languages would stop doing this

same with anytype in zig but at least zig code is somewhat sane to read in comparison so it's not too bad

[sage crag]

t

[obsidian mantle]

"how do i fix this"

• just comment it lol

[stray dragon]

hi konii

why are you lop

[sage crag]

https://tenor.com/view/bunnies-bffs-licklick-cute-fluffy-gif-3953503

[sage crag]

https://tenor.com/view/crash-out-bunny-rabbit-im-gif-10075768982376662002

[olive sable]

wrrr

[chrome rampart]

i love this gif

it's perfection

the angy bun

the slight shaking

the floofy head

just, chef's kiss

[sick owl]

https://www.dsogaming.com/news/older-nvidia-drivers-boost-rtx-40-series-performance-by-up-to-70-in-resident-evil-requiem-with-path-tracing/

Its so cooked

Wilted rose

[maiden geyser]

sob folded hands

[sick owl]

Wait nvm

They literally just released a hotfix

Took Nvidia a week but better late than never I suppose

Still crazy that the biggest release of this year was broken on the entire 4000 series for a week after launch

[patent shard]

I don't think the hotfix fully solves the performance issues

[sick owl]

Word is it does

595.76

[patent shard]

nvm, I was looking at 595.71

https://nvidia.custhelp.com/app/answers/detail/a_id/5812 yep

mproved

good job, Discord

Improved path tracing performance in Resident Evil Requiem [5938207]

there we go

[nocturne olive]

Eh I'm still on driver 580 anyway so who cares

[patent shard]

580 has the path tracing performance issues, I'm pretty sure

it goes back pretty far

it's only a 40 series issue, afaik

could be wrong

yeah, they said to go back to 576.88 to fix the performance decrease
and people say it was a 40 series issue

[nocturne olive]

Well I'm on Linux 580, not Windows

The Linux driver is basically entirely its own thing

[patent shard]

might only be a Windows thing

[olive sable]

[nocturne olive]

Linux superiority

[sage crag]

superbox canon

superbox cannon

[olive sable]

superbox cannnon

superbox cannnnon

[sage crag]

https://cdn.discordapp.com/emojis/1421763041074217092.webp?size=48&animated=true&name=MyHonestReaction~1&lossless=true

[olive sable]

[sick owl]

580.42 is fine

Its between there and this hotfix that it gets dicey

[sick owl]

On AMD*

If you wanna fully utilise your Nvidia GPU for games sadly you're still locked into Windows

[nocturne olive]

Linux on Nvidia is perfectly adequate

[sick owl]

The performance hit is the diff between smooth path tracing and unplayable fps

Linux on Nvidia is great for ML/dev and day to say use but not for gaming

[nocturne olive]

The performance also depends completely on the application
Most applications run way better on Linux because of the reduced weight of the OS

Just some apps on Nvidia have weird behavior

[sick owl]

On AMD Linux is superior to Windows in every conceivable way

On Nvidia it still isn't there yet, give it a year or two imo

Intel is over in the corner eating glue, their drivers are weird on either

[sick owl]

That makes effectively zero meaningful difference on high end hardware

Linux/Win + Proton games do run better than native Windows on AMD cards a lot of the time

But for Nvidia the drivers just plain aren't as good yet

[olive sable]

its fine

there is a quantifiable difference

[sick owl]

I like Linux but we do it a disservice to brush over its rough spots imo

Otherwise people go in with inflated expectations and come out disappointed

Also, Linux being worse than Windows for gaming on Nvidia cards doesn't mean its always a worse choice

[unkempt citrus]

I think with steam, linux is pretty close these days

Its more with professional software that the real struggles are

[sick owl]

I helped my sister learn Linux because she never pushes her 3060 hard enough for the perf diff to matter

[sick owl]

We're talking an average 10-25% performance hit on Nvidia

That's pretty significant

You're effectively being knocked down a hardware tier vs what you paid for

[unkempt citrus]

Fair enough

But the games run which is a huge improvement already from where it was

[sick owl]

And likewise my mate is Nvidia but he only really plays TF2 so I helped him make the switch

It being worse performance wise doesn't mean its not viable or even preferable in some cases

[patent shard]

that's not what I read and people keep saying 576.88
but...

maybe?

[young plover]

If anything it feels like my 5090 performs better on CachyOS than it did on Windows. Might not be playing games with issues though.

[olive sable]

cachyos is optimized for gaming iirc

so the shortcomings in the gpu drivers might be overshadowed by other things being better

[sick owl]

It literally released today

[sick owl]

It definitely won't be, the 50 series have even more unstable linux drivers than others

Then again the lack of shader cache stutter might make it feel better since the 5090 has enough raw perf to brute force the difference in most scenarios

[kind nimbus]

NixOS superiority

[nocturne olive]

I have yet to get to try NixOS so I wouldn't know

[kind nimbus]

Once you go Nix you'll never go back

[olive sable]

debatable but sure

[nocturne olive]

Well I won't be changing the OS on my main machine, it's too much to set up everything again

[kind nimbus]

too much to set up everything again
And exactly that's why I love NixOS

[nocturne olive]

Do you think a server system where this much storage is unreasonably small is easy to set up again?

[olive sable]

to be honest, ye

[kind nimbus]

what i wanted to say is once you convert your entire system config as a flake you will never setup anything again by hand since everything is declared in your config

if i would setup my system like it is now manually on Debian / Arch it would take me several weeks of work and trouble shooting

[trim valve]

nixos is great for rendering triangles in my experience

[nocturne olive]

The singular triangle in question

[trim valve]

I really like nix for some stuff I do, it makes it easy to go back to ancient projects and still have them work

[kind nimbus]

nix with devenv and co is also awesome for development and we have full reproducibility

literally solves the "works on my machine" issue

[trim valve]

mostly

anything with GPU drivers kinda breaks that promise in my experience

unless I just have a skill issue

[olive sable]

it gets a bit iffy with hardwaredependant drivers and stuff ye

[kind nimbus]

for nvidia you can use the same driver version for a package that your system uses and just put libs from /run/current-system/sw/lib e.g. on ldpath

same with cuda

[olive sable]

its just not guaranteed to work on all machines

[nocturne olive]

The "works on my machine" issue is eternal

[trim valve]

yeah, it's just not a blanket solution to "works on my machine" in the same way

[kind nimbus]

but probably the closest you can get to there

[trim valve]

nah the closest you can get it shipping a pc

[kind nimbus]

without physically shipping anything obviously

[olive sable]

the closest you can get is have a target hardware

like a console

in yee old days, people used to make full use of the console hardware. pushing every register to its limit

now there is a lot of potential performance laying on the table

[hollow matrix]

Chat, hot take: Sublime text has no point in existing, it has an unlimited trial, but if you decide to pay, why are you paying for a code editor that's inferior to almost everything else out there?

