programming

[umbral wigeon]

How long can AI era go

[void estuary]

forever

next chapter is AGI hopefully

[umbral wigeon]

How about hardware prices and resources

Also people worldwide using Ai

Including me

Data center is where Ai lives

[void estuary]

eventually supply will meet demand

i need my m5 ultra mac studio with 512gb ram already, Airi (my companion) goes OOM several times a day with only 64GB

[warped narwhal]

a big question is this: if AGI actually gets created, what happens to the humans?

[umbral wigeon]

Who design and making hardware

[warped narwhal]

AGI will

[void estuary]

https://tenor.com/view/robot-doctor-who-cybermen-gif-5034673

[warped narwhal]

Robots have already taken over most of the mass manual labour, so humans only had the intelligent things left anyway

with AGI taking away that, then what work is left over for us?

[kind nimbus]

https://tenor.com/view/neurosama-evil-neuro-embed-failure-gif-15456327638408323033

[warped narwhal]

https://tenor.com/view/neurosama-neuro-sama-vedal-vedal987-vtuber-gif-15278412527939419909

[umbral wigeon]

What will you do in future if some robot walk inside your house

[opaque wharf]

Actually the opposite. Processing raw material like mining up until a certain point is still hard to automate

[void estuary]

well, one of my managers wants an "AI CISO"......that wont have any issues......

[opaque wharf]

But factory work is highly automatable to the point the industry have terms lights out manufacturing

[void estuary]

automation != AI

[opaque wharf]

https://en.wikipedia.org/wiki/Lights_out_(manufacturing)

[opaque wharf]

See the original message I replied to

[umbral wigeon]

Are vedal vibe coding

[void estuary]

i hope not, vibe coders should be choked

[umbral wigeon]

Vibe coding with no knowledge is like gambling

From my expenerices

[void estuary]

context engineering, or open spec dev, that's all that should be allowed in the space

[vague sonnet]

not at this rate

LLMs have hit the limit

there won't be real progress if they keep using inefficient tech

[umbral wigeon]

It takes me 1 hour to make flappy bird clone (frontend) with some prompt engineering, using roughly 100+ lined of PROMPT.md
https://github.com/weenachuangkud/Flappybird_react

I let Ai pull request me and I preview before merging until got everything as expected in PROMPT.md

[opaque sigil]

I feel like they should be able to oneshot flappy bird

not that it matters

[umbral wigeon]

I wanted modular, OOP, easy to config(editing constants), yeah it gives me as I expected

Also another project

https://github.com/weenachuangkud/videolinks-to-ascii-imGui

[void estuary]

my prompts are average 1k lines, AFTER i have a planning, tasks, and rules, that are MINIMUM 1k lines

first prompt context window is like 40k tokens though, fucking brutal

[umbral wigeon]

(I got pull request by my github bot account lmao)

Gonna close issues, merge pr

[vague sonnet]

[umbral wigeon]

I don't even work tbh

My github bot account running on cloud and Ai does for me

[void estuary]

WARP+Manus (phasing out with perplexity) is like....70% of my job now, but mgmt will never know

watch YT educational shit about AI for half the work day and get paid for it, W

[umbral wigeon]

Jk I still need to work

[maiden geyser]

cool it with the anti-j*bless remarks

[opaque wharf]

U.S. of A.
https://itsfoss.com/news/os-level-age-verification-across-us/

Meanwhile France move to NixOS

[maiden geyser]

insane how obsessed everything is getting

[opaque wharf]

I can at least understand one state doing something idiotic. But the federal level

[vague sonnet]

ah yes, more anti-privacy

internet becoming more and more unusable
we should destroy the world, it's for the better

[true hemlock]

its what happens when boomers are put in charge

[nocturne olive]

Ban USA from all OS', problem solved

[opaque wharf]

What is going on in this world man. Neuro needs to take over already

[olive sable]

[vague sonnet]

not even
just a characteristic of politics, more control is better for them because they get to enforce more which eventually leads to more personal benefit (often at the cost of the general population)

[kind nimbus]

Gotta build arpanet 2.0

Call it neuronet

[true hemlock]

im so glad im still on win 10

not getting the bs update

if it were to happen

[vague sonnet]

I was actually thinking what it would take to make "internet 2"

currently beyond my capabilities

[kind nimbus]

The physical connections are already there

All you need is new protocols and software

[vague sonnet]

when I get a bigger GPU I'm gonna train an AGI (not llm-based) and it might help me

[sage crag]

when I get a bigger GPU I'm gonna train an AGI (not llm-based) and it might help me

[vague sonnet]

not a fan of linux?
I'm thinking about switching

[true hemlock]

linux only for my servers

[maiden geyser]

but the connections are the controlled thing

[sage crag]

oh hi 🦆

[true hemlock]

i still use stuff that is windows exclusive

[sage crag]

i didnt notice it was you

[opaque wharf]

The physical connection is the weak point

[vague sonnet]

is it worth the pain of using windows tho?

[true hemlock]

win 10

i aint touching win 11

[vague sonnet]

10 is shit too

[true hemlock]

its ehh

its fine

[sage crag]

for an overclocker, yes

[vague sonnet]

I would love to go back to 7 but it doesn't support anything nowadays

[true hemlock]

this tbh

overclocking ironically suck at linux

[sage crag]

win10 ltsc is basically completely fine

[true hemlock]

^

also i removed all the bloat on mine

its not worth the pain to switch when i have so much things already built around win 10 tbh

[sage crag]

its a lot better on amd than intel or nvidia, especially with recent progress, but the amount of work you go through to overclock on linux isnt really worth it when its so much easier on windows

[true hemlock]

• dont have the time

[true hemlock]

you know what's also pretty ironic

extreme oc is somewhat more stable on windows.

[sage crag]

the amount of control you get on windows for how easy it is, is something you dont get on linux

[true hemlock]

well, something like tiny 10

oc in linux is pretty much unsupported yeah

idk why

y'all are some open source gang

but somehow haven't had a proper oc interface/tool yet

are linux people just often not into hardware tweakings

[maiden geyser]

erm, if you need it, why don't you write it yourself

open source or something

[sage crag]

probably depends on what youre running, but its typical to have higher usage for the same fps on linux when on even footing so it doesnt surprise me

[true hemlock]

do i look like i have the damn time

[opaque wharf]

I like this as Internet 2
https://reticulum.network/

[opaque wharf]

To be fair, that's vendor issue most of the times not giving out the proper driver for it

[true hemlock]

i occasionally enjoy nixos a bit but not as far as just using it all the time
debian for my hpc server
ubuntu for my api/endpoint/script server

[sage crag]

they are, but a lot of overclocking stuff is hard to support in open source drivers without the companies responsible for the hardware to write their vendor specific clock management stuff

[true hemlock]

windows 10 for normal usage

[maiden geyser]

data slayer had an interesting video about it

[rough bloom]

combination of proprietary tools + nobody bothering to do both the necessary RE and build a nice interface for it

[sage crag]

essentially

its a hassle

and pointless without vendor benevolence

[opaque wharf]

Care to give tl;dw?

