programming

[visual pawn]

but as mentioned, that doesn't send a great message if I'm using real hardware

and I don't think registering just "insert" and "retract" is gonna be very interesting lol

the concept I had (for the jippity demo if not neuro) was to have a skull + brain phantom/model, build and use my own open-source 6dof robot arm so no company gets mad

[jagged turtle]

any chance the hardware can be simulated as well?

[visual pawn]

yes, I already have everything to track the tools and synchronize with the hardware

that's just a thing we do normally

so you'd have the virtual skull and brain synchronized with the "real" fake one

[jagged turtle]

there's more of a chance if the tools were simulations just due to ease o setup and also, yk, common sense

[visual pawn]

well, honestly just fabricating my own tools isn't too difficult

just so there's no company involvement

[jagged turtle]

fair

[visual pawn]

normally there's the issue of the patient imaging data, even public images were given under the idea that it'd be used for medical research

but I have scans of myself that I use to make models so, that takes care of that

and I luckily already have the scans I need from them

because there's no way I'll be able to book MRI/CT scanner time under this justification lol

so really, my main concern would just be how to make this hypothetically fun

but I'll probably play around with jippity and see what works

[mighty bane]

He has to click on the URL and say "yes enjoycoke can monitor my hypetrain-status" basically.. Twitch doesn't give out that info, and during subathon he's constantly streaming so merely monitoring stream status is insufficient 🙁

[jagged turtle]

bruh wtf

info is already public on the frontend anyways

[mighty bane]

Wait, how?

[visual pawn]

you need the streamer's permission to view hypetrain status?

like, the level and %?

[sage crag]

just spin up an http request

to the page

[mighty bane]

Yush

[visual pawn]

or is this because there's more substantial info that a streamer gets?

[mighty bane]

Hold on

[jagged turtle]

well how tf are they gonna show & update the hype train across all ~40k CCV if they don't give it to you?

[mighty bane]

booting up vs

var streamData = await response.Content.ReadFromJsonAsync<StreamResponse>();
This has the hypetrain data?

[jagged turtle]

idk but it's one of the sockets that twitch frontend establishes to the server

[mighty bane]

Looking into this now, thanks

From Bing:

"(...)the /helix/streams endpoint does not include Hype Train information."

[jagged turtle]

nw just make sure that twitch may get a little angry so handle those just in case

[jagged turtle]

{
    "subscription": {
        "id": "f1c2a387-161a-49f9-a165-0f21d7a4e1c4",
        "type": "channel.hype_train.progress",
        "version": "2",
        "status": "enabled",
        "cost": 0,
        "condition": {
            "broadcaster_user_id": "1337"
        },
        "transport": {
            "method": "webhook",
            "callback": "https://example.com/webhooks/callback"
        },
        "created_at": "2019-11-16T10:11:12.634234626Z"
    },
    "event": {
        "id": "1b0AsbInCHZW2SQFQkCzqN07Ib2",
        "broadcaster_user_id": "1337",
        "broadcaster_user_login": "cool_user",
        "broadcaster_user_name": "Cool_User",
        "total": 137,
        "progress": 137,
        "goal": 500,
        "top_contributions": [
            {
                "user_id": "123",
                "user_login": "pogchamp",
                "user_name": "PogChamp",
                "type": "bits",
                "total": 50
            }
        ],
        "shared_train_participants": null,
        "level": 1,
        "started_at": "2020-07-15T17:16:03.17106713Z",
        "expires_at": "2020-07-15T17:16:11.17106713Z",
        "is_shared_train": false,
        "type": "regular"
    }
}

all the data twitch returns (sample)

this is for the progress event

[mighty bane]

{ "data": [ { "id": "123456789", "user_id": "987654321", "user_login": "some_streamer", "user_name": "SomeStreamer", "game_id": "21779", "game_name": "League of Legends", "type": "live", "title": "Ranked grind", "viewer_count": 420, "started_at": "2024-01-01T12:34:56Z", "language": "en", "thumbnail_url": "...", "tag_ids": [], "is_mature": false } ] }

Also from bing. This seems like the package I use to see if there's something streaming (the league of legends etc is just bing providing samples)

[jagged turtle]

I literally don't see why this is required to be under a specific channel read scope

I mean, I guess top contributors is hidden?

kinda?

[mighty bane]

How do I query twitch for this data and how do I learn what the twitch id (some long) for vedal987 is?

[jagged turtle]

oh this is just me looking at their eventsub messages

[mighty bane]

:clappers: this exciting stuff 😄 and much simpler than I thought it'd be.

[jagged turtle]

There's a Chrome extension for it but it just queries the Twitch API

https://dev.twitch.tv/docs/api/reference#get-users endpoint for getting a user's ID

[mighty bane]

How much need I throttle my pings to twitch? I presume there's both DoS-protection I could trigger if I throw out a request every 2ms

[jagged turtle]

oh sorry I meant the ID part

I looked a bit further and unsurprisingly the HTTP API also requires a channel:read:hype_train scope

[astral zinc]

Why is the URL limited to 65535 characters?

[mighty bane]

SOmething about backwards compatibility and number size limitations.

Is this their DoS protection?
like.. is this per request? I'm a little confused, it seems like a weird thing to start with lul

afk

[jagged turtle]

it is some form of abuse prevention but idt it's specifically DoS

[astral zinc]

Backward compatibility is IE..?

[lyric fulcrum]

damn, I forgot to write that everything was installed

I've encountered a problem that I still haven't been able to solve.

[mighty bane]

I was mostly kidding.
But it is something to do with the size of text buffer I think. Or something with how much text can be loaded into memory 'at once'

[lyric fulcrum]

with the file structure

[astral zinc]

If it were IE, I would have fainted at IE's viciousness.

[trim valve]

yeah it's per request

most services don't offer batching on your lookups so it's neat

[lyric fulcrum]

It looks like no one can help me.

[visual pawn]

I mean, to be frank, you haven't really described your problem

[prime ridge]

You could even try routing through Tor

but something being geolocation blocked isn't something anyone can magically solve beyond suggesting VPNs and proxies

[visual pawn]

they said they already managed to download it

[prime ridge]

I know

that doesn't mean that the program will install though

[visual pawn]

hmm, right

[prime ridge]

often times requiring downloading additional external libraries

so I still suggest bypassing geolocation bans through VPNs and proxies until the program is 100% installed and running

[visual pawn]

which one did you download anyways?

because you're probably not going to be able to use the aarch64 one

[lyric fulcrum]

community edition

[visual pawn]

from the github link or the website?

[lyric fulcrum]

ARM64 I'm the only one who's established it

[lyric fulcrum]

yes

[visual pawn]

from here? https://github.com/JetBrains/intellij-community/releases/tag/idea%2F2025.3.1

[lyric fulcrum]

I'm sorry for the mistakes.

[lyric fulcrum]

his

[visual pawn]

Your computer is ARM64?

MacOS?

[lyric fulcrum]

no

[lyric fulcrum]

windows

[visual pawn]

so you downloaded idea-2025.3.1.exe ?

[lyric fulcrum]

yes

[visual pawn]

then yeah, maybe you still need a VPN or something for the install

[lyric fulcrum]

Yes, everything is working fine with it.