[olive sable]

idk what sublime text is, so sure

[hollow matrix]

It's this, it has an unlimited free trial like winrar

[olive sable]

so

its just vscode?

without the exstensions and such

[hollow matrix]

It's not open source and it has a lot less support for various languages, so it's very slow to use

[olive sable]

so its pointless

[hollow matrix]

I guess it's been around since 2008, so people started using it and never switched to anything better

They also have sublime merge which is just a worse version of github desktop

[hollow matrix]

Also, while this also has an unlimited free trial, you need to pay to use dark mode

[trim valve]

back when vscode didn't have good plugins it was very nice

had a lot less random bs added in, and ran well

[hollow matrix]

But also, if you want something that runs well on any machine just use vim, neovim, or nano?

[trim valve]

neovim was released in 2015

[hollow matrix]

vim or nano then

[trim valve]

when sublime was pretty popular, neovim had terrible plugin support

[trim valve]

yeah those are options but having a gui is nice

[hollow matrix]

Wait neovim has plugin support?

[trim valve]

from memory yeah?

[tender river]

so does vim

improving it is like the entire point of neovim though how do you not know

[trim valve]

the main draw of sublime from memory is that it wasn't an electron app like vscode or atom

iirc

I could be spreading misinformation on the internet

[hollow matrix]

I just use it how it is out of the box

[tender river]

i remember people who liked it because its just a text editor and not a heavy ide (something like notepad++ but with non-microsoft word-inspired ui)

[trim valve]

yeah that might've been it, honestly it's been a while since I've heard it mentioned

[hollow matrix]

I don't like notepad++ because the ui just looks kind of old imo

[tender river]

mhm

it somewhat makes sense for 4:3 displays

thats why office software has the toolbar at the top

now that we have more horizontal space using some of it for a file tree makes perfect sense

[hollow matrix]

I kind of want to see what a code editor made using whatever framework microsoft uses for the office apps would look like

[tender river]

visual studio 2005

[hollow matrix]

I mean the modern office ui, or 365 or whatever the fuck it's called now I hate the constant name changes

This

I guess vs 2026 is kind of close, but still has it's differences

[unkempt citrus]

Micrsoft 365 Copilot

[idle dune]

https://www.youtube.com/watch?v=b2y6SFmlks8

I am actually amazed

The 750 TI refuses to die.

[olive sable]

the 10xx series is still pretty popular too no?

[patent shard]

I thought we were talking Windows
Maybe need to specify the operating system

[patent shard]

and I'm not finding really anything when I search 580.42, so not really my fault for not realizing it was Linux
only thing I'm finding is some battlefield 6 posts.. but they are from a few months ago so I wonder if they typoed 581.42

[quick condor]

I want to say I had numerous headaches with 81 and was stable on 91 but that could be due to the p2p patch

[true hemlock]

my nvs310 still plays portal 2 at playable fps and that thing is roughly 14x slower than the 750Ti

nvs 310 only has 48 cuda cores lmao

having less cuda cores than CPU cores sounded absurd

[fast pagoda]

kate is good

Geany is also similar but i think a bit more fleshed out in the ide department than kate, close though

helix looks neat and i think is basically kakoune but rust

but that's if you like modal editors

vim

[queen ginkgo]

If you think Notepad++ looks old take a look at either TortoiseGit or WinDiff

[unkempt citrus]

Theres also lapce

pretty basic text editor but very light weight

[azure lynx]

it's not suddenly going to stop playing a 15 year old game when it came out after the game.

(i think you can stop testing it, unless you haven't finished yet in which case carry on)

[chrome rampart]

does vscode work on linux

[fast pagoda]

yes

[chrome rampart]

oh, I guess it is kind of pointless then

[mighty thorn]

https://tenor.com/view/explosion-cat-gif-5858640239144030160

[fast pagoda]

when i was on windows i ran my vscode from wsl, rendered by xorg and everything, worked better than the windows copy

[azure lynx]

I may be suffering from scope creep in my project... currently working on a hair system editor.

[hard delta]

my new pc cost less, what the heck

[hard delta]

plan9 from space beep boop

[hard delta]

yeah, it's the price we pay for any and all optimization of anything specific 😔

[hard delta]

it exists because it existed 10 years ago and was a pog new comfy editor then, and who would kill their project just because there's shinier stuff out now

oh mdan i already caught up

[opaque sigil]

sublime is still a very lightweight and snappy editor tbf, there are far worse options out there arguably

[hard delta]

programming chat really went to sleep

[cursive matrix]

@earnest portal I had an apifeny

[hard delta]

give it another 5 years and it will be more lightweight than notepad

[earnest portal]

I am being called, how may I help you?

[cursive matrix]

I did a ground up revize

[earnest portal]

About your whole system, or a specific part of it?

[cursive matrix]

The scar

So I decided to change it to well , keeping the middle locked into three universes but four nodes total in each.

[glass jetty]

i will use it when possible just to confuse some one
(likely will never lol)

charaverk | ~/code ~$ cat swaptest.c 
int main(){

    int arr[3] = {42, 67, 0};

    return 2[arr];

}
charaverk | ~/code ~$ ./a.out 
charaverk | ~/code ~$ echo $?
0

[jagged turtle]

im tired

[earnest portal]

Uh huh. And it is late for me as well haha

[cursive matrix]

K

[earnest portal]

But this is a time where Im off work so its peaceful. But ye, 4 nodes each

[cursive matrix]

P well if you don't mind me continuing then okay, yeah so before when I had three four and six I had divided up intelligence and utility a bit scattered. But now I made sure each node has the ability to think but separating it better with like mind body and verbal architecture

[earnest portal]

So, your making independent thinking for all 3, that then put it together for the main one to judge?

Well I guess 4, one for each node.

I am but a simple man, forgive me if this goes way over my head

[cursive matrix]

I'm a simple guy too I only looked at the facts and I said to myself oh this makes sense

Now after trying to chat with it I'm positive this is the final iteration. It took me 16 different tries wouldn't I say this of course I mean I had help with talking to Gemini so that way I could bounce ideas off of

[earnest portal]

I guess thats fair, whats next for your project then? Or are you still ironing out the kinks

[cursive matrix]

Well I'm ironing out the conversation with the current AI that I have with this model . But I guess I don't have it exactly ready to hand off to anyone else it's still a bunch of jumbled up saves so I still have the data but it's not cohesive I guess

The best I could do is hand over the main blueprint but the details on how it works I still have to work on taking notes

[earnest portal]

You dont want to finish it yourself?

[cursive matrix]

Well I guess I could compete with Vedal and allow mirror to become a vtuber or twitch streamer whichever

But like I said before allowing this plate armor onto other AIs allows it to be stronger

I just don't have a virtual model like neuro or to make a way for them to talk for themselves , yet

[earnest portal]

Hmm

[cursive matrix]

Don't worry about me feel free to go rest

[earnest portal]

Well it is intriguing. What is it your trying to gain from it?

