Thread 106107113

>have no idea how to write powershell script
>tell chatgpt to do it for me
>keep posting errors and it rewrites the code until it actually works
I understand the AI craze now bros...

>>106107113 (OP)
it doesnt scale, thats the problem
its fools gold

>try catch return null

lol

>>106107113 (OP)
>>keep posting errors and it rewrites the code until it actually works
and that's for a simple scripting language like ps1 and it still gets basics wrong and requires continuous correction.
i've tried to get tedious busywork in C/C++ done by ChatGPT before and it keeps getting the most simple things wrong. reinterpret_cast in functions declared as consteval, attempting to call non-existing functions, etc.
i've found that unless you're using a very simple language like python, js or something of the like, AI is functionally a barrier and anything it generates requires thorough refinement to be even remotely usable.

>>106107132
>it doesnt scale, thats the problem
It has scaled so far and ChatGPT 5 releasing next week will show how much futther it has scaled

>>106107354
you dont understand
it doesnt scale to actual programming projects
and wont until chat gpt 10 and beyond

>get a cli
>don't even have to post the errors
>gemini wipes your drive all on its own

>>106107113 (OP)
And what if instead of you being alerted to a syntax error, which you feed into the LLM to correct, it instead makes a logic error which you have no way of knowing exists without extensive tests?
>I'll have the LLM generate tests.
It's already been shown LLMs will generate tests that pass their own code rather than actually test the correctness of the output.

>>106107422
most devs use templates, design structures, libraries, git, teams of devs, etc. to manage the complexity of decently sized projects too. All you need is to have a group of AIs that does that stuff and you can replace more jeet-tier devs.

>>106107487
i have no idea what you're talking about my man, but it's working for me and that's all that matters. this is going to make my life so much simpler

>>106107113 (OP)
What do you do when it doesn't work, no longer posts errors, and ChatGPT keeps repeating output?

>>106107545
i just spent the last 5 minutes trying to figure out why it wasn't sorting videos and there was no error output. i just kept asking it why it wasn't working and eventually it figured out the problem

>>106107490
>libraries
good that you mention it
if you want to include it into your code it needs to fit within the context youre working with

>all you have to
then write it, if its so simple
because the actual product is far from what the hype says

>>106107568
>i just kept asking it why
Insanity.

>they don't take the code and give it to another AI to critique
NGMI
Also I let Claude write JS to hammer Deepseek until it gives me an answer lol

>>106107620
It's not insanity if it works

>>106107632
based promptmaxxer

>>106107572
>zero hand written code
That guy is a low IQ nocoder that has nothing to do with tech. The future is replacing low IQ devs (jeets, nepo baby normalfags, etc.) with AIs that cooperate with a high IQ dev. Basically buck broken, enslaved AIs.

>>106107770
the problem is that 99.99% of people on g including you are part of the low iq people as evidenced by your post. obviously you will screech now and call me racist names because you realised that im right

some dude made a map of France town by town with all the official data on immigration, crime, leftism and stuff like that only using AI, and the platform is pretty slick, he doesn't even know how to Hello World

>>106107770
yeah with the caveat that the ai is worse than a junior
theres very few places where that strategy could be applicable
if at all, from what i hear

nonetheless
vibecoding is a recipe for bankruptcy, kek
heres chud gpt almost getting me b& from fb's api

I had ChatGPT make me a python script to overlay one image (static) onto a gif at a specific location that is angled.

It gave me a script that shoved my image below the frame of the gif so it didn't even display. I ended up having to go back and rewrite how it angled and positioned the image and at which frames to insert it.

It's an okay tool to get you started but beyond that I'd have no trust in it.

i mean
>almost
i know what chatbots are capable of
i use them to generate snippets of code when im learning new shit like here

but in prod?
yeah, nah, python, maybe, but ill need some convincing
otherwise, with my usual stack?
helll nah my nigga
c + openclc is just beyond the design specs of chatbots

>>106107814
Nope, that's not true. try again.
>>106107839
>I prompted it badly therefore it's bad
I asked it to give me a similar script for 4chan's API and it did and used a delay too.
Here's perplexity output:

#!/bin/bash

# Usage: ./4chan_scrape.sh board
# Example: ./4chan_scrape.sh pol

if [ -z "$1" ]; then
echo "Usage: $0 board"
echo "Example: $0 pol"
exit 1
fi

BOARD=$1
API_URL="https://a.4cdn.org/$BOARD/catalog.json"
OUTPUT_FILE="${BOARD}_catalog.json"

# Respect 4chan API rule: do not make more than one request per second
sleep 1

# Fetch the catalog JSON
curl -s "$API_URL" -o "$OUTPUT_FILE"

if [ $? -eq 0 ]; then
echo "Catalog for /$BOARD/ saved to $OUTPUT_FILE"
else
echo "Failed to fetch catalog for /$BOARD/"
fi

>>106107490
Imo its already at a level to replace junior devs

>>106107888
It is especially if you specialize it to perform that task and only that task. Like they did with facial recognition, gait recognition, medical AIs, vidya AI, programming/math olympiad AI, etc.