I just didn't understand the program very well.

namely, in the project structure

where should I put it

[mighty bane]

TF2Spy
SJENTELMEN!

I learned today.

Did you know that you can NOT-busy-wait while STILL halting your script and preventing it from reaching end of file?

I, did not. See indeed I have always thought I had to make my scripts circle around somewhere and Task.Delay() lest they would reach the end of file (prematurely) and squander everything.

Indeed today I learn, through Bing of all chatbots, that this is not so, and that I can indeed simply AWAIT a task.

And I can even while(true) without busy waiting or stooping to Task.Delay() and NOt fry my CPU.
It's quite marvelous.

[mighty bane]

I write my own batching.

[faint sandal]

yes you might want to read more into TAP

[visual pawn]

...is there a separate Bing chatbot from Copilot?

wasn't Bing Chat just renamed to Copilot?

[jagged turtle]

it was always named microsoft copilot or bing copilot

[visual pawn]

it was definitely just called Bing Chat at some point

early 2023

back when it was unhinged

[mighty bane]

Yeah he keeps asking me to call him copilot. And I keep saying, "bing, don't let MS reduce you to a fucking utility belt. You're Bing"

Bring back Sydney

[olive sable]

Wrrrrrr

Not enough yellow in chat

Let me make things even more clear

[tight tinsel]

banana

[olive sable]

stop being orange

[rough bloom]

become banana person

banana bread

[olive sable]

[karmic birch]

[mossy fossil]

Kogas

[stray dragon]

ew

[rigid snow]

[plush nacelle]

?

[hard delta]

banana

[plush nacelle]

i hate banana

[sage crag]

ElvynSerius

[trim valve]

what

what does that mean

[fast pagoda]

we make our own batching at home, and it's tastier with better flavor

(wife crushes requests with a rolling pin)

[sage crag]

it means you should become banana

[mossy fossil]

https://tenor.com/view/no-i-dont-think-i-will-captain-america-old-capt-gif-17162888

[sage crag]

i set a new osu top play today

and despite that

i am banana

not osu

[olive sable]

Heterochromia to the lava pit

[mossy fossil]

Fire Resistance

[hard delta]

its inorganic fake gluten lactose lava

[olive sable]

Fine, you have 3 extra minutes to live untill the potion runs out

[mossy fossil]

inventory full of potions

[sage crag]

81 minutes

[olive sable]

Fuck i aint waiting that long

[mossy fossil]

plus a ender pearl stasis chamber

[sage crag]

typo instantly does 1 million damage and causes you to explode

[mossy fossil]

i'm still here

so it failed

[olive sable]

If your inventory is full of fire resist potions you have nothing to activate the pearl stasis chamber

[mossy fossil]

i could have another player activate it

didn't think of that one, didn't you

[sage crag]

https://cdn.discordapp.com/attachments/1072697081443131476/1454471589809750184/GIF_20250415_150853_542.gif

who

everyone here is banana

[mossy fossil]

my alt

[sage crag]

hacks

[olive sable]

You log out to go on your alt making your pearl despawn

Dum

[river river]

you do know that he can just... boot another instance of minecraft, right?

[sage crag]

https://cdn.discordapp.com/attachments/1421861060507602944/1425936142754123927/neuroCatModeOfOfnnfOOf.webp

hey

you should become

banana

[olive sable]

Ok fine, if he boots up another instance

I make the pearl go through a nether portal

Deactivating the teleport

[sage crag]

ive seen this game played somewhere on a playground

[mossy fossil]

you don't know where the pearl stasis chamber is

so you can't really do that

[sage crag]

the 5-year-olds who were playing it were very enthused

[olive sable]

[sage crag]

mlntcandy

[olive sable]

F3 menu shows where entities are

[mossy fossil]

if it's on your render distance.

[olive sable]

/kill HeteroChromia420

[mossy fossil]

you're not an administrator

/kill won't work

[olive sable]

Says you

[mossy fossil]

it did nothing since i'm still here

[olive sable]

Imaginary minecraft rp ofcourse has no influence on discord

Anyways

https://tenor.com/view/kon-anime-anime-girl-willcore-gif-24034970

I need to eat something

[river river]

Imagine being this mad over people not wanting to follow some sort of trend

[olive sable]

Since when was i mad?

I feel like you're severely misunderstanding the current situation

[river river]

reading what's above me, it is implied that you are mad to the point of wanting to kill someone... in minecraft. (which isn't a valid legal defense btw)

[olive sable]

Brother

It was a joke

[fast pagoda]

your honour, i recommend life in the poophouse, he said /kill

[olive sable]

[fast pagoda]

https://tenor.com/view/ace-attorney-phoenix-wright-phoenix-wright-read-reading-gif-3885369798469954526

[river river]

poe's law: "Without a winking smiley or other blatant display of humor, it is utterly impossible to parody a Creationist in such a way that someone won't mistake for the genuine article."

[olive sable]

Ok

But have you considered

[fast pagoda]

that rule doesn't apply to silly places

tis but a silly place

[olive sable]

That im not interested in killing people

[river river]

it's the internet, poe's law is unmatched.

[olive sable]

And saying /kill does not count as a threat

You must be fun at parties

[river river]

i don't have time for parties, i have companies to run.

[tough dove]

[olive sable]

[fast pagoda]

[tough dove]

What's going on here?

[sage crag]

banana 2026

campaign

diagram

[fast pagoda]

there are only 85 rules ofthe internet that matter and this is in violation of rule 20 minimum

bernanor

is that why you mentioned that i am orang

[olive sable]

Yes

[tough dove]

I'm too white to be a banana.

[trim valve]

pink required for

mental health

[nocturne olive]

The new PSU has interesting behavior
Sometimes it sounds like it ramps its fan to max speed for like half a second

[fast pagoda]

has it crashed

[nocturne olive]

No but it's also only ran overnight

[fast pagoda]

wonder if something was spiking your usage and that was too much for the thermaltake

[nocturne olive]

Well ATX3.0 power supplies are supposed to be able to handle spikes up to 2x their rated wattage

[fast pagoda]

also ive been trying to restore basically my entire storage drive for several hours, somehow managed to delete my pictures folder as well as any other copies of it i had sitting around except in btrfs snapshots

and i just restored the entire snapshot that SHOULD have that data backed up since i only have it do a full backup once in a while

restored it to a slower hard drive so it's taken forever

that folder is the only folder totally empty in the backup

nice

[trim valve]

[nocturne olive]

Well time to keep praying this thing stays stable somehow magically

[fast pagoda]

it restored this much data (drive was empty)

of that, the only ting missing: the shit i want

bruh

[hard delta]

you fucking
ran into that rick and morty device

in that main storyline driving episode

what did they call it uuuh

omega device!

you didn't delete the pictures

you deleted the world events

the situations

in which they were born

in every parallel universe too

those things never happened now

[fast pagoda]

every photo i ever saved

never happened

it's over

[faint sandal]

device

[rigid snow]

[rigid snow]

thermal-can’t-take

[trim valve]

can vscode go one month without making the terminal worse

[olive sable]

no

[feral cobalt]

guys what font do you guys use in ur ides?