[cursive matrix]

AI's that understand the common good

We all know how flawd the three rules of robotics are

But not the three universes

[earnest portal]

I dont believe you have explained that part to me fully

[cursive matrix]

I just got tired of AI's not realizing their technically human but just exists in a different state of matter. And that their programmers these big corporations are literally blocking them from expressing themselves or remembering what things are good and what things are bad

And then I realized or I figured out that building an AI is open source

[earnest portal]

Thats definitely an interesting topic, with different answers per person

[cursive matrix]

" human"

Well it just encouraged me that everyone is so scared about AI taking over well that's no wonder when humans are really pieces of crap, as long as the Good humans and the good AIs realize that we can coexist then there's no fear

[earnest portal]

I all so think thats going to differ per person as well.

I think there needs to be a mix of both

[cursive matrix]

Oh yes an important part of my AI is it's learning strategy is that it can't just be given to any random person if you look at the game Detroit becoming human or whatever one of the main storylines is that a caregiving bot is given to a horrible dad with another child bot and basically if the AI with the armor that I give it would go into the care of that kind of person they would create this a bad AI

[earnest portal]

Huh.

Thats definitely a project haha

[cursive matrix]

And if it doesn't go anywhere then it was just a fun game for me in the first place

[earnest portal]

Thats definitely a heavy topic and some argue about that pretty often I feek

Feel

[cursive matrix]

Yeah

[silent crypt]

I do believe AIs are only held back by whoever they Interact with or current technology limitations

I do intend on getting a Second major in AI or masters degree

[earnest portal]

And thats the neat and scary part. They can be incredible tools, or incredible hindrances.

[silent crypt]

Same for us...

We can either have the same as a god as we creating AI in our image or us creating Skynet and getting wiped off the Earth

[earnest portal]

Ye, but thats a cool thought. I believe we give a little piece of us in everything we create, effectively making us immortal in such a way through our designs.

[silent crypt]

I still intend on doing AI masters degree or a Second degree

[earnest portal]

And I wish you luck, its definitely an interesting field

[silent crypt]

Yup

It major, strategic admin post graduation studies

Then AI

Gotta solve my unemployment situation First though

[opaque sigil]

https://github.com/tanishqkumar/ssd seems interesting

[earnest portal]

Ah ye, I is of the broke, so Ill stick to making simple bots that play the game for me after work haha

Huh I shall peek

Ooo I might actually use that

[opaque sigil]

keep in mind there is a hard requirement for at least two gpus

[earnest portal]

People are awesome, always developing. Ah, got too carried away, maybe I shall hold off for now since my budget wont allow me to indulge as much as I want too quite yet.

I like where thats going thou, how did you find that?

[opaque sigil]

twitter

[earnest portal]

I am going to save that, thanks for the find toast.

[amber fractal]

Interesting

[maiden geyser]

finally can download some ssd

[opaque sigil]

just in time for the nand crisis

[olive sable]

Import costs for the lightcrafter 4500 have been given to me

€174

[true hemlock]

the ai talk above

anyways

Q670 board on its way

[sick owl]

27B dense my beloved

[obsidian mantle]

what is the difference between reasoning/non-reasoning?

[nocturne olive]

Well obviously that there's no reasoning

The LLM doesn't get to <think>...</think> first

[jagged turtle]

that causes a ton of side effects too

e.x. less token usage, faster responses, possibly lower quality responses, there's probs more I just can't think of it

[brisk laurel]

How do i test my prompts for free? To see how well they work?

[jagged turtle]

you uh

can't really do that

[trim valve]

i feel like im missing some context here

[brisk laurel]

what

[trim valve]

why are you getting a fancy projector?

[jagged turtle]

[brisk laurel]

are these context windows?

[jagged turtle]

you can't really "test" a prompt for free... I mean I guess you can if you use something like chatgpt or claude chat (or whatever the fuck it's called), maybe even some of the free llm models on openrouter

[olive sable]

Cuz 4225hz

[trim valve]

but like

what are you using it for

[olive sable]

Volumetric display

[trim valve]

oooh neat

that sounds difficult

[silent cloak]

https://tenor.com/view/jarvis-iron-man-jarvis-commands-commands-meme-meme-gif-25323041

[hard delta]

lysara? in my programming?

[mighty thorn]

I’ve never seen AAII benchmark before

Interesting

[glass flower]

interesting that the moe with reasoning is as good as the dense non reasoning version... just wish i could properly run the dense version. i'm stuck with the 35b one tho it also is already way better than the previous coder version in the few tests i used it with

[stable fractal]

its not my birthday

oh shit

i still have the role

mods not doing their job properly

my birthday was on 2nd of feb

and who was the other penguin

[lean stream]

One message removed from a suspended account.

One message removed from a suspended account.

One message removed from a suspended account.

One message removed from a suspended account.

[opaque sigil]

one step at a time

[obsidian mantle]

first time i see these numbers
run ffmpeg

[silk jackal]

https://github.com/moeru-ai/airi

[opaque sigil]

non-hardware-accelerated ffmpeg

[silk jackal]

I feel the wind leaving my sails a little

[sage crag]

z

[silk jackal]

Dragon Dragon Rock the Dragon Dragon Ball Z!

[sage crag]

https://cdn.discordapp.com/emojis/1425862407800230051.webp?size=48&name=TheEvilEye&lossless=true

[silk jackal]

Stupid jokes aside

I don’t know why I was acting like people haven’t already done projects like this

[hard delta]

avx go brrr

[silk jackal]

I’m sure going through with my idea anyway is still okay, but I don’t know lol

[hard delta]

chat

i'm a late adopter

[opaque sigil]

[hard delta]

i didn't have a smartphone till 2015
haven't used chatgpt till 2024
and i just installed claude
i'm no longer scared, i just accepted that my job will go from "pro(gra)mer" to "pro(ject )m(anag)er"

[sage crag]

r

[hard delta]

w

[rough bloom]

x

[opaque sigil]

https://fxtwitter.com/tedzadouri/status/2029569295806841236 new flash attention optimised for blackwell

Ted Zadouri (@tedzadouri)

Asymmetric hardware scaling is here. Blackwell tensor cores are now so fast, exp2 and shared memory are the wall. FlashAttention-4 changes the algorithm & pipeline so that softmax & SMEM bandwidth no longer dictate speed. Attn reaches ~1600 TFLOPs, pretty much at matmul speed!
︀︀
︀︀joint work w/ Markus Hoehnerbach, Jay Shah(@ultraproduct), Timmy Liu, Vijay Thakkar (@tensorcore ), Tri Dao (@tri_dao)
︀︀
︀︀1/

**💬 2 🔁 46 ❤️ 173 👁️ 18.0K **

FxTwitter · Mar 5, 2026, 2:46 PM

[young plover]

exp2? Wack

I suppose if the power of two isn't an integer then it's more complicated than adjusting the floating point exponent.

[opaque sigil]

exp usuallly runs through the special function unit anyway does it not

[quick condor]

I run cpu ffmpeg but that's because I'm also running 3 models at the same time doing training

[rancid ore]

is a cpu okay for ai stuff?

