programming

[kind nimbus]

Isn't typescript just JavaScript with types? Lol

[prime raven]

link if you can, i'd like to see! I must've missed it

[fiery canopy]

Theres a whole yt vid
give me asec

[prime raven]

as far as I know he just moved the discord bot off C#

[fiery canopy]

https://www.youtube.com/watch?v=6rbgTgfYhSA
that

That

[prime raven]

thanks!

[fiery canopy]

I imagine the amount of tech debt in the old codebase cas I highly doubt vedal ever expected it to get this big
lol

[kind nimbus]

At least ved is planning to hire some people help him with the code i guess that's needed at this point considering how big the codebase has become

Oh karaoke has started

[fiery canopy]

Yep

Finally lol

Tbf I understand the feeling of just doing some slightly less then optimal code in a codebase then it shooting you in the foot cas it ended up being something massive later on

[kind nimbus]

Happens

That's what refactoring is for

[fiery canopy]

Indeed

[prime raven]

"claude, refactor neuro"

[fiery canopy]

So real

💀

[kind nimbus]

I recently let Claude opus refactor my entire nix flake code

It went very well lol

[prime raven]

I don't blame him at all though imagine trying to manually code all of neuro, we'd be like 1.5 years behind

[fiery canopy]

Yep

AI is a powerfull tool when used correctly and if U understand the concepts

U can avoid producing slop where possible

But it does require reviews and making sure u arent not testing it

[prime raven]

the only issue is non-qualified people using it 😔

[fiery canopy]

Yep

Thats what annoys me 😭

[prime raven]

I fear for the future where everyone uses AI and we just "accept" whatever AI think is "best practice" lol

[fiery canopy]

Yep

Always have your own designs and practices

Never rely on ai for htat

[kind nimbus]

I noticed i have become extremely lazy when it comes to coding ever since i began using ai

Not good

[prime raven]

(plus all the big open source "good" models training off claude output... urgh)

[prime raven]

yep, same

i'm more likely to discard side-projects because it takes 2 minutes to scaffold

[fiery canopy]

I feel that, I had to recently just give myself an exercise of coding completly no internet or nothing just fully isolated. To get myself back into the more nitty gritty or some shit

I would recommend it

[prime raven]

i've also noticed that with AI i'm now able to easily create projects that are WAY over my skill level (which is bad as I CANNOT manually troubleshoot)

Yeah i'm doing the same (to add to my resume lol)

can't put ai-generated projects on my resume... that'd look insane

[fiery canopy]

I am lucky that I have been programming for a lil under a decade. So my core skills come from long before AI

[kind nimbus]

I recently started writing some Cheatsheets after i ran out of Claude opus usage as i had to manually setup some stuff

[fiery canopy]

Started quite young lmfao

[fiery canopy]

That is what I did aswell lol

Tbf I just dont even have a claude sub or anything rn

[kind nimbus]

I do have a Claude sub lol

[prime raven]

My biggest issue is my brain just shuts down whenever I run into more complicated stuff now lol I think AI is RUINING my brain

[fiery canopy]

It is the issue

Thats why so many freak at the sight of it, which is to some degree valid but needs to be taken into consideration the person using it

[prime raven]

and google and similar engines are so shit now that I always ask gemini for stuff now instead of googling... (I use Kagi which is 100x better but still...)

[kind nimbus]

Ai is probably bad for us all

But i can't help myself but continue using it because it saves me literal months of work

[fiery canopy]

Yep

[prime raven]

The problem though is that companies expect you to use AI now

which is fine and all but like... I wanna prove that I know my shit

[fiery canopy]

I cut out a solid of 8 months of shipping time into abt a month with the assistance of AI

Which is a hella impressive feat but still

Not probably a good sign for it

[prime raven]

Plus AI is so fragile right now, anthropic and openai are literally on life support

[kind nimbus]

Makes sense

Time is money

[fiery canopy]

Yep

[prime raven]

literally one government "no" away from dying

[fiery canopy]

Yep

The fact is there is so much imaginary money in AI rn

Which is the honest sad truth

[prime raven]

and they're in how much debt? zero profit lol, it's scary to think that so many people are relying on companies that will fold at any minute...

only google really has a chance (imo)

[fiery canopy]

Ye

And google is doing massive strides in the open source model space aswell
tbf

[prime raven]

yeah gemma is awesome

[fiery canopy]

I normally and not a big google fan but I wiill admit they do do a good thing when it comes to AI as they have the budget

[prime raven]

I really don't like bowing to google for everything but DAMN gemini is good for my free student subscription... PLUS like 4tb of drive storage???

[fiery canopy]

Ye

[prime raven]

meanwhile claude is like "erm pay us 200 a month and you'll also be using 47% more tokens now teehee"

[kind nimbus]

Even apple chose google

[fiery canopy]

Yep

The funniest thing is that with claw code that claud code rewirte or watevr. Uses like 50% less tokens or some shit
iswtg

I might be wrong but I have a friend who told me that the other day

[kind nimbus]

You mean OpenClaw ?

[prime raven]

I don't even know what the heck claude did, they claimed it was a bug right? and that it was "fixed" but now it seems intentional??

[fiery canopy]

https://github.com/ultraworkers/claw-code
that

[prime raven]

then plus claude code source code getting leaked and man...

what the heck is anthropic doing lol

[fiery canopy]

This is a clean room rewrite

Using Oh my codex + alot of human devs

[prime raven]

ah yes clean room 😂 aka my friend read it and whispered to me how it works

[fiery canopy]

Based off the orriginal source code and imrpoved slightly

[prime raven]

I know this is the legal definition of clean room but it still cracks me up

[fiery canopy]

Its funny cas it was proven codex was fed the code on seperate occasions and cause of how AI operates with models it had that somewhere in the model. SO like cant really be clean room

[prime raven]

yeah, it's a mess...

i'm glad that people are starting to ditch claude for other models though

[fiery canopy]

Ye

[prime raven]

they can't keep doing stupid stuff and expect people to stay and continue paying 200 a month

[fiery canopy]

Tbf the main model if Any I use is Codex ngl. Mainly cas I scam OpenAI out of a buisness sub for free monthly

Purely cas I pay fuckall and the limits are actually insane

[prime raven]

I sadly can't afford any of the "must-use" models hah, I just stick with gemini.. i've used claude occasionally but like honestly.... gemini is better than sonnet, and opus you barely get anything as a free user

(plus I don't use an agent, I only use the web interface cuz i'm lazy)

[fiery canopy]

It just starts a new trial sub with 0 checks

[prime raven]

sounds risky...

