#programming
1 messages · Page 392 of 1
I saw 5.3 codex helped a lot with its own training
Doomsday clock advancement wheeee
okay maybe if I wanted to use the tensorflow dataset pacakge I shouldn't have yeeted tensorflow 

Smol jank later and dataset aquired
how tf am I doing convs on 28x28 sized arrays
In pure technicallity, I guess it is possible to do a 7x7 conv if I can manage to get the 4x4 conv result to be in a small enough dtype
Do I actually want to do that, frick no
lucky for me, dealing with that is a future me problem
future me doesn't appreciate all the shit i've done them. they deserve the occasional impossible task being dropped on them at the last minute.
I think that is all of the code I need, let's see if it explodes
last minute bug fix and now my boi is going
it has no errors
it also didn't overflow which is an error in this version
Should have had it print how many cases it used
go again
811 cases used
YEAHHHHHHH BOIII IT WORKS
Bro, what is this ratio. 811 / 60000*49 because that is the ratio of cases per data
tensorflow exploded last tiem i installed it
tensorflows purpose is to convert to numpy
sounds like something it would explode doing
no issue as long as you put the entire dataset into ram 
I'm going to let myself have this before dealing with the training set to see how it does.
i sure love my servos randomly not working
The thing functioning at all was a slim chance
anyways I fixed it

Frodo is overrated, samwise is the mvp
sure thing samthewise graphics programmer
guys i think imma switch back
Switch back to what?
switch back to switch 1
windows 
You have unlocked new role
understandable
BUTT
and congrats, you have embed perms

just use iot ltsc dawg
that as well, if you can find it
it is hosted on Microsoft
I havent had issues with drawing tablets lmao.
Im guessing you used openTabletDriver?
NO

i uysed xp pens deafult drivers

RWKV-7 G0a3 13.3B: strongest pure RNN ever, MMLU 76.0% (+CoT=82.5%), MATH500 76.0%, GSM8K 92.3%, MMLU Pro 49.8% (+CoT=61.6%), no eval-maxxing / mid-training / post-training. Download: https://t.co/oJxpuJoeQD Ollama: https://t.co/bScaOJSoD0
Neat
Missed this
aw shit
wait a minute
my pc already had all this useful stuff downloaded onto windows
now i will not have it

does windows 11 crash out if u use it on ventoy
it's a gamble, most of the time it works
most of the time if it doesn't work, ventoy wasn't working already
rip
you should be fine because you already used it for linux
ok

