extended character set, but still
#programming
azure lynx
fast pagoda
claude gets a block character guy to represent it
he had a birthday hat for the last month
cuz it was around the time the public beta first launched lol
azure lynx
the first computer programs i wrote basically drew an animation by drawing 4 or so frames of ascii art. i did a conveyor belt moving bottles that got filled and a lid put on (i swear it was obvious that's what it was. not just after I said what it was.)
probably inspired by a "how stuff is made" or something tv program.
and i just repeated all the objects so after they had moved 4 frames, it was the same
or the next process
fast pagoda
telnet towel.blinkenlights.nl
azure lynx
they put a video game in a building, right?
with the room lights as the pixels
or was that someone else?
azure lynx
oh... i have seen that ages ago when i was at college
someone i knew said something about playing it next to the original... i don't remember if they were complaining that it didn't match up or saying that it did
/~\
|oo ) It's much too
_\=/_ rocky.
___ / _ \
/() \ //|/.\|\\
_|_____|_ || \_/ ||
| | === | | || |\ /| ||
|_| O |_| # \_ _/ #
|| O || | | |
||__*__|| | | |
|~ \___/ ~| []|[]
/=\ /=\ /=\ | | |
\_________________[_]_[_]_[_]________/_]_[_\______________________
verified it works.
fast pagoda
lol and i thought my terminal was at the wrong resolution
seems like it has been replaced with a new ipv6 version
useless video trimmer failed to remove the first part so im sure it looks like some unrelated clip
me
real sierra
he had a birthday hat for the last month
this cant be the first thing i see opening this channel
fast pagoda
happy birthday shiro claude
if i ever have kids im telling them this is the original version of star wars before they made a real movie
azure lynx
yeah. this is before all the "modern" CGI they added . computers used to be a little more limited in what they could render.
azure lynx
obviously it's all makeup and prosthetics. industrial light and magic were amazing.
fast pagoda
they really could do a lot with a little, they cooked too hard and now people can see my facial expressions though
azure lynx
if only they did ASCII Jurrasic Park so you could assure them they "Spared no expense!"
fast pagoda
they didnt have a UNIX SYSTEM or they would've been able to
azure lynx
just watched an "Everything Great About" for the original. it aged surprisingly well
fast pagoda
bzzzzzzzzzzzzzt wrrrrrrrrrrrrrr
sage crag
bzzzt
sage crag
https://tenor.com/view/termina%27s-cage-gif-2702795870681820605
there's a frame of something in there
real sierra
25 days remain
sage crag
frame of what
fast pagoda
oh shit there is lmao
sage crag
looks like a face
azure lynx
The image features the Moon from The Legend of Zelda: Majora's Mask.
sage crag
azure lynx
today i learned you can right click on a video and say "Copy this frame"
azure lynx
"it's the same picture."
fast pagoda
i guess because that's what makes the delay
it's 2 seconds long at 100 fps lmao
why
azure lynx
coz they put a 10ms delay coz they wanted it to blink quickly
but really, who can know the minds of someone who puts nicolas cage on anything?
azure lynx
shouldnt you be able to set a per-frame delay in gifs?
you can. but it's easier not to.
sage crag
Does it have to be same fps through entire gif
as far as im aware the delay for each frame can be set
obsidian mantle
Mhm
glass jetty
https://tenor.com/view/termina%27s-cage-gif-2702795870681820605
15 days
fast pagoda
heroic is what i wanted, thank god plasmasearch is forgiving
glass jetty
25 until what?
obsidian mantle
I think i saw it when i tinkered with easygif
sage crag
25 until what?
🎂
glass jetty
oh, than 15 days)
azure lynx
3 to the power of 3 is 27. so 27th of the 3rd is more better.
fast pagoda
otherwise it would've clearly had 5 stars already saved
azure lynx
or when you downloaded it if it's not actual file metadata but just disk metadata
fast pagoda
that wouldnt change the format tho
it's even titled gif in the url
but this link downloads a .mp4
even tho file extension can be meaningless
it is h264 i dont think .gif supports that
fast pagoda
so i was messing with the eza flags forgot i was in / at the time
and ran eza -lagR
was ssh'd to my minipc
and R runs recursive ofc
this action listed so much it was uploading like 15 mib
/s
fan went into turbo overdrive as well lmfao
silly
took 1min 7s to run 
benchmark against ls time inb4 i already aliased ls to eza on this (cant remember)
ls took 1 min 18s
sometimes i wonder why i dont get things done
if you have scamazon prime you apparently may have free civ 6 available to claim on luna
jagged turtle
damn
jagged turtle
how did I get one of chayleaf's yt videos recommended to me wtf
jagged turtle
I think I've gotten NS stuff also recommended to me via algo before
jagged turtle
I mean that's always been recommended to me (at least until recently, dk why it stopped) but I think I've seen some stray NS videos idk
idr tbh it was a while ago
nocturne olive
Well we'll see if the next big upcoming NS release gets recommended to anyone
jagged turtle
what's the next big upcoming NS release I haven't had the foggiest clue what it would be
nocturne olive
We'll be releasing a surprise to showcase the NS JP capabilities (finally)
NS JP is actually incredible, even better than NS EN
jagged turtle
Interesting 🤔
nocturne olive
Having some real JP Nur data really helps
jagged turtle
yeah fair
midnight sigil
I guess I can answer my questions now:
- Yes
- No
my monitor is completely unusable now 
azure lynx
guessing JP is not Jurassic Park?
nocturne olive
Japanese
azure lynx
mid answering a question the claude web interface decided "This isn't working right now. You can try again later." and won't let me see the previous reply
Can’t open this chat
It may have been deleted or you might not have permission to view it. Use the share button to share chats on Claude.
I can see it in the list on the left so it's obviously a transient problem.
azure lynx
self-modifying code is impossible to model without emulation.
i wonder if it knows (or can guess) when it's modifying itself
azure lynx
"update the step() function of the claude model to..."
except without being told that's what it's changing
or the web interface
or some other capability.
Ah,,
Claude will return soon
Claude is currently experiencing a temporary service disruption. We're working on it, please check back soon.
jagged turtle
Ah,,
> Claude will return soon
> Claude is currently experiencing a temporary se...
vibe infrastructure
azure lynx
"Solving this is taking longer than expected because... Claude isn't responding. We asked ChatGPT but it just laughed at us."
jagged turtle
You expect them to even look at ChatGPT? They probably vibecoded an internal MCP server and connected <insert agentic AI tool here> to it and just didn't bother caring
they are the ones who made the MCP spec after all
azure lynx
not really. i don't expect it to go down either though.
midnight sigil
I guess I can answer my questions now:
1. Yes
2. No
my monitor is completely un...
no fucking way the warranty is still active until 2026-11-9
‼️
glad it broke early
azure lynx
no mention of overclocking... just you went to play a game and it stopped working and got worse over time. probably.
idle dune
Don't call me out (I don't program)
I use Claude like a home assistant
Help with planning my day, mateirals lists, and schizo ranting about politics
glass jetty
for me seems like cloude in the middle
i sanity check logic full with it, so idk
I might have carpal tunnel from typing all day
opaque sigil
and this, kids, is why you use a wrist rest
idle dune
I also need an ergo keybaord cause I already have wrist issues
But no one makes a mechancial ergo keybaord
By that i mean a fancy one that lets me repalce the keycaps and switches
Wait nevermojd people do
kind nimbus
But no one makes a mechancial ergo keybaord
There are some custom ones afaik
idle dune
There are some custom ones afaik
Yeah I am looking at them now, they are expensive as fuck though. 200-300 cad
Might be worth to prevent more wrist pain
kind nimbus
Yeah custom keyboards are expensive always
You can easily reach 600-800
But if you're into ergo / orto keebs there are a bunch of good options
idle dune
i'm looking into them now, I already have to use an ergo trackball mouse for my right hand because of wrist issues
According to a friend of mine that works in cybersecuirty trackballs are a safety feature?