[sage crag]

wrrr

[true hemlock]

that shit sucks

[sage crag]

ye

[true hemlock]

its why im stuck with windoes

windoes

[opaque wharf]

It is what it is with closed source hardware

[sage crag]

riscv cluster as gpu

[true hemlock]

actually imma call it windoes from now on

fuck no

[sage crag]

windoesnt

[vague sonnet]

I guess
I would be switching for a new pc, don't overclock and wouldn't lose any important apps
I like the convenience of windows but it's just getting increasingly worse

[opaque wharf]

You may not believe this
https://www.youtube.com/watch?v=lh93FayWHqw

[sage crag]

ye i saw it

terrible gpu

[true hemlock]

its hella inefficient for pure simd

[sick owl]

Wait what

[opaque wharf]

But it is open source

[true hemlock]

but

its cool

[opaque wharf]

At least the hardware layout is

[true hemlock]

and THATS the point

its cool.

not practical but damn fuckin cool

[sage crag]

the cores in actual gpus are way way way more numerous, despite being weaker individually

[true hemlock]

unless if your project is for some reason clustering rpi

that is somewhat ew

[true hemlock]

the cores in gpu are way wider vector units

[sage crag]

clustering actual rv cores is much better than clustering full rpis

[tender river]

im reminded of the breadboard cpu

[true hemlock]

because they don't need complex pipeline and other units unlike cpu

[opaque wharf]

Breaboard GPU

[true hemlock]

so each gpu cores just have like 64 to 128 shading units

[opaque wharf]

At least VGA capable graphic unit

[sage crag]

vram is specialised too compared to traditional memory

[true hemlock]

clustering rpi is so impractical and pointless

well only usage of it that i could think of is education purposes

[sage crag]

even if this thing was decent at compute, its memory bandwidth would be

[opaque wharf]

How dense can we pack risc-v core into 5cm x 5cm die

[true hemlock]

modern vram uses pam encoding

[sage crag]

educational purposes of £120 per rpi

[true hemlock]

technically analog

[sage crag]

education in saving money

by losing a lot of money

[true hemlock]

even old epyc is way more power efficient than clusters of rpi

imma ngl

rpi ain't efficient

they're just low power

[sage crag]

small things are always inefficient

[rough bloom]

the tenstorrent in question

[sage crag]

for example

if i want to destroy you, i could throw a lot of small rocks

[trim valve]

false

[sage crag]

individually, the rocks have low energy cost

but the overall process is inefficient

but, if i threw just one large rock

[true hemlock]

i mean

[sage crag]

i'd destroy you with a lot less energy

[rough bloom]

(kinda, IIRC it's not only small RISC-V, it's small RISC-V with big SIMD or something)

[true hemlock]

it sorta depends

what if the rock is sharp

[sage crag]

if the rock is sharp then you have progressed to the stone age

congrats

[true hemlock]

mods

i want them CLUSTERED

[sage crag]

modern x86 cpus have hundreds of physical registers

[true hemlock]

but yeah if imma ngl rpi are pretty inefficient

dont use em for compute its shit

[sage crag]

do gpus have 1:1 mapping between physical and virtual registers or do they also use a tonne more

[opaque wharf]

SIMD RISC-V

[true hemlock]

sorta makes the expensive "high end" rpi pointless ngl

[sage crag]

x86 itself has like, 16 gpr and a bunch of extra simd registers

[tender river]

gpus don't even have a well defined isa

i think

the drivers translate shaders from a common language to whatever the gpu understands

[sage crag]

om

so just microcode

[true hemlock]

lmao yeah

[sage crag]

and a translation layer

[true hemlock]

anyway

[opaque wharf]

Alright, I've watched it. Nice conclusion

[tender river]

https://rocm.blogs.amd.com/software-tools-optimization/amdgcn-isa/README.html

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

[sage crag]

[vague sonnet]

however the energy cost of lifting a larger rock might introduce inefficiency
what if you threw a smaller rock with more force? which is actually the working principle of guns

[sage crag]

ok, but

if i wanted to now destroy a spaceship

i'd rather throw a big rock with more force

railguns are good!

[tender river]

spaceship can be destroyed with a pebble if you aim well

[vague sonnet]

truth, it happens in star wars

bro shoots in the right hole and whole death star explodes

[sage crag]

bad bad bad

[tender river]

bump into pebble

[sage crag]

in halo all the space ships have shields

[tender river]

orbital velocity pebble

[vague sonnet]

you need to get closer
trying to shoot through a shield is inefficient

[olive sable]

[sick owl]

That's new

[vague sonnet]

the shield needs to have holes for the ship to move
otherwise the thruster's gas would get stuck in a bubble and prevent further movement

[trim valve]

sam you know too much about hdmi on sbcs right

I have a question

on an sbc, like the one you used, is it likely possible to control when you write to the display or is that kinda out of your control

[opaque wharf]

That would require energy

As pebble would be limited by their terminal velocity

[trim valve]

basically, can you control when a frame "starts" beyond just "1/60th of a second has passed, new frame time"

[tender river]

space

spaceship

[opaque wharf]

https://tenor.com/view/space-tim-curry-gif-7935778

[tender river]

i agree that

railguns are good

[opaque wharf]

I think railguns are not good

[tender river]

huge asteroids spawn beyond aquilo

how would you deal with them

without railguns

[opaque wharf]

Pure mass

Just throw something heavier

[tender river]

https://cdn.discordapp.com/emojis/1429520293721800725.webp?size=48&animated=true

throw with

railgun

[opaque wharf]

Wait, is railguns the one with solenoid?

