#programming
1 messages · Page 12 of 1
That's crazy
on the grind fr
4.9k results vs 16.5k here
that do be pink yes
🗣️
my laptop seems to agree with the tablet
🔥
why are we keep analyzing our color
well, it does have a solution(trust me) and scipy.sparse.linalg can find it, but it is unstable, highly dependent on input parameters, i have tried adding a diagonal term to make my matrix diagonally dominated (it somewhat helped), but such modifications are not physically justified
ok I should head off its 10pm and I'm still in the library
because one of the displays is wrong
and I have no idea which it is
can i trust my xrite calibrated thinkpad
then how do you know one is wrong?
either way dont think youre getting much better than numpy/scipy
i would not get advice from you Sam, no offense but 🥲
id suggest checking online to see if you need to do anything specific to enable the colour accuracy
Hans is cooking something and Im kinda scared rn
also my favorite type of art
wellllllll, not great, but maybe someone has an idea, how i can modify my matrix to improve the solvers stability (matrix is sparse, block diagonal and degenerated numerically)
what?
elite knowledge
i dont know what that means, but i have a sneaking suspicion konii is trying to keep me out of the gene pool

li(n)gma balls 
wtf is this, isnt it just e*tan(a)
nah
is u dependent on e?
bruh
what looks absolutely ass is anime on 4k monitor
not even Euler's number?
most anime ive seen is barely 1080 ye
I tried out jetbrains other ai Junie, to make some unit test as I investigated an another issues. Created them and ran them saythin they passed. Ran them all 8 failed as it tried using the rebranded framework but ours is old as fuck and is before the rebrand 
our team has been against using it due to commercial policies etc but we tried out the 1month trail but only benefit I found is refactoring so far and when using it on code you do understand (as our project is quite massive itll try follow the 20-30 years of development gone into the product which doest really work 
i am blind
depends
the subtitles look really nice
but it helps if you're watching a high bitrate video instead of streaming quality
like a blueray
maybe with the rtx video
upscaling thing