Don't know how though
kind nimbus
?
idle dune
Don't fucking ask me
I gues sbecuase you can't move it by accident?
idk
But eitherwya I use a ergo trackball as is.
kind nimbus
Never heard of that before as someone in cyber security lmao
opaque sigil
security through obscurity 
glass jetty
good 
silent island
playing this on my laptop makes me wonder how tf did the PS4 handling these kinds of graphics lmaoooo
idle dune
PS4s at the time where equal to a midspec PC but all built on an APU iirc
silent island
PS4s at the time where equal to a midspec PC but all built on an APU iirc
im surprised they managed to make something so powerful for so cheap...
idle dune
They take intentional loses on hardware sold and make it up through game purchases through digital store fronts, subscriptions, and mtx. Cause if they didn't it would be launch PS3 prices whch didn't take a loss on hardware pricing. Which I think adjusted to inflation now launch PS3 is like 1300 bucks or something crazy.
kind nimbus
security through obscurity <a:mhm:1235974179245916170>
Unfortunately too many do that in practice
idle dune
Cause it was 699.99/599.99 when it claunceh in 2006
kind nimbus
Jokes on you my laptop is probably 5 times as powerful as a PS4
idle dune
I think my 2015 macbook pro is too.
silent island
Jokes on you my laptop is probably 5 times as powerful as a PS4
laptops nowadays are more powerful than a PS4
kind nimbus
I have a Core Ultra 9 275HX and RTX 5070 ti in my laptop
idle dune
Sick
kind nimbus
With 96 gb ddr5 ram 
silent island
what shocked me the most is that my classmates still game on PS4s after 10+ years lmaooo
shit's built to last...
kind nimbus
i just love watching the cores go vroom
kind nimbus
what shocked me the most is that my classmates still game on PS4s after 10+ year...
i mean at least they last that long after all
silent island
i mean at least they last that long after all
i just wished laptops were made the same way...
to last as long as consoles...
kind nimbus
some do
it's just that hardware advances very quickly nowadays
idle dune
i just love watching the cores go vroom
I can't wait tull i get my P920 
Gonna be a while since I am waiting for them to go down in price
kind nimbus
I can't wait tull i get my P920 <:Despairge:1010351210034896896>
specs?
ivory plinth
what shocked me the most is that my classmates still game on PS4s after 10+ year...
is it playstation airlines with that thing
idle dune
specs?
I was aiming for dual Xeon Gold 6154's, 64-128gb of ddr4 ecc, and slapping a RTX 3080 and P40 inside it.
silent island
is it playstation airlines with that thing
not rly i think?, my classmates didnt complain too much about the fan noise...
ivory plinth
kind nimbus
I was aiming for dual Xeon Gold 6154's, 64-128gb of ddr4 ecc, and slapping a RTX...
i have an RTX 3080 that i only use for hashcat lmao
maybe i should let it do some work and run an LLM on it as well
idle dune
This was going to be a multiple purpose gaming, server, and llm machine
The P40 was going to do most of my LLM stuff
silent island
not rly i think?, my classmates didnt complain too much about the fan noise...
Imagine playing dmc 5 on a PS4 lmaooo
kind nimbus
i wanted to play around with OpenClaw with vllm. Wanted to deploy it via Nix instead of Docker but had no luck yet getting it to run lol
fast pagoda
Don't call me out (I don't program)
he's good for research
like getting a bunch of info about a topic (with sources) ig research mode in the web
im too scared to use openclaw
security seems wishful at best
true hemlock
I was aiming for dual Xeon Gold 6154's, 64-128gb of ddr4 ecc, and slapping a RTX...
the xeon chips themself are cheap asf but the board is 
💸
idle dune
Its crazy that the boards are the price of awhol e-waste workstation
true hemlock
also the hdplex psu im getting accepts EPS (12-48v dynamic range) dc input and comes with 7.4*5mm to 8 pin EPS adapter
my lipo battery has xt90s output.....
i can't fucking find xt90s to 7.4x5mm there's only xt60 to 7.4x5mm... am i deadass going to do xt90s to xt60 to 7.4x5mm
true hemlock
Its crazy that the boards are the price of awhol e-waste workstation
what's your goal with the hardware btw?
silent island
playing this on my laptop makes me wonder how tf did the PS4 handling these kind...
hmm... do y'all think the PS6 would be backwards compatible with PS4 games?
cuz if it is... im switching to that from my laptop
true hemlock
i might be able to help out on an alternative 
fast pagoda
the xeon chips themself are cheap asf but the board is <:Deadge:1080155317423308...
i wasted money on one
like 2 days ago
fast pagoda
it wasnt that much but it hurt for the age of it ig
but i got xeon platinum
2x 8259cl
true hemlock
xeon gold 6138 is like ryzen 7 5700 perf multicore damn wtf
fast pagoda
it's definitely the boards that keep them cheap ig
i mean c620 boards had a huge range of pricing 2 but not much below 200 bucks and at that price it was a server board with like 0 connectivity
cheaper for you id assume but on ebay anyways
idle dune
what's your goal with the hardware btw?
Server for me and my friends, workstationf or my work, gaming, and running a basic local llm for my work. Its gonna be an alrounder machine.
chrome rampart
are you guys good with grammar, perhaps even pros at it
fast pagoda
no
gilded bay
idle dune
are you guys good with grammar, perhaps even pros at it
prehaps
glass jetty
♻️
fast pagoda
<:neuroSad2:1150144717540036729>
i used to flame people over grammar on facebook
fast pagoda
2009
glass jetty
Never talked in this chat, go tell whoever told me to post it
♻️
fast pagoda
lmao
gilded bay
glass jetty
gilded bay
It's original because I made it
fast pagoda
playing star wars battlefront II and getting into arguments with wine moms
gilded bay
Source: I lied to win
idle dune
I feel so dumb for using the youtube app on android for so long.
Because my friefox literally gives me all the premium features on mobile with a couple extensions.
fast pagoda
i have premium for free*
-# paid for it via buying the phone it came with
gilded bay
I feel so dumb for using the youtube app on android for so long.
Literally get YouTube revanced brochacho chip
gilded bay
damn had to reply it instead of react it
Not even a blocked user btw
This guy might be a buddy
idle dune
Literally get YouTube revanced brochacho chip
No clue what that is. I just do everything through my Firefox instead kf rhe android apps.
gilded bay
No clue what that is. I just do everything through my Firefox instead kf rhe and...