also tbh
this pop os stuff wasnt really doing it
im wayy more used to steam os
hell maybe i should try steam os
what about all those steam os distros
A lot of people recommend bazzite for something similar
the one that crazy man uses
a lot of steam os distros aren't directly built on it but similarly built from regular distros to act like steam os.
meaning you get the hardware support for daily usage and steam os nice to haves
steam os is ARCH btw
it's just an immutable arch distro
i usually default back to an arch distro called garuda which is out of pure laziness because it builds a lot in as far as like gaming shit and gpu related stuff goes
Okay, it is not overfitting at all :SAVED:
what even is this bro
someone teach me
i wanna be cool
guys how do i boot into the thing that lets you boot from devices
depends on your mobo but for me it's f10
del even
I usually mash delete
That works to get into the BIOS on pretty much everything, and you can get to a boot menu from there
sometimes it's del, but usually you can do "advanced startup" (type "bios" in your windows menu) and use that to restart with the bios (after another menu or 2?)
im on pop os
There's a systemd command for it, need to check
sudo systemctl reboot --firmware-setup
oh yeah
on windows
sudo shutdown /r /fw /t 0
(or admin command prompt if you don't have Windows sudo enabled)
windows also just has the advanced startup button if you hit window key and type uefi
yeah, the menu is so slow though
same works on kde actually
i dont have that issue unless you're talking about having to go into windows copium bright blue startup troubleshooter thing
shutdown does not need elevation
🤔 might depend on the arguments, I remember needing it
rebooting into firmware might need it
i know it used to at least require admin access to do it. not sure if just admin user or administrator account
garuda builds in like every gaming shit into a nice windows style idiot friendly helper thing that saves 2 seconds but feels nice to not have to even think about
in exchange all you have to do is have a 4gb iso
but is also arch so you can do whatever
somehow it has also been the single most stable distro ive used
by kinda a lot
ok yeah I just checked rebooting into firmware specifically requires elevation
but why
it's less of a Windows specific thing but more so it needs to write to the EFI vars
it writes EFI_OS_INDICATIONS_BOOT_TO_FW_UI to OsIndications in EFI vars
you obviously would want EFI vars to be protected by security mechanisms
linux does not have to because systemd boot is still running as a uefi application
it reboots the same way though, writes to OsIndications
so is the windows bootloader
well then windows must be writing from the os environment and not the bootloader when using the option
well yeah
torvalds says yolo
basically what's explained here https://superuser.com/questions/1506321/how-does-windows-bootloader-add-option-to-open-firmware-setup
my query is less about how it does it but why they've chosen that way vs not requiring elevation like systemd seems to be able to do just fine
either linux playing fast and loose with efi vars or windows being overly paranoid
or somehwere in the middle like things usually are
Linux Kernel EFI Boot Stub
guess it's this but idk why windows doesnt do similar https://docs.kernel.org/admin-guide/efi-stub.html
kde called systemd-logind via d-bus to write to efi for it
so polkit just abstracts the elevation away due to local user perms letting you do pretty much any reboot action without authentication prompt
on windows that's considered equivalent of letting a random application write to bcd which is a no-no
anything that tampers with boot in any way will require elevation
does it depend on search provider
firefox does it for me but vivaldi doesnt but i use startpage on vivaldi rn still
no
google itself does it inline in searches so i just figured maybe it was that but it appears to be a ff & derivatives thing
dont have chrome installed to see
ewwww
think this is the first time ive installed actual chrome on any linux install
to check a calculator
kek
tru
polkit bypasses it but only for local seat via checking which tty is active and which tty logind is issuing it from
polkit is nice for random things but also is just strange and probably less secure than how windows w/ uac and such handles things
it's gonna be even more locked down once adminiatrator protection rolls around
Probably with over a hundred bugs
it's fairly solid actually
Ooh, okay.
i have the free plan 
Since forever ago
what are those 9 repositories
windows, ms word, excel, powerpoint, xbox ui, ms teams, ms teams (new), ms teams (new) (2),
and this
interestingly enough
I was wondering by my MacBook Pros battery was so bad
I checked it on with upower
8.8%
with over twice the CL
2133mhz ddr3 is cl 9 to cl 11, depending on how much you want to pay
the ram there isn't even running that fast
The more I mess around with openclaw (moltbot/clawdbot) the more it stands out as an absolute vibe coded mess... I thought the only thing worse than my code was my vibe coded code but now I realize it's actually way worse
Lamo I mostly work with mistral (use Devstral2 Small, Magistral 1.2 Small, Mistral 3.2 Small, Ministral 3 3/8/14b all with and without pixtral vision) and was planning on moving to voxtral when I have time even before this came out. They aren't frontier but they release a ton of weights for models that you can actually run local (also their api OCR is stupid good)
Any integration is possible with enough effort
But yeah, turn based games effort would be lower compared to something like minecraft
Even if there are no modding sdk
You can always do screen capture
As long as there's nothing real-time in the game, Neuro can probably play it
maybe
why did your workplace wanted to get as low cl as possible to the point of down clocking the dimms
I think it's Dell doing that automatically... Doesn't like having 3 sticks to work with
even with 3 stick scuff it still won't be that low
worst you usually get is ddr4 speeds with ddr5 cl
I'll have to look into the BIOS later if Dell even allows changing these things
I need to adjust my settings or model I'm using when it's just single output inference. Right now my computer can support around 6x concurrency on what are respectability sized local models
what
In response to having to rewrite NN again, I have resorted to gambling
It was fun tho
@faint sandal what are you doing in my recommended

hi
you wouldn't be the first person whose vods got into my recommended
like literally everyone's vods show up in my recommended
even people who I've never heard of
like neuro (kinda)
mixin hell
Java or audio?
Yeah, cooking too lol
It built tools that help the human researchers evaluate metrics and monitored the training runs, I wouldn't say it's terribly close to self-improvement though
this ain't dogfooding this is cannibalism atp
can't shadow a field of a package private type

