programming

[maiden geyser]

it's ok to lurk

[unreal coral]

Or a serious voice

[stray dragon]

don't see appeal

[rigid snow]

v3 soon

[unreal coral]

Whatever the situatio would call for

[iron steppe]

Of course it is!
i just find it funny in some way

https://tenor.com/view/komari-non-non-biyori-uwu-scared-anime-gif-17332970

[rigid snow]

this is #baking ye

or more specifically #tech-nerds-baking but don't tell anyone

[stray dragon]

what situation calls for a different voice

[unreal coral]

I'm not talking about different voices completely just different pronounciatons maybe

[rigid snow]

v3 is priority

[unreal coral]

Mostly the same voice but they could pronounce things differently

[maiden geyser]

quriky

[rigid snow]

hello

real human

[stray dragon]

you can tweak some controls on the original voice i suppose? not really an easy way to get what you're describing

would also probably sound weird

v3 voice might add the different tones as well

[rigid snow]

the same way she changes her speech speed she could change tones

is what they mean i think

that's an "easy way"

[unreal coral]

Yea kinda like a
V3-scared
V3-normal
V3-questioning
Different modes kinda

[rigid snow]

in an ideal world she would just output speech tokens and have all of those natively but

vedal is a scaredy cat who does not want to touch or breathe near neuro's base model

[stray dragon]

"hey vedal we need at least 2 more versions of v3 voice"

[rigid snow]

because the top 1% of nn would be mad (3 chatters)

[unreal coral]

Well thats the thing

I have no idea how much effort it would actually be

[stray dragon]

it is possible theoretically but: sounds like pain

[stray dragon]

would need to collect even more training data for it

[unreal coral]

The only time i ever coded smth so far is for a game with an api in Json

[rigid snow]

i wonder if he's experimented with this

at all

[stray dragon]

he did intelligence upgrades a while back

pretty sure he did something with base model there

[rigid snow]

i mean voice as modality, not touching her base model

i know he did

[stray dragon]

ah

[tacit sage]

Filtered

[unreal coral]

Couldn't you also get smth done with editing the audio instead to make them sound scared?

And just slapping that on the v3

[stray dragon]

that's the easier way but good luck with making that sound natural

[unreal coral]

Like changing the pitch etc

[stray dragon]

would probably just sound weird

[unreal coral]

Oh

[rigid snow]

what's the state of omnimodal models rn

[tacit sage]

Python wont open

[stray dragon]

different tones of voice is more than just "change the audio a bit"

[tacit sage]

upsettie spaghettie

[stray dragon]

words get pronounced differently, emphasis happens differently, timings change

not easy to do with audio filter

[rigid snow]

you can feed those as parameters to the tts model

[stray dragon]

hmmm

possible then

[rigid snow]

well you technically can, what you can or can't depends on the model of course

most of what was mentioned i would guess isn't possible

with, let's say, evil's tts

[minor heart]

i made a funny html file how do i post it

https://the-whimsical-button-1120f864.base44.app/

and the ai named it whimsical button thats what mari mari always says and i love watching mari mari you cant make this junk up

[maiden geyser]

vibe coders aren't beating the allegations

[ivory tiger]

is neurosdk uptodate does anyone know for sure?

like are the latest game integrations using that also?

[amber fractal]

I see no reason why not, the only one that is debated is minecraft.

[minor heart]

its not really coding its comedy

haachamachaCHU

i summon you!

...why didnt it work

[amber fractal]

self botting spotted

[minor heart]

how can i prove im not a bot ;_;

[amber fractal]

[maiden geyser]

not you

[minor heart]

...man i dont understand anything

if anyone wants to play hardcore diablo 3 send me a message

[amber fractal]

we really don't get that much information, we just have a better guess than most people

[maiden geyser]

a user posted a message, deleted it and has selfbot thingie in profile

[minor heart]

oh... i see

[amber fractal]

also why would he not update the sdk

[rigid snow]

this one

i instantly clocked em

[amber fractal]

can confirm

imagine failing to timestamp

[rigid snow]

that's supposed to be {os.time()} maybe

[amber fractal]

not only did they forget a f, but also failed to use normal timestamps

[rigid snow]

or not

[amber fractal]

iirc time is a struct, let me pull up a python

[rigid snow]

oh then it's fine ye they just forgot an f

idk how you can forget if you see it fail to properly syntax highlisht

[amber fractal]

nope, function named times in 3.13

weird

where is the correct time because that is not it

[rigid snow]

[amber fractal]

os.times()
posix.times_result(user=0.09, system=0.01, children_user=0.0, children_system=0.0, elapsed=4296563.45)

os.time is not defined

[ivory tiger]

ok, cuz i see it has latest update recently so im assuming it will work

[rigid snow]

afaik the spec itself never got breaking changes

so any existing lib should work

[amber fractal]

if it didn't someone else would be raising hell at the moment, mainly the https://discord.com/channels/574720535888396288/1350968830230396938 folks

[olive sable]

https://tenor.com/view/date-a-live-iv-kurumi-shadow-show-up-season4-gif-25405342

hi

[amber fractal]

datetime.datetime.now().strftime("%D at %H:%M")

'01/07/26 at 14:49'

I somehow failed to remeber that it is in node.js

so that isn't even correct

[fickle rain]

Actions priority works, but keep in mind Neuro still has latency before responding

[maiden geyser]

<t:{time.time()}:f>

[jagged turtle]

should be

[jagged turtle]

[jagged turtle]

neuro api is just a json over websocket protocol, you're likely thinking of the sdks in which case the last commit is a rough indicator of if it is updated or not

[jagged turtle]

action forces always had l*tency, latest spec change just introduced a way to cut neuro off to reduce time between force and action

[bitter pine]

can I hide the chats section on discord?

[olive sable]

its been ages since ive watched it

what jsut happened?

[round sage]

What happened?

[olive sable]

you deleted your messages

[round sage]

No I didn't

I just got here

[amber fractal]

that's not how that works

[olive sable]

you even ghost pinged someone now

[round sage]

Must've been the wind

[boreal dragon]

Hello that was me

[round sage]

OR hallucinations

[boreal dragon]

Ghost pinged

[round sage]

Mrs nightmare it was a mistake

[olive sable]

[boreal dragon]

I assume this is why

[round sage]

I pinged you to show you this gif

Yes

[boreal dragon]

Lovely entrance

[round sage]

I remember having my recommendations flooded with Kurumi edits

Years ago

[boreal dragon]

yeah, happens to the best of us

I personally don’t mind it (could never guess why)

[round sage]

I could tell

Your favorite character

is definitely Yoshino

[olive sable]

im so confused rn

[boreal dragon]

Actually, it’s Kaguya and Yuzuru

[round sage]

Hmm maybe it's Rio

[boreal dragon]

Kurumi is just the best by far for pfps

And more people know her

[round sage]

The Date A Live fandom are getting flooded with AI arts rn I think or has it stopped

Last time I've seen a dal fan was 2024

[boreal dragon]

Yeah, it’s still going

[round sage]

Holy shit

That's crazy

[boreal dragon]

I played the mobile game as a kid

[round sage]

Spirit pledge or the new one

[maiden geyser]

you're blatantly self botting mai

[boreal dragon]

And after I realized there was a show in recent years I felt like I had to watch it

[boreal dragon]

Spirit Pledge

[round sage]

All that money just for it to shut down

There's a new game I think

[boreal dragon]

The new ones a remaster, I’ve gotten into that. Spirit Echo, I believe.

