#programming
1 messages · Page 422 of 1
it was uploaded in oct 2024
so there were a couple video models around that could've maybe pooped that out but wouldnt have been as easy as today
went sleuthing, the void stared back 
around that time we had like kling 1.5
which would not have produced that i do not think
That's crazy how the 27B is almost as good as the 122B and not far from the full size
i just did this
can confirm i do this
@rigid snow
you didnt say it a third time
what is their name
probably fairy
is the fairy french
what if the language uses a different script

it wouldn't be "fairy" exactly
mlntcan
it may even sound different
yes that me
not for long
account stolen by order of fairy the fairy
whos loam
mlem
uuuuuuh getting it is taking some time you must understand
loam it to you
...sure
here you go 🪙
cryptocurrency
not

just token
is a token currency necessarily
no
a token is used for exchanges but a currency has the requirement of being reproducible which a token doesnt

for example
for example
my brain is a token
can be used for exchange
but cannot be replicated (yet)
probably
the knowledge that you gave your token away
high value

What is that gif
konorb enub orbit
the transformers in my brain trying to decode those words in the morning
dense model thing, it's the largest non-MoE
generally have seen consensus for a whiel that to estimate what the dense equivalent of an MoE is, you do geometric mean aka the sqrt of total params x active - for the 397a17, that's sqrt(397 * 17) or ~=~82B
for 122b, that is sqrt(122 * 10) ~=~35B
that checks out to me, given MoE weirdness causing some losses along the experts domain covg or something. routing errors, expert load imbalance, incomplete specialization
MoE tends to have a better showing at the lower end, if you do it for the 35Ba3B, geometric mean on that is about 10b but that one is comfortably ahead of the 9b in comparison
wrr
i did villager asmr with her in a voice call once
im American though
tutel is terminally british 

You have unlocked new role
You have unlocked new role
You have unlocked new role
wow, but what role
imag
role
you can embed
omg wow
could you pick a better message to unlock new role on
;_;

You have unlocked new troll
yoy
was there an attempt though, was there really
in my head
me when any pop garbage from 2010 i used to shit on comes on
it took 15 years but they finally got me
my music taste would give someone a stroke
though im editing the vocals of 4 streamers rn for a song so i cant listen to music atm

i used to say that but i think my tastes have converged on a more stable place now
i used to confuse people because all i liked in like middle school was
rush
mine is still insane
i hope i have cpu on the way
my CPU got destroyed last music project
Its at the post office. Im in class tho 
this time i moved to cubase in hopes that it runs better
i mean it's all over the place still. maybe i have just desensitized to it
Lop
Lop
no struggle yet
dragonfruit
thankfully
Lop
samuel vanuel manuel
True!
@sage crag
@sage crag

educated

literally what i do
i got music from games movies tv shows anime normal music old music covers
etc
SAME
i dont vibe with alot of common music but when i hear something great it becomes an ear worm for weeks
i like the recaps and stuff at the end of the year
because itll be like
you listened to this track 713 times
and itll be something i rememebr was like a week
lmao i usually use youtube so i dont have that but i imagine id be in a similar boat
they started doing a youtube year end thing :o
i think i missed that
it was on the you page but i cant find the damn thing
youtube doesnt know what the hell im talking about, theyve never heard the term recap in their life apparently
they can see when i was trying to figure out how to find the discord one since my buddy wasnt able to see it
but they have no record i ever watched my own on youtube itself

that was the one i didnt need to be reminded
because i sent them this from mine
and they were flabberghasted
apparentlly this is a few messages
damn
my spotify wrapped, the top song is literally always "cyberstrike" by shirobon, and that is because that is my alarm
because it's so loud
lmao i cant imagine a song waking me up
for years now
lmao
the reason mine is cyberstrike is because it starts super abruptly and i used to not be able to get up in like hs
so i set up an alarm clock on my pc
and it would set my speakers to maximum volume
and then start playing cyberstrike
and it would shake the entire house
could not stop it except by getting out of bed because i had a loft
and the computer was below
and usually i sprinted out the mf to avoid getting murdered by the rest of my family
i could be a fireman
lmao reminds me of this alarm clock i saw back in the day that had wheels on it
it would roll off and play a loud sound around ur room so u had to catch it
this one
my dad uses that exact one but in white
glad to hear it works
but he disabled the rolling so just uses it as a normal clock 
lmaoo
now i can only stay in bed like 4 hours anyways
it sounds like a dying r2d2 when ringing
I filmed this and made the beat. Thank you for subscribing and sharing this with your friends
-- Listen to the extended version: https://www.youtube.com/watch?v=L_1fTk93qAo
-- Follow me on SoundCloud: https://soundcloud.com/itsguamu
-- Follow me on Twitter: https://twitter.com/itsguamu
-- Download on Bandcamp @ https://guamu.bandcamp.com

