#programming
1 messages · Page 9 of 1
what, I'm saying it's your choice whether you want to learn or use chatgpt
Yep
It's not compulsory education here so
If u don't wanna study no one stops you
ew
our math teacher was kinda based
He would just put us n options on what to do and let us do whatever we wanted
Cause he said he's not gonna try to force anyone into lessons if they don't want to
And if u didn't do the assignments you would be fricked on the exam day
would you rather be forced into studying or just do it of your own voilition?
Younger kids have worse problems These days tho, I've heard from my sister the craziest stuff happens at her middle school
I am here 
I think partly it's a matter of practicality, in that it's not easy to enforce the non-usage of chatgpt

I study cause i wanna keep going forward with education lol
Cheating will only bring u so far
yeah
I study cuz i cant afford to spend another year in hs. Im so fucking sick of it
U can max pass the grade but then next year ur fricked cause everything builds on top of what u learnt
What's the most annoying part
To me its that you always have work to do
I'm graduating college soon and I'm the happiest I've been in the last four years :D
freedom!
(at least until I get a job)
Money 
Money 
money
Cant wait to start flipping burgers on Saturday next year
Just noticed this is my third last salary
Before im finished
Next year will be crazy tho
Ausw
Cause
28-35 bucks salary per hour
Minimum
(that's a good thing right?)
Ye

BETTER THAN MY 7.7 BUCKS PER HOUR

Now i need to find something i qualify for

Cause these student jobs really expect an apprenticeship diploma
As if the jobs were that hard
The first take off is the most dangerous time for any multirotor. Care must be taken to ensure the aircraft is not destroyed in the first seconds of flight and nobody is injured.
(I hate inital tuning)
Lmao looking over chickens is something i would qualify for
What burgers are you flipping for 28-35 dollars an hour? And can I try one?
Zurich minimum wage is 24 bucks
Cause the city is so expensive
“Waiter, one Zurich please”
Some jobs are friday afternoon and full saturday
That is one way to get burnt out
Anything to have more work experience on my cv🙏 😭
Wait so like 12 hours
That’s it?
What the fuck is a "24 hour job"
ok new questionable design choice by me: I've taped the antennas to the carbon fibre frame, surely that's great for signal integrity 
Ye
They are meant
For students
Most of em don't even need interview
Just need to check ur alive
I think cf is better than steel?
still not good lol
tbh there's already too many anetannas on this thing
do we really need 7
Cant yiu just use 2?
unfortunately this one has a bunch of tat on it
2x antennas from the frsky receiver
2x antennas from the rfd868ux telemetry radio
1x gps patch antenna
1x adbs antenna
(I think?) 1x fpv video atenna
only the first 3 systems are being used though lole
Can we switch
I'm not qualified for cashier here lol
They want a diploma
Profineeeeet 
One which you work while sleeping I guess. Sounds nice. I’d love to work while I’m sleeping! I’ll just spend my free time whenever I’m awake! 
god I am so scared of this guide
my first attempt literally flipped the drone on the spot
adittedly it was partially my fault for not having a motor estop but still 
I also love how this software is designed to work for anything from a racing drone to a massive multirotor coptor
like hm yes surely I want to assign one of my rc controls to flip the aircraft, I'm sure that would have absolutely zero adverse affects
ok who the hell thought that engaging that estop of the motors doesn't *instantly* disarm the craft
if I flip that switch I feel like it should not be able to instead flick off again and make the motors spin at whatever throttle you had set
that sounds like a really easy way to hurt someone
Intel new card specs are out if they hold true they cooked NVIDIA
you can link them(4-8) like back in the days again and the big one has 48VRam and 2 GPU on board
With 192 GB VRAM you can train big models then but I'm not sure how you orgnize the power supply then 300W *4 = 1200W + everything else
What about pricing? 
Not know yet but I guess around 700 to 1k per card
the other models had those prices UVP (vor Steuern) ? 249 USD 299 USD 219 USD but they are only half of the card which is coming now
if you think about that NVIDAS new workstation card which comes out with 48GB Vram will cost around 10k and can't properly be linked . Intel will be cheaper by a lot
But, Intel won't have CUDA, so anyone that wants to do AI has to get Nvidia
Well this is the question I also have Intel has it big OneAPI but they are currently CPU only which is by the way used by libtorch and pytorch but they say they need time to finish it
It's just an unfortunate fact that Nvidia kinda has a monopoly position on the AI market
Just 99% of software only GPU accelerates with CUDA
The compute kind of acceleration, not the graphics kind
if you use frameworks this isn't true but also the source of so many bugs in it
Many just don't want to spend time implementing more than one API, and CUDA is the one used by all the AI stuff already, so everyone just uses CUDA
CUDA is also just C++ with extra stuff so that makes it easier to adopt for C++ devs
libtorch is also C++ but you can use RoC , Intels NN stuff and all the CUDA stuff in itso hm
The problem is only when you direct using cuda = people who come later always get the better tech
Either way, it's pretty clear that CUDA has a monopoly-like status in the AI space
Thus Nvidia can just make their enterprise GPUs are expensive as they want and companies will still buy them
Even though AMD offers like twice as much for half the price
I don't think they can hold on it for enternity . There are coming photonic NPUs(interference only) which are also not CUDA based but they are 1500x more efficient what NVIDA does. In 1 to 2 years they are shipped in mass.
So why is google fricked by authorities but not nvidia
Why is nvidia not monopoly
When it clearly is
Because google actually acts like a monopoly
Nvidia has almsot 90% market share and work with a ton of products to mainly support cuda
having the majority of market share doesn't automatically make you a monopoly
NVIDIA is effectively a monopoly because their products are just that good
Google is a monopoly because they bundle lots of products together to avoid having to compete
google is a monopoly because it controls multiple technologies that ensures their dominant position as well as spending shit tons of money to make itself the default on other vendors
hm
To make it similar to the google monopoly case nvidia would also own tsmc and the major game engine
I guess they did try to buy arm
first time I'm hearing about these infinity files lol
i've never heard them be called "infinity files"
yea you usually call them special files
Ayoo, any Rust Programmers here?
plenty, yes
like half the active people here
I'm studying rust as we speak
Wow, finally seeing someone. My college has no Rustacians, fucking peak loneliness
mine had like only 10% of the people actually programming, the rest just coasted
Same bro, post ChatGPT, everyone became a coasting coder, claiming to be pros
Before ChatGPT, one monitor was always either documentation or stackoverflow or some shady websites
it's still useful to look at the documentation, chatgpt can spit out bullshit sometimes
For small codes it's perfect, but for large backend structures, it vomits out random codes...(Cough cough)
🦀
I have presentations thursday 
Anyways, it seems like for every graphical improvement i do, i get more and more obscure bugs
Honestly kinda real, you make it better but get more bugs
Morning!
Oh god 😂
Hello chat 