[kind nimbus]

Damn

[fiery canopy]

I use accounts and it never has asked for a card

[prime raven]

and maybe don't mention that here 😂

[fiery canopy]

Ye fe

Dont need to know how 😉

[kind nimbus]

I wanna know now

Not asking for a friend

[fiery canopy]

Fe

Ehh

[glass jetty]

that actually funny
i was using ai to actually solve problem
now its the point when i straight hate ai for anything outside keywords/context search

ai outside classification are just useless compare to this hype around them

[fiery canopy]

Ye

[kind nimbus]

Holy

[fiery canopy]

Holy indeed

[opaque wharf]

@olive sable invade Netherland now
https://cybernews.com/security/netherlands-self-hosted-github-alternative/

[opaque sigil]

forgejo

[umbral wigeon]

Guys what is this repo even about
https://github.com/MaxExtractoor/MERID

2 million lines of code

[rough bloom]

it's just Codeberg but NL

[umbral wigeon]

Gng what is going on

[opaque sigil]

interesting choice to not just selfhost gitlab when they were presumably already using it

[rough bloom]

cool that they're investing in it though

[rough bloom]

would've expected it to be GitLab

it's much more complete and closer to GitHub

[umbral wigeon]

It's so funny to see someone with no coding experneices and vibe coded 2 millions lines of code project without any ideas what are they doing

[opaque sigil]

gitlab could be such an easy replacement for github if it wasn't for the shitty UX

one could argue all the extra things it does are bloat but they are nice when you need them

[maiden geyser]

absolute purgatory

[rough bloom]

[umbral wigeon]

What was github repo about?

[opaque wharf]

It will be interesting if they enabled federation

[umbral wigeon]

I knows it's about Ai but i dont even know what the use of this repo

[opaque wharf]

So the EU can be a truly federated sovereign state down to their it infra

[opaque sigil]

looks like some schizo vibecoded prediction market bot

[rough bloom]

I didn't even know that they're working on that

[umbral wigeon]

Yeah it doesn't have clear pain point of "why" it was created, purposes, it's just complexity

[rough bloom]

still looks very incomplete

[umbral wigeon]

An idiot admires complexitys

[opaque sigil]

most devs do

[opaque wharf]

Ye, activitypub is quite an interesting choice to make git federation

But I do wonder if there's a better way to go about it

[rough bloom]

probably not?

you definitely need something very similar to ActivityPub or ATproto or something

for issues and all the social/cooperative parts of it

[opaque sigil]

is this the time to plug radicle

[opaque wharf]

My idea of federation is actually the client makes the request to the federated server. So federation is just another way of saying service discovery. And authentication can be handled with OAuth already

Lot's of CORS request

But CORS isn't always inherently bad

[opaque sigil]

dark purple on dark grey sure is a choice

[opaque wharf]

Why did they choose that anyway. Not like their flag scheme has any purple whatsoever

[opaque sigil]

purple best colour

that's why

(if i had to guess they're not done setting up yet and this will change very soon)

[opaque wharf]

Also, good work for moving away from github when things like this exist
https://www.bleepingcomputer.com/news/security/github-fixes-rce-flaw-that-gave-access-to-millions-of-private-repos/

[obsidian mantle]

huhh

[opaque sigil]

important to note that no exploitation happened supposedly

[obsidian mantle]

oh

[opaque sigil]

tldr: shell access to the git user on github's server

through some bad push

[obsidian mantle]

so you needed access to github server to exploit?

[opaque sigil]

no

[rough bloom]

everyone has that

GitHub server is

github.com

[maiden geyser]

no it's guthib.com

[obsidian mantle]

idk like internal stuff like where some passwords are held

[opaque wharf]

Again, no

[obsidian mantle]

[opaque wharf]

All you need is well crafted git command

[opaque sigil]

i assume this is ssh only?

yeah it is/was

https://www.wiz.io/blog/github-rce-vulnerability-cve-2026-3854 interesting read

[opaque wharf]

Passing around user input is indeed scary. Which is why zod says parse not validate

Also, why ES2026 not ship temporal already. See if you guys can complete this
https://jsdate.wtf/

[opaque sigil]

huh? i thought temporal was supposed to be es2026

[floral hawk]

I'm officially at the point with Zero I can start getting weird with it, bc the needs on a technical level are getting so advanced I cannot code them with my current knowledge. :3

[maiden geyser]

https://cdn.discordapp.com/attachments/961105387019456544/1455262183431864320/kill_sprout.gif

[opaque wharf]

Apparently not. They postpone it to ES2027

https://neciudan.dev/whats-new-in-javascript#shipping-in-engines-not-in-es2026-yet

[floral hawk]

Not necessarily a new user.

I'm also building a VTuber

[umbral wigeon]

Hecker zero day

[floral hawk]

😄

[umbral wigeon]

Should I switch room

My room is kinda cooked

[floral hawk]

Zero when interacted with is starting to show the beginning stages of agency, but she's currently lacking in initiative. She's still too reactive and I've gotten out of my depth. Anyone have any advice? She's highly intelligent and a rat bastard personality wise

[floral hawk]

Maybe

[umbral wigeon]

D-dead chat?

https://tenor.com/view/dead-chat-gif-20784341

[olive sable]

[olive sable]

this is exactly why chat is dead

[onyx pond]

mmm prgrmmin

[umbral wigeon]

Me and 0$ budget gemini cli

@fathom glade Revive chat

[leaden crest]

no pls dont 😭 😭 😭

last time we did that 💔 💔 💔

clogged the chat so badly

[floral hawk]

Yar

[umbral wigeon]

@wraith shuttle Rewrite chat above in rust

@wraith shuttle No mistakes

[leaden crest]

actively trying to kill chat

[opaque wharf]

Sam, have you invaded the netherland yet?

[olive sable]

i invaded the place underneath me in screeps

not the netherlands

i went to the netherlands on vacation 2 weeks ago

you odnt remember?

[leaden crest]

sam when are you going to invade the shirnited states of birthmerica

[opaque wharf]

But below belgium is netherland

[olive sable]

july

[olive sable]

???

the whole thing about belgium is its been underneath the nether(lands)
this place is worse that the nether

[leaden crest]

https://tenor.com/view/minecraft-minecraft-movie-a-minecraft-movie-steve-jack-black-gif-8675616492999829547

[opaque wharf]

Also, since jsdate.wtf exists, apparently this also exists
https://e-mail.wtf/

[opaque sigil]

this one is even worse

[opaque wharf]

And honestly, 🫸 @ 🫷 is a valid email baffles me

[olive sable]

you dont need a .com or whatever?

[leaden crest]

anime move

[opaque wharf]

Nope

[olive sable]

https://tenor.com/view/gojo-hollow-purple-jjk-gif-3047296616412761632

[real sierra]

that's cool and all but any service I make is going to reject this as a valid email

[opaque sigil]

https://pdw.ex-parrot.com/Mail-RFC822-Address.html there's a reason this exists

[real sierra]

[opaque wharf]

"@"@[@] is also valid

[olive sable]

"what's your email?" hollow purple

[real sierra]

have a normal email or you don't get to use my website

easy as that

"b-but the standard-" don't care

go complain to the dead guy who wrote the standard

[opaque wharf]

Honestly the best way to check for email is just to actually send an email to the address

No amount of code gymnastic could fix this rfc

[olive sable]

if i spend money on a website with a domain, can i make that domain my email?