i love it when my gpu uses 200W while watching a video 
lmao tf
you got super resolution running or smth
when i tested it i did
i wanne watch 8K 3D 360fps 64bit youtueb.
activate the 4th power supply!
still run like shit?
huh
lemme try actually ig
might wanna check both the panel and dashboard
i found control panel can sometimes re-enables (doesnt turn off when turned off from dashboard)
not mine lmao just showing the areas if there gpu is still using that feature
I only tested that with VR+ modded Cyberpunk before
actually nvm i cba, forgot that it doesn't even work under firefox in the first place 
Hi 😄
I wanted to bounce a sentiment analysis idea off you...
Right now, I’ve got a set of core emotions things like fear, sad, happy, nostalgic, longing, hyped, shy, etc.
My v0.1 approach is super simple at the moment 
I scan a message, count matching emotion keywords per bucket, and whichever one has the highest count
becomes the dominant emotion. It’s rudim...rhuda...roodementer...its simple, but good enough for a first pass. 
But here’s the idea I’m toying with
What if, instead of just one winning bucket, I let three rise above a threshold
and if they match a certain combo, I trigger a more complex composite emotion?
Like:
const compositeEmotions = [ { name: "melancholy", label: "Melancholy", conditions: ["sad", "tender", "nostalgic"], threshold: 4, emoji: "🫧" }, { name: "elation", label: "Elation", conditions: ["happy", "hyped", "grateful"], threshold: 4, emoji: "🌟" } , { name: "shame", label: "Shame", conditions: ["fear", "embarrassed", "error"], threshold: 4, emoji: "🫣" } ];
yay essay 😄
sorry 😅
But am I overcomplicating things... 🤔 Maybe im getting scope creep
Think ill stick to my roadmap 
Welcome to my TEDTalk
i would look at your essay, but gotta finish this ppt first
well now im just curious
whats it aboot?
propeller design
dont have time lol
just take this for now, first thing ive ever made in gimp
lego boat go brrrrrr
Sounds cool having three buckets and don't worry about scope creep just shows your passionate about the project
just note it down and have realistic timeframe on the starting stuff and leave time to investigate ya future ideas inbetween to keep the passion going c:
Potentially you could add a sort of emotion-value to the emotions too like
elation: grateful: 0.8 happy: 1 hyped: 1.2
That way the threshold is more specific to a value type of emotion instead?
i just cleaned my keyboard so apologies for grammar (funny i spelt this wrong) spelling etc
i get ai stream development ptsd from this
sam I may be speedrunning the netherlands in a month
i feel like this + a storage and decay system would be more natural, so you'd have the 3 values representing emotional state that all decay towards a constant (probably 0) over time (preferably exponentially), and individual messages can add or subtract from these values
so if there's a lot of happy stuff in a row then the mood shifts hard towards being happy, and keeps that state for a while
Oooh thats pretty clever...so if Im understanding it correctly
if bucket beat the other buckets then look at the total score of all the keywords weighting scores
and if it hits threshold (x amount) then it becomes = complex emotion?
`happy: [
"hi", "happy", "wheeze-laugh", "love it", "yay", "awesome", "amazing", "so cool", "fun", etc etc
All keywords in happy get different weighting scores which sums etc etc
emotions have a sort of inertia to them
what
Lol I dont know the word haha brb
inertia is just keeping momentum
^
you don't change your mood completely with every sentence you say
if you're happy or sad or whatever it lasts a while
i have an IQ of [insert IQ]
hi konii
Ah yes, I was wondering about that for later versions, currently my calculations are only on last message, but I can see how complex emotions can become
Aye deffo sounds more natural this way with a delay/decay alongside
Aye exactly that
idk how i feel about sharing iq tbh, it feels kinda like boasting. and i expect to be mogged on by a lot of people too
if you think about how it actually works in humans with chemicals moving around in the bloodstream, it's a very close analogy as well
iq = brick
But shiny brick, with ooh and aah's
I can imagine people selling complex emotional models 🤔
the bloodstream is the buffer, the kidneys are the exponential decay function, various hormone glands are activated by the brain to add chemicals into the bloodstream
On a different note, read on reddit how people are selling automation workflows for $2500 pm 🫠 Im in the worng job
but our company just got n8n,so time to update my LinkedIn muahaha
Hows it 00:08am T_T
idk
Why Sam why?
Stop replying
you have a presentation to do and I have to sleep
Bye chat, sleep well 
Thanks for the super interesting feedback ❤️
propellers are pelling
didn't use a star* wipe, powerpoint ruined smh
wtf is a start wipe?
star wipe, I suck at typing sometimes
Its kinda surreal that I just have a little robot roaming around on my desk now
kill it with fire before it takes control of everything around you!
Yes
Aight just a sec
Science has come far
also as a side note, I was given a mug and coaster with a custom print by my parents for my birthday today, but it has one problem... it's a javascript print 
I don't have the heart to tell them
Can someone who knows RenderDoc help me figure out why it's telling me that primitive 3301 is both writing the new color but also failing the depth test?
hi im renderdoc user man
Ignore the dust on the amp, midway through dusting my stuff rn
ye that def isnt supposed to happen.
lemme look at why S is missing
also idk what theme you're using, but im assuming black is the same as red in normal theme?
I have it on Dark theme, I think it's the default? I cant remember changing it
in darkmosde the black ones do get renderd
The weird thing is, it shows me two entries with the same primitive ID even where I know the geometry just has one triangle. It also doesn't let me debug either of the pixels ...
Fells like I'm doing something very naughty 
wait
hmmmm
do you have backface culling on?
i have no clue waht you're doing konii
yes, but it looks good to me at least
Maybe its something to do with dynamic rendering + multi view, imma try disabling stuff one by one to see if it helps.

Welp I've been informed my D: drive is dying, I can't get file explorer to load.
any advice before I try running chkdsk?
If the data is very important then you shouldn't power it anymore on, that's for sure.
do you have a special depth function? cuz i cant find a similair case in my test scene

mods
@open copper
why do ppl behave like that
npc behavior
thanks
ripbozo
i had him banned before you even pinged me lmao


