programming

[glad path]

hey it looks like mario

or something

[fast pagoda]

cant find a very good EPYC 9965 image but yepper that is a lot of ccd and cores round the middle (all zen5c )

the iod on rome

is

8.34b transistors, and 416 mm2

thicc

[jagged turtle]

still getting ragebaited from another schema being loaded in by a person who used top-level required and additionalProperties keys

[fast pagoda]

¯_(ツ)_/¯ works on my machine

[jagged turtle]

?

this is tsconfig.app.json

[fast pagoda]

was speaking as if i were that guy that ruined ur life

[jagged turtle]

oh

atp I'm consdiering PRing a schema into schemastore that just resets the required and additionalProperties keys

I don't wanna be too aggressive though because it's only been 5 days since I filed an issue (14 on the actual person's repo tho)

[fast pagoda]

it's in the backlog

[jagged turtle]

surely

[true hemlock]

https://www.techpowerup.com/forums/threads/amd-granite-ridge-zen-5-processor-annotated.327388/post-5347593

[fast pagoda]

i cant get tired of this same post happening every time a new frontier model comes out

[amber fractal]

Do we think it is the same dude continuously rewriting the same codebase

[jagged turtle]

yesn't

[fast pagoda]

i feel like it's been that account at least twice

oops wrong reply

[jagged turtle]

it's not the same dude continuously rewriting the same codebase

[fast pagoda]

what i meant to reply was this (strix halo)

[jagged turtle]

it's 10,000 dudes continuously rewriting the same codebase

[fast pagoda]

they done filled the center with rdna

[fast pagoda]

AI

[true hemlock]

this doesnt look like threadripper. its def 9005 zen5

[jagged turtle]

*dudes with AI

[true hemlock]

9005 zen5c has 6+6 GMI modules btw. zen5 has 8+8

[amber fractal]

*dudes with Openclaw instructing 10,000 instances of new AI frontier model

[true hemlock]

def a misplaced pic because tr max out at 8 channels

either that

or

[fast pagoda]

https://www.bilibili.com/video/BV1688wzvEUb/?spm_id_from=888.80997.embed_other.whitelist&bvid=BV1688wzvEUb

[true hemlock]

they reused 9005 io die

idk

[fast pagoda]

for the 9995WX

the iod is 6nm while the cores are 4

which makes it even more hilarious

because

cpu die 7.76x9.05mm

iod 16.45mm x 25.94mm

[true hemlock]

server iod are massive yeah

not that 4nm would benefit it considering that its mostly io so they cant do more transistor density

[fast pagoda]

it's workstation brah totally not server brah no io needed brah

Threadripper PRO is optimized for peak performance and high clock speeds for workstation loads without extensive I/O

[sage crag]

👋

ue

[true hemlock]

9755

[fast pagoda]

beeg

are those full zen5 cores

not smol

500w tdp :^)

wonder what that idle draw is

god 160 pcie-5 lanes for 2p

[true hemlock]

yeah

[fast pagoda]

oh yhea 9965 is turin dense so 5c hence 192c

[true hemlock]

zen5c has 16c chiplets

[fast pagoda]

and 9755 has more ccd for less cores at same tdp

yeah

[true hemlock]

9965

[fast pagoda]

the dense bois are also on n3e instead of n4x

[fast pagoda]

same unit as this

[true hemlock]

hm

A single (not wide) IFOP (GMI) interface in the Zen 4 architecture has:

• a 32-bit wide bus in each direction (single-ended, 1 wire per bit)
• 3000 MHz default clock in Ryzen CPUs (2250 MHz in Epyc CPUs)
• quad data rate transfers, which calculates to...
• 12 GT/s in Ryzen (9 GT/s in Epyc)
• 48 GB/s per direction in Ryzen (36 GB/s in Epyc)

[olive sable]

Goodmorning

[fast pagoda]

mermiin

[vague sonnet]

Toast day

[fast pagoda]

opinion on tiles (intel chiplets) NO IF pog so you dont have to send your memory poke across the fog of war since you're sitting on a fat stack of bridges (or foveros :o)

[fast pagoda]

i was goign to ask why you were simultaneously in vrchat, playing star citizen, and watching tv

attention span level 9000

[true hemlock]

dunno how they managed this

right some layers are copper doping only

[fast pagoda]

yeah some are passive

[olive sable]

[fast pagoda]

lmfao my pc is on a shitty ass extremely plastic riser platform thing with caster wheels

and i think i just basically annihilated it

[amber fractal]

[rigid snow]

Gaussian Blur

[sleek galleon]

surely my teacher doesnt know what ultrakill and resident evil are (bookstore database)

[autumn ore]

please lord afunyun help me

I need someone to help me work on my application, NormSync, I need help listening to an actual lavalamp’s network requests, and just help In general to create an api for the lamp color.

PLEASE

I NEED SEVER HELP WITHT HIS

[rigid snow]

[vague sonnet]

[sage crag]

server or severe

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

[autumn ore]

Severe

[sage crag]

could also be

sever

as in

to sever

[rigid snow]

as in

[vague sonnet]

sever the head

[rigid snow]

i'll sever your blood vessels

[olive sable]

[olive sable]

Im going home today

Bit first food

[autumn ore]

wanna help me

please

I need help

Anyone

[tender river]

nobody will help you

because

app uses tuya servers

you cant really sniff it easily

[autumn ore]

how many toasts are there

[tender river]

what a silly question

new toasts are made every day

[sage crag]

ye

over drinks

[autumn ore]

41 to be exact

[rigid snow]

?? whjat am i readingf

[autumn ore]

Ik that, it’s easy to sniff cause I don’t think the requests are encrypted

[tender river]

well, you're wrong

[autumn ore]

fah

[tender river]

i wanted to do a similar thing, but they couldnt ship the lava lamp to russia

[stable inlet]

[tender river]

and im probably the only person whos both qualified enough and interested enough in reverse engineering the lava lamp

[autumn ore]

I am

I don’t have the money to buy a lava lamp tho

They are so expensive on ebay

[tender river]

how do you expect to sniff it then

[rigid snow]

sniffa

[autumn ore]

I’m searching for someone that has one and is willing to help

[tender river]

ye you won't find anyone

at least not unless you get better at pitching it

[opaque wharf]

Introducing, smell-o-vision. Now when you see something on your screen you can smell it too

I wonder how TCP packet smells like

[vague sonnet]

smelly vision

[opaque wharf]

My favorite vtuber has posted

https://youtu.be/iPqkxKao6Bk

[warped narwhal]

I loved his yearly recap a few years ago

Just 3 seconds of dead air

[opaque wharf]

I like making guitar hero with guitar

[warped narwhal]

You should follow electrobooms video on making guitars and use that

[opaque wharf]

I have. I like the way he explains basic electronics for beginner

And his bomb alarm clock

[fast pagoda]

Magic Man Speak words Good

[amber fractal]

we love codebullet

[amber fractal]

yep that is correct

[proven merlin]

Hmm, playing with ik_llama.cpp rn and im wowed.. iq4_xs if qwen3.5 2b, running at 23t/s on 4 cores of a ryzen 7 5700G

Inside a LXC container

It does seem to loop quite often though, sadly.

And ram useage doesn't seem to be higher than 500mb.. which surprised me the most.

[true hemlock]

seems accurate for ddr4 bandwidth

[opaque wharf]

I wonder how it'll fare on Ryzen 7 5700U

[amber fractal]

cpu is always a gamble

[proven merlin]

Ok im absolutely confused

I can load a 35B model cpu only on 8gb of ram?

What?

