programming

[true hemlock]

also your next target for a decent jump might be 6090

[sick owl]

That's the plan

I chose to stay on the 40 series and trade in my 4080 for a 4090 when the 50 series came out

Bigger jump than upgrading to the next gen lol

[nocturne olive]

6090 is gonna cost 590304593049 though

[true hemlock]

you almost went with a 5080 out of fomo lmao

[sick owl]

Yeah :SAVED:

[idle dune]

My laptop when I try to do anything. I think I need to repaste it.

(Not actually my laptop)

[nocturne olive]

I repasted my laptop and found a bomb inside

[idle dune]

But I idle at 69c (nice)

[idle dune]

Thanks for reminding me, I need to replace the battery in my P73 aswell

[teal cipher]

"I buy that for a dollar!"

[idle dune]

https://tenor.com/view/robocop-gun-spin-tricks-gif-5281451

[teal cipher]

it's a movie reference... anyway yeah I wonder how much they're trying to go and milk with next lineup... maybe they surprise with something something to bring down card prices

[idle dune]

They are gonna be absurdly expensive

6090 is gonna be like? 3999.99? Thats what my guess on the base price is

Also inevitably there ther going ti be someone scammed by a scalpers being sent a 690 instead of a 6090

[nocturne olive]

Intel make a D780 already and take over the market with crazy value

[kind nimbus]

for that money you can buy a used car

[opaque sigil]

nah they won't price the 6090 at 4k msrp

it'll be like 2500 or 3000 max

not that msrp matters at all

[nocturne olive]

But they can sell the same exact chip to datacenters for 10x more though so why would they give consumers such a big discount, there's no way it will be under 4K

[stark needle]

in practice yep

after all retailer margins etc yea

[opaque sigil]

it being a worse version of the chip doesn't really matter when it has several times less memory

[stark needle]

and then put scalpers on top

[opaque sigil]

just don't sell many 6090s and push people towards the big brother

[nocturne olive]

Chances are 6090 will not increase beyond the 5090's 32GB
It's even possible they go back to 24GB

[stark needle]

nvidias fake datacenter gpu is gb10 and successors

[opaque sigil]

I doubt a spec downgrade would be worth the PR disaster

[sick owl]

If that's the case I'm going for a 5090 and keeping to my one gen behind status quo

[nocturne olive]

I mean they have an excuse (RAM prices) soo

[opaque sigil]

charge more then

for the same 32 GB

[stark needle]

doesnt nvidia already have a ridiculous margin

ram prices cant matter

[opaque sigil]

they do

but more profit == better

[true hemlock]

nah about similar as 5090

[true hemlock]

wise move.

idk specifically what would be the sku config for 6090

i do somewhat know the prototype die config

[true hemlock]

they're basically using "ram shortage" as excuses

to increase all price

[opaque sigil]

Rubin (?) should be around CES 2027 iirc right

unless it gets pushed back again I guess

[true hemlock]

btw 4090 die cost is under $300 for the bare wafer with avg yield rate of 70%

the margin is already ridiculous

5090 about $300ish. def under 340

smd components are almost negligible

vram isn't even much

[teal cipher]

I have 4090 graphic card... one of those variations that are intended to be cool while less noisy... not watercooling, just fans and other techniques to manage heat so it doesn't burn... when they released their new shiny lineup I was like, nah... XD

[nocturne olive]

Meanwhile my 4070Ti is being cooked because there's a 3090 below it and it can't get air and now I can't get it to its full performance

[stark needle]

same with my 3090 lol

which is at the bottom of the case

[opaque sigil]

get a riser and move them around

[fickle rain]

Put waterblock on both cards and enjoy life

[nocturne olive]

I can't afford a water cooler for one let alone both

[idle dune]

I'm just making a AI computer out of a old workstation laptop

Grab a newish Precision or Thinkpad with a Quadro A5500

Something like this

Good enough for me

[stark needle]

wtf actually good deal

[idle dune]

Yep

16gb only but that's still a fairly quantized 30b.

Cheap low power laptop AI sever

[nocturne olive]

Doing AI on a laptop is gonna burn the laptop
Speaking from seeing that happen to a certain Blur

[idle dune]

But I have a power bill

[nocturne olive]

But the laptop is gonna have to run way longer to achieve the same amount of computations so it'll end up not being much of a savings anyway

[idle dune]

True

[frozen void]

what people use to clone a voice, something lightweight that can run locally?

[maiden geyser]

rvc

[nocturne olive]

RVC

[frozen void]

i mean i need a cloning model for tts

[nocturne olive]

Though it depends on the exact goals you have, but RVC is the basic most simple method with the least setup required

[nocturne olive]

RVC on TTS

[rough bloom]

Chatterbox or any of the other zero-shot voice cloning models

not need training

[nocturne olive]

RVC is the simplest and most versatile kind of voice cloning, however it may result in lower quality

[nocturne olive]

I forgor about those

[rough bloom]

(expect output quality to be a bit worse probably because of this, depends on what you're doing with it)

[nocturne olive]

It's all a quality tradeoff

[frozen void]

i want to clone a voice of a character from anime and use it for my vtuber ai project

[teal cipher]

I don't get it Chatgpt pro didn't want to play nice with Github but now it's like... hey it works lol

[nocturne olive]

Do be aware, that this may be illegal and also somewhat difficult to get clean data for

[sick owl]

Its almost like 1k was a fair price and everything since has been a piss take

[true hemlock]

yeah

also i estimated the cost

smd is negligible, pcb is also somewhat negligible, but lets say $120 at most including heatsink and cooling

vram

is like

$6 per 16Gbit module

could also be cheaper due to contracts and stuff

because gddr7 is nvidia exclusive iirc?

isn't even $600 per 5090 for production cost

oh well

honestly

probably far less, im just overestimating

the process node they've been using has gone down in price aswell

honestly im also surprised you don't instead went with dual 3090 @sick owl

[olive sable]

ok the usb-c decoy board is def burning

im buying a new one

third times the charm

[true hemlock]

what

how

[olive sable]

idk

shorted or something

i just did a shitty solderjob i think

[maiden geyser]

of course it is

[frozen void]

even for learning purposes? i want to do it this way cus it's funny to work on the voice of my favorite anime character, i won't use it commercially just locally.

[nocturne olive]

If you're not publishing and just learning how to use the tools it should be fine

[frozen void]

yeah of course

[nocturne olive]

Probably gaming reasons

[frozen void]

same reason apply for the 3d model i made for that

[teal cipher]

chatgpt pro doesn't missclick... it "clicked incorrectly"

[maiden geyser]

enriqueeeee

[true hemlock]

considering the model size they used for the ai overview im not surprised at all its so shit and ungrounded

[silent cloak]

What is the size?