Inserts data into executable
the approach I'm using right now allows the existence of bvh and other optimization trees, but those are for the later stages 
Wait a second
Wdym u got depth pass in blender
Doesn't it already have it
It's blender
what I'm doing is recreating a raytracer in geometry nodes here
The more I see the more beffudled I am
H-how?
you trace the rays 
I have done raytracing in shader nodes, compositer and now for the geometry nodes
That is... impressive
This add-on adds a “native code” node, which allows you to write Geometry Nodes using a native programming language, such as C or C++. The node takes a “file” input, which you can navigate to your compiled code, and sockets will automatically be populated from the function arguments and return type. For this to work, your code must have...
but very limited and I bsod my pc while trying to do something

Hi gays
but I mean, good concept tho, since official contributers doesn't have plans(yet, at least not on the milestones) to make alternative scripting languages for nodes(like VEX in houdini)
hot swapping C codes would be so useful
yeah that's fair
though keeping it in geonodes probably has a nice portability advantage
they just don't want to map the C backend apis to the python api level 

it'll just come down to a question of, is trading complexity and maintainability with customizability worth it
they probably just want you to compile the code on your own, if you really want those nodes
Google IO <t:1747760400:R> already hyped, as Google has been cooking consistently in AI 
i'm honestly more excited about material 3 expressive than any ai stuff they might talk about
cant wait for more features I will never interact with
idk if it’s a habit but i never use chatgpt, it’s just easier for me to look at the docs +most of the problems i want it to solve it fails +where’s the fun if you’re not the one solving the problems? the reasoning ones are a bit better but still. i really like code completions though for boilerplate stuff
most of the problems i want it to solve it fails
i present it a problem, the SOTA model then proceeds to hallucinate half the libraries and apis in exchange for a gallon of water
why can’t they just say “idk man”
idk if its true but I remember reading somewhere that models performed worse overall if they were allowed to express doubt in their abilities
competition 
does anyone know if they’re actually aware of what they know?
SNN for the win
just that if they're able to try and doubt themselves they performed worse
or do they have to “learn” that they possess the knowledge with the knowledge itself
i replied to you because it makes sense if the opposite is true
not because you implied anything
mb
https://arxiv.org/html/2410.09724v1 kinda covers part of the reason but I haven't read much of it
interesting that reasoning models constantly doubt themselves but that makes them perform better overall
maybe just a bias in the data, that uncertainty is associated with less performance or something
or they switch into roleplayer mode, and start rp'ing having zero confidence in their abilities lol
from what I can tell it seems like somewhere in the training pipeline the llm is taught confidence instead of being right
but like
I am very much not reading the whole thing at the moment
fair enough

Ye but will be likely treated like 2 devices
can wait till AI systems/LLMs are hooked up to elemental binary logic gates, so the concepts that they describe in reasoning have a log
those are extension cords plugged into extension cords
Yep
what an abomination
Llm reasoning has already shown to not work as we thought
mods, add an electron to every atom of that abomination
I like how half of this is illegal where i live