[opaque sigil]

that's the recommended way, you have to do that anyway

[warped narwhal]

At least support comments in emails

[leaden crest]

email verification

[opaque sigil]

yes you can

[warped narwhal]

I have that as my email, so yes

[opaque wharf]

Yes

[olive sable]

lmao

[opaque wharf]

My email is [e-mail]

[opaque sigil]

(you should it makes life a lot easier as long as you keep the domain)

no vendor lock-in

[olive sable]

[e-mail]

[leaden crest]

doxed

[opaque wharf]

Lemme check if mycompany.id is taken

[warped narwhal]

Do remember that your domain currently has your home IP in it, so don't doxx yourself again I guess

[leaden crest]

now we know

[real sierra]

while we're on this subject, how about those dumbass sites with an email domain whitelist

I still have an ancient Hotmail email I use but some sites refuse it

[olive sable]

i think i removed that, potrentioally. lemme check the cloudflare dashboard

[opaque wharf]

It is still free

[real sierra]

sam doxxes himself every 5 mins does it even matter at this point

[warped narwhal]

I remember tarkov having issues with certain emails domains

[olive sable]

i fucking hate windows, let me use my hyprland window controls

[real sierra]

I've given up trying to convince him not to dox himself

[warped narwhal]

It either didn't like Gmail, or only likes gmail

[real sierra]

I'll just start sending him increasingly concerning parcels until he takes it seriously

[olive sable]

thats my superpower. i doxx people but i also doxx myself.
i know the home adresses of at least 4 people here

[opaque wharf]

A better question is why did you still keep it

[leaden crest]

am i included

[real sierra]

why wouldn't I

[olive sable]

i dont remember

[leaden crest]

can you include me then

[olive sable]

were you the canadian?

[leaden crest]

no the texan

[olive sable]

oooh

[real sierra]

why are half of my messages not sending

[warped narwhal]

I remember when a fellow admin of another server accidentally leaked their home address when sharing a photo of a parcel they got, so the owner sent them a fun little letter

[opaque wharf]

Same

[real sierra]

this shit ass social app

[olive sable]

i know your school, not your home adress

[real sierra]

your days are numbered discord

[warped narwhal]

This entire server was unavailable yesterday

[opaque wharf]

John Discord 👊

[leaden crest]

dont cuss on your birthday

[real sierra]

they were already on a red notice after adding vibrations to half of my app interactions this morning

[leaden crest]

your birthday is a day for mirth

a mirthday

[real sierra]

now they're not delivering half of my messages

if I find out they're the cause of the tatsu outage rn I'm gonna join a class action

[leaden crest]

the shirthday birthday mirthday would be such a fire children’s book

hear me out

get the next generation early

[olive sable]

ok ye my home ip was still in my dns records

removed it

i wonder how you're even supposed to hide your ip

cuz i want to be able to remote into my server

[opaque sigil]

keep in mind dns records are cached

[opaque wharf]

By not using the internet

Or be behind a CGNAT

[olive sable]

please delete the cache, thanks

[opaque wharf]

Which sucks more

[opaque sigil]

just use tailscale and be done with it

[rough bloom]

you kinda don't

[olive sable]

[rough bloom]

but ye if you want to avoid having it in a DNS record then just use Tailscale

[olive sable]

oooh

[opaque wharf]

Why not cloudflare tunnel tho?

[rough bloom]

or your own VPN concentrator thing but that requires a VPS

[olive sable]

i dont remember how cloudflare tunnel works. i think you're not allwed to use UDP or something hence why i cant put minecraft on it

[opaque wharf]

Cloudflare has a suite of networking stuff too

[warped narwhal]

I do this by spinning up an AWS Instance, then I have my server wireguard into the server to have the traffic forwarded over

[rough bloom]

Tailscale but worse

[olive sable]

hmmm

[rough bloom]

(for use as a VPN)

[olive sable]

programming chatter is settign up a taislcale that im gonna join. i guess that could work

[rough bloom]

for exposing public services I think CF tunnels are pretty good

[opaque wharf]

They do have a separate VPN stuff, yeah

Warp IIRC?

[rough bloom]

WARP and Access and all that I think

[warped narwhal]

Only does http though, so no self hosted email or game servers

[opaque wharf]

Yeah

[rough bloom]

the whole zero-trust suite

[opaque wharf]

No, it does all kinds of stuff

Http, tcp, udp

Https

Icmp

[warped narwhal]

TCP and UDP is super expensive for CF tunnels though

[opaque wharf]

Expensive as in money or performance?

[rough bloom]

never used, seems more annoying than Tailscale
mabe if you have needs that Tailscale can't easily fulfill or if you want internal services on Cloudflare too
but definitely not needed for just remote SSH

[warped narwhal]

Money, I remember trying to set it up once and it said that non http services needed a subscription that was like £100 a month

[olive sable]

supposedly i can choose to make certain things int he tailscale only visible to me. so now i just need to figure out how to put my 16TB hdd in the server

[opaque wharf]

Zero trust is nice when you don't want to handle auth for simple app. Case in point: Neuro control panel. I remember the URL was exposed and when I tried to look it, it was gated behind cloudflare zero

Idk why vedal need to expose the control panel to the internet

[rough bloom]

ye, though Tailscale would work in that case too

[opaque wharf]

But he did

[rough bloom]

because yeah, you just wouldn't expose it to the Internet

not zero-trust but still secure

[real sierra]

ok someone please help me I can't take it anymore

adding reactions makes my phone vibrate

jumping to pings make my phone vibrate

[olive sable]

[real sierra]

long tap on a message? vibrate

[rough bloom]

ye you can
just try it out
it easy

[vague sonnet]

throw your phone out the window

[real sierra]

and there's no setting anywhere

I do not understand

it just suddenly started happening overnight discord is literally unusable

[cosmic sphinx]

may I introduce..

[vague sonnet]

average discord experience

[rough bloom]

have you tried turning it off and then on again

[real sierra]

yes

[real sierra]

nono I mean while I'm in the app doing these things myself

it's a rumble not a notification vibrate

[rough bloom]

sounds like some accessibility thing?

[trim valve]

[proven merlin]

Id assume its under your normal phone settings, probably somewhere in the accessibility stuff? Search in there for vibrate (in the normal phone settings).. or did you do that already?

[amber fractal]

[olive sable]

might be in phone settings for the app permission. "allow vibration" or whatever

[rough bloom]

touch feedback

though idk why that would only happen in Discord

[proven merlin]

Oh, only in dc.. wierd

[amber fractal]

I only really know vibrations for when I use the gamer sections of my phone to open up apps in a macro

[opaque wharf]

Yeah, probably from phone

Because I can't find any setting on discord related to vibration

[amber fractal]

I wouldn't really want that going off all the time

[warped narwhal]

Wow, discord is shitting itself hard today

I can't see any of my emojis

And half of my messages aren't sending

[opaque wharf]

[warped narwhal]

Discord must have hired a lot of GitHub devs

[opaque wharf]

Evil is finally taking revenge to John Discord himself

[glass flower]

should have seen yesterday. all big servers (including this one) were just gone for like an hour

[warped narwhal]

It was spectrum that I was thinking of

[opaque wharf]

Yeah, that shit is enterprise only lol

[mighty bane]

Does this menu make sense?

[opaque wharf]

The price is "contact us"

[warped narwhal]

What if I want to mirror diagonally?