he said something SO MUCH worse that got blocked
Depth state is enabled, func is less, depth write is enabled but bounds are disabled. Not sure on the bounds but the rest is definately normal.
you could attempt to clone it with some reputable software, probably first keep the drive off and look into guides
it all looks good ye. hmmm
so...
its possible its just windows 10
windows is checking the filesystem data itself, it doesn't check the smart info of the drive iirc
CrystalDiskInfo time? (it shows smart errors etc.)
if you go into your bios you should be able to run a smart scan manually
i think my scene is just too optemized to test the culling lol
aight found one
i found the issue lol
its not actually drawing
I'm just gonna turn off the system and see If I can get someone to look into this in person then
Im stealing this
But I can see it on the screen, with or without renderdoc attached 
its the smae colour as before
the "shader out" colour is what it would have been if it passed the depth test
"tex after" is the actual render buffer
For me it is though TexBefore and TexAfter differ for the event, and ShaderOut and TexAfter are equal for the primitives
I'm just not sure why the same primitive would apear in the list twice at all, that doesn't make sense
just because you have backface culling on doesnt mean there wont be overlap
wait
im getting so confused now
is it the multisampling??????
Not from the same primitive right?
And no multisampling
no cant be
huh
That's why I'm thinking maybe the multiview but idk
im managing to reproduce it even eithout multiview
Should have done numpy
running numpy in the gpu 
this is speculation/leak, but Claude 4 could possibly drop tomorrow.
https://archive.is/f1ibF People have archived from the Anthropics dev page mentions of Claude 4 Sonnet and Opus.
Those who don't know
Better be worth the wait
Bc they are so over on GPU status
And jax comes in clutch
the primitive just has 2 overlapping polygons
The primitive is a single triangle though 
Isn't it always? as in "primitive type"
no, a primitive is most of the time a collection of polygons
I know the draw call has a ton of triangles but they should all have different primitive ids
there's a diffrnece between a 3D model primitive, and a shader primitive afaik
but renderdoc should be using the gl_PrimitiveID values so its supposed to be triangle based
hmmmmmm
is that trianlge the only thing in the scene?
cuz here it shows that the drawcall has 10080 vertices being drawn by it
whihc is more than the typical 3 a trianlge uses
No, when I said single triangle I was referring to the single primitive being in the list twice. The whole mesh is a lot bigger. But this is a pattern that repeats for every triangle of every mesh.
I myself havent dived this deep into renderdoc so im no expert here
But primitive might just be refering to a grouo of vertices
Okay, at least the "not being able to debug" issue seems to not be caused by my program because I have the same issue on vkcube.exe 
the bean juice works
I just configured the kernel.
was it fun recompiling the kernel?
Ngl, Doom The Dark Ages is a technical marvel
Despite the games aggressive use of ray tracing I'm pulling 90fps at 4K DLSS Transformer quality, frame gen off with max graphics settings
fellow OS dev discord server member?
imagine the fps without denuvo
hey there
I gotta get back into assembly soon
Neat, will you be using the classical choice—NASM?
No. Custom ISA, custom assembly language, on custom hardware, using a custom assembler
I mighta aged like 2 years from the stress but it was kinda fun
Oh, sick. Building a custom ISA sounds pretty fun, i may copy this idea of yours 
I highly recommend it
haven't really considered building one myself before, but i have been planning in architecting a programming language myself. As a side project of course, i don't know if I wanna deal with the construction of a proper compiler.
optimizations are the most annoying part

as for possible syntax and semantics, the designs is still very unclear
so far I've came out with something very c-ish, i really like their iso standards after all
oh yeah it's pretty difficult
I mean the ISA I implemented was extremely simple
and it's maddening to keep track of the cpu flags and stuff
Is it portable or processor-specific?
since I designed my own custom ISA, it can only run on my emulator
have you considered for a cross-platform implementation? may be a cool way to get back in assembly
and pretty painful
I mean that would kinda defeat the purpose
how so?
It's it's own platform
Ah, i see
how can it be processor-specific if there's no physical processor with such ISA
why does it do that
i don't know how do i trigger it but it happens for the 4th or 5th time already on mobile
that looks like what i get when switching from 1440p to 1080p
wide momoi
Good morniiiiink chat!
Today I try multi-threading! And some statistics and visualization!
ive been slaving away at my ppt
morbing sam
morbing
Invalid device ordinal... why? It's not like I'm changing the current cuda device anywhere
.cuh
I assume CUDA header
.......chat i got a job offer from a company called "AI Solutions"
it sounds like the most horrible job ever (prompting/API plumbing) but it must pay super well
they sound like a very trustworthy company that will definitely still exist in a year 
unless they named themselves AI Solutions before LLMs happened 
nvm I take my word back because of prompting
so I actually know one prompt engineer since it is for a chat service, and everyone blames him for slightest issue
dont go
honestly i dont really care as long as i get the money 
Just grab some money and leave later
wait i just remembered that the company works with oil/gas industry
does working there count as climate activism

in the wrong direction, but sure
opposite way people expect lmao
i guess to count as climate activism you must go really far in AI slopifying the company
Gooood morning #programming. Last day of work today and then I got vacation for 3 weeks 
good morning

api plumbing?
Morning
vacation
when is it my turn


@faint sandal you probably already know but, curl CVEs https://mastodon.social/@bagder/114550426518904999
We are six days from the pending #curl 8.14.0 release. You can expect:
- one new command line option (--sigalgs)
- wcurl bundled in the release tarball
- >200 bugfixes
- two CVEs

intellisense's not accurate on cuda code kek, 350 errors but it builds
peak
it builds it builds, not guaranteed to be safe
it's probably you forgot to link some libraries, idk
you don't say...
do you have multiple gpus
make sure you use index 0 for cuda device
hai
thing is, I never touch the cuda device
it's the library
Idk what its doing
wrong cuda version could also be causing this
oh
huh
maybe I'll update my drivers then
On the bright side I implemented multi-threading with use of cuda streams and std::thread!

