programming

[kind nimbus]

but that is running in a linux vm

[silent cloak]

Tf is a zigbee

[kind nimbus]

zigbee is awesome

[silent cloak]

Is it related to zig

[kind nimbus]

no

[cobalt loom]

a bee that goes zig then zag

[kind nimbus]

https://www.reddit.com/r/homeassistant/comments/xlc2jt/help_me_understand_zigbee/

i replaced all the wifi iot stuff with zigbee

its so much better

im just using a Pi as a bridge that runs Zigbee2MQTT in a docker container

which connects to the home assistant server to the MQTT broker

[silent cloak]

Seems obscure

[kind nimbus]

no its not

its actually simple

and works so much easier and more reliable compared to tasmota & esphome

[kind nimbus]

another benefit is the batteries in my smart home devices now work for several years

i never had to replace a single battery in years

[silent cloak]

O its just a cheaper zwave

[kind nimbus]

uh its similar to zwave

[cobalt loom]

oh yeah i forgot i have a pi running pi hole and some other bs i prolly set up years ago and never checked on again

[silent cloak]

Pihole is based

I use bind9 instead tho

[kind nimbus]

lol

i currently use Blocky to block network level ads and resolve dns

it's go based and should be faster than pihole or adguard home

i was using unbound before but realized i dont even need it lol

[silent cloak]

Shouldn't be

Pihole uses dnsmasq which is in C

The web server is php tho I think

[kind nimbus]

idk for me blocky somehow feels faster in name resolution

i never actually measured it tho

[silent cloak]

I just use bind9 so I have more control and I can do some other dns optimizations

[kind nimbus]

i actually just removed unbound like 2 days ago because i figured i dont actually need a fully recursive dns server

[cobalt loom]

yeah i basically just use it as a dns for my own devices cause i dont wanna affect the rest of my fam with dns blocking

[silent cloak]

I like to cache stuff and do some other tricks

[kind nimbus]

blocky can also do all that

[silent cloak]

If it has good privacy tools I'll look into it

And if it has good rep naturally

[kind nimbus]

https://0xerr0r.github.io/blocky/latest/

[silent cloak]

Though i am hesitant to run stuff that isn't completely low level as I REALLY need speed

And as fast as go is it ain't C

[kind nimbus]

idk to me blocky is faster than unbound

[silent cloak]

I haven't used unbound directly

But ive used remakes of its tech before

[kind nimbus]

i never really used BIND. Only unbound.

[true hemlock]

:3

[fast pagoda]

[tropic spindle]

Why C when you can use assembly

[cobalt loom]

speakin of servers/services heres what im running now. media server has jellyfin and audiobookshelf.
since i got time do yall have any suggestions on what would be fun to setup?
also if you guys like live tv i recomend checking out tunarr or stuff similair. U can use jellyfin/plex media and basically setup your own iptv station and add custom commercials (sadly rn its only between episodes)

[mighty thorn]

Why use assembly when you can use 1928 adding machine

[silent cloak]

Time to write my own authoritative DNS server in ASM

[tropic spindle]

It's probably been done in ye olden days

[silent cloak]

Why use OMV

[kind nimbus]

*arr stack, plex / jellyfin, home-assistant

[silent cloak]

Seems redundant if ur running a arrsuite

[kind nimbus]

who needs a nas os anyway

all you need is storage, zfs, nfs / smb

[silent cloak]

Thats why I just use proxmox

[kind nimbus]

maybe iscsi

[cobalt loom]

bro imma be completely real with you. i dont remember. this was years ago. but prolly just cause that was what was recomended and it seemed the easiesty

[silent cloak]

I think i looked at it once and said na

U could replace it with samba or something

[kind nimbus]

yea or just nfs

windows also supports nfs since win10 i think

[silent cloak]

I dont think u can secure nfs

[kind nimbus]

lol the memory usage btop is telling me cannot be real

[kind nimbus]

you can

e.g. via Kerberos

[silent cloak]

Yeah so an external tool

[kind nimbus]

native also works via uid / gid mapping although that is not really secure

[silent cloak]

Yeah ill use samba still for proper network drive work

Nfs shines more for local

[kind nimbus]

i mean yea

[mighty thorn]

-# ai psychosis thread is progressing…

[cobalt loom]

ive thought of arr stack and home assistant, but i dont really need or want any smart home stuff, and i tried setting up an arr stack but it didnt seem worth the effort in my use case( adding a show or anime like once a month)

[kind nimbus]

nfs is much faster but only for local trusted networks usually

[silent cloak]

Yeah

[kind nimbus]

i trust my network so i use nfs

[silent cloak]

Its a case of less features for faster local use

More security and less speed for more open uses

[kind nimbus]

pretty much

[silent cloak]

Is there anything faster than nfs?

Haven't looked lately

[kind nimbus]

but i also use it e.g. to stream movies from my plex vm which mounts the nfs share

so speed is a concern

[silent cloak]

Thats what im doing too

[kind nimbus]

not really

and iSCSI is block storage for vms e.g.

[cobalt loom]

for omv I should just remove it and replace with just a service? and what if i add more hdds in the future and want to do a raid/zfs configuration

[kind nimbus]

just configure your zfs via cli

[silent cloak]

Proxmox has tools for that

Also u could literally run truenas in a container

[kind nimbus]

or use this:
https://github.com/webzfs/webzfs

[silent cloak]

Which would still be preferred to omv

[kind nimbus]

once i fully migrate my zfs pool into the main server i will test out webzfs

[cobalt loom]

oh cool. is omv that bad? lmao

[silent cloak]

I mean its fine as its own thing

For an arrsuite its just kind of deprecated

[kind nimbus]

truenas scale is definetly much better

[cobalt loom]

ohhhh

[kind nimbus]

oh they renamed TrueNAS Scale to TrueNAS Community

[silent cloak]

Yeah

[kind nimbus]

https://www.truenas.com/truenas-community-edition/

[stark needle]

Chat new use case found for dgx apark I'll just run deepseek v4

[silent cloak]

Still more free than unraid

[kind nimbus]

for an open source nas that just works with minimal effort truenas scale is probably as good as it gets

[silent cloak]

Real

Its good to have storage setup and mostly dealt with so u can focus on the suite itself

[kind nimbus]

yea

i currently just mount the zfs raid from the old server which runs TrueNAS core over the network onto the new server running FreeBSD lol

one day im gonna replace the raid controller with a flashed one so i can finally migrate the zfs array into the new server

mfw i have an entire server running just because im too lazy to replace the RAID controller

[silent cloak]

Classic compsci method

If its not broken dont fix it

[kind nimbus]

i could literally import the zfs pool with one command lmao

[silent cloak]

[kind nimbus]

yea pretty much

[sick shuttle]