It's a very customizable YT client
idle dune
Oh sick
gilded bay
Highly recommend over browser on phone
kind nimbus
I feel so dumb for using the youtube app on android for so long.
Use Revanced on Android and YouTube Plus on iOS
idle dune
I'll check out revqnced
kind nimbus
And for revanced there are malicious ones only build your own from the real website via their official app
fast pagoda
i thought revanced got shut down at one point
i guess it was likely "shut down"
kind nimbus
fast pagoda
hydra head
fast pagoda
.app is too newfangled to not seem sus to me
kind nimbus
If someone wants a compiled ipa for YouTube Plus on iOS i have that as well 
kind nimbus
https://github.com/mentalblank/YouTube-Revanced
Why not use Revanced Manager to build the APK directly on the phone?
Then you can also select all the features you want
rough bloom
ye, just use Revanced Manager from the official github
https://github.com/revanced/revanced-manager
-# or the revanced.app site, they're both legitimate
kind nimbus
If you ever used anything else change your Google password asap lmao
idle dune
Nah I never used anything else. I used the official app that came preloaded on my phone and Firefox browser. Thats it
kind nimbus
i mean any of the fake revanced apps because that's most likely malware
fast pagoda
yeah that's why it's a hard no from me on top of i imagine risking google banning me, losing my google account would be horribly inconvenient & why i don't recommend using it as lame as that sounds, worse if someone got access to it
opaque sigil
there's morphe too from some core revanced contributors now cause open source drama 
you know how it goes
idle dune
I use protonmail for everything. My Google account is a burner.
kind nimbus
you can also self host your own email server via mailcow .e.g
chrome rampart
if there's an email provider out there, I've fucking used it
kind nimbus
but i also use protonmail primarily
there is also tutanota
those two are the best ones usually
midnight sigil
I grinded 600 lines of code for a small Python project in school lul
the requirement is just a number guessing game
midnight sigil
I guess I can answer my questions now:
1. Yes
2. No
my monitor is completely un...
I was able to access my pc by connecting the HDMI cable to a capture card, and fullscreen preview with OBS on a laptop
idle dune
I was able to access my pc by connecting the HDMI cable to a capture card, and f...
Your poor laptop
tight tinsel
I was able to access my pc by connecting the HDMI cable to a capture card, and f...
shiroko bg and arona plush!!
midnight sigil
my poor laptop
midnight sigil
my poor laptop
suffered from 6 OS reinstallations
midnight sigil
shiroko bg and arona plush!!
idle dune
glass jetty
btw, i was talking about laptop-pc, here is my laptop
(big cap and small banana for compare)
tight tinsel
<:NeuroTease:1079889963287597176>
i wanna buy either a plana or arona plush but was too late 
midnight sigil
btw, i was talking about laptop-pc, here is my laptop
(big cap and small banana...
why is your mousepad burnt
olive sable
Just spent 3 hours of my life repairing the pneumatic brakes on my bike 
glass jetty
suffered from 6 OS reinstallations
erm, it is releave*(who this spells idk)* instead of suffering
glass jetty
why is your mousepad burnt
eh ah, erm... 
glass jetty
why is your mousepad burnt
definitely not pictured it on purpose
idle dune
trackball!?
Yes
midnight sigil
midnight sigil
definitely not pictured it on purpose
idle dune
I need to take a need image since I did update this setup.
But ita a Thinkpad P73 and 2015 MacBook Pro
glass jetty
Just spent 3 hours of my life repairing the pneumatic brakes on my bike <:EvilDe...
i was trying to stick up new pedal for 1.5 hour tho..
idle dune
mint and ubuntu
Both now Kubuntu since I found what I like
idle dune
Jesus christ
midnight sigil
Both now Kubuntu since I found what I like
:O why do you have a 99% medical alcohol
just, sitting on your desk
midnight sigil
battle-scarred
midnight sigil
glass jetty
superiority of not having to maintain desk 
midnight sigil
is that a 49° OLED
glass jetty
this shit is basically insane
idk how you manage to use this display
fiery anchor
is that a 49° OLED
yes. and it's a tad to small, isn't it? maybe i'll buy the 57'' one if that bocems affordable
idle dune
just, sitting on your desk
Cleaning and Gunpla. I build plamo while I listen to Neuro and Evil yap.
midnight sigil
yes. and it's a tad to small, isn't it? maybe i'll buy the 57'' one if that boce...
holy mega rich
opaque sigil
this shit is basically insane
idk how you manage to use this display
it's just two 27" monitors without the bezels in the middle
olive sable
Cleaning and Gunpla. I build plamo while I listen to Neuro and Evil yap.
is that from warhammer 40k or osmething?
glass jetty
when you have this + keybinds?
fiery anchor
this shit is basically insane
idk how you manage to use this display
most of the times i have 3 windows open. left discord, right media (youtube, crunchyroll, reddit, whatever), center game
glass jetty
super+e = open terminal, super+q - first window, super+w - other
fiery anchor
when using code, same but different. ( terminal, ide, documentation)
glass jetty
when i play i focus on game, why would i need other windows?
midnight sigil
most of the times i have 3 windows open. left discord, right media (youtube, cru...
all that but you only have 2 sticks of ram smh
olive sable
gundam isnt somehow?
midnight sigil
water cooling and super thick gpu everything but 2 sticks of ram 
opaque sigil
sth sth 2 sticks is optimal
glass jetty
it isnt even very useful for codding, i have 2 windows on work, and i cant say that it is better, i usually have discord/active browser/youtube there
and it isnt kinda 'i need it', i just have it why not use, but actual window im using is in front of me
fiery anchor
it's 32Gb in Total, fine for most Stuff.
if i need more, i'll buy the same kit again and plonk it in.
glass jetty
i have 8gb and its fine (again, why you need such amount of ram lul, i never even hit 16gb without mem leaks from discord being open 24/7 2 month or clangd cache for idk, or running qt6 build from source in parallel while other programs already use 6gb)
(it is literally only cases i hit 16gb)
olive sable
why stop at 4x?
get a threadripper and do 8x
that way you can sell of both your kidneys
opaque sigil
why stop at 8x? get an epyc and do 16x
olive sable
idle dune
gundam isnt somehow?
No what I meant is Gundam is 40k but for weebs
glass jetty
that way you can sell of both your kidneys
couldnt be enough actually
olive sable
No what I meant is Gundam is 40k but for weebs
ah lmao
i havent watched/played either
fiery anchor
i have 8gb and its fine (again, why you need such amount of ram lul, i never eve...
Yeah. Sometimes i use extensive debugging and tracing tools, that where a 20 Minute Trace can reach >15gb. But rarely. For dev purposes i see however for a new work notebook 32gb as mandatory, since we also work with virtualization and stuff. That can eat RAM quickly.
glass jetty
Yeah. Sometimes i use extensive debugging and tracing tools, that where a 20 Min...