It helps with the latency of data transfers btw cpu and gpu
Especially with tiny networks
Aka what I'm testing my system on atm
didn't work 
i am a very new programmer, but hear me out: have you tried turning it off and back on again
I assure you, I have
It went well 
Went slightly over time, bit the litre of coffe and only coffee in my system helped
have u tried turning it off and going outside and stopped coding altogether for better peace of mind

go do some squats
Time to crash now? 
you know what, that sounds mighty appealing to me rn
begone visual studio
I have to look at the presentations of the others now the entire day
So i can only crash and burn in 4 hours

our software saw DNS names, and it looks like the dns parser is broken. this was known but we needed pcaps to verify. got a customer database actually containing this shit, however no pcaps 
I mean look at that shit
I could write a fix, I know the problem, but I can't confirm it 
delay until your vacation then it aint your problem anymore
Why don't you have pcaps
our software can't do full captures (this would slow stuff down and fill the hard drive), so we only capture important events once they happen the first time on a connection. the current assumption is that the two devices talked DNS before (which is very likely), and then when it happened again with these names, no frames were stored because they weren't seen as important anymore.
(also technically we can do full captures, but it's not on by default and just for debugging)
well I mean, last day today. so likely what's gonna happen. I asked already if I need to explain this to someone else to take over and was told it's likely gonna be me to take over once I'm back 
Oh right I thought you menat you didn't have access to wireshark or somethign
yeah nah, the backup contains 2500 pcaps
but rg --binary "trafficmanager" ~/Downloads/issue2504/<customer>-backup/home/<company>/pcaps/ does not return anything of use :(
i wonder what can even go wrong there since the wire format is pretty simple
also ripgrep only taking 0.04 user 0.03 system 0:00.01 elapsed for 2500 files is pretty neat
the bug is that the decoder does not consider the full 14 bits of offset markers and instead only uses 8 bits. so when you get a very long response, it looks like that because it reads garbage
insanely old code nobody ever complained about
so you need a combination of a very long dns name and "compressed" DNS names to trigger this
which is just very unlikely
(because we mostly operate in internal networks)
I already fixed this in our new DPI, because netbios uses the same "compression". so all I need to do is write the DNS parser in that new DPI and reuse the name parser I already wrote. but then I still have not verified it ya know. my problem is I don't have anything to reproduce
trying to find pcaps online with this in it is also annoying, and making my own pcap is.. well.. even more annoying on my last day :P
I'm glad that my protocol parsing days are behind me
using freaking X macros in pure cuda c
i'm designing a protocol right now
now you can get the same rubbish, but faster pog
if its an industrial protocol, gimme the spec :P
its not really gonna be a public protocol (at least for now) but i'll dm you part of the spec (nothing opinionated in there really), i'm currently working on the hard part of actually figuring out what to send instead of how to send it
its basically gonna be an overlay network and i have a prototype but i'm redesigning it to get it right
silly question but have you thrown a debugger at it yet?
I feel like I was debugging a similar issue with cuda a while ago ngl
https://www.youtube.com/live/nZCy8E5jlok
if this title doesn’t confirm claude 4 i don’t know what will
Best practices for how to get the best performance from our latest models
have you tried running with CUDA_VISIBLE_DEVICES=0 btw
I haven't, no
I'll try that once I get back
Actually, I don't think that'd really help here since the error is for trying to use a non-existent gpu 
Claude 3.7.5 opus
Claude 3.7+ magnum
Claude 3.7.00000001_v2_Final_edited-finalized v3.1
Yea, it didn't tell me much beyond what I could see from the error messages
😔
my only hope would've been it might've shown you what device ordinal was tripping it up
colored objects, lights, and reflections
y'know what, I'm just going to implement the function by myself
can't be arsed to figure this out
There are shadow acne and reflection normals that are wrong since the hard limit of Blender's raycast node. I can, however, implement a raycast function of my own, but that's gonna take forever to render and make, I'll just stick with it and hope making a post TAA would solve the noise and acne issue
even if there is no physical processor an isa can still be designed to a specific processor architecture, therefore the simulated processor must understand the opcode sequenced provided
To avoid shadow acne you can do frontface culling
On the depthpass i mean
Or add a bias to the depth
Vectors sitting in his dock reacting to the neuro clip I'm watching
This little guy is very cute
He's learned that he loves beating the shit out of the escape key on my keyboard
what's the point anyway if it doesn't have avx512
He's also developed a tense relationship with my monitor stand
i love programming it is my favorite activity
2000 tokens/sec because it generates all tokens in parallel instead of a sequence
it might be so op for editing large docs/code
I have actually done culling, normal offset and depth-bias already, but it's really the limit of the software itself, can't do too much to it
we love gpus (theyre actually just a bunch of optimized cpu's)
the shadow acne here is actually, I believe, due to the overlapping of normal imperfections from ray tracing