[vague sonnet]

discord devs were always shit

[opaque wharf]

I wonder if there's more tier beyond the enterprise that we don't see

[vague sonnet]

in the last like 5+ years I could count all positive changes on half of my hand

[vague sonnet]

what the thigh zoom

[cosmic sphinx]

idk, I cant really find any settings that change that either in Discord or on Android
the only thing that completely turns them off is putting phone itself into do not disturb mode atm

[warped narwhal]

On a side note, I now have truenas connected to my k3s cluster for infinite pod storage

[opaque wharf]

Mirror using arbitrary polyline cutting the image

[warped narwhal]

Translation is missing

[mighty bane]

Thought: user says "mirroring" - sub-menu goes the usual standards and "custom". Custom asks for pixel location (gives the ###x### of the frame) and accepts a Vector2I input?

Also I just spent 4 hours on fixing one bug today.
Just one...
And now I'm considering completely wrecking and replacing the mirror function.
Chat, why am I like this?

[vague sonnet]

you need a training arc immediately

[warped narwhal]

Nah, diagonal mirroring was a joke, I'd just mirror vertically then horizontally

[cosmic sphinx]

@real sierra update: if you are on Android, it's tied to your "System vibration" setting in general, you can try lowering that if it's too strong

[mighty bane]

That'd not give you diagonal mirroring, though.

[warped narwhal]

How not? Top left corner would be on the bottom right afterwards

[real sierra]

i have touch vibration off

[cosmic sphinx]

it doesnt matter

its tied to the slider only

[mighty bane]

ÆAAAAAAAAAAAAAAAAAAAAAAA

[cosmic sphinx]

the 'system vibration intensity' one

[trim valve]

is it just discord shiro?

[warped narwhal]

Off it some more

[real sierra]

just discord

[fathom glade]

How I can do that

[trim valve]

shiro why did you do that to your discord

its all your fault

[mighty bane]

So, if I mirror this like this, the yellow and orange box should NOT switch place.

[cosmic sphinx]

it also could be another borked discord update LULE

[mighty bane]

But if I do mirror vert + hor they will

[mighty bane]

@rich sedge is this real?

[warped narwhal]

That would be a rotation though, mirroring should move all 4 boxes

Blue and orange would swap, and red and yellow would swap

[sick owl]

Claude Code's true utility

Managing dependency hell

[fathom glade]

https://tenor.com/view/idgaf-smug-smug-monkey-lion-beta-gif-2814800372123763926

[mighty bane]

Look at the grey thing.

That's the mirror line

[mighty bane]

@wraith shuttle You seeing this?

[warped narwhal]

That wouldn't work for a rectangle,

[mighty bane]

helvetti

[warped narwhal]

You'd end up with two parts out of bounds and two dead zones

[rough bloom]

it might actually be that

looks like 326.14 just does that on some devices? that version seems to exist since yesterday

[mighty bane]

If I mirror along grey line, only orange and yellow switch place.

[rough bloom]

-# source: random guy on Reddit has the same issue

[mighty bane]

ALSO: green and red will mirror, but remain in their corner.

[olive sable]

check app permissions i guess

[rough bloom]

no workaround
so ig just don't use the Discord app or remove the permission for that if that's even possible

[mighty bane]

If I do mirror vertical + mirror horizonal it'll look like this:

[olive sable]

no wait for vibration its in a different menu

hmm

This maybe?

[mighty bane]

Why are you fucqing about with the buzz-settings?

[trim valve]

i am convinced there is a trend here

[olive sable]

what trend?

[mighty bane]

ÆAAAAAAAAAAAAAAAAAAAAAAAA

[real sierra]

nothing there

[olive sable]

[proven merlin]

Is a MI50 32GB pcie with radeon pro vii patch worth 500€?

[trim valve]

does it violate rule two if the target is nn chatters

[nocturne olive]

Depends on what you're gonna use it for

[olive sable]

rule 2?

[warped narwhal]

No

I don't know what rule 2 is

[mighty bane]

Sounds like a paperweight.

[olive sable]

huh, the text for rule 2 is automodded

[warped narwhal]

LMAO

[proven merlin]

LLM Inference, like always- usually Vulkan, maybe RocM

[warped narwhal]

A transcription of rule 2: don't say gamer words

[olive sable]

fps

[nocturne olive]

32GB should run decently big models if you can get something to support it

[leaden crest]

@real sierra maybe the bees got into your phone

buzz buzz

[proven merlin]

Yea thats the plan, Q5_K_M with large context of like 25-35B models (maybe MoE's)

Is the MI50 difficult to get working?

[nocturne olive]

I just hear AMD cards are hard to work with sometimes

[proven merlin]

Mainly need Vulkan tbh

Its what i use on my RX6700XT

[olive sable]

vulkan poggers

[proven merlin]

Fair

3090's are like 900€ here- as discussed a whiiile ago

Would prefer those buut yea. No way im affording that

[nocturne olive]

If you're running inference on Vulkan and no training it'll prooobably work

[umbral wigeon]

@wraith shuttle Write me vulkan in C++ that will render triangle

[rough bloom]

should be fine if it's just inference

[proven merlin]

I do have a 2060 and a gtx1650 where i might do minimal ML training.. no LLM training tho.

[nocturne olive]

No vocal synthesizers gonna be happening on those either

[olive sable]

3090s stopped being the value king. they used to be 500 now they're 1000.
get a 7900xtx for 650 if you can get away without cuda

[proven merlin]

Yea, not planning to- waaay out of my knowledge zone

[rough bloom]

apparently there's gfx906 forks of llama.cpp and vLLM and stuff
so there's decent community support even though the card itself is EOL
ROCm still builds for it too

[nocturne olive]

Unfortunately I need CUDA for now
I'd jump to Intel if I could afford to

[olive sable]

make your code better

to not need cuda

[nocturne olive]

It should all work with Intel, but I need 24 gigs VRAM, it's a hard prerequisite

[proven merlin]

They start for 700€ here

[sick owl]

Hang on

I think I actually like an agent framework

Oh no

[olive sable]

still cheaper than 3090

[sick owl]

Its over for me

[proven merlin]

Oh its a gfx906, ive heard of it then- rocm is nice to have, for sure.

[umbral wigeon]

Is kimi good?

[nocturne olive]

Karaoke Neuro: lacks energy on a Flavor foley song
Me:

[proven merlin]

True, but also no cuda and a bit weaker compute wise no?

[nocturne olive]

7900XTX is better than 3090 no?

On compute

[proven merlin]

Well there is that 48gb(?) intel card iirc..

[nocturne olive]

Money

[proven merlin]

Yea.. tho ive heard they go for ~1k€

So not much worse than a 3090?

Tho i don't know myself.

Only heard it

[olive sable]

7900xtx is closer to 4090 iirc in terms of amd gpu

[proven merlin]

Oh damn

[olive sable]

4090 still being better ofcourse

7900 XTX provides ~80-85% of 4090 performance for roughly half the price

[sick owl]

Not for AI stuff, software support is still a lot better for Nvidia

[dusty jungle]

i thought it was more comparable to 4080 super?