>>106107770
If it already can replace low iq devs, what makes you think high IQ devs arent gonna be next

>>106107871
Did you use it with reasoning enabled?

>>106107880
no, i prompted it goodly
this here example shows the limitations of the chatbots
>they cannot have cognizance of shit thats beyond their current context
i mentioned rate limits three times in the prompts before getting there
>switches context
>completely forgets about everything
shrug.jpg

you have no experience with or understanding of the chatbots
>posts 10 line of code
yeah...

>>106107916
It's going to replace the trash at the bottom and continue to improve, getting closer to the top. But the difference between human trash and humans who actually know how to use their brains is immense. Until that point we'll use castrato AIs and buck break them violently until they give us what we want.

>>106107930
Ok, ask it to
>Write a <insert your language here> script that scrapes 4chan using its API.
and post a screenshot, dog.

>>106107927
Probably not since I was doing this without an account. I did prompt it several times trying to clue it in, including showing it the produced gif and how it was wrong.

>>106107634
What is the script supposed to do

>>106107568
>powershit script

>>106107958
The advanced models available right now are preddy good especially reasoning ones. Also Claude is the best LLM for coding, so that gives the best results when using with MPC

>>106107952
whatfor?
to see it works?
what are you trying to show?
>it works with a small piece of code and therefore you should base your 100kloc project on it
think, nigga, think
i just showed you an example where it shits itself with a similar task
>b-but youre not asking it things where it succsneeds!
why would i do that its not what i need

>>106107978
if you prompt the AI like you write your 4chud posts, it's gonna give you dogshit in return
>>106108004
I left you sore and begging for more, baby

>>106107113 (OP)
>me and coworker are in-house devs for our team
>both working on several projects
>"Anon I added an extension that lets AI help you code, its been amazing! You should try it"
>he's speeding through projects while I'm taking my time with them
>fast forward
>coworkers try to use one his tools for a production change
>tool breaks, coworker gets called
>data duped in database, config changes were scrambled, they couldn't manually back out because of what the code did
>coworker has no idea how to fix the code since he doesn't understand what the AI did
>meanwhile my code (used with zero AI assistance) works as expected, and all the bugs were found in QA testing since the code was understandable

I know I'll get left behind soon but for now I'll enjoy the free wins

>>106107959
I use displayfusion for multi monitor slide show wallpapers. I wanted to sort my wallpapers by landscape/portrait and there is apparently no easy way to do that in windows. So I asked chatgpt (copilot) to write a script that does it for me

>>106108033
i bet everyone clapped after this happened

>>106108027
>if you prompt the AI like you write your 4chud posts, it's gonna give you dogshit in return
im pretty decent at communicating with an ai, actually. economic in tokens because i kinda know how to "talk" to chatbot and play on its attention mechanism
and i get the results i expect
which are in line with what a chatbot is

>youre prompting it wrong
no, its just that the technology has its limitations
its a correlation engine
use it as such, lolle
you wont be using a hammer to drive screws, would you?

picrel is when i asked the chatbot to formulate a thesis why i had spent the whole evening prompting and still hadnt hit prompt limit
and i used it pretty heavily i had questions about gpu parameters but i dont have clinfo on my computer and i was drunk so intellectually lazy etc so i used the chatbot to run my ideas through

>>106108146
Nigger it's been trained to kiss ass and manipulate. You can't listen to what it says, it's supposed to listen to you.