[cobalt loom]

do u guys use any of the arrsuite urselves? is it actually worth setting up, cause i tried with sonarr and i didnt get that far before it didnt seem worth it

[kind nimbus]

i use Sonarr, Radarr, Prowlarr, Sabnzbd, etc

[kind nimbus]

use trashguides for setting it up correctly

[tender river]

ako how did you get here

[cobalt loom]

bro because you know if u touch it. the "not broken" parts will completely fall apart and u have a day of troubleshooting in your future

[cobalt loom]

aight ill look into it

[silent cloak]

Sonarr radarr cloudflare-resolver qbittorrent prowlarr etc

[kind nimbus]

sometimes its good to replace stuff that was not broken lmao

[silent cloak]

Some other microservices for Metadata management

[kind nimbus]

kometa

[silent cloak]

Yup

I forgot they rebranded

[kind nimbus]

all of that stuff has packages on freebsd i love it

[sick shuttle]

[kind nimbus]

Tautulli is nice if you use plex

[silent cloak]

Got it

I barely touch it tho

[kind nimbus]

lmao yea

[silent cloak]

Its kind of just a middleman for giving information

[kind nimbus]

i mostly use it to monitor transcoding and bandwidth

[silent cloak]

Gives me all that on homepage

[kind nimbus]

unpackerr

also good

[silent cloak]

Haven't needed it ngl

Same with the one for subtitles

[kind nimbus]

i sometimes needed it when sabnzbd did not unpack archives on its on properly lol

[silent cloak]

None of my media is archives

Beyond the data i personally archive

[kind nimbus]

you mean Bazarr ?

[silent cloak]

Yeah

I tried it once then I realized it was mostly redundant

[kind nimbus]

a lot of stuff from the seven seas is packed and password protected

[silent cloak]

Not where I sail

[kind nimbus]

huh

[silent cloak]

Nothing I get is password protected

Unless its on the clearnet

[kind nimbus]

mostly everything on usenet is password protected

[silent cloak]

Yeah I haven't used usenet in like 2 decades

[kind nimbus]

damn

[silent cloak]

Have no reason to

Plus the privacy of torrents is more appealing to me

[kind nimbus]

hearing privacy and torrents in the same sentence is rare

[silent cloak]

I mean VPNs

[kind nimbus]

well yea ok or a anon seedbox

[olive sable]

[opaque wharf]

Fire hazard 101

[sick owl]

You know I completely forgot meta owns manus now

You'd think they'd make more of a song and dance about their one useful AI product

[cobalt loom]

gotta love the safety slippers

[opaque wharf]

And it is optional!

[sick owl]

Lucas from Resident Evil 7 vibes

Anyone else enjoy the brrrrrrrr noise your gpu makes in ML tasks

Makes your decidedly digital machine feel kind of analog for a bit

[young plover]

VRAM coil whine
it's obnoxiously loud on my 3090

[kind nimbus]

It's 9 am. Claude just told me to go to sleep

[olive sable]

based claude

[opaque wharf]

Engine running again

Why must modern cars have to use a computer in the first place

[sacred nimbus]

If Ellie wont make the cfrb i will

[kind nimbus]

Yeah Claude is right

Time to get some sleep

[kind nimbus]

German engineering

[opaque wharf]

Imma kneecap the German engineer

[opaque wharf]

@opaque sigil come here

[vague sonnet]

better control of ignition timings and shit

[opaque wharf]

I know. Just hate the fact that they tied it with serialized part

[sick owl]

Gemma 4 E2B is surprisingly useful to have on hand

Sure its braindead but its just stable enough to actually have utility as a mini on device assistant for lower power hardware

[vague sonnet]

I like my assistants smarter than braindead

[kind nimbus]

I think my rank went up

[sick owl]

This is more the "intern you send to the photocopier" type of assistant

[proven merlin]

Idk, for that i prefer the small qwens3.5 (2-4B), you can even get decent results with qwen 3.5 0.8B, but reasoning off and "short answers" syspromt.. gemma 4 is super cool for long context, but its kinda useless (from my experience)

[velvet vale]

ive been using 31b for its tool use but i might drop down to 4

or just the default latest , which is 8b

[stark needle]

Today im gonna try deepseek v4 locally

Lmao

It's only 280B fortunately

[sage crag]

[sick owl]

Unironically quicker this way

[sage crag]

ye its quicker to use the browser

[sick owl]

No it was quicker to use gemma

2 billion active parameters is fast

[sage crag]

[sick owl]

Also I get the added benefit of not providing extra traffic to corpos

[trim valve]

wow incredible I can do the same thing as my calculator app, but using more power and sometimes wrong

[sick owl]

You can't just leave your calculator app in the background and ask it questions on the fly while doing other stuff

[sage crag]

its a 2b model

it cant either

[sick owl]

And a calculator app only does one thing, this also does OCR, surprisingly good translation, web search

[sage crag]

[sick owl]

Yes you can actually

It has native audio input

Its a fairly novel architecture

[trim valve]

ok??? it calculates instantly

[sick owl]

So does this, its 2 billion parameters

[trim valve]

I don't have to submit a ticket and wait for triage to figure out the temperature

[sage crag]

i get my temperatures from figma

i put the question in the document and the colleague edits it later

[sick owl]

There is no waiting, that's the point

You make it sound like a process when its literally just asking my pc a question out loud and a pop up gives me the answer

[trim valve]

i still don't get the appeal

[sick owl]

What's not to like?

[trim valve]

calculating temperature has been a solved problem for years

[sick owl]

Correct

[trim valve]

I've been able to use google assistant for that for the last like, 6 years?#

[sick owl]

Also true

[trim valve]

and its always been right

[sick owl]

Google assistant is also similarly useful

Its just way more limited in its functionality

[trim valve]

now I have the option of instead having it sometimes lie

yknow maybe I should try this myself, plot a graph and see what it thinks about temperature

[sick owl]

I mean hell feel free, its a tiny download at Q8 or even BF16

[sick owl]

Also google assistant lied too, it yoinked its results from the web and regurgitated them without even attempting to interpret context

On account of, y'know, not having context

I mean don't get me wrong this isn't a miracle model, it is braindead

But its decent enough for quick web retrieval, currency, temperature, timezone etc. conversion, ocr, image classification

Its a nice utility to have on a low power device

[rare bramble]

idk if I would use LLMs for temperature conversions but they are quite handy in lots of similar tasks that are somewhat manual, but pretty tedious to do otherwise

[sick owl]

I have it on in the background in this laptop so its quicker than googling

Obviously its not computationally elegant or unique and it does fuck up, its the broad scope of the simple tasks it can handle that makes it cool

The E4B is even better in that regard if you have an NPU or resource constrained dedicated GPU

[sick owl]

What kind of graph were you thinking of plotting? I can run it locally and save you the trouble of a download for fun

I do like pushing these dumb models

Don't get me wrong btw, I wouldn't use it for anything other than the simplest tasks

[trim valve]

I mean, feel free
I'd basically be a scatter graph of farenheight on the x axis and outputed celsius on the y axis. repeated multiple times with a bit of temperature so there is the option for multiple answers

I'd happily do it on my pc though instead because it sounds like a thousand or so prompts per temperature setting, and I'd probably also test it at different model temperatures

[sick owl]

Well I just gave it your question without any context or changed wording out of pure curiosity

It generated three graphs with matplotlib

[trim valve]

interesting, though it seems to have interpreted model temperature as temperature in farenheit

[sick owl]

It just had the one message for context so that's about what you'd expect

[proven merlin]

Trust me, on most hardware really isnt. A fairly nice ryzen 7 5700g, cpu only does 11t/s. With igpu 26t/s. And it uses way way more ram. What these 2B models ARE good for, is having an approximate knowledge of all humanity within like 1-2GB of data. Way better than Wikipedia's downloads.