20 min trace is already something extreme
fiery anchor
Heavily Multithreaded event based application written by devs that are used to tasks and interrupts. Sometimes you need 20min to reproduce the sporadic error
glass jetty
like, i kind of already unable to use valgrind with our app, cuz it slow down so dramatically that it no longer have appropriate timings and it literally cant work without manual changes
glass jetty
Heavily Multithreaded event based application written by devs that are used to t...
idk, just use log man, lul
tho trace meant to be solving manually processing everything to log something...
fast pagoda
The automaton fired off an SOS shortly before I lost contact with it around 0800
Thankfully I think he's gonna be alright I found the crash site on time, what a brave soldier tackling such a challenge alone
fiery anchor
like, i kind of already unable to use valgrind with our app, cuz it slow down so...
yes. and with QT, it's even more of a pain. We integrated, behind a global on/off toggle, https://github.com/wolfpld/tracy
It's a neat thing, easy to use, and has almost no perfomance impact. But it will definitly crash your computer if you have a full RAM / SWAP
glass jetty
idk why python framework would use 32gb just for running project, that kinda crazy, tho, python....
glass jetty
yes. and with QT, it's even more of a pain. We integrated, behind a global on/of...
wdym without qt? like you talking about qt object system making everything simple or what?
fiery anchor
nan, meant using valgrind on a app that uses qt is a pain.
a real app though, not some hello world example
glass jetty
yes. and with QT, it's even more of a pain. We integrated, behind a global on/of...
when i was really needed profiler, i used intel profiler
that was real performance question i couldnt answer with logging
whe have performance sensitive application, tho most of the work is gpu classification model inference, i just write code that not doing stupid things and acknowledge dimensions of img to perform better pre-pre processing, pre processing, visualization, etc (basically common sense stuff when you work with multi dimensions)
what do you think about adding profiler to project that actually may live perfectly fine without?
i have some free time before they implement next network i already have outer code for, but just not yet interface to attach
glass jetty
nan, meant using valgrind on a app that uses qt is a pain.
idk, just ignore indirects from qt
like, i really used to since i using only qt
idle dune
i havent watched/played either
Its alright. Some series are much better than others but I am a suckered for the model kits because its 20 bucks and I have a couple months worth of painting and modding fun.
olive sable
fiery anchor
like, i really used to since i using only qt
Maybe that's an Option, havent explored it though. And won't, since i called in sick for a month last week due to burnout.
idle dune
<:NeurOhISee:1136209034966483084>
Not Gunpla specifically but here my last miniature model kit i fully completed.
idle dune
I wanna stream my model kits building at some point
idle dune
the fastest piece of shit in the galaxy
Damn right
cosmic sphinx
@sick owl i, uh, what the fuck-
the new qwen 4B model is actually cracked 
https://fixupx.com/Alibaba_Qwen/status/2028460046510965160?s=20
idle dune
umbral wigeon
Am I cooked bro
AI able to debug faster than me lmao
And also it's all correct code (Reviewed code), passed all tests
With high performance, being able to holds 3000 projectiles moving parts multi-threaded(talking about projectile system)
fiery anchor
Am I cooked bro
nah. Step up a step on the ladder of abstraction, and you'll be fine.
mighty swift
hi any balatro coders 
we are recruiting moders for cards and fun
willow hinge
What's the salary
mighty swift
What's the salary
0 per month lmao
willow hinge
Driving a hard bargain
mighty swift
just community mod for a promise of being the best
umbral wigeon
nah. Step up a step on the ladder of abstraction, and you'll be fine.
Ai cannot write complex code for a reason it just doesn't
But it can if it a terminator robot sitting next to me
willow hinge
Especially if you are trying to make something that doesn't exist
mighty swift
#1336777692070023271 qaaap if you are interested
willow hinge
I am not
umbral wigeon
Also tiktok comments glazing ai just because they bulit something with zero code
Kinda annoying
chrome rampart
they themselves didn't code, the ai had to use code though
literally everything uses some form of code
well, everything on the computer
willow hinge
The less you understand something, the more easy it is to midnlesly praise or hate it
umbral wigeon
I saw this dude on titkok i cant tell if he ragebaiting https://vt.tiktok.com/ZSmTy318H/
bold ice
We can create the game like anime To Be Hero X?
umbral wigeon
Lmao
glass jetty
why you try to use copilot
why ppl use this shit man
it will never work mid-term ||(shorter than log term)||
willow hinge
Cuz it's free and comes with microslop stuff
glass jetty
ais are absolute shit for most tasks
willow hinge
if you aren't doing anything too complex it's good enough
glass jetty
why use microslop stuff
just use arch and nvim
unless you have unsupported laptop pins for micro, you are fine
(literally me btw, but i fixed by searching with ai with a shit ton of context so it can actually find correct solution)
obsidian mantle
how do i add new function to some class from outside 
cpp
or do i just pass the class as argument
glass jetty
except issue with not updating keyring frequently enough (so i had just to go to config and disabling checking for next run)
after instillation i have no issues with arch
glass jetty
how do i add new function to some class from outside <:thonk:602047082064510986>
reflection
see c++26
haha
what you actually need? do you have access to class source code?
or it compiled+declaration?
obsidian mantle
i need to make new function for this class
that i passed as 1st argument now and its fine
i jsut needed access to PollRateMs and m_context
i took another function as example but that function exists in class already
so it just overrides/expands/whateveritscalled
and i need new one
glass jetty
to what class?
obsidian mantle
NeuroSocket
glass jetty
i assume it is compiled or just without sources change allowed
seems like this are just public and assumed to use as public member for some reasons? that weird in term of cpp in first place
but with this api you can just inher from neuro socket and do whatever you want in new class with this members always presented
tho i really doubt that you should do like this anyway
why cant you just remember this values in your logic? if you need access for context and rate, just make value and pointer to it, context is assumed to be used as pointer anyway
just from what i see it basically values to be created once per session
glass jetty
nah it is c api purely, idk why it have namespace
obsidian mantle
so llm doesnt try to communicate with "choose_string_choice" for example (now renamed to choose_character_string_name_dont_use_unless_asked_to)
but actually if its registered only when needed
will not be a problem
same with other force commands i have
glass jetty
@obsidian mantle idk why i tried to see it as cpp
it is c api, not cpp, in this case you would wrap it into c++ class and expose only stuff you need via proper methods
try to search for actual cpp api for neuro sdk first
cuz it is surely c api
obsidian mantle
idk i just stole it from cyberpunk integration 
fickle rain
i need to make new function for this class
FYI the socket URL gets autosetup to value of some env var with libneurosdk
You don’t need to set it by hand
obsidian mantle
i think i just got confused with cpp again and now it all works
new function within class or something like that
obsidian mantle
wtf is this im scared
const char **action_names;
pointer to array of const char *?