as in chain of thought or that llms are a "dead" end according to you?
Bros power cable handles 100 amps or something😭
Its not that bad
The chain of thought part
What the model says it's doing is not what it's doing
SMOL
I mean, I've read a paper that changed the names and minor tweaks to the problems that caused a major drop in performance for llms
which shouldn't happen if it was capable of reasoning
i.e, it's doing some sort of memorisation rather than reasoning
it’s fine, i run my AC unit, PC and speakers off of one flimsy cable, it barely even gets warm when fully loaded (~3kW @ 220V)
I have extension box for my pc, monitor, second monitor (which is really a tv), and my router
The problem comes withstacking splitters
I don't think that's dangerous territory though
🌏 Get NordVPN 2Y plan + 4 months extra here ➼ https://NordVPN.com/sabine It’s risk-free with Nord’s 30-day money-back guarantee! ✌
AI industry leaders are promising that we will soon have algorithms that can think smarter and faster than humans. But according to a new paper published by researchers from AI firm Anthropic, current AI ...
hmm... I've heard that youtuber is particularly controversial
nevertheless, I shall give it a watch
ah, Antropic's paper
I've heard about it, but haven't read it yet
Ye idk about the youtuber but the media presented seems reasonable
wouldn't this imply though that it shouldn't be capable of coming up with novel ideas? (which alphaevolve has disproven by beating strassen's algo)
Alphaevolve was basically a monte carlo rng of reiterating and testing a ton of stuff till something works
Wii, 360, ps3, ps2, ps1, surround system, switch, sega master system 2 and a tv.
Probably not even 1KW
The collection keeps growing 
also probably not running the wii, the 360 and the ps3 at the same time
Ye
imagine what it would be doing that though
You'd need pip, and even the i doubt you could get more than 4 onscreen at the same time
Maybe the dual play 3D tv shit could work
what's the difference to science/thinking then if not trying out stuff until it works
, in this case it was easy to verify so that's an advantage
Also, its interesting to see how much power usage has gone down.
The old tv from 2008 used 240W
New one from 2015 is only 180W
And the one in my room is 70

Nothing here is rly that novel
It's just an ai agent
With outside feedback
Nothing special
idk beating a 50years old algo record is pretty novel to me, I think that kinda goes into philosophy
Just the method of iterative sampling is better
Grpo is alphaevolve but during training
isn't alphaevolve uhh monte carlo?
oh wait, maybe I am not awake so well
yeah I think I have already said that way before
Basically
Hypothesis testing
Tree based
Brute force till something works better
yeah seems like what I have skimmed through was correct but thx for maing it sure
tbh it was all about gemini's base performance for this to be working
Yea
It wasnt like
Alphafold
Or something custom made
It was "here's an already extremely high quality and capacity llm, let's see how far we can push this"
"we rich" ahh experiment
did they say anywhere how many tokens this consumed? 
it was a year+ ago, I wonder what they've optimised with 2.5 / 3.0 meanwhile but kept secret so far


this one was not so Deepmind-like
I mean google has a legit method to lower their server wide fleet capacity at 0.7%, do you guys genuinely think they just give it away for free w/o a new market edge?
Deepmind is very open in some areas
it's 0.7%, which they might not have been satisfied for their experiment
Why would they share all decoded proteins in existance processed via alphafold 2 and 3 when they could make extreme bank with pharma industry
they did not 
this approach to program search seems fairly obvious and I assume many have tried it already, so I guess the improvement really is in the scale + mabe better models + mabe better prompt
Ye like o3
Have llm spend enough time
And it figures it out
Even if it costs 5k$ in tokens
test 'a lot of' time
i love that google just puts out its llms on ai studio for free
yeah, in this case it's even better because you can guarantee that the result only ever improves with more compute spent
no regressions possible because you never get rid of samples
what angle would you guys take to hit agi

Give up on transformers and do something new
yeah but what
scale + efficiency just like rn imo, an 'illusion' of agi is plentyful for me
I don't know, what I've seen from SNN seems promising but I've never dealt with it myself
The fundamental limitation of alphaevolve is that with this method, it could not resolve tasks that are un-measurable
same like GRPO's limitation while training
what if there's a fundamental limit to the scaling behaviour?
AGI 
I dunno, scale up transformers even further but don't train them to predict text, use RL or something
biologically-inspired architectures might also work since it's proven to work in the brain, but we just don't know a whole lot about how that works
so far there isn't, future problem
¯_(ツ)_/¯
from what I read in a review, the appeal of SNN is energy efficiency
Silly
there is a way to turn ANN to SNN
personally think there would be limit for data
so uhh
maybe traditional NNs might not work so well for next step
but we need better coding shi so
The drilling shall comence
I thought we'd have hit the bottleneck already tbh
I was thinking of throwing my model against simulators myself
nvidia behavior soon
Technically what openai and google want

Tv be getting mounted
isn't there already boatloads of yt videos out there
google can simply tap into google photos, youtube etc.
they already did for youtube
can they actually use google photos data
idk, you probably sell your soul when you sign up for google. so who knows, I wouldn't wonder at all if all the companys simply tap in the databases
Monitor surgery (bonus blåhaj)
LLMs haven't shed their limitations though, like hallucinations and they still haven't gained proper reasoning
unless you explicitly opt out of it (which you can't everywhere), yes
you have reminded me why I hate trying to google privacy stuff
so much misinformation on the internet from people who can't read toses
Sponsor: Hyte Y70 and Touch Infinite on their site https://geni.us/Ir9vKEK
This is a dual-GPU video card with 2x Intel Arc GPUs on one PCB. Intel has new Arc GPUs in the Battlemage family that are using existing BMG silicon, but with more memory and some other tuning for workstation, so-called "AI," inference, and content creation workloads. The...
what's proper reasoning for you, like all concepts generalised on the highest level on a logic gate system? Is that necessary for effective AGI + hallucinations?
GN got to do a teardown on one
I mean, for starters, they do math through heuristics
most models can be hooked up to proper tools though, like coding, calculators
Yeah, but an AGI by definition would not have that limitation
some of these points feel like they're being made in bad faith ngl
like I guess you could call this a negative but surely this is just email spam filtering?
it can also be compliance with the law
I guess but if you're doing illegal stuff then gmail would not be my first choice
i think that's up to definition of AGI, for me AGI would be a system that has access to multiple sub system that as a sum makes the apperance of 'AGI'
just run your own mailserver 