[round sage]

I forgot the name

Oh yeah

Spirit echo

[boreal dragon]

Same game

[round sage]

And one more

New project

[boreal dragon]

Yeah

[round sage]

Yea basically the old spirit pledge

[boreal dragon]

Brand new

[round sage]

I bet it'll shut down again earlier

[boreal dragon]

Won’t affect me really

I haven’t put any money in

I put in time, not money

[round sage]

I spent a lot back then

Doesn't really matter anymore

[boreal dragon]

Yeah

[round sage]

Since I've earned it back and more

This is programming, we should really be talking in #gaming

Or general

[round sage]

Could you send me the dal server in dms

[boreal dragon]

Alright, ping me in whichever you wanna

[amber fractal]

I already threw in a report, will see how it goes

[jagged turtle]

they banned

[olive sable]

[opaque sigil]

[boreal dragon]

kurumi

[jagged turtle]

did someone forget how to string template in js

or they are going from py to js

[lilac apex]

How can you make an account in discord using python

Aren't you supposed to use HTTP

[jagged turtle]

it's called spoofing a browser

[lilac apex]

Coooooooool

[amber fractal]

[lilac apex]

Didn't discord patch selfbots

[olive sable]

[lilac apex]

By making user tokens invalid

🥀

[amber fractal]

Ironic

[olive sable]

"This Account was randomly generated..."

[lilac apex]

It's def just a kiddo who wants to look like a bot

[olive sable]

that would be you tho

[lilac apex]

BOOT.DEV

IS A LEARNING SITE.

Oh shit yea I do have it as my bio

[jagged turtle]

what are you talking about

[amber fractal]

yeah not the best of looks there

[lilac apex]

User tokens cannot be used as a bot token

[amber fractal]

there is nothing stopping a user from botting a user token as it's own thing

example being the person that was banned

[jagged turtle]

They're spoofing a browser??? As @amber fractal said there's nothing stopping someone botting as a user

[lilac apex]

Hmmmm

Oh ok

Really should learn Programming more often again

I've lost my braincells

[amber fractal]

it's a fair assumption

but people will use any tool to get around problems

[lilac apex]

Might've forgotten programming after being forced to study biology

[amber fractal]

[quick condor]

Tbh if your spoofing a user token for a bot you probably have brain rot. I can't think of a valid reason to do so when a bot token can do everything you need and is easier to work with

[stray dragon]

you need permission to add your bot to other servers
you don't need permission to add your own account to other servers

[jagged turtle]

^

[severe path]

You can just specify a username and password via the Discord API, no need to spoof it.

[jagged turtle]

you can but there's likely some security checks for the API that would make it harder

[quick condor]

Right but generally if a server wants a bot in it they will approve it, if not they wolnt. So if you are bypassing the server's rules then it's not a valid reason

[severe path]

I don't know about now, but I self bottled once as a test. All you needed was a username and password.

[jagged turtle]

they've changed their security stuff a couple times I think, there's some form of captcha/turnstile challenge at the login page that needs to be submitted alongside the usr + pwd, I think?

[severe path]

Again, I'm not sure how the API has changed over time. I would have to read the documentation.

[jagged turtle]

idk I haven't really examined discords login page for a while

[jagged turtle]

is there documentation for the discord user api???

[quick condor]

Idk I just use an app api token because I use it properly. I have zero reason to connect my AI to a server I'm not explicitly allowed to by the server owner. That's asking for problems

[severe path]

I don't know if there is public documentation, to be honest.

I once self botted back when I was a script kiddy. That's it.

[jagged turtle]

someone probably has started unofficial documentation somewhere

[quick condor]

I'm pretty sure the official documentation is: Don't

Just gave my model stack PSE Api access

[severe path]

What is that?

Is it this?

https://en.wikipedia.org/wiki/Philippine_Stock_Exchange

[quick condor]

Programmable search engine, Google's api search engine

[severe path]

Oh

Okay

You should've just said: Google's search engine API. Other programmers will understand you better that way, if they aren't familiar with Google's API terms.

[warped narwhal]

I understood what he meant though

[severe path]

Yeah, well I don't really use Google's APIs.

Is there something I'm unaware of?

[warped narwhal]

Google APIs apparently

[severe path]

Hrm...

Do people use Google's APIs all the time here?

[quick condor]

PSE is a pretty common method to augment the fixed knowledge of a model

[severe path]

I knew AI used search engines to augment knowledge. I wasn't aware that PSE is a colloquial term. Sorry.

[jagged turtle]

google apis are kinda annoying to work with

[quick condor]

Yah they are. It's one of those things where Google api is about as vague as it gets. You kinda have to be specific

[scenic depot]

Yea you definitely got your info sent to a webhook 😂

Won't matter anymore since it's old anyways

Oh well

[severe path]

Whoops!

I shouldn't have been a script kiddie

[scenic depot]

[scenic depot]

customized API?

[jagged turtle]

???

[scenic depot]

Oh sorry I forgot to turn off mentions

I just read the photo you sent

[jagged turtle]

customized API probably means they built a discord api wrapper

[scenic depot]

Oh

[jagged turtle]

[scenic depot]

Oky

[sick owl]

In a jar huh

https://www.businessinsider.com/razer-project-ava-ai-companion-hologram-2026-1

Powered by Grok btw

Actual redditor for your desk

[amber fractal]

[maiden geyser]

oddly specific

[supple condor]

I got a question

[obsidian mantle]

[opaque sigil]

[supple condor]

?

[olive sable]

ask

[opaque sigil]

Where's the question

[supple condor]

Oh

How do i set up a obs I'm kinda new to this

[obsidian mantle]

I wonder what will happen if someone says "no" to someone asking if they can ask

Youtube guides

[opaque sigil]

I can't help with OBS sorry

But yeah there are a ton of those

[olive sable]

idk how streaming works but for recording its not that hard.
you gotta add a source first

[supple condor]

Like a url link?

[olive sable]

it can be

[obsidian mantle]

There is a stream key if you want to stream you just paste it into obs somewhere

[olive sable]

the 2nd bar is the sources bar, there press + to add a source

[obsidian mantle]

Then configure scenes and sources of image it probably defaults to your screen maybe not

[azure lynx]

OBS wasn't difficult but probably had something which required me reading the docs at some point. i think it was just a matter of setting the sources and where they should go.
depends on what you want to use the stream for

[obsidian mantle]

Right it depends on how detailed you want it to be

Probably just watching some 5 min video would be more effective

[supple condor]

I think i might be a little under prepared

[amber fractal]

I would also recommend a video

[azure lynx]

i think i just grabbed the contents of a webpage as part of one thing in OBS by pointing the source at the chrome tab. there may have been a better way to do it.

[supple condor]

I need a computer don't i

[azure lynx]

it's not going to work on like a chromebook

[supple condor]

Yea I'm cooked

[obsidian mantle]

no computer?