[sage crag]

fire a railgun from a railgun

[opaque wharf]

Ohhh, railguns are not the one with solenoid

Yeah, railguns are good

That's a coilgun apparently

[vague sonnet]

fire a railgun firing railgun from a railgun

[sage crag]

ohisee

[opaque wharf]

But to be fair, coilgun is easier to handle

[rigid snow]

[tender river]

coilgun

cheap in

barotrauma

[vague sonnet]

the name doesn't go as hard

[opaque wharf]

Because they use coil not rails

Railgun shooting trains

Space travel

[tender river]

chaingun better

but

ammo 50% more expensive

[sick owl]

Thoroughly underwhelming

[opaque wharf]

I think it's more than 50% irl

Coilgun can use anything ferromagnetic

Railguns can use anything conductive. And with high enough voltage, anything is conductive

[tender river]

[opaque wharf]

Chaingun can only use special ammo. Special gun

[tender river]

levitating frog

https://tenor.com/view/frog-levitate-levitation-float-floating-gif-25339418

magnet

[opaque wharf]

That magnetic field is insane. I never thought that a safety line for metal object would be drawn

And yet it exists

[sick owl]

I assume the regressions are from safety measures since this is probably a Mythos distill

[kind nimbus]

https://gitlab.com/IsolatedOctopi/greenboost

[opaque wharf]

How would it feel to levitate in a magnetic field while still feeling gravity

[sick owl]

Note the cybergym score

[opaque wharf]

Must be a weird feeling

Or would it cancel out

[stray dragon]

depends how the force acting on the body from the magnetic field is concetrated

[opaque wharf]

That would limit the already bandwidth limited task

[stray dragon]

if it's distributed evenly throughout the body, wold feel like zero g

[vague sonnet]

gotta love vague tests that don't convert to real world applications

[opaque wharf]

Have you ever feel more than 1G for a prolonged time?

[stray dragon]

uhh no

[opaque wharf]

It's fun

[vague sonnet]

bro was on mars

we don't have artificial gravity yet, right?

[maiden geyser]

if we don't count centrifuges, no

[opaque wharf]

You can just go and break laws about speeding

[vapid pulsar]

Not yet at least

[opaque wharf]

Or do it safely on a circuit

[vapid pulsar]

100 times gravity isn't safe anw

[vague sonnet]

if goku can do it then why can't I?

he survived it because he was eating beans, just gotta be british enough

[vapid pulsar]

ah yes the beans

[kind nimbus]

Yes and here is the PoC btw:
https://github.com/Nightmare-Eclipse/RedSun

[opaque wharf]

Yep, I've read the repo and it is hilarious lmao

[kind nimbus]

it is in fact extremely hilarious

[opaque wharf]

Why in gods name would an antivirus makes sure a virus is THERE

[kind nimbus]

Idk but AV systems have so many vulnerabilities lmao

Having AV is almost more dangerous than just using brain.exe these days

[maiden geyser]

"these days"?

[kind nimbus]

well

[opaque sigil]

As part of our efforts to improve service reliability, we are streamlining our model offerings. Over the coming weeks, Opus 4.7 will replace Opus 4.5 and Opus 4.6 in the model picker for Copilot Pro+

This model is launching with a 7.5× premium request multiplier as part of promotional pricing until April 30th

surely I'm just misunderstanding and it's not billed at 7.5x right

[glass jetty]

watching ve42?

[opaque wharf]

Maybe you have 7.5x more quota from the purchase or free trial or whatever

[opaque sigil]

that'd be such weird wording then

[fickle rain]

Probably some edge case with cloud stuff that got bugreport

Then it got “fixed”

[opaque sigil]

the whole point is you get 1500 premium requests and opus is billed at 3 per prompt (arguably already not worth with codex being 3x cheaper)

time to get out of bed and check I guess

only other thing I could think of is that they added rate limits and this is for that hmmm

aint no fucking way

[opaque wharf]

Just as I was uninstalled wordpress luckily enough
https://www.youtube.com/watch?v=piah4fV_o2Q

[fickle rain]

How much does it cost from Anthropic themselves

Same as 4.6

[opaque sigil]

yup

[fickle rain]

Means GPT 5.5 is probably Soon™

[opaque sigil]

you only get medium reasoning as well, i'm impressed honestly

[vague sonnet]

7.5 * 0 = 0

😎

[opaque sigil]

[opaque wharf]

Speaking of windows exploit....

Where have I seen this picture before

[sage crag]

reze from

chainsaw man

[opaque wharf]

I think it was a profile picture of someone on Neurocord

[opaque wharf]

Ohhh

So it is a famous character

I thought it was some random pic

[sage crag]

[opaque sigil]

reze

[olive sable]

I went shopping

[rough bloom]

did the shopping succeed

[olive sable]

Yes

I bought clothes

[opaque wharf]

:NeuroBounce:

Did you mange to get 100% discount?

[nocturne olive]

[olive sable]

I did mange, i did not get 100% discount

[nocturne olive]

No free stuff

[sick owl]

I'd say large leaps are usually a good indicator but the incremental differences don't tell you anything

[safe path]

frontier benchmarks are literally worthless

they're usually only good enough to get a sense of a gap, and that does need to be quite substantial

also, it's quite funny to me GPT 5.4 has never been verified on the swe-bench verified dataset

anyone got a spare $2000 to burn to submit results?

[vague sonnet]

my point is even if all those tests hit 100% the models would still be dogshit

[nocturne olive]

I would spend it on a 4090 if I did

[safe path]

oh yeah since no one has linked it yet, here's the relevant blogpost: https://www.anthropic.com/news/claude-opus-4-7

[opaque sigil]

all the swe benchmarks are python only right

[safe path]

or well

at least the normal swebench is

there is this: https://www.swebench.com/verified.html , swebench verified, which is a subset made together with openai. makes me think why they didn't release gpt5.4 results, mebe they just didn't think about it

and swe bench multilingual, which works with other languages: https://www.swebench.com/multilingual-leaderboard.html

[opaque sigil]

[safe path]

swebench i think is the closest software engineering benchmark we have, since it actually does try to solve human PRs

that being said, the testing is behavioural. ig that's good enough

[opaque sigil]

problem is that realistically they all train on it since it's public github prs

[safe path]

poisoned dataset most likely

as time goes on they should get better at it

i think it's mebe worth only treating the numbers as a proxy to how well the LLM performs as autocomplete that is also able to call the right tools to fix problems

[opaque sigil]

yeeeeeah

[umbral wigeon]

Gemini cli is funny

With caveman extension

It can speak chinese now

[opaque sigil]

oh cool, opus 4.7 has a new tokenizer that results in much higher token usage

i was wondering why it went so high so quickly

[umbral wigeon]

You can use caveman, TOON to reduce input/output

So ai don't yap

[opaque sigil]

marginally yeah

[vague sonnet]

too bad you can't make an extension that makes it output truth

[umbral wigeon]