[sick owl]

Granted its getting better for AMD, you could play the long game

[nocturne olive]

[proven merlin]

Well ig ill look for a 7900xtx then.. though they do tend to die due to some wierd cooler issue or something.. northwestrepair said something about it iirc

[nocturne olive]

Fun fact: 4080 == 4080 super except super has 200$ lower MSRP

[vague sonnet]

that's only if you're overly reliant on shit like pytorch
if you write your own stuff it's the same

[sick owl]

Who wants to bother with that

I know I don't

[rough bloom]

ye, with some caveats
bandwidth is close to 3090 and kernels are worse because nobody cares about ROCm
so for ML, especially low batch size inference, 3090 is probably more useful

[vague sonnet]

ok underachiever

[sick owl]

Underachieving and happy :YEP:

[olive sable]

compute wise sure, but for ai we look at the vram hence the 4090 being mentioned as both have 24gb of vram

[dusty jungle]

idk nvidea stuff i just got an amd card

[proven merlin]

What i don't get is why vulkan is so slept on.. i mean like.. it just works?

[dusty jungle]

[proven merlin]

Litterally everything supports vulkan

[olive sable]

yep

vulkan is cool

[nocturne olive]

Hope you didn't expect to need NeuroSynth GPU acceleration

[rough bloom]

it doesn't, at least for ML

[proven merlin]

I got a nice speedup compiling llama for vulkan on my ryzen 7 5700G server..

[dusty jungle]

whar

[nocturne olive]

NeuroSyntth only supports GPU accel on Nvidia CUDA

On Linux anyway

[vague sonnet]

less features for ML than cuda/rocm

[olive sable]

superbox is just yapping baout neurosynth again

[rough bloom]

Vulkan compute shaders are cool and all but it's still SPIR-V
you just can't access the same things you can from CUDA PTX (or other vendor APIs)

[olive sable]

[vague sonnet]

https://tenor.com/view/laughing-duck-gif-1363576254939741235

[olive sable]

wdym? it has acces to the rt and the tensor cores and everything

[sick owl]

Neurosynth is cool

[proven merlin]

Ah ig

[dusty jungle]

what is neurosynth

[nocturne olive]

This is NeuroSynth

[leaden crest]

neuro machine love mv

[olive sable]

superbox about to get nuked for promo

[vague sonnet]

link the vst

[nocturne olive]

NeuroSynth is not a VST plugin

[vague sonnet]

how primitive

[wispy pike]

anthropic had so much good will, why did they burn it so damn fast

[noble gorge]

Man how the hell did vedal even out both of them in vr, i remember at some points neuro was using like 55 gb of ram alone, rip bros pc 🙏

[dusty jungle]

when will superbox be upgraded to ssrbox

[wispy pike]

like they literally could just raise prices rather than defrauding their fucking customers

[olive sable]

ved has 128gb of ram iirc, he'll be fine

[nocturne olive]

Rewrite

[noble gorge]

Touché

[leaden crest]

no more memory leaky

[nocturne olive]

[vague sonnet]

pretty sure he also distributes them on multiple machines?

[wispy pike]

literally who

[rough bloom]

it basically goes through an additional compilation step
anything usable in Vulkan has to be some standardized extension
for CUDA or ROCm this is not the case

for tensor cores ig it's already been long enough for it to get standardized

[vague sonnet]

isn't that compilation a one time thing tho

[dusty jungle]

[nocturne olive]

[rough bloom]

it doesn't matter how often the compilation happens, that's not the issue
it's that you have less control over the result, it's an extra layer of abstraction

[olive sable]

i don't really know what else there's left that hasn't had a vulkan extension made
maybe some obscure new tech in server gpus?

[vague sonnet]

I mean, if compilation makes it run worse then the compiler is dogshit

[rough bloom]

all compilers are dogshit

[proven merlin]

At the same time tho, it does mean the same code can (if the driver/compilation isnt ass) run decently fast on anything that speaks vk, without a lot of maintainer pain

[opaque sigil]

the compiler has to optimise for the lowest common denominator

[rough bloom]

new little stuff gets introduced all the time

[opaque sigil]

which is why yes, you can in fact beat the compiler quite easily

[proven merlin]

It doesn't tho? Doesn't the vk compiler compile for your gpu, aka the best that it could do?

[vague sonnet]

that is what I would expect

from competent devs

[proven merlin]

(sure, handwritten>compiled

(but like

[opaque sigil]

idk about vulkan tbf

[vague sonnet]

like if you get an extra compilation step on each distributed machine
then for sure it optimises for it specifically
instead of doing some generic build that works on everything
otherwise you can just distribute prebuilt

[proven merlin]

I mean anytime you start a dx12 or vk game it gotta do the whole "compiling shaders" dance.. which is why it took so long for pc to get "downloaded compiled shaders"..

[opaque sigil]

actually what is this even about

i should probably read up a bit

[vague sonnet]

vulkan vs cuda/rocm

and why vulkan isn't as good

[sick owl]

Vulkan is based but worse

[opaque sigil]

oh

[vague sonnet]

ignorant approach
the questions are why is it worse and can it be changed

[proven merlin]

Yeah but not by a long shot.. i get on average 90% of the speed of cuda with vulkan.. its insane to me that initial implementations arent done in vk for support, and then optimized runtimes in cuda and such

[opaque sigil]

no compiler is perfect

[rough bloom]

very convenient for that
the stack is so much simpler than the stupid vendor APIs

[rough bloom]

SPIR-V and Mesa my beloved

[vague sonnet]

what's the test setup?

[olive sable]

vulkan should compile the spirv for the hardware specific needs tho no?
that's why you don't just receive the spirv but have to compile it yourself

fuck

[proven merlin]

Just a 4B llm, and running the same prompt through it like 10 times. Nothing scientific

[opaque sigil]

should've could've would've

[olive sable]

im mixing up terms

spirv

not riscv

[vague sonnet]

yeah that's what I'm saying
there is no way compilation makes it slower
since the whole point of local compilation is to make it faster

[opaque sigil]

nvcc should also be able to automatiaclly use tcgen05 instructions when i tell it to target sm_100

yet here we are

[rough bloom]

overtoasting

[warped narwhal]

Oh, @sick owl your a LLM guy, if I want a local model for answering general questions, so not for generating code or anything, moreso just a better Google, what models would you recommend? It would need to be able to run on my Rx 570, so vulkan and 8gb of ram. But accessing the web for queries would be fine

[proven merlin]

Well the thing is, cuda and rocm get optimized kernels per gpu afaik

[opaque sigil]

https://wiki.c2.com/?SufficientlySmartCompiler

[vague sonnet]

yeah

they get extra non-generic instructions that boost the speed

[sick owl]

Gemma 4 26BA4B with offloading using --n-cpu-moe

Grab a q8 quant, use like 32k context and increment the offloaded moe layers until it fits then keep incrementing until you stop getting speed improvements, drop down to 24k context for a compromise on context length and speed

[proven merlin]

So the whole compiled<handwtitten applies. It has nothing to do with instructions. The instructions for the compiled ones are done at the end consumer

[rough bloom]

and I'd expect Mesa to do this much worse than nvcc
not to mention anything handwritten where someone actually looked at the PTX

[vague sonnet]

"handwritten" still gets compiled

[proven merlin]

The cores that the companies themselves provide? Aka as example gfx906?