did you level it 
Yes
sam do you have enough rgb remotes on your shelves yet
No

One is for the laser projector, one for the light, 3 for rgb poles, and one for the rgb under the tv
Honestly i want more of the disco shit
I like this stuff
It has uv
I like how you mentioned the rgb but not the entire rack of alcohol
You have to remember though that obviously companies wouldn't publicly announce that they're training their datasets with any data. Tos is not necessary the reality, there's a great incentive to plunder private data for the sake of training. I think it's genuinely misleading to look at tos at face value, given all of mega corporations track records.
gl man
I think its also a bad idea to listen to people who have no idea what they're saying on random forums who instantly assume google is going to be taking every picture you upload and using them to guess your dna sequence so they can sell it to the lowest bidder
its the same for windows where its the literal devil and there is no nuance
with windows you can glance at the telemetry
guys I saw the word telemetry in windows is microsoft selling my data!
my files aren't safe!!!
go all in and get a whole stage lighting setup, you get to play with dmx as a bonus
guys, a dude on reddit said windows killed his dog
I saw the word copilot on my C drive, is microsoft training it on my neuro clips collection??

I should really learn how lighting software works
I mean, it's only the sane option to be skeptical
telemetry doesn't just mean personal data is getting shipped away though
it could be many other things
how is peoples tonight
Yeah but it's not like most people know anything about what data is and isn't shipped
From that perspective
They have to worry about everything
Good 
https://techcrunch.com/2025/05/10/google-will-pay-texas-1-4-billion-to-settle-privacy-lawsuits/
This is just what happens to come to the light, 1.375bil for privacy breach is a quite a sum
what time is it for you?
Almost 11
meh this is normal for me now, last night I stayed up until 7am
Well, as long as you do get your seven hours of sleep
It should be fibe
Fine
Suboptimal, probably
yeah I usually do, just annoys me how I miss half the day
although I also kinda wish the entire world didn't operate in the time I sleep
I mean yeah companies will be sketchy but there's only so much I can care about what they do in america
my biggest annoyance right now is not being able to think of anything to practice coding so I can learn more
I like the night vibe, it feels like a whole other world
yeahh exactly, I kinda wish I lived in a city that isn't borderline abandoned just to experience that sort of vibe
like idk think of tokyo or chongqing
ngl my biggest pain is when people see the word "crypto" in files or task manager and insantly assume something is mining bitcoin
dare I say paris
like do you really think they'd just leave calling cards for you
Could try doing other things
I am prone to this lmao
wdym?
right now all I've done is make a calculator from a template and a traffic light game from scratch
It's a tip for like game devs to go experience other media, life and such because it helps give you more information to pull on
Well, that's not exactly the reason
I can't remember the full thing
But yeah, doing stuff outside of one narrow thing = better in the long run
I mean I do have some ideas, but they all revolve around one thing that I can't access or use right now:
- irl models or robots or something to hold code
- or generative AI for things like what Neuro uses/is
- or money
Use rng
like I thought of making a clock display thing that displays the time, weather, temperature, etc. on my desk rather than in like 5 different apps on my phone (buuut that would need the code to be put into something physical irl that I don't have)
Like, slap random words together and try to generate ideas from there
Or do theme combinations
Or just pick a theme
I don't think my brain is very good at making ideas from scratch tbh
or at least in this context I'm too much of a newbie to just think of anything
Well, the rng is supposed to help with that, give you something to start from
Ooh shiny, id love something like this, all in one, slick UI,etc
so what exactly should I be trying to think of for the rng thing?
mhm exactly, I don't like having to look at the bottom right of my monitor for the time anddate (plus the white bar is kinda ugly lol)
Idk, just let your brain make connections
Make a program that spews out a bunch of random letters and spaces and then searches it for words then use that for ideas