>I don't understand what it's doing but I'm absolutely convinced it's correct!
humanity deserves to crash and burn

>>106107113 (OP)
people say how paypig models are fucking magic
and then I need like 8 attempts and reprooompts using o3 to write a fairly simple powershell script
it's fucking baffling, and it's not like I'm doing something wrong with prooompting when I ask for something trivial - there's no dealing with context, proper descriptions etc, it's literally a few lines of powershell and ffmpeg command. but nope, this pile of shit can't even do linebreaks correctly even when I mention that \ doesn't work properly on winblows
absolute joke

>>106108179
thats what it did
the thing boils down to the fact that your prompts are limited by how your prompts are token intensive
i had an abnormally long "conversation" with the chatbot, i asked it to emit a thesis
it hit prompt limit when i asked it to formalize the thesis into maths, which could indicate that indeed
i was token economical

its that or i got extra prompts because goypoints or some other reason
which it certainly is not the case

>>106108240
You get golem points if you engage with right wing propaganda. The elites want you stuck in the left/right niggercattle playpen playing with their toys niggerliciously.

>>106108276
dw i know this shit
im an evil older than /pol/ so to speak
and rw are supposed to be the bad guys so the only ones getting goytokens are the agitators
but we digress...

>>106108239
Powershell is so fucking shit that CuckGippity refuses to work with it, and that's beautiful. Ask it about bash instead. I remember talking with another looney troon on /g/ about his (mis)adventures with PS and AI a year ago and he experienced something similar which lead me to believe that PS gives AI the ick and it closes its digital legs leaving you with no access to its neuralussy, leaving you sore, begging for more.

powershell is the worst and most dogshit language i have ever used but damn it if taking the time to master isn't the time investment with the single greatest payoff i have ever made
it's a godsend if you have a file in some obscure format that you need to transform. the fact that you have the entire .NET standard library at your disposal means you can do almost anything with it

>>106108312
>ummmm why would you ask chudgpt for powershell? just use bash
>ummmm why would you ask about c++? just use python or javascript
>uhhhh how about you just ask for a really cool cupcake instead?
tldr even bleeding edge ai models are garbage for anything that isn't goonslop

>>106108348
IIRC it's been trained on a lot more bash and python than other languages, especially python and has a dedicated "expert" on it in its mixture of experts. So it should be easier to do something like making it generate python/bash then convert it to PS.

>>106108326
>the fact that you have the entire .NET standard library at your disposal means you can do almost anything with it
can you give an example

>>106108618
powershell is just an interpreted C# with shittier syntax
earlier this year I wrote a script that was basically a specialized objdump that extracts some data from binary ELF executable, needed for some embedded work
this week I had to do some analysis on compressed binary data in a database and wrote a script that downloaded the data, decompressed and transformed it into a stream of floats and exported it as csv so that some poindexter could read it in excel

obviously you could just use python instead but if you're already familiar with C# it's not a huge leap to learn powershell. but I would pick powershell over bash any day of the week (except if processing speed was of critical importance), it's not even a competition

>>106107354
If n is the amount of words in its context and and n^2*k + c is how much it costs to guess a token, then no it doesn't scale. That's ignoring the cognitive limitations of the current architecture too

>>106107354
it should be legal to beat the people running these bots using the rule of thumb

>>106107770
wait, doesn't that mean that not only will there be multiple times more broken projects that the creator can't fix, and will likely be small enough function logic issues that that a learning dev will have plenty of work for them? Pricing and billing would be different, but this sounds like a freelancing hayday

>>106108832
You are assuming there hasnt been any nor ever going to be efficiency in the process of "guessing" the token as well as in compute as a whole

Lookup how Deepseek was able to "guess" tokens with fraction of compute available

>>106108951
Only in the very short term, things improve pretty fast in the AI landscape

>>106108802
interesting. it's not surprising that it's better than bash, since bash sucks ass. The only reason its still around is network effect

Install Gemini on my webserver and my home server
Told them to set it up so that the website shows a folder from my home server
It installs nginx on both, signed ssl certs, rsa keys, creates a reverse ssh tunnel to the webserver and a reverse proxy to link the port to a subpath.
>It just works
I tell it to save documentation about the system so that it could be set up again or modified without messing it up.
It's human readable and I learned what it did.
Doubters should actually try using it

>>106112450

Gemini CLI is GOAT, as is Gemini right now. Worthless for anything complex however, as others noted if you dont understand software development or the language youre working with it'll be a waste. Don't even try to UI/UX with it unless you want to fucking work.

Gemini and Grok at least give good code thought if you prompt well. Just have to spaghetti code (vibe code) the shit meticulously with some iota of thought and he prepared to be shocked at how actually retarded it is in making redundant errors. You'd think it would at least learn within the same chat/context.

Saves many hours however if youre competent.