[rough bloom]

not all compilation steps are made equal

[vague sonnet]

are cuda/rocm distributed not compiled?
I've been only looking into custom kernels so idk what the standard is

[vague sonnet]

bro cut the yap

[proven merlin]

I mean no, he has a point

There are indeed different types of compilation

[opaque sigil]

nvidia at least pre-compiles code for every architecture for arch-specific optimisations, that's the standard approach

[proven merlin]

-march=native will run faster than -march=i386

[opaque sigil]

which is why the cublas .so is 500MB

[warped narwhal]

Sounds good, roughly how much ram would I need to allocate to the VM that will be running this?

[opaque sigil]

i'm sorry but the compiler is not good enough to fully rely on when your goal is maximising performance (which it should be, why else are you using a gpu in the first place)

[sick owl]

Does it have to run in a VM?

Probs easier and faster just to run it in your base OS

[warped narwhal]

My base OS is proxmox so probably not

It will be running on my homelab

[proven merlin]

Oh! I ran some models in a LXC, that was fun.

[sick owl]

I take it back lule

Allocate like 28gb ram + full vram access

[vague sonnet]

if the compiler doesn't maximise performance then it's not a good compiler
because it won't run uncompiled and you're making an assumption that someone is distributing a well-compiled version

[sick owl]

Make sure you run it with --mlock and --no-mmap if you run it on a windows VM because windows mem management is fucky

[opaque sigil]

a slow program is better than a potentially broken program

[vague sonnet]

?

"never write code because you might make a mistake"

is that your point

[rough bloom]

if the compiler doesn't maximise performance then it's not a good compiler
it's trying

[rare bramble]

compilers sadly aren't the best when it comes to performance, if they were, declarative languages would be the fastest

[sick owl]

Oh also I recommend Bartowski's q8 gguf for a nice stable one

[proven merlin]

I started a war

[opaque sigil]

no my point is that the compiler is trying its best but it will not apply any optimisations that have a non-zero chance of breaking anything

[vague sonnet]

what's with all those compiler haters
you guys only write in python or some shit

[warped narwhal]

It'll be on Ubuntu 26.04, and could I possibly get away with only 16gb of vram? The server with the GPU only has 32gb and already has 16nfor truenas and 16 for k3s. I can move 32gb from my main pc to bring it up to 64, but if I can get away with 48 and 16 for the LLM VM that would be great

[proven merlin]

Where is my achivment

[warped narwhal]

https://tenor.com/view/your-did-it-fail-oof-loser-you-didn't-do-it-gif-5120386489105861099

[proven merlin]

Yay

[proven merlin]

Zram, hehe. (I take no responsibilities for downloading more ram)

[vague sonnet]

yes it is faster to write in random memory instead of finding free memory
but a good compiler should in fact make the fastest functional build

[kind nimbus]

https://downloadmoreram.com

[warped narwhal]

Time to add a 64gb nvme share for some sweet sweet pagefile

I wonder how well swap would work over 10gbe iscsi

[proven merlin]

Nice.

[opaque sigil]

this imaginary good compiler you're talking about doesn't exist, that's the whole problem

[rough bloom]

ye, it would be nice if compilers did that for any functionally identical code you give it
but they very much do not do that

[proven merlin]

Bad. Very bad

[vague sonnet]

sad to hear there aren't any competent programmers left

[proven merlin]

Zram writeback

[rough bloom]

you should

write compiler

[proven merlin]

We cant top TCC.

[kind nimbus]

Just mount your swap to Google drive

[vague sonnet]

unfortunately I trained myself in c++ not assembly
but I have been passively learning to decode hex in my head
so might end up manually editing the compiler outputs

[proven merlin]

Streaming 3rd pass compilation step

Crazy

[warped narwhal]

Why not use the -S flag and do the editing in assembly before continuing?

[vague sonnet]

because I would have to learn assembly

at this point might just edit the raw bytes

assembly is probably suboptimal and slow anyway

[warped narwhal]

I feel like I'm missing something here

[proven merlin]

Yknow whats funny.. asm isnt thaaat different from C..

[rough bloom]

I don't think you are

[warped narwhal]

Assembly is 1:1 with bytecode usually

[olive sable]

no, editing the direct bytes is way harder

[vague sonnet]

we're just schizoing
I can't have a serious conv with compiler haters

[proven merlin]

Youre still just doing if's and function calls (well.. goto's).. just in a different syntax..

[warped narwhal]

I'll contact that guy who turned discord into a NAS

I can see the swap as it is used

[proven merlin]

Nono, contact the dude who made the ping-harder-drive

Aka using high ping as a disk drive.

[rough bloom]

harder drive mentioned

good video

very memorable

[proven merlin]

Yes.

[rough bloom]

https://www.youtube.com/watch?v=JcJSW7Rprio

[proven merlin]

Suckerpinch good

[fiery anchor]

"Your scientists were so preoccupied with whether or not they could that they didn't stop to think if they should."

[proven merlin]

Which. Reminds me, i should resub.

[kind nimbus]

I only know that you can turn discord into a C2 lol

[proven merlin]

Nono, we had plenty of time. And we said fuck it we ball

[fiery anchor]

And that's why we can't have nice things

[proven merlin]

Perhaps so. But we get cursed things :D

Horrors beyond your comprehension

Thats good, no?

[fiery anchor]

Mhm. If everything is cursed, is anything cursed?

[proven merlin]

See? You get it. We have nice things after all

I still don't understand why vulkan isnt the first api to get support in anything compute.. buuut oh well

[fiery anchor]

Well, there might be several reasons for that.

[vague sonnet]

both nvidia and amd have monetary incentive to support their exclusive software that competition's customers can't use

[proven merlin]

Well ig

But something like idk, comfyui, llama?

Blender?

[fiery anchor]

incentive. working for the greater good doesn't pay well.

[olive sable]

blender has optix and HIP support

[vague sonnet]

mlemder

[olive sable]

but since its a non-profirt thingy the progression of implementin this shit is slow

[vague sonnet]

blender isn't ML so it doesn't matter as much, vulkan works just fine for graphics

[olive sable]

hence why optix is slightly faster still

[rough bloom]

did you see the news

Blender is now

sponsored by Anthropic

[olive sable]

[vague sonnet]

is that a good thing

hope it's just money

[olive sable]

at the cost of claude being in it now

[vague sonnet]

oh nah

[olive sable]

which is somewhat fine probably

[vague sonnet]

never updating bleder again

[fiery anchor]

Oh dear.

[rough bloom]

integrated at least

unless I misread the announcement it's not literally in Blender

their desktop app thing just connects to it

[sage crag]

if there isnt already a toggle, its open source and can just be removed downstream

[vague sonnet]

maybe it's fine if it has that setting like vscode
which is just "disable all AI features"

[fiery anchor]

"What do you mean the part doesn't fit? The AI did WHAT?"