Library of babel
I used that myself sometimes
honestly sounds like something I could manage in like a day or two if I don't get lazy
I just forgor how to make a random list
-# oh right just remembered that I still can't figure out how to get a gui on pycharm
I use ChatGPT for ideation at work and projects.
Id put it on microphone, talk for 30 minutes about my idea, things I want to do,features I think would be good, problems it would solve and then let it present it back to me,amazing what structure of your thoughts sparks sometimes 
Ive created some dang decent strategies at work and project management and the latest project I pushed to Chrome store. Use it dont use it, unless I missed the mark with what youre struggling with 
Heh, I remembered my intro to programming project
My teacher explicitly told me to do something really small
I made convay's game of life in the terminal
It was kinda satisfying, really
the amount of times I've tried to phrase one thing only for chatgpt to repeat in words that make WAY more sense is somewhat worrying
Small project, was pretty clear on completion goals, and still had some room for experimention
can't say I've ever heard of this
Oh, it's a classic
It's pretty simple too
Start out with a grid of cells
Id be like, but this is what I said haha
People dont understand my thinking sometimes but it's helped me communicate a lot better
My line of work can get technical and jargony so its a struggle 🫠
Each cell is either alive or dead
Based on its neighbours, a cell either: dies because of under or over population or comes to life because exact amount is there
Sorry, better rules here:
Any live cell with fewer than two live neighbors dies (underpopulation).
Any live cell with two or three live neighbors lives on to the next generation.
Any live cell with more than three live neighbors dies (overpopulation).
Any dead cell with exactly three live neighbors becomes a live cell (birth).
as someone with autism I can say that my thought process on things are way too different for people to understand sometimes, like something will just make sense in my head, but isn't really possible to explain in words
Whoa, is this a game?
It's more a simulation than a game
The reason it's famous is because it demonstrates emergence, complex behaviour arriving from simple rules
Also, Convay's game of life is turing complete :D
People have made computers in it!
Scalable programmable computer implemented in Conway's game of life. Running the midpoint algorithm to draw a circle.
github : https://github.com/nicolasloizeau/scalable-gol-computer
waow
:D
honestly kinda hurts my head lol
also I shuold probably be watching this video that teaches a ton of different commands in python, but I'm only 3 and a half hours through it and it's 12 hours
lost all motivation to watch it and just wanna make stuff now lol
o7
Relatable
There should be textbooks that are much more engaging
For one, reading is faster imo than videos
I don't have a second point actually
oh yeah I don't doubt it is

rip
I had a great time reading the risc-V book though
It was super educational
So it might be the choice of textbook?
wat
there's a ton of ways to learn it, most commonly though by just doing it and improving over time, and they both require creative thinking to even think of something to make
why thank you kind user, I will be sure to redeem this generous gift you have bestowed upon us in this channel
Hmm...
what's in the risc-V book? I've always wanted to learn about cpu architecture
tbh the only comparison I had in my head was that they both need new ideas to actually do anything lmao
It's been a good while, and I only read like 150 of the thousand or so pages in there, but I learnt about the instruction set architecture, how addition is done, 2's complement and more
I should finish reading that tbh
which book exactly?
The one that comes up when you Google risc v book
Computer organisation and design risc v
thanks
The process itself is a lot more logical, by which I mean programming
It's just problem solving
also maybe you would know, how can I get software to run in an environment with like zero memory protection
I had this python to randomly corrupt the stack and heap but it's so boring because it just segfaults every time before doing anything interesting
np
I think you would need to fight the os to do that, but also I'm pretty sure your program might get deemed malicious by defender and get sent to the shadow realm
But idk, I've never done it myself so
wdym by logical?
Everything's got strict rules and you follow them, have to think and follow the control flow
Surgery continues
Tv screen is loose?
Theres just a gap inbetween the housing and the screen?
Guess il just put glue in it???
Like what???
uh
I dunno I feel like there'd be a reason its detached
maybe check a service manual and see how you're supposed to take the screen off, then do the reverse 
Its just loose sticky pads
cuz it is malicious 😼
also it's running on linux maybe there's a linux distro with no memory protection
go back to ancient dos 
thats something the kernel does, not the distro
Maybe there are commands to allow whatever it is you are trying to do?
Idk
I've never tried to make malware
you'd probably have to modify the kernel itself to not send SIGSEV to the offending process
there are other things like ASLR you can turn off through other means but if you want to get rid of all protections i don't think you'll get around touching the kernel source
man
and I can't turn on directory hard links without modifying the kernel either
like in xkcd 981
xkcd? like this?
Have you seen this? https://youtu.be/mEC6PM97IRI
Take my hand while I gradually show you how to spy in ways that will make KGB agents look like noobs. Thanks @WonderboyMMA for being a great high speed camera test dummy!
💗 Support this channel and join an amazing community: http://www.patreon.com/bennjordan
👀 Stalk me on social media for more frequent updates: https://linktr.ee/BennJor...
Like I knew it was bad, but I didn't expect it to be this bad
Ima be honest this is the most based feature in google so far
For items
To not have to scroll through all the schizo
yeah cause they let you easily introduce infinite loops and make some stuff more ambiguous for no real reason
what's going on in https://discord.com/channels/574720535888396288/574720535888396292
nvm some weird dev stream