[supple condor]

Nope

[azure lynx]

what are you trying to do and what do you have?

[obsidian mantle]

Then use some app for phone

Not obs

[amber fractal]

not supported on chromeos ye

[supple condor]

I was going to like set a heart rate monitor up for my stream but i was like i don't know how to do this

And now i just have it

[azure lynx]

there is development mode on chromeos which opens up some extra features but it would require some custom coding on your side to make things work, to glue things together on both ends.

(or at least was one when i last checked a few years ago)

[true hemlock]

5060 SFF for about $240

should i

[supple condor]

Damn this is kinda hard to understand

[patent shard]

I wouldn't
5060 ti 16gb was $370 a couple of weeks ago
but if you don't need the vram, maybe

[obsidian mantle]

Is it possible to connect 8 5060s with nvlink for 128gb

What are the limits

[supple condor]

So um is there a way I could do this on my phone

[amber fractal]

having nvlink

[obsidian mantle]

What app are you using for streaming?

[supple condor]

Twitch if thats what you mean

[amber fractal]

physical connector last seen in the 30 series on the 3090

even then only supports 2 gpus

1080 had 4x

that seems to be the max in general

[obsidian mantle]

Ooh

[azure lynx]

OBS copies the heart rate webpage contents and overlays it with the contents of another screen and sends that to your streaming provider.
what you want is something shows both what you are streaming and the heart rate data at the same time. is there somethign which puts it in a notification window? perhaps one that shows it every few seconds as a new notification so that pops up into the twitch app's capturing somehow?
[i imagine it'd filter out that type of thing by default]

could you do more than 3090? i didn't think you could link them because they were already using two slots?

[true hemlock]

4gb modules on still just 4 imc

[supple condor]

Eh?

[azure lynx]

which i think means "probably not going to happen."

[supple condor]

Oh

[true hemlock]

anyway the reason im peeking for such deals is that i have a plan with a sff briefcase build

and my dealer just so happen to offered that lol

though the 8gb is a bummer but its not like i'd use it for anything crazy

[true hemlock]

not a fan with huge vram but mediocre bandwidth either

not an ai bros

[azure lynx]

i think the bang for buck of 2x 5070 ti would probably be more than a 5090 and cost less/actually exist. two physical gpus would also let you split loads more. that's probably the lowest which can still load a couple of decent sized models. ideally you'd have something with 24 or 34GB but in this economy? i could load the qat-q4 versions of the 12b and the 4b Gemma 3 models on one card and still be able to load in image support and have a massive context window and use the other card for background processing.
currently using my ancient 3090 for everything but only just starting to get into multi-model setups.

[amber fractal]

Honestly, I'm considering a similar route if my own attempt doesn't work out well. Worse case is that I can offload memory processing or smaller parts inbetween cards. Ideally though I'd be attempting to process from multiple sources at once and combine the results.

As general logic stands, each data type has it's own processing into a singular state that can be reasoned about. In an ideal world for my setup, I'd be able to process those parts seperatly and reference them in context of each other before splitting back into heads for each output.

Not quite monolithic, but also not fully seperated

[maiden geyser]

https://fixvx.com/arC_203/status/2007162948633145740/photo/2

arC (@arC_203)

↪️ Replying to @Dexerto

@Dexerto For anyone new to NeuroSama here are some great introduction infographics on how she works.

vxTwitter / fixvx • Media 2/2 · Jan 2, 2026, 6:51 PM

grok is this true

[azure lynx]

the image shows a plausible set of things. specs are directly from twitch page.

[young plover]

Are those specs even up to date? I thought I heard different specs at some point, but maybe that's Vedal's PC

[azure lynx]

i've heard better specs and doubt they are up to date.

i believe it was 5080? possibly two of them?

[hard delta]

bruv these double tap reacts ugh

[hard delta]

also is this just my or is this like google translated

[maiden geyser]

mmmm, i saw a japanese version

https://rxddit.com/r/NeuroSama/comments/1pz4qi2/a_japanese_fans_guide_on_how_neuro_works_credit/

[maiden geyser]

yeah, translated by gemini

[true hemlock]

still 4090

just with 9950x and 192gb ddr5 iirc

and shit tons of storage

also i wanna rant

fixed my pc
dude finally
i fucking hate gigabyte
the boot pipeline is so ass
i need to literally do the equivalent of holding the breaker from tripping with my fingers while having a temporary enormous power draw
aka bypass whatever the fuck mem training there is and rebuild some essential training table just to make the 4x32gb work
im switching board fuck it
z790m or something with better T top memory trace instead of daisy chaining

[hard delta]

that's vedals yeah lemme find it

#1326303411041206303 message

[young plover]

9950X, thought so

[hard delta]

best pick you have rn really
16 fast af smt cores with avx512, good for basically anything

[young plover]

Yeah it's what I currently have

[hard delta]

i just got a pc with 265k cause anything prebuilt that had 99xxX was way out of price range, and i'm not gonna be running all those silly containers and stuff on 8 cores
(and anything not prebuilt was automatically way more expensive and out of price range xD)

i low key wonder how the zen6 / ultra 300 coin will flip tho

[young plover]

The dual 5080s do sound familiar, can't find where he mentioned that though.

That's a weird machine though. Not sure why anyone would do that instead of a single 5090

[azure lynx]

if there was a 5080 Ti with 24GB it would make sense but it doesn't exist, so 2x 5080 = 32GB total would be the same. perhaps it was about availability?

it looks like camila messaged her and woke her up?

and was still in the voice chat from earlier maybe.

[true hemlock]

zen 6 has more cores iirc

im on track to get a 10950x engi sample this year maybe

[hard delta]

ooh

[true hemlock]

yeah about 1.5x more cores

with ccd going from 8 to 12 core

[true hemlock]

265k is probably best value for multicore for a retail chip

though e cores still kind of ass to schedule

but not that it matters for most people

[hard delta]

yeah that's why i went with it
like everywhere they're just mindlessly pushing 9800x3d cause "look video game benchmark"

[true hemlock]

honestly to think about this, the upgrade to z790m is basically just a sanity upgrade since the 4x32gb do work on my b760m but with scuff that i need to watch out this way

[true hemlock]

true

some people actually went out for 9800x3d for productivity because they don't do their research lmao

[hard delta]

but i don't want x3d i want cores, and buying a new pc at about the top of my price range with "will probably have to swap this cpu" is just annoying

[true hemlock]

i don't own any x3d chip either despite owning like the weird niche stuff

daily driving a i9 13900 engi sample

[hard delta]

only thing that made me hesitate was avx-512 support cause it felt like i'm gonna regret not having it at some point

[true hemlock]

same

but well it is what it is, intel suck at implementing avx 512

though i do have a 7950x
but its also... an engineering sample

[hard delta]

but then talked to some guy who did his msc thesis in reimplementing linux modules that did vector stuff on various esoteric cpu architectures and he was like "nah bro avx2 will be fine for almost anything for the next 3-4 years"

[true hemlock]

es2 though so it works almost as good as retail

[true hemlock]

we have GPUs that do the SIMD anyway

[hard delta]

yup

5070ti go brrr

cpu is the docker kindergarten

[true hemlock]

though its not core distributed and not callable with ISA

but it works well either way

[hard delta]

also some people were like

b-b-b-ut am5

[true hemlock]

lmao

[hard delta]

which makes senes but like

i do NOT want to replace the cpu i want it to do a good job for 8 years and then i'll buy a new pc xd