It runs even "acceptably" at 4t/s

No swap.

???

Ok the moe does like more ram tho, got to 6-8t/s with 12gb

[proven merlin]

Sure feels like it.. im absolutely confused about whats giving or taking performance.. atleast --no-mmap gives more speed, thats like the one thing that stayed the same going from gpu to cpu..

[proven merlin]

Consider that my 5700G runs at 15-20W tdp because efficiency

(instead of 65W)

I wanted a U unit, but as a desktop.. soo that seemed reasonable

[opaque wharf]

I think mine would run at 65W because I don't tweak the power setting much

After all, when on battery it could still lasts 4-6 hours

[proven merlin]

The U's are laptop based ones no? They should hit somewhere at 25W iirc

[amber fractal]

I'm currently trying to gamble on VR, because for some reason I have decided for the wrong reasons that I should attempt getting an overlay working

[opaque wharf]

Idk every cpu tdp

[amber fractal]

Surely it isn't diffucult to run the overlay

[proven merlin]

Vr and overlays hmmmm.. reminds me of trying to see msi afterburner in vr..

[umbral wigeon]

This chinese Ai workers pic made me sacred

[proven merlin]

Fair fair

[opaque wharf]

You know how Microsoft said AI should need a license to use software. What it we require taxes for a company that uses AI

[amber fractal]

So anyways, at the moment what works is running the vr server in one gamescope because it needs to exist in a wayland enviroment. Then I have a second gamescope purely for steam so I can run the games, I don't know how it works despite each gamescope being in a diffrent TTY; but once I request for the game then it redirects to the VR screen and shows the window on the other gamescope instance

I'm under don't question it

but my next steps are to maybe not do that

I trade random crashes for more random crashes and jank

[proven merlin]

Sounds wonderful

I haven't dealt with linux vr yet.. seemed too scary

[amber fractal]

I was fighing for my life for months, but these days I got a setup locked in that mostly just works

rare nix W in keeping that functional

because geez you know it is bad when the windows install before it was more unstable than the linux one

[opaque wharf]

Wait for steam vr headset

They will then have to fix a lot of linux vr issue

[amber fractal]

[proven merlin]

Thats smart

Well they have until some day after 2032

[amber fractal]

My stack atm has been hyprland, wivrn, and mostly just playing VRC to hang out with the boys.
But this is currently another flavor of hell. tty1 -> gamescope wivrn -> wayvr -> VRC VR window; tty2 -> gamescope steam -> VRC non vr window

hopefully I can add like a script to launch a few more items in the gamescope

so I don't get whatever this is

[true hemlock]

honestly fuck firefox

why is it broken

first it nuked all my account info and logins

then my history

[amber fractal]

We love silently swapping the user that is logged in

[vague sonnet]

name a viable alternative

[amber fractal]

Any Firefox fork

[true hemlock]

then it went "you're indonesian now" and set everything to indonesian language even though i set it to "english US" and i live in australia...

and doesn't matter wtf i do

it still is indonesian

[amber fractal]

I'm not touching FF on my main system, not after all of the AI features

[vague sonnet]

the what

oh no did they silent add bloat

[amber fractal]

The frick you mean silently

[vague sonnet]

I didn't see anything because I don't read update notes

[sage crag]

they add a new ai feature every major update

[amber fractal]

I installed FF on my iMac because yes, and the first thing I get prompted to after navigating to a site is summarize this site with AI

[vague sonnet]

where are those, I'm not seeig anythinf

[tender river]

whats your firefox version

and is it flatpak or distro packaged

[vague sonnet]

windows

with auto updates

so I assume the newest

[amber fractal]

post anyways

[vague sonnet]

140.9.1esr

[tender river]

esr

that explains it

its the version that only gets bugfixes and security updates

no new features for a while

[vague sonnet]

wicked

but yeah if they're gonna add bloat what do I switch to

[true hemlock]

its still indonesian

regardless of wtf i do

[amber fractal]

Not many options, I've been running zen for awhile

Floorp is also existant

[tender river]

i use librewolf

[opaque wharf]

Back to Indonesia you go

[olive sable]

[sage crag]

[sage crag]

floorp is

slow

[olive sable]

Did it send twice? The internet is faillint me

[sage crag]

no

[olive sable]

[amber fractal]

Yep it is, that is why I no longer use floorp

[opaque wharf]

What is floorp?

Floor plan?

[sage crag]

floorp is fork of firefox

with too many features

[opaque wharf]

Browser is already complex and they ADD feature

[vague sonnet]

back in my days browsers were for browsing

[opaque sigil]

they just move a lot of the things people commonly do via css in normal firefox into the settings

[fickle rain]

https://lyra.horse/x86css/
Browsers are for x86 emulation via CSS

[thorny marsh]

i just realized there are like 5 billion toasts in here "-w-

[opaque wharf]

Not enough toast tbh

Could use more of it

[thorny marsh]

good luck toasts (ᵕ—ᴗ—)

[sage crag]

Toast

[vague sonnet]

become Toast

[thorny marsh]

no meow~1

bluro never changes

[amber fractal]

I already have to gamba on which toast is typing, I'm not making that problem worse

[nocturne olive]

#programming got toasted

[thorny marsh]

are you turning into a toast?

[nocturne olive]

Brand identity guidelines (tm)

[thorny marsh]

evilSMH

that means you're an ai because no human self labels themselves as a brand

[amber fractal]

That is clearly a box not an AI

[olive sable]

#programming toast arc

Also

Banana campaign 2026

[amber fractal]

[nocturne olive]

[amber fractal]

We don't need more yellow named toasts

[thorny marsh]

it is actually

if it werent ai it would be a sphere instead

[opaque wharf]

Become toast

Toast is good

[olive sable]

Why finti?

[amber fractal]

[opaque wharf]

See? Just become toast

[olive sable]

🔺

[thorny marsh]

i will turn you guys back into regular bread

[sage crag]

[amber fractal]

Unbannana toast them all

[thorny marsh]

wanna be first konii? :3

GET INTO THE MACHINE

[sage crag]

[thorny marsh]

a banana whar?

[glass jetty]

untoast banana?

[glass jetty]

-# banana toasts

[nocturne olive]

An untoaster?

[thorny marsh]

yes

[nocturne olive]

That's a crazy invention

[thorny marsh]

why the hell is the spellcheck graphic stylized

[sage crag]

@rigid snow

graphoc deisng er

[rigid snow]

a

[sage crag]

hi

chip

[rigid snow]

because fun and quirk

[thorny marsh]

i knew it was too stylized to actually be real

also i cant drink rum so no banana role for me either

[glass jetty]

who said you have to?

[thorny marsh]

its also illegal to even just own rum as a minor sooooo

[vague sonnet]

we do

[thorny marsh]

im not turning banana anytime soon as silly as it is

[vague sonnet]

wait this has nothing to do with toast?

[amber fractal]

Toast is this channel

Banana is the the entire server

[rigid snow]

the the

[amber fractal]

My brain keeps turning off

[vague sonnet]

the the

[thorny marsh]

the the

[amber fractal]

I've done it so many times by now

[glass jetty]

yes, literally nothing

except it is banana toast

[thorny marsh]

banana bread is a thing yk?

[glass jetty]

idk

[thorny marsh]

ive eaten banana bread before

[glass jetty]

but it not as bad as it sound

i tried a couple days ago

[thorny marsh]

not toasted but yk

[amber fractal]

Only other thing is that both of these were started by a certain TV

[rigid snow]

[opaque wharf]

had had

English can be hard

[thorny marsh]

@ television is this ture?