are they separated with \0 or
rough bloom
pointer to array of strings most likely
don't need to be separated with \0, the pointers have a known size
the strings don't though, those should still be null-terminated
opaque sigil
array itself might be null-terminated too
unless they give you the length too i guess
obsidian mantle
i see i see so its just an array of pointers to const char*s
opaque sigil
if you're ever confused about some of the truly vile declarations there's https://cdecl.org/ or https://github.com/paul-j-lucas/cdecl
obsidian mantle
its a library for convenient integration development
rigid snow
hey
checking in
i'm ok
baii
just remembered i said some things to some here that might worry them because i stopped appearing
obsidian mantle
fiery anchor
sage crag
phrrr
nocturne olive
Restarting Discord be like (it was using 14GB alone)
sage crag
<:neuroAYAYARRR:1133771115651674293>
loud salmon
Does anyone know what will happen after you Windows + Ctrl + C
umbral wigeon
Relatable developer at night, staring at gigantic code
loud salmon
Relatable developer at night, staring at gigantic code
in the basement 24/7
maiden geyser
595.71
sick owl
595.71 <:neuroHypers:1135051812722982922>
Still borked on the 4000 series
cosmic sphinx
aight so..
i randomly decided to test gemini 3.1 pro on our informatics (CS) govt exam (the one that all HS graduates take to get into college/university)
it fucking passed with 95/100 score, 27/29 tasks completed, only 2 failures
are all CS graduates actually fucking cooked bro
it might aswell ace the exam when 3.5 comes out
gg wp
sick owl
aight so..
i randomly decided to test gemini 3.1 pro on our informatics (CS) go...
Saw your message about 3.5 4B, does it actually stack up to the bemchmarks though?
I could believe it with the 27B and 35BA3B since that's still a lotta params
cosmic sphinx
it might aswell ace the exam when 3.5 comes out
gg wp
^ and I didnt even give it direct access to the provided files, only gave the description to them given by the official exam text
maiden geyser
aight so..
i randomly decided to test gemini 3.1 pro on our informatics (CS) go...
informatics oge is easiest on earth even&oe
cosmic sphinx
Saw your message about 3.5 4B, does it actually stack up to the bemchmarks thoug...
need testing
sick owl
I could believe it with the 27B and 35BA3B since that's still a lotta params
And I personally tested em and found them good
cosmic sphinx
informatics oge is easiest on earth even&oe
thats ege
oge would be childs play for this thing 🗿
dont think theres a point in testing it in either physics or maths, as its still limited by the same image recognition issues, analog meters etc etc
i might acc try downloading the 4B qwen model later and give it a try
nocturne olive
And I personally tested em and found them good
You know LLMs, any good LLMs that would perform better than Nvidia Nemotron 30BA3B?
-# Xod's school project which is about making an LLM teach people using provided school materials and whatever and that's the LLM I'm providing for her right now
(has to be something in the 30B or so range for it to fit on my GPU)
quick condor
Not a huge fan of 3b expert models. honestly have been having better results with dense models
nocturne olive
So far it's worked just fine and it's also really fast
quick condor
Define fast
nocturne olive
190 t/s
quick condor
Is that sustained prefill or generation? Also quality is going to be questionable at 3b active
nocturne olive
It's the generation speed
There's also not been any major obviously model quality issues
mighty thorn
You know LLMs, any good LLMs that would perform better than Nvidia Nemotron 30BA...
MoE 
mighty thorn
I see no issue
Though I'm also no LLM enthusiast
The issue is that a 2 expert MoE is like asking someone to double the horsepower of your car and they just buy and identical one and weld the two side by side
nocturne olive
Well technically with the car it works, as long as you run both engines at once and link the steering systems
leaden crest
Well technically with the car it works, as long as you run both engines at once ...
not in parallel parking it wont
nocturne olive
Either way I see no issue with MoE because it's not just the same weights over and over again
quick condor
It faceplants on cot
nocturne olive
Hem?
quick condor
Chain of thought. 3b active is aweful at cot. The model essentially will always confidently produce the best output without considering alternatives that are actually more relevant to the question at hand
true hemlock
Chain of thought. 3b active is aweful at cot. The model essentially will always ...
sounds like temperature issues
fast pagoda
You know LLMs, any good LLMs that would perform better than Nvidia Nemotron 30BA...
I like glm4.7 flash, oss 120 (too big but good for retrieval still, tends to run very well even without much in vram but a lot in ram), qwen3 coder next (maybe too big but worth trying if you can fit it with less offloaded layers as it's really good), and devstral 2 small. Coder next is best imo for local stuff, not a thinking model but if you can fit it even at q4 etc it is very very good
Haven't gotten to try 3.5 yet
nocturne olive
I can't afford to put anything onto the CPU
There's too many mem leaks in Discord and web apps
fast pagoda
Wha
A good amount of that is probably just electron//chromium using the space cuz it's there
nocturne olive
This is meant to be a model I can just chuck onto my 3090 and forget about while Xod does her thing
fast pagoda
But idk
sick owl
You know LLMs, any good LLMs that would perform better than Nvidia Nemotron 30BA...
You looking for speed or smarts?
nocturne olive
A good amount of that is probably just electron//chromium using the space cuz it...
Eh, 14GB doesn't sound like using memory for the sake of using memory considering it also lags like stupid
nocturne olive
You looking for speed or smarts?
I don't know I like it being at least decently fast but it shouldn't be completely incompetent either
nocturne olive
Holy
sick owl
I don't know I like it being at least decently fast but it shouldn't be complete...
Qwen 3.5 27B
If that's too slow then use 35BA3B with inactive MoE layer offloading
fast pagoda
i still like gemma3 27b 
for style stuff
for fine tuning
dont really use it for anything else ig
sick owl
Qwen 3.5 27B
Easy recommend imo, absolute SOTA in that size class
fast pagoda
echo is gemma3 27b lolo
sick owl
Easy recommend imo, absolute SOTA in that size class
How much system ram do you have though because that can change things with tensor offloading
fast pagoda
he apparently ownt put any layers in ram
sick owl
he apparently ownt put any layers in ram
Then Qwen 3.5 27B is a no brainer
fast pagoda
i need to grab 3.5 rn
need to yolo an echo with it for fun and profit
i see that liquid put out a lfm2 24b (2b active) in the last week
sick owl
i see that liquid put out a lfm2 24b (2b active) in the last week
Its more or less worthless next OSS 20B sadly
fast pagoda
damn
nocturne olive
Qwen 3.5 27B
Wait this thing has image-text?
fast pagoda
oss 20b is p good at that size hard to beat that makes sense
sick owl
Wait this thing has image-text?
Its multimodal yeah
fast pagoda
Wait this thing has image-text?
qwen3 VL also has this in this size class but i imagine 3.5 is better now
nocturne olive
Well how does that work with llama.cpp?
sick owl
qwen3 VL also has this in this size class but i imagine 3.5 is better now
Qwen 3.5 27B mogs Qwen 3 VL
fast pagoda
poor qwen3
nocturne olive
Well how does that work with llama.cpp?
How does one use such stuff from the API?
fast pagoda
i still preferred 2.5 over 3 unfortunately
although qwen3 coder next is really good imo
idk how 3.5 compared to that
coder next is in 70b class tho
sick owl
Well how does that work with llama.cpp?
Bartowski and other GGUF repos have an FP16 multimodal projector as a separate file in the repos, you add it to the llama-server launch command with --mproj pathtofile
nocturne olive
Wa
fast pagoda
80b*
nocturne olive
But how do I programatically use the multimodality?