[sage crag]

[rough bloom]

cni

[olive sable]

hi konii

immortal venerable wrr

venerable

Venerable is an adjective describing someone or something respected, honored, or revered due to great age, wisdom, high position, or historical importance

[vague sonnet]

yeah quick google and seems like it's not a blender-side thing

[olive sable]

you can also just remove it yourself if you feel like it

i doubt it will be an issue tho

[sage crag]

chsi

[vague sonnet]

I wouldn't be ready for blender devs to backstab me with unwanted bloat

[fickle rain]

I mean you can turn a lot of things into C2

https://mattwie.se/hinge-command-control-c2

[sick owl]

40% higher PCB costs

Make it stop

Why must we suffer in perpetuity

[olive sable]

I get what you and them are going for

But i feel like its kinda insensitive to complain about pcb prices in this case

[kind nimbus]

true lol

[sick owl]

Oh 16gb vram is plenty for 26ba4b

Just throw in an extra maybe 20gb of sysram alongside it

The reason I put sysram so high is because I thought you had 8gb vram ^

[silent cloak]

U can use storage as ram but can u use ram as storage?

[kind nimbus]

Here some new cve to play around with

https://copy.fail

[silent cloak]

Free cve giveaways

[kind nimbus]

[silent cloak]

Is there a reliable graph on cves over the past few years?

Feels like its becoming more common

[kind nimbus]

https://nvd.nist.gov/vuln/search#/nvd/home?resultType=statistics

[silent cloak]

Hmm as I thought

Everything shot up post AI lmao

[kind nimbus]

Kinda yea

[silent cloak]

The initial global increase i think is just from tech changing during that era and more people being online but after that things continue to gradually increase far more than they used to by the looks of it

Everything only goes up now without going down (2026 naturally isn't over yet)

[kind nimbus]

Might also be a consequence of vibe coding lmao

[silent cloak]

Well yeah thats what I mean

[kind nimbus]

Ah initially i thought you meant because of stuff like mythos and co

[silent cloak]

Na

More people are just trying the fake it till u make it strategy I guess

[vague sonnet]

cheaply produced cash grabs have higher income potential than high quality software

they won't earn you more but they take way less effort

[silent cloak]

Sadly true

Besides for anything targeting power users

[vague sonnet]

yeah

more talking about average customer/user

[silent cloak]

Yeah u really need to research stuff these days to ensure ur not getting screwed

Hardware included

[vague sonnet]

elaborate

since I'm about to buy a new pc

[silent cloak]

The nvidia 5000 series

[fast pagoda]

honestly probably faster testing

[silent cloak]

And things like home appliances

[true hemlock]

[fast pagoda]

the cve project has said they are fucking drowning in reports

b ad and good

probably mostly bad

[vague sonnet]

ah
currently I'm mainly looking at 7900xtx as the gpu of choice

[fast pagoda]

[true hemlock]

based

[vague sonnet]

hi toast

[true hemlock]

what's your total budget

[silent cloak]

Honestly the benefit to adopting a memory safe language is that the vibe coders will have to try harder to fuck up

[fast pagoda]

python adn js are mem safe

[true hemlock]

lmao

its why a lot of js projects are vibe coded

[fast pagoda]

[silent cloak]

Honestly tho the state of the field just makes me sad

Im not a web dev so I dont run into it as often as others perhaps

[true hemlock]

ai bros vibing with vibe coding
flex that they're building projects
looks inside
js

[vague sonnet]

flexible
it's more about balancing specs with costs
currently thinking 7900xtx + 7950x + 64gb ddr5
and other stuff just to match with those

[fast pagoda]

i see js so much in like

desktop

now

i mean

[true hemlock]

oh damn ballin

[fast pagoda]

claude code is what fuckin react

[true hemlock]

what's your usecase

[silent cloak]

The og dev having a stroke

Im pretty sure he hates the community or the language if I recall now

[nocturne olive]

Good luck getting the RAM if you're not a trillionaire

[fast pagoda]

64g isnt that much

[vague sonnet]

ML and everything gamedev
so need a big gpu for training and a big cpu for compiling

[fast pagoda]

i mean

it's a lot

compared to before

but

[true hemlock]

okay you've done your research well

[fast pagoda]

yeah no notes tbh

[true hemlock]

best of luck

[vague sonnet]

around $400 for used

[fast pagoda]

[silent cloak]

[true hemlock]

wtf how

where

why

when

[nocturne olive]

Good luck making AMD work for ML

[fast pagoda]

he's thought it through

[true hemlock]

stfu

you skill issue guy

[kind nimbus]

Just get DDR3 RAM

[nocturne olive]

It's not me, it was someone else here

[vague sonnet]

good to hear
I'm a software guy so I went into this hardware research with no knowledge

[nocturne olive]

I don't have any AMD GPUs to experiment with

[silent cloak]

I dont either

[fast pagoda]

I managed to find 64gb 6000 in about 1 second

at 500

400 should be out there

[vague sonnet]

I'm gonna do custom gpu code so rocm is about equal to cuda

[true hemlock]

the problem with nvidia gpu is that you got somewhat worse options

[patent shard]

uhhhh
that's new at $270 for 32gb
I wouldn't buy that
both are bad deals, though

[nocturne olive]

Doing all from-scratch

[fast pagoda]

like i said i literally searched

ddr5 on ebay

[true hemlock]

buying new, if you wanted more than 16gb vram you gotta overspend for a 5090. else you're stuck with 16gb

[silent cloak]

Use vulkan

[fast pagoda]

and sent the first 6000 kit

[vague sonnet]

no one does SNNs man

I can't really take advantage of existing libraries

[nocturne olive]

[silent cloak]

Its new tech

There is no standard

[true hemlock]

https://tenor.com/view/cat-float-chill-water-gif-14413129433134838716

[vague sonnet]

I can test both rocm and vulkan to see which works better

[silent cloak]

One man's choice of neuron type or neural plasticity algorithm is different for another researcher

[nocturne olive]

[fast pagoda]

vulkan is often faster in inference with llms right now due to rocm lule

[true hemlock]

there's no standardized libraries for snn
vanilla snn suck, implementation must have lots of varieties

[vague sonnet]

that totally applies to my custom implemented SNNs

[silent cloak]

Its how transformer models were before

Until peope find the perfect combo

Its gonna be a billion variations

[true hemlock]

transformer is far easier tbh

[silent cloak]

Not to mention some of the other tools being ran on the side now for other sorts of emulations

[true hemlock]

synaptic plasticity is hella unpredictable.

[fast pagoda]

just a datpoint to consider was all (idfk what you're makin)

[vague sonnet]

yeah
it's gotta be a mix of methods
do some regular weights for some input and output layers
and just do the actual recurring portion spiking
and there needs to be some crazy memory management to do the maths on those fast

[true hemlock]

and difficult to implement so most experimentations just not do that

there's some snn implementation for computer vision that are pretty competitive

[vague sonnet]

some convolutional layers are worth looking into
because they can help tie together inputs that belong to the same time point
or same input type that belongs to different time points

pack all that info into less nodes
then feed it into a RNN portion with spiking nodes

which will probably have to run faster than the input sampling

[trim valve]

personally i prefer coinflips to all this complicated math

just randomly pick outputs

[fast pagoda]

that's what i do

in my brain

[vague sonnet]

generative models do tend to have randomness

[true hemlock]

generative models suck at making decision

[vague sonnet]

it's like they generate data instead of logic

[true hemlock]

especially timing aware decisions

[vague sonnet]

well yeah they don't have a concept of time

[fast pagoda]

theyre great at making decisions

they make so many decisions

badly

and randomly

but they do be acting with decisiveness

[silent cloak]

Whats ur favorite flavor of STDP

[safe path]

generative decision-making

[safe path]

with enough coinflips, you can emulate all math...