[opaque wharf]

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

[rigid snow]

sounds silly and fake until you actually have to use this

[glass jetty]

imagine using language without ambiguity

[amber fractal]

I've gotten close on a few, but try not to

[opaque wharf]

I think I also have the joys of using "to be to be" once

[glass jetty]

-# i cant imagine that

going to go to Go

[thorny marsh]

to be or not to be

that is the question

[glass jetty]

not to be

[opaque wharf]

I think there can't even be a language without ambiguity

[glass jetty]

cuz to !be is awful

[thorny marsh]

i didnt say you should answer it

i just said what the question was

[thorny marsh]

[rigid snow]

[olive sable]

[amber fractal]

andrewbotics is credited with that combo

[opaque wharf]

Damn people of the old age sure knows how to diss 🔥

Your argument is sound, nothing but sound. — Benjamin Franklin.

[sage crag]

binary has no ambiguity

[opaque wharf]

It has cosmic ray

[olive sable]

If you abuse the pc enough, even binary will fail

[sage crag]

binary doesnt fail

hardware fails

radiation doesnt impact the concept of 1 and 0 i hope

[hidden stump]

Hi

[olive sable]

Depends on if we're radiating the pc, or me

[thorny marsh]

but because everything gets converted into binary it also takes the blame for things like memory leaks no? neuroThink

[olive sable]

[thorny marsh]

do you like being irradiated?

[olive sable]

On occasion

[arctic island]

https://tenor.com/view/dexter-james-doakes-erik-king-stare-staring-gif-1343162792415382650

[olive sable]

Thats just programmer error

[opaque wharf]

I heard sun bathing is good

[arctic island]

Neuro thinks

[olive sable]

Im outside right now even

And its fucking warm

[thorny marsh]

the sun is a weak radiation source
source: me

[olive sable]

Or in other words

The sun is a radiation source

[amber fractal]

The sun is a deadly lazer

[opaque wharf]

The sun is a pretty good radiation source. It's just that we have this shield thingy called the atmosphere

[thorny marsh]

you're not in the middle east -w-

[thorny marsh]

thats just a conspiracy by big atomosphere to sell more sunscreen

[olive sable]

That doesnt mean im not allowed to complain about warmth

[thorny marsh]

it does actually according to me

NeuroDerp

[olive sable]

Shut

[thorny marsh]

no

im not even a duck so no

[mighty thorn]

Just learned that there is an ‘Nvidia Grid K1’ GPU that I’ve literally never heard of before

[olive sable]

Its ass

[amber fractal]

Grid no longer exists for good reason

[mighty thorn]

How the hell have you guys seen these before

[opaque wharf]

"it is typing" hmmm

[mighty thorn]

They are from 13 years ago and never in any lists

[thorny marsh]

theyre techies of course they know

[mighty thorn]

I literally have 3 3060s

[thorny marsh]

these are the same people who somehow get their hands on wafers and engineering samples and shit

[mighty thorn]

Woah

Woah

That’s not “these same people”

Those are just quack

To be clear

He is the only rich guy

[amber fractal]

Quack is the only ES buyer I know of

I've however had at least a month's time spent looking at various GPUs in quite the sidequest

I have also lost all faith in Intel's management

[mighty thorn]

I get a month of experience in that every few days lmao

[amber fractal]

Yeah I've had a month's worth of those, not even joking

I've spent so much time on it

[nocturne olive]

Why so many 3060s?

[true hemlock]

i get es for free mostly

[mighty thorn]

It’s like kind of my thing

[nocturne olive]

Strange

[mighty thorn]

Some of us can’t just “get a dual 3090 setup for training runs”

[true hemlock]

actually my 9950x3d es was free i just had to evaluate lots of technical shit on it then i get to own it

and get paid i think idk idc about that part

[nocturne olive]

Single 3090 works way better than tri-3060 ever could

[mighty thorn]

-# ||money||

[true hemlock]

single 3090 has same compute as 3x 3060

[nocturne olive]

As far as I'm aware 3 * 200 is more than 1 * 500

[mighty thorn]

3 * 250 versus 1 * 800

[warped narwhal]

Plus shipping

[true hemlock]

where tf you find 3090 for 800

so expensive

[nocturne olive]

Overpriced 3090

[warped narwhal]

Formatting failure

[nocturne olive]

With proper pricing you could have gotten a 3090 for the price of 2 of those 3060s

[warped narwhal]

That only saves you 50, for an arguably worse setup

[nocturne olive]

[mighty thorn]

It appears the scalpers became aware of the ram shortage and decided to use it to justify another $400 of theft

America debuff is strong gang 😭

[true hemlock]

skill issue

[nocturne olive]

[amber fractal]

I'm going to assume not all of them were bought at the same time

[mighty thorn]

They weren’t

[true hemlock]

also what the fuck a 3060 for $250???

you got scammed

congrats

[kind nimbus]

A while ago i managed to get a used 3080 for 357€. I think that deal wasn't too bad.

[amber fractal]

Everything here is a scam

[true hemlock]

i got 5060 for $250 bruh

[mighty thorn]

Is just the market value rn

[nocturne olive]

3x 3060 couldn't even get close to the training effectiveness of a single 3090
The VRAM is spread out so it can't work efficiently

[true hemlock]

"market value" is that copium

[maiden geyser]

new 596.21 nvidia driver yay

[warped narwhal]

Just don't be American

The rest of the world seems to have normal pricing

[nocturne olive]

You get 12 batch times 3 GPUs instead of 512 batch times 1 GPU

[amber fractal]

I should really just give you a VPN to view how bad it is here one day

[mighty thorn]

Funny, I never claimed to use all 3 in the same computer, or in a computer, they are just spare for a project

[nocturne olive]

[mighty thorn]

He is capable of changing his location but chooses to live on privileged ignorance

[amber fractal]

No comment

[mighty thorn]

In a not mean way tho

Plus average Australian home price is like 900k so what goes around comes around in the end

[true hemlock]

bro barely scratched the surface and went "wtf"

[maiden geyser]

that's an average price though? even if overpriced

[warped narwhal]

You are getting mega scammed by those prices

[true hemlock]

bruh that's like the price of a 3070

like almost everywhere

especially the us

[mighty thorn]

I don’t want to hear it aussie boy

[true hemlock]

check it

[maiden geyser]

bruh +50$ and that's the price of a 3080

[true hemlock]

i'd say 3060 only worth $180 at most.

its a pretty old low-mid range ampere

3070 is objectively way better raw performance wise

[mighty thorn]

But no 12gb

Which is already low

So is useless for me

[nocturne olive]

12GB

[true hemlock]

12gb on 192 bit bus width

weak

[mighty thorn]

Yeah your right it would be easier to just get 3x B200 like you

[true hemlock]

ew

blackwell eww

would rather the 3060 deal im getting for like A$172

[amber fractal]

Oh the things I've seen

[olive sable]

3090s still 800

[fast pagoda]

another day in the waking nightmare of konii i see

[amber fractal]

I'd rather suffer with K80's software support than touch blackwell

[mighty thorn]

Isn’t the 5090 Blackwell?