even though it compiles and runs but only crashes when the class is instantiated 
i dont want to do access widening noo noooo 
yeah ts is never coming
JUST DO IT. DON'T LET YOUR DREAMS BE DREAM
did anyone try doing games on an lcd display using arduino's liquid display?
consider it accepted 
actually
consider it shipped 
no im not implementing it
but consider it anyway

someone said they wanna buy gtx 1650 for 65 euros, i told them its overpriced i wouldn't pay more than 40 euros for that shit
bro thought im trolling so he asked other people and he went "b-but its better than my 1060 6gb" when people told him im right
ALL of us immediately crashed out when he said that 😭
huge mistake, remapping failed and tainted the sources, have to redownload everything

Isn't the 1650 worse than the 1060?
"Hmmm, how do I revert to a commit again? Ahhh, rm -rf it is"
no it was neoforge assets only for some reason but the remapping failed on the fabric jar and now i can't even repro the access widener just silently doesn't work
absurd
exactly the point why we all crashed out.
Average "bigger number better" idiot
Well there ONE point where the 1650 IS better than a 1060
nivida-open
Cuz 1650 is turing
Is it a good point? Not particularly
But it is a point
It shouldn't be but we can't have nice things
just silently doesn't work
i wonder why
i'm cooked

i don't think even reflection is saving me here it'll just be an access violation
oh right Field.setAccessible
we're so back
yeah, the 1650 felt more like e waste
the super variant is wayyyy better
I'll be waiting for the next cooked 
i run into the same problem again
at least reflection works
as a solution
well, kind of, the class is still private
FOSS "lightweight" video editor vs bloated google-infested browser
who will win
All it took was 40 minutes lol
Well, to be fair, 1.2GB for video editor is light. Could be better tho
i needed to access this one hashmap that has the values as a private class, i said fuck it and my solution is to inject into every method that adds or removes to the hashmap and mirror it in a set
a hashset as a representation of the hashmap makes sense for my usecase, i basically only need the keys (a field of the values actually but i'll say keys for simplicity’s sake) i'd rather do this than deal with that again
for the first time, I have a comp sci practice exam I can read and not feel cooked
I think I found my calling for programming. Developing Programming languages. This class is tedious, but it's not monsterously difficult
You'll get along with chay
I like functional programming 
So what are we thinking is gonna win the slop wars this time chat
Opus 4.6 or GPT 5.3 Codex
we all lose
opus has been on top for a while, surely 4.6 is substantially better than 4.5
Bankruptcy
Tried both, they felt the same as 5.2 and 4.5
and i have never had any good experience with codex models, they just feel like lobotomized versions of the main ones
They're all just overfitting to benchmarks anyway

Hem
Funny ban evader, part 2?
this is like part 67 bru
Wuh
it's not the second nor the third time i can recall
Rare Crimon in #programming
I feel like there is context that I'm missing
You know, at some point surely there'll be a way to automate it
For what? There's a person that is perma banned here but made multiple account
That's all there is to it
Is it who I think it is
discord automates it, dude is doing smth against tos to evade it
I actually wonder how good mistral's moderation ai is
Holy hell
also, i have to ask, is there really a reason to use opus when kimi k2.5 exists which is like a bit worse and like a toooon cheaper
cant even be said, its that against the rules to even be mentioned
darn
Wow that's impressive
well one could say sharing inappropriate links
ah.
My favorite hobby. Suspicious links and SQL injections
4.6 eats so many tokens it's crazy man, pro yapper
I'm wating for a decent LLM that can run on a consumer gpu.
Yeah but it's still a lot more than 4.5 by default 
I don't think it'll happen though
gpt oss 
I wish there could be an AI discord moderator tbh
You could run the entirety of Deepseek on ~12 rtx 6000s
would make things lighter
hook up clawdbot to discord 
itd be difficult to make ig, itd have to read context and such
I don't trust any AI to make decisions on its own
it'll encounter an edge case and spend 20 minutes afk doing philosophy in its reasoning traces
maybe, scanning messages for slurs, though that's not hard to implement
Devstral 2 Small is only a little behind 4.5 and 5.2
so like a normal moderator or admin?
We have a tool for that
It's called automod
That's the easy one
This is for the edge cases one
Like the ban evader
yep except a normal mod gets paid nothing ever and claude will get paid $50 for that alone
yeah of course I wouldn’t let it do everything and leave it alone unattended itd need some control panel to check things
I dont trust ai to be infallible yet
Who would host the model tho?
LLMs are best used as tools to accelerate your doing, rather than for doing everything on its own
How do you catch ban evaders anyways? I know when I moderated a Gmod server I could see incoming IP addresses. is it like that?
The infrastructure would be costly
IP Bans don't work
that's what discord does
well, this one... had the same name and pfp on all of the accounts
Yeah, I learned that the hard way
Behaviour
also, what if the person connects from a public network like University WiFi? Would you then treat every user connected to that WiFi guilty
you ban alot of people in the crossfire, and any bad actor is just... one VPN use away to do it again anyway
(or TOR)