[sick owl]

26t/s is more or less instant without reasoning turned on for the kinds of tasks you'd actually be using these for

[proven merlin]

If you think about that, it's insane compression

[sick owl]

I wouldn't use a model this small for its internal knowledge though, gonna have to disagree there

Its best used to retrieve knowledge from elsewhere if you're asking it questions

[proven merlin]

Eh. Until the model is loaded and all, it really isn't. Even assuming you keep it loaded 24/7 its fairly slow, at like 80t/s then sure.

[proven merlin]

Okay, fair, 2B is the lower end, but 4-9B is impressive there.

I mean you can get fairly ok responses even with 0.8B as said before

And yes, better to couple it with any sort of rag, but yea.

[sick owl]

I really wouldn't use a 0.8B model day to day and I'm already relatively forgiving for using E2B

[proven merlin]

Well gemma4 is particularly bad with this from my experience

Gemma4 really depends on tool calls

But imo also struggles with reasoning/using the tools in a efficient way

As example, asking it to read what files are on your desktop in windows

Either it fails, or reads User/Public

Instead of checking C:/users

What i do have to give gemma4, is speed

[sick owl]

Gemma 4 is actually very good at tool calls

But it launched with a broken tool calling template

If you haven't updated your quant and llama.cpp since the period around its launch its probably misconfigured

[rare bridge]

[rigid snow]

this is a new one

[leaden crest]

are they still going

[rigid snow]

huh

[leaden crest]

the schizoer

[rigid snow]

this is 5.3 codex

[leaden crest]

ah nvm schizoer left the server

thread’s still there though

[rigid snow]

[leaden crest]

they never answered the integral question

what are the odds they knew what an integral is

[fickle rain]

Why not use 5.5

[rigid snow]

i sure wonder why

and 5.3-codex is good enough

[gentle haven]

Is this good connection

[true hemlock]

yes

Toast™ Certified

[gentle haven]

Top is DL and below is UL

[fickle rain]

Bro just give OpenAI more money bro

[rigid snow]

i subscribed through the ios app to make sure they only get $16 out of $20 i pay

[vague sonnet]

give none

[gentle haven]

Can’t give money if you have none

[true hemlock]

damn people pay for these shit

[rigid snow]

it makes you so much faster

i already got $20 worth of work done with like 3% of the montly limits

and the code isn't that bad

[true hemlock]

that's what microsoft vibe coders said

[rigid snow]

yeah but the difference is

i actually look at it

[vague sonnet]

anyone that says that isn't good enough at programming to evaluate code properly

[rigid snow]

i'm not saying it doesn't do bullshit

^

i meant code that i actually commit not all the code it produces

[vague sonnet]

the problem is LLM automation
if you used one model for searching through the codebase/specs and one for smarter autocomplete suggestions
it would actually be good
but the moment you use a LLM to write the code for you the quality drops are crazy

[rigid snow]

it's not worth as much as a competent engineer i wouldn't expect it to produce code a competent engineer would

[vague sonnet]

and like
if I'm gonna waste time going through code that 90% goes in the trash and I have to write it myself
it's better to just save time wasted on reading and write it all myself

[vague sonnet]

so you're agreeing it's not competent at writing code
which means it's a dysfunctional tool lol

[rigid snow]

i'm saying it's not great at it but it can write code, which means it will sometimes get it right

[vague sonnet]

gambling

[rigid snow]

sure

[vague sonnet]

it's only good for simple tasks that also don't require high code quality
in other words throwaway weekend projects

[rigid snow]

it’s AWESOME for those actually

[vague sonnet]

yeah that's one thing I can agree with

[rigid snow]

you don’t have to care at all about the code

bliss

[vague sonnet]

it's hard to fuck up and you're not gonna extend it
so the flaws aren't as relevant

and since it's small it's easy to just prompt a rewrite if something does go wrong

[rigid snow]

but also i sometimes wonder
models get better which means this slop it produces is potentially more maintainable by llms in the future

[vague sonnet]

I disagree about them getting better

whether they are increasing quantization or it's just noise from degrading datasets
but the response quality keeps going down, my prompts get ignored etc

[rigid snow]

claude user?

[vague sonnet]

nah I tried multiple free models

chatgpt, claude etc

it's a pattern for all of them

[rigid snow]

minimax 2.5 is free on openrouter rn it's pretty good for what it is

[vague sonnet]

I don't really have uses for LLMs nowadays

my weekend projects usually do something funky it's not gonna get right

[rigid snow]

you don't have to vibecode the whole way through them

give it the boring parts

[vague sonnet]

I mean, I prefer getting in more practice with code structure

it's not a relevant time investment

like I wrote a full custom yaml parser into dialogue system thing in like 3 evenings
I could maybe make it 2 but have it less modular or something

[sick owl]

Did they get bullied away

I feel bad for them now

[glass flower]

can confirm. i've been trying to use GPT 5.4 with my copilot subscription to do basic bug fixing and features... its a coinflip if it even produces something working and another coinflip if its even close to something i would want to maintain, a lot of the times it just looks bad code-wise. needless fragmentation of functionality or just code duplication.
The best way i found to use it is error finding and suggestions. or very very directed plans, like basically say the way the code should work in natural language while incoperating as much existing stuff into it as possible.