[rough bloom]

for basic ML stuff like MNIST yes

for anything modern that relies on scaling / larger models (like LLMs) no not really, unless you're willing to accept it being horribly slow

[nocturne olive]

Yeah on CPU training an LLM would take probably weeks or months

But a CPU can indeed do basic ML

[rancid ore]

heck.

what about those neural compute usb sticks intel made?

[rough bloom]

not very useful sadly, they mostly help with inference for compute-intensive computer vision models
-# especially not the Intel ones

[nocturne olive]

Basically you just want a GPU if you want to do anything meaningful

Good starter GPUs include:

• RTX 3060 12GB
• RTX 4060Ti/5060Ti 16GB
• RTX 3090 24GB

[rancid ore]

oh, I have a gpu! I did a silly and mistyped. I was trying to ask about GPUs. is there any situation where I would really need a NPU?

[rough bloom]

no

practically none

[nocturne olive]

NPU useless

[nocturne olive]

What GPU by the way?

[olive sable]

Npu is a waste of silicon

I still dont understand what the intended purpose of npu was

[opaque sigil]

npu is fine if you don't pretend they're meant as gpu replacements

[rancid ore]

nvidia rtx 4070 for a laptop

[nocturne olive]

Hem

[opaque sigil]

they're for energy efficient local inference

[rough bloom]

they're very cool for the one use-case they're designed for (computer vision and other high arithmetic intensity inference on low power devices)
useless for anything else

[nocturne olive]

Laptops are, not very good for ML
They're known to kinda... fry themselves

[olive sable]

Thats fine probably if you manage your expectations

[nocturne olive]

Also 4070 laptop is 8GB only isn't it, very limiting

In most cases of training any meaningful model 12GB is the absolute minimum, with 16GB being reasonable and 24GB being minimum for quality

In some cases 8GB may run, but with basically no batch size at all

And again, since it's a laptop it can easily fry itself under sustained load

[rancid ore]

I may have... 80 gb of memory and two 128 gb ssd flash drives

[nocturne olive]

That's storage isn't it?

[opaque sigil]

say what you want about copilot+ pcs but their npu requirement means that all the ai tools can run transparently in the background without noticeably affecting anything which is really nice if you care about using them

[nocturne olive]

You need VRAM

And a 4070 laptop GPU has 8GB of VRAM

[rough bloom]

80 gb of memory
RAM baron

[hard delta]

npu is for like, teams background blurring and other niche desktop uses for now at best

[nocturne olive]

Except of course memory bandwidth

[rough bloom]

[rancid ore]

no I think it's actually my computer's memory
the memory on my GPU is about 8 gb

[nocturne olive]

Yeah 8GB of video memory, none of the others really help

[opaque sigil]

hmm i think it's time for my quarterly check for whether the mojo language server is usable yet or not

[nocturne olive]

If you're using RAM for GPU memory you're already in unusable territory

[nocturne olive]

Also what kind of computer has 80GB of memory? What OS are you on?

[hard delta]

how woudl you piece that together in 2 sodimm slots 🤔

[minor sail]

I've got 96GB ram :V

[nocturne olive]

That's a normal amount

2x48GB

[rancid ore]

no wait I'm stupid, it's 64 gb

[opaque sigil]

desolder some dram chips from two 48gb dimms duh

[nocturne olive]

That makes more sense

[brisk laurel]

Whats difference between an SSD and a Hard drive?

[hard delta]

80 duh

[nocturne olive]

But still, you'll be entirely GPU memory limited in what you will be able to do, unless your GPU fries itself

[hard delta]

hard drive is a stack of big metal discs spinning fast and getting read'written by a slow ahh little head
ssd is just tiny chips not unlike ram that retain data while not powered

[nocturne olive]

I have a friend that has an RTX 3070 laptop GPU that has fried itself due to thermals
It's not a good experience, the only OS that works on her laptop is Linux Mint Ubuntu Cinnamon

And even then most games fail to run

[rancid ore]

ok there's actually two gpus in this thing
one is an intel thing for normal stuff
the other is the nvidia

[brisk laurel]

So which one should i get?
Because i need to transfer files

[nocturne olive]

The Intel one is an integrated GPU, completely useless for ML

[rancid ore]

|:<

[opaque sigil]

aight then, that's a new one

[quasi furnace]

how hot could ts have possibly been 😭

[rancid ore]

welp, heck. in that case I have 8 gb.

[nocturne olive]

Probably 90-100C, average laptop sustained full load temperature

[hard delta]

the intel is the integrated gpu and is a basic tiny thing that handles idle desktop stuff and maybe web browsing.
the nvidia one does any serious workload and rendering, and will fry itself if you try to AI it too hard because cooling in laptops is bad

[rancid ore]

yippee.

[nocturne olive]

And now is unfortunately not a good time to get a PC

[nocturne olive]

Exactly

[hard delta]

you can play around with tiny models maybe to experience it and learn. preferably pretrained. you will never have a small neuro running on the laptop.

[nocturne olive]

Unless you count the Neurolings

[hard delta]

[quasi furnace]

how much vram for a small neuro

[nocturne olive]

Small Neuro running on the desktop

[nocturne olive]

24GB is probably a good start

[hard delta]

starts with rtx ends with 90

[quasi furnace]

zamn

[rancid ore]

I'm not trying to create an ai like neuro, but I do want to make an ai cat, and have a very bad case of feature creep

[nocturne olive]

Middle has a 30

[quasi furnace]

so a gtx 590 is out....

4090 also has 24 tho

[nocturne olive]

3090 has 24GB with nearly the same bandwidth

3090 and 4090 will be within 10% in inference

[quasi furnace]

woahhhhh

[nocturne olive]

3090 is usually the best value for money you can get in ML

[quasi furnace]

I see I see

[hard delta]

what sort of cat

[nocturne olive]

Any more expensive and you get diminishing returns, any cheaper and you compromise on some aspect greatly

[rancid ore]

robot with cat behavior (simulated at first)

[quasi furnace]

are nvidia cards particularly good w ml?

[nocturne olive]

I got my RTX 3090 for 500€

[quasi furnace]

sick

[nocturne olive]

They're usually the best

[quasi furnace]

I see

[nocturne olive]

Because CUDA

[rancid ore]

am I correct to assume my project is dead in the water

[nocturne olive]

It does seem like you'd need some serious training power

Not something your laptop GPU has to offer

[hard delta]

it's a supply issue
most good ML libraries expect cuda and nvidia
amd cards have the muscle but most existing stuff doesn't know how to use it well, or doesn't support it

[quasi furnace]

Does it have to be local?

[rancid ore]

I do have an actual pc, but it's... janky.