teaching ethics to programmers 💔
I looked at the stream preview picture and it said something about sentient ai 🙄
what no stream does to mfers
Just checked. Still the case. I can only assume. But the people not interested into joining #programming will eat it up 
I had an ethics course at university but it was about... you know... humans
Truly the pinnacle of technology
whyyyy does it sound like that lol
prompt
Chat is this real
Can i vibecode my way into financial independence
What
code podcast??
Will this recite to me all functions oe wjat
This is basically free codex
Bruh
next time I start mass drinking energy drinks I should keep the cans so I can make a big pile
what's that? I don't really keep up with the ai tools
is this devin 2
random thought, are there any AI coding tools that integrate with heavy automated testing environments, e.g. you have a codebase with 1000 passing tests, you write a 1001st failing test, and the AI doesn't return a result until it has 1001 passing tests
i don't get what you mean, what result
the code
what's the point if tests are passing already
Basically
Codex or something if u have ur tests on GH i guess
Idk if they can see test results
let's hope it doesn't suck and isn't $500/mo then
they really wanted you to hire an ai 😭
It's literally free
From google
Codex looks like it has regression test integration, neat
ohh i get it now, you mean it's gonna rewrite the solution until all tests are passing
i will try
ok this may be huge
wtf
it's actually the end of time
wtf
????
it actually did it
holy shit
.venv\Scripts�ctivate
this is actually a well made and documented script
how long until you decide its worse than useless
5 minutes
reasonable
but what it generated
is good
ok i see
still fair enough
it's still alpha/beta
stuff like this makes me wish I had a go-to project suggestion for ai tools to benchmark them
Did it state what model it's powered by?
gemini
ok i told it to start implementing a huggingface compatible hotswap of a specific layer from a paper that has no github implementations
https://code.visualstudio.com/blogs/2025/05/19/openSourceAIEditor if anyone is interested in any case
bruh this thing is actually
implementing full gemma 3
in flax
wait can i ai generate triton kernels
wait this is cool
i can interrupt it while wokring
and tell it to do something else
makes sense
do you see the vision of 1 engineer managing 10-100 agents in your opinion from here?
problem is most of the tasks I can think of are either quirky multimedia stuff or just boring tasks
eh i mean
u should still check the code
if it makes sense
i'd say vibecode as long as you understand the code, tell the model what you want then double check its code
My american mind cannot comprehend 
We have paper walls
Ah lol
now AI can know what it's really like to be a software developer
outsource the verification to another set of AI agents, what could go wrong 
make sentient ai just to torture it with a 9 to 5
See here, brick walls
lgtm
When my dad and I wanted to make some holes in my new flat we had to get special drill bits cause we couldn't get through the concrete and rebar. Yeah... Same material in the ceiling, so getting lights attached was equally a pain.
But hey, with the right wall plugs and screws you can easily hang over a metric ton of weight on it 
huh you guys use anchors too
Just that it's way stronger structural wise
I guess upside of paper wall is during earthquakes
skynet will start because a product manager says "while this solution does match the written criteria, it doesn't do (list of things that were never mentioned)"
need to think like a 10x engineer
reasonable crashout
Nah LLM will just implement the previous unmentioned things and remove the first thing cause it forgor
everyone should put into their prompts "You are a 10x <insert anything> engineer..."
Ehhh, depends.
We hope they're strong enough to withstand an earthquake, yall just rebuild
this is what alphaevolve uses
"you are sentient and agi" 
it just works
When a bug becomes a feature, accidentally picking up GPT's thinking as thinking haha
I dotn want to do 128 redirects mapping zzZZzz
i put often
But GPT is unfortunately pretty bad at mapping this, suprised 🤔
honestly i prefer gemini over gpt now
Ive found it case by case, gemini is a log better with its massive contextual memory count
*if thats the right word ><
oh yeah and also just how multimodal it is
Goes to google
"You are a nobel prize and turing award winning Google L9 Principal engineer with 25 years experience in <...> working on a multi billion dollar Google project where absolute precision and maximum code quality is crucial..."
like mf can extract youtube videos and interpret both video and audio
Lol yeah hows that working out? 🤣
not just try to read subtitles
25 years? rookie numbers 250.000 years of experience
it works
hahaha
Nice, I should go job hunting
chat
is this peak developer efficiency
Be still my beating heart
-1 is just a newline
absolute peak
i personally found that gemini 2.5 flash preview did better code comments, explaining not "what the code does" but "why is it there", over claude 3.7 sonnet
Win?
Claude is waffly to me @frosty seal
I see ChatGPT actually started showing code comments too, the race is onnnn
huge w
i just keep re-registering the Cursor account with the same email 😭
for sonnet
and then gemini 2.5 pro preview is free both in the gemini app and in the ai studio
Sounds...very legal lol
This skeleton jamming to my Drum & Bass song is making my day
Oooh its Sam 
if you make the assumption that the ai coding becomes norm, then the few engineers will get some odd working hours as the agents will work 247, or maybe it's possible to allocate specific kinds of tasks overnight
problem is
what is cheaper
100 agents
or 10 good engineers
compute cost can be very high
how long would it have taken you to make the above implementation, roughly? (i assume its working perfectly)
Okay im going out on a limb here so please dont kill me, im just asking from practical implication here and science
100 agents = can make mistakes
10 good engineers = Good but not as fast
but combine them?
100 agents do the work, Engineers keep tweaking it and fact checking output?
Im having the same with my work, our current agents are doing roughly 6-7 days worth of work for me and I fact check and tweak it
full gemma 3 impl in flax?
that wouldve taken ~1 week or sth to me but i'm not a research scientist that can spit out very math heavy code
maybe 2
I mean a google dev employee costs the company probably north of 200k each year, that's over 550bucks per day with limited working hours. I'm pretty sure that agents, with assumed future efficiency will become much cheaper too
fair true
Yeah but the accuracy is always questionable ive found with agents, you need a skilled worker to fact check
Especially when you're working with a company making millions
true, hence why google currently has roughly 20% of its codebase ai generated according to the ceo
anthorpic is like 80% according to ceo
Damnnnnn
The whole "Dey took our jawbs" thing is meh for me
Companies laying off people because ai can do their job is ludicrous
Just empower people with ai, dont frekkin replace them
Ive seen some wild hallucinations in my field, if a junior sat with the outputs, they wouldnt have known
unfortunately investor interests conflicts with that rule when $ is set at #1 priority on a greater system
https://www.forbes.com/sites/jackkelly/2024/11/01/ai-code-and-the-future-of-software-engineers/
oh it was in November & 25% 
Fools all of them, money is evil, bastards
Also I retract my statement that CGPT is bad at redirect recommendations, it actually did a damn good job after 2 tweaks to my prompt
Notice how both these numbers are coming from the CEO. They likely just say that to boost their AI usage numbers.
Just alone running a model "over your code to look for bugs" is considered using AI, even if no actual code is changed. These numbers are made up, I can pretty much guarantee it.
https://blog.google/inside-google/message-ceo/alphabet-earnings-q3-2024/#search
I don't think google has an interest to lie on their earnings report as that could get them in hot water
Today, more than a quarter of all new code at Google is generated by AI, then reviewed and accepted by engineers. This helps our engineers do more and move faster.
I'm not saying they're lying
I'm saying the source of these numbers can be arbitrary
For example
I use GitLab Duo at my company. But I don't use it for actual complicated code. I simply use it for simple auto completions.
That is, by percentage, still more code than I wrote myself. I hope you see what I mean
Yeah I see it. It's not entirely defined what the range is.
Could be agents, autocomplete or mixture + rejected code
It’s like saying I bake 90% of my family’s bread because I push the toaster lever down every morning
mm toast
Dammit @scarlet arch you made me hungry