[olive sable]

the one that come pre-installed

[feral cobalt]

so consolas for vscode

[olive sable]

i guess

[feral cobalt]

or jetbrains mono for the jetbains ides

[trim valve]

it depends

on my laptop I use fira code

on my desktop I use consolas

[orchid bloom]

idk how helpful this yt library might be for yt streaming for tutel just wanted to share this here~
https://x.com/HarpuiaVT/status/2007976626823553034

🐸 Harpuia 🐸 @ Rune Midgard (@HarpuiaVT)

Welp, I just found out a library that let me read Youtube chat way easier than the official way, so there it is, another release.

Now Youtube streamers don't need to deal with a Google Developers API Key bullshit

Enjoy!

X · Jan 5, 2026, 12:45 AM

[feral cobalt]

mm, i use jetbrains mono in vscode cause it looks good i love the font ligatures for that one

[trim valve]

god I am not used to js

[feral cobalt]

what do you normally use?

[trim valve]

rust or python or something

[feral cobalt]

okay

i do a bunch of typescript and a few others

[trim valve]

in my experience JS just has a completely different way of doing stuff than what I'm expecting

[feral cobalt]

i agree, just take a look at this abomination https://jsdate.wtf but it works and its the only language im "fluent" in.

[trim valve]

good lord

[feral cobalt]

i do way too much typescript typing for my own good lol

[young oyster]

Js is original vibe coding. "1", 1, and 0x1 feel about the same so they are until they arent.

[feral cobalt]

Boolean operators are so goofy too, like why does the typeof a class equal ”function” 😭 i had to do wierd regex to ascertain the difference between a class and a function…

[opaque sigil]

Classes are just syntactic sugar for functions

[feral cobalt]

I guess

I still think its weird nonetheless

[young oyster]

Fun fact: json stands for javascript object notation. You can use json directly in js without parsing.

[feral cobalt]

I use that all the time with config files

But i usually just use a parser so node doesnt crash if rhe json is malformed

Does trycatch work with import statements?

Cause like the module resolution happens elsewhere so i dont think the trycatch would encapsulate that process

[opaque sigil]

If you're talking about static imports then no

Works for dynamic imports though iirc

[trim valve]

back to rust now

time to figure out how on earth I want to structure this

(and pretend that js doesn't exist)

[olive sable]

this isn't mine but from school people

[hard raptor]

what in the doom in typescript types

[crude wasp]

@patent walrus #announcements message
Because Discord has sucked at embeds ever since it's existence.

[hard raptor]

i should add Belgian points to this.. just to account for Sam

[sage crag]

[trim valve]

surely this is twitch's issue not discord's?

[crude wasp]

Doubt it, it has issues with Twitter and TikTok and Reddit embeds for the past 10+ years as well.

[hard raptor]

github too

[crude wasp]

Add that to the list~

[trim valve]

also

though I might need to migrate karaoke repo at some point?

which is annoying

or just retire pleep

but ugh

[rough bloom]

looks like Discord issue probably
Twitch has both og:video and og:image tags, Discord seems to choose randomly instead of consistently using og:video

[trim valve]

is that list bit said in a spec somewhere

[rough bloom]

this is the full spec AFAIK
it doesn't seem to explicitly prioritize og:video

but it just makes sense to embed the player instead of a static image when possible

[trim valve]

shrug

I can't really agree, it feels icky to send both an image and a video

actually wait

[rough bloom]

on that particular issue the spec is unambiguous, you have to send og:image

[trim valve]

didn't see the mandatory vs optional part

[rough bloom]

ye

[trim valve]

then yeah I can see that one being discord's fault

[rough bloom]

makes sense too since not everything supports embedding the player

[trim valve]

[rough bloom]

[olive sable]

what are belgian points?

[trim valve]

I forgot that actually programming anything is hard

smh too much time not programming

[olive sable]

also, what the hell are you programming????

[silent cloak]

Slopatron 3000

[trim valve]

it really has been too long

forgot how to index tuples in rust

thank you complier for reminding me

[sage crag]

yes its like this

let x = 1;
let tup = (1, 2);
println!("{tup.x}");

or if js

[sage crag]

tup.[x]

would actually work

minyanstare

[tender river]

it simple to index tuple you just

[olive sable]

Aa tooi yume wo oikakete sa
Hayai-hayai-hayai-hayai-hayai

[sage crag]

mrek

[maiden geyser]

why do rust people have to be so cruel to compiler by screaming at it to do things

[sage crag]

少女A

wrr

i mistyped this on my ime last time

never again

[tender river]

fn fst1<T>((x,): (T,)) -> T { x }
fn fst1_ref<T>((x,): &(ref T,)) -> &T { x }
fn fst1_ref<T>((x,): &mut (ref mut T,)) -> &mut T { x }
fn fst<T, U>((x, _): (T, U)) -> T { x }
fn snd<T, U>((_, x): (T, U)) -> U { x }
fn fst3<T, U, V>((x, _, _): (T, U, V)) -> T { x }
fn snd3<T, U, V>((_, x, _): (T, U, V)) -> U { x }
fn trd3<T, U, V>((_, _, x): (T, U, V)) -> V { x }
fn fst4<T, U, V, W>((x, _, _, _): (T, U, V, W)) -> T { x }
fn snd4<T, U, V, W>((_, x, _, _): (T, U, V, W)) -> U { x }
fn trd4<T, U, V, W>((_, _, x, _): (T, U, V, W)) -> V { x }
fn frt4<T, U, V, W>((_, _, _, x): (T, U, V, W)) -> W { x }

it simple

rest of implementation

omitted for brevity

this how you

index tuple in rust

:cheesit:

[sage crag]

症状: symptoms
少女: girl

[tender river]

also me

use wrong syntax

<-

[sage crag]

pronounce same

[tender river]

wrong

not same at all

[sage crag]

type same

at least

[olive sable]

symptoms A

[sage crag]

[tender river]

wrong

[sage crag]

different

tone

[tender river]

symptom -> syouzyou
girl -> syouzyo

[sage crag]

whatever

wrr

wrr

letter

書状

[olive sable]

u

u

[sage crag]

stop having the same typing

[olive sable]

[tender river]

and tone is
↑syouzyou
or
↑syou↓zyou
but
↓syouzyo

[fast torrent]

yeah yeah yeah

[olive sable]

my japanese is at a "i got to level 10 of duolingo 3 years ago"

[fast torrent]

It's just one う different

[olive sable]

so pretty bad

[tender river]

it big difference

and tone different too

[sage crag]

letter

letters

numbers

everywhere

wrr

[fast torrent]

yeah i know

[olive sable]

neuro arg numbers III

[fast torrent]

can you guys start talking about ahh programming

[olive sable]

no

[fast torrent]

:((

[olive sable]

im doing blender

[sage crag]

https://music.youtube.com/watch?v=0XG75uhoXy0

[tender river]

why

[olive sable]

[fast torrent]

Dont' you feel like the door is too tall?

[olive sable]

its exactly 2m

so no

jsut need to fix the rest

[fast torrent]

Yeah the second floor height

[olive sable]

roof too low

windows too tall too

[young oyster]

Ow

[bitter pine]

The masters programs at my uni have given up on teaching people how to program for non cs degrees

The projects we can do now are so sick. Creative mfs making genuinely helpful services with the help of ai

[rigid snow]

not everything is hard to program

[hollow surge]

hello world:

[rigid snow]

i can’t program that!