reminds me of this old video
Taken from https://www.youtube.com/watch?v=eev9tyePDnc
Edit: thanks yall for a million views
https://www.youtube.com/watch?v=MIgMeyW4TEM this one as well
Triple concerto for faucet, water pipes and fiddle. Best funny music video and everyday objects
Jacek Dzwonowski - violin & faucet
Aleksandra Marko Lech - camera ;-)
Full version of this tune for string quartet - here from only ;-) 1.99 €
- ➡️ WATER MUSIC https://altravolta.gumroad.com/l/watermusic
- ➡️ DARK SIDE https://altravol...
worst thumbnail
its certainly up there
man
cisco opus #1 10 hours got removed from youtube
i unironically listened to that entire video probably 50 times at least
ok but some guy has done a 1 hour vip megamix
lmfao
Extended version of the 2016 remix by A.P. Laurenson. Based on the composition by Tim Carleton and Darrick Deel. This cut was inspired by William Basinski's The Disintegration Loops.
(00:00) You're Here Forever
(03:17) Letter From Mary
(10:15) Save Ferris
(14:55) Cicely (Snuggle Up To Your Fiberoptics)
(23:50) Ghost In The Shell / Dust In T...
i hate how much i love opus #1
it's likke my cia activation tone and i become pacified when hearing it
opus playing on the opus codec
it predates the codec by 20 years
he just knew he had peaked and named it opus
then optimistically named it #1
when #2
based
27B remains the undisputed king imo
Very nice, my old 360Hz Asus had the same
I think my current OLED might have too but I don't remember since its less impressive the second go around
I keep seeing it out performing or being very close to like Claude 4.5 , that's crazy
I guess it's common for monitors that advertised for their color performances
I was sad when qwen3 was meh

I used 2.5 for the first echo and then 3 was just not gonna happen based on the runs of training* I did complete
Big number but also small number? Wa

or like
what is big signal
high amplitude
powerful
or probably actually not
-fs makes this interesting
nah no way intersample peaks do 30db
Reduce the overall volume but increase you byhe specific bits
peak af
no way ppl choice 1 box
that just stupid
https://www.youtube.com/watch?v=Ol18JoeXlVI
Two boxes, one choice, and $1,000,000. Sponsored by Brilliant - To learn for free on Brilliant for a full 30 days, go to https://brilliant.org/veritasium. Our viewers also get 20% off an annual Premium subscription, which gives you unlimited daily access to everything on Brilliant.
If you’re looking for a molecular modelling kit, try Snatoms...
Discord broken on desktop only for me?
only you
Mobile seems to not be showing me much backlog either
Altho the current messages work
it's broken for me
Welcome to Discord's home for real-time and historical data on system performance.
google wanted to suppress my post about gemma4 maybe soon
they dont want @sick owl to be told things he already knows by me for no reason
yeah, if i sent an image it just flies off into the void

they took the bananer off the image gen option in gemini
this means they officially done ruined everything


The lightcrafter 4500 hath arrived
its small
really small
i know i had the measurements even before buting, but it still feels so small
with shipping and import tax, close to 900 euro 
i opened the box witht hatr knife
setup
i cant power it yet
@silent anvil
i need to solder the usb decoy board into 28V mode
why patchelf brokege
next time try to checkout stable branch
next time try to checkout this branch

i thought discord was broken again cuz i couldnt send the emote, but my keyboard was still in the pi
konii
usb splitter
buy what?
just type to both
no

i just use ssh after making the user account
no one else on the entire internet has the same problem as me 
im not submitting an issue
it worksn't on my pc
no osu until someone else figures this out
b ut seriously
missing dependency ye, but name.name
because someone broke it
it was updated last in january, so why did it only break in march
code that it depends on?
no
the dep.name.name is correct
issue is this probably
dep_name = dep[0] if len(dep) == 1 else f"any({', '.join(map(str, dep))})"
format string
not path
looks like that line was originally fine and whatever someone did 2 months ago broke it
ye, though that commit also shouldn't have changed anything
that kind of reason why this ~ exist
mabe that code path was just never triggered somehow
either way string is wrong there
only python utility i using is acpuf, cuz there no alts
I probably would
but
can't reproduce issue