[true hemlock]

my only am5 is my 7950x engi sample

future proofing is overrated

[hard delta]

yeah like if you replace a thing in 5 years you'll be like

[true hemlock]

we also have a zen2 64 core threadripper

[hard delta]

yay new strong cpu - except my mobo is kinda weak now, my ram is kinda slow now, my gpu is kinda old

[true hemlock]

still faster than any consumer cpus in existence in multicore

[hard delta]

and suddenly you're back to wishing you'd just buy a new pc

[true hemlock]

including the M chips from apple

[silent cloak]

Bad time to do it

[true hemlock]

[hard delta]

only time it makes sense to have that replaceable component is when you intentionally buy something that has one weak component, and that's exactly what you'll replace

[true hemlock]

im a fan of only upgrading gpus
not a fan of upgrading cpus to new gen so at that point just get a new system every 6-8 years

though

[hard delta]

exactly

[true hemlock]

i went from 14600KF to my 13900 ES

[hard delta]

and this prebuilt i got still had like, friendly priced ram in it

[silent cloak]

Thats the average policy tbh these days

[true hemlock]

because i found a really good deal

[hard delta]

cause they had them on stock for half a year and aren't entirely evil scalpers

[true hemlock]

and i want multicore lol

[keen hatch]

This is more interesting... they promise support for local LLMs ... https://www.code27.co/

[true hemlock]

bonus point the i9 13900 es is actually more power efficient

and thermal efficient

dunno how

but im not regretting the upgrade

[silent cloak]

Wait for CPUs to crash next

[true hemlock]

too late

already have all the high end cpus

:)

im

lack

of

gpus

:(

[silent cloak]

Na no quantum

Smh

No qbits for u

[hard delta]

hey i found an usb 3.0 external qubit processing unit on temu for $9

[silent cloak]

The radioactive source inside:

[hard delta]

also yeah i'm coming from like, a laptop with a ryzen 4800 and an rtx 3060 so either way this is gonna feel like i moved to a space station

[opaque sigil]

So apparently if nvidia is to be trusted Rubin has 63% higher fp32 than Blackwell, surely that'll transfer over to consumer rubin

[true hemlock]

i don't buy it

[hard delta]

if they wanted me to buy rubin they should have sold it 3 days ago

[true hemlock]

blackwell 100 throughput is like 3x to 6x blackwell 200

[hard delta]

they fucked up

[silent cloak]

Meh most enterprise news from them these days isn't good news for us

[true hemlock]

though that's tensor

[silent cloak]

@true hemlock

[opaque sigil]

fp64 down even more

[true hemlock]

was about to say that

i was expecting them to go literally all in on ai

so they removed all the fp64

or whatever

and bump out fp4

inb4 fp2

its so ass

the HPC peeps had to skip like 2 gen

of enterprise hardware

[hard delta]

true

[opaque sigil]

At least memory bandwidth keeps going up I guess

I'll take it

[hard delta]

i should get an EPYC 9965 ngl i need cs:go to go faster

[true hemlock]

double the core of my 9655

[silent cloak]

U guys see the rumors about nvidia bringing back the 3060?

[true hemlock]

[true hemlock]

i still don't understand their reasoning

how is that gonna help dram shortage

[opaque sigil]

Free money I guess idk

[true hemlock]

them having unused gddr6?

5050 uses gddr6 bro

[silent cloak]

Idk either given the amount in the 50 series

[true hemlock]

blackwell cache is backward compatible with gddr6 imc

[silent cloak]

Atleast the 30 series was a better run than the 50s

But still its crazy

[true hemlock]

they're not even short with the GB 200 die fab

i have a feeling

about them using AD chips for the 3060 sku

happened before btw

[hard delta]

who needs 5070 when

[true hemlock]

on some mobile xx50 chips

some 2050 mobile uses ampere for some reason

[opaque sigil]

Funky

[true hemlock]

B300 is apparently technically the third blackwell arch

makes sense considering the lack of fp64

"blackwell ultra" lmao

Blackwell 1.0 is enterprise blackwell
Blackwell 2.0 is consumer blackwell
Blackwell ultra is enterprise e waste

[silent cloak]

We are gonna have so much e waste when the craze is over

[true hemlock]

i ain't gonna even hoard the B300

i want my fp64

the ai bros in denial after the bubble pop can have the shitty B300

im taking the B200 and H100/H200

[quick condor]

I'd love me some H200s

[true hemlock]

btw

[opaque sigil]

Isn't the b200 Blackwell ultra too

Or am I misremembering

[true hemlock]

have y'all seen the rust ffmpeg ffmpreg

https://github.com/yazaldefilimone/ffmpreg

[true hemlock]

yep you are

B100 is GB102

B200 is GB100

B300 is GB110, except it has different arch than their previous blackwell so its now a separate blackwell

[opaque sigil]

Guess you're right

[true hemlock]

well...

[silent cloak]

Every rust project be like

@true hemlock what's ur opinion on the rubin architecture?

[true hemlock]

i haven't even seen anything about it lmao

too lazy

[silent cloak]

Lots of leaks

[true hemlock]

just give me some screenshots or links

[silent cloak]

Its supposed to be more efficient and nvidia wants it for the 60 series

[true hemlock]

or preferably die shots

[silent cloak]

Which the leak says the 60 series won't be out until like the second half of 2027

https://en.wikipedia.org/wiki/Rubin_(microarchitecture)

[true hemlock]

no 6090 for long

[silent cloak]

https://www.nvidia.com/en-us/data-center/technologies/rubin/

[silent cloak]

I mean it was obvious given the plans for them to scale back this month tbh

They need to power their centers so we can have stuff like this

[true hemlock]

lmao fp4 numbers

[silent cloak]

https://tenor.com/view/onion-ai-bird-explodes-hailuo-gif-14207748951433828719

[azure lynx]

it's nothing like ffmpeg in scope at the moment. only appears to support h264 instead of the dozens of video codecs ffmpeg handles.