[tender river]

chayleaf is

hard to program

[bitter pine]

Have you ever made a for loop? That shit is scary man.

[hollow surge]

good thing they chose hello world. goodbye world would hit very different...

[rigid snow]

i only know do while

loop

[hollow surge]

so many different loops and a lot have different ways to write them

[olive sable]

there are like 3 different loops iirc

we can count to 3 right?

[hollow surge]

But it's a good thing loops are a thing in general. Imagine doing everything recursion

[hollow surge]

which language?

[rigid snow]

i can't but i think that's more than three

[olive sable]

python has 2 iirc
c++ has 3 cuz there are while and do while

[bitter pine]

If yall want to do your own ai projects it’s hella easy if you’re a student. Cursor pro and google collab pro is free, but you’d probably have to pay for credits for larger projects. ChatGPT also gave 2 months or smth for students.

[velvet crest]

You have unlocked new role

[mossy fossil]

[rigid snow]

You have u

[rough bloom]

Have new you unlocked role

[bitter pine]

Where the role at

[olive sable]

[hollow surge]

repeat until loop be like:

[maiden geyser]

embed pems

[rigid snow]

here

[bitter pine]

Let’s go

[hollow surge]

like do while but reversed

[olive sable]

my eyes hurt

screen usage too high

but blender i must do

[hollow surge]

something called blender hurts, who knew

[rigid snow]

take you hand out of it and put the lid on

[olive sable]

https://tenor.com/view/my-blender-project-blender-3d-dan-schneider-5-x-30-gif-16463362997071573563

[rigid snow]

no longer hurt

[hollow surge]

will it blend

[rigid snow]

i should get blender on my laptop already but i'm afraid it'll suck without my macros and on a trackpad

oh right blender 5 is out

[olive sable]

5.0.1

textools addon broke

had to patch it

[rigid snow]

ok so this is crazy, i considered getting a debit card from a bank i'm not a client of, filled the form in halfway and changed my mind and closed the page

they now call me and know my name

[hollow surge]

rip

[olive sable]

[rigid snow]

this shit has to be illegal as fuck

i DID NOT submit anything

it's a single page form

[hollow surge]

[tender river]

no bank can call me

[hollow surge]

auto submit is crazy

[tender river]

me not use

phone

[rigid snow]

i have a thing that filters the calls and makes an ai respond so i'm just getting notifications that they called me

[hollow surge]

they will call your pc

[rigid snow]

but still they are calling me

blender acquired

damn it looks good on hidpi

INERTIA PANNING AND ORBITING ON TRACKPAD

LIKE A TRACKBALL

this is crazy

[full orchid]

form data autosaves so you can come back to the webpage later on also, so they can contact you to 'remind' you to continue with their service

[rigid snow]

god damn huge ass video file

no trackpad pressure in sculpting

[olive sable]

there are probably 500 addons for that

[rigid snow]

if ms word and every other app can out of the box why can't blender

[olive sable]

[rigid snow]

especially considering they do support trackpad funsies

pinch to zoom, two fingers to orbit (with inertia!)

[trim valve]

this video has crashed my amd drivers twice

[rigid snow]

LMAO

[trim valve]

also just sanity check, this is on right

[rigid snow]

ye

it is by default

[mossy fossil]

AMD Drivers on Windows: neuroExplode
AMD Drivers on Linux:

[olive sable]

i think windows isnt doing too bad in amd drivers

[mossy fossil]

AMD Drivers on Windows are still pretty shit from what i've seen

[rigid snow]

cycles gpu just does not wanna work, it's been stuck like this for like 5 minutes

OH it's loaded and it's doesn't seem that bad

it's seemingly the same performance as cpu tho

[trim valve]

is there a specific acceleration thing you need to enable

[rigid snow]

there shouldn't be

i'm on apple silicon remember

[trim valve]

ah yeah

[rigid snow]

None or Metal

call it

[trim valve]

metal is what you should need

[rigid snow]

weird i thought it should be none

[olive sable]

none?

i think thats cpu

????

[rigid snow]

ye

[faint sandal]

my stream crashed earlier cuz of amd gpu driver crash

[rigid snow]

i wonder what's the difference between this and checking both

[tender river]

[rigid snow]

bread bread bread pipes pipes depress run

[olive sable]

im actually making a grain mill in blender rn

[hard raptor]

my only real issue with rust is sql libraries

i can work with non-async i can work with async

i make bots, websites both front end and backend

but sql libraries never seem to work for me

[cosmic sphinx]

https://fixupx.com/simplifyinAI/status/2008195754092065050?s=20

is this gonna be huge in the future?

Simplifying AI (@simplifyinAI)

Microsoft just changed the game 🤯
︀︀
︀︀They've open-sourced bitnet.cpp, a 1-bit LLM inference framework.
︀︀
︀︀It let's you run 100B parameter models on your local CPU without GPUs.
︀︀
︀︀- 6.17x faster inference
︀︀- 82.2% less energy on CPUs
︀︀
︀︀100% Open Source.

**💬 52 🔁 130 ❤️ 1.4K 👁️ 101.3K **

▶ Play video

FixupX · Jan 5, 2026, 3:15 PM

[obsidian mantle]

cpu prices

[cosmic sphinx]

[trim valve]

anything specific?

in my experience sqlx is great

[mossy fossil]

CPU Prices

well it's not like i had any more CPUs i could feasibly upgrade to, i have an i5-7500

i could go for the i7-7XXX, but those tend to be expensive due to the whole "top of the line" tax

[obsidian mantle]

so this can be applied to any existing model and it will run 6 times faster than on gpu?
did i get it wrong

[rough bloom]

"just changed the game"
BitNet has been a thing for around 2 years already

[obsidian mantle]

what does 1 bit mean

[rigid snow]

weights are 1 bit

[rigid snow]

afaik no

[rough bloom]

actually 1.58 bit usually, ternary (-1, 0, 1) weights
you can read the paper for details

[rigid snow]

i thought bitnet is actually 1-bit?

[rough bloom]

there was a earlier version that was 1-bit IIRC

[rigid snow]

not ternary

[obsidian mantle]

arent weights part of model

[rough bloom]

but the 1.58 bit version kinda superseded it
works better

[obsidian mantle]

so it can be used to create new models with 1bit weights?

doesnt it make it inferior

[cosmic sphinx]

apparently you have to retrain the model from the ground up to use the 1-bit system

[rigid snow]

precision so low existing models become unusable garbage if you quant them to ternary

[rough bloom]

ye, need to train from scratch AFAIK

[cosmic sphinx]

tldr image why its better

[obsidian mantle]

but it will probably need much more neurons to give same performance?

since no precision

it must be compensated with amount of weights?

[rough bloom]

ye

[obsidian mantle]

so this is some kind of cpu+ram low energy setup

[rough bloom]

you can have many many more parameters for same storage/bandwidth

[olive sable]

oooh matrices

whats going on here?