How do I call it from the OpenAI compatible endpoint?
fast pagoda
Qwen 3.5 27B mogs Qwen 3 VL
have any thoughts on 122b A10b?
im aboutta try it
i want to try 397b (a17b) but it's a bit too thicc
122b moe should be an interesting mogging of oss 120 if it's already losing to the smaller models
nocturne olive
I assume this thing has tool use as well (that is also a requirement I forgor to mention)
fast pagoda
it can
sick owl
How do I call it from the OpenAI compatible endpoint?
Was getting to that, I was gonna do an overcomplicated explanation but its easier just to say same way as you would for the OpenAI API
nocturne olive
Well I don't use any other OpenAI API than llama.cpp (because others cost money) and I have never used a multimodal model
fast pagoda
Well I don't use any other OpenAI API than llama.cpp (because others cost money)...
vllm
nocturne olive
llama.cpp
fast pagoda
3.5's model card has this for vllm
vllm serve Qwen/Qwen3.5-27B --port 8000
--tensor-parallel-size 8
--max-model-len 262144
--reasoning-parser qwen3
--enable-auto-tool-choice
--tool-call-parser qwen3_coder
ollama itself (ollama run) will automatically do this
which uses llama ont he backend
sick owl
Well I don't use any other OpenAI API than llama.cpp (because others cost money)...
I find the easiest way is to convert to base64 then pass as image data
Lemme get the docs up for you
nocturne olive
How majorly would I have to modify this?
./llama-server -dev CUDA1 --port 8492 -c 131072 --context-shift -m ../../custom-models/Nemotron-3-Nano-30B-A3B-IQ4_XS.gguf --temp 0.7 -ngl 999999999 --special --verbose-prompt --jinja
the worst part about images to the model is the context
sick owl
Lemme get the docs up for you
sick owl
How majorly would I have to modify this?
```bash
./llama-server -dev CUDA1 --por...
If you're quanting down to 4 bits you don't want to be using vllm imo
nocturne olive
Well I kinda gotta go that low if I want models to fit
sick owl
How much context do you actually need?
I fit 27B in my 24gb vram at Q6 10k context on llama.cpp
nocturne olive
I have no idea but it's better to have a lot of it because I don't know what Xod is gonna do with it
sick owl
I fit 27B in my 24gb vram at Q6 10k context on llama.cpp
Dropping down to Q5 I can fit 40k context
fast pagoda
How majorly would I have to modify this?
```bash
./llama-server -dev CUDA1 --por...
-m Qwen3.5-27B-Q4_K_M.gguf \
--jinja \
--reasoning-format deepseek \
-ngl 99 \
-fa \
-c 32768 \
-n 32768 \
--temp 0.6 --top-k 20 --top-p 0.95 --min-p 0 \
--no-context-shift \
--host 0.0.0.0 --port 8000```This 30B uses only 18GB mem a this quant and context
sick owl
have any thoughts on 122b A10b?
Dumber than 27B but with far more knowledge, however ironically it actually hallucinates more
sick owl
This 30B uses only 18GB mem a this quant and context
27B uses around 24GB at Q5 with 40k context
Otherwise you really should look at offloading inactive MoE layers to RAM
sick owl
I use the 35B at Q6_K_L with 40K context at 60 tokens per second using offloading
sick owl
Then your OS is taking some of the VRAM
True, I was working off the assumption your setup was similar
nocturne olive
I have a dedicated GPU for rendering and ML separately
sick owl
I use the 35B at Q6_K_L with 40K context at 60 tokens per second using offloadin...
I stick to the 27B though since my use case is single turn anyway
nocturne olive
4070Ti + 3090
fast pagoda
wouldnt vllm be better for 2 cards
nocturne olive
I only use 1 for ML
The other is exclusively for rendering
And gaming
That way whatever happens on the ML card doesn't affect my everyday use of the PC
fast pagoda
i know that feel
sick owl
I have a dedicated GPU for rendering and ML separately
In that case you could probably fit the 27B at Q6_K_L and long context if you ran with tensor parallelism
sick owl
What is a tensor parallelism?
You can run models across multiple GPUs on setups like that and still have plenty of power left over on the lesser utilised one
nocturne olive
I don't want to split across GPUs
I need room for NeuroSynth too
I like my models Q4, very great VRAM efficiency
sick owl
If you're insistent on it fitting 24GB VRAM then run the 27B at Q6, drop down to Q5KL if you find yourself bumping into the max context you can squeeze in
Q4 isn't worth it on a lot of these more knowledge dense models, you begin to run into pretty significant degradation below Q5 nowadays
nocturne olive
I'm gonna guess I need to recompile llama.cpp first though either way
sick owl
You do yeah
fast pagoda
do
sick owl
o7 CPU
sick owl
Eh 12700K can handle it
Oh yeah but no CPU enjoys building llama.cpp
sick owl
Eh 12700K can handle it
Grab bartowskis Q6 and Q5_K_L GGUFs and F16 mmproj
And recommended sampler settings can be found here https://huggingface.co/Qwen/Qwen3.5-27B
nocturne olive
The Q6 based on file size is absolutely not gonna fit
fast pagoda
when you load 122b with old llama.cpp it is not happy
sick owl
The Q6 based on file size is absolutely not gonna fit
I fit it on my only GPU (4090) with 10k context
And 1GB free on top
But if you need plenty of context go Q5KL
Its a quality hit but not an unbearable one
nocturne olive
All that really matters is that it performs as well or better than the old 30B model
sick owl
All that really matters is that it performs as well or better than the old 30B m...
It absolutely mogs that model rest assured, go Q5KL for reliabilities sake though imo
fast pagoda
I fit it on my only GPU (4090) with 10k context
did you see the qwen3.5 with opus 4.6 reasoning distil
sick owl
It absolutely mogs that model rest assured, go Q5KL for reliabilities sake thoug...
With no OS to worry about I'd wager you can fit like 60k+ context
sick owl
did you see the qwen3.5 with opus 4.6 reasoning distil
I did but it seems a bit redundant to me
Qwen 3.5 is already a far more professional distill of Opus data
fast pagoda
it already was that 
yeah
i just saw that the benchmarks seem to have improvement in like
terminal bench
some benchmaxxing improvements really
dunno real world
sick owl
It probably benches better but I bet its way more unstable
fast pagoda
its stated purpose is to be less unstable which is why i'm interested since in qwen3 experience it was very prone to looping
im sure 3.5 is already much better at that
so idk
fast pagoda
the most fun to look at
i will have a system with 136 cores// 544 threads here at some point in the near future, i'm excited to see the longest core list of all time
sick owl
its stated purpose is to be less unstable which is why i'm interested since in q...
They say that but excessive rigidity (which is what a lot of these tunes achieve) is its own form of instability imo
fast pagoda
yeah you're not wrong
it just depends on how they executed really end of the day
could benefit from additional sft
could not
watching 122b reason throuhg something right now and its logic is already extremely structured ngl
sick owl
Circling back to that actually, one benefit of 27B is that its pretty concise in its reasoning where I found 122B is a bit yappier
fast pagoda
it's structured but it is a long ass thinking trace
sick owl
I got similar answers to hard questions in 10K tokens max from 27B where 122B took a couple thousand more
sick owl
I recommend their "precise logic tasks" preset if you don't mind it thinking longer, I find it gives a much smarter presentation
sick owl
I recommend their "precise logic tasks" preset if you don't mind it thinking lon...