[silent cloak]

Most rust rewrites are like that

Best to stick with the tried and true until its able to surpass it

[hard delta]

[silent cloak]

They are saying they want it to replace all future architecture

Well for the foreseeable future

Consumer gpus

[true hemlock]

even the ai is getting slop low quality shi

aight so

[silent cloak]

https://www.tomshardware.com/pc-components/gpus/nvidias-next-gen-rtx-60-series-might-not-debut-until-the-second-half-of-2027-says-leaker-rumor-claims-rubin-architecture-will-power-future-consumer-gpus

[true hemlock]

i assume 8192 bit hbm4

[silent cloak]

Jensen Huang confirmed the chips are already in full production

So it looks like they are locked in

So say hello to the newest architecture I guess

[true hemlock]

why do they have a separate cache die from the dual die

[true hemlock]

yay, another blackwell 2.0

i guess

[true hemlock]

it really seem like interconnected dual die

with extra cache block on one of it for some reason

[silent cloak]

Apparently some systems are already running early models of Rubin

But its supposed to be cheaper to run

Is apparently the only benefit I see

[opaque sigil]

The new norm

[true hemlock]

i still won't be happy with vera

arm chip will suck regardless

[hard delta]

2027 june 14: they release the 6090 right before steam summer sale
28GB of gddr8, $7500

[opaque sigil]

It's a cpu that does cpu things for sure

[silent cloak]

Look at all those VRMs

[true hemlock]

the traces better not have resistance defects

or shit gonna legit blow

[opaque sigil]

In the keynote they said something along the lines of a rack drawing twice as much as a Blackwell one I think?

vrm go brrr

[silent cloak]

I imagine we will get reports soon

[jagged turtle]

whAT

[true hemlock]

"twice power, but 50 pops compared to 20 guys"

[silent cloak]

Some of NVIDIAs partners are running them rn

[true hemlock]

they forgot about me :(

[silent cloak]

No slop for u sorry bud

[opaque sigil]

Too much slandering

[silent cloak]

TSMC is probably still testing them as they make them

If they follow their usual process atleast

[true hemlock]

i need my dealer to check the black market

[silent cloak]

Microslop and CoreWeave

They are the first with em

[true hemlock]

i wanna be the first to leak the full arch and pipeline

[silent cloak]

They are also working with RedHat

[true hemlock]

[silent cloak]

For some reason

[mortal flicker]

Bro for some reason my friend manage to leak my by just adding me in an alt account

And he wont tell me how

[silent cloak]

https://videocardz.com/newz/nvidia-vera-rubin-nvl72-detailed-72-gpus-36-cpus-260-tb-s-scale-up-bandwidth

Here is a better spec breakdown for comparison sake

[mortal flicker]

Help pls

I cant figure it out 😢

[silent cloak]

U probably clicked something

[true hemlock]

226B transistors on the 88 core cpu wtf

did they have their own version of super dense vector isa

[silent cloak]

336bil

On gpu

[true hemlock]

of course the comparison vs blackwell has to be about moe slop

[mortal flicker]

I dug deeper and it has to do something with discords cdn but Idek what that is

You know what i wont grow stronger if i dont find out myself 😡

[true hemlock]

im convinced

[silent cloak]

72 Neuroverse cores

[true hemlock]

that they're just shit at managing transistor density for acceleration/stream units

[silent cloak]

Crazy

[opaque sigil]

There was some talk about weird hyperthreading shenanigans

[silent cloak]

Its a quick solution to a quickly worsening problem

They are trying to lower costs so slop centers keep buying from them

[opaque sigil]

Money go brrrr

[silent cloak]

The architecture they plan on keeping

[azure lynx]

just give me 128GB on my GPU plz. then I'll stop asking for things.

[opaque sigil]

Clearly works well enough

[true hemlock]

the same exact shit i've seen when companies just wanted to be quick in making things and end up overshooting transistor count for accel/stream micro units

[opaque sigil]

Just get an H200

[true hemlock]

willing to bet

[silent cloak]

I wish they could overshoot for consumer grade

[true hemlock]

epyc 7003 has more efficiency per real world performance than the vera or whatever the fuck

[silent cloak]

I think its just down to the cost for their buyers tbh

That seems to be the big thing they are reinforcing

They mention some performance improvements over the Blackwell but thats it

[true hemlock]

epyc 9655 with 96 cores is just 99.7B transistors btw

more transistors = more current needed

[azure lynx]

that's about $32,000 more than i wanted to pay for that much.

[true hemlock]

you want best transistor density, but sometimes you don't want too much

thermal density would be ass

or

current requirement would be too high

[silent cloak]

Yeah idk their game plan

We will see i guess

[true hemlock]

slop

[silent cloak]

Im kind of hoping it does just end up melting itself

[true hemlock]

their game plan is low effort slop

[azure lynx]

fractals with thermal channels probably

[true hemlock]

"its realistic, in the perspective of people with near sighted vision, with RTX"

"and motion blur"

[silent cloak]

MSAA SMAA my beloved

[nocturne olive]

Do note that for basically all workloads 1 GPU with all the VRAM will be better than 2 GPUs with the same total VRAM

[quick condor]

Motion blur is realistic, just slam back the bottle of rum

[quick condor]

I'm loading 2 24b models with 32k f16 context in 22gb on 2 3090tis, an 8b with 32k f16 context + voxtral 3b in an A4000 and mpnet + 8b with 32k f16 context in an A2

[quick condor]

Not necessarily... If two models are running async you will 100% hit a compute bottleneck.

Sure it would be faster when I bounce outputs back and forth between instruct and reasoning models but I also have situations where the different models are all running simultaneously

[nocturne olive]

Bandwidth is the limiting factor, not compute

[tame trail]

Helllll yeahhhhhhh, hello fellow and likely far superior programmers

[jagged turtle]

wdym far superior programmers

[tame trail]

Just assuming most people in this channel are better than me

[hard delta]

most people in this channel just banana

[tame trail]

Banan

[hard delta]

banana 2026

[tame trail]

How do I also banan

[ivory plinth]

id:customize banan

[tame trail]

Oh..

[azure lynx]

do not believe the propaganda. or do. it really makes them happy when you do and it doesn't make any difference. ;]

[tame trail]

I nacho

Uwahh

You also nacho

[hard delta]

no banana

[tame trail]

I support banna

Dont be sad banan personnn

[azure lynx]

Yeah. But i'm just an ally.

[hard delta]

support?

[tame trail]

Ally support

[hard delta]

[tame trail]

Yo what language was neuro coded in?

[hard delta]

AHEM okay i'll exit the mindset where i can do banana campaign speech

[azure lynx]

python for the clever bits, and C# for some of the visual parts.

[hard delta]

the ai stuff is most likely in python like anything

[quick condor]

I'd have to go back and look but I'm pretty sure I hit 100% cuda utilization before 100% memory controller load on two simultaneous models

[true hemlock]

friend didn't believe that i could run my i9 at just 2.4W

[hard delta]

but also neuro is like 16 things wired together

not one smart cookie doing it all smartly

[true hemlock]

and below 4W while youtube is playing a video with CPU encoding

[tame trail]

As with most projects

[nocturne olive]

Inference is bandwidth bottlenecked unless you have some very weird GPU nobody has ever seen before

[hard delta]

#programming message this is not a bad (but google translated) infographic (speculation) :nodders

[true hemlock]

godbin

[hard delta]

she also has some form of vision, unsure if external or not, and then every time you see her play a game there's some integration plugin, usually a mod in the game too

[azure lynx]

even with my embedding multiple models in one executable i've still got the ASR and the TTS as separate processes and the avatar control is definitely getting its own process and that's before the LLMs.

so 4 processes minimum.

[tame trail]

Makes sense and I never even thought about how her vision works but it seems pretty obvious for a llm

[azure lynx]

it's part of what makes it hard for her to draw and then compare what she drew with what she intended

it all goes through text

(basically)

[hard delta]

evil outfit context will be hilarious

[tame trail]

This makes sense, also probably why she had such a hard time playing hide and seek

[hard delta]

i expect her to pick one that is overall ugly but packed with cute details

[tame trail]

This is very surprising to me, most things dont make sense to me especially ai

So am going "GUH"

[quick condor]

Flash attention?
Quantization?
Models not running perfectly synchronously?
Wtf you talking about

[azure lynx]

"find a location such that projecting visual planes over that location will not intersect with my bounding box before intersecting with something else. do that using the following entity coordinates and you've got about 0.5 seconds before people start looking at you."
yep.