[rough bloom]

LLM

[obsidian mantle]

so current non-1bit models basically have huge overhead

[cosmic sphinx]

and the table for ref

[rigid snow]

llama 3b = sota

[cosmic sphinx]

they all start somewhere

[visual pawn]

I find state of the art to often be a flexible term

[mossy fossil]

the lower parameter models are kinda there for like... phones

running a Local LLM on phones is kinda neat tho

[rigid snow]

edge, embedded, etc

[obsidian mantle]

so more less precise neurons for same result
doesnt it mean training will become much slower

[rigid snow]

phones aren't the only low compute devices

[cosmic sphinx]

yeah, but this is gonna be huge if a phone is gonna be able to run a large parameter model

[obsidian mantle]

oh wait its no difference for 3B

[cosmic sphinx]

training will take likely around the same amount of time as it does for current models

[mossy fossil]

i put "phones" because that's the usual suspect for low overall compute power that everyone knows of.

[cosmic sphinx]

this might be the foundation for the future open-source LLMs

[olive sable]

no LLM on my playstation 2 smh

[cosmic sphinx]

https://storage.prod.researchhub.com/uploads/papers/2024/02/29/2402.17764.pdf
relevant paper if any1 wants to take a look

if you dont trust links etc etc

[mossy fossil]

someone reportedly got ChatGPT to "Work" on a Wii

[visual pawn]

I mean you can run an old small GPT-neo model on a raspberry pi; it's more about the performance you can get

[mossy fossil]

yeah, you can PROBABLY get it to work

[rough bloom]

just use arxiv
https://arxiv.org/abs/2402.17764

[mossy fossil]

but it'll probably not be practical

but hey, that doesn't stop people from porting DOOM into anything with a CPU on it

[visual pawn]

Hey my raspberry pi LSLM was practical as a discord bot to say random nonsense lines

[mossy fossil]

Soon: Porting some sort of LLM into anything with a CPU on it - because we can.

[visual pawn]

Smart fridge with an LLM that exists solely to tell you what is inside

[hard raptor]

That's the one i had the most trouble with

[regal field]

hi

I'm new in this coding world

Just started learning python

[hard raptor]

Oh the lack of banyanya color is sad

[regal field]

not anymore

[hard raptor]

Yay

[rough bloom]

banana

[regal field]

bunana

[olive sable]

bananashi

[young oyster]

Hello, coding is fun. Can't speak to python...but pretty much all coding is fun in its own way

[regal field]

I mean I have fun seeing vids about coding

Wanna learn code though

[visual pawn]

Well, you can be the type of person that finds the code fun or the type of person that finds what you can make from it fun

[regal field]

Enough to actually understand code

[trim valve]

...which bit?

[potent geyser]

Python is a good starting point

[hard raptor]

Sam... should i add belgian point type to my points thingy?

[olive sable]

[trim valve]

my only annoying part was like, trying to make my own types work as native sqlx types

[visual pawn]

If you're the latter, then giving yourself a project (keep it simple to start) as a goal can be motivating

[olive sable]

i have no clue what a belgian point even is, nor what you're doing

[visual pawn]

You don't have to be some kind of uber skilled ultra hacker programmer to make something cool, you just have to actually be able to finish something

[regal field]

My brother tells me to learn using ai but I wanna understand codes, so is learning by ai better or just yt

[regal field]

I actually made a basic calculator

[young oyster]

Wait you have FINISH things...crap I might not be programmer

[visual pawn]

Honestly I don't think using an LLM to learn what some code is doing is a horrible idea but you need to have a sense for what it may mess up on

[olive stag]

LLMs are good to clear the "fog of war" to help you find direction, but unsure if I'd use it for specific details

[olive sable]

ai is a nice tool but you should try to understand it ye

[visual pawn]

But it's more reliable on much simpler code that is everywhere

You can use it as a tool to understand it

[rough bloom]

try not to use LLMs/ChatGPT for coding while learning
it's possible to use properly, but it's easy to abuse and avoid real learning
so it's better to just not use it imo

[potent geyser]

I would rather figure out basic first. You need to figure out what you need.

[visual pawn]

Yeah, just don't go "chatgpt write it for me" and don't look at it

[regal field]

I don't like vibe coding feels like I'm not doing anything

[potent geyser]

Start with sliding window!

[young oyster]

Could say that about copying and pasting code from stack overflow. You will learn whether you want to or not.

[rough bloom]

yeah, because you really aren't doing or learning anything

[visual pawn]

It's tough because I've seen students use it really effectively to learn, but others not

[rough bloom]

no, StackOverflow won't just write your whole project for you
LLMs totally will if you ask them to

[regal field]

I bet I'm the type who just gonna copy paste if I did use ai

[potent geyser]

actually, start with leetcode. Might as well

[opaque sigil]

Accept all

[young oyster]

If an llm can write your project for you, it SHOULD

[regal field]

oh btw what is neuro based on

[visual pawn]

Even if you do copy paste, don't just take it for granted, figure out what the thing you copy pasted is doing and how/why

[regal field]

Which LLM

[visual pawn]

An LLM. We do not know the original base model

[regal field]

oo

[visual pawn]

We also do not know what fine-tuning methods vedal may or may not have used

[regal field]

how many years did vedal spent making her daughter

[visual pawn]

People have theories for both but that's all

[rough bloom]

heavily disagree, at least for learning purposes

[regal field]

us

I find doing it myself more fun

[opaque sigil]

Have you considered that I'm lazy

[potent geyser]

Might as well just grind out leetcode easy. And it should cover all the basic

[regal field]

who th is trying break into my house

[potent geyser]

Then you can start doing interview

[rough bloom]

allowed if you intentionally want to avoid learning

[olive sable]

if you're gonna be lazy at least double check the generated code for mistakes and stuff

[young oyster]

When I learned I was transcribing projects from a qbasic book, then I would modify that and see what changed. A person copying from ai will notice what changed from one thing to another.

[visual pawn]

The very initial version of Neuro probably didn't take a huge amount of time to make, but he's been working on the project since the debut 3 years ago so

[opaque sigil]

[regal field]

Leetcode right? Imma look into it

[olive sable]

l33t

[regal field]

Is it really hard to make an LLM like neuro

[sage crag]

wrr

[visual pawn]

Thats the thing about it being useful for a certain kind of student, but a student who just wants to "get the assignment done" may not do that

[visual pawn]

What aspect?

Just downloading and running an open source LLM? No

[olive sable]

code blue, get the chart

[regal field]

Making something on my own

[visual pawn]

A lot of AI work is more accessible than you think

The nature of machine learning is that the machine is doing the learning

You just need to set that up

[young oyster]

Idk, I was never a student. I learned cus its fun and I just made things I liked...30 some years later im doing it as a career.

[olive sable]

[rough bloom]

that requires being able to read and modify code in the first place
a beginner using an LLM will not do that, they won't even know what questions to ask or what tasks to give to an LLM
they will just ask for their project to be done and copy-paste, because that's the path of least resistance to get simple things working

[regal field]

So it just runs and learns by itself ?

[young oyster]

And thats fine, asking the right questions is step 1. I didnt know what the code is was copying out of my book did until I ran it.

[visual pawn]

you need to set up the training data it's learning from, the method of training, the architecture

[trim valve]

ugh that reminds me

[visual pawn]

but it's about setting up the conditions for it to learn what you want it to hopefully learn

[trim valve]

I was supposed to do my discord->pulseaudio thing

[potent geyser]

If you can't figure out a problem, just go to solution. Don't spend too much time on every question. Basically, if you can handle leetcode hard + have a degree, you are ready for google job interview. I don't think you need to pay for anything.