im back from my 2 day break
it was originally gonna be one day for lotr, but we also completed res evil 7

i wonder
Well it depends but some ISP give new IPs everytime you reconnect and there are open wlans
is there a stable url for the latest tag in a github repo
Well wonder no more
What do you mean by latest tag?
The most recent commit on every branch?
Or branch with the most commit?
i guess what i'd like is to get the newest commit that has a git tag attached to it
actually nvm there's no point in this specific case anyway cause hashes change
I have one I made on the server for my uni soc, but it just monitors for possible problems then pings them to mods with possible actions to be taken
It can operate with a limited degree of autonomy in that it can temporarily disappear things that meet a certain suspicion theshold until mods decide on a course of action but most stuff doesn't meet that threshold by design
interesting, does it monitor like
reading the context?
or just by line
Yup, its actually quite simple to do with a Discord bot
Handles images since the model is multimodal too, and has zero data retention and extremely low inference costs since its Mistral based
what are its limitations? like can it monitor every channel in this server when it has max activity?
like during the hype train record
That would probably get it rate limited
If we're talking hype train record
Lemme have a look at my current rate limits as a non enterprise dev on Mistrals API
I never hit em since its only a server of 100 or so people
yea figured, like that translation bot in the jp section had its limits as well that were never hit
idk if they got removed when it got changed to gemini 3
only the ones who work on the bot know
Yeah you'd need an enterprise agreement to raise the concurrency cap on whatever API you used
enterprise agreement? how much $$$ are we talking about?
not discord mods getting replaced too 😭
You just gotta contact them as an org and request a rate limit raise
Of course not all APIs are subject to rate limits like that, but then you run into the issue of high costs
surely you can also just use Groq or something
Especially in times of high activity
I wonder what the difference in quality would be like with smaller models
also doesn't have to be api
DiscordModBench when
could probably set up a simple multimodal model inference
You could host an instance if you have the hardware in fairness yeah
Tbh 24-30B models are plenty smart enough here
it kinda does unless you want to buy a physical machine
resting GPU hardware is probably more expensive than API
it just depends on how effective you could make it "skim" through messages
well yeah considering that llm designated services are ran with batch inference
ye, you could batch it
but then it will probably have a pretty significant drop in quality + other weird behavior
I see, gonna check with others then thanks
i mean the bot devs
though i'd say it won't be reliable to use llm for these tasks so treat it as a way to alert mods instead of actual moderation actions
(as in, classify multiple messages using one LLM request)
Yeah the one I made just flags suspicious stuff and pings mods asking what action to take if any
vani made a tldr bot that also link messages on the messages it made tldr of
And vanishes only the most extreme stuff that passes a certain confidence threshold where there's minimal doubt
so given that the llm can manage context well, could probably work with linking messages
Mistral is great at specific tasks
That's easy yeah, mine works on a rolling window of messages so it can make context dependent judgements
yeah, I know
but very different task than classification, especially for moderation 
I feel like an LLM would start doing weird shit like comparing messages to each other too much if you chunk them like that
works well for tldr since it's naturally a single request
also, it would work well if the llm is given chance to reason which messages are problematic
if you just prompt to directly point out "problematic" messages it'll hallucinate
but then that gets expensive quick
and slow
yeah, reasoning is absolutely not happening if you want to run it for every message 