I don't understand much english, so I use caveman

[safe path]

me not like output, u do again ooga booga

[maiden geyser]

hello vanor

[umbral wigeon]

Ai kinda struggles with game dev lmao

Unlike web dev

[opaque wharf]

The lack of training data

[umbral wigeon]

It doesn't have soul

[opaque wharf]

It may be able to do individual tasks in vacuum, but comes the time to merge the art and the code it'll fall

[umbral wigeon]

Google:
AI struggles in indie game development because it lacks the holistic vision, creativity, and artistic coherence required to make a compelling, polished product. While AI accelerates tasks like coding snippets and asset generation, it often produces generic, buggy content that demands heavy human revision. Over-reliance on AI can lead to bland, inconsistent, and criticized games

[sick owl]

I thoroughly disagree

Just look in the correlation between Claude Mythos's capabilities and the benchmark leap

[opaque sigil]

tldr: they're function approximators for the average

i saw some people speculating that mythos is a looped transformer

[vague sonnet]

I know how LLMs work better than like all the enthustiasts
so I can't take that technology seriously

using a text-based model for anything logic-related is like hammering a nail with a screwdriver

if you're dedicated enough it might work out
but you're likely going to stab yourself in the hand

big part of the issue is a lot of the users aren't knowledgeable enough to know that it's not doing what it's supposed to

[rigid snow]

if you’re trying to vibecode then maybe

just don’t

easy

[nocturne olive]

Just like if you're trying to RVC
Just don't
Vocal synth instead

[rigid snow]

nuh uh

[stark needle]

sigm

[safe path]

shadowo

[olive sable]

vanorsigma

[opaque wharf]

Toast

[olive sable]

Toast

[opaque wharf]

Toast🔺

[dull egret]

https://youtu.be/83k16gfVTMk

[glass jetty]

toast?

[opaque sigil]

[glass jetty]

[rough bloom]

Toast

[dull egret]

Sourdough

[opaque sigil]

looks weird

[glass jetty]

🔪

[silent cloak]

Cookie dough

[glass jetty]

linux 7.0.0 release overnight

[opaque sigil]

it's been a couple days

[glass jetty]

ai generated instructions on fly

wait you arent saying i missed 7.0.0 release?

i really not reading news lately

[opaque sigil]

[glass jetty]

[silent cloak]

Finally Linux 7

[karmic sorrel]

I just came across someone who got burned by Gemini recommending that they use -fomit-frame-pointer for debug builds.

[glass jetty]

erm, whats wrong?

[opaque sigil]

it kind of messes with stack traces I think?

[sage crag]

dont omit the frame pointer if you need debug info

ye it messes with stack traces

[karmic sorrel]

They were struggling with debugging because they couldn't walk the stack properly.

[sage crag]

you can omit the stack frame pointer and only use the stack pointer in a lot of leaf functions

[glass jetty]

so that was just random

[safe path]

can someone do a challenge where they're only allowed to prompt with 1 character each turn, to finally get the result they want

like the wiki link game to get to a target wiki page, but with LLM single character prompts instead

[glass jetty]

[safe path]

i mean

llm prompt results are probabilistic right

there's surely a non-zero chance that's doable

[glass jetty]

[opaque sigil]

does this count

[tender river]

ye

make sure not to omit in

konii function

[sage crag]

[opaque sigil]

this is boring

[rare bramble]

kind of got it?

[glass jetty]

ai users...

[safe path]

[opaque sigil]

it makes sense tbf, you just slowly fill the context and any somewhat useful model should be able to decode your instruction and follow it

[opaque wharf]

Evil stream

[glass jetty]

[olive sable]

[stark needle]

Toast

[sage crag]

Toast

[vague sonnet]

Toast

[olive sable]

What if i like toast, but not whatever "uooohhhhh yessss mmmmmm" is supposed to mean?

[stark needle]

@Toast he just confessed to you

[olive sable]

No

Confess would be if i loved toast, which i dont

I like toast

I dont love toast

[nocturne olive]

EvilSynth (tm) soft vocal mode
-# Coming soon

[lost valley]

I request a voice changer

[nocturne olive]

[stark needle]

In japan 好きです (suki desu) is used as a formal confession

[olive sable]

Do i look Japanse to you?

[tender river]

no hiragana in font

install

[sage crag]

nrrr

noto-fonts-cjk

scary

[tender river]

shouldnt it be in maple cn

[sage crag]

does that look like the font im using

[tender river]

no

it does not

[sage crag]

its kinda nice

except for the

i

[tender river]

very round

[rough bloom]

llnes

[sage crag]

wrr

interesting

[azure lynx]

counterpoint to quote in sample text: most code is never read again after building.

[sage crag]

compared to my current font

iosevka variations

all look the same

[tender river]

same font
4 different

[azure lynx]

consider opendyslexia for ease of reading

[tender river]

and 6

[sage crag]

( also different

7 also different

[vague sonnet]

factually incorrect

[sage crag]

& also different

i still like

annotation mono

it looks like opendyslexic but

mono

code

[tender river]

qwerasd

username like the average password in 2000s

[sage crag]

ye

{qwer}tyuiop
{asd}fghjkl

[tender river]

i remember that my first password was qazwsx

back then password requirements were lax enough to allow it

[azure lynx]

mine was a sentence. like all good passwords, it was too long and made me laugh but wouldn't've been "secure" enough.

[sage crag]

all the trendy font websites break because i

turned off font downloading

[tender river]

konii resilient web development

render all text in canvas

[sage crag]

ye

see this is the optimal thing

just

a jpeg

space mono is cute

recursive monospace is nice

mononoki is cute

juliamono is interesting

i dislike how much i like ubuntu mono

[tender river]

trades size for weight

[sage crag]

ye i like heavy fonts

like cascadia code

it used to be my main font

i dont like when the fonts are too wide though

feels stretched

ye

[tender river]

its wider but shorter

[sage crag]

its a cute font

except

its

wide

[tender river]

ye

look weird

[sage crag]

super wide fonts are less readable

[tender river]

nix is a satellite of pluto

[sage crag]

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

nix is also from

ancient greek religion

like

styx

hydra

and charon

maybe kerberos idunno

[tender river]

that nyx but close enough

[sage crag]

it the same thing

[tender river]

ye, only changed from nyx to nix to avoid name collision with an asteroid

so thats why the nix ci thing is called hydra

[sage crag]

that why there a nixpkgs maintainer named

sharon

[tender river]

ohidontsee

[vague sonnet]

cerbr

[sage crag]

maple mono

fat font

i like this font it good

ye

[vague sonnet]

damn you guys font nerds

I just use the defaults
or comic sans

[vague sonnet]

I hate that 'i'

[tender river]

not sure if i like maple mono or noto sans mono more for programming

[sage crag]

maple mono

it better

i not like ligatures for programming though

so i have them disabled

[tender river]

actually i agree

some parts look about the same to me but some look better on maple mono

[kind fable]

the first one look weird, but the second one seems to have alignment issues

[tender river]

open image in new tab for best results

[opaque sigil]

everyone needs a little whimsy in their terminal

[tender river]

hello toast

[opaque sigil]

i can't go back to most mono fonts, they're so boring

[warped narwhal]

you take that back!