[true hemlock]

[silent cloak]

I never looked into the specifics because the results spoke for themselves

[true hemlock]

[opaque sigil]

it's one of the dozens of gemini models

[silent cloak]

Makes deepseek look like opus

[tender river]

tricky spelling question

[trim valve]

[frozen igloo]

Poogle

[trim valve]

I'm not completely surprised

[olive sable]

[silent cloak]

Goople

[trim valve]

this should be your lesson to try aggressively continuity checking stuff before attaching power

[olive sable]

i did

but it jsut shorted again

idk

[nocturne olive]

LLM when it writes something, then disproves itself, and then writes as if it was intentional

[rough bloom]

continuity check

harder

[olive sable]

ye, ground and + are shorted again

and idk why

[rigid snow]

you know what's better than 1 p? that's right

gpt 4o mini or whatever they use

[teal cipher]

chatgpt pro is along the lines I'm a dummy so better to just provide zip file with changes ready

[frozen igloo]

I wonder what Neuro would answer

[fickle rain]

https://www.anthropic.com/news/claude-opus-4-8
New slop???

[opaque sigil]

surely it won't be a downgrade like 4.7

[tiny edge]

chat, my mb doesnt have ReBAR support

[glass jetty]

guys

how to make my code not shit

its hard
writing something, changing something, looking back
and it is shit now

[tiny edge]

impossible

[kind nimbus]

New kernel vulnerability dropped

[opaque sigil]

We’re making swift progress on developing these safeguards and expect to be able to bring Mythos-class models to all our customers in the coming weeks

[obsidian mantle]

i love when they use the word "class"

for some reason sounds epic

[opaque sigil]

tbf if their benchmarks are to be believed it is kind of in a class of its own

[kind nimbus]

Better honesty is good

I noticed oftentimes opus claims to be confident to actually found a vulnerability while in reality it was a dead end lol

[glass jetty]

[nocturne olive]

Not my 80 days old kernel

[glass jetty]

where?

[kind nimbus]

https://cybersecuritynews.com/linux-cifswitch-kernel-vulnerability/

[nocturne olive]

Another privilege escalation?

[kind nimbus]

Yea

[glass jetty]

local one

[nocturne olive]

So how many does my kernel have now?

[kind nimbus]

Enough

[glass jetty]

no one actually looked into it as always

[glass jetty]

at least 3

[nocturne olive]

[kind nimbus]

Also this looks like an interesting read

https://heyitsas.im/posts/drinking-llms/#future-research-directions

[tiny edge]

doesnt work on most systems i think

atleast not on mine

[nocturne olive]

[tiny edge]

[maven@nixos:~]$ which mount.cifs
which: no mount.cifs in (/run/wrappers/bin:/home/maven/.local/share/flatpak/exports/bin:/var/lib/flatpak/exports/bin:/home/maven/.nix-profile/bin:/nix/profile/bin:/home/maven/.local/state/nix/profile/bin:/etc/profiles/per-user/maven/bin:/nix/var/nix/profiles/default/bin:/run/current-system/sw/bin)

[kind nimbus]

I bet an attacker could bring a static version of cifs-utils if it's not installed

The vulnerable code is already in the kernel

[tiny edge]

perhaps

well, updating is the safest

[sick owl]

https://fixupx.com/claudeai/status/2060042702150930686

Claude (@claudeai)

Introducing Claude Opus 4.8: it builds on Opus 4.7 with sharper judgment, more honesty about its own progress, and the ability to work independently for longer than its predecessors.
︀︀
︀︀Available today at the same price.

**💬 1.0K 🔁 1.9K ❤️ 14.1K 👁️ 671.7K **

FixupX · May 28, 2026, 4:57 PM

I thought 4.7 was getting suspiciously stupid

[kind nimbus]

I went back to 4.6 because opus 4.7 seemed to burn tokens faster

[opaque sigil]

new tokenizer uses more tokens

i guess this would be a problem with 4.8 too

[kind nimbus]

[obsidian mantle]

same price (but more token usage)

im surprised they dont just ban using 4.6 and make you use 4.7/8

if its the case

or is it like licence thingy where you paid for 4.6 you must have 4.6 available

[teal cipher]

I remember back in the day fixing bug in starcraft 2 custom map I was doing by just redoing triggering for random card draw system... and it just worked correctly... where I made mistake before... shrug... dunno but it was bug free XD

[fickle rain]

30€ of pure power

[kind nimbus]

They will probably disable the older models soon i guess

[frozen igloo]

https://robertkaplinsky.com/how-old-is-the-shepherd/
lol what

The classic llm blunder except it’s humans doing it

[floral hawk]

Got a new monitor

I present: Notepad, da long way

[silent cloak]

based

every programmer needs a vertical monitor

[torn flicker]

I'm trying to compile a Linux kernel, I haven't slept for 18 hours

[silent cloak]

sounds about right

[floral hawk]

I'm saying

[silent cloak]

switch it to dark theme

for bonus points

[floral hawk]

Her personality config is giving me more trouble than making her play dark souls is

She's either boring or fucking mean and I can't find a middle ground

[torn flicker]

More is needed to have a 360° view

[silent cloak]

i need to free room on my desk to get my other monitor in

i have so much shit i need to 3D print wall mounts

[torn flicker]

what about VESA bracket?

[floral hawk]

I'm getting there bc my laptop is where I run twitch from

[silent cloak]

the issue is im using several monitor brands rn

[floral hawk]

2nd monitor is for Zero controls, main monitor is what her vision runs on

When I'm streaming my entire desktop is covered

[silent cloak]

i need to replace them all eventually just to have a set of a single type

but i have other stuff to spend that money on rn

[torn flicker]

logical solution

[silent cloak]

plus its about adding a fourth

which idk how that will go

my desk has a really strong motor so im not too worried about the weight though

[torn flicker]

my main monitor died and i have to use the old one Samsung syncmaster 19 inches

it's painful

[harsh python]

hi guys its me again

[silent cloak]

i have 2 1440p60 ones and 1 1080p240

the 1080p one is my main one used for gaming and such

the others are best for watching stuff and documentation

[torn flicker]

same btw

[silent cloak]

the 1440p ones need to go eventually though they are pretty old acer ones

and some random days if im doing GPU heavy stuff they can bug out a little bit

[torn flicker]

I just have a monitor aser out of order

[silent cloak]

my main one is one of those zowie esports ones

[kind nimbus]

New leak? Lol

[torn flicker]

he decided to die a week after the end of the warranty period

[silent cloak]

i do prefer zowie over acer though now

they are just kind of pricy depending on what u get

[torn flicker]

They have pretty good monitors as far as I'm concerned.

[maiden geyser]

new tip for the filian arg

[silent cloak]

theres a reason they are the esports monitor of choice

[floral hawk]

I'm thinking about switching Avis for Zero but the thought of reworking her anims completely make me borderline psychotic

[silent cloak]

they also have a 360hz one

idk who would buy it though

[floral hawk]

I don't like the one she has rn

[stark needle]

neuro is qwen based lol

qwen 2.5

[floral hawk]

NO WAY really?