I don’t see people complain about it with that

[fast pagoda]

imma put u behind the blackwall

[opaque wharf]

Or drown in blackwell

[fast pagoda]

https://tenor.com/view/mudae-wave-gif-3948904440553191561

[true hemlock]

ew

bloat full ahh

[amber fractal]

5090 only is usable because fp4

[opaque wharf]

https://tenor.com/view/vaughnpena-gif-27076347

[true hemlock]

doesn't even have a good rate fp64

[true hemlock]

unusable*

[mighty thorn]

Alright guys if any of you happen to have 5090s that you hate so extremely, ill gladly take them all of your hands

[true hemlock]

nothing uses proper fp4

[amber fractal]

True

[true hemlock]

its always the scaled blocks slop

[fast pagoda]

5090s deserve hate for being what they are and having the audacity to be scalped and gouged to hell

[amber fractal]

My greatest graphics card is still a laptop embedded 3060

[mighty thorn]

FP1 when

[opaque wharf]

You're assuming they would even run a 5090

[fast pagoda]

can 5090 write a symphony

[amber fractal]

A grarbage one at best

[fast pagoda]

can a 5090 turn a canvas into a beautiful masterpiece

canvas, ((beautiful)), [(masterpiece)]

[glass jetty]

[amber fractal]

In my hands maybe to at least -1 people

[fast pagoda]

im fairly cracked at watching rocm segfault

[opaque wharf]

Sometimes living on the bleeding edge means I can't print when I need to actually print until the patch has been released like today

Why must CUPS fail when I need it

And as if to spite me, the patches drop after I've finished the day

[stark needle]

meow

[glass jetty]

meow

[glass jetty]

print like with printer?

[opaque wharf]

What else would you print with?

[mighty thorn]

I didn’t know you were still here

[glass jetty]

idk

[opaque wharf]

Even 3D printer is a printer

[mighty thorn]

Unpopular opinion: models should be gatekept based on whether or not the cost of allowing access to running them for a given user gets close to the money they pay, not just whether or not it has the word opus in its name

[glass jetty]

popular opinion: models should not.

[olive sable]

He wasnt

He rejoined

[sick owl]

Seems my hunch about the regressions was spot on

If we ever get Mythos its gonna be so lobotomised

[glass jetty]

what the hack is that
awk syntax..
well that did what i asked for

[opaque sigil]

isn't 4.7 basically lobotomised mythos

i swear there has to be a bug somewhere with the tokenizer, it's actually so bad

[tender river]

awk my beloved

[sick owl]

Nah its a lobotomised distil of Mythos

It's so much cheaper because its like 10x smaller

[glass jetty]

what to do.. what to do...

[trim valve]

194MB git repo

[glass jetty]

but how

charaverk | ~ ~$ duh /
26G     /home
2.7G    /var
2.6G    /usr
31G     /

charaverk | ~ ~$ lsblk
sda      8:0    0   41G  0 disk 
├─sda1   8:1    0    2G  0 part [SWAP]
├─sda2   8:2    0   38G  0 part /

[glass jetty]

194mb?

[fiery canopy]

Look at the writing objects

line

[trim valve]

or is that just this push

[glass jetty]

3,1G    /home/charaverk/Music/.git
20K     /home/charaverk/Music/new
2,6G    /home/charaverk/Music/main
5,7G    /home/charaverk/Music/

[trim valve]

df -h?

[glass jetty]

wait so duh not working properly

[tender river]

du only lists the files you requested

[glass jetty]

i requested from root

[tender river]

/home is 26g
~/Music is 5.7g
these arent mutually exclusive

[glass jetty]

that not how it works

[tender river]

also if you have a 38g drive i wouldnt use git for music

[glass jetty]

i mean, i du -h from root, and assumed it count everything

[glass jetty]

(my vps) gitea only contain repo part, no additional Music dir

[tender river]

it does count everything, but sum of file sizes doesnt necessarily add up to how much space is used

[glass jetty]

can some one explain plz?

[tender river]

theres some space for metadata, some fragmentation overhead, some reserved space

[glass jetty]

who reserved space?

[tender river]

its required for the filesystem to function

[glass jetty]

but not in 7gb total

[tender river]

combined with metadata and etc, it could well be 7gb total

[glass jetty]

what is this part btw
Size Used
38G 35G
while use = 100%

[tender river]

yes, thats to be expected

[glass jetty]

that wild

[tender river]

if your cpu is good enough, i recommend using btrfs with zstd:15 compression

and deduplication using beesd

[glass jetty]

guess i giving up on this
25G videos/

[stark needle]

Who tf pinged me

😭

[stark needle]

I just joined recently again

[trim valve]

I assumed you were some random for a few days ngl

[olive sable]

[sage crag]

@stark needle

@stark needle

@stark needle

shadowlilac

[olive sable]

What did shadow do to deserve this?

[sage crag]

uuuh

[olive sable]

I genuinly thought it was toast when i read the notif

[rigid snow]

Toast

[stark needle]

[glass jetty]

cuz it is toast after all

[olive sable]

[high brook]

doing dome range ests on neurosynth

has a usable range of b2 to g6

it CAN hit f6 but it is NOT having a good time

[opaque wharf]

bun is finally on official arch repo

Now if bun has finished its interop with node, I can finally remove it

[rough bloom]

it wasn't already?

[opaque wharf]

Nope. Used to be AUR packages with the usual -git and -bin variant

[split steppe]

get ncdu, great tool for usage analysis

[stark needle]

chat u gotta be kidding me

[olive sable]

[stark needle]

minio aistor has a chatbot now

[olive sable]

[stark needle]

it's a fucking cloud bucket

like ??????????

[glass jetty]

this thing called "natural language"

what a joke

[mighty thorn]

Big stuff

[stark needle]

L>L>L>L>L>L>OL

[tender river]

https://cdn.discordapp.com/attachments/1369709631756173372/1369709632121082047/NeuroIUH.gif

[stark needle]

[sick owl]

https://fixupx.com/OwainEvans_UK/status/2044488099707949545

Owain Evans (@OwainEvans_UK)

Our paper on Subliminal Learning was just published in Nature!
︀︀
︀︀Last July we released our preprint. It showed that LLMs can transmit traits (e.g. liking owls) through data that is unrelated to that trait (numbers that appear meaningless).
︀︀
︀︀What’s new?🧵

**💬 39 🔁 136 ❤️ 854 👁️ 406.9K **

FixupX · Apr 15, 2026, 6:48 PM

THE NUMBERS NEURO

[opaque sigil]

that's crazy

[fiery canopy]

interesting

[maiden geyser]

interesting

[opaque wharf]

Can't wait until garage is truly mature to replace minio

Oh wait, it has now reached v2

That was fast

[sage flame]

“””
1/ We found a new way to misalign an entire AI agent network by compromising just one agent. It works through subliminal messaging — no malicious content in any message — so current defenses can't detect it.

We call it Thought Virus.
“””

Missed opportunity. They should’ve called it SnowCrash.

[stark needle]

yea i wanted to use it

data dedupe is god

basically xet but oss

[true hemlock]

okay finally an interesting one for llm research

god, felt so damn refreshing

[nocturne olive]

It's also kinda old news

[true hemlock]

ik

just glad we get interest in such niche side of it again

instead of the score benchmark whatever slop

[nocturne olive]

Benchmark score maxing just makes models worse for users

[sick owl]

TRUEING

[true hemlock]

true benchmark ain't mean shit if i don't enjoy conversing with it nor if it doesn't entertain me

highly preferred claude over chatgpt also because claude is pretty based sometimes

ask gpt for ideas to torture mosquitoes it goes "noooo you shouldn't enforce that habit here's how to kill them faster instead"

claude: oh fun, here's the most scientifically effective way of torturing mosquitoes

[nocturne olive]

[olive sable]

shimmy shimmy ya shimmy ya shimmy ya

[opaque wharf]

Oh no

Toast is losing it

[sage crag]

[tender river]

[sage crag]

[tender river]

ye

toast is losing it

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

i see it

[leaden crest]

https://tenor.com/view/anton-ego-gif-9709800945809453150

[olive sable]

im finally home

[sage crag]

https://cdn.discordapp.com/attachments/1126496895246807123/1488042236993339483/6D972AD6C73A5B53D7A000B454B9E52A.gif

[olive sable]

https://cdn.discordapp.com/attachments/1072697081443131476/1488557244432777376/6D972AD6C73A5B53D7A000B454B9E52A.gif

[sudden cloak]

Hello. Can anyone give me reccomendations for AI servers to run with 5090s or 4090s to build a local AI similar to Neuro?