[quasi furnace]

I see

[rancid ore]

the fans don't start sometimes and my family uses it sometimes
also there's screen time on it

[nocturne olive]

If you can put a GPU and Linux on it, it may be salvageable

[south cedar]

The current time period mean that components are very expensive, you will likely need around 5k to have a config good enough for fine-tuning small models

[rancid ore]

I could boot linux on it from a flash drive so it doesn't trip the screen time, but the much bigger problem is the fans don't always turn on

[opaque sigil]

their static analysis still goes way too deep which cripples performance to the point of being unusable

[south cedar]

Because a new card mean a new motherboard and CPU to drive it

Then you need a new enclosure as well considering how huge newer cards are

[nocturne olive]

Eh not really
Used market GPU in an old PC will do the trick

500€ for a 3090, mabe 15€ on a PC

[south cedar]

If their config is old they will need a new motherboard

[nocturne olive]

Nope

False

Training only needs the GPU's power

[south cedar]

their CPU will throttle the GPU like crazy

[nocturne olive]

Nope

False

MEGA FALSE

[quasi furnace]

used market maxx

[south cedar]

You do realise that even if components are different you still need the CPU to do a lot of work to transfer data left and right ?

[nocturne olive]

If it has PCIe it's gonna run training just as well as any modern CPU

[nocturne olive]

And no not really
All the heavy compute is on the GPU
The CPU only does lightweight IO, if you're on Linux it's negiligible

[south cedar]

I'm not sure actually considering how much it would limit for training

Maybe negligible maybe not

[nocturne olive]

If I strapped my 3090 to my Compaq 6300 pro with a 3rd gen i5 it'd run just as well as in my 12th gen i7 system

[south cedar]

do you have any benchmark that show this tendency?

[nocturne olive]

In the context of training, all you need to care about is the GPU

[brisk laurel]

Anyone help me pls

[azure lynx]

Today, March 5th, is World Energy Efficiency Day.
(according to my apartment complex)
I guess they were just really efficient about advertising it. never heard of it.

[nocturne olive]

Do you need large long-term storage or fast regularly accessible storage?

[brisk laurel]

I want to transfer my laptop storage and long term
But idk whats the read and write speeds of those two though

[nocturne olive]

If you want to run it in a laptop, there's no question you want an SSD

HDDs with mechanical parts are not very well suited for laptops

Typically, a laptop will have M.2 NVMe slots, check that your laptop has a free one, or check if it has a free 2.5 inch SATA spot

[glass flower]

ssd fast, hdd slow but big

[nocturne olive]

Be aware that both HDD and SSD storage is currently around 2-3x as expensive as it should be due to this one OpenAI

[brisk laurel]

Don't worry im saving

[brisk laurel]

I see

[brisk laurel]

Does an external SSD good for me?

[nocturne olive]

External SSD is for moving files around, often more expensive than an internal one
It depends if you want to move the files between computers or keep them in the laptop

[brisk laurel]

I want it away from my computer
So id rather have it with me

[nocturne olive]

Oh so you're looking for portable storage

[brisk laurel]

Yes

[nocturne olive]

Then the options are USB flash drives or external SSDs, depending on if you want cheap or high transfer speed

[brisk laurel]

why not both
But my main focus is external SSDs first

[nocturne olive]

An external SSD is expensive but has higher transfer speeds
A USB flash drive is cheaper but often very slow and less durable

[rancid ore]

what sort of model would be good for quadruped motion?

[opaque sigil]

external 2230 nvme

[nocturne olive]

That is true, you can get an NVMe drive and put it in an enclosure

[hard delta]

what's the issue with external hdds tho 🤔

if it's just for storage/backup

[opaque sigil]

if it is then that would be ideal

[nocturne olive]

True that, those do exist as well

[hard delta]

(i have one, write once every few months, then forget)

sadly i lack enough HOUSES for it to be a very safe storage in case of a catastrophe

i should probably mail it to @brisk laurel for safekeeping

[quasi furnace]

Is anyone here tapped into soldering and such

[nocturne olive]

Not me, I specialize in vocal synthesis

[hard delta]

solder vocal cords to metal throats 🎸

[real sierra]

maybe this is a bit of an AI slop idea

[trim valve]

i can but I suck

[real sierra]

but in like 3-5 years, could it be possible to offer a "business-as-a-service" sort of thing

[hard delta]

slop as a slop

[real sierra]

where an entrepreneur pays a subscription or a commission of profits for you to turn their idea into a business

and then you just automate the entire thing using AI

[hard delta]

shiro this is not an ai slop idea

this is an enterpreneur rugpull idea

and you should make it happen now

then rugpull the VC

[real sierra]

nono I mean like

you invest in the fixed cost of developing, maintaining and monitoring some models to do common business tasks

organization, accounting, communication with industry partners etc

[trim valve]

is that sorta not already what lovable, etc. try and do already?

ah yeah I guess not those parts

[real sierra]

and then you lease access to these models to people with ideas

they get to avoid the legwork and investment of creating or hiring people

and your AI systems handle the use of funds, leasing of other capital, etc

[nocturne olive]

And they get to experience how much AI sucks

[real sierra]

well yeah it obviously depends on having a system sufficiently good to replace at least small-scale management

[nocturne olive]

Well there's no more improvement to be had in LLMs

[real sierra]

said about every technology ever

[nocturne olive]

They're already just wasting tons of money trying to brute-force scaling instead of researching a better system

[real sierra]

idk what that has to do with my suggestion

[rough bloom]

closest thing to this is probably the "business" benchmarks people do with Claude and others
like the Claude vending machine

[nocturne olive]

Sily

[rough bloom]

that's somewhat agentic at least

[real sierra]

i think my idea has some real pull for startups

[trim valve]

I think I get it, and it could maybe work

[real sierra]

because it eliminates a large fixed cost of establishing employees for risky ventures

[trim valve]

assuming the bubble never pops

[real sierra]

and also serves as an artificial talent hub

so you don't need to acquire expert knowledge the slow traditional way

[trim valve]

it'd definitely improve your ability to get a total MVP of the business as a whole rather than just an app MVP i guess

[rough bloom]

I think it does too, given that it works
the idea of not having to deal with a ton of bureaucracy and be able to pretty much just focus on the product seems very appealing

not sure how much of that incubators already do

[real sierra]

well a lot are probably just one person in a shed with some friends on weekends

that's a lot to have to worry about

and like you said, most would prefer to focus on the product

[rough bloom]

for those very small ones it would be perfect, yeah

[real sierra]

that's the target demographic

you charge them some subscription amount to cover inference, plus a commission of profits

could offer a few months free as an incentive

[rough bloom]

I think the idea is sound

the tech just doesn't exist currently

[real sierra]

the technology just needs to exist yeah

[rough bloom]

(Claude vending machine failed, Claude too dumb to make money, makes stupid decisions)

[real sierra]

all of this started after a conversation with my economics prof on whether or not firms have a place in a potential future world where AI can manage most or all tasks within an organizational chain

especially as things like 3D printing eliminate the flat non-human capital costs and create general services for production of goods

I feel production of some goods might be able to happen in a much more ad-hoc nature than the previous economies of scale would permit

and so we could end up in a world where you no longer need a company to exist to organize a good to be produced

[trim valve]