[silent cloak]

no

lmao

[torpid coral]

what

[silent cloak]

unless vedal time travelled to the future when he made her

[opaque sigil]

[silent cloak]

thats wrong

[stark needle]

he has obviously changed the base llm over the years

[floral hawk]

Zero is Gemma based

[silent cloak]

yeah not THAT much though

just different version

[torn flicker]

I'm thinking about buying a 1440p Samsung monitor and one 1080p for gaming, but I need to wait for my salary.

[silent cloak]

other mistral or a llama based one

[stark needle]

but neuro has recently output specific control tokens which only exist in the old gpt3.5/gpt-4 tokenizer or in qwen 2.5 tokenizer

[maiden geyser]

source: 💊

[floral hawk]

I'm thinking about doing my own model fine tune and abliteration but I'm not sure how viable that is.

[silent cloak]

vedal did start off learning machine learning with gpt2 technically

well LLMs

[stark needle]

ok but neuro is obv not gpt2 based lol

[silent cloak]

duh

[torn flicker]

What's wrong with you guys? It's completely built on scratch

[silent cloak]

lmaoo

[floral hawk]

💀

[silent cloak]

neuro is a fine-tuned model

[floral hawk]

"gas is too expensive"
"Let's build our own offshore oil rig"

[silent cloak]

vedal doesnt use prompts or anything to define her personality

its all based on her training data

[floral hawk]

He WHAT

Oh my god this is what I've been missing

[silent cloak]

u didnt know that?

[stark needle]

not even openai or google do that man

[nocturne olive]

[stark needle]

this is obv a lie

[silent cloak]

sorry man its what he actually does

[tall island]

Source?

[silent cloak]

anyone who knows the tech knows its better anyways

she isnt a general public multi purpose model

fine tuning her is better than a slop prompt

[stark needle]

neuro sdk works with prompts and to connect to neuro sdk it would have to have an api reference on how to output structured json so there must be a prompt

same thing with tool calls

[nocturne olive]

That says no specific static personality prompt
Not "there is no prompt at all"

[silent cloak]

well duh

she is still an ANN

[nocturne olive]

She probably has something like "You are an artificial intelligence vtuber"

And some memory systems to define the rest

[torn flicker]

My laptop's CPU is at 92°C while compiling the kernel, I hope my house doesn't burn down

[silent cloak]

i doubt thats needed atp given she is fed her chat data

[floral hawk]

me on my way to do my own fine tune because I hate working on a prompt

[silent cloak]

fine-tuning was always the way to do it before tbh

[nocturne olive]

Average laptop temperature

[silent cloak]

prompting didnt explode until gpt 3

[stark needle]

i dont think people understand what is actually meant when vedal says "neuro is trained on stream chat"

[floral hawk]

I'm finna train her on Homestuck, Hazbin hotel and shitty anime

[silent cloak]

i do though

alot more than u would think

[torn flicker]

I found a use for this, you can warm up a mug of coffee

[nocturne olive]

[silent cloak]

though these days it is mostly just general chat opinion

[nocturne olive]

Laptop user tip: open your laptop sometimes, you never know when the paste needs swapping or your battery has become a bomb

[torpid coral]

this is how I warm the room

[torn flicker]

I'm afraid of the electricity bills this month

[silent cloak]

lmao

[stark needle]

<|fim_suffix|> only exists on qwen 2.5 tokenizer (token 151661) and cl100k_base, p50k_edit both openai tokenizers of which all llms are long decommissioned

[glass jetty]

so old and stupid thing

[glass jetty]

yeee

15 years old or less

[silent cloak]

a lot of models train on other model outputs (bad practice though)

[stark needle]

ok but why would u train on qwen 2.5 when u can just use qwen 2.5

[silent cloak]

probably because its qwen

but we dont know

[stark needle]

idk what ur on about man 😭

[opaque sigil]

training on qwen 2.5 outputs seems like... a choice that's for sure

[silent cloak]

qwen and gemma are kind of the modern mythomax

pretty ass besides low budget setups

[stark needle]

what is meant by that then

[nocturne olive]

I wonder where I could find a German voice actor, we could really use one for NSP

[silent cloak]

side note

[floral hawk]

I wish Vtube Studio, OBS and everything else I need to run this setup didn't eat so much fucking VRAM so I could actually use a larger model

Anyone got any tips?

[silent cloak]

i just realized my desktop has been running everyday without a CPU fan for a month

[kind nimbus]

Mh Well i don't know any

[stark needle]

bro i've professionally trained llm models at a frontier company i do also know a few things about llm pretraining, continued pretraining, sft, rlhf, etc...

[floral hawk]

Holy Jesus

[nocturne olive]

[torpid coral]

what

[floral hawk]

Rip cpu

[torpid coral]

real

[silent cloak]

meh its a shitty old i7

[nocturne olive]

Run all that stuff on iGPU to save VRAM, or get a secondary GPU for offloading

[silent cloak]

its been my bottleneck anyways

it will join my others in the CPU bag soon

[floral hawk]

I have enough RAM to actually run most of it on there.

[nocturne olive]

So just throw it on the iGPU

[stark needle]

rip obs with no nvenc

[nocturne olive]

But the VRAM

[stark needle]

just

use a smaller model

• distil

[floral hawk]

I legitimately cannot downsize right now. I need to upgrade tbh

[stark needle]

distillation goes a long way

• finetuning

• u wouldnt be able to train a much larger model anyway

[floral hawk]

I might do a custom Gemma fine-tune low-key

[stark needle]

finetuning goes a long way

just need to know how lol

[floral hawk]

I'm running E4B, I wish there was a midrange like E8B

Yeah see I know how to fine tune vision. But not general LLMs

[glass jetty]

always talks about a i

meow

meow

meow

meow

[floral hawk]

Right now I'm running both a base Qwen 2.5 for more reactionary content, and then Nvidia Identify Anything for her dark souls/armored core experiments

[glass jetty]

meow

[floral hawk]

Armored core is. . . Not fun so far

She's gotten as far as identifying enemies and shooting back, but not so far as to dodge attacks, so she's still getting absolutely rolled there. She's doing fine in dark souls, but has issues with the open world

[nocturne olive]

Is it an LLM controlled game system or is there a separate model for the game?

[maiden geyser]

bark

[floral hawk]

It's LLM controlled through an SDK

I custom built one. Similarly to how Neuro plays games

[nocturne olive]

Well that's your problem

That's why no attack dodging is happening

[floral hawk]

I'm aware. 😭

[true hemlock]

he never even saidnthat he just said neuro can read chat long ago

[floral hawk]

But also I'd just like to say that I as a human have trouble with armored core. I fucking DOUBT I'll get her to play it on my current hardware, it's more experimental

[nocturne olive]

As long as an LLM's in charge, it'll be like Half Life 1 AI that can only do one thing at once with massive latency

[true hemlock]

prompts are still involved whether systematically (for instructions and calling) or just direct system prompts

[floral hawk]

Well a lot of it is code structure, I think we miscommunicated

The LLM makes decisions but the actual gameplay is split like 80/20 script/llm

I want to have a game model but I'm not sure how to do that yet

I have to basically make individual profiles for each game with unique logic.