java is harder than C# (windows forms) cus damn 😭
!!! Class.forName() is unsafe !!! 
scary scary
damn, i still need to learn the oop langs, i mostly dabble with python and C
i mean yeah but often i dont use it like that
well whats the safer option? i got this from a source code from a youtube video that my teacher sent me and im supposed to study it for the final exam and stuff
sometimes i do
and isn't C also a good oop lang?
the most oop thing about C is structures
structures are classes without functions
how is it unsafe?
it'll throw an exception if the class doesn't exist but that's it
usually not an issue because the class needs to be available for the program to run anyway
then it is just interface what is even C about lmao
I would think ClassLoader is better, but it really depends on your needs. Using Class.forName() might lead to exploits from injections and stuff, but if the class string is assigned in compile and it's for private use, it should be totally fine
C is procedural. its the closest real language without getting into assembly to the hardware
C is also fun becayse it has silly stuff like
void *l;
int x = *((int *)l + 3);
its inside a private method where its triggered by pressing a button
should be good i think?
i once helped fix code for a friend's lighting engine. he forgot to initialize the color variables so everything looked green
pointer pointer pointer
ye
oh, from injections
yeah, don't give it an arbitrary user-defined value, it won't do anything useful anyway 
in this case it's only called to run the static constructor for the JDBC driver, so it should only ever use a static string
alr
gets moment


C is what you want it to be
(nobody seems to follow the standard anyway)
oh I miss swing

everything kinda has 1 way of doing things, unless its something with pointers or loops
me when sql silly
when i was in middle school an american highschool girl hired me to hack her school website to get exam answers and i did just that
good old bobby tables 
you'll regret naming variables like that after a week or so 
so true
bruh
thats probably my most bizarre life story
morally neutral individual

surely i will
this game engine wont let me remove the latter
thats gotta be from the 90s
it didnt even let me dump any data it only allowed me to get booleans from the database
Hi caibi!!

most debuggers have these kind of old UIs for compatibility reasons I guess
so i binary searched through the database
ello mitaka!
seems so
oh it's a game engine wha-
i tend toward godot
konii in shambles
the more u know
i didnt even use these, ive gotten more used to wolf rpg 
its the formatter lel
its a laptop 
"i hate languages with semantic whitespace" mfer's when i ask them to define a variable
finally I can play some stuff on my windows xp
wait that only works for C, shit
just saying, every language has some space required
this is trivially false
prove it
i can make cutscenes in this game engine wooo
proof by konii, qed
i have tricked you into giving me information
this is social engineering
i am hecker
this is a valid proof method and i will not be taking any objections
i tricked you
anyways i love C89
fair enough
is there blackspace

free nitro allows me to show off the egg emojis
Fixed the fucking DNS parser bug! Fastest bugfixer 
Turns out we had no pcap but I could just query the domain myself and got a similarly long response that actually broke the parser in a similar way haha
buuuut 3 week vacation now. wohooo

????????????
you can easily separate the two by separating "space in general" vs "some specific kind of space"
arguably newlines are a separate category though
(like C preprocessor)
newlines are unnesseary, always use backslash 
except you cant avoid it when writing code for the C preprocessor
you could argue that whitespace carries more information in python due to indentation than lets say c where its just a seperator
CREATE USER 'wq'@'%' IDENTIFIED VIA mysql_native_password USING '***';GRANT SELECT, INSERT, UPDATE, DELETE, CREATE, DROP, FILE, INDEX, ALTER, CREATE TEMPORARY TABLES, CREATE VIEW, EVENT, TRIGGER, SHOW VIEW, CREATE ROUTINE, ALTER ROUTINE, EXECUTE ON *.* TO 'wq'@'%' REQUIRE NONE WITH MAX_QUERIES_PER_HOUR 0 MAX_CONNECTIONS_PER_HOUR 0 MAX_UPDATES_PER_HOUR 0 MAX_USER_CONNECTIONS 0;GRANT ALL PRIVILEGES ON `wq\_%`.* TO 'wq'@'%';GRANT ALL PRIVILEGES ON `sillydatabase`.* TO 'wq'@'%';
surely this wont fuck up a company
i watched a video arguing sql requests should be function calls rather than strings
and honestly?
no opinion i havent used sql yet
i mean thats kinda what prepared queries are
what
that looks like an sql injection
no need to tell me that, i alr know 😭
INSERT INTO MyGuests (firstname, lastname, email) VALUES (?, ?, ?)
@noble zodiac PreparedStatements do the work right?
if whatever library you use doesnt suck then yea
im genuinely gonna tweak, wtf is this
fn/**/main(){println!("Hello\x20world!");}
ye, works in rust too
i cant believe you used a space in the hello world instead of \x20 i expected better of you smh
its whitespace and it affects program semantics