One tip, sometimes the model almost seems like its looping but its actually checking to see if it hallucinated, after a few iterations it usually answers
fast pagoda
general tasks recommend 1.0 temp so that's good
sick owl
general tasks recommend 1.0 temp so that's good
I don't like the general tasks preset but if you're not really putting the models smarts through its paces its fine
And it is more concise thanks to that presence penalty
fast pagoda
Thinking mode for general tasks:
temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
Thinking mode for precise coding tasks (e.g., WebDev):
temperature=0.6, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0
Instruct (or non-thinking) mode for general tasks:
temperature=0.7, top_p=0.8, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
Instruct (or non-thinking) mode for reasoning tasks:
temperature=1.0, top_p=1.0, top_k=40, min_p=0.0, presence_penalty=2.0, repetition_penalty=1.0
For supported frameworks, you can adjust the presence_penalty parameter between 0 and 2 to reduce endless repetitions.
fast pagoda
I don't like the general tasks preset but if you're not really putting the model...
in this case im asking it to rewrite a song
sick owl
I don't like the general tasks preset but if you're not really putting the model...
Slightly higher hallucination rates though naturally
sick owl
in this case im asking it to rewrite a song
Yeah it'll be fine for that
nocturne olive
So how do I use the weird little extra thingy again?
sick owl
So how do I use the weird little extra thingy again?
Throw it into your launch command with --mmproj [path to file]
fast pagoda
if you just put it in the same dir you're in it's probably in the repo dir --mmproj mmproj-F16.gguf
nocturne olive
--mmproj ../../custom-models/mmproj-Qwen_Qwen3.5-27B-bf16.gguf
Like that?
sick owl
`--mmproj ../../custom-models/mmproj-Qwen_Qwen3.5-27B-bf16.gguf`
Like that?
You want FP16 not BF16 generally but yup
nocturne olive
I heard BF16 is usually better than FP16
fast pagoda
for training
nocturne olive
Does it suck for inference somehow?
fast pagoda
it's more stable for training
it's not needed
for inference
bf16's advantage is that it's better with large gradients and such which makes it better for training
sick owl
BF16 is fine tbf, but I encountered some weird bugs using it in the past
fast pagoda
it just sacrifices accuracy in small
contexts
shouldnt use contexts there given its meaning already is set for ml
sick owl
If you keep an eye out and recognise that as the cause if they do occur you're probably fine @nocturne olive
BF16 will be slightly slower and marginally more accurate at similar memory costs since you're on a modern GPU
fast pagoda
for forward passes, bf16 is not more accurate in general, it loses mantissa bit precision for exponent bits (to get the same range as fp32 for that) so your gradients dont explode
err
nocturne olive
Downloading models on 100Mbps is so slow
fast pagoda
bf16 is not more accurate
fp16 is better technically for forward pass because there's less rounding error
long sequences can suffer comparatively in bf16
it's NOT A VERY BIG DIFFERENCE tho
sick owl
bf16 is not more accurate
Qwen 3.5 is natively BF16 so I was under the impression you do lose a little accuracy in the conversion process
fast pagoda
well if that's the case then nvm it's a bit different
i'd run bf16
unless my hardware doesn't support bf16 obv
fp16 was more commonly the native format for a while so that's a bit of an antiquated thing now ig
i mainly cared about this when i was training stable diffusion vs running inference for it which was natively fp16 and bf16 would cause issues with fine detail
sick owl
I have actually heard that manually setting context cache to BF16 might reduce looping with Qwen models on llama.cpp
I don't know whether that would hold up under scientific testing though
Given llama.cpp uses FP32 accumulation for context anyway
fast pagoda
is flash attention broken with 3.5
that tends to be what happens early half the time i just havent checked
sick owl
is flash attention broken with 3.5
Its broken if you switch to BF16 context yeah, so I stick with FP16
I figure the speed loss isn't worth the edge cases (if there even is a real issue)
fast pagoda
sick owl
Oh wait hang on lmao, that means the F16 mmproj is the one you want after all
D'oh
@nocturne olive I did a dumb
nocturne olive
How weirdly silly
fast pagoda
too late he's already sat there like "this model fucking sucks wtf"
mighty thorn
Mildly unpopular opinion: PDF is the worst file format ever
mighty thorn
Downloading models on 100Mbps is so slow
fast pagoda
i hate it when my steak is too juicy and my lobster is too buttery
idle dune
I put it into Claudes memeory to argue with me and drop agreebility bias when I am engaged in work projects. Now he's all sassy and destroying my legal arguements.
fast pagoda
best instruction follower
sick owl
https://tenor.com/view/lobster-steak-too-juicy-sarcastic-shrug-gif-1224159454409...
I'd consider 100mbps to be on the slower end by uk standards at least
fast pagoda
100mbps was the first "fast" internet we had hwen i was a kid
that was in like 2010
it is the baseline for what i'd consider usable for anything more than phone tier browsing
nocturne olive
100Mbps is literally the lowest option here
nocturne olive
Eh, 1Gbps would be better value
fast pagoda
lol why
fast pagoda
Eh, 1Gbps would be better value
that's how it is here
sick owl
100Mbps is literally the lowest option here
75~ is the lowest in my area
fast pagoda
verizon asked me if i wanted FiOS and i was like sure ill see what your fiber speed is on offer
and these mfs had the audacityu
they dont have fiber here
they offered 25 mbps LTE
home modem
bruh
on the plus side xfinity is apparently offering 10g service here now i just have to wait for them to procure an xb10 (docsis 4.0 modem)
cant get those anywhere else so that's fun
fast pagoda
q5 was unhappy with long context for me w/ 32gb vram so that checks out
without any offloading anyways
q6 i mean
sick owl
Q4 it is
If you really need super long context by all means
Game of tradeoffs like anything in computing
It'll still outperform the 30B
fast pagoda
2nd go with the recommended general thinking params = 4491 tokens 
fast pagoda
agentic harness
sick owl
What’s this OpenClaw thing someone told me about?