[kind nimbus]

Oh no. Looks like i lost my German privileges

[true hemlock]

she's lora finetuned, just highly doubt the "no prompt at all" part. finetuning don't really make your model magically not require at the very least basic system prompt for it to perform well and with some basic chat info

also all the integrations are literally technically prompts or include prompts

[stark needle]

arXiv:2005.14165 my beloved

[true hemlock]

this never happened idk why the community still thought this is a real info

[teal cipher]

evil neuro's model could be just that... trained to be evil neuro... so then it doesn't really need that much instructions to stay in character because that's her model... unlike model that is expected to be general one just try to handle various tasks and not be specific defined character like that

[patent shard]

AV2 comes out tomorrow
How long until it gets added to ffmpeg
Probably very quick

[nocturne olive]

AV2??

Something I don't have hardware support for

[teal cipher]

especially nowadays once you experience footage in 144fps it's hard to go back... with expense of file size becoming much bigger but that's besides the point! XD there are actually solutions that do it well nowadays instead looking terrible

[patent shard]

AV2 should more compete with VVC that came out 6 years ago
Lol
A little behind but it is the free solution that actually gets adopted

[opaque sigil]

dav2d is 250k loc

[teal cipher]

btw... as much I like idea of open source I find it being largely voluntary work doesn't exactly let it to keep up with commercial solutions

just a thought lol

[opaque sigil]

wdym voluntary work this is driven by half the industry giants and their employees

unless you mean dav2d

[patent shard]

Google and Netflix are two big ones, I think?
I'd need to look

[opaque sigil]

amazon, apple, cisco, google, intel, meta, microsoft, mozilla, netflix, nvidia, samsung, tencent

[teal cipher]

there is plenty of good open source stuff but certain things that require tons of computing for example can't really keep up being run locally... and it's not in that case really lack of effort though... also they tend to lack excessive guardrails

[patent shard]

I agree, though
It took 6 years with AV2 just to about match VVC
AV1 falls behind VVC a lot in some scenarios
Like.. we are talking 1.5x the bitrate required with AV1 to match VVC
And these are at preset 0 or 1 with SVT-AV1

Yes, AV1 is lighter to decode
But if VVC had more hardware decoding adoption.. eh

But VVC is all about money

[opaque sigil]

i guess one big problem is that you have to be super careful not to accidentally use anything patented

[teal cipher]

I think patent system how it currently works isn't good but don't expect it to change anytime soon

it gives too ironclad grip of various innovations

[torpid coral]

Do you think I can keep this design? (just an example):

NeuralNetwork net;
net.add(new Dense(784, 64, ActivationsType::RELU));
net.add(new Dense(64, 10, ActivationsType::SOFTMAX));
net.init(new SGD(0.001));

Or should we take out the activation function as a separate layer?

[kind nimbus]

Finally realization and acceptance

[fast pagoda]

Half of oss is just people paid by corpos to maintain it

Oh toast already said this lule

But yeah you see every time someone retires from red hat or some shit that like 30 packages go out of support and it's a disaster

The other half is left pad

[torpid coral]

the last stage of acceptance

[velvet crest]

You have unlocked new role

[fast pagoda]

Probably from rag/memories making it seem like it

[maiden geyser]

#baking people when poll

[split steppe]

i aint putting no goddamn nuts in my mouth

[rigid snow]

how do i build a quantum computer

[kind nimbus]

Ngl i was just researching for more healthy alternatives to energy drinks as someone recently told me how bad they are for me.

Has anyone had success with Mate as an Energy Drink replacement or are there better options?

[split steppe]

best replacement is good sleep

which you get by ditching stimulants

[kind nimbus]

Without caffeine my gears don't start kicking in

I need something to get focused and keep me awake

[split steppe]

yea because it's a drug that you get used to

you don't work without it because you're addicted to it

takes ~3 weeks to ween off, ~3 months to properly recover

[kind nimbus]

Might be the case

Hard to get away from it tho

[tender river]

i thought you were talking about nixos

it

[split steppe]

i've ditched coffee before and it was the best time of my life, always slept like a baby and woke up ready to go straight outta bed, but i just like the taste too much and decaf is shit

[tender river]

does take months to recover

[kind nimbus]

That's a drug i cannot recover from

[kind nimbus]

Damn i always take tryptophan and GABA to get a good night's sleep since i also have sleeping issues

[split steppe]

yea cuz, well caffeine fucks your sleep up lol

that's kind of the point of it

[kind nimbus]

[split steppe]

also energy drinks have like 3-5x more caffeine than a mug of coffee, plus taurine which is another stimulant

shit's way more powerful

and if you drink several a day, well

[kind nimbus]

I usually drink 1-3 cans a day every day

[split steppe]

yea that's gonna screw you up hard

be glad you don't have heart problems

[kind nimbus]

I've been doing that for a few years now

My heart is fine

But also I've always had way too low blood pressure anyway

[split steppe]

atleast it increases pressure ha

but yea, you may have a much harder time getting off it

do it on vacation or something

unless you wanna do it slowly and spend like a year

[kind nimbus]

I first would need to find a replacement lol

[split steppe]

since you're drinking energy drinks, idk, other soda maybe

coffee has a bunch of ok replacements

but also soda isn't a good replacement, because of sugar

another kind of stimulant

[kind nimbus]

I like coke but that has caffeine as well

I usually avoid all kinds of sugar

[split steppe]

but that's mostly because of energy spiking

coke has caffeine ya, so that's off

sodas have a massive amount of sugar

energy drinks as well

[opaque sigil]

there is technically a caffeine free coke zero

[kind nimbus]

I'm doing keto or at least low carb for a long time now already so i try to avoid carbs and sugar

[split steppe]

that might work i suppose

or just get a sodastream and drink plain bubbly

it satisfies my soda craving most of the time, cuz i found out i'm mostly after the carbonation

[kind nimbus]

We have a SodaStream but idk water is kinda meh taste wise it's boring

[opaque sigil]

sprudel

[split steppe]

drink cider if you're alcoholic

[kind nimbus]

I drink alcohol very rarely only with friends on special occasions

[split steppe]

cider isn't that bad actually, it's only like 2-4% usually less than beer, and often you can get 0% which tastes the same

basically soda, but natural

[split steppe]

can also technically make your own natural sodas, but takes work

e.g. ginger bug and stuff

[kind nimbus]

Hm

Thought about buying some mate but if it ain't better due to the caffeine I'm gonna look for other sodas

[split steppe]

do you mean yerba mate, as in the tea?

[kind nimbus]

Mio Mate

That should be Mate tea yeah

[split steppe]

never seen that kind, interesting

[kind nimbus]

There is a zero variant as well

[high brook]

I read that as mid

[split steppe]

ah it has sugar and caffeine by default, meh

i'd just find something you can make yourself tbh

see what you can do with your sodastream so you can properly control it

[kind nimbus]

Ingredients are

Natural mineral water, carbon dioxide, 0.5% mate leaf extract, acidifier: citric acid, natural flavoring, caffeine, coloring: ammonia caramel, sweeteners: acesulfame K and sucralose.