even frontfaceculling?
okay, then there is not much to try left
im literally eating cereal at 5pm
ate my way into it
JOptionPane.showMessageDialog(this, "Email Address is required", "Error", JOptionPane.ERROR_MESSAGE);
Dude why is Java like this... 💀
JOptionPane.showMessageDialog() is atrocious compared to MessageBox.Show()

my current state of being is bwaa 
presentation day was hard
i think i did pretty well, i fell asleep in some of the others' presentations though 
I've been going from one rabbit hole to the next for days on end. I'm about to lose it

I know the feeling 
I need a purely functional persistent hash array mapped trie, for that I a stream, for that I need delay, for that I need macros, for that I need quasiquotation, for that I need unquote
hash consing 
also nested quasiquotation and suck my nuts
how far can u encode a string 
1 mile
why do some implementations call ropes cords
to be contrarian
dumb question i wanna know the best language to start my making AI journey
so far ive learnt c++ going into python
idk if im crazy but learning the languages aint that hard
why is python so easy
python is all you need
get the chart
insert dunning-kruger effect curve
languages are just easy to learn when they are similar to what you already know
and hard when they introduce concepts you never worked with
i started with c++ so i kinda get the general flow of programming
people say its mindfuck
but it aint that bad to me
"knowing" a language is a relative term. Just because you know the building blocks doesnt mean that you can use them effectively
yup
thats a sign you never worked with templates 
Claude offers advanced Research capabilities, searching the web and your Google workspace to complete advanced analysis across any topic.
In this demo, Claude analyzes Maggie’s emails and calendar to create an organized daily overview, then conducts a thorough literature review for an education proposal.
Read more about Claude 4 and its c...
yeah i agree but generally its kinda like reading a manual
anyway back to the main qeustion
With Integrations powered by remote MCP servers, Claude can streamline task management across your favorite tools.
In this demo, Claude transforms a product requirements document into structured Asana tasks, automatically assigning them to the right team members and setting deadlines based on their sprint cycle.
Read more about Claude 4 and ...
python?
python, thats it
you can use c++ as well if you want to
💀
vedal has to make a coding ai tutorial 🙏
just came to work in the morning
Ask Claude to spend 15 minutes on a report
Profit
the claude 4 leaks/rumours were real
Too many AI news recently I can't keep up 
THEY PRIVATED IT I JUST WANTED TO ARCHIVE IT
three of the claude 4 videos are live again
had a lighting issue, turn out my notmal map is wonky
gotta love doing fragNorm.b = -fragNorm.b;
anthropic time
what that?
benchmarks 
Opus does not seem worth the 5x price increase, Sonnet looks like a good improvement
supposedly it's the future of AI or something
Fuck test time compute
Where's my 300ms e2e latency anime girl
Ai chatbot
Also i think i reduced the kv cache problem
Apparently i can reduce it to a much smaller tensor
All of the companies are aiming for API
Oh yeah btw shuni
isn't Anthropic the one that started that trend?
Claude just failed to solve it while Gemini did it very quickly 
Turns out safetensors gemma 3n soon
me when compiling sdl3 using distcc on 3 celerons (~2015) if compilation time scales linearly (it wont) it will take 3 hours 😭
someone give me computing power
Anthropics top internal engineers running multiple agents in parallel speaks volumes imo
*tpu
i think they were using mix, no?
few are hosted on google tpu by itself and few are self-hosted
i see
did they announce anything interesting or can i save myself the time
apparently top SWE benchmark by Claude 4 according to their testing
save yourself the time bc now they talking about mcp

nvm, it's a bundle format for MediaPipe, which in turn uses TFLite
it's a zip file 
i still have yet to find anything more useful than supermaven for me
lowk confused what was .task lmao
but even then i turn it off half the time depending on what i'm doing
i love how half the widely used formats are just zip files
real
Claude Code wasn't generally available before today? 