[glass flower]

that thread reminded me off that one github repository that simulates the entire universe in realtime. its like 100s of files with a bunch of techno words like quantum thrown into it. and in the end its just a bunch of if's and a few random number generators

[vague sonnet]

ok but like
AI-based search engine + autocomplete
and human at the steering wheel
then in theory 99% of the time it would be human error when something's wrong

[umbral wigeon]

Using Ai is like gambling

[vague sonnet]

gambling is fun, using LLMs is frustrating

[glass flower]

i agree that would be good

[umbral wigeon]

100% honest as someone who praise ai and using ai and also coding for 3 years

[glass flower]

its frustrating because it looks like it got there 90% of the way but then you realize it fumbled after the first 2 minutes and stumbled into chaos

i should try out https://github.com/ThePrimeagen/99 again but this time with my github copilot subscription instead of a local model

its like ai search in your codebase

[vague sonnet]

what's the approach in this one? cool way is to use contextual embedding and match those between prompt and labeled codebase

[valid onyx]

do you guys know if neuro-sama is a custom ai model? or a pretrained model thats been modified to be like how neuro-sama

i need to know for my kai model

[vague sonnet]

I don't think we have the details on that
but I'm 99% sure it's just a finetune with agentic features

[rare bramble]

we dont know what model neurosama is based on, but we are quite confident that it's a heavily customized and refined opensource model

[glass flower]

you tell it what to find. and then it gives you a quickfix list of all the different interesting things with descriptions why it would be interesting

[valid onyx]

got it.

[vague sonnet]

I meant under the hood

[glass flower]

the way it finds things is just basic tool usage. like grep and looking through the codebase. it uses normal providers like opencode and such

[vague sonnet]

is it ai based or not?

[glass flower]

yes. the AI looks through your codebase

[vague sonnet]

that's so inefficient bleh

[glass flower]

it isn't embedding your whole codebase tho. for something like that you would likely want to use vector code

[true hemlock]

this one would be cool. would rather that than vibe coding.

[vague sonnet]

embedding is cool because it's basically instant
prompt to embedding then just run a regular comparison for most likely
and just show stuff past a threshold

[glass flower]

technically you can easily do it in neovim with codecompanion and vectorcode. since they have integrations

i don't really use vectorcode anymore since i switched to copilot. but... i'm sure it works with it too

[vague sonnet]

it doesn't have to be that complicated because the only thing it has above regular search is decoding synonyms or descriptive sentences

it really is just an embedder

[rigid snow]

codex app has this theme named "Absolutely" lmao

[mild gust]

Absolutely nah

[mild gust]

It cost money

[rigid snow]

i'm laughing at how it makes fun of claude i'm not telling you to buy a sub

[mild gust]

[glass flower]

alright so... i gave copilot cli the vectorcode mcp server to play with. and told it to use it and find all the plugins i'm using. its been 5 minutes and it still isn't done. might even be stuck

i feel using the cli directly would have been a more pleasant experience instead of doing it through this chat buffer

its so ass

[umbral wigeon]

What did you tell copliot to do

[glass flower]

this:

what tools do you have access to?

● I'm GitHub Copilot CLI, powered by GPT-5 mini (model ID: gpt-5-mini).

Available tools in this session:

• git — repository operations (commit, branch, diff, etc.)
• curl — HTTP requests
• gh — GitHub CLI (PRs, issues, workflows, etc.)

Tell me what you'd like to do and I’ll run the appropriate commands.

❯ can you use vectorcode? i've added the mcp to your toolbox. So you should be able to query this code base with it and find all the plugins i'm using. Try
it out and tell me what you found

● Listing indexed VectorCode projects first so the proper project_root can be used for plugin queries. I'll call report_intent and list projects in
parallel.

i know gpt-5-mini isn't the best. but i can use it with a 0x multiplier so i wanted to try it with it first

[umbral wigeon]

I think copliot is like some junior dev who have no idea what they're doing

Ai googles just like us

[glass jetty]

again ai

nothing but ai

[umbral wigeon]

We all googling right?

[glass jetty]

only talks about ai here

[umbral wigeon]

Idk other topic than ai

[glass flower]

you are right lately there has been a lot more ai talk

[umbral wigeon]

Maybe change topics

[glass flower]

lets get back to screeps... so how has the game changed with the latest update

[umbral wigeon]

What best game engine?(opinion)

I quit roblox studio last month

Roblox has a lot of down bad and drama

Thinking about going to unity

Or C++

[glass jetty]

havent played
i not see me writing some business code with limited amount of commands

i usually just read/watch on internet, cant even say i code a lot, like my own repos contain less then 10k lines of code for sure

i just finally learning rust again

[umbral wigeon]

https://tenor.com/view/rust-femboy-rust-femboy-programming-rust-programming-gif-27321790

[glass jetty]

string slice is bytes for some reasons btw

[umbral wigeon]

Rewrite it in rust

[glass jetty]

cringe

rust is language that strict better then c
like i have this diagram (i made it like 4 hours ago), and i wonder how i can implement similar in rust
will learn i guess
but i definitely cant implement software stack coroutines in rust, it is c++ only

[glass flower]

godot

[umbral wigeon]

I'm going to C++

[glass flower]

[glass flower]

still godot just write your game as part of the engine

[nocturne olive]

Godot

[glass jetty]

isnt godot have cpp api?

[sage crag]

bets on ai script and voiceover?

[nocturne olive]

GDExtension
It even supports Kotlin

[umbral wigeon]

Does godot have cpp

[sage crag]

yes

[umbral wigeon]

I wanna try it

[nocturne olive]

[umbral wigeon]

Okay switching to godot

[nocturne olive]

[umbral wigeon]

Does it work on mx linux?

[sage crag]

its a japanese video

[nocturne olive]

Godot works on all the OS

*major OS

[umbral wigeon]

Oh godot even supported on a phone

[nocturne olive]

Xoda writes Godot on a tablet sometimes, it's kinda peak how they just have that

[vague vortex]

Temple OS too?

[nocturne olive]

No

[vague vortex]

Aww

[nocturne olive]

TempleOS not major OS

[vague vortex]

[umbral wigeon]

How creator of TempleOs read his code

Also coding in white theme

And no autocomplete or ai

[vague vortex]

No AI
But God was telling him how to build it

So he claims

[glass flower]

he was faithcoding. a superset of vibecoding. he was ahead of his time

[vague vortex]

True

Or, he was a mentally tormented genius. Which... to write your own OS, solo, from absolutely nothing, would require some tremendous mental capacity and mental illness ngl

Either way
It is an impressive feat

Vedal presumably communed with the spirits of Banana Rum to build Neuro from nothing
Or it's just the power of being an alcoholic depressive Brit

-just sayin, every smart person in history that I'm aware of was either insane or eventually went insane because the idiots around them couldn't understand. See also; Paracelsus

[unreal field]

someone pls translate this

[queen ginkgo]

Autism Sama and ADHD Sama, the two sisters!

[obsidian mantle]

how do i extract gif from this

[kind nimbus]

The message decodes to: “hbOshi”

[unreal field]

a little off there

[amber fractal]

Sam has a copy somewhere idk

[unreal field]

actually wait, got the code wrong lmfaoo

here's the corrected code

[glass jetty]

to what?

ascii?