best idea is probably huge batch of messages, give it chance to reason which one to explain and make it format a list of message links that seem problematic every once in a while
but more than often, again, as shuni said it'll probably mix things up
pretty messy
I wonder if you can make it a two-step thing?
First run a really stupid non-LLM, maybe even non-ML filter that removes simple messages that are clearly not an issue
Then for messages that have even a tiny chance of being problematic, run the LLM
Could even queue messages to run through the LLM to fix the first filter if they disagree
context
messages context happen between multiple chats
would be really hard to manage that
ignore it. just take the message at face value.
best you get is probably an llm that detects nono words
did i miss something, why is there roblox 
for mods to check in accordingly to the chat context
the whole issue is that LLMs are slow and expensive though?
would be nice to avoid unnecessarily running that
i mean that's about your idea
someone's going to beat all moderation by having a chat script which adds "in minecraft" to the end of everything they say.
with algorithmic filtering
and let llm handle the slightest susge
emuz your job is being taken by ai

yeah
like surely you don't have to run an LLM for 90% of the messages being spammed during a hype train
you always have some randoms sneaking in some nono during that
if it's identical to the message before it, it's probably spam
The idea has merit.. and there is a lot that needs parsed... I think additional data is a good thing. The LLM could be a final stage. A lot of actions are parsing for patterns..
yeah, but you'd bias it to have more false-positives instead of false-negatives, and a few messages slipping through would be fine
just make sure you filter on whole words if you do a wordlist
I am still pondering the cost of all of this 
vue isn't it like 1 am there rn
Even one channel here is quite active 24/7

How large of a model do you think would be needed for this?
I drink way too many coffee
And it's the weekend
just use a small one they're fine for classification
So eh
Than running it shouldn't be too bad resource wise.
coffee always made me sleepy
if you just want to classify single messages then ye it should be fine to have a small model
will just be expensive regardless because of how many tokens you'd be using
It depends on my mood. Sometimes it does make me sleepy. In fact, that's what I was hoping for when I drink that many lol
-# how many tokens does Neurocord write each day
-# (yes I know it depends on the tokenizer)
probably a good idea to first just write a log bot to see how many
killed by high quality 3d modeling an animations 
wouldn't have it any other way
too many
chat, i need your help with the hardest problem in computer science
caching? 
cuCuckoo or cuCUCKOO
Doesn't the model usually come from the games data files anyway?
First one
it's the other hardest problem
If you go by the latter, you cannot have acronym
but yes, cuCuckoo
the other one is crazy
blame nvidia for their weird convention
most of the popular anime games just put their models on their websites

the other option was cuCkoo but i veto'd that
Lmao, but it's better that way I guess. Just embrace the meme that will flood

Well, time to learn a new language then
Chinese famously easy to learn 
how hard could it be 
i already get confused between german and russian. ;[
everyone should speak english. slowly and loudly while facing me.
how
it happens quicker than you'd think
"ja" and "da" come out at the wrong times
there's an easy solution to this, don't say "ja"
blyat
I haven't had much issue chunking them, but I format it as a single turn chat message in a chatlog format and have a custom system prompt + instructional prompt prepended to the log for each prompt
So it's very heavily steered
I also run it at a very low temperature, 0.15 to be precise
ya utka
having a custom system prompt + instructions would be a necessity for this anyway, yeah
nice to know that a model can technically do that then 
What's one more? You already know english, python, c, and belgium 
belgium, the language 
Learning Chinese is like learning APL
they all use the same fonts. makes it easier
To be fair, I don't know my geography 
zzz has these for example https://www.bilibili.com/blackboard/era/iI31wCpCkCkKle3I.html
i have no clue what the text on this website says at all 
I sure would hope your font can cover more than latin
I have cjk installed
some of the page says "install the billibilli app" or similar.
idk why i explicitly installed noto-cjk
belgium is the name of the country, not the language.
they speak mainly dutch and french here
must've broken something
Same, I don't speak chinese either 
So there is no such thing as belgium language?
is it like canadian french?
Alright, fuse both dialect then 
kinda
like ignoring the french french rules
i remember french was all "no, we need a french word for that" for tech stuff. do you just steal the english words instead?
idk im fromt he dutch part 
i think so ye tho
biggest difference for everyday stuff is having words for 70 and 90
i only learned to count to 39 before i decided french sucked.
french numbers make me irrationally angry every time i remember them
how drunk were these people 
i think they just never used big numbers until one time an extra person arrived and they were like... "um..... just keep counting... maybe they won't notice."
some enlightened french revolutionist:
cool so wrapping mold just broke it
belgian french jsut has "nonante" for 90
instead of "quatre-vingt-dix"
anyways, enough about languages
french numbers must be a pain for speech recognition too... with english, you just need to listen for the digit and then for the place... none of the math
Wdym? This is #programming 
the math makes it more fun tho