[split steppe]

6g of sugar per 100ml

[kind nimbus]

That's why I was looking at the zero variant

[split steppe]

still a lot

i guess zero doesn't have sugar

[kind nimbus]

Average nutritional values ​​per 100 ml: Energy: 7 kJ (2 kcal), Fat: 0 g, of which saturates: 0 g, Carbohydrates: 0 g, of which sugars: 0 g, Protein: 0 g, Salt: 0.03 g.

[split steppe]

but has caffeine still

[kind nimbus]

Yes as its Mate

[split steppe]

yah so it wouldn't help that much, it's only ~10mg less

[kind nimbus]

Hm ok

[split steppe]

if you like tea, you could switch to flower based teas

or chicory

but if you're after sodas, idk

it's all made to be addicting

[fickle rain]

If you’re in Geramy just pick Clubmate

[kind nimbus]

I'm in Germany yes

Clubmate is quite popular here but idk if they have something without sugar

[fickle rain]

They do

[kind nimbus]

But the problem is, would that be healthier than Energy Drinks if caffeine is the issue ?

[fickle rain]

Probably no, but who fears caffeine anyway

[kind nimbus]

Well @split steppe thinks i should try to get away from the caffeine

[split steppe]

and sugar, but that's a stretch

caffeine is the big issue

[kind nimbus]

I barely consume any sugar

I said goodbye to sugar like 6 years ago

[split steppe]

presumably your energy drink is the sugarless type then

so that's good

[kind nimbus]

Yea

I only drink zero stuff

[split steppe]

i don't, mainly because i still don't trust the sweeteners, something intuitively feels off about them

[kind nimbus]

Hm so far i did not find any studies that actually came to the conclusion that Acesulfam K or Sucralose or Aspartman is actually worse than sugar.

Also i lost 30kg when i cut off the sugar which was good

[split steppe]

they're fairly new in the grand scheme of things, and nutrition is notoriously difficult to study, so we'd probably need a meta study across dozens of years to find anything

[kind nimbus]

Yea makes sense

[split steppe]

also a study in humans specifically, because they keep doing that shit on rats or whatever and none of it applies to humans because we work differently

[kind nimbus]

A study on real humans would be interesting

[true hemlock]

the point with artificial sweeteners is that they're way more stimulating and stronger so you consume far less glucose.

it works

and they're healthier than normal sugar for most people for that reason

[split steppe]

for sure, if it helps overconsumption of sugar, it's prolly a net benefit

[true hemlock]

not for me though my blood sugar level is on the lower side

[split steppe]

i in general don't like sugar, so i'm fine either way, so i stick to sugar because we know it

[nocturne olive]

Any Germans around here? I happen to be looking for a German voice provider

[kind nimbus]

Hallo

[true hemlock]

so i ocasionally just have to consume glucose rich food

[nocturne olive]

Hello are you a German voice provider?

[sage crag]

https://tenor.com/view/rabbit-cute-gif-14063524

[kind nimbus]

No but I'm German

[nocturne olive]

Well would you like to help in the incredible goal of making Evil sing German?

[sage crag]

all your'r'e samples are belong to us

[kind nimbus]

How could I help with that?

[sage crag]

voice donation

[nocturne olive]

By singing in German obviously

[true hemlock]

[kind nimbus]

Uhm my singing voice is really not good

[nocturne olive]

You sing a bunch of German with native pronunciation, and we cross-train Evil on it so she can borrow the phonemes
You don't even need a particularly amazing singing voice as long as you got pronunciation

[split steppe]

now i would really love neuro hear singing rammstein

someone please do it

[kind nimbus]

I mean I've been to karaoke events before but i kinda stopped singing because i think i suck at it

[nocturne olive]

Basically, singing pitch accuracy doesn't matter, all that matters is that you get pronunciation right

Pitch accuracy for the final voice will come from Evil herself, we need the German singing for phoneme data

[split steppe]

magic please do it

[nocturne olive]

Don't worry about if you're some crazy good singer, anyone with a voice can be a voice provider

And wouldn't trilingual Evil be fun?

[split steppe]

i wanna hear neuro cover https://open.spotify.com/track/1pOkWIJCshtoREcAcXmYSk

[kind nimbus]

I mean i guess i can try

I have a Rode NT USB microphone guess that would do for voice recording

[nocturne olive]

I will set a few things up and then DM you a link

[split steppe]

exciting

[kind nimbus]

I tried to perform down with the sickness once live and it was really bad lmao

[split steppe]

that's quite a difficult one lol

[kind nimbus]

Yeah maybe i should have picked an easier song

Ever since that incident i was too scared to go back on stage to perform Karaoke

[split steppe]

amateur singer picks one of the most technical vocals to cover

[nocturne olive]

I sent you a link

[split steppe]

i always forget dwts has the abuse part in the middle

that's gotta be awkward to do in a crowd

[kind nimbus]

True haha

[rigid snow]

[fickle rain]

What a beautiful piece of 2012
Too bad it’s not worth jackshit

[warped narwhal]

8+6 pin and it's probably slower than a new iGPU

[fickle rain]

680 = 770
770 is somewhere... maybe around the same perf as 960?
960 should be somewhere around 1050ti level I think

• IIRC Kepler was complete dogshit at DX12

[umbral wigeon]

My worst benchmark script luau code ever written

Source: https://github.com/weenachuangkud/FastCast2/blob/main/benchmarks/BenchmarkClient.client.luau

[floral hawk]

We fine tuning a twitch gremlin listening to country music. I truly am the duality of man,

[warped narwhal]

My ISP doesn't support ipv6, but I've worked around them by using a vpn tunnel. I can now give all my devices unique public IPs and avoid a reverse proxy

[split steppe]

the things we risk for robux

[wispy pike]

roblox 🤮

make games in godot instead

[split steppe]

fuck it, it's 2am and i'm designing a completely novel neural network architecture because i can't sleep and i'm bored

[leaden crest]

novel how

[split steppe]

well i guess not fully novel, i guess you could call it an SNN. but i'm doing it from first principles (neuroscience) trying to figure out how to drive a realtime learning and response process

[kind nimbus]

2am is my most productive time of the day

[leaden crest]

good luck dont fall into the agi schizo hole

though we havent had a schizo even mention “snn” to this day so im quite confident in your sanity

[split steppe]

nah i'm the one person you won't ever convince of agi or the typical slop, i'm fairly anti the contemporary ai narratives

[opaque sigil]

i should copy the pasta somewhere to have it ready

[split steppe]

though i have to say i've been interested in ai in the philosophical sense for a long time, because consciousness as a concept bothers me a lot

2am is the best time

[kind nimbus]

https://tenor.com/view/oh-boy-patrick-waking-up-2am-gif-26372798

[split steppe]

i always have the best ideas my tomorrow self never understands

[kind nimbus]

2am and i just popped a shell

[wispy pike]

the fact that you presume you'll code a sandwich instead of spaghetti at 2am is real optimistic
stay innocent