[opaque sigil]

it's true though

[tender river]

niuh

[sick owl]

The LLM complimented me, I must be so clever

[leaden crest]

hiya toasts

and the potato side dish

[vague sonnet]

[sick owl]

I clicked on one of those clickbait youtube "RUN CLAUDE CODE FOR FREE WITH GEMMA 4" videos because I was curious as to how stupid they really are

Holy shit

They're running Gemma 4 E2B Q4

[jovial remnant]

model for ants

[olive sable]

[sick owl]

wot

[vague sonnet]

if you're not quantizing you are behind

[sick owl]

If you're going to crush a model that small down to Q4 you might as well not bother using AI to begin with

Not to mention the stupidity of pushing it for coding of all things

[vague sonnet]

it's a text completion algorithm

its only use should be suggestions in coding

not even writing, just suggesting

[sick owl]

Brother it is a 4 bit quantized 2 billion effective parameters model

Starting to doubt your "I know more about AI than all the enthusiasts" claim from earlier

[warped narwhal]

it will say that 1+1=3

[vague sonnet]

it's small as it should be

LLMs are a waste of electricity

[sick owl]

[olive sable]

[warped narwhal]

[olive sable]

I agree that efficiency should be a priority, but LLMs arent fully useless

[vague sonnet]

LLMs should get replaced by LMs

[mighty thorn]

Suspiciously in-a-discord-for-an-LLM shaped hypocrite

[sick owl]

Its ragebait (I think) (I hope)

[warped narwhal]

I wonder how the "small" models of today compare to the bigger models from a couple years ago

can e2b q4 with web search beat chatgpt 3.0? maybe

[vague sonnet]

from my testing a well prompted 1.3B can do better than a 7B from a few years ago at similar quantization level

[opaque wharf]

I'm impressed by managerial position to create value out of thin air

https://www.businessinsider.com/microsoft-executive-suggests-ai-agents-buy-software-licenses-seats-2026-4

[opaque sigil]

speaking of "small" models, new qwen 3.6 model dropped today apparently

[vague sonnet]

"3.6" looks inside 35B

[sick owl]

3.0 defo yeah

4 I'm actually not sure about

[warped narwhal]

how about 3.5?

I remember it being not so good

[sick owl]

I think e4b matches or beats 4

[sick owl]

I'd say it probs beats 3.5 too

It definitely hallucinates a lot less than both of those models

[sick owl]

Saw that yeah

Must admit I'm very skeptical of some of these claims

lol at including these

[opaque sigil]

good thing the claims should be pretty easy to verify

somewhat at least

[sick owl]

Troo

[faint sandal]

https://fxtwitter.com/TiHIMothee/status/2044806841935511966

TiHIMothy (@TiHIMothee)

@kenwheeler Truly incredible

**💬 2 🔁 2 ❤️ 84 👁️ 12.3K **

FxTwitter · Apr 16, 2026, 3:55 PM

[sick owl]

The departure of so much key talent from the Qwen team just has me going ↓

[faint sandal]

I don't get it

[fickle rain]

Ragebait or postirony

[real sierra]

the original question is "once in a minute, twice in a moment, never in a thousand years"

to which the correct answer is indeed "m"

the person asking the question misquoted the question but the AI knew the answer to the actual question and gave it

[faint sandal]

oh that makes far more sense

[real sierra]

so to me that screenshot makes the human look way dumber than the ai

[fickle rain]

I think it’s silly thing like counting Rs in strawberry

[faint sandal]

I was like wtf is in "second" but twice in "moment"

grok's funny though https://fxtwitter.com/TerraUnearthed/status/2044807574743024058

Terra Unearthed (@TerraUnearthed)

@TiHIMothee @kenwheeler Grok's brilliant answer 🤣

**💬 5 🔁 1 ❤️ 20 👁️ 1.8K **

FxTwitter · Apr 16, 2026, 3:58 PM

[opaque sigil]

they cranked the tism lever for opus 4.7, need to give very clear instructions

kind of annoying

[fickle rain]

So they made it like Gippity?

[opaque sigil]

i suppose

[sick owl]

It is a failure of generalisation that it gave an answer relevant to the original question and not the users edited version

Gemma 4 with a surprisingly strong answer comparatively

[kind fable]

https://fxtwitter.com/thinkingshivers/status/2044830003733365063

Shivers (@thinkingshivers)

Let’s investigate this “substantially better vision.”
︀︀
︀︀What chord am I playing?
︀︀
︀︀Gemini 3.1 Pro: C-E-G = C-major
︀︀ChatGPT 5.4: A-C-D ≈ Amin(add11)
︀︀Claude Opus 4.7: “Not enough information to say.”
︀︀
︀︀These models are still blind as bats. They have no idea what they’re looking at.

Quoting Claude (@claudeai)
︀
Opus 4.7 also has substantially better vision. It can see images at more than three times the resolution and produces higher-quality interfaces, slides, and docs as a result.

**💬 48 🔁 12 ❤️ 414 👁️ 29.3K **

FxTwitter · Apr 16, 2026, 5:27 PM

[faint sandal]

https://fxtwitter.com/promptprincess/status/2044837166916784403

♡emily♡ (@promptprincess)

oh

**💬 3 🔁 5 ❤️ 143 👁️ 3.5K **

FxTwitter · Apr 16, 2026, 5:55 PM

[opaque sigil]

[sick owl]

Gemma 4 with a surprisingly slick answer again

Wtf

I mean it slipped up but it was on the right track

[opaque sigil]

https://fxtwitter.com/PrismML/status/2044833023682896134 big if true

PrismML (@PrismML)

Today we’re announcing Ternary Bonsai: Top intelligence at 1.58 bits
︀︀
︀︀Using ternary weights {-1, 0, +1}, we built a family of models that are 9x smaller than their 16-bit counterparts while outperforming most models in their respective parameter classes on standard benchmarks.
︀︀
︀︀We’re open-sourcing the models under the Apache 2.0 license in three sizes: 8B (1.75 GB), 4B (0.86 GB), and 1.7B (0.37 GB).