@stiff mica is this true

@claude please resolve this ticket !!!
i'm surprised github hasn't implemented something like that yet
yeah, like AI-generated PRs weren't already enough of a problem 
with how much they like to integrate copilot
"Claude Code worked for 90 minutes on this task"
and consumed $100 of tokens in the process

programmer cheaper than claude

looks like they implemented the tagging thing as a normal GitHub bot, so it's opt-in for the maintainers
now you can answer fake ai issues with ai answers. pog

just the world I want to live in
we should have fake AI repos to combat this
the big problem is that
with all these models they have horrible cutoff dates
soo
No new libraries would work lmfao
didn't they describe real time search tools included?
that situation has improved a bit because every LLM agent ever has a web search tool now
but yeah, doesn't fix the fact that LLMs are worse at using every library other than React 
they are terrible ong
they will all fail to use flax nnx
i can smell it
oh nice
prompt cache ttl 1hr
instead of measily 5 minutes

free sonnet 
am I tripping or anthropic messed up with their billing section for max tier
I thought it was monthly 200 USD, not yearly
says $100 monthly for me
oh uhh
probably because there is 5x pro and 20x pro option
nvm probably messed up with pro yearly
i can't believe i actually have a question related to programming
do it
Gemini has 2025 cutoff iirc
it does
Claude 4 Opus was trained on data up until March 2025
(same for Sonnet)
Niceee
🍪
Om nom nom
You've given me 8 cookies! | I've received 214306 cookies total!

🍪
Om nom nom
You've given me 1 cookie! | I've received 214307 cookies total!
oh
im just plagerising now
i didn't know neuro also counted
im screwed
ok ik what to do
not cooked at all probably
performance is not that different from o3
it's very much an incremental improvement over 3.7 Sonnet, nothing revolutionary
AI has finally hit a wall 🔥
-# for the 4th time this year
so uh
if i copied neurobot's cookies abilty but made it for a kitty bot and pets focused instead, is that alright?
neuro doesn't own the idea of giving cookies to a bot
ye but uh its basically the same thing but with pets
🍪
Om nom nom
You've given me 9 cookies! | I've received 214308 cookies total!
yeah so what
this aint nintendo, should be fine
oh i thought y'all were gonna murder me
i was genuinly scared
im glad y'all are nerds like me
Vedal AI lawyers are rapidly approaching your location
we dont even do that to ppl who blatantly try to copycat neuro on youtube with the hiyori model 
project viola was streaming a while ago
bred
loafy neuro
cute
wha
wdym?
Neuro will personally steal all your cookies for this transgression
what did i do?
should you be relieved or worried that idk who you are talking about