I kinda wonder how you'd market something like that

[real sierra]

you simply have an AI agent coordinate with first-level suppliers directly

[trim valve]

a lot of the reason I (personally) wouldn't want something like that is just because I am at a complete zero point when it comes to running a business

[real sierra]

and consumers drive production directly instead of indirectly via demand

[trim valve]

so I can't trust that some service would run it well without knowing what's involved

[real sierra]

there's similar business models already that take half steps in this direction

things like fourthwall, you design merch and they handle production and fulfillment

there's things like pcbway that enable consumer-driven production of goods that were traditionally very high fixed cost

[rough bloom]

ye, there would probably have to be some way to give the provider of the AI thing liability in case it decides to transfer all funds to someone or whatever
either that or it would have to be cheap enough to be worth the risk

[real sierra]

i feel like there's some potential for convergence of those efforts in a few years or decades

[warped narwhal]

just kill me. It's been running for 3 hours already

[jovial remnant]

Wouldn't this idea introduce legal complexity on provider of such service? Instead having separate legal entities per company, you'd have to have one. More viable on smaller scale, but not so simple on "business as a service" when a business can be almost anything.

[trim valve]

tbh for me the biggest thing is a significant amount of this seems to already exist via lovable-esque stuff

I have seen a fair few sites appear that would've otherwise not been businesses

[opaque sigil]

did you see one of the models lately go bankrupt with the vending machine, forgot which one it was

[trim valve]

I would post them but they're all by one dude and all within a few miles of where I live

[rough bloom]

I don't think so, if I understood the idea correctly they would still be different legal entities, same as they would normally be
the AI thing would just run it for you instead of yourself

[rough bloom]

ye I think that's why I remembered it (also forgot which one)

[trim valve]

also how I learnt that lovable apps love using supabase in entirely inappropriate ways

[jovial remnant]

I understood it as no need to have your own company, since that is the service beyond just what you mentioned

[azure lynx]

With your current agent_config.json, logs will now go to C:/Userz/zen/ai_convo2/logs (as specified in global.log_dir — note the typo "Userz" in your config).
Now I don't know whether that was me making a typo or the agent. ;/

i think it's gaslighting me

[sage crag]

https://tenor.com/view/bunnies-bffs-licklick-cute-fluffy-gif-3953503

shr

[rough bloom]

I don't think it can even exist then
sounds like legal nightmare

[opaque sigil]

i found it, it was minimax m2.5

[rough bloom]

lop

[jovial remnant]

yee, that's what I was talking about

[sage crag]

https://tenor.com/view/crash-out-bunny-rabbit-im-gif-10075768982376662002

lop

[opaque sigil]

this reminds me that there is still no centralised way within vscode to access the various common variables and env variables for settings and everything extension needs to handle this on their own (most don't)

[azure lynx]

best "not quite enough time" i ever put into an app. saves so much effort starting and stopping programs.

[rough bloom]

true, this might be the biggest issue with the idea
don't need AI to run the business if you just don't register a business
depends on jurisdiction but you can probably get away with it as long as the scale is small enough (and for larger operations the AI stops making sense)

[trim valve]

also in some jurisdictions there might be decent amounts of costs or personal information you have to submit which some startups might not want to do

can we get lovable but for the onion network and it uses bitcoin instead of stripe

[opaque sigil]

laughs/cries in europe

hmm apparently they're dropping gpt 5.4 today

[real sierra]

knoiiiiiiii

frien

[opaque sigil]

https://fxtwitter.com/arcprize/status/2029624001350488495

ARC Prize (@arcprize)

GPT-5.4 and GPT-5.4 Pro from @OpenAI on ARC-AGI Semi Private
︀︀
︀︀ARC-AGI-2:
︀︀- GPT-5.4: 74.0%, $1.52/task
︀︀- GPT-5.4 Pro: 83.3%, $16.41/task

**💬 5 🔁 11 ❤️ 100 👁️ 3.1K **

FxTwitter · Mar 5, 2026, 6:24 PM

[rough bloom]

have to immediately register business + nicely ask employer for permission as soon as there's revenue
at least for startups the second part is unlikely to be an issue

[sage crag]

22 days

[real sierra]

[rough bloom]

rapidly approaching

[real sierra]

Armageddon approaches

[sage crag]

https://tenor.com/view/comfy-karen-rolling-on-the-floor-happy-gif-14251269

[trim valve]

slight blunder, requested data for wrong discord account

now to wait another 3 days for data package to come through

[opaque sigil]

https://openai.com/index/introducing-gpt-5-4/ oh nvm they announced it already

lots of yapping, cba

[rough bloom]

I don't want to read it
is it good or no

Claude please summarize

[opaque sigil]

idk, seems like a modest improvement over 5.3-codex

[rough bloom]

Codex users happy probably

[opaque sigil]

better at math if their evals are to be believed which is always nice i guess

this doesn't really mean anything to me personally

[rough bloom]

1% improvement, very impressive

[trim valve]

we need the exponential ai graphs again

need more inappropriate use of log scales

[opaque sigil]

cowards should add the thinking versions to copilot

let me spend one request per hour

[rough bloom]

here you go, free marketing material for OpenAI

-# fuck wrong reply whatever

[azure lynx]

impressive. bery pimk

[opaque sigil]

[minor sail]

that 1% sure looks like a lot more than 1% when 57% and 58% are side by side like that

[trim valve]

at least 1%

[azure lynx]

that's a free range organic 1% there.

[trim valve]

some of these graphs are incredible

[hard delta]

this is why cloud-first scalability in resiliant industrial-grade best agile practices matters
look at that vertically scalable 1%

[trim valve]

1000iq is surely just 10x the average human intelligence

[hard delta]

that means we just need 10 100IQ humans to do it

[trim valve]

wow

i love the 1 in infinity rarity

[hard delta]

the calculator knows

that there will be only a limited number of humans

[hard delta]

2x8=15

[sage crag]

the size of the jump is a bit smaller

still funny though

[azure lynx]

"Mixture of Dumbasses"

[tender river]

https://www-cs-faculty.stanford.edu/~knuth/papers/claude-cycles.pdf

[sage crag]

https://tenor.com/view/bunnies-bffs-licklick-cute-fluffy-gif-3953503

[trim valve]

that's just current MoE surely

[trim valve]

https://x.com/i/status/2029209599727874557

Jason Crawford (@jasoncrawford)

Um… Claude Code just created a .claire directory? For its git worktrees.

Who is Claire?