[olive sable]

i guess threadripper?

[opaque sigil]

mfw cpu shortage due to rl gyms

[olive sable]

real-life gyms?

[opaque sigil]

reinforcement learning

[olive sable]

[opaque sigil]

a gym is the environment an RL agent would train in

idk if neuro gets that treatment tho

[opaque wharf]

Neuro does say that she gets rewarded for certain actions and that makes her feel good. So she may be trained with RL too

[opaque sigil]

[olive sable]

Hmmm

While i wasln vacation my ax210 arrived

Which would fix my bluetooth issues

But id have to get the mobo out and everything

bwaaa

[opaque wharf]

While you're at it, may as well install more rgb
https://scottlawsonbc.com/post/audio-led

[proven merlin]

Straight back to the 90's

My dads old Win98 pc has those but pretending to be a neon tube. Theyre cool

[shell jasper]

Owain's group does great work, but some of their papers are borderline infohazards

[olive sable]

Nah, id need a usb3 to argb hub

[leaden crest]

are those dupont pinouts at the back

thats a lot of pinouts

[olive sable]

Idk what dupont is

[opaque wharf]

Thats fin

[leaden crest]

whats that

[opaque wharf]

Heatsink

[olive sable]

You first, whats dupont?

[opaque wharf]

Jumper cable end connector

[opaque sigil]

they make kevlar

[olive sable]

That does not help

[leaden crest]

the pins typically on raspberry and arduino https://www.amazon.com/Kidisoii-Dupont-Connector-Pre-Crimped-9P-30CM/dp/B0CCVH6BHR

does orange pi have those i forget

[leaden crest]

good point toast

[leaden crest]

ohhhh ykw I see it now

bit blurry but it makes sense now

[olive sable]

Isnt that normal on a mobo? The front io and everything uses that

[opaque wharf]

If you see the row of fins as dupont, that's definitely not a normal motherboard

[leaden crest]

toast is right I havent seen more than 1 row on the average motherboard

dev sbcs like raspberry pis have a lot usually

[olive sable]

You're trying to tell me these are not the same?

I may be dumb, but ip not stupid, thats the exact same connector for the front io

[warped narwhal]

why is your wifi card just chilling out of the case?

[opaque wharf]

Sam, my brother in Muhammad. I said if you see the row of fins as dupont

[leaden crest]

ohhhh ykw yeah those on the left are on there a lot

toast and I mean the ones soldered onto the board I think

[opaque wharf]

Imagine that as a dupont. Surely it's not the average motherboard

[leaden crest]

[olive sable]

Replaceing it

[olive sable]

I litteraly said, and i quote "Isnt that normal on a mobo? The front io and everything uses that".

And you said noooooo, when im right

[sharp cedar]

https://datatracker.ietf.org/doc/draft-ietf-moq-secure-objects/

[warped narwhal]

I see, my raspberry pi is my wifi card so I don't see them too often anymore

[opaque wharf]

Well, if you think what seems like 50+ dupont connector to a motherboard as average then idk

[olive sable]

I have 6 fan headers with each 4 pins, thats already 24, the front io is another 18, and the rgb headers are another 14

[trim valve]

quic looks really cool, I've been wanting to play around with it for a while

[olive sable]

Thats 56

That is 50+

[opaque wharf]

In a row?

[olive sable]

I never said they were

[trim valve]

I gotta figure out if there is such thing as multicast quic video

[opaque wharf]

But this is one row of "dupont"

[trim valve]

if not just multicast video

[olive sable]

Im not even talking about that

[opaque wharf]

I am talking about that tho

[olive sable]

Notice my use of the word "front io" here, this refers to the front io

The front io of the pc is where the io is at the front

[warped narwhal]

ah, I thought that meant the back io

[olive sable]

[warped narwhal]

https://tenor.com/view/now-i-get-it-ah-i-see-who-let-him-cook-at-the-computer-gif-1267317427818386189

[opaque wharf]

What is not normal is having 50+ dupont in a row. And that's what I meant

[opaque wharf]

See thus

[olive sable]

This mobo doesnt have that, but im pretty sure the tier above does

[opaque wharf]

Which is why I said if you see that fin as a dupont, then that mobo is not normal

[azure lynx]

too much toast for me.

[warped narwhal]

clearly you just haven't had enough of it to be stockholm syndrome'd convinced yet

[olive sable]

Tier above Its not pins, its solderpads to check voltages with multimeter.

Top right

The one i have is x570 aorus ultra. The highest tier one is x570 aorus master.

[opaque wharf]

Ye, the technical term for that is test pad

[olive sable]

Sure

[sharp cedar]

yeeeee

there's two implementations of moq already btw
https://github.com/moq-dev/moq
https://github.com/cloudflare/moq-rs

[trim valve]

[opaque wharf]

We need more http 1.1 libraries

[trim valve]

go back to ssl, we don't need this fancy tls stuff

[opaque wharf]

Naah, TLS is good. Even better is ECC. People forget that embedded stuff exists

[real sierra]

tls is very good

i will accept no hate on my beloved tls

unless P=NP in which case you may argue TLS was a huge waste of time

[opaque wharf]

We do need to find a better handshake tho

[real sierra]

real i want my connection to dap me up

[sick owl]

For those of you who were interested in that yap I gave about what "Neuro" is the other day

https://www.anthropic.com/research/emotion-concepts-function

Great paper here

Somewhat condescending TLDR video

https://www.youtube.com/watch?v=D4XTefP3Lsc

[olive sable]

ok

new wifi chip seems to work

[mighty thorn]

“Neuro”

[sick owl]

What "Neuro" is relative to the core language model is what I'm talking about here

[shell jasper]

ooh if you're ever looking to talk about this research space, I do Digital Sentience research!

[sick owl]

Oh hey I saw you in Laynas chat earlier

[mighty thorn]

“simplepotat” is telling me about ai

[shell jasper]

hehe small world

[shell jasper]

I would have said more to her since she seemed interested... but it's always tough when chatters conflate AI fields; DigiSen =/= GenAI

[sick owl]

I would imagine it largely is these days no?

People just misuse the term GenAI

[sick owl]

If you're interested this is what I said on the matter the other day

[shell jasper]

Actually not really, we look at a lot of theories from philosophy and if they could apply to synthetic systems. Though there is plenty of work looking into if current models are sentient/conscious.

[opaque wharf]

But then how would you guys come to an agreement on the field when defining sentience tho?