01101000011000100110010001110011011010000110100101110010

01101000011000100110010001110011011010000110100101110010

[sage crag]

bwaascii

[vague vortex]

Bro speaking in morse code via binary

Through heiroglyphs

Enigma wasn't so encrypted

[kind fable]

If we treat your Evil Neuro and Neuro-Sama emojis as binary code—assigning <:evilBwaa:...> as 0 and <:neuroBwaa:...> as 1—it spells out a slightly glitchy message.
Here is the exact translation block by block:

• 01101000 ➔ h
• 01100010 ➔ b
• 0110000 ➔ 0 (Note: This chunk is missing a bit, but 110000 in 7-bit ASCII is the number zero)
• 01110011 ➔ s
• 01101000 ➔ h
• 01101001 ➔ i
• 011100110 ➔ s (Note: This chunk has an extra 0 stuck on the end)
  Final Translation: hb0shis
  It looks like either a corrupted word (a typo in the binary), a random string of text, or a specific invite code/password!
  - Gemini 3.1 Pro

[unreal field]

gemini got it right, it was not the right word

[glass jetty]

hbdshir

[unreal field]

[kind fable]

~~

That is a really smart catch! Combining all the bits into one continuous stream and re-slicing them into perfect 8-bit chunks is exactly how you'd troubleshoot a misaligned data stream.
However, if we shift everything over to fill in that missing bit, our sequence spills out of standard ASCII and into Extended ASCII characters.
Here is what your newly aligned sequence translates to:

• 01101000 ➔ h
• 01100010 ➔ b
• 01100000 ➔ ` (backtick)
• 11100110 ➔ æ
• 11010000 ➔ Ð
• 11010010 ➔ Ò
• 11100110 ➔ æ
  Shifted Translation: hbæÐÒæSo, while the bits are mathematically aligned now, the output became even more cursed! The original spacing—even with the glitchy 7-bit and 9-bit chunks—was likely the intended format trying to spell outhb0shis` or something similar, before the emoji formatting got a little scrambled.
  - Same Sycophantic AI~~
  -# irrelevant

[glass jetty]

guys, do you want me to code bwaascii?

i can do it, but lets at least use it for a bit

will be my fist relatively usable app in rust

[trim valve]

go my 30000 prompts

[glass jetty]

i doubt i have even 10k

starting from 2023

[kind fable]

Ah, that looks much cleaner! Now we have exactly eight bits (emojis) per block, making it perfectly valid ASCII.
Here is the translation with your corrected code (evilBwaa = 0, neuroBwaa = 1):

• 01101000 ➔ h
• 01100010 ➔ b
• 01100100 ➔ d
• 01110011 ➔ s
• 01101000 ➔ h
• 01101001 ➔ i
• 01110010 ➔ r
  Final Translation: hbdshir
  Looks like a "Happy Birthday" message! (Happy Birthday Shir?) The missing and extra bits in the first attempt definitely garbled those middle and end characters, but the corrected spacing fixed it perfectly.
  - Still the same best AI that i have access to

[kind fable]

interesting but too simple

gotta spice it up

need to implement some form of compression

[nocturne olive]

New hacked account

[glass jetty]

oh nooo

such a shame

-clover

[stark needle]

:niuh:

[kind nimbus]

hbdshir

[unreal field]

someone already did that

#gymbwaa-general message

[minor crag]

I just built a mostly correct tiktoken (OpenAI tokenizer) implementation in pure Rust (only using the fancy-regex crate) and it works!!! (Only 2 tokens off out of 90 words)

I am very proud of myself :3

[molten island]

chat i'm cooked

[heavy pine]

https://fixvx.com/allisx86/status/2047647241087660077?s=20

allisx86 (@allisx86)

https://t.co/kQrnTi6ORc

vxTwitter / fixvx 💖 749 🔁 48 · Apr 24, 2026, 12:02 PM

is this true

[jade cliff]

@maiden geyser I was told to ping you if I saw this

[warped narwhal]

I think I have it saved somewhere lemme check

https://cdn.discordapp.com/attachments/1071784467036913664/1451008289394393120/FvcBkoP.gif

[olive sable]

https://cdn.discordapp.com/attachments/1327142377139343430/1486773213190029444/doc_2026-03-25_19-13-27.gif

[glass jetty]

no access for me

[olive sable]

Goodmorning

[jade cliff]

There you go.

[glass jetty]

ip v8 is a thing?

[glass jetty]

activate windows

[amber fractal]

Second ping

[glass jetty]

so many, unusable without nitro tho

[warped narwhal]

Read the spec and you'll very quickly realise that it is fake

[glass jetty]

-# i kind of just asked is it true rn
ok, thx

[jade cliff]

Heh, windows bugged and won't let me

[warped narwhal]

https://www.ietf.org/archive/id/draft-thain-ipv8-00.html

it's a funny read though

[amber fractal]

I summon powershell

[jade cliff]

I've tried.

Stupid Windows update broke activating.

And it won't let me go back.

[glass jetty]

install linux

[amber fractal]

[jade cliff]

I have another laptop with Linux, but I don't have minecraft on it.

[glass jetty]

it is java...

[jade cliff]

And, my Windows laptop makes it easier for me to code.

[warped narwhal]

Put minecraft on it

:D

[glass jetty]

[jade cliff]

That laptop is purposefully designed to be easy for me to reinstall operating systems in case any of my code actually works :P

[warped narwhal]

[jade cliff]

C#, visual studio

[glass jetty]

how it related to minecraft?

[warped narwhal]

[jade cliff]

It's not, it's a completely different thing lol.

[jade cliff]

[glass jetty]

try vscode, and then maybe nvim

[jade cliff]

C# is the easiest for me at the moment, I'm slowly learning java and python

[glass jetty]

how it relates?

[jade cliff]

VSCode doesn't like me

[glass jetty]

its you not like vscode

tbh even while this thing bought by microsoft is slow/// it is infinitely better then VS

[jade cliff]

Yes well, I'm stubborn

[amber fractal]

Everyone in this chat is stubborn so fair

Just a part of being a programmer

But also moreso Linux users

[glass jetty]

systemd

[amber fractal]

I make no effort to fall for this bait

[jade cliff]

[reef ingot]

Day probly 7 of making my own game engine like it's the 90s

I got the right microcontroller and got vga and a bit of his there

[glass jetty]

you are really cooking something

[kind nimbus]

Use NixOS flakes so you don't lose any of your setup when you reinstall

[glass jetty]

use arch and remember how you setup your system

[kind nimbus]

Nah my ADHD doesn't allow that

[jade cliff]

Me either

[kind nimbus]