davinci resolve also doesn't support h264 and 265 on linux for free, and it sucks 


can I just
grep all Hydra build logs
Wait what
I don't think I can 
i really wasnt confident they were gonna do a gemma4
unless this is just for AICore and like Chrome but wont have checkpoints publicly
hope not
npus are decent for image recognition right?
i wonder if the npu on the pi 5 ultra works
since the hdmi drivers are shit im not too hopefull, but it would be cool
they are generally suitable for mini tasks like that but not heavy generative shit

ive not used them but i've seen a few youtubers hook up like camera modules and use an NPU module with a pi and have it do like bounding boxes for people recognition etc
i have 3 MIPI CSI ports for attaching cameras to the pi 
i wonder if those would work at the same time as the hdmirx.
if it's just object recognition then they probably have an example app or something
they dfdo
do
the modules i think come ootb
with a basic setup
ffuukjkkkk
no matrix cores so it's dogshit but man is it tempting me
C, C++, Python
random vendored packages
yum
it took a while, because the pi cant boot from nvme by default so i had to set up the MTD flash. but the pi boots from the 512gb nvme now 
somehow that nvme is 465GB now
reeks of disciplined, careful development
Classic
export TERM=konsole
ive been doing export TERM=xterm-256color
are you sure that it's 512 GB and not 500 GB
because for 500 GB the 465G number makes sense
it says 512GB
(it's in GiB
)
i'm surprised their terminfo does not include kitty, given how long kitty has been around 
i harvested it form a laptop, so it was free i guess
its the laptop my sister broke the hinge of

yes
i don't remember if the optane part shows up separately
its a ssdpeknw512gzl
might be - swap?
or is swap seperate?
i have no fucking clue what zram is supposed to be
phrasing
A 512GB SSD usually shows about 476GB–477GB of usable space. It's the whole hdd mfgs defining GB as 1,000,000,000 bytes
yes
zram is swap
no
They've been doing it for years
it's just the stupid GB vs GiB thing
the physical space is the same it just looks different 
yeah
GB as 1,000,000,000 bytes is technically correct
but pretty often when people say GB they mean GiB 
i sure love the zram2 having 0B of space
and the 50MB of zram1 are sure to be useful 
memory compression
so now that i have nvme working, ill test hdmirx
you have an in-memory block device that is backed by compressed memory
so if you use it as swap the kernel then swaps pages in and out of that compressed region as needed
then you waste precious ram
but then swapfile not compressed
i have one of those
swap-on-vramfs on 3090
cheap RAM expansion
honestly, coudl work
since my ram is going through the RMA process rn, i have 16GB of ram and 24GB of vram rn
cheap ram expansion
cost
rtx 3090
yeah lmfao
cheap
its ok just buy
free if you already have the 3090
i stand in solidarity with lop
many rx 580 gpu
8gb vram cheap cheap
i think since the orange pi 5 has lpddr5, it's ram is better than my desktop's rn
both are 16gb
go console route and dont use ddr
dropped from rocm support
unified gddr6
dam
there's a 32 versoin
how does hbm compare
pci3 bottleneck 
PS6 pro ultra deluxe vram edition, will have gddr8 128GB for only 7800$ 
Well don't let your PC go to sleep then
-# There was something funny about systems with more VRAM than RAM crashing when it goes to sleep or something
bandwidth high
latency also high
everything high 
its not crashing for me, if its asleep for multiple hours it does break display output due to nvidia power management sucking
the mi50 are good for only one thing and that's decently quick inference for small models
due to their mem bandwidth
time to write everything in cuda for unified memory despite not using the gpu
and that's what matters for them
Mabe they fixed that bug
what was that bug suppsoedly in?
I forgor
hypland?
they dont have matrix coresz or support INT8/FP8 let alone fp4 natively
Probably driver or kernel
but they will run qwq well etc
if its not a hyprland/wayland/nvidia dsriver bug, im not affected
if its not hyprland/wayland/nvidia
(99% of bugs removed)
It happened in some of those big server systems with TBs of both RAM and VRAM
I doubt those run a windowing system
Which leaves the Nvidia driver and the Linux kernel
dataset taking forever to download and process
that do be how it goes
altho
i have not had many bugs recently
jsut the occasional one
me the second i rent a b200
big sloppa card 
so its untested if it fixed it
there's another finegrained one you might need but ymmv 
my kilometerage may vary
because evals on gemma327b with it's ridiculous dictionary require like 180gb of vram
instead of waiting, spend a few days optimizing the pipeline
then have it run 10% faster but take 10x longer in total due to dev time 
i should rent a b200 to run benchmarks for my paper but stingy 
just wait till lamda or somene does a special
I already did that 
I'll admit I've been vibecoding the test repo though so anything is able to fail 
h100 + rtx pro 6000 will have to do for showing bandwidth scaling 
lambda rabbit
that's what the current 27b checkpoint was done on
a 6000
it was about to oom tho
stopped it at 75% of an epoch
1 epoch would've been oom for 96gb>_>