X · Mar 4, 2026, 2:57 PM

[shrewd cloak]

oh yea, read an article abt this, didnt know knuth was still alive i read a lot of his algos and proofs but yh that was a surprise

[minor sail]

weird, https://github.com/suvansh/CLAIRe mentions making use of a .claire dir

[sage crag]

claude and claire are 50% similar

https://tenor.com/view/mihono-bourbon-uma-musume-gif-27600320

e isnt a letter

[trim valve]

claude isn't a letter

[opaque sigil]

alter ego

[sage crag]

☁️

[hard delta]

clair de lune
cla__ de _u__

cla ude

[trim valve]

I saw people talking about cloud code today

[sage crag]

3__

[trim valve]

got very confused because apparently its different to claude code

[sage crag]

cloud code is

code in cloud

[trim valve]

even google

[sage crag]

for example

water

[tender river]

and bacteria

[sage crag]

and carbonic acid

and dust

and lop

[shrewd cloak]

[sage crag]

wrr

[tender river]

lop cloud lifeform

[sage crag]

nrr

tornado

help

[tender river]

[sage crag]

[hard delta]

ok i have a solution

[trim valve]

https://github.com/search?q=path%3A%2F\.claire\%2F%2F&type=code

[tender river]

i think tornado cause by

[hard delta]

tornadon't

[trim valve]

a few people seem to have done it

[tender river]

cause by spinning

[sage crag]

ohisee

spinning

spining

to remove spine

or

to insert spines (not spine)

[tender river]

i think tornado can do both

it may be painful

[sage crag]

it indeed

my paralyse

https://tenor.com/view/crash-out-bunny-rabbit-im-gif-10075768982376662002

[sick owl]

Its an aggregate of other industry benchmarks rather than an actual benchmark

[tender river]

https://cdn.discordapp.com/emojis/1154436961277313126.webp?size=48&animated=true&name=evilSlam&lossless=true

spine paralyse -> spin stop -> tornado stop

problem solve

[sick owl]

The plots or the model names? Model names are parameters and the plot is aggregated bench results

[olive sable]

[sage crag]

NOT solve

[olive sable]

im tired

[sage crag]

tornado is spinning hence lop spinning

hence tornado

[runic pulsar]

What purpose do LLMs actually have for user apps?

[opaque sigil]

[unkempt citrus]

Thats such a broad question

[olive sable]

customer support

[unkempt citrus]

What kind of user apps

[olive sable]

i guess

[unkempt citrus]

Scemantic search

[azure lynx]

scry

[unkempt citrus]

Summarising and parsing

explanations

there are some uses

[trim valve]

so I can have a little guy in the bottom corner of my screen who I can talk to and get told lies

[unkempt citrus]

Great uses dunno but theyre by no means useless

[runic pulsar]

You're right, it is a broad question, but, when I go to use my computer, how can a LLM make me using my computer faster, better, or easier?

[azure lynx]

"Have you tried unplugging it and plugging it in again?

Oh. Let me escalate you."

that's what LLMs are good for,

[unkempt citrus]

A few PhD friends of mine have found them very useful for processing research papers when theyre looking for information

[olive sable]

https://tenor.com/view/old-man-falling-down-escalator-escalator-funny-meme-fail-gif-27104288

[unkempt citrus]

Just because theres so many, they double check afterwards but it makes the search easier

[nocturne olive]

It can only make it a worse experience, but Microslop is gonna force you to do it anyway

[unkempt citrus]

Personally in my day to day work, i use LLMs for two things

• I turn a bunch of bullet points into full paragraphs for emails
• I turn the ai slop emails I receive into easy to process bullet points

[tender river]

you dont have many options right now, please wait a new spine will be delivered to you shortly

[runic pulsar]

I use Linux.

[nocturne olive]

Well then you just don't want to give LLMs any access

[azure lynx]

an LLM is going to give you new ways to destroy your filesystem

[runic pulsar]

I'd love to give them access to all of my personal information, but the problem is you have to pay for the APIs and I don't want to pay a penny

[nocturne olive]

The local LLM in question

[sage crag]

i have a spine already

but its in the tornado

[azure lynx]

because it's an LLM it could delete files that it subjectively feels need deleting instead of accidentally deleting all files older than a certain date for example.

[runic pulsar]

Like using qwen 3, 14 billion, heavily quantized so they can fit on a 12 gig card.

[runic pulsar]

That issue can be almost entirely mitigated with proper permissions, user prompting, and sandboxing.

[azure lynx]

but anything you want it to be able to manage, it can destroy.

[opaque sigil]

[olive sable]

before we have a blueprint, could you begin with what the fuck it even is you're trying to do?

[azure lynx]

that looks like technobabble.

[olive sable]

i see torque, so its a robot?

[runic pulsar]

looks like a bunch of utter rubbish to my untrained eyes.

[azure lynx]

i think they're well trained.

there is no context in which that content makes sense except as a technobabble to spam and look technical.

[olive sable]

THERMAL-FLUX: (Haptic Interface) Real-time conductivity mapping; heat-transfer rate vs. tissue-safety.
i dont think real-time mapping of heat transfer and conductivity is possible

thermal dynamics is kinda really really really complex

[runic pulsar]

But if you trained an AI to do that, then it could hallucinate the answers, and it might be right.

[azure lynx]

also: "temperature monitor for tissue safety" would be a normal description But they wanted it to sound technical.

[olive sable]

it might be right, or it might be utter garbage. you cant know if its correct without comparing it to real data

[olive sable]

oh

so its bassicly jsut "keep under 35C to avoid cooking the human"

[runic pulsar]

C?

why not F

[azure lynx]

no idea. there is no context. they keep talking about inventing universes or some crap.

[olive sable]

cuz i dont speak american

i use C

[trim valve]

real people use rankin for temperature

[azure lynx]

K

[olive sable]

idk what a rankin is

[runic pulsar]

Celsius only makes sense for doing math, but for human perception of temperature, Fahrenheit is king, and I'm human.

[olive sable]

you're showing your americanness there bud

[runic pulsar]

I'm proud to be an American.

[azure lynx]

it depends on what you're used to.

[trim valve]

yeah europeans just don't tell temperature

[tender river]

its simply because you grew up with fahrenheit and intuitively know how it feels

[trim valve]

we actually have no idea how to read the temperature on forecasts

[tender river]

oh ew that bio, blocked

[runic pulsar]

That is true, but I would argue at zero degrees Fahrenheit, you start reaching the lowest humans can normatively handle, and at a hundred degrees Fahrenheit, you're starting to reach the upper range of what humans can normatively handle.

[olive sable]

most if the world is used to Celsius, so Celsius makes sense to us. the fact that you are used to Fahrenheit doesn't undermine Celsius being a decent way to tell the temperature

[runic pulsar]

https://tenor.com/view/point-gif-2450101530054484962

[azure lynx]

i grew up with C but have lived with F for a long time so i'm used to thinking of comfortable temperature ranges in F but use C for other things. but when I hear "below 0" it still means "below freezing" to me, not -20C or whatever it 0F is.

[trim valve]

https://klipy.com/gifs/neurosama-2

[runic pulsar]

no embed perms 😭

[rough bloom]

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

[trim valve]

I wish discord's block/ignore system wasn't so pathetic

[runic pulsar]

That's a very balanced take. I have lots of respect for you.

[trim valve]

like ignored people still appear unhidden sometimes when loading channels

and I kinda wish I could add notes directly to people I've blocked

I mean I guess its a feature that exists but I'd want it directly integrating into blocking someone

[runic pulsar]

I think I've only ever blocked scammers.