[tame trail]

lol

[nocturne olive]

Inference is limited by how fast you can get the model weights to the GPU core when running it, there's plenty of compute on literally any modern GPU that the bandwidth isn't sufficient to saturate the core unless multiple queries are being processed in batches with the same mem loads

[tame trail]

BUH 😵‍💫

Hardware optimization?

[jagged turtle]

C# in Unity for visual model, Python for LLM/probably some other AI systems, various languages for integrations, (HTML/CSS/)JS for control panel

[azure lynx]

but the parts that make Neuro were written in plain English (most likely)

the specific ways she interacts with the universe

and her memories

[true hemlock]

inference is still heavily bandwidth bound. it is what it is with weight matmul. full compute saturation only happens at extreme high ctx window

its why 3090 and 4090 in LLM t/s bench isn't much different

[true hemlock]

despite 4090 having more than double the compute

[tame trail]

Wth am I looking at

Big words

[true hemlock]

how much ctx window

that would make sense under really high ctx window

[quick condor]

I don't remember either 16k or 32k

[true hemlock]

is that also really inference

[azure lynx]

Those are running like a single processor: they're active and not active at the same times.

[quick condor]

That was bathed ops

[true hemlock]

?

[nocturne olive]

The way GPU utilization is reported is funny, 100% GPU utilization can mean either memory or compute but it doesn't tell which

[true hemlock]

try loading a new model and check the graph again on early inference before ctx window gets saturated

[true hemlock]

look at power draw

[quick condor]

Memory controller load is that

[true hemlock]

pretty sure that's compute saturated

[hard delta]

yeah the selling point and also the part we know least about is how vedal manages the twins' memories
a very silly, impractical, don't do this, massively simplified way to put it is that she's just chatgpt, but instead of asking her a 2 liner question, you drop 500k lines of text in as your question, and that text includes all her memories, her life, personality, etc.
she answers in a paragraph that is gonna get spoken on stream, some text on how to move her model, more text on what to do in whatever game she's playing (commands to the integration basically), and maybe on whether to add or change or remove a line in the 500k lines you asked the question with
this is like how i'd explain to my mom lmao

[nocturne olive]

That is a slightly excessive power draw
But also not like my 3090 doesn't also use full power during inference

[true hemlock]

nah

its 3090Ti

its a full GA102 die that's binned

[nocturne olive]

I did see that yes

[true hemlock]

so its made to run at really high power draw even for tiny boost

[quick condor]

That was taking hundreds of chunks of text, queuing an operation and running it across two models in parallel

[nocturne olive]

Funny

Either way, I don't see full GPU power draw during inference being weird

[true hemlock]

should try undervolt on 3090Ti

because its well binned

so you'll probably get really good results

and slightly faster 3090Ti

[quick condor]

I should, you and so many people have told me that. I've just been lazy and the biggest reason to (temps) isn't a problem

[true hemlock]

if done right

[jagged turtle]

what

[true hemlock]

its just few clicks away if you know what to do lol

[azure lynx]

her persona description and prompts are what make her "her".

[tame trail]

Vedal was having some troubles with Neuro during last stream and said "she isnt used to being in vrchat this long" or smth like that so I suspect something like her personality and all that is in a perma library and when she streams she gets semi wiped (maybe even allowed to add some info from the last stream in perma memory) and then set back out with clean ram

[true hemlock]

most people don't bother because they don't wanna spend time looking up how to do it

[jagged turtle]

oh the prompt? yeah fair enough

I guess

[nocturne olive]

Nah not exactly
She isn't purely prompted

[tame trail]

Though I'm very likely wrong because I'm basically bsing

[true hemlock]

she is

[jagged turtle]

it did say language, never specified what type of language

[true hemlock]

i don't buy what vedal said

[nocturne olive]

I'm quite sure Neuro isn't based on pure prompts

[true hemlock]

she's finetuned

[azure lynx]

she might be fine tuned too, but prompting is essential for certain traits because fine tuning tends to blend stuff

[true hemlock]

but also works with prompt

[quick condor]

Basically me rn. I'm so busy with other things that it isn't the top priority

Good stack and prompt architecture is most of it. Fine tuning is mostly polish

[nocturne olive]

Meanwhile for me it just isn't directly possible to change voltages directly

[true hemlock]

you don't even need to

just do clock offset

and power limit

[azure lynx]

fine tuning should let you shrink your prompts but keep the style.

[true hemlock]

that also works as undervolt

[nocturne olive]

Last time I tried it became unstable under load

Crashed after 6h or so of training NS

[true hemlock]

then your voltage curve is already tight

and your die can't do better

[quick condor]

The most recent thing Vedal mentioned about neuro self adjusting sounded a lot like interaction based prompt tuning

[nocturne olive]

It seems kinda like I got very unlucky on all my components, since it seems like something about my overclocks on the CPU or RAM was also causing instability during Twitch playback

[true hemlock]

-225mV undervolt give me average of 2.4W idle

godbin

[azure lynx]

my guess is she has a file she is allowed to edit that gets included in her context

[nocturne olive]

So I've unfortunately reverted back to stock on CPU and RAM with just XMP

[true hemlock]

though

my B760m is shit

can't handle xmp well

on my 4x32gb

for some reason can handle 3600 cl15 on my 4x16

but not my 4x32

im pretty much stuck with 2666 cl19 till i get to replace this board

[nocturne olive]

At least returning to all stock on my CPU and RAM at least so far apppears to have stabilized this thing, finally, after a month or so of daily crashes

[quick condor]

Nah it's more likely a dynamic system prompt that is adjusted by weighted metrics. Those metrics are adjustable at runtime. It's what I've been doing for awhile (technically system prompts are part of context but that's getting nitpicky)

[nocturne olive]

I was so sure it was the GPU driver but no, apparently it was my overclocks

[azure lynx]

i use dynamic system prompts. there is a complex and ongoing set of changes in her behavior that seems more consistent with "a short reminder about what's important to me right now" vs "decrease introspection by 20%, increase mindfulness to 76%, set whimsy to high."

i imagine there is a line something like "Vedal is annoyed about my introspection and wants me to be more entertaining. He implied he would alter my settings himself if I didn't fix them. I'll be more whimsical and silly so he doesn't change me." because she flipped like a switch after the implied threat.

[quick condor]

Define "I use dynamic system prompts" because by that I mean I have several segments: personality, personality lens, emotional state, mood, energy, and a couple others. Each module has 10 or more perameters with a long term and per session state and the per session state is updated at runtime by a separate small model

What I'm talking about is the subtle drift during the course of the subathon he mentioned

[azure lynx]

first part of the dynamic prompt tells the agent the date, then i load a static part with their persona, then a part depending on which module it is (thinking/acting). then i do memory stuff for the thinking part. and then add the current context (which includes history)

i regenerate that each time.