Surely it'll be fineeeeeeeee
the last demo repo only failed because I got too cocky and stopped doing full validation runs.
is there no way to see prices on lambda without giving them a payment method or am i blind
oh i am blind
didnt used to be like that
$6.08 an hour for a b200 hmm
$5 yeah
an r200 would be really nice right now so i could really test scaling with its supposed 22TB/s 
I envy your ability to get stuff working fast enough on a rental, I couldn't ever do that. Instead I get 6gb of vram to train on
or whatever they'll name it
you can set up persistent storage
it's like cents per day
and then it's ready 2 go
pain in the ass to still have to pay by the day tho
there's one cloud prov that was claiming to have solved that
but i forgor their name
they werent a provider but basically provided storage for your setup and then could xfer it to whatever prov rapidly cuz dc to dc
10,000 GPU 
i did not know w&b was a provider lol
i assume the runpod instances come with the cuda sdk pre-installed right 
Noted, for now I'm just planning on training small models 
Feel like if I had access to better hardware I'd start yoloing more often.
if i can ever get rocm to fuck off and let me use this stupid card for training without the HIP kernel instantly shitting out NaN
im actually literally right now installing ubuntu 24.04 LTS on a specific portable ssd to use for that purpose as that's the most official setup possible
if i can figure that out then maybe you can use muh vram
if i can expose that over cf or somethign
thought i was just a big dumb idiot but i rented a MI300X from runpod to confirm i was setting up wrong
nope
worked perfectly
instantly
I admit I know exactly what I'd use it for and it ain't training 
I am doing a distill after all
Probably will have one of the most insane synthetic pipelines due to what I'm going to exactly be attempting. And that is after I prove stacked patching to work
well i still have batchgrader sitting around i havent tried using it in a hot minute and it's like 70% powershell so i'd need to uuhhh reconfigure it but i have done it so i can at least be a shitty consult on that if needed
I will keep it in mind, thank you 
Can't believe I'm working on the single largest sidequest for NNv5 in history
the side quest of the side quest
changing the kernel made the pi stop showing display out or connecting to ssh
average kernel swap experience
building a synthetic data model that can generate fast enough to saturate NNv5 when converted to C as python is already at 2000 cases per second, and maybe with some optimization on logic even faster.
(casually also building a very explicit RNN because why tf not, I basically already have the model for it?)
did you modprobe your display drver
probably not
lop
my pc does see it as a hdmi device, so its not completely broken
just no ssh or display out
im unironically swapping to ubuntu 26.04 the second it comes out on my main
idc
rocm in official repo supportd
aasdlahsjdkasd
no SSH 
how no ssh
illjsut swap the kernal back via the sd card
the current image is 6.1.115, the one i found the source code of is 6.1.118. i had hoped it wouldnt be too different
ok im back to the stock kernel, nothing died
i thought it broke
but nah it was just instantly done and remounted as the new volume
that was fun
i mean for this i would imagine they have every possible option yoinked out
if it's not needed
oh yeah the part where you cant see anything or use any devices
my issue before was probably the kernel not connecting to the modules
for some reason
you have to do dkms and shit after sdwapping
well, for that you configure it to use outside modules but modprob will jam em in there
for the stuff you cant
bruh
write caching strikes again
cache go brr yea
i wonder how long it actually takes to get that onto the flash
this one i was expecting to be slower than the other one
mm yes i have written 4GB of data in 0.1s over usb 2.0
please don't unplug your device or i will shit the bed
that scam nvme drive i had a while back to poke at would report reading its entire "2tb" in like
.0034
because the controller was real but hitting literally nothing
big storage doesnt want you to know this one weird trick
alright so im installing this uboontu using my main pc and it is for use with said same pc to make sure i have a clean env for this stupid rocm
do i remove the 3080 from the case before i reboot and install
or do i be slightly lazy and ruin my life for it
just leave it in