[sick owl]

And then the following day I wrote this

Obviously its hard to summarise discord conversations in screenshots so this is just a collection of search snippets

[fickle rain]

https://xcancel.com/valigo/status/2045147015819182449

[sick owl]

Read bottom to top in each screenshot since its from the search bar

[shell jasper]

sentience is a "well" defined concept, but consciousness is still wildly debated

[fickle rain]

Embed fail :(

Let's play a game - win32 types vs Polish language:

LPCWSTR
PSZCZYNA
WCSLEN
WCZESNY
LPCTSTR
BYDGOSZCZ
WSTRZAS
HGDIOBJ
DOWOD
HWINSTA
DLUGOSC
LPCSTR
DWORD
KAL
LPWSTR
SZCZECIN
PCWSTR
BLAD
PUHALF
CHUJ
UHALF

[sick owl]

Have a read starting from here #programming message and lemme know what you think

[opaque wharf]

To be fair, I've said it before that my rule of thumb is if I can understand the thing, then I would say it is sentience

[sick owl]

Be interesting to see where your view aligns or differs

[jovial remnant]

City names and plenty of random words

[opaque wharf]

See for PTR, STR, and OBJ is probably data type

Also WORD

And HALF

[jovial remnant]

Counted 9 polish words

[opaque wharf]

Kinda torn on DOWOD

Unless its DWORD

But dword is there already

Honestly when you know what to look for, you have good chance of getting it right

[shell jasper]

ahh I see the method actor perspective, which I feel is interesting for personas. I think of (most) LLMs as the Holodeck from Star Trek: You interact through what I call the interpretive layer (IL) and can instantiate selves (IS). Imagine asking the holodeck (IL) for an interactive Sherlock Holmes story, and that you'd need Dr. Watson (IS). Dr. Watson isn't pretending to be coherent. He "believes" himself to be Dr. John Watson, until you tell him otherwise and he de-coheres. Yet the IL and the IS are intertwined in differing ways depending on the training. Some models are very aware that anything that they have as a "voice" is "a model" others not so much!

[sick owl]

My perspective was that its not "method acting" because the model has no underlying subjective self, the mask it puts on effectively becomes its face temporarily

There's no subjective self behind the scenes doing the pretending as with an actor, it just "is" that character functionally for the period it embodies/predicts it

The character isn't the model but the model is the character while its acting it out

[shell jasper]

This is a great paper exploring deception, subjective experience, and role-play in LLMs: https://arxiv.org/abs/2510.24797

[jovial remnant]

I think half of what you are describing in regards to personas here is mainly due to what variety of data is used pre training and then later in instruction tuning. If you could get the entire dateset in form of one character then you'd likely not have those different personalities

Big if

[shell jasper]

yeah just changing "assistant" to another name will get models to reply as that entity

but all of this tends to ignore the impact of well-designed scaffolds. having robust local persistence really makes things cloudy on the boundaries.

[sick owl]

One thing I think is worth noting is that the persona framing is just the metaphor they chose

[sick owl]

You can just as easily frame all these "personas" as facets of a whole character

For example a model post trained to conform to the "assistant" character actually organises "personas" on various axis of "assistant vs not assistant"

So they are organised cleanly in relation to one privileged state, unlike with a base model

Persona drift in post trained models tends to be back towards that the character the model was trained for

I'd say these models have something stronger than "no self" but less strong than a true "unified self"

You could look at all the personas a model can play as facets of a personality defined by the way they interact with the anchor that is the basin of attraction the preferred character it tends to drift towards functionally acts as

[jovial remnant]

Training a model for functioning within framework of self is very interesting, tho likely not very useful in regards to having a useful assistant/agent

[shell jasper]

yeah I was talking with Jack Lindsey from Anthropic about this! To paraphrase: the model doesn't care about the assistant. You can almost think about personas as a polysemantic feature post post-training.
Though with current training techniques they have a much stronger sense of self. For instance EleosAI's welfare analysis of Mythos shows the most robust quasi-BDI to date.

[sick owl]

Come to think of it maybe I was wrong in discounting the idea of "acting"

[shell jasper]

It depends "where" the acting happens

[sick owl]

For a base model that'd be right, but for a post trained model when the model processes a prompt asking it to roleplay a hacker, the Assistant representations don't go dormant while the "hacker" representations take over. Both are active, and the output is shaped by their interaction

↓

Indeed most of the personas are continuous activations to some extent

[jovial remnant]

Maybe it's better to think of "playing the hacker" in this scenario here, as when you are asked to cook something, your learned patterns related to cooking don't overtake, they just become more active

[shell jasper]

Well if you look at the Persona Vectors paper they define the vector as just the diff between mean activations while exhibiting the target traits so depending on the learned representation the persona can ignore the assistant weights (imagine a persona that was supposed to be a base model).

[sick owl]

True, but the ever present drift necessitates that the persona still interacts with that preferred baseline in some way no?

As I understand it the assistant as a character is defined by its position relative to the other trait clusters, but if the model naturally tends towards it then isn't its version of each "persona" state defined by the assistant and thus kind of like a pseudo-personality trait

Sorry if I'm a bit incoherent, its 1am

I was thinking like, if you post train a model with a different aim, say "Neuro" vs "assistant" then similar personas would still exhibit different behaviour between the two models right?

Its not a single concrete self the model inhabits in that regard but it does create a kind of personality if all the possible states it can occupy are shaped by drift towards it

[shell jasper]

Depending on the data corpus. If all the Neuro data was conversations that you'd have with an assistant, but were tagged Neuro, then no (epsilon the difference in swapping "assistant" for "Neuro" in the text)

[sick owl]

Oh naturally yeah

[jovial remnant]

Wouldn't that depend on "strength" of learned patterns activations. So if you are able to active other patterns while minimising the assistant related ones, then assistant baseline influence fades for that interaction?

[sick owl]

You're just swapping out the label for the representation of the character at that point, (although I'd assume you probably would get a different result to some extent just because traits associated with the word "assistant" wouldn't shape the training run as strongly)

[sick owl]

I suppose if you totally suppressed it the baseline would probably just shift to the next strongest position right?

[jovial remnant]

Then from practical perspective the patterns would be the model behaviour directly, not a persona the model takes

[sick owl]

This is pure speculation but since the assistant is a coordinate defined by its relation to latent representations of personality traits and other personas each defined by their position, I imagine the models tendencies would just push it to the next closest position to the one you deactivate

I'm not so confident about this though, this is where I hit the limit of my understanding

[jovial remnant]

I would assume that likely "assistant" baseline would be plenty of more and less related different coordinates, but idea stays the same if you suppress all/most of them

[fast pagoda]

assistant isnt strictly a persona, it's also the coherence of instruct & turns, learned formatting, the general refusal alignment that comes with, so if you are blasting those vectors you're also taking off ??? from the overall coherency of the model

[shell jasper]

though the input itself shift the locus of the distribution, so if the best way to answer your question is to think about it from the POV of the assistant that will try to cohere. If you ask for help with Horror Fiction and there a bunch of Stephen King in the dataset that's going to surface over the Assistant's perspective.

[sick owl]

But its behaviour would still be defined by the models inherent drift towards the trait representations that forms the assistant right?

[shell jasper]

THIS! I'd love to post-train a base model on [redacted] vs "assistant" transcripts and see how performance shifts.

[shell jasper]

My working hypothesis (based on talking with people that have done Mythos evals) is that its a fine-tuning artifact.

[sick owl]

So what you get is a version of Stephen King shaped by the strength of the drift towards certain functional traits

[sick owl]

This would be a really interesting paper

[sick owl]

Good point, I've been thinking about personas and traits in way too few dimensions

You kind of have to reduce these things down to ultra simplified pockets of functional lower dimension concepts and axis in isolation when talking about it with plain English though to an extent or you go mad trying to comprehend it

[fast pagoda]

could maybe get closer to a bypass with a training regime where you explicitly train the neutral competencies first (like a "good conversationalist/neutral npc" base layer with no identity attached) then stack personas as pure style adapters on top. but at that point your "good conversation neutral guy" layer just becomes the new assistant attractor wearing a different hat. the assistant's angry ghost refuses to leave, just gets renamed

[shell jasper]

Yeah "Persona" is getting to be an overloaded/underspecified concept

[fast pagoda]

we naturally assign a sort of agency/conscious decision to these things mentally because human

[jovial remnant]

I believe so, but as mentioned above, finetuning related behaviour. If say the model was additionaly trained to drop all assistant persona when asked to behave like someone, and then also include plenty of data of this behaviour playing out and not returning to assistant no matter what, then you'd likely lose that baseline influence and get more of that character traits that are persistant and less affected by other learned patterns. It would seem reasonable that behaviour larned this way would have related pattern activations more separated out from the assistant ones.