A security hole you willingly install
rigid snow
you should give it access to your whole life to let it ruin it for you
silk jackal
That’s what I was suspecting
fast pagoda
like swiss cheese
rigid snow
the holes ARE the value prop
fast pagoda
it's running up against the problem
of agents
they cant be useful without the holes
without a lot of manual intervention
silk jackal
I heard nobody’s overseeing the GitHub repo 😬
fast pagoda
he got hired by openAI so who knows
rigid snow
the holes ARE the value prop
it's basically "what if i give ai access to everything and the ability to do anything?"
as in, delete your whole gmail inbox
fast pagoda
it's fine there's no way an agent might come across some adversarial prompts on the internet surely not
or just go off the rails
rigid snow
as in, delete your whole gmail inbox
or message your ex you miss her
fast pagoda
he's just tryna helpp
rigid snow
most do i would imagine
fast pagoda
probably knows there are flaws
might not know the extent or risk
also the software itself has a shitload of its own vulnerabilities
it's not just the agent going berserk
the docker container had like 2000 vulnerabilities or something dumb
Reddit
silk jackal
I’ll just keep going down the road I was already on.
ftr I was the guy asking about an AI chat buddy for streamers who aren’t used to talking live
fast pagoda
I’ll just keep going down the road I was already on.
ftr I was the guy asking ab...
that can be done many safer ways
silk jackal
that can be done many safer ways
Ho ho yeah
sick owl
the docker container had like 2000 vulnerabilities or something dumb
Vibe coded LLM powered sysadmin at scale, what could go wrong 
silk jackal
I believe that
rigid snow
top comment 
is that referring to the prediction market / crypto trading bots people are setting up
fast pagoda
web3 baby
rigid snow
fast pagoda
but my nft newsletter that no-one reads is automated and generates intrinsic value (somehow)
rigid snow
https://fxtwitter.com/durov/status/2028455440862830970
i don't know any openclaw users is this an accurate depiction of how they type?
fast pagoda
yes
wow the agentic future is incredible
gimme gimme gimme
somehow i dont see how this is unsustainable, just tell the bot to build me wealthy 
well done
entirely helpless without asking grok in the replies too
epic
prompt engineers
silk jackal
Good news (?)
The person is aware of Open Claws flaws
fast pagoda
isnt your 30b an a3b
27b is all loaded in vram in the dense model
for the 30b it's loading not nearly as much
it only has 3b active paramters
nocturne olive
Hooly mem usage
Q4 barely fits
Ok it doesn't really fit NeuroSynth won't run
fast pagoda
offload like 2 layers to cpu
nocturne olive
No
fast pagoda
i knwo you dont want to but that's like nothing
nocturne olive
I'll just drop a few CTX I guess
Looks like NeuroSynth needs 2GB VRAM at inference time
fast pagoda
wouldnt you just run that on your other gpu
nocturne olive
Taking off 16K CTX from 128K is enough to get it running
fast pagoda
you must be bigly gaming at all times
nocturne olive
I'll see if I swap to this model or not
This one's way more annoying to deal with VRAM wise
fast pagoda
what about the 3.5 equivalent of your 30b?
the dense model is gonna be more resource hungry which is obviously a concern this one should be very much more similar to the 30b with 3b active
nocturne olive
Well what does the @sick owl think?
fast pagoda
with 3.5 35b a3b loaded entire on to GPU and set to 128k context, im using 27.9 gb vram with it loaded
and getting 67 tok/sec dayum
fast pagoda
well this is with 128k context
nocturne olive
Well 128K is what I was running Nemotron 30B at
fast pagoda
i have a few things enabled that probably increase vram hold on
25.5 with kv quant welp idk it's just a bit bigger than 30b cuz 5b more i guess
i still think you're overestimating just how much offloading like 2 layers to system ram would harm your overall pc experience, probably significantly less than if one of your GPU is sitting at 100% vram
nocturne olive
I want the model to be completely invisible to me when I use my PC
Otherwise I can't exactly keep it up for Xod more than absolutely needed
fast pagoda
i think it'd be more invisible than having a full gpu tbh
when one of mine is maxxed out it causes the rest of the pc to be a bit less happy presumably due to bumping up against the point it might need to use some system ram but not quite
nocturne olive
what about the 3.5 equivalent of your 30b?
I used an IQ4 quant by the way instead of a regular Q4_K_M
It's a bit smaller
fast pagoda
with 5 model layers offloaded the mem use of lm studio is 22gb on vram and 3gigs in system ram
it's pretty nothing
but that also would probably help idk if there are any imatrix quants of this yet but people are usually fast
closed vivaldi kekw
didnt even have much open
went down to 37 tok/s with 5 layer offload & 128k context 22.1gb vram used
lm studio in the interface will estimate vram/ram usage as you adjust parameters before loading which is nice
nocturne olive
I just raw terminal llama.cpp
fast pagoda
yea i mean that's what i do for vllm
but for this purpose
rather than loading/unloading over and over to see
not that it takes forever
fast pagoda
isnt it just based on when you first used it that week
er the last week
i dont use codex much
fickle rain
Well the ratelimit reset for me was on March 5
I checked it a few mins ago and it reset to 100%
fast pagoda
i found some threads of that happening to people randomly a few months ago
maybe they just felt like resetting it "as a gift"
oh wait
fickle rain
Attempt to retain userbase after uh, misadvertisement?
fast pagoda
gpt 5.4 is out in codex randomly
in the last hour
seems like they may be blasting 5.4 codex out and reset it for that + they were having issues when doing so
fickle rain
Probably some A/B testing stuff
fast pagoda
it is
the model name is literally
ab
lmfao
interesting that some info can be guessed from this name string
fast pagoda
agi
?
fickle rain
OK maybe if mister Altman routes this Codex instance to GPT5.4 it can solve this crypto challenge
Lol my 2nd acct got kicked out of verified?
fast pagoda
5.4 - obviously new version incremented from 5.3, still gpt5 pretrain
ab - a/b test there's prob another variant or 2
arm1 - seems like arm arch but idk what arm1 is unless it's talking about some sort of branching a/b with "arms"
1020 - october 20th? maybe the original checkpoint date, or 1020 steps, or something else. original gpt 5.1 announcement was nov 12th so i could see this being the first 5.x training run date or something
1p - 1 pass? something with precision?
codexswic - codex line, swic could be a few things, maybe software in context or something if i schizo, or just switch
ev3 - eval 3?
oo found a 5.1-ab-arm3-840-1p-codexswic-ev3
arm1/arm3 arent arm arch then i guess
1020/840 probably not dates
maybe thye are context length
1020k vs 840k
fast pagoda
> arm1
https://www.reddit.com/r/codex/comments/1rjcwli/usage_limit_reset/ seems like there was an error with 5.3 that was fixed
seems liek possible deepseek v4 imminent
fast pagoda
@sick owl interestingly enough with both 122b and 35b MoE, the thinking trace making any sense or focusing on things that matter for the task is inversely correlated with the final output quality
the best final answers came from the unhinged thinking traces that spent all their time on nothing of value
true hemlock
fast pagoda
fast pagoda
<a:pogs:1100369978324156496>
is that 1u/2u atx?
fast pagoda
true hemlock
3.2
fast pagoda
on amazon, $12
depending on level of sketchy unsurity about getting what it's listed as, as low as $4
on other sites
those seem not real tho
idk
i should buy one to see
theyre free anyways
fast pagoda
that's abouit the verified price i could find
true hemlock
there are a lot of overpriced $20+ ones
fast pagoda
yeah there are
true hemlock
in fact i just got one for $10
because i realized all my usb drive are 32gb
so why not ig
fast pagoda
holy
too small
well, i used those kinds for lik
bios
and small transfers ig but if i am putting something on a usb stick it's usually something big
true hemlock
for bios most 64MB-128MB already works lol
true hemlock
i do
true hemlock
the one my father years ago got