I wonder how CachyOS would fair on my MacBook
And we're not stopping here. Geist Serif is already in progress.
also can't help but notice they fucked up the og image triangle is actually circle 💔
What do you MEAN?! ( 🐢 )
Okay after 2.5h of dev I gave up on neuros ONi mod improvment I will try this late
Now Minecraft AI Agent stuff
time to harass it to blow its cover
my favorite
Like every AI ever
that could either mean everything or nothing
like every ai from the last year or two thinks it's claude since so many distill it
on top of that i reeeeeally doubt anthropic would do a stealth model
This is actually really nice huh
Surely this is sonnet 5 right 
that's more like it
glm 5 

but yeah i got it to say it's glm twice now
probably safe to say it is then
yeah all other evidence points to it being glm
:classic: distill, no one can actually afford to train large models so just use one that exists already.
Destroying a tree like a player wood is really a pain
dawg i didn't know these could have identity crisis
I finally thought I found a good subtitle editor, KDE Subtitle Composer, but it turns out that the Windows build for it has been broken since November of 2023 because no one can be bothered to fix it 🫠
least slop modern model
is aegisub bad?
My goal is a subtitle editor that doesn't have a ancient ui
Also the website for aegisub is fucked rn

i want to make better russian utena fansubs some day
is it also moe

were you taught how to count days in a particular month (30 or 31) or multiply by 9 using your fingers?
9
20
3
neither
1127581915084492880
vedalNeuroHUH


nuero caret
playing "mario cart"
guys what silly app do i make next (please not something so profesional, i make them in Mobile so i are a little límites)
A chromium-based browser with vertical tabs
😭
man
these scam messages are getting too advanced
just saw a novel iteration of one and i can't even begin to figure out how we're gonna automod this
Whats so different about it
You can add a filter that inmediatly when people without embed perms post a link it checks though a list of safe websites and if it is not there it deleted the message and also times out the user, or if they do it x amount of times it bans/kicks them
you dont understand
Was it that discord invite link that was posted in genchat not long after I gave you a birthday ping
yeah
oh no wdym by that-
here you know what
They used like the euro as an E I think
oh its much worse than that

*Submit your questions / Issues below*
**0PEN-A-TlCK£T HERE**:point_down:
> ><ht
> tp
>
> ://di
> sco
> rdapp。
> com\invite\linkToScamServerHere>>

oh now i see
Oh my god how the fuck do you automod something like that
how this renders as a clickable link, i could not even begin to fathom
the longer you look the worse it gets
the japanese period in the link
the backslashes instead of forward slashes

Ok, well for starters, if you don't allow discord links at all you could maybe automod rdapp?
personally i'd just put TlCK£T in the filter and see if they try anything different
I wouldn't see another reason to have that

because honestly i doubt they've prepared a randomizer or alternative phrases
they probably will just change the message
oh my god
yeah I assume
id rather have something that blocked this shitty way of splitting links
it is so cursed
that requires me to see these messages when they're posted

i dont wanna stare at chat 24/7
If it's on seperate lines it should never be able to create links as a rule I feel like
yes, make the trial mods do it instead
I just noticed the "O" is actually a 0 too, they just did everything to prevent detection
oh my god that's horrid it actually takes all that and turns it into a clickable link
the scammers are getting much more advanced
yes
that's why it's so bad
what's the fucking limit
oh there isn't one.
you can split up the message over as many lines as you want
it doesn't change a thing
that's... so awful
im
the regex you'll need for this is gonna be so long
How tho, even when typing it it looks like it should be separate lines, why does it get compressed to one
i don't know
👍 "feature"
and the issues with the link also get corrected
yeah
what
it's intentially added
discord takes this:
*Submit your questions / Issues below*
**0PEN-A-TlCK£T HERE**👇
> ><ht
> tp
>
> ://di
> sco
> rdapp。
> com\invite\linkToScamServerHere>>
and turns it into this:
wtf-
Discord has some sort of link autocorrect feature, and it really should not be autocorrecting links under any circumstances
moderating this server requires a compsci degree
"just a bit"
a lot of regex