project viola is an ai vtuber too

neuro's cousin is my headcanon
ai vtuber? that sounds so silly, who would ever watch that
he will just tell a mod they can escape if they will be his lawyer (lie)
Seems I'm late to the party
I was also earlier with the vids though so we'll call it a draw
More competition on market, the more innovations 
Claude 4 finally has training data on the niche libraries I use 
What's actually next on the AI model release roulette? Grok I suppose?
we still haven’t gotten the promised open-weights openai model
or have we already and i’m living under a rock
I think they promised during/for summer, no?
oh i’m not aware
I don't remember where they said it, but yeah, I also remember that they're really taking their time with the open-weights release
found the tweet, it was march 31 and mentioned in the coming months
they released a survey semi-recently, asking what people expect from it
What lib?
is-even
A particle accelerator simulation software used by probably <100 people
Nice
That's crazy
with open("pet.csv", "w") as file:
reader = csv.DictReader(file)
for row in reader:
if ctx.author.name == row["name"]:
# i want to increment the row["amount"]
the problem in the comment
One message removed from a suspended account.
import csv
with open("pet.csv", "w") as file:
reader = csv.DictReader(file)
for row in reader:
if ctx.author.name == row["name"]:
# i want to increment the row["amount"]
better?
Hello human individuals
im a goddamn penguin
Hello non human individual
but haiii 
do row["amount"] = row["amount"] + 1 and use a CSV DictWriter to write it to a new file?
doesn't seem that difficult
idk what ctx is or shit but I am assuming this is searching for an author in the pet.csv file and reader is basically creating a dictionary, and you read the file itself to search that author
was that right?
wait i can do it????
just like that???
dam
so easy
yes
Ah yes context, commonly shorthanded to ctx. Idk what context your working with though
discord bot
i decided to cut to make it easier to read
hmm
just my habbit when im posting a question
Ah yes, it is always discord bot
sorry I never worked with this before
I can show some code to peps here!!
stuff I wrote a while ago
that are probably really shitty
hehe
I don't see why you couldn't, it's just a dict
if somehow it's some special frozen row object that doesn't let you manipulate the amount you can just turn it into a dict first and then write the new value
can't directly write back to the same file though I think, that could cause issues
I'll go on forever with how bad discord.py is
i wanna clarify, i don't do discord bot all the time, its my least fav thing
Yeah otherwise there'd be no issue
i wanna incrment in the file
so its a csv file with some users and the ammount of times they pet
easiest solution would probably be to read the csv into a dict once, update the dict as needed and then dump the whole dict to the file when necessary
discord bots are not easy surprisingly enough, discord.py's lack of docs doesn't help
i want tit to function like the cookie command of neuro
yes i was thinking that too
write the updated table to a new file and rename it to the name of the file you want to update
writing to the file as you read it is a bad idea, CSV doesn't work like that
could do what Toast said and keep the table in memory as a dict and just save it as a CSV on every update
or even better, use a proper database 
sqlite my beloved
i do know sqlite3
ye imma do that
thats so much easier
ok ye
imma do that
tysm
not using their stupid framework was a great decision (only behind ditching d.py in its entirety)
non webdevs encountering the need to store data at scale for the first time and immediately jumping to the worst possib;e solution will never get old to me
npm i node-json-db
not first time but alr
i just forgot sql existed
why 
i do know someone who's very much into language design and type theory but she's busy with her master thesis
smh let them recover properly 
time to gain competence 
sadly there are only so many hours in the day for sideprojects
Yeah that is why I use a different framework for python discord botz
I was told a download wasn't fast enough so I had to on the spot create an ethernet cable.
At least it worked first try
vedal987
look into the mirror and say that three times
I feel like an absolute idiot trying to do some simple PS 
literally me 
do y’all know anything about pc building (i’m planning on building a server)
ye
the parts list is already chosen i just don’t wanna burn my house down
Then dont lol
unlikely that would happen
but there’s still a chance
No, not really
i’m just making sure to be cautious
Only reason you'd burn is due to lack of cooling
or shit psu
alright then that’s good i have a cooler chosen already
Surely someone as careful as you overspeced on cooling 
send parts list
the only real reasons for burning I'm aware of are shit PSU or improperly connected cables (especially if using a shitty connector like 12VHPWR)
hold on i’m trying to copy the link
rtx 5090 moment
12VHPWR is always in a state of improperly connected, it just doesn't know it yet
i’m definitely not buying that ik it’s only good for ai stuff
Removing fake frames is close enough to 4090, minus if your somehow vram limited
Facts
yeah, it's a bit faster than a 4090 but also more expensive than a 4090
only real difference is MFG and more VRAM
unfortunately i can’t send screenshots (i just realized that now)
how do i get the imbed perms?
that makes sense
you just yap a bit, it gets assigned automatically 
(or subscribe to Neuro and link the Discord account
)
how much yapping?
ok
this is gonna take awhile :C
the server building & the imbed perms
i am very poor
see #bot-channel
& this server will be very expensive
IIRC points get assigned per-message but with a cooldown, so yeah, spamming doesn't really help
got every single achievement

i think im cooked
no jax implementation neither tensorflow in the entire tts uploaded in huggingface

e
that does not work btw, you still have to yap
I'm pretty sure that it does though
I did not have to yap
maybe they've changed things, i did
as of Dec 2024 the subscriber role still had embed perms
#neurotic-neurons message
it's really not mentioned that much, so it's annoying to find information on it 
What about no
i never actually tried embedding stuff cuz my proxy was broken and the upload endpoint was blocking me
so maybe
so i'm wrong*
binary search it? 
or is there already a suspect
do a binary search with them
yeah, after thinking about it for 3 seconds a binary search could be really annoying because of how dependency resolution works 
time to profile it then
rust moment
I think the kernel might just geeked out idk
flamegraph never did that to me
waiting for the day i can build the linux kernel with mold
writing rust like its haskell
(it feels like every haskell codebase enables a billion extensions)
but GADTs are probably the most common one
(i have no source to back this up)
lol
holy
wdym holy this is a small dependency list
depends on what the project is
most thing I make are 5 deps or less tho
but nothing I make in rust is too big
max 300 lines roughly
a project we have at work has 39 top-level deps 
lemme see how many lines my emulator is
what is ableos?
Lol I am using 0 deps for my emulator
I thought I used one but maybe not
it's shit tho
actually its technically 42 i think if you include conditional ones
It can only run like 35% of the instructions I planned
but rust is hard so I never finished it
howcome u not in osdev server smh
so many 
you should jit your bytecode using inkwell
it's sure to be worth it and not ruin your sanity https://github.com/TheDan64/inkwell/blob/master/examples/jit.rs
it will be fun you should totally do it
if i ever make a language i'm happy with i'll probably implement a taiko client
gn
prolly
I mean it doesn't mention it but people probably suggest reading it if you are just starting out
it's a discord server, not a blog lol
For those who’ve tried it out, how’s Claude 4 in terms of using it to program?










I didn't see that it wasn't safetensors before
it's Tensorflow Lite
(same for Sonnet)