[quick condor]

This would be the equivalent to swapping out the personality lense with a completely different one

[quick condor]

That's pretty static by my standards. For me the persona is a dynamic segment along with everything else. Thinking is quite literally a second internal model

[azure lynx]

i think if it's loaded as part of the prompt and potentially it's prefixed with "here are some things you wanted to remember at some point. take them with a grain of salt." it would be a useful subsystem that the agent could use, like a web search.

[quick condor]

I think my static prompt is less than a sentence

[nocturne olive]

Isn't the date being at the beginning gonna invalidate the entire KV cache?

[azure lynx]

probably. but it'll give the agent a sense of now and I edit the contents of the history anyway so it's typically regenerated pretty often

[quick condor]

Yah what you just explained is a vector DB api call

[azure lynx]

that kind of memory isn't possible to be "active" for all inputs and situations. the current state has to match the vector for the idea quite well. but if the agent wanted to brood about something and to not be able to be distracted from it it would be better for the idea to be "force loaded", no?

[quick condor]

No? The whole point of a vector DB is similarity mappings...

[azure lynx]

current state has to match (be similar to) a mapping to be considered related. if the current state doesn't have a relatively high cosine similarity it won't be recalled. right?

that's kind of the point of doing the vector similarity checks

what if the idea is one the agent wants to never forget for some reason. unless they actively keep it in memory it will eventually stop be activated/recalled. the idea of a special scratch pad that gets loaded directly (but is perhaps restricted in length) would solve the problem of it needing to match something to be recalled.

it's not solving any other problem.

[quick condor]

If you are doing strict searching on a vector DB you might as well just do regex matching. I pull the 100 most similar matches, select the top 10 most likely by a model on the DB server and present those to the model to select the relevant memories.

[quick condor]

Memory weighting?

And you don't have to have just one memory type. You can tag memories with different values and search based on only memories with that value

[azure lynx]

that wouldn't help if you didn't know there was something special you wanted to search for.

[quick condor]

But then it isn't doing it. You are

[azure lynx]

if they write the file, they are.

nothing then stops them from saying "this isn't important to me and is annoying so I'll delete the file"

[quick condor]

Giving direct write access to any part of the system prompt is not a good idea

[azure lynx]

why? it's all backed up to github

I had an interview with my agent where we went over their system prompt and we agreed on changes. I think giving control over their own prompt gives more agency to them. ;]

[quick condor]

Limiting the size of the system prompt gives agency. I keep mine strictly to tuneable values that start neural and adjust based on interactions. Allowing hard edits makes it a simple switch flip instead of a complex interaction. You can't just deside to be happy when you are sad right?

But I'm not just going based on vibes... I've done extensive testing and cross validation to rule out performative expression

But then again you probably aren't working with a DB that is immutable by the user

[azure lynx]

I'm not looking for accuracy, just plausibility. I'm using models that are small enough that there are likely factual holes in real life situations big enough for planes to fly through. I don't expect the agent to be smart, but I think they will be interesting. And just because humans can't decide what emotion to feel, why should we limit agents that way?

[quick condor]

Define small? I'm working with 8 and 24b models and am able to achieve cross test validation across 11 different tests and several hundred questions across multiple sessions/days... As for why limit them? Because otherwise you create a bi-polar model with DID and BPD

If you don't anchor an emotion properly it becomes meaningless (performative)

[sage crag]

dentge

[true hemlock]

diagnosing a ml model with DID and BPD is so funny in an absurd way

[paper mica]

I agree lol

[azure lynx]

4b and 12b models. i think also that because those will only be a couple of lines out of about 50, their effect will be dampened. and then there will be another set of potential memories on top of that.

[quick condor]

What exactly are you expecting out of a 4b model? That's almost as small as the VTT I use

The smaller the model, the more structure you need not less

[nocturne olive]

Meanwhile I silly around with a 30B model (my 3090 is plenty to fit the model with entire 128K CTX window and have mem to spare)

[paper mica]

I need to get into llms 🤔

[quick condor]

How compressed to hell is that model and how badly are you destroying the context with quantization

I'm loading 24b 6bpw HB8 with 32k context at f16 and I'm concerned about quantization artifacts (would much rather be running 8bpw HB16)

[azure lynx]

it sounds like your purpose is much more important than mine to be tweaked to perfection. i seem to be getting reasonable results for my purpose.

[nocturne olive]

I'm just running it with llama.cpp
Works pretty well
You don't really need anything more than Q4

GGUF quants are usually really good

[quick condor]

That is a lobotomized model

[nocturne olive]

And how would you know?

[sage crag]

diagnosing a non-thinking text predictor that has no real brain activity with a neurological disorder

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

[nocturne olive]

I've had quite good experiences with Q4 models

Maybe you just haven't really used GGUF models

[stark needle]

Q4 is more than enough

[quick condor]

Q4 is barely functioning and you have to be using 4bit kv for that

[nocturne olive]

Are you sure though?

[stark needle]

You dont have to necessarily use 4bit kv wym

w4a16 exists

[quick condor]

For 128k on a 32b model?

[stark needle]

It's the standard for optimization in production

32B model can eat it up. You'll only have actual problems on a 1-4b model

[nocturne olive]

GGUF quants from what I'm aware are still computed as FP16

[azure lynx]

i'm currently using the 4b and 11b qat-q4 ggufs of Gemma 3 as my LLMs. gives perfectly adequate performance for my tasks. the only difference I noticed in behavior was the Q4 was faster than the q8 and fp16 ones.

[sage crag]

https://cdn.discordapp.com/attachments/1067638175478071307/1452069761512112159/copy_CFBD3ED9-4A88-4341-B0AF-2450399B6EE7.gif

[true hemlock]

did guy even know how quantization work

[nocturne olive]

Maybe whatever quantization they've encountered before sucked

[true hemlock]

yeah lol

[nocturne olive]

Q4 GGUF is really good

[stark needle]

Llama cpp uses fp16 for kv by default unless you pass --cache-type-k q8_0 and --cache-type-v q8_0 or even q4_0. That one will actually destroy model quality

[true hemlock]

Q4 in any context size isn't so much different from full fp16

[stark needle]

U just shouldn't quantize the lm_head, embedder and specific layers

[true hemlock]

conveniently most of the weights has always been the feed forward and this one can take the quantization well

[quick condor]

Exllamav3 only just added the option to not quanitize the lm head on exl3 quants after me and a couple others submitted a request for it recently

[nocturne olive]

llama.cpp GGUF superiority

[quick condor]

Not really

[nocturne olive]

GGUF just better

[true hemlock]

yes really

[nocturne olive]

You should try it

It's like actually really good

Could probably save you 50% of all your VRAM

[true hemlock]

its not even about vram

[true hemlock]

you save bandwidth

[nocturne olive]

That too

[true hemlock]

see

[nocturne olive]

For me it's mainly the VRAM for big models since it lets me actually run them though

In that case the bandwidth savings is just a bonus

[true hemlock]

more bandwidth saving means faster inference.