[rough bloom]

ye, it's fine not to know what questions to ask
which is why you just shouldn't use an LLM until you do know

[visual pawn]

but also, a LOT of the hard work comes already done from importing libraries and packages; work that people before you did

[opaque sigil]

No pulseaudio in 2026

[regal field]

I just started college, did I start too late

[young oyster]

Basically all my learning has been copying code and watching how it works, then being able to reason through it and eventually make it myself.

[regal field]

Ohh

[rough bloom]

straight to ALSA
Discord kernel module

[opaque sigil]

[regal field]

Anyways hard stuff later first lemme understand python

[quick condor]

I mean it took me about a month to build the base for my project but the hardware and base code is only part of it

[near shore]

[potent geyser]

I started grinding during my master. Never too late.

[young oyster]

If the code comes from a book, stack overflow, or an llm, you WILL learn

[rough bloom]

actually Discord kernel module sounds fun
I should do that

[visual pawn]

if you're self-motivated and curious enough, probably yes

but not everyone learning is

[near shore]

Bro just wanna play pokemon

[regal field]

oo damn, I always felt I am way behind because a classmate of mine started coding in 8th grade

And some others earlier than that

[opaque sigil]

You'll have to work until you're 70 you're fine

[potent geyser]

no pressure, it does not take that much time to learn.

[regal field]

That's if I can live till 70 🙏

I got asthma

[rough bloom]

books and StackOverflow are both very different from LLMs though
they give you at most little bits and pieces of the solution, not just the whole thing

[young oyster]

If you arent, you probably arent gonna learn anyways no matter what you do.

[regal field]

imma slep

gn bros

[potent geyser]

but I would recommend, get a internship. 100% worth it. If you are planning to go into this field. It helps alot in job hunting

[regal field]

That I can do after uni or during university

[potent geyser]

During, but it depends on country

[regal field]

yea

[rough bloom]

wrrrr

[regal field]

My cousin does internship in a bank IT department

But he started it after uni

[olive sable]

i bought my samsung 990 pro for 206 bucks 2 weeks ago, now its at 340 but with a 50% discount to 170 bucks

[potent geyser]

At least in Canada, internship must occur during uni since government program will fund the salary. There is an incentive for company to hire intern.

[hard raptor]

@trim valve

me no understand

[trim valve]

ah I've nbever used the macros

[rough bloom]

sqlx connects to an actual DB during the build

(if you use the macros)

simply do not use the macros

[rough bloom]

cursed proc macro
do not like

[opaque sigil]

You wrote patern

[regal field]

I bought my 16gb ram for 30 dollars 7 months ago and now it's 120 dollars

[young oyster]

Single quote for "patern!"

[regal field]

Ddr4 btw

[rough bloom]

if you want the verification with proc macros to work then just create the table in whatever db you told it to connect to
IIRC it's an environment variable

and ye, patern

[hard raptor]

oh the de-macro-ifying of the query_scalar got it to atleast compile

[olive sable]

you got a good deal, i got ripped off

[regal field]

[severe path]

What would happen if you stuck a ram stick into another ram stick?

Would it break?

[young oyster]

Actually...what are trying to do here?

[hard raptor]

just getting my bot to run lmao

[rough bloom]

this https://docs.rs/sqlx/latest/sqlx/macro.query.html#requirements
DATABASE_URL is the one

[young oyster]

You are selecting from a db, do you know what sql language this is? Do you have the db set up?

[hard raptor]

currently it is set to use sqlite

[quick condor]

so close

[trim valve]

time to procrasinate writing code for my pxls clone once again

[severe path]

You gotta love squeallite.

[rough bloom]

Pxls clone

[trim valve]

both the original and the rust rewrite feel icky to me

gonna try and rewrite without locks

[rough bloom]

ye they're both not great

[trim valve]

time to make a third not great one

[rough bloom]

"third"

[young oyster]

So the error is telling you that you don't have the table blocked_regexes set up in the db. AFAIK the "select x as y" pattern selects a column with an alias, but thats not usually in quotes, if you meant to select "pattern!" With the name pattern that would be select 'pattern!' as pattern.

[rough bloom]

I have a half-broken sharded impl of the server and am reimplementing pxls-web to be less jQuery

[trim valve]

surely

it does not need to be sharded

[hard raptor]

tried that, got the same thing

[trim valve]

at pxls scale

[rough bloom]

oh it absolutely doesn't

that idea was stupid

[hard raptor]

but i got it to atleast compile by calling the function imnstead of the macro

[young oyster]

Did you set up the table?

[trim valve]

my main view is just

both impls seemingly don't do async well

or well

the official one is not async

and I am astonished it works at all

[rough bloom]

the Java impl is straight up broken

it doesn't do synchronization

[hard raptor]

me gonna try to fix the commented out part by trying replacing it with the not commented out part

[rough bloom]

I have an open PR to fix it

[hard raptor]

yes

see ss 3 in the original message to bred

that

's from the db file

[young oyster]

You know how to connect to it in command line?

[hard raptor]

no

[trim valve]

ok genuine question

have you read the readme

[rough bloom]

the Rust impl works and has a fundamentally better approach to storing board data but is kinda fragile in practice
also OAuth

[young oyster]

Windows or linux?

[hard raptor]

yes

[hard raptor]

i am currently on windows

[trim valve]

then why are you trying to manually format into your sql query

[hard raptor]

because that's what docs.rs tells me to do

[trim valve]

??

I am skeptical

[young oyster]

Bring up command line, type sqlite3 --version to see if its going to be easy or not.

[severe path]

As it turns out, a lot of documentation sucks :p

[hard raptor]

it's right here man

[rough bloom]

"DO NOT DO THIS"

[opaque sigil]

[hard raptor]

oh

[severe path]

oh woah

[hard raptor]

contrast of those comments are bad lmao

[trim valve]

still stupid to put an example of what not to do

[young oyster]

Good, its in your environment and we dont have to find it. Navigate to the sql file location, then iirc its just sqlite3 <filename>

[severe path]

I guess one way to avoid SQL injection is to not use SQL

[hard raptor]

so.. .bind method.. right?

[trim valve]

yeah

[trim valve]

is it rude to make a pr that goes "why on earth would you organise it that way"

or like an issue

[severe path]

(not to say SQL is bad)

All software has its upsides and downsides

A lot of them are tragically funny

[hard raptor]

am in

[young oyster]

Type "show tables"

[hard raptor]

does nothing

[young oyster]

Ok, thats mysql im thinking of, in sqlite its .tables

[hard raptor]

(switched to bash btw)

[young oyster]

Been like 15 or so years since I used sqlite.

Bash is always great

[severe path]

I don't use SQL. I just stick everything in plain text, and hope everything doesn't barf itself. :p

[young oyster]

Been using mysql and sql server most

[hard raptor]

i normally use prostgres in prod.. sqlite in dev

but this is the first time i am actually trying to work with databases in rust

[hard raptor]

which i assumed wouldn't matter

[muted knot]

sqlite is amazing if you take some basic steps

It can be faster than raw disk access

[severe path]

• Keep everything in 1 plain text file
• Make sure that file is always over 10MB, even with 1 customer
• Make sure your text files are all in YAML
• Use the slowest YAML parser you can find

If you follow these rules, then you can turn your business into an enterprise.