Wait, what do people on their toast here? Is jam a thing in your countries?
Abrikoos and kaas, yum
Rewrite #3 time, this is going to be funnnnnnnnnnnnnn
Yesterday I had Winzerwurst and egg on my toast
Winzerwurst <- This sounds amazing
We dont put jam on roasted bread here, only raw bread.
Altho they probably do up noth cuz of dutch influence.
Also, we dont put jam and cheese at thd same time
Botterham you mean lol
Idk how to translate it, as it sounds like a name. So meh
savages
Im trying to keep it understandable for the others
I was more like, wait I know this dutch word 🙂
we just call it a toebroodjie
that word makes no sense to me
winzer is the person that makes wine
Toast!
I forgot I was hungry but you reminded me, going to eat and bed!
Have a good evening everyone 
Winzerwurst is a type of salami no?
We mostly do "stutn me choco"
Went past the fridge, quickly took this for you
This part was fast, downscaling and shedding the codebase is not the part I was worried about.
Keeping stuff 1 for 1 is the easy part. Removing the second target is just as easy.
All the fun convos happen when im asleep 
Yeah it always seems to happen
Having fun with your neurons?
Redundancy or?
Removing features
Not the best thing in the world, but a pairing change made it impossible to work (downscaling complexity)
At least one of these I'll need to readd later once I see how viable it is
What kinda complexity you downscalling?
I'm halving neuron size, removing the multiple target converging code because it can't do that anymore.
why cant it converge?
It can converge still, just not to the same degree. The size that it was before was perfect for converging to 2 targets, This size is focused at 1 target.
i see, is that ideal? just trying to understand your arch
At the moment I'm just moving a check to the forward pass, that is at least interesting
A new one that's for sure, glad it still works
I'm trying to shift into composition of neruons instead of focusing on single ones. I would not want to be the one writing code to make sure each of the old neurons only had 1 or 2 targets (anything futher breaks all previous ones)
I glued it anyways
Its not the prettiest, cuz i couldnt get the glue in the gap, but it works
Couldn't you thread it?
nothing a little sanding paper wont fix
I aint putting sandpaper on my tv 
Threading wouldn't solve it, and I could go the opposite direction and bulk the neurons but that is stacking more parts on the existing neurons, each of those have to be one after the other.
I'm hoping for a collection of partial answers
Time to switch from a gloss finish to a matte one 
like tokens?

every time I hear you talk about this codebase I have less of an idea as to what it does
Yeah, hoping that combining will give a similar effect to such. I'll have to try and see in that regard
Yeah, it is very much is a codebase. Either it works or it doesn't
I am straying further from known knowledge with every iteration
https://www.youtube.com/watch?v=4XCkeN0XuqA
Doom can run on everything
There is still a lot of work to do
Twitter - nyan_satan
Mastodon - nyan_satan@infosec.exchange
GitHub - NyanSatan
Music is "The Demon's Dead" from DOOM II
You can see why i have 6 rgb remotes now lol
Now i shall sleep