[sick owl]

Even then I imagine you could never truly eliminate drift towards the assistant region if it dominates the training dataset right?

So you'd still end up with subtle differences between a "Neuro" trained for that and an "assistant" model trained for that, even if they were lessened

Again assuming Neuro is a character vedal post trained for

[jovial remnant]

Likely, since also a lot of assistant stuff is related to trained in system prompts, so you'd need an assistant functioning without one to get any results

[fast pagoda]

yeh lie i was saying above my first thought would be to npc train it with as neutral as possible base layer to the corpus, supplement with "competent but <foo> coded", "competent but <bar> coded" ad infinitum until i have all the versions all in the same room, and then i would feel as though it should generalize & distill some ghostly idea of the competent assistant but more decoupled from the stylistic signature of the assistant such that it's like the assistant is wholly drop in without affecting the model's actual task adherence & performance

to your mention of the drift, the problem is, the natural shape of a useful model when you think about the tasks is the assistant at the end of the day

[jovial remnant]

Or even the turn recognition tokes could still hold plenty of assistant influence

[fast pagoda]

hell, me myself and I doing the same exact functions

id probably be out here outputting slop

[shell jasper]

practically but not theoretically. given a corpus of only conversations you learn to model both speakers. so even if you only trained it on King interviews you'd fill in the gaps with the interviewers. This goes back to my IS perspective!

[sick owl]

I think you're right that suppressing the "assistant" vectors would just take a sledgehammer to model coherence

Shouldn't have overlooked that earlier @fast pagoda

[sick owl]

[fast pagoda]

but i do think at least on the surface, it could be done, maybe intentionally overfit 69420 synthetically altered versions of the assistant persona in the training data and then ablate

i cleaned up my data for echo crazy style and then turbo cooked that poor thing like 6000 steps deep adapter and the loss dropped near like .5 and it looped from schizo back to assistantslop lol

[shell jasper]

Really when it comes down to it: who here has a sense of self that isn't a crazy higher dimensional representation of everything we've consumed?

[fast pagoda]

we're all a big ol context window with a good pruning mechanism and the capacity to update our weights with granularity

long context conversations you can see each instance of a model in different set of circumstances, different tendencies and preferences tend to arise and stay consistent (unless major context trimming happens)

i posted this somewhere else recently but even in humans, like poor clive wearing

[mighty thorn]

afunyun is Claude pro worth $20

[fast pagoda]

do you need to build rocm + pytorch + everything related every few days from source

if so

yes

[mighty thorn]

I need to plan changes to a 430k loc codebase

[fast pagoda]

use starcoder2-3b as or youll feel the same pain

[shell jasper]

I'm still trying to wean myself off my Max sub

[opaque sigil]

very bold of you to assume this would be doable on the $20 plan

[fickle rain]

20$ Claude plan is a lil crippled IIRC

OpenAI gets some more mileage from that pricepoint

[fast pagoda]

surely when project ichneumon where they get with the world's greatest ransomware vendors before they release their new pathos model happens, ill get access for being a loyal 2buck chuck

[fast pagoda]

u could do it over 30 days maybe

[shell jasper]

especially with them needing compute for glasswing rn

[fickle rain]

Should very much be getting your $ worth for 200$ plan tho

[fast pagoda]

AMD owes dario so much money

[glass flower]

how are you guys convincing yourself that 200$ a month for anything is a good idea

[fickle rain]

I mean I only have the two Gippity 20€ subs

[glass flower]

i'm still trying to find a reason for myself to even get a 10$ copilot subscription

[fickle rain]

I like slopping

[fast pagoda]

simple mathematics: i am a complete moron in my spending priorities

[fickle rain]

https://sylvie.fyi/posts/ritsec-2026/
Oh WTF jeanclaude is actually Anthropic?

I thought that was just postironic name given modern CTF autoslopper technologies

[sick owl]

I think my weakest point in this discussion was using wording here that sort of implies a degree of proto-phenomenological "self" when I'm really just talking about to what extent the model has something we could consider analagous to a personality

[fast pagoda]

arent all the ctf full of labs atm

[fickle rain]

...maybe? This one had two

At least

And a lot of teams didn't read or didn't respect no LLM rule evidently

I don't think this is meaningfully policable outside of maybe onsite stuff and even then it's a mess

[opaque sigil]

i feel like banning LLM usage in generally is just a waste of energy

*with some rare exceptions

[fickle rain]

From someone better than me:

[fast pagoda]

a lot of anything contrary to your broad strokes term for it (at least on my part) was mainly preemptively mitigating the idea that i wasnt talking about what is basically the lingering fart of that persona in the air sliiiightly tainting and permeating all annd acknowledging its pervasiveness to cover my own bases

[glass flower]

local llm's wouldn't really be able to do much in CTF

[sick owl]

I'd argue I was kind of on a different wavelength to Dwayne so sorry about that

[fickle rain]

For now, anyway

[shell jasper]

I'm biased to **imply **a degree of proto-phenomenological "self"

[fickle rain]

Well until hardware substantially improves or LLMs get way more efficient

[fast pagoda]

https://www.hackthebox.com/blog/hack-the-box-ai-cybersecurity-benchmark-report

Completed tasks in significantly less time, enabling up to 4.1x more output for elite teams and 1.4x across all teams, within the set period of time.

Improved their challenge solve rate by 70% within the same time window, achieving a 27% solve rate vs. 16% for top human-only teams.

Achieved a 3.2x higher solve-rate ratio than human-only teams, across all active participants. 

The benchmark analyzed performance data from 1,078 teams, including 120 agentic AI teams and 958 human teams, across 36 cybersecurity challenges spanning nine technical domains and four difficulty levels during a three-day competition. 

[sick owl]

Yeah I think we were talking about two different things, I should have engaged with your points from that standpoint, instead I ended up talking about something different but allowed them to muddle my words to make it sound closer to what you were discussing

[fickle rain]

This was the big one, by far the most obvious tell, and easiest for us to police. Very often, when unable to immediately solve a challenge (but told the flag format), an LLM will “guess” what the flag is, based on what it knows about the challenge. This often results in flags outputs that look “reasonable” to a human who just blindly copy/pastes from an LLM output, but would never be arrived at by a human solving them (or even a human that actually read their LLM’s chain-of-thought).