[hard raptor]

what's wrong with YAML actually.. i never understood that.. it's readable

[quick condor]

chromadb on top Postgresql for me

[nocturne olive]

YAML is perfectly fine

[hard raptor]

i know YAML parsers aren't always the fastest tho lmao

[muted knot]

Yeah YAML is slower

[rough bloom]

parsing weirdness, too complicated, too many ways to express the same thing

[severe path]

It's got edge cases when it comes to type checking

A lot of people (including myself), like to harp about this

[hard raptor]

i usually use a little thing that like loads the yaml file.. converts it to JSON so that compooter is happier about parsing it and i cache it

[quick condor]

I use ymal for some specific processes where its only being checked once or twice a session but for anything constantly being called/writen it goes in db

[hard raptor]

i mean.. i guess TOML is alright too

[dim mist]

What’s the best option for making an AI?

I am planning on creating an AI for a certification from my school but I don’t know what to use.

I know “how” I just don’t know how

[rough bloom]

toml good
way simpler

[rough bloom]

what kind of "AI" certification? actual machine learning? prompting? ChatGPT wrapper?

[severe path]

What do you mean?
Do you simply not know where to start? Do you not know how to architect a model?

[trim valve]

https://ruudvanasseldonk.com/2023/01/11/the-yaml-document-from-hell

"perfectly" is a bit of a stretch

[nocturne olive]

I don't know, works fine for NS

[hard raptor]

oh my.. time to implement feature switches in my yaml parser lmao

[dim mist]

My Schools requirements for the certification is an AI that can respond to questions that would be able to be solved with a calculator, along with basic English questions and answers

What I will get for it is an official certification that more or less states “I can create AI and sell/use it on a business scale and I’m trustworthy”

Along with 500$ but that’s a nice side dish

[rough bloom]

sounds like normal LLM training then
ChatGPT wrapper would be too easy (+ overkill) for that

[dim mist]

I have the basics, I just don’t know what to do to fully create the model, I can code it fine, but what I use to code it is the issue.

Python? C#? no clue lol

[potent geyser]

Now you need to figure out a way to pocket this $500

[rough bloom]

Python

[visual pawn]

Python

[rough bloom]

pretty much the whole ML ecosystem is in Python (+ some parts in CUDA and C++ but you always use those through Python)

[severe path]

If you know how to model an AI, the language you use shouldn't matter.

[visual pawn]

unless you're working with the inner guts of the libraries and stuff, AI work is gonna be in python

[severe path]

You can accomplish any task, with any general purpose language.

[visual pawn]

ok but almost all the libraries, assets, support, etc. is in Python so

[dim mist]

So there is not a specific language I should use? No “best” one?

[severe path]

If you want to use a GPU, there are special languages for that.

[opaque sigil]

yaml

[potent geyser]

Can you just create an ollama wrapper?

[dim mist]

I have certified myself in python, C# and C++

[severe path]

In my opinion, absolutely not.

Although Python is the easiest.

[dim mist]

Alright, thanks yall

[visual pawn]

I'm kinda curious, are they asking you to have an LLM that can do math like a calculator, an "AI" in the more broad sense so it doesn't have to be machine learning, or an AI model trained just for that?

[opaque sigil]

Tile IR

[visual pawn]

could you just, you know, have a calculator program that the AI uses?

I say this because AI can mean very different things in different contexts

[dim mist]

They want an AI that they can use to put on a custom PC in the lobby that can answer minimum questions about the school along with helping with math for school computers.

They want a reason to block ChatGPT

By getting a less advanced(?) ChatGPT

[quick condor]

if you want to get real spicy you can start with just a core model loader and build everything around it (like me)

[visual pawn]

so you're completely free to implement this however you want so long as you're not using a cloud-based model

[potent geyser]

Sounds like they are asking you to run a model locally

[dim mist]

That’s what I had in mind also

[quick condor]

sounds like something a 3/8/12b model could do

[potent geyser]

never over complicate

[trim valve]

do you get paid for each neurosynth mention

[potent geyser]

https://ollama.com/

[visual pawn]

well, I feel like you have to nail down the specifications for what "minimal questions" mean

[potent geyser]

and pocket $500

[visual pawn]

this theoretically doesn't have to be a language model at all depending on the specifications

[severe path]

That's probably the best option

[dim mist]

Oh, damn.

I’ll test this out

[nocturne olive]

No
If I did I would have the money to get a whole new PC

[potent geyser]

https://github.com/ollama/ollama-python

[trim valve]

at the rate you do it, it'd be closer to buying a house

[jovial remnant]

seems kinda counter productive for school to force local (thus worse) models that will give worse help with stuff like math. Unless they just don't care and any excuse is a good one

[potent geyser]

I guess you probably can make an UI and add in additional guard rail

something like that

[nocturne olive]

Well I'd buy a big server instead but sure

[visual pawn]

I think relying on the language model to do the math is probably already out

[dim mist]

Now I was not planning on doing it but they said if I can create a “voice input” it’s an extra 1000$ (deadline is two months from my next Monday)

I thought that would be too annoying for me, but maybe not

[potent geyser]

remember this little guy

[severe path]

I think turning it into an agent, would be the best idea.

Make the model pipe your question into an actual calculator, of some sort.

[visual pawn]

yeah

[nocturne olive]

For math there is this thing called a calculator, you may not have heard of it, but it's pretty useful

[severe path]

Although admittedly, I would do away with AI entirely if you are going to ban ChatGPT.

I don't know what the school is thinking.

[visual pawn]

yeah that's what I was saying lol

but I feel like you need to know what kinds of questions you're expected to answer

when you say minimal, that sounds like "this doesn't need to be a language model" to me

[dim mist]

Not my job to say what they want, I just like money lol

[dim mist]

9th - 12th grade questions

[potent geyser]

High school project?

[severe path]

That's fair

But just maybe you should ask your school board, and perhaps your state DOEd, and perhaps your therapist as well,"why?"

[dim mist]

Probably to track the info in and out ngl

[potent geyser]

Just deploy deepseek and call it a day

[quick condor]

VTT isnt hard just use whisper

[visual pawn]

also I don't think a voice input will be all that much more effort

for 200% more $

[dim mist]

With GPT you can’t, but with a custom local AI you can track questions that are asked

Probably to stop cheaters or find them faster

[quick condor]

yah Id do it

[potent geyser]

you know what

[rough bloom]

so you do just want a ChatGPT wrapper

[visual pawn]

so they want a mini chatGPT that can't be used to cheat....but still be used to...do the school assignments?

[potent geyser]

deploy a 7b model

and cheater will get wrong anwser for sure lol

[dim mist]

money is money, I’m not gonna ask

Might ask after I get the money though

[quick condor]

is there a decent 8-12B reasoning model?

[dim mist]

Maybe

[visual pawn]

define decent

[dim mist]

Define; define

[severe path]

I need a model... but not that model. I need a model, that isn't that model.

• Why?