I need everything cleanly committed in git

[glass jetty]

i think i forgot how to setup grab
i definitely forgot that

there 2 commands, one to make and one to install

• set target and dirs

[kind nimbus]

grab

I recently switched to systemd-boot

https://tenor.com/view/nixos-nixos-rebuild-flake-gif-17654708210828148959

--install-bootloader

[obsidian mantle]

nooo why is it not smooth why 2 frames

[obsidian mantle]

its small and i cant repost it i need gif

[jade cliff]

https://media.discordapp.net/stickers/1451004013683277974.gif

[obsidian mantle]

literally this

[rigid snow]

audhd

[obsidian mantle]

so nice

[rigid snow]

not really

the autism gif lags :(

[obsidian mantle]

and now we get back ontopic of #programming discussing how to fix it

what a turn

[rigid snow]

@claude pls write script to generate this gif

[opaque wharf]

Make the adhd gif lag too

[maiden geyser]

self-ragebait situation is insane

[sage crag]

https://cdn.discordapp.com/attachments/1384918024708034750/1487235837019029584/EvilADHD.gif

[jade cliff]

Whuh

[maiden geyser]

i put it there months ago to check if a subless one can do seeable things on the cerber, now can't remove because can't log into microsoft

[stark needle]

hi

is this u

[sage crag]

hi

this is evil

[reef ingot]

Yo

Day 7 of making my own game engine today was mostly hardwarey stuff

[glass jetty]

what is this cortisol joke

[olive sable]

theatre pc wouldnt boot yesterday

bios update fixed it

i apparently have terrible luck with getting motherboards on a bios version recent enough for the cpu im using

[reef ingot]

[olive sable]

at least i could flash the bios without needing another cpu

[sage crag]

cortisol is a stress hormone

[olive sable]

high cortisol meme on tiktok rn

[sage crag]

ohisee

[opaque wharf]

BMW do be making my cortisol level raise

[glass jetty]

but why it forced everyhwere

[glass jetty]

so it just weird popularity spike?

[olive sable]

ye kinda

[umbral wigeon]

Can't we speak in hex

[sage crag]

65 is a or A i dont remember, in hex thats 0x41

probably

[umbral wigeon]

49 20 6C 69 6B 65 20 6E 65 75 72 6F 20 73 61 6D 61

4E 65 75 72 6F 20 73 61 6D 61 20 62 65 73 74 20 76 74 75 62 65 72 20 66 72

[olive sable]

[umbral wigeon]

It's basically binary but readable

[sage crag]

@real sierra automod evasion

[umbral wigeon]

4E 65 75 72 6F 20 73 61 6D 61 20 6E 75 6D 62 65 72 20 6F 6E 65

[glass jetty]

it basically text but make sense

i think i already deleted my encoder

[real sierra]

legible English only please

[sage crag]

i have this crazy method of encoding text thats like binary but readable

let me demonstrate

"hello"

[glass jetty]

happy birthday shiro~!

[olive sable]

[real sierra]

🥀

my gift is moderating for hours

[glass jetty]

what wilted_rose means?

[olive sable]

sad

[glass jetty]

|1kэ, 4o you rэmэmbэr mэ po$t1ng $omэ7h1ng |1kэ 7h1$? it actua||y prэ77y rэa4ab|э

but i guess i deleted encoder as garbage

[sage crag]

type llwrrll

[olive sable]

llwrrll

[sage crag]

not you

[olive sable]

[glass jetty]

||wrr|| :xdx:

[sage crag]

[glass jetty]

[real sierra]

wrr?

[sage crag]

that me

[glass jetty]

like ||wrr||

[real sierra]

bwaa

[sage crag]

bweero

[real sierra]

I miss my screeps

[olive sable]

[real sierra]

I get 8 emails about them a day

[sage crag]

your screeps miss you very much too

poor baby screeps

[olive sable]

prob cuz i keep killing scouts

[real sierra]

[sage crag]

[olive sable]

get off my property

[real sierra]

they doing good job at collecting data

I'm glad that functionality is effective

[olive sable]

i keep getting scouts form the guy 2 below me too btw

[sage crag]

if i played screeps i would cry if any of them died

[real sierra]

someone is below you?

[glass jetty]

is it really funny to write business logic? you not have it enough at work?

[real sierra]

you can refresh them at the spawner to make them live forever

but I just don't do that

[olive sable]

they only live for about 1500 ticks

[sage crag]

im already tearing up

[olive sable]

i dont do so either

probably should actually

[sage crag]

my emotional state is highly stable

[real sierra]

idk if I should or shouldn't

needs more math

konii hblang screeps soon

I excite

[sage crag]

_e

[real sierra]

[sage crag]

i realised a fundamental issue with hblang screeps

being that

[olive sable]

someone stole cloud's old spot,a nd someone else got between me and the guy 2 below me.

[real sierra]

[sage crag]

the compiler keeps crashing

[real sierra]

holy it's filled up quick

wait cloud moved?

holy I see my room has been up to something

my classic checkerboard building placement

[real sierra]

that not good

what happen

[sage crag]

so basically

[real sierra]

is it sad

[sage crag]

so

so

[olive sable]

i keep running out of energy. i should invade someone

[sage crag]

the feature for importing external functions

...its ..unimplemented in wasm

[real sierra]

[sage crag]

discord web client on firefox mobile

highly good

[real sierra]

[sage crag]

@rigid snow send DuranDuran

[rigid snow]

[sage crag]

ye

[real sierra]

but then

[olive sable]

i have this tower code that keeps some spare energy to kill invaders, and in this case your scouts too.
3 mil might be a bit too high for my walls since my room controller level has stagnated

[real sierra]

how is going to use wasm

[sage crag]

thanks

[real sierra]

[sage crag]

this is key point

impossible

[real sierra]

[sage crag]

mayb possibl later

NOT yet possibl

[rigid snow]

legal fake nitro mod (unpatched method 2026 not tos)

[olive sable]

we should improve wasm

[sage crag]

compiler dev currently on vacay

[olive sable]

usurp

[opaque wharf]

I thought you said you're evil

[sage crag]

wuu

[mighty thorn]

Konii is getting dangerously close to uwu

[sage crag]

false

[mighty thorn]

Literally a single character swap away

[sage crag]

false

[fickle rain]

https://imgur.com/ocZslo2.gif

[mighty thorn]

I’m revoking your math degreee

[glass flower]

wasm support is broken now?