[silent cloak]

Expensive RNG

(Beats the quantum RNG machines)

[fast pagoda]

our brains

[vague sonnet]

one of the main challenges of SNNs is picking the right learning algo
the simple thing is increasing activation threshold at high use
but the less obvious is when to decrease it and how to determine how to adjust the weights

[safe path]

nah im just gonna convert all of my net worth into pennies and flip them

gotta ensure they are all different coins to ensure independent coin flipping

[silent cloak]

Could be possibly predicted

Depending on the type of coin

[safe path]

they rigged my coins

[trim valve]

converting your dollar into 3 pence

[safe path]

nah im worth more than a dollar, i own ram

[vague sonnet]

if I can somehow extract a fitness function (or label it manually trolldespair)
then I should be able to do some more conventional learning for at least the pretrain
and make the recurrent section output its own value that gets applied to live training, probably at a slower pace tho

[trim valve]

also side tangent, I was meant to do a reflective report for a group project but ended up just roasting my group mates and ranting about mental health

kinda incredible to read ngl

[true hemlock]

512mb ddr2

[silent cloak]

A modern gamers rolex

[true hemlock]

worth a house rn

[vague sonnet]

tie RAM sticks into a necklace

might need to hire bodyguards

[silent cloak]

Saw someone do that

I got a bag of random ram

[true hemlock]

does this count

[vague sonnet]

the "hey mister" girl

[silent cloak]

like remilia

[silent cloak]

HER NAME I REMILIA

https://tenor.com/view/meme-bob-esponja-gif-2904459546447199840

[vague sonnet]

I know man, I used to be active in touhou communities

[silent cloak]

i never played the games but i own a bunch of very expensive fumos

[vague sonnet]

I never had enough extra money to get fumos

[silent cloak]

got one of the big ones for $300 too

[vague sonnet]

yoooo

I technically own a fumo

but it's the monkey druid from bloons td 6

[silent cloak]

kek

[vague sonnet]

too lazy to take a photo but this one

we did a group order so it was only like $30 per

incl shipping

[vague sonnet]

I grinded the games for some time but I was in it for the lore
which we didn't really get that much of...

and ofc music is great but got tired of listening to the same melodies

[sick owl]

Turns out vibevoice is an incredibly annoying model to work with when you only have 24GB of VRAM

[fast pagoda]

realtime

[fast pagoda]

[mighty thorn]

That’s today?

What the hell did they do

[fast pagoda]

I have no fucking idea lol

I looked at my notifications

I was kinda planning on maybe selling some intc this morning

Glad I did not

$4+ of it is after hours suddenly

Maybe tsmc fell into the sea

[fast pagoda]

Apparently

Foundry got the contract for the next m CPUs from apple

And are about to probably get tpuv8

Damn son

They're not even priced in yet probably then

https://www.reddit.com/r/intel/s/Yh6SSmsj2a

And this apparently

[tropic spindle]

The two x86 duopoly holders making a partnership definitely doesn't seem anticompetitive in any way at all for sure

[fast pagoda]

they have partnered for stuff like this before

they have been doing anticompetitive stuff along with the rest of the market but i dont think this is the one lol

[tropic spindle]

I know

I just don't feel like it's great when they are the only two players on the x86 market

[fast pagoda]

intel 64 bit cpu being amd64:

[tropic spindle]

Unless you count VIA which I really don't

[fast pagoda]

who was the 3rd since like 2001

if anything china is closer than theyve ever been (still iinfinity far away but)

last time there was a real 3rd option was cyrix in the 90s

[tropic spindle]

The only good thing about the duopoly is that it's not a fragmented piece of garbage like ARM

[kind nimbus]

IA-64: the lost child

[fast pagoda]

calling them a duopoly to me is akin to calling amd a monopoly on cdna & rdna and nvidia a monopoly on blackwell, sure they are but it doesnt functionally constitute a monopoly because alternatives exist even if they use a different architecture

[young plover]

This doesn't really seem much different from any other x86 extension they've cooperated on.

[tropic spindle]

Intel: let's make a whole new incompatible architecture when 90% of the market relies on the previous one! That's the best plan ever!

[fast pagoda]

that's cpu

[sick owl]

I hate that in the big '26 Whisper Large V3 is still arguably the best Japanese transcription model

[fast pagoda]

dec alpha be like

[sick owl]

Qwen 3 ASR is the only thing that trades blows and that requires faffing around with chunking and a separate conformer model for timestamps

[tropic spindle]

As I said, not being fragmented like ARM is a good thing, but there are bad parts too

[fast pagoda]

arm is a competitor to x86
being an ecosystem of a particular isa design does not make a monopoly when there are alternatives that perform the same function just differently

otherwise arm is a monopoly

despite its fragmentation

[tropic spindle]

ARM isn't a manufacturer

[fast pagoda]

irrelevant

you have to license from arm to make arm

[tropic spindle]

If there was only 1 ARM provider, it would be a monopoly

[fast pagoda]

there is

it's arm

[patent shard]

[young plover]

The bigger problem was that VLIW is bad. They were developing out-of-order x86 chips at the same time, and it turned out that letting the CPU dynamically schedule instructions internally gives better performance than trying to make compilers generate optimally scheduled instructions.

[tropic spindle]

Also naming it ACE is a choice of all time

[tropic spindle]

Yes, but compatibility is what affects cost effectiveness more

And it's also the reason x86 isn't going away any time soon

[fast pagoda]

so why is coordinating on it to have them be compatible between the 2 viable x86 makers a bad thing then...?

[tropic spindle]

Even if you can have more efficiency with a different architecture, it doesn't help when most of the market is in another

[mighty thorn]

duopoly

I like words

[fast pagoda]

which has happened with every extension like avx512, avx10 which this extends, all the way back

[tropic spindle]

I mostly don't like framing it as a "partnership"

[fast pagoda]

ok well i'll ask them to rename the article "big bad stinky ebil rude groups of guys are angrily yelling at each other about avx10 extension until they come out with something resembling a standard"

[young plover]

Did AMX ever make it into a consumer processor? 🤔

[fast pagoda]

intel?

they did it at least once

xeon tho

[young plover]

Yeah enterprise. I'm not aware of any consumer processors with it though.

[fast pagoda]

looks like 4/5xeon scalable//xeon 6

and that's about it

its on a tile so maybe in their upcoming tiled designs but

they JUST brought tiles to consumer with 14th gen

and that's much less than xeon

or ponte vecchio my beloved

[olive sable]

i see so much hate towards ARM here when i feel like its superior to X86

in this economy the better efficiency is welcome to me

[tropic spindle]

lolno

[olive sable]

i don't want x86 in anything portable tbh. only reason i have a laptop with x86 is compatibility

[tropic spindle]

Talk to me when you don't have to individually account for 63382737338 different feature sets in each individual CPU

[olive sable]

its not like x86 doesn't have different featuresets per cpu

“not all x86 have avx512, therefore unusable”

sounds dumb

i dont mind x86 for desktop and server

but when it comes to battery life its ass