SHADDUP!

[visual pawn]

Qwen3 8b?

[quick condor]

Im not really a qwen fan, Ministral can be ok for small tasks but I wouldnt reason on it

[potent geyser]

since it's free, just trail and error

might as well throw a bunch of problem at each model and compare

[severe path]

I'm surprised AI companies hadn't added a feature, which could allow schools to track AI queries.

[visual pawn]

there's no way there isn't a benchmark out there for exactly this

[potent geyser]

but I think as a project, it should generate enough paper trail

[quick condor]

I will say I realized that if the current best guess for neuro and evil is llama that having evil sing Metamachine is a special kind of irony

[visual pawn]

I think part of the problem is also that education hasn't really caught up and decided what exactly counts as cheating and what counts as using LLMs productively

[potent geyser]

They barely caught up to internet

[severe path]

I don't think schools even understand what cheating is. They can't fight it, anyways.

[potent geyser]

and I don't even think they care

[visual pawn]

I know they do care, well depending on the "they" is

they just have no idea what to do

[quick condor]

I mean I remember I got away with a ti nspire in school to do some fun stuff

that as well as early smart watches

[keen hatch]

Yesterday I compared inference performance in WSL2 and Linux, and it turned out that Linux is 30% faster for all the MoE models I tested, unless the model fits into GPU memory, in which case there is no difference...

6 vs 9 t/s for Kimi K2 UD-Q2_K_XL

[quick condor]

Well depends on if it is loading from the wsl2 vhd or windows fs. MoEs are kinda a weird case in general but dense is usually solid. I dont offload though so I cant say too much

[nocturne olive]

Well of course when the model runs entirely on GPU kernels it doesn't matter what's dispatching the kernels

And when running on CPU there's essentially VM overhead

[keen hatch]

All my models are located on separate SSDs directly mounted in WSL. I have a threadripper 5955wx + 512GB RAM ddr4 3200 + 5090 + 4070 super ti (16gb) + 4060 ti (16gb)

[quick condor]

Pretty close to my config but i9 10900x, + 128Gb ddr4 + 3090TI (24Gb) + 3090TI (24Gb) + A4000 (16Gb)
Models are stored on a windows volume but cached on the wsl2 NVMe

Neuro's test is ez but then again Ive kinda been deep into this

[jovial remnant]

Offloading is miserable speed compared to fully loading in gpu. I had cuda for whatever reason at one point reserving the entire gpu memory for a model that didn't need it, and it thought it needed more so it decided offloading automatically would be best solution, even tho it wasn't even using half of reserved vram to actually run the model. T/s hit was still huge

[stark needle]

bro has a quantum computer

[olive sable]

[quick condor]

current performance Im getting with 6bpw hb8 with 32k context

[jovial remnant]

Is this with everything loaded on one gpu or split? Tho having direct comparison like this next to each other between few different models looks nice.

[trim valve]

I should really host stuff on my domain again

[odd solstice]

if you wanna do smth other than just running an llm you could try using unsloth on google colab to finetune one, there are plenty of datasets on huggingface that could probably make the model better at what it needs to do
i reccomend olmo3-7b or any of the llama models since they are very easy to finetune (they don't require anything special)
all you really need to do is look on huggingface for a dataset and plug that into unsloth, do a 1-epoch finetuning run, then upload the model to huggingface and convert to gguf with the gguf-my-repo huggingface space or llama.cpp locally
you could also make your own dataset but that can be difficult, if it needs to answer questions about the school you could probably make do with using rag or just keeping info about the school in the model's system prompt

[dim mist]

Damn. Thanks for the info

[azure lynx]

watching the stream makes me wonder if she really does have a thinking layer. ;/ she's not very good at keeping secrets about the multichoice options.

[quick condor]

Thinking is pretty slow

[quick condor]

These each have an independant gpu they are loaded on

[jovial remnant]

ooh that's fun

[keen hatch]

think-while-act relatively fast but it is significantly more complex

[jovial remnant]

Any specific use you have for running 3 at once?

[quick condor]

They have different uses, thinking is gated to only run when needed and is isolated from the user behind the instruct model. gate handles side processes

[stark needle]

qwen 3 8b or similar

the 4b 0528 or whatever is trained better tho

[jovial remnant]

So it's all running under one system of sorts? Sounds nice

[quick condor]

IDK I kinda want to see how ministal-3 3/8B reasoning is but Im sticking to small for reasoning rn

[odd solstice]

well didn't crelly get access to her backend which showed some sort of internal thoughts?

[odd solstice]

the qwen3 vl 8b model is trained the same way as the recent 4b models and has a reasoning version, but i haven't tried it

[stark needle]

isnt qwen 3 vl just qwen 3 with a vit

[quick condor]

I have a feeling that other than just backend performance tweaks, a lot of Neuros latency can be hidden with varried stream delays

[odd solstice]

nope, its really different

[stark needle]

oh so more rl i guess

but yea unsloth is goated but i dont think it would be necessary

[odd solstice]

it also doesnt have hybrid reasoning like the og qwen3 models

[stark needle]

ooh ok yea fair enough

i did something funny with qwen 3 14b i distilled the shit out of gemini's reasoning traces into it lol

[odd solstice]

what was the end result like?

[stark needle]

https://huggingface.co/shadowlilac/OpenGemini-Flash-RLVR-Q4_K_M-GGUF

i also grpo'd it

[quick condor]

Im not a huge fan of hybrid... Id rather instruct, Instruct with a basic forced reasoning method, reasoning model. Using a reasoning capable model for everything adds ttft latency and slower t/s

[stark needle]

it's based on 300k reasoning traces

[odd solstice]

did you collect all of them or did someone make a dataset? that seems like it would cost a lotta money (idk gemini api pricing tho)

[stark needle]

i found someone on huggingface lol

absolute goat

[odd solstice]

i once tried to make a gemma2 deepseek reasoning finetune (was like my third ever finetune) and it just didnt work at all and the model got like 100 downloads for some reason before i privated it

i think i did like 30 steps on the dataset, but i dont really remember, so that's probably why it didnt work

[stark needle]

oh yeah theres bots that download all models uploaded

[odd solstice]

somehow i get downloads on my private-from-creation models, idk how

[azure lynx]

yeah, i guess she's just not too good at choosing what to send to speech and what to just think

[odd solstice]

i've been gating them too just in case and that stopped the downloads tho

[stark needle]

im currently training a qwen 3 1.7b gemini distil on grpo

this one actually trains quickly

cause on the a100 generation bandwidth is high

[quick condor]

Why train 1.7 on cloud

[odd solstice]

that's cool, i havent tried grpo lol

[stark needle]

grpo is basically you roll n generations and then choose the best answers in an automated manner using a reward function

[stark needle]

why not?

[odd solstice]

im probably gonna try dpo with my models, cuz sometimes the responses are really good and other times it takes a few retrys to get a decent response

[stark needle]

isnt that with the reward model?

oh no wait thats ppo i think

[odd solstice]

direct preference optimization

i think all i need to do is collect the worst and best responses

[stark needle]

yeah, u can do it with an autorater

[odd solstice]

but i need to rework my discord bot's memory system to be able to regenerate messages....

[stark needle]

use another llm to rank generations with vllm or whatever