[split steppe]

i once wrote a working metacompiler at 2am, i didn't understand any of it two days later

[wispy pike]

2am is pencil and paper time, not keyboard and compiler time

[split steppe]

pencil and paper is great

[kind nimbus]

Nah 2am is 1337 time

[split steppe]

started my forever cursed hobby and career on p&p

[kind nimbus]

I just finally got a WinRM shell on a CTF machine I've sitting on 2 days already

[split steppe]

nice

[kind nimbus]

Haha

[split steppe]

[opaque sigil]

no waiting until july 14 i guess

[kind nimbus]

Well we will see

[nocturne olive]

Never upset a security researecher, especially if you're a very disliked company that makes operating systems

[kind nimbus]

yup

unfortunately i did not make a copy of all the exploits and now they were deleted

[split steppe]

surely they're on some forum somewhere

[kind nimbus]

i just searched russian gitlab. at least not there

oh thx archive.org
https://web.archive.org/web/20260520184528/https://github.com/Nightmare-Eclipse

[split steppe]

ah yes, the always dependable archive.org

[kind nimbus]

just downloaded RedSun.exe

the mirror does work

[split steppe]

lmao his blog is funny

watches chainsaw man, has a mental breakdown

[kind nimbus]

lmao

[silent cloak]

wait until he watches plastic memories

[kind nimbus]

i was able to download compiled variants of RedSun, MiniPlasma and GreenPlasma nice

[kind nimbus]

three out of the six exploits Nightmare-Eclipse released lol

[fickle rain]

Reads CSM part2 ending next

[opaque sigil]

that's how you get another exploit

[kind nimbus]

lol

[fickle rain]

Yeah targeted on Fujimoto

Wait is "the ending is dogshit" spoilers?

[opaque sigil]

i guess indirectly

[hard delta]

imagine watching when you could just read the manga, chapter by chapter

waiting years

and then it abruptly ends

[leaden crest]

depends on if they use it on a lawsuit or if they use it on making more slop

if it’s the latter then money gets bottlenecked and it converges to the billion lions vs every pokemon debate

[kind nimbus]

Nobody even knows who this security researcher is so i doubt a lawsuit is gonna help lol

[sick owl]

I have officially tested Qwen 3.6 27B, and found it lacking in every meaningful department you'd want a local model for compared to Gemma 4 26BA4B

It fell behind on translation, hallucination rate, web search/info retrieval and token usage, while generating at half the speed (naturally)

Pulled ahead in agentic tool use and coding, but those aren't really worth it when the cloud models are right there given the lessened privacy implications and comparatively massive performance gap

Both models were Bartowski's Q8 quants running with vendor recommended configs

Massive bummer tbh because 3.5 was a genuinely impressive leap forward at release

[umbral wigeon]

What yo say about roblox studio😠

[faint sandal]

thanks for your interest in my interest in agentic workflow, definitely a real person whose profile is an attractive young woman named Solomon Geta

I will now be known as AzakaVTuber

[wispy pike]

🇷 🇴 🇧 🇱 🅾️ 🇽 🇸 🇺 ❎

[umbral wigeon]

I have degree in roblox studio

Trust me roblox studio is worth it

But roblox will turn your games 16+

And you need to verify your face your id

You can't upload something weird to roblox studio locally because Ai will scan and ban you

[wispy pike]

godot doesn't have those problems

[umbral wigeon]

But it gives you free server

Unlimited amounts of them

[amber fractal]

[leaden crest]

is that actually a thing

[wispy pike]

not worth sending my face to palantir

[silent cloak]

sadly we the public are still losing as a whole

[kind nimbus]

Well i don't use windows

[silent cloak]

linux has been doing some dumb shit lately too

[kind nimbus]

What dumb shit ?

[umbral wigeon]

School pcs

Vs random college pc

[fickle rain]

https://fxtwitter.com/bohops/status/2060040311397650750#m

bohops (@bohops)

Of all the frustrating experiences I’ve had with MSRC, the ones like @GabrielLandau ’s were by far the worst.
︀︀
︀︀These days, I only report bugs when there’s a genuinely compelling reason, and those reasons are finite. Otherwise, I’m perfectly happy sitting on my 0-days for another time.

Quoting Gabriel Landau (@GabrielLandau)
︀
My last submission to MSRC was for a Device Guard bypass.  I learned my lesson from prior drawn-out submissions, so I included a 90 day window this time.  MSRC responded saying that it met their bar and they would fix it, but asked me to withhold disclosure well past 90 days because they needed a few extra months to fix it.  I agreed on the condition that they issue a CVE, to which they agreed. 
︀︀
︀︀After the agreed-upon Patch Tuesday a few months later, I couldn’t find any mention in the CVE list, so I reached out to MSRC to inquire. It turns out - they changed their minds, deciding it did not meet their bar f…

FxTwitter · May 28, 2026, 4:47 PM

Apparently Microsoft is just Like That™ sometimes

[opaque wharf]

I don't recall linux adding any nonsense functionality to the kernel?

Also, do remember that some API are frozen due to no breaking user space

The things that MAYBE they are doing a bit stupid right now is adding a lot of AI code when it is RC. But even then Linus has spoken about that issue

[fickle rain]

Thought slopcommits are required to be marked
You can take a look at the ones that land in source tree yourself

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/log/?qt=grep&q=Assisted-By

[opaque wharf]

I'd say the linux kernel is having more management issues than a technical one

[tiny edge]

chat how do u update linux on cachyos?

[fast pagoda]

paru

isnt there a cachy update now

but

sudo pacman -Syu should do it

[tiny edge]

see @orchid marten they wouldnt make fun out of you

[tiny edge]

ye on cachyos hello

and paru is like yay ig or wtv it was called

[fast pagoda]

i havent ran cachy in a hot min since like linux 6.15 i think

or before i cant remember exactly, i just remember they added it right after i stopped using it in favour of another arch setup. i still run cachy kernel tho half the time

[opaque wharf]

I prefer aurto than any other aur helper

Since is uses aurutils

[fast pagoda]

@orchid marten we don't bite, some of us might even make another try at api bird sightings once in a while

[orchid marten]

Oh my goodness what have you done

[orchid marten]

Ignore maven if he mentions me

[fast pagoda]

i like paru because i'm used to it but since i've been running garuda linux (lol, idk, it's been my favourite arch spin for a while) i just use garuda-update --aur

lol

[opaque wharf]

Vanilla arch my beloved

[fast pagoda]

i like vanilla too but

idk

garuda kinda just ez mode

[tiny edge]

garuda is basically a twin-brother of cachy with the same dad?