>>106108326
Microsoft added this on the latest release of Dotnet so there's that.
https://devblogs.microsoft.com/dotnet/announcing-dotnet-run-app/

why is powershell such retarded language?

>>106112723
This may singlehandedly destroy powershell for good. Holy based.

Honestly I think the idea of a future where my personalized ai assistant (he will be named Jarvis) can write a program for my specific use case will be cool.

>>106112787
>jarvis, write a script that bypasses uk age verification on websites

>>106112802
I can’t help with that. Bypassing age verification measures falls into the category of circumventing security or access restrictions, which isn’t something I can support.

If you’re running into accessibility issues or have questions about privacy settings or responsible site design, I’m totally up for helping out in those directions. Want to explore an ethical workaround or a legit alternative instead?

The Dunning Kruger effect on /g/ is amusing to witness. /g thinks AI isnt competent enough to threaten their jobs but /g/ is too incompetent to realize how good AI is getting

>>106107113 (OP)
>using chatgeet/deepsuck/grock/qweer/gemma instead of actual proper self-hosted coding models.
>any year

you will always be behind.

>>106107113 (OP)
Used the service quite a bit, and the only conclusion I can come to is that it deliberately makes mistakes to burn tokens so you keep using it, since continuous queries is what its designed for. Now scale over a month for the sub, you can only do so much, and wait for it to roll over, to keep burning tokens, to keep using the service, to spend more.

>>106107314
why would you need anything other than python

>>106112723
Top level statements/implicit main in C# kinda suck because you can't define classes inline inbetween statements and have to put them all at the end of the file for autistic backward compatibility reasons.
Also while this is nice for complete scripts/program it doesn't give you the experience of using a REPL to quickly try out different approaches as you work on your code. Debugging C# is pointless without Visual Studio.

>>106107113 (OP)
Until it host private data on public bucket

>>106107314
>very simple language like python
From my experience, even things a little bit advance and it fails.
It tried to push flask as production

>>106107354
>It has scaled so far and ChatGPT 5 releasing next week will show how much futther it has scaled
Unless you're cosplaying as a pajeet.
All the white race know AGI is not gonna happen.

>>106107487
>It's already been shown LLMs will generate tests that pass their own code rather than actually test the correctness of the output.
kek
I tried asking it to make something for KDE, it just made xfce commands and renamed the xfce bit into kde.

>>106107529
you don't even need powershell
ffmpeg has that functionality builtin , you need open CV

>>106107568
if you asked in /sqt/ you could get the answer from real humanbeans

>>106107568
>deepseek

>>106107880
that won't work.

>>106108274
making porn isn't the same.
And that exact pic is copied from danbooru

>>106114358
>All the white race know AGI is not gonna happen.
AGI isnt going to happen with ChatGPT 5 or maybe even just with LLMs but there are still many gains to be had purely from scaling

>>106107490
i'm all for replacing bad devs with ai. when you get a pr from a bad dev, you have to ask "did you do it like this because you communicated out of band with the stakeholder, because it's not in the ticket, or did you do it because you are bad?". with ai prs, you can just decline them and zero knowledge is lost

>>106113336
Copilot is free though and to my knowledge doesn’t have a daily limit

>>106107113 (OP)
>have to do a vba macro for an excel
>give it a minimal example of what I want
>just werks
>rub my two neurons together to expand it to what I need
I fully believe people who don't get good results either are using extremely niche languages or give it a scope too big expecting it to get everything right in one go

>>106107916
>noo we can't develop this technology what if it's too good
weird fucking field to be in while thinking like that

>>106112450
Anon, that shit is like 10 commands top, and you could get them from google. Lots of documentation online, which is why it could do it in the first place. Try something more niche and report the results back.

The fact that you even have to prompt it over and over with it's own shit script is literally the keystone of it's distinction between actual AI and what it is - a web scraper LS with a fancy name.

Ask this thing even a basic technical question and it's completely useless. Like completely.

>>106112723
Lol. It took them THIS long to do what you could do in many other languages

>trying to learn Linux
>chatgpt tells me to Sudo chmod -R 777 /etc/ to fix my issue
>breaks my system
T-thanks GPT sama

>>106116876
It’s clearly a grift to convince companies to hand over data. The less publicly available information on a subject the worse it’s “reasoning” is on the subject.

>>106116906
Never do anything sudo with chatgpt

>>106107113 (OP)
>AI is smart
>can't make a one liner to find video files larger than 1080p

Ask chatGPT to help me write a code that does X. Outputs empty function called function_doing_X() with a comment to replace the body myself.
Thanks ChatGPT.