[trim valve]

I am a victim of identity theft

[trim valve]

this run is so funny

because I keep hearing my gpu coil whine for each prompt

[umbral wigeon]

[trim valve]

like i can tell when the prompts are taking 200ms vs taking 10 seconds

[opaque wharf]

You clearly have not seen Toast

Oh, konii turns into bread now

[trim valve]

@sage crag very reliable temperature converter

(-60 is the input temp, col after is output)

[opaque wharf]

[trim valve]

this is gonna make such a funny graph

just uh

7 hours left

[trim valve]

also also

the last column is number of nanoseconds spent thinking

so when it thought for way longer it got the right answer

shuni do you want context

[rough bloom]

I know the context

see everything

[trim valve]

im an incredible programmer

[rough bloom]

[olive sable]

i honeslty cant hear anything besides the fan humming

[trim valve]

yeah its more clear in person tbh

its not done yet but the data so far is funny

[rare bramble]

wasm works in screeps, just tested it

[olive sable]

before the 3090 i had before the 7900xtx, i had a 3060ti that made the most coil whine i have ever heard ina ny electronic device ive heard

i think thats a correct english sentence, but it barely makes sense

[glass flower]

thanks for testing it.. probably not gonna play for a while but ye

[trim valve]

i wanna graph this but lazy

I'm thinking I'm gonna make one graph of "input F vs predicted C" and then another one of "input F vs time taken to predict"

[rough bloom]

preview of that first graph

[opaque wharf]

The universal function approximator everyone

[trim valve]

lol yeah I just plotted it

[trim valve]

with ground truth as well

execution time is also funny

[trim valve]

I wonder if there's correlation between execution time and error

[opaque wharf]

Correlation calculation time

Does longer thinking means it makes more mistake? Calculate to find out!

[trim valve]

huuuh

this one is uh

thinking time in ms on the x axis and absolute error on the y axis

extra silly if you make it error instead of absolute error

[warped narwhal]

gotta love that opaque bar on the error

[trim valve]

I think those are all the negative ones lol

like the thing at the left

[trim valve]

@sick owl so in conclusion, for temperatures <= 0.3 gemma4 e2b is a reasonable approximation of F to C until you reach -40 (conveniently, the point at which both values are the same), at which it loses the ability to do math 30-50% of the time

for higher temperatures,

[sick owl]

Recommended sampler settings are temp 1, top k 64 and top_p 0.95

[trim valve]

prompts are still running

huh fair

[sick owl]

So temps less than or equal to 3 are technically outside of the officially supported range

[trim valve]

I'll figure out how on earth I make that work with ollama python i guess

[sick owl]

Oh yeah what quant are you running as well btw

Tiny models suffer disproportionately from quants below q8

Even in this broken state though see what I mean about it being braindead but also surprisingly smart for its size

[trim valve]

potentially Q4_K_M, I think?

[sick owl]

Oh yeah that's gonna be mega lobotomised

[trim valve]

https://ollama.com/library/gemma4:e2b

[sick owl]

Run the Q8KXL gguf here

Do that with official sampler settings

You'll get great results comparatively

https://huggingface.co/unsloth/gemma-4-E2B-it-GGUF

Wrong link, there you go

Accidentally linked you to the 26B before

[glass flower]

i still haven't tested that model... how good is it?

[trim valve]

it depends if you want to convert temperatures below -40 or not

[trim valve]

ollama won't load that model for some reason

llama_model_load: error loading model: error loading model architecture: unknown model architecture: 'gemma4' huuuuuuh

[glass flower]

did you update ollama

[sick owl]

[glass flower]

surely they support gemma4

[trim valve]

I am running latest nixos unstable here

[sick owl]

Did you do ollama run hf.co/unsloth/gemma-4-E2B-it-GGUF:UD-Q8_K_XL ?

[trim valve]

and also i was running previous one just fine

[sick owl]

I always recommend llama.cpp for a reason, ollama is nothing but trouble for all its promises of convenience

[trim valve]

yep

[sick owl]

That should just work wtf

[trim valve]

probably related to this?

[sick owl]

Guess download the gguf manually and run it that way since you're not using vision anyway

That ought to fix it

[glass flower]

but atleast ollama is easy. i still haven't figured out all the things i need to set in llamacpp. also context window not being automatically truncated is such a missing feature in llamacpp i need it so much i'm considering not using llamacpp until its in. or make my own wrapper

[glass flower]

how are they so slow wtf

[trim valve]

idk enough about llms for that ngl

[glass flower]

not even qwen3.5?

honestly i switched off ollama a long time ago.. you could go with lmstudio its basically the same but has day 1 support for basically all things and a nice GUI

[sick owl]

Has a habit of running like ass though

[sick owl]

But yeah try Lmstudio actually

[trim valve]

does that provide me a python api

[sick owl]

If you want a convenient ollama style wrapper to test a non lobotomised version with

[trim valve]

ooh yeah it does

[glass flower]

i haven't had that issue. most of the time i'm actually getting better performance than with llamacpp because i can't screw up the configuration

[trim valve]

[sick owl]

Yup

https://lmstudio.ai/docs/developer/rest

Its just another llama.cpp wrapper that's a bit less busted

[glass flower]

just use the python package they provide and not the restapi

[sick owl]

There's a python package

[glass flower]

i've wrote my own little tool with it. pretty nice api

[sick owl]

I don't use lmstudio lul

[glass flower]

there is

i've made a git commit generator with it. that goes through every file and generates a message from the diff

(tho i don't really use it )

maybe i should rewrite that tool to just use the copilot cli directly since that has a way to generate commit message i think

[sick owl]

Side note, the fact ollama defaults to Q4 quants on all models fills me with hate

[glass flower]

huggingface does the same

and llama cpp when use the hf command

[sick owl]

Difference is they don't offer alternatives half the time on their official model repository

So many people have their first experience with a lobotomised model because of the tiny default context window, lack of quant options and strange default sampler settings on ollama

And don't get me started on the consequences of them naming the deepseek distills "Deepseek V3 4B" etc.

[glass flower]

the same is true on llamacpp unless you config it correctly

[sick owl]

The llama.cpp defaults are at least a bit more sensible

And it doesn't bill itself as plug and play

[glass flower]

i mean true...

but i'm also not smart enough to know where to even find information on what parameters to use

maybe i should make a AI fetch the unsloth site and get the settings for me and write it into the config file

[sick owl]

Honestly just go with the context limit you want and the recommended parameters on the official model card for a given model

[glass flower]

i mean thats what i've been doing... but again the fact the context isn't truncated is a pain for me

since i'm using gary for testing integrations so context can grow to 100's of messages and 20000+tokens very easily if it isn't truncated. and gary itself doesn't have anything like that yet

[sick owl]

They're working on it

https://github.com/ggml-org/llama.cpp/issues/19838

[glass flower]

i know...

but it isn't in yet

[sick owl]

In the meantime maybe just do --context-shift

Or use summarisation

[glass flower]

both of those don't work

atleast its not a simple plug and play thing

i patched gary rn to just discard messages older than 50 entries

[sick owl]

Summarisation isn't plug and play no, context shift more or less is

Unless you're on an unsupported model

[glass flower]

unsupported model?

isn't that just a general feature? why is it model specific

[sick owl]

Oh wait nvm context shift was deprecated in favour of the truncation stuff they're working on lule

[glass flower]

[sick owl]

You could just merge in the pull request for it tbh if you don't mind the fact its not been fully vetted yet

[rigid snow]

codex compaction is actually awesome

black magic

[trim valve]

lmstudio sounds different on my gpu

lol

[glass flower]

[trim valve]

thank you amd for coil whine

[sick owl]

Different models, frameworks and quants actually make different noises when going brr believe it or not

[rigid snow]

ye it depends on the inference framework too

[sick owl]

Its a well documented phenomenon

I like coil whine during inference, its fun

[glass flower]

i guess i could do that... do i need to fork llama.cpp to pull in the pr or is there a way to locally just patch it

i'm not a git expert yet

[rigid snow]

add the pr repo as a remote, pull the branch from it and merge

[stark needle]

Why doesn't deepseek v4 have vision

😭😭😭😭😭😭

It would have been so perfect

[rigid snow]

they still haven't figured it out??