**💬 47 🔁 103 ❤️ 899 👁️ 106.0K **

FxTwitter · Apr 16, 2026, 5:39 PM

[sick owl]

Emphasis on if true

[opaque wharf]

All roads leads to therapy

[opaque sigil]

indeed

[sick owl]

1 bit bonsai looked alright in benchmarks but in practice was more impressive just for the fact that it worked at all rather than for any practical utility

[opaque sigil]

worth checking out at the very least

[glass flower]

could be a insane model for local development if true like inline assistant or autocomplete etc.

i'll wait until gguf versions

[vague sonnet]

what format did they upload in

[glass flower]

mlx

[sick owl]

If I had to guess what went wrong here I'd assume the image tokens just aren't granular enough to properly differentiate the keys

Ha, claude agrees

[vague sonnet]

there are ggufs for the 1bit non-ternary model

[opaque sigil]

the ternary ones are mlx apparently

[vague sonnet]

yeah

[opaque sigil]

hmm

[vague sonnet]

https://huggingface.co/collections/prism-ml/ternary-bonsai

damn they doing benchmarks on phones for those
now that's the kind of language model research I can get behind

[opaque sigil]

it's been allegedly compiling shaders for about 10 minutes now, surely it'll finish soon right

[vague sonnet]

damn bro what are you compiling

[opaque sigil]

their web demo

Program Sub requires f16 but the device does not support it
hmm

i wonder if it's trying to use the cpu for some reason

[vague sonnet]

the demo is just for 1bit right?

or am I blind

[opaque sigil]

https://huggingface.co/spaces/webml-community/bonsai-ternary-webgpu

[sick owl]

Gonna give it a try in a sec

Curious about how the 8B one holds up to scrutiny

[opaque sigil]

webgpu doesn't seem to detect fp16 support

weird

even switched to chrome for this

[vague sonnet]

2 prompts for the smallest model and 2 lies 😔

typical LLM behaviour

[opaque sigil]

idk who to blame so i'll just blame nvidia

[sick owl]

Yeah its not very bright

[vague sonnet]

"A heartbeat is a very rare event in the context of human life, occurring only once every few hundred years or so, and even less frequently in the vast time scale of a thousand years"
BANGER

[opaque sigil]

[vague sonnet]

ah boring

[sick owl]

Hold up

For two gigs thats not half bad

[opaque sigil]

that's probably 1:1 in the training data tbf

[sick owl]

Even contamination surviving a quant process that aggressive is impressive

[vague sonnet]

is it even quantized tho

[twilit vigil]

Remove that refusal through further training.

[opaque sigil]

wasn't the whole thing about bitnet that you have to train the model for it already

or am i misremembering

would assume it's the same here

[vague sonnet]

their paper doesn't mention any quantization so

I assume it was just trained like that already

oh wait does ctrl+f not work on pdf on github

[sick owl]

As far as I can tell it uses QAT on top of the Qwen 3 8B base model

Lemme actually go read the report though

[vague sonnet]

shared weights are an interesting approach

idk why they keep the 0 tho
could just prune those probably

I guess pruning might make it awkward to read

[dusk trout]

I just got Copilot CLI set up w/ my git+perforce workspace (code is handled in git, data/assets in perforce), and it is pretty crazy. Never used AI for work till now

Using Claude Opus 4.6 (3x token rate) currently and its reasoning ability is pretty spot on. If I point it to relevant files and give it context about a bug I need to fix it’s been pretty quick about identifying issues and suggesting fixes. Even got it to follow my C++ conventions lol

I feel like my job is gonna be half prompting AI very soon

[vague sonnet]

[dusk trout]

anyone else have any experience with Copilot CLI? its much more effective than I thought it would be. curious what people have tried with it. The direct copilot integration into Visual Studio has been slower and much worse in my (brief) experience

[opaque sigil]

it do be like that

[frozen hollow]

Nah I’ve used opencode Claude code and Gemini for CLI

[glass flower]

just tested the 8b model for bonsai... for the size quite alright i think

[kind nimbus]

I just found this

https://sylve.io

looks like the proxmox web ui. i'm gonna try that out

[kind fable]

Meow meow meow
It's 1AM i can't sleep so i'm making Web dev tutorials on my phone
Very very fun
https://julienraptor01.ddns.net/Modern Theming.html
Like and Subscribe or something like that thanks

[safe path]

https://openai.com/index/codex-for-almost-everything/

codex can u give me a billion dollars thanks

why is everything being released today

[vague sonnet]

the scheme

[sharp nacelle]

I used codex to rebuild a web page, it had a really hard time with the graphics from an image, claude code was much better at that (so far)

[jagged turtle]

codex please shut down my power grid

[sick owl]

This guy did a very good job of summarising why I'm suspicious of Qwen 3.6

[sick owl]

I found it to be more or less useless not gonna lie

Much like 1-bit it seems more like a proof of concept than anything

[glass flower]

it is definitely not great but i'm sure in some very very controlled situation you can probably get it to do something useful

[glass flower]

well.. but is 3.6 better than 3.5?

[hard delta]

i'm better than all ai models cause we live in a simulation and i'm hallucinating all of this and regularly i'm surprised by my own hallucinations

[fickle rain]

Sometimes I think someone should take my computer privileges away

[sick owl]

I genuinely think FunctionGemma at 350M parameters probably has more practical utility

[opaque sigil]

i agree, putting brackets on their own line

https://www.youtube.com/watch?v=xSla5vfGi3A

[timber dagger]

i think i am slowly losing my marbles over python and html generation

i think python just doesnt like me
the fact that I miss R is crazy

[warped narwhal]

I miss R
Have you considered getting professional help?

[timber dagger]

i should

like i really dont think linux is the issue

[warped narwhal]

It's likely a problem between the keyboard and the chair

[timber dagger]

i spent today like 3 hours just loading a stupid text file into the html
mistral of all things was the only one that actually helped in debugging

[timber dagger]

always has been

this is what happens when you force mechE to freaking code
man

[opaque sigil]

https://ziglang.org/download/0.16.0/release-notes.html finally

[fickle rain]

Completed Migration to NtDll

On Windows, all standard library functionality is now implemented based on calls to the lowest level stable syscall API. The remaining extern functions in the standard library which make calls to Windows DLLs are:

extern "kernel32" fn CreateProcessW(
extern "crypt32" fn CertOpenStore(
extern "crypt32" fn CertCloseStore(
extern "crypt32" fn CertEnumCertificatesInStore(
extern "crypt32" fn CertFreeCertificateContext(
extern "crypt32" fn CertAddEncodedCertificateToStore(
extern "crypt32" fn CertOpenSystemStoreW(
extern "crypt32" fn CertGetCertificateChain(
extern "crypt32" fn CertFreeCertificateChain(
extern "crypt32" fn CertVerifyCertificateChainPolicy(

This avoids bugs, performance pitfalls, and missing functionality on Windows, making Zig programs more robust, lean, and fast than other programming languages that target this platform.