no no no dont say that word
"a bit" 
why not
regex is good
"easy"
prism out
too much regex talk
i never really had trouble with it
and i started learning it from nothing
idk
it confuses me
why is my laptop fucking boiling
i can boil eggs on this thing
im barely running anything 😭
regex any emoji
ok tbf it is a macbook air
It's not just discord, there seems to be something else happening with the message
(h[>\s]+ttps?)|(ht[>\s]+tps?)|(htt[>\s]+ps?)|(http[>\s]+s?)
do you figure that'd work
no probably needs more wildcards
Not a chance lol
i want my mbp back
Regex emoji 
(h[>\s]+t[>\s]*t[>\s]*p[>\s]*s?)|(h[>\s]*t[>\s]+t[>\s]*p[>\s]*s?)|(h[>\s]*t[>\s]*t[>\s]+p[>\s]*s?)|(h[>\s]*t[>\s]*t[>\s]*p[>\s]+s?)
there we go
that might work
you still have a few + in there
instead of *
intended
but yeah looks better
it's not desired to block https everywhere
The fact that it's combining them doesn't seem normal
only where it's split between many lines
ahhh
using weird quote stuff
ok that's clever
waiit
but what if they don't split it.
well if they type their links like normal people it gets caught by the normal invite link regex
if they split another part of the link
no i don't mean that
...uhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhh
i mean the https stays unsplit
the rest of the link is split
I think posting the regex rule here defeats the purpose a bit
they're not going to check
it's beautiful
white space
white spacewhite spacewhite spacewhite spacewhite spacewhite spacewhite spacewhite spacewhite spacewhite spacewhite spacewhite spacewhite spacewhite spacewhite spacewhite spacewhite spacewhite spacewhite spacewhite spacewhite spacewhite spacewhite spacewhite space
spontaneous I4 appears
watch them pull out another even wackier bypass once this one gets patched
would be hilarious
the limit instead of an i om
Just so you know, the sent link was http not https
thats true
shiro accounts for that
s is optional

that's why it appears in the regex as s?
yeah true
the s would be helpful if you need to do that
hopefully a better regex comes along that lets me replace whatever the hell this is
Always better to account for more scenerios 
i honestly don't see what you could do to compact it more lol
the length of this regex is O(n^2) with the number of fixed characters
which isnt great
if i send it right now will i get automodded
is there any way to remove or edit parts of the message before you run regex?
actually it's discord who am i kidding
it's probably just a text box that you put regex into
and that's the whole automod
i shit you not
we have a limit of 10 regexes

why do i have an ffmpeg command copied to my clipboard
soon these scammers are gonna start splitting the characters themseles in half
are they going to develop a new form of atomic bomb
fucking discord will probably add a feature that lets people dump arbitrary sequences of bytes in the textbox and specify the encoding later
file uploads
will add?
it's probably already there smh
wait, actually true...
the regex works 👍
Its not something you can do immediately, but wouldn't like a sentiment analysis pick it up?
but anyhow, embed perms fix that
let me check it
does discord automod even apply to text file uploads?
only one way to find out