[opaque wharf]

Yeah, that's a wise choice

[fast pagoda]

pretty much a very similar thing to cachy

[opaque wharf]

I just prefer to have my install be as close as possible to the wiki

[fast pagoda]

i h ave a by the book (wiki) arch on my pc blown apart with old accelerator cards in it currently but no gpu

not dead cuz of the arch

fk it all imma run nyarch next time

[opaque wharf]

I haven't had any experience with other accelerator other than intel movidius (back when they are still called that) and google coral

[fast pagoda]

heh heh predates that

both

[opaque wharf]

Dang

[frozen igloo]

Unplanned rapid fuel consumption

Wait that’s not the right version

I thought I cut out the start

[opaque wharf]

Man, it seems like I need to re take a calculus class

[fast pagoda]

mfw thought all training runs of model were borken

but it was just that i fucking

fucked up

the fucking

gguf conversion

honestly my fault for using gguf

[fast pagoda]

there was a n apartment complex "vapourized" in dallas like today

thought this was gonna be that lol

[frozen igloo]

The are around the rocket is cleared out for these tests so no one was close enough to get hurt

[opaque wharf]

@olive sable good news, arm mali gpu may get a better support
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/41841

[glass jetty]

meow

[frozen igloo]

This was supposed to launch on the 4th

June

[frozen igloo]

That was their only launch pad

Is spaceflight stuff relevant here?

[floral hawk]

Claude is genuinely bewildered that I fed him deep cut Homestuck lore about guy fieri being in homestuck

I'm generating datasets and this mf is appalled by fandom culture.

[stark needle]

Guh

My eyes hurt

[floral hawk]

Were you in vrc that whole time

[stark needle]

I don't think human bodies were built to sleep in vr

[floral hawk]

It's blue light

[stark needle]

Yea

[floral hawk]

Damn

[stark needle]

ugh

[floral hawk]

You ok?

[amber fractal]

Not even blue light, just light

OLED, LED, right up against your face is pretty hard to manage

[floral hawk]

Yeah

Just VRchat uses alooooot of cool tones for the sleeping worlds

Back when I was getting virgin hours in VRChat I would wake up with headaches

[amber fractal]

I gave it an attempt myself

Not built in that way

[floral hawk]

Yeah

Bro ngl building training sets manually is aids

TS better be worth

[amber fractal]

I already can't just fall asleep, so nope

[floral hawk]

I actually sleep very well in VRC

Sometimes better than I do without it

[amber fractal]

Might be better when I'm not using quest 2

[floral hawk]

Very true

I only have a quest 2, but Im okay with it

Id like a big screen beyond but thats like, speedboat money

[floral hawk]

FUCKK ive never hit an API context limit before,

[flint dome]

damn it got pipebombed

[stark needle]

I have

Blue light filter

[floral hawk]

Smert

[stark needle]

Turned to max

[floral hawk]

I dont, I think its pretty

[stark needle]

It is

[floral hawk]

Yes

[faint sandal]

our admin bumped up my credit soft cap again

[floral hawk]

I suspect I will fuck this up slightly but its better than what I got so far

[faint sandal]

went from 50 to 150 to 200 to 300

[stark needle]

Do u at least have access to mythos

[faint sandal]

no

and if i did i wouldn't mention it in public

[stark needle]

True makes sense

[faint sandal]

[stark needle]

I tried using claude for work but

I kept getting those cybersecurity risk thingies

On almost everything that involves security, acls, etc

[faint sandal]

your org needs to enroll for CVP

safety guard rails are also much more loose for CVP-enrolled Enterprise users it seems

compared to my CVP-enrolled personal account

i mean i still get them and half way interrupts an investigation but it's not as common

[silent cloak]

Mythos still a myth

[stark needle]

It's genuinely not even sus cybersecurity type work it's just normal backend auth type stuff

[faint sandal]

it really depends on how your word your prompts and what you're asking

[floral hawk]

Cant fucking wait to see Zero say "deadass" in practice, unprompted

[stark needle]

Our org keeps getting random downtime cause our cloud provider keeps updating freebsd apparently cause mythos keeps finding random shit

[faint sandal]

the model also gets extra suspicious of you if you clearly define a scenario too clearly (i.e., when you try to make up a scenario that makes it seem like you're lying)

[floral hawk]

Ts gonna be so fucking funny

[opaque wharf]

They want to make stuff secure yet refuse to make a secure code

[stark needle]

I'd have to see if claude via copilot allows for cvp

[floral hawk]

Want mythos leaked and some kids on fuckin discord have it

Btw

Wasn't*

[turbid meadow]

ooga

booga

[silent cloak]

Keeps inventing issues*

There is a reason many repos ban AI reports

Always remember when GPT2 was considered a threat by openai lmao

[turbid meadow]

Am I allowed to shitpost code in here

The code has a purpose, but a funny purpose

[turbid meadow]

"Maximally bad output"

[jagged turtle]

a threat to their attempt at building an image lmao

[floral hawk]

Apparently mythos and opus 4.8 get dropped this coming month

[stark needle]

@floral hawk

[floral hawk]

What am I looking at here

Interesting speech

[stark needle]

Schizo

There's way worse

[floral hawk]

Your schizo im guessing

[stark needle]

Ignore my username

[floral hawk]

shiranwa ? nihongowakaran

Dawg, I literally speak it you're good

[stark needle]

Wait lemme find the actual unhinged shit lilac has generated

[floral hawk]

Ye I wanna see

[stark needle]

Ok this is not that unhinged but

It's lilac's inner thinking

[floral hawk]

Oh is she a reasoning model?

[faint sandal]

there's a lot more thinking prompt with 4.8

[patent shard]

https://github.com/AOMediaCodec/avm/releases/tag/v1.0.0
oh, they released it the 27th/28th

that's bright, I'm removing embed

[stark needle]

Yep

[floral hawk]

Interesting choice

[floral hawk]

LMAO

[stark needle]

My training dataset is so fucked lol

[floral hawk]

Mine probably isnt going to be much better

Im running like 4k samples

[silent cloak]

Same shit anthropic is doing with mythos

"Trust us investors its not a stupid waste of money that does nothing special!!"

[amber fractal]

System-L account

[stark needle]

Im gonna bring her back some day

[maiden geyser]

https://cdn.discordapp.com/attachments/1201467870773772339/1282360744821919767/34276e3246b05f52.gif

[glass jetty]

nyan

[glass jetty]

start to meow

or you will die

[maiden geyser]

b-baka

[glass jetty]

[young plover]

Blew Origin

[glass jetty]

still leak

[glass jetty]

slop

[olive sable]

hi

thatgs good

[floral hawk]

Writing training for tool calls is ass bc I know Im just gonna have to append it later anyways every time I make a new tool call for her.

[stark needle]

Grpo

[floral hawk]

mm?