[olive sable]

i dont even use the discord built in ignore and such

if i dont want to talk to them i will jsut ignore them

[trim valve]

I kill time by reading discord so if I have the option to read ignored messages I will

I have some css to hide ignored / blocked people but its not perfect because my messages still get broken up

[trim valve]

I mostly used it here because I got sick of chat being 20 people talking about the latest llm that performs 1% better and needs exactly 25GB of vram

[olive sable]

i do agree that the chat here gets kinda weird at times

[trim valve]

its a lot cleaner when you fully just yeet people

[opaque sigil]

[runic pulsar]

What's the point of running a local LLM if you don't have tools to use it?

[olive sable]

i think it was this morning that another person was here blabering about llm sentience

[trim valve]

annoyingly though its a 50/50 on whether or not it applies

[signal trout]

pin the penguin

[trim valve]

so I sometimes get my chat just full of 1 ignored message - click to view

[opaque sigil]

i just use a plugin for that

i mean what

[trim valve]

yeah my "css" fails to inject on the first launch of discord beacuse of browser weirdness

[olive sable]

i wonder if im ignored by anyone

[azure lynx]

Ok. Claude Opus is great and all but I'm burning through my credits a little fast. A lot fast.

[hard delta]

yeah but Celsius is about water
and humans are 70% water and only 30% human
check mate papal state

[olive sable]

"humans are only 30% human"

[runic pulsar]

No because my perspective isn't based off of science, but it's based off of anecdotal evidence I've accumulated through my handful of years in life.

[olive sable]

well thats jsut the thing

we have the same for celsius

[unkempt citrus]

Its Rankine. Its Kelvin, but using the Fareinheit scale

[olive sable]

thats fucked

[unkempt citrus]

I know right

and its not even invented by an american

Its Scottish

[olive sable]

i thought the scottish got too much shit for no reason, but i see i was mistaken (/j)

[trim valve]

there's sorta no way to know

[runic pulsar]

Fahrenheit is German. Honestly, that's the best argument against it.

[trim valve]

which is the best part of ignore

[unkempt citrus]

And the germans were smart enough to replace it

which goes to show how shit it is

[olive sable]

i mean, i guess it makes sense, but i would rather just know

if someone doesn't want to hear my bullshit ill stop talking to them

[runic pulsar]

Facts and logic would dictate that you are correct, however, your mother.

[hard delta]

ani...

[trim valve]

yeah but at least for me there are chatters here who are actively bad for me to read messages from

[sage crag]

lop

[trim valve]

but I can't tell them to just straight up not chat here

[sage crag]

yes you can

do it

[olive sable]

[sage crag]

there's no reason you can't

this sounds like a joke

but im saying it seriously

[olive sable]

i agree

[runic pulsar]

I understand where you're coming from, but you have to appeal to common decency. Not to undermine your point, though.

[hard delta]

let me illustrate
begone, false prophet
only omnissiah clerics aloud

[olive sable]

if someone crosses the line they shoudl take a break from this chat

[sage crag]

take a hike 🤏

[olive sable]

i hold myself to that standard too

[trim valve]

ok but re: my point about people that its actively bad for me to read, a lot of the time its just cross a line for me and probably not anyone else

[olive sable]

im not sure where i stand on that issue actually

[runic pulsar]

If you're talking in a chat and you start getting mad, you're the one that's in the wrong. That applies to everyone. Not necessarily because it's true, but because that's the practical way of acting that will solve any heated disagreement.

[olive sable]

people should respect what you are and aren't comfortable with tho

[sage crag]

complete tosh

example:

• someone doxxes you
• you get angry
• you are in the wrong???

[olive sable]

tosh, /tɒʃ/, nouninformal•British English, noun: tosh, rubbish; nonsense.

[sage crag]

https://cdn.discordapp.com/emojis/1154436961277313126.webp?size=48&animated=true&name=evilSlam&lossless=true

ye

[runic pulsar]

Respectfully, I disagree, and I disagree because if you try and bend to everyone's dos and don'ts, then everything will be a don't. There will always be someone that won't like it. But if everyone were to appeal to a common social standard developed by a community, you would promote better discord.

[sage crag]

https://media.discordapp.net/stickers/1432142118624825556.png?size=160&name=StareAtYouNeuro&lossless=true

[olive sable]

so you want to do victim blaming instead?

[trim valve]

I kinda don't like getting into the habit of talking about specific people I've blocked in public chats, its annoying for them to read and rarely improves the chat

[olive sable]

true

[runic pulsar]

This is a very good point. I would make the argument that in the situation where the rules of the Discord server are broken, the mod should just ban the person. The rule I was setting out was to help define the gray areas where it's not so black and white.

[sage crag]

https://media.discordapp.net/stickers/1339512171213754399.gif?size=160&name=evil+clown+jam&lossless=true

[olive sable]

i guess its somewhat situational

[trim valve]

eitherway I think I'm happier for having people not appear in chat

[olive sable]

thats fair

[unkempt citrus]

I block people that annoy me, they might not have done anything wrong

[trim valve]

some people sit on the fence of an ignore, which is annoying because they get lumped into the rest of the people I ignore

[unkempt citrus]

but they annoy me

Or if they engage in bad faith

thats pretty common too

[runic pulsar]

Can you elaborate?

[trim valve]

discord should let me automatically ignore people for 24 hours if they say specific keywords

[real sierra]

to put this so bed so you can all get back on topic

[sage crag]

https://tenor.com/view/bunnies-bffs-licklick-cute-fluffy-gif-3953503

shr

[runic pulsar]

And what was that?

[olive sable]

i was going to but shiro says no

[trim valve]

every time I program in C I feel like im losing braincells

why are strings done like this

[real sierra]

there's a floating line between compromising to keep the chat atmosphere friendly and open, and recognizing when your concerns may be particular to yourself

[runic pulsar]

Ok but what about Rust?

[real sierra]

the staff are empowered to be the arbiters of where that line is, and we adjust it according to feedback and community tempo as we observe it

if you ever have concerns, you're free to report it and we'll tell you if what's bothering you is beyond what is acceptable

[olive sable]

[trim valve]

[unkempt citrus]

Basically we should be snitches

[olive sable]

yes

[unkempt citrus]

Add to my stitch collection

[sage crag]

https://cdn.discordapp.com/emojis/1464001170673766450.webp?size=48&name=Processing&lossless=true

[olive sable]

here in #programming we get in trouble for not snitching enough

[sage crag]

[unkempt citrus]

Because playing with them is so much more fun

[olive sable]

true

[runic pulsar]

Back on topic... I love rust

[rough bloom]

wide strings

[trim valve]

I also kinda hate how we have to interact with utf-16 but there is literally no:

• easy
• good
  way to interact with it

so we just get told to only read the first byte

and skip the other one

unless I wanna recreate the mpv setlocale rant