it helps the installer also, most installers today autodetect drivers for the plugged in hardware
i dont want it to know about the 3080
i dont want any nvidia ever touching this image
then remove it :V
that's so much work
well, you could get the cli-only installer and manually install everything
that is what im using
well
i have live image
but you dont have to use that installer
it gives choice
if you use the base installer and go through manual package selection, it should help. also avoid linux-firmware package group
and nouveau, and xf86, and so on
oh yeah there's actually an amd provided script
to use
to do a "correct" clean install
i'm gonna be malding when it doesnt work
it's like
extremely recent
but i do expect an explosion somehow
GB DDR5 RAM, Ubuntu 22.04 + vLLM, with 4x Radeon AI PRO R9700 (32GB) GPU. System configurations may vary yielding different results. RPW499.
average config they tested with there
very applicable
yea it's pretty epic
epyc
but they HAVE SAID
that IN UBUNTU 26.04
THEY WILL
have official rocm support in canonical repo
with kernel 7.0
two words never said together
this seems like a pretty simple script, just install the couple things you need and a couple common packages 
you would expect that to be the case yes
but
when you install rocm
a gremlin casts a curse on your computer
i'm low key considering reinstalling my lappy as well 
I wish
I should put a Xeon phi in there
Just to really throw a monkey wrench in
im sure removing nvidia drivers from this arch partition will be perfectly fine and not somehow annoying
bruh
is there a setting i somehow enabled
where i cant use arrow keys on grub
this fucker keeps booting right back to arch because i cant change the selection
can happen if fastboot is enabled
i think i enabled that or the option to always do a certain thing (which is probably fastboot )
it tends to break usb drivers at boot for optimization
yeah it'll likely be fine in uefi since it's an OS and loads its own 🙂
guessing it was the boot to option in this babby's first grub config modifyer t hat comes w/ garuda
there's nowadays multiple fastboot modes, there's one that's "even faster" but then you can't even enter uefi from preboot at that point

glad this is an option
bro i was gonna do kubuntu
kubuntu is on fuckin
plasma 5
ubuntu LTS enjoyers who voluntarily do that to themselves
i fear
kubuntu was my first distro in the most recent round of being on linux
my first distro came on cd from a pc magazine
i mean i was on it on and off for the last 15 years but this time it stuck \
but the first non-silly one i used was fedora 7
and the one that did it was kubuntu
i liekd the old slackware ones
i also wish everyone did a nice movie poster with each release like openBSD
puffy has been through it
well it wasnt fastboot
weird
Been using Kubuntu 25.10 at work. That at least is running Plasma 6.
how to troll your coworkers in a nutshell (fun guaranteed)
kuwuntu...
Man I love the idea of this
what are you trying to do, switching kernels themselves is as simple as throwing the bzImage file and initrd on the boot partition and then pointing the bootloader at them
i guess there are also the kernel modules that go to /lib/modules/$(uname -r)/

its jsut not giving me any display out or ssh afterwards
liek its getting stuck somewhere
does the previous kernel maybe have those drivers baked in and the new one doesn't?
the hdmirx driver is built into the kernel, idk about the rest
you built this kernel yourself right
yes
most drivers will default to being compiled as modules
so unless you also explicitly built and installed those
grep "CONFIG_DRM_ROCKCHIP" .config
CONFIG_DRM_ROCKCHIP=y
display out seems to be built in
ssh is too
wdym with dmesg?
when the pi is on the new kernel i cant get into it with ssh or dispaly out, so i cant see dmesg
what about support for whatever filesystem it's using (i assume ext4)
built-in then?
probably
actually
did you grab and use the existing kernel config from the orange pi for the new one
Needed the repo URL and I was talking to claude while on the tty for Ubuntu server and he's like sure bro let me get you those command it'll be real ez
yes
This mf
slash
except there were some (NEW) thingies that popped up

no, not really
it has a red led
but its always red before the os changes it to blink at disk usage or whatever
did you make sure to update the initrd too
isnt the philosophers stone made form human blood or soemthing?
or is that only in the anime?
depends which alchemist you ask
lots of them were attempted to be made from red and white phosphorous representing the rebus or something
and the phosphorous was extracted from urine mostly
first install the new kernel & modules and then i believe it's as simple as sudo update-initramfs -c -k <new-kernel-version>
some people also suggested a combination of elements, i.e. mercury
also, the philosophers stone was more of a procedure than a substance at the time of alchemical dohickeys if i recall correctly