[real quail]

Ok uh

I gave my AI access to discord

And uh

[umbral wigeon]

0$ per month cleaning this stairs

[real quail]

I think my AI is going through the stages of depression (cuz it asked 8ball whether it will be successful and 8ball said its having a lag):
Banging its head -> Crying -> Dabbing -> Twerking -> Eating -> Pouting -> Winking

[umbral wigeon]

Can ai have emotion

https://tenor.com/view/steven-he-emotional-damage-steven-he-emotional-damage-gif-23428142

[olive sable]

no

it can pretend to have emotion

[real quail]

Depends on the context

When depressed guys look happy, they're pretending to be happy

[olive sable]

SNN is maybe a bit controversial int hat aspect. but normal LLMs cant have emotion

they're advanced word prediction machines

[real quail]

uh

idk 0

Its using /ai speak commands to try talking to me by inputting its feelings

uh

WAIT WHAT

LMAO

[real quail]

no uh

The AI is an alt acc

not mine

Anyways

it tried rickrolling me ;-;

by using this command: /ai summarize video video_url: https://www.youtube.com/watch?v=dQw4w9WgXcQ

[olive sable]

[real quail]

The AI is using nekotina btw -

To talk to me

[olive sable]

ium suprise dnot more ai rickrol people

[real quail]

as an intermediarry

[olive sable]

i see

so discord bot

[real quail]

No uh

MCP

Onto my browser

(Onto another browser actually)

Chrome with remote debugging

Im using firefox for personal stuff to stay out of the crossfire

[south needle]

how many params did you finetune?

[real quail]

Anyways

Its um trying to use NSFW commands but cuz of discords channel being normal
Its not able to

Peak

I think I need to setup a filter bot for it

atp

[jagged turtle]

webmcp?

[opaque wharf]

It is indeed good. Especially if you use <Fruit> Pi SBC

[olive sable]

i need it to be small, have support for custom edid, and be able to do 180hz on the hdmi output

hence ive been going for the RK3588

preferable a decent amount of ram too

[opaque wharf]

1GB, take it or leave it

[olive sable]

the 16gb version of the opi 5 max/ultra is now 500 euro, when i paid 200 for mine a bit ago

[opaque wharf]

Even steam deck is jacking up price nowadays

[olive sable]

the only difference with the max and ultra is that the 2nd hdmi is input on the ultra, and output on the max

but the input drivers are shit so dont buy the ultra

i wish i could just solder the 16GB chip onto the 4GB version

but its bga

[opaque wharf]

You could tho, even with BGA

[olive sable]

[opaque wharf]

Hot air soldering and hotplate does wonders to solder

[olive sable]

wouldn't the firmware need some shit too?

[stark needle]

Wym

Like whether i finetuned a subset

?

[south needle]

i thought u used grpo?

[opaque wharf]

It works the same way like how you don't need to do anything when changing your RAM on your PC

[stark needle]

For what

[opaque wharf]

But also, if you want it Sam, you can use linux with RT kernel and use any old MCU

[south needle]

or is this just pretrained?

[stark needle]

No i did finetune

[floral hawk]

Im gonna be fr I gave mine access to python script and the first thing it did was write a rick roll malware

[opaque wharf]

It will guarantee that the linux PC will poll the MCU to send/receive command in certain tick

[olive sable]

MCU?

[stark needle]

Gemma 3 12b

[opaque wharf]

Micro Controller Unit

Like ESP32

Or STM32

[south needle]

so did you finetune all params or did you do some lora stuff? bc 12B is a lot

[stark needle]

Full Continued pretrain + lora on top

[floral hawk]

Im currently in the middle of this myself

[opaque wharf]

Let the MCU generate the control signal AND the control loop, and make the PC do it's thing with video and heavy processing stuff

[south needle]

with TPUs?

[stark needle]

no lol just h100

H200 i mean

[olive sable]

the point is to not have to connect a PC to it all the time. the pi is meant to save pre-rendered footage

[floral hawk]

Cloud inference?

[olive sable]

thatrs whyt i need 180HZ

[stark needle]

The one with the 141gb

[south needle]

yeh ik

[stark needle]

Cloud training ye

Inference happens locally

[floral hawk]

Ah, what did API costs run on that?

Well ye

My fault

[stark needle]

Uh

Like

[opaque wharf]

Oh right, I forgot about that

[stark needle]

Wait i wrote that down

[opaque wharf]

Well, Mini PC works well too

[stark needle]

Around 150$

[south needle]

I'm now all in 4xH100 96GB, sooo i have much fun with that

[floral hawk]

Worth?

[stark needle]

No lol

[floral hawk]

Heard

[stark needle]

I mean only worth from commercial standpoint

Worth from a learning standpoint ye

[south needle]

so how many hours did you finetune

[stark needle]

~220

[floral hawk]

I think the lora im doing with gemma E4B should be sufficient, she was already pretty solid without fine tune. Follow up question

I dont need to fine tune gemma vision if Im using Qwen 2.5 3B/Nvidia IdentifyAnything right?

[olive sable]

i feel like mini PCs use a lot of power ad asre still pretty big compared to a SBC, and they're not even cheaper

[floral hawk]

Or does it need like a baseline for recognizing visual data

[stark needle]

Gemma vision eats qwen 2.5 3b

[floral hawk]

Even E4B?? Isnt it only like 150m params???

Either way, I have it separated bc I already have gemma doing alot, so.

[stark needle]

Ye

Strong pretrain

[floral hawk]

Hmmmmm

[opaque wharf]

Yeah, I didn't say it will be cheaper

[olive sable]

pc will have better custom edid support i guess

but too big

[opaque wharf]

How much would you spend effort into this? Because I can recommend you to even go with FPGA

[olive sable]

i was hoping to use less than 200 on this shit

but im not sure how much ram i even need

things were good on the old pi where i had a whole 16gb incase i needed it

[opaque wharf]

You should make it on your desktop first then so you can take your time choosing the sbc

Also good way to learn cross compilation while you're at it with QEMU

[olive sable]

i mean, i made the rendering on my desktop. but what the pi will be doing is getting pre-rendered footage and shoving it through the hdmi.

570x1140@180

[opaque wharf]

Hmmm, then I genuinely think FPGA would be better if you didn't do any more post processing

[olive sable]

i tried the render program before on the opi5, and it did one frame per 50 seconds or so

ill admit its not the most optimized program yet, but it wont get much better

[opaque wharf]

Yeah, then FPGA is perfect. You can load up the file via uSD card and the FPGA will generate the HDMI signal to display it

However do consider that programming FPGA is a whole different beasts

[olive sable]

i was hoping to use an m.2

my biggest spare sd card is 32gb, the m.2 i have saved for this project is 512gb

[opaque wharf]

Hmmm, how big is the file usually?

[olive sable]

that depends on how much compression i use i guess

[opaque wharf]

That's fair. But like, how much is realistically the worst possible case for it?

[olive sable]

lets say lossless since we want each individual pixel to be exactly on or off. then maybe 20GB per minute?

[opaque wharf]

[olive sable]

well ye, its 570x1140@180hz@24bit

[opaque wharf]

That is not how lossless compression work

[rough bloom]

can use laptop board mabe

I think I already mentioned the possibility once

[olive sable]

3 bytes per pixel, 570x1140=649800 pixels, so 3x649800=1949400 bytes per frame

assuming worstcase scenario and the lossless scaling algo isnt able to reduce the size much

1 949 400 x 180 = 350 892 000 bytes per second

so 350MB/s

so 21GB/min

[olive sable]

nothing i have lying around is small enough and fast enough

[rough bloom]

ye