I don't actually know
skibidi
ok i think it does work actually
so it is
shiro cooked
Why is that banned but not porn
nope don't work
LOL
Amazing 
wait do links work in the preview for these
Too early to say that 
here have arbitrary bytes
that genuinely looks like a virus LOL
well thank you for helping test
<NEOPRISMA>‡n �����
b�'‚����b�K∆¬!����
�Q����
�ÜéD����
�(*èI����
�$ÉN����
�ÜËãS����
�YQìX����
�ôa†]����
�L≈Üb����
�Høwg����
�î∏ul����
�¢ÌÉq����
�∫›}v����
�„r{����
�÷
k����
�¸ÄnÖ����
�¿sä����
�˝ê����
�µÕËë����
��\„ˇó������WQ]Œ����
�ò∂“�����»€q���
�$-I����fl±=(���
�{F���
�™D
K����»x5b���
��Scbr�����∫MCz���
��\å(ò���
��§ù���
��ìf¬û�����)PN£���
��ÌÒ˘•�����˛ëË´���
��ÊÕ¸¥�����®¢ø���
��ıÖU∆�����ÿ ‡œ���
��`7,›�����P”.5���
��ïË›;�����âFa���
��áñ˙g�����p1fin���
��√ûów���
�˚¬Ä���
�ÔÇ�����ˆ|Åã���
��—B€ì�����ûÀô���
�X›®ú���
�¥•Åü����ÿØ���
�*ªØ���
�"ª;ƒ���
�NÃçÕ���
�îaá“���
�£#{◊���
�‡5l‹���
�◊}t·���
�4–q‰���
�™´˝Á����à¶Ó���
�wr!˙����o@Œˇ���
�gØ����ò≥õ���
��Â}*�����]H���
�e EN����ä®T���
�+ ‰Y���
�„Ád���
�Œœ/i���
�G!Uq���
�ǵ«ã����⁄e/è���
�ò?.î���
�¿`$ô���
�¨O$û���
�~9£���
�G[®���
�·ä≠���
�ÛT�≤���
�®∑���
�}:Dº���
�õ毿���
�’;6∆���
�ΩÂ&À���
�_
e–���
�°öÓ‘���
�è†$⁄���
�å!fl���
�è;‰���
�(‘1È���
�◊ÜÿÌ���
�êØ3Ô���
�R
Bˆ����›"����
�⁄EŸ���
�ìYÏ���
�·"���
�óÌ'���
�áU¡(���
�—‰‘4����ªZC���
�¯ÏÉT���
�s51a���
�ˆ:f���
�R´Vk���
�€∂7p���
�t_u���
�≥Ô)z���
�§9���
�&:���
�J%��
�ƒé���
�q≥5è���
�Êç=ì����[ü˚¨���
�≈ ≤���
�«˘£¥���
�!äN���
;�y k���
b�
damn it i just pulled a rickroll alternative link to test it
arbitrary bytes part two
Squares 👀
erm actually those are rectangles
Wtf, I have emoji and cjk and even I still get tofu
anyway
Is that ancient sumerian bytes or something
wait what does it look like for you hold on
stop filter evading now
its just a macro recording for my app
if i catch any of you filter evading with files ill just ban you 
ive already been banned once please no
oh wait
okay
I wouldn't fileter evade with a text file
so you would just
embed perms does gate file uploads so we're mostly safe there
Ture
*true
ELOM MUSKTE CRYTPO SCAM
👇
@real sierra I FOUND A LIMIT
there
is that so
yes, it isn't working if i split it up at all
i could work with that potentially
with :// on its own line it works fine
could potentially just check for a line starting with > and containing a ://
well no
Backreading chat 
> [a-zA-Z><]:\/\/
well but this isnt what we want to block
<https
://g
oogle.
com\teapot
honestly i think this is a "nuke all links in the quote markdown" situation
automodding neurobot 
!discord
discord has fucked the feature
why not
MORE REGEX 
schizo
oh noes
o7
welp
we are so cooked
no https either
don't worry

that's... horrid
Aintnoway the vanity links don't need it
Yeah, @real sierra I assume you don't get automodded so you wouldn't know
lemme find a vanity like we can trust

Lies
Deception
i'd say maybe this is something to report to discord but they'd never do anything about it
one moment
<a
b
stfu markdownlint

hi
just automod anything that doesnt conform to markdownlint conventions
i do think that removing the clickable link fixes the issue
and it seems that it doesn't like multiline links without the <>
<line1
line2
well
ok that's just
if it must have the ://
making it non-clickable will work alright i think
wait what
<line1
line2
edited message
gonna contact some other people and see if we can devise... something... for this
atleast it doesnt embed
oh my god it works
yeah
at least it doesnt
shiro do you think it would be possible to use a bot for this
instead of automod
yeah was thinking about that
thats a bigger conversation
That's probably why it's autocorrecting actually