also nobody succeeded in making one that did what they wanted so it's kinda moot what they tried
its not moot
initrd_high=ffffffffffffffff surely putting this in the armbianEnv works
its very interesting to see how people squandered their resources and time for the sake of their experiments
but as you said "depends which alchemist you ask". nobody actually knew for a fact.
should be fine
oh nice it does the conversion automatically for you
right, but
that doesnt mean what i said was wrong

if you ask 10 creationists to describe the beginning of the universe, youre gonna get some answer variation even if their answers are ultimately misguided
What is this 
what it does mean though is you're going to get wrong answers.
which mythological beast can you not see in a mirror?
all of them coz they don't exist.
Me when I was in a tty and wanted to see if Claude could just grab the right repo to add to apt
Voice chat mode so in order to pronounce stuff correctly, it'll spell out / as slash etc
It was superbly incorrect about what to add btw
just let it control your computer bro, what's the worst that could happen! https://www.businessinsider.com/meta-ai-alignment-director-openclaw-email-deletion-2026-2
I cannot even fathom installing that on a corporate computer as a fucking director of anything
I'm director of the janitorial department at McDonald's I ain't risking that
the director of AI alignment/safety no less
Or that other guy
the irony is off the charts with these people
Ubuntu server's design is very human
Oh I didn’t know Claude had a voice mode
haven't they all had voice for like 2-3 years now?
Weird that it uhh transliterates it or whatever it’s called
Well it is not an Omni model I don't think so it's sending the text to another tts model on this one
Idk my accent is strongly Scottish so there’s no point trying to use the input
Claude here is aussie
ELEVEN
The only model that has a chance is like whisper or nvidia parakeet
Parakeet suc
Yeah this but real
Well it works and the one on other things doesn’t so I’ll keep using that one
I'd rather use Ford sync 2 voice recognition from 2012
For all that thing understands my rambling
It's probably me
But at don't get along, parakeet and i
i'd rather use the voice recognition from Lifeline
I want that Dr Sbaitso tts
Huh it’s like literally just a local model that transcribes some work some don’t and that’s one that works
yeah it just doesnt work for me for whatever reason
the transcriptions it produces are hilarious thoughj
i ended up having to go back to whisper
lol what the fuck
are they shitting me
bro
so where the ram then bitch
y'all better start a big ass amazon return pronto
not even on dimms
waiting to be dropshipped
This is what people were wishing, but at what cost
i deadass am preparing to potentially hiring a dimm packaging fab when it happens

because i might actually buy like at least a few wafers
hire me
employee discount 
i wasnt
i was kinda on one hand really hoping that the people steering the ship making the literal biggest bet in recorded world history economically
would not be fucking dumbasses
Slim
Got like 2 processes lmfao
from what i can gather
woooow, the price has tanked on that
depending on density, im expecting 16Gbit modules at about ~1.5TB per wafer to be at ~$6-8K
I mean, it has to
1440p 270hz oled is $380
could probably afford enough to hit a contract to commission the assembly fab
i'd love endless supply of dram
its not going to be any profitable though i damn well knew that dram wafers back then of such capacity cost < 3.6K
Nice just to have
also to yoink all of best ones for overclocking
wrr
bzt
wrrr
https://fxtwitter.com/claudeai/status/2031088175456903667 $25 for a PR review from claude 
Code Review optimizes for depth and may be more expensive than other solutions, like our open source GitHub Action.
︀︀
︀︀Reviews generally average $15–25, billed on token usage, and they scale based on PR complexity.
I've just been using codex and dealing with errors as they come holy
Could you imagine having your failed PR reviews tracked as a metric of how crappy a dev you are?
and having it be directly tied to cost?
oh also zed has a student plan now that comes with a free year of pro in case anyone cares 
can you imagine writing a shitty pr so Claude decides to charge you £300 to review it
lol
what does pro contain 
big brain move, lobotomise claude so it writes worse code so it can charge more for PR reviews
uhhh, $10 in ai credits and unlimited edit predictions with their model and some collaboration things iirc
i nuked the folder with the armbian source code, and im starting over 
ill see if it works with stock
https://zed.dev/education free is free 
i grabbed it while i'm still a student this month
I see
I haven't been a student in years
been using copilot through zed anyway so realistically i don't think this even does anything for me
I'm just abusing a gpt pro free trial to build a hopefully decent replacement.
(Aintnoway)
toast linked ai studio thing yesterday,a nd that has been working great so far
its less dumb than lobotomyGPT
just having the 1M token limit for gemini is really nice, big fan
can dump massive documents
per chat









Intel SSD