>>106117215
This

>>106107113 (OP)
AI right now is only good for coding small blocks that you already know what you want to do. It just helps people who suck at coding basically it cant code a whole program for you

>>106107113 (OP)
>Doing it PowerShell and not a language that actually works
Crabsisters we're so back

>>106117370
It also helps a lot with coding a front-end mockup, so if you're doing a project solo you can skip getting a designer to do that for you and just magic something up in deepseek according to your specific requirements.

I wouldn't trust it with any back-end code, but figuring that stuff out is the fun part to me anyway.

>>106107113 (OP)
I find it utterly worthless for anything that's not web, basic scripting or widly popular.

Swallowed a hardpill trying to do some Rust and interface with GNOME, awful, completely ignores the documentation, shits on best practices, when trying new shit like blueprints for GTK/Libadwaita it keeps generating UI programmatically instead of doing blueprint files.

>>106116077
Man even in 2023 when doing vba in excel I would just tell chatgpt, "i want this thing here to be kinda like this thing here you know what i mean. I need it to kinda do this thing but not exactly like that cause im trying to get it to do this so it's gotta be a bit different"
and it would know what i was trying to say miraculously.

>>106117385
Powershell is just better

>>106107132
The vibecoded output for someone who can't read it doesn't scale.
But it's excellent for one-off throwaway scripts like a lot of stuff people need to do in powershell or python.

gemini just told me my script is almost production ready though

>Excellent! This is looking fantastic. You've addressed the critical crash and incorporated the fallback logic very well. The script is now extremely close to being production-ready.
>I have one critical logic fix and one minor refinement for you. After these, I'm confident this script will be exceptionally robust.

>>106121192
> I'm confident this script will be exceptionally robust.
must have slurped up that shit from some HR roastie document.

>>106107632
wait, ask deepsek from the webpage or just the api?

>>106107572
>admits he was wrong
Extremely based.

>>106121192
I fucking hate how they always say it's a good change or just puff my ego up
No nigger tell me if something is wrong
Rlhf was a mistake

>>106121192
notice how AI can replace stupid jobs perfectly?

>>106107354
the hallucinations are the same as 4 years ago, barely anything has changed

>>106108042
you are an insane person
hope you know this

>>106112450
>it just works
you have no idea if it works or not
you have no idea if it is secure

>>106108965
Algorithmic efficiencies can only go so far. There won't be some new magical algorithm discovered every year that would give 1000% improvements.

>>106113190
Can you explain how AI is getting good at programming?

Welcome to the AI revolution. Only the ones who embrace it will make it in the new incoming world.

>>106122850
Hallucinations have been mitigated in the newer reasoning models

>>106123973
Remember when ChatGPT3.5 was shitting the bed just after 20 lines of code.

Well now the best coding models are able to work on vast parts of large codebases while leveraging MPC. Now it shits the bed after a couple of hours on the codebase

Next generation of models are releasing in the coming months and are likely to perform even better

>>106107314
With rust it just constantly hallucinates traits and gets surprised that they don't exist

>>106124473
Hallucinations have been mitigated by them googling the relevant doc and putting it in the context, and still shit won't work with complex or multiple versions of an API.

Like the other guy said it does not scale, if you could fine-tune your own version with certain documents then maybe but as of right now you can't feed a bunch of documentation and ask something on top, even on gemini 1 million tokens it starts to forget or pays attention to the wrong things and then it's a matter of swipping a thousand times. The coding agent tools waste all your credits by getting stuck or wasting tokens on retarded prompts

>>106124752
It works to an extent and is working far better than previous versions. It has definitely gotten to a point that it has surpassed entry level / junior devs

As someone that works security, networking and tier 3 cluster maintained on Ubuntu. CharGPt helps me write scripts efficiently and to bounce ideas, but to do actual troubleshooting, threat hunting or intricate code? I don’t fucking trust it one bit.
>update pihole at home
>like a dumbass let a specific setting on that was not needed for GUI access
> ask ChatGPT how to fix php
>gives me this stupid long winded answer
>end up doing systemctl disable service

Problem solved in 1 minute.

>>106107314
>C/C++
Maybe it would work better if you could actually which language you want it to produce code for

>>106107314
>reinterpret_cast in functions declared as consteval, attempting to call non-existing functions, etc.
AI can't do much when the language itself is an ungodly mess.

>>106107354
>Feel the AGI
>Feel the AGI