Notably the Batch API and Cancelation have full Windows support with efficient implementations thanks to these efforts.

Users who wish to target older versions of Windows such as XP, or for whatever reason would rather their applications use higher level DLLs such as kernel32 are encouraged to collaborate on a third-party I/O implementation that eschews NtDll.

There are no plans to migrate away from using the above listed functions.

That seems like an interesting idea

Well """interesting"""

[young plover]

The Windows kernel API is better when it's stable. You just have to pray they don't change it.

[fickle rain]

P sure antimalware might get mad at this cuz generally executables don't only import from NTDLL

[kind fable]

I have a stupid idea for how we could add translation support to html
https://julienraptor01.ddns.net/Modern Theming and Translation.html
It's 3AM and i should sleep instead of reinventing the world
But hey with my idea no need to use js to have translations on local web pages and PWA
And the polyfill literally is doable as well i did it (it doesn't work entirely in Firefox tho cause can't theme dropdowns)

[fickle rain]

Well it's just what clang-format Microsoft style does IIRC

Actually I don't remember

[opaque sigil]

you're probably right

[fickle rain]

But I carry my .clang-format everywhere

[opaque sigil]

same except it's based on chromium

finally i can neatly pattern match on binary messages

[fickle rain]

I actually need to finish the challenge now
I mean I got the "hard" part done I guess

[faint sandal]

not really most sane EDRs don't use that as a metric

and if that were the case a bunch of legitimate native images will get flagged

[kind fable]

Demo

[opaque sigil]

always love to see having a smarter compiler

[fast pagoda]

i started reading the text between the code blocks as if it was just output from the compiler liek a total dumbass

"damn, that compiler is writing a fucking blog post, it really is smart"

[opaque sigil]

[fast pagoda]

look inside

microsoft phi 3

for some reason

nother day another nickel

[opaque sigil]

phi is kind of endearing

how utterly useless it is

[fast pagoda]

yes yes i agree and definitely am discussing the same phi

[opaque sigil]

the language model

[fast pagoda]

i feel this way about granite

[opaque sigil]

i shouldn't be surprised that IBM has a line of models

but there's probably a reason i did not know this until today

[fast pagoda]

yeah it was big blue going "waow, that's a lotta copyright infringement" and wanting something the bobs can enjoy

it's pretty damn good at staring at charts and shit

"enterprise document domains"

they LOOOOOOOOOOOOOOOOOOVE documents

this baby'll read the OS/2 promotional material and digitize that shit into the most dogshit ancient FileNet bucket of all time, lickety split

[sick owl]

Yeah they're pretty much braindead

But they are extremely predictable and like funyun said good at OCR/charts

[fast pagoda]

theyre good little corpos

[opaque sigil]

oh i thought you were joking about the documents part

[fast pagoda]

no

they fucking love documents

[opaque sigil]

i should expect nothing less from IBM tbf

this is on me

[fast pagoda]

most ibm possible llm yeah

the new one is apache2 as well which is nice

rock

that is all

[kind fable]

IPv6 is so bad that we looking at IPv8 lmao
https://www.ietf.org/archive/id/draft-thain-ipv8-00.html

[void estuary]

lol

will be interesting to say the least

[kind fable]

Who even tought that colons in IP adresses would be a good ideas
Only dots and numbers allowed

[young plover]

We talked about this a bit yesterday. It's one dude submitting a bad draft.

[opaque sigil]

wdym, nothing wrong about it

this solves all the problems

(just ignore the 10 new problems)

[kind fable]

It's cooler than IPv6 tho
The name is better and the way to write IPs is better, i'm sold

[fast pagoda]

click it n ship it

garry wouldve pushed that shit yesterday

[kind fable]

I mean we should just ask Claude to review it and push it

We'll see if it's down tomorrow

[young plover]

It wants the firewall to block any outgoing packets which don't have a preceding DNS request for a valid https domain. :dentge:

[kind fable]

It's obv a joke tbh

It being not decentralized is an instant fail

[fast pagoda]

https://llanfairpwllgwyngyllgogerychwyrndrobwllllantysiliogogogoch.co.uk/

[amber fractal]

why is it protected by cloudflare

[kind fable]

Bruh i'm blocked

[jagged turtle]

because yes

[amber fractal]

I should reveal my most stupid project that I'm activly putting in dev time to, I'm making an image generator because it would be funny to troll AI purists with

If the argument is I'm stealing from others then I will just use my own photos

if the argument is that I'm taking photos of things that have a copyright to them, then I just won't take photos

That means the entire dataset is some custom patterns applied to gradients with some extra noise on top

checkmate

Also means I'll have one of the few completly legal models

even if the output is

[opaque sigil]

[amber fractal]

orb

How this works under the hood is abusing gimp3 for all it is worth, every alteration is a filter mask on the orignal image including the noise

so we get the very fun thing to where if I reroll the noise, the entire image changes while the basic concept stays the same

Plus this is unreasonably very affected by the inital color, so color rotation is another axis I can use

So I can actually get quite a bit out of each of them

and they only take about an hour for me to make each

[safe path]

[leaden crest]

nooo iggly dont abstract pls

[amber fractal]

It is inevitable, the abstract can't be categorized or claimed

[amber fractal]

TLDR: yes

[fast pagoda]

too late

wow hf has dogshit download speed today

im going to die of old age

[kind nimbus]

who needs sleep anyway

it's 6:41 am and i'm still awake

[fast pagoda]

me yesterday morning at 6:41 am

https://tenor.com/view/festthefestivefest-morgone-jones-morgan-jones-tweaking-tweak-gif-14491782401479632016

how not tired whatsoever

[mighty thorn]

Wait qwen3.6 is already out

What is this shit

[fast pagoda]

yea claude 4.7 too

[mighty thorn]