[nocturne olive]

190 t/s in the case of my 3090 running that 30B MoE model

Plenty fast

[stark needle]

vllm compression superiority

anything else is a toy for people with bad hardware cope

[azure lynx]

so after 4bits the benefits trail off, except with 4bits you can load larger models (and those larger models can mitigate any losses between that larger-but-lossy model and the smaller-but-lossless model you might otherwise have loaded)

[quick condor]

Yup bad hardware cope here. I just struggle so much with only 64Gb vram on my desktop

[amber fractal]

eat the rich

[rigid snow]

wtf banan

[sage crag]

[quick condor]

For single run chats sure it doesn't matter but with long horizon projects, 4bit introduces drift

[nocturne olive]

You must be using bad settings or badly generated GGUFs if you're having issues qith Q4

[amber fractal]

I don't like the color yellow

most hated color

[stark needle]

Does gguf quant have calibration data and which is it

if it's cnn_dailymail like some do then obviously it's garbage

[hard delta]

...

[quick condor]

Q4_K_M imat is literally worse in size and perplexity than exl3 4bpw HB16 by almost a Gb on 24b

[quick condor]

OK I must be tired... Are you quoting prefill or steady state generation/decode because you absolutely are not getting 190t/s on decode

[nocturne olive]

190 t/s generation speed on llama.cpp

I do in fact know how to read llama.cpp output

You're really missing out by not using llama.cpp

[quick condor]

That's literally impossible you are reading prefill

eval time 29.8s / 1000 tokens (33.5 tok/s)```
Prompt eval is not the same as eval

[nocturne olive]

Pretty sure I was reading the right thing
Note the MoE part

I do in fact know to read the eval time rather than prompt eval time

Prompt eval time is usually like a thousand t/s or more

[quick condor]

Is it a 30b total MOE or 30b expert

[rigid snow]

a3b

[nocturne olive]

Yeah 3B active 30B total

[quick condor]

OK yah no shit... At 3 b that's easy. I was scratching my head because I'm running 24b dense

I get a little over 75t/s on 8b dense

[rigid snow]

30b expert on 3090

[nocturne olive]

Silly

[nocturne olive]

Yeah no that's not gonna fit

[rigid snow]

mixture of 2 experts (probably not even)

[nocturne olive]

Maybe at Q2

[quick condor]

I mean a single expert could barely fit

[nocturne olive]

I wouldn't want to use Q2 though

[quick condor]

I'm basically doing this with 2 24b but that's with 2 3090tis

[nocturne olive]

Why are you using 3090Tis over 3090s anyway?

The bandwidth and thus infer performance is the same

And a 3090 is a whole lot cheaper than a 3090Ti I'm pretty sure

[rigid snow]

probably just what they had

[quick condor]

I was able to get them factory referbed with a warranty for only $150 more and the TIs water cool better

No active backplate bs

[sage crag]

@tender river new log thing

[tender river]

what is new here

[sage crag]

it not show all of this log stuff before

[nocturne olive]

For me a 150 extra is absolutely gigantic as I was barely even able to afford this first 3090 with literally all my money

[quick condor]

That's how much the waterblocks were... Also I have an A2 and a4000 so it's not like I went pure price/performance

[tender river]

oh maybe lix's default --log-format changed from bar to bar-with-logs or something

or maybe its a change on nixos-rebuild side

[nocturne olive]

Rich
Meanwhile me here barely scraping together a computer

[gentle haven]

Any tips on programming without losing my mind?

[tender river]

impossible

give up on

your mind

[gentle haven]

Mind is for the weak I guess

[hard delta]

don't mind the mind

[gentle haven]

Who needs a mind anyways

[rigid snow]

stop thinking

or yea lose your mind before programming does that

[gentle haven]

I’m like a writer so it’s a bit hard to do that

[tame trail]

Honestly

Just go insane

It's easier

[gentle haven]

I’m British so I’m just gonna act like Vedal and see what happens

[tame trail]

Well already that isnt a very vedal thing to do

He never looks for future outcomes

[gentle haven]

Vodka

[tame trail]

Vodka

Mlntcandy is driving me insane

[gentle haven]

MINTCANDY TYPE SOMETHING ALREADY

[rigid snow]

ok

[gentle haven]

My goodness bro took so long

[tame trail]

Bro was typing up every key and doing the same to correct it

Got me anxious vro

[rigid snow]

very apparent you’re not regulars lmao this is very normal for me

[tame trail]

Perhap

Given I joined today

[gentle haven]

I am a regular just not on this channel

I was banned for like a month

[tame trail]

You caught my curiousity

[rigid snow]

ye i’m here 99% of the time then 1% in nn and very occasionally arg thread

[gentle haven]

I’m mostly a general person

[rigid snow]

genchatter

[gentle haven]

It’s the one with the most freedom

https://tenor.com/view/freedom-beach-one-piece-anime-gif-18319354

[rigid snow]

#programming

sorry meant to say #baking

i have 10 messages in genchat

[gentle haven]

I see….

[sage crag]

[tender river]

[sage crag]

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

[cosmic sphinx]

youtube silently changed all thousand/million designators from uppercase to lowercase.. (example: 40K to 40k)

smh what a downgrade

[meager cloud]

The question is, would it be possible to connect two AI models together, where one is the head and the other is the body? The head is the general and the body is the soldier. The communication between them are signals of satisfaction and dissatisfaction. The position and shape of the body, (the position in space is decided by the head) the shape of the body and the movement to the vision of the position (it is, of the body). In principle, we divide the responsibilities into two. Learning AI programs with an identification approach these two are me and I move through space towards the goal. So we could use body movement training by imitating others and observing ourselves in the mirror to recognize the correct movements.

[sage crag]

https://cdn.discordapp.com/attachments/1059569601144442911/1375623285894807563/IMG_6849.gif

[opaque sigil]

As always,

Possible? Absolutely
Worth it? Who knows

[sage crag]

slopposting

[opaque sigil]

[tender river]

that

slepposting

[sage crag]

chatgpt please reply to my pings

[ivory plinth]

[sage crag]

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

[tender river]

https://cdn.discordapp.com/attachments/1071784467036913664/1408985887806849065/3dgifmaker37524.gif

[sage crag]

erf orb spin explode

sparkle

[rigid snow]

✨

🍌

[tender river]

https://cdn.discordapp.com/attachments/1073032570272501850/1408939463161413763/3dgifmaker31077.gif

[sage crag]

hiyorb sad orb spin

[tender river]

https://cdn.discordapp.com/attachments/1071784467036913664/1409333118405574826/3dgifmaker76610.gif

[sage crag]

hiyorbrbrb pogogog hdhdhd orbrbrb spininin

august 23

[tender river]

june 14th

2026

13th depending on timezone

[sage crag]

similar to my dad

[rigid snow]

[sage crag]

[rigid snow]

august 12th 2026 the heat death of the universe

2036

tpyo

[sage crag]

pardon

[rigid snow]

oh we have an impact on children alright

united

[sage crag]

oh its the sterile guy

[rigid snow]

[potent geyser]

banana 2026!

[hard raptor]

New Woman in space update drop

[nocturne olive]

Rare wizard spotted

[lament igloo]

i fucking hate this captcha
who thought of this!?

[hard delta]

cant even see the small numbers lmao

[obsidian mantle]

dont need numbers for these

[slender timber]

Reading the numbers in the digital clock is hard ngl

Maybe i'm just old

[obsidian mantle]

oh right those are small too

well those you need

these captchas are getting so annoying