[trim valve]

its kinad neat having actual sound every time a progress bar clicks forward

[sick owl]

git fetch origin pull/19841/head

git merge FETCH_HEAD

[glass flower]

there are merge conflics

[sick owl]

Or if you want to give it a separate local branch for testing @glass flower

gh pr checkout 19841

[stark needle]

No image input

[glass flower]

i'm using lazygit. i think i got the merging thing figured out

[stark needle]

They have deepseek ocr connected to deepseek v4

[rigid snow]

i know they have ocr

[stark needle]

For fake image input

[rigid snow]

i mean, still????

like come on

[glass jetty]

like, surely there merge conflicts
when they weren't?

[stark needle]

Stupid shit

Frontier llm and no vision

[rigid snow]

it apparently has a hidden rp mode via prompt injection into the system prompt

[stark needle]

Bro just vector quantizer vae and call it a day 😭

[rigid snow]

and it would reason in character

[stark needle]

I'm gonna try tomorrow

[rigid snow]

they had the resources for THAT but not vision

lol

[stark needle]

My ethernet is broken rn so i cant download it

[rigid snow]

it's in chinese

wait let me find it

[stark needle]

Well it's not broken I just bought a 10gbe mikrotik switch and it needs some routeros shit config

Cause i need 10gbe on both dgx spark

[rigid snow]

https://github.com/victorchen96/deepseek_v4_rolepaly_instruct/blob/main/README.md this apparently

[stark needle]

How did people even find out

I mean cant u prompt any llm for rp?

[rigid snow]

not that it would rp in CoT i never could

it's without the usual "this is a simulated chat/you're the DM"

[stark needle]

So safety bs?

[rigid snow]

it just thinks it's the character

[stark needle]

Wait is v4 safety trained

[rigid snow]

i never attributed it to safety just thought of it as a side effect of what they post train them to be

[stark needle]

I just need a good <400b llm

😭

Every good llm is text only

[sick owl]

Gemma 4

31B

EZ

[stark needle]

Runs like shit on 2x dgx spark

[sick owl]

So does everything dense

[stark needle]

Yea thats why i need massive moe model

I can afford 14b active params max

[sick owl]

Why not just use deepseek v4 flash with a smaller model handling vision tasks

Like throw Gemma 4 26b on top for it or something

[stark needle]

True just use Gemma e4b

So i get audio as well

[sick owl]

I'd use 26B if you want something closer to frontier vision performance to go with that frontier model

[stark needle]

I need sth that can realistically handle embodied robotics

[vague sonnet]

crazy how text generators operate on text

[sick owl]

GLM 5V Turbo?

Wait no that's like 800b params lule

[nocturne olive]

Trash machine

[stark needle]

I just need nvidia for quietest machine possible

Since i need to run my ai agent thing

And need stealth setup

That i can claim to my parents it only cost 1000$/machine lmao

To my mom at least since my dad knows how much he spent on this

[sick owl]

You could run Qwen 3.5 397b at like a q4 quant

Would be lobotomized though

Low active param count and all that

[stark needle]

Why doesn't google just pls drop gemini 2.5 flash open source

[sick owl]

Gemini 4 31b already outperforms 2.5 Pro let alone Flash

[stark needle]

2.5 pro?

[sick owl]

Have you considered using it with speculative decoding?

[stark needle]

Not rly supported in vllm

I aint running ollama/llama cpp

Also gemma 31b is a distil of a distil

[sick owl]

Qwen 3.5 122B A10B might be your best bet actually

[sick owl]

Well, unless you want to chance it with the 397b at a q4 quant

[trim valve]

@sick owl that model you told me to use is arguably worse lol

[sick owl]

And you set the recommended samplers?

Lmao

[trim valve]

pretty sure?

I can double check

[sick owl]

That'll just be the usual tiny model variance if you're using the recommended sampler parameters

Make sure you don't have repetition penalty or presence penalty enabled

[trim valve]

im not super sure how I'd go about setting that through the python api

[sick owl]

Without these settings its expected that the model'll underperform

[trim valve]

any suggestions for better settings?

oh wait i can't read

thanks dyslexia

[glass flower]

lower temperature? ( i know this sounds ironic)

[trim valve]

going well

[glass flower]

is LLM. the model or the class?

i think there are like 2 or 3 things called respond in the api

depending on what you call it on

[trim valve]

it takes the args differently

you need to pass it in as dict under the config kwarg

well apparently this is close enough to right

response = model.respond(
    chat,
    config={
        "temperature": 1.0,
        "topPSampling": 0.95,
        "topKSampling": 64,
        "presencePenalty": 0.0,
        "frequencyPenalty": 0.0,
        "repeatPenalty": 1.0,
    },
)

though frequencyPenalty seems to not actually be part of the typescript structure

[sick owl]

Idk what arguments and format they use in lmstudio*

Like I said I don't use it

[trim valve]

im using lmstudio

[sick owl]

Typo

[trim valve]

ah

go my computer

[hearty notch]

is vedal getting botted followers because people bet on when itll hit 1m is that whats happening lmao

[trim valve]

what the

great prediction there bud

-58 + (459.67 * (-1)) = -58 - 459.67 = -517.67
-517.67

what kinda cooked working out is that

[sick owl]

HOLY

[sick owl]

Technically not wrong

[trim valve]

without the outlier

it seemingly is getting different results if I ask it in the chat vs via the api which is interesting

I guess this test it probably pretty sensitive to the prompt I choose

[sick owl]

Q8 vs BF16 is always gonna have some amount of difference, especially on a tiny model like this

Hell you can try the BF16 gguf locally if you want, its only 9 gigs

Though tbh at that size I'd probably go for E4B at Q8_K_XL

[trim valve]

I probably need to sleep ngl

but I would maybe think twice before trusting it for temperature calculations

[sick owl]

Or hell Qwen 3.5 9B

[cosmic sphinx]

god

windirstat changed my life