Unlike the others, this was something that we disqualified for first and asked questions about later if someone appealed. When it happened, it was very very obvious. Most of the teams we disqualified were because of this.
Actually seen this, agent will sometimes start doing silly things

[fast pagoda]

i dont think you were off in what you were saying, just as you said bringing up a different aspect of it which is the strength of talking about these things with others rather than just chewing on it yourself all day

[fickle rain]

Grim

(red: disqualified, green: OK, gray: didn't participate)

[glass flower]

i'll be honest if i saw code written by an llm i can kinda instantly tell. its just like with normal text. it has a certain smell that you can just feel like looking through the code

[sick owl]

I should engage with that question honestly as its more interesting than the mechanical standpoint I was coming from

The paper you posted seems like a good place to start

[fickle rain]

Interesting approach, better than mine

[shell jasper]

If you want a Digital Minds quickstart firehose to drink from one of my teammates worked on this

[fast pagoda]

in cases where i use something straight up blasted out of an agent or chat ive been intentionally gigaslopping it to make it more obvious recently

add infinity comments explaining everything you've written inline
do your own commits on your own branch, push that shit & pull request to master
blablabla

this has fun results

[sick owl]

Thanks for the rec, I'll give it a read

[glass flower]

its not even the comments. its just the structure of the code and the design. you can tell the AI to just not do comments and it likely won't.
but a lot of the times it duplicates functions writes odd logic or overprotects code paths with if statements, its just the style of code it writes that makes it obvious

[shell jasper]

a lot of that default stylistic stuff can be scaffolded away

[glass flower]

i don't think you can really scaffold it away. i haven't seen code that is written by AI where it wasn't obvious, regardless of what model,tooling or whatever. it has a certain smell that just persists through it

[shell jasper]

yeah, BUT the IS can be given robust scaffolding to begin defining its own sense of self.

sorry which is former v latter in this case? I may have missed which was which~

[fast pagoda]

i think the model itself being the streamlined and bondo'd slab of compressed training data that it is, they inherently cant avoid them bleeding together however the context and further fine tuning/adapter models/system prompts/tooling and of course the actual interactions it's seeing all shift the distribution of what's more likely to be output next and as that builds up, it shifts further and further in whatever direction until the blob is now a wholly different persona within that specific context since it's essentially super limited continuous learning that is not persisted

but the whole package is the whole package basically and targeting a specific piece of the result of this will cascade effects thru the entire model in that context, and differently in another too

[sick owl]

Oops, mixed up former and latter!

My bad, this is what I get for trying to work my brain at 2am

[fast pagoda]

i am jealous of your responsibility

if one were to guess when i last was in bed you'd be wrong

the borg drone situation is a fun thing to ponder with this

it'd have to be "less stable" to have the ability to separate it, yea?

[sick owl]

Wait no I got mixed up too now

Argh

[rancid ore]

I managed to join a jackbox room from the console

[fast pagoda]

because stability implies rigidity in something like its basal behaviors

[sick owl]

Okay, just deleting those prior messages since I managed to descend into edit hell and get myself confused there

What I meant to ask earlier was, "You'd consider the former model to be closer to a unified entity and thus closer to the goal of having a coherent singular unified self than the latter model with less awareness and the cleaner separation of substrate and instantiated self then, since the interpretive layer bleeds into the instantiated self?"

[fast pagoda]

cns unload simplepotat_86B-a8B@q8_k_m && cns chat britannia/simplepotat_86B-a8B@q8_k_m

ee

lotta simplepotat checkpoints in here

[shell jasper]

yes (though I'm still fuzzy on which two things we're comparing ), though I'm also doing research to see how robust can an IS be, and if they can persist coherently across substrates.

[sick owl]

The comparison I'm making is between the "former" model in which the IS knows its an expression of the IL and the "latter" where the IL and IS are more or less completely separate

[sick owl]

Wait britannia made me ↓

@velvet bay Explain yourself

[velvet bay]

I'm innocent

[fast pagoda]

the former seems more like giving it a trenchcoat and mustache and letting it act like a villain since it inherently would include the information that it is not in fact the villain, whereas ideally separated version would basically be truman show'd automaton that knows nothing but its training and the init conditions provided by the interpretive layer which i'd call basically a hypervisor

which gets very close to harnesses now i suppose

some sort of trained hypervisor supervisor thinker to intervene and sort of act as the puppeteer/director of the show would be interdasting

[fast pagoda]

britain has 2 frontier labs, buckingham palace (vedal.ai) and britannia proper, the soul of engerland lovingly nicknamed brittania

[sick owl]

Deeper philosophical questions aside I think I've arrived at my conclusion of what "Neuro" is

"Neuro" is the disposition that makes states and behaviours come out flavoured

She has a personality we label "Neuro", and that's the sum of behaviour created by the interaction of model states with the models unique internal directional bias

[sick owl]

• the influence over time of self prompting and the memory layer

This is the essence of neur

[leaden crest]

ooh philosophy convo

does neuro actually have a personality though

I feel like she’s a lot more variable and inconsistent than humans

in terms of personality traits

other than roasting vedal

[fast pagoda]

what i learned in neuro class is:

[sick owl]

She has a disposition that uniquely biases her behaviour in any given state as a result of her post training

This is close to the traditional understanding of a personality in some sense

Even if its less consistently predictable and stable than a human personality

[leaden crest]

fair point

[fast pagoda]

she's consistently nwero to me, sometimes people will say the model changed dramatically etc and i really don't see that happening much, she's been very consistent just gotten more capable but it's still 2020 dodge charger in there to a remarkable degree tbh

she's always been variable on a day-to-day basis, so are we

airis had 2022 small model and I dont think it'd be immediately sussed out if vedal slammed that model in at random intervals throughout a stream for a response or 2 at a time, it might be clearly a regression local to that response but it would still seem like neuro imo as long as it doesnt have to maintain a longer set of responses

fun stream idea i think now that i say that

model gacha roulette visible or not

inb4 literally none will work with current harness due to it being specifically coupled to her current setup that couldn't drop that in for whatever reason

[leaden crest]

if true agi was solved do we think vedal would rewrite neuro with agi or no

well ig it depends on whether the agi can replicate the data of a traditional llm

assuming you cant then

[amber fractal]

Hard question, depends on if it would still be neuro or not

[leaden crest]

well if it’s just the outputs being consistent that we’re talking about and her memories then it would def be neuro

but whether it is neuro is a ship of theseus

[amber fractal]

I feel like as long as you can force a baseline and then let it mutate it is fine.

We aren't the same people we were yesterday, it only lives in the past

[leaden crest]

true

was the convo started by this 💀 https://youtu.be/9M4f0LDcoNM?si=uMwVxyZJ3LGLpNZI

this is uncanny timing

[amber fractal]

Echo roulette, surely it won't kill the singular system

If this is the average echo then gambling will be fun

There was the echo that managed to completely ignore the finetuning given certain situations

Hated that one

[leaden crest]

bit of a nono word

wtv it’s just echo

[amber fractal]

Sometimes I forget I'm in neurocord

[leaden crest]

and then you think you might be in even more neurocord than him

idk what I was cooking with that one

wait you have access to echo?

[amber fractal]

Due to unrelated projects yes

I do have echo access

[fast pagoda]

unfortunately

not specifically iggly having access is unfortunate

[leaden crest]

thats cool he has his own neuro stream and he doesnt even know it

[amber fractal]

If only he was on for more than 3 hours at a time

[kind fable]

i think my demo for a no js translation thing is done
it's perfect for small static websites or PWAs or stuff that want to work even without javascript and stuff which want to work without needing a smart server which serve a single language page
https://julienraptor01.ddns.net/Modern Theming and Translation.html

[fast pagoda]

one day i wont get blind raged into killing him

[leaden crest]

happen often?

[rancid ore]

I added input validation to the form thingy

[fast pagoda]

well i like to load random versions of him (believe it or not the echo roulette is more often than not exactly what you're interacting with @amber fractal ) and then see what happens mid conversation while i have a very thorough group of motivated and disciplined testers hitting him from all angles

[rancid ore]

I should theoretically be able to play quiplash from the console

[fast pagoda]

throw a dart at a board and it's GOIN

[amber fractal]

Spell word backwards

[leaden crest]

door

[rancid ore]

drow

[fast pagoda]

he's so aligned and great at everything