Bro

3.5 27 dense clears this MoEslop on almost every benchmark

[fast pagoda]

yezzir

[true hemlock]

https://wccftech.com/apples-new-performance-cores-within-the-m5-pro-chip-are-outcompeting-similar-cores-within-intels-panther-lake-chips/

i wonder how hard did apple go on clock gating

spec integer rate...

i see

spec doesn't necessarily forbids vectorization

[fast pagoda]

notebookcheck pulled these out of somewhere

prob their ass

[true hemlock]

but well, integer rate itself naturally isn't really vector dominant

its somewhat branchy and latency sensitive

and pretty irregular

apple had advantage with their insane IPC

[fast pagoda]

4295 points single 28436 points multi for m5 pro w/18cores on geekbench6

[true hemlock]

though that also means apple did implement clock gating on micro ops component level

explains the 2.5w draw

because spec benchmark doesn't stress all the alu 100% util because its more branching and latency heavy

[fast pagoda]

is it finally time for arm to kill x86

surely

[true hemlock]

i mean

the issue with arm vs x86 is more of versatility and compatibility issue imma ngl

both are fine

[fast pagoda]

different beasts but nobody cares if it's way faster for many uses if it's cisc/risc/weird compat

can finger it out ya kno

[true hemlock]

x86 is more versatile and compatible, arm needs its own ecosystem

both can be great

but i really really feel like arm isn't the reason its efficient.

[opaque wharf]

I see one big glaring issue. Licensing

[fast pagoda]

Xx86_ARM_68xX

[true hemlock]

power draw still gets eaten by clocking traces and processing

cant cheat that one out

can only make it look good by clock gating

its why apple hit sub 2W idle
and so did my tuned i9 13900

idk why amd didn't do it i can't hit below 20W

on my 9950x3d

[true hemlock]

this needs to be pinned

actually pin that in atomic

man don't you love it when headlines NEVER goes deep into these details and just throw in "woah apple new chip beat <insert other chip here> while being very efficient"

[fast pagoda]

that's because 99% of people even in the audience will see that and go "aawagga" and then forget they ever saw it

[fast pagoda]

ccd gets stuck on

[true hemlock]

now you have misinfo like "arm is efficient"

[fast pagoda]

i beliebe most zen5 idle draw is gonna be a stinky io die

and the wonderful infinity fabric you love

[true hemlock]

actually lets test that out

[fast pagoda]

i think the apu dont have the problem cuz iod isnt a separate unit

[true hemlock]

all 16 cores itself is at 0.128w lmao

core ccd total 3.7w

[fast pagoda]

what's it run the cores at min on wandows

[true hemlock]

soc and misc draws a lot more

[fast pagoda]

linux amd-pstate min is 400

[true hemlock]

actually, brb, imma tweak some c states

or p state

[fast pagoda]

'soc muh misc' - amd

400mhz seems to be about .3W/core with the voltage drop that comes w/ it

[true hemlock]

infinity fabric my foe

[fast pagoda]

got curiouis about strix halo and that bastard runs 10-12w at idle anyways kekw, supposedly a memory bus/IF(classic, it's dummy thicc on strix halo because of the gpu)

[true hemlock]

10w idle is still high lol

[fast pagoda]

yeah it is

that was calling it out for being dumb high for what i'd hope

[fast pagoda]

strix point does a bit better

[true hemlock]

intel my beloved

[fast pagoda]

like hx370

manages like 5-7w idles but even that is silly

wonder what the idle on my 8745hs is prease hodl

[kind fable]

Someone who invent HTML spec out of nowhere at 3AM does need sleep
#programming message
#programming message

[opaque wharf]

JS is made in 1 week

1 WEEK

[fast pagoda]

stinky

[fast pagoda]

[tender river]

i'd say it's still reasonably well designed, though i'd like to see the alternate universe where we ended up with scheme instead

[fast pagoda]

we did end up with scheme

just nobody use

[true hemlock]

https://x.com/IntCyberDigest/status/2044568738767294652

International Cyber Digest (@IntCyberDigest)

🖥️🔥 Two inmates at an Ohio prison built a secret hacking operation from behind bars, using computers they were supposed to be recycling, they downloaded and sold porn in return for snacks, built a hacker toolkit with Kali Linux and password crackers, and created fake passes to

X · Apr 16, 2026, 12:09 AM

[fast pagoda]

holy based

[fast pagoda]

randy sudofresne

[true hemlock]

a bit lower now

[fast pagoda]

cell block nmap

[true hemlock]

apparently core powers is an average

bruh

2.2w on all core is nice

but soc still fucks it up

[fast pagoda]

............ escape from /dev/cellblk13

the count of monte crontab

im out of ideas

[true hemlock]

this was taken 4 months ago with my i9 13900

[fast pagoda]

cant believe how dogwater linux powermonitoring is

time to reboot to the forbidden drive

[true hemlock]

theres a reason why i never bother overclocking my servers

[true hemlock]

1.2w on all core

amd could never

i don't think amd bothered clock gating the io die much

[fast pagoda]

from what ive read (not a lot mind you) it's something something chiplets something something iodie separate and always on asf because it's separate because ?????? amdentge

the reason of course naturally being "FUCK you" - amd

[true hemlock]

its to save cost

[fast pagoda]

that's the chiplet way

[true hemlock]

because big die are somewhat non linearly more expensive

[fast pagoda]

ye

[true hemlock]

because something something defect densities and yield rate

[fast pagoda]

i just never figured the direct reason that it being a chiplet requires always on iod, i guess to avoid wake issues and shit

[true hemlock]

but the power issue isn't because chiplets

[fast pagoda]

infinity fabric is called infinity brah

it must

run

at

infinity

[true hemlock]

though chiplet is also ass because extra latency

[fast pagoda]

what you dont want your instructions to run an olympic marathon always

[true hemlock]

its not as simple as wiring traces through the package you still need to go through amplification stage and signal shit

latencies between chiplets are ass that way

[fast pagoda]

part of the power problem as well

[true hemlock]

fair

misc itself is 6 watts so yeah

but the io die

istg

[fast pagoda]

and it's prob not helping all the io retain any form of steady idle

[true hemlock]

amd can you learn how to properly implement clock gating

[fast pagoda]

turn on x3dturbo

it'll half ya power use :^)

well maybe not fully half

but it will killean the 2nd ccd

and also smt

fking 486 turbo button ass name

[true hemlock]

yeah kills off 1.1w outta the still chunky 26w

[fast pagoda]

what does amd do in epycs and shit for iod

cant possibly still be one iod for all of it

oh it is all of it

mf 12 ccds to 1 io

[true hemlock]

hahaha

[fast pagoda]

tr 9995