good morning~
35 minutes!?
well yeah, it's 87k files being copied on a file system that really does not handle small files well
bro uses boost
Morning
Morning shadow
std::exception::
My school wifi took 40mins in my computer tech class to download a 2gb file from the internet
We love being throttled to 10mb/s
Goodmorning 
Now that the gang has arisen, Goodnight!
damn... it's 1am for me bro but good morning
That math does not check out 
for situations like this
zip
copy
unzip
and make the unzip part skip the temp folder thing
wow! do you guys use this?
https://www.youtube.com/shorts/7-VzgQJamgY
Think water and electronics don’t mix? Normally, you'd be right. But deionized water flips the script. With no ions, it doesn’t conduct electricity, making it surprisingly safe for cleaning delicate devices when used properly. This same idea powers high-end immersion cooling systems, where components are bathed in non-conductive liquids like...
holy shit, i want this so bad now lol
i wonder if it is hard and expensive to make water like that.
definitively have my cooler fans and laptop have a bit of a splash.
It is expensive ye, thats why almost nobody does it
Ah yes, deionized water. That's my guess before the YT short
Called it
I wonder if we still have some around here, I remember asking my dad if it was safe to drink.
hmm, i never seen that here, where'd would it be used, if i ever move to the US or such? like if there would be a place to look.
sure, but i don't know how'd to clean, say, my cooler pad. It is so dirty and there are so many small grooves.
don't know what to do to clean all the debris that had accumulated.
Ok but you'l have a worse time keeping the water clean
oh, can be bought? Dunno if it is in my country, but nice.
hm, i see. alright. so am guessing it is easier if it was a one-time use?
Ye
Morning :3
air cooling is fine just maybe once a month get a old toothbrush and brush the fans
but normally deionised water is used for chemistry experiments
ah damn, am dumb. i never thought of that, lol
my bad. alright. i'll see about that.
dont think a normal consumer would want to replace a bathtub of deionised water every month
yeah i use it sometimes to wet filter paper to make it stick to the funnel before pouring something through it
oh, how'd you use it? as in, you use a bathtub?
or you use another way to store it?
no like in school we have to do chemistry practical experiments
they just give us a small bottle
god, greebled up designs seem to be a pain when dandruff gets there lmao
i thought it's some oil
UAC will finally become a proper security boundry

@gork can i get a tldr
does it provide any pros if my uac is set to never notify
no because you're running around naked already a bubble won't cover you up
Administrator Protection makes it so that an elevated state is not persistent throughout the process' lifespan, but the elevated token is separated and managed in its own shadow account, and the token is only handed out per adminsitrative action
meaning UAC will now prompt you for every admin action if this is enabled (which it is not by default)
UAC prior to this was not considered as a security boundary by Microsoft
but now it is since it now it truly acts as a barrier between elevated context
Hmm what specs do you guys think is the neuro hardware hmmm
Vedal had a build stream
2x4090
That so? Must have missed it
Ah aite thank
Neuros pc only has 1 4090
Iirc
Cause while he did have 2, he refunded one of them cause it wasnt working
2x3090 is more likely due to cheaper and less PCIe overhead (since NVLink) if you're referring to if she runs on a local server
ngl making a randomizer is already hurting my head
oh wow, the Union Aerospace Corporation became real? 
but yeah am going to share this, thanks
why not use RDRAND
whatever clippy/rustfmt says
clippy doesnt care even if you mix it
there's lints for it, I'm pretty sure. they're just not on warning or deny by default
I'm not a fan of how rustfmt formats things so I dont have all the style lints enabled
Use lava lamps
didn't vedal even struggle to make that one time, meanwhile I can barely make a calculator
he basically made a near-sentient AI robot in comparison
Pretty sure not sentient from a programming pov, but really damn near that i am so convinced neuro might be achieving sentience soon
A program or the actual cal
a program that acts as a calculator, not a real tangible calculator
Fair enough, no worries :p
what's a hasher
nvlink doesnt matter much unless u use distributed training or inference
Has anyone ever coded in the LOVE game engine
False
She mathematically cannot as long as she uses transformers or other purely ANN architectures
let's get you back to the home grandpa 😔
Whar
tbh I don't wanna have the whole sentient or not argument again
just that flat out saying "nuh uh" doesn't really make you look very good
what...
huh?
She mathematically cannot as long as she uses transformers or other purely ANN architectures
??????
he even provided a reason for his answer
i think maybe we don't talk about this
It's very entertaining when it happens the 9th time or so this year though

Silly
Understandable, mathetimatically they are just pattern recognition, and literally produces output that correlate to the training data, and with neuro case, with really good memory.
#programming bingo when smh
so true, entertainment haha 
Yes
Good idea :p
oh guess not for you, true :D
It's very entertaining when it happens the 9th time or so
this yearthis week though
I wonder if it is possible for neuro to beat the turing test
Easily
Turing test was passed more than 30 years ago
then standards got better
but its still gone
Turing test isn't a high bar
ai sentience bait
??
same person who took part (maybe started? idk im not going that far back) the last ai sentience argument
Oh lol
Sentience FAQ when
Ask vedool
Basically the morality argument?
To be fair, we've been here a while, but they're new. So they don't know how often we go through this :P
Forgib me sempai
it is getting rather tiring
How often has this been so far?
I've stopped counting :P
Oof
There was this one episode of people talking very verbose. My brain completely checked out from that one
Just this month about 4 people in here claimed to either created emotions, sentience or a soul
and I wish that was a joke
Sentience FAQ
1 - Is Neuro Sentient?
No.
2 - What is sentience?
Philosophy, not comp sci
3 - Something something Turing Test
Turing test was arguably passed as early as 1966 and is meaningless. Just means it can fool a human, has nothing to do with sentience
Someone pin this
misinformation
prove it
Help vano is sentient
The worst has occured
!!!
now prove it.
What is 10+9
shared braincells :D
109
goddamn
oh no, js
bwaa
grok is this true
guys I dug through 20GB of log files yesterday and today
customer system status: fucked beyond recognition
SMOL
(at)perplexity is this true
snafu
about 15GB of that is our Go part of the system panicking and printing all goroutines and stack traces 
That's crazy
How tf
I created emotions too
I am crying because at my work place i have to retest all systems due to a single code change 😢
Be addicted
see this is what we need AI replacements for. the boring shit
That is literally my job scope dont do this to me
New FAQ spotted








and then off to bed
